1 // SPDX-License-Identifier: GPL-2.0-only <<
2 /* 1 /*
3 * kernel/workqueue.c - generic async executio 2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * 3 *
5 * Copyright (C) 2002 Ingo Molnar 4 * Copyright (C) 2002 Ingo Molnar
6 * 5 *
7 * Derived from the taskqueue/keventd code b 6 * Derived from the taskqueue/keventd code by:
8 * David Woodhouse <dwmw2@infradead.org> 7 * David Woodhouse <dwmw2@infradead.org>
9 * Andrew Morton 8 * Andrew Morton
10 * Kai Petzke <wpp@marie.physik.tu-berlin. 9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
11 * Theodore Ts'o <tytso@mit.edu> 10 * Theodore Ts'o <tytso@mit.edu>
12 * 11 *
13 * Made to use alloc_percpu by Christoph Lamet 12 * Made to use alloc_percpu by Christoph Lameter.
14 * 13 *
15 * Copyright (C) 2010 SUSE Linux Pro 14 * Copyright (C) 2010 SUSE Linux Products GmbH
16 * Copyright (C) 2010 Tejun Heo <tj@ 15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * 16 *
18 * This is the generic async execution mechani 17 * This is the generic async execution mechanism. Work items as are
19 * executed in process context. The worker po 18 * executed in process context. The worker pool is shared and
20 * automatically managed. There are two worke !! 19 * automatically managed. There is one worker pool for each CPU and
21 * normal work items and the other for high pr !! 20 * one extra for works which are better served by workers which are
22 * pools for workqueues which are not bound to !! 21 * not bound to any specific CPU.
23 * number of these backing pools is dynamic. <<
24 * 22 *
25 * Please read Documentation/core-api/workqueu !! 23 * Please read Documentation/workqueue.txt for details.
26 */ 24 */
27 25
28 #include <linux/export.h> 26 #include <linux/export.h>
29 #include <linux/kernel.h> 27 #include <linux/kernel.h>
30 #include <linux/sched.h> 28 #include <linux/sched.h>
31 #include <linux/init.h> 29 #include <linux/init.h>
32 #include <linux/interrupt.h> <<
33 #include <linux/signal.h> 30 #include <linux/signal.h>
34 #include <linux/completion.h> 31 #include <linux/completion.h>
35 #include <linux/workqueue.h> 32 #include <linux/workqueue.h>
36 #include <linux/slab.h> 33 #include <linux/slab.h>
37 #include <linux/cpu.h> 34 #include <linux/cpu.h>
38 #include <linux/notifier.h> 35 #include <linux/notifier.h>
39 #include <linux/kthread.h> 36 #include <linux/kthread.h>
40 #include <linux/hardirq.h> 37 #include <linux/hardirq.h>
41 #include <linux/mempolicy.h> 38 #include <linux/mempolicy.h>
42 #include <linux/freezer.h> 39 #include <linux/freezer.h>
>> 40 #include <linux/kallsyms.h>
43 #include <linux/debug_locks.h> 41 #include <linux/debug_locks.h>
44 #include <linux/lockdep.h> 42 #include <linux/lockdep.h>
45 #include <linux/idr.h> 43 #include <linux/idr.h>
46 #include <linux/jhash.h> 44 #include <linux/jhash.h>
47 #include <linux/hashtable.h> 45 #include <linux/hashtable.h>
48 #include <linux/rculist.h> 46 #include <linux/rculist.h>
49 #include <linux/nodemask.h> 47 #include <linux/nodemask.h>
50 #include <linux/moduleparam.h> 48 #include <linux/moduleparam.h>
51 #include <linux/uaccess.h> 49 #include <linux/uaccess.h>
52 #include <linux/sched/isolation.h> <<
53 #include <linux/sched/debug.h> <<
54 #include <linux/nmi.h> <<
55 #include <linux/kvm_para.h> <<
56 #include <linux/delay.h> <<
57 #include <linux/irq_work.h> <<
58 50
59 #include "workqueue_internal.h" 51 #include "workqueue_internal.h"
60 52
61 enum worker_pool_flags { !! 53 enum {
62 /* 54 /*
63 * worker_pool flags 55 * worker_pool flags
64 * 56 *
65 * A bound pool is either associated o 57 * A bound pool is either associated or disassociated with its CPU.
66 * While associated (!DISASSOCIATED), 58 * While associated (!DISASSOCIATED), all workers are bound to the
67 * CPU and none has %WORKER_UNBOUND se 59 * CPU and none has %WORKER_UNBOUND set and concurrency management
68 * is in effect. 60 * is in effect.
69 * 61 *
70 * While DISASSOCIATED, the cpu may be 62 * While DISASSOCIATED, the cpu may be offline and all workers have
71 * %WORKER_UNBOUND set and concurrency 63 * %WORKER_UNBOUND set and concurrency management disabled, and may
72 * be executing on any CPU. The pool 64 * be executing on any CPU. The pool behaves as an unbound one.
73 * 65 *
74 * Note that DISASSOCIATED should be f 66 * Note that DISASSOCIATED should be flipped only while holding
75 * wq_pool_attach_mutex to avoid chang !! 67 * manager_mutex to avoid changing binding state while
76 * worker_attach_to_pool() is in progr !! 68 * create_worker() is in progress.
77 * <<
78 * As there can only be one concurrent <<
79 * BH pool is per-CPU and always DISAS <<
80 */ 69 */
81 POOL_BH = 1 << 0, !! 70 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
82 POOL_MANAGER_ACTIVE = 1 << 1, <<
83 POOL_DISASSOCIATED = 1 << 2, 71 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
84 POOL_BH_DRAINING = 1 << 3, !! 72 POOL_FREEZING = 1 << 3, /* freeze in progress */
85 }; <<
86 73
87 enum worker_flags { <<
88 /* worker flags */ 74 /* worker flags */
>> 75 WORKER_STARTED = 1 << 0, /* started */
89 WORKER_DIE = 1 << 1, 76 WORKER_DIE = 1 << 1, /* die die die */
90 WORKER_IDLE = 1 << 2, 77 WORKER_IDLE = 1 << 2, /* is idle */
91 WORKER_PREP = 1 << 3, 78 WORKER_PREP = 1 << 3, /* preparing to run works */
92 WORKER_CPU_INTENSIVE = 1 << 6, 79 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
93 WORKER_UNBOUND = 1 << 7, 80 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
94 WORKER_REBOUND = 1 << 8, 81 WORKER_REBOUND = 1 << 8, /* worker was rebound */
95 82
96 WORKER_NOT_RUNNING = WORKER_PREP 83 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
97 WORKER_UNBOU 84 WORKER_UNBOUND | WORKER_REBOUND,
98 }; <<
99 85
100 enum work_cancel_flags { <<
101 WORK_CANCEL_DELAYED = 1 << 0, <<
102 WORK_CANCEL_DISABLE = 1 << 1, <<
103 }; <<
104 <<
105 enum wq_internal_consts { <<
106 NR_STD_WORKER_POOLS = 2, 86 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
107 87
108 UNBOUND_POOL_HASH_ORDER = 6, 88 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
109 BUSY_WORKER_HASH_ORDER = 6, 89 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
110 90
111 MAX_IDLE_WORKERS_RATIO = 4, 91 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
112 IDLE_WORKER_TIMEOUT = 300 * HZ, 92 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
113 93
114 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 94 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
115 95 /* call for help after 10ms
116 96 (min two ticks) */
117 MAYDAY_INTERVAL = HZ / 10, 97 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
118 CREATE_COOLDOWN = HZ, 98 CREATE_COOLDOWN = HZ, /* time to breath after fail */
119 99
120 /* 100 /*
121 * Rescue workers are used only on eme 101 * Rescue workers are used only on emergencies and shared by
122 * all cpus. Give MIN_NICE. !! 102 * all cpus. Give -20.
123 */ 103 */
124 RESCUER_NICE_LEVEL = MIN_NICE, !! 104 RESCUER_NICE_LEVEL = -20,
125 HIGHPRI_NICE_LEVEL = MIN_NICE, !! 105 HIGHPRI_NICE_LEVEL = -20,
126 106
127 WQ_NAME_LEN = 32, !! 107 WQ_NAME_LEN = 24,
128 WORKER_ID_LEN = 10 + WQ_NAME <<
129 }; 108 };
130 109
131 /* 110 /*
132 * We don't want to trap softirq for too long. <<
133 * MAX_SOFTIRQ_RESTART in kernel/softirq.c. Th <<
134 * msecs_to_jiffies() can't be an initializer. <<
135 */ <<
136 #define BH_WORKER_JIFFIES msecs_to_jiffi <<
137 #define BH_WORKER_RESTARTS 10 <<
138 <<
139 /* <<
140 * Structure fields follow one of the followin 111 * Structure fields follow one of the following exclusion rules.
141 * 112 *
142 * I: Modifiable by initialization/destruction 113 * I: Modifiable by initialization/destruction paths and read-only for
143 * everyone else. 114 * everyone else.
144 * 115 *
145 * P: Preemption protected. Disabling preempt 116 * P: Preemption protected. Disabling preemption is enough and should
146 * only be modified and accessed from the l 117 * only be modified and accessed from the local cpu.
147 * 118 *
148 * L: pool->lock protected. Access with pool- 119 * L: pool->lock protected. Access with pool->lock held.
149 * 120 *
150 * LN: pool->lock and wq_node_nr_active->lock !! 121 * X: During normal operation, modification requires pool->lock and should
151 * reads. !! 122 * be done only from local cpu. Either disabling preemption on local
152 * !! 123 * cpu or grabbing pool->lock is enough for read access. If
153 * K: Only modified by worker while holding po !! 124 * POOL_DISASSOCIATED is set, it's identical to L.
154 * self, while holding pool->lock or from I <<
155 * kworker. <<
156 * 125 *
157 * S: Only modified by worker self. !! 126 * MG: pool->manager_mutex and pool->lock protected. Writes require both
158 * !! 127 * locks. Reads can happen under either lock.
159 * A: wq_pool_attach_mutex protected. <<
160 * 128 *
161 * PL: wq_pool_mutex protected. 129 * PL: wq_pool_mutex protected.
162 * 130 *
163 * PR: wq_pool_mutex protected for writes. RC !! 131 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
164 * <<
165 * PW: wq_pool_mutex and wq->mutex protected f <<
166 * <<
167 * PWR: wq_pool_mutex and wq->mutex protected <<
168 * RCU for reads. <<
169 * 132 *
170 * WQ: wq->mutex protected. 133 * WQ: wq->mutex protected.
171 * 134 *
172 * WR: wq->mutex protected for writes. RCU pr !! 135 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
173 * <<
174 * WO: wq->mutex protected for writes. Updated <<
175 * with READ_ONCE() without locking. <<
176 * 136 *
177 * MD: wq_mayday_lock protected. 137 * MD: wq_mayday_lock protected.
178 * <<
179 * WD: Used internally by the watchdog. <<
180 */ 138 */
181 139
182 /* struct worker is defined in workqueue_inter 140 /* struct worker is defined in workqueue_internal.h */
183 141
184 struct worker_pool { 142 struct worker_pool {
185 raw_spinlock_t lock; !! 143 spinlock_t lock; /* the pool lock */
186 int cpu; 144 int cpu; /* I: the associated cpu */
187 int node; 145 int node; /* I: the associated node ID */
188 int id; 146 int id; /* I: pool ID */
189 unsigned int flags; !! 147 unsigned int flags; /* X: flags */
190 <<
191 unsigned long watchdog_ts; <<
192 bool cpu_stall; <<
193 <<
194 /* <<
195 * The counter is incremented in a pro <<
196 * w/ preemption disabled, and decreme <<
197 * but w/ pool->lock held. The readers <<
198 * guaranteed to see if the counter re <<
199 */ <<
200 int nr_running; <<
201 148
202 struct list_head worklist; 149 struct list_head worklist; /* L: list of pending works */
203 <<
204 int nr_workers; 150 int nr_workers; /* L: total number of workers */
205 int nr_idle; <<
206 151
207 struct list_head idle_list; !! 152 /* nr_idle includes the ones off idle_list for rebinding */
208 struct timer_list idle_timer; !! 153 int nr_idle; /* L: currently idle ones */
209 struct work_struct idle_cull_work <<
210 154
211 struct timer_list mayday_timer; !! 155 struct list_head idle_list; /* X: list of idle workers */
>> 156 struct timer_list idle_timer; /* L: worker idle timeout */
>> 157 struct timer_list mayday_timer; /* L: SOS timer for workers */
212 158
213 /* a workers is either on busy_hash or 159 /* a workers is either on busy_hash or idle_list, or the manager */
214 DECLARE_HASHTABLE(busy_hash, BUSY_WORK 160 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
215 161 /* L: hash of busy workers */
216 162
217 struct worker *manager; !! 163 /* see manage_workers() for details on the two manager mutexes */
218 struct list_head workers; !! 164 struct mutex manager_arb; /* manager arbitration */
219 !! 165 struct mutex manager_mutex; /* manager exclusion */
220 struct ida worker_ida; !! 166 struct idr worker_idr; /* MG: worker IDs and iteration */
221 167
222 struct workqueue_attrs *attrs; 168 struct workqueue_attrs *attrs; /* I: worker attributes */
223 struct hlist_node hash_node; 169 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
224 int refcnt; 170 int refcnt; /* PL: refcnt for unbound pools */
225 171
226 /* 172 /*
227 * Destruction of pool is RCU protecte !! 173 * The current concurrency level. As it's likely to be accessed
>> 174 * from other CPUs during try_to_wake_up(), put it in a separate
>> 175 * cacheline.
>> 176 */
>> 177 atomic_t nr_running ____cacheline_aligned_in_smp;
>> 178
>> 179 /*
>> 180 * Destruction of pool is sched-RCU protected to allow dereferences
228 * from get_work_pool(). 181 * from get_work_pool().
229 */ 182 */
230 struct rcu_head rcu; 183 struct rcu_head rcu;
231 }; !! 184 } ____cacheline_aligned_in_smp;
232 185
233 /* 186 /*
234 * Per-pool_workqueue statistics. These can be !! 187 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
235 * tools/workqueue/wq_monitor.py. <<
236 */ <<
237 enum pool_workqueue_stats { <<
238 PWQ_STAT_STARTED, /* work items <<
239 PWQ_STAT_COMPLETED, /* work items <<
240 PWQ_STAT_CPU_TIME, /* total CPU t <<
241 PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_inte <<
242 PWQ_STAT_CM_WAKEUP, /* concurrency <<
243 PWQ_STAT_REPATRIATED, /* unbound wor <<
244 PWQ_STAT_MAYDAY, /* maydays to <<
245 PWQ_STAT_RESCUED, /* linked work <<
246 <<
247 PWQ_NR_STATS, <<
248 }; <<
249 <<
250 /* <<
251 * The per-pool workqueue. While queued, bits <<
252 * of work_struct->data are used for flags and 188 * of work_struct->data are used for flags and the remaining high bits
253 * point to the pwq; thus, pwqs need to be ali 189 * point to the pwq; thus, pwqs need to be aligned at two's power of the
254 * number of flag bits. 190 * number of flag bits.
255 */ 191 */
256 struct pool_workqueue { 192 struct pool_workqueue {
257 struct worker_pool *pool; 193 struct worker_pool *pool; /* I: the associated pool */
258 struct workqueue_struct *wq; 194 struct workqueue_struct *wq; /* I: the owning workqueue */
259 int work_color; 195 int work_color; /* L: current color */
260 int flush_color; 196 int flush_color; /* L: flushing color */
261 int refcnt; 197 int refcnt; /* L: reference count */
262 int nr_in_flight[W 198 int nr_in_flight[WORK_NR_COLORS];
263 199 /* L: nr of in_flight works */
264 bool plugged; <<
265 <<
266 /* <<
267 * nr_active management and WORK_STRUC <<
268 * <<
269 * When pwq->nr_active >= max_active, <<
270 * pwq->inactive_works instead of pool <<
271 * WORK_STRUCT_INACTIVE. <<
272 * <<
273 * All work items marked with WORK_STR <<
274 * nr_active and all work items in pwq <<
275 * WORK_STRUCT_INACTIVE. But not all W <<
276 * in pwq->inactive_works. Some of the <<
277 * pool->worklist or worker->scheduled <<
278 * wq_barrier which is used for flush_ <<
279 * in nr_active. For non-barrier work <<
280 * WORK_STRUCT_INACTIVE iff it is in p <<
281 */ <<
282 int nr_active; 200 int nr_active; /* L: nr of active works */
283 struct list_head inactive_works !! 201 int max_active; /* L: max active works */
284 struct list_head pending_node; !! 202 struct list_head delayed_works; /* L: delayed works */
285 struct list_head pwqs_node; 203 struct list_head pwqs_node; /* WR: node on wq->pwqs */
286 struct list_head mayday_node; 204 struct list_head mayday_node; /* MD: node on wq->maydays */
287 205
288 u64 stats[PWQ_NR_S <<
289 <<
290 /* 206 /*
291 * Release of unbound pwq is punted to !! 207 * Release of unbound pwq is punted to system_wq. See put_pwq()
292 * and pwq_release_workfn() for detail !! 208 * and pwq_unbound_release_workfn() for details. pool_workqueue
293 * RCU protected so that the first pwq !! 209 * itself is also sched-RCU protected so that the first pwq can be
294 * grabbing wq->mutex. !! 210 * determined without grabbing wq->mutex.
295 */ 211 */
296 struct kthread_work release_work; !! 212 struct work_struct unbound_release_work;
297 struct rcu_head rcu; 213 struct rcu_head rcu;
298 } __aligned(1 << WORK_STRUCT_PWQ_SHIFT); !! 214 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
299 215
300 /* 216 /*
301 * Structure used to wait for workqueue flush. 217 * Structure used to wait for workqueue flush.
302 */ 218 */
303 struct wq_flusher { 219 struct wq_flusher {
304 struct list_head list; 220 struct list_head list; /* WQ: list of flushers */
305 int flush_color; 221 int flush_color; /* WQ: flush color waiting for */
306 struct completion done; 222 struct completion done; /* flush completion */
307 }; 223 };
308 224
309 struct wq_device; 225 struct wq_device;
310 226
311 /* 227 /*
312 * Unlike in a per-cpu workqueue where max_act <<
313 * on each CPU, in an unbound workqueue, max_a <<
314 * As sharing a single nr_active across multip <<
315 * the counting and enforcement is per NUMA no <<
316 * <<
317 * The following struct is used to enforce per <<
318 * to start executing a work item, it should i <<
319 * tryinc_node_nr_active(). If acquisition fai <<
320 * ->max, the pwq is queued on ->pending_pwqs. <<
321 * and decrement ->nr, node_activate_pending_p <<
322 * round-robin order. <<
323 */ <<
324 struct wq_node_nr_active { <<
325 int max; <<
326 atomic_t nr; <<
327 raw_spinlock_t lock; <<
328 struct list_head pending_pwqs; <<
329 }; <<
330 <<
331 /* <<
332 * The externally visible workqueue. It relay 228 * The externally visible workqueue. It relays the issued work items to
333 * the appropriate worker_pool through its poo 229 * the appropriate worker_pool through its pool_workqueues.
334 */ 230 */
335 struct workqueue_struct { 231 struct workqueue_struct {
336 struct list_head pwqs; 232 struct list_head pwqs; /* WR: all pwqs of this wq */
337 struct list_head list; !! 233 struct list_head list; /* PL: list of all workqueues */
338 234
339 struct mutex mutex; 235 struct mutex mutex; /* protects this wq */
340 int work_color; 236 int work_color; /* WQ: current work color */
341 int flush_color; 237 int flush_color; /* WQ: current flush color */
342 atomic_t nr_pwqs_to_flu 238 atomic_t nr_pwqs_to_flush; /* flush in progress */
343 struct wq_flusher *first_flusher 239 struct wq_flusher *first_flusher; /* WQ: first flusher */
344 struct list_head flusher_queue; 240 struct list_head flusher_queue; /* WQ: flush waiters */
345 struct list_head flusher_overfl 241 struct list_head flusher_overflow; /* WQ: flush overflow list */
346 242
347 struct list_head maydays; 243 struct list_head maydays; /* MD: pwqs requesting rescue */
348 struct worker *rescuer; !! 244 struct worker *rescuer; /* I: rescue worker */
349 245
350 int nr_drainers; 246 int nr_drainers; /* WQ: drain in progress */
>> 247 int saved_max_active; /* WQ: saved pwq max_active */
351 248
352 /* See alloc_workqueue() function comm !! 249 struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */
353 int max_active; !! 250 struct pool_workqueue *dfl_pwq; /* WQ: only for unbound wqs */
354 int min_active; <<
355 int saved_max_acti <<
356 int saved_min_acti <<
357 <<
358 struct workqueue_attrs *unbound_attrs <<
359 struct pool_workqueue __rcu *dfl_pwq; <<
360 251
361 #ifdef CONFIG_SYSFS 252 #ifdef CONFIG_SYSFS
362 struct wq_device *wq_dev; 253 struct wq_device *wq_dev; /* I: for sysfs interface */
363 #endif 254 #endif
364 #ifdef CONFIG_LOCKDEP 255 #ifdef CONFIG_LOCKDEP
365 char *lock_name; <<
366 struct lock_class_key key; <<
367 struct lockdep_map lockdep_map; 256 struct lockdep_map lockdep_map;
368 #endif 257 #endif
369 char name[WQ_NAME_L 258 char name[WQ_NAME_LEN]; /* I: workqueue name */
370 259
371 /* <<
372 * Destruction of workqueue_struct is <<
373 * the workqueues list without grabbin <<
374 * This is used to dump all workqueues <<
375 */ <<
376 struct rcu_head rcu; <<
377 <<
378 /* hot fields used during command issu 260 /* hot fields used during command issue, aligned to cacheline */
379 unsigned int flags ____cach 261 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
380 struct pool_workqueue __rcu * __percpu !! 262 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
381 struct wq_node_nr_active *node_nr_acti !! 263 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */
382 }; 264 };
383 265
384 /* !! 266 static struct kmem_cache *pwq_cache;
385 * Each pod type describes how CPUs should be <<
386 * See the comment above workqueue_attrs->affn <<
387 */ <<
388 struct wq_pod_type { <<
389 int nr_pods; <<
390 cpumask_var_t *pod_cpus; <<
391 int *pod_node; <<
392 int *cpu_pod; <<
393 }; <<
394 <<
395 struct work_offq_data { <<
396 u32 pool_id; <<
397 u32 disable; <<
398 u32 flags; <<
399 }; <<
400 <<
401 static const char *wq_affn_names[WQ_AFFN_NR_TY <<
402 [WQ_AFFN_DFL] = "default", <<
403 [WQ_AFFN_CPU] = "cpu", <<
404 [WQ_AFFN_SMT] = "smt", <<
405 [WQ_AFFN_CACHE] = "cache", <<
406 [WQ_AFFN_NUMA] = "numa", <<
407 [WQ_AFFN_SYSTEM] = "system", <<
408 }; <<
409 <<
410 /* <<
411 * Per-cpu work items which run for longer tha <<
412 * automatically considered CPU intensive and <<
413 * management to prevent them from noticeably <<
414 * ULONG_MAX indicates that the user hasn't ov <<
415 * The actual value is initialized in wq_cpu_i <<
416 */ <<
417 static unsigned long wq_cpu_intensive_thresh_u <<
418 module_param_named(cpu_intensive_thresh_us, wq <<
419 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT <<
420 static unsigned int wq_cpu_intensive_warning_t <<
421 module_param_named(cpu_intensive_warning_thres <<
422 #endif <<
423 <<
424 /* see the comment above the definition of WQ_ <<
425 static bool wq_power_efficient = IS_ENABLED(CO <<
426 module_param_named(power_efficient, wq_power_e <<
427 267
428 static bool wq_online; /* can !! 268 static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */
429 static bool wq_topo_initialized __read_mostly !! 269 static cpumask_var_t *wq_numa_possible_cpumask;
>> 270 /* possible CPUs of each node */
430 271
431 static struct kmem_cache *pwq_cache; !! 272 static bool wq_disable_numa;
>> 273 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
432 274
433 static struct wq_pod_type wq_pod_types[WQ_AFFN !! 275 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
434 static enum wq_affn_scope wq_affn_dfl = WQ_AFF <<
435 276
436 /* buf for wq_update_unbound_pod_attrs(), prot !! 277 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
437 static struct workqueue_attrs *unbound_wq_upda !! 278 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
438 279
439 static DEFINE_MUTEX(wq_pool_mutex); /* pro 280 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
440 static DEFINE_MUTEX(wq_pool_attach_mutex); /* !! 281 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
441 static DEFINE_RAW_SPINLOCK(wq_mayday_lock); <<
442 /* wait for manager to go away */ <<
443 static struct rcuwait manager_wait = __RCUWAIT <<
444 282
445 static LIST_HEAD(workqueues); /* PR: !! 283 static LIST_HEAD(workqueues); /* PL: list of all workqueues */
446 static bool workqueue_freezing; /* PL: 284 static bool workqueue_freezing; /* PL: have wqs started freezing? */
447 285
448 /* PL: mirror the cpu_online_mask excluding th <<
449 static cpumask_var_t wq_online_cpumask; <<
450 <<
451 /* PL&A: allowable cpus for unbound wqs and wo <<
452 static cpumask_var_t wq_unbound_cpumask; <<
453 <<
454 /* PL: user requested unbound cpumask via sysf <<
455 static cpumask_var_t wq_requested_unbound_cpum <<
456 <<
457 /* PL: isolated cpumask to be excluded from un <<
458 static cpumask_var_t wq_isolated_cpumask; <<
459 <<
460 /* for further constrain wq_unbound_cpumask by <<
461 static struct cpumask wq_cmdline_cpumask __ini <<
462 <<
463 /* CPU where unbound work was last round robin <<
464 static DEFINE_PER_CPU(int, wq_rr_cpu_last); <<
465 <<
466 /* <<
467 * Local execution of unbound work items is no <<
468 * following always forces round-robin CPU sel <<
469 * to uncover usages which depend on it. <<
470 */ <<
471 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU <<
472 static bool wq_debug_force_rr_cpu = true; <<
473 #else <<
474 static bool wq_debug_force_rr_cpu = false; <<
475 #endif <<
476 module_param_named(debug_force_rr_cpu, wq_debu <<
477 <<
478 /* to raise softirq for the BH worker pools on <<
479 static DEFINE_PER_CPU_SHARED_ALIGNED(struct ir <<
480 bh_pool_i <<
481 <<
482 /* the BH worker pools */ <<
483 static DEFINE_PER_CPU_SHARED_ALIGNED(struct wo <<
484 bh_worker <<
485 <<
486 /* the per-cpu worker pools */ 286 /* the per-cpu worker pools */
487 static DEFINE_PER_CPU_SHARED_ALIGNED(struct wo 287 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
488 cpu_worke 288 cpu_worker_pools);
489 289
490 static DEFINE_IDR(worker_pool_idr); /* PR: 290 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
491 291
492 /* PL: hash of all unbound pools keyed by pool 292 /* PL: hash of all unbound pools keyed by pool->attrs */
493 static DEFINE_HASHTABLE(unbound_pool_hash, UNB 293 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
494 294
495 /* I: attributes used when instantiating stand 295 /* I: attributes used when instantiating standard unbound pools on demand */
496 static struct workqueue_attrs *unbound_std_wq_ 296 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
497 297
498 /* I: attributes used when instantiating order 298 /* I: attributes used when instantiating ordered pools on demand */
499 static struct workqueue_attrs *ordered_wq_attr 299 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
500 300
501 /* !! 301 struct workqueue_struct *system_wq __read_mostly;
502 * I: kthread_worker to release pwq's. pwq rel <<
503 * process context while holding a pool lock. <<
504 * worker to avoid A-A deadlocks. <<
505 */ <<
506 static struct kthread_worker *pwq_release_work <<
507 <<
508 struct workqueue_struct *system_wq __ro_after_ <<
509 EXPORT_SYMBOL(system_wq); 302 EXPORT_SYMBOL(system_wq);
510 struct workqueue_struct *system_highpri_wq __r !! 303 struct workqueue_struct *system_highpri_wq __read_mostly;
511 EXPORT_SYMBOL_GPL(system_highpri_wq); 304 EXPORT_SYMBOL_GPL(system_highpri_wq);
512 struct workqueue_struct *system_long_wq __ro_a !! 305 struct workqueue_struct *system_long_wq __read_mostly;
513 EXPORT_SYMBOL_GPL(system_long_wq); 306 EXPORT_SYMBOL_GPL(system_long_wq);
514 struct workqueue_struct *system_unbound_wq __r !! 307 struct workqueue_struct *system_unbound_wq __read_mostly;
515 EXPORT_SYMBOL_GPL(system_unbound_wq); 308 EXPORT_SYMBOL_GPL(system_unbound_wq);
516 struct workqueue_struct *system_freezable_wq _ !! 309 struct workqueue_struct *system_freezable_wq __read_mostly;
517 EXPORT_SYMBOL_GPL(system_freezable_wq); 310 EXPORT_SYMBOL_GPL(system_freezable_wq);
518 struct workqueue_struct *system_power_efficien <<
519 EXPORT_SYMBOL_GPL(system_power_efficient_wq); <<
520 struct workqueue_struct *system_freezable_powe <<
521 EXPORT_SYMBOL_GPL(system_freezable_power_effic <<
522 struct workqueue_struct *system_bh_wq; <<
523 EXPORT_SYMBOL_GPL(system_bh_wq); <<
524 struct workqueue_struct *system_bh_highpri_wq; <<
525 EXPORT_SYMBOL_GPL(system_bh_highpri_wq); <<
526 311
527 static int worker_thread(void *__worker); 312 static int worker_thread(void *__worker);
528 static void workqueue_sysfs_unregister(struct !! 313 static void copy_workqueue_attrs(struct workqueue_attrs *to,
529 static void show_pwq(struct pool_workqueue *pw !! 314 const struct workqueue_attrs *from);
530 static void show_one_worker_pool(struct worker <<
531 315
532 #define CREATE_TRACE_POINTS 316 #define CREATE_TRACE_POINTS
533 #include <trace/events/workqueue.h> 317 #include <trace/events/workqueue.h>
534 318
535 #define assert_rcu_or_pool_mutex() 319 #define assert_rcu_or_pool_mutex() \
536 RCU_LOCKDEP_WARN(!rcu_read_lock_any_he !! 320 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
537 !lockdep_is_held(&wq_ !! 321 lockdep_is_held(&wq_pool_mutex), \
538 "RCU or wq_pool_mutex !! 322 "sched RCU or wq_pool_mutex should be held")
539 !! 323
540 #define assert_rcu_or_wq_mutex_or_pool_mutex(w !! 324 #define assert_rcu_or_wq_mutex(wq) \
541 RCU_LOCKDEP_WARN(!rcu_read_lock_any_he !! 325 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
542 !lockdep_is_held(&wq- !! 326 lockdep_is_held(&wq->mutex), \
543 !lockdep_is_held(&wq_ !! 327 "sched RCU or wq->mutex should be held")
544 "RCU, wq->mutex or wq !! 328
545 !! 329 #ifdef CONFIG_LOCKDEP
546 #define for_each_bh_worker_pool(pool, cpu) !! 330 #define assert_manager_or_pool_lock(pool) \
547 for ((pool) = &per_cpu(bh_worker_pools !! 331 WARN_ONCE(debug_locks && \
548 (pool) < &per_cpu(bh_worker_pools !! 332 !lockdep_is_held(&(pool)->manager_mutex) && \
549 (pool)++) !! 333 !lockdep_is_held(&(pool)->lock), \
>> 334 "pool->manager_mutex or ->lock should be held")
>> 335 #else
>> 336 #define assert_manager_or_pool_lock(pool) do { } while (0)
>> 337 #endif
550 338
551 #define for_each_cpu_worker_pool(pool, cpu) 339 #define for_each_cpu_worker_pool(pool, cpu) \
552 for ((pool) = &per_cpu(cpu_worker_pool 340 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
553 (pool) < &per_cpu(cpu_worker_pool 341 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
554 (pool)++) 342 (pool)++)
555 343
556 /** 344 /**
557 * for_each_pool - iterate through all worker_ 345 * for_each_pool - iterate through all worker_pools in the system
558 * @pool: iteration cursor 346 * @pool: iteration cursor
559 * @pi: integer used for iteration 347 * @pi: integer used for iteration
560 * 348 *
561 * This must be called either with wq_pool_mut !! 349 * This must be called either with wq_pool_mutex held or sched RCU read
562 * locked. If the pool needs to be used beyon 350 * locked. If the pool needs to be used beyond the locking in effect, the
563 * caller is responsible for guaranteeing that 351 * caller is responsible for guaranteeing that the pool stays online.
564 * 352 *
565 * The if/else clause exists only for the lock 353 * The if/else clause exists only for the lockdep assertion and can be
566 * ignored. 354 * ignored.
567 */ 355 */
568 #define for_each_pool(pool, pi) 356 #define for_each_pool(pool, pi) \
569 idr_for_each_entry(&worker_pool_idr, p 357 idr_for_each_entry(&worker_pool_idr, pool, pi) \
570 if (({ assert_rcu_or_pool_mute 358 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
571 else 359 else
572 360
573 /** 361 /**
574 * for_each_pool_worker - iterate through all 362 * for_each_pool_worker - iterate through all workers of a worker_pool
575 * @worker: iteration cursor 363 * @worker: iteration cursor
>> 364 * @wi: integer used for iteration
576 * @pool: worker_pool to iterate workers of 365 * @pool: worker_pool to iterate workers of
577 * 366 *
578 * This must be called with wq_pool_attach_mut !! 367 * This must be called with either @pool->manager_mutex or ->lock held.
579 * 368 *
580 * The if/else clause exists only for the lock 369 * The if/else clause exists only for the lockdep assertion and can be
581 * ignored. 370 * ignored.
582 */ 371 */
583 #define for_each_pool_worker(worker, pool) !! 372 #define for_each_pool_worker(worker, wi, pool) \
584 list_for_each_entry((worker), &(pool)- !! 373 idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \
585 if (({ lockdep_assert_held(&wq !! 374 if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
586 else 375 else
587 376
588 /** 377 /**
589 * for_each_pwq - iterate through all pool_wor 378 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
590 * @pwq: iteration cursor 379 * @pwq: iteration cursor
591 * @wq: the target workqueue 380 * @wq: the target workqueue
592 * 381 *
593 * This must be called either with wq->mutex h !! 382 * This must be called either with wq->mutex held or sched RCU read locked.
594 * If the pwq needs to be used beyond the lock 383 * If the pwq needs to be used beyond the locking in effect, the caller is
595 * responsible for guaranteeing that the pwq s 384 * responsible for guaranteeing that the pwq stays online.
596 * 385 *
597 * The if/else clause exists only for the lock 386 * The if/else clause exists only for the lockdep assertion and can be
598 * ignored. 387 * ignored.
599 */ 388 */
600 #define for_each_pwq(pwq, wq) 389 #define for_each_pwq(pwq, wq) \
601 list_for_each_entry_rcu((pwq), &(wq)-> !! 390 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
602 lockdep_is_he !! 391 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
>> 392 else
603 393
604 #ifdef CONFIG_DEBUG_OBJECTS_WORK 394 #ifdef CONFIG_DEBUG_OBJECTS_WORK
605 395
606 static const struct debug_obj_descr work_debug !! 396 static struct debug_obj_descr work_debug_descr;
607 397
608 static void *work_debug_hint(void *addr) 398 static void *work_debug_hint(void *addr)
609 { 399 {
610 return ((struct work_struct *) addr)-> 400 return ((struct work_struct *) addr)->func;
611 } 401 }
612 402
613 static bool work_is_static_object(void *addr) !! 403 /*
>> 404 * fixup_init is called when:
>> 405 * - an active object is initialized
>> 406 */
>> 407 static int work_fixup_init(void *addr, enum debug_obj_state state)
614 { 408 {
615 struct work_struct *work = addr; 409 struct work_struct *work = addr;
616 410
617 return test_bit(WORK_STRUCT_STATIC_BIT !! 411 switch (state) {
>> 412 case ODEBUG_STATE_ACTIVE:
>> 413 cancel_work_sync(work);
>> 414 debug_object_init(work, &work_debug_descr);
>> 415 return 1;
>> 416 default:
>> 417 return 0;
>> 418 }
618 } 419 }
619 420
620 /* 421 /*
621 * fixup_init is called when: !! 422 * fixup_activate is called when:
622 * - an active object is initialized !! 423 * - an active object is activated
>> 424 * - an unknown object is activated (might be a statically initialized object)
623 */ 425 */
624 static bool work_fixup_init(void *addr, enum d !! 426 static int work_fixup_activate(void *addr, enum debug_obj_state state)
625 { 427 {
626 struct work_struct *work = addr; 428 struct work_struct *work = addr;
627 429
628 switch (state) { 430 switch (state) {
>> 431
>> 432 case ODEBUG_STATE_NOTAVAILABLE:
>> 433 /*
>> 434 * This is not really a fixup. The work struct was
>> 435 * statically initialized. We just make sure that it
>> 436 * is tracked in the object tracker.
>> 437 */
>> 438 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
>> 439 debug_object_init(work, &work_debug_descr);
>> 440 debug_object_activate(work, &work_debug_descr);
>> 441 return 0;
>> 442 }
>> 443 WARN_ON_ONCE(1);
>> 444 return 0;
>> 445
629 case ODEBUG_STATE_ACTIVE: 446 case ODEBUG_STATE_ACTIVE:
630 cancel_work_sync(work); !! 447 WARN_ON(1);
631 debug_object_init(work, &work_ !! 448
632 return true; <<
633 default: 449 default:
634 return false; !! 450 return 0;
635 } 451 }
636 } 452 }
637 453
638 /* 454 /*
639 * fixup_free is called when: 455 * fixup_free is called when:
640 * - an active object is freed 456 * - an active object is freed
641 */ 457 */
642 static bool work_fixup_free(void *addr, enum d !! 458 static int work_fixup_free(void *addr, enum debug_obj_state state)
643 { 459 {
644 struct work_struct *work = addr; 460 struct work_struct *work = addr;
645 461
646 switch (state) { 462 switch (state) {
647 case ODEBUG_STATE_ACTIVE: 463 case ODEBUG_STATE_ACTIVE:
648 cancel_work_sync(work); 464 cancel_work_sync(work);
649 debug_object_free(work, &work_ 465 debug_object_free(work, &work_debug_descr);
650 return true; !! 466 return 1;
651 default: 467 default:
652 return false; !! 468 return 0;
653 } 469 }
654 } 470 }
655 471
656 static const struct debug_obj_descr work_debug !! 472 static struct debug_obj_descr work_debug_descr = {
657 .name = "work_struct", 473 .name = "work_struct",
658 .debug_hint = work_debug_hint, 474 .debug_hint = work_debug_hint,
659 .is_static_object = work_is_static_obj <<
660 .fixup_init = work_fixup_init, 475 .fixup_init = work_fixup_init,
>> 476 .fixup_activate = work_fixup_activate,
661 .fixup_free = work_fixup_free, 477 .fixup_free = work_fixup_free,
662 }; 478 };
663 479
664 static inline void debug_work_activate(struct 480 static inline void debug_work_activate(struct work_struct *work)
665 { 481 {
666 debug_object_activate(work, &work_debu 482 debug_object_activate(work, &work_debug_descr);
667 } 483 }
668 484
669 static inline void debug_work_deactivate(struc 485 static inline void debug_work_deactivate(struct work_struct *work)
670 { 486 {
671 debug_object_deactivate(work, &work_de 487 debug_object_deactivate(work, &work_debug_descr);
672 } 488 }
673 489
674 void __init_work(struct work_struct *work, int 490 void __init_work(struct work_struct *work, int onstack)
675 { 491 {
676 if (onstack) 492 if (onstack)
677 debug_object_init_on_stack(wor 493 debug_object_init_on_stack(work, &work_debug_descr);
678 else 494 else
679 debug_object_init(work, &work_ 495 debug_object_init(work, &work_debug_descr);
680 } 496 }
681 EXPORT_SYMBOL_GPL(__init_work); 497 EXPORT_SYMBOL_GPL(__init_work);
682 498
683 void destroy_work_on_stack(struct work_struct 499 void destroy_work_on_stack(struct work_struct *work)
684 { 500 {
685 debug_object_free(work, &work_debug_de 501 debug_object_free(work, &work_debug_descr);
686 } 502 }
687 EXPORT_SYMBOL_GPL(destroy_work_on_stack); 503 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
688 504
689 void destroy_delayed_work_on_stack(struct dela <<
690 { <<
691 destroy_timer_on_stack(&work->timer); <<
692 debug_object_free(&work->work, &work_d <<
693 } <<
694 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stac <<
695 <<
696 #else 505 #else
697 static inline void debug_work_activate(struct 506 static inline void debug_work_activate(struct work_struct *work) { }
698 static inline void debug_work_deactivate(struc 507 static inline void debug_work_deactivate(struct work_struct *work) { }
699 #endif 508 #endif
700 509
701 /** !! 510 /* allocate ID and assign it to @pool */
702 * worker_pool_assign_id - allocate ID and ass <<
703 * @pool: the pool pointer of interest <<
704 * <<
705 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) <<
706 * successfully, -errno on failure. <<
707 */ <<
708 static int worker_pool_assign_id(struct worker 511 static int worker_pool_assign_id(struct worker_pool *pool)
709 { 512 {
710 int ret; 513 int ret;
711 514
712 lockdep_assert_held(&wq_pool_mutex); 515 lockdep_assert_held(&wq_pool_mutex);
713 516
714 ret = idr_alloc(&worker_pool_idr, pool !! 517 ret = idr_alloc(&worker_pool_idr, pool, 0, 0, GFP_KERNEL);
715 GFP_KERNEL); <<
716 if (ret >= 0) { 518 if (ret >= 0) {
717 pool->id = ret; 519 pool->id = ret;
718 return 0; 520 return 0;
719 } 521 }
720 return ret; 522 return ret;
721 } 523 }
722 524
723 static struct pool_workqueue __rcu ** <<
724 unbound_pwq_slot(struct workqueue_struct *wq, <<
725 { <<
726 if (cpu >= 0) <<
727 return per_cpu_ptr(wq->cpu_pwq, <<
728 else <<
729 return &wq->dfl_pwq; <<
730 } <<
731 <<
732 /* @cpu < 0 for dfl_pwq */ <<
733 static struct pool_workqueue *unbound_pwq(stru <<
734 { <<
735 return rcu_dereference_check(*unbound_ <<
736 lockdep_i <<
737 lockdep_i <<
738 } <<
739 <<
740 /** 525 /**
741 * unbound_effective_cpumask - effective cpuma !! 526 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
742 * @wq: workqueue of interest !! 527 * @wq: the target workqueue
>> 528 * @node: the node ID
743 * 529 *
744 * @wq->unbound_attrs->cpumask contains the cp !! 530 * This must be called either with pwq_lock held or sched RCU read locked.
745 * is masked with wq_unbound_cpumask to determ !! 531 * If the pwq needs to be used beyond the locking in effect, the caller is
746 * default pwq is always mapped to the pool wi !! 532 * responsible for guaranteeing that the pwq stays online.
747 */ 533 */
748 static struct cpumask *unbound_effective_cpuma !! 534 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
>> 535 int node)
749 { 536 {
750 return unbound_pwq(wq, -1)->pool->attr !! 537 assert_rcu_or_wq_mutex(wq);
>> 538 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
751 } 539 }
752 540
753 static unsigned int work_color_to_flags(int co 541 static unsigned int work_color_to_flags(int color)
754 { 542 {
755 return color << WORK_STRUCT_COLOR_SHIF 543 return color << WORK_STRUCT_COLOR_SHIFT;
756 } 544 }
757 545
758 static int get_work_color(unsigned long work_d !! 546 static int get_work_color(struct work_struct *work)
759 { 547 {
760 return (work_data >> WORK_STRUCT_COLOR !! 548 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
761 ((1 << WORK_STRUCT_COLOR_BITS) 549 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
762 } 550 }
763 551
764 static int work_next_color(int color) 552 static int work_next_color(int color)
765 { 553 {
766 return (color + 1) % WORK_NR_COLORS; 554 return (color + 1) % WORK_NR_COLORS;
767 } 555 }
768 556
769 static unsigned long pool_offq_flags(struct wo <<
770 { <<
771 return (pool->flags & POOL_BH) ? WORK_ <<
772 } <<
773 <<
774 /* 557 /*
775 * While queued, %WORK_STRUCT_PWQ is set and n 558 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
776 * contain the pointer to the queued pwq. Onc 559 * contain the pointer to the queued pwq. Once execution starts, the flag
777 * is cleared and the high bits contain OFFQ f 560 * is cleared and the high bits contain OFFQ flags and pool ID.
778 * 561 *
779 * set_work_pwq(), set_work_pool_and_clear_pen !! 562 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
780 * can be used to set the pwq, pool or clear w !! 563 * and clear_work_data() can be used to set the pwq, pool or clear
781 * only be called while the work is owned - ie !! 564 * work->data. These functions should only be called while the work is
>> 565 * owned - ie. while the PENDING bit is set.
782 * 566 *
783 * get_work_pool() and get_work_pwq() can be u 567 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
784 * corresponding to a work. Pool is available 568 * corresponding to a work. Pool is available once the work has been
785 * queued anywhere after initialization until 569 * queued anywhere after initialization until it is sync canceled. pwq is
786 * available only while the work item is queue 570 * available only while the work item is queued.
>> 571 *
>> 572 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
>> 573 * canceled. While being canceled, a work item may have its PENDING set
>> 574 * but stay off timer and worklist for arbitrarily long and nobody should
>> 575 * try to steal the PENDING bit.
787 */ 576 */
788 static inline void set_work_data(struct work_s !! 577 static inline void set_work_data(struct work_struct *work, unsigned long data,
>> 578 unsigned long flags)
789 { 579 {
790 WARN_ON_ONCE(!work_pending(work)); 580 WARN_ON_ONCE(!work_pending(work));
791 atomic_long_set(&work->data, data | wo !! 581 atomic_long_set(&work->data, data | flags | work_static(work));
792 } 582 }
793 583
794 static void set_work_pwq(struct work_struct *w 584 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
795 unsigned long flags) !! 585 unsigned long extra_flags)
796 { 586 {
797 set_work_data(work, (unsigned long)pwq !! 587 set_work_data(work, (unsigned long)pwq,
798 WORK_STRUCT_PWQ | flags) !! 588 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
799 } 589 }
800 590
801 static void set_work_pool_and_keep_pending(str 591 static void set_work_pool_and_keep_pending(struct work_struct *work,
802 int !! 592 int pool_id)
803 { 593 {
804 set_work_data(work, ((unsigned long)po !! 594 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
805 WORK_STRUCT_PENDING | fl !! 595 WORK_STRUCT_PENDING);
806 } 596 }
807 597
808 static void set_work_pool_and_clear_pending(st 598 static void set_work_pool_and_clear_pending(struct work_struct *work,
809 in !! 599 int pool_id)
810 { 600 {
811 /* 601 /*
812 * The following wmb is paired with th 602 * The following wmb is paired with the implied mb in
813 * test_and_set_bit(PENDING) and ensur 603 * test_and_set_bit(PENDING) and ensures all updates to @work made
814 * here are visible to and precede any 604 * here are visible to and precede any updates by the next PENDING
815 * owner. 605 * owner.
816 */ 606 */
817 smp_wmb(); 607 smp_wmb();
818 set_work_data(work, ((unsigned long)po !! 608 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
819 flags); <<
820 /* 609 /*
821 * The following mb guarantees that pr 610 * The following mb guarantees that previous clear of a PENDING bit
822 * will not be reordered with any spec 611 * will not be reordered with any speculative LOADS or STORES from
823 * work->current_func, which is execut 612 * work->current_func, which is executed afterwards. This possible
824 * reordering can lead to a missed exe !! 613 * reordering can lead to a missed execution on attempt to qeueue
825 * the same @work. E.g. consider this 614 * the same @work. E.g. consider this case:
826 * 615 *
827 * CPU#0 CPU 616 * CPU#0 CPU#1
828 * ---------------------------- --- 617 * ---------------------------- --------------------------------
829 * 618 *
830 * 1 STORE event_indicated 619 * 1 STORE event_indicated
831 * 2 queue_work_on() { 620 * 2 queue_work_on() {
832 * 3 test_and_set_bit(PENDING) 621 * 3 test_and_set_bit(PENDING)
833 * 4 } set 622 * 4 } set_..._and_clear_pending() {
834 * 5 s 623 * 5 set_work_data() # clear bit
835 * 6 s 624 * 6 smp_mb()
836 * 7 wor 625 * 7 work->current_func() {
837 * 8 626 * 8 LOAD event_indicated
838 * } 627 * }
839 * 628 *
840 * Without an explicit full barrier sp 629 * Without an explicit full barrier speculative LOAD on line 8 can
841 * be executed before CPU#0 does STORE 630 * be executed before CPU#0 does STORE on line 1. If that happens,
842 * CPU#0 observes the PENDING bit is s 631 * CPU#0 observes the PENDING bit is still set and new execution of
843 * a @work is not queued in a hope, th 632 * a @work is not queued in a hope, that CPU#1 will eventually
844 * finish the queued @work. Meanwhile 633 * finish the queued @work. Meanwhile CPU#1 does not see
845 * event_indicated is set, because spe 634 * event_indicated is set, because speculative LOAD was executed
846 * before actual STORE. 635 * before actual STORE.
847 */ 636 */
848 smp_mb(); 637 smp_mb();
849 } 638 }
850 639
851 static inline struct pool_workqueue *work_stru !! 640 static void clear_work_data(struct work_struct *work)
852 { 641 {
853 return (struct pool_workqueue *)(data !! 642 smp_wmb(); /* see set_work_pool_and_clear_pending() */
>> 643 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
854 } 644 }
855 645
856 static struct pool_workqueue *get_work_pwq(str 646 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
857 { 647 {
858 unsigned long data = atomic_long_read( 648 unsigned long data = atomic_long_read(&work->data);
859 649
860 if (data & WORK_STRUCT_PWQ) 650 if (data & WORK_STRUCT_PWQ)
861 return work_struct_pwq(data); !! 651 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
862 else 652 else
863 return NULL; 653 return NULL;
864 } 654 }
865 655
866 /** 656 /**
867 * get_work_pool - return the worker_pool a gi 657 * get_work_pool - return the worker_pool a given work was associated with
868 * @work: the work item of interest 658 * @work: the work item of interest
869 * 659 *
>> 660 * Return the worker_pool @work was last associated with. %NULL if none.
>> 661 *
870 * Pools are created and destroyed under wq_po 662 * Pools are created and destroyed under wq_pool_mutex, and allows read
871 * access under RCU read lock. As such, this !! 663 * access under sched-RCU read lock. As such, this function should be
872 * called under wq_pool_mutex or inside of a r !! 664 * called under wq_pool_mutex or with preemption disabled.
873 * 665 *
874 * All fields of the returned pool are accessi 666 * All fields of the returned pool are accessible as long as the above
875 * mentioned locking is in effect. If the ret 667 * mentioned locking is in effect. If the returned pool needs to be used
876 * beyond the critical section, the caller is 668 * beyond the critical section, the caller is responsible for ensuring the
877 * returned pool is and stays online. 669 * returned pool is and stays online.
878 * <<
879 * Return: The worker_pool @work was last asso <<
880 */ 670 */
881 static struct worker_pool *get_work_pool(struc 671 static struct worker_pool *get_work_pool(struct work_struct *work)
882 { 672 {
883 unsigned long data = atomic_long_read( 673 unsigned long data = atomic_long_read(&work->data);
884 int pool_id; 674 int pool_id;
885 675
886 assert_rcu_or_pool_mutex(); 676 assert_rcu_or_pool_mutex();
887 677
888 if (data & WORK_STRUCT_PWQ) 678 if (data & WORK_STRUCT_PWQ)
889 return work_struct_pwq(data)-> !! 679 return ((struct pool_workqueue *)
>> 680 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
890 681
891 pool_id = data >> WORK_OFFQ_POOL_SHIFT 682 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
892 if (pool_id == WORK_OFFQ_POOL_NONE) 683 if (pool_id == WORK_OFFQ_POOL_NONE)
893 return NULL; 684 return NULL;
894 685
895 return idr_find(&worker_pool_idr, pool 686 return idr_find(&worker_pool_idr, pool_id);
896 } 687 }
897 688
898 static unsigned long shift_and_mask(unsigned l !! 689 /**
>> 690 * get_work_pool_id - return the worker pool ID a given work is associated with
>> 691 * @work: the work item of interest
>> 692 *
>> 693 * Return the worker_pool ID @work was last associated with.
>> 694 * %WORK_OFFQ_POOL_NONE if none.
>> 695 */
>> 696 static int get_work_pool_id(struct work_struct *work)
899 { 697 {
900 return (v >> shift) & ((1U << bits) - !! 698 unsigned long data = atomic_long_read(&work->data);
>> 699
>> 700 if (data & WORK_STRUCT_PWQ)
>> 701 return ((struct pool_workqueue *)
>> 702 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
>> 703
>> 704 return data >> WORK_OFFQ_POOL_SHIFT;
901 } 705 }
902 706
903 static void work_offqd_unpack(struct work_offq !! 707 static void mark_work_canceling(struct work_struct *work)
904 { 708 {
905 WARN_ON_ONCE(data & WORK_STRUCT_PWQ); !! 709 unsigned long pool_id = get_work_pool_id(work);
906 710
907 offqd->pool_id = shift_and_mask(data, !! 711 pool_id <<= WORK_OFFQ_POOL_SHIFT;
908 WORK_O !! 712 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
909 offqd->disable = shift_and_mask(data, <<
910 WORK_O <<
911 offqd->flags = data & WORK_OFFQ_FLAG_M <<
912 } 713 }
913 714
914 static unsigned long work_offqd_pack_flags(str !! 715 static bool work_is_canceling(struct work_struct *work)
915 { 716 {
916 return ((unsigned long)offqd->disable !! 717 unsigned long data = atomic_long_read(&work->data);
917 ((unsigned long)offqd->flags); !! 718
>> 719 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
918 } 720 }
919 721
920 /* 722 /*
921 * Policy functions. These define the policie 723 * Policy functions. These define the policies on how the global worker
922 * pools are managed. Unless noted otherwise, 724 * pools are managed. Unless noted otherwise, these functions assume that
923 * they're being called with pool->lock held. 725 * they're being called with pool->lock held.
924 */ 726 */
925 727
>> 728 static bool __need_more_worker(struct worker_pool *pool)
>> 729 {
>> 730 return !atomic_read(&pool->nr_running);
>> 731 }
>> 732
926 /* 733 /*
927 * Need to wake up a worker? Called from anyt 734 * Need to wake up a worker? Called from anything but currently
928 * running workers. 735 * running workers.
929 * 736 *
930 * Note that, because unbound workers never co 737 * Note that, because unbound workers never contribute to nr_running, this
931 * function will always return %true for unbou 738 * function will always return %true for unbound pools as long as the
932 * worklist isn't empty. 739 * worklist isn't empty.
