lustre/quota/qsd_handler.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, 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 "qsd_internal.h"
  34
  35 /**
  36  * helper function bumping lqe_pending_req if there is no quota request in
  37  * flight for the lquota entry \a lqe. Otherwise, EBUSY is returned.
  38  */
  39 static inline int qsd_request_enter(struct lquota_entry *lqe)
  40 {
  41         /* is there already a quota request in flight? */
  42         if (lqe->lqe_pending_req != 0) {
  43                 LQUOTA_DEBUG(lqe, "already a request in flight");
  44                 return -EBUSY;
  45         }
  46
  47         if (lqe->lqe_pending_rel != 0) {
  48                 LQUOTA_ERROR(lqe, "no request in flight with pending_rel="LPU64,
  49                              lqe->lqe_pending_rel);
  50                 LBUG();
  51         }
  52
  53         lqe->lqe_pending_req++;
  54         return 0;
  55 }
  56
  57 /**
  58  * Companion of qsd_request_enter() dropping lqe_pending_req to 0.
  59  */
  60 static inline void qsd_request_exit(struct lquota_entry *lqe)
  61 {
  62         if (lqe->lqe_pending_req != 1) {
  63                 LQUOTA_ERROR(lqe, "lqe_pending_req != 1!!!");
  64                 LBUG();
  65         }
  66         lqe->lqe_pending_req--;
  67         lqe->lqe_pending_rel = 0;
  68         cfs_waitq_broadcast(&lqe->lqe_waiters);
  69 }
  70
  71 /**
  72  * Check whether a qsd instance is all set to send quota request to master.
  73  * This includes checking whether:
  74  * - the connection to master is set up and usable,
  75  * - the qsd isn't stopping
  76  * - reintegration has been successfully completed and all indexes are
  77  *   up-to-date
  78  *
  79  * \param lqe - is the lquota entry for which we would like to send an quota
  80  *              request
  81  * \param lockh - is the remote handle of the global lock returned on success
  82  *
  83  * \retval 0 on success, appropriate error on failure
  84  */
  85 static int qsd_ready(struct lquota_entry *lqe, struct lustre_handle *lockh)
  86 {
  87         struct qsd_qtype_info   *qqi = lqe2qqi(lqe);
  88         struct qsd_instance     *qsd = qqi->qqi_qsd;
  89         struct obd_import       *imp = NULL;
  90         struct ldlm_lock        *lock;
  91         ENTRY;
  92
  93         read_lock(&qsd->qsd_lock);
  94         /* is the qsd about to shut down? */
  95         if (qsd->qsd_stopping) {
  96                 read_unlock(&qsd->qsd_lock);
  97                 LQUOTA_DEBUG(lqe, "dropping quota req since qsd is stopping");
  98                 /* Target is about to shut down, client will retry */
  99                 RETURN(-EINPROGRESS);
 100         }
 101
 102         /* is the connection to the quota master ready? */
 103         if (qsd->qsd_exp_valid)
 104                 imp = class_exp2cliimp(qsd->qsd_exp);
 105         if (imp == NULL || imp->imp_invalid) {
 106                 read_unlock(&qsd->qsd_lock);
 107                 LQUOTA_DEBUG(lqe, "connection to master not ready");
 108                 RETURN(-ENOTCONN);
 109         }
 110
 111         /* In most case, reintegration must have been triggered (when enable
 112          * quota or on OST start), however, in rare race condition (enabling
 113          * quota when starting OSTs), we might miss triggering reintegration
 114          * for some qqi.
 115          *
 116          * If the previous reintegration failed for some reason, we'll
 117          * re-trigger it here as well. */
 118         if (!qqi->qqi_glb_uptodate || !qqi->qqi_slv_uptodate) {
 119                 read_unlock(&qsd->qsd_lock);
 120                 LQUOTA_DEBUG(lqe, "not up-to-date, dropping request and "
 121                              "kicking off reintegration");
 122                 qsd_start_reint_thread(qqi);
 123                 RETURN(-EINPROGRESS);
 124         }
 125
 126         /* Fill the remote global lock handle, master will check this handle
 127          * to see if the slave is sending request with stale lock */
 128         lustre_handle_copy(lockh, &qqi->qqi_lockh);
 129         read_unlock(&qsd->qsd_lock);
 130
 131         if (!lustre_handle_is_used(lockh))
 132                 RETURN(-ENOLCK);
 133
 134         lock = ldlm_handle2lock(lockh);
 135         if (lock == NULL)
 136                 RETURN(-ENOLCK);
 137
 138         /* return remote lock handle to be packed in quota request */
 139         lustre_handle_copy(lockh, &lock->l_remote_handle);
 140         LDLM_LOCK_PUT(lock);
 141
 142         RETURN(0);
 143 }
 144
 145 /**
 146  * Check whether any quota space adjustment (pre-acquire/release/report) is
 147  * needed for a given quota ID. If a non-null \a qbody is passed, then the
 148  * \a qbody structure (qb_count/flags/usage) is filled with appropriate data
 149  * to be packed in the quota request.
 150  *
 151  * \param lqe   - is the lquota entry for which we would like to adjust quota
 152  *                space.
 153  * \param qbody - is the quota body to fill, if not NULL.
 154  *
 155  * \retval true  - space adjustment is required and \a qbody is filled, if not
 156  *                 NULL
 157  * \retval false - no space adjustment required
 158  */
 159 static bool qsd_calc_adjust(struct lquota_entry *lqe, struct quota_body *qbody)
 160 {
 161         __u64   usage, granted;
 162         ENTRY;
 163
 164         usage   = lqe->lqe_usage;
 165         usage  += lqe->lqe_pending_write + lqe->lqe_waiting_write;
 166         granted = lqe->lqe_granted;
 167
 168         if (qbody != NULL)
 169                 qbody->qb_flags = 0;
 170
 171         if (!lqe->lqe_enforced) {
 172                 /* quota not enforced any more for this ID */
 173                 if (granted != 0) {
 174                         /* release all quota space unconditionally */
 175                         LQUOTA_DEBUG(lqe, "not enforced, releasing all space");
 176                         if (qbody != NULL) {
 177                                 qbody->qb_count = granted;
 178                                 qbody->qb_flags = QUOTA_DQACQ_FL_REL;
 179                         }
 180                         RETURN(true);
 181                 }
 182                 RETURN(false);
 183         }
 184
 185         if (!lustre_handle_is_used(&lqe->lqe_lockh)) {
 186                 /* No valid per-ID lock
 187                  * When reporting quota (during reintegration or on setquota
 188                  * glimpse), we should release granted space if usage is 0.
