kernel/padata.c

   1 /*
   2  * padata.c - generic interface to process data streams in parallel
   3  *
   4  * Copyright (C) 2008, 2009 secunet Security Networks AG
   5  * Copyright (C) 2008, 2009 Steffen Klassert <steffen.klassert@secunet.com>
   6  *
   7  * This program is free software; you can redistribute it and/or modify it
   8  * under the terms and conditions of the GNU General Public License,
   9  * version 2, as published by the Free Software Foundation.
  10  *
  11  * This program is distributed in the hope it will be useful, but WITHOUT
  12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  14  * more details.
  15  *
  16  * You should have received a copy of the GNU General Public License along with
  17  * this program; if not, write to the Free Software Foundation, Inc.,
  18  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
  19  */
  20
  21 #include <linux/module.h>
  22 #include <linux/cpumask.h>
  23 #include <linux/err.h>
  24 #include <linux/cpu.h>
  25 #include <linux/padata.h>
  26 #include <linux/mutex.h>
  27 #include <linux/sched.h>
  28 #include <linux/slab.h>
  29 #include <linux/rcupdate.h>
  30
  31 #define MAX_SEQ_NR INT_MAX - NR_CPUS
  32 #define MAX_OBJ_NUM 10000 * NR_CPUS
  33
  34 static int padata_index_to_cpu(struct parallel_data *pd, int cpu_index)
  35 {
  36         int cpu, target_cpu;
  37
  38         target_cpu = cpumask_first(pd->cpumask);
  39         for (cpu = 0; cpu < cpu_index; cpu++)
  40                 target_cpu = cpumask_next(target_cpu, pd->cpumask);
  41
  42         return target_cpu;
  43 }
  44
  45 static int padata_cpu_hash(struct padata_priv *padata)
  46 {
  47         int cpu_index;
  48         struct parallel_data *pd;
  49
  50         pd =  padata->pd;
  51
  52         /*
  53          * Hash the sequence numbers to the cpus by taking
  54          * seq_nr mod. number of cpus in use.
  55          */
  56         cpu_index =  padata->seq_nr % cpumask_weight(pd->cpumask);
  57
  58         return padata_index_to_cpu(pd, cpu_index);
  59 }
  60
  61 static void padata_parallel_worker(struct work_struct *work)
  62 {
  63         struct padata_queue *queue;
  64         struct parallel_data *pd;
  65         struct padata_instance *pinst;
  66         LIST_HEAD(local_list);
  67
  68         local_bh_disable();
  69         queue = container_of(work, struct padata_queue, pwork);
  70         pd = queue->pd;
  71         pinst = pd->pinst;
  72
  73         spin_lock(&queue->parallel.lock);
  74         list_replace_init(&queue->parallel.list, &local_list);
  75         spin_unlock(&queue->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 cpumask of padata.
  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_queue *queue;
 108         struct parallel_data *pd;
 109
 110         rcu_read_lock_bh();
 111
 112         pd = rcu_dereference(pinst->pd);
 113
 114         err = 0;
 115         if (!(pinst->flags & PADATA_INIT))
 116                 goto out;
 117
 118         err =  -EBUSY;
 119         if ((pinst->flags & PADATA_RESET))
 120                 goto out;
 121
 122         if (atomic_read(&pd->refcnt) >= MAX_OBJ_NUM)
 123                 goto out;
 124
 125         err = -EINVAL;
 126         if (!cpumask_test_cpu(cb_cpu, pd->cpumask))
 127                 goto out;
 128
 129         err = -EINPROGRESS;
 130         atomic_inc(&pd->refcnt);
 131         padata->pd = pd;
 132         padata->cb_cpu = cb_cpu;
 133
 134         if (unlikely(atomic_read(&pd->seq_nr) == pd->max_seq_nr))
 135                 atomic_set(&pd->seq_nr, -1);
 136
 137         padata->seq_nr = atomic_inc_return(&pd->seq_nr);
 138
 139         target_cpu = padata_cpu_hash(padata);
 140         queue = per_cpu_ptr(pd->queue, target_cpu);
 141
 142         spin_lock(&queue->parallel.lock);
 143         list_add_tail(&padata->list, &queue->parallel.list);
 144         spin_unlock(&queue->parallel.lock);
 145
 146         queue_work_on(target_cpu, pinst->wq, &queue->pwork);
 147
 148 out:
 149         rcu_read_unlock_bh();
 150
 151         return err;
 152 }
 153 EXPORT_SYMBOL(padata_do_parallel);
 154
 155 static struct padata_priv *padata_get_next(struct parallel_data *pd)
 156 {
 157         int cpu, num_cpus, empty, calc_seq_nr;
 158         int seq_nr, next_nr, overrun, next_overrun;
 159         struct padata_queue *queue, *next_queue;
 160         struct padata_priv *padata;
 161         struct padata_list *reorder;
 162
 163         empty = 0;
 164         next_nr = -1;
 165         next_overrun = 0;
 166         next_queue = NULL;
 167
 168         num_cpus = cpumask_weight(pd->cpumask);
 169
 170         for_each_cpu(cpu, pd->cpumask) {
 171                 queue = per_cpu_ptr(pd->queue, cpu);
 172                 reorder = &queue->reorder;
 173
 174                 /*
 175                  * Calculate the seq_nr of the object that should be
 176                  * next in this queue.
