1 // SPDX-License-Identifier: GPL-2.0-or-later <<
2 /* 1 /*
3 * kernel/stop_machine.c 2 * kernel/stop_machine.c
4 * 3 *
5 * Copyright (C) 2008, 2005 IBM Corporatio 4 * Copyright (C) 2008, 2005 IBM Corporation.
6 * Copyright (C) 2008, 2005 Rusty Russell 5 * Copyright (C) 2008, 2005 Rusty Russell rusty@rustcorp.com.au
7 * Copyright (C) 2010 SUSE Linux Pro 6 * Copyright (C) 2010 SUSE Linux Products GmbH
8 * Copyright (C) 2010 Tejun Heo <tj@ 7 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
>> 8 *
>> 9 * This file is released under the GPLv2 and any later version.
9 */ 10 */
10 #include <linux/compiler.h> <<
11 #include <linux/completion.h> 11 #include <linux/completion.h>
12 #include <linux/cpu.h> 12 #include <linux/cpu.h>
13 #include <linux/init.h> 13 #include <linux/init.h>
14 #include <linux/kthread.h> 14 #include <linux/kthread.h>
15 #include <linux/export.h> 15 #include <linux/export.h>
16 #include <linux/percpu.h> 16 #include <linux/percpu.h>
17 #include <linux/sched.h> 17 #include <linux/sched.h>
18 #include <linux/stop_machine.h> 18 #include <linux/stop_machine.h>
19 #include <linux/interrupt.h> 19 #include <linux/interrupt.h>
20 #include <linux/kallsyms.h> 20 #include <linux/kallsyms.h>
21 #include <linux/smpboot.h> 21 #include <linux/smpboot.h>
22 #include <linux/atomic.h> 22 #include <linux/atomic.h>
23 #include <linux/nmi.h> 23 #include <linux/nmi.h>
24 #include <linux/sched/wake_q.h> 24 #include <linux/sched/wake_q.h>
25 25
26 /* 26 /*
27 * Structure to determine completion condition 27 * Structure to determine completion condition and record errors. May
28 * be shared by works on different cpus. 28 * be shared by works on different cpus.
29 */ 29 */
30 struct cpu_stop_done { 30 struct cpu_stop_done {
31 atomic_t nr_todo; 31 atomic_t nr_todo; /* nr left to execute */
32 int ret; 32 int ret; /* collected return value */
33 struct completion completion; 33 struct completion completion; /* fired if nr_todo reaches 0 */
34 }; 34 };
35 35
36 /* the actual stopper, one per every possible 36 /* the actual stopper, one per every possible cpu, enabled on online cpus */
37 struct cpu_stopper { 37 struct cpu_stopper {
38 struct task_struct *thread; 38 struct task_struct *thread;
39 39
40 raw_spinlock_t lock; 40 raw_spinlock_t lock;
41 bool enabled; 41 bool enabled; /* is this stopper enabled? */
42 struct list_head works; 42 struct list_head works; /* list of pending works */
43 43
44 struct cpu_stop_work stop_work; 44 struct cpu_stop_work stop_work; /* for stop_cpus */
45 unsigned long caller; <<
46 cpu_stop_fn_t fn; <<
47 }; 45 };
48 46
49 static DEFINE_PER_CPU(struct cpu_stopper, cpu_ 47 static DEFINE_PER_CPU(struct cpu_stopper, cpu_stopper);
50 static bool stop_machine_initialized = false; 48 static bool stop_machine_initialized = false;
51 49
52 void print_stop_info(const char *log_lvl, stru <<
53 { <<
54 /* <<
55 * If @task is a stopper task, it cann <<
56 * stable. <<
57 */ <<
58 struct cpu_stopper *stopper = per_cpu_ <<
59 <<
60 if (task != stopper->thread) <<
61 return; <<
62 <<
63 printk("%sStopper: %pS <- %pS\n", log_ <<
64 } <<
65 <<
66 /* static data for stop_cpus */ 50 /* static data for stop_cpus */
67 static DEFINE_MUTEX(stop_cpus_mutex); 51 static DEFINE_MUTEX(stop_cpus_mutex);
68 static bool stop_cpus_in_progress; 52 static bool stop_cpus_in_progress;
69 53
70 static void cpu_stop_init_done(struct cpu_stop 54 static void cpu_stop_init_done(struct cpu_stop_done *done, unsigned int nr_todo)
71 { 55 {
72 memset(done, 0, sizeof(*done)); 56 memset(done, 0, sizeof(*done));
73 atomic_set(&done->nr_todo, nr_todo); 57 atomic_set(&done->nr_todo, nr_todo);
74 init_completion(&done->completion); 58 init_completion(&done->completion);
75 } 59 }
76 60
77 /* signal completion unless @done is NULL */ 61 /* signal completion unless @done is NULL */
78 static void cpu_stop_signal_done(struct cpu_st 62 static void cpu_stop_signal_done(struct cpu_stop_done *done)
79 { 63 {
80 if (atomic_dec_and_test(&done->nr_todo 64 if (atomic_dec_and_test(&done->nr_todo))
81 complete(&done->completion); 65 complete(&done->completion);
82 } 66 }
83 67
84 static void __cpu_stop_queue_work(struct cpu_s 68 static void __cpu_stop_queue_work(struct cpu_stopper *stopper,
85 struct 69 struct cpu_stop_work *work,
86 struct 70 struct wake_q_head *wakeq)
87 { 71 {
88 list_add_tail(&work->list, &stopper->w 72 list_add_tail(&work->list, &stopper->works);
89 wake_q_add(wakeq, stopper->thread); 73 wake_q_add(wakeq, stopper->thread);
90 } 74 }
91 75
92 /* queue @work to @stopper. if offline, @work 76 /* queue @work to @stopper. if offline, @work is completed immediately */
93 static bool cpu_stop_queue_work(unsigned int c 77 static bool cpu_stop_queue_work(unsigned int cpu, struct cpu_stop_work *work)
94 { 78 {
95 struct cpu_stopper *stopper = &per_cpu 79 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
