xref: /openbmc/linux/kernel/sched/membarrier.c (revision 6c9111bc)
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 "sched.h"
8 
9 /*
10  * Bitmask made from a "or" of all commands within enum membarrier_cmd,
11  * except MEMBARRIER_CMD_QUERY.
12  */
13 #ifdef CONFIG_ARCH_HAS_MEMBARRIER_SYNC_CORE
14 #define MEMBARRIER_PRIVATE_EXPEDITED_SYNC_CORE_BITMASK			\
15 	(MEMBARRIER_CMD_PRIVATE_EXPEDITED_SYNC_CORE			\
16 	| MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_SYNC_CORE)
17 #else
18 #define MEMBARRIER_PRIVATE_EXPEDITED_SYNC_CORE_BITMASK	0
19 #endif
20 
21 #ifdef CONFIG_RSEQ
22 #define MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ_BITMASK		\
23 	(MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ			\
24 	| MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_RSEQ_BITMASK)
25 #else
26 #define MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ_BITMASK	0
27 #endif
28 
29 #define MEMBARRIER_CMD_BITMASK						\
30 	(MEMBARRIER_CMD_GLOBAL | MEMBARRIER_CMD_GLOBAL_EXPEDITED	\
31 	| MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED			\
32 	| MEMBARRIER_CMD_PRIVATE_EXPEDITED				\
33 	| MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED			\
34 	| MEMBARRIER_PRIVATE_EXPEDITED_SYNC_CORE_BITMASK)
35 
36 static void ipi_mb(void *info)
37 {
38 	smp_mb();	/* IPIs should be serializing but paranoid. */
39 }
40 
41 static void ipi_rseq(void *info)
42 {
43 	rseq_preempt(current);
44 }
45 
46 static void ipi_sync_rq_state(void *info)
47 {
48 	struct mm_struct *mm = (struct mm_struct *) info;
49 
50 	if (current->mm != mm)
51 		return;
52 	this_cpu_write(runqueues.membarrier_state,
53 		       atomic_read(&mm->membarrier_state));
54 	/*
55 	 * Issue a memory barrier after setting
56 	 * MEMBARRIER_STATE_GLOBAL_EXPEDITED in the current runqueue to
57 	 * guarantee that no memory access following registration is reordered
58 	 * before registration.
59 	 */
60 	smp_mb();
61 }
62 
63 void membarrier_exec_mmap(struct mm_struct *mm)
64 {
65 	/*
66 	 * Issue a memory barrier before clearing membarrier_state to
67 	 * guarantee that no memory access prior to exec is reordered after
68 	 * clearing this state.
69 	 */
70 	smp_mb();
71 	atomic_set(&mm->membarrier_state, 0);
72 	/*
73 	 * Keep the runqueue membarrier_state in sync with this mm
74 	 * membarrier_state.
75 	 */
76 	this_cpu_write(runqueues.membarrier_state, 0);
77 }
78 
79 static int membarrier_global_expedited(void)
80 {
81 	int cpu;
82 	cpumask_var_t tmpmask;
83 
84 	if (num_online_cpus() == 1)
85 		return 0;
86 
87 	/*
88 	 * Matches memory barriers around rq->curr modification in
89 	 * scheduler.
90 	 */
91 	smp_mb();	/* system call entry is not a mb. */
92 
93 	if (!zalloc_cpumask_var(&tmpmask, GFP_KERNEL))
94 		return -ENOMEM;
95 
96 	cpus_read_lock();
97 	rcu_read_lock();
98 	for_each_online_cpu(cpu) {
99 		struct task_struct *p;
100 
101 		/*
102 		 * Skipping the current CPU is OK even through we can be
103 		 * migrated at any point. The current CPU, at the point
104 		 * where we read raw_smp_processor_id(), is ensured to
105 		 * be in program order with respect to the caller
106 		 * thread. Therefore, we can skip this CPU from the
107 		 * iteration.
108 		 */
109 		if (cpu == raw_smp_processor_id())
110 			continue;
111 
112 		if (!(READ_ONCE(cpu_rq(cpu)->membarrier_state) &
113 		    MEMBARRIER_STATE_GLOBAL_EXPEDITED))
114 			continue;
115 
116 		/*
117 		 * Skip the CPU if it runs a kernel thread. The scheduler
118 		 * leaves the prior task mm in place as an optimization when
119 		 * scheduling a kthread.