933 */ 740 */
934 static bool need_more_worker(struct worker_poo 741 static bool need_more_worker(struct worker_pool *pool)
935 { 742 {
936 return !list_empty(&pool->worklist) && !! 743 return !list_empty(&pool->worklist) && __need_more_worker(pool);
937 } 744 }
938 745
939 /* Can I start working? Called from busy but 746 /* Can I start working? Called from busy but !running workers. */
940 static bool may_start_working(struct worker_po 747 static bool may_start_working(struct worker_pool *pool)
941 { 748 {
942 return pool->nr_idle; 749 return pool->nr_idle;
943 } 750 }
944 751
945 /* Do I need to keep working? Called from cur 752 /* Do I need to keep working? Called from currently running workers. */
946 static bool keep_working(struct worker_pool *p 753 static bool keep_working(struct worker_pool *pool)
947 { 754 {
948 return !list_empty(&pool->worklist) && !! 755 return !list_empty(&pool->worklist) &&
>> 756 atomic_read(&pool->nr_running) <= 1;
949 } 757 }
950 758
951 /* Do we need a new worker? Called from manag 759 /* Do we need a new worker? Called from manager. */
952 static bool need_to_create_worker(struct worke 760 static bool need_to_create_worker(struct worker_pool *pool)
953 { 761 {
954 return need_more_worker(pool) && !may_ 762 return need_more_worker(pool) && !may_start_working(pool);
955 } 763 }
956 764
>> 765 /* Do I need to be the manager? */
>> 766 static bool need_to_manage_workers(struct worker_pool *pool)
>> 767 {
>> 768 return need_to_create_worker(pool) ||
>> 769 (pool->flags & POOL_MANAGE_WORKERS);
>> 770 }
>> 771
957 /* Do we have too many workers and should some 772 /* Do we have too many workers and should some go away? */
958 static bool too_many_workers(struct worker_poo 773 static bool too_many_workers(struct worker_pool *pool)
959 { 774 {
960 bool managing = pool->flags & POOL_MAN !! 775 bool managing = mutex_is_locked(&pool->manager_arb);
961 int nr_idle = pool->nr_idle + managing 776 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
962 int nr_busy = pool->nr_workers - nr_id 777 int nr_busy = pool->nr_workers - nr_idle;
963 778
>> 779 /*
>> 780 * nr_idle and idle_list may disagree if idle rebinding is in
>> 781 * progress. Never return %true if idle_list is empty.
>> 782 */
>> 783 if (list_empty(&pool->idle_list))
>> 784 return false;
>> 785
964 return nr_idle > 2 && (nr_idle - 2) * 786 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
965 } 787 }
966 788
967 /** !! 789 /*
968 * worker_set_flags - set worker flags and adj !! 790 * Wake up functions.
969 * @worker: self <<
970 * @flags: flags to set <<
971 * <<
972 * Set @flags in @worker->flags and adjust nr_ <<
973 */ 791 */
974 static inline void worker_set_flags(struct wor !! 792
>> 793 /* Return the first worker. Safe with preemption disabled */
>> 794 static struct worker *first_worker(struct worker_pool *pool)
975 { 795 {
976 struct worker_pool *pool = worker->poo !! 796 if (unlikely(list_empty(&pool->idle_list)))
>> 797 return NULL;
977 798
978 lockdep_assert_held(&pool->lock); !! 799 return list_first_entry(&pool->idle_list, struct worker, entry);
>> 800 }
979 801
980 /* If transitioning into NOT_RUNNING, !! 802 /**
981 if ((flags & WORKER_NOT_RUNNING) && !! 803 * wake_up_worker - wake up an idle worker
982 !(worker->flags & WORKER_NOT_RUNNI !! 804 * @pool: worker pool to wake worker from
983 pool->nr_running--; !! 805 *
984 } !! 806 * Wake up the first idle worker of @pool.
>> 807 *
>> 808 * CONTEXT:
>> 809 * spin_lock_irq(pool->lock).
>> 810 */
>> 811 static void wake_up_worker(struct worker_pool *pool)
>> 812 {
>> 813 struct worker *worker = first_worker(pool);
985 814
986 worker->flags |= flags; !! 815 if (likely(worker))
>> 816 wake_up_process(worker->task);
987 } 817 }
988 818
989 /** 819 /**
990 * worker_clr_flags - clear worker flags and a !! 820 * wq_worker_waking_up - a worker is waking up
991 * @worker: self !! 821 * @task: task waking up
992 * @flags: flags to clear !! 822 * @cpu: CPU @task is waking up to
993 * 823 *
994 * Clear @flags in @worker->flags and adjust n !! 824 * This function is called during try_to_wake_up() when a worker is
>> 825 * being awoken.
>> 826 *
>> 827 * CONTEXT:
>> 828 * spin_lock_irq(rq->lock)
995 */ 829 */
996 static inline void worker_clr_flags(struct wor !! 830 void wq_worker_waking_up(struct task_struct *task, int cpu)
997 { 831 {
998 struct worker_pool *pool = worker->poo !! 832 struct worker *worker = kthread_data(task);
999 unsigned int oflags = worker->flags; <<
1000 833
1001 lockdep_assert_held(&pool->lock); !! 834 if (!(worker->flags & WORKER_NOT_RUNNING)) {
>> 835 WARN_ON_ONCE(worker->pool->cpu != cpu);
>> 836 atomic_inc(&worker->pool->nr_running);
>> 837 }
>> 838 }
1002 839
1003 worker->flags &= ~flags; !! 840 /**
>> 841 * wq_worker_sleeping - a worker is going to sleep
>> 842 * @task: task going to sleep
>> 843 * @cpu: CPU in question, must be the current CPU number
>> 844 *
>> 845 * This function is called during schedule() when a busy worker is
>> 846 * going to sleep. Worker on the same cpu can be woken up by
>> 847 * returning pointer to its task.
>> 848 *
>> 849 * CONTEXT:
>> 850 * spin_lock_irq(rq->lock)
>> 851 *
>> 852 * RETURNS:
>> 853 * Worker task on @cpu to wake up, %NULL if none.
>> 854 */
>> 855 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
>> 856 {
>> 857 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
>> 858 struct worker_pool *pool;
1004 859
1005 /* 860 /*
1006 * If transitioning out of NOT_RUNNIN !! 861 * Rescuers, which may not have all the fields set up like normal
1007 * that the nested NOT_RUNNING is not !! 862 * workers, also reach here, let's not access anything before
1008 * of multiple flags, not a single fl !! 863 * checking NOT_RUNNING.
1009 */ 864 */
1010 if ((flags & WORKER_NOT_RUNNING) && ( !! 865 if (worker->flags & WORKER_NOT_RUNNING)
1011 if (!(worker->flags & WORKER_ !! 866 return NULL;
1012 pool->nr_running++; <<
1013 } <<
1014 867
1015 /* Return the first idle worker. Called with !! 868 pool = worker->pool;
1016 static struct worker *first_idle_worker(struc !! 869
1017 { !! 870 /* this can only happen on the local cpu */
1018 if (unlikely(list_empty(&pool->idle_l !! 871 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
1019 return NULL; 872 return NULL;
1020 873
1021 return list_first_entry(&pool->idle_l !! 874 /*
>> 875 * The counterpart of the following dec_and_test, implied mb,
>> 876 * worklist not empty test sequence is in insert_work().
>> 877 * Please read comment there.
>> 878 *
>> 879 * NOT_RUNNING is clear. This means that we're bound to and
>> 880 * running on the local cpu w/ rq lock held and preemption
>> 881 * disabled, which in turn means that none else could be
>> 882 * manipulating idle_list, so dereferencing idle_list without pool
>> 883 * lock is safe.
>> 884 */
>> 885 if (atomic_dec_and_test(&pool->nr_running) &&
>> 886 !list_empty(&pool->worklist))
>> 887 to_wakeup = first_worker(pool);
>> 888 return to_wakeup ? to_wakeup->task : NULL;
1022 } 889 }
1023 890
1024 /** 891 /**
1025 * worker_enter_idle - enter idle state !! 892 * worker_set_flags - set worker flags and adjust nr_running accordingly
1026 * @worker: worker which is entering idle sta !! 893 * @worker: self
>> 894 * @flags: flags to set
>> 895 * @wakeup: wakeup an idle worker if necessary
1027 * 896 *
1028 * @worker is entering idle state. Update st !! 897 * Set @flags in @worker->flags and adjust nr_running accordingly. If
1029 * necessary. !! 898 * nr_running becomes zero and @wakeup is %true, an idle worker is
>> 899 * woken up.
1030 * 900 *
1031 * LOCKING: !! 901 * CONTEXT:
1032 * raw_spin_lock_irq(pool->lock). !! 902 * spin_lock_irq(pool->lock)
1033 */ 903 */
1034 static void worker_enter_idle(struct worker * !! 904 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
>> 905 bool wakeup)
1035 { 906 {
1036 struct worker_pool *pool = worker->po 907 struct worker_pool *pool = worker->pool;
1037 908
1038 if (WARN_ON_ONCE(worker->flags & WORK !! 909 WARN_ON_ONCE(worker->task != current);
1039 WARN_ON_ONCE(!list_empty(&worker- <<
1040 (worker->hentry.next <<
1041 return; <<
1042 <<
1043 /* can't use worker_set_flags(), also <<
1044 worker->flags |= WORKER_IDLE; <<
1045 pool->nr_idle++; <<
1046 worker->last_active = jiffies; <<
1047 910
1048 /* idle_list is LIFO */ !! 911 /*
1049 list_add(&worker->entry, &pool->idle_ !! 912 * If transitioning into NOT_RUNNING, adjust nr_running and
1050 !! 913 * wake up an idle worker as necessary if requested by
1051 if (too_many_workers(pool) && !timer_ !! 914 * @wakeup.
1052 mod_timer(&pool->idle_timer, !! 915 */
>> 916 if ((flags & WORKER_NOT_RUNNING) &&
>> 917 !(worker->flags & WORKER_NOT_RUNNING)) {
>> 918 if (wakeup) {
>> 919 if (atomic_dec_and_test(&pool->nr_running) &&
>> 920 !list_empty(&pool->worklist))
>> 921 wake_up_worker(pool);
>> 922 } else
>> 923 atomic_dec(&pool->nr_running);
>> 924 }
1053 925
1054 /* Sanity check nr_running. */ !! 926 worker->flags |= flags;
1055 WARN_ON_ONCE(pool->nr_workers == pool <<
1056 } 927 }
1057 928
1058 /** 929 /**
1059 * worker_leave_idle - leave idle state !! 930 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
1060 * @worker: worker which is leaving idle stat !! 931 * @worker: self
>> 932 * @flags: flags to clear
1061 * 933 *
1062 * @worker is leaving idle state. Update sta !! 934 * Clear @flags in @worker->flags and adjust nr_running accordingly.
1063 * 935 *
1064 * LOCKING: !! 936 * CONTEXT:
1065 * raw_spin_lock_irq(pool->lock). !! 937 * spin_lock_irq(pool->lock)
1066 */ 938 */
1067 static void worker_leave_idle(struct worker * !! 939 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
1068 { 940 {
1069 struct worker_pool *pool = worker->po 941 struct worker_pool *pool = worker->pool;
>> 942 unsigned int oflags = worker->flags;
1070 943
1071 if (WARN_ON_ONCE(!(worker->flags & WO !! 944 WARN_ON_ONCE(worker->task != current);
1072 return; !! 945
1073 worker_clr_flags(worker, WORKER_IDLE) !! 946 worker->flags &= ~flags;
1074 pool->nr_idle--; !! 947
1075 list_del_init(&worker->entry); !! 948 /*
>> 949 * If transitioning out of NOT_RUNNING, increment nr_running. Note
>> 950 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
>> 951 * of multiple flags, not a single flag.
>> 952 */
>> 953 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
>> 954 if (!(worker->flags & WORKER_NOT_RUNNING))
>> 955 atomic_inc(&pool->nr_running);
1076 } 956 }
1077 957
1078 /** 958 /**
1079 * find_worker_executing_work - find worker w 959 * find_worker_executing_work - find worker which is executing a work
1080 * @pool: pool of interest 960 * @pool: pool of interest
1081 * @work: work to find worker for 961 * @work: work to find worker for
1082 * 962 *
1083 * Find a worker which is executing @work on 963 * Find a worker which is executing @work on @pool by searching
1084 * @pool->busy_hash which is keyed by the add 964 * @pool->busy_hash which is keyed by the address of @work. For a worker
1085 * to match, its current execution should mat 965 * to match, its current execution should match the address of @work and
1086 * its work function. This is to avoid unwan 966 * its work function. This is to avoid unwanted dependency between
1087 * unrelated work executions through a work i 967 * unrelated work executions through a work item being recycled while still
1088 * being executed. 968 * being executed.
1089 * 969 *
1090 * This is a bit tricky. A work item may be 970 * This is a bit tricky. A work item may be freed once its execution
1091 * starts and nothing prevents the freed area 971 * starts and nothing prevents the freed area from being recycled for
1092 * another work item. If the same work item 972 * another work item. If the same work item address ends up being reused
1093 * before the original execution finishes, wo 973 * before the original execution finishes, workqueue will identify the
1094 * recycled work item as currently executing 974 * recycled work item as currently executing and make it wait until the
1095 * current execution finishes, introducing an 975 * current execution finishes, introducing an unwanted dependency.
1096 * 976 *
1097 * This function checks the work item address 977 * This function checks the work item address and work function to avoid
1098 * false positives. Note that this isn't com 978 * false positives. Note that this isn't complete as one may construct a
1099 * work function which can introduce dependen 979 * work function which can introduce dependency onto itself through a
1100 * recycled work item. Well, if somebody wan 980 * recycled work item. Well, if somebody wants to shoot oneself in the
1101 * foot that badly, there's only so much we c 981 * foot that badly, there's only so much we can do, and if such deadlock
1102 * actually occurs, it should be easy to loca 982 * actually occurs, it should be easy to locate the culprit work function.
1103 * 983 *
1104 * CONTEXT: 984 * CONTEXT:
1105 * raw_spin_lock_irq(pool->lock). !! 985 * spin_lock_irq(pool->lock).
1106 * 986 *
1107 * Return: !! 987 * RETURNS:
1108 * Pointer to worker which is executing @work !! 988 * Pointer to worker which is executing @work if found, NULL
1109 * otherwise. 989 * otherwise.
1110 */ 990 */
1111 static struct worker *find_worker_executing_w 991 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1112 992 struct work_struct *work)
1113 { 993 {
1114 struct worker *worker; 994 struct worker *worker;
1115 995
1116 hash_for_each_possible(pool->busy_has 996 hash_for_each_possible(pool->busy_hash, worker, hentry,
1117 (unsigned long 997 (unsigned long)work)
1118 if (worker->current_work == w 998 if (worker->current_work == work &&
1119 worker->current_func == w 999 worker->current_func == work->func)
1120 return worker; 1000 return worker;
1121 1001
1122 return NULL; 1002 return NULL;
1123 } 1003 }
1124 1004
1125 /** 1005 /**
1126 * move_linked_works - move linked works to a 1006 * move_linked_works - move linked works to a list
1127 * @work: start of series of works to be sche 1007 * @work: start of series of works to be scheduled
1128 * @head: target list to append @work to 1008 * @head: target list to append @work to
1129 * @nextp: out parameter for nested worklist !! 1009 * @nextp: out paramter for nested worklist walking
1130 * 1010 *
1131 * Schedule linked works starting from @work !! 1011 * Schedule linked works starting from @work to @head. Work series to
1132 * scheduled starts at @work and includes any !! 1012 * be scheduled starts at @work and includes any consecutive work with
1133 * WORK_STRUCT_LINKED set in its predecessor. !! 1013 * WORK_STRUCT_LINKED set in its predecessor.
1134 * @nextp. !! 1014 *
>> 1015 * If @nextp is not NULL, it's updated to point to the next work of
>> 1016 * the last scheduled work. This allows move_linked_works() to be
>> 1017 * nested inside outer list_for_each_entry_safe().
1135 * 1018 *
1136 * CONTEXT: 1019 * CONTEXT:
1137 * raw_spin_lock_irq(pool->lock). !! 1020 * spin_lock_irq(pool->lock).
1138 */ 1021 */
1139 static void move_linked_works(struct work_str 1022 static void move_linked_works(struct work_struct *work, struct list_head *head,
1140 struct work_str 1023 struct work_struct **nextp)
1141 { 1024 {
1142 struct work_struct *n; 1025 struct work_struct *n;
1143 1026
1144 /* 1027 /*
1145 * Linked worklist will always end be 1028 * Linked worklist will always end before the end of the list,
1146 * use NULL for list head. 1029 * use NULL for list head.
1147 */ 1030 */
1148 list_for_each_entry_safe_from(work, n 1031 list_for_each_entry_safe_from(work, n, NULL, entry) {
1149 list_move_tail(&work->entry, 1032 list_move_tail(&work->entry, head);
1150 if (!(*work_data_bits(work) & 1033 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1151 break; 1034 break;
1152 } 1035 }
1153 1036
1154 /* 1037 /*
1155 * If we're already inside safe list 1038 * If we're already inside safe list traversal and have moved
1156 * multiple works to the scheduled qu 1039 * multiple works to the scheduled queue, the next position
1157 * needs to be updated. 1040 * needs to be updated.
1158 */ 1041 */
1159 if (nextp) 1042 if (nextp)
1160 *nextp = n; 1043 *nextp = n;
1161 } 1044 }
1162 1045
1163 /** 1046 /**
1164 * assign_work - assign a work item and its l <<
1165 * @work: work to assign <<
1166 * @worker: worker to assign to <<
1167 * @nextp: out parameter for nested worklist <<
1168 * <<
1169 * Assign @work and its linked work items to <<
1170 * executed by another worker in the same poo <<
1171 * <<
1172 * If @nextp is not NULL, it's updated to poi <<
1173 * scheduled work. This allows assign_work() <<
1174 * list_for_each_entry_safe(). <<
1175 * <<
1176 * Returns %true if @work was successfully as <<
1177 * was punted to another worker already execu <<
1178 */ <<
1179 static bool assign_work(struct work_struct *w <<
1180 struct work_struct ** <<
1181 { <<
1182 struct worker_pool *pool = worker->po <<
1183 struct worker *collision; <<
1184 <<
1185 lockdep_assert_held(&pool->lock); <<
1186 <<
1187 /* <<
1188 * A single work shouldn't be execute <<
1189 * __queue_work() ensures that @work <<
1190 * while still running in the previou <<
1191 * @work is not executed concurrently <<
1192 * pool. Check whether anyone is alre <<
1193 * defer the work to the currently ex <<
1194 */ <<
1195 collision = find_worker_executing_wor <<
1196 if (unlikely(collision)) { <<
1197 move_linked_works(work, &coll <<
1198 return false; <<
1199 } <<
1200 <<
1201 move_linked_works(work, &worker->sche <<
1202 return true; <<
1203 } <<
1204 <<
1205 static struct irq_work *bh_pool_irq_work(stru <<
1206 { <<
1207 int high = pool->attrs->nice == HIGHP <<
1208 <<
1209 return &per_cpu(bh_pool_irq_works, po <<
1210 } <<
1211 <<
1212 static void kick_bh_pool(struct worker_pool * <<
1213 { <<
1214 #ifdef CONFIG_SMP <<
1215 /* see drain_dead_softirq_workfn() fo <<
1216 if (unlikely(pool->cpu != smp_process <<
1217 !(pool->flags & POOL_BH_ <<
1218 irq_work_queue_on(bh_pool_irq <<
1219 return; <<
1220 } <<
1221 #endif <<
1222 if (pool->attrs->nice == HIGHPRI_NICE <<
1223 raise_softirq_irqoff(HI_SOFTI <<
1224 else <<
1225 raise_softirq_irqoff(TASKLET_ <<
1226 } <<
1227 <<
1228 /** <<
1229 * kick_pool - wake up an idle worker if nece <<
1230 * @pool: pool to kick <<
1231 * <<
1232 * @pool may have pending work items. Wake up <<
1233 * whether a worker was woken up. <<
1234 */ <<
1235 static bool kick_pool(struct worker_pool *poo <<
1236 { <<
1237 struct worker *worker = first_idle_wo <<
1238 struct task_struct *p; <<
1239 <<
1240 lockdep_assert_held(&pool->lock); <<
1241 <<
1242 if (!need_more_worker(pool) || !worke <<
1243 return false; <<
1244 <<
1245 if (pool->flags & POOL_BH) { <<
1246 kick_bh_pool(pool); <<
1247 return true; <<
1248 } <<
1249 <<
1250 p = worker->task; <<
1251 <<
1252 #ifdef CONFIG_SMP <<
1253 /* <<
1254 * Idle @worker is about to execute @ <<
1255 * opportunity to migrate @worker at <<
1256 * wake_cpu field. Let's see if we wa <<
1257 * execution locality. <<
1258 * <<
1259 * We're waking the worker that went <<
1260 * chance that @worker is marked idle <<
1261 * so, setting the wake_cpu won't do <<
1262 * optimization and the race window i <<
1263 * now. If this becomes pronounced, w <<
1264 * still on cpu when picking an idle <<
1265 * <<
1266 * If @pool has non-strict affinity, <<
1267 * its affinity scope. Repatriate. <<
1268 */ <<
1269 if (!pool->attrs->affn_strict && <<
1270 !cpumask_test_cpu(p->wake_cpu, po <<
1271 struct work_struct *work = li <<
1272 <<
1273 int wake_cpu = cpumask_any_an <<
1274 <<
1275 if (wake_cpu < nr_cpu_ids) { <<
1276 p->wake_cpu = wake_cp <<
1277 get_work_pwq(work)->s <<
1278 } <<
1279 } <<
1280 #endif <<
1281 wake_up_process(p); <<
1282 return true; <<
1283 } <<
1284 <<
1285 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT <<
1286 <<
1287 /* <<
1288 * Concurrency-managed per-cpu work items tha <<
1289 * wq_cpu_intensive_thresh_us trigger the aut <<
1290 * which prevents them from stalling other co <<
1291 * work function keeps triggering this mechan <<
1292 * should be using an unbound workqueue inste <<
1293 * <<
1294 * wq_cpu_intensive_report() tracks work func <<
1295 * and report them so that they can be examin <<
1296 * workqueues as appropriate. To avoid floodi <<
1297 * function is tracked and reported with expo <<
1298 */ <<
1299 #define WCI_MAX_ENTS 128 <<
1300 <<
1301 struct wci_ent { <<
1302 work_func_t func; <<
1303 atomic64_t cnt; <<
1304 struct hlist_node hash_node; <<
1305 }; <<
1306 <<
1307 static struct wci_ent wci_ents[WCI_MAX_ENTS]; <<
1308 static int wci_nr_ents; <<
1309 static DEFINE_RAW_SPINLOCK(wci_lock); <<
1310 static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_M <<
1311 <<
1312 static struct wci_ent *wci_find_ent(work_func <<
1313 { <<
1314 struct wci_ent *ent; <<
1315 <<
1316 hash_for_each_possible_rcu(wci_hash, <<
1317 (unsigned <<
1318 if (ent->func == func) <<
1319 return ent; <<
1320 } <<
1321 return NULL; <<
1322 } <<
1323 <<
1324 static void wq_cpu_intensive_report(work_func <<
1325 { <<
1326 struct wci_ent *ent; <<
1327 <<
1328 restart: <<
1329 ent = wci_find_ent(func); <<
1330 if (ent) { <<
1331 u64 cnt; <<
1332 <<
1333 /* <<
1334 * Start reporting from the w <<
1335 * exponentially. <<
1336 */ <<
1337 cnt = atomic64_inc_return_rel <<
1338 if (wq_cpu_intensive_warning_ <<
1339 cnt >= wq_cpu_intensive_w <<
1340 is_power_of_2(cnt + 1 - w <<
1341 printk_deferred(KERN_ <<
1342 ent-> <<
1343 atomi <<
1344 return; <<
1345 } <<
1346 <<
1347 /* <<
1348 * @func is a new violation. Allocate <<
1349 * is exhausted, something went reall <<
1350 * noise already. <<
1351 */ <<
1352 if (wci_nr_ents >= WCI_MAX_ENTS) <<
1353 return; <<
1354 <<
1355 raw_spin_lock(&wci_lock); <<
1356 <<
1357 if (wci_nr_ents >= WCI_MAX_ENTS) { <<
1358 raw_spin_unlock(&wci_lock); <<
1359 return; <<
1360 } <<
1361 <<
1362 if (wci_find_ent(func)) { <<
1363 raw_spin_unlock(&wci_lock); <<
1364 goto restart; <<
1365 } <<
1366 <<
1367 ent = &wci_ents[wci_nr_ents++]; <<
1368 ent->func = func; <<
1369 atomic64_set(&ent->cnt, 0); <<
1370 hash_add_rcu(wci_hash, &ent->hash_nod <<
1371 <<
1372 raw_spin_unlock(&wci_lock); <<
1373 <<
1374 goto restart; <<
1375 } <<
1376 <<
1377 #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ <<
1378 static void wq_cpu_intensive_report(work_func <<
1379 #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ <<
1380 <<
1381 /** <<
1382 * wq_worker_running - a worker is running ag <<
1383 * @task: task waking up <<
1384 * <<
1385 * This function is called when a worker retu <<
1386 */ <<
1387 void wq_worker_running(struct task_struct *ta <<
1388 { <<
1389 struct worker *worker = kthread_data( <<
1390 <<
1391 if (!READ_ONCE(worker->sleeping)) <<
1392 return; <<
1393 <<
1394 /* <<
1395 * If preempted by unbind_workers() b <<
1396 * and the nr_running increment below <<
1397 * and leave with an unexpected pool- <<
1398 * pool. Protect against such race. <<
1399 */ <<
1400 preempt_disable(); <<
1401 if (!(worker->flags & WORKER_NOT_RUNN <<
1402 worker->pool->nr_running++; <<
1403 preempt_enable(); <<
1404 <<
1405 /* <<
1406 * CPU intensive auto-detection cares <<
1407 * CPU without sleeping. Reset the st <<
1408 */ <<
1409 worker->current_at = worker->task->se <<
1410 <<
1411 WRITE_ONCE(worker->sleeping, 0); <<
1412 } <<
1413 <<
1414 /** <<
1415 * wq_worker_sleeping - a worker is going to <<
1416 * @task: task going to sleep <<
1417 * <<
1418 * This function is called from schedule() wh <<
1419 * going to sleep. <<
1420 */ <<
1421 void wq_worker_sleeping(struct task_struct *t <<
1422 { <<
1423 struct worker *worker = kthread_data( <<
1424 struct worker_pool *pool; <<
1425 <<
1426 /* <<
1427 * Rescuers, which may not have all t <<
1428 * workers, also reach here, let's no <<
1429 * checking NOT_RUNNING. <<
1430 */ <<
1431 if (worker->flags & WORKER_NOT_RUNNIN <<
1432 return; <<
1433 <<
1434 pool = worker->pool; <<
1435 <<
1436 /* Return if preempted before wq_work <<
1437 if (READ_ONCE(worker->sleeping)) <<
1438 return; <<
1439 <<
1440 WRITE_ONCE(worker->sleeping, 1); <<
1441 raw_spin_lock_irq(&pool->lock); <<
1442 <<
1443 /* <<
1444 * Recheck in case unbind_workers() p <<
1445 * want to decrement nr_running after <<
1446 * and nr_running has been reset. <<
1447 */ <<
1448 if (worker->flags & WORKER_NOT_RUNNIN <<
1449 raw_spin_unlock_irq(&pool->lo <<
1450 return; <<
1451 } <<
1452 <<
1453 pool->nr_running--; <<
1454 if (kick_pool(pool)) <<
1455 worker->current_pwq->stats[PW <<
1456 <<
1457 raw_spin_unlock_irq(&pool->lock); <<
1458 } <<
1459 <<
1460 /** <<
1461 * wq_worker_tick - a scheduler tick occurred <<
1462 * @task: task currently running <<
1463 * <<
1464 * Called from sched_tick(). We're in the IRQ <<
1465 * worker's fields which follow the 'K' locki <<
1466 */ <<
1467 void wq_worker_tick(struct task_struct *task) <<
1468 { <<
1469 struct worker *worker = kthread_data( <<
1470 struct pool_workqueue *pwq = worker-> <<
1471 struct worker_pool *pool = worker->po <<
1472 <<
1473 if (!pwq) <<
1474 return; <<
1475 <<
1476 pwq->stats[PWQ_STAT_CPU_TIME] += TICK <<
1477 <<
1478 if (!wq_cpu_intensive_thresh_us) <<
1479 return; <<
1480 <<
1481 /* <<
1482 * If the current worker is concurren <<
1483 * longer than wq_cpu_intensive_thres <<
1484 * CPU_INTENSIVE to avoid stalling ot <<
1485 * <<
1486 * Set @worker->sleeping means that @ <<
1487 * switching out voluntarily and won' <<
1488 * @pool->nr_running until it wakes u <<
1489 * decrements ->nr_running, setting C <<
1490 * double decrements. The task is rel <<
1491 * We probably want to make this pret <<
1492 */ <<
1493 if ((worker->flags & WORKER_NOT_RUNNI <<
1494 worker->task->se.sum_exec_runtime <<
1495 wq_cpu_intensive_thresh_us * NSEC <<
1496 return; <<
1497 <<
1498 raw_spin_lock(&pool->lock); <<
1499 <<
1500 worker_set_flags(worker, WORKER_CPU_I <<
1501 wq_cpu_intensive_report(worker->curre <<
1502 pwq->stats[PWQ_STAT_CPU_INTENSIVE]++; <<
1503 <<
1504 if (kick_pool(pool)) <<
1505 pwq->stats[PWQ_STAT_CM_WAKEUP <<
1506 <<
1507 raw_spin_unlock(&pool->lock); <<
1508 } <<
1509 <<
1510 /** <<
1511 * wq_worker_last_func - retrieve worker's la <<
1512 * @task: Task to retrieve last work function <<
1513 * <<
1514 * Determine the last function a worker execu <<
1515 * the scheduler to get a worker's last known <<
1516 * <<
1517 * CONTEXT: <<
1518 * raw_spin_lock_irq(rq->lock) <<
1519 * <<
1520 * This function is called during schedule() <<
1521 * to sleep. It's used by psi to identify agg <<
1522 * dequeuing, to allow periodic aggregation t <<
1523 * worker is the last task in the system or c <<
1524 * <<
1525 * As this function doesn't involve any workq <<
1526 * only returns stable values when called fro <<
1527 * queuing and dequeuing paths, when @task, w <<
1528 * is guaranteed to not be processing any wor <<
1529 * <<
1530 * Return: <<
1531 * The last work function %current executed a <<
1532 * hasn't executed any work yet. <<
1533 */ <<
1534 work_func_t wq_worker_last_func(struct task_s <<
1535 { <<
1536 struct worker *worker = kthread_data( <<
1537 <<
1538 return worker->last_func; <<
1539 } <<
1540 <<
1541 /** <<
1542 * wq_node_nr_active - Determine wq_node_nr_a <<
1543 * @wq: workqueue of interest <<
1544 * @node: NUMA node, can be %NUMA_NO_NODE <<
1545 * <<
1546 * Determine wq_node_nr_active to use for @wq <<
1547 * <<
1548 * - %NULL for per-cpu workqueues as they don <<
1549 * <<
1550 * - node_nr_active[nr_node_ids] if @node is <<
1551 * <<
1552 * - Otherwise, node_nr_active[@node]. <<
1553 */ <<
1554 static struct wq_node_nr_active *wq_node_nr_a <<
1555 <<
1556 { <<
1557 if (!(wq->flags & WQ_UNBOUND)) <<
1558 return NULL; <<
1559 <<
1560 if (node == NUMA_NO_NODE) <<
1561 node = nr_node_ids; <<
1562 <<
1563 return wq->node_nr_active[node]; <<
1564 } <<
1565 <<
1566 /** <<
1567 * wq_update_node_max_active - Update per-nod <<
1568 * @wq: workqueue to update <<
1569 * @off_cpu: CPU that's going down, -1 if a C <<
1570 * <<
1571 * Update @wq->node_nr_active[]->max. @wq mus <<
1572 * distributed among nodes according to the p <<
1573 * cpus. The result is always between @wq->mi <<
1574 */ <<
1575 static void wq_update_node_max_active(struct <<
1576 { <<
1577 struct cpumask *effective = unbound_e <<
1578 int min_active = READ_ONCE(wq->min_ac <<
1579 int max_active = READ_ONCE(wq->max_ac <<
1580 int total_cpus, node; <<
1581 <<
1582 lockdep_assert_held(&wq->mutex); <<
1583 <<
1584 if (!wq_topo_initialized) <<
1585 return; <<
1586 <<
1587 if (off_cpu >= 0 && !cpumask_test_cpu <<
1588 off_cpu = -1; <<
1589 <<
1590 total_cpus = cpumask_weight_and(effec <<
1591 if (off_cpu >= 0) <<
1592 total_cpus--; <<
1593 <<
1594 /* If all CPUs of the wq get offline, <<
1595 if (unlikely(!total_cpus)) { <<
1596 for_each_node(node) <<
1597 wq_node_nr_active(wq, <<
1598 <<
1599 wq_node_nr_active(wq, NUMA_NO <<
1600 return; <<
1601 } <<
1602 <<
1603 for_each_node(node) { <<
1604 int node_cpus; <<
1605 <<
1606 node_cpus = cpumask_weight_an <<
1607 if (off_cpu >= 0 && cpu_to_no <<
1608 node_cpus--; <<
1609 <<
1610 wq_node_nr_active(wq, node)-> <<
1611 clamp(DIV_ROUND_UP(ma <<
1612 min_active, max <<
1613 } <<
1614 <<
1615 wq_node_nr_active(wq, NUMA_NO_NODE)-> <<
1616 } <<
1617 <<
1618 /** <<
1619 * get_pwq - get an extra reference on the sp 1047 * get_pwq - get an extra reference on the specified pool_workqueue
1620 * @pwq: pool_workqueue to get 1048 * @pwq: pool_workqueue to get
1621 * 1049 *
1622 * Obtain an extra reference on @pwq. The ca 1050 * Obtain an extra reference on @pwq. The caller should guarantee that
1623 * @pwq has positive refcnt and be holding th 1051 * @pwq has positive refcnt and be holding the matching pool->lock.
1624 */ 1052 */
1625 static void get_pwq(struct pool_workqueue *pw 1053 static void get_pwq(struct pool_workqueue *pwq)
1626 { 1054 {
1627 lockdep_assert_held(&pwq->pool->lock) 1055 lockdep_assert_held(&pwq->pool->lock);
1628 WARN_ON_ONCE(pwq->refcnt <= 0); 1056 WARN_ON_ONCE(pwq->refcnt <= 0);
1629 pwq->refcnt++; 1057 pwq->refcnt++;
1630 } 1058 }
1631 1059
1632 /** 1060 /**
1633 * put_pwq - put a pool_workqueue reference 1061 * put_pwq - put a pool_workqueue reference
1634 * @pwq: pool_workqueue to put 1062 * @pwq: pool_workqueue to put
1635 * 1063 *
1636 * Drop a reference of @pwq. If its refcnt r 1064 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1637 * destruction. The caller should be holding 1065 * destruction. The caller should be holding the matching pool->lock.
1638 */ 1066 */
1639 static void put_pwq(struct pool_workqueue *pw 1067 static void put_pwq(struct pool_workqueue *pwq)
1640 { 1068 {
1641 lockdep_assert_held(&pwq->pool->lock) 1069 lockdep_assert_held(&pwq->pool->lock);
1642 if (likely(--pwq->refcnt)) 1070 if (likely(--pwq->refcnt))
1643 return; 1071 return;
>> 1072 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
>> 1073 return;
1644 /* 1074 /*
1645 * @pwq can't be released under pool- !! 1075 * @pwq can't be released under pool->lock, bounce to
1646 * kthread_worker to avoid A-A deadlo !! 1076 * pwq_unbound_release_workfn(). This never recurses on the same
>> 1077 * pool->lock as this path is taken only for unbound workqueues and
>> 1078 * the release work item is scheduled on a per-cpu workqueue. To
>> 1079 * avoid lockdep warning, unbound pool->locks are given lockdep
>> 1080 * subclass of 1 in get_unbound_pool().
1647 */ 1081 */
1648 kthread_queue_work(pwq_release_worker !! 1082 schedule_work(&pwq->unbound_release_work);
1649 } 1083 }
1650 1084
1651 /** 1085 /**
1652 * put_pwq_unlocked - put_pwq() with surround 1086 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1653 * @pwq: pool_workqueue to put (can be %NULL) 1087 * @pwq: pool_workqueue to put (can be %NULL)
1654 * 1088 *
1655 * put_pwq() with locking. This function als 1089 * put_pwq() with locking. This function also allows %NULL @pwq.
1656 */ 1090 */
1657 static void put_pwq_unlocked(struct pool_work 1091 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1658 { 1092 {
1659 if (pwq) { 1093 if (pwq) {
1660 /* 1094 /*
1661 * As both pwqs and pools are !! 1095 * As both pwqs and pools are sched-RCU protected, the
1662 * following lock operations 1096 * following lock operations are safe.
1663 */ 1097 */
1664 raw_spin_lock_irq(&pwq->pool- !! 1098 spin_lock_irq(&pwq->pool->lock);
1665 put_pwq(pwq); 1099 put_pwq(pwq);
1666 raw_spin_unlock_irq(&pwq->poo !! 1100 spin_unlock_irq(&pwq->pool->lock);
1667 } 1101 }
1668 } 1102 }
1669 1103
1670 static bool pwq_is_empty(struct pool_workqueu !! 1104 static void pwq_activate_delayed_work(struct work_struct *work)
1671 { <<
1672 return !pwq->nr_active && list_empty( <<
1673 } <<
1674 <<
1675 static void __pwq_activate_work(struct pool_w <<
1676 struct work_s <<
1677 { 1105 {
1678 unsigned long *wdb = work_data_bits(w !! 1106 struct pool_workqueue *pwq = get_work_pwq(work);
1679 1107
1680 WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INA <<
1681 trace_workqueue_activate_work(work); 1108 trace_workqueue_activate_work(work);
1682 if (list_empty(&pwq->pool->worklist)) <<
1683 pwq->pool->watchdog_ts = jiff <<
1684 move_linked_works(work, &pwq->pool->w 1109 move_linked_works(work, &pwq->pool->worklist, NULL);
1685 __clear_bit(WORK_STRUCT_INACTIVE_BIT, !! 1110 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1686 } !! 1111 pwq->nr_active++;
1687 <<
1688 static bool tryinc_node_nr_active(struct wq_n <<
1689 { <<
1690 int max = READ_ONCE(nna->max); <<
1691 <<
1692 while (true) { <<
1693 int old, tmp; <<
1694 <<
1695 old = atomic_read(&nna->nr); <<
1696 if (old >= max) <<
1697 return false; <<
1698 tmp = atomic_cmpxchg_relaxed( <<
1699 if (tmp == old) <<
1700 return true; <<
1701 } <<
1702 } 1112 }
1703 1113
1704 /** !! 1114 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1705 * pwq_tryinc_nr_active - Try to increment nr <<
1706 * @pwq: pool_workqueue of interest <<
1707 * @fill: max_active may have increased, try <<
1708 * <<
1709 * Try to increment nr_active for @pwq. Retur <<
1710 * successfully obtained. %false otherwise. <<
1711 */ <<
1712 static bool pwq_tryinc_nr_active(struct pool_ <<
1713 { 1115 {
1714 struct workqueue_struct *wq = pwq->wq !! 1116 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1715 struct worker_pool *pool = pwq->pool; !! 1117 struct work_struct, entry);
1716 struct wq_node_nr_active *nna = wq_no <<
1717 bool obtained = false; <<
1718 <<
1719 lockdep_assert_held(&pool->lock); <<
1720 <<
1721 if (!nna) { <<
1722 /* BH or per-cpu workqueue, p <<
1723 obtained = pwq->nr_active < R <<
1724 goto out; <<
1725 } <<
1726 <<
1727 if (unlikely(pwq->plugged)) <<
1728 return false; <<
1729 <<
1730 /* <<
1731 * Unbound workqueue uses per-node sh <<
1732 * already waiting on $nna, pwq_dec_n <<
1733 * concurrency level. Don't jump the <<
1734 * <<
1735 * We need to ignore the pending test <<
1736 * pwq_dec_nr_active() can only maint <<
1737 * increase it. This is indicated by <<
1738 */ <<
1739 if (!list_empty(&pwq->pending_node) & <<
1740 goto out; <<
1741 <<
1742 obtained = tryinc_node_nr_active(nna) <<
1743 if (obtained) <<
1744 goto out; <<
1745 <<
1746 /* <<
1747 * Lockless acquisition failed. Lock, <<
1748 * and try again. The smp_mb() is pai <<
1749 * of atomic_dec_return() in pwq_dec_ <<
1750 * we see the decremented $nna->nr or <<
1751 * $nna->pending_pwqs. <<
1752 */ <<
1753 raw_spin_lock(&nna->lock); <<
1754 <<
1755 if (list_empty(&pwq->pending_node)) <<
1756 list_add_tail(&pwq->pending_n <<
1757 else if (likely(!fill)) <<
1758 goto out_unlock; <<
1759 <<
1760 smp_mb(); <<
1761 1118
1762 obtained = tryinc_node_nr_active(nna) !! 1119 pwq_activate_delayed_work(work);
1763 <<
1764 /* <<
1765 * If @fill, @pwq might have already <<
1766 * pending in cold paths doesn't affe <<
1767 */ <<
1768 if (obtained && likely(!fill)) <<
1769 list_del_init(&pwq->pending_n <<
1770 <<
1771 out_unlock: <<
1772 raw_spin_unlock(&nna->lock); <<
1773 out: <<
1774 if (obtained) <<
1775 pwq->nr_active++; <<
1776 return obtained; <<
1777 } <<
1778 <<
1779 /** <<
1780 * pwq_activate_first_inactive - Activate the <<
1781 * @pwq: pool_workqueue of interest <<
1782 * @fill: max_active may have increased, try <<
1783 * <<
1784 * Activate the first inactive work item of @ <<
1785 * max_active limit. <<
1786 * <<
1787 * Returns %true if an inactive work item has <<
1788 * inactive work item is found or max_active <<
1789 */ <<
1790 static bool pwq_activate_first_inactive(struc <<
1791 { <<
1792 struct work_struct *work = <<
1793 list_first_entry_or_null(&pwq <<
1794 stru <<
1795 <<
1796 if (work && pwq_tryinc_nr_active(pwq, <<
1797 __pwq_activate_work(pwq, work <<
1798 return true; <<
1799 } else { <<
1800 return false; <<
1801 } <<
1802 } <<
1803 <<
1804 /** <<
1805 * unplug_oldest_pwq - unplug the oldest pool <<
1806 * @wq: workqueue_struct where its oldest pwq <<
1807 * <<
1808 * This function should only be called for or <<
1809 * oldest pwq is unplugged, the others are pl <<
1810 * ensure proper work item ordering:: <<
1811 * <<
1812 * dfl_pwq --------------+ [P] - plugg <<
1813 * | <<
1814 * v <<
1815 * pwqs -> A -> B [P] -> C [P] (newest) <<
1816 * | | | <<
1817 * 1 3 5 <<
1818 * | | | <<
1819 * 2 4 6 <<
1820 * <<
1821 * When the oldest pwq is drained and removed <<
1822 * to unplug the next oldest one to start its <<
1823 * pwq's are linked into wq->pwqs with the ol <<
1824 * the list is the oldest. <<
1825 */ <<
1826 static void unplug_oldest_pwq(struct workqueu <<
1827 { <<
1828 struct pool_workqueue *pwq; <<
1829 <<
1830 lockdep_assert_held(&wq->mutex); <<
1831 <<
1832 /* Caller should make sure that pwqs <<
1833 pwq = list_first_entry_or_null(&wq->p <<
1834 pwqs_n <<
1835 raw_spin_lock_irq(&pwq->pool->lock); <<
1836 if (pwq->plugged) { <<
1837 pwq->plugged = false; <<
1838 if (pwq_activate_first_inacti <<
1839 kick_pool(pwq->pool); <<
1840 } <<
1841 raw_spin_unlock_irq(&pwq->pool->lock) <<
1842 } <<
1843 <<
1844 /** <<
1845 * node_activate_pending_pwq - Activate a pen <<
1846 * @nna: wq_node_nr_active to activate a pend <<
1847 * @caller_pool: worker_pool the caller is lo <<
1848 * <<
1849 * Activate a pwq in @nna->pending_pwqs. Call <<
1850 * @caller_pool may be unlocked and relocked <<
1851 */ <<
1852 static void node_activate_pending_pwq(struct <<
1853 struct <<
1854 { <<
1855 struct worker_pool *locked_pool = cal <<
1856 struct pool_workqueue *pwq; <<
1857 struct work_struct *work; <<
1858 <<
1859 lockdep_assert_held(&caller_pool->loc <<
1860 <<
1861 raw_spin_lock(&nna->lock); <<
1862 retry: <<
1863 pwq = list_first_entry_or_null(&nna-> <<
1864 struct <<
1865 if (!pwq) <<
1866 goto out_unlock; <<
1867 <<
1868 /* <<
1869 * If @pwq is for a different pool th <<
1870 * @pwq->pool->lock. Let's trylock fi <<
1871 * / lock dance. For that, we also ne <<
1872 * nested inside pool locks. <<
1873 */ <<
1874 if (pwq->pool != locked_pool) { <<
1875 raw_spin_unlock(&locked_pool- <<
1876 locked_pool = pwq->pool; <<
1877 if (!raw_spin_trylock(&locked <<
1878 raw_spin_unlock(&nna- <<
1879 raw_spin_lock(&locked <<
1880 raw_spin_lock(&nna->l <<
1881 goto retry; <<
1882 } <<
1883 } <<
1884 <<
1885 /* <<
1886 * $pwq may not have any inactive wor <<
1887 * Drop it from pending_pwqs and see <<
1888 */ <<
1889 work = list_first_entry_or_null(&pwq- <<
1890 struc <<
1891 if (!work) { <<
1892 list_del_init(&pwq->pending_n <<
1893 goto retry; <<
1894 } <<
1895 <<
1896 /* <<
1897 * Acquire an nr_active count and act <<
1898 * $pwq still has inactive work items <<
1899 * pending_pwqs so that we round-robi <<
1900 * inactive work items are not activa <<
1901 * given that there has never been an <<
1902 */ <<
1903 if (likely(tryinc_node_nr_active(nna) <<
1904 pwq->nr_active++; <<
1905 __pwq_activate_work(pwq, work <<
1906 <<
1907 if (list_empty(&pwq->inactive <<
1908 list_del_init(&pwq->p <<
1909 else <<
1910 list_move_tail(&pwq-> <<
1911 <<
1912 /* if activating a foreign po <<
1913 if (pwq->pool != caller_pool) <<
1914 kick_pool(pwq->pool); <<
1915 } <<
1916 <<
1917 out_unlock: <<
1918 raw_spin_unlock(&nna->lock); <<
1919 if (locked_pool != caller_pool) { <<
1920 raw_spin_unlock(&locked_pool- <<
1921 raw_spin_lock(&caller_pool->l <<
1922 } <<
1923 } <<
1924 <<
1925 /** <<
1926 * pwq_dec_nr_active - Retire an active count <<
1927 * @pwq: pool_workqueue of interest <<
1928 * <<
1929 * Decrement @pwq's nr_active and try to acti <<
1930 * For unbound workqueues, this function may <<
1931 */ <<
1932 static void pwq_dec_nr_active(struct pool_wor <<
1933 { <<
1934 struct worker_pool *pool = pwq->pool; <<
1935 struct wq_node_nr_active *nna = wq_no <<
1936 <<
1937 lockdep_assert_held(&pool->lock); <<
1938 <<
1939 /* <<
1940 * @pwq->nr_active should be decremen <<
1941 * workqueues. <<
1942 */ <<
1943 pwq->nr_active--; <<
1944 <<
1945 /* <<
1946 * For a percpu workqueue, it's simpl <<
1947 * inactive work item on @pwq itself. <<
1948 */ <<
1949 if (!nna) { <<
1950 pwq_activate_first_inactive(p <<
1951 return; <<
1952 } <<
1953 <<
1954 /* <<
1955 * If @pwq is for an unbound workqueu <<
1956 * multiple pwqs and pools may be sha <<
1957 * pwq needs to wait for an nr_active <<
1958 * $nna->pending_pwqs. The following <<
1959 * memory barrier is paired with smp_ <<
1960 * guarantee that either we see non-e <<
1961 * decremented $nna->nr. <<
1962 * <<
1963 * $nna->max may change as CPUs come <<
1964 * max_active gets updated. However, <<
1965 * larger than @pwq->wq->min_active w <<
1966 * This maintains the forward progres <<
1967 */ <<
1968 if (atomic_dec_return(&nna->nr) >= RE <<
1969 return; <<
1970 <<
1971 if (!list_empty(&nna->pending_pwqs)) <<
1972 node_activate_pending_pwq(nna <<
1973 } 1120 }
1974 1121
1975 /** 1122 /**
1976 * pwq_dec_nr_in_flight - decrement pwq's nr_ 1123 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1977 * @pwq: pwq of interest 1124 * @pwq: pwq of interest
1978 * @work_data: work_data of work which left t !! 1125 * @color: color of work which left the queue
1979 * 1126 *
1980 * A work either has completed or is removed 1127 * A work either has completed or is removed from pending queue,
1981 * decrement nr_in_flight of its pwq and hand 1128 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1982 * 1129 *
1983 * NOTE: <<
1984 * For unbound workqueues, this function may <<
1985 * and thus should be called after all other <<
1986 * work item is complete. <<
1987 * <<
1988 * CONTEXT: 1130 * CONTEXT:
1989 * raw_spin_lock_irq(pool->lock). !! 1131 * spin_lock_irq(pool->lock).
1990 */ 1132 */
1991 static void pwq_dec_nr_in_flight(struct pool_ !! 1133 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1992 { 1134 {
1993 int color = get_work_color(work_data) !! 1135 /* uncolored work items don't participate in flushing or nr_active */
1994 !! 1136 if (color == WORK_NO_COLOR)
1995 if (!(work_data & WORK_STRUCT_INACTIV !! 1137 goto out_put;
1996 pwq_dec_nr_active(pwq); <<
1997 1138
1998 pwq->nr_in_flight[color]--; 1139 pwq->nr_in_flight[color]--;
1999 1140
>> 1141 pwq->nr_active--;
>> 1142 if (!list_empty(&pwq->delayed_works)) {
>> 1143 /* one down, submit a delayed one */
>> 1144 if (pwq->nr_active < pwq->max_active)
>> 1145 pwq_activate_first_delayed(pwq);
>> 1146 }
>> 1147
2000 /* is flush in progress and are we at 1148 /* is flush in progress and are we at the flushing tip? */
2001 if (likely(pwq->flush_color != color) 1149 if (likely(pwq->flush_color != color))
2002 goto out_put; 1150 goto out_put;
2003 1151
2004 /* are there still in-flight works? * 1152 /* are there still in-flight works? */
2005 if (pwq->nr_in_flight[color]) 1153 if (pwq->nr_in_flight[color])
2006 goto out_put; 1154 goto out_put;
2007 1155
2008 /* this pwq is done, clear flush_colo 1156 /* this pwq is done, clear flush_color */
2009 pwq->flush_color = -1; 1157 pwq->flush_color = -1;
2010 1158
2011 /* 1159 /*
2012 * If this was the last pwq, wake up 1160 * If this was the last pwq, wake up the first flusher. It
2013 * will handle the rest. 1161 * will handle the rest.
2014 */ 1162 */
2015 if (atomic_dec_and_test(&pwq->wq->nr_ 1163 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
2016 complete(&pwq->wq->first_flus 1164 complete(&pwq->wq->first_flusher->done);
2017 out_put: 1165 out_put:
2018 put_pwq(pwq); 1166 put_pwq(pwq);
2019 } 1167 }
2020 1168
2021 /** 1169 /**
2022 * try_to_grab_pending - steal work item from 1170 * try_to_grab_pending - steal work item from worklist and disable irq
2023 * @work: work item to steal 1171 * @work: work item to steal
2024 * @cflags: %WORK_CANCEL_ flags !! 1172 * @is_dwork: @work is a delayed_work
2025 * @irq_flags: place to store irq state !! 1173 * @flags: place to store irq state
2026 * 1174 *
2027 * Try to grab PENDING bit of @work. This fu 1175 * Try to grab PENDING bit of @work. This function can handle @work in any
2028 * stable state - idle, on timer or on workli !! 1176 * stable state - idle, on timer or on worklist. Return values are
2029 * 1177 *
2030 * Return: <<
2031 * <<
2032 * ======== ============================= <<
2033 * 1 if @work was pending and we s 1178 * 1 if @work was pending and we successfully stole PENDING
2034 * 0 if @work was idle and we clai 1179 * 0 if @work was idle and we claimed PENDING
2035 * -EAGAIN if PENDING couldn't be grabbe 1180 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
2036 * ======== ============================= !! 1181 * -ENOENT if someone else is canceling @work, this state may persist
>> 1182 * for arbitrarily long
2037 * 1183 *
2038 * Note: <<
2039 * On >= 0 return, the caller owns @work's PE 1184 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
2040 * interrupted while holding PENDING and @wor 1185 * interrupted while holding PENDING and @work off queue, irq must be
2041 * disabled on entry. This, combined with de 1186 * disabled on entry. This, combined with delayed_work->timer being
2042 * irqsafe, ensures that we return -EAGAIN fo 1187 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
2043 * 1188 *
2044 * On successful return, >= 0, irq is disable 1189 * On successful return, >= 0, irq is disabled and the caller is
2045 * responsible for releasing it using local_i !! 1190 * responsible for releasing it using local_irq_restore(*@flags).
2046 * 1191 *
2047 * This function is safe to call from any con 1192 * This function is safe to call from any context including IRQ handler.
2048 */ 1193 */
2049 static int try_to_grab_pending(struct work_st !! 1194 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
2050 unsigned long !! 1195 unsigned long *flags)
2051 { 1196 {
2052 struct worker_pool *pool; 1197 struct worker_pool *pool;
2053 struct pool_workqueue *pwq; 1198 struct pool_workqueue *pwq;
2054 1199
2055 local_irq_save(*irq_flags); !! 1200 local_irq_save(*flags);
2056 1201
2057 /* try to steal the timer if it exist 1202 /* try to steal the timer if it exists */
2058 if (cflags & WORK_CANCEL_DELAYED) { !! 1203 if (is_dwork) {
2059 struct delayed_work *dwork = 1204 struct delayed_work *dwork = to_delayed_work(work);
2060 1205
2061 /* 1206 /*
2062 * dwork->timer is irqsafe. 1207 * dwork->timer is irqsafe. If del_timer() fails, it's
2063 * guaranteed that the timer 1208 * guaranteed that the timer is not queued anywhere and not
2064 * running on the local CPU. 1209 * running on the local CPU.
2065 */ 1210 */
2066 if (likely(del_timer(&dwork-> 1211 if (likely(del_timer(&dwork->timer)))
2067 return 1; 1212 return 1;
2068 } 1213 }
2069 1214
2070 /* try to claim PENDING the normal wa 1215 /* try to claim PENDING the normal way */
2071 if (!test_and_set_bit(WORK_STRUCT_PEN 1216 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2072 return 0; 1217 return 0;
2073 1218
2074 rcu_read_lock(); <<
2075 /* 1219 /*
2076 * The queueing is in progress, or it 1220 * The queueing is in progress, or it is already queued. Try to
2077 * steal it from ->worklist without c 1221 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2078 */ 1222 */
2079 pool = get_work_pool(work); 1223 pool = get_work_pool(work);
2080 if (!pool) 1224 if (!pool)
2081 goto fail; 1225 goto fail;
2082 1226
2083 raw_spin_lock(&pool->lock); !! 1227 spin_lock(&pool->lock);
2084 /* 1228 /*
2085 * work->data is guaranteed to point 1229 * work->data is guaranteed to point to pwq only while the work
2086 * item is queued on pwq->wq, and bot 1230 * item is queued on pwq->wq, and both updating work->data to point
2087 * to pwq on queueing and to pool on 1231 * to pwq on queueing and to pool on dequeueing are done under
2088 * pwq->pool->lock. This in turn gua 1232 * pwq->pool->lock. This in turn guarantees that, if work->data
2089 * points to pwq which is associated 1233 * points to pwq which is associated with a locked pool, the work
2090 * item is currently queued on that p 1234 * item is currently queued on that pool.
2091 */ 1235 */
2092 pwq = get_work_pwq(work); 1236 pwq = get_work_pwq(work);
2093 if (pwq && pwq->pool == pool) { 1237 if (pwq && pwq->pool == pool) {
2094 unsigned long work_data = *wo <<
2095 <<
2096 debug_work_deactivate(work); 1238 debug_work_deactivate(work);
2097 1239
2098 /* 1240 /*
2099 * A cancelable inactive work !! 1241 * A delayed work item cannot be grabbed directly because
2100 * pwq->inactive_works since !! 1242 * it might have linked NO_COLOR work items which, if left
2101 * canceled (see the comments !! 1243 * on the delayed_list, will confuse pwq->nr_active
2102 * !! 1244 * management later on and cause stall. Make sure the work
2103 * An inactive work item cann !! 1245 * item is activated before grabbing.
