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 static int filter_quota_getflag(struct obd_device *obd, struct obdo *oa)
 181 {
 182         struct obd_device_target *obt = &obd->u.obt;
 183         struct lustre_quota_ctxt *qctxt = &obt->obt_qctxt;
 184         int err, cnt, rc = 0;
 185         struct obd_quotactl *oqctl;
 186         ENTRY;
 187
 188         if (!sb_any_quota_enabled(obt->obt_sb))
 189                 RETURN(0);
 190
 191         OBD_ALLOC_PTR(oqctl);
 192         if (!oqctl) {
 193                 CERROR("Not enough memory!");
 194                 RETURN(-ENOMEM);
 195         }
 196
 197         /* set over quota flags for a uid/gid */
 198         oa->o_valid |= OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA;
 199         oa->o_flags &= ~(OBD_FL_NO_USRQUOTA | OBD_FL_NO_GRPQUOTA);
 200
 201         for (cnt = 0; cnt < MAXQUOTAS; cnt++) {
 202                 struct quota_adjust_qunit oqaq_tmp;
 203                 struct lustre_qunit_size *lqs = NULL;
 204
 205                 oqaq_tmp.qaq_flags = cnt;
 206                 oqaq_tmp.qaq_id = (cnt == USRQUOTA) ? oa->o_uid : oa->o_gid;
 207
 208                 quota_search_lqs(NULL, &oqaq_tmp, qctxt, &lqs);
 209                 if (lqs) {
 210                         spin_lock(&lqs->lqs_lock);
 211                         if (lqs->lqs_bunit_sz <= qctxt->lqc_sync_blk) {
 212                                 oa->o_flags |= (cnt == USRQUOTA) ?
 213                                         OBD_FL_NO_USRQUOTA : OBD_FL_NO_GRPQUOTA;
 214                                 spin_unlock(&lqs->lqs_lock);
 215                                 CDEBUG(D_QUOTA, "set sync flag: bunit(%lu), "
 216                                        "sync_blk(%d)\n", lqs->lqs_bunit_sz,
 217                                        qctxt->lqc_sync_blk);
 218                                 /* this is for quota_search_lqs */
 219                                 lqs_putref(lqs);
 220                                 continue;
 221                         }
 222                         spin_unlock(&lqs->lqs_lock);
 223                         /* this is for quota_search_lqs */
 224                         lqs_putref(lqs);
 225                 }
 226
 227                 memset(oqctl, 0, sizeof(*oqctl));
 228
 229                 oqctl->qc_cmd = Q_GETQUOTA;
 230                 oqctl->qc_type = cnt;
 231                 oqctl->qc_id = (cnt == USRQUOTA) ? oa->o_uid : oa->o_gid;
 232                 err = fsfilt_quotactl(obd, obt->obt_sb, oqctl);
 233                 if (err) {
 234                         if (!rc)
 235                                 rc = err;
 236                         oa->o_valid &= ~((cnt == USRQUOTA) ? OBD_MD_FLUSRQUOTA :
 237                                                              OBD_MD_FLGRPQUOTA);
 238                         continue;
 239                 }
 240
 241                 if (oqctl->qc_dqblk.dqb_bhardlimit &&
 242                    (toqb(oqctl->qc_dqblk.dqb_curspace) >=
 243                     oqctl->qc_dqblk.dqb_bhardlimit))
 244                         oa->o_flags |= (cnt == USRQUOTA) ?
