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