blob: 084c8b3a26812d78e3eb2da63d46091e491a7e23 [file] [log] [blame]
Jens Axboe3d442232008-06-26 11:21:34 +02001/*
2 * Generic helpers for smp ipi calls
3 *
4 * (C) Jens Axboe <[email protected]> 2008
Jens Axboe3d442232008-06-26 11:21:34 +02005 */
Michael Ellermanca7dfdb2016-10-26 16:37:53 +11006
7#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
8
Frederic Weisbecker47885012014-05-08 01:37:48 +02009#include <linux/irq_work.h>
Jens Axboe3d442232008-06-26 11:21:34 +020010#include <linux/rcupdate.h>
Linus Torvalds59190f4212008-07-15 14:02:33 -070011#include <linux/rculist.h>
Ingo Molnar641cd4c2009-03-13 10:47:34 +010012#include <linux/kernel.h>
Paul Gortmaker9984de12011-05-23 14:51:41 -040013#include <linux/export.h>
Ingo Molnar0b13fda2009-02-25 16:52:11 +010014#include <linux/percpu.h>
15#include <linux/init.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090016#include <linux/gfp.h>
Jens Axboe3d442232008-06-26 11:21:34 +020017#include <linux/smp.h>
Peter Zijlstra8969a5e2009-02-25 13:59:47 +010018#include <linux/cpu.h>
Chuansheng Liuc6f44592014-09-04 15:17:54 +080019#include <linux/sched.h>
Ingo Molnar4c822692017-02-01 16:36:40 +010020#include <linux/sched/idle.h>
Juergen Gross47ae4b02016-08-29 08:48:43 +020021#include <linux/hypervisor.h>
Jens Axboe3d442232008-06-26 11:21:34 +020022
Suresh Siddha3bb5d2e2012-04-20 17:08:50 -070023#include "smpboot.h"
24
Jens Axboe3d442232008-06-26 11:21:34 +020025enum {
Peter Zijlstra6e275632009-02-25 13:59:48 +010026 CSD_FLAG_LOCK = 0x01,
Linus Torvalds80538712015-02-11 12:42:10 -080027 CSD_FLAG_SYNCHRONOUS = 0x02,
Jens Axboe3d442232008-06-26 11:21:34 +020028};
29
30struct call_function_data {
Ying Huang966a9672017-08-08 12:30:00 +080031 call_single_data_t __percpu *csd;
Ingo Molnar0b13fda2009-02-25 16:52:11 +010032 cpumask_var_t cpumask;
Aaron Lu3fc5b3b2017-05-19 15:53:31 +080033 cpumask_var_t cpumask_ipi;
Jens Axboe3d442232008-06-26 11:21:34 +020034};
35
Milton Millere03bcb62010-01-18 13:00:51 +110036static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data, cfd_data);
37
Christoph Hellwig6897fc22014-01-30 15:45:47 -080038static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue);
Peter Zijlstra8969a5e2009-02-25 13:59:47 +010039
Srivatsa S. Bhat8d056c42014-06-23 13:22:02 -070040static void flush_smp_call_function_queue(bool warn_cpu_offline);
41
Richard Weinberger31487f82016-07-13 17:17:01 +000042int smpcfd_prepare_cpu(unsigned int cpu)
Peter Zijlstra8969a5e2009-02-25 13:59:47 +010043{
Peter Zijlstra8969a5e2009-02-25 13:59:47 +010044 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
45
Richard Weinberger31487f82016-07-13 17:17:01 +000046 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
47 cpu_to_node(cpu)))
48 return -ENOMEM;
Aaron Lu3fc5b3b2017-05-19 15:53:31 +080049 if (!zalloc_cpumask_var_node(&cfd->cpumask_ipi, GFP_KERNEL,
50 cpu_to_node(cpu))) {
51 free_cpumask_var(cfd->cpumask);
52 return -ENOMEM;
53 }
Ying Huang966a9672017-08-08 12:30:00 +080054 cfd->csd = alloc_percpu(call_single_data_t);
Richard Weinberger31487f82016-07-13 17:17:01 +000055 if (!cfd->csd) {
Peter Zijlstra8969a5e2009-02-25 13:59:47 +010056 free_cpumask_var(cfd->cpumask);
Aaron Lu3fc5b3b2017-05-19 15:53:31 +080057 free_cpumask_var(cfd->cpumask_ipi);
Richard Weinberger31487f82016-07-13 17:17:01 +000058 return -ENOMEM;
59 }
Srivatsa S. Bhat8d056c42014-06-23 13:22:02 -070060
Richard Weinberger31487f82016-07-13 17:17:01 +000061 return 0;
Peter Zijlstra8969a5e2009-02-25 13:59:47 +010062}
63
Richard Weinberger31487f82016-07-13 17:17:01 +000064int smpcfd_dead_cpu(unsigned int cpu)
65{
66 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
67
68 free_cpumask_var(cfd->cpumask);
Aaron Lu3fc5b3b2017-05-19 15:53:31 +080069 free_cpumask_var(cfd->cpumask_ipi);
Richard Weinberger31487f82016-07-13 17:17:01 +000070 free_percpu(cfd->csd);
71 return 0;
72}
73
74int smpcfd_dying_cpu(unsigned int cpu)
75{
76 /*
77 * The IPIs for the smp-call-function callbacks queued by other
78 * CPUs might arrive late, either due to hardware latencies or
79 * because this CPU disabled interrupts (inside stop-machine)
80 * before the IPIs were sent. So flush out any pending callbacks
81 * explicitly (without waiting for the IPIs to arrive), to
82 * ensure that the outgoing CPU doesn't go offline with work
83 * still pending.
