xref: /linux/kernel/sched/membarrier.c (revision bf76f23aa1c178e9115eba17f699fa726aed669b)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Copyright (C) 2010-2017 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
4  *
5  * membarrier system call
6  */
7 #include <uapi/linux/membarrier.h>
8 #include "sched.h"
9 
10 /*
11  * For documentation purposes, here are some membarrier ordering
12  * scenarios to keep in mind:
13  *
14  * A) Userspace thread execution after IPI vs membarrier's memory
15  *    barrier before sending the IPI
16  *
17  * Userspace variables:
18  *
19  * int x = 0, y = 0;
20  *
21  * The memory barrier at the start of membarrier() on CPU0 is necessary in
22  * order to enforce the guarantee that any writes occurring on CPU0 before
23  * the membarrier() is executed will be visible to any code executing on
24  * CPU1 after the IPI-induced memory barrier:
25  *
26  *         CPU0                              CPU1
27  *
28  *         x = 1
29  *         membarrier():
30  *           a: smp_mb()
31  *           b: send IPI                       IPI-induced mb
32  *           c: smp_mb()
33  *         r2 = y
34  *                                           y = 1
35  *                                           barrier()
36  *                                           r1 = x
37  *
38  *                     BUG_ON(r1 == 0 && r2 == 0)
39  *
40  * The write to y and load from x by CPU1 are unordered by the hardware,
41  * so it's possible to have "r1 = x" reordered before "y = 1" at any
42  * point after (b).  If the memory barrier at (a) is omitted, then "x = 1"
43  * can be reordered after (a) (although not after (c)), so we get r1 == 0
44  * and r2 == 0.  This violates the guarantee that membarrier() is
45  * supposed by provide.
46  *
47  * The timing of the memory barrier at (a) has to ensure that it executes
48  * before the IPI-induced memory barrier on CPU1.
49  *
50  * B) Userspace thread execution before IPI vs membarrier's memory
51  *    barrier after completing the IPI
52  *
53  * Userspace variables:
54  *
55  * int x = 0, y = 0;
56  *
57  * The memory barrier at the end of membarrier() on CPU0 is necessary in
58  * order to enforce the guarantee that any writes occurring on CPU1 before
59  * the membarrier() is executed will be visible to any code executing on
60  * CPU0 after the membarrier():
61  *
62  *         CPU0                              CPU1
63  *
64  *                                           x = 1
65  *                                           barrier()
66  *                                           y = 1
67  *         r2 = y
68  *         membarrier():
69  *           a: smp_mb()
70  *           b: send IPI                       IPI-induced mb
71  *           c: smp_mb()
72  *         r1 = x
73  *         BUG_ON(r1 == 0 && r2 == 1)
74  *
75  * The writes to x and y are unordered by the hardware, so it's possible to
76  * have "r2 = 1" even though the write to x doesn't execute until (b).  If
77  * the memory barrier at (c) is omitted then "r1 = x" can be reordered
78  * before (b) (although not before (a)), so we get "r1 = 0".  This violates
79  * the guarantee that membarrier() is supposed to provide.
80  *
81  * The timing of the memory barrier at (c) has to ensure that it executes
82  * after the IPI-induced memory barrier on CPU1.
83  *
84  * C) Scheduling userspace thread -> kthread -> userspace thread vs membarrier
85  *
86  *           CPU0                            CPU1
87  *
88  *           membarrier():
89  *           a: smp_mb()
90  *                                           d: switch to kthread (includes mb)
91  *           b: read rq->curr->mm == NULL
92  *                                           e: switch to user (includes mb)
93  *           c: smp_mb()
94  *
95  * Using the scenario from (A), we can show that (a) needs to be paired
96  * with (e). Using the scenario from (B), we can show that (c) needs to
97  * be paired with (d).
98  *
99  * D) exit_mm vs membarrier
100  *
101  * Two thread groups are created, A and B.  Thread group B is created by
102  * issuing clone from group A with flag CLONE_VM set, but not CLONE_THREAD.