 189                  * Otherwise, if the usage is less than granted, we need to
 190                  * acquire the per-ID lock to make sure the unused grant can be
 191                  * reclaimed by per-ID lock glimpse. */
 192                 if (usage == 0) {
 193                         /* no on-disk usage and no outstanding activity, release
 194                          * space */
 195                         if (granted != 0) {
 196                                 LQUOTA_DEBUG(lqe, "no usage, releasing all "
 197                                              "space");
 198                                 if (qbody != NULL) {
 199                                         qbody->qb_count = granted;
 200                                         qbody->qb_flags = QUOTA_DQACQ_FL_REL;
 201                                 }
 202                                 RETURN(true);
 203                         }
 204                         LQUOTA_DEBUG(lqe, "no usage + no granted, nothing to "
 205                                      "do");
 206                         RETURN(false);
 207                 }
 208
 209                 if (lqe->lqe_usage < lqe->lqe_granted) {
 210                         /* holding quota space w/o any lock, enqueue per-ID lock
 211                          * again */
 212                         LQUOTA_DEBUG(lqe, "(re)acquiring per-ID lock");
 213                         if (qbody != NULL) {
 214                                 qbody->qb_count = 0;
 215                                 qbody->qb_flags = QUOTA_DQACQ_FL_ACQ;
 216                         }
 217                         RETURN(true);
 218                 }
 219
 220                 if (lqe->lqe_usage > lqe->lqe_granted) {
 221                         /* quota overrun, report usage */
 222                         LQUOTA_DEBUG(lqe, "overrun, reporting usage");
 223                         if (qbody != NULL) {
 224                                 qbody->qb_usage = lqe->lqe_usage;
 225                                 qbody->qb_flags = QUOTA_DQACQ_FL_REPORT;
 226                         }
 227                         RETURN(true);
 228                 }
 229                 LQUOTA_DEBUG(lqe, "granted matches usage, nothing to do");
 230                 RETURN(false);
 231         }
 232
 233         /* valid per-ID lock
 234          * Apply good old quota qunit adjustment logic which has been around
 235          * since lustre 1.4:
 236          * 1. release spare quota space? */
 237         if (granted > usage + lqe->lqe_qunit) {
 238                 /* pre-release quota space */
 239                 if (qbody == NULL)
 240                         RETURN(true);
 241                 qbody->qb_count = granted - usage;
 242                 /* if usage == 0, release all granted space */
 243                 if (usage) {
 244                         /* try to keep one qunit of quota space */
 245                         qbody->qb_count -= lqe->lqe_qunit;
 246                         /* but don't release less than qtune to avoid releasing
 247                          * space too often */
 248                         if (qbody->qb_count < lqe->lqe_qtune)
 249                                 qbody->qb_count = lqe->lqe_qtune;
 250                 }
 251                 qbody->qb_flags = QUOTA_DQACQ_FL_REL;
 252                 RETURN(true);
 253         }
 254
 255         /* 2. Any quota overrun? */
 256         if (lqe->lqe_usage > lqe->lqe_granted) {
 257                 /* we overconsumed quota space, we report usage in request so
 258                  * that master can adjust it unconditionally */
 259                 if (qbody == NULL)
 260                         RETURN(true);
 261                 qbody->qb_usage = lqe->lqe_usage;
 262                 granted         = lqe->lqe_usage;
 263                 qbody->qb_flags = QUOTA_DQACQ_FL_REPORT;
 264         }
 265
 266         /* 3. Time to pre-acquire? */
 267         if (!lqe->lqe_edquot && !lqe->lqe_nopreacq && usage > 0 &&
 268             lqe->lqe_qunit != 0 && granted < usage + lqe->lqe_qtune) {
 269                 /* To pre-acquire quota space, we report how much spare quota
 270                  * space the slave currently owns, then the master will grant us
 271                  * back how much we can pretend given the current state of
 272                  * affairs */
 273                 if (qbody == NULL)
 274                         RETURN(true);
 275                 if (granted <= usage)
 276                         qbody->qb_count = 0;
 277                 else
 278                         qbody->qb_count = granted - usage;
 279                 qbody->qb_flags |= QUOTA_DQACQ_FL_PREACQ;
 280                 RETURN(true);
 281         }
 282
 283         if (qbody != NULL)
 284                 RETURN(qbody->qb_flags != 0);
 285         else
 286                 RETURN(false);
 287 }
 288
 289 /**
 290  * Helper function returning true when quota space need to be adjusted (some
 291  * unused space should be free or pre-acquire) and false otherwise.