2104 * it might have linked barri <<
2105 * on the inactive_works list <<
2106 * management later on and ca <<
2107 * barrier work items to the <<
2108 * item. Also keep WORK_STRUC <<
2109 * it doesn't participate in <<
2110 * pwq_dec_nr_in_flight(). <<
2111 */ 1246 */
2112 if (work_data & WORK_STRUCT_I !! 1247 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
2113 move_linked_works(wor !! 1248 pwq_activate_delayed_work(work);
2114 1249
2115 list_del_init(&work->entry); 1250 list_del_init(&work->entry);
>> 1251 pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
2116 1252
2117 /* !! 1253 /* work->data points to pwq iff queued, point to pool */
2118 * work->data points to pwq i !! 1254 set_work_pool_and_keep_pending(work, pool->id);
2119 * this destroys work->data n <<
2120 */ <<
2121 set_work_pool_and_keep_pendin <<
2122 <<
2123 <<
2124 /* must be the last step, see <<
2125 pwq_dec_nr_in_flight(pwq, wor <<
2126 1255
2127 raw_spin_unlock(&pool->lock); !! 1256 spin_unlock(&pool->lock);
2128 rcu_read_unlock(); <<
2129 return 1; 1257 return 1;
2130 } 1258 }
2131 raw_spin_unlock(&pool->lock); !! 1259 spin_unlock(&pool->lock);
2132 fail: 1260 fail:
2133 rcu_read_unlock(); !! 1261 local_irq_restore(*flags);
2134 local_irq_restore(*irq_flags); !! 1262 if (work_is_canceling(work))
>> 1263 return -ENOENT;
>> 1264 cpu_relax();
2135 return -EAGAIN; 1265 return -EAGAIN;
2136 } 1266 }
2137 1267
2138 /** 1268 /**
2139 * work_grab_pending - steal work item from w <<
2140 * @work: work item to steal <<
2141 * @cflags: %WORK_CANCEL_ flags <<
2142 * @irq_flags: place to store IRQ state <<
2143 * <<
2144 * Grab PENDING bit of @work. @work can be in <<
2145 * or on worklist. <<
2146 * <<
2147 * Can be called from any context. IRQ is dis <<
2148 * stored in *@irq_flags. The caller is respo <<
2149 * local_irq_restore(). <<
2150 * <<
2151 * Returns %true if @work was pending. %false <<
2152 */ <<
2153 static bool work_grab_pending(struct work_str <<
2154 unsigned long * <<
2155 { <<
2156 int ret; <<
2157 <<
2158 while (true) { <<
2159 ret = try_to_grab_pending(wor <<
2160 if (ret >= 0) <<
2161 return ret; <<
2162 cpu_relax(); <<
2163 } <<
2164 } <<
2165 <<
2166 /** <<
2167 * insert_work - insert a work into a pool 1269 * insert_work - insert a work into a pool
2168 * @pwq: pwq @work belongs to 1270 * @pwq: pwq @work belongs to
2169 * @work: work to insert 1271 * @work: work to insert
2170 * @head: insertion point 1272 * @head: insertion point
2171 * @extra_flags: extra WORK_STRUCT_* flags to 1273 * @extra_flags: extra WORK_STRUCT_* flags to set
2172 * 1274 *
2173 * Insert @work which belongs to @pwq after @ 1275 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
2174 * work_struct flags. 1276 * work_struct flags.
2175 * 1277 *
2176 * CONTEXT: 1278 * CONTEXT:
2177 * raw_spin_lock_irq(pool->lock). !! 1279 * spin_lock_irq(pool->lock).
2178 */ 1280 */
2179 static void insert_work(struct pool_workqueue 1281 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
2180 struct list_head *hea 1282 struct list_head *head, unsigned int extra_flags)
2181 { 1283 {
2182 debug_work_activate(work); !! 1284 struct worker_pool *pool = pwq->pool;
2183 <<
2184 /* record the work call stack in orde <<
2185 kasan_record_aux_stack_noalloc(work); <<
2186 1285
2187 /* we own @work, set data and link */ 1286 /* we own @work, set data and link */
2188 set_work_pwq(work, pwq, extra_flags); 1287 set_work_pwq(work, pwq, extra_flags);
2189 list_add_tail(&work->entry, head); 1288 list_add_tail(&work->entry, head);
2190 get_pwq(pwq); 1289 get_pwq(pwq);
>> 1290
>> 1291 /*
>> 1292 * Ensure either wq_worker_sleeping() sees the above
>> 1293 * list_add_tail() or we see zero nr_running to avoid workers lying
>> 1294 * around lazily while there are works to be processed.
>> 1295 */
>> 1296 smp_mb();
>> 1297
>> 1298 if (__need_more_worker(pool))
>> 1299 wake_up_worker(pool);
2191 } 1300 }
2192 1301
2193 /* 1302 /*
2194 * Test whether @work is being queued from an 1303 * Test whether @work is being queued from another work executing on the
2195 * same workqueue. 1304 * same workqueue.
2196 */ 1305 */
2197 static bool is_chained_work(struct workqueue_ 1306 static bool is_chained_work(struct workqueue_struct *wq)
2198 { 1307 {
2199 struct worker *worker; 1308 struct worker *worker;
2200 1309
2201 worker = current_wq_worker(); 1310 worker = current_wq_worker();
2202 /* 1311 /*
2203 * Return %true iff I'm a worker exec !! 1312 * Return %true iff I'm a worker execuing a work item on @wq. If
2204 * I'm @worker, it's safe to derefere 1313 * I'm @worker, it's safe to dereference it without locking.
2205 */ 1314 */
2206 return worker && worker->current_pwq- 1315 return worker && worker->current_pwq->wq == wq;
2207 } 1316 }
2208 1317
2209 /* <<
2210 * When queueing an unbound work item to a wq <<
2211 * by wq_unbound_cpumask. Otherwise, round r <<
2212 * avoid perturbing sensitive tasks. <<
2213 */ <<
2214 static int wq_select_unbound_cpu(int cpu) <<
2215 { <<
2216 int new_cpu; <<
2217 <<
2218 if (likely(!wq_debug_force_rr_cpu)) { <<
2219 if (cpumask_test_cpu(cpu, wq_ <<
2220 return cpu; <<
2221 } else { <<
2222 pr_warn_once("workqueue: roun <<
2223 } <<
2224 <<
2225 new_cpu = __this_cpu_read(wq_rr_cpu_l <<
2226 new_cpu = cpumask_next_and(new_cpu, w <<
2227 if (unlikely(new_cpu >= nr_cpu_ids)) <<
2228 new_cpu = cpumask_first_and(w <<
2229 if (unlikely(new_cpu >= nr_cp <<
2230 return cpu; <<
2231 } <<
2232 __this_cpu_write(wq_rr_cpu_last, new_ <<
2233 <<
2234 return new_cpu; <<
2235 } <<
2236 <<
2237 static void __queue_work(int cpu, struct work 1318 static void __queue_work(int cpu, struct workqueue_struct *wq,
2238 struct work_struct * 1319 struct work_struct *work)
2239 { 1320 {
2240 struct pool_workqueue *pwq; 1321 struct pool_workqueue *pwq;
2241 struct worker_pool *last_pool, *pool; !! 1322 struct worker_pool *last_pool;
>> 1323 struct list_head *worklist;
2242 unsigned int work_flags; 1324 unsigned int work_flags;
2243 unsigned int req_cpu = cpu; 1325 unsigned int req_cpu = cpu;
2244 1326
2245 /* 1327 /*
2246 * While a work item is PENDING && of 1328 * While a work item is PENDING && off queue, a task trying to
2247 * steal the PENDING will busy-loop w 1329 * steal the PENDING will busy-loop waiting for it to either get
2248 * queued or lose PENDING. Grabbing 1330 * queued or lose PENDING. Grabbing PENDING and queueing should
2249 * happen with IRQ disabled. 1331 * happen with IRQ disabled.
2250 */ 1332 */
2251 lockdep_assert_irqs_disabled(); !! 1333 WARN_ON_ONCE(!irqs_disabled());
2252 1334
2253 /* !! 1335 debug_work_activate(work);
2254 * For a draining wq, only works from !! 1336
2255 * allowed. The __WQ_DESTROYING helps !! 1337 /* if dying, only works from the same workqueue are allowed */
2256 * queues a new work item to a wq aft !! 1338 if (unlikely(wq->flags & __WQ_DRAINING) &&
2257 */ !! 1339 WARN_ON_ONCE(!is_chained_work(wq)))
2258 if (unlikely(wq->flags & (__WQ_DESTRO <<
2259 WARN_ON_ONCE(!is_chained <<
2260 return; 1340 return;
2261 rcu_read_lock(); <<
2262 retry: 1341 retry:
2263 /* pwq which will be used unless @wor !! 1342 if (req_cpu == WORK_CPU_UNBOUND)
2264 if (req_cpu == WORK_CPU_UNBOUND) { !! 1343 cpu = raw_smp_processor_id();
2265 if (wq->flags & WQ_UNBOUND) <<
2266 cpu = wq_select_unbou <<
2267 else <<
2268 cpu = raw_smp_process <<
2269 } <<
2270 1344
2271 pwq = rcu_dereference(*per_cpu_ptr(wq !! 1345 /* pwq which will be used unless @work is executing elsewhere */
2272 pool = pwq->pool; !! 1346 if (!(wq->flags & WQ_UNBOUND))
>> 1347 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
>> 1348 else
>> 1349 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
2273 1350
2274 /* 1351 /*
2275 * If @work was previously on a diffe 1352 * If @work was previously on a different pool, it might still be
2276 * running there, in which case the w 1353 * running there, in which case the work needs to be queued on that
2277 * pool to guarantee non-reentrancy. 1354 * pool to guarantee non-reentrancy.
2278 * <<
2279 * For ordered workqueue, work items <<
2280 * for accurate order management. Gu <<
2281 * non-reentrancy. See the comments <<
2282 */ 1355 */
2283 last_pool = get_work_pool(work); 1356 last_pool = get_work_pool(work);
2284 if (last_pool && last_pool != pool && !! 1357 if (last_pool && last_pool != pwq->pool) {
2285 struct worker *worker; 1358 struct worker *worker;
2286 1359
2287 raw_spin_lock(&last_pool->loc !! 1360 spin_lock(&last_pool->lock);
2288 1361
2289 worker = find_worker_executin 1362 worker = find_worker_executing_work(last_pool, work);
2290 1363
2291 if (worker && worker->current 1364 if (worker && worker->current_pwq->wq == wq) {
2292 pwq = worker->current 1365 pwq = worker->current_pwq;
2293 pool = pwq->pool; <<
2294 WARN_ON_ONCE(pool != <<
2295 } else { 1366 } else {
2296 /* meh... not running 1367 /* meh... not running there, queue here */
2297 raw_spin_unlock(&last !! 1368 spin_unlock(&last_pool->lock);
2298 raw_spin_lock(&pool-> !! 1369 spin_lock(&pwq->pool->lock);
2299 } 1370 }
2300 } else { 1371 } else {
2301 raw_spin_lock(&pool->lock); !! 1372 spin_lock(&pwq->pool->lock);
2302 } 1373 }
2303 1374
2304 /* 1375 /*
2305 * pwq is determined and locked. For !! 1376 * pwq is determined and locked. For unbound pools, we could have
2306 * with pwq release and it could alre !! 1377 * raced with pwq release and it could already be dead. If its
2307 * repeat pwq selection. Note that un !! 1378 * refcnt is zero, repeat pwq selection. Note that pwqs never die
2308 * another pwq replacing it in cpu_pw !! 1379 * without another pwq replacing it in the numa_pwq_tbl or while
2309 * on it, so the retrying is guarante !! 1380 * work items are executing on it, so the retrying is guaranteed to
>> 1381 * make forward-progress.
2310 */ 1382 */
2311 if (unlikely(!pwq->refcnt)) { 1383 if (unlikely(!pwq->refcnt)) {
2312 if (wq->flags & WQ_UNBOUND) { 1384 if (wq->flags & WQ_UNBOUND) {
2313 raw_spin_unlock(&pool !! 1385 spin_unlock(&pwq->pool->lock);
2314 cpu_relax(); 1386 cpu_relax();
2315 goto retry; 1387 goto retry;
2316 } 1388 }
2317 /* oops */ 1389 /* oops */
2318 WARN_ONCE(true, "workqueue: p 1390 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
2319 wq->name, cpu); 1391 wq->name, cpu);
2320 } 1392 }
2321 1393
2322 /* pwq determined, queue */ 1394 /* pwq determined, queue */
2323 trace_workqueue_queue_work(req_cpu, p 1395 trace_workqueue_queue_work(req_cpu, pwq, work);
2324 1396
2325 if (WARN_ON(!list_empty(&work->entry) !! 1397 if (WARN_ON(!list_empty(&work->entry))) {
2326 goto out; !! 1398 spin_unlock(&pwq->pool->lock);
>> 1399 return;
>> 1400 }
2327 1401
2328 pwq->nr_in_flight[pwq->work_color]++; 1402 pwq->nr_in_flight[pwq->work_color]++;
2329 work_flags = work_color_to_flags(pwq- 1403 work_flags = work_color_to_flags(pwq->work_color);
2330 1404
2331 /* !! 1405 if (likely(pwq->nr_active < pwq->max_active)) {
2332 * Limit the number of concurrently a <<
2333 * @work must also queue behind exist <<
2334 * ordering when max_active changes. <<
2335 */ <<
2336 if (list_empty(&pwq->inactive_works) <<
2337 if (list_empty(&pool->worklis <<
2338 pool->watchdog_ts = j <<
2339 <<
2340 trace_workqueue_activate_work 1406 trace_workqueue_activate_work(work);
2341 insert_work(pwq, work, &pool- !! 1407 pwq->nr_active++;
2342 kick_pool(pool); !! 1408 worklist = &pwq->pool->worklist;
2343 } else { 1409 } else {
2344 work_flags |= WORK_STRUCT_INA !! 1410 work_flags |= WORK_STRUCT_DELAYED;
2345 insert_work(pwq, work, &pwq-> !! 1411 worklist = &pwq->delayed_works;
2346 } 1412 }
2347 1413
2348 out: !! 1414 insert_work(pwq, work, worklist, work_flags);
2349 raw_spin_unlock(&pool->lock); <<
2350 rcu_read_unlock(); <<
2351 } <<
2352 <<
2353 static bool clear_pending_if_disabled(struct <<
2354 { <<
2355 unsigned long data = *work_data_bits( <<
2356 struct work_offq_data offqd; <<
2357 1415
2358 if (likely((data & WORK_STRUCT_PWQ) | !! 1416 spin_unlock(&pwq->pool->lock);
2359 !(data & WORK_OFFQ_DISABLE <<
2360 return false; <<
2361 <<
2362 work_offqd_unpack(&offqd, data); <<
2363 set_work_pool_and_clear_pending(work, <<
2364 work_ <<
2365 return true; <<
2366 } 1417 }
2367 1418
2368 /** 1419 /**
2369 * queue_work_on - queue work on specific cpu 1420 * queue_work_on - queue work on specific cpu
2370 * @cpu: CPU number to execute work on 1421 * @cpu: CPU number to execute work on
2371 * @wq: workqueue to use 1422 * @wq: workqueue to use
2372 * @work: work to queue 1423 * @work: work to queue
2373 * 1424 *
2374 * We queue the work to a specific CPU, the c !! 1425 * Returns %false if @work was already on a queue, %true otherwise.
2375 * can't go away. Callers that fail to ensur <<
2376 * CPU cannot go away will execute on a rando <<
2377 * But note well that callers specifying a CP <<
2378 * online will get a splat. <<
2379 * 1426 *
2380 * Return: %false if @work was already on a q !! 1427 * We queue the work to a specific CPU, the caller must ensure it
>> 1428 * can't go away.
2381 */ 1429 */
2382 bool queue_work_on(int cpu, struct workqueue_ 1430 bool queue_work_on(int cpu, struct workqueue_struct *wq,
2383 struct work_struct *work) 1431 struct work_struct *work)
2384 { 1432 {
2385 bool ret = false; 1433 bool ret = false;
2386 unsigned long irq_flags; !! 1434 unsigned long flags;
2387 1435
2388 local_irq_save(irq_flags); !! 1436 local_irq_save(flags);
2389 1437
2390 if (!test_and_set_bit(WORK_STRUCT_PEN !! 1438 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2391 !clear_pending_if_disabled(work)) <<
2392 __queue_work(cpu, wq, work); 1439 __queue_work(cpu, wq, work);
2393 ret = true; 1440 ret = true;
2394 } 1441 }
2395 1442
2396 local_irq_restore(irq_flags); !! 1443 local_irq_restore(flags);
2397 return ret; 1444 return ret;
2398 } 1445 }
2399 EXPORT_SYMBOL(queue_work_on); 1446 EXPORT_SYMBOL(queue_work_on);
2400 1447
2401 /** !! 1448 void delayed_work_timer_fn(unsigned long __data)
2402 * select_numa_node_cpu - Select a CPU based <<
2403 * @node: NUMA node ID that we want to select <<
2404 * <<
2405 * This function will attempt to find a "rand <<
2406 * node. If there are no CPUs available on th <<
2407 * WORK_CPU_UNBOUND indicating that we should <<
2408 * available CPU if we need to schedule this <<
2409 */ <<
2410 static int select_numa_node_cpu(int node) <<
2411 { <<
2412 int cpu; <<
2413 <<
2414 /* Delay binding to CPU if node is no <<
2415 if (node < 0 || node >= MAX_NUMNODES <<
2416 return WORK_CPU_UNBOUND; <<
2417 <<
2418 /* Use local node/cpu if we are alrea <<
2419 cpu = raw_smp_processor_id(); <<
2420 if (node == cpu_to_node(cpu)) <<
2421 return cpu; <<
2422 <<
2423 /* Use "random" otherwise know as "fi <<
2424 cpu = cpumask_any_and(cpumask_of_node <<
2425 <<
2426 /* If CPU is valid return that, other <<
2427 return cpu < nr_cpu_ids ? cpu : WORK_ <<
2428 } <<
2429 <<
2430 /** <<
2431 * queue_work_node - queue work on a "random" <<
2432 * @node: NUMA node that we are targeting the <<
2433 * @wq: workqueue to use <<
2434 * @work: work to queue <<
2435 * <<
2436 * We queue the work to a "random" CPU within <<
2437 * idea here is to provide a way to somehow a <<
2438 * NUMA node. <<
2439 * <<
2440 * This function will only make a best effort <<
2441 * the right NUMA node. If no node is request <<
2442 * offline then we just fall back to standard <<
2443 * <<
2444 * Currently the "random" CPU ends up being t <<
2445 * intersection of cpu_online_mask and the cp <<
2446 * are running on the node. In that case we j <<
2447 * <<
2448 * Return: %false if @work was already on a q <<
2449 */ <<
2450 bool queue_work_node(int node, struct workque <<
2451 struct work_struct *work <<
2452 { <<
2453 unsigned long irq_flags; <<
2454 bool ret = false; <<
2455 <<
2456 /* <<
2457 * This current implementation is spe <<
2458 * Specifically we only return the fi <<
2459 * node instead of cycling through in <<
2460 * <<
2461 * If this is used with a per-cpu wor <<
2462 * workqueue_select_cpu_near would ne <<
2463 * some round robin type logic. <<
2464 */ <<
2465 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND <<
2466 <<
2467 local_irq_save(irq_flags); <<
2468 <<
2469 if (!test_and_set_bit(WORK_STRUCT_PEN <<
2470 !clear_pending_if_disabled(work)) <<
2471 int cpu = select_numa_node_cp <<
2472 <<
2473 __queue_work(cpu, wq, work); <<
2474 ret = true; <<
2475 } <<
2476 <<
2477 local_irq_restore(irq_flags); <<
2478 return ret; <<
2479 } <<
2480 EXPORT_SYMBOL_GPL(queue_work_node); <<
2481 <<
2482 void delayed_work_timer_fn(struct timer_list <<
2483 { 1449 {
2484 struct delayed_work *dwork = from_tim !! 1450 struct delayed_work *dwork = (struct delayed_work *)__data;
2485 1451
2486 /* should have been called from irqsa 1452 /* should have been called from irqsafe timer with irq already off */
2487 __queue_work(dwork->cpu, dwork->wq, & 1453 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2488 } 1454 }
2489 EXPORT_SYMBOL(delayed_work_timer_fn); 1455 EXPORT_SYMBOL(delayed_work_timer_fn);
2490 1456
2491 static void __queue_delayed_work(int cpu, str 1457 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
2492 struct delaye 1458 struct delayed_work *dwork, unsigned long delay)
2493 { 1459 {
2494 struct timer_list *timer = &dwork->ti 1460 struct timer_list *timer = &dwork->timer;
2495 struct work_struct *work = &dwork->wo 1461 struct work_struct *work = &dwork->work;
2496 1462
2497 WARN_ON_ONCE(!wq); !! 1463 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
2498 WARN_ON_ONCE(timer->function != delay !! 1464 timer->data != (unsigned long)dwork);
2499 WARN_ON_ONCE(timer_pending(timer)); 1465 WARN_ON_ONCE(timer_pending(timer));
2500 WARN_ON_ONCE(!list_empty(&work->entry 1466 WARN_ON_ONCE(!list_empty(&work->entry));
2501 1467
2502 /* 1468 /*
2503 * If @delay is 0, queue @dwork->work 1469 * If @delay is 0, queue @dwork->work immediately. This is for
2504 * both optimization and correctness. 1470 * both optimization and correctness. The earliest @timer can
2505 * expire is on the closest next tick 1471 * expire is on the closest next tick and delayed_work users depend
2506 * on that there's no such delay when 1472 * on that there's no such delay when @delay is 0.
2507 */ 1473 */
2508 if (!delay) { 1474 if (!delay) {
2509 __queue_work(cpu, wq, &dwork- 1475 __queue_work(cpu, wq, &dwork->work);
2510 return; 1476 return;
2511 } 1477 }
2512 1478
>> 1479 timer_stats_timer_set_start_info(&dwork->timer);
>> 1480
2513 dwork->wq = wq; 1481 dwork->wq = wq;
2514 dwork->cpu = cpu; 1482 dwork->cpu = cpu;
2515 timer->expires = jiffies + delay; 1483 timer->expires = jiffies + delay;
2516 1484
2517 if (housekeeping_enabled(HK_TYPE_TIME !! 1485 if (unlikely(cpu != WORK_CPU_UNBOUND))
2518 /* If the current cpu is a ho <<
2519 cpu = smp_processor_id(); <<
2520 if (!housekeeping_test_cpu(cp <<
2521 cpu = housekeeping_an <<
2522 add_timer_on(timer, cpu); 1486 add_timer_on(timer, cpu);
2523 } else { !! 1487 else
2524 if (likely(cpu == WORK_CPU_UN !! 1488 add_timer(timer);
2525 add_timer_global(time <<
2526 else <<
2527 add_timer_on(timer, c <<
2528 } <<
2529 } 1489 }
2530 1490
2531 /** 1491 /**
2532 * queue_delayed_work_on - queue work on spec 1492 * queue_delayed_work_on - queue work on specific CPU after delay
2533 * @cpu: CPU number to execute work on 1493 * @cpu: CPU number to execute work on
2534 * @wq: workqueue to use 1494 * @wq: workqueue to use
2535 * @dwork: work to queue 1495 * @dwork: work to queue
2536 * @delay: number of jiffies to wait before q 1496 * @delay: number of jiffies to wait before queueing
2537 * 1497 *
2538 * Return: %false if @work was already on a q !! 1498 * Returns %false if @work was already on a queue, %true otherwise. If
2539 * @delay is zero and @dwork is idle, it will 1499 * @delay is zero and @dwork is idle, it will be scheduled for immediate
2540 * execution. 1500 * execution.
2541 */ 1501 */
2542 bool queue_delayed_work_on(int cpu, struct wo 1502 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
2543 struct delayed_wor 1503 struct delayed_work *dwork, unsigned long delay)
2544 { 1504 {
2545 struct work_struct *work = &dwork->wo 1505 struct work_struct *work = &dwork->work;
2546 bool ret = false; 1506 bool ret = false;
2547 unsigned long irq_flags; !! 1507 unsigned long flags;
2548 1508
2549 /* read the comment in __queue_work() 1509 /* read the comment in __queue_work() */
2550 local_irq_save(irq_flags); !! 1510 local_irq_save(flags);
2551 1511
2552 if (!test_and_set_bit(WORK_STRUCT_PEN !! 1512 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2553 !clear_pending_if_disabled(work)) <<
2554 __queue_delayed_work(cpu, wq, 1513 __queue_delayed_work(cpu, wq, dwork, delay);
2555 ret = true; 1514 ret = true;
2556 } 1515 }
2557 1516
2558 local_irq_restore(irq_flags); !! 1517 local_irq_restore(flags);
2559 return ret; 1518 return ret;
2560 } 1519 }
2561 EXPORT_SYMBOL(queue_delayed_work_on); 1520 EXPORT_SYMBOL(queue_delayed_work_on);
2562 1521
2563 /** 1522 /**
2564 * mod_delayed_work_on - modify delay of or q 1523 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
2565 * @cpu: CPU number to execute work on 1524 * @cpu: CPU number to execute work on
2566 * @wq: workqueue to use 1525 * @wq: workqueue to use
2567 * @dwork: work to queue 1526 * @dwork: work to queue
2568 * @delay: number of jiffies to wait before q 1527 * @delay: number of jiffies to wait before queueing
2569 * 1528 *
2570 * If @dwork is idle, equivalent to queue_del 1529 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
2571 * modify @dwork's timer so that it expires a 1530 * modify @dwork's timer so that it expires after @delay. If @delay is
2572 * zero, @work is guaranteed to be scheduled 1531 * zero, @work is guaranteed to be scheduled immediately regardless of its
2573 * current state. 1532 * current state.
2574 * 1533 *
2575 * Return: %false if @dwork was idle and queu !! 1534 * Returns %false if @dwork was idle and queued, %true if @dwork was
2576 * pending and its timer was modified. 1535 * pending and its timer was modified.
2577 * 1536 *
2578 * This function is safe to call from any con 1537 * This function is safe to call from any context including IRQ handler.
2579 * See try_to_grab_pending() for details. 1538 * See try_to_grab_pending() for details.
2580 */ 1539 */
2581 bool mod_delayed_work_on(int cpu, struct work 1540 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
2582 struct delayed_work 1541 struct delayed_work *dwork, unsigned long delay)
2583 { 1542 {
2584 unsigned long irq_flags; !! 1543 unsigned long flags;
2585 bool ret; !! 1544 int ret;
2586 1545
2587 ret = work_grab_pending(&dwork->work, !! 1546 do {
>> 1547 ret = try_to_grab_pending(&dwork->work, true, &flags);
>> 1548 } while (unlikely(ret == -EAGAIN));
2588 1549
2589 if (!clear_pending_if_disabled(&dwork !! 1550 if (likely(ret >= 0)) {
2590 __queue_delayed_work(cpu, wq, 1551 __queue_delayed_work(cpu, wq, dwork, delay);
>> 1552 local_irq_restore(flags);
>> 1553 }
2591 1554
2592 local_irq_restore(irq_flags); !! 1555 /* -ENOENT from try_to_grab_pending() becomes %true */
2593 return ret; 1556 return ret;
2594 } 1557 }
2595 EXPORT_SYMBOL_GPL(mod_delayed_work_on); 1558 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
2596 1559
2597 static void rcu_work_rcufn(struct rcu_head *r <<
2598 { <<
2599 struct rcu_work *rwork = container_of <<
2600 <<
2601 /* read the comment in __queue_work() <<
2602 local_irq_disable(); <<
2603 __queue_work(WORK_CPU_UNBOUND, rwork- <<
2604 local_irq_enable(); <<
2605 } <<
2606 <<
2607 /** 1560 /**
2608 * queue_rcu_work - queue work after a RCU gr !! 1561 * worker_enter_idle - enter idle state
2609 * @wq: workqueue to use !! 1562 * @worker: worker which is entering idle state
2610 * @rwork: work to queue <<
2611 * 1563 *
2612 * Return: %false if @rwork was already pendi !! 1564 * @worker is entering idle state. Update stats and idle timer if
2613 * that a full RCU grace period is guaranteed !! 1565 * necessary.
2614 * While @rwork is guaranteed to be executed !! 1566 *
2615 * execution may happen before a full RCU gra !! 1567 * LOCKING:
>> 1568 * spin_lock_irq(pool->lock).
2616 */ 1569 */
2617 bool queue_rcu_work(struct workqueue_struct * !! 1570 static void worker_enter_idle(struct worker *worker)
2618 { 1571 {
2619 struct work_struct *work = &rwork->wo !! 1572 struct worker_pool *pool = worker->pool;
2620 1573
2621 /* !! 1574 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
2622 * rcu_work can't be canceled or disa !! 1575 WARN_ON_ONCE(!list_empty(&worker->entry) &&
2623 * inside @rwork and disabled the inn !! 1576 (worker->hentry.next || worker->hentry.pprev)))
2624 */ !! 1577 return;
2625 if (!test_and_set_bit(WORK_STRUCT_PEN <<
2626 !WARN_ON_ONCE(clear_pending_if_di <<
2627 rwork->wq = wq; <<
2628 call_rcu_hurry(&rwork->rcu, r <<
2629 return true; <<
2630 } <<
2631 1578
2632 return false; !! 1579 /* can't use worker_set_flags(), also called from start_worker() */
2633 } !! 1580 worker->flags |= WORKER_IDLE;
2634 EXPORT_SYMBOL(queue_rcu_work); !! 1581 pool->nr_idle++;
>> 1582 worker->last_active = jiffies;
2635 1583
2636 static struct worker *alloc_worker(int node) !! 1584 /* idle_list is LIFO */
2637 { !! 1585 list_add(&worker->entry, &pool->idle_list);
2638 struct worker *worker; <<
2639 1586
2640 worker = kzalloc_node(sizeof(*worker) !! 1587 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
2641 if (worker) { !! 1588 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
2642 INIT_LIST_HEAD(&worker->entry <<
2643 INIT_LIST_HEAD(&worker->sched <<
2644 INIT_LIST_HEAD(&worker->node) <<
2645 /* on creation a worker is in <<
2646 worker->flags = WORKER_PREP; <<
2647 } <<
2648 return worker; <<
2649 } <<
2650 1589
2651 static cpumask_t *pool_allowed_cpus(struct wo !! 1590 /*
2652 { !! 1591 * Sanity check nr_running. Because wq_unbind_fn() releases
2653 if (pool->cpu < 0 && pool->attrs->aff !! 1592 * pool->lock between setting %WORKER_UNBOUND and zapping
2654 return pool->attrs->__pod_cpu !! 1593 * nr_running, the warning may trigger spuriously. Check iff
2655 else !! 1594 * unbind is not in progress.
2656 return pool->attrs->cpumask; !! 1595 */
>> 1596 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
>> 1597 pool->nr_workers == pool->nr_idle &&
>> 1598 atomic_read(&pool->nr_running));
2657 } 1599 }
2658 1600
2659 /** 1601 /**
2660 * worker_attach_to_pool() - attach a worker !! 1602 * worker_leave_idle - leave idle state
2661 * @worker: worker to be attached !! 1603 * @worker: worker which is leaving idle state
2662 * @pool: the target pool !! 1604 *
>> 1605 * @worker is leaving idle state. Update stats.
2663 * 1606 *
2664 * Attach @worker to @pool. Once attached, t !! 1607 * LOCKING:
2665 * cpu-binding of @worker are kept coordinate !! 1608 * spin_lock_irq(pool->lock).
2666 * cpu-[un]hotplugs. <<
2667 */ 1609 */
2668 static void worker_attach_to_pool(struct work !! 1610 static void worker_leave_idle(struct worker *worker)
2669 struct work <<
2670 { <<
2671 mutex_lock(&wq_pool_attach_mutex); <<
2672 <<
2673 /* <<
2674 * The wq_pool_attach_mutex ensures % <<
2675 * across this function. See the comm <<
2676 * details. BH workers are, while per <<
2677 */ <<
2678 if (pool->flags & POOL_DISASSOCIATED) <<
2679 worker->flags |= WORKER_UNBOU <<
2680 } else { <<
2681 WARN_ON_ONCE(pool->flags & PO <<
2682 kthread_set_per_cpu(worker->t <<
2683 } <<
2684 <<
2685 if (worker->rescue_wq) <<
2686 set_cpus_allowed_ptr(worker-> <<
2687 <<
2688 list_add_tail(&worker->node, &pool->w <<
2689 worker->pool = pool; <<
2690 <<
2691 mutex_unlock(&wq_pool_attach_mutex); <<
2692 } <<
2693 <<
2694 static void unbind_worker(struct worker *work <<
2695 { <<
2696 lockdep_assert_held(&wq_pool_attach_m <<
2697 <<
2698 kthread_set_per_cpu(worker->task, -1) <<
2699 if (cpumask_intersects(wq_unbound_cpu <<
2700 WARN_ON_ONCE(set_cpus_allowed <<
2701 else <<
2702 WARN_ON_ONCE(set_cpus_allowed <<
2703 } <<
2704 <<
2705 <<
2706 static void detach_worker(struct worker *work <<
2707 { 1611 {
2708 lockdep_assert_held(&wq_pool_attach_m !! 1612 struct worker_pool *pool = worker->pool;
2709 1613
2710 unbind_worker(worker); !! 1614 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
2711 list_del(&worker->node); !! 1615 return;
>> 1616 worker_clr_flags(worker, WORKER_IDLE);
>> 1617 pool->nr_idle--;
>> 1618 list_del_init(&worker->entry);
2712 } 1619 }
2713 1620
2714 /** 1621 /**
2715 * worker_detach_from_pool() - detach a worke !! 1622 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
2716 * @worker: worker which is attached to its p !! 1623 * @pool: target worker_pool
2717 * 1624 *
2718 * Undo the attaching which had been done in !! 1625 * Bind %current to the cpu of @pool if it is associated and lock @pool.
2719 * caller worker shouldn't access to the pool !! 1626 *
2720 * other reference to the pool. !! 1627 * Works which are scheduled while the cpu is online must at least be
>> 1628 * scheduled to a worker which is bound to the cpu so that if they are
>> 1629 * flushed from cpu callbacks while cpu is going down, they are
>> 1630 * guaranteed to execute on the cpu.
>> 1631 *
>> 1632 * This function is to be used by unbound workers and rescuers to bind
>> 1633 * themselves to the target cpu and may race with cpu going down or
>> 1634 * coming online. kthread_bind() can't be used because it may put the
>> 1635 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
>> 1636 * verbatim as it's best effort and blocking and pool may be
>> 1637 * [dis]associated in the meantime.
>> 1638 *
>> 1639 * This function tries set_cpus_allowed() and locks pool and verifies the
>> 1640 * binding against %POOL_DISASSOCIATED which is set during
>> 1641 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
>> 1642 * enters idle state or fetches works without dropping lock, it can
>> 1643 * guarantee the scheduling requirement described in the first paragraph.
>> 1644 *
>> 1645 * CONTEXT:
>> 1646 * Might sleep. Called without any lock but returns with pool->lock
>> 1647 * held.
>> 1648 *
>> 1649 * RETURNS:
>> 1650 * %true if the associated pool is online (@worker is successfully
>> 1651 * bound), %false if offline.
2721 */ 1652 */
2722 static void worker_detach_from_pool(struct wo !! 1653 static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
>> 1654 __acquires(&pool->lock)
2723 { 1655 {
2724 struct worker_pool *pool = worker->po !! 1656 while (true) {
2725 !! 1657 /*
2726 /* there is one permanent BH worker p !! 1658 * The following call may fail, succeed or succeed
2727 WARN_ON_ONCE(pool->flags & POOL_BH); !! 1659 * without actually migrating the task to the cpu if
>> 1660 * it races with cpu hotunplug operation. Verify
>> 1661 * against POOL_DISASSOCIATED.
>> 1662 */
>> 1663 if (!(pool->flags & POOL_DISASSOCIATED))
>> 1664 set_cpus_allowed_ptr(current, pool->attrs->cpumask);
2728 1665
2729 mutex_lock(&wq_pool_attach_mutex); !! 1666 spin_lock_irq(&pool->lock);
2730 detach_worker(worker); !! 1667 if (pool->flags & POOL_DISASSOCIATED)
2731 worker->pool = NULL; !! 1668 return false;
2732 mutex_unlock(&wq_pool_attach_mutex); !! 1669 if (task_cpu(current) == pool->cpu &&
>> 1670 cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask))
>> 1671 return true;
>> 1672 spin_unlock_irq(&pool->lock);
2733 1673
2734 /* clear leftover flags without pool- !! 1674 /*
2735 worker->flags &= ~(WORKER_UNBOUND | W !! 1675 * We've raced with CPU hot[un]plug. Give it a breather
>> 1676 * and retry migration. cond_resched() is required here;
>> 1677 * otherwise, we might deadlock against cpu_stop trying to
>> 1678 * bring down the CPU on non-preemptive kernel.
>> 1679 */
>> 1680 cpu_relax();
>> 1681 cond_resched();
>> 1682 }
2736 } 1683 }
2737 1684
2738 static int format_worker_id(char *buf, size_t !! 1685 static struct worker *alloc_worker(void)
2739 struct worker_poo <<
2740 { 1686 {
2741 if (worker->rescue_wq) !! 1687 struct worker *worker;
2742 return scnprintf(buf, size, " <<
2743 worker->resc <<
2744 1688
2745 if (pool) { !! 1689 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
2746 if (pool->cpu >= 0) !! 1690 if (worker) {
2747 return scnprintf(buf, !! 1691 INIT_LIST_HEAD(&worker->entry);
2748 pool !! 1692 INIT_LIST_HEAD(&worker->scheduled);
2749 pool !! 1693 /* on creation a worker is in !idle && prep state */
2750 else !! 1694 worker->flags = WORKER_PREP;
2751 return scnprintf(buf, <<
2752 pool <<
2753 } else { <<
2754 return scnprintf(buf, size, " <<
2755 } 1695 }
>> 1696 return worker;
2756 } 1697 }
2757 1698
2758 /** 1699 /**
2759 * create_worker - create a new workqueue wor 1700 * create_worker - create a new workqueue worker
2760 * @pool: pool the new worker will belong to 1701 * @pool: pool the new worker will belong to
2761 * 1702 *
2762 * Create and start a new worker which is att !! 1703 * Create a new worker which is bound to @pool. The returned worker
>> 1704 * can be started by calling start_worker() or destroyed using
>> 1705 * destroy_worker().
2763 * 1706 *
2764 * CONTEXT: 1707 * CONTEXT:
2765 * Might sleep. Does GFP_KERNEL allocations. 1708 * Might sleep. Does GFP_KERNEL allocations.
2766 * 1709 *
2767 * Return: !! 1710 * RETURNS:
2768 * Pointer to the newly created worker. 1711 * Pointer to the newly created worker.
2769 */ 1712 */
2770 static struct worker *create_worker(struct wo 1713 static struct worker *create_worker(struct worker_pool *pool)
2771 { 1714 {
2772 struct worker *worker; !! 1715 struct worker *worker = NULL;
2773 int id; !! 1716 int id = -1;
>> 1717 char id_buf[16];
2774 1718
2775 /* ID is needed to determine kthread !! 1719 lockdep_assert_held(&pool->manager_mutex);
2776 id = ida_alloc(&pool->worker_ida, GFP <<
2777 if (id < 0) { <<
2778 pr_err_once("workqueue: Faile <<
2779 ERR_PTR(id)); <<
2780 return NULL; <<
2781 } <<
2782 1720
2783 worker = alloc_worker(pool->node); !! 1721 /*
2784 if (!worker) { !! 1722 * ID is needed to determine kthread name. Allocate ID first
2785 pr_err_once("workqueue: Faile !! 1723 * without installing the pointer.
2786 goto fail; !! 1724 */
2787 } !! 1725 idr_preload(GFP_KERNEL);
>> 1726 spin_lock_irq(&pool->lock);
2788 1727
2789 worker->id = id; !! 1728 id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT);
2790 1729
2791 if (!(pool->flags & POOL_BH)) { !! 1730 spin_unlock_irq(&pool->lock);
2792 char id_buf[WORKER_ID_LEN]; !! 1731 idr_preload_end();
>> 1732 if (id < 0)
>> 1733 goto fail;
2793 1734
2794 format_worker_id(id_buf, size !! 1735 worker = alloc_worker();
2795 worker->task = kthread_create !! 1736 if (!worker)
2796 !! 1737 goto fail;
2797 if (IS_ERR(worker->task)) { <<
2798 if (PTR_ERR(worker->t <<
2799 pr_err("workq <<
2800 id_buf <<
2801 } else { <<
2802 pr_err_once(" <<
2803 w <<
2804 } <<
2805 goto fail; <<
2806 } <<
2807 1738
2808 set_user_nice(worker->task, p !! 1739 worker->pool = pool;
2809 kthread_bind_mask(worker->tas !! 1740 worker->id = id;
2810 } !! 1741
>> 1742 if (pool->cpu >= 0)
>> 1743 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
>> 1744 pool->attrs->nice < 0 ? "H" : "");
>> 1745 else
>> 1746 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
2811 1747
2812 /* successful, attach the worker to t !! 1748 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
2813 worker_attach_to_pool(worker, pool); !! 1749 "kworker/%s", id_buf);
>> 1750 if (IS_ERR(worker->task))
>> 1751 goto fail;
2814 1752
2815 /* start the newly created worker */ !! 1753 /*
2816 raw_spin_lock_irq(&pool->lock); !! 1754 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
>> 1755 * online CPUs. It'll be re-applied when any of the CPUs come up.
>> 1756 */
>> 1757 set_user_nice(worker->task, pool->attrs->nice);
>> 1758 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
2817 1759
2818 worker->pool->nr_workers++; !! 1760 /* prevent userland from meddling with cpumask of workqueue workers */
2819 worker_enter_idle(worker); !! 1761 worker->task->flags |= PF_NO_SETAFFINITY;
2820 1762
2821 /* 1763 /*
2822 * @worker is waiting on a completion !! 1764 * The caller is responsible for ensuring %POOL_DISASSOCIATED
2823 * check if not woken up soon. As kic !! 1765 * remains stable across this function. See the comments above the
2824 * wake it up explicitly. !! 1766 * flag definition for details.
2825 */ 1767 */
2826 if (worker->task) !! 1768 if (pool->flags & POOL_DISASSOCIATED)
2827 wake_up_process(worker->task) !! 1769 worker->flags |= WORKER_UNBOUND;
2828 1770
2829 raw_spin_unlock_irq(&pool->lock); !! 1771 /* successful, commit the pointer to idr */
>> 1772 spin_lock_irq(&pool->lock);
>> 1773 idr_replace(&pool->worker_idr, worker, worker->id);
>> 1774 spin_unlock_irq(&pool->lock);
2830 1775
2831 return worker; 1776 return worker;
2832 1777
2833 fail: 1778 fail:
2834 ida_free(&pool->worker_ida, id); !! 1779 if (id >= 0) {
>> 1780 spin_lock_irq(&pool->lock);
>> 1781 idr_remove(&pool->worker_idr, id);
>> 1782 spin_unlock_irq(&pool->lock);
>> 1783 }
2835 kfree(worker); 1784 kfree(worker);
2836 return NULL; 1785 return NULL;
2837 } 1786 }
2838 1787
2839 static void detach_dying_workers(struct list_ !! 1788 /**
>> 1789 * start_worker - start a newly created worker
>> 1790 * @worker: worker to start
>> 1791 *
>> 1792 * Make the pool aware of @worker and start it.
>> 1793 *
>> 1794 * CONTEXT:
>> 1795 * spin_lock_irq(pool->lock).
>> 1796 */
>> 1797 static void start_worker(struct worker *worker)
2840 { 1798 {
2841 struct worker *worker; !! 1799 worker->flags |= WORKER_STARTED;
2842 !! 1800 worker->pool->nr_workers++;
2843 list_for_each_entry(worker, cull_list !! 1801 worker_enter_idle(worker);
2844 detach_worker(worker); !! 1802 wake_up_process(worker->task);
2845 } 1803 }
2846 1804
2847 static void reap_dying_workers(struct list_he !! 1805 /**
>> 1806 * create_and_start_worker - create and start a worker for a pool
>> 1807 * @pool: the target pool
>> 1808 *
>> 1809 * Grab the managership of @pool and create and start a new worker for it.
>> 1810 */
>> 1811 static int create_and_start_worker(struct worker_pool *pool)
2848 { 1812 {
2849 struct worker *worker, *tmp; !! 1813 struct worker *worker;
>> 1814
>> 1815 mutex_lock(&pool->manager_mutex);
2850 1816
2851 list_for_each_entry_safe(worker, tmp, !! 1817 worker = create_worker(pool);
2852 list_del_init(&worker->entry) !! 1818 if (worker) {
2853 kthread_stop_put(worker->task !! 1819 spin_lock_irq(&pool->lock);
2854 kfree(worker); !! 1820 start_worker(worker);
>> 1821 spin_unlock_irq(&pool->lock);
2855 } 1822 }
>> 1823
>> 1824 mutex_unlock(&pool->manager_mutex);
>> 1825
>> 1826 return worker ? 0 : -ENOMEM;
2856 } 1827 }
2857 1828
2858 /** 1829 /**
2859 * set_worker_dying - Tag a worker for destru !! 1830 * destroy_worker - destroy a workqueue worker
2860 * @worker: worker to be destroyed 1831 * @worker: worker to be destroyed
2861 * @list: transfer worker away from its pool- <<
2862 * 1832 *
2863 * Tag @worker for destruction and adjust @po !! 1833 * Destroy @worker and adjust @pool stats accordingly.
2864 * should be idle. <<
2865 * 1834 *
2866 * CONTEXT: 1835 * CONTEXT:
2867 * raw_spin_lock_irq(pool->lock). !! 1836 * spin_lock_irq(pool->lock) which is released and regrabbed.
2868 */ 1837 */
2869 static void set_worker_dying(struct worker *w !! 1838 static void destroy_worker(struct worker *worker)
2870 { 1839 {
2871 struct worker_pool *pool = worker->po 1840 struct worker_pool *pool = worker->pool;
2872 1841
>> 1842 lockdep_assert_held(&pool->manager_mutex);
2873 lockdep_assert_held(&pool->lock); 1843 lockdep_assert_held(&pool->lock);
2874 lockdep_assert_held(&wq_pool_attach_m <<
2875 1844
2876 /* sanity check frenzy */ 1845 /* sanity check frenzy */
2877 if (WARN_ON(worker->current_work) || 1846 if (WARN_ON(worker->current_work) ||
2878 WARN_ON(!list_empty(&worker->sche !! 1847 WARN_ON(!list_empty(&worker->scheduled)))
2879 WARN_ON(!(worker->flags & WORKER_ <<
2880 return; 1848 return;
2881 1849
2882 pool->nr_workers--; !! 1850 if (worker->flags & WORKER_STARTED)
2883 pool->nr_idle--; !! 1851 pool->nr_workers--;
>> 1852 if (worker->flags & WORKER_IDLE)
>> 1853 pool->nr_idle--;
2884 1854
2885 worker->flags |= WORKER_DIE; !! 1855 /*
2886 !! 1856 * Once WORKER_DIE is set, the kworker may destroy itself at any
2887 list_move(&worker->entry, list); !! 1857 * point. Pin to ensure the task stays until we're done with it.
2888 !! 1858 */
2889 /* get an extra task struct reference <<
2890 get_task_struct(worker->task); 1859 get_task_struct(worker->task);
2891 } <<
2892 <<
2893 /** <<
2894 * idle_worker_timeout - check if some idle w <<
2895 * @t: The pool's idle_timer that just expire <<
2896 * <<
2897 * The timer is armed in worker_enter_idle(). <<
2898 * worker_leave_idle(), as a worker flicking <<
2899 * pool is at the too_many_workers() tipping <<
2900 * housekeeping overhead. Since IDLE_WORKER_T <<
2901 * it expire and re-evaluate things from ther <<
2902 */ <<
2903 static void idle_worker_timeout(struct timer_ <<
2904 { <<
2905 struct worker_pool *pool = from_timer <<
2906 bool do_cull = false; <<
2907 1860
2908 if (work_pending(&pool->idle_cull_wor !! 1861 list_del_init(&worker->entry);
2909 return; !! 1862 worker->flags |= WORKER_DIE;
2910 <<
2911 raw_spin_lock_irq(&pool->lock); <<
2912 1863
2913 if (too_many_workers(pool)) { !! 1864 idr_remove(&pool->worker_idr, worker->id);
2914 struct worker *worker; <<
2915 unsigned long expires; <<
2916 1865
2917 /* idle_list is kept in LIFO !! 1866 spin_unlock_irq(&pool->lock);
2918 worker = list_last_entry(&poo <<
2919 expires = worker->last_active <<
2920 do_cull = !time_before(jiffie <<
2921 1867
2922 if (!do_cull) !! 1868 kthread_stop(worker->task);
2923 mod_timer(&pool->idle !! 1869 put_task_struct(worker->task);
2924 } !! 1870 kfree(worker);
2925 raw_spin_unlock_irq(&pool->lock); <<
2926 1871
2927 if (do_cull) !! 1872 spin_lock_irq(&pool->lock);
2928 queue_work(system_unbound_wq, <<
2929 } 1873 }
2930 1874
2931 /** !! 1875 static void idle_worker_timeout(unsigned long __pool)
2932 * idle_cull_fn - cull workers that have been <<
2933 * @work: the pool's work for handling these <<
2934 * <<
2935 * This goes through a pool's idle workers an <<
2936 * idle for at least IDLE_WORKER_TIMEOUT seco <<
2937 * <<
2938 * We don't want to disturb isolated CPUs bec <<
2939 * culled, so this also resets worker affinit <<
2940 * context, hence the split between timer cal <<
2941 */ <<
2942 static void idle_cull_fn(struct work_struct * <<
2943 { 1876 {
2944 struct worker_pool *pool = container_ !! 1877 struct worker_pool *pool = (void *)__pool;
2945 LIST_HEAD(cull_list); <<
2946 1878
2947 /* !! 1879 spin_lock_irq(&pool->lock);
2948 * Grabbing wq_pool_attach_mutex here <<
2949 * cannot proceed beyong set_pf_worke <<
2950 * This is required as a previously-p <<
2951 * set_worker_dying() has happened bu <<
2952 */ <<
2953 mutex_lock(&wq_pool_attach_mutex); <<
2954 raw_spin_lock_irq(&pool->lock); <<
2955 1880
2956 while (too_many_workers(pool)) { !! 1881 if (too_many_workers(pool)) {
2957 struct worker *worker; 1882 struct worker *worker;
2958 unsigned long expires; 1883 unsigned long expires;
2959 1884
2960 worker = list_last_entry(&poo !! 1885 /* idle_list is kept in LIFO order, check the last one */
>> 1886 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2961 expires = worker->last_active 1887 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2962 1888
2963 if (time_before(jiffies, expi !! 1889 if (time_before(jiffies, expires))
2964 mod_timer(&pool->idle 1890 mod_timer(&pool->idle_timer, expires);
2965 break; !! 1891 else {
>> 1892 /* it's been idle for too long, wake up manager */
>> 1893 pool->flags |= POOL_MANAGE_WORKERS;
>> 1894 wake_up_worker(pool);
2966 } 1895 }
2967 <<
2968 set_worker_dying(worker, &cul <<
2969 } 1896 }
2970 1897
2971 raw_spin_unlock_irq(&pool->lock); !! 1898 spin_unlock_irq(&pool->lock);
2972 detach_dying_workers(&cull_list); <<
2973 mutex_unlock(&wq_pool_attach_mutex); <<
2974 <<
2975 reap_dying_workers(&cull_list); <<
2976 } 1899 }
2977 1900
2978 static void send_mayday(struct work_struct *w 1901 static void send_mayday(struct work_struct *work)
2979 { 1902 {
2980 struct pool_workqueue *pwq = get_work 1903 struct pool_workqueue *pwq = get_work_pwq(work);
2981 struct workqueue_struct *wq = pwq->wq 1904 struct workqueue_struct *wq = pwq->wq;
2982 1905
2983 lockdep_assert_held(&wq_mayday_lock); 1906 lockdep_assert_held(&wq_mayday_lock);
2984 1907
2985 if (!wq->rescuer) 1908 if (!wq->rescuer)
2986 return; 1909 return;
2987 1910
2988 /* mayday mayday mayday */ 1911 /* mayday mayday mayday */
2989 if (list_empty(&pwq->mayday_node)) { 1912 if (list_empty(&pwq->mayday_node)) {
2990 /* 1913 /*
2991 * If @pwq is for an unbound 1914 * If @pwq is for an unbound wq, its base ref may be put at
2992 * any time due to an attribu 1915 * any time due to an attribute change. Pin @pwq until the
2993 * rescuer is done with it. 1916 * rescuer is done with it.
2994 */ 1917 */
2995 get_pwq(pwq); 1918 get_pwq(pwq);
2996 list_add_tail(&pwq->mayday_no 1919 list_add_tail(&pwq->mayday_node, &wq->maydays);
2997 wake_up_process(wq->rescuer-> 1920 wake_up_process(wq->rescuer->task);
2998 pwq->stats[PWQ_STAT_MAYDAY]++ <<
2999 } 1921 }
3000 } 1922 }
3001 1923
3002 static void pool_mayday_timeout(struct timer_ !! 1924 static void pool_mayday_timeout(unsigned long __pool)
3003 { 1925 {
3004 struct worker_pool *pool = from_timer !! 1926 struct worker_pool *pool = (void *)__pool;
3005 struct work_struct *work; 1927 struct work_struct *work;
3006 1928
3007 raw_spin_lock_irq(&pool->lock); !! 1929 spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */
3008 raw_spin_lock(&wq_mayday_lock); !! 1930 spin_lock(&pool->lock);
3009 1931
3010 if (need_to_create_worker(pool)) { 1932 if (need_to_create_worker(pool)) {
3011 /* 1933 /*
3012 * We've been trying to creat 1934 * We've been trying to create a new worker but
3013 * haven't been successful. 1935 * haven't been successful. We might be hitting an
3014 * allocation deadlock. Send 1936 * allocation deadlock. Send distress signals to
3015 * rescuers. 1937 * rescuers.
3016 */ 1938 */
3017 list_for_each_entry(work, &po 1939 list_for_each_entry(work, &pool->worklist, entry)
3018 send_mayday(work); 1940 send_mayday(work);
3019 } 1941 }
3020 1942
3021 raw_spin_unlock(&wq_mayday_lock); !! 1943 spin_unlock(&pool->lock);
3022 raw_spin_unlock_irq(&pool->lock); !! 1944 spin_unlock_irq(&wq_mayday_lock);
3023 1945
3024 mod_timer(&pool->mayday_timer, jiffie 1946 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
3025 } 1947 }
3026 1948
3027 /** 1949 /**
3028 * maybe_create_worker - create a new worker 1950 * maybe_create_worker - create a new worker if necessary
3029 * @pool: pool to create a new worker for 1951 * @pool: pool to create a new worker for
3030 * 1952 *
3031 * Create a new worker for @pool if necessary 1953 * Create a new worker for @pool if necessary. @pool is guaranteed to
3032 * have at least one idle worker on return fr 1954 * have at least one idle worker on return from this function. If
3033 * creating a new worker takes longer than MA 1955 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
3034 * sent to all rescuers with works scheduled 1956 * sent to all rescuers with works scheduled on @pool to resolve
3035 * possible allocation deadlock. 1957 * possible allocation deadlock.
3036 * 1958 *
3037 * On return, need_to_create_worker() is guar 1959 * On return, need_to_create_worker() is guaranteed to be %false and
3038 * may_start_working() %true. 1960 * may_start_working() %true.