 177                  */
 178                 overrun = 0;
 179                 calc_seq_nr = (atomic_read(&queue->num_obj) * num_cpus)
 180                                + queue->cpu_index;
 181
 182                 if (unlikely(calc_seq_nr > pd->max_seq_nr)) {
 183                         calc_seq_nr = calc_seq_nr - pd->max_seq_nr - 1;
 184                         overrun = 1;
 185                 }
 186
 187                 if (!list_empty(&reorder->list)) {
 188                         padata = list_entry(reorder->list.next,
 189                                             struct padata_priv, list);
 190
 191                         seq_nr  = padata->seq_nr;
 192                         BUG_ON(calc_seq_nr != seq_nr);
 193                 } else {
 194                         seq_nr = calc_seq_nr;
 195                         empty++;
 196                 }
 197
 198                 if (next_nr < 0 || seq_nr < next_nr
 199                     || (next_overrun && !overrun)) {
 200                         next_nr = seq_nr;
 201                         next_overrun = overrun;
 202                         next_queue = queue;
 203                 }
 204         }
 205
 206         padata = NULL;
 207
 208         if (empty == num_cpus)
 209                 goto out;
 210
 211         reorder = &next_queue->reorder;
 212
 213         if (!list_empty(&reorder->list)) {
 214                 padata = list_entry(reorder->list.next,
 215                                     struct padata_priv, list);
 216
 217                 if (unlikely(next_overrun)) {
 218                         for_each_cpu(cpu, pd->cpumask) {
 219                                 queue = per_cpu_ptr(pd->queue, cpu);
 220                                 atomic_set(&queue->num_obj, 0);
 221                         }
 222                 }
 223
 224                 spin_lock(&reorder->lock);
 225                 list_del_init(&padata->list);
 226                 atomic_dec(&pd->reorder_objects);
 227                 spin_unlock(&reorder->lock);
 228
 229                 atomic_inc(&next_queue->num_obj);
 230
 231                 goto out;
 232         }
 233
 234         if (next_nr % num_cpus == next_queue->cpu_index) {
 235                 padata = ERR_PTR(-ENODATA);
 236                 goto out;
 237         }
 238
 239         padata = ERR_PTR(-EINPROGRESS);
 240 out:
 241         return padata;
 242 }
 243
 244 static void padata_reorder(struct parallel_data *pd)
 245 {
 246         struct padata_priv *padata;
 247         struct padata_queue *queue;
 248         struct padata_instance *pinst = pd->pinst;
 249
 250 try_again:
 251         if (!spin_trylock_bh(&pd->lock))
 252                 goto out;
 253
 254         while (1) {
 255                 padata = padata_get_next(pd);
 256
 257                 if (!padata || PTR_ERR(padata) == -EINPROGRESS)
 258                         break;
 259
 260                 if (PTR_ERR(padata) == -ENODATA) {
 261                         spin_unlock_bh(&pd->lock);
 262                         goto out;
 263                 }
 264
 265                 queue = per_cpu_ptr(pd->queue, padata->cb_cpu);
 266
 267                 spin_lock(&queue->serial.lock);
 268                 list_add_tail(&padata->list, &queue->serial.list);
 269                 spin_unlock(&queue->serial.lock);
 270
 271                 queue_work_on(padata->cb_cpu, pinst->wq, &queue->swork);
 272         }
 273
 274         spin_unlock_bh(&pd->lock);
 275
 276         if (atomic_read(&pd->reorder_objects))
 277                 goto try_again;
 278
 279 out:
 280         return;
 281 }
 282
 283 static void padata_serial_worker(struct work_struct *work)
 284 {
 285         struct padata_queue *queue;
 286         struct parallel_data *pd;
 287         LIST_HEAD(local_list);
 288
 289         local_bh_disable();
 290         queue = container_of(work, struct padata_queue, swork);
 291         pd = queue->pd;
 292
 293         spin_lock(&queue->serial.lock);
 294         list_replace_init(&queue->serial.list, &local_list);
 295         spin_unlock(&queue->serial.lock);
 296
 297         while (!list_empty(&local_list)) {
 298                 struct padata_priv *padata;
 299
 300                 padata = list_entry(local_list.next,
 301                                     struct padata_priv, list);
 302
 303                 list_del_init(&padata->list);
 304
 305                 padata->serial(padata);
 306                 atomic_dec(&pd->refcnt);
 307         }
 308         local_bh_enable();
 309 }
 310
 311 /*
 312  * padata_do_serial - padata serialization function
 313  *
 314  * @padata: object to be serialized.
