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, 2014, 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 <linux/kthread.h>
  45 #include <libcfs/libcfs.h>
  46
  47 #define CFS_WS_NAME_LEN         16
  48
  49 struct cfs_wi_sched {
  50         struct list_head                ws_list;        /* chain on global list */
  51         /** serialised workitems */
  52         spinlock_t                      ws_lock;
  53         /** where schedulers sleep */
  54         wait_queue_head_t               ws_waitq;
  55         /** concurrent workitems */
  56         struct list_head                ws_runq;
  57         /** rescheduled running-workitems, a workitem can be rescheduled
  58          * while running in wi_action(), but we don't to execute it again
  59          * unless it returns from wi_action(), so we put it on ws_rerunq
  60          * while rescheduling, and move it to runq after it returns
  61          * from wi_action() */
  62         struct list_head                ws_rerunq;
  63         /** CPT-table for this scheduler */
  64         struct cfs_cpt_table    *ws_cptab;
  65         /** CPT id for affinity */
  66         int                     ws_cpt;
  67         /** number of scheduled workitems */
  68         int                     ws_nscheduled;
  69         /** started scheduler thread, protected by cfs_wi_data::wi_glock */
  70         unsigned int            ws_nthreads:30;
  71         /** shutting down, protected by cfs_wi_data::wi_glock */
  72         unsigned int            ws_stopping:1;
  73         /** serialize starting thread, protected by cfs_wi_data::wi_glock */
  74         unsigned int            ws_starting:1;
  75         /** scheduler name */
  76         char                    ws_name[CFS_WS_NAME_LEN];
  77 };
  78
  79 static struct cfs_workitem_data {
  80         /** serialize */
  81         spinlock_t              wi_glock;
  82         /** list of all schedulers */
  83         struct list_head        wi_scheds;
  84         /** WI module is initialized */
  85         int                     wi_init;
  86         /** shutting down the whole WI module */
  87         int                     wi_stopping;
  88 } cfs_wi_data;
  89
  90 static inline int
  91 cfs_wi_sched_cansleep(struct cfs_wi_sched *sched)
  92 {
  93         spin_lock(&sched->ws_lock);
  94         if (sched->ws_stopping) {
  95                 spin_unlock(&sched->ws_lock);
  96                 return 0;
  97         }
  98
  99         if (!list_empty(&sched->ws_runq)) {
 100                 spin_unlock(&sched->ws_lock);
 101                 return 0;
 102         }
 103         spin_unlock(&sched->ws_lock);
 104         return 1;
 105 }
 106
 107 /* XXX:
 108  * 0. it only works when called from wi->wi_action.
 109  * 1. when it returns no one shall try to schedule the workitem.
 110  */
 111 void
 112 cfs_wi_exit(struct cfs_wi_sched *sched, struct cfs_workitem *wi)
 113 {
 114         LASSERT(!in_interrupt()); /* because we use plain spinlock */
 115         LASSERT(!sched->ws_stopping);
 116
 117         spin_lock(&sched->ws_lock);
 118
 119         LASSERT(wi->wi_running);
 120
 121         if (wi->wi_scheduled) { /* cancel pending schedules */
 122                 LASSERT(!list_empty(&wi->wi_list));
 123                 list_del_init(&wi->wi_list);
 124
 125                 LASSERT(sched->ws_nscheduled > 0);
 126                 sched->ws_nscheduled--;
 127         }
 128
 129         LASSERT(list_empty(&wi->wi_list));
 130
 131         wi->wi_scheduled = 1; /* LBUG future schedule attempts */
 132         spin_unlock(&sched->ws_lock);
 133
 134         return;
 135 }
 136 EXPORT_SYMBOL(cfs_wi_exit);
 137
 138 /**
 139  * cancel schedule request of workitem \a wi
 140  */
 141 int
 142 cfs_wi_deschedule(struct cfs_wi_sched *sched, struct cfs_workitem *wi)
 143 {
 144         int     rc;
 145
 146         LASSERT(!in_interrupt()); /* because we use plain spinlock */
 147         LASSERT(!sched->ws_stopping);
 148
 149         /*
 150          * return 0 if it's running already, otherwise return 1, which
 151          * means the workitem will not be scheduled and will not have
 152          * any race with wi_action.