96 DEFINE_WAKE_Q(wakeq); 80 DEFINE_WAKE_Q(wakeq);
97 unsigned long flags; 81 unsigned long flags;
98 bool enabled; 82 bool enabled;
99 83
100 preempt_disable(); <<
101 raw_spin_lock_irqsave(&stopper->lock, 84 raw_spin_lock_irqsave(&stopper->lock, flags);
102 enabled = stopper->enabled; 85 enabled = stopper->enabled;
103 if (enabled) 86 if (enabled)
104 __cpu_stop_queue_work(stopper, 87 __cpu_stop_queue_work(stopper, work, &wakeq);
105 else if (work->done) 88 else if (work->done)
106 cpu_stop_signal_done(work->don 89 cpu_stop_signal_done(work->done);
107 raw_spin_unlock_irqrestore(&stopper->l 90 raw_spin_unlock_irqrestore(&stopper->lock, flags);
108 91
109 wake_up_q(&wakeq); 92 wake_up_q(&wakeq);
110 preempt_enable(); <<
111 93
112 return enabled; 94 return enabled;
113 } 95 }
114 96
115 /** 97 /**
116 * stop_one_cpu - stop a cpu 98 * stop_one_cpu - stop a cpu
117 * @cpu: cpu to stop 99 * @cpu: cpu to stop
118 * @fn: function to execute 100 * @fn: function to execute
119 * @arg: argument to @fn 101 * @arg: argument to @fn
120 * 102 *
121 * Execute @fn(@arg) on @cpu. @fn is run in a 103 * Execute @fn(@arg) on @cpu. @fn is run in a process context with
122 * the highest priority preempting any task on 104 * the highest priority preempting any task on the cpu and
123 * monopolizing it. This function returns aft 105 * monopolizing it. This function returns after the execution is
124 * complete. 106 * complete.
125 * 107 *
126 * This function doesn't guarantee @cpu stays 108 * This function doesn't guarantee @cpu stays online till @fn
127 * completes. If @cpu goes down in the middle 109 * completes. If @cpu goes down in the middle, execution may happen
128 * partially or fully on different cpus. @fn 110 * partially or fully on different cpus. @fn should either be ready
129 * for that or the caller should ensure that @ 111 * for that or the caller should ensure that @cpu stays online until
130 * this function completes. 112 * this function completes.
131 * 113 *
132 * CONTEXT: 114 * CONTEXT:
133 * Might sleep. 115 * Might sleep.
134 * 116 *
135 * RETURNS: 117 * RETURNS:
136 * -ENOENT if @fn(@arg) was not executed becau 118 * -ENOENT if @fn(@arg) was not executed because @cpu was offline;
137 * otherwise, the return value of @fn. 119 * otherwise, the return value of @fn.
138 */ 120 */
139 int stop_one_cpu(unsigned int cpu, cpu_stop_fn 121 int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
140 { 122 {
141 struct cpu_stop_done done; 123 struct cpu_stop_done done;
142 struct cpu_stop_work work = { .fn = fn !! 124 struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done };
143 125
144 cpu_stop_init_done(&done, 1); 126 cpu_stop_init_done(&done, 1);
145 if (!cpu_stop_queue_work(cpu, &work)) 127 if (!cpu_stop_queue_work(cpu, &work))
146 return -ENOENT; 128 return -ENOENT;
147 /* 129 /*
148 * In case @cpu == smp_proccessor_id() 130 * In case @cpu == smp_proccessor_id() we can avoid a sleep+wakeup
149 * cycle by doing a preemption: 131 * cycle by doing a preemption:
150 */ 132 */
151 cond_resched(); 133 cond_resched();
152 wait_for_completion(&done.completion); 134 wait_for_completion(&done.completion);
153 return done.ret; 135 return done.ret;
154 } 136 }
155 137
156 /* This controls the threads on each CPU. */ 138 /* This controls the threads on each CPU. */
157 enum multi_stop_state { 139 enum multi_stop_state {
158 /* Dummy starting state for thread. */ 140 /* Dummy starting state for thread. */
159 MULTI_STOP_NONE, 141 MULTI_STOP_NONE,
160 /* Awaiting everyone to be scheduled. 142 /* Awaiting everyone to be scheduled. */
161 MULTI_STOP_PREPARE, 143 MULTI_STOP_PREPARE,
162 /* Disable interrupts. */ 144 /* Disable interrupts. */
163 MULTI_STOP_DISABLE_IRQ, 145 MULTI_STOP_DISABLE_IRQ,
164 /* Run the function */ 146 /* Run the function */
165 MULTI_STOP_RUN, 147 MULTI_STOP_RUN,
166 /* Exit */ 148 /* Exit */
167 MULTI_STOP_EXIT, 149 MULTI_STOP_EXIT,
168 }; 150 };
169 151
170 struct multi_stop_data { 152 struct multi_stop_data {
171 cpu_stop_fn_t fn; 153 cpu_stop_fn_t fn;
172 void *data; 154 void *data;
173 /* Like num_online_cpus(), but hotplug 155 /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */
174 unsigned int num_threads; 156 unsigned int num_threads;
175 const struct cpumask *active_cpus; 157 const struct cpumask *active_cpus;
176 158
177 enum multi_stop_state state; 159 enum multi_stop_state state;
178 atomic_t thread_ack; 160 atomic_t thread_ack;
179 }; 161 };
180 162
181 static void set_state(struct multi_stop_data * 163 static void set_state(struct multi_stop_data *msdata,
182 enum multi_stop_state ne 164 enum multi_stop_state newstate)
183 { 165 {
184 /* Reset ack counter. */ 166 /* Reset ack counter. */
185 atomic_set(&msdata->thread_ack, msdata 167 atomic_set(&msdata->thread_ack, msdata->num_threads);
186 smp_wmb(); 168 smp_wmb();