3039 * 1961 *
3040 * LOCKING: 1962 * LOCKING:
3041 * raw_spin_lock_irq(pool->lock) which may be !! 1963 * spin_lock_irq(pool->lock) which may be released and regrabbed
3042 * multiple times. Does GFP_KERNEL allocatio 1964 * multiple times. Does GFP_KERNEL allocations. Called only from
3043 * manager. 1965 * manager.
3044 */ 1966 */
3045 static void maybe_create_worker(struct worker 1967 static void maybe_create_worker(struct worker_pool *pool)
3046 __releases(&pool->lock) 1968 __releases(&pool->lock)
3047 __acquires(&pool->lock) 1969 __acquires(&pool->lock)
3048 { 1970 {
>> 1971 if (!need_to_create_worker(pool))
>> 1972 return;
3049 restart: 1973 restart:
3050 raw_spin_unlock_irq(&pool->lock); !! 1974 spin_unlock_irq(&pool->lock);
3051 1975
3052 /* if we don't make progress in MAYDA 1976 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
3053 mod_timer(&pool->mayday_timer, jiffie 1977 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
3054 1978
3055 while (true) { 1979 while (true) {
3056 if (create_worker(pool) || !n !! 1980 struct worker *worker;
>> 1981
>> 1982 worker = create_worker(pool);
>> 1983 if (worker) {
>> 1984 del_timer_sync(&pool->mayday_timer);
>> 1985 spin_lock_irq(&pool->lock);
>> 1986 start_worker(worker);
>> 1987 if (WARN_ON_ONCE(need_to_create_worker(pool)))
>> 1988 goto restart;
>> 1989 return;
>> 1990 }
>> 1991
>> 1992 if (!need_to_create_worker(pool))
3057 break; 1993 break;
3058 1994
3059 schedule_timeout_interruptibl !! 1995 __set_current_state(TASK_INTERRUPTIBLE);
>> 1996 schedule_timeout(CREATE_COOLDOWN);
3060 1997
3061 if (!need_to_create_worker(po 1998 if (!need_to_create_worker(pool))
3062 break; 1999 break;
3063 } 2000 }
3064 2001
3065 del_timer_sync(&pool->mayday_timer); 2002 del_timer_sync(&pool->mayday_timer);
3066 raw_spin_lock_irq(&pool->lock); !! 2003 spin_lock_irq(&pool->lock);
3067 /* <<
3068 * This is necessary even after a new <<
3069 * created as @pool->lock was dropped <<
3070 * already become busy. <<
3071 */ <<
3072 if (need_to_create_worker(pool)) 2004 if (need_to_create_worker(pool))
3073 goto restart; 2005 goto restart;
>> 2006 return;
>> 2007 }
>> 2008
>> 2009 /**
>> 2010 * maybe_destroy_worker - destroy workers which have been idle for a while
>> 2011 * @pool: pool to destroy workers for
>> 2012 *
>> 2013 * Destroy @pool workers which have been idle for longer than
>> 2014 * IDLE_WORKER_TIMEOUT.
>> 2015 *
>> 2016 * LOCKING:
>> 2017 * spin_lock_irq(pool->lock) which may be released and regrabbed
>> 2018 * multiple times. Called only from manager.
>> 2019 */
>> 2020 static void maybe_destroy_workers(struct worker_pool *pool)
>> 2021 {
>> 2022 while (too_many_workers(pool)) {
>> 2023 struct worker *worker;
>> 2024 unsigned long expires;
>> 2025
>> 2026 worker = list_entry(pool->idle_list.prev, struct worker, entry);
>> 2027 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
>> 2028
>> 2029 if (time_before(jiffies, expires)) {
>> 2030 mod_timer(&pool->idle_timer, expires);
>> 2031 break;
>> 2032 }
>> 2033
>> 2034 destroy_worker(worker);
>> 2035 }
3074 } 2036 }
3075 2037
3076 /** 2038 /**
3077 * manage_workers - manage worker pool 2039 * manage_workers - manage worker pool
3078 * @worker: self 2040 * @worker: self
3079 * 2041 *
3080 * Assume the manager role and manage the wor 2042 * Assume the manager role and manage the worker pool @worker belongs
3081 * to. At any given time, there can be only 2043 * to. At any given time, there can be only zero or one manager per
3082 * pool. The exclusion is handled automatica 2044 * pool. The exclusion is handled automatically by this function.
3083 * 2045 *
3084 * The caller can safely start processing wor 2046 * The caller can safely start processing works on false return. On
3085 * true return, it's guaranteed that need_to_ 2047 * true return, it's guaranteed that need_to_create_worker() is false
3086 * and may_start_working() is true. 2048 * and may_start_working() is true.
3087 * 2049 *
3088 * CONTEXT: 2050 * CONTEXT:
3089 * raw_spin_lock_irq(pool->lock) which may be !! 2051 * spin_lock_irq(pool->lock) which may be released and regrabbed
3090 * multiple times. Does GFP_KERNEL allocatio 2052 * multiple times. Does GFP_KERNEL allocations.
3091 * 2053 *
3092 * Return: !! 2054 * RETURNS:
3093 * %false if the pool doesn't need management 2055 * %false if the pool doesn't need management and the caller can safely
3094 * start processing works, %true if managemen 2056 * start processing works, %true if management function was performed and
3095 * the conditions that the caller verified be 2057 * the conditions that the caller verified before calling the function may
3096 * no longer be true. 2058 * no longer be true.
3097 */ 2059 */
3098 static bool manage_workers(struct worker *wor 2060 static bool manage_workers(struct worker *worker)
3099 { 2061 {
3100 struct worker_pool *pool = worker->po 2062 struct worker_pool *pool = worker->pool;
3101 2063
3102 if (pool->flags & POOL_MANAGER_ACTIVE !! 2064 /*
>> 2065 * Managership is governed by two mutexes - manager_arb and
>> 2066 * manager_mutex. manager_arb handles arbitration of manager role.
>> 2067 * Anyone who successfully grabs manager_arb wins the arbitration
>> 2068 * and becomes the manager. mutex_trylock() on pool->manager_arb
>> 2069 * failure while holding pool->lock reliably indicates that someone
>> 2070 * else is managing the pool and the worker which failed trylock
>> 2071 * can proceed to executing work items. This means that anyone
>> 2072 * grabbing manager_arb is responsible for actually performing
>> 2073 * manager duties. If manager_arb is grabbed and released without
>> 2074 * actual management, the pool may stall indefinitely.
>> 2075 *
>> 2076 * manager_mutex is used for exclusion of actual management
>> 2077 * operations. The holder of manager_mutex can be sure that none
>> 2078 * of management operations, including creation and destruction of
>> 2079 * workers, won't take place until the mutex is released. Because
>> 2080 * manager_mutex doesn't interfere with manager role arbitration,
>> 2081 * it is guaranteed that the pool's management, while may be
>> 2082 * delayed, won't be disturbed by someone else grabbing
>> 2083 * manager_mutex.
>> 2084 */
>> 2085 if (!mutex_trylock(&pool->manager_arb))
3103 return false; 2086 return false;
3104 2087
3105 pool->flags |= POOL_MANAGER_ACTIVE; !! 2088 /*
3106 pool->manager = worker; !! 2089 * With manager arbitration won, manager_mutex would be free in
>> 2090 * most cases. trylock first without dropping @pool->lock.
>> 2091 */
>> 2092 if (unlikely(!mutex_trylock(&pool->manager_mutex))) {
>> 2093 spin_unlock_irq(&pool->lock);
>> 2094 mutex_lock(&pool->manager_mutex);
>> 2095 spin_lock_irq(&pool->lock);
>> 2096 }
>> 2097
>> 2098 pool->flags &= ~POOL_MANAGE_WORKERS;
3107 2099
>> 2100 /*
>> 2101 * Destroy and then create so that may_start_working() is true
>> 2102 * on return.
>> 2103 */
>> 2104 maybe_destroy_workers(pool);
3108 maybe_create_worker(pool); 2105 maybe_create_worker(pool);
3109 2106
3110 pool->manager = NULL; !! 2107 mutex_unlock(&pool->manager_mutex);
3111 pool->flags &= ~POOL_MANAGER_ACTIVE; !! 2108 mutex_unlock(&pool->manager_arb);
3112 rcuwait_wake_up(&manager_wait); <<
3113 return true; 2109 return true;
3114 } 2110 }
3115 2111
3116 /** 2112 /**
3117 * process_one_work - process single work 2113 * process_one_work - process single work
3118 * @worker: self 2114 * @worker: self
3119 * @work: work to process 2115 * @work: work to process
3120 * 2116 *
3121 * Process @work. This function contains all 2117 * Process @work. This function contains all the logics necessary to
3122 * process a single work including synchroniz 2118 * process a single work including synchronization against and
3123 * interaction with other workers on the same 2119 * interaction with other workers on the same cpu, queueing and
3124 * flushing. As long as context requirement 2120 * flushing. As long as context requirement is met, any worker can
3125 * call this function to process a work. 2121 * call this function to process a work.
3126 * 2122 *
3127 * CONTEXT: 2123 * CONTEXT:
3128 * raw_spin_lock_irq(pool->lock) which is rel !! 2124 * spin_lock_irq(pool->lock) which is released and regrabbed.
3129 */ 2125 */
3130 static void process_one_work(struct worker *w 2126 static void process_one_work(struct worker *worker, struct work_struct *work)
3131 __releases(&pool->lock) 2127 __releases(&pool->lock)
3132 __acquires(&pool->lock) 2128 __acquires(&pool->lock)
3133 { 2129 {
3134 struct pool_workqueue *pwq = get_work 2130 struct pool_workqueue *pwq = get_work_pwq(work);
3135 struct worker_pool *pool = worker->po 2131 struct worker_pool *pool = worker->pool;
3136 unsigned long work_data; !! 2132 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
3137 int lockdep_start_depth, rcu_start_de !! 2133 int work_color;
3138 bool bh_draining = pool->flags & POOL !! 2134 struct worker *collision;
3139 #ifdef CONFIG_LOCKDEP 2135 #ifdef CONFIG_LOCKDEP
3140 /* 2136 /*
3141 * It is permissible to free the stru 2137 * It is permissible to free the struct work_struct from
3142 * inside the function that is called 2138 * inside the function that is called from it, this we need to
3143 * take into account for lockdep too. 2139 * take into account for lockdep too. To avoid bogus "held
3144 * lock freed" warnings as well as pr 2140 * lock freed" warnings as well as problems when looking into
3145 * work->lockdep_map, make a copy and 2141 * work->lockdep_map, make a copy and use that here.
3146 */ 2142 */
3147 struct lockdep_map lockdep_map; 2143 struct lockdep_map lockdep_map;
3148 2144
3149 lockdep_copy_map(&lockdep_map, &work- 2145 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
3150 #endif 2146 #endif
3151 /* ensure we're on the correct CPU */ !! 2147 /*
3152 WARN_ON_ONCE(!(pool->flags & POOL_DIS !! 2148 * Ensure we're on the correct CPU. DISASSOCIATED test is
>> 2149 * necessary to avoid spurious warnings from rescuers servicing the
>> 2150 * unbound or a disassociated pool.
>> 2151 */
>> 2152 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
>> 2153 !(pool->flags & POOL_DISASSOCIATED) &&
3153 raw_smp_processor_id() ! 2154 raw_smp_processor_id() != pool->cpu);
3154 2155
>> 2156 /*
>> 2157 * A single work shouldn't be executed concurrently by
>> 2158 * multiple workers on a single cpu. Check whether anyone is
>> 2159 * already processing the work. If so, defer the work to the
>> 2160 * currently executing one.
>> 2161 */
>> 2162 collision = find_worker_executing_work(pool, work);
>> 2163 if (unlikely(collision)) {
>> 2164 move_linked_works(work, &collision->scheduled, NULL);
>> 2165 return;
>> 2166 }
>> 2167
3155 /* claim and dequeue */ 2168 /* claim and dequeue */
3156 debug_work_deactivate(work); 2169 debug_work_deactivate(work);
3157 hash_add(pool->busy_hash, &worker->he 2170 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
3158 worker->current_work = work; 2171 worker->current_work = work;
3159 worker->current_func = work->func; 2172 worker->current_func = work->func;
3160 worker->current_pwq = pwq; 2173 worker->current_pwq = pwq;
3161 if (worker->task) !! 2174 work_color = get_work_color(work);
3162 worker->current_at = worker-> <<
3163 work_data = *work_data_bits(work); <<
3164 worker->current_color = get_work_colo <<
3165 <<
3166 /* <<
3167 * Record wq name for cmdline and deb <<
3168 * overridden through set_worker_desc <<
3169 */ <<
3170 strscpy(worker->desc, pwq->wq->name, <<
3171 2175
3172 list_del_init(&work->entry); 2176 list_del_init(&work->entry);
3173 2177
3174 /* 2178 /*
3175 * CPU intensive works don't particip !! 2179 * CPU intensive works don't participate in concurrency
3176 * They're the scheduler's responsibi !! 2180 * management. They're the scheduler's responsibility.
3177 * of concurrency management and the <<
3178 * execution of the pending work item <<
3179 */ 2181 */
3180 if (unlikely(pwq->wq->flags & WQ_CPU_ !! 2182 if (unlikely(cpu_intensive))
3181 worker_set_flags(worker, WORK !! 2183 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
3182 2184
3183 /* 2185 /*
3184 * Kick @pool if necessary. It's alwa !! 2186 * Unbound pool isn't concurrency managed and work items should be
3185 * since nr_running would always be > !! 2187 * executed ASAP. Wake up another worker if necessary.
3186 * chain execution of the pending wor <<
3187 * workers such as the UNBOUND and CP <<
3188 */ 2188 */
3189 kick_pool(pool); !! 2189 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
>> 2190 wake_up_worker(pool);
3190 2191
3191 /* 2192 /*
3192 * Record the last pool and clear PEN 2193 * Record the last pool and clear PENDING which should be the last
3193 * update to @work. Also, do this in 2194 * update to @work. Also, do this inside @pool->lock so that
3194 * PENDING and queued state changes h 2195 * PENDING and queued state changes happen together while IRQ is
3195 * disabled. 2196 * disabled.
3196 */ 2197 */
3197 set_work_pool_and_clear_pending(work, !! 2198 set_work_pool_and_clear_pending(work, pool->id);
3198 2199
3199 pwq->stats[PWQ_STAT_STARTED]++; !! 2200 spin_unlock_irq(&pool->lock);
3200 raw_spin_unlock_irq(&pool->lock); <<
3201 2201
3202 rcu_start_depth = rcu_preempt_depth() !! 2202 lock_map_acquire_read(&pwq->wq->lockdep_map);
3203 lockdep_start_depth = lockdep_depth(c <<
3204 /* see drain_dead_softirq_workfn() */ <<
3205 if (!bh_draining) <<
3206 lock_map_acquire(&pwq->wq->lo <<
3207 lock_map_acquire(&lockdep_map); 2203 lock_map_acquire(&lockdep_map);
3208 /* <<
3209 * Strictly speaking we should mark t <<
3210 * any locks, that is, before these t <<
3211 * <<
3212 * However, that would result in: <<
3213 * <<
3214 * A(W1) <<
3215 * WFC(C) <<
3216 * A(W1) <<
3217 * C(C) <<
3218 * <<
3219 * Which would create W1->C->W1 depen <<
3220 * actual deadlock possible. There ar <<
3221 * read-recursive acquire on the work <<
3222 * hit the lockdep limitation on recu <<
3223 * these locks. <<
3224 * <<
3225 * AFAICT there is no possible deadlo <<
3226 * flush_work() and complete() primit <<
3227 * workqueues), so hiding them isn't <<
3228 */ <<
3229 lockdep_invariant_state(true); <<
3230 trace_workqueue_execute_start(work); 2204 trace_workqueue_execute_start(work);
3231 worker->current_func(work); 2205 worker->current_func(work);
3232 /* 2206 /*
3233 * While we must be careful to not us 2207 * While we must be careful to not use "work" after this, the trace
3234 * point will only record its address 2208 * point will only record its address.
3235 */ 2209 */
3236 trace_workqueue_execute_end(work, wor !! 2210 trace_workqueue_execute_end(work);
3237 pwq->stats[PWQ_STAT_COMPLETED]++; <<
3238 lock_map_release(&lockdep_map); 2211 lock_map_release(&lockdep_map);
3239 if (!bh_draining) !! 2212 lock_map_release(&pwq->wq->lockdep_map);
3240 lock_map_release(&pwq->wq->lo <<
3241 2213
3242 if (unlikely((worker->task && in_atom !! 2214 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
3243 lockdep_depth(current) ! !! 2215 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
3244 rcu_preempt_depth() != r !! 2216 " last function: %pf\n",
3245 pr_err("BUG: workqueue leaked !! 2217 current->comm, preempt_count(), task_pid_nr(current),
3246 " preempt=0x%08x l <<
3247 current->comm, task_pi <<
3248 lockdep_start_depth, l <<
3249 rcu_start_depth, rcu_p <<
3250 worker->current_func); 2218 worker->current_func);
3251 debug_show_held_locks(current 2219 debug_show_held_locks(current);
3252 dump_stack(); 2220 dump_stack();
3253 } 2221 }
3254 2222
3255 /* 2223 /*
3256 * The following prevents a kworker f !! 2224 * The following prevents a kworker from hogging CPU on !PREEMPT
3257 * kernels, where a requeueing work i 2225 * kernels, where a requeueing work item waiting for something to
3258 * happen could deadlock with stop_ma 2226 * happen could deadlock with stop_machine as such work item could
3259 * indefinitely requeue itself while 2227 * indefinitely requeue itself while all other CPUs are trapped in
3260 * stop_machine. At the same time, re !! 2228 * stop_machine.
3261 * the same condition doesn't freeze <<
3262 */ 2229 */
3263 if (worker->task) !! 2230 cond_resched();
3264 cond_resched(); <<
3265 <<
3266 raw_spin_lock_irq(&pool->lock); <<
3267 2231
3268 /* !! 2232 spin_lock_irq(&pool->lock);
3269 * In addition to %WQ_CPU_INTENSIVE, <<
3270 * CPU intensive by wq_worker_tick() <<
3271 * wq_cpu_intensive_thresh_us. Clear <<
3272 */ <<
3273 worker_clr_flags(worker, WORKER_CPU_I <<
3274 2233
3275 /* tag the worker for identification !! 2234 /* clear cpu intensive status */
3276 worker->last_func = worker->current_f !! 2235 if (unlikely(cpu_intensive))
>> 2236 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
3277 2237
3278 /* we're done with it, release */ 2238 /* we're done with it, release */
3279 hash_del(&worker->hentry); 2239 hash_del(&worker->hentry);
3280 worker->current_work = NULL; 2240 worker->current_work = NULL;
3281 worker->current_func = NULL; 2241 worker->current_func = NULL;
3282 worker->current_pwq = NULL; 2242 worker->current_pwq = NULL;
3283 worker->current_color = INT_MAX; !! 2243 worker->desc_valid = false;
3284 !! 2244 pwq_dec_nr_in_flight(pwq, work_color);
3285 /* must be the last step, see the fun <<
3286 pwq_dec_nr_in_flight(pwq, work_data); <<
3287 } 2245 }
3288 2246
3289 /** 2247 /**
3290 * process_scheduled_works - process schedule 2248 * process_scheduled_works - process scheduled works
3291 * @worker: self 2249 * @worker: self
3292 * 2250 *
3293 * Process all scheduled works. Please note 2251 * Process all scheduled works. Please note that the scheduled list
3294 * may change while processing a work, so thi 2252 * may change while processing a work, so this function repeatedly
3295 * fetches a work from the top and executes i 2253 * fetches a work from the top and executes it.
3296 * 2254 *
3297 * CONTEXT: 2255 * CONTEXT:
3298 * raw_spin_lock_irq(pool->lock) which may be !! 2256 * spin_lock_irq(pool->lock) which may be released and regrabbed
3299 * multiple times. 2257 * multiple times.
3300 */ 2258 */
3301 static void process_scheduled_works(struct wo 2259 static void process_scheduled_works(struct worker *worker)
3302 { 2260 {
3303 struct work_struct *work; !! 2261 while (!list_empty(&worker->scheduled)) {
3304 bool first = true; !! 2262 struct work_struct *work = list_first_entry(&worker->scheduled,
3305 !! 2263 struct work_struct, entry);
3306 while ((work = list_first_entry_or_nu <<
3307 <<
3308 if (first) { <<
3309 worker->pool->watchdo <<
3310 first = false; <<
3311 } <<
3312 process_one_work(worker, work 2264 process_one_work(worker, work);
3313 } 2265 }
3314 } 2266 }
3315 2267
3316 static void set_pf_worker(bool val) <<
3317 { <<
3318 mutex_lock(&wq_pool_attach_mutex); <<
3319 if (val) <<
3320 current->flags |= PF_WQ_WORKE <<
3321 else <<
3322 current->flags &= ~PF_WQ_WORK <<
3323 mutex_unlock(&wq_pool_attach_mutex); <<
3324 } <<
3325 <<
3326 /** 2268 /**
3327 * worker_thread - the worker thread function 2269 * worker_thread - the worker thread function
3328 * @__worker: self 2270 * @__worker: self
3329 * 2271 *
3330 * The worker thread function. All workers b 2272 * The worker thread function. All workers belong to a worker_pool -
3331 * either a per-cpu one or dynamic unbound on 2273 * either a per-cpu one or dynamic unbound one. These workers process all
3332 * work items regardless of their specific ta 2274 * work items regardless of their specific target workqueue. The only
3333 * exception is work items which belong to wo 2275 * exception is work items which belong to workqueues with a rescuer which
3334 * will be explained in rescuer_thread(). 2276 * will be explained in rescuer_thread().
3335 * <<
3336 * Return: 0 <<
3337 */ 2277 */
3338 static int worker_thread(void *__worker) 2278 static int worker_thread(void *__worker)
3339 { 2279 {
3340 struct worker *worker = __worker; 2280 struct worker *worker = __worker;
3341 struct worker_pool *pool = worker->po 2281 struct worker_pool *pool = worker->pool;
3342 2282
3343 /* tell the scheduler that this is a 2283 /* tell the scheduler that this is a workqueue worker */
3344 set_pf_worker(true); !! 2284 worker->task->flags |= PF_WQ_WORKER;
3345 woke_up: 2285 woke_up:
3346 raw_spin_lock_irq(&pool->lock); !! 2286 spin_lock_irq(&pool->lock);
3347 2287
3348 /* am I supposed to die? */ 2288 /* am I supposed to die? */
3349 if (unlikely(worker->flags & WORKER_D 2289 if (unlikely(worker->flags & WORKER_DIE)) {
3350 raw_spin_unlock_irq(&pool->lo !! 2290 spin_unlock_irq(&pool->lock);
3351 set_pf_worker(false); !! 2291 WARN_ON_ONCE(!list_empty(&worker->entry));
3352 /* !! 2292 worker->task->flags &= ~PF_WQ_WORKER;
3353 * The worker is dead and PF_ <<
3354 * shouldn't be accessed, res <<
3355 */ <<
3356 worker->pool = NULL; <<
3357 ida_free(&pool->worker_ida, w <<
3358 return 0; 2293 return 0;
3359 } 2294 }
3360 2295
3361 worker_leave_idle(worker); 2296 worker_leave_idle(worker);
3362 recheck: 2297 recheck:
3363 /* no more worker necessary? */ 2298 /* no more worker necessary? */
3364 if (!need_more_worker(pool)) 2299 if (!need_more_worker(pool))
3365 goto sleep; 2300 goto sleep;
3366 2301
3367 /* do we need to manage? */ 2302 /* do we need to manage? */
3368 if (unlikely(!may_start_working(pool) 2303 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
3369 goto recheck; 2304 goto recheck;
3370 2305
3371 /* 2306 /*
3372 * ->scheduled list can only be fille 2307 * ->scheduled list can only be filled while a worker is
3373 * preparing to process a work or act 2308 * preparing to process a work or actually processing it.
3374 * Make sure nobody diddled with it w 2309 * Make sure nobody diddled with it while I was sleeping.
3375 */ 2310 */
3376 WARN_ON_ONCE(!list_empty(&worker->sch 2311 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3377 2312
3378 /* 2313 /*
3379 * Finish PREP stage. We're guarante 2314 * Finish PREP stage. We're guaranteed to have at least one idle
3380 * worker or that someone else has al 2315 * worker or that someone else has already assumed the manager
3381 * role. This is where @worker start 2316 * role. This is where @worker starts participating in concurrency
3382 * management if applicable and concu 2317 * management if applicable and concurrency management is restored
3383 * after being rebound. See rebind_w 2318 * after being rebound. See rebind_workers() for details.
3384 */ 2319 */
3385 worker_clr_flags(worker, WORKER_PREP 2320 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3386 2321
3387 do { 2322 do {
3388 struct work_struct *work = 2323 struct work_struct *work =
3389 list_first_entry(&poo 2324 list_first_entry(&pool->worklist,
3390 stru 2325 struct work_struct, entry);
3391 2326
3392 if (assign_work(work, worker, !! 2327 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
>> 2328 /* optimization path, not strictly necessary */
>> 2329 process_one_work(worker, work);
>> 2330 if (unlikely(!list_empty(&worker->scheduled)))
>> 2331 process_scheduled_works(worker);
>> 2332 } else {
>> 2333 move_linked_works(work, &worker->scheduled, NULL);
3393 process_scheduled_wor 2334 process_scheduled_works(worker);
>> 2335 }
3394 } while (keep_working(pool)); 2336 } while (keep_working(pool));
3395 2337
3396 worker_set_flags(worker, WORKER_PREP) !! 2338 worker_set_flags(worker, WORKER_PREP, false);
3397 sleep: 2339 sleep:
>> 2340 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
>> 2341 goto recheck;
>> 2342
3398 /* 2343 /*
3399 * pool->lock is held and there's no 2344 * pool->lock is held and there's no work to process and no need to
3400 * manage, sleep. Workers are woken 2345 * manage, sleep. Workers are woken up only while holding
3401 * pool->lock or from local cpu, so s 2346 * pool->lock or from local cpu, so setting the current state
3402 * before releasing pool->lock is eno 2347 * before releasing pool->lock is enough to prevent losing any
3403 * event. 2348 * event.
3404 */ 2349 */
3405 worker_enter_idle(worker); 2350 worker_enter_idle(worker);
3406 __set_current_state(TASK_IDLE); !! 2351 __set_current_state(TASK_INTERRUPTIBLE);
3407 raw_spin_unlock_irq(&pool->lock); !! 2352 spin_unlock_irq(&pool->lock);
3408 schedule(); 2353 schedule();
3409 goto woke_up; 2354 goto woke_up;
3410 } 2355 }
3411 2356
3412 /** 2357 /**
3413 * rescuer_thread - the rescuer thread functi 2358 * rescuer_thread - the rescuer thread function
3414 * @__rescuer: self 2359 * @__rescuer: self
3415 * 2360 *
3416 * Workqueue rescuer thread function. There' 2361 * Workqueue rescuer thread function. There's one rescuer for each
3417 * workqueue which has WQ_MEM_RECLAIM set. 2362 * workqueue which has WQ_MEM_RECLAIM set.
3418 * 2363 *
3419 * Regular work processing on a pool may bloc 2364 * Regular work processing on a pool may block trying to create a new
3420 * worker which uses GFP_KERNEL allocation wh 2365 * worker which uses GFP_KERNEL allocation which has slight chance of
3421 * developing into deadlock if some works cur 2366 * developing into deadlock if some works currently on the same queue
3422 * need to be processed to satisfy the GFP_KE 2367 * need to be processed to satisfy the GFP_KERNEL allocation. This is
3423 * the problem rescuer solves. 2368 * the problem rescuer solves.
3424 * 2369 *
3425 * When such condition is possible, the pool 2370 * When such condition is possible, the pool summons rescuers of all
3426 * workqueues which have works queued on the 2371 * workqueues which have works queued on the pool and let them process
3427 * those works so that forward progress can b 2372 * those works so that forward progress can be guaranteed.
3428 * 2373 *
3429 * This should happen rarely. 2374 * This should happen rarely.
3430 * <<
3431 * Return: 0 <<
3432 */ 2375 */
3433 static int rescuer_thread(void *__rescuer) 2376 static int rescuer_thread(void *__rescuer)
3434 { 2377 {
3435 struct worker *rescuer = __rescuer; 2378 struct worker *rescuer = __rescuer;
3436 struct workqueue_struct *wq = rescuer 2379 struct workqueue_struct *wq = rescuer->rescue_wq;
>> 2380 struct list_head *scheduled = &rescuer->scheduled;
3437 bool should_stop; 2381 bool should_stop;
3438 2382
3439 set_user_nice(current, RESCUER_NICE_L 2383 set_user_nice(current, RESCUER_NICE_LEVEL);
3440 2384
3441 /* 2385 /*
3442 * Mark rescuer as worker too. As WO 2386 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
3443 * doesn't participate in concurrency 2387 * doesn't participate in concurrency management.
3444 */ 2388 */
3445 set_pf_worker(true); !! 2389 rescuer->task->flags |= PF_WQ_WORKER;
3446 repeat: 2390 repeat:
3447 set_current_state(TASK_IDLE); !! 2391 set_current_state(TASK_INTERRUPTIBLE);
3448 2392
3449 /* 2393 /*
3450 * By the time the rescuer is request 2394 * By the time the rescuer is requested to stop, the workqueue
3451 * shouldn't have any work pending, b 2395 * shouldn't have any work pending, but @wq->maydays may still have
3452 * pwq(s) queued. This can happen by 2396 * pwq(s) queued. This can happen by non-rescuer workers consuming
3453 * all the work items before the resc 2397 * all the work items before the rescuer got to them. Go through
3454 * @wq->maydays processing before act 2398 * @wq->maydays processing before acting on should_stop so that the
3455 * list is always empty on exit. 2399 * list is always empty on exit.
3456 */ 2400 */
3457 should_stop = kthread_should_stop(); 2401 should_stop = kthread_should_stop();
3458 2402
3459 /* see whether any pwq is asking for 2403 /* see whether any pwq is asking for help */
3460 raw_spin_lock_irq(&wq_mayday_lock); !! 2404 spin_lock_irq(&wq_mayday_lock);
3461 2405
3462 while (!list_empty(&wq->maydays)) { 2406 while (!list_empty(&wq->maydays)) {
3463 struct pool_workqueue *pwq = 2407 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
3464 struc 2408 struct pool_workqueue, mayday_node);
3465 struct worker_pool *pool = pw 2409 struct worker_pool *pool = pwq->pool;
3466 struct work_struct *work, *n; 2410 struct work_struct *work, *n;
3467 2411
3468 __set_current_state(TASK_RUNN 2412 __set_current_state(TASK_RUNNING);
3469 list_del_init(&pwq->mayday_no 2413 list_del_init(&pwq->mayday_node);
3470 2414
3471 raw_spin_unlock_irq(&wq_mayda !! 2415 spin_unlock_irq(&wq_mayday_lock);
3472 2416
3473 worker_attach_to_pool(rescuer !! 2417 /* migrate to the target cpu if possible */
3474 !! 2418 worker_maybe_bind_and_lock(pool);
3475 raw_spin_lock_irq(&pool->lock !! 2419 rescuer->pool = pool;
3476 2420
3477 /* 2421 /*
3478 * Slurp in all works issued 2422 * Slurp in all works issued via this workqueue and
3479 * process'em. 2423 * process'em.
3480 */ 2424 */
3481 WARN_ON_ONCE(!list_empty(&res 2425 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
3482 list_for_each_entry_safe(work !! 2426 list_for_each_entry_safe(work, n, &pool->worklist, entry)
3483 if (get_work_pwq(work !! 2427 if (get_work_pwq(work) == pwq)
3484 assign_work(work, !! 2428 move_linked_works(work, scheduled, &n);
3485 pwq->stats[PW <<
3486 } <<
3487 2429
3488 if (!list_empty(&rescuer->sch !! 2430 process_scheduled_works(rescuer);
3489 process_scheduled_wor <<
3490 <<
3491 /* <<
3492 * The above executio <<
3493 * have created more <<
3494 * pwq_activate_first <<
3495 * queueing. Let's p <<
3496 * that such back-to- <<
3497 * being used to reli <<
3498 * incur MAYDAY_INTER <<
3499 */ <<
3500 if (pwq->nr_active && <<
3501 raw_spin_lock <<
3502 /* <<
3503 * Queue iff <<
3504 * and somebo <<
3505 */ <<
3506 if (wq->rescu <<
3507 get_p <<
3508 list_ <<
3509 } <<
3510 raw_spin_unlo <<
3511 } <<
3512 } <<
3513 2431
3514 /* 2432 /*
3515 * Put the reference grabbed 2433 * Put the reference grabbed by send_mayday(). @pool won't
3516 * go away while we're still !! 2434 * go away while we're holding its lock.
3517 */ 2435 */
3518 put_pwq(pwq); 2436 put_pwq(pwq);
3519 2437
3520 /* 2438 /*
3521 * Leave this pool. Notify re !! 2439 * Leave this pool. If keep_working() is %true, notify a
3522 * with 0 concurrency and sta !! 2440 * regular worker; otherwise, we end up with 0 concurrency
>> 2441 * and stalling the execution.
3523 */ 2442 */
3524 kick_pool(pool); !! 2443 if (keep_working(pool))
3525 !! 2444 wake_up_worker(pool);
3526 raw_spin_unlock_irq(&pool->lo <<
3527 <<
3528 worker_detach_from_pool(rescu <<
3529 2445
3530 raw_spin_lock_irq(&wq_mayday_ !! 2446 rescuer->pool = NULL;
>> 2447 spin_unlock(&pool->lock);
>> 2448 spin_lock(&wq_mayday_lock);
3531 } 2449 }
3532 2450
3533 raw_spin_unlock_irq(&wq_mayday_lock); !! 2451 spin_unlock_irq(&wq_mayday_lock);
3534 2452
3535 if (should_stop) { 2453 if (should_stop) {
3536 __set_current_state(TASK_RUNN 2454 __set_current_state(TASK_RUNNING);
3537 set_pf_worker(false); !! 2455 rescuer->task->flags &= ~PF_WQ_WORKER;
3538 return 0; 2456 return 0;
3539 } 2457 }
3540 2458
3541 /* rescuers should never participate 2459 /* rescuers should never participate in concurrency management */
3542 WARN_ON_ONCE(!(rescuer->flags & WORKE 2460 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
3543 schedule(); 2461 schedule();
3544 goto repeat; 2462 goto repeat;
3545 } 2463 }
3546 2464
3547 static void bh_worker(struct worker *worker) <<
3548 { <<
3549 struct worker_pool *pool = worker->po <<
3550 int nr_restarts = BH_WORKER_RESTARTS; <<
3551 unsigned long end = jiffies + BH_WORK <<
3552 <<
3553 raw_spin_lock_irq(&pool->lock); <<
3554 worker_leave_idle(worker); <<
3555 <<
3556 /* <<
3557 * This function follows the structur <<
3558 * explanations on each step. <<
3559 */ <<
3560 if (!need_more_worker(pool)) <<
3561 goto done; <<
3562 <<
3563 WARN_ON_ONCE(!list_empty(&worker->sch <<
3564 worker_clr_flags(worker, WORKER_PREP <<
3565 <<
3566 do { <<
3567 struct work_struct *work = <<
3568 list_first_entry(&poo <<
3569 stru <<
3570 <<
3571 if (assign_work(work, worker, <<
3572 process_scheduled_wor <<
3573 } while (keep_working(pool) && <<
3574 --nr_restarts && time_before <<
3575 <<
3576 worker_set_flags(worker, WORKER_PREP) <<
3577 done: <<
3578 worker_enter_idle(worker); <<
3579 kick_pool(pool); <<
3580 raw_spin_unlock_irq(&pool->lock); <<
3581 } <<
3582 <<
3583 /* <<
3584 * TODO: Convert all tasklet users to workque <<
3585 * <<
3586 * This is currently called from tasklet[_hi] <<
3587 * whenever there are tasklets to run. Let's <<
3588 * queued. Once conversion from tasklet is co <<
3589 * can be dropped. <<
3590 * <<
3591 * After full conversion, we'll add worker->s <<
3592 * softirq action and obtain the worker point <<
3593 */ <<
3594 void workqueue_softirq_action(bool highpri) <<
3595 { <<
3596 struct worker_pool *pool = <<
3597 &per_cpu(bh_worker_pools, smp <<
3598 if (need_more_worker(pool)) <<
3599 bh_worker(list_first_entry(&p <<
3600 } <<
3601 <<
3602 struct wq_drain_dead_softirq_work { <<
3603 struct work_struct work; <<
3604 struct worker_pool *pool; <<
3605 struct completion done; <<
3606 }; <<
3607 <<
3608 static void drain_dead_softirq_workfn(struct <<
3609 { <<
3610 struct wq_drain_dead_softirq_work *de <<
3611 container_of(work, struct wq_ <<
3612 struct worker_pool *pool = dead_work- <<
3613 bool repeat; <<
3614 <<
3615 /* <<
3616 * @pool's CPU is dead and we want to <<
3617 * items from this BH work item which <<
3618 * its CPU is dead, @pool can't be ki <<
3619 * will be nested, a lockdep annotati <<
3620 * @pool with %POOL_BH_DRAINING for t <<
3621 */ <<
3622 raw_spin_lock_irq(&pool->lock); <<
3623 pool->flags |= POOL_BH_DRAINING; <<
3624 raw_spin_unlock_irq(&pool->lock); <<
3625 <<
3626 bh_worker(list_first_entry(&pool->wor <<
3627 <<
3628 raw_spin_lock_irq(&pool->lock); <<
3629 pool->flags &= ~POOL_BH_DRAINING; <<
3630 repeat = need_more_worker(pool); <<
3631 raw_spin_unlock_irq(&pool->lock); <<
3632 <<
3633 /* <<
3634 * bh_worker() might hit consecutive <<
3635 * still are pending work items, resc <<
3636 * don't hog this CPU's BH. <<
3637 */ <<
3638 if (repeat) { <<
3639 if (pool->attrs->nice == HIGH <<
3640 queue_work(system_bh_ <<
3641 else <<
3642 queue_work(system_bh_ <<
3643 } else { <<
3644 complete(&dead_work->done); <<
3645 } <<
3646 } <<
3647 <<
3648 /* <<
3649 * @cpu is dead. Drain the remaining BH work <<
3650 * possible to allocate dead_work per CPU and <<
3651 * have to worry about draining overlapping w <<
3652 * nesting (one CPU's dead_work queued on ano <<
3653 * on). Let's keep it simple and drain them s <<
3654 * items which shouldn't be requeued on the s <<
3655 */ <<
3656 void workqueue_softirq_dead(unsigned int cpu) <<
3657 { <<
3658 int i; <<
3659 <<
3660 for (i = 0; i < NR_STD_WORKER_POOLS; <<
3661 struct worker_pool *pool = &p <<
3662 struct wq_drain_dead_softirq_ <<
3663 <<
3664 if (!need_more_worker(pool)) <<
3665 continue; <<
3666 <<
3667 INIT_WORK_ONSTACK(&dead_work. <<
3668 dead_work.pool = pool; <<
3669 init_completion(&dead_work.do <<
3670 <<
3671 if (pool->attrs->nice == HIGH <<
3672 queue_work(system_bh_ <<
3673 else <<
3674 queue_work(system_bh_ <<
3675 <<
3676 wait_for_completion(&dead_wor <<
3677 destroy_work_on_stack(&dead_w <<
3678 } <<
3679 } <<
3680 <<
3681 /** <<
3682 * check_flush_dependency - check for flush d <<
3683 * @target_wq: workqueue being flushed <<
3684 * @target_work: work item being flushed (NUL <<
3685 * <<
3686 * %current is trying to flush the whole @tar <<
3687 * If @target_wq doesn't have %WQ_MEM_RECLAIM <<
3688 * reclaiming memory or running on a workqueu <<
3689 * %WQ_MEM_RECLAIM as that can break forward- <<
3690 * a deadlock. <<
3691 */ <<
3692 static void check_flush_dependency(struct wor <<
3693 struct wor <<
3694 { <<
3695 work_func_t target_func = target_work <<
3696 struct worker *worker; <<
3697 <<
3698 if (target_wq->flags & WQ_MEM_RECLAIM <<
3699 return; <<
3700 <<
3701 worker = current_wq_worker(); <<
3702 <<
3703 WARN_ONCE(current->flags & PF_MEMALLO <<
3704 "workqueue: PF_MEMALLOC tas <<
3705 current->pid, current->comm <<
3706 WARN_ONCE(worker && ((worker->current <<
3707 (WQ_MEM_RECLAIM <<
3708 "workqueue: WQ_MEM_RECLAIM <<
3709 worker->current_pwq->wq->na <<
3710 target_wq->name, target_fun <<
3711 } <<
3712 <<
3713 struct wq_barrier { 2465 struct wq_barrier {
3714 struct work_struct work; 2466 struct work_struct work;
3715 struct completion done; 2467 struct completion done;
3716 struct task_struct *task; /* pu <<
3717 }; 2468 };
3718 2469
3719 static void wq_barrier_func(struct work_struc 2470 static void wq_barrier_func(struct work_struct *work)
3720 { 2471 {
3721 struct wq_barrier *barr = container_o 2472 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
3722 complete(&barr->done); 2473 complete(&barr->done);
3723 } 2474 }
3724 2475
3725 /** 2476 /**
3726 * insert_wq_barrier - insert a barrier work 2477 * insert_wq_barrier - insert a barrier work
3727 * @pwq: pwq to insert barrier into 2478 * @pwq: pwq to insert barrier into
3728 * @barr: wq_barrier to insert 2479 * @barr: wq_barrier to insert
3729 * @target: target work to attach @barr to 2480 * @target: target work to attach @barr to
3730 * @worker: worker currently executing @targe 2481 * @worker: worker currently executing @target, NULL if @target is not executing
3731 * 2482 *
3732 * @barr is linked to @target such that @barr 2483 * @barr is linked to @target such that @barr is completed only after
3733 * @target finishes execution. Please note t 2484 * @target finishes execution. Please note that the ordering
3734 * guarantee is observed only with respect to 2485 * guarantee is observed only with respect to @target and on the local
3735 * cpu. 2486 * cpu.
3736 * 2487 *
3737 * Currently, a queued barrier can't be cance 2488 * Currently, a queued barrier can't be canceled. This is because
3738 * try_to_grab_pending() can't determine whet 2489 * try_to_grab_pending() can't determine whether the work to be
3739 * grabbed is at the head of the queue and th 2490 * grabbed is at the head of the queue and thus can't clear LINKED
3740 * flag of the previous work while there must 2491 * flag of the previous work while there must be a valid next work
3741 * after a work with LINKED flag set. 2492 * after a work with LINKED flag set.
3742 * 2493 *
3743 * Note that when @worker is non-NULL, @targe 2494 * Note that when @worker is non-NULL, @target may be modified
3744 * underneath us, so we can't reliably determ 2495 * underneath us, so we can't reliably determine pwq from @target.
3745 * 2496 *
3746 * CONTEXT: 2497 * CONTEXT:
3747 * raw_spin_lock_irq(pool->lock). !! 2498 * spin_lock_irq(pool->lock).
3748 */ 2499 */
3749 static void insert_wq_barrier(struct pool_wor 2500 static void insert_wq_barrier(struct pool_workqueue *pwq,
3750 struct wq_barri 2501 struct wq_barrier *barr,
3751 struct work_str 2502 struct work_struct *target, struct worker *worker)
3752 { 2503 {
3753 static __maybe_unused struct lock_cla <<
3754 unsigned int work_flags = 0; <<
3755 unsigned int work_color; <<
3756 struct list_head *head; 2504 struct list_head *head;
>> 2505 unsigned int linked = 0;
3757 2506
3758 /* 2507 /*
3759 * debugobject calls are safe here ev 2508 * debugobject calls are safe here even with pool->lock locked
3760 * as we know for sure that this will 2509 * as we know for sure that this will not trigger any of the
3761 * checks and call back into the fixu 2510 * checks and call back into the fixup functions where we
3762 * might deadlock. 2511 * might deadlock.
3763 * <<
3764 * BH and threaded workqueues need se <<
3765 * spuriously triggering "inconsisten <<
3766 * usage". <<
3767 */ 2512 */
3768 INIT_WORK_ONSTACK_KEY(&barr->work, wq !! 2513 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
3769 (pwq->wq->flags <<
3770 __set_bit(WORK_STRUCT_PENDING_BIT, wo 2514 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
3771 !! 2515 init_completion(&barr->done);
3772 init_completion_map(&barr->done, &tar <<
3773 <<
3774 barr->task = current; <<
3775 <<
3776 /* The barrier work item does not par <<
3777 work_flags |= WORK_STRUCT_INACTIVE; <<
3778 2516
3779 /* 2517 /*
3780 * If @target is currently being exec 2518 * If @target is currently being executed, schedule the
3781 * barrier to the worker; otherwise, 2519 * barrier to the worker; otherwise, put it after @target.
3782 */ 2520 */
3783 if (worker) { !! 2521 if (worker)
3784 head = worker->scheduled.next 2522 head = worker->scheduled.next;
3785 work_color = worker->current_ !! 2523 else {
3786 } else { <<
3787 unsigned long *bits = work_da 2524 unsigned long *bits = work_data_bits(target);
3788 2525
3789 head = target->entry.next; 2526 head = target->entry.next;
3790 /* there can already be other 2527 /* there can already be other linked works, inherit and set */
3791 work_flags |= *bits & WORK_ST !! 2528 linked = *bits & WORK_STRUCT_LINKED;
3792 work_color = get_work_color(* <<
3793 __set_bit(WORK_STRUCT_LINKED_ 2529 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
3794 } 2530 }
3795 2531
3796 pwq->nr_in_flight[work_color]++; !! 2532 debug_work_activate(&barr->work);
3797 work_flags |= work_color_to_flags(wor !! 2533 insert_work(pwq, &barr->work, head,
3798 !! 2534 work_color_to_flags(WORK_NO_COLOR) | linked);
3799 insert_work(pwq, &barr->work, head, w <<
3800 } 2535 }
3801 2536
3802 /** 2537 /**
3803 * flush_workqueue_prep_pwqs - prepare pwqs f 2538 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
3804 * @wq: workqueue being flushed 2539 * @wq: workqueue being flushed
3805 * @flush_color: new flush color, < 0 for no- 2540 * @flush_color: new flush color, < 0 for no-op
3806 * @work_color: new work color, < 0 for no-op 2541 * @work_color: new work color, < 0 for no-op
3807 * 2542 *
3808 * Prepare pwqs for workqueue flushing. 2543 * Prepare pwqs for workqueue flushing.
3809 * 2544 *
3810 * If @flush_color is non-negative, flush_col 2545 * If @flush_color is non-negative, flush_color on all pwqs should be
3811 * -1. If no pwq has in-flight commands at t 2546 * -1. If no pwq has in-flight commands at the specified color, all
3812 * pwq->flush_color's stay at -1 and %false i 2547 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
3813 * has in flight commands, its pwq->flush_col 2548 * has in flight commands, its pwq->flush_color is set to
3814 * @flush_color, @wq->nr_pwqs_to_flush is upd 2549 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
3815 * wakeup logic is armed and %true is returne 2550 * wakeup logic is armed and %true is returned.
3816 * 2551 *
3817 * The caller should have initialized @wq->fi 2552 * The caller should have initialized @wq->first_flusher prior to
3818 * calling this function with non-negative @f 2553 * calling this function with non-negative @flush_color. If
3819 * @flush_color is negative, no flush color u 2554 * @flush_color is negative, no flush color update is done and %false
3820 * is returned. 2555 * is returned.
3821 * 2556 *
3822 * If @work_color is non-negative, all pwqs s 2557 * If @work_color is non-negative, all pwqs should have the same
3823 * work_color which is previous to @work_colo 2558 * work_color which is previous to @work_color and all will be
3824 * advanced to @work_color. 2559 * advanced to @work_color.
3825 * 2560 *
3826 * CONTEXT: 2561 * CONTEXT:
3827 * mutex_lock(wq->mutex). 2562 * mutex_lock(wq->mutex).
3828 * 2563 *
3829 * Return: !! 2564 * RETURNS:
3830 * %true if @flush_color >= 0 and there's som 2565 * %true if @flush_color >= 0 and there's something to flush. %false
3831 * otherwise. 2566 * otherwise.
3832 */ 2567 */
3833 static bool flush_workqueue_prep_pwqs(struct 2568 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
3834 int flu 2569 int flush_color, int work_color)
3835 { 2570 {
3836 bool wait = false; 2571 bool wait = false;
3837 struct pool_workqueue *pwq; 2572 struct pool_workqueue *pwq;
3838 2573
3839 if (flush_color >= 0) { 2574 if (flush_color >= 0) {
3840 WARN_ON_ONCE(atomic_read(&wq- 2575 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
3841 atomic_set(&wq->nr_pwqs_to_fl 2576 atomic_set(&wq->nr_pwqs_to_flush, 1);
3842 } 2577 }
3843 2578
3844 for_each_pwq(pwq, wq) { 2579 for_each_pwq(pwq, wq) {
3845 struct worker_pool *pool = pw 2580 struct worker_pool *pool = pwq->pool;
3846 2581
3847 raw_spin_lock_irq(&pool->lock !! 2582 spin_lock_irq(&pool->lock);
3848 2583
3849 if (flush_color >= 0) { 2584 if (flush_color >= 0) {
3850 WARN_ON_ONCE(pwq->flu 2585 WARN_ON_ONCE(pwq->flush_color != -1);
3851 2586
3852 if (pwq->nr_in_flight 2587 if (pwq->nr_in_flight[flush_color]) {
3853 pwq->flush_co 2588 pwq->flush_color = flush_color;
3854 atomic_inc(&w 2589 atomic_inc(&wq->nr_pwqs_to_flush);
3855 wait = true; 2590 wait = true;
3856 } 2591 }
3857 } 2592 }
3858 2593
3859 if (work_color >= 0) { 2594 if (work_color >= 0) {
3860 WARN_ON_ONCE(work_col 2595 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
3861 pwq->work_color = wor 2596 pwq->work_color = work_color;
3862 } 2597 }
3863 2598
3864 raw_spin_unlock_irq(&pool->lo !! 2599 spin_unlock_irq(&pool->lock);
3865 } 2600 }
3866 2601
3867 if (flush_color >= 0 && atomic_dec_an 2602 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
3868 complete(&wq->first_flusher-> 2603 complete(&wq->first_flusher->done);
3869 2604
3870 return wait; 2605 return wait;
3871 } 2606 }
3872 2607
3873 static void touch_wq_lockdep_map(struct workq <<
3874 { <<
3875 #ifdef CONFIG_LOCKDEP <<
3876 if (wq->flags & WQ_BH) <<
3877 local_bh_disable(); <<
3878 <<
3879 lock_map_acquire(&wq->lockdep_map); <<
3880 lock_map_release(&wq->lockdep_map); <<
3881 <<
3882 if (wq->flags & WQ_BH) <<
3883 local_bh_enable(); <<
3884 #endif <<
3885 } <<
3886 <<
3887 static void touch_work_lockdep_map(struct wor <<
3888 struct wor <<
3889 { <<
3890 #ifdef CONFIG_LOCKDEP <<
3891 if (wq->flags & WQ_BH) <<
3892 local_bh_disable(); <<
3893 <<
3894 lock_map_acquire(&work->lockdep_map); <<
3895 lock_map_release(&work->lockdep_map); <<
3896 <<
3897 if (wq->flags & WQ_BH) <<
3898 local_bh_enable(); <<
3899 #endif <<
3900 } <<
3901 <<
3902 /** 2608 /**
3903 * __flush_workqueue - ensure that any schedu !! 2609 * flush_workqueue - ensure that any scheduled work has run to completion.
3904 * @wq: workqueue to flush 2610 * @wq: workqueue to flush
3905 * 2611 *
3906 * This function sleeps until all work items 2612 * This function sleeps until all work items which were queued on entry
3907 * have finished execution, but it is not liv 2613 * have finished execution, but it is not livelocked by new incoming ones.
3908 */ 2614 */
3909 void __flush_workqueue(struct workqueue_struc !! 2615 void flush_workqueue(struct workqueue_struct *wq)
3910 { 2616 {
3911 struct wq_flusher this_flusher = { 2617 struct wq_flusher this_flusher = {
3912 .list = LIST_HEAD_INIT(this_f 2618 .list = LIST_HEAD_INIT(this_flusher.list),
3913 .flush_color = -1, 2619 .flush_color = -1,
3914 .done = COMPLETION_INITIALIZE !! 2620 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
3915 }; 2621 };
3916 int next_color; 2622 int next_color;
3917 2623
3918 if (WARN_ON(!wq_online)) !! 2624 lock_map_acquire(&wq->lockdep_map);
3919 return; !! 2625 lock_map_release(&wq->lockdep_map);
3920 <<
3921 touch_wq_lockdep_map(wq); <<
3922 2626
3923 mutex_lock(&wq->mutex); 2627 mutex_lock(&wq->mutex);
3924 2628
3925 /* 2629 /*
3926 * Start-to-wait phase 2630 * Start-to-wait phase
3927 */ 2631 */
3928 next_color = work_next_color(wq->work 2632 next_color = work_next_color(wq->work_color);
3929 2633
3930 if (next_color != wq->flush_color) { 2634 if (next_color != wq->flush_color) {
3931 /* 2635 /*
3932 * Color space is not full. 2636 * Color space is not full. The current work_color
3933 * becomes our flush_color an 2637 * becomes our flush_color and work_color is advanced
3934 * by one. 2638 * by one.
3935 */ 2639 */
3936 WARN_ON_ONCE(!list_empty(&wq- 2640 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
3937 this_flusher.flush_color = wq 2641 this_flusher.flush_color = wq->work_color;
3938 wq->work_color = next_color; 2642 wq->work_color = next_color;
3939 2643
3940 if (!wq->first_flusher) { 2644 if (!wq->first_flusher) {
3941 /* no flush in progre 2645 /* no flush in progress, become the first flusher */
3942 WARN_ON_ONCE(wq->flus 2646 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3943 2647
3944 wq->first_flusher = & 2648 wq->first_flusher = &this_flusher;
3945 2649
3946 if (!flush_workqueue_ 2650 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
3947 2651 wq->work_color)) {
3948 /* nothing to 2652 /* nothing to flush, done */
3949 wq->flush_col 2653 wq->flush_color = next_color;
3950 wq->first_flu 2654 wq->first_flusher = NULL;
3951 goto out_unlo 2655 goto out_unlock;
3952 } 2656 }
3953 } else { 2657 } else {
3954 /* wait in queue */ 2658 /* wait in queue */
3955 WARN_ON_ONCE(wq->flus 2659 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
3956 list_add_tail(&this_f 2660 list_add_tail(&this_flusher.list, &wq->flusher_queue);
3957 flush_workqueue_prep_ 2661 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3958 } 2662 }
3959 } else { 2663 } else {
3960 /* 2664 /*
3961 * Oops, color space is full, 2665 * Oops, color space is full, wait on overflow queue.
3962 * The next flush completion 2666 * The next flush completion will assign us
3963 * flush_color and transfer t 2667 * flush_color and transfer to flusher_queue.
3964 */ 2668 */
3965 list_add_tail(&this_flusher.l 2669 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
3966 } 2670 }
3967 2671
3968 check_flush_dependency(wq, NULL); <<
3969 <<
3970 mutex_unlock(&wq->mutex); 2672 mutex_unlock(&wq->mutex);
3971 2673
3972 wait_for_completion(&this_flusher.don 2674 wait_for_completion(&this_flusher.done);
3973 2675
3974 /* 2676 /*
3975 * Wake-up-and-cascade phase 2677 * Wake-up-and-cascade phase
3976 * 2678 *
3977 * First flushers are responsible for 2679 * First flushers are responsible for cascading flushes and
3978 * handling overflow. Non-first flus 2680 * handling overflow. Non-first flushers can simply return.
