xref: /linux/kernel/sched/membarrier.c (revision 79790b6818e96c58fe2bffee1b418c16e64e7b80)
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 
8 /*
9  * For documentation purposes, here are some membarrier ordering
10  * scenarios to keep in mind:
11  *
12  * A) Userspace thread execution after IPI vs membarrier's memory
13  *    barrier before sending the IPI
14  *
15  * Userspace variables:
16  *
17  * int x = 0, y = 0;
18  *
19  * The memory barrier at the start of membarrier() on CPU0 is necessary in
20  * order to enforce the guarantee that any writes occurring on CPU0 before
21  * the membarrier() is executed will be visible to any code executing on
22  * CPU1 after the IPI-induced memory barrier:
23  *
24  *         CPU0                              CPU1
25  *
26  *         x = 1
27  *         membarrier():
28  *           a: smp_mb()
29  *           b: send IPI                       IPI-induced mb
30  *           c: smp_mb()
31  *         r2 = y
32  *                                           y = 1
33  *                                           barrier()
34  *                                           r1 = x
35  *
36  *                     BUG_ON(r1 == 0 && r2 == 0)
37  *
38  * The write to y and load from x by CPU1 are unordered by the hardware,
39  * so it's possible to have "r1 = x" reordered before "y = 1" at any
40  * point after (b).  If the memory barrier at (a) is omitted, then "x = 1"
41  * can be reordered after (a) (although not after (c)), so we get r1 == 0
42  * and r2 == 0.  This violates the guarantee that membarrier() is
43  * supposed by provide.
44  *
45  * The timing of the memory barrier at (a) has to ensure that it executes
46  * before the IPI-induced memory barrier on CPU1.
47  *
48  * B) Userspace thread execution before IPI vs membarrier's memory
49  *    barrier after completing the IPI
50  *
51  * Userspace variables:
52  *
53  * int x = 0, y = 0;
54  *
55  * The memory barrier at the end of membarrier() on CPU0 is necessary in
56  * order to enforce the guarantee that any writes occurring on CPU1 before
57  * the membarrier() is executed will be visible to any code executing on
58  * CPU0 after the membarrier():
59  *
60  *         CPU0                              CPU1
61  *
62  *                                           x = 1
63  *                                           barrier()
64  *                                           y = 1
65  *         r2 = y
66  *         membarrier():
67  *           a: smp_mb()
68  *           b: send IPI                       IPI-induced mb
69  *           c: smp_mb()
70  *         r1 = x
71  *         BUG_ON(r1 == 0 && r2 == 1)
72  *
73  * The writes to x and y are unordered by the hardware, so it's possible to
74  * have "r2 = 1" even though the write to x doesn't execute until (b).  If
75  * the memory barrier at (c) is omitted then "r1 = x" can be reordered
76  * before (b) (although not before (a)), so we get "r1 = 0".  This violates
77  * the guarantee that membarrier() is supposed to provide.
78  *
79  * The timing of the memory barrier at (c) has to ensure that it executes
80  * after the IPI-induced memory barrier on CPU1.
81  *
82  * C) Scheduling userspace thread -> kthread -> userspace thread vs membarrier
83  *
84  *           CPU0                            CPU1
85  *
86  *           membarrier():
87  *           a: smp_mb()
88  *                                           d: switch to kthread (includes mb)
89  *           b: read rq->curr->mm == NULL
90  *                                           e: switch to user (includes mb)
91  *           c: smp_mb()
92  *
93  * Using the scenario from (A), we can show that (a) needs to be paired
94  * with (e). Using the scenario from (B), we can show that (c) needs to
95  * be paired with (d).
96  *
97  * D) exit_mm vs membarrier
98  *
99  * Two thread groups are created, A and B.  Thread group B is created by
100  * issuing clone from group A with flag CLONE_VM set, but not CLONE_THREAD.
101  * Let's assume we have a single thread within each thread group (Thread A
102  * and Thread B).  Thread A runs on CPU0, Thread B runs on CPU1.
