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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2013, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * libcfs/libcfs/workitem.c
  33  *
  34  * Author: Isaac Huang <isaac@clusterfs.com>
  35  *         Liang Zhen  <zhen.liang@sun.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_LNET
  39
  40 #include <linux/kthread.h>
  41 #include <libcfs/libcfs.h>
  42
  43 #define CFS_WS_NAME_LEN         16
  44
  45 struct cfs_wi_sched {
  46         struct list_head                ws_list;        /* chain on global list */
  47         /** serialised workitems */
  48         spinlock_t                      ws_lock;
  49         /** where schedulers sleep */
  50         wait_queue_head_t               ws_waitq;
  51         /** concurrent workitems */
  52         struct list_head                ws_runq;
  53         /** rescheduled running-workitems, a workitem can be rescheduled
  54          * while running in wi_action(), but we don't to execute it again
  55          * unless it returns from wi_action(), so we put it on ws_rerunq
  56          * while rescheduling, and move it to runq after it returns
  57          * from wi_action() */
  58         struct list_head                ws_rerunq;
  59         /** CPT-table for this scheduler */
  60         struct cfs_cpt_table    *ws_cptab;
  61         /** CPT id for affinity */
  62         int                     ws_cpt;
  63         /** number of scheduled workitems */
  64         int                     ws_nscheduled;
  65         /** started scheduler thread, protected by cfs_wi_data::wi_glock */
  66         unsigned int            ws_nthreads:30;
  67         /** shutting down, protected by cfs_wi_data::wi_glock */
  68         unsigned int            ws_stopping:1;
  69         /** serialize starting thread, protected by cfs_wi_data::wi_glock */
  70         unsigned int            ws_starting:1;
  71         /** scheduler name */
  72         char                    ws_name[CFS_WS_NAME_LEN];
  73 };
  74
  75 static struct cfs_workitem_data {
  76         /** serialize */
  77         spinlock_t              wi_glock;
  78         /** list of all schedulers */
  79         struct list_head        wi_scheds;
  80         /** WI module is initialized */
  81         int                     wi_init;
  82         /** shutting down the whole WI module */
  83         int                     wi_stopping;
  84 } cfs_wi_data;
  85
  86 static inline int
  87 cfs_wi_sched_cansleep(struct cfs_wi_sched *sched)
  88 {
  89         spin_lock(&sched->ws_lock);
  90         if (sched->ws_stopping) {
  91                 spin_unlock(&sched->ws_lock);
  92                 return 0;
  93         }
  94
  95         if (!list_empty(&sched->ws_runq)) {
  96                 spin_unlock(&sched->ws_lock);
  97                 return 0;
  98         }
  99         spin_unlock(&sched->ws_lock);
 100         return 1;
 101 }
 102
 103 /* XXX:
 104  * 0. it only works when called from wi->wi_action.
 105  * 1. when it returns no one shall try to schedule the workitem.
 106  */
 107 void
 108 cfs_wi_exit(struct cfs_wi_sched *sched, struct cfs_workitem *wi)
 109 {
 110         LASSERT(!in_interrupt()); /* because we use plain spinlock */
 111         LASSERT(!sched->ws_stopping);
 112
 113         spin_lock(&sched->ws_lock);
 114
 115         LASSERT(wi->wi_running);
 116
 117         if (wi->wi_scheduled) { /* cancel pending schedules */
 118                 LASSERT(!list_empty(&wi->wi_list));
 119                 list_del_init(&wi->wi_list);
 120
 121                 LASSERT(sched->ws_nscheduled > 0);
 122                 sched->ws_nscheduled--;
 123         }
 124
 125         LASSERT(list_empty(&wi->wi_list));
 126
 127         wi->wi_scheduled = 1; /* LBUG future schedule attempts */
 128         spin_unlock(&sched->ws_lock);
 129 }
 130 EXPORT_SYMBOL(cfs_wi_exit);
 131
 132 /**
 133  * cancel schedule request of workitem \a wi
 134  */
 135 int
 136 cfs_wi_deschedule(struct cfs_wi_sched *sched, struct cfs_workitem *wi)
 137 {
 138         int     rc;
 139
 140         LASSERT(!in_interrupt()); /* because we use plain spinlock */
 141         LASSERT(!sched->ws_stopping);
 142
 143         /*
 144          * return 0 if it's running already, otherwise return 1, which
 145          * means the workitem will not be scheduled and will not have
 146          * any race with wi_action.
