libcfs/libcfs/workitem.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * libcfs/libcfs/workitem.c
  37  *
  38  * Author: Isaac Huang <isaac@clusterfs.com>
  39  *         Liang Zhen  <zhen.liang@sun.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_LNET
  43
  44 #include <libcfs/libcfs.h>
  45
  46 #define CFS_WS_NAME_LEN         16
  47
  48 typedef struct cfs_wi_sched {
  49         cfs_list_t              ws_list;        /* chain on global list */
  50 #ifdef __KERNEL__
  51         /** serialised workitems */
  52         spinlock_t              ws_lock;
  53         /** where schedulers sleep */
  54         wait_queue_head_t               ws_waitq;
  55 #endif
  56         /** concurrent workitems */
  57         cfs_list_t              ws_runq;
  58         /** rescheduled running-workitems, a workitem can be rescheduled
  59          * while running in wi_action(), but we don't to execute it again
  60          * unless it returns from wi_action(), so we put it on ws_rerunq
  61          * while rescheduling, and move it to runq after it returns
  62          * from wi_action() */
  63         cfs_list_t              ws_rerunq;
  64         /** CPT-table for this scheduler */
  65         struct cfs_cpt_table    *ws_cptab;
  66         /** CPT id for affinity */
  67         int                     ws_cpt;
  68         /** number of scheduled workitems */
  69         int                     ws_nscheduled;
  70         /** started scheduler thread, protected by cfs_wi_data::wi_glock */
  71         unsigned int            ws_nthreads:30;
  72         /** shutting down, protected by cfs_wi_data::wi_glock */
  73         unsigned int            ws_stopping:1;
  74         /** serialize starting thread, protected by cfs_wi_data::wi_glock */
  75         unsigned int            ws_starting:1;
  76         /** scheduler name */
  77         char                    ws_name[CFS_WS_NAME_LEN];
  78 } cfs_wi_sched_t;
  79
  80 struct cfs_workitem_data {
  81         /** serialize */
  82         spinlock_t              wi_glock;
  83         /** list of all schedulers */
  84         cfs_list_t              wi_scheds;
  85         /** WI module is initialized */
  86         int                     wi_init;
  87         /** shutting down the whole WI module */
  88         int                     wi_stopping;
  89 } cfs_wi_data;
  90
  91 #ifdef __KERNEL__
  92 static inline void
  93 cfs_wi_sched_lock(cfs_wi_sched_t *sched)
  94 {
  95         spin_lock(&sched->ws_lock);
  96 }
  97
  98 static inline void
  99 cfs_wi_sched_unlock(cfs_wi_sched_t *sched)
 100 {
 101         spin_unlock(&sched->ws_lock);
 102 }
 103
 104 static inline int
 105 cfs_wi_sched_cansleep(cfs_wi_sched_t *sched)
 106 {
 107         cfs_wi_sched_lock(sched);
 108         if (sched->ws_stopping) {
 109                 cfs_wi_sched_unlock(sched);
 110                 return 0;
 111         }
 112
 113         if (!cfs_list_empty(&sched->ws_runq)) {
 114                 cfs_wi_sched_unlock(sched);
 115                 return 0;
 116         }
 117         cfs_wi_sched_unlock(sched);
 118         return 1;
 119 }
 120
 121 #else /* !__KERNEL__ */
 122
 123 static inline void
 124 cfs_wi_sched_lock(cfs_wi_sched_t *sched)
 125 {
 126         spin_lock(&cfs_wi_data.wi_glock);
 127 }
 128
 129 static inline void
 130 cfs_wi_sched_unlock(cfs_wi_sched_t *sched)
 131 {
 132         spin_unlock(&cfs_wi_data.wi_glock);
 133 }
 134
 135 #endif /* __KERNEL__ */
 136
 137 /* XXX:
 138  * 0. it only works when called from wi->wi_action.
 139  * 1. when it returns no one shall try to schedule the workitem.