187 WRITE_ONCE(msdata->state, newstate); !! 169 msdata->state = newstate;
188 } 170 }
189 171
190 /* Last one to ack a state moves to the next s 172 /* Last one to ack a state moves to the next state. */
191 static void ack_state(struct multi_stop_data * 173 static void ack_state(struct multi_stop_data *msdata)
192 { 174 {
193 if (atomic_dec_and_test(&msdata->threa 175 if (atomic_dec_and_test(&msdata->thread_ack))
194 set_state(msdata, msdata->stat 176 set_state(msdata, msdata->state + 1);
195 } 177 }
196 178
197 notrace void __weak stop_machine_yield(const s <<
198 { <<
199 cpu_relax(); <<
200 } <<
201 <<
202 /* This is the cpu_stop function which stops t 179 /* This is the cpu_stop function which stops the CPU. */
203 static int multi_cpu_stop(void *data) 180 static int multi_cpu_stop(void *data)
204 { 181 {
205 struct multi_stop_data *msdata = data; 182 struct multi_stop_data *msdata = data;
206 enum multi_stop_state newstate, cursta !! 183 enum multi_stop_state curstate = MULTI_STOP_NONE;
207 int cpu = smp_processor_id(), err = 0; 184 int cpu = smp_processor_id(), err = 0;
208 const struct cpumask *cpumask; <<
209 unsigned long flags; 185 unsigned long flags;
210 bool is_active; 186 bool is_active;
211 187
212 /* 188 /*
213 * When called from stop_machine_from_ 189 * When called from stop_machine_from_inactive_cpu(), irq might
214 * already be disabled. Save the stat 190 * already be disabled. Save the state and restore it on exit.
215 */ 191 */
216 local_save_flags(flags); 192 local_save_flags(flags);
217 193
218 if (!msdata->active_cpus) { !! 194 if (!msdata->active_cpus)
219 cpumask = cpu_online_mask; !! 195 is_active = cpu == cpumask_first(cpu_online_mask);
220 is_active = cpu == cpumask_fir !! 196 else
221 } else { !! 197 is_active = cpumask_test_cpu(cpu, msdata->active_cpus);
222 cpumask = msdata->active_cpus; <<
223 is_active = cpumask_test_cpu(c <<
224 } <<
225 198
226 /* Simple state machine */ 199 /* Simple state machine */
227 do { 200 do {
228 /* Chill out and ensure we re- 201 /* Chill out and ensure we re-read multi_stop_state. */
229 stop_machine_yield(cpumask); !! 202 cpu_relax_yield();
230 newstate = READ_ONCE(msdata->s !! 203 if (msdata->state != curstate) {
231 if (newstate != curstate) { !! 204 curstate = msdata->state;
232 curstate = newstate; <<
233 switch (curstate) { 205 switch (curstate) {
234 case MULTI_STOP_DISABL 206 case MULTI_STOP_DISABLE_IRQ:
235 local_irq_disa 207 local_irq_disable();
236 hard_irq_disab 208 hard_irq_disable();
237 break; 209 break;
238 case MULTI_STOP_RUN: 210 case MULTI_STOP_RUN:
239 if (is_active) 211 if (is_active)
240 err = 212 err = msdata->fn(msdata->data);
241 break; 213 break;
242 default: 214 default:
243 break; 215 break;
244 } 216 }
245 ack_state(msdata); 217 ack_state(msdata);
246 } else if (curstate > MULTI_ST 218 } else if (curstate > MULTI_STOP_PREPARE) {
247 /* 219 /*
248 * At this stage all o 220 * At this stage all other CPUs we depend on must spin
249 * in the same loop. A 221 * in the same loop. Any reason for hard-lockup should
250 * be detected and rep 222 * be detected and reported on their side.
251 */ 223 */
252 touch_nmi_watchdog(); 224 touch_nmi_watchdog();
253 } 225 }
254 rcu_momentary_dyntick_idle(); <<
255 } while (curstate != MULTI_STOP_EXIT); 226 } while (curstate != MULTI_STOP_EXIT);
256 227
257 local_irq_restore(flags); 228 local_irq_restore(flags);
258 return err; 229 return err;
259 } 230 }
260 231
261 static int cpu_stop_queue_two_works(int cpu1, 232 static int cpu_stop_queue_two_works(int cpu1, struct cpu_stop_work *work1,
262 int cpu2, 233 int cpu2, struct cpu_stop_work *work2)
263 { 234 {
264 struct cpu_stopper *stopper1 = per_cpu 235 struct cpu_stopper *stopper1 = per_cpu_ptr(&cpu_stopper, cpu1);
265 struct cpu_stopper *stopper2 = per_cpu 236 struct cpu_stopper *stopper2 = per_cpu_ptr(&cpu_stopper, cpu2);
266 DEFINE_WAKE_Q(wakeq); 237 DEFINE_WAKE_Q(wakeq);
267 int err; 238 int err;
268 <<
269 retry: 239 retry:
270 /* <<
271 * The waking up of stopper threads ha <<
272 * scheduling context as the queueing. <<
273 * possibility of one of the above sto <<
274 * CPU, and preempting us. This will c <<
275 * stopper forever. <<
276 */ <<
277 preempt_disable(); <<
278 raw_spin_lock_irq(&stopper1->lock); 240 raw_spin_lock_irq(&stopper1->lock);
279 raw_spin_lock_nested(&stopper2->lock, 241 raw_spin_lock_nested(&stopper2->lock, SINGLE_DEPTH_NESTING);
280 242
281 if (!stopper1->enabled || !stopper2->e !! 243 err = -ENOENT;
282 err = -ENOENT; !! 244 if (!stopper1->enabled || !stopper2->enabled)
283 goto unlock; 245 goto unlock;
284 } <<
285 <<
286 /* 246 /*
287 * Ensure that if we race with __stop_ 247 * Ensure that if we race with __stop_cpus() the stoppers won't get
288 * queued up in reverse order leading 248 * queued up in reverse order leading to system deadlock.
