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