103  *
104  *           CPU0                            CPU1
105  *
106  *           membarrier():
107  *             a: smp_mb()
108  *                                           exit_mm():
109  *                                             d: smp_mb()
110  *                                             e: current->mm = NULL
111  *             b: read rq->curr->mm == NULL
112  *             c: smp_mb()
113  *
114  * Using scenario (B), we can show that (c) needs to be paired with (d).
115  *
116  * E) kthread_{use,unuse}_mm vs membarrier
117  *
118  *           CPU0                            CPU1
119  *
120  *           membarrier():
121  *           a: smp_mb()
122  *                                           kthread_unuse_mm()
123  *                                             d: smp_mb()
124  *                                             e: current->mm = NULL
125  *           b: read rq->curr->mm == NULL
126  *                                           kthread_use_mm()
127  *                                             f: current->mm = mm
128  *                                             g: smp_mb()
129  *           c: smp_mb()
130  *
131  * Using the scenario from (A), we can show that (a) needs to be paired
132  * with (g). Using the scenario from (B), we can show that (c) needs to
133  * be paired with (d).
134  */
135 
136 /*
137  * Bitmask made from a "or" of all commands within enum membarrier_cmd,
138  * except MEMBARRIER_CMD_QUERY.
139  */
140 #ifdef CONFIG_ARCH_HAS_MEMBARRIER_SYNC_CORE
141 #define MEMBARRIER_PRIVATE_EXPEDITED_SYNC_CORE_BITMASK			\
142 	(MEMBARRIER_CMD_PRIVATE_EXPEDITED_SYNC_CORE			\
143 	| MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_SYNC_CORE)
144 #else
145 #define MEMBARRIER_PRIVATE_EXPEDITED_SYNC_CORE_BITMASK	0
146 #endif
147 
148 #ifdef CONFIG_RSEQ
149 #define MEMBARRIER_PRIVATE_EXPEDITED_RSEQ_BITMASK		\
150 	(MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ			\
151 	| MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_RSEQ)
152 #else
153 #define MEMBARRIER_PRIVATE_EXPEDITED_RSEQ_BITMASK	0
154 #endif
155 
156 #define MEMBARRIER_CMD_BITMASK						\
157 	(MEMBARRIER_CMD_GLOBAL | MEMBARRIER_CMD_GLOBAL_EXPEDITED	\
158 	| MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED			\
159 	| MEMBARRIER_CMD_PRIVATE_EXPEDITED				\
160 	| MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED			\
161 	| MEMBARRIER_PRIVATE_EXPEDITED_SYNC_CORE_BITMASK		\
162 	| MEMBARRIER_PRIVATE_EXPEDITED_RSEQ_BITMASK			\
163 	| MEMBARRIER_CMD_GET_REGISTRATIONS)
164 
165 static DEFINE_MUTEX(membarrier_ipi_mutex);
166 #define SERIALIZE_IPI() guard(mutex)(&membarrier_ipi_mutex)
167 
ipi_mb(void * info)168 static void ipi_mb(void *info)
169 {
170 	smp_mb();	/* IPIs should be serializing but paranoid. */
171 }
172 
ipi_sync_core(void * info)173 static void ipi_sync_core(void *info)
174 {
175 	/*
176 	 * The smp_mb() in membarrier after all the IPIs is supposed to
177 	 * ensure that memory on remote CPUs that occur before the IPI
178 	 * become visible to membarrier()'s caller -- see scenario B in
179 	 * the big comment at the top of this file.
180 	 *
181 	 * A sync_core() would provide this guarantee, but
182 	 * sync_core_before_usermode() might end up being deferred until
183 	 * after membarrier()'s smp_mb().
184 	 */
185 	smp_mb();	/* IPIs should be serializing but paranoid. */
186 
187 	sync_core_before_usermode();
188 }
189 
ipi_rseq(void * info)190 static void ipi_rseq(void *info)
191 {
192 	/*
193 	 * Ensure that all stores done by the calling thread are visible
194 	 * to the current task before the current task resumes.  We could
195 	 * probably optimize this away on most architectures, but by the
196 	 * time we've already sent an IPI, the cost of the extra smp_mb()
197 	 * is negligible.
