lustre/quota/qsd_writeback.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, write to the
  18  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  19  * Boston, MA 021110-1307, USA
  20  *
  21  * GPL HEADER END
  22  */
  23 /*
  24  * Copyright (c) 2012, 2014, Intel Corporation.
  25  * Use is subject to license terms.
  26  *
  27  * Author: Johann Lombardi <johann.lombardi@intel.com>
  28  * Author: Niu    Yawei    <yawei.niu@intel.com>
  29  */
  30
  31 #define DEBUG_SUBSYSTEM S_LQUOTA
  32
  33 #include <linux/kthread.h>
  34 #include "qsd_internal.h"
  35
  36 extern struct kmem_cache *upd_kmem;
  37
  38 /*
  39  * Allocate and fill an qsd_upd_rec structure to be processed by the writeback
  40  * thread.
  41  *
  42  * \param qqi - is the qsd_qtype_info structure relevant to the update
  43  * \param lqe - is the lquota entry subject to the update
  44  * \param qid - is the identifier subject to the update
  45  * \param rec - is the record storing the new quota settings
  46  * \param ver - is the version associated with the update
  47  * \param global - is a boolean set to true if this is an update of the global
  48  *                 index and false for a slave index.
  49  */
  50 static struct qsd_upd_rec *qsd_upd_alloc(struct qsd_qtype_info *qqi,
  51                                          struct lquota_entry *lqe,
  52                                          union lquota_id *qid,
  53                                          union lquota_rec *rec, __u64 ver,
  54                                          bool global)
  55 {
  56         struct qsd_upd_rec      *upd;
  57
  58         OBD_SLAB_ALLOC_PTR_GFP(upd, upd_kmem, GFP_NOFS);
  59         if (upd == NULL) {
  60                 return NULL;
  61         }
  62
  63         /* fill it */
  64         INIT_LIST_HEAD(&upd->qur_link);
  65         upd->qur_qqi = qqi;
  66         upd->qur_lqe = lqe;
  67         if (lqe)
  68                 lqe_getref(lqe);
  69         upd->qur_qid    = *qid;
  70         upd->qur_rec    = *rec;
  71         upd->qur_ver    = ver;
  72         upd->qur_global = global;
  73
  74         return upd;
  75 }
  76
  77 static void qsd_upd_free(struct qsd_upd_rec *upd)
  78 {
  79         if (upd->qur_lqe)
  80                 lqe_putref(upd->qur_lqe);
  81         OBD_SLAB_FREE_PTR(upd, upd_kmem);
  82 }
  83
  84 /* must hold the qsd_lock */
  85 static void qsd_upd_add(struct qsd_instance *qsd, struct qsd_upd_rec *upd)
  86 {
  87         if (!qsd->qsd_stopping) {
  88                 list_add_tail(&upd->qur_link, &qsd->qsd_upd_list);
  89                 /* wake up the upd thread */
  90                 wake_up(&qsd->qsd_upd_thread.t_ctl_waitq);
  91         } else {
  92                 CWARN("%s: discard update.\n", qsd->qsd_svname);
  93                 if (upd->qur_lqe)
  94                         LQUOTA_WARN(upd->qur_lqe, "discard update.");
  95                 qsd_upd_free(upd);
  96         }
  97 }
  98
  99 /* must hold the qsd_lock */
 100 static void qsd_add_deferred(struct qsd_instance *qsd, struct list_head *list,
 101                              struct qsd_upd_rec *upd)
 102 {
 103         struct qsd_upd_rec      *tmp, *n;
 104
 105         if (qsd->qsd_stopping) {
 106                 CWARN("%s: discard deferred udpate.\n", qsd->qsd_svname);
 107                 if (upd->qur_lqe)
 108                         LQUOTA_WARN(upd->qur_lqe, "discard deferred update.");
 109                 qsd_upd_free(upd);
