kernel/padata.c

   1 /*
   2  * padata.c - generic interface to process data streams in parallel
   3  *
   4  * See Documentation/padata.txt for an api documentation.
   5  *
   6  * Copyright (C) 2008, 2009 secunet Security Networks AG
   7  * Copyright (C) 2008, 2009 Steffen Klassert <steffen.klassert@secunet.com>
   8  *
   9  * This program is free software; you can redistribute it and/or modify it
  10  * under the terms and conditions of the GNU General Public License,
  11  * version 2, as published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope it will be useful, but WITHOUT
  14  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  15  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  16  * more details.
  17  *
  18  * You should have received a copy of the GNU General Public License along with
  19  * this program; if not, write to the Free Software Foundation, Inc.,
  20  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
  21  */
  22
  23 #include <linux/export.h>
  24 #include <linux/cpumask.h>
  25 #include <linux/err.h>
  26 #include <linux/cpu.h>
  27 #include <linux/padata.h>
  28 #include <linux/mutex.h>
  29 #include <linux/sched.h>
  30 #include <linux/slab.h>
  31 #include <linux/sysfs.h>
  32 #include <linux/rcupdate.h>
  33
  34 #define MAX_OBJ_NUM 1000
  35
  36 static int padata_index_to_cpu(struct parallel_data *pd, int cpu_index)
  37 {
  38         int cpu, target_cpu;
  39
  40         target_cpu = cpumask_first(pd->cpumask.pcpu);
  41         for (cpu = 0; cpu < cpu_index; cpu++)
  42                 target_cpu = cpumask_next(target_cpu, pd->cpumask.pcpu);
  43
  44         return target_cpu;
  45 }
  46
  47 static int padata_cpu_hash(struct parallel_data *pd)
  48 {
  49         unsigned int seq_nr;
  50         int cpu_index;
  51
  52         /*
  53          * Hash the sequence numbers to the cpus by taking
  54          * seq_nr mod. number of cpus in use.
  55          */
  56
  57         seq_nr = atomic_inc_return(&pd->seq_nr);
  58         cpu_index = seq_nr % cpumask_weight(pd->cpumask.pcpu);
  59
  60         return padata_index_to_cpu(pd, cpu_index);
  61 }
  62
  63 static void padata_parallel_worker(struct work_struct *parallel_work)
  64 {
  65         struct padata_parallel_queue *pqueue;
  66         struct parallel_data *pd;
  67         struct padata_instance *pinst;
  68         LIST_HEAD(local_list);
  69
  70         local_bh_disable();
  71         pqueue = container_of(parallel_work,
  72                               struct padata_parallel_queue, work);
  73         pd = pqueue->pd;
  74         pinst = pd->pinst;
  75
  76         spin_lock(&pqueue->parallel.lock);
  77         list_replace_init(&pqueue->parallel.list, &local_list);
  78         spin_unlock(&pqueue->parallel.lock);
  79
  80         while (!list_empty(&local_list)) {
  81                 struct padata_priv *padata;
  82
  83                 padata = list_entry(local_list.next,
  84                                     struct padata_priv, list);
  85
  86                 list_del_init(&padata->list);
  87
  88                 padata->parallel(padata);
  89         }
  90
  91         local_bh_enable();
  92 }
  93
  94 /**
  95  * padata_do_parallel - padata parallelization function
  96  *
  97  * @pinst: padata instance
  98  * @padata: object to be parallelized
  99  * @cb_cpu: cpu the serialization callback function will run on,
 100  *          must be in the serial cpumask of padata(i.e. cpumask.cbcpu).
 101  *
 102  * The parallelization callback function will run with BHs off.
 103  * Note: Every object which is parallelized by padata_do_parallel
 104  * must be seen by padata_do_serial.
 105  */
 106 int padata_do_parallel(struct padata_instance *pinst,
 107                        struct padata_priv *padata, int cb_cpu)
 108 {
 109         int target_cpu, err;
 110         struct padata_parallel_queue *queue;
 111         struct parallel_data *pd;
 112
 113         rcu_read_lock_bh();
 114
 115         pd = rcu_dereference_bh(pinst->pd);
 116
 117         err = -EINVAL;
 118         if (!(pinst->flags & PADATA_INIT) || pinst->flags & PADATA_INVALID)
 119                 goto out;
 120
 121         if (!cpumask_test_cpu(cb_cpu, pd->cpumask.cbcpu))
 122                 goto out;
 123
 124         err =  -EBUSY;
 125         if ((pinst->flags & PADATA_RESET))
 126                 goto out;
 127
 128         if (atomic_read(&pd->refcnt) >= MAX_OBJ_NUM)
 129                 goto out;
 130
 131         err = 0;
 132         atomic_inc(&pd->refcnt);
 133         padata->pd = pd;
 134         padata->cb_cpu = cb_cpu;
 135
 136         target_cpu = padata_cpu_hash(pd);
 137         queue = per_cpu_ptr(pd->pqueue, target_cpu);
 138
 139         spin_lock(&queue->parallel.lock);
 140         list_add_tail(&padata->list, &queue->parallel.list);
 141         spin_unlock(&queue->parallel.lock);
 142
 143         queue_work_on(target_cpu, pinst->wq, &queue->work);
 144
 145 out:
 146         rcu_read_unlock_bh();
 147
 148         return err;
 149 }
 150 EXPORT_SYMBOL(padata_do_parallel);
 151
 152 /*
 153  * padata_get_next - Get the next object that needs serialization.