 292  */
 293 static inline bool qsd_adjust_needed(struct lquota_entry *lqe)
 294 {
 295         return qsd_calc_adjust(lqe, NULL);
 296 }
 297
 298 /**
 299  * Callback function called when an acquire/release request sent to the master
 300  * is completed
 301  */
 302 static void qsd_req_completion(const struct lu_env *env,
 303                                struct qsd_qtype_info *qqi,
 304                                struct quota_body *reqbody,
 305                                struct quota_body *repbody,
 306                                struct lustre_handle *lockh,
 307                                struct lquota_lvb *lvb,
 308                                void *arg, int ret)
 309 {
 310         struct lquota_entry     *lqe = (struct lquota_entry *)arg;
 311         struct qsd_thread_info  *qti;
 312         int                      rc;
 313         bool                     adjust = false, cancel = false;
 314         ENTRY;
 315
 316         LASSERT(qqi != NULL && lqe != NULL);
 317
 318         /* environment passed by ptlrpcd is mostly used by CLIO and hasn't the
 319          * DT tags set. */
 320         rc = lu_env_refill_by_tags((struct lu_env *)env, LCT_DT_THREAD, 0);
 321         if (rc) {
 322                 LQUOTA_ERROR(lqe, "failed to refill environmnent %d", rc);
 323                 lqe_write_lock(lqe);
 324                 /* can't afford to adjust quota space with no suitable lu_env */
 325                 GOTO(out_noadjust, rc);
 326         }
 327         qti = qsd_info(env);
 328
 329         lqe_write_lock(lqe);
 330         LQUOTA_DEBUG(lqe, "DQACQ returned %d, flags:0x%x", ret,
 331                      reqbody->qb_flags);
 332
 333         /* despite -EDQUOT & -EINPROGRESS errors, the master might still
 334          * grant us back quota space to adjust quota overrun */
 335         if (ret != 0 && ret != -EDQUOT && ret != -EINPROGRESS) {
 336                 if (ret != -ETIMEDOUT && ret != -ENOTCONN &&
 337                    ret != -ESHUTDOWN && ret != -EAGAIN)
 338                         /* print errors only if return code is unexpected */
 339                         LQUOTA_ERROR(lqe, "DQACQ failed with %d, flags:0x%x",
 340                                      ret, reqbody->qb_flags);
 341                 GOTO(out, ret);
 342         }
 343
 344         /* Set the lqe_lockh */
 345         if (lustre_handle_is_used(lockh) &&
 346             !lustre_handle_equal(lockh, &lqe->lqe_lockh))
 347                 lustre_handle_copy(&lqe->lqe_lockh, lockh);
 348
 349         /* If the replied qb_count is zero, it means master didn't process
 350          * the DQACQ since the limit for this ID has been removed, so we
 351          * should not update quota entry & slave index copy neither. */
 352         if (repbody != NULL && repbody->qb_count != 0) {
 353                 LQUOTA_DEBUG(lqe, "DQACQ qb_count:"LPU64, repbody->qb_count);
 354
 355                 if (req_is_rel(reqbody->qb_flags)) {
 356                         if (lqe->lqe_granted < repbody->qb_count) {
 357                                 LQUOTA_ERROR(lqe, "can't release more space "
 358                                              "than owned "LPU64"<"LPU64,
 359                                              lqe->lqe_granted,
 360                                              repbody->qb_count);
 361                                 lqe->lqe_granted = 0;
 362                         } else {
 363                                 lqe->lqe_granted -= repbody->qb_count;
 364                         }
 365                         /* Cancel the per-ID lock initiatively when there
 366                          * isn't any usage & grant, which can avoid master
 367                          * sending glimpse unnecessarily to this slave on
 368                          * quota revoking */
 369                         if (!lqe->lqe_pending_write && !lqe->lqe_granted &&
 370                             !lqe->lqe_waiting_write && !lqe->lqe_usage)
 371                                 cancel = true;
 372                 } else {
 373                         lqe->lqe_granted += repbody->qb_count;
 374                 }
 375                 qti->qti_rec.lqr_slv_rec.qsr_granted = lqe->lqe_granted;
 376                 lqe_write_unlock(lqe);
 377
 378                 /* Update the slave index file in the dedicated thread. So far,
 379                  * We don't update the version of slave index copy on DQACQ.
 380                  * No locking is necessary since nobody can change
 381                  * lqe->lqe_granted while lqe->lqe_pending_req > 0 */
 382                 qsd_upd_schedule(qqi, lqe, &lqe->lqe_id, &qti->qti_rec, 0,
 383                                  false);
 384                 lqe_write_lock(lqe);
 385         }
 386
 387         /* extract information from lvb */
 388         if (ret == 0 && lvb != 0) {
 389                 if (lvb->lvb_id_qunit != 0)
 390                         qsd_set_qunit(lqe, lvb->lvb_id_qunit);
 391                 if (lvb->lvb_flags & LQUOTA_FL_EDQUOT)
 392                         lqe->lqe_edquot = true;
 393                 else
 394                         lqe->lqe_edquot = false;
 395         } else if (repbody != NULL && repbody->qb_qunit != 0) {
 396                 qsd_set_qunit(lqe, repbody->qb_qunit);
 397         }
 398
 399         /* turn off pre-acquire if it failed with -EDQUOT. This is done to avoid
 400          * flooding the master with acquire request. Pre-acquire will be turned
 401          * on again as soon as qunit is modified */
 402         if (req_is_preacq(reqbody->qb_flags) && ret == -EDQUOT)
 403                 lqe->lqe_nopreacq = true;
 404 out:
 405         adjust = qsd_adjust_needed(lqe);
 406         if (reqbody && req_is_acq(reqbody->qb_flags) && ret != -EDQUOT) {
 407                 lqe->lqe_acq_rc = ret;
 408                 lqe->lqe_acq_time = cfs_time_current_64();
 409         }
 410 out_noadjust:
 411         qsd_request_exit(lqe);
 412         lqe_write_unlock(lqe);
 413
 414         /* release reference on per-ID lock */
 415         if (lustre_handle_is_used(lockh))
 416                 ldlm_lock_decref(lockh, qsd_id_einfo.ei_mode);
 417
 418         if (cancel) {
 419                 qsd_adjust_schedule(lqe, false, true);
 420         } else if (adjust) {
 421                 if (!ret || ret == -EDQUOT)
 422                         qsd_adjust_schedule(lqe, false, false);
 423                 else
 424                         qsd_adjust_schedule(lqe, true, false);
 425         }
 426         lqe_putref(lqe);
 427
 428         if (lvb)
 429                 OBD_FREE_PTR(lvb);
 430         EXIT;
 431 }
 432
 433 /**
 434  * Try to consume local quota space.