120 		 */
121 		p = rcu_dereference(cpu_rq(cpu)->curr);
122 		if (p->flags & PF_KTHREAD)
123 			continue;
124 
125 		__cpumask_set_cpu(cpu, tmpmask);
126 	}
127 	rcu_read_unlock();
128 
129 	preempt_disable();
130 	smp_call_function_many(tmpmask, ipi_mb, NULL, 1);
131 	preempt_enable();
132 
133 	free_cpumask_var(tmpmask);
134 	cpus_read_unlock();
135 
136 	/*
137 	 * Memory barrier on the caller thread _after_ we finished
138 	 * waiting for the last IPI. Matches memory barriers around
139 	 * rq->curr modification in scheduler.
140 	 */
141 	smp_mb();	/* exit from system call is not a mb */
142 	return 0;
143 }
144 
145 static int membarrier_private_expedited(int flags, int cpu_id)
146 {
147 	cpumask_var_t tmpmask;
148 	struct mm_struct *mm = current->mm;
149 	smp_call_func_t ipi_func = ipi_mb;
150 
151 	if (flags == MEMBARRIER_FLAG_SYNC_CORE) {
152 		if (!IS_ENABLED(CONFIG_ARCH_HAS_MEMBARRIER_SYNC_CORE))
153 			return -EINVAL;
154 		if (!(atomic_read(&mm->membarrier_state) &
155 		      MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE_READY))
156 			return -EPERM;
157 	} else if (flags == MEMBARRIER_FLAG_RSEQ) {
158 		if (!IS_ENABLED(CONFIG_RSEQ))
159 			return -EINVAL;
160 		if (!(atomic_read(&mm->membarrier_state) &
161 		      MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ_READY))
162 			return -EPERM;
163 		ipi_func = ipi_rseq;
164 	} else {
165 		WARN_ON_ONCE(flags);
166 		if (!(atomic_read(&mm->membarrier_state) &
167 		      MEMBARRIER_STATE_PRIVATE_EXPEDITED_READY))
168 			return -EPERM;
169 	}
170 
171 	if (atomic_read(&mm->mm_users) == 1 || num_online_cpus() == 1)
172 		return 0;
173 
174 	/*
175 	 * Matches memory barriers around rq->curr modification in
176 	 * scheduler.
177 	 */
178 	smp_mb();	/* system call entry is not a mb. */
179 
180 	if (cpu_id < 0 && !zalloc_cpumask_var(&tmpmask, GFP_KERNEL))
181 		return -ENOMEM;
182 
183 	cpus_read_lock();
184 
185 	if (cpu_id >= 0) {
186 		struct task_struct *p;
187 
188 		if (cpu_id >= nr_cpu_ids || !cpu_online(cpu_id))
189 			goto out;
190 		if (cpu_id == raw_smp_processor_id())
191 			goto out;
192 		rcu_read_lock();
193 		p = rcu_dereference(cpu_rq(cpu_id)->curr);
194 		if (!p || p->mm != mm) {
195 			rcu_read_unlock();
196 			goto out;
197 		}
198 		rcu_read_unlock();
199 	} else {
200 		int cpu;
201 
202 		rcu_read_lock();
203 		for_each_online_cpu(cpu) {
204 			struct task_struct *p;
205 
206 			/*
207 			 * Skipping the current CPU is OK even through we can be
208 			 * migrated at any point. The current CPU, at the point
209 			 * where we read raw_smp_processor_id(), is ensured to
210 			 * be in program order with respect to the caller
211 			 * thread. Therefore, we can skip this CPU from the
212 			 * iteration.
213 			 */
214 			if (cpu == raw_smp_processor_id())
215 				continue;
216 			p = rcu_dereference(cpu_rq(cpu)->curr);
217 			if (p && p->mm == mm)
218 				__cpumask_set_cpu(cpu, tmpmask);
219 		}
220 		rcu_read_unlock();
221 	}
222 
223 	preempt_disable();
224 	if (cpu_id >= 0)
225 		smp_call_function_single(cpu_id, ipi_func, NULL, 1);
226 	else
227 		smp_call_function_many(tmpmask, ipi_func, NULL, 1);
228 	preempt_enable();
229 
230 out:
231 	if (cpu_id < 0)
232 		free_cpumask_var(tmpmask);
233 	cpus_read_unlock();
234 
235 	/*
236 	 * Memory barrier on the caller thread _after_ we finished
237 	 * waiting for the last IPI. Matches memory barriers around
238 	 * rq->curr modification in scheduler.