 154  *
 155  * Return values are:
 156  *
 157  * A pointer to the control struct of the next object that needs
 158  * serialization, if present in one of the percpu reorder queues.
 159  *
 160  * NULL, if all percpu reorder queues are empty.
 161  *
 162  * -EINPROGRESS, if the next object that needs serialization will
 163  *  be parallel processed by another cpu and is not yet present in
 164  *  the cpu's reorder queue.
 165  *
 166  * -ENODATA, if this cpu has to do the parallel processing for
 167  *  the next object.
 168  */
 169 static struct padata_priv *padata_get_next(struct parallel_data *pd)
 170 {
 171         int cpu, num_cpus;
 172         unsigned int next_nr, next_index;
 173         struct padata_parallel_queue *next_queue;
 174         struct padata_priv *padata;
 175         struct padata_list *reorder;
 176
 177         num_cpus = cpumask_weight(pd->cpumask.pcpu);
 178
 179         /*
 180          * Calculate the percpu reorder queue and the sequence
 181          * number of the next object.
 182          */
 183         next_nr = pd->processed;
 184         next_index = next_nr % num_cpus;
 185         cpu = padata_index_to_cpu(pd, next_index);
 186         next_queue = per_cpu_ptr(pd->pqueue, cpu);
 187
 188         padata = NULL;
 189
 190         reorder = &next_queue->reorder;
 191
 192         if (!list_empty(&reorder->list)) {
 193                 padata = list_entry(reorder->list.next,
 194                                     struct padata_priv, list);
 195
 196                 spin_lock(&reorder->lock);
 197                 list_del_init(&padata->list);
 198                 atomic_dec(&pd->reorder_objects);
 199                 spin_unlock(&reorder->lock);
 200
 201                 pd->processed++;
 202
 203                 goto out;
 204         }
 205
 206         if (__this_cpu_read(pd->pqueue->cpu_index) == next_queue->cpu_index) {
 207                 padata = ERR_PTR(-ENODATA);
 208                 goto out;
 209         }
 210
 211         padata = ERR_PTR(-EINPROGRESS);
 212 out:
 213         return padata;
 214 }
 215
 216 static void padata_reorder(struct parallel_data *pd)
 217 {
 218         int cb_cpu;
 219         struct padata_priv *padata;
 220         struct padata_serial_queue *squeue;
 221         struct padata_instance *pinst = pd->pinst;
 222
 223         /*
 224          * We need to ensure that only one cpu can work on dequeueing of
 225          * the reorder queue the time. Calculating in which percpu reorder
 226          * queue the next object will arrive takes some time. A spinlock
 227          * would be highly contended. Also it is not clear in which order
 228          * the objects arrive to the reorder queues. So a cpu could wait to
 229          * get the lock just to notice that there is nothing to do at the
 230          * moment. Therefore we use a trylock and let the holder of the lock
 231          * care for all the objects enqueued during the holdtime of the lock.
 232          */
 233         if (!spin_trylock_bh(&pd->lock))
 234                 return;
 235
 236         while (1) {
 237                 padata = padata_get_next(pd);
 238
 239                 /*
 240                  * All reorder queues are empty, or the next object that needs
 241                  * serialization is parallel processed by another cpu and is
 242                  * still on it's way to the cpu's reorder queue, nothing to
 243                  * do for now.
 244                  */
 245                 if (!padata || PTR_ERR(padata) == -EINPROGRESS)
 246                         break;
 247
 248                 /*
 249                  * This cpu has to do the parallel processing of the next
 250                  * object. It's waiting in the cpu's parallelization queue,
 251                  * so exit immediately.
 252                  */
 253                 if (PTR_ERR(padata) == -ENODATA) {
 254                         del_timer(&pd->timer);
 255                         spin_unlock_bh(&pd->lock);
 256                         return;
 257                 }
 258
 259                 cb_cpu = padata->cb_cpu;
 260                 squeue = per_cpu_ptr(pd->squeue, cb_cpu);
 261
 262                 spin_lock(&squeue->serial.lock);
 263                 list_add_tail(&padata->list, &squeue->serial.list);
 264                 spin_unlock(&squeue->serial.lock);
 265
 266                 queue_work_on(cb_cpu, pinst->wq, &squeue->work);
 267         }
 268
 269         spin_unlock_bh(&pd->lock);
 270
 271         /*
 272          * The next object that needs serialization might have arrived to
 273          * the reorder queues in the meantime, we will be called again
 274          * from the timer function if no one else cares for it.
