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         raw_spinlock_t          lock;
  20 } call_function __cacheline_aligned_in_smp =
  21         {
  22                 .queue          = LIST_HEAD_INIT(call_function.queue),
  23                 .lock           = __RAW_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         atomic_t                refs;
  33         cpumask_var_t           cpumask;
  34 };
  35
  36 struct call_single_queue {
  37         struct list_head        list;
  38         raw_spinlock_t          lock;
  39 };
  40
  41 static DEFINE_PER_CPU(struct call_function_data, cfd_data);
  42
  43 static int
  44 hotplug_cfd(struct notifier_block *nfb, unsigned long action, void *hcpu)
  45 {
  46         long cpu = (long)hcpu;
  47         struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
  48
  49         switch (action) {
  50         case CPU_UP_PREPARE:
  51         case CPU_UP_PREPARE_FROZEN:
  52                 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
  53                                 cpu_to_node(cpu)))
  54                         return NOTIFY_BAD;
  55                 break;
  56
  57 #ifdef CONFIG_HOTPLUG_CPU
  58         case CPU_UP_CANCELED:
  59         case CPU_UP_CANCELED_FROZEN:
  60
  61         case CPU_DEAD:
  62         case CPU_DEAD_FROZEN:
  63                 free_cpumask_var(cfd->cpumask);
  64                 break;
  65 #endif
  66         };
  67
  68         return NOTIFY_OK;
  69 }
  70
  71 static struct notifier_block __cpuinitdata hotplug_cfd_notifier = {
  72         .notifier_call          = hotplug_cfd,
  73 };
  74
  75 static int __cpuinit init_call_single_data(void)
  76 {
  77         void *cpu = (void *)(long)smp_processor_id();
  78         int i;
  79
  80         for_each_possible_cpu(i) {
  81                 struct call_single_queue *q = &per_cpu(call_single_queue, i);
  82
  83                 raw_spin_lock_init(&q->lock);
  84                 INIT_LIST_HEAD(&q->list);
  85         }
  86
  87         hotplug_cfd(&hotplug_cfd_notifier, CPU_UP_PREPARE, cpu);
  88         register_cpu_notifier(&hotplug_cfd_notifier);
  89
  90         return 0;
  91 }
  92 early_initcall(init_call_single_data);
  93
  94 /*
  95  * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
  96  *
  97  * For non-synchronous ipi calls the csd can still be in use by the
  98  * previous function call. For multi-cpu calls its even more interesting
  99  * as we'll have to ensure no other cpu is observing our csd.
 100  */
 101 static void csd_lock_wait(struct call_single_data *data)
 102 {
 103         while (data->flags & CSD_FLAG_LOCK)
 104                 cpu_relax();
 105 }
 106
 107 static void csd_lock(struct call_single_data *data)
 108 {
 109         csd_lock_wait(data);
 110         data->flags = CSD_FLAG_LOCK;
 111
 112         /*
 113          * prevent CPU from reordering the above assignment
 114          * to ->flags with any subsequent assignments to other
 115          * fields of the specified call_single_data structure:
 116          */
 117         smp_mb();
 118 }
 119
 120 static void csd_unlock(struct call_single_data *data)
 121 {
 122         WARN_ON(!(data->flags & CSD_FLAG_LOCK));
 123
 124         /*
 125          * ensure we're all done before releasing data:
 126          */
 127         smp_mb();
 128
 129         data->flags &= ~CSD_FLAG_LOCK;
 130 }
 131
 132 /*
 133  * Insert a previously allocated call_single_data element
 134  * for execution on the given CPU. data must already have
 135  * ->func, ->info, and ->flags set.
 136  */
 137 static
 138 void generic_exec_single(int cpu, struct call_single_data *data, int wait)
 139 {
 140         struct call_single_queue *dst = &per_cpu(call_single_queue, cpu);
 141         unsigned long flags;
 142         int ipi;
 143
 144         raw_spin_lock_irqsave(&dst->lock, flags);
 145         ipi = list_empty(&dst->list);
 146         list_add_tail(&data->list, &dst->list);
 147         raw_spin_unlock_irqrestore(&dst->lock, flags);
 148
 149         /*
 150          * The list addition should be visible before sending the IPI
 151          * handler locks the list to pull the entry off it because of
 152          * normal cache coherency rules implied by spinlocks.
