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                         if (tmp->qur_lqe)
 125                                 LQUOTA_WARN(tmp->qur_lqe, "Found a conflict "
 126                                             "record with ver:%llu",
 127                                             tmp->qur_ver);
 128                         else
 129                                 CWARN("%s: Found a conflict record with ver: "
 130                                       "%llu\n", qsd->qsd_svname, tmp->qur_ver);
 131
 132                         list_del_init(&tmp->qur_link);
 133                         qsd_upd_free(tmp);
 134                 } else if (upd->qur_ver < tmp->qur_ver) {
 135                         continue;
 136                 } else {
 137                         list_add_tail(&upd->qur_link, &tmp->qur_link);
 138                         return;
 139                 }
 140         }
 141         list_add(&upd->qur_link, list);
 142 }
 143
 144 /* must hold the qsd_lock */
 145 static void qsd_kickoff_deferred(struct qsd_qtype_info *qqi,
 146                                  struct list_head *list, __u64 ver)
 147 {
 148         struct qsd_upd_rec      *upd, *tmp;
 149         ENTRY;
 150
 151         /* Get the first update record in the list, which has the smallest
 152          * version, discard all records with versions smaller than the current
 153          * one */
 154         list_for_each_entry_safe(upd, tmp, list, qur_link) {
 155                 if (upd->qur_ver <= ver) {
 156                         /* drop this update */
 157                         list_del_init(&upd->qur_link);
 158                         CDEBUG(D_QUOTA, "%s: skipping deferred update ver:"
 159                                "%llu/%llu, global:%d, qid:%llu\n",
 160                                qqi->qqi_qsd->qsd_svname, upd->qur_ver, ver,
 161                                upd->qur_global, upd->qur_qid.qid_uid);
 162                         qsd_upd_free(upd);
 163                 } else {
 164                         break;
 165                 }
 166         }
 167
 168         /* No remaining deferred update */
 169         if (list_empty(list))
 170                 RETURN_EXIT;
 171
 172         CDEBUG(D_QUOTA, "%s: found deferred update record. "
 173                "version:%llu/%llu, global:%d, qid:%llu\n",
 174                qqi->qqi_qsd->qsd_svname, upd->qur_ver, ver,
 175                upd->qur_global, upd->qur_qid.qid_uid);
 176
 177         LASSERTF(upd->qur_ver > ver, "lur_ver:%llu, cur_ver:%llu\n",
 178                  upd->qur_ver, ver);
 179
 180         /* Kick off the deferred udpate */
 181         if (upd->qur_ver == ver + 1) {
 182                 list_del_init(&upd->qur_link);
 183                 qsd_upd_add(qqi->qqi_qsd, upd);
 184         }
 185         EXIT;
 186 }
 187
 188 /* Bump version of global or slave index copy
 189  *
 190  * \param qqi    - qsd_qtype_info
 191  * \param ver    - version to be bumped to
 192  * \param global - global or slave index copy?
