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 #include <linux/module.h>
  34
  35 #define MAX_OBJ_NUM 1000
  36
  37 static int padata_index_to_cpu(struct parallel_data *pd, int cpu_index)
  38 {
  39         int cpu, target_cpu;
  40
  41         target_cpu = cpumask_first(pd->cpumask.pcpu);
  42         for (cpu = 0; cpu < cpu_index; cpu++)
  43                 target_cpu = cpumask_next(target_cpu, pd->cpumask.pcpu);
  44
  45         return target_cpu;
  46 }
  47
  48 static int padata_cpu_hash(struct parallel_data *pd)
  49 {
  50         unsigned int seq_nr;
  51         int cpu_index;
  52
  53         /*
  54          * Hash the sequence numbers to the cpus by taking
  55          * seq_nr mod. number of cpus in use.
  56          */
  57
  58         seq_nr = atomic_inc_return(&pd->seq_nr);
  59         cpu_index = seq_nr % cpumask_weight(pd->cpumask.pcpu);
  60
  61         return padata_index_to_cpu(pd, cpu_index);
  62 }
  63
  64 static void padata_parallel_worker(struct work_struct *parallel_work)
  65 {
  66         struct padata_parallel_queue *pqueue;
  67         struct parallel_data *pd;
  68         struct padata_instance *pinst;
  69         LIST_HEAD(local_list);
  70
  71         local_bh_disable();
  72         pqueue = container_of(parallel_work,
  73                               struct padata_parallel_queue, work);
  74         pd = pqueue->pd;
  75         pinst = pd->pinst;
  76
  77         spin_lock(&pqueue->parallel.lock);
  78         list_replace_init(&pqueue->parallel.list, &local_list);
  79         spin_unlock(&pqueue->parallel.lock);
  80
  81         while (!list_empty(&local_list)) {
  82                 struct padata_priv *padata;
  83
  84                 padata = list_entry(local_list.next,
  85                                     struct padata_priv, list);
  86
  87                 list_del_init(&padata->list);
  88
  89                 padata->parallel(padata);
  90         }
  91
  92         local_bh_enable();
  93 }
  94
  95 /**
  96  * padata_do_parallel - padata parallelization function
  97  *
  98  * @pinst: padata instance
  99  * @padata: object to be parallelized
 100  * @cb_cpu: cpu the serialization callback function will run on,
 101  *          must be in the serial cpumask of padata(i.e. cpumask.cbcpu).
 102  *
 103  * The parallelization callback function will run with BHs off.
 104  * Note: Every object which is parallelized by padata_do_parallel
 105  * must be seen by padata_do_serial.
 106  */
 107 int padata_do_parallel(struct padata_instance *pinst,
 108                        struct padata_priv *padata, int cb_cpu)
 109 {
 110         int target_cpu, err;
 111         struct padata_parallel_queue *queue;
 112         struct parallel_data *pd;
 113
 114         rcu_read_lock_bh();
 115
 116         pd = rcu_dereference_bh(pinst->pd);
 117
 118         err = -EINVAL;
 119         if (!(pinst->flags & PADATA_INIT) || pinst->flags & PADATA_INVALID)
 120                 goto out;
 121
 122         if (!cpumask_test_cpu(cb_cpu, pd->cpumask.cbcpu))
 123                 goto out;
 124
 125         err =  -EBUSY;
 126         if ((pinst->flags & PADATA_RESET))
 127                 goto out;
 128
 129         if (atomic_read(&pd->refcnt) >= MAX_OBJ_NUM)
 130                 goto out;
 131
 132         err = 0;
 133         atomic_inc(&pd->refcnt);
 134         padata->pd = pd;
 135         padata->cb_cpu = cb_cpu;
 136
 137         target_cpu = padata_cpu_hash(pd);
 138         queue = per_cpu_ptr(pd->pqueue, target_cpu);
 139
 140         spin_lock(&queue->parallel.lock);
 141         list_add_tail(&padata->list, &queue->parallel.list);
 142         spin_unlock(&queue->parallel.lock);
 143
 144         queue_work_on(target_cpu, pinst->wq, &queue->work);
 145
 146 out:
 147         rcu_read_unlock_bh();
 148
 149         return err;
 150 }
 151 EXPORT_SYMBOL(padata_do_parallel);
 152
 153 /*
 154  * padata_get_next - Get the next object that needs serialization.
 155  *
 156  * Return values are:
 157  *
 158  * A pointer to the control struct of the next object that needs
 159  * serialization, if present in one of the percpu reorder queues.
 160  *
 161  * NULL, if all percpu reorder queues are empty.
 162  *
 163  * -EINPROGRESS, if the next object that needs serialization will
 164  *  be parallel processed by another cpu and is not yet present in
 165  *  the cpu's reorder queue.