198 	 */
199 	smp_mb();
200 	rseq_preempt(current);
201 }
202 
ipi_sync_rq_state(void * info)203 static void ipi_sync_rq_state(void *info)
204 {
205 	struct mm_struct *mm = (struct mm_struct *) info;
206 
207 	if (current->mm != mm)
208 		return;
209 	this_cpu_write(runqueues.membarrier_state,
210 		       atomic_read(&mm->membarrier_state));
211 	/*
212 	 * Issue a memory barrier after setting
213 	 * MEMBARRIER_STATE_GLOBAL_EXPEDITED in the current runqueue to
214 	 * guarantee that no memory access following registration is reordered
215 	 * before registration.
216 	 */
217 	smp_mb();
218 }
219 
membarrier_exec_mmap(struct mm_struct * mm)220 void membarrier_exec_mmap(struct mm_struct *mm)
221 {
222 	/*
223 	 * Issue a memory barrier before clearing membarrier_state to
224 	 * guarantee that no memory access prior to exec is reordered after
225 	 * clearing this state.
226 	 */
227 	smp_mb();
228 	atomic_set(&mm->membarrier_state, 0);
229 	/*
230 	 * Keep the runqueue membarrier_state in sync with this mm
231 	 * membarrier_state.
232 	 */
233 	this_cpu_write(runqueues.membarrier_state, 0);
234 }
235 
membarrier_update_current_mm(struct mm_struct * next_mm)236 void membarrier_update_current_mm(struct mm_struct *next_mm)
237 {
238 	struct rq *rq = this_rq();
239 	int membarrier_state = 0;
240 
241 	if (next_mm)
242 		membarrier_state = atomic_read(&next_mm->membarrier_state);
243 	if (READ_ONCE(rq->membarrier_state) == membarrier_state)
244 		return;
245 	WRITE_ONCE(rq->membarrier_state, membarrier_state);
246 }
247 
membarrier_global_expedited(void)248 static int membarrier_global_expedited(void)
249 {
250 	int cpu;
251 	cpumask_var_t tmpmask;
252 
253 	if (num_online_cpus() == 1)
254 		return 0;
255 
256 	/*
257 	 * Matches memory barriers after rq->curr modification in
258 	 * scheduler.
259 	 */
260 	smp_mb();	/* system call entry is not a mb. */
261 
262 	if (!zalloc_cpumask_var(&tmpmask, GFP_KERNEL))
263 		return -ENOMEM;
264 
265 	SERIALIZE_IPI();
266 	cpus_read_lock();
267 	rcu_read_lock();
268 	for_each_online_cpu(cpu) {
269 		struct task_struct *p;
270 
271 		/*
272 		 * Skipping the current CPU is OK even through we can be
273 		 * migrated at any point. The current CPU, at the point
274 		 * where we read raw_smp_processor_id(), is ensured to
275 		 * be in program order with respect to the caller
276 		 * thread. Therefore, we can skip this CPU from the
277 		 * iteration.
278 		 */
279 		if (cpu == raw_smp_processor_id())
280 			continue;
281 
282 		if (!(READ_ONCE(cpu_rq(cpu)->membarrier_state) &
283 		    MEMBARRIER_STATE_GLOBAL_EXPEDITED))
284 			continue;
285 
286 		/*
287 		 * Skip the CPU if it runs a kernel thread which is not using
288 		 * a task mm.
289 		 */
290 		p = rcu_dereference(cpu_rq(cpu)->curr);
291 		if (!p->mm)
292 			continue;
293 
294 		__cpumask_set_cpu(cpu, tmpmask);
295 	}
296 	rcu_read_unlock();
297 
298 	preempt_disable();
299 	smp_call_function_many(tmpmask, ipi_mb, NULL, 1);
300 	preempt_enable();
301 
302 	free_cpumask_var(tmpmask);
303 	cpus_read_unlock();
304 
305 	/*
306 	 * Memory barrier on the caller thread _after_ we finished
307 	 * waiting for the last IPI. Matches memory barriers before
308 	 * rq->curr modification in scheduler.