84 */
85 flush_smp_call_function_queue(false);
86 return 0;
87}
Peter Zijlstra8969a5e2009-02-25 13:59:47 +010088
Takao Indohd8ad7d12011-03-29 12:35:04 -040089void __init call_function_init(void)
Jens Axboe3d442232008-06-26 11:21:34 +020090{
91 int i;
92
Christoph Hellwig6897fc22014-01-30 15:45:47 -080093 for_each_possible_cpu(i)
94 init_llist_head(&per_cpu(call_single_queue, i));
Peter Zijlstra8969a5e2009-02-25 13:59:47 +010095
Richard Weinberger31487f82016-07-13 17:17:01 +000096 smpcfd_prepare_cpu(smp_processor_id());
Jens Axboe3d442232008-06-26 11:21:34 +020097}
98
Peter Zijlstra8969a5e2009-02-25 13:59:47 +010099/*
Peter Zijlstra8969a5e2009-02-25 13:59:47 +0100100 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
101 *
Ingo Molnar0b13fda2009-02-25 16:52:11 +0100102 * For non-synchronous ipi calls the csd can still be in use by the
103 * previous function call. For multi-cpu calls its even more interesting
104 * as we'll have to ensure no other cpu is observing our csd.
Peter Zijlstra8969a5e2009-02-25 13:59:47 +0100105 */
Ying Huang966a9672017-08-08 12:30:00 +0800106static __always_inline void csd_lock_wait(call_single_data_t *csd)
Peter Zijlstra8969a5e2009-02-25 13:59:47 +0100107{
Peter Zijlstra1f03e8d2016-04-04 10:57:12 +0200108 smp_cond_load_acquire(&csd->flags, !(VAL & CSD_FLAG_LOCK));
Peter Zijlstra6e275632009-02-25 13:59:48 +0100109}
110
Ying Huang966a9672017-08-08 12:30:00 +0800111static __always_inline void csd_lock(call_single_data_t *csd)
Peter Zijlstra6e275632009-02-25 13:59:48 +0100112{
Andrew Mortone1d12f32013-04-30 15:27:28 -0700113 csd_lock_wait(csd);
114 csd->flags |= CSD_FLAG_LOCK;
Peter Zijlstra8969a5e2009-02-25 13:59:47 +0100115
116 /*
Ingo Molnar0b13fda2009-02-25 16:52:11 +0100117 * prevent CPU from reordering the above assignment
118 * to ->flags with any subsequent assignments to other
Ying Huang966a9672017-08-08 12:30:00 +0800119 * fields of the specified call_single_data_t structure:
Peter Zijlstra8969a5e2009-02-25 13:59:47 +0100120 */
Linus Torvalds80538712015-02-11 12:42:10 -0800121 smp_wmb();
Peter Zijlstra8969a5e2009-02-25 13:59:47 +0100122}
123
Ying Huang966a9672017-08-08 12:30:00 +0800124static __always_inline void csd_unlock(call_single_data_t *csd)
Peter Zijlstra8969a5e2009-02-25 13:59:47 +0100125{
Linus Torvalds80538712015-02-11 12:42:10 -0800126 WARN_ON(!(csd->flags & CSD_FLAG_LOCK));
Ingo Molnar0b13fda2009-02-25 16:52:11 +0100127
Peter Zijlstra8969a5e2009-02-25 13:59:47 +0100128 /*
Ingo Molnar0b13fda2009-02-25 16:52:11 +0100129 * ensure we're all done before releasing data:
Peter Zijlstra8969a5e2009-02-25 13:59:47 +0100130 */
Linus Torvalds80538712015-02-11 12:42:10 -0800131 smp_store_release(&csd->flags, 0);
Jens Axboe3d442232008-06-26 11:21:34 +0200132}
133
Ying Huang966a9672017-08-08 12:30:00 +0800134static DEFINE_PER_CPU_SHARED_ALIGNED(call_single_data_t, csd_data);
Frederic Weisbecker8b284992014-02-24 16:39:58 +0100135
Jens Axboe3d442232008-06-26 11:21:34 +0200136/*
Ying Huang966a9672017-08-08 12:30:00 +0800137 * Insert a previously allocated call_single_data_t element
Ingo Molnar0b13fda2009-02-25 16:52:11 +0100138 * for execution on the given CPU. data must already have
139 * ->func, ->info, and ->flags set.
Jens Axboe3d442232008-06-26 11:21:34 +0200140 */
Ying Huang966a9672017-08-08 12:30:00 +0800141static int generic_exec_single(int cpu, call_single_data_t *csd,
Linus Torvalds80538712015-02-11 12:42:10 -0800142 smp_call_func_t func, void *info)
Jens Axboe3d442232008-06-26 11:21:34 +0200143{
Frederic Weisbecker8b284992014-02-24 16:39:58 +0100144 if (cpu == smp_processor_id()) {
Linus Torvalds80538712015-02-11 12:42:10 -0800145 unsigned long flags;
146
147 /*
148 * We can unlock early even for the synchronous on-stack case,
149 * since we're doing this from the same CPU..