3979 */ 2681 */
3980 if (READ_ONCE(wq->first_flusher) != & !! 2682 if (wq->first_flusher != &this_flusher)
3981 return; 2683 return;
3982 2684
3983 mutex_lock(&wq->mutex); 2685 mutex_lock(&wq->mutex);
3984 2686
3985 /* we might have raced, check again w 2687 /* we might have raced, check again with mutex held */
3986 if (wq->first_flusher != &this_flushe 2688 if (wq->first_flusher != &this_flusher)
3987 goto out_unlock; 2689 goto out_unlock;
3988 2690
3989 WRITE_ONCE(wq->first_flusher, NULL); !! 2691 wq->first_flusher = NULL;
3990 2692
3991 WARN_ON_ONCE(!list_empty(&this_flushe 2693 WARN_ON_ONCE(!list_empty(&this_flusher.list));
3992 WARN_ON_ONCE(wq->flush_color != this_ 2694 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3993 2695
3994 while (true) { 2696 while (true) {
3995 struct wq_flusher *next, *tmp 2697 struct wq_flusher *next, *tmp;
3996 2698
3997 /* complete all the flushers 2699 /* complete all the flushers sharing the current flush color */
3998 list_for_each_entry_safe(next 2700 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
3999 if (next->flush_color 2701 if (next->flush_color != wq->flush_color)
4000 break; 2702 break;
4001 list_del_init(&next-> 2703 list_del_init(&next->list);
4002 complete(&next->done) 2704 complete(&next->done);
4003 } 2705 }
4004 2706
4005 WARN_ON_ONCE(!list_empty(&wq- 2707 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
4006 wq->flush_color 2708 wq->flush_color != work_next_color(wq->work_color));
4007 2709
4008 /* this flush_color is finish 2710 /* this flush_color is finished, advance by one */
4009 wq->flush_color = work_next_c 2711 wq->flush_color = work_next_color(wq->flush_color);
4010 2712
4011 /* one color has been freed, 2713 /* one color has been freed, handle overflow queue */
4012 if (!list_empty(&wq->flusher_ 2714 if (!list_empty(&wq->flusher_overflow)) {
4013 /* 2715 /*
4014 * Assign the same co 2716 * Assign the same color to all overflowed
4015 * flushers, advance 2717 * flushers, advance work_color and append to
4016 * flusher_queue. Th 2718 * flusher_queue. This is the start-to-wait
4017 * phase for these ov 2719 * phase for these overflowed flushers.
4018 */ 2720 */
4019 list_for_each_entry(t 2721 list_for_each_entry(tmp, &wq->flusher_overflow, list)
4020 tmp->flush_co 2722 tmp->flush_color = wq->work_color;
4021 2723
4022 wq->work_color = work 2724 wq->work_color = work_next_color(wq->work_color);
4023 2725
4024 list_splice_tail_init 2726 list_splice_tail_init(&wq->flusher_overflow,
4025 2727 &wq->flusher_queue);
4026 flush_workqueue_prep_ 2728 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
4027 } 2729 }
4028 2730
4029 if (list_empty(&wq->flusher_q 2731 if (list_empty(&wq->flusher_queue)) {
4030 WARN_ON_ONCE(wq->flus 2732 WARN_ON_ONCE(wq->flush_color != wq->work_color);
4031 break; 2733 break;
4032 } 2734 }
4033 2735
4034 /* 2736 /*
4035 * Need to flush more colors. 2737 * Need to flush more colors. Make the next flusher
4036 * the new first flusher and 2738 * the new first flusher and arm pwqs.
4037 */ 2739 */
4038 WARN_ON_ONCE(wq->flush_color 2740 WARN_ON_ONCE(wq->flush_color == wq->work_color);
4039 WARN_ON_ONCE(wq->flush_color 2741 WARN_ON_ONCE(wq->flush_color != next->flush_color);
4040 2742
4041 list_del_init(&next->list); 2743 list_del_init(&next->list);
4042 wq->first_flusher = next; 2744 wq->first_flusher = next;
4043 2745
4044 if (flush_workqueue_prep_pwqs 2746 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
4045 break; 2747 break;
4046 2748
4047 /* 2749 /*
4048 * Meh... this color is alrea 2750 * Meh... this color is already done, clear first
4049 * flusher and repeat cascadi 2751 * flusher and repeat cascading.
4050 */ 2752 */
4051 wq->first_flusher = NULL; 2753 wq->first_flusher = NULL;
4052 } 2754 }
4053 2755
4054 out_unlock: 2756 out_unlock:
4055 mutex_unlock(&wq->mutex); 2757 mutex_unlock(&wq->mutex);
4056 } 2758 }
4057 EXPORT_SYMBOL(__flush_workqueue); !! 2759 EXPORT_SYMBOL_GPL(flush_workqueue);
4058 2760
4059 /** 2761 /**
4060 * drain_workqueue - drain a workqueue 2762 * drain_workqueue - drain a workqueue
4061 * @wq: workqueue to drain 2763 * @wq: workqueue to drain
4062 * 2764 *
4063 * Wait until the workqueue becomes empty. W 2765 * Wait until the workqueue becomes empty. While draining is in progress,
4064 * only chain queueing is allowed. IOW, only 2766 * only chain queueing is allowed. IOW, only currently pending or running
4065 * work items on @wq can queue further work i 2767 * work items on @wq can queue further work items on it. @wq is flushed
4066 * repeatedly until it becomes empty. The nu !! 2768 * repeatedly until it becomes empty. The number of flushing is detemined
4067 * by the depth of chaining and should be rel 2769 * by the depth of chaining and should be relatively short. Whine if it
4068 * takes too long. 2770 * takes too long.
4069 */ 2771 */
4070 void drain_workqueue(struct workqueue_struct 2772 void drain_workqueue(struct workqueue_struct *wq)
4071 { 2773 {
4072 unsigned int flush_cnt = 0; 2774 unsigned int flush_cnt = 0;
4073 struct pool_workqueue *pwq; 2775 struct pool_workqueue *pwq;
4074 2776
4075 /* 2777 /*
4076 * __queue_work() needs to test wheth 2778 * __queue_work() needs to test whether there are drainers, is much
4077 * hotter than drain_workqueue() and 2779 * hotter than drain_workqueue() and already looks at @wq->flags.
4078 * Use __WQ_DRAINING so that queue do 2780 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
4079 */ 2781 */
4080 mutex_lock(&wq->mutex); 2782 mutex_lock(&wq->mutex);
4081 if (!wq->nr_drainers++) 2783 if (!wq->nr_drainers++)
4082 wq->flags |= __WQ_DRAINING; 2784 wq->flags |= __WQ_DRAINING;
4083 mutex_unlock(&wq->mutex); 2785 mutex_unlock(&wq->mutex);
4084 reflush: 2786 reflush:
4085 __flush_workqueue(wq); !! 2787 flush_workqueue(wq);
4086 2788
4087 mutex_lock(&wq->mutex); 2789 mutex_lock(&wq->mutex);
4088 2790
4089 for_each_pwq(pwq, wq) { 2791 for_each_pwq(pwq, wq) {
4090 bool drained; 2792 bool drained;
4091 2793
4092 raw_spin_lock_irq(&pwq->pool- !! 2794 spin_lock_irq(&pwq->pool->lock);
4093 drained = pwq_is_empty(pwq); !! 2795 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
4094 raw_spin_unlock_irq(&pwq->poo !! 2796 spin_unlock_irq(&pwq->pool->lock);
4095 2797
4096 if (drained) 2798 if (drained)
4097 continue; 2799 continue;
4098 2800
4099 if (++flush_cnt == 10 || 2801 if (++flush_cnt == 10 ||
4100 (flush_cnt % 100 == 0 && 2802 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
4101 pr_warn("workqueue %s !! 2803 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
4102 wq->name, __f !! 2804 wq->name, flush_cnt);
4103 2805
4104 mutex_unlock(&wq->mutex); 2806 mutex_unlock(&wq->mutex);
4105 goto reflush; 2807 goto reflush;
4106 } 2808 }
4107 2809
4108 if (!--wq->nr_drainers) 2810 if (!--wq->nr_drainers)
4109 wq->flags &= ~__WQ_DRAINING; 2811 wq->flags &= ~__WQ_DRAINING;
4110 mutex_unlock(&wq->mutex); 2812 mutex_unlock(&wq->mutex);
4111 } 2813 }
4112 EXPORT_SYMBOL_GPL(drain_workqueue); 2814 EXPORT_SYMBOL_GPL(drain_workqueue);
4113 2815
4114 static bool start_flush_work(struct work_stru !! 2816 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
4115 bool from_cancel <<
4116 { 2817 {
4117 struct worker *worker = NULL; 2818 struct worker *worker = NULL;
4118 struct worker_pool *pool; 2819 struct worker_pool *pool;
4119 struct pool_workqueue *pwq; 2820 struct pool_workqueue *pwq;
4120 struct workqueue_struct *wq; <<
4121 2821
4122 rcu_read_lock(); !! 2822 might_sleep();
>> 2823
>> 2824 local_irq_disable();
4123 pool = get_work_pool(work); 2825 pool = get_work_pool(work);
4124 if (!pool) { 2826 if (!pool) {
4125 rcu_read_unlock(); !! 2827 local_irq_enable();
4126 return false; 2828 return false;
4127 } 2829 }
4128 2830
4129 raw_spin_lock_irq(&pool->lock); !! 2831 spin_lock(&pool->lock);
4130 /* see the comment in try_to_grab_pen 2832 /* see the comment in try_to_grab_pending() with the same code */
4131 pwq = get_work_pwq(work); 2833 pwq = get_work_pwq(work);
4132 if (pwq) { 2834 if (pwq) {
4133 if (unlikely(pwq->pool != poo 2835 if (unlikely(pwq->pool != pool))
4134 goto already_gone; 2836 goto already_gone;
4135 } else { 2837 } else {
4136 worker = find_worker_executin 2838 worker = find_worker_executing_work(pool, work);
4137 if (!worker) 2839 if (!worker)
4138 goto already_gone; 2840 goto already_gone;
4139 pwq = worker->current_pwq; 2841 pwq = worker->current_pwq;
4140 } 2842 }
4141 2843
4142 wq = pwq->wq; <<
4143 check_flush_dependency(wq, work); <<
4144 <<
4145 insert_wq_barrier(pwq, barr, work, wo 2844 insert_wq_barrier(pwq, barr, work, worker);
4146 raw_spin_unlock_irq(&pool->lock); !! 2845 spin_unlock_irq(&pool->lock);
4147 <<
4148 touch_work_lockdep_map(work, wq); <<
4149 2846
4150 /* 2847 /*
4151 * Force a lock recursion deadlock wh !! 2848 * If @max_active is 1 or rescuer is in use, flushing another work
4152 * single-threaded or rescuer equippe !! 2849 * item on the same workqueue may lead to deadlock. Make sure the
4153 * !! 2850 * flusher is not running on the same workqueue by verifying write
4154 * For single threaded workqueues the !! 2851 * access.
4155 * is after the work issuing the flus <<
4156 * workqueues the deadlock happens wh <<
4157 * forward progress. <<
4158 */ 2852 */
4159 if (!from_cancel && (wq->saved_max_ac !! 2853 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
4160 touch_wq_lockdep_map(wq); !! 2854 lock_map_acquire(&pwq->wq->lockdep_map);
>> 2855 else
>> 2856 lock_map_acquire_read(&pwq->wq->lockdep_map);
>> 2857 lock_map_release(&pwq->wq->lockdep_map);
4161 2858
4162 rcu_read_unlock(); <<
4163 return true; 2859 return true;
4164 already_gone: 2860 already_gone:
4165 raw_spin_unlock_irq(&pool->lock); !! 2861 spin_unlock_irq(&pool->lock);
4166 rcu_read_unlock(); <<
4167 return false; 2862 return false;
4168 } 2863 }
4169 2864
4170 static bool __flush_work(struct work_struct * <<
4171 { <<
4172 struct wq_barrier barr; <<
4173 <<
4174 if (WARN_ON(!wq_online)) <<
4175 return false; <<
4176 <<
4177 if (WARN_ON(!work->func)) <<
4178 return false; <<
4179 <<
4180 if (!start_flush_work(work, &barr, fr <<
4181 return false; <<
4182 <<
4183 /* <<
4184 * start_flush_work() returned %true. <<
4185 * that @work must have been executin <<
4186 * can't currently be queued. Its dat <<
4187 * was queued on a BH workqueue, we a <<
4188 * BH context and thus can be busy-wa <<
4189 */ <<
4190 if (from_cancel) { <<
4191 unsigned long data = *work_da <<
4192 <<
4193 if (!WARN_ON_ONCE(data & WORK <<
4194 (data & WORK_OFFQ_BH)) { <<
4195 /* <<
4196 * On RT, prevent a l <<
4197 * soft interrupt pro <<
4198 * running by keeping <<
4199 * runs on a differen <<
4200 * than doing the BH <<
4201 * This is copied fro <<
4202 * kernel/softirq.c:: <<
4203 */ <<
4204 while (!try_wait_for_ <<
4205 if (IS_ENABLE <<
4206 local <<
4207 local <<
4208 } else { <<
4209 cpu_r <<
4210 } <<
4211 } <<
4212 goto out_destroy; <<
4213 } <<
4214 } <<
4215 <<
4216 wait_for_completion(&barr.done); <<
4217 <<
4218 out_destroy: <<
4219 destroy_work_on_stack(&barr.work); <<
4220 return true; <<
4221 } <<
4222 <<
4223 /** 2865 /**
4224 * flush_work - wait for a work to finish exe 2866 * flush_work - wait for a work to finish executing the last queueing instance
4225 * @work: the work to flush 2867 * @work: the work to flush
4226 * 2868 *
4227 * Wait until @work has finished execution. 2869 * Wait until @work has finished execution. @work is guaranteed to be idle
4228 * on return if it hasn't been requeued since 2870 * on return if it hasn't been requeued since flush started.
4229 * 2871 *
4230 * Return: !! 2872 * RETURNS:
4231 * %true if flush_work() waited for the work 2873 * %true if flush_work() waited for the work to finish execution,
4232 * %false if it was already idle. 2874 * %false if it was already idle.
4233 */ 2875 */
4234 bool flush_work(struct work_struct *work) 2876 bool flush_work(struct work_struct *work)
4235 { 2877 {
4236 might_sleep(); !! 2878 struct wq_barrier barr;
4237 return __flush_work(work, false); <<
4238 } <<
4239 EXPORT_SYMBOL_GPL(flush_work); <<
4240 2879
4241 /** !! 2880 lock_map_acquire(&work->lockdep_map);
4242 * flush_delayed_work - wait for a dwork to f !! 2881 lock_map_release(&work->lockdep_map);
4243 * @dwork: the delayed work to flush <<
4244 * <<
4245 * Delayed timer is cancelled and the pending <<
4246 * immediate execution. Like flush_work(), t <<
4247 * considers the last queueing instance of @d <<
4248 * <<
4249 * Return: <<
4250 * %true if flush_work() waited for the work <<
4251 * %false if it was already idle. <<
4252 */ <<
4253 bool flush_delayed_work(struct delayed_work * <<
4254 { <<
4255 local_irq_disable(); <<
4256 if (del_timer_sync(&dwork->timer)) <<
4257 __queue_work(dwork->cpu, dwor <<
4258 local_irq_enable(); <<
4259 return flush_work(&dwork->work); <<
4260 } <<
4261 EXPORT_SYMBOL(flush_delayed_work); <<
4262 2882
4263 /** !! 2883 if (start_flush_work(work, &barr)) {
4264 * flush_rcu_work - wait for a rwork to finis !! 2884 wait_for_completion(&barr.done);
4265 * @rwork: the rcu work to flush !! 2885 destroy_work_on_stack(&barr.work);
4266 * <<
4267 * Return: <<
4268 * %true if flush_rcu_work() waited for the w <<
4269 * %false if it was already idle. <<
4270 */ <<
4271 bool flush_rcu_work(struct rcu_work *rwork) <<
4272 { <<
4273 if (test_bit(WORK_STRUCT_PENDING_BIT, <<
4274 rcu_barrier(); <<
4275 flush_work(&rwork->work); <<
4276 return true; 2886 return true;
4277 } else { 2887 } else {
4278 return flush_work(&rwork->wor !! 2888 return false;
4279 } 2889 }
4280 } 2890 }
4281 EXPORT_SYMBOL(flush_rcu_work); !! 2891 EXPORT_SYMBOL_GPL(flush_work);
4282 <<
4283 static void work_offqd_disable(struct work_of <<
4284 { <<
4285 const unsigned long max = (1lu << WOR <<
4286 2892
4287 if (likely(offqd->disable < max)) !! 2893 struct cwt_wait {
4288 offqd->disable++; !! 2894 wait_queue_t wait;
4289 else !! 2895 struct work_struct *work;
4290 WARN_ONCE(true, "workqueue: w !! 2896 };
4291 } <<
4292 2897
4293 static void work_offqd_enable(struct work_off !! 2898 static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key)
4294 { 2899 {
4295 if (likely(offqd->disable > 0)) !! 2900 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
4296 offqd->disable--; !! 2901
4297 else !! 2902 if (cwait->work != key)
4298 WARN_ONCE(true, "workqueue: w !! 2903 return 0;
>> 2904 return autoremove_wake_function(wait, mode, sync, key);
4299 } 2905 }
4300 2906
4301 static bool __cancel_work(struct work_struct !! 2907 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
4302 { 2908 {
4303 struct work_offq_data offqd; !! 2909 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
4304 unsigned long irq_flags; !! 2910 unsigned long flags;
4305 int ret; 2911 int ret;
4306 2912
4307 ret = work_grab_pending(work, cflags, !! 2913 do {
4308 !! 2914 ret = try_to_grab_pending(work, is_dwork, &flags);
4309 work_offqd_unpack(&offqd, *work_data_ !! 2915 /*
4310 !! 2916 * If someone else is already canceling, wait for it to
4311 if (cflags & WORK_CANCEL_DISABLE) !! 2917 * finish. flush_work() doesn't work for PREEMPT_NONE
4312 work_offqd_disable(&offqd); !! 2918 * because we may get scheduled between @work's completion
>> 2919 * and the other canceling task resuming and clearing
>> 2920 * CANCELING - flush_work() will return false immediately
>> 2921 * as @work is no longer busy, try_to_grab_pending() will
>> 2922 * return -ENOENT as @work is still being canceled and the
>> 2923 * other canceling task won't be able to clear CANCELING as
>> 2924 * we're hogging the CPU.
>> 2925 *
>> 2926 * Let's wait for completion using a waitqueue. As this
>> 2927 * may lead to the thundering herd problem, use a custom
>> 2928 * wake function which matches @work along with exclusive
>> 2929 * wait and wakeup.
>> 2930 */
>> 2931 if (unlikely(ret == -ENOENT)) {
>> 2932 struct cwt_wait cwait;
4313 2933
4314 set_work_pool_and_clear_pending(work, !! 2934 init_wait(&cwait.wait);
4315 work_ !! 2935 cwait.wait.func = cwt_wakefn;
4316 local_irq_restore(irq_flags); !! 2936 cwait.work = work;
4317 return ret; <<
4318 } <<
4319 2937
4320 static bool __cancel_work_sync(struct work_st !! 2938 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
4321 { !! 2939 TASK_UNINTERRUPTIBLE);
4322 bool ret; !! 2940 if (work_is_canceling(work))
>> 2941 schedule();
>> 2942 finish_wait(&cancel_waitq, &cwait.wait);
>> 2943 }
>> 2944 } while (unlikely(ret < 0));
4323 2945
4324 ret = __cancel_work(work, cflags | WO !! 2946 /* tell other tasks trying to grab @work to back off */
>> 2947 mark_work_canceling(work);
>> 2948 local_irq_restore(flags);
4325 2949
4326 if (*work_data_bits(work) & WORK_OFFQ !! 2950 flush_work(work);
4327 WARN_ON_ONCE(in_hardirq()); !! 2951 clear_work_data(work);
4328 else <<
4329 might_sleep(); <<
4330 2952
4331 /* 2953 /*
4332 * Skip __flush_work() during early b !! 2954 * Paired with prepare_to_wait() above so that either
4333 * executing. This allows canceling d !! 2955 * waitqueue_active() is visible here or !work_is_canceling() is
>> 2956 * visible there.
4334 */ 2957 */
4335 if (wq_online) !! 2958 smp_mb();
4336 __flush_work(work, true); !! 2959 if (waitqueue_active(&cancel_waitq))
4337 !! 2960 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
4338 if (!(cflags & WORK_CANCEL_DISABLE)) <<
4339 enable_work(work); <<
4340 2961
4341 return ret; 2962 return ret;
4342 } 2963 }
4343 2964
4344 /* <<
4345 * See cancel_delayed_work() <<
4346 */ <<
4347 bool cancel_work(struct work_struct *work) <<
4348 { <<
4349 return __cancel_work(work, 0); <<
4350 } <<
4351 EXPORT_SYMBOL(cancel_work); <<
4352 <<
4353 /** 2965 /**
4354 * cancel_work_sync - cancel a work and wait 2966 * cancel_work_sync - cancel a work and wait for it to finish
4355 * @work: the work to cancel 2967 * @work: the work to cancel
4356 * 2968 *
4357 * Cancel @work and wait for its execution to !! 2969 * Cancel @work and wait for its execution to finish. This function
4358 * even if the work re-queues itself or migra !! 2970 * can be used even if the work re-queues itself or migrates to
4359 * from this function, @work is guaranteed to !! 2971 * another workqueue. On return from this function, @work is
4360 * CPU as long as there aren't racing enqueue !! 2972 * guaranteed to be not pending or executing on any CPU.
4361 * !! 2973 *
4362 * cancel_work_sync(&delayed_work->work) must !! 2974 * cancel_work_sync(&delayed_work->work) must not be used for
4363 * Use cancel_delayed_work_sync() instead. !! 2975 * delayed_work's. Use cancel_delayed_work_sync() instead.
4364 * <<
4365 * Must be called from a sleepable context if <<
4366 * workqueue. Can also be called from non-har <<
4367 * if @work was last queued on a BH workqueue <<
4368 * 2976 *
4369 * Returns %true if @work was pending, %false !! 2977 * The caller must ensure that the workqueue on which @work was last
>> 2978 * queued can't be destroyed before this function returns.
>> 2979 *
>> 2980 * RETURNS:
>> 2981 * %true if @work was pending, %false otherwise.
4370 */ 2982 */
4371 bool cancel_work_sync(struct work_struct *wor 2983 bool cancel_work_sync(struct work_struct *work)
4372 { 2984 {
4373 return __cancel_work_sync(work, 0); !! 2985 return __cancel_work_timer(work, false);
4374 } 2986 }
4375 EXPORT_SYMBOL_GPL(cancel_work_sync); 2987 EXPORT_SYMBOL_GPL(cancel_work_sync);
4376 2988
4377 /** 2989 /**
4378 * cancel_delayed_work - cancel a delayed wor !! 2990 * flush_delayed_work - wait for a dwork to finish executing the last queueing
4379 * @dwork: delayed_work to cancel !! 2991 * @dwork: the delayed work to flush
4380 * 2992 *
4381 * Kill off a pending delayed_work. !! 2993 * Delayed timer is cancelled and the pending work is queued for
>> 2994 * immediate execution. Like flush_work(), this function only
>> 2995 * considers the last queueing instance of @dwork.
4382 * 2996 *
4383 * Return: %true if @dwork was pending and ca !! 2997 * RETURNS:
4384 * pending. !! 2998 * %true if flush_work() waited for the work to finish execution,
>> 2999 * %false if it was already idle.
>> 3000 */
>> 3001 bool flush_delayed_work(struct delayed_work *dwork)
>> 3002 {
>> 3003 local_irq_disable();
>> 3004 if (del_timer_sync(&dwork->timer))
>> 3005 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
>> 3006 local_irq_enable();
>> 3007 return flush_work(&dwork->work);
>> 3008 }
>> 3009 EXPORT_SYMBOL(flush_delayed_work);
>> 3010
>> 3011 /**
>> 3012 * cancel_delayed_work - cancel a delayed work
>> 3013 * @dwork: delayed_work to cancel
4385 * 3014 *
4386 * Note: !! 3015 * Kill off a pending delayed_work. Returns %true if @dwork was pending
4387 * The work callback function may still be ru !! 3016 * and canceled; %false if wasn't pending. Note that the work callback
4388 * it returns %true and the work doesn't re-a !! 3017 * function may still be running on return, unless it returns %true and the
4389 * use cancel_delayed_work_sync() to wait on !! 3018 * work doesn't re-arm itself. Explicitly flush or use
>> 3019 * cancel_delayed_work_sync() to wait on it.
4390 * 3020 *
4391 * This function is safe to call from any con 3021 * This function is safe to call from any context including IRQ handler.
4392 */ 3022 */
4393 bool cancel_delayed_work(struct delayed_work 3023 bool cancel_delayed_work(struct delayed_work *dwork)
4394 { 3024 {
4395 return __cancel_work(&dwork->work, WO !! 3025 unsigned long flags;
>> 3026 int ret;
>> 3027
>> 3028 do {
>> 3029 ret = try_to_grab_pending(&dwork->work, true, &flags);
>> 3030 } while (unlikely(ret == -EAGAIN));
>> 3031
>> 3032 if (unlikely(ret < 0))
>> 3033 return false;
>> 3034
>> 3035 set_work_pool_and_clear_pending(&dwork->work,
>> 3036 get_work_pool_id(&dwork->work));
>> 3037 local_irq_restore(flags);
>> 3038 return ret;
4396 } 3039 }
4397 EXPORT_SYMBOL(cancel_delayed_work); 3040 EXPORT_SYMBOL(cancel_delayed_work);
4398 3041
4399 /** 3042 /**
4400 * cancel_delayed_work_sync - cancel a delaye 3043 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
4401 * @dwork: the delayed work cancel 3044 * @dwork: the delayed work cancel
4402 * 3045 *
4403 * This is cancel_work_sync() for delayed wor 3046 * This is cancel_work_sync() for delayed works.
4404 * 3047 *
4405 * Return: !! 3048 * RETURNS:
4406 * %true if @dwork was pending, %false otherw 3049 * %true if @dwork was pending, %false otherwise.
4407 */ 3050 */
4408 bool cancel_delayed_work_sync(struct delayed_ 3051 bool cancel_delayed_work_sync(struct delayed_work *dwork)
4409 { 3052 {
4410 return __cancel_work_sync(&dwork->wor !! 3053 return __cancel_work_timer(&dwork->work, true);
4411 } 3054 }
4412 EXPORT_SYMBOL(cancel_delayed_work_sync); 3055 EXPORT_SYMBOL(cancel_delayed_work_sync);
4413 3056
4414 /** 3057 /**
4415 * disable_work - Disable and cancel a work i <<
4416 * @work: work item to disable <<
4417 * <<
4418 * Disable @work by incrementing its disable <<
4419 * pending. As long as the disable count is n <<
4420 * will fail and return %false. The maximum s <<
4421 * power of %WORK_OFFQ_DISABLE_BITS, currentl <<
4422 * <<
4423 * Can be called from any context. Returns %t <<
4424 * otherwise. <<
4425 */ <<
4426 bool disable_work(struct work_struct *work) <<
4427 { <<
4428 return __cancel_work(work, WORK_CANCE <<
4429 } <<
4430 EXPORT_SYMBOL_GPL(disable_work); <<
4431 <<
4432 /** <<
4433 * disable_work_sync - Disable, cancel and dr <<
4434 * @work: work item to disable <<
4435 * <<
4436 * Similar to disable_work() but also wait fo <<
4437 * executing. <<
4438 * <<
4439 * Must be called from a sleepable context if <<
4440 * workqueue. Can also be called from non-har <<
4441 * if @work was last queued on a BH workqueue <<
4442 * <<
4443 * Returns %true if @work was pending, %false <<
4444 */ <<
4445 bool disable_work_sync(struct work_struct *wo <<
4446 { <<
4447 return __cancel_work_sync(work, WORK_ <<
4448 } <<
4449 EXPORT_SYMBOL_GPL(disable_work_sync); <<
4450 <<
4451 /** <<
4452 * enable_work - Enable a work item <<
4453 * @work: work item to enable <<
4454 * <<
4455 * Undo disable_work[_sync]() by decrementing <<
4456 * only be queued if its disable count is 0. <<
4457 * <<
4458 * Can be called from any context. Returns %t <<
4459 * Otherwise, %false. <<
4460 */ <<
4461 bool enable_work(struct work_struct *work) <<
4462 { <<
4463 struct work_offq_data offqd; <<
4464 unsigned long irq_flags; <<
4465 <<
4466 work_grab_pending(work, 0, &irq_flags <<
4467 <<
4468 work_offqd_unpack(&offqd, *work_data_ <<
4469 work_offqd_enable(&offqd); <<
4470 set_work_pool_and_clear_pending(work, <<
4471 work_ <<
4472 local_irq_restore(irq_flags); <<
4473 <<
4474 return !offqd.disable; <<
4475 } <<
4476 EXPORT_SYMBOL_GPL(enable_work); <<
4477 <<
4478 /** <<
4479 * disable_delayed_work - Disable and cancel <<
4480 * @dwork: delayed work item to disable <<
4481 * <<
4482 * disable_work() for delayed work items. <<
4483 */ <<
4484 bool disable_delayed_work(struct delayed_work <<
4485 { <<
4486 return __cancel_work(&dwork->work, <<
4487 WORK_CANCEL_DELA <<
4488 } <<
4489 EXPORT_SYMBOL_GPL(disable_delayed_work); <<
4490 <<
4491 /** <<
4492 * disable_delayed_work_sync - Disable, cance <<
4493 * @dwork: delayed work item to disable <<
4494 * <<
4495 * disable_work_sync() for delayed work items <<
4496 */ <<
4497 bool disable_delayed_work_sync(struct delayed <<
4498 { <<
4499 return __cancel_work_sync(&dwork->wor <<
4500 WORK_CANCEL <<
4501 } <<
4502 EXPORT_SYMBOL_GPL(disable_delayed_work_sync); <<
4503 <<
4504 /** <<
4505 * enable_delayed_work - Enable a delayed wor <<
4506 * @dwork: delayed work item to enable <<
4507 * <<
4508 * enable_work() for delayed work items. <<
4509 */ <<
4510 bool enable_delayed_work(struct delayed_work <<
4511 { <<
4512 return enable_work(&dwork->work); <<
4513 } <<
4514 EXPORT_SYMBOL_GPL(enable_delayed_work); <<
4515 <<
4516 /** <<
4517 * schedule_on_each_cpu - execute a function 3058 * schedule_on_each_cpu - execute a function synchronously on each online CPU
4518 * @func: the function to call 3059 * @func: the function to call
4519 * 3060 *
4520 * schedule_on_each_cpu() executes @func on e 3061 * schedule_on_each_cpu() executes @func on each online CPU using the
4521 * system workqueue and blocks until all CPUs 3062 * system workqueue and blocks until all CPUs have completed.
4522 * schedule_on_each_cpu() is very slow. 3063 * schedule_on_each_cpu() is very slow.
4523 * 3064 *
4524 * Return: !! 3065 * RETURNS:
4525 * 0 on success, -errno on failure. 3066 * 0 on success, -errno on failure.
4526 */ 3067 */
4527 int schedule_on_each_cpu(work_func_t func) 3068 int schedule_on_each_cpu(work_func_t func)
4528 { 3069 {
4529 int cpu; 3070 int cpu;
4530 struct work_struct __percpu *works; 3071 struct work_struct __percpu *works;
4531 3072
4532 works = alloc_percpu(struct work_stru 3073 works = alloc_percpu(struct work_struct);
4533 if (!works) 3074 if (!works)
4534 return -ENOMEM; 3075 return -ENOMEM;
4535 3076
4536 cpus_read_lock(); !! 3077 get_online_cpus();
4537 3078
4538 for_each_online_cpu(cpu) { 3079 for_each_online_cpu(cpu) {
4539 struct work_struct *work = pe 3080 struct work_struct *work = per_cpu_ptr(works, cpu);
4540 3081
4541 INIT_WORK(work, func); 3082 INIT_WORK(work, func);
4542 schedule_work_on(cpu, work); 3083 schedule_work_on(cpu, work);
4543 } 3084 }
4544 3085
4545 for_each_online_cpu(cpu) 3086 for_each_online_cpu(cpu)
4546 flush_work(per_cpu_ptr(works, 3087 flush_work(per_cpu_ptr(works, cpu));
4547 3088
4548 cpus_read_unlock(); !! 3089 put_online_cpus();
4549 free_percpu(works); 3090 free_percpu(works);
4550 return 0; 3091 return 0;
4551 } 3092 }
4552 3093
4553 /** 3094 /**
>> 3095 * flush_scheduled_work - ensure that any scheduled work has run to completion.
>> 3096 *
>> 3097 * Forces execution of the kernel-global workqueue and blocks until its
>> 3098 * completion.
>> 3099 *
>> 3100 * Think twice before calling this function! It's very easy to get into
>> 3101 * trouble if you don't take great care. Either of the following situations
>> 3102 * will lead to deadlock:
>> 3103 *
>> 3104 * One of the work items currently on the workqueue needs to acquire
>> 3105 * a lock held by your code or its caller.
>> 3106 *
>> 3107 * Your code is running in the context of a work routine.
>> 3108 *
>> 3109 * They will be detected by lockdep when they occur, but the first might not
>> 3110 * occur very often. It depends on what work items are on the workqueue and
>> 3111 * what locks they need, which you have no control over.
>> 3112 *
>> 3113 * In most situations flushing the entire workqueue is overkill; you merely
>> 3114 * need to know that a particular work item isn't queued and isn't running.
>> 3115 * In such cases you should use cancel_delayed_work_sync() or
>> 3116 * cancel_work_sync() instead.
>> 3117 */
>> 3118 void flush_scheduled_work(void)
>> 3119 {
>> 3120 flush_workqueue(system_wq);
>> 3121 }
>> 3122 EXPORT_SYMBOL(flush_scheduled_work);
>> 3123
>> 3124 /**
4554 * execute_in_process_context - reliably exec 3125 * execute_in_process_context - reliably execute the routine with user context
4555 * @fn: the function to execute 3126 * @fn: the function to execute
4556 * @ew: guaranteed storage for the ex 3127 * @ew: guaranteed storage for the execute work structure (must
4557 * be available when the work ex 3128 * be available when the work executes)
4558 * 3129 *
4559 * Executes the function immediately if proce 3130 * Executes the function immediately if process context is available,
4560 * otherwise schedules the function for delay 3131 * otherwise schedules the function for delayed execution.
4561 * 3132 *
4562 * Return: 0 - function was executed !! 3133 * Returns: 0 - function was executed
4563 * 1 - function was scheduled fo 3134 * 1 - function was scheduled for execution
4564 */ 3135 */
4565 int execute_in_process_context(work_func_t fn 3136 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
4566 { 3137 {
4567 if (!in_interrupt()) { 3138 if (!in_interrupt()) {
4568 fn(&ew->work); 3139 fn(&ew->work);
4569 return 0; 3140 return 0;
4570 } 3141 }
4571 3142
4572 INIT_WORK(&ew->work, fn); 3143 INIT_WORK(&ew->work, fn);
4573 schedule_work(&ew->work); 3144 schedule_work(&ew->work);
4574 3145
4575 return 1; 3146 return 1;
4576 } 3147 }
4577 EXPORT_SYMBOL_GPL(execute_in_process_context) 3148 EXPORT_SYMBOL_GPL(execute_in_process_context);
4578 3149
>> 3150 #ifdef CONFIG_SYSFS
>> 3151 /*
>> 3152 * Workqueues with WQ_SYSFS flag set is visible to userland via
>> 3153 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
>> 3154 * following attributes.
>> 3155 *
>> 3156 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
>> 3157 * max_active RW int : maximum number of in-flight work items
>> 3158 *
>> 3159 * Unbound workqueues have the following extra attributes.
>> 3160 *
>> 3161 * id RO int : the associated pool ID
>> 3162 * nice RW int : nice value of the workers
>> 3163 * cpumask RW mask : bitmask of allowed CPUs for the workers
>> 3164 */
>> 3165 struct wq_device {
>> 3166 struct workqueue_struct *wq;
>> 3167 struct device dev;
>> 3168 };
>> 3169
>> 3170 static struct workqueue_struct *dev_to_wq(struct device *dev)
>> 3171 {
>> 3172 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
>> 3173
>> 3174 return wq_dev->wq;
>> 3175 }
>> 3176
>> 3177 static ssize_t wq_per_cpu_show(struct device *dev,
>> 3178 struct device_attribute *attr, char *buf)
>> 3179 {
>> 3180 struct workqueue_struct *wq = dev_to_wq(dev);
>> 3181
>> 3182 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
>> 3183 }
>> 3184
>> 3185 static ssize_t wq_max_active_show(struct device *dev,
>> 3186 struct device_attribute *attr, char *buf)
>> 3187 {
>> 3188 struct workqueue_struct *wq = dev_to_wq(dev);
>> 3189
>> 3190 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
>> 3191 }
>> 3192
>> 3193 static ssize_t wq_max_active_store(struct device *dev,
>> 3194 struct device_attribute *attr,
>> 3195 const char *buf, size_t count)
>> 3196 {
>> 3197 struct workqueue_struct *wq = dev_to_wq(dev);
>> 3198 int val;
>> 3199
>> 3200 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
>> 3201 return -EINVAL;
>> 3202
>> 3203 workqueue_set_max_active(wq, val);
>> 3204 return count;
>> 3205 }
>> 3206
>> 3207 static struct device_attribute wq_sysfs_attrs[] = {
>> 3208 __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL),
>> 3209 __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store),
>> 3210 __ATTR_NULL,
>> 3211 };
>> 3212
>> 3213 static ssize_t wq_pool_ids_show(struct device *dev,
>> 3214 struct device_attribute *attr, char *buf)
>> 3215 {
>> 3216 struct workqueue_struct *wq = dev_to_wq(dev);
>> 3217 const char *delim = "";
>> 3218 int node, written = 0;
>> 3219
>> 3220 rcu_read_lock_sched();
>> 3221 for_each_node(node) {
>> 3222 written += scnprintf(buf + written, PAGE_SIZE - written,
>> 3223 "%s%d:%d", delim, node,
>> 3224 unbound_pwq_by_node(wq, node)->pool->id);
>> 3225 delim = " ";
>> 3226 }
>> 3227 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
>> 3228 rcu_read_unlock_sched();
>> 3229
>> 3230 return written;
>> 3231 }
>> 3232
>> 3233 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
>> 3234 char *buf)
>> 3235 {
>> 3236 struct workqueue_struct *wq = dev_to_wq(dev);
>> 3237 int written;
>> 3238
>> 3239 mutex_lock(&wq->mutex);
>> 3240 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
>> 3241 mutex_unlock(&wq->mutex);
>> 3242
>> 3243 return written;
>> 3244 }
>> 3245
>> 3246 /* prepare workqueue_attrs for sysfs store operations */
>> 3247 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
>> 3248 {
>> 3249 struct workqueue_attrs *attrs;
>> 3250
>> 3251 attrs = alloc_workqueue_attrs(GFP_KERNEL);
>> 3252 if (!attrs)
>> 3253 return NULL;
>> 3254
>> 3255 mutex_lock(&wq->mutex);
>> 3256 copy_workqueue_attrs(attrs, wq->unbound_attrs);
>> 3257 mutex_unlock(&wq->mutex);
>> 3258 return attrs;
>> 3259 }
>> 3260
>> 3261 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
>> 3262 const char *buf, size_t count)
>> 3263 {
>> 3264 struct workqueue_struct *wq = dev_to_wq(dev);
>> 3265 struct workqueue_attrs *attrs;
>> 3266 int ret;
>> 3267
>> 3268 attrs = wq_sysfs_prep_attrs(wq);
>> 3269 if (!attrs)
>> 3270 return -ENOMEM;
>> 3271
>> 3272 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
>> 3273 attrs->nice >= -20 && attrs->nice <= 19)
>> 3274 ret = apply_workqueue_attrs(wq, attrs);
>> 3275 else
>> 3276 ret = -EINVAL;
>> 3277
>> 3278 free_workqueue_attrs(attrs);
>> 3279 return ret ?: count;
>> 3280 }
>> 3281
>> 3282 static ssize_t wq_cpumask_show(struct device *dev,
>> 3283 struct device_attribute *attr, char *buf)
>> 3284 {
>> 3285 struct workqueue_struct *wq = dev_to_wq(dev);
>> 3286 int written;
>> 3287
>> 3288 mutex_lock(&wq->mutex);
>> 3289 written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask);
>> 3290 mutex_unlock(&wq->mutex);
>> 3291
>> 3292 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
>> 3293 return written;
>> 3294 }
>> 3295
>> 3296 static ssize_t wq_cpumask_store(struct device *dev,
>> 3297 struct device_attribute *attr,
>> 3298 const char *buf, size_t count)
>> 3299 {
>> 3300 struct workqueue_struct *wq = dev_to_wq(dev);
>> 3301 struct workqueue_attrs *attrs;
>> 3302 int ret;
>> 3303
>> 3304 attrs = wq_sysfs_prep_attrs(wq);
>> 3305 if (!attrs)
>> 3306 return -ENOMEM;
>> 3307
>> 3308 ret = cpumask_parse(buf, attrs->cpumask);
>> 3309 if (!ret)
>> 3310 ret = apply_workqueue_attrs(wq, attrs);
>> 3311
>> 3312 free_workqueue_attrs(attrs);
>> 3313 return ret ?: count;
>> 3314 }
>> 3315
>> 3316 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
>> 3317 char *buf)
>> 3318 {
>> 3319 struct workqueue_struct *wq = dev_to_wq(dev);
>> 3320 int written;
>> 3321
>> 3322 mutex_lock(&wq->mutex);
>> 3323 written = scnprintf(buf, PAGE_SIZE, "%d\n",
>> 3324 !wq->unbound_attrs->no_numa);
>> 3325 mutex_unlock(&wq->mutex);
>> 3326
>> 3327 return written;
>> 3328 }
>> 3329
>> 3330 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
>> 3331 const char *buf, size_t count)
>> 3332 {
>> 3333 struct workqueue_struct *wq = dev_to_wq(dev);
>> 3334 struct workqueue_attrs *attrs;
>> 3335 int v, ret;
>> 3336
>> 3337 attrs = wq_sysfs_prep_attrs(wq);
>> 3338 if (!attrs)
>> 3339 return -ENOMEM;
>> 3340
>> 3341 ret = -EINVAL;
>> 3342 if (sscanf(buf, "%d", &v) == 1) {
>> 3343 attrs->no_numa = !v;
>> 3344 ret = apply_workqueue_attrs(wq, attrs);
>> 3345 }
>> 3346
>> 3347 free_workqueue_attrs(attrs);
>> 3348 return ret ?: count;
>> 3349 }
>> 3350
>> 3351 static struct device_attribute wq_sysfs_unbound_attrs[] = {
>> 3352 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
>> 3353 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
>> 3354 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
>> 3355 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
>> 3356 __ATTR_NULL,
>> 3357 };
>> 3358
>> 3359 static struct bus_type wq_subsys = {
>> 3360 .name = "workqueue",
>> 3361 .dev_attrs = wq_sysfs_attrs,
>> 3362 };
>> 3363
>> 3364 static int __init wq_sysfs_init(void)
>> 3365 {
>> 3366 return subsys_virtual_register(&wq_subsys, NULL);
>> 3367 }
>> 3368 core_initcall(wq_sysfs_init);
>> 3369
>> 3370 static void wq_device_release(struct device *dev)
>> 3371 {
>> 3372 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
>> 3373
>> 3374 kfree(wq_dev);
>> 3375 }
>> 3376
>> 3377 /**
>> 3378 * workqueue_sysfs_register - make a workqueue visible in sysfs
>> 3379 * @wq: the workqueue to register
>> 3380 *
>> 3381 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
>> 3382 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
>> 3383 * which is the preferred method.
>> 3384 *
>> 3385 * Workqueue user should use this function directly iff it wants to apply
>> 3386 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
>> 3387 * apply_workqueue_attrs() may race against userland updating the
>> 3388 * attributes.
>> 3389 *
>> 3390 * Returns 0 on success, -errno on failure.
>> 3391 */
>> 3392 int workqueue_sysfs_register(struct workqueue_struct *wq)
>> 3393 {
>> 3394 struct wq_device *wq_dev;
>> 3395 int ret;
>> 3396
>> 3397 /*
>> 3398 * Adjusting max_active or creating new pwqs by applyting
>> 3399 * attributes breaks ordering guarantee. Disallow exposing ordered
>> 3400 * workqueues.
>> 3401 */
>> 3402 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
>> 3403 return -EINVAL;
>> 3404
>> 3405 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
>> 3406 if (!wq_dev)
>> 3407 return -ENOMEM;
>> 3408
>> 3409 wq_dev->wq = wq;
>> 3410 wq_dev->dev.bus = &wq_subsys;
>> 3411 wq_dev->dev.init_name = wq->name;
>> 3412 wq_dev->dev.release = wq_device_release;
>> 3413
>> 3414 /*
>> 3415 * unbound_attrs are created separately. Suppress uevent until
>> 3416 * everything is ready.
>> 3417 */
>> 3418 dev_set_uevent_suppress(&wq_dev->dev, true);
>> 3419
>> 3420 ret = device_register(&wq_dev->dev);
>> 3421 if (ret) {
>> 3422 kfree(wq_dev);
>> 3423 wq->wq_dev = NULL;
>> 3424 return ret;
>> 3425 }
>> 3426
>> 3427 if (wq->flags & WQ_UNBOUND) {
>> 3428 struct device_attribute *attr;
>> 3429
>> 3430 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
>> 3431 ret = device_create_file(&wq_dev->dev, attr);
>> 3432 if (ret) {
>> 3433 device_unregister(&wq_dev->dev);
>> 3434 wq->wq_dev = NULL;
>> 3435 return ret;
>> 3436 }
>> 3437 }
>> 3438 }
>> 3439
>> 3440 dev_set_uevent_suppress(&wq_dev->dev, false);
>> 3441 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
>> 3442 return 0;
>> 3443 }
>> 3444
>> 3445 /**
>> 3446 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
>> 3447 * @wq: the workqueue to unregister
>> 3448 *
>> 3449 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
>> 3450 */
>> 3451 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
>> 3452 {
>> 3453 struct wq_device *wq_dev = wq->wq_dev;
>> 3454
>> 3455 if (!wq->wq_dev)
>> 3456 return;
>> 3457
>> 3458 wq->wq_dev = NULL;
>> 3459 device_unregister(&wq_dev->dev);
>> 3460 }
>> 3461 #else /* CONFIG_SYSFS */
>> 3462 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
>> 3463 #endif /* CONFIG_SYSFS */
>> 3464
4579 /** 3465 /**
4580 * free_workqueue_attrs - free a workqueue_at 3466 * free_workqueue_attrs - free a workqueue_attrs
4581 * @attrs: workqueue_attrs to free 3467 * @attrs: workqueue_attrs to free
4582 * 3468 *
4583 * Undo alloc_workqueue_attrs(). 3469 * Undo alloc_workqueue_attrs().
4584 */ 3470 */
4585 void free_workqueue_attrs(struct workqueue_at 3471 void free_workqueue_attrs(struct workqueue_attrs *attrs)
4586 { 3472 {
4587 if (attrs) { 3473 if (attrs) {
4588 free_cpumask_var(attrs->cpuma 3474 free_cpumask_var(attrs->cpumask);
4589 free_cpumask_var(attrs->__pod <<
4590 kfree(attrs); 3475 kfree(attrs);
4591 } 3476 }
4592 } 3477 }
4593 3478
4594 /** 3479 /**
4595 * alloc_workqueue_attrs - allocate a workque 3480 * alloc_workqueue_attrs - allocate a workqueue_attrs
>> 3481 * @gfp_mask: allocation mask to use
4596 * 3482 *
4597 * Allocate a new workqueue_attrs, initialize 3483 * Allocate a new workqueue_attrs, initialize with default settings and
4598 * return it. !! 3484 * return it. Returns NULL on failure.
4599 * <<
4600 * Return: The allocated new workqueue_attr o <<
4601 */ 3485 */
4602 struct workqueue_attrs *alloc_workqueue_attrs !! 3486 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
4603 { 3487 {
4604 struct workqueue_attrs *attrs; 3488 struct workqueue_attrs *attrs;
4605 3489
4606 attrs = kzalloc(sizeof(*attrs), GFP_K !! 3490 attrs = kzalloc(sizeof(*attrs), gfp_mask);
4607 if (!attrs) 3491 if (!attrs)
4608 goto fail; 3492 goto fail;
4609 if (!alloc_cpumask_var(&attrs->cpumas !! 3493 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
4610 goto fail; <<
4611 if (!alloc_cpumask_var(&attrs->__pod_ <<
4612 goto fail; 3494 goto fail;
4613 3495
4614 cpumask_copy(attrs->cpumask, cpu_poss 3496 cpumask_copy(attrs->cpumask, cpu_possible_mask);
4615 attrs->affn_scope = WQ_AFFN_DFL; <<
4616 return attrs; 3497 return attrs;
4617 fail: 3498 fail:
4618 free_workqueue_attrs(attrs); 3499 free_workqueue_attrs(attrs);
4619 return NULL; 3500 return NULL;
4620 } 3501 }
4621 3502
4622 static void copy_workqueue_attrs(struct workq 3503 static void copy_workqueue_attrs(struct workqueue_attrs *to,
4623 const struct 3504 const struct workqueue_attrs *from)
4624 { 3505 {
4625 to->nice = from->nice; 3506 to->nice = from->nice;
4626 cpumask_copy(to->cpumask, from->cpuma 3507 cpumask_copy(to->cpumask, from->cpumask);
4627 cpumask_copy(to->__pod_cpumask, from- <<
4628 to->affn_strict = from->affn_strict; <<
4629 <<
4630 /* 3508 /*
4631 * Unlike hash and equality test, cop !! 3509 * Unlike hash and equality test, this function doesn't ignore
4632 * fields as copying is used for both !! 3510 * ->no_numa as it is used for both pool and wq attrs. Instead,
4633 * get_unbound_pool() explicitly clea !! 3511 * get_unbound_pool() explicitly clears ->no_numa after copying.
4634 */ 3512 */
4635 to->affn_scope = from->affn_scope; !! 3513 to->no_numa = from->no_numa;
4636 to->ordered = from->ordered; <<
4637 } <<
4638 <<
4639 /* <<
4640 * Some attrs fields are workqueue-only. Clea <<
4641 * comments in 'struct workqueue_attrs' defin <<
4642 */ <<
4643 static void wqattrs_clear_for_pool(struct wor <<
4644 { <<
4645 attrs->affn_scope = WQ_AFFN_NR_TYPES; <<
4646 attrs->ordered = false; <<
4647 if (attrs->affn_strict) <<
4648 cpumask_copy(attrs->cpumask, <<
4649 } 3514 }
4650 3515
4651 /* hash value of the content of @attr */ 3516 /* hash value of the content of @attr */
4652 static u32 wqattrs_hash(const struct workqueu 3517 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
4653 { 3518 {
4654 u32 hash = 0; 3519 u32 hash = 0;
4655 3520
4656 hash = jhash_1word(attrs->nice, hash) 3521 hash = jhash_1word(attrs->nice, hash);
4657 hash = jhash_1word(attrs->affn_strict !! 3522 hash = jhash(cpumask_bits(attrs->cpumask),
4658 hash = jhash(cpumask_bits(attrs->__po <<
4659 BITS_TO_LONGS(nr_cpumask 3523 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4660 if (!attrs->affn_strict) <<
4661 hash = jhash(cpumask_bits(att <<
4662 BITS_TO_LONGS(nr <<
4663 return hash; 3524 return hash;
4664 } 3525 }
4665 3526
4666 /* content equality test */ 3527 /* content equality test */
4667 static bool wqattrs_equal(const struct workqu 3528 static bool wqattrs_equal(const struct workqueue_attrs *a,
4668 const struct workqu 3529 const struct workqueue_attrs *b)
4669 { 3530 {
4670 if (a->nice != b->nice) 3531 if (a->nice != b->nice)
4671 return false; 3532 return false;
4672 if (a->affn_strict != b->affn_strict) !! 3533 if (!cpumask_equal(a->cpumask, b->cpumask))
4673 return false; <<
4674 if (!cpumask_equal(a->__pod_cpumask, <<
4675 return false; <<
4676 if (!a->affn_strict && !cpumask_equal <<
4677 return false; 3534 return false;
4678 return true; 3535 return true;
4679 } 3536 }
4680 3537
4681 /* Update @attrs with actually available CPUs <<
4682 static void wqattrs_actualize_cpumask(struct <<
4683 const c <<
4684 { <<
4685 /* <<
4686 * Calculate the effective CPU mask o <<
4687 * @attrs->cpumask doesn't overlap wi <<
4688 * @unbound_cpumask. <<
4689 */ <<
4690 cpumask_and(attrs->cpumask, attrs->cp <<
4691 if (unlikely(cpumask_empty(attrs->cpu <<
4692 cpumask_copy(attrs->cpumask, <<
4693 } <<
4694 <<
4695 /* find wq_pod_type to use for @attrs */ <<
4696 static const struct wq_pod_type * <<
4697 wqattrs_pod_type(const struct workqueue_attrs <<
4698 { <<
4699 enum wq_affn_scope scope; <<
4700 struct wq_pod_type *pt; <<
4701 <<
4702 /* to synchronize access to wq_affn_d <<
4703 lockdep_assert_held(&wq_pool_mutex); <<
4704 <<
4705 if (attrs->affn_scope == WQ_AFFN_DFL) <<
4706 scope = wq_affn_dfl; <<
4707 else <<
4708 scope = attrs->affn_scope; <<
4709 <<
4710 pt = &wq_pod_types[scope]; <<
4711 <<
4712 if (!WARN_ON_ONCE(attrs->affn_scope = <<
4713 likely(pt->nr_pods)) <<
4714 return pt; <<
4715 <<
4716 /* <<
4717 * Before workqueue_init_topology(), <<
4718 * initialized in workqueue_init_earl <<
4719 */ <<
4720 pt = &wq_pod_types[WQ_AFFN_SYSTEM]; <<
4721 BUG_ON(!pt->nr_pods); <<
4722 return pt; <<
4723 } <<
4724 <<
4725 /** 3538 /**
4726 * init_worker_pool - initialize a newly zall 3539 * init_worker_pool - initialize a newly zalloc'd worker_pool
4727 * @pool: worker_pool to initialize 3540 * @pool: worker_pool to initialize
4728 * 3541 *
4729 * Initialize a newly zalloc'd @pool. It als !! 3542 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
4730 * !! 3543 * Returns 0 on success, -errno on failure. Even on failure, all fields
4731 * Return: 0 on success, -errno on failure. <<
4732 * inside @pool proper are initialized and pu 3544 * inside @pool proper are initialized and put_unbound_pool() can be called
4733 * on @pool safely to release it. 3545 * on @pool safely to release it.