 166  *
 167  * -ENODATA, if this cpu has to do the parallel processing for
 168  *  the next object.
 169  */
 170 static struct padata_priv *padata_get_next(struct parallel_data *pd)
 171 {
 172         int cpu, num_cpus;
 173         unsigned int next_nr, next_index;
 174         struct padata_parallel_queue *next_queue;
 175         struct padata_priv *padata;
 176         struct padata_list *reorder;
 177
 178         num_cpus = cpumask_weight(pd->cpumask.pcpu);
 179
 180         /*
 181          * Calculate the percpu reorder queue and the sequence
 182          * number of the next object.
 183          */
 184         next_nr = pd->processed;
 185         next_index = next_nr % num_cpus;
 186         cpu = padata_index_to_cpu(pd, next_index);
 187         next_queue = per_cpu_ptr(pd->pqueue, cpu);
 188
 189         padata = NULL;
 190
 191         reorder = &next_queue->reorder;
 192
 193         spin_lock(&reorder->lock);
 194         if (!list_empty(&reorder->list)) {
 195                 padata = list_entry(reorder->list.next,
 196                                     struct padata_priv, list);
 197
 198                 list_del_init(&padata->list);
 199                 atomic_dec(&pd->reorder_objects);
 200
 201                 pd->processed++;
 202
 203                 spin_unlock(&reorder->lock);
 204                 goto out;
 205         }
 206         spin_unlock(&reorder->lock);
 207
 208         if (__this_cpu_read(pd->pqueue->cpu_index) == next_queue->cpu_index) {
 209                 padata = ERR_PTR(-ENODATA);
 210                 goto out;
 211         }
 212
 213         padata = ERR_PTR(-EINPROGRESS);
 214 out:
 215         return padata;
 216 }
 217
 218 static void padata_reorder(struct parallel_data *pd)
 219 {
 220         int cb_cpu;
 221         struct padata_priv *padata;
 222         struct padata_serial_queue *squeue;
 223         struct padata_instance *pinst = pd->pinst;
 224
 225         /*
 226          * We need to ensure that only one cpu can work on dequeueing of
 227          * the reorder queue the time. Calculating in which percpu reorder
 228          * queue the next object will arrive takes some time. A spinlock
 229          * would be highly contended. Also it is not clear in which order
 230          * the objects arrive to the reorder queues. So a cpu could wait to
 231          * get the lock just to notice that there is nothing to do at the
 232          * moment. Therefore we use a trylock and let the holder of the lock
 233          * care for all the objects enqueued during the holdtime of the lock.
 234          */
 235         if (!spin_trylock_bh(&pd->lock))
 236                 return;
 237
 238         while (1) {
 239                 padata = padata_get_next(pd);
 240
 241                 /*
 242                  * All reorder queues are empty, or the next object that needs
 243                  * serialization is parallel processed by another cpu and is
 244                  * still on it's way to the cpu's reorder queue, nothing to
 245                  * do for now.
 246                  */
 247                 if (!padata || PTR_ERR(padata) == -EINPROGRESS)
 248                         break;
 249
 250                 /*
 251                  * This cpu has to do the parallel processing of the next
 252                  * object. It's waiting in the cpu's parallelization queue,
 253                  * so exit immediately.
 254                  */
 255                 if (PTR_ERR(padata) == -ENODATA) {
 256                         del_timer(&pd->timer);
 257                         spin_unlock_bh(&pd->lock);
 258                         return;
 259                 }
 260
 261                 cb_cpu = padata->cb_cpu;
 262                 squeue = per_cpu_ptr(pd->squeue, cb_cpu);
 263
 264                 spin_lock(&squeue->serial.lock);
 265                 list_add_tail(&padata->list, &squeue->serial.list);
 266                 spin_unlock(&squeue->serial.lock);
 267
 268                 queue_work_on(cb_cpu, pinst->wq, &squeue->work);
 269         }
 270
 271         spin_unlock_bh(&pd->lock);
 272
 273         /*
 274          * The next object that needs serialization might have arrived to
 275          * the reorder queues in the meantime, we will be called again
 276          * from the timer function if no one else cares for it.
