kernel/smp.c

   1 /*
   2  * Generic helpers for smp ipi calls
   3  *
   4  * (C) Jens Axboe <jens.axboe@oracle.com> 2008
   5  *
   6  */
   7 #include <linux/init.h>
   8 #include <linux/module.h>
   9 #include <linux/percpu.h>
  10 #include <linux/rcupdate.h>
  11 #include <linux/rculist.h>
  12 #include <linux/smp.h>
  13
  14 static DEFINE_PER_CPU(struct call_single_queue, call_single_queue);
  15 static LIST_HEAD(call_function_queue);
  16 __cacheline_aligned_in_smp DEFINE_SPINLOCK(call_function_lock);
  17
  18 enum {
  19         CSD_FLAG_WAIT           = 0x01,
  20         CSD_FLAG_ALLOC          = 0x02,
  21 };
  22
  23 struct call_function_data {
  24         struct call_single_data csd;
  25         spinlock_t lock;
  26         unsigned int refs;
  27         cpumask_t cpumask;
  28         struct rcu_head rcu_head;
  29 };
  30
  31 struct call_single_queue {
  32         struct list_head list;
  33         spinlock_t lock;
  34 };
  35
  36 static int __cpuinit init_call_single_data(void)
  37 {
  38         int i;
  39
  40         for_each_possible_cpu(i) {
  41                 struct call_single_queue *q = &per_cpu(call_single_queue, i);
  42
  43                 spin_lock_init(&q->lock);
  44                 INIT_LIST_HEAD(&q->list);
  45         }
  46         return 0;
  47 }
  48 early_initcall(init_call_single_data);
  49
  50 static void csd_flag_wait(struct call_single_data *data)
  51 {
  52         /* Wait for response */
  53         do {
  54                 if (!(data->flags & CSD_FLAG_WAIT))
  55                         break;
  56                 cpu_relax();
  57         } while (1);
  58 }
  59
  60 /*
  61  * Insert a previously allocated call_single_data element for execution
  62  * on the given CPU. data must already have ->func, ->info, and ->flags set.
  63  */
  64 static void generic_exec_single(int cpu, struct call_single_data *data)
  65 {
  66         struct call_single_queue *dst = &per_cpu(call_single_queue, cpu);
  67         int wait = data->flags & CSD_FLAG_WAIT, ipi;
  68         unsigned long flags;
  69
  70         spin_lock_irqsave(&dst->lock, flags);
  71         ipi = list_empty(&dst->list);
  72         list_add_tail(&data->list, &dst->list);
  73         spin_unlock_irqrestore(&dst->lock, flags);
  74
  75         /*
  76          * Make the list addition visible before sending the ipi.
  77          */
  78         smp_mb();
  79
  80         if (ipi)
  81                 arch_send_call_function_single_ipi(cpu);
  82
  83         if (wait)
  84                 csd_flag_wait(data);
  85 }
  86
  87 static void rcu_free_call_data(struct rcu_head *head)
  88 {
  89         struct call_function_data *data;
  90
  91         data = container_of(head, struct call_function_data, rcu_head);
  92
  93         kfree(data);
  94 }
  95
  96 /*
  97  * Invoked by arch to handle an IPI for call function. Must be called with
  98  * interrupts disabled.
  99  */
 100 void generic_smp_call_function_interrupt(void)
 101 {
 102         struct call_function_data *data;
 103         int cpu = get_cpu();
 104
 105         /*
 106          * It's ok to use list_for_each_rcu() here even though we may delete
 107          * 'pos', since list_del_rcu() doesn't clear ->next
 108          */
 109         rcu_read_lock();
 110         list_for_each_entry_rcu(data, &call_function_queue, csd.list) {
 111                 int refs;
 112
 113                 if (!cpu_isset(cpu, data->cpumask))
 114                         continue;
 115
 116                 data->csd.func(data->csd.info);
 117
 118                 spin_lock(&data->lock);
 119                 cpu_clear(cpu, data->cpumask);
 120                 WARN_ON(data->refs == 0);
 121                 data->refs--;
 122                 refs = data->refs;
 123                 spin_unlock(&data->lock);
 124
 125                 if (refs)
 126                         continue;
 127
 128                 spin_lock(&call_function_lock);
 129                 list_del_rcu(&data->csd.list);
 130                 spin_unlock(&call_function_lock);
 131
 132                 if (data->csd.flags & CSD_FLAG_WAIT) {
 133                         /*
 134                          * serialize stores to data with the flag clear
 135                          * and wakeup
 136                          */
 137                         smp_wmb();
 138                         data->csd.flags &= ~CSD_FLAG_WAIT;
 139                 }
 140                 if (data->csd.flags & CSD_FLAG_ALLOC)
 141                         call_rcu(&data->rcu_head, rcu_free_call_data);
 142         }
 143         rcu_read_unlock();
 144
 145         put_cpu();
 146 }
 147
 148 /*
 149  * Invoked by arch to handle an IPI for call function single. Must be called
 150  * from the arch with interrupts disabled.
