lustre/quota/quota_interface.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #ifndef EXPORT_SYMTAB
  38 # define EXPORT_SYMTAB
  39 #endif
  40 #define DEBUG_SUBSYSTEM S_LQUOTA
  41
  42 #ifdef __KERNEL__
  43 # include <linux/version.h>
  44 # include <linux/module.h>
  45 # include <linux/init.h>
  46 # include <linux/fs.h>
  47 # include <linux/jbd.h>
  48 # include <linux/smp_lock.h>
  49 # include <linux/buffer_head.h>
  50 # include <linux/workqueue.h>
  51 # include <linux/mount.h>
  52 #else /* __KERNEL__ */
  53 # include <liblustre.h>
  54 #endif
  55
  56 #include <obd_class.h>
  57 #include <lustre_mds.h>
  58 #include <lustre_dlm.h>
  59 #include <lustre_cfg.h>
  60 #include <obd_ost.h>
  61 #include <lustre_fsfilt.h>
  62 #include <lustre_quota.h>
  63 #include <lprocfs_status.h>
  64 #include "quota_internal.h"
  65
  66 #ifdef __KERNEL__
  67
  68 #ifdef HAVE_QUOTA_SUPPORT
  69
  70 static cfs_time_t last_print = 0;
  71 static spinlock_t last_print_lock = SPIN_LOCK_UNLOCKED;
  72
  73 static int filter_quota_setup(struct obd_device *obd)
  74 {
  75         int rc = 0;
  76         struct obd_device_target *obt = &obd->u.obt;
  77         ENTRY;
  78
  79         init_rwsem(&obt->obt_rwsem);
  80         obt->obt_qfmt = LUSTRE_QUOTA_V2;
  81         atomic_set(&obt->obt_quotachecking, 1);
  82         rc = qctxt_init(obd, NULL);
  83         if (rc)
  84                 CERROR("initialize quota context failed! (rc:%d)\n", rc);
  85
  86         RETURN(rc);
  87 }
  88
  89 static int filter_quota_cleanup(struct obd_device *obd)
  90 {
  91         ENTRY;
  92         qctxt_cleanup(&obd->u.obt.obt_qctxt, 0);
  93         RETURN(0);
  94 }
  95
  96 static int filter_quota_setinfo(struct obd_device *obd, void *data)
  97 {
  98         struct obd_export *exp = data;
  99         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
 100         struct obd_import *imp = exp->exp_imp_reverse;
 101         ENTRY;
 102
 103         LASSERT(imp != NULL);
 104
 105         /* setup the quota context import */
 106         spin_lock(&qctxt->lqc_lock);
 107         if (qctxt->lqc_import != NULL) {
 108                 spin_unlock(&qctxt->lqc_lock);
 109                 if (qctxt->lqc_import == imp)
 110                         CDEBUG(D_WARNING, "%s: lqc_import(%p) of obd(%p) was "
 111                                "activated already.\n", obd->obd_name, imp, obd);
 112                 else
 113                         CDEBUG(D_ERROR, "%s: lqc_import(%p:%p) of obd(%p) was "
 114                                "activated by others.\n", obd->obd_name,
 115                                qctxt->lqc_import, imp, obd);
 116         } else {
 117                 qctxt->lqc_import = imp;
 118                 /* make imp's connect flags equal relative exp's connect flags
 119                  * adding it to avoid the scan export list */
 120                 imp->imp_connect_data.ocd_connect_flags |=
 121                                 (exp->exp_connect_flags &
 122                                  (OBD_CONNECT_QUOTA64 | OBD_CONNECT_CHANGE_QS));
 123                 spin_unlock(&qctxt->lqc_lock);
 124                 CDEBUG(D_QUOTA, "%s: lqc_import(%p) of obd(%p) is reactivated "
 125                        "now.\n", obd->obd_name, imp, obd);
 126
 127                 cfs_waitq_signal(&qctxt->lqc_wait_for_qmaster);
 128                 /* start quota slave recovery thread. (release high limits) */
 129                 qslave_start_recovery(obd, qctxt);
 130         }
 131         RETURN(0);
 132 }
 133
 134 static int filter_quota_clearinfo(struct obd_export *exp, struct obd_device *obd)
 135 {
 136         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
 137         struct obd_import *imp = exp->exp_imp_reverse;
 138         ENTRY;
 139
 140         /* lquota may be not set up before destroying export, b=14896 */
 141         if (!obd->obd_set_up)
 142                 RETURN(0);
 143
 144         if (unlikely(imp == NULL))
 145                 RETURN(0);
 146
 147         /* when exp->exp_imp_reverse is destroyed, the corresponding lqc_import
 148          * should be invalid b=12374 */
 149         spin_lock(&qctxt->lqc_lock);