 435  *
 436  * \param lqe   - is the qid entry to be processed
 437  * \param space - is the amount of quota space needed to complete the operation
 438  *
 439  * \retval 0       - success
 440  * \retval -EDQUOT - out of quota
 441  * \retval -EAGAIN - need to acquire space from master
 442  */
 443 static int qsd_acquire_local(struct lquota_entry *lqe, __u64 space)
 444 {
 445         __u64   usage;
 446         int     rc;
 447         ENTRY;
 448
 449         if (!lqe->lqe_enforced)
 450                 /* not enforced any more, we are good */
 451                 RETURN(-ESRCH);
 452
 453         lqe_write_lock(lqe);
 454         /* use latest usage */
 455         usage = lqe->lqe_usage;
 456         /* take pending write into account */
 457         usage += lqe->lqe_pending_write;
 458
 459         if (space + usage <= lqe->lqe_granted - lqe->lqe_pending_rel) {
 460                 /* Yay! we got enough space */
 461                 lqe->lqe_pending_write += space;
 462                 lqe->lqe_waiting_write -= space;
 463                 rc = 0;
 464         } else if (lqe->lqe_edquot) {
 465                 rc = -EDQUOT;
 466         } else {
 467                 rc = -EAGAIN;
 468         }
 469         lqe_write_unlock(lqe);
 470
 471         RETURN(rc);
 472 }
 473
 474 /**
 475  * Compute how much quota space should be acquire from the master based
 476  * on how much is currently granted to this slave and pending/waiting
 477  * operations.
 478  *
 479  * \param lqe - is the lquota entry for which we would like to adjust quota
 480  *              space.
 481  * \param qbody - is the quota body of the acquire request to fill
 482  *
 483  * \retval true  - space acquisition is needed and qbody is filled
 484  * \retval false - no space acquisition required
 485  */
 486 static inline bool qsd_calc_acquire(struct lquota_entry *lqe,
 487                                     struct quota_body *qbody)
 488 {
 489         __u64   usage, granted;
 490
 491         usage   = lqe->lqe_usage;
 492         usage  += lqe->lqe_pending_write + lqe->lqe_waiting_write;
 493         granted = lqe->lqe_granted;
 494
 495         qbody->qb_flags = 0;
 496
 497         /* if we overconsumed quota space, we report usage in request so that
 498          * master can adjust it unconditionally */
 499         if (lqe->lqe_usage > lqe->lqe_granted) {
 500                 qbody->qb_usage = lqe->lqe_usage;
 501                 qbody->qb_flags = QUOTA_DQACQ_FL_REPORT;
 502                 granted = lqe->lqe_usage;
 503         }
 504
 505         /* acquire as much as needed, but not more */
 506         if (usage > granted) {
 507                 qbody->qb_count  = usage - granted;
 508                 qbody->qb_flags |= QUOTA_DQACQ_FL_ACQ;
 509         }
 510
 511         return qbody->qb_flags != 0;
 512 }
 513
 514 /**
 515  * Acquire quota space from master.
 516  * There are at most 1 in-flight dqacq/dqrel.
 517  *
 518  * \param env    - the environment passed by the caller
 519  * \param lqe    - is the qid entry to be processed
 520  *
 521  * \retval 0            - success
 522  * \retval -EDQUOT      - out of quota
 523  * \retval -EINPROGRESS - inform client to retry write/create
 524  * \retval -EBUSY       - already a quota request in flight
 525  * \retval -ve          - other appropriate errors
 526  */
 527 static int qsd_acquire_remote(const struct lu_env *env,
 528                               struct lquota_entry *lqe)
 529 {
 530         struct qsd_thread_info  *qti = qsd_info(env);
 531         struct quota_body       *qbody = &qti->qti_body;
 532         struct qsd_instance     *qsd;
 533         struct qsd_qtype_info   *qqi;
 534         int                      rc;
 535         ENTRY;
 536
 537         memset(qbody, 0, sizeof(*qbody));
 538         rc = qsd_ready(lqe, &qbody->qb_glb_lockh);
 539         if (rc)
 540                 RETURN(rc);
 541
 542         qqi = lqe2qqi(lqe);
 543         qsd = qqi->qqi_qsd;
 544
 545         lqe_write_lock(lqe);
 546
 547         /* is quota really enforced for this id? */
 548         if (!lqe->lqe_enforced) {
 549                 lqe_write_unlock(lqe);
 550                 LQUOTA_DEBUG(lqe, "quota not enforced any more");
 551                 RETURN(0);
 552         }
 553
 554         /* fill qb_count & qb_flags */
 555         if (!qsd_calc_acquire(lqe, qbody)) {
 556                 lqe_write_unlock(lqe);
 557                 LQUOTA_DEBUG(lqe, "No acquire required");
 558                 RETURN(0);
 559         }
 560
 561         /* check whether an acquire request completed recently */
 562         if (lqe->lqe_acq_rc != 0 &&
 563             cfs_time_before_64(cfs_time_shift_64(-1), lqe->lqe_acq_time)) {
 564                 lqe_write_unlock(lqe);
 565                 LQUOTA_DEBUG(lqe, "using cached return code %d", lqe->lqe_acq_rc);
 566                 RETURN(lqe->lqe_acq_rc);
 567         }
 568
 569         /* only 1 quota request in flight for a given ID is allowed */
 570         rc = qsd_request_enter(lqe);
 571         if (rc) {
 572                 lqe_write_unlock(lqe);
 573                 RETURN(rc);
 574         }
 575
 576         lustre_handle_copy(&qti->qti_lockh, &lqe->lqe_lockh);
 577         lqe_write_unlock(lqe);
 578
 579         /* hold a refcount until completion */
 580         lqe_getref(lqe);
 581
 582         /* fill other quota body fields */
 583         qbody->qb_fid = qqi->qqi_fid;
 584         qbody->qb_id  = lqe->lqe_id;
 585
 586         /* check whether we already own a valid lock for this ID */
 587         rc = qsd_id_lock_match(&qti->qti_lockh, &qbody->qb_lockh);
 588         if (rc) {
 589                 struct lquota_lvb *lvb;
 590
 591                 OBD_ALLOC_PTR(lvb);
 592                 if (lvb == NULL) {
 593                         rc = -ENOMEM;
 594                         qsd_req_completion(env, qqi, qbody, NULL,
 595                                            &qti->qti_lockh, NULL, lqe, rc);
 596                         RETURN(rc);
 597                 }
 598                 /* no lock found, should use intent */
 599                 rc = qsd_intent_lock(env, qsd->qsd_exp, qbody, true,
 600                                      IT_QUOTA_DQACQ, qsd_req_completion,
 601                                      qqi, lvb, (void *)lqe);
 602         } else {
 603                 /* lock found, should use regular dqacq */
 604                 rc = qsd_send_dqacq(env, qsd->qsd_exp, qbody, true,
 605                                     qsd_req_completion, qqi, &qti->qti_lockh,
 606                                     lqe);
 607         }
 608
 609         /* the completion function will be called by qsd_send_dqacq or
 610          * qsd_intent_lock */
 611         RETURN(rc);
 612 }
 613
 614 /**
 615  * Acquire \a space of quota space in order to complete an operation.