 151  */
 152 void generic_smp_call_function_single_interrupt(void)
 153 {
 154         struct call_single_queue *q = &__get_cpu_var(call_single_queue);
 155         LIST_HEAD(list);
 156
 157         /*
 158          * Need to see other stores to list head for checking whether
 159          * list is empty without holding q->lock
 160          */
 161         smp_read_barrier_depends();
 162         while (!list_empty(&q->list)) {
 163                 unsigned int data_flags;
 164
 165                 spin_lock(&q->lock);
 166                 list_replace_init(&q->list, &list);
 167                 spin_unlock(&q->lock);
 168
 169                 while (!list_empty(&list)) {
 170                         struct call_single_data *data;
 171
 172                         data = list_entry(list.next, struct call_single_data,
 173                                                 list);
 174                         list_del(&data->list);
 175
 176                         /*
 177                          * 'data' can be invalid after this call if
 178                          * flags == 0 (when called through
 179                          * generic_exec_single(), so save them away before
 180                          * making the call.
 181                          */
 182                         data_flags = data->flags;
 183
 184                         data->func(data->info);
 185
 186                         if (data_flags & CSD_FLAG_WAIT) {
 187                                 smp_wmb();
 188                                 data->flags &= ~CSD_FLAG_WAIT;
 189                         } else if (data_flags & CSD_FLAG_ALLOC)
 190                                 kfree(data);
 191                 }
 192                 /*
 193                  * See comment on outer loop
 194                  */
 195                 smp_read_barrier_depends();
 196         }
 197 }
 198
 199 /*
 200  * smp_call_function_single - Run a function on a specific CPU
 201  * @func: The function to run. This must be fast and non-blocking.
 202  * @info: An arbitrary pointer to pass to the function.
 203  * @wait: If true, wait until function has completed on other CPUs.
 204  *
 205  * Returns 0 on success, else a negative status code. Note that @wait
 206  * will be implicitly turned on in case of allocation failures, since
 207  * we fall back to on-stack allocation.
 208  */
 209 int smp_call_function_single(int cpu, void (*func) (void *info), void *info,
 210                              int wait)
 211 {
 212         struct call_single_data d;
 213         unsigned long flags;
 214         /* prevent preemption and reschedule on another processor,
 215            as well as CPU removal */
 216         int me = get_cpu();
 217         int err = 0;
 218
 219         /* Can deadlock when called with interrupts disabled */
 220         WARN_ON(irqs_disabled());
 221
 222         if (cpu == me) {
 223                 local_irq_save(flags);
 224                 func(info);
 225                 local_irq_restore(flags);
 226         } else if ((unsigned)cpu < NR_CPUS && cpu_online(cpu)) {
 227                 struct call_single_data *data = NULL;
 228
 229                 if (!wait) {
 230                         data = kmalloc(sizeof(*data), GFP_ATOMIC);
 231                         if (data)
 232                                 data->flags = CSD_FLAG_ALLOC;
 233                 }
 234                 if (!data) {
 235                         data = &d;
 236                         data->flags = CSD_FLAG_WAIT;
 237                 }
 238
 239                 data->func = func;
 240                 data->info = info;
 241                 generic_exec_single(cpu, data);
 242         } else {
 243                 err = -ENXIO;   /* CPU not online */
 244         }
 245
 246         put_cpu();
 247         return err;
 248 }
 249 EXPORT_SYMBOL(smp_call_function_single);
 250
 251 /**
 252  * __smp_call_function_single(): Run a function on another CPU
 253  * @cpu: The CPU to run on.
 254  * @data: Pre-allocated and setup data structure
 255  *
 256  * Like smp_call_function_single(), but allow caller to pass in a pre-allocated
 257  * data structure. Useful for embedding @data inside other structures, for
 258  * instance.
 259  *
 260  */
 261 void __smp_call_function_single(int cpu, struct call_single_data *data)
 262 {
 263         /* Can deadlock when called with interrupts disabled */
 264         WARN_ON((data->flags & CSD_FLAG_WAIT) && irqs_disabled());
 265
 266         generic_exec_single(cpu, data);
 267 }
 268
 269 /* Dummy function */
 270 static void quiesce_dummy(void *unused)
 271 {
 272 }
 273
 274 /*
 275  * Ensure stack based data used in call function mask is safe to free.
 276  *
 277  * This is needed by smp_call_function_mask when using on-stack data, because
 278  * a single call function queue is shared by all CPUs, and any CPU may pick up
 279  * the data item on the queue at any time before it is deleted. So we need to
 280  * ensure that all CPUs have transitioned through a quiescent state after
 281  * this call.
 282  *
 283  * This is a very slow function, implemented by sending synchronous IPIs to
 284  * all possible CPUs. For this reason, we have to alloc data rather than use
 285  * stack based data even in the case of synchronous calls. The stack based
 286  * data is then just used for deadlock/oom fallback which will be very rare.
 287  *
 288  * If a faster scheme can be made, we could go back to preferring stack based
 289  * data -- the data allocation/free is non-zero cost.