 150         if (qctxt->lqc_import == imp) {
 151                 qctxt->lqc_import = NULL;
 152                 spin_unlock(&qctxt->lqc_lock);
 153                 CDEBUG(D_QUOTA, "%s: lqc_import(%p) of obd(%p) is invalid now.\n",
 154                        obd->obd_name, imp, obd);
 155                 ptlrpc_cleanup_imp(imp);
 156                 dqacq_interrupt(qctxt);
 157         } else {
 158                 spin_unlock(&qctxt->lqc_lock);
 159         }
 160         RETURN(0);
 161 }
 162
 163 static int filter_quota_enforce(struct obd_device *obd, unsigned int ignore)
 164 {
 165         ENTRY;
 166
 167         if (!sb_any_quota_enabled(obd->u.obt.obt_sb))
 168                 RETURN(0);
 169
 170         if (ignore) {
 171                 CDEBUG(D_QUOTA, "blocks will be written with ignoring quota.\n");
 172                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 173         } else {
 174                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
 175         }
 176
 177         RETURN(0);
 178 }
 179
 180 #define GET_OA_ID(flag, oa) (flag == USRQUOTA ? oa->o_uid : oa->o_gid)
 181 static int filter_quota_getflag(struct obd_device *obd, struct obdo *oa)
 182 {
 183         struct obd_device_target *obt = &obd->u.obt;
 184         struct lustre_quota_ctxt *qctxt = &obt->obt_qctxt;
 185         int err, cnt, rc = 0;
 186         struct obd_quotactl *oqctl;
 187         ENTRY;
 188
 189         if (!sb_any_quota_enabled(obt->obt_sb))
 190                 RETURN(0);
 191
 192         OBD_ALLOC_PTR(oqctl);
 193         if (!oqctl) {
 194                 CERROR("Not enough memory!");
 195                 RETURN(-ENOMEM);
 196         }
 197
 198         /* set over quota flags for a uid/gid */
 199         oa->o_valid |= OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA;
 200         oa->o_flags &= ~(OBD_FL_NO_USRQUOTA | OBD_FL_NO_GRPQUOTA);
 201
 202         for (cnt = 0; cnt < MAXQUOTAS; cnt++) {
 203                 struct lustre_qunit_size *lqs = NULL;
 204
 205                 lqs = quota_search_lqs(LQS_KEY(cnt, GET_OA_ID(cnt, oa)),
 206                                        qctxt, 0);
 207                 if (lqs == NULL || IS_ERR(lqs)) {
 208                         rc = PTR_ERR(lqs);
 209                         break;
 210                 } else {
 211                         spin_lock(&lqs->lqs_lock);
 212                         if (lqs->lqs_bunit_sz <= qctxt->lqc_sync_blk) {
 213                                 oa->o_flags |= (cnt == USRQUOTA) ?
 214                                         OBD_FL_NO_USRQUOTA : OBD_FL_NO_GRPQUOTA;
 215                                 spin_unlock(&lqs->lqs_lock);
 216                                 CDEBUG(D_QUOTA, "set sync flag: bunit(%lu), "
 217                                        "sync_blk(%d)\n", lqs->lqs_bunit_sz,
 218                                        qctxt->lqc_sync_blk);
 219                                 /* this is for quota_search_lqs */
 220                                 lqs_putref(lqs);
 221                                 continue;
 222                         }
 223                         spin_unlock(&lqs->lqs_lock);
 224                         /* this is for quota_search_lqs */
 225                         lqs_putref(lqs);
 226                 }
 227
 228                 memset(oqctl, 0, sizeof(*oqctl));
 229
 230                 oqctl->qc_cmd = Q_GETQUOTA;
 231                 oqctl->qc_type = cnt;
 232                 oqctl->qc_id = (cnt == USRQUOTA) ? oa->o_uid : oa->o_gid;
 233                 err = fsfilt_quotactl(obd, obt->obt_sb, oqctl);
 234                 if (err) {
 235                         if (!rc)
 236                                 rc = err;
 237                         oa->o_valid &= ~((cnt == USRQUOTA) ? OBD_MD_FLUSRQUOTA :
 238                                                              OBD_MD_FLGRPQUOTA);
 239                         continue;
 240                 }
 241
 242                 if (oqctl->qc_dqblk.dqb_bhardlimit &&
 243                    (toqb(oqctl->qc_dqblk.dqb_curspace) >=
 244                     oqctl->qc_dqblk.dqb_bhardlimit))
 245                         oa->o_flags |= (cnt == USRQUOTA) ?