 153          */
 154         spin_lock(&sched->ws_lock);
 155
 156         rc = !(wi->wi_running);
 157
 158         if (wi->wi_scheduled) { /* cancel pending schedules */
 159                 LASSERT(!list_empty(&wi->wi_list));
 160                 list_del_init(&wi->wi_list);
 161
 162                 LASSERT(sched->ws_nscheduled > 0);
 163                 sched->ws_nscheduled--;
 164
 165                 wi->wi_scheduled = 0;
 166         }
 167
 168         LASSERT (list_empty(&wi->wi_list));
 169
 170         spin_unlock(&sched->ws_lock);
 171         return rc;
 172 }
 173 EXPORT_SYMBOL(cfs_wi_deschedule);
 174
 175 /*
 176  * Workitem scheduled with (serial == 1) is strictly serialised not only with
 177  * itself, but also with others scheduled this way.
 178  *
 179  * Now there's only one static serialised queue, but in the future more might
 180  * be added, and even dynamic creation of serialised queues might be supported.
 181  */
 182 void
 183 cfs_wi_schedule(struct cfs_wi_sched *sched, struct cfs_workitem *wi)
 184 {
 185         LASSERT(!in_interrupt()); /* because we use plain spinlock */
 186         LASSERT(!sched->ws_stopping);
 187
 188         spin_lock(&sched->ws_lock);
 189
 190         if (!wi->wi_scheduled) {
 191                 LASSERT (list_empty(&wi->wi_list));
 192
 193                 wi->wi_scheduled = 1;
 194                 sched->ws_nscheduled++;
 195                 if (!wi->wi_running) {
 196                         list_add_tail(&wi->wi_list, &sched->ws_runq);
 197                         wake_up(&sched->ws_waitq);
 198                 } else {
 199                         list_add(&wi->wi_list, &sched->ws_rerunq);
 200                 }
 201         }
 202
 203         LASSERT (!list_empty(&wi->wi_list));
 204         spin_unlock(&sched->ws_lock);
 205         return;
 206 }
 207 EXPORT_SYMBOL(cfs_wi_schedule);
 208
 209 static int
 210 cfs_wi_scheduler(void *arg)
 211 {
 212         struct cfs_wi_sched *sched = (struct cfs_wi_sched *)arg;
 213
 214         cfs_block_allsigs();
 215
 216         /* CPT affinity scheduler? */
 217         if (sched->ws_cptab != NULL)
 218                 if (cfs_cpt_bind(sched->ws_cptab, sched->ws_cpt) != 0)
 219                         CWARN("Failed to bind %s on CPT %d\n",
 220                                 sched->ws_name, sched->ws_cpt);
 221
 222         spin_lock(&cfs_wi_data.wi_glock);
 223
 224         LASSERT(sched->ws_starting == 1);
 225         sched->ws_starting--;
 226         sched->ws_nthreads++;
 227
 228         spin_unlock(&cfs_wi_data.wi_glock);
 229
 230         spin_lock(&sched->ws_lock);
 231
 232         while (!sched->ws_stopping) {
 233                 int             nloops = 0;
 234                 int             rc;
 235                 struct cfs_workitem *wi;
 236
 237                 while (!list_empty(&sched->ws_runq) &&
 238                        nloops < CFS_WI_RESCHED) {
 239                         wi = list_entry(sched->ws_runq.next,
 240                                         struct cfs_workitem, wi_list);
 241                         LASSERT(wi->wi_scheduled && !wi->wi_running);
 242
 243                         list_del_init(&wi->wi_list);
 244
 245                         LASSERT(sched->ws_nscheduled > 0);
 246                         sched->ws_nscheduled--;
 247
 248                         wi->wi_running   = 1;
 249                         wi->wi_scheduled = 0;
 250
 251                         spin_unlock(&sched->ws_lock);
 252                         nloops++;
 253
 254                         rc = (*wi->wi_action) (wi);
 255
 256                         spin_lock(&sched->ws_lock);
 257                         if (rc != 0) /* WI should be dead, even be freed! */
 258                                 continue;
 259
 260                         wi->wi_running = 0;
 261                         if (list_empty(&wi->wi_list))
 262                                 continue;