 140  */
 141 void
 142 cfs_wi_exit(struct cfs_wi_sched *sched, cfs_workitem_t *wi)
 143 {
 144         LASSERT(!cfs_in_interrupt()); /* because we use plain spinlock */
 145         LASSERT(!sched->ws_stopping);
 146
 147         cfs_wi_sched_lock(sched);
 148
 149 #ifdef __KERNEL__
 150         LASSERT(wi->wi_running);
 151 #endif
 152         if (wi->wi_scheduled) { /* cancel pending schedules */
 153                 LASSERT(!cfs_list_empty(&wi->wi_list));
 154                 cfs_list_del_init(&wi->wi_list);
 155
 156                 LASSERT(sched->ws_nscheduled > 0);
 157                 sched->ws_nscheduled--;
 158         }
 159
 160         LASSERT(cfs_list_empty(&wi->wi_list));
 161
 162         wi->wi_scheduled = 1; /* LBUG future schedule attempts */
 163         cfs_wi_sched_unlock(sched);
 164
 165         return;
 166 }
 167 EXPORT_SYMBOL(cfs_wi_exit);
 168
 169 /**
 170  * cancel schedule request of workitem \a wi
 171  */
 172 int
 173 cfs_wi_deschedule(struct cfs_wi_sched *sched, cfs_workitem_t *wi)
 174 {
 175         int     rc;
 176
 177         LASSERT(!cfs_in_interrupt()); /* because we use plain spinlock */
 178         LASSERT(!sched->ws_stopping);
 179
 180         /*
 181          * return 0 if it's running already, otherwise return 1, which
 182          * means the workitem will not be scheduled and will not have
 183          * any race with wi_action.
 184          */
 185         cfs_wi_sched_lock(sched);
 186
 187         rc = !(wi->wi_running);
 188
 189         if (wi->wi_scheduled) { /* cancel pending schedules */
 190                 LASSERT(!cfs_list_empty(&wi->wi_list));
 191                 cfs_list_del_init(&wi->wi_list);
 192
 193                 LASSERT(sched->ws_nscheduled > 0);
 194                 sched->ws_nscheduled--;
 195
 196                 wi->wi_scheduled = 0;
 197         }
 198
 199         LASSERT (cfs_list_empty(&wi->wi_list));
 200
 201         cfs_wi_sched_unlock(sched);
 202         return rc;
 203 }
 204 EXPORT_SYMBOL(cfs_wi_deschedule);
 205
 206 /*
 207  * Workitem scheduled with (serial == 1) is strictly serialised not only with
 208  * itself, but also with others scheduled this way.
 209  *
 210  * Now there's only one static serialised queue, but in the future more might
 211  * be added, and even dynamic creation of serialised queues might be supported.
 212  */
 213 void
 214 cfs_wi_schedule(struct cfs_wi_sched *sched, cfs_workitem_t *wi)
 215 {
 216         LASSERT(!cfs_in_interrupt()); /* because we use plain spinlock */
 217         LASSERT(!sched->ws_stopping);
 218
 219         cfs_wi_sched_lock(sched);
 220
 221         if (!wi->wi_scheduled) {
 222                 LASSERT (cfs_list_empty(&wi->wi_list));
 223
 224                 wi->wi_scheduled = 1;
 225                 sched->ws_nscheduled++;
 226                 if (!wi->wi_running) {
 227                         cfs_list_add_tail(&wi->wi_list, &sched->ws_runq);
 228 #ifdef __KERNEL__
 229                         wake_up(&sched->ws_waitq);
 230 #endif
 231                 } else {
 232                         cfs_list_add(&wi->wi_list, &sched->ws_rerunq);
 233                 }
 234         }
 235
 236         LASSERT (!cfs_list_empty(&wi->wi_list));
 237         cfs_wi_sched_unlock(sched);
 238         return;
 239 }
 240 EXPORT_SYMBOL(cfs_wi_schedule);
 241
 242 #ifdef __KERNEL__
 243
 244 static int
 245 cfs_wi_scheduler (void *arg)
 246 {
 247         struct cfs_wi_sched     *sched = (cfs_wi_sched_t *)arg;
 248
 249         cfs_block_allsigs();
 250
 251         /* CPT affinity scheduler? */
 252         if (sched->ws_cptab != NULL)
 253                 if (cfs_cpt_bind(sched->ws_cptab, sched->ws_cpt) != 0)