 246                                 OBD_FL_NO_USRQUOTA : OBD_FL_NO_GRPQUOTA;
 247         }
 248         OBD_FREE_PTR(oqctl);
 249         RETURN(rc);
 250 }
 251
 252 /**
 253  * check whether the left quota of certain uid and gid can satisfy a block_write
 254  * or inode_create rpc. When need to acquire quota, return QUOTA_RET_ACQUOTA
 255  */
 256 static int quota_check_common(struct obd_device *obd, const unsigned int id[],
 257                               int pending[], int count, int cycle, int isblk,
 258                               struct inode *inode, int frags)
 259 {
 260         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
 261         int i;
 262         struct qunit_data qdata[MAXQUOTAS];
 263         int mb = 0;
 264         int rc = 0, rc2[2] = { 0, 0 };
 265         ENTRY;
 266
 267         CLASSERT(MAXQUOTAS < 4);
 268         if (!sb_any_quota_enabled(qctxt->lqc_sb))
 269                 RETURN(rc);
 270
 271         spin_lock(&qctxt->lqc_lock);
 272         if (!qctxt->lqc_valid){
 273                 spin_unlock(&qctxt->lqc_lock);
 274                 RETURN(rc);
 275         }
 276         spin_unlock(&qctxt->lqc_lock);
 277
 278         for (i = 0; i < MAXQUOTAS; i++) {
 279                 struct lustre_qunit_size *lqs = NULL;
 280
 281                 qdata[i].qd_id = id[i];
 282                 qdata[i].qd_flags = i;
 283                 if (isblk)
 284                         QDATA_SET_BLK(&qdata[i]);
 285                 qdata[i].qd_count = 0;
 286
 287                 /* ignore root user */
 288                 if (qdata[i].qd_id == 0 && !QDATA_IS_GRP(&qdata[i]))
 289                         continue;
 290
 291                 lqs = quota_search_lqs(LQS_KEY(i, id[i]), qctxt, 0);
 292                 if (lqs == NULL || IS_ERR(lqs))
 293                         continue;
 294
 295                 if (IS_ERR(lqs)) {
 296                         CERROR("can not find lqs for check_common: "
 297                                "[id %u] [%c] [isblk %d] [count %d] [rc %ld]\n",
 298                                id[i], i % 2 ? 'g': 'u', isblk, count,
 299                                PTR_ERR(lqs));
 300                         RETURN(PTR_ERR(lqs));
 301                 }
 302
 303                 rc2[i] = compute_remquota(obd, qctxt, &qdata[i], isblk);
 304                 spin_lock(&lqs->lqs_lock);
 305                 if (!cycle) {
 306                         if (isblk) {
 307                                 pending[i] = count * CFS_PAGE_SIZE;
 308                                 /* in order to complete this write, we need extra
 309                                  * meta blocks. This function can get it through
 310                                  * data needed to be written b=16542 */
 311                                 if (inode) {
 312                                         mb = pending[i];
 313                                         rc = fsfilt_get_mblk(obd, qctxt->lqc_sb,
 314                                                              &mb, inode,frags);
 315                                         if (rc)
 316                                                 CDEBUG(D_ERROR,
 317                                                        "can't get extra "
 318                                                        "meta blocks.\n");
 319                                         else
 320                                                 pending[i] += mb;
 321                                 }
 322                                 lqs->lqs_bwrite_pending += pending[i];
 323                         } else {
 324                                 pending[i] = count;
 325                                 lqs->lqs_iwrite_pending += pending[i];
 326                         }
 327                 }
 328
 329                 /* if xx_rec < 0, that means quota are releasing,
 330                  * and it may return before we use quota. So if
 331                  * we find this situation, we assuming it has
 332                  * returned b=18491 */
 333                 if (isblk && lqs->lqs_blk_rec < 0) {
 334                         if (qdata[i].qd_count < -lqs->lqs_blk_rec)
 335                                 qdata[i].qd_count = 0;
 336                         else
 337                                 qdata[i].qd_count += lqs->lqs_blk_rec;
 338                 }
 339                 if (!isblk && lqs->lqs_ino_rec < 0) {
 340                         if (qdata[i].qd_count < -lqs->lqs_ino_rec)
 341                                 qdata[i].qd_count = 0;
 342                         else
 343                                 qdata[i].qd_count += lqs->lqs_ino_rec;
 344                 }
 345
 346                 CDEBUG(D_QUOTA, "[id %u] [%c] [isblk %d] [count %d]"
 347                        " [lqs pending: %lu] [qd_count: "LPU64"] [metablocks: %d]"
 348                        " [pending: %d]\n", id[i], i % 2 ? 'g': 'u', isblk, count,
 349                        isblk ? lqs->lqs_bwrite_pending : lqs->lqs_iwrite_pending,
 350                        qdata[i].qd_count, mb, pending[i]);
 351                 if (rc2[i] == QUOTA_RET_OK) {
 352                         if (isblk && qdata[i].qd_count < lqs->lqs_bwrite_pending)
 353                                 rc2[i] = QUOTA_RET_ACQUOTA;
 354                         if (!isblk && qdata[i].qd_count <
 355                             lqs->lqs_iwrite_pending)
 356                                 rc2[i] = QUOTA_RET_ACQUOTA;
 357                 }
 358
 359                 spin_unlock(&lqs->lqs_lock);
 360
 361                 if (lqs->lqs_blk_rec  < 0 &&
 362                     qdata[i].qd_count <