 616  * Try to consume local quota space first and send acquire request to quota
 617  * master if required.
 618  *
 619  * \param env   - the environment passed by the caller
 620  * \param lqe   - is the qid entry to be processed
 621  * \param space - is the amount of quota required for the operation
 622  * \param ret   - is the return code (-EDQUOT, -EINPROGRESS, ...)
 623  *
 624  * \retval true  - exit from l_wait_event and real return value in \a ret
 625  * \retval false - continue waiting
 626  */
 627 static bool qsd_acquire(const struct lu_env *env, struct lquota_entry *lqe,
 628                         long long space, int *ret)
 629 {
 630         int rc = 0, count;
 631         ENTRY;
 632
 633         for (count = 0; rc == 0; count++) {
 634                 LQUOTA_DEBUG(lqe, "acquiring:"LPD64 " count=%d", space, count);
 635
 636                 if (lqe2qqi(lqe)->qqi_qsd->qsd_stopping) {
 637                         rc = -EINPROGRESS;
 638                         break;
 639                 }
 640
 641                 /* refresh disk usage */
 642                 rc = qsd_refresh_usage(env, lqe);
 643                 if (rc)
 644                         break;
 645
 646                 /* try to consume local quota space first */
 647                 rc = qsd_acquire_local(lqe, space);
 648                 if (rc != -EAGAIN)
 649                         /* rc == 0, Wouhou! enough local quota space
 650                          * rc < 0, something bad happened */
 651                          break;
 652
 653                 /* need to acquire more quota space from master */
 654                 rc = qsd_acquire_remote(env, lqe);
 655         }
 656
 657         if (rc == -EBUSY)
 658                 /* already a request in flight, continue waiting */
 659                 RETURN(false);
 660         *ret = rc;
 661         RETURN(true); /* exit from l_wait_event */
 662 }
 663
 664 /**
 665  * Quota enforcement handler. If local quota can satisfy this operation,
 666  * return success, otherwise, acquire more quota from master.
 667  * (for write operation, if master isn't available at this moment, return
 668  * -EINPROGRESS to inform client to retry the write)
 669  *
 670  * \param env   - the environment passed by the caller
 671  * \param qsd   - is the qsd instance associated with the device in charge
 672  *                of the operation.
 673  * \param qid   - is the qid information attached in the transaction handle
 674  * \param space - is the space required by the operation
 675  * \param flags - if the operation is write, return caller no user/group
 676  *                and sync commit flags
 677  *
 678  * \retval 0            - success
 679  * \retval -EDQUOT      - out of quota
 680  * \retval -EINPROGRESS - inform client to retry write
 681  * \retval -ve          - other appropriate errors
 682  */
 683 static int qsd_op_begin0(const struct lu_env *env, struct qsd_qtype_info *qqi,
 684                          struct lquota_id_info *qid, long long space,
 685                          int *flags)
 686 {
 687         struct lquota_entry     *lqe;
 688         int                      rc, ret = -EINPROGRESS;
 689         struct l_wait_info       lwi;
 690         ENTRY;
 691
 692         if (qid->lqi_qentry != NULL) {
 693                 /* we already had to deal with this id for this transaction */
 694                 lqe = qid->lqi_qentry;
 695                 if (!lqe->lqe_enforced)
 696                         RETURN(0);
 697         } else {
 698                 /* look up lquota entry associated with qid */
 699                 lqe = lqe_locate(env, qqi->qqi_site, &qid->lqi_id);
 700                 if (IS_ERR(lqe))
 701                         RETURN(PTR_ERR(lqe));
 702                 if (!lqe->lqe_enforced) {
 703                         lqe_putref(lqe);
 704                         RETURN(0);
 705                 }
 706                 qid->lqi_qentry = lqe;
 707                 /* lqe will be released in qsd_op_end() */
 708         }
 709
 710         if (space <= 0) {
 711                 /* when space is negative or null, we don't need to consume
 712                  * quota space. That said, we still want to perform space
 713                  * adjustments in qsd_op_end, so we return here, but with
 714                  * a reference on the lqe */
 715                 if (flags != NULL) {
 716                         rc = qsd_refresh_usage(env, lqe);
 717                         GOTO(out_flags, rc);
 718                 }
 719                 RETURN(0);
 720         }
 721
 722         LQUOTA_DEBUG(lqe, "op_begin space:"LPD64, space);
 723
 724         lqe_write_lock(lqe);
 725         lqe->lqe_waiting_write += space;
 726         lqe_write_unlock(lqe);
 727
 728         /* acquire quota space for the operation, cap overall wait time to
 729          * prevent a service thread from being stuck for too long */
 730         lwi = LWI_TIMEOUT(cfs_time_seconds(qsd_wait_timeout(qqi->qqi_qsd)),
 731                           NULL, NULL);
 732         rc = l_wait_event(lqe->lqe_waiters, qsd_acquire(env, lqe, space, &ret),
 733                           &lwi);
 734
 735         if (rc == 0 && ret == 0) {
 736                 qid->lqi_space += space;
 737         } else {
 738                 if (rc == 0)
 739                         rc = ret;
 740
 741                 LQUOTA_DEBUG(lqe, "acquire quota failed:%d", rc);
 742
 743                 lqe_write_lock(lqe);
 744                 lqe->lqe_waiting_write -= space;
 745
 746                 if (flags && lqe->lqe_pending_write != 0)
 747                         /* Inform OSD layer that there are pending writes.