 147          */
 148         spin_lock(&sched->ws_lock);
 149
 150         rc = !(wi->wi_running);
 151
 152         if (wi->wi_scheduled) { /* cancel pending schedules */
 153                 LASSERT(!list_empty(&wi->wi_list));
 154                 list_del_init(&wi->wi_list);
 155
 156                 LASSERT(sched->ws_nscheduled > 0);
 157                 sched->ws_nscheduled--;
 158
 159                 wi->wi_scheduled = 0;
 160         }
 161
 162         LASSERT (list_empty(&wi->wi_list));
 163
 164         spin_unlock(&sched->ws_lock);
 165         return rc;
 166 }
 167 EXPORT_SYMBOL(cfs_wi_deschedule);
 168
 169 /*
 170  * Workitem scheduled with (serial == 1) is strictly serialised not only with
 171  * itself, but also with others scheduled this way.
 172  *
 173  * Now there's only one static serialised queue, but in the future more might
 174  * be added, and even dynamic creation of serialised queues might be supported.
 175  */
 176 void
 177 cfs_wi_schedule(struct cfs_wi_sched *sched, struct cfs_workitem *wi)
 178 {
 179         LASSERT(!in_interrupt()); /* because we use plain spinlock */
 180         LASSERT(!sched->ws_stopping);
 181
 182         spin_lock(&sched->ws_lock);
 183
 184         if (!wi->wi_scheduled) {
 185                 LASSERT (list_empty(&wi->wi_list));
 186
 187                 wi->wi_scheduled = 1;
 188                 sched->ws_nscheduled++;
 189                 if (!wi->wi_running) {
 190                         list_add_tail(&wi->wi_list, &sched->ws_runq);
 191                         wake_up(&sched->ws_waitq);
 192                 } else {
 193                         list_add(&wi->wi_list, &sched->ws_rerunq);
 194                 }
 195         }
 196
 197         LASSERT (!list_empty(&wi->wi_list));
 198         spin_unlock(&sched->ws_lock);
 199 }
 200 EXPORT_SYMBOL(cfs_wi_schedule);
 201
 202 static int
 203 cfs_wi_scheduler(void *arg)
 204 {
 205         struct cfs_wi_sched *sched = (struct cfs_wi_sched *)arg;
 206
 207         /* CPT affinity scheduler? */
 208         if (sched->ws_cptab != NULL)
 209                 if (cfs_cpt_bind(sched->ws_cptab, sched->ws_cpt) != 0)
 210                         CWARN("Unable to bind %s on CPU partition %d\n",
 211                                 sched->ws_name, sched->ws_cpt);
 212
 213         spin_lock(&cfs_wi_data.wi_glock);
 214
 215         LASSERT(sched->ws_starting == 1);
 216         sched->ws_starting--;
 217         sched->ws_nthreads++;
 218
 219         spin_unlock(&cfs_wi_data.wi_glock);
 220
 221         spin_lock(&sched->ws_lock);
 222
 223         while (!sched->ws_stopping) {
 224                 int             nloops = 0;
 225                 int             rc;
 226                 struct cfs_workitem *wi;
 227
 228                 while (!list_empty(&sched->ws_runq) &&
 229                        nloops < CFS_WI_RESCHED) {
 230                         wi = list_entry(sched->ws_runq.next,
 231                                         struct cfs_workitem, wi_list);
 232                         LASSERT(wi->wi_scheduled && !wi->wi_running);
 233
 234                         list_del_init(&wi->wi_list);
 235
 236                         LASSERT(sched->ws_nscheduled > 0);
 237                         sched->ws_nscheduled--;
 238
 239                         wi->wi_running   = 1;
 240                         wi->wi_scheduled = 0;
 241
 242                         spin_unlock(&sched->ws_lock);
 243                         nloops++;
 244
 245                         rc = (*wi->wi_action) (wi);
 246
 247                         spin_lock(&sched->ws_lock);
 248                         if (rc != 0) /* WI should be dead, even be freed! */
 249                                 continue;
 250
 251                         wi->wi_running = 0;
 252                         if (list_empty(&wi->wi_list))
 253                                 continue;
 254