 277          */
 278         if (atomic_read(&pd->reorder_objects)
 279                         && !(pinst->flags & PADATA_RESET))
 280                 mod_timer(&pd->timer, jiffies + HZ);
 281         else
 282                 del_timer(&pd->timer);
 283
 284         return;
 285 }
 286
 287 static void padata_reorder_timer(unsigned long arg)
 288 {
 289         struct parallel_data *pd = (struct parallel_data *)arg;
 290
 291         padata_reorder(pd);
 292 }
 293
 294 static void padata_serial_worker(struct work_struct *serial_work)
 295 {
 296         struct padata_serial_queue *squeue;
 297         struct parallel_data *pd;
 298         LIST_HEAD(local_list);
 299
 300         local_bh_disable();
 301         squeue = container_of(serial_work, struct padata_serial_queue, work);
 302         pd = squeue->pd;
 303
 304         spin_lock(&squeue->serial.lock);
 305         list_replace_init(&squeue->serial.list, &local_list);
 306         spin_unlock(&squeue->serial.lock);
 307
 308         while (!list_empty(&local_list)) {
 309                 struct padata_priv *padata;
 310
 311                 padata = list_entry(local_list.next,
 312                                     struct padata_priv, list);
 313
 314                 list_del_init(&padata->list);
 315
 316                 padata->serial(padata);
 317                 atomic_dec(&pd->refcnt);
 318         }
 319         local_bh_enable();
 320 }
 321
 322 /**
 323  * padata_do_serial - padata serialization function
 324  *
 325  * @padata: object to be serialized.
 326  *
 327  * padata_do_serial must be called for every parallelized object.
 328  * The serialization callback function will run with BHs off.
 329  */
 330 void padata_do_serial(struct padata_priv *padata)
 331 {
 332         int cpu;
 333         struct padata_parallel_queue *pqueue;
 334         struct parallel_data *pd;
 335
 336         pd = padata->pd;
 337
 338         cpu = get_cpu();
 339         pqueue = per_cpu_ptr(pd->pqueue, cpu);
 340
 341         spin_lock(&pqueue->reorder.lock);
 342         atomic_inc(&pd->reorder_objects);
 343         list_add_tail(&padata->list, &pqueue->reorder.list);
 344         spin_unlock(&pqueue->reorder.lock);
 345
 346         put_cpu();
 347
 348         padata_reorder(pd);
 349 }
 350 EXPORT_SYMBOL(padata_do_serial);
 351
 352 static int padata_setup_cpumasks(struct parallel_data *pd,
 353                                  const struct cpumask *pcpumask,
 354                                  const struct cpumask *cbcpumask)
 355 {
 356         if (!alloc_cpumask_var(&pd->cpumask.pcpu, GFP_KERNEL))
 357                 return -ENOMEM;
 358
 359         cpumask_and(pd->cpumask.pcpu, pcpumask, cpu_online_mask);
 360         if (!alloc_cpumask_var(&pd->cpumask.cbcpu, GFP_KERNEL)) {
 361                 free_cpumask_var(pd->cpumask.pcpu);
 362                 return -ENOMEM;
 363         }
 364
 365         cpumask_and(pd->cpumask.cbcpu, cbcpumask, cpu_online_mask);
 366         return 0;
 367 }
 368
 369 static void __padata_list_init(struct padata_list *pd_list)
 370 {
 371         INIT_LIST_HEAD(&pd_list->list);
 372         spin_lock_init(&pd_list->lock);
 373 }
 374
 375 /* Initialize all percpu queues used by serial workers */
 376 static void padata_init_squeues(struct parallel_data *pd)
 377 {
 378         int cpu;
 379         struct padata_serial_queue *squeue;
 380
 381         for_each_cpu(cpu, pd->cpumask.cbcpu) {
 382                 squeue = per_cpu_ptr(pd->squeue, cpu);
 383                 squeue->pd = pd;
 384                 __padata_list_init(&squeue->serial);
 385                 INIT_WORK(&squeue->work, padata_serial_worker);
 386         }
 387 }
 388
 389 /* Initialize all percpu queues used by parallel workers */
 390 static void padata_init_pqueues(struct parallel_data *pd)
 391 {
 392         int cpu_index, cpu;
 393         struct padata_parallel_queue *pqueue;
 394
 395         cpu_index = 0;
 396         for_each_cpu(cpu, pd->cpumask.pcpu) {
 397                 pqueue = per_cpu_ptr(pd->pqueue, cpu);
 398                 pqueue->pd = pd;
 399                 pqueue->cpu_index = cpu_index;
 400                 cpu_index++;
 401
 402                 __padata_list_init(&pqueue->reorder);
 403                 __padata_list_init(&pqueue->parallel);
 404                 INIT_WORK(&pqueue->work, padata_parallel_worker);
 405                 atomic_set(&pqueue->num_obj, 0);
 406         }
 407 }
 408