 290  */
 291 static void smp_call_function_mask_quiesce_stack(cpumask_t mask)
 292 {
 293         struct call_single_data data;
 294         int cpu;
 295
 296         data.func = quiesce_dummy;
 297         data.info = NULL;
 298
 299         for_each_cpu_mask(cpu, mask) {
 300                 data.flags = CSD_FLAG_WAIT;
 301                 generic_exec_single(cpu, &data);
 302         }
 303 }
 304
 305 /**
 306  * smp_call_function_mask(): Run a function on a set of other CPUs.
 307  * @mask: The set of cpus to run on.
 308  * @func: The function to run. This must be fast and non-blocking.
 309  * @info: An arbitrary pointer to pass to the function.
 310  * @wait: If true, wait (atomically) until function has completed on other CPUs.
 311  *
 312  * Returns 0 on success, else a negative status code.
 313  *
 314  * If @wait is true, then returns once @func has returned. Note that @wait
 315  * will be implicitly turned on in case of allocation failures, since
 316  * we fall back to on-stack allocation.
 317  *
 318  * You must not call this function with disabled interrupts or from a
 319  * hardware interrupt handler or from a bottom half handler. Preemption
 320  * must be disabled when calling this function.
 321  */
 322 int smp_call_function_mask(cpumask_t mask, void (*func)(void *), void *info,
 323                            int wait)
 324 {
 325         struct call_function_data d;
 326         struct call_function_data *data = NULL;
 327         cpumask_t allbutself;
 328         unsigned long flags;
 329         int cpu, num_cpus;
 330         int slowpath = 0;
 331
 332         /* Can deadlock when called with interrupts disabled */
 333         WARN_ON(irqs_disabled());
 334
 335         cpu = smp_processor_id();
 336         allbutself = cpu_online_map;
 337         cpu_clear(cpu, allbutself);
 338         cpus_and(mask, mask, allbutself);
 339         num_cpus = cpus_weight(mask);
 340
 341         /*
 342          * If zero CPUs, return. If just a single CPU, turn this request
 343          * into a targetted single call instead since it's faster.
 344          */
 345         if (!num_cpus)
 346                 return 0;
 347         else if (num_cpus == 1) {
 348                 cpu = first_cpu(mask);
 349                 return smp_call_function_single(cpu, func, info, wait);
 350         }
 351
 352         data = kmalloc(sizeof(*data), GFP_ATOMIC);
 353         if (data) {
 354                 data->csd.flags = CSD_FLAG_ALLOC;
 355                 if (wait)
 356                         data->csd.flags |= CSD_FLAG_WAIT;
 357         } else {
 358                 data = &d;
 359                 data->csd.flags = CSD_FLAG_WAIT;
 360                 wait = 1;
 361                 slowpath = 1;
 362         }
 363
 364         spin_lock_init(&data->lock);
 365         data->csd.func = func;
 366         data->csd.info = info;
 367         data->refs = num_cpus;
 368         data->cpumask = mask;
 369
 370         spin_lock_irqsave(&call_function_lock, flags);
 371         list_add_tail_rcu(&data->csd.list, &call_function_queue);
 372         spin_unlock_irqrestore(&call_function_lock, flags);
 373
 374         /*
 375          * Make the list addition visible before sending the ipi.
 376          */
 377         smp_mb();
 378
 379         /* Send a message to all CPUs in the map */
 380         arch_send_call_function_ipi(mask);
 381
 382         /* optionally wait for the CPUs to complete */
 383         if (wait) {
 384                 csd_flag_wait(&data->csd);
 385                 if (unlikely(slowpath))
 386                         smp_call_function_mask_quiesce_stack(mask);
 387         }
 388
 389         return 0;
 390 }
 391 EXPORT_SYMBOL(smp_call_function_mask);
 392
 393 /**
 394  * smp_call_function(): Run a function on all other CPUs.
 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 on other CPUs.
 398  *
 399  * Returns 0 on success, else a negative status code.
 400  *
 401  * If @wait is true, then returns once @func has returned; otherwise
 402  * it returns just before the target cpu calls @func. In case of allocation
 403  * failure, @wait will be implicitly turned on.
 404  *
 405  * You must not call this function with disabled interrupts or from a
 406  * hardware interrupt handler or from a bottom half handler.
 407  */
 408 int smp_call_function(void (*func)(void *), void *info, int wait)
 409 {
 410         int ret;
 411
 412         preempt_disable();
 413         ret = smp_call_function_mask(cpu_online_map, func, info, wait);
 414         preempt_enable();
 415         return ret;
 416 }
 417 EXPORT_SYMBOL(smp_call_function);
 418
 419 void ipi_call_lock(void)
 420 {
 421         spin_lock(&call_function_lock);
 422 }
 423
 424 void ipi_call_unlock(void)
 425 {
 426         spin_unlock(&call_function_lock);
 427 }
 428
 429 void ipi_call_lock_irq(void)
 430 {
 431         spin_lock_irq(&call_function_lock);
 432 }
 433
 434 void ipi_call_unlock_irq(void)
 435 {
 436         spin_unlock_irq(&call_function_lock);
 437 }