 275          */
 276         if (atomic_read(&pd->reorder_objects)
 277                         && !(pinst->flags & PADATA_RESET))
 278                 mod_timer(&pd->timer, jiffies + HZ);
 279         else
 280                 del_timer(&pd->timer);
 281
 282         return;
 283 }
 284
 285 static void padata_reorder_timer(unsigned long arg)
 286 {
 287         struct parallel_data *pd = (struct parallel_data *)arg;
 288
 289         padata_reorder(pd);
 290 }
 291
 292 static void padata_serial_worker(struct work_struct *serial_work)
 293 {
 294         struct padata_serial_queue *squeue;
 295         struct parallel_data *pd;
 296         LIST_HEAD(local_list);
 297
 298         local_bh_disable();
 299         squeue = container_of(serial_work, struct padata_serial_queue, work);
 300         pd = squeue->pd;
 301
 302         spin_lock(&squeue->serial.lock);
 303         list_replace_init(&squeue->serial.list, &local_list);
 304         spin_unlock(&squeue->serial.lock);
 305
 306         while (!list_empty(&local_list)) {
 307                 struct padata_priv *padata;
 308
 309                 padata = list_entry(local_list.next,
 310                                     struct padata_priv, list);
 311
 312                 list_del_init(&padata->list);
 313
 314                 padata->serial(padata);
 315                 atomic_dec(&pd->refcnt);
 316         }
 317         local_bh_enable();
 318 }
 319
 320 /**
 321  * padata_do_serial - padata serialization function
 322  *
 323  * @padata: object to be serialized.
 324  *
 325  * padata_do_serial must be called for every parallelized object.
 326  * The serialization callback function will run with BHs off.
 327  */
 328 void padata_do_serial(struct padata_priv *padata)
 329 {
 330         int cpu;
 331         struct padata_parallel_queue *pqueue;
 332         struct parallel_data *pd;
 333
 334         pd = padata->pd;
 335
 336         cpu = get_cpu();
 337         pqueue = per_cpu_ptr(pd->pqueue, cpu);
 338
 339         spin_lock(&pqueue->reorder.lock);
 340         atomic_inc(&pd->reorder_objects);
 341         list_add_tail(&padata->list, &pqueue->reorder.list);
 342         spin_unlock(&pqueue->reorder.lock);
 343
 344         put_cpu();
 345
 346         padata_reorder(pd);
 347 }
 348 EXPORT_SYMBOL(padata_do_serial);
 349
 350 static int padata_setup_cpumasks(struct parallel_data *pd,
 351                                  const struct cpumask *pcpumask,
 352                                  const struct cpumask *cbcpumask)
 353 {
 354         if (!alloc_cpumask_var(&pd->cpumask.pcpu, GFP_KERNEL))
 355                 return -ENOMEM;
 356
 357         cpumask_and(pd->cpumask.pcpu, pcpumask, cpu_online_mask);
 358         if (!alloc_cpumask_var(&pd->cpumask.cbcpu, GFP_KERNEL)) {
 359                 free_cpumask_var(pd->cpumask.cbcpu);
 360                 return -ENOMEM;
 361         }
 362
 363         cpumask_and(pd->cpumask.cbcpu, cbcpumask, cpu_online_mask);
 364         return 0;
 365 }
 366
 367 static void __padata_list_init(struct padata_list *pd_list)
 368 {
 369         INIT_LIST_HEAD(&pd_list->list);
 370         spin_lock_init(&pd_list->lock);
 371 }
 372
 373 /* Initialize all percpu queues used by serial workers */
 374 static void padata_init_squeues(struct parallel_data *pd)
 375 {
 376         int cpu;
 377         struct padata_serial_queue *squeue;
 378
 379         for_each_cpu(cpu, pd->cpumask.cbcpu) {
 380                 squeue = per_cpu_ptr(pd->squeue, cpu);
 381                 squeue->pd = pd;
 382                 __padata_list_init(&squeue->serial);
 383                 INIT_WORK(&squeue->work, padata_serial_worker);
 384         }
 385 }
 386
 387 /* Initialize all percpu queues used by parallel workers */
 388 static void padata_init_pqueues(struct parallel_data *pd)
 389 {
 390         int cpu_index, cpu;
 391         struct padata_parallel_queue *pqueue;
 392
 393         cpu_index = 0;
 394         for_each_cpu(cpu, pd->cpumask.pcpu) {
 395                 pqueue = per_cpu_ptr(pd->pqueue, cpu);
 396                 pqueue->pd = pd;
 397                 pqueue->cpu_index = cpu_index;
 398                 cpu_index++;
 399
 400                 __padata_list_init(&pqueue->reorder);
 401                 __padata_list_init(&pqueue->parallel);
 402                 INIT_WORK(&pqueue->work, padata_parallel_worker);
 403                 atomic_set(&pqueue->num_obj, 0);