 254                         CWARN("Failed to bind %s on CPT %d\n",
 255                                 sched->ws_name, sched->ws_cpt);
 256
 257         spin_lock(&cfs_wi_data.wi_glock);
 258
 259         LASSERT(sched->ws_starting == 1);
 260         sched->ws_starting--;
 261         sched->ws_nthreads++;
 262
 263         spin_unlock(&cfs_wi_data.wi_glock);
 264
 265         cfs_wi_sched_lock(sched);
 266
 267         while (!sched->ws_stopping) {
 268                 int             nloops = 0;
 269                 int             rc;
 270                 cfs_workitem_t *wi;
 271
 272                 while (!cfs_list_empty(&sched->ws_runq) &&
 273                        nloops < CFS_WI_RESCHED) {
 274                         wi = cfs_list_entry(sched->ws_runq.next,
 275                                             cfs_workitem_t, wi_list);
 276                         LASSERT(wi->wi_scheduled && !wi->wi_running);
 277
 278                         cfs_list_del_init(&wi->wi_list);
 279
 280                         LASSERT(sched->ws_nscheduled > 0);
 281                         sched->ws_nscheduled--;
 282
 283                         wi->wi_running   = 1;
 284                         wi->wi_scheduled = 0;
 285
 286
 287                         cfs_wi_sched_unlock(sched);
 288                         nloops++;
 289
 290                         rc = (*wi->wi_action) (wi);
 291
 292                         cfs_wi_sched_lock(sched);
 293                         if (rc != 0) /* WI should be dead, even be freed! */
 294                                 continue;
 295
 296                         wi->wi_running = 0;
 297                         if (cfs_list_empty(&wi->wi_list))
 298                                 continue;
 299
 300                         LASSERT(wi->wi_scheduled);
 301                         /* wi is rescheduled, should be on rerunq now, we
 302                          * move it to runq so it can run action now */
 303                         cfs_list_move_tail(&wi->wi_list, &sched->ws_runq);
 304                 }
 305
 306                 if (!cfs_list_empty(&sched->ws_runq)) {
 307                         cfs_wi_sched_unlock(sched);
 308                         /* don't sleep because some workitems still
 309                          * expect me to come back soon */
 310                         cond_resched();
 311                         cfs_wi_sched_lock(sched);
 312                         continue;
 313                 }
 314
 315                 cfs_wi_sched_unlock(sched);
 316                 rc = wait_event_interruptible_exclusive(sched->ws_waitq,
 317                                 !cfs_wi_sched_cansleep(sched));
 318                 cfs_wi_sched_lock(sched);
 319         }
 320
 321         cfs_wi_sched_unlock(sched);
 322
 323         spin_lock(&cfs_wi_data.wi_glock);
 324         sched->ws_nthreads--;
 325         spin_unlock(&cfs_wi_data.wi_glock);
 326
 327         return 0;
 328 }
 329
 330 #else /* __KERNEL__ */
 331
 332 int
 333 cfs_wi_check_events (void)
 334 {
 335         int               n = 0;
 336         cfs_workitem_t   *wi;
 337
 338         spin_lock(&cfs_wi_data.wi_glock);
 339
 340         for (;;) {
 341                 struct cfs_wi_sched     *sched = NULL;
 342                 struct cfs_wi_sched     *tmp;
 343
 344                 /** rerunq is always empty for userspace */
 345                 cfs_list_for_each_entry(tmp,
 346                                         &cfs_wi_data.wi_scheds, ws_list) {
 347                         if (!cfs_list_empty(&tmp->ws_runq)) {
 348                                 sched = tmp;