4734 */ 3546 */
4735 static int init_worker_pool(struct worker_poo 3547 static int init_worker_pool(struct worker_pool *pool)
4736 { 3548 {
4737 raw_spin_lock_init(&pool->lock); !! 3549 spin_lock_init(&pool->lock);
4738 pool->id = -1; 3550 pool->id = -1;
4739 pool->cpu = -1; 3551 pool->cpu = -1;
4740 pool->node = NUMA_NO_NODE; 3552 pool->node = NUMA_NO_NODE;
4741 pool->flags |= POOL_DISASSOCIATED; 3553 pool->flags |= POOL_DISASSOCIATED;
4742 pool->watchdog_ts = jiffies; <<
4743 INIT_LIST_HEAD(&pool->worklist); 3554 INIT_LIST_HEAD(&pool->worklist);
4744 INIT_LIST_HEAD(&pool->idle_list); 3555 INIT_LIST_HEAD(&pool->idle_list);
4745 hash_init(pool->busy_hash); 3556 hash_init(pool->busy_hash);
4746 3557
4747 timer_setup(&pool->idle_timer, idle_w !! 3558 init_timer_deferrable(&pool->idle_timer);
4748 INIT_WORK(&pool->idle_cull_work, idle !! 3559 pool->idle_timer.function = idle_worker_timeout;
4749 !! 3560 pool->idle_timer.data = (unsigned long)pool;
4750 timer_setup(&pool->mayday_timer, pool !! 3561
4751 !! 3562 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
4752 INIT_LIST_HEAD(&pool->workers); !! 3563 (unsigned long)pool);
>> 3564
>> 3565 mutex_init(&pool->manager_arb);
>> 3566 mutex_init(&pool->manager_mutex);
>> 3567 idr_init(&pool->worker_idr);
4753 3568
4754 ida_init(&pool->worker_ida); <<
4755 INIT_HLIST_NODE(&pool->hash_node); 3569 INIT_HLIST_NODE(&pool->hash_node);
4756 pool->refcnt = 1; 3570 pool->refcnt = 1;
4757 3571
4758 /* shouldn't fail above this point */ 3572 /* shouldn't fail above this point */
4759 pool->attrs = alloc_workqueue_attrs() !! 3573 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
4760 if (!pool->attrs) 3574 if (!pool->attrs)
4761 return -ENOMEM; 3575 return -ENOMEM;
4762 <<
4763 wqattrs_clear_for_pool(pool->attrs); <<
4764 <<
4765 return 0; <<
4766 } <<
4767 <<
4768 #ifdef CONFIG_LOCKDEP <<
4769 static void wq_init_lockdep(struct workqueue_ <<
4770 { <<
4771 char *lock_name; <<
4772 <<
4773 lockdep_register_key(&wq->key); <<
4774 lock_name = kasprintf(GFP_KERNEL, "%s <<
4775 if (!lock_name) <<
4776 lock_name = wq->name; <<
4777 <<
4778 wq->lock_name = lock_name; <<
4779 lockdep_init_map(&wq->lockdep_map, lo <<
4780 } <<
4781 <<
4782 static void wq_unregister_lockdep(struct work <<
4783 { <<
4784 lockdep_unregister_key(&wq->key); <<
4785 } <<
4786 <<
4787 static void wq_free_lockdep(struct workqueue_ <<
4788 { <<
4789 if (wq->lock_name != wq->name) <<
4790 kfree(wq->lock_name); <<
4791 } <<
4792 #else <<
4793 static void wq_init_lockdep(struct workqueue_ <<
4794 { <<
4795 } <<
4796 <<
4797 static void wq_unregister_lockdep(struct work <<
4798 { <<
4799 } <<
4800 <<
4801 static void wq_free_lockdep(struct workqueue_ <<
4802 { <<
4803 } <<
4804 #endif <<
4805 <<
4806 static void free_node_nr_active(struct wq_nod <<
4807 { <<
4808 int node; <<
4809 <<
4810 for_each_node(node) { <<
4811 kfree(nna_ar[node]); <<
4812 nna_ar[node] = NULL; <<
4813 } <<
4814 <<
4815 kfree(nna_ar[nr_node_ids]); <<
4816 nna_ar[nr_node_ids] = NULL; <<
4817 } <<
4818 <<
4819 static void init_node_nr_active(struct wq_nod <<
4820 { <<
4821 nna->max = WQ_DFL_MIN_ACTIVE; <<
4822 atomic_set(&nna->nr, 0); <<
4823 raw_spin_lock_init(&nna->lock); <<
4824 INIT_LIST_HEAD(&nna->pending_pwqs); <<
4825 } <<
4826 <<
4827 /* <<
4828 * Each node's nr_active counter will be acce <<
4829 * should be allocated in the node. <<
4830 */ <<
4831 static int alloc_node_nr_active(struct wq_nod <<
4832 { <<
4833 struct wq_node_nr_active *nna; <<
4834 int node; <<
4835 <<
4836 for_each_node(node) { <<
4837 nna = kzalloc_node(sizeof(*nn <<
4838 if (!nna) <<
4839 goto err_free; <<
4840 init_node_nr_active(nna); <<
4841 nna_ar[node] = nna; <<
4842 } <<
4843 <<
4844 /* [nr_node_ids] is used as the fallb <<
4845 nna = kzalloc_node(sizeof(*nna), GFP_ <<
4846 if (!nna) <<
4847 goto err_free; <<
4848 init_node_nr_active(nna); <<
4849 nna_ar[nr_node_ids] = nna; <<
4850 <<
4851 return 0; 3576 return 0;
4852 <<
4853 err_free: <<
4854 free_node_nr_active(nna_ar); <<
4855 return -ENOMEM; <<
4856 } <<
4857 <<
4858 static void rcu_free_wq(struct rcu_head *rcu) <<
4859 { <<
4860 struct workqueue_struct *wq = <<
4861 container_of(rcu, struct work <<
4862 <<
4863 if (wq->flags & WQ_UNBOUND) <<
4864 free_node_nr_active(wq->node_ <<
4865 <<
4866 wq_free_lockdep(wq); <<
4867 free_percpu(wq->cpu_pwq); <<
4868 free_workqueue_attrs(wq->unbound_attr <<
4869 kfree(wq); <<
4870 } 3577 }
4871 3578
4872 static void rcu_free_pool(struct rcu_head *rc 3579 static void rcu_free_pool(struct rcu_head *rcu)
4873 { 3580 {
4874 struct worker_pool *pool = container_ 3581 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
4875 3582
4876 ida_destroy(&pool->worker_ida); !! 3583 idr_destroy(&pool->worker_idr);
4877 free_workqueue_attrs(pool->attrs); 3584 free_workqueue_attrs(pool->attrs);
4878 kfree(pool); 3585 kfree(pool);
4879 } 3586 }
4880 3587
4881 /** 3588 /**
4882 * put_unbound_pool - put a worker_pool 3589 * put_unbound_pool - put a worker_pool
4883 * @pool: worker_pool to put 3590 * @pool: worker_pool to put
4884 * 3591 *
4885 * Put @pool. If its refcnt reaches zero, it !! 3592 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
4886 * safe manner. get_unbound_pool() calls thi 3593 * safe manner. get_unbound_pool() calls this function on its failure path
4887 * and this function should be able to releas 3594 * and this function should be able to release pools which went through,
4888 * successfully or not, init_worker_pool(). 3595 * successfully or not, init_worker_pool().
4889 * 3596 *
4890 * Should be called with wq_pool_mutex held. 3597 * Should be called with wq_pool_mutex held.
4891 */ 3598 */
4892 static void put_unbound_pool(struct worker_po 3599 static void put_unbound_pool(struct worker_pool *pool)
4893 { 3600 {
4894 struct worker *worker; 3601 struct worker *worker;
4895 LIST_HEAD(cull_list); <<
4896 3602
4897 lockdep_assert_held(&wq_pool_mutex); 3603 lockdep_assert_held(&wq_pool_mutex);
4898 3604
4899 if (--pool->refcnt) 3605 if (--pool->refcnt)
4900 return; 3606 return;
4901 3607
4902 /* sanity checks */ 3608 /* sanity checks */
4903 if (WARN_ON(!(pool->cpu < 0)) || !! 3609 if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
4904 WARN_ON(!list_empty(&pool->workli 3610 WARN_ON(!list_empty(&pool->worklist)))
4905 return; 3611 return;
4906 3612
4907 /* release id and unhash */ 3613 /* release id and unhash */
4908 if (pool->id >= 0) 3614 if (pool->id >= 0)
4909 idr_remove(&worker_pool_idr, 3615 idr_remove(&worker_pool_idr, pool->id);
4910 hash_del(&pool->hash_node); 3616 hash_del(&pool->hash_node);
4911 3617
4912 /* 3618 /*
4913 * Become the manager and destroy all !! 3619 * Become the manager and destroy all workers. Grabbing
4914 * @pool's workers from blocking on a !! 3620 * manager_arb prevents @pool's workers from blocking on
4915 * manager and @pool gets freed with !! 3621 * manager_mutex.
4916 * !! 3622 */
4917 * Having a concurrent manager is qui !! 3623 mutex_lock(&pool->manager_arb);
4918 * only get here with !! 3624 mutex_lock(&pool->manager_mutex);
4919 * pwq->refcnt == pool->refcnt == 0 !! 3625 spin_lock_irq(&pool->lock);
4920 * which implies no work queued to th <<
4921 * become the manager. However a work <<
4922 * manager before the refcnts dropped <<
4923 * drops pool->lock <<
4924 */ <<
4925 while (true) { <<
4926 rcuwait_wait_event(&manager_w <<
4927 !(pool->fl <<
4928 TASK_UNINT <<
4929 <<
4930 mutex_lock(&wq_pool_attach_mu <<
4931 raw_spin_lock_irq(&pool->lock <<
4932 if (!(pool->flags & POOL_MANA <<
4933 pool->flags |= POOL_M <<
4934 break; <<
4935 } <<
4936 raw_spin_unlock_irq(&pool->lo <<
4937 mutex_unlock(&wq_pool_attach_ <<
4938 } <<
4939 3626
4940 while ((worker = first_idle_worker(po !! 3627 while ((worker = first_worker(pool)))
4941 set_worker_dying(worker, &cul !! 3628 destroy_worker(worker);
4942 WARN_ON(pool->nr_workers || pool->nr_ 3629 WARN_ON(pool->nr_workers || pool->nr_idle);
4943 raw_spin_unlock_irq(&pool->lock); <<
4944 <<
4945 detach_dying_workers(&cull_list); <<
4946 3630
4947 mutex_unlock(&wq_pool_attach_mutex); !! 3631 spin_unlock_irq(&pool->lock);
4948 !! 3632 mutex_unlock(&pool->manager_mutex);
4949 reap_dying_workers(&cull_list); !! 3633 mutex_unlock(&pool->manager_arb);
4950 3634
4951 /* shut down the timers */ 3635 /* shut down the timers */
4952 del_timer_sync(&pool->idle_timer); 3636 del_timer_sync(&pool->idle_timer);
4953 cancel_work_sync(&pool->idle_cull_wor <<
4954 del_timer_sync(&pool->mayday_timer); 3637 del_timer_sync(&pool->mayday_timer);
4955 3638
4956 /* RCU protected to allow dereference !! 3639 /* sched-RCU protected to allow dereferences from get_work_pool() */
4957 call_rcu(&pool->rcu, rcu_free_pool); !! 3640 call_rcu_sched(&pool->rcu, rcu_free_pool);
4958 } 3641 }
4959 3642
4960 /** 3643 /**
4961 * get_unbound_pool - get a worker_pool with 3644 * get_unbound_pool - get a worker_pool with the specified attributes
4962 * @attrs: the attributes of the worker_pool 3645 * @attrs: the attributes of the worker_pool to get
4963 * 3646 *
4964 * Obtain a worker_pool which has the same at 3647 * Obtain a worker_pool which has the same attributes as @attrs, bump the
4965 * reference count and return it. If there a 3648 * reference count and return it. If there already is a matching
4966 * worker_pool, it will be used; otherwise, t 3649 * worker_pool, it will be used; otherwise, this function attempts to
4967 * create a new one. !! 3650 * create a new one. On failure, returns NULL.
4968 * 3651 *
4969 * Should be called with wq_pool_mutex held. 3652 * Should be called with wq_pool_mutex held.
4970 * <<
4971 * Return: On success, a worker_pool with the <<
4972 * On failure, %NULL. <<
4973 */ 3653 */
4974 static struct worker_pool *get_unbound_pool(c 3654 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
4975 { 3655 {
4976 struct wq_pod_type *pt = &wq_pod_type <<
4977 u32 hash = wqattrs_hash(attrs); 3656 u32 hash = wqattrs_hash(attrs);
4978 struct worker_pool *pool; 3657 struct worker_pool *pool;
4979 int pod, node = NUMA_NO_NODE; !! 3658 int node;
4980 3659
4981 lockdep_assert_held(&wq_pool_mutex); 3660 lockdep_assert_held(&wq_pool_mutex);
4982 3661
4983 /* do we already have a matching pool 3662 /* do we already have a matching pool? */
4984 hash_for_each_possible(unbound_pool_h 3663 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
4985 if (wqattrs_equal(pool->attrs 3664 if (wqattrs_equal(pool->attrs, attrs)) {
4986 pool->refcnt++; 3665 pool->refcnt++;
4987 return pool; !! 3666 goto out_unlock;
4988 } <<
4989 } <<
4990 <<
4991 /* If __pod_cpumask is contained insi <<
4992 for (pod = 0; pod < pt->nr_pods; pod+ <<
4993 if (cpumask_subset(attrs->__p <<
4994 node = pt->pod_node[p <<
4995 break; <<
4996 } 3667 }
4997 } 3668 }
4998 3669
4999 /* nope, create a new one */ 3670 /* nope, create a new one */
5000 pool = kzalloc_node(sizeof(*pool), GF !! 3671 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
5001 if (!pool || init_worker_pool(pool) < 3672 if (!pool || init_worker_pool(pool) < 0)
5002 goto fail; 3673 goto fail;
5003 3674
5004 pool->node = node; !! 3675 if (workqueue_freezing)
>> 3676 pool->flags |= POOL_FREEZING;
>> 3677
>> 3678 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
5005 copy_workqueue_attrs(pool->attrs, att 3679 copy_workqueue_attrs(pool->attrs, attrs);
5006 wqattrs_clear_for_pool(pool->attrs); !! 3680
>> 3681 /*
>> 3682 * no_numa isn't a worker_pool attribute, always clear it. See
>> 3683 * 'struct workqueue_attrs' comments for detail.
>> 3684 */
>> 3685 pool->attrs->no_numa = false;
>> 3686
>> 3687 /* if cpumask is contained inside a NUMA node, we belong to that node */
>> 3688 if (wq_numa_enabled) {
>> 3689 for_each_node(node) {
>> 3690 if (cpumask_subset(pool->attrs->cpumask,
>> 3691 wq_numa_possible_cpumask[node])) {
>> 3692 pool->node = node;
>> 3693 break;
>> 3694 }
>> 3695 }
>> 3696 }
5007 3697
5008 if (worker_pool_assign_id(pool) < 0) 3698 if (worker_pool_assign_id(pool) < 0)
5009 goto fail; 3699 goto fail;
5010 3700
5011 /* create and start the initial worke 3701 /* create and start the initial worker */
5012 if (wq_online && !create_worker(pool) !! 3702 if (create_and_start_worker(pool) < 0)
5013 goto fail; 3703 goto fail;
5014 3704
5015 /* install */ 3705 /* install */
5016 hash_add(unbound_pool_hash, &pool->ha 3706 hash_add(unbound_pool_hash, &pool->hash_node, hash);
5017 !! 3707 out_unlock:
5018 return pool; 3708 return pool;
5019 fail: 3709 fail:
5020 if (pool) 3710 if (pool)
5021 put_unbound_pool(pool); 3711 put_unbound_pool(pool);
5022 return NULL; 3712 return NULL;
5023 } 3713 }
5024 3714
>> 3715 static void rcu_free_pwq(struct rcu_head *rcu)
>> 3716 {
>> 3717 kmem_cache_free(pwq_cache,
>> 3718 container_of(rcu, struct pool_workqueue, rcu));
>> 3719 }
>> 3720
5025 /* 3721 /*
5026 * Scheduled on pwq_release_worker by put_pwq !! 3722 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
5027 * refcnt and needs to be destroyed. !! 3723 * and needs to be destroyed.
5028 */ 3724 */
5029 static void pwq_release_workfn(struct kthread !! 3725 static void pwq_unbound_release_workfn(struct work_struct *work)
5030 { 3726 {
5031 struct pool_workqueue *pwq = containe 3727 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
5032 !! 3728 unbound_release_work);
5033 struct workqueue_struct *wq = pwq->wq 3729 struct workqueue_struct *wq = pwq->wq;
5034 struct worker_pool *pool = pwq->pool; 3730 struct worker_pool *pool = pwq->pool;
5035 bool is_last = false; !! 3731 bool is_last;
>> 3732
>> 3733 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
>> 3734 return;
5036 3735
5037 /* 3736 /*
5038 * When @pwq is not linked, it doesn' !! 3737 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
5039 * @wq, and @wq is invalid to access. !! 3738 * necessary on release but do it anyway. It's easier to verify
>> 3739 * and consistent with the linking path.
5040 */ 3740 */
5041 if (!list_empty(&pwq->pwqs_node)) { !! 3741 mutex_lock(&wq->mutex);
5042 mutex_lock(&wq->mutex); !! 3742 list_del_rcu(&pwq->pwqs_node);
5043 list_del_rcu(&pwq->pwqs_node) !! 3743 is_last = list_empty(&wq->pwqs);
5044 is_last = list_empty(&wq->pwq !! 3744 mutex_unlock(&wq->mutex);
5045 3745
5046 /* !! 3746 mutex_lock(&wq_pool_mutex);
5047 * For ordered workqueue with !! 3747 put_unbound_pool(pool);
5048 */ !! 3748 mutex_unlock(&wq_pool_mutex);
5049 if (!is_last && (wq->flags & <<
5050 unplug_oldest_pwq(wq) <<
5051 3749
5052 mutex_unlock(&wq->mutex); !! 3750 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
5053 } <<
5054 3751
5055 if (wq->flags & WQ_UNBOUND) { !! 3752 /*
5056 mutex_lock(&wq_pool_mutex); !! 3753 * If we're the last pwq going away, @wq is already dead and no one
5057 put_unbound_pool(pool); !! 3754 * is gonna access it anymore. Free it.
5058 mutex_unlock(&wq_pool_mutex); !! 3755 */
>> 3756 if (is_last) {
>> 3757 free_workqueue_attrs(wq->unbound_attrs);
>> 3758 kfree(wq);
5059 } 3759 }
>> 3760 }
5060 3761
5061 if (!list_empty(&pwq->pending_node)) !! 3762 /**
5062 struct wq_node_nr_active *nna !! 3763 * pwq_adjust_max_active - update a pwq's max_active to the current setting
5063 wq_node_nr_active(pwq !! 3764 * @pwq: target pool_workqueue
>> 3765 *
>> 3766 * If @pwq isn't freezing, set @pwq->max_active to the associated
>> 3767 * workqueue's saved_max_active and activate delayed work items
>> 3768 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
>> 3769 */
>> 3770 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
>> 3771 {
>> 3772 struct workqueue_struct *wq = pwq->wq;
>> 3773 bool freezable = wq->flags & WQ_FREEZABLE;
5064 3774
5065 raw_spin_lock_irq(&nna->lock) !! 3775 /* for @wq->saved_max_active */
5066 list_del_init(&pwq->pending_n !! 3776 lockdep_assert_held(&wq->mutex);
5067 raw_spin_unlock_irq(&nna->loc <<
5068 } <<
5069 3777
5070 kfree_rcu(pwq, rcu); !! 3778 /* fast exit for non-freezable wqs */
>> 3779 if (!freezable && pwq->max_active == wq->saved_max_active)
>> 3780 return;
5071 3781
5072 /* !! 3782 spin_lock_irq(&pwq->pool->lock);
5073 * If we're the last pwq going away, !! 3783
5074 * is gonna access it anymore. Sched !! 3784 if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) {
5075 */ !! 3785 pwq->max_active = wq->saved_max_active;
5076 if (is_last) { !! 3786
5077 wq_unregister_lockdep(wq); !! 3787 while (!list_empty(&pwq->delayed_works) &&
5078 call_rcu(&wq->rcu, rcu_free_w !! 3788 pwq->nr_active < pwq->max_active)
>> 3789 pwq_activate_first_delayed(pwq);
>> 3790
>> 3791 /*
>> 3792 * Need to kick a worker after thawed or an unbound wq's
>> 3793 * max_active is bumped. It's a slow path. Do it always.
>> 3794 */
>> 3795 wake_up_worker(pwq->pool);
>> 3796 } else {
>> 3797 pwq->max_active = 0;
5079 } 3798 }
>> 3799
>> 3800 spin_unlock_irq(&pwq->pool->lock);
5080 } 3801 }
5081 3802
5082 /* initialize newly allocated @pwq which is a !! 3803 /* initialize newly alloced @pwq which is associated with @wq and @pool */
5083 static void init_pwq(struct pool_workqueue *p 3804 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
5084 struct worker_pool *pool 3805 struct worker_pool *pool)
5085 { 3806 {
5086 BUG_ON((unsigned long)pwq & ~WORK_STR !! 3807 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
5087 3808
5088 memset(pwq, 0, sizeof(*pwq)); 3809 memset(pwq, 0, sizeof(*pwq));
5089 3810
5090 pwq->pool = pool; 3811 pwq->pool = pool;
5091 pwq->wq = wq; 3812 pwq->wq = wq;
5092 pwq->flush_color = -1; 3813 pwq->flush_color = -1;
5093 pwq->refcnt = 1; 3814 pwq->refcnt = 1;
5094 INIT_LIST_HEAD(&pwq->inactive_works); !! 3815 INIT_LIST_HEAD(&pwq->delayed_works);
5095 INIT_LIST_HEAD(&pwq->pending_node); <<
5096 INIT_LIST_HEAD(&pwq->pwqs_node); 3816 INIT_LIST_HEAD(&pwq->pwqs_node);
5097 INIT_LIST_HEAD(&pwq->mayday_node); 3817 INIT_LIST_HEAD(&pwq->mayday_node);
5098 kthread_init_work(&pwq->release_work, !! 3818 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
5099 } 3819 }
5100 3820
5101 /* sync @pwq with the current state of its as 3821 /* sync @pwq with the current state of its associated wq and link it */
5102 static void link_pwq(struct pool_workqueue *p 3822 static void link_pwq(struct pool_workqueue *pwq)
5103 { 3823 {
5104 struct workqueue_struct *wq = pwq->wq 3824 struct workqueue_struct *wq = pwq->wq;
5105 3825
5106 lockdep_assert_held(&wq->mutex); 3826 lockdep_assert_held(&wq->mutex);
5107 3827
5108 /* may be called multiple times, igno 3828 /* may be called multiple times, ignore if already linked */
5109 if (!list_empty(&pwq->pwqs_node)) 3829 if (!list_empty(&pwq->pwqs_node))
5110 return; 3830 return;
5111 3831
5112 /* set the matching work_color */ !! 3832 /*
>> 3833 * Set the matching work_color. This is synchronized with
>> 3834 * wq->mutex to avoid confusing flush_workqueue().
>> 3835 */
5113 pwq->work_color = wq->work_color; 3836 pwq->work_color = wq->work_color;
5114 3837
>> 3838 /* sync max_active to the current setting */
>> 3839 pwq_adjust_max_active(pwq);
>> 3840
5115 /* link in @pwq */ 3841 /* link in @pwq */
5116 list_add_tail_rcu(&pwq->pwqs_node, &w !! 3842 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
5117 } 3843 }
5118 3844
5119 /* obtain a pool matching @attr and create a 3845 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
5120 static struct pool_workqueue *alloc_unbound_p 3846 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
5121 const 3847 const struct workqueue_attrs *attrs)
5122 { 3848 {
5123 struct worker_pool *pool; 3849 struct worker_pool *pool;
5124 struct pool_workqueue *pwq; 3850 struct pool_workqueue *pwq;
5125 3851
5126 lockdep_assert_held(&wq_pool_mutex); 3852 lockdep_assert_held(&wq_pool_mutex);
5127 3853
5128 pool = get_unbound_pool(attrs); 3854 pool = get_unbound_pool(attrs);
5129 if (!pool) 3855 if (!pool)
5130 return NULL; 3856 return NULL;
5131 3857
5132 pwq = kmem_cache_alloc_node(pwq_cache 3858 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
5133 if (!pwq) { 3859 if (!pwq) {
5134 put_unbound_pool(pool); 3860 put_unbound_pool(pool);
5135 return NULL; 3861 return NULL;
5136 } 3862 }
5137 3863
5138 init_pwq(pwq, wq, pool); 3864 init_pwq(pwq, wq, pool);
5139 return pwq; 3865 return pwq;
5140 } 3866 }
5141 3867
5142 static void apply_wqattrs_lock(void) !! 3868 /* undo alloc_unbound_pwq(), used only in the error path */
>> 3869 static void free_unbound_pwq(struct pool_workqueue *pwq)
5143 { 3870 {
5144 mutex_lock(&wq_pool_mutex); !! 3871 lockdep_assert_held(&wq_pool_mutex);
5145 } <<
5146 3872
5147 static void apply_wqattrs_unlock(void) !! 3873 if (pwq) {
5148 { !! 3874 put_unbound_pool(pwq->pool);
5149 mutex_unlock(&wq_pool_mutex); !! 3875 kmem_cache_free(pwq_cache, pwq);
>> 3876 }
5150 } 3877 }
5151 3878
5152 /** 3879 /**
5153 * wq_calc_pod_cpumask - calculate a wq_attrs !! 3880 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
5154 * @attrs: the wq_attrs of the default pwq of !! 3881 * @attrs: the wq_attrs of interest
5155 * @cpu: the target CPU !! 3882 * @node: the target NUMA node
>> 3883 * @cpu_going_down: if >= 0, the CPU to consider as offline
>> 3884 * @cpumask: outarg, the resulting cpumask
5156 * 3885 *
5157 * Calculate the cpumask a workqueue with @at !! 3886 * Calculate the cpumask a workqueue with @attrs should use on @node. If
5158 * The result is stored in @attrs->__pod_cpum !! 3887 * @cpu_going_down is >= 0, that cpu is considered offline during
>> 3888 * calculation. The result is stored in @cpumask. This function returns
>> 3889 * %true if the resulting @cpumask is different from @attrs->cpumask,
>> 3890 * %false if equal.
5159 * 3891 *
5160 * If pod affinity is not enabled, @attrs->cp !! 3892 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
5161 * and @pod has online CPUs requested by @att !! 3893 * enabled and @node has online CPUs requested by @attrs, the returned
5162 * intersection of the possible CPUs of @pod !! 3894 * cpumask is the intersection of the possible CPUs of @node and
>> 3895 * @attrs->cpumask.
5163 * 3896 *
5164 * The caller is responsible for ensuring tha !! 3897 * The caller is responsible for ensuring that the cpumask of @node stays
>> 3898 * stable.
5165 */ 3899 */
5166 static void wq_calc_pod_cpumask(struct workqu !! 3900 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
>> 3901 int cpu_going_down, cpumask_t *cpumask)
5167 { 3902 {
5168 const struct wq_pod_type *pt = wqattr !! 3903 if (!wq_numa_enabled || attrs->no_numa)
5169 int pod = pt->cpu_pod[cpu]; !! 3904 goto use_dfl;
5170 3905
5171 /* calculate possible CPUs in @pod th !! 3906 /* does @node have any online CPUs @attrs wants? */
5172 cpumask_and(attrs->__pod_cpumask, pt- !! 3907 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
5173 /* does @pod have any online CPUs @at !! 3908 if (cpu_going_down >= 0)
5174 if (!cpumask_intersects(attrs->__pod_ !! 3909 cpumask_clear_cpu(cpu_going_down, cpumask);
5175 cpumask_copy(attrs->__pod_cpu !! 3910
5176 return; !! 3911 if (cpumask_empty(cpumask))
5177 } !! 3912 goto use_dfl;
>> 3913
>> 3914 /* yeap, return possible CPUs in @node that @attrs wants */
>> 3915 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
>> 3916 return !cpumask_equal(cpumask, attrs->cpumask);
>> 3917
>> 3918 use_dfl:
>> 3919 cpumask_copy(cpumask, attrs->cpumask);
>> 3920 return false;
5178 } 3921 }
5179 3922
5180 /* install @pwq into @wq and return the old p !! 3923 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
5181 static struct pool_workqueue *install_unbound !! 3924 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
5182 int c !! 3925 int node,
>> 3926 struct pool_workqueue *pwq)
5183 { 3927 {
5184 struct pool_workqueue __rcu **slot = <<
5185 struct pool_workqueue *old_pwq; 3928 struct pool_workqueue *old_pwq;
5186 3929
5187 lockdep_assert_held(&wq_pool_mutex); <<
5188 lockdep_assert_held(&wq->mutex); 3930 lockdep_assert_held(&wq->mutex);
5189 3931
5190 /* link_pwq() can handle duplicate ca 3932 /* link_pwq() can handle duplicate calls */
5191 link_pwq(pwq); 3933 link_pwq(pwq);
5192 3934
5193 old_pwq = rcu_access_pointer(*slot); !! 3935 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
5194 rcu_assign_pointer(*slot, pwq); !! 3936 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
5195 return old_pwq; 3937 return old_pwq;
5196 } 3938 }
5197 3939
5198 /* context to store the prepared attrs & pwqs !! 3940 /**
5199 struct apply_wqattrs_ctx { !! 3941 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
5200 struct workqueue_struct *wq; !! 3942 * @wq: the target workqueue
5201 struct workqueue_attrs *attrs; !! 3943 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
5202 struct list_head list; !! 3944 *
5203 struct pool_workqueue *dfl_pwq; !! 3945 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
5204 struct pool_workqueue *pwq_tbl[]; !! 3946 * machines, this function maps a separate pwq to each NUMA node with
5205 }; !! 3947 * possibles CPUs in @attrs->cpumask so that work items are affine to the
5206 !! 3948 * NUMA node it was issued on. Older pwqs are released as in-flight work
5207 /* free the resources after success or abort !! 3949 * items finish. Note that a work item which repeatedly requeues itself
5208 static void apply_wqattrs_cleanup(struct appl !! 3950 * back-to-back will stay on its current pwq.
>> 3951 *
>> 3952 * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on
>> 3953 * failure.
>> 3954 */
>> 3955 int apply_workqueue_attrs(struct workqueue_struct *wq,
>> 3956 const struct workqueue_attrs *attrs)
5209 { 3957 {
5210 if (ctx) { !! 3958 struct workqueue_attrs *new_attrs, *tmp_attrs;
5211 int cpu; !! 3959 struct pool_workqueue **pwq_tbl, *dfl_pwq;
>> 3960 int node, ret;
5212 3961
5213 for_each_possible_cpu(cpu) !! 3962 /* only unbound workqueues can change attributes */
5214 put_pwq_unlocked(ctx- !! 3963 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
5215 put_pwq_unlocked(ctx->dfl_pwq !! 3964 return -EINVAL;
5216 <<
5217 free_workqueue_attrs(ctx->att <<
5218 3965
5219 kfree(ctx); !! 3966 /* creating multiple pwqs breaks ordering guarantee */
>> 3967 if (!list_empty(&wq->pwqs)) {
>> 3968 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
>> 3969 return -EINVAL;
>> 3970
>> 3971 wq->flags &= ~__WQ_ORDERED;
5220 } 3972 }
5221 } <<
5222 3973
5223 /* allocate the attrs and pwqs for later inst !! 3974 pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL);
5224 static struct apply_wqattrs_ctx * !! 3975 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
5225 apply_wqattrs_prepare(struct workqueue_struct !! 3976 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
5226 const struct workqueue_ !! 3977 if (!pwq_tbl || !new_attrs || !tmp_attrs)
5227 const cpumask_var_t unb !! 3978 goto enomem;
5228 { <<
5229 struct apply_wqattrs_ctx *ctx; <<
5230 struct workqueue_attrs *new_attrs; <<
5231 int cpu; <<
5232 3979
5233 lockdep_assert_held(&wq_pool_mutex); !! 3980 /* make a copy of @attrs and sanitize it */
>> 3981 copy_workqueue_attrs(new_attrs, attrs);
>> 3982 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
>> 3983
>> 3984 /*
>> 3985 * We may create multiple pwqs with differing cpumasks. Make a
>> 3986 * copy of @new_attrs which will be modified and used to obtain
>> 3987 * pools.
>> 3988 */
>> 3989 copy_workqueue_attrs(tmp_attrs, new_attrs);
>> 3990
>> 3991 /*
>> 3992 * CPUs should stay stable across pwq creations and installations.
>> 3993 * Pin CPUs, determine the target cpumask for each node and create
>> 3994 * pwqs accordingly.
>> 3995 */
>> 3996 get_online_cpus();
5234 3997
5235 if (WARN_ON(attrs->affn_scope < 0 || !! 3998 mutex_lock(&wq_pool_mutex);
5236 attrs->affn_scope >= WQ_A <<
5237 return ERR_PTR(-EINVAL); <<
5238 <<
5239 ctx = kzalloc(struct_size(ctx, pwq_tb <<
5240 <<
5241 new_attrs = alloc_workqueue_attrs(); <<
5242 if (!ctx || !new_attrs) <<
5243 goto out_free; <<
5244 3999
5245 /* 4000 /*
5246 * If something goes wrong during CPU 4001 * If something goes wrong during CPU up/down, we'll fall back to
5247 * the default pwq covering whole @at 4002 * the default pwq covering whole @attrs->cpumask. Always create
5248 * it even if we don't use it immedia 4003 * it even if we don't use it immediately.
5249 */ 4004 */
5250 copy_workqueue_attrs(new_attrs, attrs !! 4005 dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
5251 wqattrs_actualize_cpumask(new_attrs, !! 4006 if (!dfl_pwq)
5252 cpumask_copy(new_attrs->__pod_cpumask !! 4007 goto enomem_pwq;
5253 ctx->dfl_pwq = alloc_unbound_pwq(wq, <<
5254 if (!ctx->dfl_pwq) <<
5255 goto out_free; <<
5256 4008
5257 for_each_possible_cpu(cpu) { !! 4009 for_each_node(node) {
5258 if (new_attrs->ordered) { !! 4010 if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) {
5259 ctx->dfl_pwq->refcnt+ !! 4011 pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
5260 ctx->pwq_tbl[cpu] = c !! 4012 if (!pwq_tbl[node])
>> 4013 goto enomem_pwq;
5261 } else { 4014 } else {
5262 wq_calc_pod_cpumask(n !! 4015 dfl_pwq->refcnt++;
5263 ctx->pwq_tbl[cpu] = a !! 4016 pwq_tbl[node] = dfl_pwq;
5264 if (!ctx->pwq_tbl[cpu <<
5265 goto out_free <<
5266 } 4017 }
5267 } 4018 }
5268 4019
5269 /* save the user configured attrs and !! 4020 mutex_unlock(&wq_pool_mutex);
5270 copy_workqueue_attrs(new_attrs, attrs <<
5271 cpumask_and(new_attrs->cpumask, new_a <<
5272 cpumask_copy(new_attrs->__pod_cpumask <<
5273 ctx->attrs = new_attrs; <<
5274 <<
5275 /* <<
5276 * For initialized ordered workqueues <<
5277 * (dfl_pwq). Set the plugged flag of <<
5278 * of newly queued work items until e <<
5279 * the old pwq's have completed. <<
5280 */ <<
5281 if ((wq->flags & __WQ_ORDERED) && !li <<
5282 ctx->dfl_pwq->plugged = true; <<
5283 <<
5284 ctx->wq = wq; <<
5285 return ctx; <<
5286 <<
5287 out_free: <<
5288 free_workqueue_attrs(new_attrs); <<
5289 apply_wqattrs_cleanup(ctx); <<
5290 return ERR_PTR(-ENOMEM); <<
5291 } <<
5292 <<
5293 /* set attrs and install prepared pwqs, @ctx <<
5294 static void apply_wqattrs_commit(struct apply <<
5295 { <<
5296 int cpu; <<
5297 4021
5298 /* all pwqs have been created success 4022 /* all pwqs have been created successfully, let's install'em */
5299 mutex_lock(&ctx->wq->mutex); !! 4023 mutex_lock(&wq->mutex);
5300 <<
5301 copy_workqueue_attrs(ctx->wq->unbound <<
5302 <<
5303 /* save the previous pwqs and install <<
5304 for_each_possible_cpu(cpu) <<
5305 ctx->pwq_tbl[cpu] = install_u <<
5306 <<
5307 ctx->dfl_pwq = install_unbound_pwq(ct <<
5308 <<
5309 /* update node_nr_active->max */ <<
5310 wq_update_node_max_active(ctx->wq, -1 <<
5311 4024
5312 /* rescuer needs to respect wq cpumas !! 4025 copy_workqueue_attrs(wq->unbound_attrs, new_attrs);
5313 if (ctx->wq->rescuer) <<
5314 set_cpus_allowed_ptr(ctx->wq- <<
5315 unbound_ <<
5316 4026
5317 mutex_unlock(&ctx->wq->mutex); !! 4027 /* save the previous pwq and install the new one */
5318 } !! 4028 for_each_node(node)
>> 4029 pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]);
>> 4030
>> 4031 /* @dfl_pwq might not have been used, ensure it's linked */
>> 4032 link_pwq(dfl_pwq);
>> 4033 swap(wq->dfl_pwq, dfl_pwq);
5319 4034
5320 static int apply_workqueue_attrs_locked(struc !! 4035 mutex_unlock(&wq->mutex);
5321 const <<
5322 { <<
5323 struct apply_wqattrs_ctx *ctx; <<
5324 <<
5325 /* only unbound workqueues can change <<
5326 if (WARN_ON(!(wq->flags & WQ_UNBOUND) <<
5327 return -EINVAL; <<
5328 <<
5329 ctx = apply_wqattrs_prepare(wq, attrs <<
5330 if (IS_ERR(ctx)) <<
5331 return PTR_ERR(ctx); <<
5332 <<
5333 /* the ctx has been prepared successf <<
5334 apply_wqattrs_commit(ctx); <<
5335 apply_wqattrs_cleanup(ctx); <<
5336 <<
5337 return 0; <<
5338 } <<
5339 4036
5340 /** !! 4037 /* put the old pwqs */
5341 * apply_workqueue_attrs - apply new workqueu !! 4038 for_each_node(node)
5342 * @wq: the target workqueue !! 4039 put_pwq_unlocked(pwq_tbl[node]);
5343 * @attrs: the workqueue_attrs to apply, allo !! 4040 put_pwq_unlocked(dfl_pwq);
5344 * !! 4041
5345 * Apply @attrs to an unbound workqueue @wq. !! 4042 put_online_cpus();
5346 * a separate pwq to each CPU pod with possib !! 4043 ret = 0;
5347 * work items are affine to the pod it was is !! 4044 /* fall through */
5348 * in-flight work items finish. Note that a w !! 4045 out_free:
5349 * itself back-to-back will stay on its curre !! 4046 free_workqueue_attrs(tmp_attrs);
5350 * !! 4047 free_workqueue_attrs(new_attrs);
5351 * Performs GFP_KERNEL allocations. !! 4048 kfree(pwq_tbl);
5352 * !! 4049 return ret;
5353 * Return: 0 on success and -errno on failure <<
5354 */ <<
5355 int apply_workqueue_attrs(struct workqueue_st <<
5356 const struct workqu <<
5357 { <<
5358 int ret; <<
5359 4050
5360 mutex_lock(&wq_pool_mutex); !! 4051 enomem_pwq:
5361 ret = apply_workqueue_attrs_locked(wq !! 4052 free_unbound_pwq(dfl_pwq);
>> 4053 for_each_node(node)
>> 4054 if (pwq_tbl && pwq_tbl[node] != dfl_pwq)
>> 4055 free_unbound_pwq(pwq_tbl[node]);
5362 mutex_unlock(&wq_pool_mutex); 4056 mutex_unlock(&wq_pool_mutex);
5363 !! 4057 put_online_cpus();
5364 return ret; !! 4058 enomem:
>> 4059 ret = -ENOMEM;
>> 4060 goto out_free;
5365 } 4061 }
5366 4062
5367 /** 4063 /**
5368 * unbound_wq_update_pwq - update a pwq slot !! 4064 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
5369 * @wq: the target workqueue 4065 * @wq: the target workqueue
5370 * @cpu: the CPU to update the pwq slot for !! 4066 * @cpu: the CPU coming up or going down
>> 4067 * @online: whether @cpu is coming up or going down
5371 * 4068 *
5372 * This function is to be called from %CPU_DO 4069 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
5373 * %CPU_DOWN_FAILED. @cpu is in the same pod !! 4070 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
5374 * !! 4071 * @wq accordingly.
5375 * <<
5376 * If pod affinity can't be adjusted due to m <<
5377 * back to @wq->dfl_pwq which may not be opti <<
5378 * 4072 *
5379 * Note that when the last allowed CPU of a p !! 4073 * If NUMA affinity can't be adjusted due to memory allocation failure, it
5380 * with a cpumask spanning multiple pods, the !! 4074 * falls back to @wq->dfl_pwq which may not be optimal but is always
5381 * executing the work items for the workqueue !! 4075 * correct.
5382 * may execute on any CPU. This is similar to !! 4076 *
5383 * CPU_DOWN. If a workqueue user wants strict !! 4077 * Note that when the last allowed CPU of a NUMA node goes offline for a
5384 * responsibility to flush the work item from !! 4078 * workqueue with a cpumask spanning multiple nodes, the workers which were
>> 4079 * already executing the work items for the workqueue will lose their CPU
>> 4080 * affinity and may execute on any CPU. This is similar to how per-cpu
>> 4081 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
>> 4082 * affinity, it's the user's responsibility to flush the work item from
>> 4083 * CPU_DOWN_PREPARE.
5385 */ 4084 */
5386 static void unbound_wq_update_pwq(struct work !! 4085 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
>> 4086 bool online)
5387 { 4087 {
>> 4088 int node = cpu_to_node(cpu);
>> 4089 int cpu_off = online ? -1 : cpu;
5388 struct pool_workqueue *old_pwq = NULL 4090 struct pool_workqueue *old_pwq = NULL, *pwq;
5389 struct workqueue_attrs *target_attrs; 4091 struct workqueue_attrs *target_attrs;
>> 4092 cpumask_t *cpumask;
5390 4093
5391 lockdep_assert_held(&wq_pool_mutex); 4094 lockdep_assert_held(&wq_pool_mutex);
5392 4095
5393 if (!(wq->flags & WQ_UNBOUND) || wq-> !! 4096 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
5394 return; 4097 return;
5395 4098
5396 /* 4099 /*
5397 * We don't wanna alloc/free wq_attrs 4100 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
5398 * Let's use a preallocated one. The 4101 * Let's use a preallocated one. The following buf is protected by
5399 * CPU hotplug exclusion. 4102 * CPU hotplug exclusion.
5400 */ 4103 */
5401 target_attrs = unbound_wq_update_pwq_ !! 4104 target_attrs = wq_update_unbound_numa_attrs_buf;
>> 4105 cpumask = target_attrs->cpumask;
>> 4106
>> 4107 mutex_lock(&wq->mutex);
>> 4108 if (wq->unbound_attrs->no_numa)
>> 4109 goto out_unlock;
5402 4110
5403 copy_workqueue_attrs(target_attrs, wq 4111 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
5404 wqattrs_actualize_cpumask(target_attr !! 4112 pwq = unbound_pwq_by_node(wq, node);
5405 4113
5406 /* nothing to do if the target cpumas !! 4114 /*
5407 wq_calc_pod_cpumask(target_attrs, cpu !! 4115 * Let's determine what needs to be done. If the target cpumask is
5408 if (wqattrs_equal(target_attrs, unbou !! 4116 * different from wq's, we need to compare it to @pwq's and create
5409 return; !! 4117 * a new one if they don't match. If the target cpumask equals
>> 4118 * wq's, the default pwq should be used. If @pwq is already the
>> 4119 * default one, nothing to do; otherwise, install the default one.
>> 4120 */
>> 4121 if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) {
>> 4122 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
>> 4123 goto out_unlock;
>> 4124 } else {
>> 4125 if (pwq == wq->dfl_pwq)
>> 4126 goto out_unlock;
>> 4127 else
>> 4128 goto use_dfl_pwq;
>> 4129 }
>> 4130
>> 4131 mutex_unlock(&wq->mutex);
5410 4132
5411 /* create a new pwq */ 4133 /* create a new pwq */
5412 pwq = alloc_unbound_pwq(wq, target_at 4134 pwq = alloc_unbound_pwq(wq, target_attrs);
5413 if (!pwq) { 4135 if (!pwq) {
5414 pr_warn("workqueue: allocatio !! 4136 pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
5415 wq->name); !! 4137 wq->name);
>> 4138 mutex_lock(&wq->mutex);
5416 goto use_dfl_pwq; 4139 goto use_dfl_pwq;
5417 } 4140 }
5418 4141
5419 /* Install the new pwq. */ !! 4142 /*
>> 4143 * Install the new pwq. As this function is called only from CPU
>> 4144 * hotplug callbacks and applying a new attrs is wrapped with
>> 4145 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
>> 4146 * inbetween.
>> 4147 */
5420 mutex_lock(&wq->mutex); 4148 mutex_lock(&wq->mutex);
5421 old_pwq = install_unbound_pwq(wq, cpu !! 4149 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
5422 goto out_unlock; 4150 goto out_unlock;
5423 4151
5424 use_dfl_pwq: 4152 use_dfl_pwq:
5425 mutex_lock(&wq->mutex); !! 4153 spin_lock_irq(&wq->dfl_pwq->pool->lock);
5426 pwq = unbound_pwq(wq, -1); !! 4154 get_pwq(wq->dfl_pwq);
5427 raw_spin_lock_irq(&pwq->pool->lock); !! 4155 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
5428 get_pwq(pwq); !! 4156 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
5429 raw_spin_unlock_irq(&pwq->pool->lock) <<
5430 old_pwq = install_unbound_pwq(wq, cpu <<
5431 out_unlock: 4157 out_unlock:
5432 mutex_unlock(&wq->mutex); 4158 mutex_unlock(&wq->mutex);
5433 put_pwq_unlocked(old_pwq); 4159 put_pwq_unlocked(old_pwq);
5434 } 4160 }
5435 4161
5436 static int alloc_and_link_pwqs(struct workque 4162 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
5437 { 4163 {
5438 bool highpri = wq->flags & WQ_HIGHPRI 4164 bool highpri = wq->flags & WQ_HIGHPRI;
5439 int cpu, ret; 4165 int cpu, ret;
5440 4166
5441 lockdep_assert_held(&wq_pool_mutex); <<
5442 <<
5443 wq->cpu_pwq = alloc_percpu(struct poo <<
5444 if (!wq->cpu_pwq) <<
5445 goto enomem; <<
5446 <<
5447 if (!(wq->flags & WQ_UNBOUND)) { 4167 if (!(wq->flags & WQ_UNBOUND)) {
5448 struct worker_pool __percpu * !! 4168 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
5449 !! 4169 if (!wq->cpu_pwqs)
5450 if (wq->flags & WQ_BH) !! 4170 return -ENOMEM;
5451 pools = bh_worker_poo <<
5452 else <<
5453 pools = cpu_worker_po <<
5454 4171
5455 for_each_possible_cpu(cpu) { 4172 for_each_possible_cpu(cpu) {
5456 struct pool_workqueue !! 4173 struct pool_workqueue *pwq =
5457 struct worker_pool *p !! 4174 per_cpu_ptr(wq->cpu_pwqs, cpu);
>> 4175 struct worker_pool *cpu_pools =
>> 4176 per_cpu(cpu_worker_pools, cpu);
5458 4177
5459 pool = &(per_cpu_ptr( !! 4178 init_pwq(pwq, wq, &cpu_pools[highpri]);
5460 pwq_p = per_cpu_ptr(w <<
5461 <<
5462 *pwq_p = kmem_cache_a <<
5463 <<
5464 if (!*pwq_p) <<
5465 goto enomem; <<
5466 <<
5467 init_pwq(*pwq_p, wq, <<
5468 4179
5469 mutex_lock(&wq->mutex 4180 mutex_lock(&wq->mutex);
5470 link_pwq(*pwq_p); !! 4181 link_pwq(pwq);
5471 mutex_unlock(&wq->mut 4182 mutex_unlock(&wq->mutex);
5472 } 4183 }
5473 return 0; 4184 return 0;
5474 } !! 4185 } else if (wq->flags & __WQ_ORDERED) {
5475 !! 4186 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
5476 if (wq->flags & __WQ_ORDERED) { <<
5477 struct pool_workqueue *dfl_pw <<
5478 <<
5479 ret = apply_workqueue_attrs_l <<
5480 /* there should only be singl 4187 /* there should only be single pwq for ordering guarantee */
5481 dfl_pwq = rcu_access_pointer( !! 4188 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
5482 WARN(!ret && (wq->pwqs.next ! !! 4189 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
5483 wq->pwqs.prev ! <<
5484 "ordering guarantee brok 4190 "ordering guarantee broken for workqueue %s\n", wq->name);
>> 4191 return ret;
5485 } else { 4192 } else {
5486 ret = apply_workqueue_attrs_l !! 4193 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
5487 } <<
5488 <<
5489 return ret; <<
5490 <<
5491 enomem: <<
5492 if (wq->cpu_pwq) { <<
5493 for_each_possible_cpu(cpu) { <<
5494 struct pool_workqueue <<
5495 <<
5496 if (pwq) <<
5497 kmem_cache_fr <<
5498 } <<
5499 free_percpu(wq->cpu_pwq); <<
5500 wq->cpu_pwq = NULL; <<
5501 } 4194 }
5502 return -ENOMEM; <<
5503 } 4195 }
5504 4196
5505 static int wq_clamp_max_active(int max_active 4197 static int wq_clamp_max_active(int max_active, unsigned int flags,
5506 const char *na 4198 const char *name)
5507 { 4199 {
5508 if (max_active < 1 || max_active > WQ !! 4200 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
5509 pr_warn("workqueue: max_activ <<
5510 max_active, name, 1, <<
5511 <<
5512 return clamp_val(max_active, 1, WQ_MA <<
5513 } <<
5514 <<
5515 /* <<
5516 * Workqueues which may be used during memory <<
5517 * to guarantee forward progress. <<
5518 */ <<
5519 static int init_rescuer(struct workqueue_stru <<
5520 { <<
5521 struct worker *rescuer; <<
5522 char id_buf[WORKER_ID_LEN]; <<
5523 int ret; <<
5524 <<
5525 lockdep_assert_held(&wq_pool_mutex); <<
5526 <<
5527 if (!(wq->flags & WQ_MEM_RECLAIM)) <<
5528 return 0; <<
5529 <<
5530 rescuer = alloc_worker(NUMA_NO_NODE); <<
5531 if (!rescuer) { <<
5532 pr_err("workqueue: Failed to <<
5533 wq->name); <<
5534 return -ENOMEM; <<
5535 } <<
5536 <<
5537 rescuer->rescue_wq = wq; <<
5538 format_worker_id(id_buf, sizeof(id_bu <<
5539 4201
5540 rescuer->task = kthread_create(rescue !! 4202 if (max_active < 1 || max_active > lim)
5541 if (IS_ERR(rescuer->task)) { !! 4203 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
5542 ret = PTR_ERR(rescuer->task); !! 4204 max_active, name, 1, lim);
5543 pr_err("workqueue: Failed to <<
5544 wq->name, ERR_PTR(ret) <<
5545 kfree(rescuer); <<
5546 return ret; <<
5547 } <<
5548 <<
5549 wq->rescuer = rescuer; <<
5550 if (wq->flags & WQ_UNBOUND) <<
5551 kthread_bind_mask(rescuer->ta <<
5552 else <<
5553 kthread_bind_mask(rescuer->ta <<
5554 wake_up_process(rescuer->task); <<
5555 <<
5556 return 0; <<
5557 } <<
5558 <<
5559 /** <<
5560 * wq_adjust_max_active - update a wq's max_a <<
5561 * @wq: target workqueue <<
5562 * <<
5563 * If @wq isn't freezing, set @wq->max_active <<
5564 * activate inactive work items accordingly. <<
5565 * @wq->max_active to zero. <<
5566 */ <<
5567 static void wq_adjust_max_active(struct workq <<
5568 { <<
5569 bool activated; <<
5570 int new_max, new_min; <<
5571 <<
5572 lockdep_assert_held(&wq->mutex); <<
5573 <<
5574 if ((wq->flags & WQ_FREEZABLE) && wor <<
5575 new_max = 0; <<
5576 new_min = 0; <<
5577 } else { <<
5578 new_max = wq->saved_max_activ <<
5579 new_min = wq->saved_min_activ <<
5580 } <<
5581 <<
5582 if (wq->max_active == new_max && wq-> <<
5583 return; <<
5584 <<
5585 /* <<
5586 * Update @wq->max/min_active and the <<
5587 * active work items are allowed. Thi <<
5588 * because new work items are always <<
5589 * work items if there are any. <<
5590 */ <<
5591 WRITE_ONCE(wq->max_active, new_max); <<
5592 WRITE_ONCE(wq->min_active, new_min); <<
5593 <<
5594 if (wq->flags & WQ_UNBOUND) <<
5595 wq_update_node_max_active(wq, <<
5596 <<
5597 if (new_max == 0) <<
5598 return; <<
5599 <<
5600 /* <<
5601 * Round-robin through pwq's activati <<
5602 * until max_active is filled. <<
5603 */ <<
5604 do { <<
5605 struct pool_workqueue *pwq; <<
5606 <<
5607 activated = false; <<
5608 for_each_pwq(pwq, wq) { <<
5609 unsigned long irq_fla <<
5610 4205
5611 /* can be called duri !! 4206 return clamp_val(max_active, 1, lim);
5612 raw_spin_lock_irqsave <<
5613 if (pwq_activate_firs <<
5614 activated = t <<
5615 kick_pool(pwq <<
5616 } <<
5617 raw_spin_unlock_irqre <<
5618 } <<
5619 } while (activated); <<
5620 } 4207 }
5621 4208
5622 __printf(1, 4) !! 4209 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
5623 struct workqueue_struct *alloc_workqueue(cons !! 4210 unsigned int flags,
5624 unsi !! 4211 int max_active,
5625 int !! 4212 struct lock_class_key *key,
>> 4213 const char *lock_name, ...)
5626 { 4214 {
>> 4215 size_t tbl_size = 0;
5627 va_list args; 4216 va_list args;
5628 struct workqueue_struct *wq; 4217 struct workqueue_struct *wq;
5629 size_t wq_size; !! 4218 struct pool_workqueue *pwq;
5630 int name_len; <<
5631 <<
5632 if (flags & WQ_BH) { <<
5633 if (WARN_ON_ONCE(flags & ~__W <<
5634 return NULL; <<
5635 if (WARN_ON_ONCE(max_active)) <<
5636 return NULL; <<
5637 } <<
5638 4219
5639 /* see the comment above the definiti !! 4220 /*
5640 if ((flags & WQ_POWER_EFFICIENT) && w !! 4221 * Unbound && max_active == 1 used to imply ordered, which is no
5641 flags |= WQ_UNBOUND; !! 4222 * longer the case on NUMA machines due to per-node pools. While
>> 4223 * alloc_ordered_workqueue() is the right way to create an ordered
>> 4224 * workqueue, keep the previous behavior to avoid subtle breakages
>> 4225 * on NUMA.