 363                     lqs->lqs_bwrite_pending - lqs->lqs_blk_rec - mb)
 364                         OBD_FAIL_TIMEOUT(OBD_FAIL_QUOTA_DELAY_REL, 5);
 365
 366                 /* When cycle is zero, lqs_*_pending will be changed. We will
 367                  * get reference of the lqs here and put reference of lqs in
 368                  * quota_pending_commit b=14784 */
 369                 if (!cycle)
 370                         lqs_getref(lqs);
 371
 372                 /* this is for quota_search_lqs */
 373                 lqs_putref(lqs);
 374         }
 375
 376         if (rc2[0] == QUOTA_RET_ACQUOTA || rc2[1] == QUOTA_RET_ACQUOTA)
 377                 RETURN(QUOTA_RET_ACQUOTA);
 378         else
 379                 RETURN(rc);
 380 }
 381
 382 static int quota_chk_acq_common(struct obd_device *obd, const unsigned int id[],
 383                                 int pending[], int count, quota_acquire acquire,
 384                                 struct obd_trans_info *oti, int isblk,
 385                                 struct inode *inode, int frags)
 386 {
 387         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
 388         struct timeval work_start;
 389         struct timeval work_end;
 390         long timediff;
 391         struct l_wait_info lwi = { 0 };
 392         int rc = 0, cycle = 0, count_err = 1;
 393         ENTRY;
 394
 395         CDEBUG(D_QUOTA, "check quota for %s\n", obd->obd_name);
 396         pending[USRQUOTA] = pending[GRPQUOTA] = 0;
 397         /* Unfortunately, if quota master is too busy to handle the
 398          * pre-dqacq in time and quota hash on ost is used up, we
 399          * have to wait for the completion of in flight dqacq/dqrel,
 400          * in order to get enough quota for write b=12588 */
 401         do_gettimeofday(&work_start);
 402         while ((rc = quota_check_common(obd, id, pending, count, cycle, isblk,
 403                                         inode, frags)) &
 404                QUOTA_RET_ACQUOTA) {
 405
 406                 spin_lock(&qctxt->lqc_lock);
 407                 if (!qctxt->lqc_import && oti) {
 408                         spin_unlock(&qctxt->lqc_lock);
 409
 410                         LASSERT(oti && oti->oti_thread &&
 411                                 oti->oti_thread->t_watchdog);
 412
 413                         lc_watchdog_disable(oti->oti_thread->t_watchdog);
 414                         CDEBUG(D_QUOTA, "sleep for quota master\n");
 415                         l_wait_event(qctxt->lqc_wait_for_qmaster, check_qm(qctxt),
 416                                      &lwi);
 417                         CDEBUG(D_QUOTA, "wake up when quota master is back\n");
 418                         lc_watchdog_touch(oti->oti_thread->t_watchdog,
 419                                  GET_TIMEOUT(oti->oti_thread->t_svc));
 420                 } else {
 421                         spin_unlock(&qctxt->lqc_lock);
 422                 }
 423
 424                 cycle++;
 425                 if (isblk)
 426                         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_HOLD_WRITE_RPC, 90);
 427                 /* after acquire(), we should run quota_check_common again
 428                  * so that we confirm there are enough quota to finish write */
 429                 rc = acquire(obd, id, oti, isblk);
 430
 431                 /* please reference to dqacq_completion for the below */
 432                 /* a new request is finished, try again */
 433                 if (rc == QUOTA_REQ_RETURNED) {
 434                         CDEBUG(D_QUOTA, "finish a quota req, try again\n");
 435                         continue;
 436                 }
 437
 438                 /* it is out of quota already */
 439                 if (rc == -EDQUOT) {
 440                         CDEBUG(D_QUOTA, "out of quota,  return -EDQUOT\n");
 441                         break;
 442                 }
 443
 444                 /* -EBUSY and others, wait a second and try again */
 445                 if (rc < 0) {
 446                         cfs_waitq_t        waitq;
 447                         struct l_wait_info lwi;
 448
 449                         if (oti && oti->oti_thread && oti->oti_thread->t_watchdog)
 450                                 lc_watchdog_touch(oti->oti_thread->t_watchdog,
 451                                          GET_TIMEOUT(oti->oti_thread->t_svc));
 452                         CDEBUG(D_QUOTA, "rc: %d, count_err: %d\n", rc,
 453                                count_err++);
 454
 455                         init_waitqueue_head(&waitq);
 456                         lwi = LWI_TIMEOUT(cfs_time_seconds(min(cycle, 10)), NULL,
 457                                           NULL);
 458                         l_wait_event(waitq, 0, &lwi);
 459                 }
 460
 461                 if (rc < 0 || cycle % 10 == 2) {
 462                         spin_lock(&last_print_lock);
 463                         if (last_print == 0 ||
 464                             cfs_time_before((last_print + cfs_time_seconds(30)),
 465                                             cfs_time_current())) {
 466                                 last_print = cfs_time_current();
 467                                 spin_unlock(&last_print_lock);
 468                                 CWARN("still haven't managed to acquire quota "
 469                                       "space from the quota master after %d "
 470                                       "retries (err=%d, rc=%d)\n",
 471                                       cycle, count_err - 1, rc);
 472                         } else {
 473                                 spin_unlock(&last_print_lock);