 404         }
 405 }
 406
 407 /* Allocate and initialize the internal cpumask dependend resources. */
 408 static struct parallel_data *padata_alloc_pd(struct padata_instance *pinst,
 409                                              const struct cpumask *pcpumask,
 410                                              const struct cpumask *cbcpumask)
 411 {
 412         struct parallel_data *pd;
 413
 414         pd = kzalloc(sizeof(struct parallel_data), GFP_KERNEL);
 415         if (!pd)
 416                 goto err;
 417
 418         pd->pqueue = alloc_percpu(struct padata_parallel_queue);
 419         if (!pd->pqueue)
 420                 goto err_free_pd;
 421
 422         pd->squeue = alloc_percpu(struct padata_serial_queue);
 423         if (!pd->squeue)
 424                 goto err_free_pqueue;
 425         if (padata_setup_cpumasks(pd, pcpumask, cbcpumask) < 0)
 426                 goto err_free_squeue;
 427
 428         padata_init_pqueues(pd);
 429         padata_init_squeues(pd);
 430         setup_timer(&pd->timer, padata_reorder_timer, (unsigned long)pd);
 431         atomic_set(&pd->seq_nr, -1);
 432         atomic_set(&pd->reorder_objects, 0);
 433         atomic_set(&pd->refcnt, 0);
 434         pd->pinst = pinst;
 435         spin_lock_init(&pd->lock);
 436
 437         return pd;
 438
 439 err_free_squeue:
 440         free_percpu(pd->squeue);
 441 err_free_pqueue:
 442         free_percpu(pd->pqueue);
 443 err_free_pd:
 444         kfree(pd);
 445 err:
 446         return NULL;
 447 }
 448
 449 static void padata_free_pd(struct parallel_data *pd)
 450 {
 451         free_cpumask_var(pd->cpumask.pcpu);
 452         free_cpumask_var(pd->cpumask.cbcpu);
 453         free_percpu(pd->pqueue);
 454         free_percpu(pd->squeue);
 455         kfree(pd);
 456 }
 457
 458 /* Flush all objects out of the padata queues. */
 459 static void padata_flush_queues(struct parallel_data *pd)
 460 {
 461         int cpu;
 462         struct padata_parallel_queue *pqueue;
 463         struct padata_serial_queue *squeue;
 464
 465         for_each_cpu(cpu, pd->cpumask.pcpu) {
 466                 pqueue = per_cpu_ptr(pd->pqueue, cpu);
 467                 flush_work(&pqueue->work);
 468         }
 469
 470         del_timer_sync(&pd->timer);
 471
 472         if (atomic_read(&pd->reorder_objects))
 473                 padata_reorder(pd);
 474
 475         for_each_cpu(cpu, pd->cpumask.cbcpu) {
 476                 squeue = per_cpu_ptr(pd->squeue, cpu);
 477                 flush_work(&squeue->work);
 478         }
 479
 480         BUG_ON(atomic_read(&pd->refcnt) != 0);
 481 }
 482
 483 static void __padata_start(struct padata_instance *pinst)
 484 {
 485         pinst->flags |= PADATA_INIT;
 486 }
 487
 488 static void __padata_stop(struct padata_instance *pinst)
 489 {
 490         if (!(pinst->flags & PADATA_INIT))
 491                 return;
 492
 493         pinst->flags &= ~PADATA_INIT;
 494
 495         synchronize_rcu();
 496
 497         get_online_cpus();
 498         padata_flush_queues(pinst->pd);
 499         put_online_cpus();
 500 }
 501
 502 /* Replace the internal control structure with a new one. */
 503 static void padata_replace(struct padata_instance *pinst,
 504                            struct parallel_data *pd_new)
 505 {
 506         struct parallel_data *pd_old = pinst->pd;
 507         int notification_mask = 0;
 508
 509         pinst->flags |= PADATA_RESET;
 510
 511         rcu_assign_pointer(pinst->pd, pd_new);
 512
 513         synchronize_rcu();
 514
 515         if (!cpumask_equal(pd_old->cpumask.pcpu, pd_new->cpumask.pcpu))
 516                 notification_mask |= PADATA_CPU_PARALLEL;
 517         if (!cpumask_equal(pd_old->cpumask.cbcpu, pd_new->cpumask.cbcpu))
 518                 notification_mask |= PADATA_CPU_SERIAL;
 519
 520         padata_flush_queues(pd_old);
 521         padata_free_pd(pd_old);
 522
 523         if (notification_mask)
 524                 blocking_notifier_call_chain(&pinst->cpumask_change_notifier,
 525                                              notification_mask,
 526                                              &pd_new->cpumask);
 527
 528         pinst->flags &= ~PADATA_RESET;
 529 }
 530
 531 /**
 532  * padata_register_cpumask_notifier - Registers a notifier that will be called
 533  *                             if either pcpu or cbcpu or both cpumasks change.