289 * 249 *
290 * We can't miss stop_cpus_in_progress 250 * We can't miss stop_cpus_in_progress if queue_stop_cpus_work() has
291 * queued a work on cpu1 but not on cp 251 * queued a work on cpu1 but not on cpu2, we hold both locks.
292 * 252 *
293 * It can be falsely true but it is sa 253 * It can be falsely true but it is safe to spin until it is cleared,
294 * queue_stop_cpus_work() does everyth 254 * queue_stop_cpus_work() does everything under preempt_disable().
295 */ 255 */
296 if (unlikely(stop_cpus_in_progress)) { !! 256 err = -EDEADLK;
297 err = -EDEADLK; !! 257 if (unlikely(stop_cpus_in_progress))
298 goto unlock; !! 258 goto unlock;
299 } <<
300 259
301 err = 0; 260 err = 0;
302 __cpu_stop_queue_work(stopper1, work1, 261 __cpu_stop_queue_work(stopper1, work1, &wakeq);
303 __cpu_stop_queue_work(stopper2, work2, 262 __cpu_stop_queue_work(stopper2, work2, &wakeq);
304 !! 263 /*
>> 264 * The waking up of stopper threads has to happen
>> 265 * in the same scheduling context as the queueing.
>> 266 * Otherwise, there is a possibility of one of the
>> 267 * above stoppers being woken up by another CPU,
>> 268 * and preempting us. This will cause us to n ot
>> 269 * wake up the other stopper forever.
>> 270 */
>> 271 preempt_disable();
305 unlock: 272 unlock:
306 raw_spin_unlock(&stopper2->lock); 273 raw_spin_unlock(&stopper2->lock);
307 raw_spin_unlock_irq(&stopper1->lock); 274 raw_spin_unlock_irq(&stopper1->lock);
308 275
309 if (unlikely(err == -EDEADLK)) { 276 if (unlikely(err == -EDEADLK)) {
310 preempt_enable(); <<
311 <<
312 while (stop_cpus_in_progress) 277 while (stop_cpus_in_progress)
313 cpu_relax(); 278 cpu_relax();
314 <<
315 goto retry; 279 goto retry;
316 } 280 }
317 281
318 wake_up_q(&wakeq); !! 282 if (!err) {
319 preempt_enable(); !! 283 wake_up_q(&wakeq);
>> 284 preempt_enable();
>> 285 }
320 286
321 return err; 287 return err;
322 } 288 }
323 /** 289 /**
324 * stop_two_cpus - stops two cpus 290 * stop_two_cpus - stops two cpus
325 * @cpu1: the cpu to stop 291 * @cpu1: the cpu to stop
326 * @cpu2: the other cpu to stop 292 * @cpu2: the other cpu to stop
327 * @fn: function to execute 293 * @fn: function to execute
328 * @arg: argument to @fn 294 * @arg: argument to @fn
329 * 295 *
330 * Stops both the current and specified CPU an 296 * Stops both the current and specified CPU and runs @fn on one of them.
331 * 297 *
332 * returns when both are completed. 298 * returns when both are completed.
333 */ 299 */
334 int stop_two_cpus(unsigned int cpu1, unsigned 300 int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg)
335 { 301 {
336 struct cpu_stop_done done; 302 struct cpu_stop_done done;
337 struct cpu_stop_work work1, work2; 303 struct cpu_stop_work work1, work2;
338 struct multi_stop_data msdata; 304 struct multi_stop_data msdata;
339 305
340 msdata = (struct multi_stop_data){ 306 msdata = (struct multi_stop_data){
341 .fn = fn, 307 .fn = fn,
342 .data = arg, 308 .data = arg,
343 .num_threads = 2, 309 .num_threads = 2,
344 .active_cpus = cpumask_of(cpu1 310 .active_cpus = cpumask_of(cpu1),
345 }; 311 };
346 312
347 work1 = work2 = (struct cpu_stop_work) 313 work1 = work2 = (struct cpu_stop_work){
348 .fn = multi_cpu_stop, 314 .fn = multi_cpu_stop,
349 .arg = &msdata, 315 .arg = &msdata,
350 .done = &done, !! 316 .done = &done
351 .caller = _RET_IP_, <<
352 }; 317 };
353 318
354 cpu_stop_init_done(&done, 2); 319 cpu_stop_init_done(&done, 2);
355 set_state(&msdata, MULTI_STOP_PREPARE) 320 set_state(&msdata, MULTI_STOP_PREPARE);
356 321
357 if (cpu1 > cpu2) 322 if (cpu1 > cpu2)
358 swap(cpu1, cpu2); 323 swap(cpu1, cpu2);
359 if (cpu_stop_queue_two_works(cpu1, &wo 324 if (cpu_stop_queue_two_works(cpu1, &work1, cpu2, &work2))
360 return -ENOENT; 325 return -ENOENT;
361 326
362 wait_for_completion(&done.completion); 327 wait_for_completion(&done.completion);
363 return done.ret; 328 return done.ret;
364 } 329 }
365 330
366 /** 331 /**
367 * stop_one_cpu_nowait - stop a cpu but don't 332 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
368 * @cpu: cpu to stop 333 * @cpu: cpu to stop
369 * @fn: function to execute 334 * @fn: function to execute
370 * @arg: argument to @fn 335 * @arg: argument to @fn
371 * @work_buf: pointer to cpu_stop_work structu 336 * @work_buf: pointer to cpu_stop_work structure
372 * 337 *
373 * Similar to stop_one_cpu() but doesn't wait 338 * Similar to stop_one_cpu() but doesn't wait for completion. The
374 * caller is responsible for ensuring @work_bu 339 * caller is responsible for ensuring @work_buf is currently unused
375 * and will remain untouched until stopper sta 340 * and will remain untouched until stopper starts executing @fn.