 245                                 OBD_FL_NO_USRQUOTA : OBD_FL_NO_GRPQUOTA;
 246         }
 247         OBD_FREE_PTR(oqctl);
 248         RETURN(rc);
 249 }
 250
 251 /**
 252  * check whether the left quota of certain uid and gid can satisfy a block_write
 253  * or inode_create rpc. When need to acquire quota, return QUOTA_RET_ACQUOTA
 254  */
 255 static int quota_check_common(struct obd_device *obd, const unsigned int id[],
 256                               int pending[], int count, int cycle, int isblk,
 257                               struct inode *inode, int frags)
 258 {
 259         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
 260         int i;
 261         struct qunit_data qdata[MAXQUOTAS];
 262         int mb = 0;
 263         int rc = 0, rc2[2] = { 0, 0 };
 264         ENTRY;
 265
 266         CLASSERT(MAXQUOTAS < 4);
 267         if (!sb_any_quota_enabled(qctxt->lqc_sb))
 268                 RETURN(rc);
 269
 270         spin_lock(&qctxt->lqc_lock);
 271         if (!qctxt->lqc_valid){
 272                 spin_unlock(&qctxt->lqc_lock);
 273                 RETURN(rc);
 274         }
 275         spin_unlock(&qctxt->lqc_lock);
 276
 277         for (i = 0; i < MAXQUOTAS; i++) {
 278                 struct lustre_qunit_size *lqs = NULL;
 279
 280                 qdata[i].qd_id = id[i];
 281                 qdata[i].qd_flags = i;
 282                 if (isblk)
 283                         QDATA_SET_BLK(&qdata[i]);
 284                 qdata[i].qd_count = 0;
 285
 286                 /* ignore root user */
 287                 if (qdata[i].qd_id == 0 && !QDATA_IS_GRP(&qdata[i]))
 288                         continue;
 289
 290                 quota_search_lqs(&qdata[i], NULL, qctxt, &lqs);
 291                 if (!lqs)
 292                         continue;
 293
 294                 if (IS_ERR(lqs)) {
 295                         CERROR("can not find lqs for check_common: "
 296                                "[id %u] [%c] [isblk %d] [count %d] [rc %ld]\n",
 297                                id[i], i % 2 ? 'g': 'u', isblk, count,
 298                                PTR_ERR(lqs));
 299                         RETURN(PTR_ERR(lqs));
 300                 }
 301
 302                 rc2[i] = compute_remquota(obd, qctxt, &qdata[i], isblk);
 303                 spin_lock(&lqs->lqs_lock);
 304                 if (!cycle) {
 305                         if (isblk) {
 306                                 pending[i] = count * CFS_PAGE_SIZE;
 307                                 /* in order to complete this write, we need extra
 308                                  * meta blocks. This function can get it through
 309                                  * data needed to be written b=16542 */
 310                                 if (inode) {
 311                                         mb = pending[i];
 312                                         rc = fsfilt_get_mblk(obd, qctxt->lqc_sb,
 313                                                              &mb, inode,frags);
 314                                         if (rc)
 315                                                 CDEBUG(D_ERROR,
 316                                                        "can't get extra "
 317                                                        "meta blocks.\n");
 318                                         else
 319                                                 pending[i] += mb;
 320                                 }
 321                                 lqs->lqs_bwrite_pending += pending[i];
 322                         } else {
 323                                 pending[i] = count;
 324                                 lqs->lqs_iwrite_pending += pending[i];
 325                         }
 326                 }
 327
 328                 /* if xx_rec < 0, that means quota are releasing,
 329                  * and it may return before we use quota. So if
 330                  * we find this situation, we assuming it has
 331                  * returned b=18491 */