 153          *
 154          * If IPIs can go out of order to the cache coherency protocol
 155          * in an architecture, sufficient synchronisation should be added
 156          * to arch code to make it appear to obey cache coherency WRT
 157          * locking and barrier primitives. Generic code isn't really
 158          * equipped to do the right thing...
 159          */
 160         if (ipi)
 161                 arch_send_call_function_single_ipi(cpu);
 162
 163         if (wait)
 164                 csd_lock_wait(data);
 165 }
 166
 167 /*
 168  * Invoked by arch to handle an IPI for call function. Must be called with
 169  * interrupts disabled.
 170  */
 171 void generic_smp_call_function_interrupt(void)
 172 {
 173         struct call_function_data *data;
 174         int cpu = smp_processor_id();
 175
 176         /*
 177          * Shouldn't receive this interrupt on a cpu that is not yet online.
 178          */
 179         WARN_ON_ONCE(!cpu_online(cpu));
 180
 181         /*
 182          * Ensure entry is visible on call_function_queue after we have
 183          * entered the IPI. See comment in smp_call_function_many.
 184          * If we don't have this, then we may miss an entry on the list
 185          * and never get another IPI to process it.
 186          */
 187         smp_mb();
 188
 189         /*
 190          * It's ok to use list_for_each_rcu() here even though we may
 191          * delete 'pos', since list_del_rcu() doesn't clear ->next
 192          */
 193         list_for_each_entry_rcu(data, &call_function.queue, csd.list) {
 194                 int refs;
 195
 196                 if (!cpumask_test_and_clear_cpu(cpu, data->cpumask))
 197                         continue;
 198
 199                 data->csd.func(data->csd.info);
 200
 201                 refs = atomic_dec_return(&data->refs);
 202                 WARN_ON(refs < 0);
 203                 if (!refs) {
 204                         raw_spin_lock(&call_function.lock);
 205                         list_del_rcu(&data->csd.list);
 206                         raw_spin_unlock(&call_function.lock);
 207                 }
 208
 209                 if (refs)
 210                         continue;
 211
 212                 csd_unlock(&data->csd);
 213         }
 214
 215 }
 216
 217 /*
 218  * Invoked by arch to handle an IPI for call function single. Must be
 219  * called from the arch with interrupts disabled.
 220  */
 221 void generic_smp_call_function_single_interrupt(void)
 222 {
 223         struct call_single_queue *q = &__get_cpu_var(call_single_queue);
 224         unsigned int data_flags;
 225         LIST_HEAD(list);
 226
 227         /*
 228          * Shouldn't receive this interrupt on a cpu that is not yet online.
 229          */
 230         WARN_ON_ONCE(!cpu_online(smp_processor_id()));
 231
 232         raw_spin_lock(&q->lock);
 233         list_replace_init(&q->list, &list);
 234         raw_spin_unlock(&q->lock);
 235
 236         while (!list_empty(&list)) {
 237                 struct call_single_data *data;
 238
 239                 data = list_entry(list.next, struct call_single_data, list);
 240                 list_del(&data->list);
 241
 242                 /*
 243                  * 'data' can be invalid after this call if flags == 0
 244                  * (when called through generic_exec_single()),
 245                  * so save them away before making the call:
 246                  */
 247                 data_flags = data->flags;
 248
 249                 data->func(data->info);
 250
 251                 /*
 252                  * Unlocked CSDs are valid through generic_exec_single():
 253                  */
 254                 if (data_flags & CSD_FLAG_LOCK)
 255                         csd_unlock(data);
 256         }
 257 }
 258
 259 static DEFINE_PER_CPU(struct call_single_data, csd_data);
 260
 261 /*
 262  * smp_call_function_single - Run a function on a specific CPU
 263  * @func: The function to run. This must be fast and non-blocking.
 264  * @info: An arbitrary pointer to pass to the function.
 265  * @wait: If true, wait until function has completed on other CPUs.
 266  *
 267  * Returns 0 on success, else a negative status code.
 268  */
 269 int smp_call_function_single(int cpu, void (*func) (void *info), void *info,
 270                              int wait)
 271 {
 272         struct call_single_data d = {
 273                 .flags = 0,
 274         };
 275         unsigned long flags;
 276         int this_cpu;
 277         int err = 0;
 278
 279         /*
 280          * prevent preemption and reschedule on another processor,
 281          * as well as CPU removal
 282          */
 283         this_cpu = get_cpu();
 284
 285         /*
 286          * Can deadlock when called with interrupts disabled.