 110                 return;
 111         }
 112
 113         /* Sort the updates in ascending order */
 114         list_for_each_entry_safe_reverse(tmp, n, list, qur_link) {
 115
 116                 /* There could be some legacy records which have duplicated
 117                  * version. Imagine following scenario: slave received global
 118                  * glimpse and queued a record in the deferred list, then
 119                  * master crash and rollback to an ealier version, then the
 120                  * version of queued record will be conflicting with later
 121                  * updates. We should just delete the legacy record in such
 122                  * case. */
 123                 if (upd->qur_ver == tmp->qur_ver) {
 124                         LASSERT(tmp->qur_lqe);
 125                         LQUOTA_ERROR(tmp->qur_lqe, "Found a conflict record "
 126                                      "with ver:%llu", tmp->qur_ver);
 127                         list_del_init(&tmp->qur_link);
 128                         qsd_upd_free(tmp);
 129                 } else if (upd->qur_ver < tmp->qur_ver) {
 130                         continue;
 131                 } else {
 132                         list_add_tail(&upd->qur_link, &tmp->qur_link);
 133                         return;
 134                 }
 135         }
 136         list_add(&upd->qur_link, list);
 137 }
 138
 139 /* must hold the qsd_lock */
 140 static void qsd_kickoff_deferred(struct qsd_qtype_info *qqi,
 141                                  struct list_head *list, __u64 ver)
 142 {
 143         struct qsd_upd_rec      *upd, *tmp;
 144         ENTRY;
 145
 146         /* Get the first update record in the list, which has the smallest
 147          * version, discard all records with versions smaller than the current
 148          * one */
 149         list_for_each_entry_safe(upd, tmp, list, qur_link) {
 150                 if (upd->qur_ver <= ver) {
 151                         /* drop this update */
 152                         list_del_init(&upd->qur_link);
 153                         CDEBUG(D_QUOTA, "%s: skipping deferred update ver:"
 154                                "%llu/%llu, global:%d, qid:%llu\n",
 155                                qqi->qqi_qsd->qsd_svname, upd->qur_ver, ver,
 156                                upd->qur_global, upd->qur_qid.qid_uid);
 157                         qsd_upd_free(upd);
 158                 } else {
 159                         break;
 160                 }
 161         }
 162
 163         /* No remaining deferred update */
 164         if (list_empty(list))
 165                 RETURN_EXIT;
 166
 167         CDEBUG(D_QUOTA, "%s: found deferred update record. "
 168                "version:%llu/%llu, global:%d, qid:%llu\n",
 169                qqi->qqi_qsd->qsd_svname, upd->qur_ver, ver,
 170                upd->qur_global, upd->qur_qid.qid_uid);
 171
 172         LASSERTF(upd->qur_ver > ver, "lur_ver:%llu, cur_ver:%llu\n",
 173                  upd->qur_ver, ver);
 174
 175         /* Kick off the deferred udpate */
 176         if (upd->qur_ver == ver + 1) {
 177                 list_del_init(&upd->qur_link);
 178                 qsd_upd_add(qqi->qqi_qsd, upd);
 179         }
 180         EXIT;
 181 }
 182
 183 /* Bump version of global or slave index copy
 184  *
 185  * \param qqi    - qsd_qtype_info
 186  * \param ver    - version to be bumped to
 187  * \param global - global or slave index copy?