 255                         LASSERT(wi->wi_scheduled);
 256                         /* wi is rescheduled, should be on rerunq now, we
 257                          * move it to runq so it can run action now */
 258                         list_move_tail(&wi->wi_list, &sched->ws_runq);
 259                 }
 260
 261                 if (!list_empty(&sched->ws_runq)) {
 262                         spin_unlock(&sched->ws_lock);
 263                         /* don't sleep because some workitems still
 264                          * expect me to come back soon */
 265                         cond_resched();
 266                         spin_lock(&sched->ws_lock);
 267                         continue;
 268                 }
 269
 270                 spin_unlock(&sched->ws_lock);
 271                 rc = wait_event_interruptible_exclusive(sched->ws_waitq,
 272                                 !cfs_wi_sched_cansleep(sched));
 273                 spin_lock(&sched->ws_lock);
 274         }
 275
 276         spin_unlock(&sched->ws_lock);
 277
 278         spin_lock(&cfs_wi_data.wi_glock);
 279         sched->ws_nthreads--;
 280         spin_unlock(&cfs_wi_data.wi_glock);
 281
 282         return 0;
 283 }
 284
 285 void
 286 cfs_wi_sched_destroy(struct cfs_wi_sched *sched)
 287 {
 288         LASSERT(cfs_wi_data.wi_init);
 289         LASSERT(!cfs_wi_data.wi_stopping);
 290
 291         spin_lock(&cfs_wi_data.wi_glock);
 292         if (sched->ws_stopping) {
 293                 CDEBUG(D_INFO, "%s is in progress of stopping\n",
 294                        sched->ws_name);
 295                 spin_unlock(&cfs_wi_data.wi_glock);
 296                 return;
 297         }
 298
 299         LASSERT(!list_empty(&sched->ws_list));
 300         sched->ws_stopping = 1;
 301
 302         spin_unlock(&cfs_wi_data.wi_glock);
 303
 304         wake_up_all(&sched->ws_waitq);
 305
 306         spin_lock(&cfs_wi_data.wi_glock);
 307         {
 308                 int i = 2;
 309
 310                 while (sched->ws_nthreads > 0) {
 311                         CDEBUG(is_power_of_2(++i / 20) ? D_WARNING : D_NET,
 312                                "waiting %us for %d %s worker threads to exit\n",
 313                                i / 20, sched->ws_nthreads, sched->ws_name);
 314
 315                         spin_unlock(&cfs_wi_data.wi_glock);
 316                         schedule_timeout_uninterruptible(cfs_time_seconds(1)
 317                                                          / 20);
 318                         spin_lock(&cfs_wi_data.wi_glock);
 319                 }
 320         }
 321
 322         list_del(&sched->ws_list);
 323
 324         spin_unlock(&cfs_wi_data.wi_glock);
 325
 326         LASSERT(sched->ws_nscheduled == 0);
 327
 328         LIBCFS_FREE(sched, sizeof(*sched));
 329 }
 330 EXPORT_SYMBOL(cfs_wi_sched_destroy);
 331
 332 int
 333 cfs_wi_sched_create(char *name, struct cfs_cpt_table *cptab,
 334                     int cpt, int nthrs, struct cfs_wi_sched **sched_pp)
 335 {
 336         struct cfs_wi_sched     *sched;
 337
 338         LASSERT(cfs_wi_data.wi_init);
 339         LASSERT(!cfs_wi_data.wi_stopping);
 340         LASSERT(cptab == NULL || cpt == CFS_CPT_ANY ||
 341                 (cpt >= 0 && cpt < cfs_cpt_number(cptab)));
 342
 343         LIBCFS_ALLOC(sched, sizeof(*sched));
 344         if (sched == NULL)
 345                 return -ENOMEM;
 346
 347         if (strlen(name) > sizeof(sched->ws_name)-1) {
 348                 LIBCFS_FREE(sched, sizeof(*sched));
 349                 return -E2BIG;
 350         }
 351         strlcpy(sched->ws_name, name, sizeof(sched->ws_name));
 352
 353         sched->ws_cptab = cptab;
 354         sched->ws_cpt = cpt;
 355
 356         spin_lock_init(&sched->ws_lock);
 357         init_waitqueue_head(&sched->ws_waitq);
 358
 359         INIT_LIST_HEAD(&sched->ws_runq);
 360         INIT_LIST_HEAD(&sched->ws_rerunq);