 534  *
 535  * @pinst: A poineter to padata instance
 536  * @nblock: A pointer to notifier block.
 537  */
 538 int padata_register_cpumask_notifier(struct padata_instance *pinst,
 539                                      struct notifier_block *nblock)
 540 {
 541         return blocking_notifier_chain_register(&pinst->cpumask_change_notifier,
 542                                                 nblock);
 543 }
 544 EXPORT_SYMBOL(padata_register_cpumask_notifier);
 545
 546 /**
 547  * padata_unregister_cpumask_notifier - Unregisters cpumask notifier
 548  *        registered earlier  using padata_register_cpumask_notifier
 549  *
 550  * @pinst: A pointer to data instance.
 551  * @nlock: A pointer to notifier block.
 552  */
 553 int padata_unregister_cpumask_notifier(struct padata_instance *pinst,
 554                                        struct notifier_block *nblock)
 555 {
 556         return blocking_notifier_chain_unregister(
 557                 &pinst->cpumask_change_notifier,
 558                 nblock);
 559 }
 560 EXPORT_SYMBOL(padata_unregister_cpumask_notifier);
 561
 562
 563 /* If cpumask contains no active cpu, we mark the instance as invalid. */
 564 static bool padata_validate_cpumask(struct padata_instance *pinst,
 565                                     const struct cpumask *cpumask)
 566 {
 567         if (!cpumask_intersects(cpumask, cpu_online_mask)) {
 568                 pinst->flags |= PADATA_INVALID;
 569                 return false;
 570         }
 571
 572         pinst->flags &= ~PADATA_INVALID;
 573         return true;
 574 }
 575
 576 static int __padata_set_cpumasks(struct padata_instance *pinst,
 577                                  cpumask_var_t pcpumask,
 578                                  cpumask_var_t cbcpumask)
 579 {
 580         int valid;
 581         struct parallel_data *pd;
 582
 583         valid = padata_validate_cpumask(pinst, pcpumask);
 584         if (!valid) {
 585                 __padata_stop(pinst);
 586                 goto out_replace;
 587         }
 588
 589         valid = padata_validate_cpumask(pinst, cbcpumask);
 590         if (!valid)
 591                 __padata_stop(pinst);
 592
 593 out_replace:
 594         pd = padata_alloc_pd(pinst, pcpumask, cbcpumask);
 595         if (!pd)
 596                 return -ENOMEM;
 597
 598         cpumask_copy(pinst->cpumask.pcpu, pcpumask);
 599         cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
 600
 601         padata_replace(pinst, pd);
 602
 603         if (valid)
 604                 __padata_start(pinst);
 605
 606         return 0;
 607 }
 608
 609 /**
 610  * padata_set_cpumask: Sets specified by @cpumask_type cpumask to the value
 611  *                     equivalent to @cpumask.
 612  *
 613  * @pinst: padata instance
 614  * @cpumask_type: PADATA_CPU_SERIAL or PADATA_CPU_PARALLEL corresponding
 615  *                to parallel and serial cpumasks respectively.
 616  * @cpumask: the cpumask to use
 617  */
 618 int padata_set_cpumask(struct padata_instance *pinst, int cpumask_type,
 619                        cpumask_var_t cpumask)
 620 {
 621         struct cpumask *serial_mask, *parallel_mask;
 622         int err = -EINVAL;
 623
 624         mutex_lock(&pinst->lock);
 625         get_online_cpus();
 626
 627         switch (cpumask_type) {
 628         case PADATA_CPU_PARALLEL:
 629                 serial_mask = pinst->cpumask.cbcpu;
 630                 parallel_mask = cpumask;
 631                 break;
 632         case PADATA_CPU_SERIAL:
 633                 parallel_mask = pinst->cpumask.pcpu;
 634                 serial_mask = cpumask;
 635                 break;
 636         default:
 637                  goto out;
 638         }
 639
 640         err =  __padata_set_cpumasks(pinst, parallel_mask, serial_mask);
 641
 642 out:
 643         put_online_cpus();
 644         mutex_unlock(&pinst->lock);
 645
 646         return err;
 647 }
 648 EXPORT_SYMBOL(padata_set_cpumask);
 649
 650 /**
 651  * padata_start - start the parallel processing
 652  *
 653  * @pinst: padata instance to start
 654  */
 655 int padata_start(struct padata_instance *pinst)
 656 {
 657         int err = 0;
 658
 659         mutex_lock(&pinst->lock);
 660
 661         if (pinst->flags & PADATA_INVALID)
 662                 err = -EINVAL;
 663
 664          __padata_start(pinst);
 665
 666         mutex_unlock(&pinst->lock);
 667
 668         return err;
 669 }
 670 EXPORT_SYMBOL(padata_start);
 671
 672 /**
 673  * padata_stop - stop the parallel processing
 674  *
 675  * @pinst: padata instance to stop
 676  */
 677 void padata_stop(struct padata_instance *pinst)
 678 {
 679         mutex_lock(&pinst->lock);
 680         __padata_stop(pinst);
 681         mutex_unlock(&pinst->lock);
 682 }