309 	 */
310 	smp_mb();	/* exit from system call is not a mb */
311 	return 0;
312 }
313 
membarrier_private_expedited(int flags,int cpu_id)314 static int membarrier_private_expedited(int flags, int cpu_id)
315 {
316 	cpumask_var_t tmpmask;
317 	struct mm_struct *mm = current->mm;
318 	smp_call_func_t ipi_func = ipi_mb;
319 
320 	if (flags == MEMBARRIER_FLAG_SYNC_CORE) {
321 		if (!IS_ENABLED(CONFIG_ARCH_HAS_MEMBARRIER_SYNC_CORE))
322 			return -EINVAL;
323 		if (!(atomic_read(&mm->membarrier_state) &
324 		      MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE_READY))
325 			return -EPERM;
326 		ipi_func = ipi_sync_core;
327 		prepare_sync_core_cmd(mm);
328 	} else if (flags == MEMBARRIER_FLAG_RSEQ) {
329 		if (!IS_ENABLED(CONFIG_RSEQ))
330 			return -EINVAL;
331 		if (!(atomic_read(&mm->membarrier_state) &
332 		      MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ_READY))
333 			return -EPERM;
334 		ipi_func = ipi_rseq;
335 	} else {
336 		WARN_ON_ONCE(flags);
337 		if (!(atomic_read(&mm->membarrier_state) &
338 		      MEMBARRIER_STATE_PRIVATE_EXPEDITED_READY))
339 			return -EPERM;
340 	}
341 
342 	if (flags != MEMBARRIER_FLAG_SYNC_CORE &&
343 	    (atomic_read(&mm->mm_users) == 1 || num_online_cpus() == 1))
344 		return 0;
345 
346 	/*
347 	 * Matches memory barriers after rq->curr modification in
348 	 * scheduler.
349 	 *
350 	 * On RISC-V, this barrier pairing is also needed for the
351 	 * SYNC_CORE command when switching between processes, cf.
352 	 * the inline comments in membarrier_arch_switch_mm().
353 	 */
354 	smp_mb();	/* system call entry is not a mb. */
355 
356 	if (cpu_id < 0 && !zalloc_cpumask_var(&tmpmask, GFP_KERNEL))
357 		return -ENOMEM;
358 
359 	SERIALIZE_IPI();
360 	cpus_read_lock();
361 
362 	if (cpu_id >= 0) {
363 		struct task_struct *p;
364 
365 		if (cpu_id >= nr_cpu_ids || !cpu_online(cpu_id))
366 			goto out;
367 		rcu_read_lock();
368 		p = rcu_dereference(cpu_rq(cpu_id)->curr);
369 		if (!p || p->mm != mm) {
370 			rcu_read_unlock();
371 			goto out;
372 		}
373 		rcu_read_unlock();
374 	} else {
375 		int cpu;
376 
377 		rcu_read_lock();
378 		for_each_online_cpu(cpu) {
379 			struct task_struct *p;
380 
381 			p = rcu_dereference(cpu_rq(cpu)->curr);
382 			if (p && p->mm == mm)
383 				__cpumask_set_cpu(cpu, tmpmask);
384 		}
385 		rcu_read_unlock();
386 	}
387 
388 	if (cpu_id >= 0) {
389 		/*
390 		 * smp_call_function_single() will call ipi_func() if cpu_id
391 		 * is the calling CPU.
392 		 */
393 		smp_call_function_single(cpu_id, ipi_func, NULL, 1);
394 	} else {
395 		/*
396 		 * For regular membarrier, we can save a few cycles by
397 		 * skipping the current cpu -- we're about to do smp_mb()
398 		 * below, and if we migrate to a different cpu, this cpu
399 		 * and the new cpu will execute a full barrier in the
400 		 * scheduler.