 315  *
 316  * padata_do_serial must be called for every parallelized object.
 317  * The serialization callback function will run with BHs off.
 318  */
 319 void padata_do_serial(struct padata_priv *padata)
 320 {
 321         int cpu;
 322         struct padata_queue *queue;
 323         struct parallel_data *pd;
 324
 325         pd = padata->pd;
 326
 327         cpu = get_cpu();
 328         queue = per_cpu_ptr(pd->queue, cpu);
 329
 330         spin_lock(&queue->reorder.lock);
 331         atomic_inc(&pd->reorder_objects);
 332         list_add_tail(&padata->list, &queue->reorder.list);
 333         spin_unlock(&queue->reorder.lock);
 334
 335         put_cpu();
 336
 337         padata_reorder(pd);
 338 }
 339 EXPORT_SYMBOL(padata_do_serial);
 340
 341 static struct parallel_data *padata_alloc_pd(struct padata_instance *pinst,
 342                                              const struct cpumask *cpumask)
 343 {
 344         int cpu, cpu_index, num_cpus;
 345         struct padata_queue *queue;
 346         struct parallel_data *pd;
 347
 348         cpu_index = 0;
 349
 350         pd = kzalloc(sizeof(struct parallel_data), GFP_KERNEL);
 351         if (!pd)
 352                 goto err;
 353
 354         pd->queue = alloc_percpu(struct padata_queue);
 355         if (!pd->queue)
 356                 goto err_free_pd;
 357
 358         if (!alloc_cpumask_var(&pd->cpumask, GFP_KERNEL))
 359                 goto err_free_queue;
 360
 361         for_each_possible_cpu(cpu) {
 362                 queue = per_cpu_ptr(pd->queue, cpu);
 363
 364                 queue->pd = pd;
 365
 366                 if (cpumask_test_cpu(cpu, cpumask)
 367                     && cpumask_test_cpu(cpu, cpu_active_mask)) {
 368                         queue->cpu_index = cpu_index;
 369                         cpu_index++;
 370                 } else
 371                         queue->cpu_index = -1;
 372
 373                 INIT_LIST_HEAD(&queue->reorder.list);
 374                 INIT_LIST_HEAD(&queue->parallel.list);
 375                 INIT_LIST_HEAD(&queue->serial.list);
 376                 spin_lock_init(&queue->reorder.lock);
 377                 spin_lock_init(&queue->parallel.lock);
 378                 spin_lock_init(&queue->serial.lock);
 379
 380                 INIT_WORK(&queue->pwork, padata_parallel_worker);
 381                 INIT_WORK(&queue->swork, padata_serial_worker);
 382                 atomic_set(&queue->num_obj, 0);
 383         }
 384
 385         cpumask_and(pd->cpumask, cpumask, cpu_active_mask);
 386
 387         num_cpus = cpumask_weight(pd->cpumask);
 388         pd->max_seq_nr = (MAX_SEQ_NR / num_cpus) * num_cpus - 1;
 389
 390         atomic_set(&pd->seq_nr, -1);
 391         atomic_set(&pd->reorder_objects, 0);
 392         atomic_set(&pd->refcnt, 0);
 393         pd->pinst = pinst;
 394         spin_lock_init(&pd->lock);
 395
 396         return pd;
 397
 398 err_free_queue:
 399         free_percpu(pd->queue);
 400 err_free_pd:
 401         kfree(pd);
 402 err:
 403         return NULL;
 404 }
 405
 406 static void padata_free_pd(struct parallel_data *pd)
 407 {
 408         free_cpumask_var(pd->cpumask);
 409         free_percpu(pd->queue);
 410         kfree(pd);
 411 }
 412
 413 static void padata_replace(struct padata_instance *pinst,
 414                            struct parallel_data *pd_new)
 415 {
 416         struct parallel_data *pd_old = pinst->pd;
 417
 418         pinst->flags |= PADATA_RESET;
 419
 420         rcu_assign_pointer(pinst->pd, pd_new);
 421
 422         synchronize_rcu();
 423
 424         while (atomic_read(&pd_old->refcnt) != 0)
 425                 yield();
 426
 427         flush_workqueue(pinst->wq);
 428
 429         padata_free_pd(pd_old);
 430
 431         pinst->flags &= ~PADATA_RESET;
 432 }
 433
 434 /*