239 	 */
240 	smp_mb();	/* exit from system call is not a mb */
241 
242 	return 0;
243 }
244 
245 static int sync_runqueues_membarrier_state(struct mm_struct *mm)
246 {
247 	int membarrier_state = atomic_read(&mm->membarrier_state);
248 	cpumask_var_t tmpmask;
249 	int cpu;
250 
251 	if (atomic_read(&mm->mm_users) == 1 || num_online_cpus() == 1) {
252 		this_cpu_write(runqueues.membarrier_state, membarrier_state);
253 
254 		/*
255 		 * For single mm user, we can simply issue a memory barrier
256 		 * after setting MEMBARRIER_STATE_GLOBAL_EXPEDITED in the
257 		 * mm and in the current runqueue to guarantee that no memory
258 		 * access following registration is reordered before
259 		 * registration.
260 		 */
261 		smp_mb();
262 		return 0;
263 	}
264 
265 	if (!zalloc_cpumask_var(&tmpmask, GFP_KERNEL))
266 		return -ENOMEM;
267 
268 	/*
269 	 * For mm with multiple users, we need to ensure all future
270 	 * scheduler executions will observe @mm's new membarrier
271 	 * state.
272 	 */
273 	synchronize_rcu();
274 
275 	/*
276 	 * For each cpu runqueue, if the task's mm match @mm, ensure that all
277 	 * @mm's membarrier state set bits are also set in in the runqueue's
278 	 * membarrier state. This ensures that a runqueue scheduling
279 	 * between threads which are users of @mm has its membarrier state
280 	 * updated.
281 	 */
282 	cpus_read_lock();
283 	rcu_read_lock();
284 	for_each_online_cpu(cpu) {
285 		struct rq *rq = cpu_rq(cpu);
286 		struct task_struct *p;
287 
288 		p = rcu_dereference(rq->curr);
289 		if (p && p->mm == mm)
290 			__cpumask_set_cpu(cpu, tmpmask);
291 	}
292 	rcu_read_unlock();
293 
294 	preempt_disable();
295 	smp_call_function_many(tmpmask, ipi_sync_rq_state, mm, 1);
296 	preempt_enable();
297 
298 	free_cpumask_var(tmpmask);
299 	cpus_read_unlock();
300 
301 	return 0;
302 }
303 
304 static int membarrier_register_global_expedited(void)
305 {
306 	struct task_struct *p = current;
307 	struct mm_struct *mm = p->mm;
308 	int ret;
309 
310 	if (atomic_read(&mm->membarrier_state) &
311 	    MEMBARRIER_STATE_GLOBAL_EXPEDITED_READY)
312 		return 0;
313 	atomic_or(MEMBARRIER_STATE_GLOBAL_EXPEDITED, &mm->membarrier_state);
314 	ret = sync_runqueues_membarrier_state(mm);
315 	if (ret)
316 		return ret;
317 	atomic_or(MEMBARRIER_STATE_GLOBAL_EXPEDITED_READY,
318 		  &mm->membarrier_state);
319 
320 	return 0;
321 }
322 
323 static int membarrier_register_private_expedited(int flags)
324 {
325 	struct task_struct *p = current;
326 	struct mm_struct *mm = p->mm;
327 	int ready_state = MEMBARRIER_STATE_PRIVATE_EXPEDITED_READY,
328 	    set_state = MEMBARRIER_STATE_PRIVATE_EXPEDITED,
329 	    ret;
330 
331 	if (flags == MEMBARRIER_FLAG_SYNC_CORE) {
332 		if (!IS_ENABLED(CONFIG_ARCH_HAS_MEMBARRIER_SYNC_CORE))
333 			return -EINVAL;
334 		ready_state =
335 			MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE_READY;
336 	} else if (flags == MEMBARRIER_FLAG_RSEQ) {
337 		if (!IS_ENABLED(CONFIG_RSEQ))
338 			return -EINVAL;
339 		ready_state =
340 			MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ_READY;
341 	} else {
342 		WARN_ON_ONCE(flags);
343 	}
344 
345 	/*
346 	 * We need to consider threads belonging to different thread
347 	 * groups, which use the same mm. (CLONE_VM but not
348 	 * CLONE_THREAD).