150 */
151 csd_unlock(csd);
Frederic Weisbecker8b284992014-02-24 16:39:58 +0100152 local_irq_save(flags);
153 func(info);
154 local_irq_restore(flags);
155 return 0;
156 }
157
158
Linus Torvalds5224b962015-04-19 04:56:03 -0400159 if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu)) {
160 csd_unlock(csd);
Frederic Weisbecker8b284992014-02-24 16:39:58 +0100161 return -ENXIO;
Linus Torvalds5224b962015-04-19 04:56:03 -0400162 }
Frederic Weisbecker8b284992014-02-24 16:39:58 +0100163
Frederic Weisbecker8b284992014-02-24 16:39:58 +0100164 csd->func = func;
165 csd->info = info;
166
Suresh Siddha561920a02008-10-30 18:28:41 +0100167 /*
Nick Piggin15d0d3b2009-02-25 06:22:45 +0100168 * The list addition should be visible before sending the IPI
169 * handler locks the list to pull the entry off it because of
170 * normal cache coherency rules implied by spinlocks.
171 *
172 * If IPIs can go out of order to the cache coherency protocol
173 * in an architecture, sufficient synchronisation should be added
174 * to arch code to make it appear to obey cache coherency WRT
Ingo Molnar0b13fda2009-02-25 16:52:11 +0100175 * locking and barrier primitives. Generic code isn't really
176 * equipped to do the right thing...
Suresh Siddha561920a02008-10-30 18:28:41 +0100177 */
Christoph Hellwig6897fc22014-01-30 15:45:47 -0800178 if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
Jens Axboe3d442232008-06-26 11:21:34 +0200179 arch_send_call_function_single_ipi(cpu);
180
Frederic Weisbecker8b284992014-02-24 16:39:58 +0100181 return 0;
Jens Axboe3d442232008-06-26 11:21:34 +0200182}
183
Srivatsa S. Bhat8d056c42014-06-23 13:22:02 -0700184/**
185 * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks
186 *
187 * Invoked by arch to handle an IPI for call function single.
188 * Must be called with interrupts disabled.
Jens Axboe3d442232008-06-26 11:21:34 +0200189 */
190void generic_smp_call_function_single_interrupt(void)
191{
Srivatsa S. Bhat8d056c42014-06-23 13:22:02 -0700192 flush_smp_call_function_queue(true);
193}
194
195/**
196 * flush_smp_call_function_queue - Flush pending smp-call-function callbacks
197 *
198 * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an
199 * offline CPU. Skip this check if set to 'false'.
200 *
201 * Flush any pending smp-call-function callbacks queued on this CPU. This is
202 * invoked by the generic IPI handler, as well as by a CPU about to go offline,
203 * to ensure that all pending IPI callbacks are run before it goes completely
204 * offline.
205 *
206 * Loop through the call_single_queue and run all the queued callbacks.
207 * Must be called with interrupts disabled.
208 */
209static void flush_smp_call_function_queue(bool warn_cpu_offline)
210{
211 struct llist_head *head;
Jan Kara5fd77592014-02-24 16:39:55 +0100212 struct llist_node *entry;
Ying Huang966a9672017-08-08 12:30:00 +0800213 call_single_data_t *csd, *csd_next;
Srivatsa S. Bhata219ccf2014-06-06 14:37:05 -0700214 static bool warned;
215
Frederic Weisbecker83efcbd2017-11-06 16:01:22 +0100216 lockdep_assert_irqs_disabled();
Srivatsa S. Bhat8d056c42014-06-23 13:22:02 -0700217
Christoph Lameterbb964a92014-08-17 12:30:24 -0500218 head = this_cpu_ptr(&call_single_queue);
Srivatsa S. Bhat8d056c42014-06-23 13:22:02 -0700219 entry = llist_del_all(head);
Srivatsa S. Bhata219ccf2014-06-06 14:37:05 -0700220 entry = llist_reverse_order(entry);
Jens Axboe3d442232008-06-26 11:21:34 +0200221
Srivatsa S. Bhat8d056c42014-06-23 13:22:02 -0700222 /* There shouldn't be any pending callbacks on an offline CPU. */
223 if (unlikely(warn_cpu_offline && !cpu_online(smp_processor_id()) &&
224 !warned && !llist_empty(head))) {
Srivatsa S. Bhata219ccf2014-06-06 14:37:05 -0700225 warned = true;
226 WARN(1, "IPI on offline CPU %d\n", smp_processor_id());
Suresh Siddha269c8612009-08-19 18:05:35 -0700227
Srivatsa S. Bhata219ccf2014-06-06 14:37:05 -0700228 /*
229 * We don't have to use the _safe() variant here
230 * because we are not invoking the IPI handlers yet.
231 */
232 llist_for_each_entry(csd, entry, llist)
233 pr_warn("IPI callback %pS sent to offline CPU\n",
234 csd->func);
235 }
Jens Axboe3d442232008-06-26 11:21:34 +0200236
Jan Kara5fd77592014-02-24 16:39:55 +0100237 llist_for_each_entry_safe(csd, csd_next, entry, llist) {
Linus Torvalds80538712015-02-11 12:42:10 -0800238 smp_call_func_t func = csd->func;
239 void *info = csd->info;
240
241 /* Do we wait until *after* callback? */
242 if (csd->flags & CSD_FLAG_SYNCHRONOUS) {
243 func(info);
244 csd_unlock(csd);
245 } else {
246 csd_unlock(csd);
247 func(info);
248 }
Jens Axboe3d442232008-06-26 11:21:34 +0200249 }
Frederic Weisbecker47885012014-05-08 01:37:48 +0200250
251 /*
252 * Handle irq works queued remotely by irq_work_queue_on().
253 * Smp functions above are typically synchronous so they
254 * better run first since some other CPUs may be busy waiting
255 * for them.
256 */
257 irq_work_run();
Jens Axboe3d442232008-06-26 11:21:34 +0200258}
259
260/*
261 * smp_call_function_single - Run a function on a specific CPU
262 * @func: The function to run. This must be fast and non-blocking.