 435  * padata_set_cpumask - set the cpumask that padata should use
 436  *
 437  * @pinst: padata instance
 438  * @cpumask: the cpumask to use
 439  */
 440 int padata_set_cpumask(struct padata_instance *pinst,
 441                         cpumask_var_t cpumask)
 442 {
 443         struct parallel_data *pd;
 444         int err = 0;
 445
 446         might_sleep();
 447
 448         mutex_lock(&pinst->lock);
 449
 450         pd = padata_alloc_pd(pinst, cpumask);
 451         if (!pd) {
 452                 err = -ENOMEM;
 453                 goto out;
 454         }
 455
 456         cpumask_copy(pinst->cpumask, cpumask);
 457
 458         padata_replace(pinst, pd);
 459
 460 out:
 461         mutex_unlock(&pinst->lock);
 462
 463         return err;
 464 }
 465 EXPORT_SYMBOL(padata_set_cpumask);
 466
 467 static int __padata_add_cpu(struct padata_instance *pinst, int cpu)
 468 {
 469         struct parallel_data *pd;
 470
 471         if (cpumask_test_cpu(cpu, cpu_active_mask)) {
 472                 pd = padata_alloc_pd(pinst, pinst->cpumask);
 473                 if (!pd)
 474                         return -ENOMEM;
 475
 476                 padata_replace(pinst, pd);
 477         }
 478
 479         return 0;
 480 }
 481
 482 /*
 483  * padata_add_cpu - add a cpu to the padata cpumask
 484  *
 485  * @pinst: padata instance
 486  * @cpu: cpu to add
 487  */
 488 int padata_add_cpu(struct padata_instance *pinst, int cpu)
 489 {
 490         int err;
 491
 492         might_sleep();
 493
 494         mutex_lock(&pinst->lock);
 495
 496         cpumask_set_cpu(cpu, pinst->cpumask);
 497         err = __padata_add_cpu(pinst, cpu);
 498
 499         mutex_unlock(&pinst->lock);
 500
 501         return err;
 502 }
 503 EXPORT_SYMBOL(padata_add_cpu);
 504
 505 static int __padata_remove_cpu(struct padata_instance *pinst, int cpu)
 506 {
 507         struct parallel_data *pd;
 508
 509         if (cpumask_test_cpu(cpu, cpu_online_mask)) {
 510                 pd = padata_alloc_pd(pinst, pinst->cpumask);
 511                 if (!pd)
 512                         return -ENOMEM;
 513
 514                 padata_replace(pinst, pd);
 515         }
 516
 517         return 0;
 518 }
 519
 520 /*
 521  * padata_remove_cpu - remove a cpu from the padata cpumask
 522  *
 523  * @pinst: padata instance
 524  * @cpu: cpu to remove
 525  */
 526 int padata_remove_cpu(struct padata_instance *pinst, int cpu)
 527 {
 528         int err;
 529
 530         might_sleep();
 531
 532         mutex_lock(&pinst->lock);
 533
 534         cpumask_clear_cpu(cpu, pinst->cpumask);
 535         err = __padata_remove_cpu(pinst, cpu);
 536
 537         mutex_unlock(&pinst->lock);
 538
 539         return err;
 540 }
 541 EXPORT_SYMBOL(padata_remove_cpu);
 542
 543 /*
 544  * padata_start - start the parallel processing
 545  *
 546  * @pinst: padata instance to start
 547  */
 548 void padata_start(struct padata_instance *pinst)
 549 {
 550         might_sleep();
 551
 552         mutex_lock(&pinst->lock);
 553         pinst->flags |= PADATA_INIT;
 554         mutex_unlock(&pinst->lock);
 555 }
 556 EXPORT_SYMBOL(padata_start);
 557
 558 /*
 559  * padata_stop - stop the parallel processing
 560  *
 561  * @pinst: padata instance to stop
 562  */
 563 void padata_stop(struct padata_instance *pinst)
 564 {
 565         might_sleep();
 566
 567         mutex_lock(&pinst->lock);
 568         pinst->flags &= ~PADATA_INIT;
 569         mutex_unlock(&pinst->lock);
 570 }
 571 EXPORT_SYMBOL(padata_stop);
 572
 573 static int __cpuinit padata_cpu_callback(struct notifier_block *nfb,
 574                                          unsigned long action, void *hcpu)
 575 {
 576         int err;
 577         struct padata_instance *pinst;
 578         int cpu = (unsigned long)hcpu;
 579