 748                          * It might want to retry after a sync if appropriate */
 749                          *flags |= QUOTA_FL_SYNC;
 750                 lqe_write_unlock(lqe);
 751
 752                 /* convert recoverable error into -EINPROGRESS, client will
 753                  * retry */
 754                 if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ENOLCK ||
 755                     rc == -EAGAIN || rc == -EINTR) {
 756                         rc = -EINPROGRESS;
 757                 } else if (rc == -ESRCH) {
 758                         rc = 0;
 759                         LQUOTA_ERROR(lqe, "ID isn't enforced on master, it "
 760                                      "probably due to a legeal race, if this "
 761                                      "message is showing up constantly, there "
 762                                      "could be some inconsistence between "
 763                                      "master & slave, and quota reintegration "
 764                                      "needs be re-triggered.");
 765                 }
 766         }
 767
 768         if (flags != NULL) {
 769 out_flags:
 770                 LASSERT(qid->lqi_is_blk);
 771                 if (rc != 0) {
 772                         *flags |= LQUOTA_OVER_FL(qqi->qqi_qtype);
 773                 } else {
 774                         __u64   usage;
 775
 776                         lqe_read_lock(lqe);
 777                         usage  = lqe->lqe_usage;
 778                         usage += lqe->lqe_pending_write;
 779                         usage += lqe->lqe_waiting_write;
 780                         usage += qqi->qqi_qsd->qsd_sync_threshold;
 781
 782                         /* if we should notify client to start sync write */
 783                         if (usage >= lqe->lqe_granted - lqe->lqe_pending_rel)
 784                                 *flags |= LQUOTA_OVER_FL(qqi->qqi_qtype);
 785                         else
 786                                 *flags &= ~LQUOTA_OVER_FL(qqi->qqi_qtype);
 787                         lqe_read_unlock(lqe);
 788                 }
 789         }
 790         RETURN(rc);
 791 }
 792
 793 /**
 794  * helper function comparing two lquota_id_info structures
 795  */
 796 static inline bool qid_equal(struct lquota_id_info *q1,
 797                              struct lquota_id_info *q2)
 798 {
 799         if (q1->lqi_type != q2->lqi_type)
 800                 return false;
 801         return (q1->lqi_id.qid_uid == q2->lqi_id.qid_uid) ? true : false;
 802 }
 803
 804 /**
 805  * Enforce quota, it's called in the declaration of each operation.
 806  * qsd_op_end() will then be called later once all the operations have been
 807  * completed in order to release/adjust the quota space.
 808  *
 809  * \param env   - the environment passed by the caller
 810  * \param qsd   - is the qsd instance associated with the device in charge of
 811  *                the operation.
 812  * \param trans - is the quota transaction information
 813  * \param qi    - qid & space required by current operation
 814  * \param flags - if the operation is write, return caller no user/group and
 815  *                sync commit flags
 816  *
 817  * \retval 0            - success
 818  * \retval -EDQUOT      - out of quota
 819  * \retval -EINPROGRESS - inform client to retry write
 820  * \retval -ve          - other appropriate errors
 821  */
 822 int qsd_op_begin(const struct lu_env *env, struct qsd_instance *qsd,
 823                  struct lquota_trans *trans, struct lquota_id_info *qi,
 824                  int *flags)
 825 {
 826         int     i, rc;
 827         bool    found = false;
 828         ENTRY;
 829
 830         if (unlikely(qsd == NULL))
 831                 RETURN(0);
 832
 833         /* We don't enforce quota until the qsd_instance is started */
 834         read_lock(&qsd->qsd_lock);
 835         if (!qsd->qsd_started) {
 836                 read_unlock(&qsd->qsd_lock);
 837                 RETURN(0);
 838         }
 839         read_unlock(&qsd->qsd_lock);
 840
 841         /* ignore block quota on MDTs, ignore inode quota on OSTs */
 842         if ((!qsd->qsd_is_md && !qi->lqi_is_blk) ||
 843             (qsd->qsd_is_md && qi->lqi_is_blk))
 844                 RETURN(0);
 845
 846         /* ignore quota enforcement request when:
 847          *    - quota isn't enforced for this quota type
 848          * or - the user/group is root
 849          * or - quota accounting isn't enabled */
 850         if (!qsd_type_enabled(qsd, qi->lqi_type) || qi->lqi_id.qid_uid == 0 ||
 851             qsd->qsd_acct_failed)
 852                 RETURN(0);
 853
 854         LASSERTF(trans->lqt_id_cnt <= QUOTA_MAX_TRANSIDS, "id_cnt=%d",
 855                  trans->lqt_id_cnt);
 856         /* check whether we already allocated a slot for this id */
 857         for (i = 0; i < trans->lqt_id_cnt; i++) {
 858                 if (qid_equal(qi, &trans->lqt_ids[i])) {
 859                         found = true;
 860                         /* make sure we are not mixing inodes & blocks */
 861                         LASSERT(trans->lqt_ids[i].lqi_is_blk == qi->lqi_is_blk);
 862                         break;
 863                 }
 864         }
 865
 866         if (!found) {
 867                 if (unlikely(i >= QUOTA_MAX_TRANSIDS)) {
 868                         CERROR("%s: more than %d qids enforced for a "
 869                                "transaction?\n", qsd->qsd_svname, i);
 870                         RETURN(-EINVAL);
 871                 }
 872
 873                 /* fill new slot */
 874                 trans->lqt_ids[i].lqi_id     = qi->lqi_id;
 875                 trans->lqt_ids[i].lqi_type   = qi->lqi_type;
 876                 trans->lqt_ids[i].lqi_is_blk = qi->lqi_is_blk;
 877                 trans->lqt_id_cnt++;
 878         }
 879
 880         /* manage quota enforcement for this ID */
 881         rc = qsd_op_begin0(env, qsd->qsd_type_array[qi->lqi_type],
 882                            &trans->lqt_ids[i], qi->lqi_space, flags);
 883         RETURN(rc);
 884 }
 885 EXPORT_SYMBOL(qsd_op_begin);
 886
 887 /**
 888  * Adjust quota space (by acquiring or releasing) hold by the quota slave.