 409 /* Allocate and initialize the internal cpumask dependend resources. */
 410 static struct parallel_data *padata_alloc_pd(struct padata_instance *pinst,
 411                                              const struct cpumask *pcpumask,
 412                                              const struct cpumask *cbcpumask)
 413 {
 414         struct parallel_data *pd;
 415
 416         pd = kzalloc(sizeof(struct parallel_data), GFP_KERNEL);
 417         if (!pd)
 418                 goto err;
 419
 420         pd->pqueue = alloc_percpu(struct padata_parallel_queue);
 421         if (!pd->pqueue)
 422                 goto err_free_pd;
 423
 424         pd->squeue = alloc_percpu(struct padata_serial_queue);
 425         if (!pd->squeue)
 426                 goto err_free_pqueue;
 427         if (padata_setup_cpumasks(pd, pcpumask, cbcpumask) < 0)
 428                 goto err_free_squeue;
 429
 430         padata_init_pqueues(pd);
 431         padata_init_squeues(pd);
 432         setup_timer(&pd->timer, padata_reorder_timer, (unsigned long)pd);
 433         atomic_set(&pd->seq_nr, -1);
 434         atomic_set(&pd->reorder_objects, 0);
 435         atomic_set(&pd->refcnt, 0);
 436         pd->pinst = pinst;
 437         spin_lock_init(&pd->lock);
 438
 439         return pd;
 440
 441 err_free_squeue:
 442         free_percpu(pd->squeue);
 443 err_free_pqueue:
 444         free_percpu(pd->pqueue);
 445 err_free_pd:
 446         kfree(pd);
 447 err:
 448         return NULL;
 449 }
 450
 451 static void padata_free_pd(struct parallel_data *pd)
 452 {
 453         free_cpumask_var(pd->cpumask.pcpu);
 454         free_cpumask_var(pd->cpumask.cbcpu);
 455         free_percpu(pd->pqueue);
 456         free_percpu(pd->squeue);
 457         kfree(pd);
 458 }
 459
 460 /* Flush all objects out of the padata queues. */
 461 static void padata_flush_queues(struct parallel_data *pd)
 462 {
 463         int cpu;
 464         struct padata_parallel_queue *pqueue;
 465         struct padata_serial_queue *squeue;
 466
 467         for_each_cpu(cpu, pd->cpumask.pcpu) {
 468                 pqueue = per_cpu_ptr(pd->pqueue, cpu);
 469                 flush_work(&pqueue->work);
 470         }
 471
 472         del_timer_sync(&pd->timer);
 473
 474         if (atomic_read(&pd->reorder_objects))
 475                 padata_reorder(pd);
 476
 477         for_each_cpu(cpu, pd->cpumask.cbcpu) {
 478                 squeue = per_cpu_ptr(pd->squeue, cpu);
 479                 flush_work(&squeue->work);
 480         }
 481
 482         BUG_ON(atomic_read(&pd->refcnt) != 0);
 483 }
 484
 485 static void __padata_start(struct padata_instance *pinst)
 486 {
 487         pinst->flags |= PADATA_INIT;
 488 }
 489
 490 static void __padata_stop(struct padata_instance *pinst)
 491 {
 492         if (!(pinst->flags & PADATA_INIT))
 493                 return;
 494
 495         pinst->flags &= ~PADATA_INIT;
 496
 497         synchronize_rcu();
 498
 499         get_online_cpus();
 500         padata_flush_queues(pinst->pd);
 501         put_online_cpus();
 502 }
 503
 504 /* Replace the internal control structure with a new one. */
 505 static void padata_replace(struct padata_instance *pinst,
 506                            struct parallel_data *pd_new)
 507 {
 508         struct parallel_data *pd_old = pinst->pd;
 509         int notification_mask = 0;
 510
 511         pinst->flags |= PADATA_RESET;
 512
 513         rcu_assign_pointer(pinst->pd, pd_new);
 514
 515         synchronize_rcu();
 516
 517         if (!cpumask_equal(pd_old->cpumask.pcpu, pd_new->cpumask.pcpu))
 518                 notification_mask |= PADATA_CPU_PARALLEL;
 519         if (!cpumask_equal(pd_old->cpumask.cbcpu, pd_new->cpumask.cbcpu))
 520                 notification_mask |= PADATA_CPU_SERIAL;
 521
 522         padata_flush_queues(pd_old);
 523         padata_free_pd(pd_old);
 524
 525         if (notification_mask)
 526                 blocking_notifier_call_chain(&pinst->cpumask_change_notifier,
 527                                              notification_mask,
 528                                              &pd_new->cpumask);
 529
 530         pinst->flags &= ~PADATA_RESET;
 531 }
 532
 533 /**
 534  * padata_register_cpumask_notifier - Registers a notifier that will be called
 535  *                             if either pcpu or cbcpu or both cpumasks change.
 536  *
 537  * @pinst: A poineter to padata instance
 538  * @nblock: A pointer to notifier block.