103  * Let's assume we have a single thread within each thread group (Thread A
104  * and Thread B).  Thread A runs on CPU0, Thread B runs on CPU1.
105  *
106  *           CPU0                            CPU1
107  *
108  *           membarrier():
109  *             a: smp_mb()
110  *                                           exit_mm():
111  *                                             d: smp_mb()
112  *                                             e: current->mm = NULL
113  *             b: read rq->curr->mm == NULL
114  *             c: smp_mb()
115  *
116  * Using scenario (B), we can show that (c) needs to be paired with (d).
117  *
118  * E) kthread_{use,unuse}_mm vs membarrier
119  *
120  *           CPU0                            CPU1
121  *
122  *           membarrier():
123  *           a: smp_mb()
124  *                                           kthread_unuse_mm()
125  *                                             d: smp_mb()
126  *                                             e: current->mm = NULL
127  *           b: read rq->curr->mm == NULL
128  *                                           kthread_use_mm()
129  *                                             f: current->mm = mm
130  *                                             g: smp_mb()
131  *           c: smp_mb()
132  *
133  * Using the scenario from (A), we can show that (a) needs to be paired
134  * with (g). Using the scenario from (B), we can show that (c) needs to
135  * be paired with (d).
136  */
137 
138 /*
139  * Bitmask made from a "or" of all commands within enum membarrier_cmd,
140  * except MEMBARRIER_CMD_QUERY.
141  */
142 #ifdef CONFIG_ARCH_HAS_MEMBARRIER_SYNC_CORE
143 #define MEMBARRIER_PRIVATE_EXPEDITED_SYNC_CORE_BITMASK			\
144 	(MEMBARRIER_CMD_PRIVATE_EXPEDITED_SYNC_CORE			\
145 	| MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_SYNC_CORE)
146 #else
147 #define MEMBARRIER_PRIVATE_EXPEDITED_SYNC_CORE_BITMASK	0
148 #endif
149 
150 #ifdef CONFIG_RSEQ
151 #define MEMBARRIER_PRIVATE_EXPEDITED_RSEQ_BITMASK		\
152 	(MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ			\
153 	| MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_RSEQ)
154 #else
155 #define MEMBARRIER_PRIVATE_EXPEDITED_RSEQ_BITMASK	0
156 #endif
157 
158 #define MEMBARRIER_CMD_BITMASK						\
159 	(MEMBARRIER_CMD_GLOBAL | MEMBARRIER_CMD_GLOBAL_EXPEDITED	\
160 	| MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED			\
161 	| MEMBARRIER_CMD_PRIVATE_EXPEDITED				\
162 	| MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED			\
163 	| MEMBARRIER_PRIVATE_EXPEDITED_SYNC_CORE_BITMASK		\
164 	| MEMBARRIER_PRIVATE_EXPEDITED_RSEQ_BITMASK			\
165 	| MEMBARRIER_CMD_GET_REGISTRATIONS)
166 
167 static DEFINE_MUTEX(membarrier_ipi_mutex);
168 #define SERIALIZE_IPI() guard(mutex)(&membarrier_ipi_mutex)
169 
ipi_mb(void * info)170 static void ipi_mb(void *info)
171 {
172 	smp_mb();	/* IPIs should be serializing but paranoid. */
173 }
174 
ipi_sync_core(void * info)175 static void ipi_sync_core(void *info)
176 {
177 	/*
178 	 * The smp_mb() in membarrier after all the IPIs is supposed to
179 	 * ensure that memory on remote CPUs that occur before the IPI
180 	 * become visible to membarrier()'s caller -- see scenario B in
181 	 * the big comment at the top of this file.
182 	 *
183 	 * A sync_core() would provide this guarantee, but
184 	 * sync_core_before_usermode() might end up being deferred until
185 	 * after membarrier()'s smp_mb().