 188  */
 189 void qsd_bump_version(struct qsd_qtype_info *qqi, __u64 ver, bool global)
 190 {
 191         struct list_head *list;
 192         __u64            *idx_ver;
 193
 194         idx_ver = global ? &qqi->qqi_glb_ver : &qqi->qqi_slv_ver;
 195         list    = global ? &qqi->qqi_deferred_glb : &qqi->qqi_deferred_slv;
 196
 197         write_lock(&qqi->qqi_qsd->qsd_lock);
 198         *idx_ver = ver;
 199         if (global)
 200                 qqi->qqi_glb_uptodate = 1;
 201         else
 202                 qqi->qqi_slv_uptodate = 1;
 203         qsd_kickoff_deferred(qqi, list, ver);
 204         write_unlock(&qqi->qqi_qsd->qsd_lock);
 205 }
 206
 207 /*
 208  * Schedule a commit of a lquota entry
 209  *
 210  * \param  qqi   - qsd_qtype_info
 211  * \param  lqe   - lquota_entry
 212  * \param  qid   - quota id
 213  * \param  rec   - global or slave record to be updated to disk
 214  * \param  ver   - new index file version
 215  * \param  global- true: master record; false: slave record
 216  */
 217 void qsd_upd_schedule(struct qsd_qtype_info *qqi, struct lquota_entry *lqe,
 218                       union lquota_id *qid, union lquota_rec *rec, __u64 ver,
 219                       bool global)
 220 {
 221         struct qsd_upd_rec      *upd;
 222         struct qsd_instance     *qsd = qqi->qqi_qsd;
 223         __u64                    cur_ver;
 224         ENTRY;
 225
 226         CDEBUG(D_QUOTA, "%s: schedule update. global:%s, version:%llu\n",
 227                qsd->qsd_svname, global ? "true" : "false", ver);
 228
 229         upd = qsd_upd_alloc(qqi, lqe, qid, rec, ver, global);
 230         if (upd == NULL)
 231                 RETURN_EXIT;
 232
 233         /* If we don't want update index version, no need to sort the
 234          * records in version order, just schedule the updates instantly. */
 235         if (ver == 0) {
 236                 write_lock(&qsd->qsd_lock);
 237                 qsd_upd_add(qsd, upd);
 238                 write_unlock(&qsd->qsd_lock);
 239                 RETURN_EXIT;
 240         }
 241
 242         write_lock(&qsd->qsd_lock);
 243
 244         cur_ver = global ? qqi->qqi_glb_ver : qqi->qqi_slv_ver;
 245
 246         if (ver <= cur_ver) {
 247                 if (global)
 248                         /* legitimate race between glimpse AST and
 249                          * reintegration */
 250                         CDEBUG(D_QUOTA, "%s: discarding glb update from glimpse"
 251                                " ver:%llu local ver:%llu\n",
 252                                qsd->qsd_svname, ver, cur_ver);
 253                 else
 254                         CERROR("%s: discard slv update, ver:%llu local ver:"
 255                                "%llu\n", qsd->qsd_svname, ver, cur_ver);
 256                 qsd_upd_free(upd);
 257         } else if ((ver == cur_ver + 1) && qqi->qqi_glb_uptodate &&
 258                    qqi->qqi_slv_uptodate) {
 259                 /* In order update, and reintegration has been done. */
 260                 qsd_upd_add(qsd, upd);
 261         } else {
 262                 /* Out of order update (the one with smaller version hasn't
 263                  * reached slave or hasn't been flushed to disk yet), or
 264                  * the reintegration is in progress. Defer the update. */
 265                 struct list_head *list = global ? &qqi->qqi_deferred_glb :
 266                                                   &qqi->qqi_deferred_slv;
 267                 qsd_add_deferred(qsd, list, upd);
 268         }
 269
 270         write_unlock(&qsd->qsd_lock);
 271
 272         EXIT;
 273 }
 274
 275 static int qsd_process_upd(const struct lu_env *env, struct qsd_upd_rec *upd)
 276 {
 277         struct lquota_entry     *lqe = upd->qur_lqe;
 278         struct qsd_qtype_info   *qqi = upd->qur_qqi;
 279         int                      rc;
 280         ENTRY;
 281
 282         if (lqe == NULL) {
 283                 lqe = lqe_locate(env, qqi->qqi_site, &upd->qur_qid);