263 * @info: An arbitrary pointer to pass to the function.
Jens Axboe3d442232008-06-26 11:21:34 +0200264 * @wait: If true, wait until function has completed on other CPUs.
265 *
Sheng Yang72f279b2009-10-22 19:19:34 +0800266 * Returns 0 on success, else a negative status code.
Jens Axboe3d442232008-06-26 11:21:34 +0200267 */
David Howells3a5f65df2010-10-27 17:28:36 +0100268int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
Jens Axboe8691e5a2008-06-06 11:18:06 +0200269 int wait)
Jens Axboe3d442232008-06-26 11:21:34 +0200270{
Ying Huang966a9672017-08-08 12:30:00 +0800271 call_single_data_t *csd;
272 call_single_data_t csd_stack = {
273 .flags = CSD_FLAG_LOCK | CSD_FLAG_SYNCHRONOUS,
274 };
Ingo Molnar0b13fda2009-02-25 16:52:11 +0100275 int this_cpu;
Frederic Weisbecker8b284992014-02-24 16:39:58 +0100276 int err;
Jens Axboe3d442232008-06-26 11:21:34 +0200277
Ingo Molnar0b13fda2009-02-25 16:52:11 +0100278 /*
279 * prevent preemption and reschedule on another processor,
280 * as well as CPU removal
281 */
282 this_cpu = get_cpu();
283
Suresh Siddha269c8612009-08-19 18:05:35 -0700284 /*
285 * Can deadlock when called with interrupts disabled.
286 * We allow cpu's that are not yet online though, as no one else can
287 * send smp call function interrupt to this cpu and as such deadlocks
288 * can't happen.
289 */
290 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
291 && !oops_in_progress);
Jens Axboe3d442232008-06-26 11:21:34 +0200292
Linus Torvalds80538712015-02-11 12:42:10 -0800293 csd = &csd_stack;
294 if (!wait) {
295 csd = this_cpu_ptr(&csd_data);
296 csd_lock(csd);
297 }
298
299 err = generic_exec_single(cpu, csd, func, info);
300
301 if (wait)
302 csd_lock_wait(csd);
Jens Axboe3d442232008-06-26 11:21:34 +0200303
304 put_cpu();
Ingo Molnar0b13fda2009-02-25 16:52:11 +0100305
H. Peter Anvinf73be6de2008-08-25 17:07:14 -0700306 return err;
Jens Axboe3d442232008-06-26 11:21:34 +0200307}
308EXPORT_SYMBOL(smp_call_function_single);
309
Frederic Weisbeckerd7877c02014-02-24 16:39:59 +0100310/**
Frederic Weisbeckerc46fff22014-02-24 16:40:02 +0100311 * smp_call_function_single_async(): Run an asynchronous function on a
312 * specific CPU.
Frederic Weisbeckerd7877c02014-02-24 16:39:59 +0100313 * @cpu: The CPU to run on.
314 * @csd: Pre-allocated and setup data structure
Frederic Weisbeckerd7877c02014-02-24 16:39:59 +0100315 *
Frederic Weisbeckerc46fff22014-02-24 16:40:02 +0100316 * Like smp_call_function_single(), but the call is asynchonous and
317 * can thus be done from contexts with disabled interrupts.
318 *
319 * The caller passes his own pre-allocated data structure
320 * (ie: embedded in an object) and is responsible for synchronizing it
321 * such that the IPIs performed on the @csd are strictly serialized.
322 *
323 * NOTE: Be careful, there is unfortunately no current debugging facility to
324 * validate the correctness of this serialization.
Frederic Weisbeckerd7877c02014-02-24 16:39:59 +0100325 */
Ying Huang966a9672017-08-08 12:30:00 +0800326int smp_call_function_single_async(int cpu, call_single_data_t *csd)
Frederic Weisbeckerd7877c02014-02-24 16:39:59 +0100327{
328 int err = 0;
Frederic Weisbeckerd7877c02014-02-24 16:39:59 +0100329
Frederic Weisbeckerfce8ad12014-02-24 16:40:01 +0100330 preempt_disable();
Linus Torvalds80538712015-02-11 12:42:10 -0800331
332 /* We could deadlock if we have to wait here with interrupts disabled! */
333 if (WARN_ON_ONCE(csd->flags & CSD_FLAG_LOCK))
334 csd_lock_wait(csd);
335
336 csd->flags = CSD_FLAG_LOCK;
337 smp_wmb();
338
339 err = generic_exec_single(cpu, csd, csd->func, csd->info);
Frederic Weisbeckerfce8ad12014-02-24 16:40:01 +0100340 preempt_enable();
Frederic Weisbeckerd7877c02014-02-24 16:39:59 +0100341
342 return err;
343}
Frederic Weisbeckerc46fff22014-02-24 16:40:02 +0100344EXPORT_SYMBOL_GPL(smp_call_function_single_async);
Frederic Weisbeckerd7877c02014-02-24 16:39:59 +0100345
Rusty Russell2ea6dec2009-11-17 14:27:27 -0800346/*
347 * smp_call_function_any - Run a function on any of the given cpus
348 * @mask: The mask of cpus it can run on.
349 * @func: The function to run. This must be fast and non-blocking.
350 * @info: An arbitrary pointer to pass to the function.
351 * @wait: If true, wait until function has completed.
352 *
353 * Returns 0 on success, else a negative status code (if no cpus were online).