 539  */
 540 int padata_register_cpumask_notifier(struct padata_instance *pinst,
 541                                      struct notifier_block *nblock)
 542 {
 543         return blocking_notifier_chain_register(&pinst->cpumask_change_notifier,
 544                                                 nblock);
 545 }
 546 EXPORT_SYMBOL(padata_register_cpumask_notifier);
 547
 548 /**
 549  * padata_unregister_cpumask_notifier - Unregisters cpumask notifier
 550  *        registered earlier  using padata_register_cpumask_notifier
 551  *
 552  * @pinst: A pointer to data instance.
 553  * @nlock: A pointer to notifier block.
 554  */
 555 int padata_unregister_cpumask_notifier(struct padata_instance *pinst,
 556                                        struct notifier_block *nblock)
 557 {
 558         return blocking_notifier_chain_unregister(
 559                 &pinst->cpumask_change_notifier,
 560                 nblock);
 561 }
 562 EXPORT_SYMBOL(padata_unregister_cpumask_notifier);
 563
 564
 565 /* If cpumask contains no active cpu, we mark the instance as invalid. */
 566 static bool padata_validate_cpumask(struct padata_instance *pinst,
 567                                     const struct cpumask *cpumask)
 568 {
 569         if (!cpumask_intersects(cpumask, cpu_online_mask)) {
 570                 pinst->flags |= PADATA_INVALID;
 571                 return false;
 572         }
 573
 574         pinst->flags &= ~PADATA_INVALID;
 575         return true;
 576 }
 577
 578 static int __padata_set_cpumasks(struct padata_instance *pinst,
 579                                  cpumask_var_t pcpumask,
 580                                  cpumask_var_t cbcpumask)
 581 {
 582         int valid;
 583         struct parallel_data *pd;
 584
 585         valid = padata_validate_cpumask(pinst, pcpumask);
 586         if (!valid) {
 587                 __padata_stop(pinst);
 588                 goto out_replace;
 589         }
 590
 591         valid = padata_validate_cpumask(pinst, cbcpumask);
 592         if (!valid)
 593                 __padata_stop(pinst);
 594
 595 out_replace:
 596         pd = padata_alloc_pd(pinst, pcpumask, cbcpumask);
 597         if (!pd)
 598                 return -ENOMEM;
 599
 600         cpumask_copy(pinst->cpumask.pcpu, pcpumask);
 601         cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
 602
 603         padata_replace(pinst, pd);
 604
 605         if (valid)
 606                 __padata_start(pinst);
 607
 608         return 0;
 609 }
 610
 611 /**
 612  * padata_set_cpumask: Sets specified by @cpumask_type cpumask to the value
 613  *                     equivalent to @cpumask.
 614  *
 615  * @pinst: padata instance
 616  * @cpumask_type: PADATA_CPU_SERIAL or PADATA_CPU_PARALLEL corresponding
 617  *                to parallel and serial cpumasks respectively.
 618  * @cpumask: the cpumask to use
 619  */
 620 int padata_set_cpumask(struct padata_instance *pinst, int cpumask_type,
 621                        cpumask_var_t cpumask)
 622 {
 623         struct cpumask *serial_mask, *parallel_mask;
 624         int err = -EINVAL;
 625
 626         mutex_lock(&pinst->lock);
 627         get_online_cpus();
 628
 629         switch (cpumask_type) {
 630         case PADATA_CPU_PARALLEL:
 631                 serial_mask = pinst->cpumask.cbcpu;
 632                 parallel_mask = cpumask;
 633                 break;
 634         case PADATA_CPU_SERIAL:
 635                 parallel_mask = pinst->cpumask.pcpu;
 636                 serial_mask = cpumask;
 637                 break;
 638         default:
 639                  goto out;
 640         }
 641
 642         err =  __padata_set_cpumasks(pinst, parallel_mask, serial_mask);
 643
 644 out:
 645         put_online_cpus();
 646         mutex_unlock(&pinst->lock);
 647
 648         return err;
 649 }
 650 EXPORT_SYMBOL(padata_set_cpumask);
 651
 652 /**
 653  * padata_start - start the parallel processing
 654  *
 655  * @pinst: padata instance to start
 656  */
 657 int padata_start(struct padata_instance *pinst)
 658 {
 659         int err = 0;
 660
 661         mutex_lock(&pinst->lock);
 662
 663         if (pinst->flags & PADATA_INVALID)
 664                 err = -EINVAL;
 665
 666          __padata_start(pinst);
 667
 668         mutex_unlock(&pinst->lock);
 669
 670         return err;
 671 }
 672 EXPORT_SYMBOL(padata_start);
 673
 674 /**
 675  * padata_stop - stop the parallel processing
 676  *
 677  * @pinst: padata instance to stop
 678  */
 679 void padata_stop(struct padata_instance *pinst)
 680 {
 681         mutex_lock(&pinst->lock);