 474                         }
 475                 }
 476
 477                 CDEBUG(D_QUOTA, "recheck quota with rc: %d, cycle: %d\n", rc,
 478                        cycle);
 479         }
 480         do_gettimeofday(&work_end);
 481         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
 482         lprocfs_counter_add(qctxt->lqc_stats,
 483                             isblk ? LQUOTA_WAIT_FOR_CHK_BLK :
 484                                     LQUOTA_WAIT_FOR_CHK_INO,
 485                             timediff);
 486
 487         RETURN(rc);
 488 }
 489
 490 /**
 491  * when a block_write or inode_create rpc is finished, adjust the record for
 492  * pending blocks and inodes
 493  */
 494 static int quota_pending_commit(struct obd_device *obd, const unsigned int id[],
 495                                 int pending[], int isblk)
 496 {
 497         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
 498         struct timeval work_start;
 499         struct timeval work_end;
 500         long timediff;
 501         int i;
 502         struct qunit_data qdata[MAXQUOTAS];
 503         ENTRY;
 504
 505         CDEBUG(D_QUOTA, "commit pending quota for  %s\n", obd->obd_name);
 506         CLASSERT(MAXQUOTAS < 4);
 507         if (!sb_any_quota_enabled(qctxt->lqc_sb))
 508                 RETURN(0);
 509
 510         do_gettimeofday(&work_start);
 511         for (i = 0; i < MAXQUOTAS; i++) {
 512                 struct lustre_qunit_size *lqs = NULL;
 513
 514                 LASSERT(pending[i] >= 0);
 515                 if (pending[i] == 0)
 516                         continue;
 517
 518                 qdata[i].qd_id = id[i];
 519                 qdata[i].qd_flags = i;
 520                 if (isblk)
 521                         QDATA_SET_BLK(&qdata[i]);
 522                 qdata[i].qd_count = 0;
 523
 524                 if (qdata[i].qd_id == 0 && !QDATA_IS_GRP(&qdata[i]))
 525                         continue;
 526
 527                 lqs = quota_search_lqs(LQS_KEY(i, qdata[i].qd_id), qctxt, 0);
 528                 if (lqs == NULL || IS_ERR(lqs)) {
 529                         CERROR("can not find lqs for pending_commit: "
 530                                "[id %u] [%c] [pending %u] [isblk %d] (rc %ld), "
 531                                "maybe cause unexpected lqs refcount error!\n",
 532                                id[i], i ? 'g': 'u', pending[i], isblk,
 533                                lqs ? PTR_ERR(lqs) : -1);
 534                         continue;
 535                 }
 536
 537                 spin_lock(&lqs->lqs_lock);
 538                 if (isblk) {
 539                         LASSERTF(lqs->lqs_bwrite_pending >= pending[i],
 540                                  "there are too many blocks! [id %u] [%c] "
 541                                  "[bwrite_pending %lu] [pending %u]\n",
 542                                  id[i], i % 2 ? 'g' : 'u',
 543                                  lqs->lqs_bwrite_pending, pending[i]);
 544
 545                         lqs->lqs_bwrite_pending -= pending[i];
 546                 } else {
 547                         LASSERTF(lqs->lqs_iwrite_pending >= pending[i],
 548                                 "there are too many files! [id %u] [%c] "
 549                                 "[iwrite_pending %lu] [pending %u]\n",
 550                                 id[i], i % 2 ? 'g' : 'u',
 551                                 lqs->lqs_iwrite_pending, pending[i]);
 552
 553                         lqs->lqs_iwrite_pending -= pending[i];
 554                 }
 555                 CDEBUG(D_QUOTA, "%s: lqs_pending=%lu pending[%d]=%d isblk=%d\n",
 556                        obd->obd_name,
 557                        isblk ? lqs->lqs_bwrite_pending : lqs->lqs_iwrite_pending,
 558                        i, pending[i], isblk);
 559                 spin_unlock(&lqs->lqs_lock);
 560
 561                 /* for quota_search_lqs in pending_commit */
 562                 lqs_putref(lqs);
 563                 /* for quota_search_lqs in quota_check */
 564                 lqs_putref(lqs);
 565         }
 566         do_gettimeofday(&work_end);
 567         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
 568         lprocfs_counter_add(qctxt->lqc_stats,
 569                             isblk ? LQUOTA_WAIT_FOR_COMMIT_BLK :
 570                                     LQUOTA_WAIT_FOR_COMMIT_INO,
 571                             timediff);
 572
 573         RETURN(0);
 574 }
 575
 576 static int mds_quota_init(void)
 577 {
 578         return lustre_dquot_init();
 579 }
 580
 581 static int mds_quota_exit(void)
 582 {
 583         lustre_dquot_exit();
 584         return 0;
 585 }
 586
 587 static int mds_quota_setup(struct obd_device *obd)
 588 {
 589         struct obd_device_target *obt = &obd->u.obt;
 590         struct mds_obd *mds = &obd->u.mds;
 591         int rc;
 592         ENTRY;
 593
 594         if (unlikely(mds->mds_quota)) {
 595                 CWARN("try to reinitialize quota context!\n");
 596                 RETURN(0);
 597         }
 598
 599         init_rwsem(&obt->obt_rwsem);
 600         obt->obt_qfmt = LUSTRE_QUOTA_V2;
 601         mds->mds_quota_info.qi_version = LUSTRE_QUOTA_V2;
 602         atomic_set(&obt->obt_quotachecking, 1);
 603         /* initialize quota master and quota context */
 604         sema_init(&mds->mds_qonoff_sem, 1);
 605         rc = qctxt_init(obd, dqacq_handler);
 606         if (rc) {
 607                 CERROR("initialize quota context failed! (rc:%d)\n", rc);
 608                 RETURN(rc);
 609         }
 610         mds->mds_quota = 1;
 611         RETURN(rc);
 612 }
 613
