kernel/smp.c

   1 /*
   2  * Generic helpers for smp ipi calls
   3  *
   4  * (C) Jens Axboe <jens.axboe@oracle.com> 2008
   5  */
   6 #include <linux/rcupdate.h>
   7 #include <linux/rculist.h>
   8 #include <linux/kernel.h>
   9 #include <linux/module.h>
  10 #include <linux/percpu.h>
  11 #include <linux/init.h>
  12 #include <linux/smp.h>
  13 #include <linux/cpu.h>
  14
  15 static DEFINE_PER_CPU(struct call_single_queue, call_single_queue);
  16
  17 static struct {
  18         struct list_head        queue;
  19         spinlock_t              lock;
  20 } call_function __cacheline_aligned_in_smp =
  21         {
  22                 .queue          = LIST_HEAD_INIT(call_function.queue),
  23                 .lock           = __SPIN_LOCK_UNLOCKED(call_function.lock),
  24         };
  25
  26 enum {
  27         CSD_FLAG_LOCK           = 0x01,
  28 };
  29
  30 struct call_function_data {
  31         struct call_single_data csd;
  32         spinlock_t              lock;
  33         unsigned int            refs;
  34         cpumask_var_t           cpumask;
  35 };
  36
  37 struct call_single_queue {
  38         struct list_head        list;
  39         spinlock_t              lock;
  40 };
  41
  42 static DEFINE_PER_CPU(struct call_function_data, cfd_data) = {
  43         .lock                   = __SPIN_LOCK_UNLOCKED(cfd_data.lock),
  44 };
  45
  46 static int
  47 hotplug_cfd(struct notifier_block *nfb, unsigned long action, void *hcpu)
  48 {
  49         long cpu = (long)hcpu;
  50         struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
  51
  52         switch (action) {
  53         case CPU_UP_PREPARE:
  54         case CPU_UP_PREPARE_FROZEN:
  55                 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
  56                                 cpu_to_node(cpu)))
  57                         return NOTIFY_BAD;
  58                 break;
  59
  60 #ifdef CONFIG_HOTPLUG_CPU
  61         case CPU_UP_CANCELED:
  62         case CPU_UP_CANCELED_FROZEN:
  63
  64         case CPU_DEAD:
  65         case CPU_DEAD_FROZEN:
  66                 free_cpumask_var(cfd->cpumask);
  67                 break;
  68 #endif
  69         };
  70
  71         return NOTIFY_OK;
  72 }
  73
  74 static struct notifier_block __cpuinitdata hotplug_cfd_notifier = {
  75         .notifier_call          = hotplug_cfd,
  76 };
  77
  78 static int __cpuinit init_call_single_data(void)
  79 {
  80         void *cpu = (void *)(long)smp_processor_id();
  81         int i;
  82
  83         for_each_possible_cpu(i) {
  84                 struct call_single_queue *q = &per_cpu(call_single_queue, i);
  85
  86                 spin_lock_init(&q->lock);
  87                 INIT_LIST_HEAD(&q->list);
  88         }
  89
  90         hotplug_cfd(&hotplug_cfd_notifier, CPU_UP_PREPARE, cpu);
  91         register_cpu_notifier(&hotplug_cfd_notifier);
  92
  93         return 0;
  94 }
  95 early_initcall(init_call_single_data);
  96
  97 /*
  98  * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
  99  *
 100  * For non-synchronous ipi calls the csd can still be in use by the
 101  * previous function call. For multi-cpu calls its even more interesting
 102  * as we'll have to ensure no other cpu is observing our csd.
 103  */
 104 static void csd_lock_wait(struct call_single_data *data)
 105 {
 106         while (data->flags & CSD_FLAG_LOCK)
 107                 cpu_relax();
 108 }
 109
 110 static void csd_lock(struct call_single_data *data)
 111 {
 112         csd_lock_wait(data);
 113         data->flags = CSD_FLAG_LOCK;
 114
 115         /*
 116          * prevent CPU from reordering the above assignment
 117          * to ->flags with any subsequent assignments to other
 118          * fields of the specified call_single_data structure:
 119          */
 120         smp_mb();
 121 }
 122
 123 static void csd_unlock(struct call_single_data *data)
 124 {
 125         WARN_ON(!(data->flags & CSD_FLAG_LOCK));
 126
 127         /*
 128          * ensure we're all done before releasing data:
 129          */
 130         smp_mb();
 131
 132         data->flags &= ~CSD_FLAG_LOCK;
 133 }
 134
 135 /*
 136  * Insert a previously allocated call_single_data element
 137  * for execution on the given CPU. data must already have
 138  * ->func, ->info, and ->flags set.