 682         __padata_stop(pinst);
 683         mutex_unlock(&pinst->lock);
 684 }
 685 EXPORT_SYMBOL(padata_stop);
 686
 687 #ifdef CONFIG_HOTPLUG_CPU
 688
 689 static int __padata_add_cpu(struct padata_instance *pinst, int cpu)
 690 {
 691         struct parallel_data *pd;
 692
 693         if (cpumask_test_cpu(cpu, cpu_online_mask)) {
 694                 pd = padata_alloc_pd(pinst, pinst->cpumask.pcpu,
 695                                      pinst->cpumask.cbcpu);
 696                 if (!pd)
 697                         return -ENOMEM;
 698
 699                 padata_replace(pinst, pd);
 700
 701                 if (padata_validate_cpumask(pinst, pinst->cpumask.pcpu) &&
 702                     padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
 703                         __padata_start(pinst);
 704         }
 705
 706         return 0;
 707 }
 708
 709 static int __padata_remove_cpu(struct padata_instance *pinst, int cpu)
 710 {
 711         struct parallel_data *pd = NULL;
 712
 713         if (cpumask_test_cpu(cpu, cpu_online_mask)) {
 714
 715                 if (!padata_validate_cpumask(pinst, pinst->cpumask.pcpu) ||
 716                     !padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
 717                         __padata_stop(pinst);
 718
 719                 pd = padata_alloc_pd(pinst, pinst->cpumask.pcpu,
 720                                      pinst->cpumask.cbcpu);
 721                 if (!pd)
 722                         return -ENOMEM;
 723
 724                 padata_replace(pinst, pd);
 725
 726                 cpumask_clear_cpu(cpu, pd->cpumask.cbcpu);
 727                 cpumask_clear_cpu(cpu, pd->cpumask.pcpu);
 728         }
 729
 730         return 0;
 731 }
 732
 733  /**
 734  * padata_remove_cpu - remove a cpu from the one or both(serial and parallel)
 735  *                     padata cpumasks.
 736  *
 737  * @pinst: padata instance
 738  * @cpu: cpu to remove
 739  * @mask: bitmask specifying from which cpumask @cpu should be removed
 740  *        The @mask may be any combination of the following flags:
 741  *          PADATA_CPU_SERIAL   - serial cpumask
 742  *          PADATA_CPU_PARALLEL - parallel cpumask
 743  */
 744 int padata_remove_cpu(struct padata_instance *pinst, int cpu, int mask)
 745 {
 746         int err;
 747
 748         if (!(mask & (PADATA_CPU_SERIAL | PADATA_CPU_PARALLEL)))
 749                 return -EINVAL;
 750
 751         mutex_lock(&pinst->lock);
 752
 753         get_online_cpus();
 754         if (mask & PADATA_CPU_SERIAL)
 755                 cpumask_clear_cpu(cpu, pinst->cpumask.cbcpu);
 756         if (mask & PADATA_CPU_PARALLEL)
 757                 cpumask_clear_cpu(cpu, pinst->cpumask.pcpu);
 758
 759         err = __padata_remove_cpu(pinst, cpu);
 760         put_online_cpus();
 761
 762         mutex_unlock(&pinst->lock);
 763
 764         return err;
 765 }
 766 EXPORT_SYMBOL(padata_remove_cpu);
 767
 768 static inline int pinst_has_cpu(struct padata_instance *pinst, int cpu)
 769 {
 770         return cpumask_test_cpu(cpu, pinst->cpumask.pcpu) ||
 771                 cpumask_test_cpu(cpu, pinst->cpumask.cbcpu);
 772 }
 773
 774 static int padata_cpu_online(unsigned int cpu, struct hlist_node *node)
 775 {
 776         struct padata_instance *pinst;
 777         int ret;
 778
 779         pinst = hlist_entry_safe(node, struct padata_instance, node);
 780         if (!pinst_has_cpu(pinst, cpu))
 781                 return 0;
 782
 783         mutex_lock(&pinst->lock);
 784         ret = __padata_add_cpu(pinst, cpu);
 785         mutex_unlock(&pinst->lock);
 786         return ret;
 787 }
 788
 789 static int padata_cpu_prep_down(unsigned int cpu, struct hlist_node *node)
 790 {
 791         struct padata_instance *pinst;
 792         int ret;
 793
 794         pinst = hlist_entry_safe(node, struct padata_instance, node);
 795         if (!pinst_has_cpu(pinst, cpu))
 796                 return 0;
 797
 798         mutex_lock(&pinst->lock);
 799         ret = __padata_remove_cpu(pinst, cpu);
 800         mutex_unlock(&pinst->lock);
 801         return ret;
 802 }
 803
 804 static enum cpuhp_state hp_online;
 805 #endif
 806
 807 static void __padata_free(struct padata_instance *pinst)
 808 {
 809 #ifdef CONFIG_HOTPLUG_CPU
 810         cpuhp_state_remove_instance_nocalls(hp_online, &pinst->node);
 811 #endif
 812
 813         padata_stop(pinst);
 814         padata_free_pd(pinst->pd);
 815         free_cpumask_var(pinst->cpumask.pcpu);
 816         free_cpumask_var(pinst->cpumask.cbcpu);
 817         kfree(pinst);
 818 }
 819