 614 static int mds_quota_cleanup(struct obd_device *obd)
 615 {
 616         ENTRY;
 617         if (unlikely(!obd->u.mds.mds_quota))
 618                 RETURN(0);
 619
 620         qctxt_cleanup(&obd->u.obt.obt_qctxt, 0);
 621         RETURN(0);
 622 }
 623
 624 static int mds_quota_setinfo(struct obd_device *obd, void *data)
 625 {
 626         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
 627         ENTRY;
 628
 629         if (unlikely(!obd->u.mds.mds_quota))
 630                 RETURN(0);
 631
 632         if (data != NULL)
 633                 QUOTA_MASTER_READY(qctxt);
 634         else
 635                 QUOTA_MASTER_UNREADY(qctxt);
 636         RETURN(0);
 637 }
 638
 639 static int mds_quota_fs_cleanup(struct obd_device *obd)
 640 {
 641         struct mds_obd *mds = &obd->u.mds;
 642         struct obd_quotactl oqctl;
 643         ENTRY;
 644
 645         if (unlikely(!mds->mds_quota))
 646                 RETURN(0);
 647
 648         mds->mds_quota = 0;
 649         memset(&oqctl, 0, sizeof(oqctl));
 650         oqctl.qc_type = UGQUOTA;
 651
 652         down(&mds->mds_qonoff_sem);
 653         mds_admin_quota_off(obd, &oqctl);
 654         up(&mds->mds_qonoff_sem);
 655         RETURN(0);
 656 }
 657
 658 static int quota_acquire_common(struct obd_device *obd, const unsigned int id[],
 659                                 struct obd_trans_info *oti, int isblk)
 660 {
 661         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
 662         int rc;
 663         ENTRY;
 664
 665         rc = qctxt_adjust_qunit(obd, qctxt, id, isblk, 1, oti);
 666         RETURN(rc);
 667 }
 668
 669 #endif /* HAVE_QUOTA_SUPPORT */
 670 #endif /* __KERNEL__ */
 671
 672 struct osc_quota_info {
 673         struct list_head        oqi_hash;       /* hash list */
 674         struct client_obd      *oqi_cli;        /* osc obd */
 675         unsigned int            oqi_id;         /* uid/gid of a file */
 676         short                   oqi_type;       /* quota type */
 677 };
 678
 679 spinlock_t qinfo_list_lock = SPIN_LOCK_UNLOCKED;
 680
 681 static struct list_head qinfo_hash[NR_DQHASH];
 682 /* SLAB cache for client quota context */
 683 cfs_mem_cache_t *qinfo_cachep = NULL;
 684
 685 static inline int hashfn(struct client_obd *cli, unsigned long id, int type)
 686                          __attribute__((__const__));
 687
 688 static inline int hashfn(struct client_obd *cli, unsigned long id, int type)
 689 {
 690         unsigned long tmp = ((unsigned long)cli>>6) ^ id;
 691         tmp = (tmp * (MAXQUOTAS - type)) % NR_DQHASH;
 692         return tmp;
 693 }
 694
 695 /* caller must hold qinfo_list_lock */
 696 static inline void insert_qinfo_hash(struct osc_quota_info *oqi)
 697 {
 698         struct list_head *head = qinfo_hash +
 699                 hashfn(oqi->oqi_cli, oqi->oqi_id, oqi->oqi_type);
 700
 701         LASSERT_SPIN_LOCKED(&qinfo_list_lock);
 702         list_add(&oqi->oqi_hash, head);
 703 }
 704
 705 /* caller must hold qinfo_list_lock */
 706 static inline void remove_qinfo_hash(struct osc_quota_info *oqi)
 707 {
 708         LASSERT_SPIN_LOCKED(&qinfo_list_lock);
 709         list_del_init(&oqi->oqi_hash);
 710 }
 711
 712 /* caller must hold qinfo_list_lock */
 713 static inline struct osc_quota_info *find_qinfo(struct client_obd *cli,
 714                                                 unsigned int id, int type)
 715 {
 716         unsigned int hashent = hashfn(cli, id, type);
 717         struct osc_quota_info *oqi;
 718         ENTRY;
 719
 720         LASSERT_SPIN_LOCKED(&qinfo_list_lock);
 721         list_for_each_entry(oqi, &qinfo_hash[hashent], oqi_hash) {
 722                 if (oqi->oqi_cli == cli &&
 723                     oqi->oqi_id == id && oqi->oqi_type == type)
 724                         return oqi;
 725         }
 726         RETURN(NULL);
 727 }
 728
 729 static struct osc_quota_info *alloc_qinfo(struct client_obd *cli,
 730                                           unsigned int id, int type)
 731 {
 732         struct osc_quota_info *oqi;
 733         ENTRY;
 734
 735         OBD_SLAB_ALLOC(oqi, qinfo_cachep, CFS_ALLOC_STD, sizeof(*oqi));
 736         if(!oqi)
 737                 RETURN(NULL);
 738
 739         CFS_INIT_LIST_HEAD(&oqi->oqi_hash);
 740         oqi->oqi_cli = cli;
 741         oqi->oqi_id = id;
 742         oqi->oqi_type = type;
 743
 744         RETURN(oqi);
 745 }
 746
 747 static void free_qinfo(struct osc_quota_info *oqi)
 748 {
 749         OBD_SLAB_FREE(oqi, qinfo_cachep, sizeof(*oqi));
 750 }
 751
 752 int osc_quota_chkdq(struct client_obd *cli, const unsigned int qid[])
 753 {
 754         unsigned int id;
 755         int cnt, rc = QUOTA_OK;
 756         ENTRY;
 757
 758         spin_lock(&qinfo_list_lock);
 759         for (cnt = 0; cnt < MAXQUOTAS; cnt++) {
 760                 struct osc_quota_info *oqi = NULL;
 761
 762                 id = (cnt == USRQUOTA) ? qid[USRQUOTA] : qid[GRPQUOTA];
 763                 oqi = find_qinfo(cli, id, cnt);
 764                 if (oqi) {
 765                         rc = NO_QUOTA;
 766                         break;
 767                 }
 768         }
 769         spin_unlock(&qinfo_list_lock);
 770
 771         RETURN(rc);
 772 }
 773
 774 int osc_quota_setdq(struct client_obd *cli, const unsigned int qid[],
 775                     obd_flag valid, obd_flag flags)
 776 {
 777         unsigned int id;
 778         obd_flag noquota;
 779         int cnt, rc = 0;
 780         ENTRY;
 781
 782
 783         for (cnt = 0; cnt < MAXQUOTAS; cnt++) {
 784                 struct osc_quota_info *oqi, *old;
 785
 786                 if (!(valid & ((cnt == USRQUOTA) ?