401 		 *
402 		 * For SYNC_CORE, we do need a barrier on the current cpu --
403 		 * otherwise, if we are migrated and replaced by a different
404 		 * task in the same mm just before, during, or after
405 		 * membarrier, we will end up with some thread in the mm
406 		 * running without a core sync.
407 		 *
408 		 * For RSEQ, don't rseq_preempt() the caller.  User code
409 		 * is not supposed to issue syscalls at all from inside an
410 		 * rseq critical section.
411 		 */
412 		if (flags != MEMBARRIER_FLAG_SYNC_CORE) {
413 			preempt_disable();
414 			smp_call_function_many(tmpmask, ipi_func, NULL, true);
415 			preempt_enable();
416 		} else {
417 			on_each_cpu_mask(tmpmask, ipi_func, NULL, true);
418 		}
419 	}
420 
421 out:
422 	if (cpu_id < 0)
423 		free_cpumask_var(tmpmask);
424 	cpus_read_unlock();
425 
426 	/*
427 	 * Memory barrier on the caller thread _after_ we finished
428 	 * waiting for the last IPI. Matches memory barriers before
429 	 * rq->curr modification in scheduler.
430 	 */
431 	smp_mb();	/* exit from system call is not a mb */
432 
433 	return 0;
434 }
435 
sync_runqueues_membarrier_state(struct mm_struct * mm)436 static int sync_runqueues_membarrier_state(struct mm_struct *mm)
437 {
438 	int membarrier_state = atomic_read(&mm->membarrier_state);
439 	cpumask_var_t tmpmask;
440 	int cpu;
441 
442 	if (atomic_read(&mm->mm_users) == 1 || num_online_cpus() == 1) {
443 		this_cpu_write(runqueues.membarrier_state, membarrier_state);
444 
445 		/*
446 		 * For single mm user, we can simply issue a memory barrier
447 		 * after setting MEMBARRIER_STATE_GLOBAL_EXPEDITED in the
448 		 * mm and in the current runqueue to guarantee that no memory
449 		 * access following registration is reordered before
450 		 * registration.
451 		 */
452 		smp_mb();
453 		return 0;
454 	}
455 
456 	if (!zalloc_cpumask_var(&tmpmask, GFP_KERNEL))
457 		return -ENOMEM;
458 
459 	/*
460 	 * For mm with multiple users, we need to ensure all future
461 	 * scheduler executions will observe @mm's new membarrier
462 	 * state.
463 	 */
464 	synchronize_rcu();
465 
466 	/*
467 	 * For each cpu runqueue, if the task's mm match @mm, ensure that all
468 	 * @mm's membarrier state set bits are also set in the runqueue's
469 	 * membarrier state. This ensures that a runqueue scheduling
470 	 * between threads which are users of @mm has its membarrier state
471 	 * updated.
472 	 */
473 	SERIALIZE_IPI();
474 	cpus_read_lock();
475 	rcu_read_lock();
476 	for_each_online_cpu(cpu) {
477 		struct rq *rq = cpu_rq(cpu);
478 		struct task_struct *p;
479 
480 		p = rcu_dereference(rq->curr);
481 		if (p && p->mm == mm)
482 			__cpumask_set_cpu(cpu, tmpmask);
483 	}
484 	rcu_read_unlock();
485 
486 	on_each_cpu_mask(tmpmask, ipi_sync_rq_state, mm, true);
487 
488 	free_cpumask_var(tmpmask);
489 	cpus_read_unlock();
490 
491 	return 0;
492 }
493 
membarrier_register_global_expedited(void)494 static int membarrier_register_global_expedited(void)
495 {
496 	struct task_struct *p = current;
497 	struct mm_struct *mm = p->mm;
498 	int ret;
499 
500 	if (atomic_read(&mm->membarrier_state) &
501 	    MEMBARRIER_STATE_GLOBAL_EXPEDITED_READY)
502 		return 0;
503 	atomic_or(MEMBARRIER_STATE_GLOBAL_EXPEDITED, &mm->membarrier_state);
504 	ret = sync_runqueues_membarrier_state(mm);
505 	if (ret)
506 		return ret;
507 	atomic_or(MEMBARRIER_STATE_GLOBAL_EXPEDITED_READY,
508 		  &mm->membarrier_state);
509 
510 	return 0;