 139  */
 140 static
 141 void generic_exec_single(int cpu, struct call_single_data *data, int wait)
 142 {
 143         struct call_single_queue *dst = &per_cpu(call_single_queue, cpu);
 144         unsigned long flags;
 145         int ipi;
 146
 147         spin_lock_irqsave(&dst->lock, flags);
 148         ipi = list_empty(&dst->list);
 149         list_add_tail(&data->list, &dst->list);
 150         spin_unlock_irqrestore(&dst->lock, flags);
 151
 152         /*
 153          * The list addition should be visible before sending the IPI
 154          * handler locks the list to pull the entry off it because of
 155          * normal cache coherency rules implied by spinlocks.
 156          *
 157          * If IPIs can go out of order to the cache coherency protocol
 158          * in an architecture, sufficient synchronisation should be added
 159          * to arch code to make it appear to obey cache coherency WRT
 160          * locking and barrier primitives. Generic code isn't really
 161          * equipped to do the right thing...
 162          */
 163         if (ipi)
 164                 arch_send_call_function_single_ipi(cpu);
 165
 166         if (wait)
 167                 csd_lock_wait(data);
 168 }
 169
 170 /*
 171  * Invoked by arch to handle an IPI for call function. Must be called with
 172  * interrupts disabled.
 173  */
 174 void generic_smp_call_function_interrupt(void)
 175 {
 176         struct call_function_data *data;
 177         int cpu = get_cpu();
 178
 179         /*
 180          * Shouldn't receive this interrupt on a cpu that is not yet online.
 181          */
 182         WARN_ON_ONCE(!cpu_online(cpu));
 183
 184         /*
 185          * Ensure entry is visible on call_function_queue after we have
 186          * entered the IPI. See comment in smp_call_function_many.
 187          * If we don't have this, then we may miss an entry on the list
 188          * and never get another IPI to process it.
 189          */
 190         smp_mb();
 191
 192         /*
 193          * It's ok to use list_for_each_rcu() here even though we may
 194          * delete 'pos', since list_del_rcu() doesn't clear ->next
 195          */
 196         list_for_each_entry_rcu(data, &call_function.queue, csd.list) {
 197                 int refs;
 198
 199                 spin_lock(&data->lock);
 200                 if (!cpumask_test_cpu(cpu, data->cpumask)) {
 201                         spin_unlock(&data->lock);
 202                         continue;
 203                 }
 204                 cpumask_clear_cpu(cpu, data->cpumask);
 205                 spin_unlock(&data->lock);
 206
 207                 data->csd.func(data->csd.info);
 208
 209                 spin_lock(&data->lock);
 210                 WARN_ON(data->refs == 0);
 211                 refs = --data->refs;
 212                 if (!refs) {
 213                         spin_lock(&call_function.lock);
 214                         list_del_rcu(&data->csd.list);
 215                         spin_unlock(&call_function.lock);
 216                 }
 217                 spin_unlock(&data->lock);
 218
 219                 if (refs)
 220                         continue;
 221
 222                 csd_unlock(&data->csd);
 223         }
 224
 225         put_cpu();
 226 }
 227
 228 /*
 229  * Invoked by arch to handle an IPI for call function single. Must be
 230  * called from the arch with interrupts disabled.
 231  */
 232 void generic_smp_call_function_single_interrupt(void)
 233 {
 234         struct call_single_queue *q = &__get_cpu_var(call_single_queue);
 235         unsigned int data_flags;
 236         LIST_HEAD(list);
 237
 238         /*
 239          * Shouldn't receive this interrupt on a cpu that is not yet online.