 787                     OBD_MD_FLUSRQUOTA : OBD_MD_FLGRPQUOTA)))
 788                         continue;
 789
 790                 id = (cnt == USRQUOTA) ? qid[USRQUOTA] : qid[GRPQUOTA];
 791                 noquota = (cnt == USRQUOTA) ?
 792                     (flags & OBD_FL_NO_USRQUOTA) : (flags & OBD_FL_NO_GRPQUOTA);
 793
 794                 oqi = alloc_qinfo(cli, id, cnt);
 795                 if (oqi) {
 796                         spin_lock(&qinfo_list_lock);
 797
 798                         old = find_qinfo(cli, id, cnt);
 799                         if (old && !noquota)
 800                                 remove_qinfo_hash(old);
 801                         else if (!old && noquota)
 802                                 insert_qinfo_hash(oqi);
 803
 804                         spin_unlock(&qinfo_list_lock);
 805
 806                         if (old || !noquota)
 807                                 free_qinfo(oqi);
 808                         if (old && !noquota)
 809                                 free_qinfo(old);
 810                 } else {
 811                         CERROR("not enough mem!\n");
 812                         rc = -ENOMEM;
 813                         break;
 814                 }
 815         }
 816
 817         RETURN(rc);
 818 }
 819
 820 int osc_quota_cleanup(struct obd_device *obd)
 821 {
 822         struct client_obd *cli = &obd->u.cli;
 823         struct osc_quota_info *oqi, *n;
 824         int i;
 825         ENTRY;
 826
 827         spin_lock(&qinfo_list_lock);
 828         for (i = 0; i < NR_DQHASH; i++) {
 829                 list_for_each_entry_safe(oqi, n, &qinfo_hash[i], oqi_hash) {
 830                         if (oqi->oqi_cli != cli)
 831                                 continue;
 832                         remove_qinfo_hash(oqi);
 833                         free_qinfo(oqi);
 834                 }
 835         }
 836         spin_unlock(&qinfo_list_lock);
 837
 838         RETURN(0);
 839 }
 840
 841 int osc_quota_init(void)
 842 {
 843         int i;
 844         ENTRY;
 845
 846         LASSERT(qinfo_cachep == NULL);
 847         qinfo_cachep = cfs_mem_cache_create("osc_quota_info",
 848                                             sizeof(struct osc_quota_info),
 849                                             0, 0);
 850         if (!qinfo_cachep)
 851                 RETURN(-ENOMEM);
 852
 853         for (i = 0; i < NR_DQHASH; i++)
 854                 CFS_INIT_LIST_HEAD(qinfo_hash + i);
 855
 856         RETURN(0);
 857 }
 858
 859 int osc_quota_exit(void)
 860 {
 861         struct osc_quota_info *oqi, *n;
 862         int i, rc;
 863         ENTRY;
 864
 865         spin_lock(&qinfo_list_lock);
 866         for (i = 0; i < NR_DQHASH; i++) {
 867                 list_for_each_entry_safe(oqi, n, &qinfo_hash[i], oqi_hash) {
 868                         remove_qinfo_hash(oqi);
 869                         free_qinfo(oqi);
 870                 }
 871         }
 872         spin_unlock(&qinfo_list_lock);
 873
 874         rc = cfs_mem_cache_destroy(qinfo_cachep);
 875         LASSERTF(rc == 0, "couldn't destory qinfo_cachep slab\n");
 876         qinfo_cachep = NULL;
 877
 878         RETURN(0);
 879 }
 880
 881 #ifdef __KERNEL__
 882 #ifdef HAVE_QUOTA_SUPPORT
 883 quota_interface_t mds_quota_interface = {
 884         .quota_init     = mds_quota_init,
 885         .quota_exit     = mds_quota_exit,
 886         .quota_setup    = mds_quota_setup,
 887         .quota_cleanup  = mds_quota_cleanup,
 888         .quota_check    = target_quota_check,
 889         .quota_ctl      = mds_quota_ctl,
 890         .quota_setinfo  = mds_quota_setinfo,
 891         .quota_fs_cleanup = mds_quota_fs_cleanup,
 892         .quota_recovery = mds_quota_recovery,
 893         .quota_adjust   = mds_quota_adjust,
 894         .quota_chkquota = quota_chk_acq_common,
 895         .quota_acquire  = quota_acquire_common,
 896         .quota_pending_commit = quota_pending_commit,
 897 };
 898
 899 quota_interface_t filter_quota_interface = {
 900         .quota_setup    = filter_quota_setup,