511 }
512 
membarrier_register_private_expedited(int flags)513 static int membarrier_register_private_expedited(int flags)
514 {
515 	struct task_struct *p = current;
516 	struct mm_struct *mm = p->mm;
517 	int ready_state = MEMBARRIER_STATE_PRIVATE_EXPEDITED_READY,
518 	    set_state = MEMBARRIER_STATE_PRIVATE_EXPEDITED,
519 	    ret;
520 
521 	if (flags == MEMBARRIER_FLAG_SYNC_CORE) {
522 		if (!IS_ENABLED(CONFIG_ARCH_HAS_MEMBARRIER_SYNC_CORE))
523 			return -EINVAL;
524 		ready_state =
525 			MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE_READY;
526 	} else if (flags == MEMBARRIER_FLAG_RSEQ) {
527 		if (!IS_ENABLED(CONFIG_RSEQ))
528 			return -EINVAL;
529 		ready_state =
530 			MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ_READY;
531 	} else {
532 		WARN_ON_ONCE(flags);
533 	}
534 
535 	/*
536 	 * We need to consider threads belonging to different thread
537 	 * groups, which use the same mm. (CLONE_VM but not
538 	 * CLONE_THREAD).
539 	 */
540 	if ((atomic_read(&mm->membarrier_state) & ready_state) == ready_state)
541 		return 0;
542 	if (flags & MEMBARRIER_FLAG_SYNC_CORE)
543 		set_state |= MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE;
544 	if (flags & MEMBARRIER_FLAG_RSEQ)
545 		set_state |= MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ;
546 	atomic_or(set_state, &mm->membarrier_state);
547 	ret = sync_runqueues_membarrier_state(mm);
548 	if (ret)
549 		return ret;
550 	atomic_or(ready_state, &mm->membarrier_state);
551 
552 	return 0;
553 }
554 
membarrier_get_registrations(void)555 static int membarrier_get_registrations(void)
556 {
557 	struct task_struct *p = current;
558 	struct mm_struct *mm = p->mm;
559 	int registrations_mask = 0, membarrier_state, i;
560 	static const int states[] = {
561 		MEMBARRIER_STATE_GLOBAL_EXPEDITED |
562 			MEMBARRIER_STATE_GLOBAL_EXPEDITED_READY,
563 		MEMBARRIER_STATE_PRIVATE_EXPEDITED |
564 			MEMBARRIER_STATE_PRIVATE_EXPEDITED_READY,
565 		MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE |
566 			MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE_READY,
567 		MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ |
568 			MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ_READY
569 	};
570 	static const int registration_cmds[] = {
571 		MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED,
572 		MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED,
573 		MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_SYNC_CORE,
574 		MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_RSEQ
575 	};
576 	BUILD_BUG_ON(ARRAY_SIZE(states) != ARRAY_SIZE(registration_cmds));
577 
578 	membarrier_state = atomic_read(&mm->membarrier_state);
579 	for (i = 0; i < ARRAY_SIZE(states); ++i) {
580 		if (membarrier_state & states[i]) {
581 			registrations_mask |= registration_cmds[i];
582 			membarrier_state &= ~states[i];
583 		}
584 	}
585 	WARN_ON_ONCE(membarrier_state != 0);
586 	return registrations_mask;
587 }
588 
589 /**
590  * sys_membarrier - issue memory barriers on a set of threads
591  * @cmd:    Takes command values defined in enum membarrier_cmd.
592  * @flags:  Currently needs to be 0 for all commands other than
593  *          MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ: in the latter
594  *          case it can be MEMBARRIER_CMD_FLAG_CPU, indicating that @cpu_id
595  *          contains the CPU on which to interrupt (= restart)
596  *          the RSEQ critical section.