 683 EXPORT_SYMBOL(padata_stop);
 684
 685 #ifdef CONFIG_HOTPLUG_CPU
 686
 687 static int __padata_add_cpu(struct padata_instance *pinst, int cpu)
 688 {
 689         struct parallel_data *pd;
 690
 691         if (cpumask_test_cpu(cpu, cpu_online_mask)) {
 692                 pd = padata_alloc_pd(pinst, pinst->cpumask.pcpu,
 693                                      pinst->cpumask.cbcpu);
 694                 if (!pd)
 695                         return -ENOMEM;
 696
 697                 padata_replace(pinst, pd);
 698
 699                 if (padata_validate_cpumask(pinst, pinst->cpumask.pcpu) &&
 700                     padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
 701                         __padata_start(pinst);
 702         }
 703
 704         return 0;
 705 }
 706
 707 static int __padata_remove_cpu(struct padata_instance *pinst, int cpu)
 708 {
 709         struct parallel_data *pd = NULL;
 710
 711         if (cpumask_test_cpu(cpu, cpu_online_mask)) {
 712
 713                 if (!padata_validate_cpumask(pinst, pinst->cpumask.pcpu) ||
 714                     !padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
 715                         __padata_stop(pinst);
 716
 717                 pd = padata_alloc_pd(pinst, pinst->cpumask.pcpu,
 718                                      pinst->cpumask.cbcpu);
 719                 if (!pd)
 720                         return -ENOMEM;
 721
 722                 padata_replace(pinst, pd);
 723
 724                 cpumask_clear_cpu(cpu, pd->cpumask.cbcpu);
 725                 cpumask_clear_cpu(cpu, pd->cpumask.pcpu);
 726         }
 727
 728         return 0;
 729 }
 730
 731  /**
 732  * padata_remove_cpu - remove a cpu from the one or both(serial and parallel)
 733  *                     padata cpumasks.
 734  *
 735  * @pinst: padata instance
 736  * @cpu: cpu to remove
 737  * @mask: bitmask specifying from which cpumask @cpu should be removed
 738  *        The @mask may be any combination of the following flags:
 739  *          PADATA_CPU_SERIAL   - serial cpumask
 740  *          PADATA_CPU_PARALLEL - parallel cpumask
 741  */
 742 int padata_remove_cpu(struct padata_instance *pinst, int cpu, int mask)
 743 {
 744         int err;
 745
 746         if (!(mask & (PADATA_CPU_SERIAL | PADATA_CPU_PARALLEL)))
 747                 return -EINVAL;
 748
 749         mutex_lock(&pinst->lock);
 750
 751         get_online_cpus();
 752         if (mask & PADATA_CPU_SERIAL)
 753                 cpumask_clear_cpu(cpu, pinst->cpumask.cbcpu);
 754         if (mask & PADATA_CPU_PARALLEL)
 755                 cpumask_clear_cpu(cpu, pinst->cpumask.pcpu);
 756
 757         err = __padata_remove_cpu(pinst, cpu);
 758         put_online_cpus();
 759
 760         mutex_unlock(&pinst->lock);
 761
 762         return err;
 763 }
 764 EXPORT_SYMBOL(padata_remove_cpu);
 765
 766 static inline int pinst_has_cpu(struct padata_instance *pinst, int cpu)
 767 {
 768         return cpumask_test_cpu(cpu, pinst->cpumask.pcpu) ||
 769                 cpumask_test_cpu(cpu, pinst->cpumask.cbcpu);
 770 }
 771
 772
 773 static int padata_cpu_callback(struct notifier_block *nfb,
 774                                unsigned long action, void *hcpu)
 775 {
 776         int err;
 777         struct padata_instance *pinst;
 778         int cpu = (unsigned long)hcpu;
 779
 780         pinst = container_of(nfb, struct padata_instance, cpu_notifier);
 781
 782         switch (action) {
 783         case CPU_ONLINE:
 784         case CPU_ONLINE_FROZEN:
 785         case CPU_DOWN_FAILED:
 786         case CPU_DOWN_FAILED_FROZEN:
 787                 if (!pinst_has_cpu(pinst, cpu))
 788                         break;
 789                 mutex_lock(&pinst->lock);
 790                 err = __padata_add_cpu(pinst, cpu);
 791                 mutex_unlock(&pinst->lock);
 792                 if (err)
 793                         return notifier_from_errno(err);
 794                 break;
 795
 796         case CPU_DOWN_PREPARE:
 797         case CPU_DOWN_PREPARE_FROZEN:
 798         case CPU_UP_CANCELED:
 799         case CPU_UP_CANCELED_FROZEN:
 800                 if (!pinst_has_cpu(pinst, cpu))
 801                         break;
 802                 mutex_lock(&pinst->lock);
 803                 err = __padata_remove_cpu(pinst, cpu);
 804                 mutex_unlock(&pinst->lock);
 805                 if (err)
 806                         return notifier_from_errno(err);
 807                 break;
 808         }
 809
 810         return NOTIFY_OK;
 811 }
 812 #endif
 813
 814 static void __padata_free(struct padata_instance *pinst)
 815 {
 816 #ifdef CONFIG_HOTPLUG_CPU
 817         unregister_hotcpu_notifier(&pinst->cpu_notifier);
 818 #endif
 819
 820         padata_stop(pinst);
 821         padata_free_pd(pinst->pd);