 901         .quota_cleanup  = filter_quota_cleanup,
 902         .quota_check    = target_quota_check,
 903         .quota_ctl      = filter_quota_ctl,
 904         .quota_setinfo  = filter_quota_setinfo,
 905         .quota_clearinfo = filter_quota_clearinfo,
 906         .quota_enforce  = filter_quota_enforce,
 907         .quota_getflag  = filter_quota_getflag,
 908         .quota_acquire  = quota_acquire_common,
 909         .quota_adjust   = filter_quota_adjust,
 910         .quota_chkquota = quota_chk_acq_common,
 911         .quota_adjust_qunit   = filter_quota_adjust_qunit,
 912         .quota_pending_commit = quota_pending_commit,
 913 };
 914 #endif
 915 #endif /* __KERNEL__ */
 916
 917 quota_interface_t mdc_quota_interface = {
 918         .quota_ctl      = client_quota_ctl,
 919         .quota_check    = client_quota_check,
 920         .quota_poll_check = client_quota_poll_check,
 921 };
 922
 923 quota_interface_t lmv_quota_interface = {
 924         .quota_ctl      = lmv_quota_ctl,
 925         .quota_check    = lmv_quota_check,
 926 };
 927
 928 quota_interface_t osc_quota_interface = {
 929         .quota_ctl      = client_quota_ctl,
 930         .quota_check    = client_quota_check,
 931         .quota_poll_check = client_quota_poll_check,
 932         .quota_init     = osc_quota_init,
 933         .quota_exit     = osc_quota_exit,
 934         .quota_chkdq    = osc_quota_chkdq,
 935         .quota_setdq    = osc_quota_setdq,
 936         .quota_cleanup  = osc_quota_cleanup,
 937         .quota_adjust_qunit = client_quota_adjust_qunit,
 938 };
 939
 940 quota_interface_t lov_quota_interface = {
 941         .quota_ctl      = lov_quota_ctl,
 942         .quota_check    = lov_quota_check,
 943         .quota_adjust_qunit = lov_quota_adjust_qunit,
 944 };
 945
 946 #ifdef __KERNEL__
 947
 948 cfs_proc_dir_entry_t *lquota_type_proc_dir = NULL;
 949
 950 static int __init init_lustre_quota(void)
 951 {
 952 #ifdef HAVE_QUOTA_SUPPORT
 953         int rc = 0;
 954
 955         lquota_type_proc_dir = lprocfs_register(OBD_LQUOTA_DEVICENAME,
 956                                                 proc_lustre_root,
 957                                                 NULL, NULL);
 958         if (IS_ERR(lquota_type_proc_dir)) {
 959                 CERROR("LProcFS failed in lquota-init\n");
 960                 rc = PTR_ERR(lquota_type_proc_dir);
 961                 return rc;
 962         }
 963
 964         rc = qunit_cache_init();
 965         if (rc)
 966                 return rc;
 967
 968         PORTAL_SYMBOL_REGISTER(filter_quota_interface);
 969         PORTAL_SYMBOL_REGISTER(mds_quota_interface);
 970 #endif
 971         PORTAL_SYMBOL_REGISTER(mdc_quota_interface);
 972         PORTAL_SYMBOL_REGISTER(lmv_quota_interface);
 973         PORTAL_SYMBOL_REGISTER(osc_quota_interface);
 974         PORTAL_SYMBOL_REGISTER(lov_quota_interface);
 975         return 0;
 976 }
 977
 978 static void /*__exit*/ exit_lustre_quota(void)
 979 {
 980         PORTAL_SYMBOL_UNREGISTER(mdc_quota_interface);
 981         PORTAL_SYMBOL_UNREGISTER(lmv_quota_interface);
 982         PORTAL_SYMBOL_UNREGISTER(osc_quota_interface);
 983         PORTAL_SYMBOL_UNREGISTER(lov_quota_interface);
 984 #ifdef HAVE_QUOTA_SUPPORT
 985         PORTAL_SYMBOL_UNREGISTER(filter_quota_interface);
 986         PORTAL_SYMBOL_UNREGISTER(mds_quota_interface);
 987
 988         qunit_cache_cleanup();
 989
 990         if (lquota_type_proc_dir)
 991                 lprocfs_remove(&lquota_type_proc_dir);
 992 #endif
 993 }
 994
 995 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
 996 MODULE_DESCRIPTION("Lustre Quota");
 997 MODULE_LICENSE("GPL");
 998
 999 cfs_module(lquota, "1.0.0", init_lustre_quota, exit_lustre_quota);
1000
1001 #ifdef HAVE_QUOTA_SUPPORT
1002 EXPORT_SYMBOL(mds_quota_interface);
1003 EXPORT_SYMBOL(filter_quota_interface);
1004 #endif
1005 EXPORT_SYMBOL(mdc_quota_interface);
1006 EXPORT_SYMBOL(lmv_quota_interface);
1007 EXPORT_SYMBOL(osc_quota_interface);
1008 EXPORT_SYMBOL(lov_quota_interface);
1009 #endif /* __KERNEL */