Rusty Russell2ea6dec2009-11-17 14:27:27 -0800354 *
355 * Selection preference:
356 * 1) current cpu if in @mask
357 * 2) any cpu of current node if in @mask
358 * 3) any other online cpu in @mask
359 */
360int smp_call_function_any(const struct cpumask *mask,
David Howells3a5f65df2010-10-27 17:28:36 +0100361 smp_call_func_t func, void *info, int wait)
Rusty Russell2ea6dec2009-11-17 14:27:27 -0800362{
363 unsigned int cpu;
364 const struct cpumask *nodemask;
365 int ret;
366
367 /* Try for same CPU (cheapest) */
368 cpu = get_cpu();
369 if (cpumask_test_cpu(cpu, mask))
370 goto call;
371
372 /* Try for same node. */
David Johnaf2422c2010-01-15 17:01:23 -0800373 nodemask = cpumask_of_node(cpu_to_node(cpu));
Rusty Russell2ea6dec2009-11-17 14:27:27 -0800374 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
375 cpu = cpumask_next_and(cpu, nodemask, mask)) {
376 if (cpu_online(cpu))
377 goto call;
378 }
379
380 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
381 cpu = cpumask_any_and(mask, cpu_online_mask);
382call:
383 ret = smp_call_function_single(cpu, func, info, wait);
384 put_cpu();
385 return ret;
386}
387EXPORT_SYMBOL_GPL(smp_call_function_any);
388
Jens Axboe3d442232008-06-26 11:21:34 +0200389/**
Rusty Russell54b11e62008-12-30 09:05:16 +1030390 * smp_call_function_many(): Run a function on a set of other CPUs.
391 * @mask: The set of cpus to run on (only runs on online subset).
Jens Axboe3d442232008-06-26 11:21:34 +0200392 * @func: The function to run. This must be fast and non-blocking.
393 * @info: An arbitrary pointer to pass to the function.
Ingo Molnar0b13fda2009-02-25 16:52:11 +0100394 * @wait: If true, wait (atomically) until function has completed
395 * on other CPUs.
Jens Axboe3d442232008-06-26 11:21:34 +0200396 *
Sheng Yang72f279b2009-10-22 19:19:34 +0800397 * If @wait is true, then returns once @func has returned.
Jens Axboe3d442232008-06-26 11:21:34 +0200398 *
399 * You must not call this function with disabled interrupts or from a
400 * hardware interrupt handler or from a bottom half handler. Preemption
401 * must be disabled when calling this function.
402 */
Rusty Russell54b11e62008-12-30 09:05:16 +1030403void smp_call_function_many(const struct cpumask *mask,
David Howells3a5f65df2010-10-27 17:28:36 +0100404 smp_call_func_t func, void *info, bool wait)
Jens Axboe3d442232008-06-26 11:21:34 +0200405{
Andrew Mortone1d12f32013-04-30 15:27:28 -0700406 struct call_function_data *cfd;
Shaohua Li9a46ad62013-02-21 16:43:03 -0800407 int cpu, next_cpu, this_cpu = smp_processor_id();
Jens Axboe3d442232008-06-26 11:21:34 +0200408
Suresh Siddha269c8612009-08-19 18:05:35 -0700409 /*
410 * Can deadlock when called with interrupts disabled.
411 * We allow cpu's that are not yet online though, as no one else can
412 * send smp call function interrupt to this cpu and as such deadlocks
413 * can't happen.
414 */
415 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
Tejun Heobd924e82011-01-20 12:07:13 +0100416 && !oops_in_progress && !early_boot_irqs_disabled);
Jens Axboe3d442232008-06-26 11:21:34 +0200417
Milton Miller723aae22011-03-15 13:27:17 -0600418 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
Rusty Russell54b11e62008-12-30 09:05:16 +1030419 cpu = cpumask_first_and(mask, cpu_online_mask);
Ingo Molnar0b13fda2009-02-25 16:52:11 +0100420 if (cpu == this_cpu)
Rusty Russell54b11e62008-12-30 09:05:16 +1030421 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
Ingo Molnar0b13fda2009-02-25 16:52:11 +0100422
Rusty Russell54b11e62008-12-30 09:05:16 +1030423 /* No online cpus? We're done. */
424 if (cpu >= nr_cpu_ids)
425 return;
Jens Axboe3d442232008-06-26 11:21:34 +0200426
Rusty Russell54b11e62008-12-30 09:05:16 +1030427 /* Do we have another CPU which isn't us? */
428 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
Ingo Molnar0b13fda2009-02-25 16:52:11 +0100429 if (next_cpu == this_cpu)
Rusty Russell54b11e62008-12-30 09:05:16 +1030430 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
431
432 /* Fastpath: do that cpu by itself. */
433 if (next_cpu >= nr_cpu_ids) {
434 smp_call_function_single(cpu, func, info, wait);
435 return;
Jens Axboe3d442232008-06-26 11:21:34 +0200436 }
437