 287          * We allow cpu's that are not yet online though, as no one else can
 288          * send smp call function interrupt to this cpu and as such deadlocks
 289          * can't happen.
 290          */
 291         WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
 292                      && !oops_in_progress);
 293
 294         if (cpu == this_cpu) {
 295                 local_irq_save(flags);
 296                 func(info);
 297                 local_irq_restore(flags);
 298         } else {
 299                 if ((unsigned)cpu < nr_cpu_ids && cpu_online(cpu)) {
 300                         struct call_single_data *data = &d;
 301
 302                         if (!wait)
 303                                 data = &__get_cpu_var(csd_data);
 304
 305                         csd_lock(data);
 306
 307                         data->func = func;
 308                         data->info = info;
 309                         generic_exec_single(cpu, data, wait);
 310                 } else {
 311                         err = -ENXIO;   /* CPU not online */
 312                 }
 313         }
 314
 315         put_cpu();
 316
 317         return err;
 318 }
 319 EXPORT_SYMBOL(smp_call_function_single);
 320
 321 /*
 322  * smp_call_function_any - Run a function on any of the given cpus
 323  * @mask: The mask of cpus it can run on.
 324  * @func: The function to run. This must be fast and non-blocking.
 325  * @info: An arbitrary pointer to pass to the function.
 326  * @wait: If true, wait until function has completed.
 327  *
 328  * Returns 0 on success, else a negative status code (if no cpus were online).
 329  * Note that @wait will be implicitly turned on in case of allocation failures,
 330  * since we fall back to on-stack allocation.
 331  *
 332  * Selection preference:
 333  *      1) current cpu if in @mask
 334  *      2) any cpu of current node if in @mask
 335  *      3) any other online cpu in @mask
 336  */
 337 int smp_call_function_any(const struct cpumask *mask,
 338                           void (*func)(void *info), void *info, int wait)
 339 {
 340         unsigned int cpu;
 341         const struct cpumask *nodemask;
 342         int ret;
 343
 344         /* Try for same CPU (cheapest) */
 345         cpu = get_cpu();
 346         if (cpumask_test_cpu(cpu, mask))
 347                 goto call;
 348
 349         /* Try for same node. */
 350         nodemask = cpumask_of_node(cpu_to_node(cpu));
 351         for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
 352              cpu = cpumask_next_and(cpu, nodemask, mask)) {
 353                 if (cpu_online(cpu))
 354                         goto call;
 355         }
 356
 357         /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
 358         cpu = cpumask_any_and(mask, cpu_online_mask);
 359 call:
 360         ret = smp_call_function_single(cpu, func, info, wait);
 361         put_cpu();
 362         return ret;
 363 }
 364 EXPORT_SYMBOL_GPL(smp_call_function_any);
 365
 366 /**
 367  * __smp_call_function_single(): Run a function on another CPU
 368  * @cpu: The CPU to run on.
 369  * @data: Pre-allocated and setup data structure
 370  *
 371  * Like smp_call_function_single(), but allow caller to pass in a
 372  * pre-allocated data structure. Useful for embedding @data inside
 373  * other structures, for instance.
 374  */
 375 void __smp_call_function_single(int cpu, struct call_single_data *data,
 376                                 int wait)
 377 {
 378         csd_lock(data);
 379
 380         /*
 381          * Can deadlock when called with interrupts disabled.
 382          * We allow cpu's that are not yet online though, as no one else can
 383          * send smp call function interrupt to this cpu and as such deadlocks
 384          * can't happen.
 385          */
 386         WARN_ON_ONCE(cpu_online(smp_processor_id()) && wait && irqs_disabled()
 387                      && !oops_in_progress);
 388
 389         generic_exec_single(cpu, data, wait);
 390 }
 391
 392 /**
 393  * smp_call_function_many(): Run a function on a set of other CPUs.
 394  * @mask: The set of cpus to run on (only runs on online subset).
 395  * @func: The function to run. This must be fast and non-blocking.
 396  * @info: An arbitrary pointer to pass to the function.
 397  * @wait: If true, wait (atomically) until function has completed
 398  *        on other CPUs.
 399  *
 400  * If @wait is true, then returns once @func has returned.