 349                                 break;
 350                         }
 351                 }
 352
 353                 if (sched == NULL)
 354                         break;
 355
 356                 wi = cfs_list_entry(sched->ws_runq.next,
 357                                     cfs_workitem_t, wi_list);
 358                 cfs_list_del_init(&wi->wi_list);
 359
 360                 LASSERT(sched->ws_nscheduled > 0);
 361                 sched->ws_nscheduled--;
 362
 363                 LASSERT(wi->wi_scheduled);
 364                 wi->wi_scheduled = 0;
 365                 spin_unlock(&cfs_wi_data.wi_glock);
 366
 367                 n++;
 368                 (*wi->wi_action) (wi);
 369
 370                 spin_lock(&cfs_wi_data.wi_glock);
 371         }
 372
 373         spin_unlock(&cfs_wi_data.wi_glock);
 374         return n;
 375 }
 376
 377 #endif
 378
 379 void
 380 cfs_wi_sched_destroy(struct cfs_wi_sched *sched)
 381 {
 382         LASSERT(cfs_wi_data.wi_init);
 383         LASSERT(!cfs_wi_data.wi_stopping);
 384
 385         spin_lock(&cfs_wi_data.wi_glock);
 386         if (sched->ws_stopping) {
 387                 CDEBUG(D_INFO, "%s is in progress of stopping\n",
 388                        sched->ws_name);
 389                 spin_unlock(&cfs_wi_data.wi_glock);
 390                 return;
 391         }
 392
 393         LASSERT(!cfs_list_empty(&sched->ws_list));
 394         sched->ws_stopping = 1;
 395
 396         spin_unlock(&cfs_wi_data.wi_glock);
 397
 398 #ifdef __KERNEL__
 399         wake_up_all(&sched->ws_waitq);
 400
 401         spin_lock(&cfs_wi_data.wi_glock);
 402         {
 403                 int i = 2;
 404
 405                 while (sched->ws_nthreads > 0) {
 406                         CDEBUG(IS_PO2(++i) ? D_WARNING : D_NET,
 407                                "waiting for %d threads of WI sched[%s] to "
 408                                "terminate\n", sched->ws_nthreads,
 409                                sched->ws_name);
 410
 411                         spin_unlock(&cfs_wi_data.wi_glock);
 412                         cfs_pause(cfs_time_seconds(1) / 20);
 413                         spin_lock(&cfs_wi_data.wi_glock);
 414                 }
 415         }
 416
 417         cfs_list_del(&sched->ws_list);
 418
 419         spin_unlock(&cfs_wi_data.wi_glock);
 420 #endif
 421         LASSERT(sched->ws_nscheduled == 0);
 422
 423         LIBCFS_FREE(sched, sizeof(*sched));
 424 }
 425 EXPORT_SYMBOL(cfs_wi_sched_destroy);
 426
 427 int
 428 cfs_wi_sched_create(char *name, struct cfs_cpt_table *cptab,
 429                     int cpt, int nthrs, struct cfs_wi_sched **sched_pp)
 430 {
 431         struct cfs_wi_sched     *sched;
 432
 433         LASSERT(cfs_wi_data.wi_init);
 434         LASSERT(!cfs_wi_data.wi_stopping);
 435         LASSERT(cptab == NULL || cpt == CFS_CPT_ANY ||
 436                 (cpt >= 0 && cpt < cfs_cpt_number(cptab)));
 437
 438         LIBCFS_ALLOC(sched, sizeof(*sched));
 439         if (sched == NULL)
 440                 return -ENOMEM;
 441
 442         strncpy(sched->ws_name, name, CFS_WS_NAME_LEN);
 443         sched->ws_cptab = cptab;
 444         sched->ws_cpt = cpt;
 445
 446 #ifdef __KERNEL__
 447         spin_lock_init(&sched->ws_lock);
 448         init_waitqueue_head(&sched->ws_waitq);
 449 #endif
 450         CFS_INIT_LIST_HEAD(&sched->ws_runq);
 451         CFS_INIT_LIST_HEAD(&sched->ws_rerunq);
 452         CFS_INIT_LIST_HEAD(&sched->ws_list);
 453
 454 #ifdef __KERNEL__
 455         for (; nthrs > 0; nthrs--)  {
 456                 char            name[16];
 457                 cfs_task_t      *task;
 458