 193  */
 194 void qsd_bump_version(struct qsd_qtype_info *qqi, __u64 ver, bool global)
 195 {
 196         struct list_head *list;
 197         __u64            *idx_ver;
 198
 199         idx_ver = global ? &qqi->qqi_glb_ver : &qqi->qqi_slv_ver;
 200         list    = global ? &qqi->qqi_deferred_glb : &qqi->qqi_deferred_slv;
 201
 202         write_lock(&qqi->qqi_qsd->qsd_lock);
 203         *idx_ver = ver;
 204         if (global)
 205                 qqi->qqi_glb_uptodate = 1;
 206         else
 207                 qqi->qqi_slv_uptodate = 1;
 208         qsd_kickoff_deferred(qqi, list, ver);
 209         write_unlock(&qqi->qqi_qsd->qsd_lock);
 210 }
 211
 212 /*
 213  * Schedule a commit of a lquota entry
 214  *
 215  * \param  qqi   - qsd_qtype_info
 216  * \param  lqe   - lquota_entry
 217  * \param  qid   - quota id
 218  * \param  rec   - global or slave record to be updated to disk
 219  * \param  ver   - new index file version
 220  * \param  global- true: master record; false: slave record
 221  */
 222 void qsd_upd_schedule(struct qsd_qtype_info *qqi, struct lquota_entry *lqe,
 223                       union lquota_id *qid, union lquota_rec *rec, __u64 ver,
 224                       bool global)
 225 {
 226         struct qsd_upd_rec      *upd;
 227         struct qsd_instance     *qsd = qqi->qqi_qsd;
 228         __u64                    cur_ver;
 229         ENTRY;
 230
 231         CDEBUG(D_QUOTA, "%s: schedule update. global:%s, version:%llu\n",
 232                qsd->qsd_svname, global ? "true" : "false", ver);
 233
 234         upd = qsd_upd_alloc(qqi, lqe, qid, rec, ver, global);
 235         if (upd == NULL)
 236                 RETURN_EXIT;
 237
 238         /* If we don't want update index version, no need to sort the
 239          * records in version order, just schedule the updates instantly. */
 240         if (ver == 0) {
 241                 write_lock(&qsd->qsd_lock);
 242                 qsd_upd_add(qsd, upd);
 243                 write_unlock(&qsd->qsd_lock);
 244                 RETURN_EXIT;
 245         }
 246
 247         write_lock(&qsd->qsd_lock);
 248
 249         cur_ver = global ? qqi->qqi_glb_ver : qqi->qqi_slv_ver;
 250
 251         if (ver <= cur_ver) {
 252                 if (global)
 253                         /* legitimate race between glimpse AST and
 254                          * reintegration */
 255                         CDEBUG(D_QUOTA, "%s: discarding glb update from glimpse"
 256                                " ver:%llu local ver:%llu\n",
 257                                qsd->qsd_svname, ver, cur_ver);
 258                 else
 259                         CERROR("%s: discard slv update, ver:%llu local ver:"
 260                                "%llu\n", qsd->qsd_svname, ver, cur_ver);
 261                 qsd_upd_free(upd);
 262         } else if ((ver == cur_ver + 1) && qqi->qqi_glb_uptodate &&
 263                    qqi->qqi_slv_uptodate) {
 264                 /* In order update, and reintegration has been done. */
 265                 qsd_upd_add(qsd, upd);
 266         } else {
 267                 /* Out of order update (the one with smaller version hasn't
 268                  * reached slave or hasn't been flushed to disk yet), or
 269                  * the reintegration is in progress. Defer the update. */
 270                 struct list_head *list = global ? &qqi->qqi_deferred_glb :
 271                                                   &qqi->qqi_deferred_slv;
 272                 qsd_add_deferred(qsd, list, upd);
 273         }
 274
 275         write_unlock(&qsd->qsd_lock);
 276
 277         EXIT;
 278 }
 279
 280 static int qsd_process_upd(const struct lu_env *env, struct qsd_upd_rec *upd)
 281 {
 282         struct lquota_entry     *lqe = upd->qur_lqe;
 283         struct qsd_qtype_info   *qqi = upd->qur_qqi;
 284         int                      rc;
 285         ENTRY;
 286
 287         if (lqe == NULL) {