376 * 341 *
377 * CONTEXT: 342 * CONTEXT:
378 * Don't care. 343 * Don't care.
379 * 344 *
380 * RETURNS: 345 * RETURNS:
381 * true if cpu_stop_work was queued successful 346 * true if cpu_stop_work was queued successfully and @fn will be called,
382 * false otherwise. 347 * false otherwise.
383 */ 348 */
384 bool stop_one_cpu_nowait(unsigned int cpu, cpu 349 bool stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
385 struct cpu_stop_work * 350 struct cpu_stop_work *work_buf)
386 { 351 {
387 *work_buf = (struct cpu_stop_work){ .f !! 352 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
388 return cpu_stop_queue_work(cpu, work_b 353 return cpu_stop_queue_work(cpu, work_buf);
389 } 354 }
390 355
391 static bool queue_stop_cpus_work(const struct 356 static bool queue_stop_cpus_work(const struct cpumask *cpumask,
392 cpu_stop_fn_t 357 cpu_stop_fn_t fn, void *arg,
393 struct cpu_st 358 struct cpu_stop_done *done)
394 { 359 {
395 struct cpu_stop_work *work; 360 struct cpu_stop_work *work;
396 unsigned int cpu; 361 unsigned int cpu;
397 bool queued = false; 362 bool queued = false;
398 363
399 /* 364 /*
400 * Disable preemption while queueing t 365 * Disable preemption while queueing to avoid getting
401 * preempted by a stopper which might 366 * preempted by a stopper which might wait for other stoppers
402 * to enter @fn which can lead to dead 367 * to enter @fn which can lead to deadlock.
403 */ 368 */
404 preempt_disable(); 369 preempt_disable();
405 stop_cpus_in_progress = true; 370 stop_cpus_in_progress = true;
406 barrier(); <<
407 for_each_cpu(cpu, cpumask) { 371 for_each_cpu(cpu, cpumask) {
408 work = &per_cpu(cpu_stopper.st 372 work = &per_cpu(cpu_stopper.stop_work, cpu);
409 work->fn = fn; 373 work->fn = fn;
410 work->arg = arg; 374 work->arg = arg;
411 work->done = done; 375 work->done = done;
412 work->caller = _RET_IP_; <<
413 if (cpu_stop_queue_work(cpu, w 376 if (cpu_stop_queue_work(cpu, work))
414 queued = true; 377 queued = true;
415 } 378 }
416 barrier(); <<
417 stop_cpus_in_progress = false; 379 stop_cpus_in_progress = false;
418 preempt_enable(); 380 preempt_enable();
419 381
420 return queued; 382 return queued;
421 } 383 }
422 384
423 static int __stop_cpus(const struct cpumask *c 385 static int __stop_cpus(const struct cpumask *cpumask,
424 cpu_stop_fn_t fn, void 386 cpu_stop_fn_t fn, void *arg)
425 { 387 {
426 struct cpu_stop_done done; 388 struct cpu_stop_done done;
427 389
428 cpu_stop_init_done(&done, cpumask_weig 390 cpu_stop_init_done(&done, cpumask_weight(cpumask));
429 if (!queue_stop_cpus_work(cpumask, fn, 391 if (!queue_stop_cpus_work(cpumask, fn, arg, &done))
430 return -ENOENT; 392 return -ENOENT;
431 wait_for_completion(&done.completion); 393 wait_for_completion(&done.completion);
432 return done.ret; 394 return done.ret;
433 } 395 }
434 396
435 /** 397 /**
436 * stop_cpus - stop multiple cpus 398 * stop_cpus - stop multiple cpus
437 * @cpumask: cpus to stop 399 * @cpumask: cpus to stop
438 * @fn: function to execute 400 * @fn: function to execute
439 * @arg: argument to @fn 401 * @arg: argument to @fn
440 * 402 *
441 * Execute @fn(@arg) on online cpus in @cpumas 403 * Execute @fn(@arg) on online cpus in @cpumask. On each target cpu,
442 * @fn is run in a process context with the hi 404 * @fn is run in a process context with the highest priority
443 * preempting any task on the cpu and monopoli 405 * preempting any task on the cpu and monopolizing it. This function
444 * returns after all executions are complete. 406 * returns after all executions are complete.