 332                 if (isblk && lqs->lqs_blk_rec < 0) {
 333                         if (qdata[i].qd_count < -lqs->lqs_blk_rec)
 334                                 qdata[i].qd_count = 0;
 335                         else
 336                                 qdata[i].qd_count += lqs->lqs_blk_rec;
 337                 }
 338                 if (!isblk && lqs->lqs_ino_rec < 0) {
 339                         if (qdata[i].qd_count < -lqs->lqs_ino_rec)
 340                                 qdata[i].qd_count = 0;
 341                         else
 342                                 qdata[i].qd_count += lqs->lqs_ino_rec;
 343                 }
 344
 345                 CDEBUG(D_QUOTA, "[id %u] [%c] [isblk %d] [count %d]"
 346                        " [lqs pending: %lu] [qd_count: "LPU64"] [metablocks: %d]"
 347                        " [pending: %d]\n", id[i], i % 2 ? 'g': 'u', isblk, count,
 348                        isblk ? lqs->lqs_bwrite_pending : lqs->lqs_iwrite_pending,
 349                        qdata[i].qd_count, mb, pending[i]);
 350                 if (rc2[i] == QUOTA_RET_OK) {
 351                         if (isblk && qdata[i].qd_count < lqs->lqs_bwrite_pending)
 352                                 rc2[i] = QUOTA_RET_ACQUOTA;
 353                         if (!isblk && qdata[i].qd_count <
 354                             lqs->lqs_iwrite_pending)
 355                                 rc2[i] = QUOTA_RET_ACQUOTA;
 356                 }
 357
 358                 spin_unlock(&lqs->lqs_lock);
 359
 360                 if (lqs->lqs_blk_rec  < 0 &&
 361                     qdata[i].qd_count <
 362                     lqs->lqs_bwrite_pending - lqs->lqs_blk_rec - mb)
 363                         OBD_FAIL_TIMEOUT(OBD_FAIL_QUOTA_DELAY_REL, 5);
 364
 365                 /* When cycle is zero, lqs_*_pending will be changed. We will
 366                  * get reference of the lqs here and put reference of lqs in
 367                  * quota_pending_commit b=14784 */
 368                 if (!cycle)
 369                         lqs_getref(lqs);
 370
 371                 /* this is for quota_search_lqs */
 372                 lqs_putref(lqs);
 373         }
 374
 375         if (rc2[0] == QUOTA_RET_ACQUOTA || rc2[1] == QUOTA_RET_ACQUOTA)
 376                 RETURN(QUOTA_RET_ACQUOTA);
 377         else
 378                 RETURN(rc);
 379 }
 380
 381 static int quota_chk_acq_common(struct obd_device *obd, const unsigned int id[],
 382                                 int pending[], int count, quota_acquire acquire,
 383                                 struct obd_trans_info *oti, int isblk,
 384                                 struct inode *inode, int frags)
 385 {
 386         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
 387         struct timeval work_start;
 388         struct timeval work_end;
 389         long timediff;
 390         struct l_wait_info lwi = { 0 };
 391         int rc = 0, cycle = 0, count_err = 1;
 392         ENTRY;
 393
 394         CDEBUG(D_QUOTA, "check quota for %s\n", obd->obd_name);
 395         pending[USRQUOTA] = pending[GRPQUOTA] = 0;
 396         /* Unfortunately, if quota master is too busy to handle the
 397          * pre-dqacq in time and quota hash on ost is used up, we
 398          * have to wait for the completion of in flight dqacq/dqrel,
 399          * in order to get enough quota for write b=12588 */
 400         do_gettimeofday(&work_start);
 401         while ((rc = quota_check_common(obd, id, pending, count, cycle, isblk,
 402                                         inode, frags)) &
 403                QUOTA_RET_ACQUOTA) {
 404
 405                 spin_lock(&qctxt->lqc_lock);
 406                 if (!qctxt->lqc_import && oti) {
 407                         spin_unlock(&qctxt->lqc_lock);
 408
 409                         LASSERT(oti && oti->oti_thread &&
 410                                 oti->oti_thread->t_watchdog);
 411
 412                         lc_watchdog_disable(oti->oti_thread->t_watchdog);