 288                 lqe = lqe_locate(env, qqi->qqi_site, &upd->qur_qid);
 289                 if (IS_ERR(lqe))
 290                         GOTO(out, rc = PTR_ERR(lqe));
 291         }
 292
 293         /* The in-memory lqe update for slave index copy isn't deferred,
 294          * we shouldn't touch it here. */
 295         if (upd->qur_global) {
 296                 rc = qsd_update_lqe(env, lqe, upd->qur_global, &upd->qur_rec);
 297                 if (rc)
 298                         GOTO(out, rc);
 299                 /* refresh usage */
 300                 qsd_refresh_usage(env, lqe);
 301                 /* Report usage asynchronously */
 302                 rc = qsd_adjust(env, lqe);
 303                 if (rc)
 304                         LQUOTA_ERROR(lqe, "failed to report usage, rc:%d", rc);
 305         }
 306
 307         rc = qsd_update_index(env, qqi, &upd->qur_qid, upd->qur_global,
 308                               upd->qur_ver, &upd->qur_rec);
 309 out:
 310         if (lqe && !IS_ERR(lqe)) {
 311                 lqe_putref(lqe);
 312                 upd->qur_lqe = NULL;
 313         }
 314         RETURN(rc);
 315 }
 316
 317 void qsd_adjust_schedule(struct lquota_entry *lqe, bool defer, bool cancel)
 318 {
 319         struct qsd_instance     *qsd = lqe2qqi(lqe)->qqi_qsd;
 320         bool                     added = false;
 321
 322         read_lock(&qsd->qsd_lock);
 323         if (qsd->qsd_stopping) {
 324                 read_unlock(&qsd->qsd_lock);
 325                 return;
 326         }
 327         read_unlock(&qsd->qsd_lock);
 328
 329         lqe_getref(lqe);
 330         spin_lock(&qsd->qsd_adjust_lock);
 331
 332         /* the lqe is being queued for the per-ID lock cancel, we should
 333          * cancel the lock cancel and re-add it for quota adjust */
 334         if (!list_empty(&lqe->lqe_link) &&
 335             lqe->lqe_adjust_time == 0) {
 336                 list_del_init(&lqe->lqe_link);
 337                 lqe_putref(lqe);
 338         }
 339
 340         if (list_empty(&lqe->lqe_link)) {
 341                 if (!cancel) {
 342                         lqe->lqe_adjust_time = ktime_get_seconds();
 343                         if (defer)
 344                                 lqe->lqe_adjust_time += QSD_WB_INTERVAL;
 345                 } else {
 346                         lqe->lqe_adjust_time = 0;
 347                 }
 348
 349                 /* lqe reference transferred to list */
 350                 if (defer)
 351                         list_add_tail(&lqe->lqe_link,
 352                                           &qsd->qsd_adjust_list);
 353                 else
 354                         list_add(&lqe->lqe_link, &qsd->qsd_adjust_list);
 355                 added = true;
 356         }
 357         spin_unlock(&qsd->qsd_adjust_lock);
 358
 359         if (added)
 360                 wake_up(&qsd->qsd_upd_thread.t_ctl_waitq);
 361         else
 362                 lqe_putref(lqe);
 363 }
 364
 365 /* return true if there is pending writeback records or the pending
 366  * adjust requests */
 367 static bool qsd_job_pending(struct qsd_instance *qsd, struct list_head *upd,
 368                             bool *uptodate)
 369 {
 370         bool    job_pending = false;
 371         int     qtype;
 372
 373         LASSERT(list_empty(upd));
 374         *uptodate = true;
 375
 376         spin_lock(&qsd->qsd_adjust_lock);
 377         if (!list_empty(&qsd->qsd_adjust_list)) {
 378                 struct lquota_entry *lqe;
 379                 lqe = list_entry(qsd->qsd_adjust_list.next,
 380                                      struct lquota_entry, lqe_link);
 381                 if (ktime_get_seconds() > lqe->lqe_adjust_time)
 382                         job_pending = true;
 383         }
 384         spin_unlock(&qsd->qsd_adjust_lock);
 385
 386         write_lock(&qsd->qsd_lock);
 387         if (!list_empty(&qsd->qsd_upd_list)) {
 388                 list_splice_init(&qsd->qsd_upd_list, upd);