445 * 407 *
446 * This function doesn't guarantee the cpus in 408 * This function doesn't guarantee the cpus in @cpumask stay online
447 * till @fn completes. If some cpus go down i 409 * till @fn completes. If some cpus go down in the middle, execution
448 * on the cpu may happen partially or fully on 410 * on the cpu may happen partially or fully on different cpus. @fn
449 * should either be ready for that or the call 411 * should either be ready for that or the caller should ensure that
450 * the cpus stay online until this function co 412 * the cpus stay online until this function completes.
451 * 413 *
452 * All stop_cpus() calls are serialized making 414 * All stop_cpus() calls are serialized making it safe for @fn to wait
453 * for all cpus to start executing it. 415 * for all cpus to start executing it.
454 * 416 *
455 * CONTEXT: 417 * CONTEXT:
456 * Might sleep. 418 * Might sleep.
457 * 419 *
458 * RETURNS: 420 * RETURNS:
459 * -ENOENT if @fn(@arg) was not executed at al 421 * -ENOENT if @fn(@arg) was not executed at all because all cpus in
460 * @cpumask were offline; otherwise, 0 if all 422 * @cpumask were offline; otherwise, 0 if all executions of @fn
461 * returned 0, any non zero return value if an 423 * returned 0, any non zero return value if any returned non zero.
462 */ 424 */
463 static int stop_cpus(const struct cpumask *cpu !! 425 int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
464 { 426 {
465 int ret; 427 int ret;
466 428
467 /* static works are used, process one 429 /* static works are used, process one request at a time */
468 mutex_lock(&stop_cpus_mutex); 430 mutex_lock(&stop_cpus_mutex);
469 ret = __stop_cpus(cpumask, fn, arg); 431 ret = __stop_cpus(cpumask, fn, arg);
470 mutex_unlock(&stop_cpus_mutex); 432 mutex_unlock(&stop_cpus_mutex);
471 return ret; 433 return ret;
472 } 434 }
473 435
>> 436 /**
>> 437 * try_stop_cpus - try to stop multiple cpus
>> 438 * @cpumask: cpus to stop
>> 439 * @fn: function to execute
>> 440 * @arg: argument to @fn
>> 441 *
>> 442 * Identical to stop_cpus() except that it fails with -EAGAIN if
>> 443 * someone else is already using the facility.
>> 444 *
>> 445 * CONTEXT:
>> 446 * Might sleep.
>> 447 *
>> 448 * RETURNS:
>> 449 * -EAGAIN if someone else is already stopping cpus, -ENOENT if
>> 450 * @fn(@arg) was not executed at all because all cpus in @cpumask were
>> 451 * offline; otherwise, 0 if all executions of @fn returned 0, any non
>> 452 * zero return value if any returned non zero.
>> 453 */
>> 454 int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
>> 455 {
>> 456 int ret;
>> 457
>> 458 /* static works are used, process one request at a time */
>> 459 if (!mutex_trylock(&stop_cpus_mutex))
>> 460 return -EAGAIN;
>> 461 ret = __stop_cpus(cpumask, fn, arg);
>> 462 mutex_unlock(&stop_cpus_mutex);
>> 463 return ret;
>> 464 }
>> 465
474 static int cpu_stop_should_run(unsigned int cp 466 static int cpu_stop_should_run(unsigned int cpu)
475 { 467 {
476 struct cpu_stopper *stopper = &per_cpu 468 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
477 unsigned long flags; 469 unsigned long flags;
478 int run; 470 int run;
479 471
480 raw_spin_lock_irqsave(&stopper->lock, 472 raw_spin_lock_irqsave(&stopper->lock, flags);
481 run = !list_empty(&stopper->works); 473 run = !list_empty(&stopper->works);
482 raw_spin_unlock_irqrestore(&stopper->l 474 raw_spin_unlock_irqrestore(&stopper->lock, flags);
483 return run; 475 return run;
484 } 476 }
485 477
486 static void cpu_stopper_thread(unsigned int cp 478 static void cpu_stopper_thread(unsigned int cpu)
487 { 479 {
488 struct cpu_stopper *stopper = &per_cpu 480 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
489 struct cpu_stop_work *work; 481 struct cpu_stop_work *work;
490 482
491 repeat: 483 repeat:
492 work = NULL; 484 work = NULL;
493 raw_spin_lock_irq(&stopper->lock); 485 raw_spin_lock_irq(&stopper->lock);
494 if (!list_empty(&stopper->works)) { 486 if (!list_empty(&stopper->works)) {
495 work = list_first_entry(&stopp 487 work = list_first_entry(&stopper->works,
496 struct 488 struct cpu_stop_work, list);
497 list_del_init(&work->list); 489 list_del_init(&work->list);
498 } 490 }
499 raw_spin_unlock_irq(&stopper->lock); 491 raw_spin_unlock_irq(&stopper->lock);
500 492
501 if (work) { 493 if (work) {
502 cpu_stop_fn_t fn = work->fn; 494 cpu_stop_fn_t fn = work->fn;
503 void *arg = work->arg; 495 void *arg = work->arg;
504 struct cpu_stop_done *done = w 496 struct cpu_stop_done *done = work->done;
505 int ret; 497 int ret;
506 498
507 /* cpu stop callbacks must not 499 /* cpu stop callbacks must not sleep, make in_atomic() == T */
508 stopper->caller = work->caller <<
509 stopper->fn = fn; <<
510 preempt_count_inc(); 500 preempt_count_inc();
511 ret = fn(arg); 501 ret = fn(arg);
512 if (done) { 502 if (done) {
513 if (ret) 503 if (ret)
514 done->ret = re 504 done->ret = ret;
515 cpu_stop_signal_done(d 505 cpu_stop_signal_done(done);
516 } 506 }
517 preempt_count_dec(); 507 preempt_count_dec();
518 stopper->fn = NULL; <<
519 stopper->caller = 0; <<
520 WARN_ONCE(preempt_count(), 508 WARN_ONCE(preempt_count(),
521 "cpu_stop: %ps(%p) l !! 509 "cpu_stop: %pf(%p) leaked preempt count\n", fn, arg);
522 goto repeat; 510 goto repeat;
523 } 511 }
524 } 512 }
525 513
526 void stop_machine_park(int cpu) 514 void stop_machine_park(int cpu)
527 { 515 {
528 struct cpu_stopper *stopper = &per_cpu 516 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
529 /* 517 /*
530 * Lockless. cpu_stopper_thread() will 518 * Lockless. cpu_stopper_thread() will take stopper->lock and flush
531 * the pending works before it parks, 519 * the pending works before it parks, until then it is fine to queue
532 * the new works. 520 * the new works.