 413                         CDEBUG(D_QUOTA, "sleep for quota master\n");
 414                         l_wait_event(qctxt->lqc_wait_for_qmaster, check_qm(qctxt),
 415                                      &lwi);
 416                         CDEBUG(D_QUOTA, "wake up when quota master is back\n");
 417                         lc_watchdog_touch(oti->oti_thread->t_watchdog,
 418                                  GET_TIMEOUT(oti->oti_thread->t_svc));
 419                 } else {
 420                         spin_unlock(&qctxt->lqc_lock);
 421                 }
 422
 423                 cycle++;
 424                 if (isblk)
 425                         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_HOLD_WRITE_RPC, 90);
 426                 /* after acquire(), we should run quota_check_common again
 427                  * so that we confirm there are enough quota to finish write */
 428                 rc = acquire(obd, id, oti, isblk);
 429
 430                 /* please reference to dqacq_completion for the below */
 431                 /* a new request is finished, try again */
 432                 if (rc == QUOTA_REQ_RETURNED) {
 433                         CDEBUG(D_QUOTA, "finish a quota req, try again\n");
 434                         continue;
 435                 }
 436
 437                 /* it is out of quota already */
 438                 if (rc == -EDQUOT) {
 439                         CDEBUG(D_QUOTA, "out of quota,  return -EDQUOT\n");
 440                         break;
 441                 }
 442
 443                 /* -EBUSY and others, wait a second and try again */
 444                 if (rc < 0) {
 445                         cfs_waitq_t        waitq;
 446                         struct l_wait_info lwi;
 447
 448                         if (oti && oti->oti_thread && oti->oti_thread->t_watchdog)
 449                                 lc_watchdog_touch(oti->oti_thread->t_watchdog,
 450                                          GET_TIMEOUT(oti->oti_thread->t_svc));
 451                         CDEBUG(D_QUOTA, "rc: %d, count_err: %d\n", rc,
 452                                count_err++);
 453
 454                         init_waitqueue_head(&waitq);
 455                         lwi = LWI_TIMEOUT(cfs_time_seconds(min(cycle, 10)), NULL,
 456                                           NULL);
 457                         l_wait_event(waitq, 0, &lwi);
 458                 }
 459
 460                 if (rc < 0 || cycle % 10 == 2) {
 461                         spin_lock(&last_print_lock);
 462                         if (last_print == 0 ||
 463                             cfs_time_before((last_print + cfs_time_seconds(30)),
 464                                             cfs_time_current())) {
 465                                 last_print = cfs_time_current();
 466                                 spin_unlock(&last_print_lock);
 467                                 CWARN("still haven't managed to acquire quota "
 468                                       "space from the quota master after %d "
 469                                       "retries (err=%d, rc=%d)\n",
 470                                       cycle, count_err - 1, rc);
 471                         } else {
 472                                 spin_unlock(&last_print_lock);
 473                         }
 474                 }
 475
 476                 CDEBUG(D_QUOTA, "recheck quota with rc: %d, cycle: %d\n", rc,
 477                        cycle);
 478         }
 479         do_gettimeofday(&work_end);
 480         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
 481         lprocfs_counter_add(qctxt->lqc_stats,
 482                             isblk ? LQUOTA_WAIT_FOR_CHK_BLK :
 483                                     LQUOTA_WAIT_FOR_CHK_INO,
 484                             timediff);
 485
 486         RETURN(rc);
 487 }
 488
 489 /**
 490  * when a block_write or inode_create rpc is finished, adjust the record for