 889  * This function is called after each quota request completion and during
 890  * reintegration in order to report usage or re-acquire quota locks.
 891  * Space adjustment is aborted if there is already a quota request in flight
 892  * for this ID.
 893  *
 894  * \param env    - the environment passed by the caller
 895  * \param lqe    - is the qid entry to be processed
 896  *
 897  * \retval 0 on success, appropriate errors on failure
 898  */
 899 int qsd_adjust(const struct lu_env *env, struct lquota_entry *lqe)
 900 {
 901         struct qsd_thread_info  *qti = qsd_info(env);
 902         struct quota_body       *qbody = &qti->qti_body;
 903         struct qsd_instance     *qsd;
 904         struct qsd_qtype_info   *qqi;
 905         int                      rc;
 906         bool                     intent = false;
 907         ENTRY;
 908
 909         memset(qbody, 0, sizeof(*qbody));
 910         rc = qsd_ready(lqe, &qbody->qb_glb_lockh);
 911         if (rc) {
 912                 /* add to adjust list again to trigger adjustment later when
 913                  * slave is ready */
 914                 LQUOTA_DEBUG(lqe, "delaying adjustment since qsd isn't ready");
 915                 qsd_adjust_schedule(lqe, true, false);
 916                 RETURN(0);
 917         }
 918
 919         qqi = lqe2qqi(lqe);
 920         qsd = qqi->qqi_qsd;
 921
 922         lqe_write_lock(lqe);
 923
 924         /* fill qb_count & qb_flags */
 925         if (!qsd_calc_adjust(lqe, qbody)) {
 926                 lqe_write_unlock(lqe);
 927                 LQUOTA_DEBUG(lqe, "no adjustment required");
 928                 RETURN(0);
 929         }
 930
 931         /* only 1 quota request in flight for a given ID is allowed */
 932         rc = qsd_request_enter(lqe);
 933         if (rc) {
 934                 /* already a request in flight, space adjustment will be run
 935                  * again on request completion */
 936                 lqe_write_unlock(lqe);
 937                 RETURN(0);
 938         }
 939
 940         if (req_is_rel(qbody->qb_flags))
 941                 lqe->lqe_pending_rel = qbody->qb_count;
 942         lustre_handle_copy(&qti->qti_lockh, &lqe->lqe_lockh);
 943         lqe_write_unlock(lqe);
 944
 945         /* hold a refcount until completion */
 946         lqe_getref(lqe);
 947
 948         /* fill other quota body fields */
 949         qbody->qb_fid = qqi->qqi_fid;
 950         qbody->qb_id  = lqe->lqe_id;
 951
 952         if (req_is_acq(qbody->qb_flags) || req_is_preacq(qbody->qb_flags)) {
 953                 /* check whether we own a valid lock for this ID */
 954                 rc = qsd_id_lock_match(&qti->qti_lockh, &qbody->qb_lockh);
 955                 if (rc) {
 956                         memset(&qti->qti_lockh, 0, sizeof(qti->qti_lockh));
 957                         if (req_is_preacq(qbody->qb_flags)) {
 958                                 if (req_has_rep(qbody->qb_flags))
 959                                         /* still want to report usage */
 960                                         qbody->qb_flags = QUOTA_DQACQ_FL_REPORT;
 961                                 else
 962                                         /* no pre-acquire if no per-ID lock */
 963                                         GOTO(out, rc = -ENOLCK);
 964                         } else {
 965                                 /* no lock found, should use intent */
 966                                 intent = true;
 967                         }
 968                 } else if (req_is_acq(qbody->qb_flags) &&
 969                            qbody->qb_count == 0) {
 970                         /* found cached lock, no need to acquire */
 971                         GOTO(out, rc = 0);
 972                 }
 973         } else {
 974                 /* release and report don't need a per-ID lock */
 975                 memset(&qti->qti_lockh, 0, sizeof(qti->qti_lockh));
 976         }
 977
 978         if (!intent) {
 979                 rc = qsd_send_dqacq(env, qsd->qsd_exp, qbody, false,
 980                                     qsd_req_completion, qqi, &qti->qti_lockh,
 981                                     lqe);
 982         } else {
 983                 struct lquota_lvb *lvb;
 984
 985                 OBD_ALLOC_PTR(lvb);
 986                 if (lvb == NULL)
 987                         GOTO(out, rc = -ENOMEM);
 988
 989                 rc = qsd_intent_lock(env, qsd->qsd_exp, qbody, false,
 990                                      IT_QUOTA_DQACQ, qsd_req_completion,
 991                                      qqi, lvb, (void *)lqe);
 992         }
 993         /* the completion function will be called by qsd_send_dqacq or
 994          * qsd_intent_lock */
 995         RETURN(rc);
 996 out:
 997         qsd_req_completion(env, qqi, qbody, NULL, &qti->qti_lockh, NULL, lqe,
 998                            rc);
 999         return rc;
1000 }
1001
1002 /**
1003  * Post quota operation, pre-acquire/release quota from master.
1004  *
1005  * \param  env  - the environment passed by the caller
1006  * \param  qsd  - is the qsd instance attached to the OSD device which
1007  *                is handling the operation.