>> 4226 */
>> 4227 if ((flags & WQ_UNBOUND) && max_active == 1)
>> 4228 flags |= __WQ_ORDERED;
5642 4229
5643 /* allocate wq and format name */ 4230 /* allocate wq and format name */
5644 if (flags & WQ_UNBOUND) 4231 if (flags & WQ_UNBOUND)
5645 wq_size = struct_size(wq, nod !! 4232 tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]);
5646 else <<
5647 wq_size = sizeof(*wq); <<
5648 4233
5649 wq = kzalloc(wq_size, GFP_KERNEL); !! 4234 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
5650 if (!wq) 4235 if (!wq)
5651 return NULL; 4236 return NULL;
5652 4237
5653 if (flags & WQ_UNBOUND) { 4238 if (flags & WQ_UNBOUND) {
5654 wq->unbound_attrs = alloc_wor !! 4239 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
5655 if (!wq->unbound_attrs) 4240 if (!wq->unbound_attrs)
5656 goto err_free_wq; 4241 goto err_free_wq;
5657 } 4242 }
5658 4243
5659 va_start(args, max_active); !! 4244 va_start(args, lock_name);
5660 name_len = vsnprintf(wq->name, sizeof !! 4245 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
5661 va_end(args); 4246 va_end(args);
5662 4247
5663 if (name_len >= WQ_NAME_LEN) !! 4248 max_active = max_active ?: WQ_DFL_ACTIVE;
5664 pr_warn_once("workqueue: name !! 4249 max_active = wq_clamp_max_active(max_active, flags, wq->name);
5665 wq->name); <<
5666 <<
5667 if (flags & WQ_BH) { <<
5668 /* <<
5669 * BH workqueues always share <<
5670 * and don't impose any max_a <<
5671 */ <<
5672 max_active = INT_MAX; <<
5673 } else { <<
5674 max_active = max_active ?: WQ <<
5675 max_active = wq_clamp_max_act <<
5676 } <<
5677 4250
5678 /* init wq */ 4251 /* init wq */
5679 wq->flags = flags; 4252 wq->flags = flags;
5680 wq->max_active = max_active; !! 4253 wq->saved_max_active = max_active;
5681 wq->min_active = min(max_active, WQ_D <<
5682 wq->saved_max_active = wq->max_active <<
5683 wq->saved_min_active = wq->min_active <<
5684 mutex_init(&wq->mutex); 4254 mutex_init(&wq->mutex);
5685 atomic_set(&wq->nr_pwqs_to_flush, 0); 4255 atomic_set(&wq->nr_pwqs_to_flush, 0);
5686 INIT_LIST_HEAD(&wq->pwqs); 4256 INIT_LIST_HEAD(&wq->pwqs);
5687 INIT_LIST_HEAD(&wq->flusher_queue); 4257 INIT_LIST_HEAD(&wq->flusher_queue);
5688 INIT_LIST_HEAD(&wq->flusher_overflow) 4258 INIT_LIST_HEAD(&wq->flusher_overflow);
5689 INIT_LIST_HEAD(&wq->maydays); 4259 INIT_LIST_HEAD(&wq->maydays);
5690 4260
5691 wq_init_lockdep(wq); !! 4261 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
5692 INIT_LIST_HEAD(&wq->list); 4262 INIT_LIST_HEAD(&wq->list);
5693 4263
5694 if (flags & WQ_UNBOUND) { !! 4264 if (alloc_and_link_pwqs(wq) < 0)
5695 if (alloc_node_nr_active(wq-> !! 4265 goto err_free_wq;
5696 goto err_unreg_lockde <<
5697 } <<
5698 4266
5699 /* 4267 /*
5700 * wq_pool_mutex protects the workque !! 4268 * Workqueues which may be used during memory reclaim should
5701 * and the global freeze state. !! 4269 * have a rescuer to guarantee forward progress.
5702 */ 4270 */
5703 apply_wqattrs_lock(); !! 4271 if (flags & WQ_MEM_RECLAIM) {
5704 !! 4272 struct worker *rescuer;
5705 if (alloc_and_link_pwqs(wq) < 0) <<
5706 goto err_unlock_free_node_nr_ <<
5707 <<
5708 mutex_lock(&wq->mutex); <<
5709 wq_adjust_max_active(wq); <<
5710 mutex_unlock(&wq->mutex); <<
5711 4273
5712 list_add_tail_rcu(&wq->list, &workque !! 4274 rescuer = alloc_worker();
>> 4275 if (!rescuer)
>> 4276 goto err_destroy;
5713 4277
5714 if (wq_online && init_rescuer(wq) < 0 !! 4278 rescuer->rescue_wq = wq;
5715 goto err_unlock_destroy; !! 4279 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
>> 4280 wq->name);
>> 4281 if (IS_ERR(rescuer->task)) {
>> 4282 kfree(rescuer);
>> 4283 goto err_destroy;
>> 4284 }
5716 4285
5717 apply_wqattrs_unlock(); !! 4286 wq->rescuer = rescuer;
>> 4287 rescuer->task->flags |= PF_NO_SETAFFINITY;
>> 4288 wake_up_process(rescuer->task);
>> 4289 }
5718 4290
5719 if ((wq->flags & WQ_SYSFS) && workque 4291 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
5720 goto err_destroy; 4292 goto err_destroy;
5721 4293
5722 return wq; <<
5723 <<
5724 err_unlock_free_node_nr_active: <<
5725 apply_wqattrs_unlock(); <<
5726 /* 4294 /*
5727 * Failed alloc_and_link_pwqs() may l !! 4295 * wq_pool_mutex protects global freeze state and workqueues list.
5728 * flushing the pwq_release_worker en !! 4296 * Grab it, adjust max_active and add the new @wq to workqueues
5729 * completes before calling kfree(wq) !! 4297 * list.
5730 */ 4298 */
5731 if (wq->flags & WQ_UNBOUND) { !! 4299 mutex_lock(&wq_pool_mutex);
5732 kthread_flush_worker(pwq_rele !! 4300
5733 free_node_nr_active(wq->node_ !! 4301 mutex_lock(&wq->mutex);
5734 } !! 4302 for_each_pwq(pwq, wq)
5735 err_unreg_lockdep: !! 4303 pwq_adjust_max_active(pwq);
5736 wq_unregister_lockdep(wq); !! 4304 mutex_unlock(&wq->mutex);
5737 wq_free_lockdep(wq); !! 4305
>> 4306 list_add(&wq->list, &workqueues);
>> 4307
>> 4308 mutex_unlock(&wq_pool_mutex);
>> 4309
>> 4310 return wq;
>> 4311
5738 err_free_wq: 4312 err_free_wq:
5739 free_workqueue_attrs(wq->unbound_attr 4313 free_workqueue_attrs(wq->unbound_attrs);
5740 kfree(wq); 4314 kfree(wq);
5741 return NULL; 4315 return NULL;
5742 err_unlock_destroy: <<
5743 apply_wqattrs_unlock(); <<
5744 err_destroy: 4316 err_destroy:
5745 destroy_workqueue(wq); 4317 destroy_workqueue(wq);
5746 return NULL; 4318 return NULL;
5747 } 4319 }
5748 EXPORT_SYMBOL_GPL(alloc_workqueue); !! 4320 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
5749 <<
5750 static bool pwq_busy(struct pool_workqueue *p <<
5751 { <<
5752 int i; <<
5753 <<
5754 for (i = 0; i < WORK_NR_COLORS; i++) <<
5755 if (pwq->nr_in_flight[i]) <<
5756 return true; <<
5757 <<
5758 if ((pwq != rcu_access_pointer(pwq->w <<
5759 return true; <<
5760 if (!pwq_is_empty(pwq)) <<
5761 return true; <<
5762 <<
5763 return false; <<
5764 } <<
5765 4321
5766 /** 4322 /**
5767 * destroy_workqueue - safely terminate a wor 4323 * destroy_workqueue - safely terminate a workqueue
5768 * @wq: target workqueue 4324 * @wq: target workqueue
5769 * 4325 *
5770 * Safely destroy a workqueue. All work curre 4326 * Safely destroy a workqueue. All work currently pending will be done first.
5771 */ 4327 */
5772 void destroy_workqueue(struct workqueue_struc 4328 void destroy_workqueue(struct workqueue_struct *wq)
5773 { 4329 {
5774 struct pool_workqueue *pwq; 4330 struct pool_workqueue *pwq;
5775 int cpu; !! 4331 int node;
5776 <<
5777 /* <<
5778 * Remove it from sysfs first so that <<
5779 * lead to sysfs name conflicts. <<
5780 */ <<
5781 workqueue_sysfs_unregister(wq); <<
5782 <<
5783 /* mark the workqueue destruction is <<
5784 mutex_lock(&wq->mutex); <<
5785 wq->flags |= __WQ_DESTROYING; <<
5786 mutex_unlock(&wq->mutex); <<
5787 4332
5788 /* drain it before proceeding with de 4333 /* drain it before proceeding with destruction */
5789 drain_workqueue(wq); 4334 drain_workqueue(wq);
5790 4335
5791 /* kill rescuer, if sanity checks fai !! 4336 /* sanity checks */
5792 if (wq->rescuer) { <<
5793 struct worker *rescuer = wq-> <<
5794 <<
5795 /* this prevents new queueing <<
5796 raw_spin_lock_irq(&wq_mayday_ <<
5797 wq->rescuer = NULL; <<
5798 raw_spin_unlock_irq(&wq_mayda <<
5799 <<
5800 /* rescuer will empty maydays <<
5801 kthread_stop(rescuer->task); <<
5802 kfree(rescuer); <<
5803 } <<
5804 <<
5805 /* <<
5806 * Sanity checks - grab all the locks <<
5807 * in-flight operations which may do <<
5808 */ <<
5809 mutex_lock(&wq_pool_mutex); <<
5810 mutex_lock(&wq->mutex); 4337 mutex_lock(&wq->mutex);
5811 for_each_pwq(pwq, wq) { 4338 for_each_pwq(pwq, wq) {
5812 raw_spin_lock_irq(&pwq->pool- !! 4339 int i;
5813 if (WARN_ON(pwq_busy(pwq))) { !! 4340
5814 pr_warn("%s: %s has t !! 4341 for (i = 0; i < WORK_NR_COLORS; i++) {
5815 __func__, wq- !! 4342 if (WARN_ON(pwq->nr_in_flight[i])) {
5816 show_pwq(pwq); !! 4343 mutex_unlock(&wq->mutex);
5817 raw_spin_unlock_irq(& !! 4344 return;
>> 4345 }
>> 4346 }
>> 4347
>> 4348 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
>> 4349 WARN_ON(pwq->nr_active) ||
>> 4350 WARN_ON(!list_empty(&pwq->delayed_works))) {
5818 mutex_unlock(&wq->mut 4351 mutex_unlock(&wq->mutex);
5819 mutex_unlock(&wq_pool <<
5820 show_one_workqueue(wq <<
5821 return; 4352 return;
5822 } 4353 }
5823 raw_spin_unlock_irq(&pwq->poo <<
5824 } 4354 }
5825 mutex_unlock(&wq->mutex); 4355 mutex_unlock(&wq->mutex);
5826 4356
5827 /* 4357 /*
5828 * wq list is used to freeze wq, remo 4358 * wq list is used to freeze wq, remove from list after
5829 * flushing is complete in case freez 4359 * flushing is complete in case freeze races us.
5830 */ 4360 */
5831 list_del_rcu(&wq->list); !! 4361 mutex_lock(&wq_pool_mutex);
>> 4362 list_del_init(&wq->list);
5832 mutex_unlock(&wq_pool_mutex); 4363 mutex_unlock(&wq_pool_mutex);
5833 4364
5834 /* !! 4365 workqueue_sysfs_unregister(wq);
5835 * We're the sole accessor of @wq. Di <<
5836 * to put the base refs. @wq will be <<
5837 * pwq_put. RCU read lock prevents @w <<
5838 */ <<
5839 rcu_read_lock(); <<
5840 4366
5841 for_each_possible_cpu(cpu) { !! 4367 if (wq->rescuer) {
5842 put_pwq_unlocked(unbound_pwq( !! 4368 kthread_stop(wq->rescuer->task);
5843 RCU_INIT_POINTER(*unbound_pwq !! 4369 kfree(wq->rescuer);
>> 4370 wq->rescuer = NULL;
5844 } 4371 }
5845 4372
5846 put_pwq_unlocked(unbound_pwq(wq, -1)) !! 4373 if (!(wq->flags & WQ_UNBOUND)) {
5847 RCU_INIT_POINTER(*unbound_pwq_slot(wq !! 4374 /*
>> 4375 * The base ref is never dropped on per-cpu pwqs. Directly
>> 4376 * free the pwqs and wq.
>> 4377 */
>> 4378 free_percpu(wq->cpu_pwqs);
>> 4379 kfree(wq);
>> 4380 } else {
>> 4381 /*
>> 4382 * We're the sole accessor of @wq at this point. Directly
>> 4383 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
>> 4384 * @wq will be freed when the last pwq is released.
>> 4385 */
>> 4386 for_each_node(node) {
>> 4387 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
>> 4388 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
>> 4389 put_pwq_unlocked(pwq);
>> 4390 }
5848 4391
5849 rcu_read_unlock(); !! 4392 /*
>> 4393 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
>> 4394 * put. Don't access it afterwards.
>> 4395 */
>> 4396 pwq = wq->dfl_pwq;
>> 4397 wq->dfl_pwq = NULL;
>> 4398 put_pwq_unlocked(pwq);
>> 4399 }
5850 } 4400 }
5851 EXPORT_SYMBOL_GPL(destroy_workqueue); 4401 EXPORT_SYMBOL_GPL(destroy_workqueue);
5852 4402
5853 /** 4403 /**
5854 * workqueue_set_max_active - adjust max_acti 4404 * workqueue_set_max_active - adjust max_active of a workqueue
5855 * @wq: target workqueue 4405 * @wq: target workqueue
5856 * @max_active: new max_active value. 4406 * @max_active: new max_active value.
5857 * 4407 *
5858 * Set max_active of @wq to @max_active. See !! 4408 * Set max_active of @wq to @max_active.
5859 * comment. <<
5860 * 4409 *
5861 * CONTEXT: 4410 * CONTEXT:
5862 * Don't call from IRQ context. 4411 * Don't call from IRQ context.
5863 */ 4412 */
5864 void workqueue_set_max_active(struct workqueu 4413 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
5865 { 4414 {
5866 /* max_active doesn't mean anything f !! 4415 struct pool_workqueue *pwq;
5867 if (WARN_ON(wq->flags & WQ_BH)) !! 4416
5868 return; <<
5869 /* disallow meddling with max_active 4417 /* disallow meddling with max_active for ordered workqueues */
5870 if (WARN_ON(wq->flags & __WQ_ORDERED) !! 4418 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5871 return; 4419 return;
5872 4420
5873 max_active = wq_clamp_max_active(max_ 4421 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
5874 4422
5875 mutex_lock(&wq->mutex); 4423 mutex_lock(&wq->mutex);
5876 4424
>> 4425 wq->flags &= ~__WQ_ORDERED;
5877 wq->saved_max_active = max_active; 4426 wq->saved_max_active = max_active;
5878 if (wq->flags & WQ_UNBOUND) <<
5879 wq->saved_min_active = min(wq <<
5880 4427
5881 wq_adjust_max_active(wq); !! 4428 for_each_pwq(pwq, wq)
>> 4429 pwq_adjust_max_active(pwq);
5882 4430
5883 mutex_unlock(&wq->mutex); 4431 mutex_unlock(&wq->mutex);
5884 } 4432 }
5885 EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4433 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
5886 4434
5887 /** 4435 /**
5888 * workqueue_set_min_active - adjust min_acti <<
5889 * @wq: target unbound workqueue <<
5890 * @min_active: new min_active value <<
5891 * <<
5892 * Set min_active of an unbound workqueue. Un <<
5893 * unbound workqueue is not guaranteed to be <<
5894 * interdependent work items. Instead, an unb <<
5895 * able to process min_active number of inter <<
5896 * %WQ_DFL_MIN_ACTIVE by default. <<
5897 * <<
5898 * Use this function to adjust the min_active <<
5899 * max_active. <<
5900 */ <<
5901 void workqueue_set_min_active(struct workqueu <<
5902 { <<
5903 /* min_active is only meaningful for <<
5904 if (WARN_ON((wq->flags & (WQ_BH | WQ_ <<
5905 WQ_UNBOUND)) <<
5906 return; <<
5907 <<
5908 mutex_lock(&wq->mutex); <<
5909 wq->saved_min_active = clamp(min_acti <<
5910 wq_adjust_max_active(wq); <<
5911 mutex_unlock(&wq->mutex); <<
5912 } <<
5913 <<
5914 /** <<
5915 * current_work - retrieve %current task's wo <<
5916 * <<
5917 * Determine if %current task is a workqueue <<
5918 * Useful to find out the context that the %c <<
5919 * <<
5920 * Return: work struct if %current task is a <<
5921 */ <<
5922 struct work_struct *current_work(void) <<
5923 { <<
5924 struct worker *worker = current_wq_wo <<
5925 <<
5926 return worker ? worker->current_work <<
5927 } <<
5928 EXPORT_SYMBOL(current_work); <<
5929 <<
5930 /** <<
5931 * current_is_workqueue_rescuer - is %current 4436 * current_is_workqueue_rescuer - is %current workqueue rescuer?
5932 * 4437 *
5933 * Determine whether %current is a workqueue 4438 * Determine whether %current is a workqueue rescuer. Can be used from
5934 * work functions to determine whether it's b 4439 * work functions to determine whether it's being run off the rescuer task.
5935 * <<
5936 * Return: %true if %current is a workqueue r <<
5937 */ 4440 */
5938 bool current_is_workqueue_rescuer(void) 4441 bool current_is_workqueue_rescuer(void)
5939 { 4442 {
5940 struct worker *worker = current_wq_wo 4443 struct worker *worker = current_wq_worker();
5941 4444
5942 return worker && worker->rescue_wq; 4445 return worker && worker->rescue_wq;
5943 } 4446 }
5944 4447
5945 /** 4448 /**
5946 * workqueue_congested - test whether a workq 4449 * workqueue_congested - test whether a workqueue is congested
5947 * @cpu: CPU in question 4450 * @cpu: CPU in question
5948 * @wq: target workqueue 4451 * @wq: target workqueue
5949 * 4452 *
5950 * Test whether @wq's cpu workqueue for @cpu 4453 * Test whether @wq's cpu workqueue for @cpu is congested. There is
5951 * no synchronization around this function an 4454 * no synchronization around this function and the test result is
5952 * unreliable and only useful as advisory hin 4455 * unreliable and only useful as advisory hints or for debugging.
5953 * 4456 *
5954 * If @cpu is WORK_CPU_UNBOUND, the test is p 4457 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
>> 4458 * Note that both per-cpu and unbound workqueues may be associated with
>> 4459 * multiple pool_workqueues which have separate congested states. A
>> 4460 * workqueue being congested on one CPU doesn't mean the workqueue is also
>> 4461 * contested on other CPUs / NUMA nodes.
5955 * 4462 *
5956 * With the exception of ordered workqueues, !! 4463 * RETURNS:
5957 * pool_workqueues, each with its own congest <<
5958 * congested on one CPU doesn't mean that the <<
5959 * other CPUs. <<
5960 * <<
5961 * Return: <<
5962 * %true if congested, %false otherwise. 4464 * %true if congested, %false otherwise.
5963 */ 4465 */
5964 bool workqueue_congested(int cpu, struct work 4466 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
5965 { 4467 {
5966 struct pool_workqueue *pwq; 4468 struct pool_workqueue *pwq;
5967 bool ret; 4469 bool ret;
5968 4470
5969 rcu_read_lock(); !! 4471 rcu_read_lock_sched();
5970 preempt_disable(); <<
5971 4472
5972 if (cpu == WORK_CPU_UNBOUND) 4473 if (cpu == WORK_CPU_UNBOUND)
5973 cpu = smp_processor_id(); 4474 cpu = smp_processor_id();
5974 4475
5975 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); !! 4476 if (!(wq->flags & WQ_UNBOUND))
5976 ret = !list_empty(&pwq->inactive_work !! 4477 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
>> 4478 else
>> 4479 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
5977 4480
5978 preempt_enable(); !! 4481 ret = !list_empty(&pwq->delayed_works);
5979 rcu_read_unlock(); !! 4482 rcu_read_unlock_sched();
5980 4483
5981 return ret; 4484 return ret;
5982 } 4485 }
5983 EXPORT_SYMBOL_GPL(workqueue_congested); 4486 EXPORT_SYMBOL_GPL(workqueue_congested);
5984 4487
5985 /** 4488 /**
5986 * work_busy - test whether a work is current 4489 * work_busy - test whether a work is currently pending or running
5987 * @work: the work to be tested 4490 * @work: the work to be tested
5988 * 4491 *
5989 * Test whether @work is currently pending or 4492 * Test whether @work is currently pending or running. There is no
5990 * synchronization around this function and t 4493 * synchronization around this function and the test result is
5991 * unreliable and only useful as advisory hin 4494 * unreliable and only useful as advisory hints or for debugging.
5992 * 4495 *
5993 * Return: !! 4496 * RETURNS:
5994 * OR'd bitmask of WORK_BUSY_* bits. 4497 * OR'd bitmask of WORK_BUSY_* bits.
5995 */ 4498 */
5996 unsigned int work_busy(struct work_struct *wo 4499 unsigned int work_busy(struct work_struct *work)
5997 { 4500 {
5998 struct worker_pool *pool; 4501 struct worker_pool *pool;
5999 unsigned long irq_flags; !! 4502 unsigned long flags;
6000 unsigned int ret = 0; 4503 unsigned int ret = 0;
6001 4504
6002 if (work_pending(work)) 4505 if (work_pending(work))
6003 ret |= WORK_BUSY_PENDING; 4506 ret |= WORK_BUSY_PENDING;
6004 4507
6005 rcu_read_lock(); !! 4508 local_irq_save(flags);
6006 pool = get_work_pool(work); 4509 pool = get_work_pool(work);
6007 if (pool) { 4510 if (pool) {
6008 raw_spin_lock_irqsave(&pool-> !! 4511 spin_lock(&pool->lock);
6009 if (find_worker_executing_wor 4512 if (find_worker_executing_work(pool, work))
6010 ret |= WORK_BUSY_RUNN 4513 ret |= WORK_BUSY_RUNNING;
6011 raw_spin_unlock_irqrestore(&p !! 4514 spin_unlock(&pool->lock);
6012 } 4515 }
6013 rcu_read_unlock(); !! 4516 local_irq_restore(flags);
6014 4517
6015 return ret; 4518 return ret;
6016 } 4519 }
6017 EXPORT_SYMBOL_GPL(work_busy); 4520 EXPORT_SYMBOL_GPL(work_busy);
6018 4521
6019 /** 4522 /**
6020 * set_worker_desc - set description for the 4523 * set_worker_desc - set description for the current work item
6021 * @fmt: printf-style format string 4524 * @fmt: printf-style format string
6022 * @...: arguments for the format string 4525 * @...: arguments for the format string
6023 * 4526 *
6024 * This function can be called by a running w 4527 * This function can be called by a running work function to describe what
6025 * the work item is about. If the worker tas 4528 * the work item is about. If the worker task gets dumped, this
6026 * information will be printed out together t 4529 * information will be printed out together to help debugging. The
6027 * description can be at most WORKER_DESC_LEN 4530 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
6028 */ 4531 */
6029 void set_worker_desc(const char *fmt, ...) 4532 void set_worker_desc(const char *fmt, ...)
6030 { 4533 {
6031 struct worker *worker = current_wq_wo 4534 struct worker *worker = current_wq_worker();
6032 va_list args; 4535 va_list args;
6033 4536
6034 if (worker) { 4537 if (worker) {
6035 va_start(args, fmt); 4538 va_start(args, fmt);
6036 vsnprintf(worker->desc, sizeo 4539 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
6037 va_end(args); 4540 va_end(args);
>> 4541 worker->desc_valid = true;
6038 } 4542 }
6039 } 4543 }
6040 EXPORT_SYMBOL_GPL(set_worker_desc); <<
6041 4544
6042 /** 4545 /**
6043 * print_worker_info - print out worker infor 4546 * print_worker_info - print out worker information and description
6044 * @log_lvl: the log level to use when printi 4547 * @log_lvl: the log level to use when printing
6045 * @task: target task 4548 * @task: target task
6046 * 4549 *
6047 * If @task is a worker and currently executi 4550 * If @task is a worker and currently executing a work item, print out the
6048 * name of the workqueue being serviced and w 4551 * name of the workqueue being serviced and worker description set with
6049 * set_worker_desc() by the currently executi 4552 * set_worker_desc() by the currently executing work item.
6050 * 4553 *
6051 * This function can be safely called on any 4554 * This function can be safely called on any task as long as the
6052 * task_struct itself is accessible. While s 4555 * task_struct itself is accessible. While safe, this function isn't
6053 * synchronized and may print out mixups or g 4556 * synchronized and may print out mixups or garbages of limited length.
6054 */ 4557 */
6055 void print_worker_info(const char *log_lvl, s 4558 void print_worker_info(const char *log_lvl, struct task_struct *task)
6056 { 4559 {
6057 work_func_t *fn = NULL; 4560 work_func_t *fn = NULL;
6058 char name[WQ_NAME_LEN] = { }; 4561 char name[WQ_NAME_LEN] = { };
6059 char desc[WORKER_DESC_LEN] = { }; 4562 char desc[WORKER_DESC_LEN] = { };
6060 struct pool_workqueue *pwq = NULL; 4563 struct pool_workqueue *pwq = NULL;
6061 struct workqueue_struct *wq = NULL; 4564 struct workqueue_struct *wq = NULL;
>> 4565 bool desc_valid = false;
6062 struct worker *worker; 4566 struct worker *worker;
6063 4567
6064 if (!(task->flags & PF_WQ_WORKER)) 4568 if (!(task->flags & PF_WQ_WORKER))
6065 return; 4569 return;
6066 4570
6067 /* 4571 /*
6068 * This function is called without an 4572 * This function is called without any synchronization and @task
6069 * could be in any state. Be careful 4573 * could be in any state. Be careful with dereferences.
6070 */ 4574 */
6071 worker = kthread_probe_data(task); !! 4575 worker = probe_kthread_data(task);
6072 4576
6073 /* 4577 /*
6074 * Carefully copy the associated work !! 4578 * Carefully copy the associated workqueue's workfn and name. Keep
6075 * Keep the original last '\0' in cas !! 4579 * the original last '\0' in case the original contains garbage.
6076 */ 4580 */
6077 copy_from_kernel_nofault(&fn, &worker !! 4581 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
6078 copy_from_kernel_nofault(&pwq, &worke !! 4582 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
6079 copy_from_kernel_nofault(&wq, &pwq->w !! 4583 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
6080 copy_from_kernel_nofault(name, wq->na !! 4584 probe_kernel_read(name, wq->name, sizeof(name) - 1);
6081 copy_from_kernel_nofault(desc, worker !! 4585
>> 4586 /* copy worker description */
>> 4587 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
>> 4588 if (desc_valid)
>> 4589 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
6082 4590
6083 if (fn || name[0] || desc[0]) { 4591 if (fn || name[0] || desc[0]) {
6084 printk("%sWorkqueue: %s %ps", !! 4592 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
6085 if (strcmp(name, desc)) !! 4593 if (desc[0])
6086 pr_cont(" (%s)", desc 4594 pr_cont(" (%s)", desc);
6087 pr_cont("\n"); 4595 pr_cont("\n");
6088 } 4596 }
6089 } 4597 }
6090 4598
6091 static void pr_cont_pool_info(struct worker_p <<
6092 { <<
6093 pr_cont(" cpus=%*pbl", nr_cpumask_bit <<
6094 if (pool->node != NUMA_NO_NODE) <<
6095 pr_cont(" node=%d", pool->nod <<
6096 pr_cont(" flags=0x%x", pool->flags); <<
6097 if (pool->flags & POOL_BH) <<
6098 pr_cont(" bh%s", <<
6099 pool->attrs->nice == <<
6100 else <<
6101 pr_cont(" nice=%d", pool->att <<
6102 } <<
6103 <<
6104 static void pr_cont_worker_id(struct worker * <<
6105 { <<
6106 struct worker_pool *pool = worker->po <<
6107 <<
6108 if (pool->flags & WQ_BH) <<
6109 pr_cont("bh%s", <<
6110 pool->attrs->nice == <<
6111 else <<
6112 pr_cont("%d%s", task_pid_nr(w <<
6113 worker->rescue_wq ? " <<
6114 } <<
6115 <<
6116 struct pr_cont_work_struct { <<
6117 bool comma; <<
6118 work_func_t func; <<
6119 long ctr; <<
6120 }; <<
6121 <<
6122 static void pr_cont_work_flush(bool comma, wo <<
6123 { <<
6124 if (!pcwsp->ctr) <<
6125 goto out_record; <<
6126 if (func == pcwsp->func) { <<
6127 pcwsp->ctr++; <<
6128 return; <<
6129 } <<
6130 if (pcwsp->ctr == 1) <<
6131 pr_cont("%s %ps", pcwsp->comm <<
6132 else <<
6133 pr_cont("%s %ld*%ps", pcwsp-> <<
6134 pcwsp->ctr = 0; <<
6135 out_record: <<
6136 if ((long)func == -1L) <<
6137 return; <<
6138 pcwsp->comma = comma; <<
6139 pcwsp->func = func; <<
6140 pcwsp->ctr = 1; <<
6141 } <<
6142 <<
6143 static void pr_cont_work(bool comma, struct w <<
6144 { <<
6145 if (work->func == wq_barrier_func) { <<
6146 struct wq_barrier *barr; <<
6147 <<
6148 barr = container_of(work, str <<
6149 <<
6150 pr_cont_work_flush(comma, (wo <<
6151 pr_cont("%s BAR(%d)", comma ? <<
6152 task_pid_nr(barr->tas <<
6153 } else { <<
6154 if (!comma) <<
6155 pr_cont_work_flush(co <<
6156 pr_cont_work_flush(comma, wor <<
6157 } <<
6158 } <<
6159 <<
6160 static void show_pwq(struct pool_workqueue *p <<
6161 { <<
6162 struct pr_cont_work_struct pcws = { . <<
6163 struct worker_pool *pool = pwq->pool; <<
6164 struct work_struct *work; <<
6165 struct worker *worker; <<
6166 bool has_in_flight = false, has_pendi <<
6167 int bkt; <<
6168 <<
6169 pr_info(" pwq %d:", pool->id); <<
6170 pr_cont_pool_info(pool); <<
6171 <<
6172 pr_cont(" active=%d refcnt=%d%s\n", <<
6173 pwq->nr_active, pwq->refcnt, <<
6174 !list_empty(&pwq->mayday_node <<
6175 <<
6176 hash_for_each(pool->busy_hash, bkt, w <<
6177 if (worker->current_pwq == pw <<
6178 has_in_flight = true; <<
6179 break; <<
6180 } <<
6181 } <<
6182 if (has_in_flight) { <<
6183 bool comma = false; <<
6184 <<
6185 pr_info(" in-flight:"); <<
6186 hash_for_each(pool->busy_hash <<
6187 if (worker->current_p <<
6188 continue; <<
6189 <<
6190 pr_cont(" %s", comma <<
6191 pr_cont_worker_id(wor <<
6192 pr_cont(":%ps", worke <<
6193 list_for_each_entry(w <<
6194 pr_cont_work( <<
6195 pr_cont_work_flush(co <<
6196 comma = true; <<
6197 } <<
6198 pr_cont("\n"); <<
6199 } <<
6200 <<
6201 list_for_each_entry(work, &pool->work <<
6202 if (get_work_pwq(work) == pwq <<
6203 has_pending = true; <<
6204 break; <<
6205 } <<
6206 } <<
6207 if (has_pending) { <<
6208 bool comma = false; <<
6209 <<
6210 pr_info(" pending:"); <<
6211 list_for_each_entry(work, &po <<
6212 if (get_work_pwq(work <<
6213 continue; <<
6214 <<
6215 pr_cont_work(comma, w <<
6216 comma = !(*work_data_ <<
6217 } <<
6218 pr_cont_work_flush(comma, (wo <<
6219 pr_cont("\n"); <<
6220 } <<
6221 <<
6222 if (!list_empty(&pwq->inactive_works) <<
6223 bool comma = false; <<
6224 <<
6225 pr_info(" inactive:"); <<
6226 list_for_each_entry(work, &pw <<
6227 pr_cont_work(comma, w <<
6228 comma = !(*work_data_ <<
6229 } <<
6230 pr_cont_work_flush(comma, (wo <<
6231 pr_cont("\n"); <<
6232 } <<
6233 } <<
6234 <<
6235 /** <<
6236 * show_one_workqueue - dump state of specifi <<
6237 * @wq: workqueue whose state will be printed <<
6238 */ <<
6239 void show_one_workqueue(struct workqueue_stru <<
6240 { <<
6241 struct pool_workqueue *pwq; <<
6242 bool idle = true; <<
6243 unsigned long irq_flags; <<
6244 <<
6245 for_each_pwq(pwq, wq) { <<
6246 if (!pwq_is_empty(pwq)) { <<
6247 idle = false; <<
6248 break; <<
6249 } <<
6250 } <<
6251 if (idle) /* Nothing to print for idl <<
6252 return; <<
6253 <<
6254 pr_info("workqueue %s: flags=0x%x\n", <<
6255 <<
6256 for_each_pwq(pwq, wq) { <<
6257 raw_spin_lock_irqsave(&pwq->p <<
6258 if (!pwq_is_empty(pwq)) { <<
6259 /* <<
6260 * Defer printing to <<
6261 * drivers that queue <<
6262 * also taken in thei <<
6263 */ <<
6264 printk_deferred_enter <<
6265 show_pwq(pwq); <<
6266 printk_deferred_exit( <<
6267 } <<
6268 raw_spin_unlock_irqrestore(&p <<
6269 /* <<
6270 * We could be printing a lot <<
6271 * sysrq-t -> show_all_workqu <<
6272 * hard lockup. <<
6273 */ <<
6274 touch_nmi_watchdog(); <<
6275 } <<
6276 <<
6277 } <<
6278 <<
6279 /** <<
6280 * show_one_worker_pool - dump state of speci <<
6281 * @pool: worker pool whose state will be pri <<
6282 */ <<
6283 static void show_one_worker_pool(struct worke <<
6284 { <<
6285 struct worker *worker; <<
6286 bool first = true; <<
6287 unsigned long irq_flags; <<
6288 unsigned long hung = 0; <<
6289 <<
6290 raw_spin_lock_irqsave(&pool->lock, ir <<
6291 if (pool->nr_workers == pool->nr_idle <<
6292 goto next_pool; <<
6293 <<
6294 /* How long the first pending work is <<
6295 if (!list_empty(&pool->worklist)) <<
6296 hung = jiffies_to_msecs(jiffi <<
6297 <<
6298 /* <<
6299 * Defer printing to avoid deadlocks <<
6300 * queue work while holding locks als <<
6301 * paths. <<
6302 */ <<
6303 printk_deferred_enter(); <<
6304 pr_info("pool %d:", pool->id); <<
6305 pr_cont_pool_info(pool); <<
6306 pr_cont(" hung=%lus workers=%d", hung <<
6307 if (pool->manager) <<
6308 pr_cont(" manager: %d", <<
6309 task_pid_nr(pool->man <<
6310 list_for_each_entry(worker, &pool->id <<
6311 pr_cont(" %s", first ? "idle: <<
6312 pr_cont_worker_id(worker); <<
6313 first = false; <<
6314 } <<
6315 pr_cont("\n"); <<
6316 printk_deferred_exit(); <<
6317 next_pool: <<
6318 raw_spin_unlock_irqrestore(&pool->loc <<
6319 /* <<
6320 * We could be printing a lot from at <<
6321 * sysrq-t -> show_all_workqueues(). <<
6322 * hard lockup. <<
6323 */ <<
6324 touch_nmi_watchdog(); <<
6325 <<
6326 } <<
6327 <<
6328 /** <<
6329 * show_all_workqueues - dump workqueue state <<
6330 * <<
6331 * Called from a sysrq handler and prints out <<
6332 */ <<
6333 void show_all_workqueues(void) <<
6334 { <<
6335 struct workqueue_struct *wq; <<
6336 struct worker_pool *pool; <<
6337 int pi; <<
6338 <<
6339 rcu_read_lock(); <<
6340 <<
6341 pr_info("Showing busy workqueues and <<
6342 <<
6343 list_for_each_entry_rcu(wq, &workqueu <<
6344 show_one_workqueue(wq); <<
6345 <<
6346 for_each_pool(pool, pi) <<
6347 show_one_worker_pool(pool); <<
6348 <<
6349 rcu_read_unlock(); <<
6350 } <<
6351 <<
6352 /** <<
6353 * show_freezable_workqueues - dump freezable <<
6354 * <<
6355 * Called from try_to_freeze_tasks() and prin <<
6356 * still busy. <<
6357 */ <<
6358 void show_freezable_workqueues(void) <<
6359 { <<
6360 struct workqueue_struct *wq; <<
6361 <<
6362 rcu_read_lock(); <<
6363 <<
6364 pr_info("Showing freezable workqueues <<
6365 <<
6366 list_for_each_entry_rcu(wq, &workqueu <<
6367 if (!(wq->flags & WQ_FREEZABL <<
6368 continue; <<
6369 show_one_workqueue(wq); <<
6370 } <<
6371 <<
6372 rcu_read_unlock(); <<
6373 } <<
6374 <<
6375 /* used to show worker information through /p <<
6376 void wq_worker_comm(char *buf, size_t size, s <<
6377 { <<
6378 /* stabilize PF_WQ_WORKER and worker <<
6379 mutex_lock(&wq_pool_attach_mutex); <<
6380 <<
6381 if (task->flags & PF_WQ_WORKER) { <<
6382 struct worker *worker = kthre <<
6383 struct worker_pool *pool = wo <<
6384 int off; <<
6385 <<
6386 off = format_worker_id(buf, s <<
6387 <<
6388 if (pool) { <<
6389 raw_spin_lock_irq(&po <<
6390 /* <<
6391 * ->desc tracks info <<
6392 * set_worker_desc()) <<
6393 * current, prepend ' <<
6394 */ <<
6395 if (worker->desc[0] ! <<
6396 if (worker->c <<
6397 scnpr <<
6398 <<
6399 else <<
6400 scnpr <<
6401 <<
6402 } <<
6403 raw_spin_unlock_irq(& <<
6404 } <<
6405 } else { <<
6406 strscpy(buf, task->comm, size <<
6407 } <<
6408 <<
6409 mutex_unlock(&wq_pool_attach_mutex); <<
6410 } <<
6411 <<
6412 #ifdef CONFIG_SMP <<
6413 <<
6414 /* 4599 /*
6415 * CPU hotplug. 4600 * CPU hotplug.
6416 * 4601 *
6417 * There are two challenges in supporting CPU 4602 * There are two challenges in supporting CPU hotplug. Firstly, there
6418 * are a lot of assumptions on strong associa 4603 * are a lot of assumptions on strong associations among work, pwq and
6419 * pool which make migrating pending and sche 4604 * pool which make migrating pending and scheduled works very
6420 * difficult to implement without impacting h 4605 * difficult to implement without impacting hot paths. Secondly,
6421 * worker pools serve mix of short, long and 4606 * worker pools serve mix of short, long and very long running works making
6422 * blocked draining impractical. 4607 * blocked draining impractical.
6423 * 4608 *
6424 * This is solved by allowing the pools to be 4609 * This is solved by allowing the pools to be disassociated from the CPU
6425 * running as an unbound one and allowing it 4610 * running as an unbound one and allowing it to be reattached later if the
6426 * cpu comes back online. 4611 * cpu comes back online.
6427 */ 4612 */
6428 4613
6429 static void unbind_workers(int cpu) !! 4614 static void wq_unbind_fn(struct work_struct *work)
6430 { 4615 {
>> 4616 int cpu = smp_processor_id();
6431 struct worker_pool *pool; 4617 struct worker_pool *pool;
6432 struct worker *worker; 4618 struct worker *worker;
>> 4619 int wi;
6433 4620
6434 for_each_cpu_worker_pool(pool, cpu) { 4621 for_each_cpu_worker_pool(pool, cpu) {
6435 mutex_lock(&wq_pool_attach_mu !! 4622 WARN_ON_ONCE(cpu != smp_processor_id());
6436 raw_spin_lock_irq(&pool->lock !! 4623
>> 4624 mutex_lock(&pool->manager_mutex);
>> 4625 spin_lock_irq(&pool->lock);
6437 4626
6438 /* 4627 /*
6439 * We've blocked all attach/d !! 4628 * We've blocked all manager operations. Make all workers
6440 * unbound and set DISASSOCIA 4629 * unbound and set DISASSOCIATED. Before this, all workers
6441 * must be on the cpu. After !! 4630 * except for the ones which are still executing works from
6442 * And the preemption disable !! 4631 * before the last CPU down must be on the cpu. After
6443 * are guaranteed to see WORK !! 4632 * this, they may become diasporas.
6444 * is on the same cpu. <<
6445 */ 4633 */
6446 for_each_pool_worker(worker, !! 4634 for_each_pool_worker(worker, wi, pool)
6447 worker->flags |= WORK 4635 worker->flags |= WORKER_UNBOUND;
6448 4636
6449 pool->flags |= POOL_DISASSOCI 4637 pool->flags |= POOL_DISASSOCIATED;
6450 4638
>> 4639 spin_unlock_irq(&pool->lock);
>> 4640 mutex_unlock(&pool->manager_mutex);
>> 4641
6451 /* 4642 /*
6452 * The handling of nr_running !! 4643 * Call schedule() so that we cross rq->lock and thus can
6453 * now. Zap nr_running. Aft !! 4644 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
6454 * need_more_worker() and kee !! 4645 * This is necessary as scheduler callbacks may be invoked
6455 * long as the worklist is no !! 4646 * from other cpus.
6456 * an unbound (in terms of co !! 4647 */
>> 4648 schedule();
>> 4649
>> 4650 /*
>> 4651 * Sched callbacks are disabled now. Zap nr_running.
>> 4652 * After this, nr_running stays zero and need_more_worker()
>> 4653 * and keep_working() are always true as long as the
>> 4654 * worklist is not empty. This pool now behaves as an
>> 4655 * unbound (in terms of concurrency management) pool which
6457 * are served by workers tied 4656 * are served by workers tied to the pool.
6458 */ 4657 */
6459 pool->nr_running = 0; !! 4658 atomic_set(&pool->nr_running, 0);
6460 4659
6461 /* 4660 /*
6462 * With concurrency managemen 4661 * With concurrency management just turned off, a busy
6463 * worker blocking could lead 4662 * worker blocking could lead to lengthy stalls. Kick off
6464 * unbound chain execution of 4663 * unbound chain execution of currently pending work items.
6465 */ 4664 */
6466 kick_pool(pool); !! 4665 spin_lock_irq(&pool->lock);
6467 !! 4666 wake_up_worker(pool);
6468 raw_spin_unlock_irq(&pool->lo !! 4667 spin_unlock_irq(&pool->lock);
6469 <<
6470 for_each_pool_worker(worker, <<
6471 unbind_worker(worker) <<
6472 <<
6473 mutex_unlock(&wq_pool_attach_ <<
6474 } 4668 }
6475 } 4669 }
6476 4670
6477 /** 4671 /**
6478 * rebind_workers - rebind all workers of a p 4672 * rebind_workers - rebind all workers of a pool to the associated CPU
6479 * @pool: pool of interest 4673 * @pool: pool of interest
6480 * 4674 *
6481 * @pool->cpu is coming online. Rebind all w 4675 * @pool->cpu is coming online. Rebind all workers to the CPU.
6482 */ 4676 */
6483 static void rebind_workers(struct worker_pool 4677 static void rebind_workers(struct worker_pool *pool)
6484 { 4678 {
6485 struct worker *worker; 4679 struct worker *worker;
>> 4680 int wi;
6486 4681
6487 lockdep_assert_held(&wq_pool_attach_m !! 4682 lockdep_assert_held(&pool->manager_mutex);
6488 4683
6489 /* 4684 /*
6490 * Restore CPU affinity of all worker 4685 * Restore CPU affinity of all workers. As all idle workers should
6491 * be on the run-queue of the associa 4686 * be on the run-queue of the associated CPU before any local
6492 * wake-ups for concurrency managemen !! 4687 * wake-ups for concurrency management happen, restore CPU affinty
6493 * of all workers first and then clea 4688 * of all workers first and then clear UNBOUND. As we're called
6494 * from CPU_ONLINE, the following sho 4689 * from CPU_ONLINE, the following shouldn't fail.
6495 */ 4690 */
6496 for_each_pool_worker(worker, pool) { !! 4691 for_each_pool_worker(worker, wi, pool)
6497 kthread_set_per_cpu(worker->t <<
6498 WARN_ON_ONCE(set_cpus_allowed 4692 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
6499 !! 4693 pool->attrs->cpumask) < 0);
6500 } <<
6501 4694
6502 raw_spin_lock_irq(&pool->lock); !! 4695 spin_lock_irq(&pool->lock);
6503 4696
6504 pool->flags &= ~POOL_DISASSOCIATED; !! 4697 for_each_pool_worker(worker, wi, pool) {
6505 <<
6506 for_each_pool_worker(worker, pool) { <<
6507 unsigned int worker_flags = w 4698 unsigned int worker_flags = worker->flags;
6508 4699
6509 /* 4700 /*
>> 4701 * A bound idle worker should actually be on the runqueue
>> 4702 * of the associated CPU for local wake-ups targeting it to
>> 4703 * work. Kick all idle workers so that they migrate to the
>> 4704 * associated CPU. Doing this in the same loop as
>> 4705 * replacing UNBOUND with REBOUND is safe as no worker will
>> 4706 * be bound before @pool->lock is released.
>> 4707 */
>> 4708 if (worker_flags & WORKER_IDLE)
>> 4709 wake_up_process(worker->task);
>> 4710
>> 4711 /*
6510 * We want to clear UNBOUND b 4712 * We want to clear UNBOUND but can't directly call
6511 * worker_clr_flags() or adju 4713 * worker_clr_flags() or adjust nr_running. Atomically
6512 * replace UNBOUND with anoth 4714 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
6513 * @worker will clear REBOUND 4715 * @worker will clear REBOUND using worker_clr_flags() when
6514 * it initiates the next exec 4716 * it initiates the next execution cycle thus restoring
6515 * concurrency management. N 4717 * concurrency management. Note that when or whether
6516 * @worker clears REBOUND doe 4718 * @worker clears REBOUND doesn't affect correctness.
6517 * 4719 *
6518 * WRITE_ONCE() is necessary !! 4720 * ACCESS_ONCE() is necessary because @worker->flags may be
6519 * tested without holding any 4721 * tested without holding any lock in
6520 * wq_worker_running(). With !! 4722 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
6521 * fail incorrectly leading t 4723 * fail incorrectly leading to premature concurrency
6522 * management operations. 4724 * management operations.
6523 */ 4725 */
6524 WARN_ON_ONCE(!(worker_flags & 4726 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
6525 worker_flags |= WORKER_REBOUN 4727 worker_flags |= WORKER_REBOUND;
6526 worker_flags &= ~WORKER_UNBOU 4728 worker_flags &= ~WORKER_UNBOUND;
6527 WRITE_ONCE(worker->flags, wor !! 4729 ACCESS_ONCE(worker->flags) = worker_flags;
6528 } 4730 }
6529 4731
6530 raw_spin_unlock_irq(&pool->lock); !! 4732 spin_unlock_irq(&pool->lock);
6531 } 4733 }
6532 4734
6533 /** 4735 /**
6534 * restore_unbound_workers_cpumask - restore 4736 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
6535 * @pool: unbound pool of interest 4737 * @pool: unbound pool of interest
6536 * @cpu: the CPU which is coming up 4738 * @cpu: the CPU which is coming up
6537 * 4739 *
6538 * An unbound pool may end up with a cpumask 4740 * An unbound pool may end up with a cpumask which doesn't have any online
6539 * CPUs. When a worker of such pool get sche 4741 * CPUs. When a worker of such pool get scheduled, the scheduler resets
6540 * its cpus_allowed. If @cpu is in @pool's c 4742 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
6541 * online CPU before, cpus_allowed of all its 4743 * online CPU before, cpus_allowed of all its workers should be restored.
6542 */ 4744 */
6543 static void restore_unbound_workers_cpumask(s 4745 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
6544 { 4746 {
6545 static cpumask_t cpumask; 4747 static cpumask_t cpumask;
6546 struct worker *worker; 4748 struct worker *worker;
>> 4749 int wi;
6547 4750
6548 lockdep_assert_held(&wq_pool_attach_m !! 4751 lockdep_assert_held(&pool->manager_mutex);
6549 4752
6550 /* is @cpu allowed for @pool? */ 4753 /* is @cpu allowed for @pool? */
6551 if (!cpumask_test_cpu(cpu, pool->attr 4754 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
6552 return; 4755 return;
6553 4756
>> 4757 /* is @cpu the only online CPU? */
6554 cpumask_and(&cpumask, pool->attrs->cp 4758 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
>> 4759 if (cpumask_weight(&cpumask) != 1)
>> 4760 return;
6555 4761
6556 /* as we're called from CPU_ONLINE, t 4762 /* as we're called from CPU_ONLINE, the following shouldn't fail */
6557 for_each_pool_worker(worker, pool) !! 4763 for_each_pool_worker(worker, wi, pool)
6558 WARN_ON_ONCE(set_cpus_allowed !! 4764 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
6559 } !! 4765 pool->attrs->cpumask) < 0);
6560 <<
6561 int workqueue_prepare_cpu(unsigned int cpu) <<
6562 { <<
6563 struct worker_pool *pool; <<
6564 <<
6565 for_each_cpu_worker_pool(pool, cpu) { <<
6566 if (pool->nr_workers) <<
6567 continue; <<
6568 if (!create_worker(pool)) <<
6569 return -ENOMEM; <<
6570 } <<
6571 return 0; <<
6572 } 4766 }
6573 4767
6574 int workqueue_online_cpu(unsigned int cpu) !! 4768 /*
>> 4769 * Workqueues should be brought up before normal priority CPU notifiers.
>> 4770 * This will be registered high priority CPU notifier.
>> 4771 */
>> 4772 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
>> 4773 unsigned long action,
>> 4774 void *hcpu)
6575 { 4775 {
>> 4776 int cpu = (unsigned long)hcpu;
6576 struct worker_pool *pool; 4777 struct worker_pool *pool;
6577 struct workqueue_struct *wq; 4778 struct workqueue_struct *wq;
6578 int pi; 4779 int pi;
6579 4780
6580 mutex_lock(&wq_pool_mutex); !! 4781 switch (action & ~CPU_TASKS_FROZEN) {
6581 !! 4782 case CPU_UP_PREPARE:
6582 cpumask_set_cpu(cpu, wq_online_cpumas !! 4783 for_each_cpu_worker_pool(pool, cpu) {
>> 4784 if (pool->nr_workers)
>> 4785 continue;
>> 4786 if (create_and_start_worker(pool) < 0)
>> 4787 return NOTIFY_BAD;
>> 4788 }
>> 4789 break;
6583 4790
6584 for_each_pool(pool, pi) { !! 4791 case CPU_DOWN_FAILED:
6585 /* BH pools aren't affected b !! 4792 case CPU_ONLINE:
6586 if (pool->flags & POOL_BH) !! 4793 mutex_lock(&wq_pool_mutex);
6587 continue; <<
6588 4794
6589 mutex_lock(&wq_pool_attach_mu !! 4795 for_each_pool(pool, pi) {
6590 if (pool->cpu == cpu) !! 4796 mutex_lock(&pool->manager_mutex);
6591 rebind_workers(pool); <<
6592 else if (pool->cpu < 0) <<
6593 restore_unbound_worke <<
6594 mutex_unlock(&wq_pool_attach_ <<
6595 } <<
6596 4797
6597 /* update pod affinity of unbound wor !! 4798 if (pool->cpu == cpu) {
6598 list_for_each_entry(wq, &workqueues, !! 4799 spin_lock_irq(&pool->lock);
6599 struct workqueue_attrs *attrs !! 4800 pool->flags &= ~POOL_DISASSOCIATED;
>> 4801 spin_unlock_irq(&pool->lock);
>> 4802
>> 4803 rebind_workers(pool);
>> 4804 } else if (pool->cpu < 0) {
>> 4805 restore_unbound_workers_cpumask(pool, cpu);
>> 4806 }
6600 4807
6601 if (attrs) { !! 4808 mutex_unlock(&pool->manager_mutex);
6602 const struct wq_pod_t !! 4809 }
6603 int tcpu; <<
6604 4810
6605 for_each_cpu(tcpu, pt !! 4811 /* update NUMA affinity of unbound workqueues */
6606 unbound_wq_up !! 4812 list_for_each_entry(wq, &workqueues, list)
>> 4813 wq_update_unbound_numa(wq, cpu, true);
6607 4814
6608 mutex_lock(&wq->mutex !! 4815 mutex_unlock(&wq_pool_mutex);
6609 wq_update_node_max_ac !! 4816 break;
6610 mutex_unlock(&wq->mut <<
6611 } <<
6612 } 4817 }
6613 !! 4818 return NOTIFY_OK;
6614 mutex_unlock(&wq_pool_mutex); <<
6615 return 0; <<
6616 } 4819 }
6617 4820
6618 int workqueue_offline_cpu(unsigned int cpu) !! 4821 /*
>> 4822 * Workqueues should be brought down after normal priority CPU notifiers.
>> 4823 * This will be registered as low priority CPU notifier.
>> 4824 */
>> 4825 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
>> 4826 unsigned long action,
>> 4827 void *hcpu)
6619 { 4828 {
>> 4829 int cpu = (unsigned long)hcpu;
>> 4830 struct work_struct unbind_work;
6620 struct workqueue_struct *wq; 4831 struct workqueue_struct *wq;
6621 4832
6622 /* unbinding per-cpu workers should h !! 4833 switch (action & ~CPU_TASKS_FROZEN) {
6623 if (WARN_ON(cpu != smp_processor_id() !! 4834 case CPU_DOWN_PREPARE:
6624 return -1; !! 4835 /* unbinding per-cpu workers should happen on the local CPU */
6625 !! 4836 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
6626 unbind_workers(cpu); !! 4837 queue_work_on(cpu, system_highpri_wq, &unbind_work);
6627 <<
6628 /* update pod affinity of unbound wor <<
6629 mutex_lock(&wq_pool_mutex); <<
6630 <<
6631 cpumask_clear_cpu(cpu, wq_online_cpum <<
6632 4838
6633 list_for_each_entry(wq, &workqueues, !! 4839 /* update NUMA affinity of unbound workqueues */
6634 struct workqueue_attrs *attrs !! 4840 mutex_lock(&wq_pool_mutex);
6635 !! 4841 list_for_each_entry(wq, &workqueues, list)
6636 if (attrs) { !! 4842 wq_update_unbound_numa(wq, cpu, false);
6637 const struct wq_pod_t !! 4843 mutex_unlock(&wq_pool_mutex);
6638 int tcpu; <<
6639 <<
6640 for_each_cpu(tcpu, pt <<
6641 unbound_wq_up <<
6642 4844
6643 mutex_lock(&wq->mutex !! 4845 /* wait for per-cpu unbinding to finish */
6644 wq_update_node_max_ac !! 4846 flush_work(&unbind_work);
6645 mutex_unlock(&wq->mut !! 4847 break;
6646 } <<
6647 } 4848 }
6648 mutex_unlock(&wq_pool_mutex); !! 4849 return NOTIFY_OK;
6649 <<
6650 return 0; <<
6651 } 4850 }
6652 4851
>> 4852 #ifdef CONFIG_SMP
>> 4853
6653 struct work_for_cpu { 4854 struct work_for_cpu {
6654 struct work_struct work; 4855 struct work_struct work;
6655 long (*fn)(void *); 4856 long (*fn)(void *);
6656 void *arg; 4857 void *arg;
6657 long ret; 4858 long ret;
6658 }; 4859 };
6659 4860
6660 static void work_for_cpu_fn(struct work_struc 4861 static void work_for_cpu_fn(struct work_struct *work)
6661 { 4862 {
6662 struct work_for_cpu *wfc = container_ 4863 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
6663 4864
6664 wfc->ret = wfc->fn(wfc->arg); 4865 wfc->ret = wfc->fn(wfc->arg);
6665 } 4866 }
6666 4867
6667 /** 4868 /**
6668 * work_on_cpu_key - run a function in thread !! 4869 * work_on_cpu - run a function in user context on a particular cpu
6669 * @cpu: the cpu to run on 4870 * @cpu: the cpu to run on
6670 * @fn: the function to run 4871 * @fn: the function to run
6671 * @arg: the function arg 4872 * @arg: the function arg
6672 * @key: The lock class key for lock debuggin <<
6673 * 4873 *
>> 4874 * This will return the value @fn returns.
6674 * It is up to the caller to ensure that the 4875 * It is up to the caller to ensure that the cpu doesn't go offline.
6675 * The caller must not hold any locks which w 4876 * The caller must not hold any locks which would prevent @fn from completing.
