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