 401  *
 402  * You must not call this function with disabled interrupts or from a
 403  * hardware interrupt handler or from a bottom half handler. Preemption
 404  * must be disabled when calling this function.
 405  */
 406 void smp_call_function_many(const struct cpumask *mask,
 407                             void (*func)(void *), void *info, bool wait)
 408 {
 409         struct call_function_data *data;
 410         unsigned long flags;
 411         int cpu, next_cpu, this_cpu = smp_processor_id();
 412
 413         /*
 414          * Can deadlock when called with interrupts disabled.
 415          * We allow cpu's that are not yet online though, as no one else can
 416          * send smp call function interrupt to this cpu and as such deadlocks
 417          * can't happen.
 418          */
 419         WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
 420                      && !oops_in_progress);
 421
 422         /* So, what's a CPU they want? Ignoring this one. */
 423         cpu = cpumask_first_and(mask, cpu_online_mask);
 424         if (cpu == this_cpu)
 425                 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
 426
 427         /* No online cpus?  We're done. */
 428         if (cpu >= nr_cpu_ids)
 429                 return;
 430
 431         /* Do we have another CPU which isn't us? */
 432         next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
 433         if (next_cpu == this_cpu)
 434                 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
 435
 436         /* Fastpath: do that cpu by itself. */
 437         if (next_cpu >= nr_cpu_ids) {
 438                 smp_call_function_single(cpu, func, info, wait);
 439                 return;
 440         }
 441
 442         data = &__get_cpu_var(cfd_data);
 443         csd_lock(&data->csd);
 444
 445         data->csd.func = func;
 446         data->csd.info = info;
 447         cpumask_and(data->cpumask, mask, cpu_online_mask);
 448         cpumask_clear_cpu(this_cpu, data->cpumask);
 449         atomic_set(&data->refs, cpumask_weight(data->cpumask));
 450
 451         raw_spin_lock_irqsave(&call_function.lock, flags);
 452         /*
 453          * Place entry at the _HEAD_ of the list, so that any cpu still
 454          * observing the entry in generic_smp_call_function_interrupt()
 455          * will not miss any other list entries:
 456          */
 457         list_add_rcu(&data->csd.list, &call_function.queue);
 458         raw_spin_unlock_irqrestore(&call_function.lock, flags);
 459
 460         /*
 461          * Make the list addition visible before sending the ipi.
 462          * (IPIs must obey or appear to obey normal Linux cache
 463          * coherency rules -- see comment in generic_exec_single).
 464          */
 465         smp_mb();
 466
 467         /* Send a message to all CPUs in the map */
 468         arch_send_call_function_ipi_mask(data->cpumask);
 469
 470         /* Optionally wait for the CPUs to complete */
 471         if (wait)
 472                 csd_lock_wait(&data->csd);
 473 }
 474 EXPORT_SYMBOL(smp_call_function_many);
 475
 476 /**
 477  * smp_call_function(): Run a function on all other CPUs.
 478  * @func: The function to run. This must be fast and non-blocking.
 479  * @info: An arbitrary pointer to pass to the function.
 480  * @wait: If true, wait (atomically) until function has completed
 481  *        on other CPUs.
 482  *
 483  * Returns 0.
 484  *
 485  * If @wait is true, then returns once @func has returned; otherwise
 486  * it returns just before the target cpu calls @func.
 487  *
 488  * You must not call this function with disabled interrupts or from a
 489  * hardware interrupt handler or from a bottom half handler.
 490  */
 491 int smp_call_function(void (*func)(void *), void *info, int wait)
 492 {
 493         preempt_disable();
 494         smp_call_function_many(cpu_online_mask, func, info, wait);
 495         preempt_enable();
 496
 497         return 0;
 498 }
 499 EXPORT_SYMBOL(smp_call_function);
 500
 501 void ipi_call_lock(void)
 502 {
 503         raw_spin_lock(&call_function.lock);
 504 }
 505
 506 void ipi_call_unlock(void)
 507 {
 508         raw_spin_unlock(&call_function.lock);
 509 }
 510
 511 void ipi_call_lock_irq(void)
 512 {
 513         raw_spin_lock_irq(&call_function.lock);
 514 }
 515
 516 void ipi_call_unlock_irq(void)
 517 {
 518         raw_spin_unlock_irq(&call_function.lock);
 519 }