 361         INIT_LIST_HEAD(&sched->ws_list);
 362
 363         for (; nthrs > 0; nthrs--)  {
 364                 char                    name[16];
 365                 struct task_struct      *task;
 366
 367                 spin_lock(&cfs_wi_data.wi_glock);
 368                 while (sched->ws_starting > 0) {
 369                         spin_unlock(&cfs_wi_data.wi_glock);
 370                         schedule();
 371                         spin_lock(&cfs_wi_data.wi_glock);
 372                 }
 373
 374                 sched->ws_starting++;
 375                 spin_unlock(&cfs_wi_data.wi_glock);
 376
 377                 if (sched->ws_cptab != NULL && sched->ws_cpt >= 0) {
 378                         snprintf(name, sizeof(name), "%s_%02d_%02d",
 379                                  sched->ws_name, sched->ws_cpt,
 380                                  sched->ws_nthreads);
 381                 } else {
 382                         snprintf(name, sizeof(name), "%s_%02d",
 383                                  sched->ws_name, sched->ws_nthreads);
 384                 }
 385
 386                 task = kthread_run(cfs_wi_scheduler, sched, "%s", name);
 387                 if (IS_ERR(task)) {
 388                         int rc = PTR_ERR(task);
 389
 390                         CERROR("Failed to create thread for "
 391                                 "WI scheduler %s: %d\n", name, rc);
 392
 393                         spin_lock(&cfs_wi_data.wi_glock);
 394
 395                         /* make up for cfs_wi_sched_destroy */
 396                         list_add(&sched->ws_list, &cfs_wi_data.wi_scheds);
 397                         sched->ws_starting--;
 398
 399                         spin_unlock(&cfs_wi_data.wi_glock);
 400
 401                         cfs_wi_sched_destroy(sched);
 402                         return rc;
 403                 }
 404         }
 405
 406         spin_lock(&cfs_wi_data.wi_glock);
 407         list_add(&sched->ws_list, &cfs_wi_data.wi_scheds);
 408         spin_unlock(&cfs_wi_data.wi_glock);
 409
 410         *sched_pp = sched;
 411         return 0;
 412 }
 413 EXPORT_SYMBOL(cfs_wi_sched_create);
 414
 415 int
 416 cfs_wi_startup(void)
 417 {
 418         memset(&cfs_wi_data, 0, sizeof(struct cfs_workitem_data));
 419
 420         spin_lock_init(&cfs_wi_data.wi_glock);
 421         INIT_LIST_HEAD(&cfs_wi_data.wi_scheds);
 422         cfs_wi_data.wi_init = 1;
 423
 424         return 0;
 425 }
 426
 427 void
 428 cfs_wi_shutdown (void)
 429 {
 430         struct cfs_wi_sched     *sched;
 431
 432         spin_lock(&cfs_wi_data.wi_glock);
 433         cfs_wi_data.wi_stopping = 1;
 434         spin_unlock(&cfs_wi_data.wi_glock);
 435
 436         /* nobody should contend on this list */
 437         list_for_each_entry(sched, &cfs_wi_data.wi_scheds, ws_list) {
 438                 sched->ws_stopping = 1;
 439                 wake_up_all(&sched->ws_waitq);
 440         }
 441
 442         list_for_each_entry(sched, &cfs_wi_data.wi_scheds, ws_list) {
 443                 spin_lock(&cfs_wi_data.wi_glock);
 444
 445                 while (sched->ws_nthreads != 0) {
 446                         spin_unlock(&cfs_wi_data.wi_glock);
 447                         schedule_timeout_uninterruptible(cfs_time_seconds(1)
 448                                                          / 20);
 449                         spin_lock(&cfs_wi_data.wi_glock);
 450                 }
 451                 spin_unlock(&cfs_wi_data.wi_glock);
 452         }
 453
 454         while (!list_empty(&cfs_wi_data.wi_scheds)) {
 455                 sched = list_entry(cfs_wi_data.wi_scheds.next,
 456                                        struct cfs_wi_sched, ws_list);
 457                 list_del(&sched->ws_list);
 458                 LIBCFS_FREE(sched, sizeof(*sched));
 459         }
 460
 461         cfs_wi_data.wi_stopping = 0;
 462         cfs_wi_data.wi_init = 0;
 463 }