186 	 */
187 	smp_mb();	/* IPIs should be serializing but paranoid. */
188 
189 	sync_core_before_usermode();
190 }
191 
ipi_rseq(void * info)192 static void ipi_rseq(void *info)
193 {
194 	/*
195 	 * Ensure that all stores done by the calling thread are visible
196 	 * to the current task before the current task resumes.  We could
197 	 * probably optimize this away on most architectures, but by the
198 	 * time we've already sent an IPI, the cost of the extra smp_mb()
199 	 * is negligible.
200 	 */
201 	smp_mb();
202 	rseq_preempt(current);
203 }
204 
ipi_sync_rq_state(void * info)205 static void ipi_sync_rq_state(void *info)
206 {
207 	struct mm_struct *mm = (struct mm_struct *) info;
208 
209 	if (current->mm != mm)
210 		return;
211 	this_cpu_write(runqueues.membarrier_state,
212 		       atomic_read(&mm->membarrier_state));
213 	/*
214 	 * Issue a memory barrier after setting
215 	 * MEMBARRIER_STATE_GLOBAL_EXPEDITED in the current runqueue to
216 	 * guarantee that no memory access following registration is reordered
217 	 * before registration.
218 	 */
219 	smp_mb();
220 }
221 
membarrier_exec_mmap(struct mm_struct * mm)222 void membarrier_exec_mmap(struct mm_struct *mm)
223 {
224 	/*
225 	 * Issue a memory barrier before clearing membarrier_state to
226 	 * guarantee that no memory access prior to exec is reordered after
227 	 * clearing this state.
228 	 */
229 	smp_mb();
230 	atomic_set(&mm->membarrier_state, 0);
231 	/*
232 	 * Keep the runqueue membarrier_state in sync with this mm
233 	 * membarrier_state.
234 	 */
235 	this_cpu_write(runqueues.membarrier_state, 0);
236 }
237 
membarrier_update_current_mm(struct mm_struct * next_mm)238 void membarrier_update_current_mm(struct mm_struct *next_mm)
239 {
240 	struct rq *rq = this_rq();
241 	int membarrier_state = 0;
242 
243 	if (next_mm)
244 		membarrier_state = atomic_read(&next_mm->membarrier_state);
245 	if (READ_ONCE(rq->membarrier_state) == membarrier_state)
246 		return;
247 	WRITE_ONCE(rq->membarrier_state, membarrier_state);
248 }
249 
membarrier_global_expedited(void)250 static int membarrier_global_expedited(void)
251 {
252 	int cpu;
253 	cpumask_var_t tmpmask;
254 
255 	if (num_online_cpus() == 1)
256 		return 0;
257 
258 	/*
259 	 * Matches memory barriers after rq->curr modification in
260 	 * scheduler.
261 	 */
262 	smp_mb();	/* system call entry is not a mb. */
263 
264 	if (!zalloc_cpumask_var(&tmpmask, GFP_KERNEL))
265 		return -ENOMEM;
266 
267 	SERIALIZE_IPI();
268 	cpus_read_lock();
269 	rcu_read_lock();
270 	for_each_online_cpu(cpu) {
271 		struct task_struct *p;
272 
273 		/*
274 		 * Skipping the current CPU is OK even through we can be
275 		 * migrated at any point. The current CPU, at the point
276 		 * where we read raw_smp_processor_id(), is ensured to
277 		 * be in program order with respect to the caller
278 		 * thread. Therefore, we can skip this CPU from the
279 		 * iteration.
280 		 */
281 		if (cpu == raw_smp_processor_id())
282 			continue;
283 
284 		if (!(READ_ONCE(cpu_rq(cpu)->membarrier_state) &
285 		    MEMBARRIER_STATE_GLOBAL_EXPEDITED))
286 			continue;
287 
288 		/*
289 		 * Skip the CPU if it runs a kernel thread which is not using
290 		 * a task mm.
291 		 */
292 		p = rcu_dereference(cpu_rq(cpu)->curr);
293 		if (!p->mm)
294 			continue;
295 
296 		__cpumask_set_cpu(cpu, tmpmask);
297 	}
298 	rcu_read_unlock();
299 
300 	preempt_disable();
301 	smp_call_function_many(tmpmask, ipi_mb, NULL, 1);
302 	preempt_enable();
303 
304 	free_cpumask_var(tmpmask);
305 	cpus_read_unlock();
306 
307 	/*
308 	 * Memory barrier on the caller thread _after_ we finished
309 	 * waiting for the last IPI. Matches memory barriers before
310 	 * rq->curr modification in scheduler.