597  * @cpu_id: if @flags == MEMBARRIER_CMD_FLAG_CPU, indicates the cpu on which
598  *          RSEQ CS should be interrupted (@cmd must be
599  *          MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ).
600  *
601  * If this system call is not implemented, -ENOSYS is returned. If the
602  * command specified does not exist, not available on the running
603  * kernel, or if the command argument is invalid, this system call
604  * returns -EINVAL. For a given command, with flags argument set to 0,
605  * if this system call returns -ENOSYS or -EINVAL, it is guaranteed to
606  * always return the same value until reboot. In addition, it can return
607  * -ENOMEM if there is not enough memory available to perform the system
608  * call.
609  *
610  * All memory accesses performed in program order from each targeted thread
611  * is guaranteed to be ordered with respect to sys_membarrier(). If we use
612  * the semantic "barrier()" to represent a compiler barrier forcing memory
613  * accesses to be performed in program order across the barrier, and
614  * smp_mb() to represent explicit memory barriers forcing full memory
615  * ordering across the barrier, we have the following ordering table for
616  * each pair of barrier(), sys_membarrier() and smp_mb():
617  *
618  * The pair ordering is detailed as (O: ordered, X: not ordered):
619  *
620  *                        barrier()   smp_mb() sys_membarrier()
621  *        barrier()          X           X            O
622  *        smp_mb()           X           O            O
623  *        sys_membarrier()   O           O            O
624  */
SYSCALL_DEFINE3(membarrier,int,cmd,unsigned int,flags,int,cpu_id)625 SYSCALL_DEFINE3(membarrier, int, cmd, unsigned int, flags, int, cpu_id)
626 {
627 	switch (cmd) {
628 	case MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ:
629 		if (unlikely(flags && flags != MEMBARRIER_CMD_FLAG_CPU))
630 			return -EINVAL;
631 		break;
632 	default:
633 		if (unlikely(flags))
634 			return -EINVAL;
635 	}
636 
637 	if (!(flags & MEMBARRIER_CMD_FLAG_CPU))
638 		cpu_id = -1;
639 
640 	switch (cmd) {
641 	case MEMBARRIER_CMD_QUERY:
642 	{
643 		int cmd_mask = MEMBARRIER_CMD_BITMASK;
644 
645 		if (tick_nohz_full_enabled())
646 			cmd_mask &= ~MEMBARRIER_CMD_GLOBAL;
647 		return cmd_mask;
648 	}
649 	case MEMBARRIER_CMD_GLOBAL:
650 		/* MEMBARRIER_CMD_GLOBAL is not compatible with nohz_full. */
651 		if (tick_nohz_full_enabled())
652 			return -EINVAL;
653 		if (num_online_cpus() > 1)
654 			synchronize_rcu();
655 		return 0;
656 	case MEMBARRIER_CMD_GLOBAL_EXPEDITED:
657 		return membarrier_global_expedited();
658 	case MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED:
659 		return membarrier_register_global_expedited();
660 	case MEMBARRIER_CMD_PRIVATE_EXPEDITED:
661 		return membarrier_private_expedited(0, cpu_id);
662 	case MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED:
663 		return membarrier_register_private_expedited(0);
664 	case MEMBARRIER_CMD_PRIVATE_EXPEDITED_SYNC_CORE:
665 		return membarrier_private_expedited(MEMBARRIER_FLAG_SYNC_CORE, cpu_id);
666 	case MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_SYNC_CORE:
667 		return membarrier_register_private_expedited(MEMBARRIER_FLAG_SYNC_CORE);
668 	case MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ:
669 		return membarrier_private_expedited(MEMBARRIER_FLAG_RSEQ, cpu_id);
670 	case MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_RSEQ:
671 		return membarrier_register_private_expedited(MEMBARRIER_FLAG_RSEQ);
672 	case MEMBARRIER_CMD_GET_REGISTRATIONS:
673 		return membarrier_get_registrations();
674 	default:
675 		return -EINVAL;
676 	}
677 }
678