Christoph Lameterbb964a92014-08-17 12:30:24 -0500438 cfd = this_cpu_ptr(&cfd_data);
Milton Miller45a57912011-03-15 13:27:16 -0600439
Andrew Mortone1d12f32013-04-30 15:27:28 -0700440 cpumask_and(cfd->cpumask, mask, cpu_online_mask);
Peter Zijlstra6c8557b2017-05-19 12:58:25 +0200441 __cpumask_clear_cpu(this_cpu, cfd->cpumask);
Milton Miller723aae22011-03-15 13:27:17 -0600442
443 /* Some callers race with other cpus changing the passed mask */
Andrew Mortone1d12f32013-04-30 15:27:28 -0700444 if (unlikely(!cpumask_weight(cfd->cpumask)))
Milton Miller723aae22011-03-15 13:27:17 -0600445 return;
Anton Blanchard6dc19892011-01-20 14:44:33 -0800446
Aaron Lu3fc5b3b2017-05-19 15:53:31 +0800447 cpumask_clear(cfd->cpumask_ipi);
Andrew Mortone1d12f32013-04-30 15:27:28 -0700448 for_each_cpu(cpu, cfd->cpumask) {
Ying Huang966a9672017-08-08 12:30:00 +0800449 call_single_data_t *csd = per_cpu_ptr(cfd->csd, cpu);
Shaohua Li9a46ad62013-02-21 16:43:03 -0800450
451 csd_lock(csd);
Linus Torvalds80538712015-02-11 12:42:10 -0800452 if (wait)
453 csd->flags |= CSD_FLAG_SYNCHRONOUS;
Shaohua Li9a46ad62013-02-21 16:43:03 -0800454 csd->func = func;
455 csd->info = info;
Aaron Lu3fc5b3b2017-05-19 15:53:31 +0800456 if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
Peter Zijlstra6c8557b2017-05-19 12:58:25 +0200457 __cpumask_set_cpu(cpu, cfd->cpumask_ipi);
Shaohua Li9a46ad62013-02-21 16:43:03 -0800458 }
Suresh Siddha561920a02008-10-30 18:28:41 +0100459
Jens Axboe3d442232008-06-26 11:21:34 +0200460 /* Send a message to all CPUs in the map */
Aaron Lu3fc5b3b2017-05-19 15:53:31 +0800461 arch_send_call_function_ipi_mask(cfd->cpumask_ipi);
Jens Axboe3d442232008-06-26 11:21:34 +0200462
Shaohua Li9a46ad62013-02-21 16:43:03 -0800463 if (wait) {
Andrew Mortone1d12f32013-04-30 15:27:28 -0700464 for_each_cpu(cpu, cfd->cpumask) {
Ying Huang966a9672017-08-08 12:30:00 +0800465 call_single_data_t *csd;
Andrew Mortone1d12f32013-04-30 15:27:28 -0700466
467 csd = per_cpu_ptr(cfd->csd, cpu);
Shaohua Li9a46ad62013-02-21 16:43:03 -0800468 csd_lock_wait(csd);
469 }
470 }
Jens Axboe3d442232008-06-26 11:21:34 +0200471}
Rusty Russell54b11e62008-12-30 09:05:16 +1030472EXPORT_SYMBOL(smp_call_function_many);
Jens Axboe3d442232008-06-26 11:21:34 +0200473
474/**
475 * smp_call_function(): Run a function on all other CPUs.
476 * @func: The function to run. This must be fast and non-blocking.
477 * @info: An arbitrary pointer to pass to the function.
Ingo Molnar0b13fda2009-02-25 16:52:11 +0100478 * @wait: If true, wait (atomically) until function has completed
479 * on other CPUs.
Jens Axboe3d442232008-06-26 11:21:34 +0200480 *
Rusty Russell54b11e62008-12-30 09:05:16 +1030481 * Returns 0.
Jens Axboe3d442232008-06-26 11:21:34 +0200482 *
483 * If @wait is true, then returns once @func has returned; otherwise
Sheng Yang72f279b2009-10-22 19:19:34 +0800484 * it returns just before the target cpu calls @func.
Jens Axboe3d442232008-06-26 11:21:34 +0200485 *
486 * You must not call this function with disabled interrupts or from a
487 * hardware interrupt handler or from a bottom half handler.
488 */
David Howells3a5f65df2010-10-27 17:28:36 +0100489int smp_call_function(smp_call_func_t func, void *info, int wait)
Jens Axboe3d442232008-06-26 11:21:34 +0200490{
Jens Axboe3d442232008-06-26 11:21:34 +0200491 preempt_disable();
Rusty Russell54b11e62008-12-30 09:05:16 +1030492 smp_call_function_many(cpu_online_mask, func, info, wait);
Jens Axboe3d442232008-06-26 11:21:34 +0200493 preempt_enable();
Ingo Molnar0b13fda2009-02-25 16:52:11 +0100494
Rusty Russell54b11e62008-12-30 09:05:16 +1030495 return 0;
Jens Axboe3d442232008-06-26 11:21:34 +0200496}
497EXPORT_SYMBOL(smp_call_function);
Amerigo Wang351f8f82011-01-12 16:59:39 -0800498
Amerigo Wang34db18a02011-03-22 16:34:06 -0700499/* Setup configured maximum number of CPUs to activate */
500unsigned int setup_max_cpus = NR_CPUS;
501EXPORT_SYMBOL(setup_max_cpus);
502
503
504/*
505 * Setup routine for controlling SMP activation
506 *
507 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
508 * activation entirely (the MPS table probe still happens, though).
509 *
510 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
511 * greater than 0, limits the maximum number of CPUs activated in
512 * SMP mode to <NUM>.