311 	 */
312 	smp_mb();	/* exit from system call is not a mb */
313 	return 0;
314 }
315 
membarrier_private_expedited(int flags,int cpu_id)316 static int membarrier_private_expedited(int flags, int cpu_id)
317 {
318 	cpumask_var_t tmpmask;
319 	struct mm_struct *mm = current->mm;
320 	smp_call_func_t ipi_func = ipi_mb;
321 
322 	if (flags == MEMBARRIER_FLAG_SYNC_CORE) {
323 		if (!IS_ENABLED(CONFIG_ARCH_HAS_MEMBARRIER_SYNC_CORE))
324 			return -EINVAL;
325 		if (!(atomic_read(&mm->membarrier_state) &
326 		      MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE_READY))
327 			return -EPERM;
328 		ipi_func = ipi_sync_core;
329 		prepare_sync_core_cmd(mm);
330 	} else if (flags == MEMBARRIER_FLAG_RSEQ) {
331 		if (!IS_ENABLED(CONFIG_RSEQ))
332 			return -EINVAL;
333 		if (!(atomic_read(&mm->membarrier_state) &
334 		      MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ_READY))
335 			return -EPERM;
336 		ipi_func = ipi_rseq;
337 	} else {
338 		WARN_ON_ONCE(flags);
339 		if (!(atomic_read(&mm->membarrier_state) &
340 		      MEMBARRIER_STATE_PRIVATE_EXPEDITED_READY))
341 			return -EPERM;
342 	}
343 
344 	if (flags != MEMBARRIER_FLAG_SYNC_CORE &&
345 	    (atomic_read(&mm->mm_users) == 1 || num_online_cpus() == 1))
346 		return 0;
347 
348 	/*
349 	 * Matches memory barriers after rq->curr modification in
350 	 * scheduler.
351 	 *
352 	 * On RISC-V, this barrier pairing is also needed for the
353 	 * SYNC_CORE command when switching between processes, cf.
354 	 * the inline comments in membarrier_arch_switch_mm().
355 	 */
356 	smp_mb();	/* system call entry is not a mb. */
357 
358 	if (cpu_id < 0 && !zalloc_cpumask_var(&tmpmask, GFP_KERNEL))
359 		return -ENOMEM;
360 
361 	SERIALIZE_IPI();
362 	cpus_read_lock();
363 
364 	if (cpu_id >= 0) {
365 		struct task_struct *p;
366 
367 		if (cpu_id >= nr_cpu_ids || !cpu_online(cpu_id))
368 			goto out;
369 		rcu_read_lock();
370 		p = rcu_dereference(cpu_rq(cpu_id)->curr);
371 		if (!p || p->mm != mm) {
372 			rcu_read_unlock();
373 			goto out;
374 		}
375 		rcu_read_unlock();
376 	} else {
377 		int cpu;
378 
379 		rcu_read_lock();
380 		for_each_online_cpu(cpu) {
381 			struct task_struct *p;
382 
383 			p = rcu_dereference(cpu_rq(cpu)->curr);
384 			if (p && p->mm == mm)
385 				__cpumask_set_cpu(cpu, tmpmask);
386 		}
387 		rcu_read_unlock();
388 	}
389 
390 	if (cpu_id >= 0) {
391 		/*
392 		 * smp_call_function_single() will call ipi_func() if cpu_id
393 		 * is the calling CPU.
394 		 */
395 		smp_call_function_single(cpu_id, ipi_func, NULL, 1);
396 	} else {
397 		/*
398 		 * For regular membarrier, we can save a few cycles by
399 		 * skipping the current cpu -- we're about to do smp_mb()
400 		 * below, and if we migrate to a different cpu, this cpu
401 		 * and the new cpu will execute a full barrier in the
402 		 * scheduler.