 491  * pending blocks and inodes
 492  */
 493 static int quota_pending_commit(struct obd_device *obd, const unsigned int id[],
 494                                 int pending[], int isblk)
 495 {
 496         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
 497         struct timeval work_start;
 498         struct timeval work_end;
 499         long timediff;
 500         int i;
 501         struct qunit_data qdata[MAXQUOTAS];
 502         ENTRY;
 503
 504         CDEBUG(D_QUOTA, "commit pending quota for  %s\n", obd->obd_name);
 505         CLASSERT(MAXQUOTAS < 4);
 506         if (!sb_any_quota_enabled(qctxt->lqc_sb))
 507                 RETURN(0);
 508
 509         do_gettimeofday(&work_start);
 510         for (i = 0; i < MAXQUOTAS; i++) {
 511                 struct lustre_qunit_size *lqs = NULL;
 512
 513                 LASSERT(pending[i] >= 0);
 514                 if (pending[i] == 0)
 515                         continue;
 516
 517                 qdata[i].qd_id = id[i];
 518                 qdata[i].qd_flags = i;
 519                 if (isblk)
 520                         QDATA_SET_BLK(&qdata[i]);
 521                 qdata[i].qd_count = 0;
 522
 523                 if (qdata[i].qd_id == 0 && !QDATA_IS_GRP(&qdata[i]))
 524                         continue;
 525
 526                 quota_search_lqs(&qdata[i], NULL, qctxt, &lqs);
 527                 if (lqs == NULL || IS_ERR(lqs)) {
 528                         CERROR("can not find lqs for pending_commit: "
 529                                "[id %u] [%c] [pending %u] [isblk %d] (rc %ld), "
 530                                "maybe cause unexpected lqs refcount error!\n",
 531                                id[i], i % 2 ? 'g': 'u', pending[i], isblk,
 532                                lqs ? PTR_ERR(lqs) : -1);
 533                         continue;
 534                 }
 535
 536                 spin_lock(&lqs->lqs_lock);
 537                 if (isblk) {
 538                         LASSERTF(lqs->lqs_bwrite_pending >= pending[i],
 539                                  "there are too many blocks! [id %u] [%c] "
 540                                  "[bwrite_pending %lu] [pending %u]\n",
 541                                  id[i], i % 2 ? 'g' : 'u',
 542                                  lqs->lqs_bwrite_pending, pending[i]);
 543
 544                         lqs->lqs_bwrite_pending -= pending[i];
 545                 } else {
 546                         LASSERTF(lqs->lqs_iwrite_pending >= pending[i],
 547                                 "there are too many files! [id %u] [%c] "
 548                                 "[iwrite_pending %lu] [pending %u]\n",
 549                                 id[i], i % 2 ? 'g' : 'u',
 550                                 lqs->lqs_iwrite_pending, pending[i]);
 551
 552                         lqs->lqs_iwrite_pending -= pending[i];
 553                 }
 554                 CDEBUG(D_QUOTA, "id: %u, %c, lqs pending: %lu, pending: %d, "
 555                        "isblk: %d.\n", id[i], i % 2 ? 'g' : 'u',
 556                        isblk ? lqs->lqs_bwrite_pending: lqs->lqs_iwrite_pending,
 557                        pending[i], isblk);
 558                 spin_unlock(&lqs->lqs_lock);
 559
 560                 /* for quota_search_lqs in pending_commit */
 561                 lqs_putref(lqs);
 562                 /* for quota_search_lqs in quota_check */
 563                 lqs_putref(lqs);
 564         }
 565         do_gettimeofday(&work_end);
 566         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
 567         lprocfs_counter_add(qctxt->lqc_stats,
 568                             isblk ? LQUOTA_WAIT_FOR_COMMIT_BLK :
 569                                     LQUOTA_WAIT_FOR_COMMIT_INO,
 570                             timediff);
 571
 572         RETURN(0);
 573 }
 574
 575 static int mds_quota_init(void)
 576 {
 577         return lustre_dquot_init();
 578 }
 579
 580 static int mds_quota_exit(void)
 581 {
 582         lustre_dquot_exit();