349 	 */
350 	if ((atomic_read(&mm->membarrier_state) & ready_state) == ready_state)
351 		return 0;
352 	if (flags & MEMBARRIER_FLAG_SYNC_CORE)
353 		set_state |= MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE;
354 	if (flags & MEMBARRIER_FLAG_RSEQ)
355 		set_state |= MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ;
356 	atomic_or(set_state, &mm->membarrier_state);
357 	ret = sync_runqueues_membarrier_state(mm);
358 	if (ret)
359 		return ret;
360 	atomic_or(ready_state, &mm->membarrier_state);
361 
362 	return 0;
363 }
364 
365 /**
366  * sys_membarrier - issue memory barriers on a set of threads
367  * @cmd:    Takes command values defined in enum membarrier_cmd.
368  * @flags:  Currently needs to be 0 for all commands other than
369  *          MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ: in the latter
370  *          case it can be MEMBARRIER_CMD_FLAG_CPU, indicating that @cpu_id
371  *          contains the CPU on which to interrupt (= restart)
372  *          the RSEQ critical section.
373  * @cpu_id: if @flags == MEMBARRIER_CMD_FLAG_CPU, indicates the cpu on which
374  *          RSEQ CS should be interrupted (@cmd must be
375  *          MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ).
376  *
377  * If this system call is not implemented, -ENOSYS is returned. If the
378  * command specified does not exist, not available on the running
379  * kernel, or if the command argument is invalid, this system call
380  * returns -EINVAL. For a given command, with flags argument set to 0,
381  * if this system call returns -ENOSYS or -EINVAL, it is guaranteed to
382  * always return the same value until reboot. In addition, it can return
383  * -ENOMEM if there is not enough memory available to perform the system
384  * call.
385  *
386  * All memory accesses performed in program order from each targeted thread
387  * is guaranteed to be ordered with respect to sys_membarrier(). If we use
388  * the semantic "barrier()" to represent a compiler barrier forcing memory
389  * accesses to be performed in program order across the barrier, and
390  * smp_mb() to represent explicit memory barriers forcing full memory
391  * ordering across the barrier, we have the following ordering table for
392  * each pair of barrier(), sys_membarrier() and smp_mb():
393  *
394  * The pair ordering is detailed as (O: ordered, X: not ordered):
395  *
396  *                        barrier()   smp_mb() sys_membarrier()
397  *        barrier()          X           X            O
398  *        smp_mb()           X           O            O
399  *        sys_membarrier()   O           O            O
400  */
401 SYSCALL_DEFINE3(membarrier, int, cmd, unsigned int, flags, int, cpu_id)
402 {
403 	switch (cmd) {
404 	case MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ:
405 		if (unlikely(flags && flags != MEMBARRIER_CMD_FLAG_CPU))
406 			return -EINVAL;
407 		break;
408 	default:
409 		if (unlikely(flags))
410 			return -EINVAL;
411 	}
412 
413 	if (!(flags & MEMBARRIER_CMD_FLAG_CPU))
414 		cpu_id = -1;
415 
416 	switch (cmd) {
417 	case MEMBARRIER_CMD_QUERY:
418 	{
419 		int cmd_mask = MEMBARRIER_CMD_BITMASK;
420 
421 		if (tick_nohz_full_enabled())
422 			cmd_mask &= ~MEMBARRIER_CMD_GLOBAL;
423 		return cmd_mask;
424 	}
425 	case MEMBARRIER_CMD_GLOBAL:
426 		/* MEMBARRIER_CMD_GLOBAL is not compatible with nohz_full. */
427 		if (tick_nohz_full_enabled())
428 			return -EINVAL;
429 		if (num_online_cpus() > 1)
430 			synchronize_rcu();
431 		return 0;
432 	case MEMBARRIER_CMD_GLOBAL_EXPEDITED:
433 		return membarrier_global_expedited();
434 	case MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED:
435 		return membarrier_register_global_expedited();
436 	case MEMBARRIER_CMD_PRIVATE_EXPEDITED:
437 		return membarrier_private_expedited(0, cpu_id);
438 	case MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED:
439 		return membarrier_register_private_expedited(0);
440 	case MEMBARRIER_CMD_PRIVATE_EXPEDITED_SYNC_CORE:
441 		return membarrier_private_expedited(MEMBARRIER_FLAG_SYNC_CORE, cpu_id);
442 	case MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_SYNC_CORE:
443 		return membarrier_register_private_expedited(MEMBARRIER_FLAG_SYNC_CORE);
444 	case MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ:
445 		return membarrier_private_expedited(MEMBARRIER_FLAG_RSEQ, cpu_id);
446 	case MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_RSEQ:
447 		return membarrier_register_private_expedited(MEMBARRIER_FLAG_RSEQ);
448 	default:
449 		return -EINVAL;
450 	}
451 }
452