 459                 spin_lock(&cfs_wi_data.wi_glock);
 460                 while (sched->ws_starting > 0) {
 461                         spin_unlock(&cfs_wi_data.wi_glock);
 462                         schedule();
 463                         spin_lock(&cfs_wi_data.wi_glock);
 464                 }
 465
 466                 sched->ws_starting++;
 467                 spin_unlock(&cfs_wi_data.wi_glock);
 468
 469                 if (sched->ws_cptab != NULL && sched->ws_cpt >= 0) {
 470                         snprintf(name, sizeof(name), "%s_%02d_%02d",
 471                                  sched->ws_name, sched->ws_cpt,
 472                                  sched->ws_nthreads);
 473                 } else {
 474                         snprintf(name, sizeof(name), "%s_%02d",
 475                                  sched->ws_name, sched->ws_nthreads);
 476                 }
 477
 478                 task = kthread_run(cfs_wi_scheduler, sched, name);
 479                 if (IS_ERR(task)) {
 480                         int rc = PTR_ERR(task);
 481
 482                         CERROR("Failed to create thread for "
 483                                 "WI scheduler %s: %d\n", name, rc);
 484
 485                         spin_lock(&cfs_wi_data.wi_glock);
 486
 487                         /* make up for cfs_wi_sched_destroy */
 488                         cfs_list_add(&sched->ws_list, &cfs_wi_data.wi_scheds);
 489                         sched->ws_starting--;
 490
 491                         spin_unlock(&cfs_wi_data.wi_glock);
 492
 493                         cfs_wi_sched_destroy(sched);
 494                         return rc;
 495                 }
 496         }
 497 #endif
 498         spin_lock(&cfs_wi_data.wi_glock);
 499         cfs_list_add(&sched->ws_list, &cfs_wi_data.wi_scheds);
 500         spin_unlock(&cfs_wi_data.wi_glock);
 501
 502         *sched_pp = sched;
 503         return 0;
 504 }
 505 EXPORT_SYMBOL(cfs_wi_sched_create);
 506
 507 int
 508 cfs_wi_startup(void)
 509 {
 510         memset(&cfs_wi_data, 0, sizeof(cfs_wi_data));
 511
 512         spin_lock_init(&cfs_wi_data.wi_glock);
 513         CFS_INIT_LIST_HEAD(&cfs_wi_data.wi_scheds);
 514         cfs_wi_data.wi_init = 1;
 515
 516         return 0;
 517 }
 518
 519 void
 520 cfs_wi_shutdown (void)
 521 {
 522         struct cfs_wi_sched     *sched;
 523
 524         spin_lock(&cfs_wi_data.wi_glock);
 525         cfs_wi_data.wi_stopping = 1;
 526         spin_unlock(&cfs_wi_data.wi_glock);
 527
 528 #ifdef __KERNEL__
 529         /* nobody should contend on this list */
 530         cfs_list_for_each_entry(sched, &cfs_wi_data.wi_scheds, ws_list) {
 531                 sched->ws_stopping = 1;
 532                 wake_up_all(&sched->ws_waitq);
 533         }
 534
 535         cfs_list_for_each_entry(sched, &cfs_wi_data.wi_scheds, ws_list) {
 536                 spin_lock(&cfs_wi_data.wi_glock);
 537
 538                 while (sched->ws_nthreads != 0) {
 539                         spin_unlock(&cfs_wi_data.wi_glock);
 540                         cfs_pause(cfs_time_seconds(1) / 20);
 541                         spin_lock(&cfs_wi_data.wi_glock);
 542                 }
 543                 spin_unlock(&cfs_wi_data.wi_glock);
 544         }
 545 #endif
 546         while (!cfs_list_empty(&cfs_wi_data.wi_scheds)) {
 547                 sched = cfs_list_entry(cfs_wi_data.wi_scheds.next,
 548                                        struct cfs_wi_sched, ws_list);
 549                 cfs_list_del(&sched->ws_list);
 550                 LIBCFS_FREE(sched, sizeof(*sched));
 551         }
 552
 553         cfs_wi_data.wi_stopping = 0;
 554         cfs_wi_data.wi_init = 0;
 555 }