 389                 job_pending = true;
 390         }
 391
 392         for (qtype = USRQUOTA; qtype < LL_MAXQUOTAS; qtype++) {
 393                 struct qsd_qtype_info *qqi = qsd->qsd_type_array[qtype];
 394
 395                 /* don't bother kicking off reintegration if space accounting
 396                  * failed to be enabled */
 397                 if (qqi->qqi_acct_failed)
 398                         continue;
 399
 400                 if (!qsd_type_enabled(qsd, qtype))
 401                         continue;
 402
 403                 if ((!qqi->qqi_glb_uptodate || !qqi->qqi_slv_uptodate) &&
 404                      !qqi->qqi_reint)
 405                         /* global or slave index not up to date and reint
 406                          * thread not running */
 407                         *uptodate = false;
 408         }
 409
 410         write_unlock(&qsd->qsd_lock);
 411         return job_pending;
 412 }
 413
 414 static int qsd_upd_thread(void *arg)
 415 {
 416         struct qsd_instance     *qsd = (struct qsd_instance *)arg;
 417         struct ptlrpc_thread    *thread = &qsd->qsd_upd_thread;
 418         struct l_wait_info       lwi;
 419         struct list_head         queue;
 420         struct qsd_upd_rec      *upd, *n;
 421         struct lu_env           *env;
 422         int                      qtype, rc = 0;
 423         bool                     uptodate;
 424         struct lquota_entry     *lqe;
 425         time64_t cur_time;
 426         ENTRY;
 427
 428         OBD_ALLOC_PTR(env);
 429         if (env == NULL)
 430                 RETURN(-ENOMEM);
 431
 432         rc = lu_env_init(env, LCT_DT_THREAD);
 433         if (rc) {
 434                 CERROR("%s: cannot init env: rc = %d\n", qsd->qsd_svname, rc);
 435                 OBD_FREE_PTR(env);
 436                 RETURN(rc);
 437         }
 438
 439         thread_set_flags(thread, SVC_RUNNING);
 440         wake_up(&thread->t_ctl_waitq);
 441
 442         INIT_LIST_HEAD(&queue);
 443         lwi = LWI_TIMEOUT(cfs_time_seconds(QSD_WB_INTERVAL), NULL, NULL);
 444         while (1) {
 445                 l_wait_event(thread->t_ctl_waitq,
 446                              qsd_job_pending(qsd, &queue, &uptodate) ||
 447                              !thread_is_running(thread), &lwi);
 448
 449                 list_for_each_entry_safe(upd, n, &queue, qur_link) {
 450                         list_del_init(&upd->qur_link);
 451                         qsd_process_upd(env, upd);
 452                         qsd_upd_free(upd);
 453                 }
 454
 455                 spin_lock(&qsd->qsd_adjust_lock);
 456                 cur_time = ktime_get_seconds();
 457                 while (!list_empty(&qsd->qsd_adjust_list)) {
 458                         lqe = list_entry(qsd->qsd_adjust_list.next,
 459                                          struct lquota_entry, lqe_link);
 460                         /* deferred items are sorted by time */
 461                         if (lqe->lqe_adjust_time > cur_time)
 462                                 break;
 463
 464                         list_del_init(&lqe->lqe_link);
 465                         spin_unlock(&qsd->qsd_adjust_lock);
 466
 467                         if (thread_is_running(thread) && uptodate) {
 468                                 qsd_refresh_usage(env, lqe);
 469                                 if (lqe->lqe_adjust_time == 0)
 470                                         qsd_id_lock_cancel(env, lqe);
 471                                 else
 472                                         qsd_adjust(env, lqe);
 473                         }
 474
 475                         lqe_putref(lqe);
 476                         spin_lock(&qsd->qsd_adjust_lock);
 477                 }
 478                 spin_unlock(&qsd->qsd_adjust_lock);
 479
 480                 if (!thread_is_running(thread))
 481                         break;
 482
 483                 if (uptodate)