 263
 264                         LASSERT(wi->wi_scheduled);
 265                         /* wi is rescheduled, should be on rerunq now, we
 266                          * move it to runq so it can run action now */
 267                         list_move_tail(&wi->wi_list, &sched->ws_runq);
 268                 }
 269
 270                 if (!list_empty(&sched->ws_runq)) {
 271                         spin_unlock(&sched->ws_lock);
 272                         /* don't sleep because some workitems still
 273                          * expect me to come back soon */
 274                         cond_resched();
 275                         spin_lock(&sched->ws_lock);
 276                         continue;
 277                 }
 278
 279                 spin_unlock(&sched->ws_lock);
 280                 rc = wait_event_interruptible_exclusive(sched->ws_waitq,
 281                                 !cfs_wi_sched_cansleep(sched));
 282                 spin_lock(&sched->ws_lock);
 283         }
 284
 285         spin_unlock(&sched->ws_lock);
 286
 287         spin_lock(&cfs_wi_data.wi_glock);
 288         sched->ws_nthreads--;
 289         spin_unlock(&cfs_wi_data.wi_glock);
 290
 291         return 0;
 292 }
 293
 294 void
 295 cfs_wi_sched_destroy(struct cfs_wi_sched *sched)
 296 {
 297         LASSERT(cfs_wi_data.wi_init);
 298         LASSERT(!cfs_wi_data.wi_stopping);
 299
 300         spin_lock(&cfs_wi_data.wi_glock);
 301         if (sched->ws_stopping) {
 302                 CDEBUG(D_INFO, "%s is in progress of stopping\n",
 303                        sched->ws_name);
 304                 spin_unlock(&cfs_wi_data.wi_glock);
 305                 return;
 306         }
 307
 308         LASSERT(!list_empty(&sched->ws_list));
 309         sched->ws_stopping = 1;
 310
 311         spin_unlock(&cfs_wi_data.wi_glock);
 312
 313         wake_up_all(&sched->ws_waitq);
 314
 315         spin_lock(&cfs_wi_data.wi_glock);
 316         {
 317                 int i = 2;
 318
 319                 while (sched->ws_nthreads > 0) {
 320                         CDEBUG(IS_PO2(++i) ? D_WARNING : D_NET,
 321                                "waiting for %d threads of WI sched[%s] to "
 322                                "terminate\n", sched->ws_nthreads,
 323                                sched->ws_name);
 324
 325                         spin_unlock(&cfs_wi_data.wi_glock);
 326                         set_current_state(TASK_UNINTERRUPTIBLE);
 327                         schedule_timeout(cfs_time_seconds(1) / 20);
 328                         spin_lock(&cfs_wi_data.wi_glock);
 329                 }
 330         }
 331
 332         list_del(&sched->ws_list);
 333
 334         spin_unlock(&cfs_wi_data.wi_glock);
 335
 336         LASSERT(sched->ws_nscheduled == 0);
 337
 338         LIBCFS_FREE(sched, sizeof(*sched));
 339 }
 340 EXPORT_SYMBOL(cfs_wi_sched_destroy);
 341
 342 int
 343 cfs_wi_sched_create(char *name, struct cfs_cpt_table *cptab,
 344                     int cpt, int nthrs, struct cfs_wi_sched **sched_pp)
 345 {
 346         struct cfs_wi_sched     *sched;
 347
 348         LASSERT(cfs_wi_data.wi_init);
 349         LASSERT(!cfs_wi_data.wi_stopping);
 350         LASSERT(cptab == NULL || cpt == CFS_CPT_ANY ||
 351                 (cpt >= 0 && cpt < cfs_cpt_number(cptab)));
 352
 353         LIBCFS_ALLOC(sched, sizeof(*sched));
 354         if (sched == NULL)
 355                 return -ENOMEM;
 356
 357         if (strlen(name) > sizeof(sched->ws_name)-1) {
 358                 LIBCFS_FREE(sched, sizeof(*sched));
 359                 return -E2BIG;
 360         }
 361         strlcpy(sched->ws_name, name, sizeof(sched->ws_name));
 362
 363         sched->ws_cptab = cptab;
 364         sched->ws_cpt = cpt;
 365
 366         spin_lock_init(&sched->ws_lock);
 367         init_waitqueue_head(&sched->ws_waitq);
 368
 369         INIT_LIST_HEAD(&sched->ws_runq);