 822         free_cpumask_var(pinst->cpumask.pcpu);
 823         free_cpumask_var(pinst->cpumask.cbcpu);
 824         kfree(pinst);
 825 }
 826
 827 #define kobj2pinst(_kobj)                                       \
 828         container_of(_kobj, struct padata_instance, kobj)
 829 #define attr2pentry(_attr)                                      \
 830         container_of(_attr, struct padata_sysfs_entry, attr)
 831
 832 static void padata_sysfs_release(struct kobject *kobj)
 833 {
 834         struct padata_instance *pinst = kobj2pinst(kobj);
 835         __padata_free(pinst);
 836 }
 837
 838 struct padata_sysfs_entry {
 839         struct attribute attr;
 840         ssize_t (*show)(struct padata_instance *, struct attribute *, char *);
 841         ssize_t (*store)(struct padata_instance *, struct attribute *,
 842                          const char *, size_t);
 843 };
 844
 845 static ssize_t show_cpumask(struct padata_instance *pinst,
 846                             struct attribute *attr,  char *buf)
 847 {
 848         struct cpumask *cpumask;
 849         ssize_t len;
 850
 851         mutex_lock(&pinst->lock);
 852         if (!strcmp(attr->name, "serial_cpumask"))
 853                 cpumask = pinst->cpumask.cbcpu;
 854         else
 855                 cpumask = pinst->cpumask.pcpu;
 856
 857         len = snprintf(buf, PAGE_SIZE, "%*pb\n",
 858                        nr_cpu_ids, cpumask_bits(cpumask));
 859         mutex_unlock(&pinst->lock);
 860         return len < PAGE_SIZE ? len : -EINVAL;
 861 }
 862
 863 static ssize_t store_cpumask(struct padata_instance *pinst,
 864                              struct attribute *attr,
 865                              const char *buf, size_t count)
 866 {
 867         cpumask_var_t new_cpumask;
 868         ssize_t ret;
 869         int mask_type;
 870
 871         if (!alloc_cpumask_var(&new_cpumask, GFP_KERNEL))
 872                 return -ENOMEM;
 873
 874         ret = bitmap_parse(buf, count, cpumask_bits(new_cpumask),
 875                            nr_cpumask_bits);
 876         if (ret < 0)
 877                 goto out;
 878
 879         mask_type = !strcmp(attr->name, "serial_cpumask") ?
 880                 PADATA_CPU_SERIAL : PADATA_CPU_PARALLEL;
 881         ret = padata_set_cpumask(pinst, mask_type, new_cpumask);
 882         if (!ret)
 883                 ret = count;
 884
 885 out:
 886         free_cpumask_var(new_cpumask);
 887         return ret;
 888 }
 889
 890 #define PADATA_ATTR_RW(_name, _show_name, _store_name)          \
 891         static struct padata_sysfs_entry _name##_attr =         \
 892                 __ATTR(_name, 0644, _show_name, _store_name)
 893 #define PADATA_ATTR_RO(_name, _show_name)               \
 894         static struct padata_sysfs_entry _name##_attr = \
 895                 __ATTR(_name, 0400, _show_name, NULL)
 896
 897 PADATA_ATTR_RW(serial_cpumask, show_cpumask, store_cpumask);
 898 PADATA_ATTR_RW(parallel_cpumask, show_cpumask, store_cpumask);
 899
 900 /*
 901  * Padata sysfs provides the following objects:
 902  * serial_cpumask   [RW] - cpumask for serial workers
 903  * parallel_cpumask [RW] - cpumask for parallel workers
 904  */
 905 static struct attribute *padata_default_attrs[] = {
 906         &serial_cpumask_attr.attr,
 907         &parallel_cpumask_attr.attr,
 908         NULL,
 909 };
 910
 911 static ssize_t padata_sysfs_show(struct kobject *kobj,
 912                                  struct attribute *attr, char *buf)
 913 {
 914         struct padata_instance *pinst;
 915         struct padata_sysfs_entry *pentry;
 916         ssize_t ret = -EIO;
 917
 918         pinst = kobj2pinst(kobj);
 919         pentry = attr2pentry(attr);
 920         if (pentry->show)
 921                 ret = pentry->show(pinst, attr, buf);
 922
 923         return ret;
 924 }
 925
 926 static ssize_t padata_sysfs_store(struct kobject *kobj, struct attribute *attr,
 927                                   const char *buf, size_t count)
 928 {
 929         struct padata_instance *pinst;
 930         struct padata_sysfs_entry *pentry;
 931         ssize_t ret = -EIO;
 932
 933         pinst = kobj2pinst(kobj);
 934         pentry = attr2pentry(attr);
 935         if (pentry->show)
 936                 ret = pentry->store(pinst, attr, buf, count);
 937
 938         return ret;
 939 }
 940
 941 static const struct sysfs_ops padata_sysfs_ops = {
 942         .show = padata_sysfs_show,
 943         .store = padata_sysfs_store,
 944 };
 945
 946 static struct kobj_type padata_attr_type = {
 947         .sysfs_ops = &padata_sysfs_ops,
 948         .default_attrs = padata_default_attrs,
 949         .release = padata_sysfs_release,
 950 };
 951
 952 /**
 953  * padata_alloc_possible - Allocate and initialize padata instance.