403 		 *
404 		 * For SYNC_CORE, we do need a barrier on the current cpu --
405 		 * otherwise, if we are migrated and replaced by a different
406 		 * task in the same mm just before, during, or after
407 		 * membarrier, we will end up with some thread in the mm
408 		 * running without a core sync.
409 		 *
410 		 * For RSEQ, don't rseq_preempt() the caller.  User code
411 		 * is not supposed to issue syscalls at all from inside an
412 		 * rseq critical section.
413 		 */
414 		if (flags != MEMBARRIER_FLAG_SYNC_CORE) {
415 			preempt_disable();
416 			smp_call_function_many(tmpmask, ipi_func, NULL, true);
417 			preempt_enable();
418 		} else {
419 			on_each_cpu_mask(tmpmask, ipi_func, NULL, true);
420 		}
421 	}
422 
423 out:
424 	if (cpu_id < 0)
425 		free_cpumask_var(tmpmask);
426 	cpus_read_unlock();
427 
428 	/*
429 	 * Memory barrier on the caller thread _after_ we finished
430 	 * waiting for the last IPI. Matches memory barriers before
431 	 * rq->curr modification in scheduler.
432 	 */
433 	smp_mb();	/* exit from system call is not a mb */
434 
435 	return 0;
436 }
437 
sync_runqueues_membarrier_state(struct mm_struct * mm)438 static int sync_runqueues_membarrier_state(struct mm_struct *mm)
439 {
440 	int membarrier_state = atomic_read(&mm->membarrier_state);
441 	cpumask_var_t tmpmask;
442 	int cpu;
443 
444 	if (atomic_read(&mm->mm_users) == 1 || num_online_cpus() == 1) {
445 		this_cpu_write(runqueues.membarrier_state, membarrier_state);
446 
447 		/*
448 		 * For single mm user, we can simply issue a memory barrier
449 		 * after setting MEMBARRIER_STATE_GLOBAL_EXPEDITED in the
450 		 * mm and in the current runqueue to guarantee that no memory
451 		 * access following registration is reordered before
452 		 * registration.
453 		 */
454 		smp_mb();
455 		return 0;
456 	}
457 
458 	if (!zalloc_cpumask_var(&tmpmask, GFP_KERNEL))
459 		return -ENOMEM;
460 
461 	/*
462 	 * For mm with multiple users, we need to ensure all future
463 	 * scheduler executions will observe @mm's new membarrier
464 	 * state.
465 	 */
466 	synchronize_rcu();
467 
468 	/*
469 	 * For each cpu runqueue, if the task's mm match @mm, ensure that all
470 	 * @mm's membarrier state set bits are also set in the runqueue's
471 	 * membarrier state. This ensures that a runqueue scheduling
472 	 * between threads which are users of @mm has its membarrier state
473 	 * updated.