 580         pinst = container_of(nfb, struct padata_instance, cpu_notifier);
 581
 582         switch (action) {
 583         case CPU_ONLINE:
 584         case CPU_ONLINE_FROZEN:
 585                 if (!cpumask_test_cpu(cpu, pinst->cpumask))
 586                         break;
 587                 mutex_lock(&pinst->lock);
 588                 err = __padata_add_cpu(pinst, cpu);
 589                 mutex_unlock(&pinst->lock);
 590                 if (err)
 591                         return NOTIFY_BAD;
 592                 break;
 593
 594         case CPU_DOWN_PREPARE:
 595         case CPU_DOWN_PREPARE_FROZEN:
 596                 if (!cpumask_test_cpu(cpu, pinst->cpumask))
 597                         break;
 598                 mutex_lock(&pinst->lock);
 599                 err = __padata_remove_cpu(pinst, cpu);
 600                 mutex_unlock(&pinst->lock);
 601                 if (err)
 602                         return NOTIFY_BAD;
 603                 break;
 604
 605         case CPU_UP_CANCELED:
 606         case CPU_UP_CANCELED_FROZEN:
 607                 if (!cpumask_test_cpu(cpu, pinst->cpumask))
 608                         break;
 609                 mutex_lock(&pinst->lock);
 610                 __padata_remove_cpu(pinst, cpu);
 611                 mutex_unlock(&pinst->lock);
 612
 613         case CPU_DOWN_FAILED:
 614         case CPU_DOWN_FAILED_FROZEN:
 615                 if (!cpumask_test_cpu(cpu, pinst->cpumask))
 616                         break;
 617                 mutex_lock(&pinst->lock);
 618                 __padata_add_cpu(pinst, cpu);
 619                 mutex_unlock(&pinst->lock);
 620         }
 621
 622         return NOTIFY_OK;
 623 }
 624
 625 /*
 626  * padata_alloc - allocate and initialize a padata instance
 627  *
 628  * @cpumask: cpumask that padata uses for parallelization
 629  * @wq: workqueue to use for the allocated padata instance
 630  */
 631 struct padata_instance *padata_alloc(const struct cpumask *cpumask,
 632                                      struct workqueue_struct *wq)
 633 {
 634         int err;
 635         struct padata_instance *pinst;
 636         struct parallel_data *pd;
 637
 638         pinst = kzalloc(sizeof(struct padata_instance), GFP_KERNEL);
 639         if (!pinst)
 640                 goto err;
 641
 642         pd = padata_alloc_pd(pinst, cpumask);
 643         if (!pd)
 644                 goto err_free_inst;
 645
 646         if (!alloc_cpumask_var(&pinst->cpumask, GFP_KERNEL))
 647                 goto err_free_pd;
 648
 649         rcu_assign_pointer(pinst->pd, pd);
 650
 651         pinst->wq = wq;
 652
 653         cpumask_copy(pinst->cpumask, cpumask);
 654
 655         pinst->flags = 0;
 656
 657         pinst->cpu_notifier.notifier_call = padata_cpu_callback;
 658         pinst->cpu_notifier.priority = 0;
 659         err = register_hotcpu_notifier(&pinst->cpu_notifier);
 660         if (err)
 661                 goto err_free_cpumask;
 662
 663         mutex_init(&pinst->lock);
 664
 665         return pinst;
 666
 667 err_free_cpumask:
 668         free_cpumask_var(pinst->cpumask);
 669 err_free_pd:
 670         padata_free_pd(pd);
 671 err_free_inst:
 672         kfree(pinst);
 673 err:
 674         return NULL;
 675 }
 676 EXPORT_SYMBOL(padata_alloc);
 677
 678 /*
 679  * padata_free - free a padata instance
 680  *
 681  * @ padata_inst: padata instance to free
 682  */
 683 void padata_free(struct padata_instance *pinst)
 684 {
 685         padata_stop(pinst);
 686
 687         synchronize_rcu();
 688
 689         while (atomic_read(&pinst->pd->refcnt) != 0)
 690                 yield();
 691
 692         unregister_hotcpu_notifier(&pinst->cpu_notifier);
 693         padata_free_pd(pinst->pd);
 694         free_cpumask_var(pinst->cpumask);
 695         kfree(pinst);
 696 }
 697 EXPORT_SYMBOL(padata_free);