 240          */
 241         WARN_ON_ONCE(!cpu_online(smp_processor_id()));
 242
 243         spin_lock(&q->lock);
 244         list_replace_init(&q->list, &list);
 245         spin_unlock(&q->lock);
 246
 247         while (!list_empty(&list)) {
 248                 struct call_single_data *data;
 249
 250                 data = list_entry(list.next, struct call_single_data, list);
 251                 list_del(&data->list);
 252
 253                 /*
 254                  * 'data' can be invalid after this call if flags == 0
 255                  * (when called through generic_exec_single()),
 256                  * so save them away before making the call:
 257                  */
 258                 data_flags = data->flags;
 259
 260                 data->func(data->info);
 261
 262                 /*
 263                  * Unlocked CSDs are valid through generic_exec_single():
 264                  */
 265                 if (data_flags & CSD_FLAG_LOCK)
 266                         csd_unlock(data);
 267         }
 268 }
 269
 270 static DEFINE_PER_CPU(struct call_single_data, csd_data);
 271
 272 /*
 273  * smp_call_function_single - Run a function on a specific CPU
 274  * @func: The function to run. This must be fast and non-blocking.
 275  * @info: An arbitrary pointer to pass to the function.
 276  * @wait: If true, wait until function has completed on other CPUs.
 277  *
 278  * Returns 0 on success, else a negative status code. Note that @wait
 279  * will be implicitly turned on in case of allocation failures, since
 280  * we fall back to on-stack allocation.
 281  */
 282 int smp_call_function_single(int cpu, void (*func) (void *info), void *info,
 283                              int wait)
 284 {
 285         struct call_single_data d = {
 286                 .flags = 0,
 287         };
 288         unsigned long flags;
 289         int this_cpu;
 290         int err = 0;
 291
 292         /*
 293          * prevent preemption and reschedule on another processor,
 294          * as well as CPU removal
 295          */
 296         this_cpu = get_cpu();
 297
 298         /*
 299          * Can deadlock when called with interrupts disabled.
 300          * We allow cpu's that are not yet online though, as no one else can
 301          * send smp call function interrupt to this cpu and as such deadlocks
 302          * can't happen.
 303          */
 304         WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
 305                      && !oops_in_progress);
 306
 307         if (cpu == this_cpu) {
 308                 local_irq_save(flags);
 309                 func(info);
 310                 local_irq_restore(flags);
 311         } else {
 312                 if ((unsigned)cpu < nr_cpu_ids && cpu_online(cpu)) {
 313                         struct call_single_data *data = &d;
 314
 315                         if (!wait)
 316                                 data = &__get_cpu_var(csd_data);
 317
 318                         csd_lock(data);
 319
 320                         data->func = func;
 321                         data->info = info;
 322                         generic_exec_single(cpu, data, wait);
 323                 } else {
 324                         err = -ENXIO;   /* CPU not online */
 325                 }
 326         }
 327
 328         put_cpu();
 329
 330         return err;
 331 }
 332 EXPORT_SYMBOL(smp_call_function_single);
 333
 334 /**
 335  * __smp_call_function_single(): Run a function on another CPU
 336  * @cpu: The CPU to run on.
 337  * @data: Pre-allocated and setup data structure
 338  *
 339  * Like smp_call_function_single(), but allow caller to pass in a
 340  * pre-allocated data structure. Useful for embedding @data inside
 341  * other structures, for instance.
 342  */
 343 void __smp_call_function_single(int cpu, struct call_single_data *data,
 344                                 int wait)
 345 {
 346         csd_lock(data);
 347
 348         /*
 349          * Can deadlock when called with interrupts disabled.
 350          * We allow cpu's that are not yet online though, as no one else can
 351          * send smp call function interrupt to this cpu and as such deadlocks
 352          * can't happen.
 353          */
 354         WARN_ON_ONCE(cpu_online(smp_processor_id()) && wait && irqs_disabled()
 355                      && !oops_in_progress);
 356
 357         generic_exec_single(cpu, data, wait);
 358 }
 359
 360 /* Deprecated: shim for archs using old arch_send_call_function_ipi API. */
 361
 362 #ifndef arch_send_call_function_ipi_mask
 363 # define arch_send_call_function_ipi_mask(maskp) \
 364          arch_send_call_function_ipi(*(maskp))
 365 #endif
 366
 367 /**
 368  * smp_call_function_many(): Run a function on a set of other CPUs.
 369  * @mask: The set of cpus to run on (only runs on online subset).
 370  * @func: The function to run. This must be fast and non-blocking.
 371  * @info: An arbitrary pointer to pass to the function.
 372  * @wait: If true, wait (atomically) until function has completed
 373  *        on other CPUs.
 374  *
 375  * If @wait is true, then returns once @func has returned. Note that @wait
 376  * will be implicitly turned on in case of allocation failures, since
 377  * we fall back to on-stack allocation.
