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