 484                         continue;
 485
 486                 for (qtype = USRQUOTA; qtype < LL_MAXQUOTAS; qtype++)
 487                         qsd_start_reint_thread(qsd->qsd_type_array[qtype]);
 488         }
 489         lu_env_fini(env);
 490         OBD_FREE_PTR(env);
 491         thread_set_flags(thread, SVC_STOPPED);
 492         wake_up(&thread->t_ctl_waitq);
 493         RETURN(rc);
 494 }
 495
 496 int qsd_start_upd_thread(struct qsd_instance *qsd)
 497 {
 498         struct ptlrpc_thread    *thread = &qsd->qsd_upd_thread;
 499         struct l_wait_info       lwi = { 0 };
 500         struct task_struct              *task;
 501         ENTRY;
 502
 503         task = kthread_run(qsd_upd_thread, (void *)qsd,
 504                            "lquota_wb_%s", qsd->qsd_svname);
 505         if (IS_ERR(task)) {
 506                 CERROR("fail to start quota update thread: rc = %ld\n",
 507                         PTR_ERR(task));
 508                 thread_set_flags(thread, SVC_STOPPED);
 509                 RETURN(PTR_ERR(task));
 510         }
 511
 512         l_wait_event(thread->t_ctl_waitq,
 513                      thread_is_running(thread) || thread_is_stopped(thread),
 514                      &lwi);
 515         RETURN(0);
 516 }
 517
 518 static void qsd_cleanup_deferred(struct qsd_instance *qsd)
 519 {
 520         int     qtype;
 521
 522         for (qtype = USRQUOTA; qtype < LL_MAXQUOTAS; qtype++) {
 523                 struct qsd_upd_rec      *upd, *tmp;
 524                 struct qsd_qtype_info   *qqi = qsd->qsd_type_array[qtype];
 525
 526                 if (qqi == NULL)
 527                         continue;
 528
 529                 write_lock(&qsd->qsd_lock);
 530                 list_for_each_entry_safe(upd, tmp, &qqi->qqi_deferred_glb,
 531                                          qur_link) {
 532                         CWARN("%s: Free global deferred upd: ID:%llu, "
 533                               "ver:%llu/%llu\n", qsd->qsd_svname,
 534                               upd->qur_qid.qid_uid, upd->qur_ver,
 535                               qqi->qqi_glb_ver);
 536                         list_del_init(&upd->qur_link);
 537                         qsd_upd_free(upd);
 538                 }
 539                 list_for_each_entry_safe(upd, tmp, &qqi->qqi_deferred_slv,
 540                                          qur_link) {
 541                         CWARN("%s: Free slave deferred upd: ID:%llu, "
 542                               "ver:%llu/%llu\n", qsd->qsd_svname,
 543                               upd->qur_qid.qid_uid, upd->qur_ver,
 544                               qqi->qqi_slv_ver);
 545                         list_del_init(&upd->qur_link);
 546                         qsd_upd_free(upd);
 547                 }
 548                 write_unlock(&qsd->qsd_lock);
 549         }
 550 }
 551
 552 static void qsd_cleanup_adjust(struct qsd_instance *qsd)
 553 {
 554         struct lquota_entry     *lqe;
 555
 556         spin_lock(&qsd->qsd_adjust_lock);
 557         while (!list_empty(&qsd->qsd_adjust_list)) {
 558                 lqe = list_entry(qsd->qsd_adjust_list.next,
 559                                  struct lquota_entry, lqe_link);
 560                 list_del_init(&lqe->lqe_link);
 561                 lqe_putref(lqe);
 562         }
 563         spin_unlock(&qsd->qsd_adjust_lock);
 564 }
 565
 566 void qsd_stop_upd_thread(struct qsd_instance *qsd)
 567 {
 568         struct ptlrpc_thread    *thread = &qsd->qsd_upd_thread;
 569         struct l_wait_info       lwi    = { 0 };
 570
 571         if (!thread_is_stopped(thread)) {
 572                 thread_set_flags(thread, SVC_STOPPING);
 573                 wake_up(&thread->t_ctl_waitq);
 574
 575                 l_wait_event(thread->t_ctl_waitq, thread_is_stopped(thread),
 576                              &lwi);
 577         }
 578         qsd_cleanup_deferred(qsd);
 579         qsd_cleanup_adjust(qsd);
 580 }