 583         return 0;
 584 }
 585
 586 static int mds_quota_setup(struct obd_device *obd)
 587 {
 588         struct obd_device_target *obt = &obd->u.obt;
 589         struct mds_obd *mds = &obd->u.mds;
 590         int rc;
 591         ENTRY;
 592
 593         if (unlikely(mds->mds_quota)) {
 594                 CWARN("try to reinitialize quota context!\n");
 595                 RETURN(0);
 596         }
 597
 598         init_rwsem(&obt->obt_rwsem);
 599         obt->obt_qfmt = LUSTRE_QUOTA_V2;
 600         mds->mds_quota_info.qi_version = LUSTRE_QUOTA_V2;
 601         atomic_set(&obt->obt_quotachecking, 1);
 602         /* initialize quota master and quota context */
 603         sema_init(&mds->mds_qonoff_sem, 1);
 604         rc = qctxt_init(obd, dqacq_handler);
 605         if (rc) {
 606                 CERROR("initialize quota context failed! (rc:%d)\n", rc);
 607                 RETURN(rc);
 608         }
 609         mds->mds_quota = 1;
 610         RETURN(rc);
 611 }
 612
 613 static int mds_quota_cleanup(struct obd_device *obd)
 614 {
 615         ENTRY;
 616         if (unlikely(!obd->u.mds.mds_quota))
 617                 RETURN(0);
 618
 619         qctxt_cleanup(&obd->u.obt.obt_qctxt, 0);
 620         RETURN(0);
 621 }
 622
 623 static int mds_quota_setinfo(struct obd_device *obd, void *data)
 624 {
 625         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
 626         ENTRY;
 627
 628         if (unlikely(!obd->u.mds.mds_quota))
 629                 RETURN(0);
 630
 631         if (data != NULL)
 632                 QUOTA_MASTER_READY(qctxt);
 633         else
 634                 QUOTA_MASTER_UNREADY(qctxt);
 635         RETURN(0);
 636 }
 637
 638 static int mds_quota_fs_cleanup(struct obd_device *obd)
 639 {
 640         struct mds_obd *mds = &obd->u.mds;
 641         struct obd_quotactl oqctl;
 642         ENTRY;
 643
 644         if (unlikely(!mds->mds_quota))
 645                 RETURN(0);
 646
 647         mds->mds_quota = 0;
 648         memset(&oqctl, 0, sizeof(oqctl));
 649         oqctl.qc_type = UGQUOTA;
 650
 651         down(&mds->mds_qonoff_sem);
 652         mds_admin_quota_off(obd, &oqctl);
 653         up(&mds->mds_qonoff_sem);
 654         RETURN(0);
 655 }
 656
 657 static int quota_acquire_common(struct obd_device *obd, const unsigned int id[],
 658                                 struct obd_trans_info *oti, int isblk)
 659 {
 660         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
 661         int rc;
 662         ENTRY;
 663
 664         rc = qctxt_adjust_qunit(obd, qctxt, id, isblk, 1, oti);
 665         RETURN(rc);
 666 }
 667
 668 #endif /* HAVE_QUOTA_SUPPORT */
 669 #endif /* __KERNEL__ */
 670
 671 struct osc_quota_info {
 672         struct list_head        oqi_hash;       /* hash list */
 673         struct client_obd      *oqi_cli;        /* osc obd */
 674         unsigned int            oqi_id;         /* uid/gid of a file */
 675         short                   oqi_type;       /* quota type */
 676 };
 677
 678 spinlock_t qinfo_list_lock = SPIN_LOCK_UNLOCKED;
 679
 680 static struct list_head qinfo_hash[NR_DQHASH];
 681 /* SLAB cache for client quota context */
 682 cfs_mem_cache_t *qinfo_cachep = NULL;
 683
 684 static inline int hashfn(struct client_obd *cli, unsigned long id, int type)
 685                          __attribute__((__const__));
 686
 687 static inline int hashfn(struct client_obd *cli, unsigned long id, int type)
 688 {
 689         unsigned long tmp = ((unsigned long)cli>>6) ^ id;
 690         tmp = (tmp * (MAXQUOTAS - type)) % NR_DQHASH;
 691         return tmp;
 692 }
 693
 694 /* caller must hold qinfo_list_lock */
 695 static inline void insert_qinfo_hash(struct osc_quota_info *oqi)
 696 {
 697         struct list_head *head = qinfo_hash +
 698                 hashfn(oqi->oqi_cli, oqi->oqi_id, oqi->oqi_type);
 699
 700         LASSERT_SPIN_LOCKED(&qinfo_list_lock);
 701         list_add(&oqi->oqi_hash, head);
 702 }
 703