513 */
514
515void __weak arch_disable_smp_support(void) { }
516
517static int __init nosmp(char *str)
518{
519 setup_max_cpus = 0;
520 arch_disable_smp_support();
521
522 return 0;
523}
524
525early_param("nosmp", nosmp);
526
527/* this is hard limit */
528static int __init nrcpus(char *str)
529{
530 int nr_cpus;
531
532 get_option(&str, &nr_cpus);
533 if (nr_cpus > 0 && nr_cpus < nr_cpu_ids)
534 nr_cpu_ids = nr_cpus;
535
536 return 0;
537}
538
539early_param("nr_cpus", nrcpus);
540
541static int __init maxcpus(char *str)
542{
543 get_option(&str, &setup_max_cpus);
544 if (setup_max_cpus == 0)
545 arch_disable_smp_support();
546
547 return 0;
548}
549
550early_param("maxcpus", maxcpus);
551
552/* Setup number of possible processor ids */
Alexey Dobriyan9b130ad2017-09-08 16:14:18 -0700553unsigned int nr_cpu_ids __read_mostly = NR_CPUS;
Amerigo Wang34db18a02011-03-22 16:34:06 -0700554EXPORT_SYMBOL(nr_cpu_ids);
555
556/* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
557void __init setup_nr_cpu_ids(void)
558{
559 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
560}
561
562/* Called by boot processor to activate the rest. */
563void __init smp_init(void)
564{
Michael Ellerman92b23272016-10-26 16:37:54 +1100565 int num_nodes, num_cpus;
Amerigo Wang34db18a02011-03-22 16:34:06 -0700566 unsigned int cpu;
567
Suresh Siddha3bb5d2e2012-04-20 17:08:50 -0700568 idle_threads_init();
Thomas Gleixner4cb28ce2016-02-26 18:43:38 +0000569 cpuhp_threads_init();
Suresh Siddha3bb5d2e2012-04-20 17:08:50 -0700570
Michael Ellerman51111dc2016-10-26 16:37:55 +1100571 pr_info("Bringing up secondary CPUs ...\n");
572
Amerigo Wang34db18a02011-03-22 16:34:06 -0700573 /* FIXME: This should be done in userspace --RR */
574 for_each_present_cpu(cpu) {
575 if (num_online_cpus() >= setup_max_cpus)
576 break;
577 if (!cpu_online(cpu))
578 cpu_up(cpu);
579 }
580
Michael Ellerman92b23272016-10-26 16:37:54 +1100581 num_nodes = num_online_nodes();
582 num_cpus = num_online_cpus();
583 pr_info("Brought up %d node%s, %d CPU%s\n",
584 num_nodes, (num_nodes > 1 ? "s" : ""),
585 num_cpus, (num_cpus > 1 ? "s" : ""));
586
Amerigo Wang34db18a02011-03-22 16:34:06 -0700587 /* Any cleanup work */
Amerigo Wang34db18a02011-03-22 16:34:06 -0700588 smp_cpus_done(setup_max_cpus);
589}
590
Amerigo Wang351f8f82011-01-12 16:59:39 -0800591/*
Tejun Heobd924e82011-01-20 12:07:13 +0100592 * Call a function on all processors. May be used during early boot while
593 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
594 * of local_irq_disable/enable().
Amerigo Wang351f8f82011-01-12 16:59:39 -0800595 */
596int on_each_cpu(void (*func) (void *info), void *info, int wait)
597{
Tejun Heobd924e82011-01-20 12:07:13 +0100598 unsigned long flags;
Amerigo Wang351f8f82011-01-12 16:59:39 -0800599 int ret = 0;
600
601 preempt_disable();
602 ret = smp_call_function(func, info, wait);
Tejun Heobd924e82011-01-20 12:07:13 +0100603 local_irq_save(flags);
Amerigo Wang351f8f82011-01-12 16:59:39 -0800604 func(info);
Tejun Heobd924e82011-01-20 12:07:13 +0100605 local_irq_restore(flags);
Amerigo Wang351f8f82011-01-12 16:59:39 -0800606 preempt_enable();
607 return ret;
608}
609EXPORT_SYMBOL(on_each_cpu);
Gilad Ben-Yossef3fc498f2012-03-28 14:42:43 -0700610
611/**
612 * on_each_cpu_mask(): Run a function on processors specified by
613 * cpumask, which may include the local processor.
614 * @mask: The set of cpus to run on (only runs on online subset).
615 * @func: The function to run. This must be fast and non-blocking.
616 * @info: An arbitrary pointer to pass to the function.
617 * @wait: If true, wait (atomically) until function has completed
618 * on other CPUs.
619 *
620 * If @wait is true, then returns once @func has returned.
621 *
David Daney202da402013-09-11 14:23:29 -0700622 * You must not call this function with disabled interrupts or from a
623 * hardware interrupt handler or from a bottom half handler. The
624 * exception is that it may be used during early boot while
625 * early_boot_irqs_disabled is set.
Gilad Ben-Yossef3fc498f2012-03-28 14:42:43 -0700626 */
627void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
628 void *info, bool wait)
629{
630 int cpu = get_cpu();
631
632 smp_call_function_many(mask, func, info, wait);
633 if (cpumask_test_cpu(cpu, mask)) {
David Daney202da402013-09-11 14:23:29 -0700634 unsigned long flags;
635 local_irq_save(flags);
Gilad Ben-Yossef3fc498f2012-03-28 14:42:43 -0700636 func(info);
David Daney202da402013-09-11 14:23:29 -0700637 local_irq_restore(flags);
Gilad Ben-Yossef3fc498f2012-03-28 14:42:43 -0700638 }
639 put_cpu();
640}
641EXPORT_SYMBOL(on_each_cpu_mask);
Gilad Ben-Yossefb3a7e982012-03-28 14:42:43 -0700642
643/*
644 * on_each_cpu_cond(): Call a function on each processor for which
645 * the supplied function cond_func returns true, optionally waiting
646 * for all the required CPUs to finish. This may include the local
647 * processor.
648 * @cond_func: A callback function that is passed a cpu id and
649 * the the info parameter. The function is called
650 * with preemption disabled. The function should
651 * return a blooean value indicating whether to IPI
652 * the specified CPU.