 820 #define kobj2pinst(_kobj)                                       \
 821         container_of(_kobj, struct padata_instance, kobj)
 822 #define attr2pentry(_attr)                                      \
 823         container_of(_attr, struct padata_sysfs_entry, attr)
 824
 825 static void padata_sysfs_release(struct kobject *kobj)
 826 {
 827         struct padata_instance *pinst = kobj2pinst(kobj);
 828         __padata_free(pinst);
 829 }
 830
 831 struct padata_sysfs_entry {
 832         struct attribute attr;
 833         ssize_t (*show)(struct padata_instance *, struct attribute *, char *);
 834         ssize_t (*store)(struct padata_instance *, struct attribute *,
 835                          const char *, size_t);
 836 };
 837
 838 static ssize_t show_cpumask(struct padata_instance *pinst,
 839                             struct attribute *attr,  char *buf)
 840 {
 841         struct cpumask *cpumask;
 842         ssize_t len;
 843
 844         mutex_lock(&pinst->lock);
 845         if (!strcmp(attr->name, "serial_cpumask"))
 846                 cpumask = pinst->cpumask.cbcpu;
 847         else
 848                 cpumask = pinst->cpumask.pcpu;
 849
 850         len = snprintf(buf, PAGE_SIZE, "%*pb\n",
 851                        nr_cpu_ids, cpumask_bits(cpumask));
 852         mutex_unlock(&pinst->lock);
 853         return len < PAGE_SIZE ? len : -EINVAL;
 854 }
 855
 856 static ssize_t store_cpumask(struct padata_instance *pinst,
 857                              struct attribute *attr,
 858                              const char *buf, size_t count)
 859 {
 860         cpumask_var_t new_cpumask;
 861         ssize_t ret;
 862         int mask_type;
 863
 864         if (!alloc_cpumask_var(&new_cpumask, GFP_KERNEL))
 865                 return -ENOMEM;
 866
 867         ret = bitmap_parse(buf, count, cpumask_bits(new_cpumask),
 868                            nr_cpumask_bits);
 869         if (ret < 0)
 870                 goto out;
 871
 872         mask_type = !strcmp(attr->name, "serial_cpumask") ?
 873                 PADATA_CPU_SERIAL : PADATA_CPU_PARALLEL;
 874         ret = padata_set_cpumask(pinst, mask_type, new_cpumask);
 875         if (!ret)
 876                 ret = count;
 877
 878 out:
 879         free_cpumask_var(new_cpumask);
 880         return ret;
 881 }
 882
 883 #define PADATA_ATTR_RW(_name, _show_name, _store_name)          \
 884         static struct padata_sysfs_entry _name##_attr =         \
 885                 __ATTR(_name, 0644, _show_name, _store_name)
 886 #define PADATA_ATTR_RO(_name, _show_name)               \
 887         static struct padata_sysfs_entry _name##_attr = \
 888                 __ATTR(_name, 0400, _show_name, NULL)
 889
 890 PADATA_ATTR_RW(serial_cpumask, show_cpumask, store_cpumask);
 891 PADATA_ATTR_RW(parallel_cpumask, show_cpumask, store_cpumask);
 892
 893 /*
 894  * Padata sysfs provides the following objects:
 895  * serial_cpumask   [RW] - cpumask for serial workers
 896  * parallel_cpumask [RW] - cpumask for parallel workers
 897  */
 898 static struct attribute *padata_default_attrs[] = {
 899         &serial_cpumask_attr.attr,
 900         &parallel_cpumask_attr.attr,
 901         NULL,
 902 };
 903
 904 static ssize_t padata_sysfs_show(struct kobject *kobj,
 905                                  struct attribute *attr, char *buf)
 906 {
 907         struct padata_instance *pinst;
 908         struct padata_sysfs_entry *pentry;
 909         ssize_t ret = -EIO;
 910
 911         pinst = kobj2pinst(kobj);
 912         pentry = attr2pentry(attr);
 913         if (pentry->show)
 914                 ret = pentry->show(pinst, attr, buf);
 915
 916         return ret;
 917 }
 918
 919 static ssize_t padata_sysfs_store(struct kobject *kobj, struct attribute *attr,
 920                                   const char *buf, size_t count)
 921 {
 922         struct padata_instance *pinst;
 923         struct padata_sysfs_entry *pentry;
 924         ssize_t ret = -EIO;
 925
 926         pinst = kobj2pinst(kobj);
 927         pentry = attr2pentry(attr);
 928         if (pentry->show)
 929                 ret = pentry->store(pinst, attr, buf, count);
 930
 931         return ret;
 932 }
 933
 934 static const struct sysfs_ops padata_sysfs_ops = {
 935         .show = padata_sysfs_show,
 936         .store = padata_sysfs_store,
 937 };
 938
 939 static struct kobj_type padata_attr_type = {
 940         .sysfs_ops = &padata_sysfs_ops,
 941         .default_attrs = padata_default_attrs,
 942         .release = padata_sysfs_release,
 943 };
 944
 945 /**
 946  * padata_alloc_possible - Allocate and initialize padata instance.