 378  *
 379  * You must not call this function with disabled interrupts or from a
 380  * hardware interrupt handler or from a bottom half handler. Preemption
 381  * must be disabled when calling this function.
 382  */
 383 void smp_call_function_many(const struct cpumask *mask,
 384                             void (*func)(void *), void *info, bool wait)
 385 {
 386         struct call_function_data *data;
 387         unsigned long flags;
 388         int cpu, next_cpu, this_cpu = smp_processor_id();
 389
 390         /*
 391          * Can deadlock when called with interrupts disabled.
 392          * We allow cpu's that are not yet online though, as no one else can
 393          * send smp call function interrupt to this cpu and as such deadlocks
 394          * can't happen.
 395          */
 396         WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
 397                      && !oops_in_progress);
 398
 399         /* So, what's a CPU they want? Ignoring this one. */
 400         cpu = cpumask_first_and(mask, cpu_online_mask);
 401         if (cpu == this_cpu)
 402                 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
 403
 404         /* No online cpus?  We're done. */
 405         if (cpu >= nr_cpu_ids)
 406                 return;
 407
 408         /* Do we have another CPU which isn't us? */
 409         next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
 410         if (next_cpu == this_cpu)
 411                 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
 412
 413         /* Fastpath: do that cpu by itself. */
 414         if (next_cpu >= nr_cpu_ids) {
 415                 smp_call_function_single(cpu, func, info, wait);
 416                 return;
 417         }
 418
 419         data = &__get_cpu_var(cfd_data);
 420         csd_lock(&data->csd);
 421
 422         spin_lock_irqsave(&data->lock, flags);
 423         data->csd.func = func;
 424         data->csd.info = info;
 425         cpumask_and(data->cpumask, mask, cpu_online_mask);
 426         cpumask_clear_cpu(this_cpu, data->cpumask);
 427         data->refs = cpumask_weight(data->cpumask);
 428
 429         spin_lock(&call_function.lock);
 430         /*
 431          * Place entry at the _HEAD_ of the list, so that any cpu still
 432          * observing the entry in generic_smp_call_function_interrupt()
 433          * will not miss any other list entries:
 434          */
 435         list_add_rcu(&data->csd.list, &call_function.queue);
 436         spin_unlock(&call_function.lock);
 437
 438         spin_unlock_irqrestore(&data->lock, flags);
 439
 440         /*
 441          * Make the list addition visible before sending the ipi.
 442          * (IPIs must obey or appear to obey normal Linux cache
 443          * coherency rules -- see comment in generic_exec_single).
 444          */
 445         smp_mb();
 446
 447         /* Send a message to all CPUs in the map */
 448         arch_send_call_function_ipi_mask(data->cpumask);
 449
 450         /* Optionally wait for the CPUs to complete */
 451         if (wait)
 452                 csd_lock_wait(&data->csd);
 453 }
 454 EXPORT_SYMBOL(smp_call_function_many);
 455
 456 /**
 457  * smp_call_function(): Run a function on all other CPUs.
 458  * @func: The function to run. This must be fast and non-blocking.
 459  * @info: An arbitrary pointer to pass to the function.
 460  * @wait: If true, wait (atomically) until function has completed
 461  *        on other CPUs.
 462  *
 463  * Returns 0.
 464  *
 465  * If @wait is true, then returns once @func has returned; otherwise
 466  * it returns just before the target cpu calls @func. In case of allocation
 467  * failure, @wait will be implicitly turned on.
 468  *
 469  * You must not call this function with disabled interrupts or from a
 470  * hardware interrupt handler or from a bottom half handler.
 471  */
 472 int smp_call_function(void (*func)(void *), void *info, int wait)
 473 {
 474         preempt_disable();
 475         smp_call_function_many(cpu_online_mask, func, info, wait);
 476         preempt_enable();
 477
 478         return 0;
 479 }
 480 EXPORT_SYMBOL(smp_call_function);
 481
 482 void ipi_call_lock(void)
 483 {
 484         spin_lock(&call_function.lock);
 485 }
 486
 487 void ipi_call_unlock(void)
 488 {
 489         spin_unlock(&call_function.lock);
 490 }
 491
 492 void ipi_call_lock_irq(void)
 493 {
 494         spin_lock_irq(&call_function.lock);
 495 }
 496
 497 void ipi_call_unlock_irq(void)
 498 {
 499         spin_unlock_irq(&call_function.lock);
 500 }