 284                 if (IS_ERR(lqe))
 285                         GOTO(out, rc = PTR_ERR(lqe));
 286         }
 287
 288         /* The in-memory lqe update for slave index copy isn't deferred,
 289          * we shouldn't touch it here. */
 290         if (upd->qur_global) {
 291                 rc = qsd_update_lqe(env, lqe, upd->qur_global, &upd->qur_rec);
 292                 if (rc)
 293                         GOTO(out, rc);
 294                 /* refresh usage */
 295                 qsd_refresh_usage(env, lqe);
 296                 /* Report usage asynchronously */
 297                 rc = qsd_adjust(env, lqe);
 298                 if (rc)
 299                         LQUOTA_ERROR(lqe, "failed to report usage, rc:%d", rc);
 300         }
 301
 302         rc = qsd_update_index(env, qqi, &upd->qur_qid, upd->qur_global,
 303                               upd->qur_ver, &upd->qur_rec);
 304 out:
 305         if (lqe && !IS_ERR(lqe)) {
 306                 lqe_putref(lqe);
 307                 upd->qur_lqe = NULL;
 308         }
 309         RETURN(rc);
 310 }
 311
 312 void qsd_adjust_schedule(struct lquota_entry *lqe, bool defer, bool cancel)
 313 {
 314         struct qsd_instance     *qsd = lqe2qqi(lqe)->qqi_qsd;
 315         bool                     added = false;
 316
 317         read_lock(&qsd->qsd_lock);
 318         if (qsd->qsd_stopping) {
 319                 read_unlock(&qsd->qsd_lock);
 320                 return;
 321         }
 322         read_unlock(&qsd->qsd_lock);
 323
 324         lqe_getref(lqe);
 325         spin_lock(&qsd->qsd_adjust_lock);
 326
 327         /* the lqe is being queued for the per-ID lock cancel, we should
 328          * cancel the lock cancel and re-add it for quota adjust */
 329         if (!list_empty(&lqe->lqe_link) &&
 330             lqe->lqe_adjust_time == 0) {
 331                 list_del_init(&lqe->lqe_link);
 332                 lqe_putref(lqe);
 333         }
 334
 335         if (list_empty(&lqe->lqe_link)) {
 336                 if (cancel)
 337                         lqe->lqe_adjust_time = 0;
 338                 else
 339                         lqe->lqe_adjust_time = defer ?
 340                                 cfs_time_shift_64(QSD_WB_INTERVAL) :
 341                                 cfs_time_current_64();
 342                 /* lqe reference transferred to list */
 343                 if (defer)
 344                         list_add_tail(&lqe->lqe_link,
 345                                           &qsd->qsd_adjust_list);
 346                 else
 347                         list_add(&lqe->lqe_link, &qsd->qsd_adjust_list);
 348                 added = true;
 349         }
 350         spin_unlock(&qsd->qsd_adjust_lock);
 351
 352         if (added)
 353                 wake_up(&qsd->qsd_upd_thread.t_ctl_waitq);
 354         else
 355                 lqe_putref(lqe);
 356 }
 357
 358 /* return true if there is pending writeback records or the pending
 359  * adjust requests */
 360 static bool qsd_job_pending(struct qsd_instance *qsd, struct list_head *upd,
 361                             bool *uptodate)
 362 {
 363         bool    job_pending = false;
 364         int     qtype;
 365
 366         LASSERT(list_empty(upd));
 367         *uptodate = true;
 368
 369         spin_lock(&qsd->qsd_adjust_lock);
 370         if (!list_empty(&qsd->qsd_adjust_list)) {
 371                 struct lquota_entry *lqe;
 372                 lqe = list_entry(qsd->qsd_adjust_list.next,
 373                                      struct lquota_entry, lqe_link);
 374                 if (cfs_time_beforeq_64(lqe->lqe_adjust_time,
 375                                         cfs_time_current_64()))
 376                         job_pending = true;
 377         }
 378         spin_unlock(&qsd->qsd_adjust_lock);
 379
 380         write_lock(&qsd->qsd_lock);
 381         if (!list_empty(&qsd->qsd_upd_list)) {
 382                 list_splice_init(&qsd->qsd_upd_list, upd);
 383                 job_pending = true;