533 */ 521 */
534 stopper->enabled = false; 522 stopper->enabled = false;
535 kthread_park(stopper->thread); 523 kthread_park(stopper->thread);
536 } 524 }
537 525
>> 526 extern void sched_set_stop_task(int cpu, struct task_struct *stop);
>> 527
538 static void cpu_stop_create(unsigned int cpu) 528 static void cpu_stop_create(unsigned int cpu)
539 { 529 {
540 sched_set_stop_task(cpu, per_cpu(cpu_s 530 sched_set_stop_task(cpu, per_cpu(cpu_stopper.thread, cpu));
541 } 531 }
542 532
543 static void cpu_stop_park(unsigned int cpu) 533 static void cpu_stop_park(unsigned int cpu)
544 { 534 {
545 struct cpu_stopper *stopper = &per_cpu 535 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
546 536
547 WARN_ON(!list_empty(&stopper->works)); 537 WARN_ON(!list_empty(&stopper->works));
548 } 538 }
549 539
550 void stop_machine_unpark(int cpu) 540 void stop_machine_unpark(int cpu)
551 { 541 {
552 struct cpu_stopper *stopper = &per_cpu 542 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
553 543
554 stopper->enabled = true; 544 stopper->enabled = true;
555 kthread_unpark(stopper->thread); 545 kthread_unpark(stopper->thread);
556 } 546 }
557 547
558 static struct smp_hotplug_thread cpu_stop_thre 548 static struct smp_hotplug_thread cpu_stop_threads = {
559 .store = &cpu_stopper 549 .store = &cpu_stopper.thread,
560 .thread_should_run = cpu_stop_sho 550 .thread_should_run = cpu_stop_should_run,
561 .thread_fn = cpu_stopper_ 551 .thread_fn = cpu_stopper_thread,
562 .thread_comm = "migration/% 552 .thread_comm = "migration/%u",
563 .create = cpu_stop_cre 553 .create = cpu_stop_create,
564 .park = cpu_stop_par 554 .park = cpu_stop_park,
565 .selfparking = true, 555 .selfparking = true,
566 }; 556 };
567 557
568 static int __init cpu_stop_init(void) 558 static int __init cpu_stop_init(void)
569 { 559 {
570 unsigned int cpu; 560 unsigned int cpu;
571 561
572 for_each_possible_cpu(cpu) { 562 for_each_possible_cpu(cpu) {
573 struct cpu_stopper *stopper = 563 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
574 564
575 raw_spin_lock_init(&stopper->l 565 raw_spin_lock_init(&stopper->lock);
576 INIT_LIST_HEAD(&stopper->works 566 INIT_LIST_HEAD(&stopper->works);
577 } 567 }
578 568
579 BUG_ON(smpboot_register_percpu_thread( 569 BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads));
580 stop_machine_unpark(raw_smp_processor_ 570 stop_machine_unpark(raw_smp_processor_id());
581 stop_machine_initialized = true; 571 stop_machine_initialized = true;
582 return 0; 572 return 0;
583 } 573 }
584 early_initcall(cpu_stop_init); 574 early_initcall(cpu_stop_init);
585 575
586 int stop_machine_cpuslocked(cpu_stop_fn_t fn, 576 int stop_machine_cpuslocked(cpu_stop_fn_t fn, void *data,
587 const struct cpuma 577 const struct cpumask *cpus)
588 { 578 {
589 struct multi_stop_data msdata = { 579 struct multi_stop_data msdata = {
590 .fn = fn, 580 .fn = fn,
591 .data = data, 581 .data = data,
592 .num_threads = num_online_cpus 582 .num_threads = num_online_cpus(),
593 .active_cpus = cpus, 583 .active_cpus = cpus,
594 }; 584 };
595 585
596 lockdep_assert_cpus_held(); 586 lockdep_assert_cpus_held();
597 587
598 if (!stop_machine_initialized) { 588 if (!stop_machine_initialized) {
599 /* 589 /*
600 * Handle the case where stop_ 590 * Handle the case where stop_machine() is called
601 * early in boot before stop_m 591 * early in boot before stop_machine() has been
602 * initialized. 592 * initialized.