474 	 */
475 	SERIALIZE_IPI();
476 	cpus_read_lock();
477 	rcu_read_lock();
478 	for_each_online_cpu(cpu) {
479 		struct rq *rq = cpu_rq(cpu);
480 		struct task_struct *p;
481 
482 		p = rcu_dereference(rq->curr);
483 		if (p && p->mm == mm)
484 			__cpumask_set_cpu(cpu, tmpmask);
485 	}
486 	rcu_read_unlock();
487 
488 	on_each_cpu_mask(tmpmask, ipi_sync_rq_state, mm, true);
489 
490 	free_cpumask_var(tmpmask);
491 	cpus_read_unlock();
492 
493 	return 0;
494 }
495 
membarrier_register_global_expedited(void)496 static int membarrier_register_global_expedited(void)
497 {
498 	struct task_struct *p = current;
499 	struct mm_struct *mm = p->mm;
500 	int ret;
501 
502 	if (atomic_read(&mm->membarrier_state) &
503 	    MEMBARRIER_STATE_GLOBAL_EXPEDITED_READY)
504 		return 0;
505 	atomic_or(MEMBARRIER_STATE_GLOBAL_EXPEDITED, &mm->membarrier_state);
506 	ret = sync_runqueues_membarrier_state(mm);
507 	if (ret)
508 		return ret;
509 	atomic_or(MEMBARRIER_STATE_GLOBAL_EXPEDITED_READY,
510 		  &mm->membarrier_state);
511 
512 	return 0;
513 }
514 
membarrier_register_private_expedited(int flags)515 static int membarrier_register_private_expedited(int flags)
516 {
517 	struct task_struct *p = current;
518 	struct mm_struct *mm = p->mm;
519 	int ready_state = MEMBARRIER_STATE_PRIVATE_EXPEDITED_READY,
520 	    set_state = MEMBARRIER_STATE_PRIVATE_EXPEDITED,
521 	    ret;
522 
523 	if (flags == MEMBARRIER_FLAG_SYNC_CORE) {
524 		if (!IS_ENABLED(CONFIG_ARCH_HAS_MEMBARRIER_SYNC_CORE))
525 			return -EINVAL;
526 		ready_state =
527 			MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE_READY;
528 	} else if (flags == MEMBARRIER_FLAG_RSEQ) {
529 		if (!IS_ENABLED(CONFIG_RSEQ))
530 			return -EINVAL;
531 		ready_state =
532 			MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ_READY;
533 	} else {
534 		WARN_ON_ONCE(flags);
535 	}
536 
537 	/*
538 	 * We need to consider threads belonging to different thread
539 	 * groups, which use the same mm. (CLONE_VM but not
540 	 * CLONE_THREAD).
541 	 */
542 	if ((atomic_read(&mm->membarrier_state) & ready_state) == ready_state)
543 		return 0;
544 	if (flags & MEMBARRIER_FLAG_SYNC_CORE)
545 		set_state |= MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE;
546 	if (flags & MEMBARRIER_FLAG_RSEQ)
547 		set_state |= MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ;
548 	atomic_or(set_state, &mm->membarrier_state);
549 	ret = sync_runqueues_membarrier_state(mm);
550 	if (ret)
551 		return ret;
552 	atomic_or(ready_state, &mm->membarrier_state);
553 
554 	return 0;
555 }
556 
membarrier_get_registrations(void)557 static int membarrier_get_registrations(void)
558 {
559 	struct task_struct *p = current;
560 	struct mm_struct *mm = p->mm;
561 	int registrations_mask = 0, membarrier_state, i;
562 	static const int states[] = {
563 		MEMBARRIER_STATE_GLOBAL_EXPEDITED |
564 			MEMBARRIER_STATE_GLOBAL_EXPEDITED_READY,
565 		MEMBARRIER_STATE_PRIVATE_EXPEDITED |
566 			MEMBARRIER_STATE_PRIVATE_EXPEDITED_READY,
567 		MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE |
568 			MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE_READY,
569 		MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ |
570 			MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ_READY
571 	};
572 	static const int registration_cmds[] = {
573 		MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED,
574 		MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED,
575 		MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_SYNC_CORE,
576 		MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_RSEQ
577 	};
578 	BUILD_BUG_ON(ARRAY_SIZE(states) != ARRAY_SIZE(registration_cmds));
579 
580 	membarrier_state = atomic_read(&mm->membarrier_state);
581 	for (i = 0; i < ARRAY_SIZE(states); ++i) {
582 		if (membarrier_state & states[i]) {
583 			registrations_mask |= registration_cmds[i];
584 			membarrier_state &= ~states[i];
585 		}
586 	}
587 	WARN_ON_ONCE(membarrier_state != 0);
588 	return registrations_mask;
589 }
590 
591 /**
592  * sys_membarrier - issue memory barriers on a set of threads
593  * @cmd:    Takes command values defined in enum membarrier_cmd.
594  * @flags:  Currently needs to be 0 for all commands other than
595  *          MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ: in the latter
596  *          case it can be MEMBARRIER_CMD_FLAG_CPU, indicating that @cpu_id
597  *          contains the CPU on which to interrupt (= restart)
598  *          the RSEQ critical section.