 384         }
 385
 386         if (qsd->qsd_acct_failed) {
 387                 /* don't bother kicking off reintegration if space accounting
 388                  * failed to be enabled */
 389                 write_unlock(&qsd->qsd_lock);
 390                 return job_pending;
 391         }
 392
 393         for (qtype = USRQUOTA; qtype < LL_MAXQUOTAS; qtype++) {
 394                 struct qsd_qtype_info *qqi = qsd->qsd_type_array[qtype];
 395
 396                 if (!qsd_type_enabled(qsd, qtype))
 397                         continue;
 398
 399                 if ((!qqi->qqi_glb_uptodate || !qqi->qqi_slv_uptodate) &&
 400                      !qqi->qqi_reint)
 401                         /* global or slave index not up to date and reint
 402                          * thread not running */
 403                         *uptodate = false;
 404         }
 405
 406         write_unlock(&qsd->qsd_lock);
 407         return job_pending;
 408 }
 409
 410 static int qsd_upd_thread(void *arg)
 411 {
 412         struct qsd_instance     *qsd = (struct qsd_instance *)arg;
 413         struct ptlrpc_thread    *thread = &qsd->qsd_upd_thread;
 414         struct l_wait_info       lwi;
 415         struct list_head         queue;
 416         struct qsd_upd_rec      *upd, *n;
 417         struct lu_env           *env;
 418         int                      qtype, rc = 0;
 419         bool                     uptodate;
 420         struct lquota_entry     *lqe;
 421         __u64                    cur_time;
 422         ENTRY;
 423
 424         OBD_ALLOC_PTR(env);
 425         if (env == NULL)
 426                 RETURN(-ENOMEM);
 427
 428         rc = lu_env_init(env, LCT_DT_THREAD);
 429         if (rc) {
 430                 CERROR("%s: cannot init env: rc = %d\n", qsd->qsd_svname, rc);
 431                 OBD_FREE_PTR(env);
 432                 RETURN(rc);
 433         }
 434
 435         thread_set_flags(thread, SVC_RUNNING);
 436         wake_up(&thread->t_ctl_waitq);
 437
 438         INIT_LIST_HEAD(&queue);
 439         lwi = LWI_TIMEOUT(cfs_time_seconds(QSD_WB_INTERVAL), NULL, NULL);
 440         while (1) {
 441                 l_wait_event(thread->t_ctl_waitq,
 442                              qsd_job_pending(qsd, &queue, &uptodate) ||
 443                              !thread_is_running(thread), &lwi);
 444
 445                 list_for_each_entry_safe(upd, n, &queue, qur_link) {
 446                         list_del_init(&upd->qur_link);
 447                         qsd_process_upd(env, upd);
 448                         qsd_upd_free(upd);
 449                 }
 450
 451                 spin_lock(&qsd->qsd_adjust_lock);
 452                 cur_time = cfs_time_current_64();
 453                 while (!list_empty(&qsd->qsd_adjust_list)) {
 454                         lqe = list_entry(qsd->qsd_adjust_list.next,
 455                                          struct lquota_entry, lqe_link);
 456                         /* deferred items are sorted by time */
 457                         if (!cfs_time_beforeq_64(lqe->lqe_adjust_time,
 458                                                  cur_time))
 459                                 break;
 460
 461                         list_del_init(&lqe->lqe_link);
 462                         spin_unlock(&qsd->qsd_adjust_lock);
 463
 464                         if (thread_is_running(thread) && uptodate) {
 465                                 qsd_refresh_usage(env, lqe);
 466                                 if (lqe->lqe_adjust_time == 0)
 467                                         qsd_id_lock_cancel(env, lqe);
 468                                 else
 469                                         qsd_adjust(env, lqe);
 470                         }
 471
 472                         lqe_putref(lqe);
 473                         spin_lock(&qsd->qsd_adjust_lock);
 474                 }
 475                 spin_unlock(&qsd->qsd_adjust_lock);
 476
 477                 if (!thread_is_running(thread))
 478                         break;