603 */ 593 */
604 unsigned long flags; 594 unsigned long flags;
605 int ret; 595 int ret;
606 596
607 WARN_ON_ONCE(msdata.num_thread 597 WARN_ON_ONCE(msdata.num_threads != 1);
608 598
609 local_irq_save(flags); 599 local_irq_save(flags);
610 hard_irq_disable(); 600 hard_irq_disable();
611 ret = (*fn)(data); 601 ret = (*fn)(data);
612 local_irq_restore(flags); 602 local_irq_restore(flags);
613 603
614 return ret; 604 return ret;
615 } 605 }
616 606
617 /* Set the initial state and stop all 607 /* Set the initial state and stop all online cpus. */
618 set_state(&msdata, MULTI_STOP_PREPARE) 608 set_state(&msdata, MULTI_STOP_PREPARE);
619 return stop_cpus(cpu_online_mask, mult 609 return stop_cpus(cpu_online_mask, multi_cpu_stop, &msdata);
620 } 610 }
621 611
622 int stop_machine(cpu_stop_fn_t fn, void *data, 612 int stop_machine(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus)
623 { 613 {
624 int ret; 614 int ret;
625 615
626 /* No CPUs can come up or down during 616 /* No CPUs can come up or down during this. */
627 cpus_read_lock(); 617 cpus_read_lock();
628 ret = stop_machine_cpuslocked(fn, data 618 ret = stop_machine_cpuslocked(fn, data, cpus);
629 cpus_read_unlock(); 619 cpus_read_unlock();
630 return ret; 620 return ret;
631 } 621 }
632 EXPORT_SYMBOL_GPL(stop_machine); 622 EXPORT_SYMBOL_GPL(stop_machine);
633 <<
634 #ifdef CONFIG_SCHED_SMT <<
635 int stop_core_cpuslocked(unsigned int cpu, cpu <<
636 { <<
637 const struct cpumask *smt_mask = cpu_s <<
638 <<
639 struct multi_stop_data msdata = { <<
640 .fn = fn, <<
641 .data = data, <<
642 .num_threads = cpumask_weight( <<
643 .active_cpus = smt_mask, <<
644 }; <<
645 <<
646 lockdep_assert_cpus_held(); <<
647 <<
648 /* Set the initial state and stop all <<
649 set_state(&msdata, MULTI_STOP_PREPARE) <<
650 return stop_cpus(smt_mask, multi_cpu_s <<
651 } <<
652 EXPORT_SYMBOL_GPL(stop_core_cpuslocked); <<
653 #endif <<
654 623
655 /** 624 /**
656 * stop_machine_from_inactive_cpu - stop_machi 625 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
657 * @fn: the function to run 626 * @fn: the function to run
658 * @data: the data ptr for the @fn() 627 * @data: the data ptr for the @fn()
659 * @cpus: the cpus to run the @fn() on (NULL = 628 * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
660 * 629 *
661 * This is identical to stop_machine() but can 630 * This is identical to stop_machine() but can be called from a CPU which
662 * is not active. The local CPU is in the pro 631 * is not active. The local CPU is in the process of hotplug (so no other
663 * CPU hotplug can start) and not marked activ 632 * CPU hotplug can start) and not marked active and doesn't have enough
664 * context to sleep. 633 * context to sleep.
665 * 634 *
666 * This function provides stop_machine() funct 635 * This function provides stop_machine() functionality for such state by
667 * using busy-wait for synchronization and exe 636 * using busy-wait for synchronization and executing @fn directly for local
668 * CPU. 637 * CPU.
669 * 638 *
670 * CONTEXT: 639 * CONTEXT:
671 * Local CPU is inactive. Temporarily stops a 640 * Local CPU is inactive. Temporarily stops all active CPUs.
672 * 641 *
673 * RETURNS: 642 * RETURNS:
674 * 0 if all executions of @fn returned 0, any 643 * 0 if all executions of @fn returned 0, any non zero return value if any
675 * returned non zero. 644 * returned non zero.
676 */ 645 */
677 int stop_machine_from_inactive_cpu(cpu_stop_fn 646 int stop_machine_from_inactive_cpu(cpu_stop_fn_t fn, void *data,
678 const struct 647 const struct cpumask *cpus)
679 { 648 {
680 struct multi_stop_data msdata = { .fn 649 struct multi_stop_data msdata = { .fn = fn, .data = data,
681 .a 650 .active_cpus = cpus };
682 struct cpu_stop_done done; 651 struct cpu_stop_done done;
683 int ret; 652 int ret;
684 653
685 /* Local CPU must be inactive and CPU 654 /* Local CPU must be inactive and CPU hotplug in progress. */
686 BUG_ON(cpu_active(raw_smp_processor_id 655 BUG_ON(cpu_active(raw_smp_processor_id()));
687 msdata.num_threads = num_active_cpus() 656 msdata.num_threads = num_active_cpus() + 1; /* +1 for local */
688 657
689 /* No proper task established and can' 658 /* No proper task established and can't sleep - busy wait for lock. */
690 while (!mutex_trylock(&stop_cpus_mutex 659 while (!mutex_trylock(&stop_cpus_mutex))
691 cpu_relax(); 660 cpu_relax();
692 661
693 /* Schedule work on other CPUs and exe 662 /* Schedule work on other CPUs and execute directly for local CPU */
694 set_state(&msdata, MULTI_STOP_PREPARE) 663 set_state(&msdata, MULTI_STOP_PREPARE);
695 cpu_stop_init_done(&done, num_active_c 664 cpu_stop_init_done(&done, num_active_cpus());
696 queue_stop_cpus_work(cpu_active_mask, 665 queue_stop_cpus_work(cpu_active_mask, multi_cpu_stop, &msdata,
697 &done); 666 &done);
698 ret = multi_cpu_stop(&msdata); 667 ret = multi_cpu_stop(&msdata);
699 668
700 /* Busy wait for completion. */ 669 /* Busy wait for completion. */
701 while (!completion_done(&done.completi 670 while (!completion_done(&done.completion))
702 cpu_relax(); 671 cpu_relax();
703 672
704 mutex_unlock(&stop_cpus_mutex); 673 mutex_unlock(&stop_cpus_mutex);
705 return ret ?: done.ret; 674 return ret ?: done.ret;
706 } 675 }
707 676
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.