 370         INIT_LIST_HEAD(&sched->ws_rerunq);
 371         INIT_LIST_HEAD(&sched->ws_list);
 372
 373         for (; nthrs > 0; nthrs--)  {
 374                 char                    name[16];
 375                 struct task_struct      *task;
 376
 377                 spin_lock(&cfs_wi_data.wi_glock);
 378                 while (sched->ws_starting > 0) {
 379                         spin_unlock(&cfs_wi_data.wi_glock);
 380                         schedule();
 381                         spin_lock(&cfs_wi_data.wi_glock);
 382                 }
 383
 384                 sched->ws_starting++;
 385                 spin_unlock(&cfs_wi_data.wi_glock);
 386
 387                 if (sched->ws_cptab != NULL && sched->ws_cpt >= 0) {
 388                         snprintf(name, sizeof(name), "%s_%02d_%02d",
 389                                  sched->ws_name, sched->ws_cpt,
 390                                  sched->ws_nthreads);
 391                 } else {
 392                         snprintf(name, sizeof(name), "%s_%02d",
 393                                  sched->ws_name, sched->ws_nthreads);
 394                 }
 395
 396                 task = kthread_run(cfs_wi_scheduler, sched, name);
 397                 if (IS_ERR(task)) {
 398                         int rc = PTR_ERR(task);
 399
 400                         CERROR("Failed to create thread for "
 401                                 "WI scheduler %s: %d\n", name, rc);
 402
 403                         spin_lock(&cfs_wi_data.wi_glock);
 404
 405                         /* make up for cfs_wi_sched_destroy */
 406                         list_add(&sched->ws_list, &cfs_wi_data.wi_scheds);
 407                         sched->ws_starting--;
 408
 409                         spin_unlock(&cfs_wi_data.wi_glock);
 410
 411                         cfs_wi_sched_destroy(sched);
 412                         return rc;
 413                 }
 414         }
 415
 416         spin_lock(&cfs_wi_data.wi_glock);
 417         list_add(&sched->ws_list, &cfs_wi_data.wi_scheds);
 418         spin_unlock(&cfs_wi_data.wi_glock);
 419
 420         *sched_pp = sched;
 421         return 0;
 422 }
 423 EXPORT_SYMBOL(cfs_wi_sched_create);
 424
 425 int
 426 cfs_wi_startup(void)
 427 {
 428         memset(&cfs_wi_data, 0, sizeof(struct cfs_workitem_data));
 429
 430         spin_lock_init(&cfs_wi_data.wi_glock);
 431         INIT_LIST_HEAD(&cfs_wi_data.wi_scheds);
 432         cfs_wi_data.wi_init = 1;
 433
 434         return 0;
 435 }
 436
 437 void
 438 cfs_wi_shutdown (void)
 439 {
 440         struct cfs_wi_sched     *sched;
 441
 442         spin_lock(&cfs_wi_data.wi_glock);
 443         cfs_wi_data.wi_stopping = 1;
 444         spin_unlock(&cfs_wi_data.wi_glock);
 445
 446         /* nobody should contend on this list */
 447         list_for_each_entry(sched, &cfs_wi_data.wi_scheds, ws_list) {
 448                 sched->ws_stopping = 1;
 449                 wake_up_all(&sched->ws_waitq);
 450         }
 451
 452         list_for_each_entry(sched, &cfs_wi_data.wi_scheds, ws_list) {
 453                 spin_lock(&cfs_wi_data.wi_glock);
 454
 455                 while (sched->ws_nthreads != 0) {
 456                         spin_unlock(&cfs_wi_data.wi_glock);
 457                         set_current_state(TASK_UNINTERRUPTIBLE);
 458                         schedule_timeout(cfs_time_seconds(1) / 20);
 459                         spin_lock(&cfs_wi_data.wi_glock);
 460                 }
 461                 spin_unlock(&cfs_wi_data.wi_glock);
 462         }
 463
 464         while (!list_empty(&cfs_wi_data.wi_scheds)) {
 465                 sched = list_entry(cfs_wi_data.wi_scheds.next,
 466                                        struct cfs_wi_sched, ws_list);
 467                 list_del(&sched->ws_list);
 468                 LIBCFS_FREE(sched, sizeof(*sched));
 469         }
 470
 471         cfs_wi_data.wi_stopping = 0;
 472         cfs_wi_data.wi_init = 0;
 473 }