 479
 480                 if (uptodate)
 481                         continue;
 482
 483                 for (qtype = USRQUOTA; qtype < LL_MAXQUOTAS; qtype++)
 484                         qsd_start_reint_thread(qsd->qsd_type_array[qtype]);
 485         }
 486         lu_env_fini(env);
 487         OBD_FREE_PTR(env);
 488         thread_set_flags(thread, SVC_STOPPED);
 489         wake_up(&thread->t_ctl_waitq);
 490         RETURN(rc);
 491 }
 492
 493 int qsd_start_upd_thread(struct qsd_instance *qsd)
 494 {
 495         struct ptlrpc_thread    *thread = &qsd->qsd_upd_thread;
 496         struct l_wait_info       lwi = { 0 };
 497         struct task_struct              *task;
 498         ENTRY;
 499
 500         task = kthread_run(qsd_upd_thread, (void *)qsd,
 501                            "lquota_wb_%s", qsd->qsd_svname);
 502         if (IS_ERR(task)) {
 503                 CERROR("fail to start quota update thread: rc = %ld\n",
 504                         PTR_ERR(task));
 505                 thread_set_flags(thread, SVC_STOPPED);
 506                 RETURN(PTR_ERR(task));
 507         }
 508
 509         l_wait_event(thread->t_ctl_waitq,
 510                      thread_is_running(thread) || thread_is_stopped(thread),
 511                      &lwi);
 512         RETURN(0);
 513 }
 514
 515 static void qsd_cleanup_deferred(struct qsd_instance *qsd)
 516 {
 517         int     qtype;
 518
 519         for (qtype = USRQUOTA; qtype < LL_MAXQUOTAS; qtype++) {
 520                 struct qsd_upd_rec      *upd, *tmp;
 521                 struct qsd_qtype_info   *qqi = qsd->qsd_type_array[qtype];
 522
 523                 if (qqi == NULL)
 524                         continue;
 525
 526                 write_lock(&qsd->qsd_lock);
 527                 list_for_each_entry_safe(upd, tmp, &qqi->qqi_deferred_glb,
 528                                          qur_link) {
 529                         CWARN("%s: Free global deferred upd: ID:%llu, "
 530                               "ver:%llu/%llu\n", qsd->qsd_svname,
 531                               upd->qur_qid.qid_uid, upd->qur_ver,
 532                               qqi->qqi_glb_ver);
 533                         list_del_init(&upd->qur_link);
 534                         qsd_upd_free(upd);
 535                 }
 536                 list_for_each_entry_safe(upd, tmp, &qqi->qqi_deferred_slv,
 537                                          qur_link) {
 538                         CWARN("%s: Free slave deferred upd: ID:%llu, "
 539                               "ver:%llu/%llu\n", qsd->qsd_svname,
 540                               upd->qur_qid.qid_uid, upd->qur_ver,
 541                               qqi->qqi_slv_ver);
 542                         list_del_init(&upd->qur_link);
 543                         qsd_upd_free(upd);
 544                 }
 545                 write_unlock(&qsd->qsd_lock);
 546         }
 547 }
 548
 549 static void qsd_cleanup_adjust(struct qsd_instance *qsd)
 550 {
 551         struct lquota_entry     *lqe;
 552
 553         spin_lock(&qsd->qsd_adjust_lock);
 554         while (!list_empty(&qsd->qsd_adjust_list)) {
 555                 lqe = list_entry(qsd->qsd_adjust_list.next,
 556                                  struct lquota_entry, lqe_link);
 557                 list_del_init(&lqe->lqe_link);
 558                 lqe_putref(lqe);
 559         }
 560         spin_unlock(&qsd->qsd_adjust_lock);
 561 }
 562
 563 void qsd_stop_upd_thread(struct qsd_instance *qsd)
 564 {
 565         struct ptlrpc_thread    *thread = &qsd->qsd_upd_thread;
 566         struct l_wait_info       lwi    = { 0 };
 567
 568         if (!thread_is_stopped(thread)) {
 569                 thread_set_flags(thread, SVC_STOPPING);
 570                 wake_up(&thread->t_ctl_waitq);
 571
 572                 l_wait_event(thread->t_ctl_waitq, thread_is_stopped(thread),
 573                              &lwi);
 574         }
 575         qsd_cleanup_deferred(qsd);
 576         qsd_cleanup_adjust(qsd);
 577 }