653 * @func: The function to run on all applicable CPUs.
654 * This must be fast and non-blocking.
655 * @info: An arbitrary pointer to pass to both functions.
656 * @wait: If true, wait (atomically) until function has
657 * completed on other CPUs.
658 * @gfp_flags: GFP flags to use when allocating the cpumask
659 * used internally by the function.
660 *
661 * The function might sleep if the GFP flags indicates a non
662 * atomic allocation is allowed.
663 *
664 * Preemption is disabled to protect against CPUs going offline but not online.
665 * CPUs going online during the call will not be seen or sent an IPI.
666 *
667 * You must not call this function with disabled interrupts or
668 * from a hardware interrupt handler or from a bottom half handler.
669 */
670void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
671 smp_call_func_t func, void *info, bool wait,
672 gfp_t gfp_flags)
673{
674 cpumask_var_t cpus;
675 int cpu, ret;
676
Mel Gormand0164ad2015-11-06 16:28:21 -0800677 might_sleep_if(gfpflags_allow_blocking(gfp_flags));
Gilad Ben-Yossefb3a7e982012-03-28 14:42:43 -0700678
679 if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
680 preempt_disable();
681 for_each_online_cpu(cpu)
682 if (cond_func(cpu, info))
683 cpumask_set_cpu(cpu, cpus);
684 on_each_cpu_mask(cpus, func, info, wait);
685 preempt_enable();
686 free_cpumask_var(cpus);
687 } else {
688 /*
689 * No free cpumask, bother. No matter, we'll
690 * just have to IPI them one by one.
691 */
692 preempt_disable();
693 for_each_online_cpu(cpu)
694 if (cond_func(cpu, info)) {
695 ret = smp_call_function_single(cpu, func,
696 info, wait);
Sasha Levin618fde872014-08-06 16:08:14 -0700697 WARN_ON_ONCE(ret);
Gilad Ben-Yossefb3a7e982012-03-28 14:42:43 -0700698 }
699 preempt_enable();
700 }
701}
702EXPORT_SYMBOL(on_each_cpu_cond);
Thomas Gleixnerf37f4352012-05-07 17:59:48 +0000703
704static void do_nothing(void *unused)
705{
706}
707
708/**
709 * kick_all_cpus_sync - Force all cpus out of idle
710 *
711 * Used to synchronize the update of pm_idle function pointer. It's
712 * called after the pointer is updated and returns after the dummy
713 * callback function has been executed on all cpus. The execution of
714 * the function can only happen on the remote cpus after they have
715 * left the idle function which had been called via pm_idle function
716 * pointer. So it's guaranteed that nothing uses the previous pointer
717 * anymore.
718 */
719void kick_all_cpus_sync(void)
720{
721 /* Make sure the change is visible before we kick the cpus */
722 smp_mb();
723 smp_call_function(do_nothing, NULL, 1);
724}
725EXPORT_SYMBOL_GPL(kick_all_cpus_sync);
Chuansheng Liuc6f44592014-09-04 15:17:54 +0800726
727/**
728 * wake_up_all_idle_cpus - break all cpus out of idle
729 * wake_up_all_idle_cpus try to break all cpus which is in idle state even
730 * including idle polling cpus, for non-idle cpus, we will do nothing
731 * for them.
732 */
733void wake_up_all_idle_cpus(void)
734{
735 int cpu;
736
737 preempt_disable();
738 for_each_online_cpu(cpu) {
739 if (cpu == smp_processor_id())
740 continue;
741
742 wake_up_if_idle(cpu);
743 }
744 preempt_enable();
745}
746EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus);
Juergen Grossdf8ce9d2016-08-29 08:48:44 +0200747
748/**
749 * smp_call_on_cpu - Call a function on a specific cpu
750 *
751 * Used to call a function on a specific cpu and wait for it to return.
752 * Optionally make sure the call is done on a specified physical cpu via vcpu
753 * pinning in order to support virtualized environments.
754 */
755struct smp_call_on_cpu_struct {
756 struct work_struct work;
757 struct completion done;
758 int (*func)(void *);
759 void *data;
760 int ret;
761 int cpu;
762};
763
764static void smp_call_on_cpu_callback(struct work_struct *work)
765{
766 struct smp_call_on_cpu_struct *sscs;
767
768 sscs = container_of(work, struct smp_call_on_cpu_struct, work);
769 if (sscs->cpu >= 0)
770 hypervisor_pin_vcpu(sscs->cpu);
771 sscs->ret = sscs->func(sscs->data);
772 if (sscs->cpu >= 0)
773 hypervisor_pin_vcpu(-1);
774
775 complete(&sscs->done);
776}
777
778int smp_call_on_cpu(unsigned int cpu, int (*func)(void *), void *par, bool phys)
779{
780 struct smp_call_on_cpu_struct sscs = {
Juergen Grossdf8ce9d2016-08-29 08:48:44 +0200781 .done = COMPLETION_INITIALIZER_ONSTACK(sscs.done),
782 .func = func,
783 .data = par,
784 .cpu = phys ? cpu : -1,
785 };
786
Peter Zijlstra8db54942016-09-11 10:36:26 +0200787 INIT_WORK_ONSTACK(&sscs.work, smp_call_on_cpu_callback);
788
Juergen Grossdf8ce9d2016-08-29 08:48:44 +0200789 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
790 return -ENXIO;
791
792 queue_work_on(cpu, system_wq, &sscs.work);
793 wait_for_completion(&sscs.done);
794
795 return sscs.ret;
796}
797EXPORT_SYMBOL_GPL(smp_call_on_cpu);