6676 * <<
6677 * Return: The value @fn returns. <<
6678 */ 4877 */
6679 long work_on_cpu_key(int cpu, long (*fn)(void !! 4878 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
6680 void *arg, struct lock_c <<
6681 { 4879 {
6682 struct work_for_cpu wfc = { .fn = fn, 4880 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
6683 4881
6684 INIT_WORK_ONSTACK_KEY(&wfc.work, work !! 4882 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
6685 schedule_work_on(cpu, &wfc.work); 4883 schedule_work_on(cpu, &wfc.work);
6686 flush_work(&wfc.work); 4884 flush_work(&wfc.work);
6687 destroy_work_on_stack(&wfc.work); <<
6688 return wfc.ret; 4885 return wfc.ret;
6689 } 4886 }
6690 EXPORT_SYMBOL_GPL(work_on_cpu_key); !! 4887 EXPORT_SYMBOL_GPL(work_on_cpu);
6691 <<
6692 /** <<
6693 * work_on_cpu_safe_key - run a function in t <<
6694 * @cpu: the cpu to run on <<
6695 * @fn: the function to run <<
6696 * @arg: the function argument <<
6697 * @key: The lock class key for lock debuggin <<
6698 * <<
6699 * Disables CPU hotplug and calls work_on_cpu <<
6700 * any locks which would prevent @fn from com <<
6701 * <<
6702 * Return: The value @fn returns. <<
6703 */ <<
6704 long work_on_cpu_safe_key(int cpu, long (*fn) <<
6705 void *arg, struct l <<
6706 { <<
6707 long ret = -ENODEV; <<
6708 <<
6709 cpus_read_lock(); <<
6710 if (cpu_online(cpu)) <<
6711 ret = work_on_cpu_key(cpu, fn <<
6712 cpus_read_unlock(); <<
6713 return ret; <<
6714 } <<
6715 EXPORT_SYMBOL_GPL(work_on_cpu_safe_key); <<
6716 #endif /* CONFIG_SMP */ 4888 #endif /* CONFIG_SMP */
6717 4889
6718 #ifdef CONFIG_FREEZER 4890 #ifdef CONFIG_FREEZER
6719 4891
6720 /** 4892 /**
6721 * freeze_workqueues_begin - begin freezing w 4893 * freeze_workqueues_begin - begin freezing workqueues
6722 * 4894 *
6723 * Start freezing workqueues. After this fun 4895 * Start freezing workqueues. After this function returns, all freezable
6724 * workqueues will queue new works to their i !! 4896 * workqueues will queue new works to their delayed_works list instead of
6725 * pool->worklist. 4897 * pool->worklist.
6726 * 4898 *
6727 * CONTEXT: 4899 * CONTEXT:
6728 * Grabs and releases wq_pool_mutex, wq->mute 4900 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6729 */ 4901 */
6730 void freeze_workqueues_begin(void) 4902 void freeze_workqueues_begin(void)
6731 { 4903 {
>> 4904 struct worker_pool *pool;
6732 struct workqueue_struct *wq; 4905 struct workqueue_struct *wq;
>> 4906 struct pool_workqueue *pwq;
>> 4907 int pi;
6733 4908
6734 mutex_lock(&wq_pool_mutex); 4909 mutex_lock(&wq_pool_mutex);
6735 4910
6736 WARN_ON_ONCE(workqueue_freezing); 4911 WARN_ON_ONCE(workqueue_freezing);
6737 workqueue_freezing = true; 4912 workqueue_freezing = true;
6738 4913
>> 4914 /* set FREEZING */
>> 4915 for_each_pool(pool, pi) {
>> 4916 spin_lock_irq(&pool->lock);
>> 4917 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
>> 4918 pool->flags |= POOL_FREEZING;
>> 4919 spin_unlock_irq(&pool->lock);
>> 4920 }
>> 4921
6739 list_for_each_entry(wq, &workqueues, 4922 list_for_each_entry(wq, &workqueues, list) {
6740 mutex_lock(&wq->mutex); 4923 mutex_lock(&wq->mutex);
6741 wq_adjust_max_active(wq); !! 4924 for_each_pwq(pwq, wq)
>> 4925 pwq_adjust_max_active(pwq);
6742 mutex_unlock(&wq->mutex); 4926 mutex_unlock(&wq->mutex);
6743 } 4927 }
6744 4928
6745 mutex_unlock(&wq_pool_mutex); 4929 mutex_unlock(&wq_pool_mutex);
6746 } 4930 }
6747 4931
6748 /** 4932 /**
6749 * freeze_workqueues_busy - are freezable wor 4933 * freeze_workqueues_busy - are freezable workqueues still busy?
6750 * 4934 *
6751 * Check whether freezing is complete. This 4935 * Check whether freezing is complete. This function must be called
6752 * between freeze_workqueues_begin() and thaw 4936 * between freeze_workqueues_begin() and thaw_workqueues().
6753 * 4937 *
6754 * CONTEXT: 4938 * CONTEXT:
6755 * Grabs and releases wq_pool_mutex. 4939 * Grabs and releases wq_pool_mutex.
6756 * 4940 *
6757 * Return: !! 4941 * RETURNS:
6758 * %true if some freezable workqueues are sti 4942 * %true if some freezable workqueues are still busy. %false if freezing
6759 * is complete. 4943 * is complete.
6760 */ 4944 */
6761 bool freeze_workqueues_busy(void) 4945 bool freeze_workqueues_busy(void)
6762 { 4946 {
6763 bool busy = false; 4947 bool busy = false;
6764 struct workqueue_struct *wq; 4948 struct workqueue_struct *wq;
6765 struct pool_workqueue *pwq; 4949 struct pool_workqueue *pwq;
6766 4950
6767 mutex_lock(&wq_pool_mutex); 4951 mutex_lock(&wq_pool_mutex);
6768 4952
6769 WARN_ON_ONCE(!workqueue_freezing); 4953 WARN_ON_ONCE(!workqueue_freezing);
6770 4954
6771 list_for_each_entry(wq, &workqueues, 4955 list_for_each_entry(wq, &workqueues, list) {
6772 if (!(wq->flags & WQ_FREEZABL 4956 if (!(wq->flags & WQ_FREEZABLE))
6773 continue; 4957 continue;
6774 /* 4958 /*
6775 * nr_active is monotonically 4959 * nr_active is monotonically decreasing. It's safe
6776 * to peek without lock. 4960 * to peek without lock.
6777 */ 4961 */
6778 rcu_read_lock(); !! 4962 rcu_read_lock_sched();
6779 for_each_pwq(pwq, wq) { 4963 for_each_pwq(pwq, wq) {
6780 WARN_ON_ONCE(pwq->nr_ 4964 WARN_ON_ONCE(pwq->nr_active < 0);
6781 if (pwq->nr_active) { 4965 if (pwq->nr_active) {
6782 busy = true; 4966 busy = true;
6783 rcu_read_unlo !! 4967 rcu_read_unlock_sched();
6784 goto out_unlo 4968 goto out_unlock;
6785 } 4969 }
6786 } 4970 }
6787 rcu_read_unlock(); !! 4971 rcu_read_unlock_sched();
6788 } 4972 }
6789 out_unlock: 4973 out_unlock:
6790 mutex_unlock(&wq_pool_mutex); 4974 mutex_unlock(&wq_pool_mutex);
6791 return busy; 4975 return busy;
6792 } 4976 }
6793 4977
6794 /** 4978 /**
6795 * thaw_workqueues - thaw workqueues 4979 * thaw_workqueues - thaw workqueues
6796 * 4980 *
6797 * Thaw workqueues. Normal queueing is resto 4981 * Thaw workqueues. Normal queueing is restored and all collected
6798 * frozen works are transferred to their resp 4982 * frozen works are transferred to their respective pool worklists.
6799 * 4983 *
6800 * CONTEXT: 4984 * CONTEXT:
6801 * Grabs and releases wq_pool_mutex, wq->mute 4985 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6802 */ 4986 */
6803 void thaw_workqueues(void) 4987 void thaw_workqueues(void)
6804 { 4988 {
6805 struct workqueue_struct *wq; 4989 struct workqueue_struct *wq;
>> 4990 struct pool_workqueue *pwq;
>> 4991 struct worker_pool *pool;
>> 4992 int pi;
6806 4993
6807 mutex_lock(&wq_pool_mutex); 4994 mutex_lock(&wq_pool_mutex);
6808 4995
6809 if (!workqueue_freezing) 4996 if (!workqueue_freezing)
6810 goto out_unlock; 4997 goto out_unlock;
6811 4998
6812 workqueue_freezing = false; !! 4999 /* clear FREEZING */
>> 5000 for_each_pool(pool, pi) {
>> 5001 spin_lock_irq(&pool->lock);
>> 5002 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
>> 5003 pool->flags &= ~POOL_FREEZING;
>> 5004 spin_unlock_irq(&pool->lock);
>> 5005 }
6813 5006
6814 /* restore max_active and repopulate 5007 /* restore max_active and repopulate worklist */
6815 list_for_each_entry(wq, &workqueues, 5008 list_for_each_entry(wq, &workqueues, list) {
6816 mutex_lock(&wq->mutex); 5009 mutex_lock(&wq->mutex);
6817 wq_adjust_max_active(wq); !! 5010 for_each_pwq(pwq, wq)
>> 5011 pwq_adjust_max_active(pwq);
6818 mutex_unlock(&wq->mutex); 5012 mutex_unlock(&wq->mutex);
6819 } 5013 }
6820 5014
>> 5015 workqueue_freezing = false;
6821 out_unlock: 5016 out_unlock:
6822 mutex_unlock(&wq_pool_mutex); 5017 mutex_unlock(&wq_pool_mutex);
6823 } 5018 }
6824 #endif /* CONFIG_FREEZER */ 5019 #endif /* CONFIG_FREEZER */
6825 5020
6826 static int workqueue_apply_unbound_cpumask(co !! 5021 static void __init wq_numa_init(void)
6827 { 5022 {
6828 LIST_HEAD(ctxs); !! 5023 cpumask_var_t *tbl;
6829 int ret = 0; !! 5024 int node, cpu;
6830 struct workqueue_struct *wq; <<
6831 struct apply_wqattrs_ctx *ctx, *n; <<
6832 <<
6833 lockdep_assert_held(&wq_pool_mutex); <<
6834 <<
6835 list_for_each_entry(wq, &workqueues, <<
6836 if (!(wq->flags & WQ_UNBOUND) <<
6837 continue; <<
6838 5025
6839 ctx = apply_wqattrs_prepare(w !! 5026 /* determine NUMA pwq table len - highest node id + 1 */
6840 if (IS_ERR(ctx)) { !! 5027 for_each_node(node)
6841 ret = PTR_ERR(ctx); !! 5028 wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1);
6842 break; <<
6843 } <<
6844 5029
6845 list_add_tail(&ctx->list, &ct !! 5030 if (num_possible_nodes() <= 1)
6846 } !! 5031 return;
6847 <<
6848 list_for_each_entry_safe(ctx, n, &ctx <<
6849 if (!ret) <<
6850 apply_wqattrs_commit( <<
6851 apply_wqattrs_cleanup(ctx); <<
6852 } <<
6853 5032
6854 if (!ret) { !! 5033 if (wq_disable_numa) {
6855 mutex_lock(&wq_pool_attach_mu !! 5034 pr_info("workqueue: NUMA affinity support disabled\n");
6856 cpumask_copy(wq_unbound_cpuma !! 5035 return;
6857 mutex_unlock(&wq_pool_attach_ <<
6858 } 5036 }
6859 return ret; <<
6860 } <<
6861 5037
6862 /** !! 5038 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
6863 * workqueue_unbound_exclude_cpumask - Exclud !! 5039 BUG_ON(!wq_update_unbound_numa_attrs_buf);
6864 * @exclude_cpumask: the cpumask to be exclud <<
6865 * <<
6866 * This function can be called from cpuset co <<
6867 * CPUs that should be excluded from wq_unbou <<
6868 */ <<
6869 int workqueue_unbound_exclude_cpumask(cpumask <<
6870 { <<
6871 cpumask_var_t cpumask; <<
6872 int ret = 0; <<
6873 <<
6874 if (!zalloc_cpumask_var(&cpumask, GFP <<
6875 return -ENOMEM; <<
6876 <<
6877 mutex_lock(&wq_pool_mutex); <<
6878 5040
6879 /* 5041 /*
6880 * If the operation fails, it will fa !! 5042 * We want masks of possible CPUs of each node which isn't readily
6881 * wq_requested_unbound_cpumask which !! 5043 * available. Build one from cpu_to_node() which should have been
6882 * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) ho !! 5044 * fully initialized by now.
6883 * by any subsequent write to workque <<
6884 */ <<
6885 if (!cpumask_andnot(cpumask, wq_reque <<
6886 cpumask_copy(cpumask, wq_requ <<
6887 if (!cpumask_equal(cpumask, wq_unboun <<
6888 ret = workqueue_apply_unbound <<
6889 <<
6890 /* Save the current isolated cpumask <<
6891 if (!ret) <<
6892 cpumask_copy(wq_isolated_cpum <<
6893 <<
6894 mutex_unlock(&wq_pool_mutex); <<
6895 free_cpumask_var(cpumask); <<
6896 return ret; <<
6897 } <<
6898 <<
6899 static int parse_affn_scope(const char *val) <<
6900 { <<
6901 int i; <<
6902 <<
6903 for (i = 0; i < ARRAY_SIZE(wq_affn_na <<
6904 if (!strncasecmp(val, wq_affn <<
6905 return i; <<
6906 } <<
6907 return -EINVAL; <<
6908 } <<
6909 <<
6910 static int wq_affn_dfl_set(const char *val, c <<
6911 { <<
6912 struct workqueue_struct *wq; <<
6913 int affn, cpu; <<
6914 <<
6915 affn = parse_affn_scope(val); <<
6916 if (affn < 0) <<
6917 return affn; <<
6918 if (affn == WQ_AFFN_DFL) <<
6919 return -EINVAL; <<
6920 <<
6921 cpus_read_lock(); <<
6922 mutex_lock(&wq_pool_mutex); <<
6923 <<
6924 wq_affn_dfl = affn; <<
6925 <<
6926 list_for_each_entry(wq, &workqueues, <<
6927 for_each_online_cpu(cpu) <<
6928 unbound_wq_update_pwq <<
6929 } <<
6930 <<
6931 mutex_unlock(&wq_pool_mutex); <<
6932 cpus_read_unlock(); <<
6933 <<
6934 return 0; <<
6935 } <<
6936 <<
6937 static int wq_affn_dfl_get(char *buffer, cons <<
6938 { <<
6939 return scnprintf(buffer, PAGE_SIZE, " <<
6940 } <<
6941 <<
6942 static const struct kernel_param_ops wq_affn_ <<
6943 .set = wq_affn_dfl_set, <<
6944 .get = wq_affn_dfl_get, <<
6945 }; <<
6946 <<
6947 module_param_cb(default_affinity_scope, &wq_a <<
6948 <<
6949 #ifdef CONFIG_SYSFS <<
6950 /* <<
6951 * Workqueues with WQ_SYSFS flag set is visib <<
6952 * /sys/bus/workqueue/devices/WQ_NAME. All v <<
6953 * following attributes. <<
6954 * <<
6955 * per_cpu RO bool : whether the <<
6956 * max_active RW int : maximum num <<
6957 * <<
6958 * Unbound workqueues have the following extr <<
6959 * <<
6960 * nice RW int : nice value <<
6961 * cpumask RW mask : bitmask of <<
6962 * affinity_scope RW str : worker CPU <<
6963 * affinity_strict RW bool : worker CPU <<
6964 */ <<
6965 struct wq_device { <<
6966 struct workqueue_struct *wq; <<
6967 struct device dev; <<
6968 }; <<
6969 <<
6970 static struct workqueue_struct *dev_to_wq(str <<
6971 { <<
6972 struct wq_device *wq_dev = container_ <<
6973 <<
6974 return wq_dev->wq; <<
6975 } <<
6976 <<
6977 static ssize_t per_cpu_show(struct device *de <<
6978 char *buf) <<
6979 { <<
6980 struct workqueue_struct *wq = dev_to_ <<
6981 <<
6982 return scnprintf(buf, PAGE_SIZE, "%d\ <<
6983 } <<
6984 static DEVICE_ATTR_RO(per_cpu); <<
6985 <<
6986 static ssize_t max_active_show(struct device <<
6987 struct device_ <<
6988 { <<
6989 struct workqueue_struct *wq = dev_to_ <<
6990 <<
6991 return scnprintf(buf, PAGE_SIZE, "%d\ <<
6992 } <<
6993 <<
6994 static ssize_t max_active_store(struct device <<
6995 struct device <<
6996 size_t count) <<
6997 { <<
6998 struct workqueue_struct *wq = dev_to_ <<
6999 int val; <<
7000 <<
7001 if (sscanf(buf, "%d", &val) != 1 || v <<
7002 return -EINVAL; <<
7003 <<
7004 workqueue_set_max_active(wq, val); <<
7005 return count; <<
7006 } <<
7007 static DEVICE_ATTR_RW(max_active); <<
7008 <<
7009 static struct attribute *wq_sysfs_attrs[] = { <<
7010 &dev_attr_per_cpu.attr, <<
7011 &dev_attr_max_active.attr, <<
7012 NULL, <<
7013 }; <<
7014 ATTRIBUTE_GROUPS(wq_sysfs); <<
7015 <<
7016 static ssize_t wq_nice_show(struct device *de <<
7017 char *buf) <<
7018 { <<
7019 struct workqueue_struct *wq = dev_to_ <<
7020 int written; <<
7021 <<
7022 mutex_lock(&wq->mutex); <<
7023 written = scnprintf(buf, PAGE_SIZE, " <<
7024 mutex_unlock(&wq->mutex); <<
7025 <<
7026 return written; <<
7027 } <<
7028 <<
7029 /* prepare workqueue_attrs for sysfs store op <<
7030 static struct workqueue_attrs *wq_sysfs_prep_ <<
7031 { <<
7032 struct workqueue_attrs *attrs; <<
7033 <<
7034 lockdep_assert_held(&wq_pool_mutex); <<
7035 <<
7036 attrs = alloc_workqueue_attrs(); <<
7037 if (!attrs) <<
7038 return NULL; <<
7039 <<
7040 copy_workqueue_attrs(attrs, wq->unbou <<
7041 return attrs; <<
7042 } <<
7043 <<
7044 static ssize_t wq_nice_store(struct device *d <<
7045 const char *buf, <<
7046 { <<
7047 struct workqueue_struct *wq = dev_to_ <<
7048 struct workqueue_attrs *attrs; <<
7049 int ret = -ENOMEM; <<
7050 <<
7051 apply_wqattrs_lock(); <<
7052 <<
7053 attrs = wq_sysfs_prep_attrs(wq); <<
7054 if (!attrs) <<
7055 goto out_unlock; <<
7056 <<
7057 if (sscanf(buf, "%d", &attrs->nice) = <<
7058 attrs->nice >= MIN_NICE && attrs- <<
7059 ret = apply_workqueue_attrs_l <<
7060 else <<
7061 ret = -EINVAL; <<
7062 <<
7063 out_unlock: <<
7064 apply_wqattrs_unlock(); <<
7065 free_workqueue_attrs(attrs); <<
7066 return ret ?: count; <<
7067 } <<
7068 <<
7069 static ssize_t wq_cpumask_show(struct device <<
7070 struct device_ <<
7071 { <<
7072 struct workqueue_struct *wq = dev_to_ <<
7073 int written; <<
7074 <<
7075 mutex_lock(&wq->mutex); <<
7076 written = scnprintf(buf, PAGE_SIZE, " <<
7077 cpumask_pr_args(w <<
7078 mutex_unlock(&wq->mutex); <<
7079 return written; <<
7080 } <<
7081 <<
7082 static ssize_t wq_cpumask_store(struct device <<
7083 struct device <<
7084 const char *b <<
7085 { <<
7086 struct workqueue_struct *wq = dev_to_ <<
7087 struct workqueue_attrs *attrs; <<
7088 int ret = -ENOMEM; <<
7089 <<
7090 apply_wqattrs_lock(); <<
7091 <<
7092 attrs = wq_sysfs_prep_attrs(wq); <<
7093 if (!attrs) <<
7094 goto out_unlock; <<
7095 <<
7096 ret = cpumask_parse(buf, attrs->cpuma <<
7097 if (!ret) <<
7098 ret = apply_workqueue_attrs_l <<
7099 <<
7100 out_unlock: <<
7101 apply_wqattrs_unlock(); <<
7102 free_workqueue_attrs(attrs); <<
7103 return ret ?: count; <<
7104 } <<
7105 <<
7106 static ssize_t wq_affn_scope_show(struct devi <<
7107 struct devi <<
7108 { <<
7109 struct workqueue_struct *wq = dev_to_ <<
7110 int written; <<
7111 <<
7112 mutex_lock(&wq->mutex); <<
7113 if (wq->unbound_attrs->affn_scope == <<
7114 written = scnprintf(buf, PAGE <<
7115 wq_affn_n <<
7116 wq_affn_n <<
7117 else <<
7118 written = scnprintf(buf, PAGE <<
7119 wq_affn_n <<
7120 mutex_unlock(&wq->mutex); <<
7121 <<
7122 return written; <<
7123 } <<
7124 <<
7125 static ssize_t wq_affn_scope_store(struct dev <<
7126 struct dev <<
7127 const char <<
7128 { <<
7129 struct workqueue_struct *wq = dev_to_ <<
7130 struct workqueue_attrs *attrs; <<
7131 int affn, ret = -ENOMEM; <<
7132 <<
7133 affn = parse_affn_scope(buf); <<
7134 if (affn < 0) <<
7135 return affn; <<
7136 <<
7137 apply_wqattrs_lock(); <<
7138 attrs = wq_sysfs_prep_attrs(wq); <<
7139 if (attrs) { <<
7140 attrs->affn_scope = affn; <<
7141 ret = apply_workqueue_attrs_l <<
7142 } <<
7143 apply_wqattrs_unlock(); <<
7144 free_workqueue_attrs(attrs); <<
7145 return ret ?: count; <<
7146 } <<
7147 <<
7148 static ssize_t wq_affinity_strict_show(struct <<
7149 struct <<
7150 { <<
7151 struct workqueue_struct *wq = dev_to_ <<
7152 <<
7153 return scnprintf(buf, PAGE_SIZE, "%d\ <<
7154 wq->unbound_attrs->a <<
7155 } <<
7156 <<
7157 static ssize_t wq_affinity_strict_store(struc <<
7158 struc <<
7159 const <<
7160 { <<
7161 struct workqueue_struct *wq = dev_to_ <<
7162 struct workqueue_attrs *attrs; <<
7163 int v, ret = -ENOMEM; <<
7164 <<
7165 if (sscanf(buf, "%d", &v) != 1) <<
7166 return -EINVAL; <<
7167 <<
7168 apply_wqattrs_lock(); <<
7169 attrs = wq_sysfs_prep_attrs(wq); <<
7170 if (attrs) { <<
7171 attrs->affn_strict = (bool)v; <<
7172 ret = apply_workqueue_attrs_l <<
7173 } <<
7174 apply_wqattrs_unlock(); <<
7175 free_workqueue_attrs(attrs); <<
7176 return ret ?: count; <<
7177 } <<
7178 <<
7179 static struct device_attribute wq_sysfs_unbou <<
7180 __ATTR(nice, 0644, wq_nice_show, wq_n <<
7181 __ATTR(cpumask, 0644, wq_cpumask_show <<
7182 __ATTR(affinity_scope, 0644, wq_affn_ <<
7183 __ATTR(affinity_strict, 0644, wq_affi <<
7184 __ATTR_NULL, <<
7185 }; <<
7186 <<
7187 static const struct bus_type wq_subsys = { <<
7188 .name = "wo <<
7189 .dev_groups = wq_ <<
7190 }; <<
7191 <<
7192 /** <<
7193 * workqueue_set_unbound_cpumask - Set the l <<
7194 * @cpumask: the cpumask to set <<
7195 * <<
7196 * The low-level workqueues cpumask is a glo <<
7197 * the affinity of all unbound workqueues. <<
7198 * and apply it to all unbound workqueues an <<
7199 * <<
7200 * Return: 0 - Success <<
7201 * -EINVAL - Invalid @cpumask <<
7202 * -ENOMEM - Failed to allocate <<
7203 */ <<
7204 static int workqueue_set_unbound_cpumask(cpum <<
7205 { <<
7206 int ret = -EINVAL; <<
7207 <<
7208 /* <<
7209 * Not excluding isolated cpus on pur <<
7210 * If the user wishes to include them <<
7211 */ <<
7212 cpumask_and(cpumask, cpumask, cpu_pos <<
7213 if (!cpumask_empty(cpumask)) { <<
7214 ret = 0; <<
7215 apply_wqattrs_lock(); <<
7216 if (!cpumask_equal(cpumask, w <<
7217 ret = workqueue_apply <<
7218 if (!ret) <<
7219 cpumask_copy(wq_reque <<
7220 apply_wqattrs_unlock(); <<
7221 } <<
7222 <<
7223 return ret; <<
7224 } <<
7225 <<
7226 static ssize_t __wq_cpumask_show(struct devic <<
7227 struct device_attribute *attr <<
7228 { <<
7229 int written; <<
7230 <<
7231 mutex_lock(&wq_pool_mutex); <<
7232 written = scnprintf(buf, PAGE_SIZE, " <<
7233 mutex_unlock(&wq_pool_mutex); <<
7234 <<
7235 return written; <<
7236 } <<
7237 <<
7238 static ssize_t cpumask_requested_show(struct <<
7239 struct device_attribute *attr <<
7240 { <<
7241 return __wq_cpumask_show(dev, attr, b <<
7242 } <<
7243 static DEVICE_ATTR_RO(cpumask_requested); <<
7244 <<
7245 static ssize_t cpumask_isolated_show(struct d <<
7246 struct device_attribute *attr <<
7247 { <<
7248 return __wq_cpumask_show(dev, attr, b <<
7249 } <<
7250 static DEVICE_ATTR_RO(cpumask_isolated); <<
7251 <<
7252 static ssize_t cpumask_show(struct device *de <<
7253 struct device_attribute *attr <<
7254 { <<
7255 return __wq_cpumask_show(dev, attr, b <<
7256 } <<
7257 <<
7258 static ssize_t cpumask_store(struct device *d <<
7259 struct device_attribute *attr <<
7260 { <<
7261 cpumask_var_t cpumask; <<
7262 int ret; <<
7263 <<
7264 if (!zalloc_cpumask_var(&cpumask, GFP <<
7265 return -ENOMEM; <<
7266 <<
7267 ret = cpumask_parse(buf, cpumask); <<
7268 if (!ret) <<
7269 ret = workqueue_set_unbound_c <<
7270 <<
7271 free_cpumask_var(cpumask); <<
7272 return ret ? ret : count; <<
7273 } <<
7274 static DEVICE_ATTR_RW(cpumask); <<
7275 <<
7276 static struct attribute *wq_sysfs_cpumask_att <<
7277 &dev_attr_cpumask.attr, <<
7278 &dev_attr_cpumask_requested.attr, <<
7279 &dev_attr_cpumask_isolated.attr, <<
7280 NULL, <<
7281 }; <<
7282 ATTRIBUTE_GROUPS(wq_sysfs_cpumask); <<
7283 <<
7284 static int __init wq_sysfs_init(void) <<
7285 { <<
7286 return subsys_virtual_register(&wq_su <<
7287 } <<
7288 core_initcall(wq_sysfs_init); <<
7289 <<
7290 static void wq_device_release(struct device * <<
7291 { <<
7292 struct wq_device *wq_dev = container_ <<
7293 <<
7294 kfree(wq_dev); <<
7295 } <<
7296 <<
7297 /** <<
7298 * workqueue_sysfs_register - make a workqueu <<
7299 * @wq: the workqueue to register <<
7300 * <<
7301 * Expose @wq in sysfs under /sys/bus/workque <<
7302 * alloc_workqueue*() automatically calls thi <<
7303 * which is the preferred method. <<
7304 * <<
7305 * Workqueue user should use this function di <<
7306 * workqueue_attrs before making the workqueu <<
7307 * apply_workqueue_attrs() may race against u <<
7308 * attributes. <<
7309 * <<
7310 * Return: 0 on success, -errno on failure. <<
7311 */ <<
7312 int workqueue_sysfs_register(struct workqueue <<
7313 { <<
7314 struct wq_device *wq_dev; <<
7315 int ret; <<
7316 <<
7317 /* <<
7318 * Adjusting max_active breaks orderi <<
7319 * ordered workqueues. <<
7320 */ <<
7321 if (WARN_ON(wq->flags & __WQ_ORDERED) <<
7322 return -EINVAL; <<
7323 <<
7324 wq->wq_dev = wq_dev = kzalloc(sizeof( <<
7325 if (!wq_dev) <<
7326 return -ENOMEM; <<
7327 <<
7328 wq_dev->wq = wq; <<
7329 wq_dev->dev.bus = &wq_subsys; <<
7330 wq_dev->dev.release = wq_device_relea <<
7331 dev_set_name(&wq_dev->dev, "%s", wq-> <<
7332 <<
7333 /* <<
7334 * unbound_attrs are created separate <<
7335 * everything is ready. <<
7336 */ 5045 */
7337 dev_set_uevent_suppress(&wq_dev->dev, !! 5046 tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL);
7338 !! 5047 BUG_ON(!tbl);
7339 ret = device_register(&wq_dev->dev); <<
7340 if (ret) { <<
7341 put_device(&wq_dev->dev); <<
7342 wq->wq_dev = NULL; <<
7343 return ret; <<
7344 } <<
7345 <<
7346 if (wq->flags & WQ_UNBOUND) { <<
7347 struct device_attribute *attr <<
7348 <<
7349 for (attr = wq_sysfs_unbound_ <<
7350 ret = device_create_f <<
7351 if (ret) { <<
7352 device_unregi <<
7353 wq->wq_dev = <<
7354 return ret; <<
7355 } <<
7356 } <<
7357 } <<
7358 5048
7359 dev_set_uevent_suppress(&wq_dev->dev, !! 5049 for_each_node(node)
7360 kobject_uevent(&wq_dev->dev.kobj, KOB !! 5050 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
7361 return 0; !! 5051 node_online(node) ? node : NUMA_NO_NODE));
7362 } <<
7363 5052
7364 /** !! 5053 for_each_possible_cpu(cpu) {
7365 * workqueue_sysfs_unregister - undo workqueu !! 5054 node = cpu_to_node(cpu);
7366 * @wq: the workqueue to unregister !! 5055 if (WARN_ON(node == NUMA_NO_NODE)) {
7367 * !! 5056 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
7368 * If @wq is registered to sysfs by workqueue !! 5057 /* happens iff arch is bonkers, let's just proceed */
7369 */ !! 5058 return;
7370 static void workqueue_sysfs_unregister(struct <<
7371 { <<
7372 struct wq_device *wq_dev = wq->wq_dev <<
7373 <<
7374 if (!wq->wq_dev) <<
7375 return; <<
7376 <<
7377 wq->wq_dev = NULL; <<
7378 device_unregister(&wq_dev->dev); <<
7379 } <<
7380 #else /* CONFIG_SYSFS */ <<
7381 static void workqueue_sysfs_unregister(struct <<
7382 #endif /* CONFIG_SYSFS */ <<
7383 <<
7384 /* <<
7385 * Workqueue watchdog. <<
7386 * <<
7387 * Stall may be caused by various bugs - miss <<
7388 * flush dependency, a concurrency managed wo <<
7389 * indefinitely. Workqueue stalls can be ver <<
7390 * usual warning mechanisms don't trigger and <<
7391 * largely opaque. <<
7392 * <<
7393 * Workqueue watchdog monitors all worker poo <<
7394 * state if some pools failed to make forward <<
7395 * forward progress is defined as the first i <<
7396 * <<
7397 * This mechanism is controlled through the k <<
7398 * "workqueue.watchdog_thresh" which can be u <<
7399 * corresponding sysfs parameter file. <<
7400 */ <<
7401 #ifdef CONFIG_WQ_WATCHDOG <<
7402 <<
7403 static unsigned long wq_watchdog_thresh = 30; <<
7404 static struct timer_list wq_watchdog_timer; <<
7405 <<
7406 static unsigned long wq_watchdog_touched = IN <<
7407 static DEFINE_PER_CPU(unsigned long, wq_watch <<
7408 <<
7409 /* <<
7410 * Show workers that might prevent the proces <<
7411 * The only candidates are CPU-bound workers <<
7412 * Pending work items should be handled by an <<
7413 * in all other situations. <<
7414 */ <<
7415 static void show_cpu_pool_hog(struct worker_p <<
7416 { <<
7417 struct worker *worker; <<
7418 unsigned long irq_flags; <<
7419 int bkt; <<
7420 <<
7421 raw_spin_lock_irqsave(&pool->lock, ir <<
7422 <<
7423 hash_for_each(pool->busy_hash, bkt, w <<
7424 if (task_is_running(worker->t <<
7425 /* <<
7426 * Defer printing to <<
7427 * drivers that queue <<
7428 * also taken in thei <<
7429 */ <<
7430 printk_deferred_enter <<
7431 <<
7432 pr_info("pool %d:\n", <<
7433 sched_show_task(worke <<
7434 <<
7435 printk_deferred_exit( <<
7436 } <<
7437 } <<
7438 <<
7439 raw_spin_unlock_irqrestore(&pool->loc <<
7440 } <<
7441 <<
7442 static void show_cpu_pools_hogs(void) <<
7443 { <<
7444 struct worker_pool *pool; <<
7445 int pi; <<
7446 <<
7447 pr_info("Showing backtraces of runnin <<
7448 <<
7449 rcu_read_lock(); <<
7450 <<
7451 for_each_pool(pool, pi) { <<
7452 if (pool->cpu_stall) <<
7453 show_cpu_pool_hog(poo <<
7454 <<
7455 } <<
7456 <<
7457 rcu_read_unlock(); <<
7458 } <<
7459 <<
7460 static void wq_watchdog_reset_touched(void) <<
7461 { <<
7462 int cpu; <<
7463 <<
7464 wq_watchdog_touched = jiffies; <<
7465 for_each_possible_cpu(cpu) <<
7466 per_cpu(wq_watchdog_touched_c <<
7467 } <<
7468 <<
7469 static void wq_watchdog_timer_fn(struct timer <<
7470 { <<
7471 unsigned long thresh = READ_ONCE(wq_w <<
7472 bool lockup_detected = false; <<
7473 bool cpu_pool_stall = false; <<
7474 unsigned long now = jiffies; <<
7475 struct worker_pool *pool; <<
7476 int pi; <<
7477 <<
7478 if (!thresh) <<
7479 return; <<
7480 <<
7481 rcu_read_lock(); <<
7482 <<
7483 for_each_pool(pool, pi) { <<
7484 unsigned long pool_ts, touche <<
7485 <<
7486 pool->cpu_stall = false; <<
7487 if (list_empty(&pool->worklis <<
7488 continue; <<
7489 <<
7490 /* <<
7491 * If a virtual machine is st <<
7492 * the watchdog like a stall. <<
7493 */ <<
7494 kvm_check_and_clear_guest_pau <<
7495 <<
7496 /* get the latest of pool and <<
7497 if (pool->cpu >= 0) <<
7498 touched = READ_ONCE(p <<
7499 else <<
7500 touched = READ_ONCE(w <<
7501 pool_ts = READ_ONCE(pool->wat <<
7502 <<
7503 if (time_after(pool_ts, touch <<
7504 ts = pool_ts; <<
7505 else <<
7506 ts = touched; <<
7507 <<
7508 /* did we stall? */ <<
7509 if (time_after(now, ts + thre <<
7510 lockup_detected = tru <<
7511 if (pool->cpu >= 0 && <<
7512 pool->cpu_sta <<
7513 cpu_pool_stal <<
7514 } <<
7515 pr_emerg("BUG: workqu <<
7516 pr_cont_pool_info(poo <<
7517 pr_cont(" stuck for % <<
7518 jiffies_to_ms <<
7519 } 5059 }
7520 !! 5060 cpumask_set_cpu(cpu, tbl[node]);
7521 <<
7522 } 5061 }
7523 5062
7524 rcu_read_unlock(); !! 5063 wq_numa_possible_cpumask = tbl;
7525 !! 5064 wq_numa_enabled = true;
7526 if (lockup_detected) <<
7527 show_all_workqueues(); <<
7528 <<
7529 if (cpu_pool_stall) <<
7530 show_cpu_pools_hogs(); <<
7531 <<
7532 wq_watchdog_reset_touched(); <<
7533 mod_timer(&wq_watchdog_timer, jiffies <<
7534 } 5065 }
7535 5066
7536 notrace void wq_watchdog_touch(int cpu) !! 5067 static int __init init_workqueues(void)
7537 { 5068 {
7538 unsigned long thresh = READ_ONCE(wq_w <<
7539 unsigned long touch_ts = READ_ONCE(wq <<
7540 unsigned long now = jiffies; <<
7541 <<
7542 if (cpu >= 0) <<
7543 per_cpu(wq_watchdog_touched_c <<
7544 else <<
7545 WARN_ONCE(1, "%s should be ca <<
7546 <<
7547 /* Don't unnecessarily store to globa <<
7548 if (time_after(now, touch_ts + thresh <<
7549 WRITE_ONCE(wq_watchdog_touche <<
7550 } <<
7551 <<
7552 static void wq_watchdog_set_thresh(unsigned l <<
7553 { <<
7554 wq_watchdog_thresh = 0; <<
7555 del_timer_sync(&wq_watchdog_timer); <<
7556 <<
7557 if (thresh) { <<
7558 wq_watchdog_thresh = thresh; <<
7559 wq_watchdog_reset_touched(); <<
7560 mod_timer(&wq_watchdog_timer, <<
7561 } <<
7562 } <<
7563 <<
7564 static int wq_watchdog_param_set_thresh(const <<
7565 const <<
7566 { <<
7567 unsigned long thresh; <<
7568 int ret; <<
7569 <<
7570 ret = kstrtoul(val, 0, &thresh); <<
7571 if (ret) <<
7572 return ret; <<
7573 <<
7574 if (system_wq) <<
7575 wq_watchdog_set_thresh(thresh <<
7576 else <<
7577 wq_watchdog_thresh = thresh; <<
7578 <<
7579 return 0; <<
7580 } <<
7581 <<
7582 static const struct kernel_param_ops wq_watch <<
7583 .set = wq_watchdog_param_set_thres <<
7584 .get = param_get_ulong, <<
7585 }; <<
7586 <<
7587 module_param_cb(watchdog_thresh, &wq_watchdog <<
7588 0644); <<
7589 <<
7590 static void wq_watchdog_init(void) <<
7591 { <<
7592 timer_setup(&wq_watchdog_timer, wq_wa <<
7593 wq_watchdog_set_thresh(wq_watchdog_th <<
7594 } <<
7595 <<
7596 #else /* CONFIG_WQ_WATCHDOG */ <<
7597 <<
7598 static inline void wq_watchdog_init(void) { } <<
7599 <<
7600 #endif /* CONFIG_WQ_WATCHDOG */ <<
7601 <<
7602 static void bh_pool_kick_normal(struct irq_wo <<
7603 { <<
7604 raise_softirq_irqoff(TASKLET_SOFTIRQ) <<
7605 } <<
7606 <<
7607 static void bh_pool_kick_highpri(struct irq_w <<
7608 { <<
7609 raise_softirq_irqoff(HI_SOFTIRQ); <<
7610 } <<
7611 <<
7612 static void __init restrict_unbound_cpumask(c <<
7613 { <<
7614 if (!cpumask_intersects(wq_unbound_cp <<
7615 pr_warn("workqueue: Restricti <<
7616 cpumask_pr_args(wq_un <<
7617 return; <<
7618 } <<
7619 <<
7620 cpumask_and(wq_unbound_cpumask, wq_un <<
7621 } <<
7622 <<
7623 static void __init init_cpu_worker_pool(struc <<
7624 { <<
7625 BUG_ON(init_worker_pool(pool)); <<
7626 pool->cpu = cpu; <<
7627 cpumask_copy(pool->attrs->cpumask, cp <<
7628 cpumask_copy(pool->attrs->__pod_cpuma <<
7629 pool->attrs->nice = nice; <<
7630 pool->attrs->affn_strict = true; <<
7631 pool->node = cpu_to_node(cpu); <<
7632 <<
7633 /* alloc pool ID */ <<
7634 mutex_lock(&wq_pool_mutex); <<
7635 BUG_ON(worker_pool_assign_id(pool)); <<
7636 mutex_unlock(&wq_pool_mutex); <<
7637 } <<
7638 <<
7639 /** <<
7640 * workqueue_init_early - early init for work <<
7641 * <<
7642 * This is the first step of three-staged wor <<
7643 * invoked as soon as the bare basics - memor <<
7644 * up. It sets up all the data structures and <<
7645 * boot code to create workqueues and queue/c <<
7646 * execution starts only after kthreads can b <<
7647 * before early initcalls. <<
7648 */ <<
7649 void __init workqueue_init_early(void) <<
7650 { <<
7651 struct wq_pod_type *pt = &wq_pod_type <<
7652 int std_nice[NR_STD_WORKER_POOLS] = { 5069 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
7653 void (*irq_work_fns[2])(struct irq_wo <<
7654 <<
7655 int i, cpu; 5070 int i, cpu;
7656 5071
7657 BUILD_BUG_ON(__alignof__(struct pool_ !! 5072 /* make sure we have enough bits for OFFQ pool ID */
>> 5073 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
>> 5074 WORK_CPU_END * NR_STD_WORKER_POOLS);
7658 5075
7659 BUG_ON(!alloc_cpumask_var(&wq_online_ !! 5076 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
7660 BUG_ON(!alloc_cpumask_var(&wq_unbound <<
7661 BUG_ON(!alloc_cpumask_var(&wq_request <<
7662 BUG_ON(!zalloc_cpumask_var(&wq_isolat <<
7663 <<
7664 cpumask_copy(wq_online_cpumask, cpu_o <<
7665 cpumask_copy(wq_unbound_cpumask, cpu_ <<
7666 restrict_unbound_cpumask("HK_TYPE_WQ" <<
7667 restrict_unbound_cpumask("HK_TYPE_DOM <<
7668 if (!cpumask_empty(&wq_cmdline_cpumas <<
7669 restrict_unbound_cpumask("wor <<
7670 <<
7671 cpumask_copy(wq_requested_unbound_cpu <<
7672 5077
7673 pwq_cache = KMEM_CACHE(pool_workqueue 5078 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
7674 5079
7675 unbound_wq_update_pwq_attrs_buf = all !! 5080 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
7676 BUG_ON(!unbound_wq_update_pwq_attrs_b !! 5081 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
7677 5082
7678 /* !! 5083 wq_numa_init();
7679 * If nohz_full is enabled, set power <<
7680 * This allows workqueue items to be <<
7681 */ <<
7682 if (housekeeping_enabled(HK_TYPE_TICK <<
7683 wq_power_efficient = true; <<
7684 5084
7685 /* initialize WQ_AFFN_SYSTEM pods */ !! 5085 /* initialize CPU pools */
7686 pt->pod_cpus = kcalloc(1, sizeof(pt-> <<
7687 pt->pod_node = kcalloc(1, sizeof(pt-> <<
7688 pt->cpu_pod = kcalloc(nr_cpu_ids, siz <<
7689 BUG_ON(!pt->pod_cpus || !pt->pod_node <<
7690 <<
7691 BUG_ON(!zalloc_cpumask_var_node(&pt-> <<
7692 <<
7693 pt->nr_pods = 1; <<
7694 cpumask_copy(pt->pod_cpus[0], cpu_pos <<
7695 pt->pod_node[0] = NUMA_NO_NODE; <<
7696 pt->cpu_pod[0] = 0; <<
7697 <<
7698 /* initialize BH and CPU pools */ <<
7699 for_each_possible_cpu(cpu) { 5086 for_each_possible_cpu(cpu) {
7700 struct worker_pool *pool; 5087 struct worker_pool *pool;
7701 5088
7702 i = 0; 5089 i = 0;
7703 for_each_bh_worker_pool(pool, !! 5090 for_each_cpu_worker_pool(pool, cpu) {
7704 init_cpu_worker_pool( !! 5091 BUG_ON(init_worker_pool(pool));
7705 pool->flags |= POOL_B !! 5092 pool->cpu = cpu;
7706 init_irq_work(bh_pool !! 5093 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
7707 i++; !! 5094 pool->attrs->nice = std_nice[i++];
>> 5095 pool->node = cpu_to_node(cpu);
>> 5096
>> 5097 /* alloc pool ID */
>> 5098 mutex_lock(&wq_pool_mutex);
>> 5099 BUG_ON(worker_pool_assign_id(pool));
>> 5100 mutex_unlock(&wq_pool_mutex);
7708 } 5101 }
>> 5102 }
7709 5103
7710 i = 0; !! 5104 /* create the initial worker */
7711 for_each_cpu_worker_pool(pool !! 5105 for_each_online_cpu(cpu) {
7712 init_cpu_worker_pool( !! 5106 struct worker_pool *pool;
>> 5107
>> 5108 for_each_cpu_worker_pool(pool, cpu) {
>> 5109 pool->flags &= ~POOL_DISASSOCIATED;
>> 5110 BUG_ON(create_and_start_worker(pool) < 0);
>> 5111 }
7713 } 5112 }
7714 5113
7715 /* create default unbound and ordered 5114 /* create default unbound and ordered wq attrs */
7716 for (i = 0; i < NR_STD_WORKER_POOLS; 5115 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
7717 struct workqueue_attrs *attrs 5116 struct workqueue_attrs *attrs;
7718 5117
7719 BUG_ON(!(attrs = alloc_workqu !! 5118 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
7720 attrs->nice = std_nice[i]; 5119 attrs->nice = std_nice[i];
7721 unbound_std_wq_attrs[i] = att 5120 unbound_std_wq_attrs[i] = attrs;
7722 5121
7723 /* 5122 /*
7724 * An ordered wq should have 5123 * An ordered wq should have only one pwq as ordering is
7725 * guaranteed by max_active w 5124 * guaranteed by max_active which is enforced by pwqs.
>> 5125 * Turn off NUMA so that dfl_pwq is used for all nodes.
7726 */ 5126 */
7727 BUG_ON(!(attrs = alloc_workqu !! 5127 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
7728 attrs->nice = std_nice[i]; 5128 attrs->nice = std_nice[i];
7729 attrs->ordered = true; !! 5129 attrs->no_numa = true;
7730 ordered_wq_attrs[i] = attrs; 5130 ordered_wq_attrs[i] = attrs;
7731 } 5131 }
7732 5132
7733 system_wq = alloc_workqueue("events", 5133 system_wq = alloc_workqueue("events", 0, 0);
7734 system_highpri_wq = alloc_workqueue(" 5134 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
7735 system_long_wq = alloc_workqueue("eve 5135 system_long_wq = alloc_workqueue("events_long", 0, 0);
7736 system_unbound_wq = alloc_workqueue(" 5136 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
7737 W !! 5137 WQ_UNBOUND_MAX_ACTIVE);
7738 system_freezable_wq = alloc_workqueue 5138 system_freezable_wq = alloc_workqueue("events_freezable",
7739 5139 WQ_FREEZABLE, 0);
7740 system_power_efficient_wq = alloc_wor <<
7741 <<
7742 system_freezable_power_efficient_wq = <<
7743 <<
7744 <<
7745 system_bh_wq = alloc_workqueue("event <<
7746 system_bh_highpri_wq = alloc_workqueu <<
7747 <<
7748 BUG_ON(!system_wq || !system_highpri_ 5140 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
7749 !system_unbound_wq || !system_ !! 5141 !system_unbound_wq || !system_freezable_wq);
7750 !system_power_efficient_wq || !! 5142 return 0;
7751 !system_freezable_power_effici <<
7752 !system_bh_wq || !system_bh_hi <<
7753 } <<
7754 <<
7755 static void __init wq_cpu_intensive_thresh_in <<
7756 { <<
7757 unsigned long thresh; <<
7758 unsigned long bogo; <<
7759 <<
7760 pwq_release_worker = kthread_create_w <<
7761 BUG_ON(IS_ERR(pwq_release_worker)); <<
7762 <<
7763 /* if the user set it to a specific v <<
7764 if (wq_cpu_intensive_thresh_us != ULO <<
7765 return; <<
7766 <<
7767 /* <<
7768 * The default of 10ms is derived fro <<
7769 * 2023) processors can do a lot in 1 <<
7770 * most consider human-perceivable. H <<
7771 * lot slower CPUs including microcon <<
7772 * too low. <<
7773 * <<
7774 * Let's scale up the threshold upto <<
7775 * This is by no means accurate but i <<
7776 * is still useful even when the thre <<
7777 * the reports would usually be appli <<
7778 * operating on longer thresholds won <<
7779 * usefulness. <<
7780 */ <<
7781 thresh = 10 * USEC_PER_MSEC; <<
7782 <<
7783 /* see init/calibrate.c for lpj -> Bo <<
7784 bogo = max_t(unsigned long, loops_per <<
7785 if (bogo < 4000) <<
7786 thresh = min_t(unsigned long, <<
7787 <<
7788 pr_debug("wq_cpu_intensive_thresh: lp <<
7789 loops_per_jiffy, bogo, thres <<
7790 <<
7791 wq_cpu_intensive_thresh_us = thresh; <<
7792 } <<
7793 <<
7794 /** <<
7795 * workqueue_init - bring workqueue subsystem <<
7796 * <<
7797 * This is the second step of three-staged wo <<
7798 * and invoked as soon as kthreads can be cre <<
7799 * been created and work items queued on them <<
7800 * executing the work items yet. Populate the <<
7801 * workers and enable future kworker creation <<
7802 */ <<
7803 void __init workqueue_init(void) <<
7804 { <<
7805 struct workqueue_struct *wq; <<
7806 struct worker_pool *pool; <<
7807 int cpu, bkt; <<
7808 <<
7809 wq_cpu_intensive_thresh_init(); <<
7810 <<
7811 mutex_lock(&wq_pool_mutex); <<
7812 <<
7813 /* <<
7814 * Per-cpu pools created earlier coul <<
7815 * up. Also, create a rescuer for wor <<
7816 */ <<
7817 for_each_possible_cpu(cpu) { <<
7818 for_each_bh_worker_pool(pool, <<
7819 pool->node = cpu_to_n <<
7820 for_each_cpu_worker_pool(pool <<
7821 pool->node = cpu_to_n <<
7822 } <<
7823 <<
7824 list_for_each_entry(wq, &workqueues, <<
7825 WARN(init_rescuer(wq), <<
7826 "workqueue: failed to cr <<
7827 wq->name); <<
7828 } <<
7829 <<
7830 mutex_unlock(&wq_pool_mutex); <<
7831 <<
7832 /* <<
7833 * Create the initial workers. A BH p <<
7834 * represents the shared BH execution <<
7835 * affected by hotplug events. Create <<
7836 * possible CPUs here. <<
7837 */ <<
7838 for_each_possible_cpu(cpu) <<
7839 for_each_bh_worker_pool(pool, <<
7840 BUG_ON(!create_worker <<
7841 <<
7842 for_each_online_cpu(cpu) { <<
7843 for_each_cpu_worker_pool(pool <<
7844 pool->flags &= ~POOL_ <<
7845 BUG_ON(!create_worker <<
7846 } <<
7847 } <<
7848 <<
7849 hash_for_each(unbound_pool_hash, bkt, <<
7850 BUG_ON(!create_worker(pool)); <<
7851 <<
7852 wq_online = true; <<
7853 wq_watchdog_init(); <<
7854 } <<
7855 <<
7856 /* <<
7857 * Initialize @pt by first initializing @pt-> <<
7858 * @cpu_shares_pod(). Each subset of CPUs tha <<
7859 * and consecutive pod ID. The rest of @pt is <<
7860 */ <<
7861 static void __init init_pod_type(struct wq_po <<
7862 bool (*cpus_ <<
7863 { <<
7864 int cur, pre, cpu, pod; <<
7865 <<
7866 pt->nr_pods = 0; <<
7867 <<
7868 /* init @pt->cpu_pod[] according to @ <<
7869 pt->cpu_pod = kcalloc(nr_cpu_ids, siz <<
7870 BUG_ON(!pt->cpu_pod); <<
7871 <<
7872 for_each_possible_cpu(cur) { <<
7873 for_each_possible_cpu(pre) { <<
7874 if (pre >= cur) { <<
7875 pt->cpu_pod[c <<
7876 break; <<
7877 } <<
7878 if (cpus_share_pod(cu <<
7879 pt->cpu_pod[c <<
7880 break; <<
7881 } <<
7882 } <<
7883 } <<
7884 <<
7885 /* init the rest to match @pt->cpu_po <<
7886 pt->pod_cpus = kcalloc(pt->nr_pods, s <<
7887 pt->pod_node = kcalloc(pt->nr_pods, s <<
7888 BUG_ON(!pt->pod_cpus || !pt->pod_node <<
7889 <<
7890 for (pod = 0; pod < pt->nr_pods; pod+ <<
7891 BUG_ON(!zalloc_cpumask_var(&p <<
7892 <<
7893 for_each_possible_cpu(cpu) { <<
7894 cpumask_set_cpu(cpu, pt->pod_ <<
7895 pt->pod_node[pt->cpu_pod[cpu] <<
7896 } <<
7897 } <<
7898 <<
7899 static bool __init cpus_dont_share(int cpu0, <<
7900 { <<
7901 return false; <<
7902 } <<
7903 <<
7904 static bool __init cpus_share_smt(int cpu0, i <<
7905 { <<
7906 #ifdef CONFIG_SCHED_SMT <<
7907 return cpumask_test_cpu(cpu0, cpu_smt <<
7908 #else <<
7909 return false; <<
7910 #endif <<
7911 } <<
7912 <<
7913 static bool __init cpus_share_numa(int cpu0, <<
7914 { <<
7915 return cpu_to_node(cpu0) == cpu_to_no <<
7916 } <<
7917 <<
7918 /** <<
7919 * workqueue_init_topology - initialize CPU p <<
7920 * <<
7921 * This is the third step of three-staged wor <<
7922 * invoked after SMP and topology information <<
7923 * initializes the unbound CPU pods according <<
7924 */ <<
7925 void __init workqueue_init_topology(void) <<
7926 { <<
7927 struct workqueue_struct *wq; <<
7928 int cpu; <<
7929 <<
7930 init_pod_type(&wq_pod_types[WQ_AFFN_C <<
7931 init_pod_type(&wq_pod_types[WQ_AFFN_S <<
7932 init_pod_type(&wq_pod_types[WQ_AFFN_C <<
7933 init_pod_type(&wq_pod_types[WQ_AFFN_N <<
7934 <<
7935 wq_topo_initialized = true; <<
7936 <<
7937 mutex_lock(&wq_pool_mutex); <<
7938 <<
7939 /* <<
7940 * Workqueues allocated earlier would <<
7941 * worker pool. Explicitly call unbou <<
7942 * and CPU combinations to apply per- <<
7943 */ <<
7944 list_for_each_entry(wq, &workqueues, <<
7945 for_each_online_cpu(cpu) <<
7946 unbound_wq_update_pwq <<
7947 if (wq->flags & WQ_UNBOUND) { <<
7948 mutex_lock(&wq->mutex <<
7949 wq_update_node_max_ac <<
7950 mutex_unlock(&wq->mut <<
7951 } <<
7952 } <<
7953 <<
7954 mutex_unlock(&wq_pool_mutex); <<
7955 } <<
7956 <<
7957 void __warn_flushing_systemwide_wq(void) <<
7958 { <<
7959 pr_warn("WARNING: Flushing system-wid <<
7960 dump_stack(); <<
7961 } <<
7962 EXPORT_SYMBOL(__warn_flushing_systemwide_wq); <<
7963 <<
7964 static int __init workqueue_unbound_cpus_setu <<
7965 { <<
7966 if (cpulist_parse(str, &wq_cmdline_cp <<
7967 cpumask_clear(&wq_cmdline_cpu <<
7968 pr_warn("workqueue.unbound_cp <<
7969 } <<
7970 <<
7971 return 1; <<
7972 } 5143 }
7973 __setup("workqueue.unbound_cpus=", workqueue_ !! 5144 early_initcall(init_workqueues);
7974 5145
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.