1008  * \param  qqi  - is the qsd_qtype_info structure associated with the quota ID
1009  *                subject to the operation
1010  * \param  qid  - stores information related to his ID for the operation
1011  *                which has just completed
1012  *
1013  * \retval 0    - success
1014  * \retval -ve  - failure
1015  */
1016 static void qsd_op_end0(const struct lu_env *env, struct qsd_qtype_info *qqi,
1017                         struct lquota_id_info *qid)
1018 {
1019         struct lquota_entry     *lqe;
1020         bool                     adjust;
1021         ENTRY;
1022
1023         lqe = qid->lqi_qentry;
1024         if (lqe == NULL)
1025                 RETURN_EXIT;
1026         qid->lqi_qentry = NULL;
1027
1028         /* refresh cached usage if a suitable environment is passed */
1029         if (env != NULL)
1030                 qsd_refresh_usage(env, lqe);
1031
1032         lqe_write_lock(lqe);
1033         if (qid->lqi_space > 0)
1034                 lqe->lqe_pending_write -= qid->lqi_space;
1035         if (env != NULL)
1036                 adjust = qsd_adjust_needed(lqe);
1037         else
1038                 adjust = true;
1039         lqe_write_unlock(lqe);
1040
1041         if (adjust) {
1042                 /* pre-acquire/release quota space is needed */
1043                 if (env != NULL)
1044                         qsd_adjust(env, lqe);
1045                 else
1046                         /* no suitable environment, handle adjustment in
1047                          * separate thread context */
1048                         qsd_adjust_schedule(lqe, false, false);
1049         }
1050         lqe_putref(lqe);
1051         EXIT;
1052 }
1053
1054 /**
1055  * Post quota operation. It's called after each operation transaction stopped.
1056  *
1057  * \param  env   - the environment passed by the caller
1058  * \param  qsd   - is the qsd instance associated with device which is handling
1059  *                 the operation.
1060  * \param  qids  - all qids information attached in the transaction handle
1061  * \param  count - is the number of qid entries in the qids array.
1062  *
1063  * \retval 0     - success
1064  * \retval -ve   - failure
1065  */
1066 void qsd_op_end(const struct lu_env *env, struct qsd_instance *qsd,
1067                 struct lquota_trans *trans)
1068 {
1069         int i;
1070         ENTRY;
1071
1072         if (unlikely(qsd == NULL))
1073                 RETURN_EXIT;
1074
1075         /* We don't enforce quota until the qsd_instance is started */
1076         read_lock(&qsd->qsd_lock);
1077         if (!qsd->qsd_started) {
1078                 read_unlock(&qsd->qsd_lock);
1079                 RETURN_EXIT;
1080         }
1081         read_unlock(&qsd->qsd_lock);
1082
1083         LASSERT(trans != NULL);
1084
1085         for (i = 0; i < trans->lqt_id_cnt; i++) {
1086                 struct qsd_qtype_info *qqi;
1087
1088                 if (trans->lqt_ids[i].lqi_qentry == NULL)
1089                         continue;
1090
1091                 qqi = qsd->qsd_type_array[trans->lqt_ids[i].lqi_type];
1092                 qsd_op_end0(env, qqi, &trans->lqt_ids[i]);
1093         }
1094
1095         /* reset id_count to 0 so that a second accidental call to qsd_op_end()
1096          * does not result in failure */
1097         trans->lqt_id_cnt = 0;
1098         EXIT;
1099 }
1100 EXPORT_SYMBOL(qsd_op_end);
1101
1102 /**
1103  * Trigger pre-acquire/release if necessary.
1104  * It's only used by ldiskfs osd so far. When unlink a file in ldiskfs, the
1105  * quota accounting isn't updated when the transaction stopped. Instead, it'll
1106  * be updated on the final iput, so qsd_op_adjust() will be called then (in
1107  * osd_object_delete()) to trigger quota release if necessary.
1108  *
1109  * \param env - the environment passed by the caller
1110  * \param qsd - is the qsd instance associated with the device in charge
1111  *              of the operation.
1112  * \param qid - is the lquota ID of the user/group for which to trigger
1113  *              quota space adjustment
1114  * \param qtype - is the quota type (USRQUOTA or GRPQUOTA)
1115  */
1116 void qsd_op_adjust(const struct lu_env *env, struct qsd_instance *qsd,
1117                    union lquota_id *qid, int qtype)
1118 {
1119         struct lquota_entry    *lqe;
1120         struct qsd_qtype_info  *qqi;
1121         bool                    adjust;
1122         ENTRY;
1123
1124         if (unlikely(qsd == NULL))
1125                 RETURN_EXIT;
1126
1127         /* We don't enforce quota until the qsd_instance is started */
1128         read_lock(&qsd->qsd_lock);
1129         if (!qsd->qsd_started) {
1130                 read_unlock(&qsd->qsd_lock);
1131                 RETURN_EXIT;
1132         }
1133         read_unlock(&qsd->qsd_lock);
1134
1135         qqi = qsd->qsd_type_array[qtype];
1136         LASSERT(qqi);
1137
1138         if (!qsd_type_enabled(qsd, qtype) || qqi->qqi_acct_obj == NULL ||
1139             qid->qid_uid == 0)
1140                 RETURN_EXIT;
1141
1142         read_lock(&qsd->qsd_lock);
1143         if (!qsd->qsd_started) {
1144                 read_unlock(&qsd->qsd_lock);
1145                 RETURN_EXIT;
1146         }
1147         read_unlock(&qsd->qsd_lock);
1148
1149         lqe = lqe_locate(env, qqi->qqi_site, qid);
1150         if (IS_ERR(lqe)) {
1151                 CERROR("%s: fail to locate lqe for id:"LPU64", type:%d\n",
1152                        qsd->qsd_svname, qid->qid_uid, qtype);
1153                 RETURN_EXIT;
1154         }
1155
1156         qsd_refresh_usage(env, lqe);
1157
1158         lqe_read_lock(lqe);
1159         adjust = qsd_adjust_needed(lqe);
1160         lqe_read_unlock(lqe);
1161
1162         if (adjust)
1163                 qsd_adjust(env, lqe);
1164
1165         lqe_putref(lqe);
1166         EXIT;
1167 }
1168 EXPORT_SYMBOL(qsd_op_adjust);