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         cfs_block_allsigs();
 208
 209         /* CPT affinity scheduler? */
 210         if (sched->ws_cptab != NULL)
 211                 if (cfs_cpt_bind(sched->ws_cptab, sched->ws_cpt) != 0)
 212                         CWARN("Unable to bind %s on CPU partition %d\n",
 213                                 sched->ws_name, sched->ws_cpt);
 214
 215         spin_lock(&cfs_wi_data.wi_glock);
 216
 217         LASSERT(sched->ws_starting == 1);
 218         sched->ws_starting--;
 219         sched->ws_nthreads++;
 220
 221         spin_unlock(&cfs_wi_data.wi_glock);
 222
 223         spin_lock(&sched->ws_lock);
 224
 225         while (!sched->ws_stopping) {
 226                 int             nloops = 0;
 227                 int             rc;
 228                 struct cfs_workitem *wi;
 229
 230                 while (!list_empty(&sched->ws_runq) &&
 231                        nloops < CFS_WI_RESCHED) {
 232                         wi = list_entry(sched->ws_runq.next,
 233                                         struct cfs_workitem, wi_list);
 234                         LASSERT(wi->wi_scheduled && !wi->wi_running);
 235
 236                         list_del_init(&wi->wi_list);
 237
 238                         LASSERT(sched->ws_nscheduled > 0);
 239                         sched->ws_nscheduled--;
 240
 241                         wi->wi_running   = 1;
 242                         wi->wi_scheduled = 0;
 243
 244                         spin_unlock(&sched->ws_lock);
 245                         nloops++;
 246
 247                         rc = (*wi->wi_action) (wi);
 248
 249                         spin_lock(&sched->ws_lock);
 250                         if (rc != 0) /* WI should be dead, even be freed! */
 251                                 continue;
 252
 253                         wi->wi_running = 0;
 254                         if (list_empty(&wi->wi_list))
 255                                 continue;
 256
 257                         LASSERT(wi->wi_scheduled);
 258                         /* wi is rescheduled, should be on rerunq now, we
 259                          * move it to runq so it can run action now */
 260                         list_move_tail(&wi->wi_list, &sched->ws_runq);
 261                 }
 262
 263                 if (!list_empty(&sched->ws_runq)) {
 264                         spin_unlock(&sched->ws_lock);
 265                         /* don't sleep because some workitems still
 266                          * expect me to come back soon */
 267                         cond_resched();
 268                         spin_lock(&sched->ws_lock);
 269                         continue;
 270                 }
 271
 272                 spin_unlock(&sched->ws_lock);
 273                 rc = wait_event_interruptible_exclusive(sched->ws_waitq,
 274                                 !cfs_wi_sched_cansleep(sched));
 275                 spin_lock(&sched->ws_lock);
 276         }
 277
 278         spin_unlock(&sched->ws_lock);
 279
 280         spin_lock(&cfs_wi_data.wi_glock);
 281         sched->ws_nthreads--;
 282         spin_unlock(&cfs_wi_data.wi_glock);
 283
 284         return 0;
 285 }
 286
 287 void
 288 cfs_wi_sched_destroy(struct cfs_wi_sched *sched)
 289 {
 290         LASSERT(cfs_wi_data.wi_init);
 291         LASSERT(!cfs_wi_data.wi_stopping);
 292
 293         spin_lock(&cfs_wi_data.wi_glock);
 294         if (sched->ws_stopping) {
 295                 CDEBUG(D_INFO, "%s is in progress of stopping\n",
 296                        sched->ws_name);
 297                 spin_unlock(&cfs_wi_data.wi_glock);
 298                 return;
 299         }
 300
 301         LASSERT(!list_empty(&sched->ws_list));
 302         sched->ws_stopping = 1;
 303
 304         spin_unlock(&cfs_wi_data.wi_glock);
 305
 306         wake_up_all(&sched->ws_waitq);
 307
 308         spin_lock(&cfs_wi_data.wi_glock);
 309         {
 310                 int i = 2;
 311
 312                 while (sched->ws_nthreads > 0) {
 313                         CDEBUG(is_power_of_2(++i / 20) ? D_WARNING : D_NET,
 314                                "waiting %us for %d %s worker threads to exit\n",
 315                                i / 20, sched->ws_nthreads, sched->ws_name);
 316
 317                         spin_unlock(&cfs_wi_data.wi_glock);
 318                         set_current_state(TASK_UNINTERRUPTIBLE);
 319                         schedule_timeout(cfs_time_seconds(1) / 20);
 320                         spin_lock(&cfs_wi_data.wi_glock);
 321                 }
 322         }
 323
 324         list_del(&sched->ws_list);
 325
 326         spin_unlock(&cfs_wi_data.wi_glock);
 327
 328         LASSERT(sched->ws_nscheduled == 0);
 329
 330         LIBCFS_FREE(sched, sizeof(*sched));
 331 }
 332 EXPORT_SYMBOL(cfs_wi_sched_destroy);
 333
 334 int
 335 cfs_wi_sched_create(char *name, struct cfs_cpt_table *cptab,