 954  *                         Use the cpu_possible_mask for serial and
 955  *                         parallel workers.
 956  *
 957  * @wq: workqueue to use for the allocated padata instance
 958  */
 959 struct padata_instance *padata_alloc_possible(struct workqueue_struct *wq)
 960 {
 961         return padata_alloc(wq, cpu_possible_mask, cpu_possible_mask);
 962 }
 963 EXPORT_SYMBOL(padata_alloc_possible);
 964
 965 /**
 966  * padata_alloc - allocate and initialize a padata instance and specify
 967  *                cpumasks for serial and parallel workers.
 968  *
 969  * @wq: workqueue to use for the allocated padata instance
 970  * @pcpumask: cpumask that will be used for padata parallelization
 971  * @cbcpumask: cpumask that will be used for padata serialization
 972  */
 973 struct padata_instance *padata_alloc(struct workqueue_struct *wq,
 974                                      const struct cpumask *pcpumask,
 975                                      const struct cpumask *cbcpumask)
 976 {
 977         struct padata_instance *pinst;
 978         struct parallel_data *pd = NULL;
 979
 980         pinst = kzalloc(sizeof(struct padata_instance), GFP_KERNEL);
 981         if (!pinst)
 982                 goto err;
 983
 984         get_online_cpus();
 985         if (!alloc_cpumask_var(&pinst->cpumask.pcpu, GFP_KERNEL))
 986                 goto err_free_inst;
 987         if (!alloc_cpumask_var(&pinst->cpumask.cbcpu, GFP_KERNEL)) {
 988                 free_cpumask_var(pinst->cpumask.pcpu);
 989                 goto err_free_inst;
 990         }
 991         if (!padata_validate_cpumask(pinst, pcpumask) ||
 992             !padata_validate_cpumask(pinst, cbcpumask))
 993                 goto err_free_masks;
 994
 995         pd = padata_alloc_pd(pinst, pcpumask, cbcpumask);
 996         if (!pd)
 997                 goto err_free_masks;
 998
 999         rcu_assign_pointer(pinst->pd, pd);
1000
1001         pinst->wq = wq;
1002
1003         cpumask_copy(pinst->cpumask.pcpu, pcpumask);
1004         cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
1005
1006         pinst->flags = 0;
1007
1008         put_online_cpus();
1009
1010         BLOCKING_INIT_NOTIFIER_HEAD(&pinst->cpumask_change_notifier);
1011         kobject_init(&pinst->kobj, &padata_attr_type);
1012         mutex_init(&pinst->lock);
1013
1014 #ifdef CONFIG_HOTPLUG_CPU
1015         pinst->cpu_notifier.notifier_call = padata_cpu_callback;
1016         pinst->cpu_notifier.priority = 0;
1017         register_hotcpu_notifier(&pinst->cpu_notifier);
1018 #endif
1019
1020         return pinst;
1021
1022 err_free_masks:
1023         free_cpumask_var(pinst->cpumask.pcpu);
1024         free_cpumask_var(pinst->cpumask.cbcpu);
1025 err_free_inst:
1026         kfree(pinst);
1027         put_online_cpus();
1028 err:
1029         return NULL;
1030 }
1031
1032 /**
1033  * padata_free - free a padata instance
1034  *
1035  * @padata_inst: padata instance to free
1036  */
1037 void padata_free(struct padata_instance *pinst)
1038 {
1039         kobject_put(&pinst->kobj);
1040 }
1041 EXPORT_SYMBOL(padata_free);