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