 336                     int cpt, int nthrs, struct cfs_wi_sched **sched_pp)
 337 {
 338         struct cfs_wi_sched     *sched;
 339
 340         LASSERT(cfs_wi_data.wi_init);
 341         LASSERT(!cfs_wi_data.wi_stopping);
 342         LASSERT(cptab == NULL || cpt == CFS_CPT_ANY ||
 343                 (cpt >= 0 && cpt < cfs_cpt_number(cptab)));
 344
 345         LIBCFS_ALLOC(sched, sizeof(*sched));
 346         if (sched == NULL)
 347                 return -ENOMEM;
 348
 349         if (strlen(name) > sizeof(sched->ws_name)-1) {
 350                 LIBCFS_FREE(sched, sizeof(*sched));
 351                 return -E2BIG;
 352         }
 353         strlcpy(sched->ws_name, name, sizeof(sched->ws_name));
 354
 355         sched->ws_cptab = cptab;
 356         sched->ws_cpt = cpt;
 357
 358         spin_lock_init(&sched->ws_lock);
 359         init_waitqueue_head(&sched->ws_waitq);
 360
 361         INIT_LIST_HEAD(&sched->ws_runq);
 362         INIT_LIST_HEAD(&sched->ws_rerunq);
 363         INIT_LIST_HEAD(&sched->ws_list);
 364
 365         for (; nthrs > 0; nthrs--)  {
 366                 char                    name[16];
 367                 struct task_struct      *task;
 368
 369                 spin_lock(&cfs_wi_data.wi_glock);
 370                 while (sched->ws_starting > 0) {
 371                         spin_unlock(&cfs_wi_data.wi_glock);
 372                         schedule();
 373                         spin_lock(&cfs_wi_data.wi_glock);
 374                 }
 375
 376                 sched->ws_starting++;
 377                 spin_unlock(&cfs_wi_data.wi_glock);
 378
 379                 if (sched->ws_cptab != NULL && sched->ws_cpt >= 0) {
 380                         snprintf(name, sizeof(name), "%s_%02d_%02d",
 381                                  sched->ws_name, sched->ws_cpt,
 382                                  sched->ws_nthreads);
 383                 } else {
 384                         snprintf(name, sizeof(name), "%s_%02d",
 385                                  sched->ws_name, sched->ws_nthreads);
 386                 }
 387
 388                 task = kthread_run(cfs_wi_scheduler, sched, name);
 389                 if (IS_ERR(task)) {
 390                         int rc = PTR_ERR(task);
 391
 392                         CERROR("Failed to create thread for "
 393                                 "WI scheduler %s: %d\n", name, rc);
 394
 395                         spin_lock(&cfs_wi_data.wi_glock);
 396
 397                         /* make up for cfs_wi_sched_destroy */
 398                         list_add(&sched->ws_list, &cfs_wi_data.wi_scheds);
 399                         sched->ws_starting--;
 400
 401                         spin_unlock(&cfs_wi_data.wi_glock);
 402
 403                         cfs_wi_sched_destroy(sched);
 404                         return rc;
 405                 }
 406         }
 407
 408         spin_lock(&cfs_wi_data.wi_glock);
 409         list_add(&sched->ws_list, &cfs_wi_data.wi_scheds);
 410         spin_unlock(&cfs_wi_data.wi_glock);
 411
 412         *sched_pp = sched;
 413         return 0;
 414 }
 415 EXPORT_SYMBOL(cfs_wi_sched_create);
 416
 417 int
 418 cfs_wi_startup(void)
 419 {
 420         memset(&cfs_wi_data, 0, sizeof(struct cfs_workitem_data));
 421
 422         spin_lock_init(&cfs_wi_data.wi_glock);
 423         INIT_LIST_HEAD(&cfs_wi_data.wi_scheds);
 424         cfs_wi_data.wi_init = 1;
 425
 426         return 0;
 427 }
 428
 429 void
 430 cfs_wi_shutdown (void)
 431 {
 432         struct cfs_wi_sched     *sched;
 433
 434         spin_lock(&cfs_wi_data.wi_glock);
 435         cfs_wi_data.wi_stopping = 1;
 436         spin_unlock(&cfs_wi_data.wi_glock);
 437
 438         /* nobody should contend on this list */
 439         list_for_each_entry(sched, &cfs_wi_data.wi_scheds, ws_list) {
 440                 sched->ws_stopping = 1;
 441                 wake_up_all(&sched->ws_waitq);
 442         }
 443
 444         list_for_each_entry(sched, &cfs_wi_data.wi_scheds, ws_list) {
 445                 spin_lock(&cfs_wi_data.wi_glock);
 446
 447                 while (sched->ws_nthreads != 0) {
 448                         spin_unlock(&cfs_wi_data.wi_glock);
 449                         set_current_state(TASK_UNINTERRUPTIBLE);
 450                         schedule_timeout(cfs_time_seconds(1) / 20);
 451                         spin_lock(&cfs_wi_data.wi_glock);
 452                 }
 453                 spin_unlock(&cfs_wi_data.wi_glock);
 454         }
 455
 456         while (!list_empty(&cfs_wi_data.wi_scheds)) {
 457                 sched = list_entry(cfs_wi_data.wi_scheds.next,
 458                                        struct cfs_wi_sched, ws_list);
 459                 list_del(&sched->ws_list);
 460                 LIBCFS_FREE(sched, sizeof(*sched));
 461         }
 462
 463         cfs_wi_data.wi_stopping = 0;
 464         cfs_wi_data.wi_init = 0;
 465 }