599  * @cpu_id: if @flags == MEMBARRIER_CMD_FLAG_CPU, indicates the cpu on which
600  *          RSEQ CS should be interrupted (@cmd must be
601  *          MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ).
602  *
603  * If this system call is not implemented, -ENOSYS is returned. If the
604  * command specified does not exist, not available on the running
605  * kernel, or if the command argument is invalid, this system call
606  * returns -EINVAL. For a given command, with flags argument set to 0,
607  * if this system call returns -ENOSYS or -EINVAL, it is guaranteed to
608  * always return the same value until reboot. In addition, it can return
609  * -ENOMEM if there is not enough memory available to perform the system
610  * call.
611  *
612  * All memory accesses performed in program order from each targeted thread
613  * is guaranteed to be ordered with respect to sys_membarrier(). If we use
614  * the semantic "barrier()" to represent a compiler barrier forcing memory
615  * accesses to be performed in program order across the barrier, and
616  * smp_mb() to represent explicit memory barriers forcing full memory
617  * ordering across the barrier, we have the following ordering table for
618  * each pair of barrier(), sys_membarrier() and smp_mb():
619  *
620  * The pair ordering is detailed as (O: ordered, X: not ordered):
621  *
622  *                        barrier()   smp_mb() sys_membarrier()
623  *        barrier()          X           X            O
624  *        smp_mb()           X           O            O
625  *        sys_membarrier()   O           O            O
626  */
SYSCALL_DEFINE3(membarrier,int,cmd,unsigned int,flags,int,cpu_id)627 SYSCALL_DEFINE3(membarrier, int, cmd, unsigned int, flags, int, cpu_id)
628 {
629 	switch (cmd) {
630 	case MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ:
631 		if (unlikely(flags && flags != MEMBARRIER_CMD_FLAG_CPU))
632 			return -EINVAL;
633 		break;
634 	default:
635 		if (unlikely(flags))
636 			return -EINVAL;
637 	}
638 
639 	if (!(flags & MEMBARRIER_CMD_FLAG_CPU))
640 		cpu_id = -1;
641 
642 	switch (cmd) {
643 	case MEMBARRIER_CMD_QUERY:
644 	{
645 		int cmd_mask = MEMBARRIER_CMD_BITMASK;
646 
647 		if (tick_nohz_full_enabled())
648 			cmd_mask &= ~MEMBARRIER_CMD_GLOBAL;
649 		return cmd_mask;
650 	}
651 	case MEMBARRIER_CMD_GLOBAL:
652 		/* MEMBARRIER_CMD_GLOBAL is not compatible with nohz_full. */
653 		if (tick_nohz_full_enabled())
654 			return -EINVAL;
655 		if (num_online_cpus() > 1)
656 			synchronize_rcu();
657 		return 0;
658 	case MEMBARRIER_CMD_GLOBAL_EXPEDITED:
659 		return membarrier_global_expedited();
660 	case MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED:
661 		return membarrier_register_global_expedited();
662 	case MEMBARRIER_CMD_PRIVATE_EXPEDITED:
663 		return membarrier_private_expedited(0, cpu_id);
664 	case MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED:
665 		return membarrier_register_private_expedited(0);
666 	case MEMBARRIER_CMD_PRIVATE_EXPEDITED_SYNC_CORE:
667 		return membarrier_private_expedited(MEMBARRIER_FLAG_SYNC_CORE, cpu_id);
668 	case MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_SYNC_CORE:
669 		return membarrier_register_private_expedited(MEMBARRIER_FLAG_SYNC_CORE);
670 	case MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ:
671 		return membarrier_private_expedited(MEMBARRIER_FLAG_RSEQ, cpu_id);
672 	case MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_RSEQ:
673 		return membarrier_register_private_expedited(MEMBARRIER_FLAG_RSEQ);
674 	case MEMBARRIER_CMD_GET_REGISTRATIONS:
675 		return membarrier_get_registrations();
676 	default:
677 		return -EINVAL;
678 	}
679 }
680