 947  *                         Use the cpu_possible_mask for serial and
 948  *                         parallel workers.
 949  *
 950  * @wq: workqueue to use for the allocated padata instance
 951  */
 952 struct padata_instance *padata_alloc_possible(struct workqueue_struct *wq)
 953 {
 954         return padata_alloc(wq, cpu_possible_mask, cpu_possible_mask);
 955 }
 956 EXPORT_SYMBOL(padata_alloc_possible);
 957
 958 /**
 959  * padata_alloc - allocate and initialize a padata instance and specify
 960  *                cpumasks for serial and parallel workers.
 961  *
 962  * @wq: workqueue to use for the allocated padata instance
 963  * @pcpumask: cpumask that will be used for padata parallelization
 964  * @cbcpumask: cpumask that will be used for padata serialization
 965  */
 966 struct padata_instance *padata_alloc(struct workqueue_struct *wq,
 967                                      const struct cpumask *pcpumask,
 968                                      const struct cpumask *cbcpumask)
 969 {
 970         struct padata_instance *pinst;
 971         struct parallel_data *pd = NULL;
 972
 973         pinst = kzalloc(sizeof(struct padata_instance), GFP_KERNEL);
 974         if (!pinst)
 975                 goto err;
 976
 977         get_online_cpus();
 978         if (!alloc_cpumask_var(&pinst->cpumask.pcpu, GFP_KERNEL))
 979                 goto err_free_inst;
 980         if (!alloc_cpumask_var(&pinst->cpumask.cbcpu, GFP_KERNEL)) {
 981                 free_cpumask_var(pinst->cpumask.pcpu);
 982                 goto err_free_inst;
 983         }
 984         if (!padata_validate_cpumask(pinst, pcpumask) ||
 985             !padata_validate_cpumask(pinst, cbcpumask))
 986                 goto err_free_masks;
 987
 988         pd = padata_alloc_pd(pinst, pcpumask, cbcpumask);
 989         if (!pd)
 990                 goto err_free_masks;
 991
 992         rcu_assign_pointer(pinst->pd, pd);
 993
 994         pinst->wq = wq;
 995
 996         cpumask_copy(pinst->cpumask.pcpu, pcpumask);
 997         cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
 998
 999         pinst->flags = 0;
1000
1001         put_online_cpus();
1002
1003         BLOCKING_INIT_NOTIFIER_HEAD(&pinst->cpumask_change_notifier);
1004         kobject_init(&pinst->kobj, &padata_attr_type);
1005         mutex_init(&pinst->lock);
1006
1007 #ifdef CONFIG_HOTPLUG_CPU
1008         cpuhp_state_add_instance_nocalls(hp_online, &pinst->node);
1009 #endif
1010         return pinst;
1011
1012 err_free_masks:
1013         free_cpumask_var(pinst->cpumask.pcpu);
1014         free_cpumask_var(pinst->cpumask.cbcpu);
1015 err_free_inst:
1016         kfree(pinst);
1017         put_online_cpus();
1018 err:
1019         return NULL;
1020 }
1021
1022 /**
1023  * padata_free - free a padata instance
1024  *
1025  * @padata_inst: padata instance to free
1026  */
1027 void padata_free(struct padata_instance *pinst)
1028 {
1029         kobject_put(&pinst->kobj);
1030 }
1031 EXPORT_SYMBOL(padata_free);
1032
1033 #ifdef CONFIG_HOTPLUG_CPU
1034
1035 static __init int padata_driver_init(void)
1036 {
1037         int ret;
1038
1039         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "padata:online",
1040                                       padata_cpu_online,
1041                                       padata_cpu_prep_down);
1042         if (ret < 0)
1043                 return ret;
1044         hp_online = ret;
1045         return 0;
1046 }
1047 module_init(padata_driver_init);
1048
1049 static __exit void padata_driver_exit(void)
1050 {
1051         cpuhp_remove_multi_state(hp_online);
1052 }
1053 module_exit(padata_driver_exit);
1054 #endif