 704 /* caller must hold qinfo_list_lock */
 705 static inline void remove_qinfo_hash(struct osc_quota_info *oqi)
 706 {
 707         LASSERT_SPIN_LOCKED(&qinfo_list_lock);
 708         list_del_init(&oqi->oqi_hash);
 709 }
 710
 711 /* caller must hold qinfo_list_lock */
 712 static inline struct osc_quota_info *find_qinfo(struct client_obd *cli,
 713                                                 unsigned int id, int type)
 714 {
 715         unsigned int hashent = hashfn(cli, id, type);
 716         struct osc_quota_info *oqi;
 717         ENTRY;
 718
 719         LASSERT_SPIN_LOCKED(&qinfo_list_lock);
 720         list_for_each_entry(oqi, &qinfo_hash[hashent], oqi_hash) {
 721                 if (oqi->oqi_cli == cli &&
 722                     oqi->oqi_id == id && oqi->oqi_type == type)
 723                         return oqi;
 724         }
 725         RETURN(NULL);
 726 }
 727
 728 static struct osc_quota_info *alloc_qinfo(struct client_obd *cli,
 729                                           unsigned int id, int type)
 730 {
 731         struct osc_quota_info *oqi;
 732         ENTRY;
 733
 734         OBD_SLAB_ALLOC(oqi, qinfo_cachep, CFS_ALLOC_STD, sizeof(*oqi));
 735         if(!oqi)
 736                 RETURN(NULL);
 737
 738         CFS_INIT_LIST_HEAD(&oqi->oqi_hash);
 739         oqi->oqi_cli = cli;
 740         oqi->oqi_id = id;
 741         oqi->oqi_type = type;
 742
 743         RETURN(oqi);
 744 }
 745
 746 static void free_qinfo(struct osc_quota_info *oqi)
 747 {
 748         OBD_SLAB_FREE(oqi, qinfo_cachep, sizeof(*oqi));
 749 }
 750
 751 int osc_quota_chkdq(struct client_obd *cli, const unsigned int qid[])
 752 {
 753         unsigned int id;
 754         int cnt, rc = QUOTA_OK;
 755         ENTRY;
 756
 757         spin_lock(&qinfo_list_lock);
 758         for (cnt = 0; cnt < MAXQUOTAS; cnt++) {
 759                 struct osc_quota_info *oqi = NULL;
 760
 761                 id = (cnt == USRQUOTA) ? qid[USRQUOTA] : qid[GRPQUOTA];
 762                 oqi = find_qinfo(cli, id, cnt);
 763                 if (oqi) {
 764                         rc = NO_QUOTA;
 765                         break;
 766                 }
 767         }
 768         spin_unlock(&qinfo_list_lock);
 769
 770         RETURN(rc);
 771 }
 772
 773 int osc_quota_setdq(struct client_obd *cli, const unsigned int qid[],
 774                     obd_flag valid, obd_flag flags)
 775 {
 776         unsigned int id;
 777         obd_flag noquota;
 778         int cnt, rc = 0;
 779         ENTRY;
 780
 781
 782         for (cnt = 0; cnt < MAXQUOTAS; cnt++) {
 783                 struct osc_quota_info *oqi, *old;
 784
 785                 if (!(valid & ((cnt == USRQUOTA) ?
 786                     OBD_MD_FLUSRQUOTA : OBD_MD_FLGRPQUOTA)))
 787                         continue;
 788
 789                 id = (cnt == USRQUOTA) ? qid[USRQUOTA] : qid[GRPQUOTA];
 790                 noquota = (cnt == USRQUOTA) ?
 791                     (flags & OBD_FL_NO_USRQUOTA) : (flags & OBD_FL_NO_GRPQUOTA);
 792
 793                 oqi = alloc_qinfo(cli, id, cnt);
 794                 if (oqi) {
 795                         spin_lock(&qinfo_list_lock);
 796
 797                         old = find_qinfo(cli, id, cnt);
 798                         if (old && !noquota)
 799                                 remove_qinfo_hash(old);
 800                         else if (!old && noquota)
 801                                 insert_qinfo_hash(oqi);
 802
 803                         spin_unlock(&qinfo_list_lock);
 804
 805                         if (old || !noquota)
 806                                 free_qinfo(oqi);
 807                         if (old && !noquota)
 808                                 free_qinfo(old);
 809                 } else {
 810                         CERROR("not enough mem!\n");
 811                         rc = -ENOMEM;
 812                         break;
 813                 }
 814         }
 815
 816         RETURN(rc);
 817 }
 818
 819 int osc_quota_cleanup(struct obd_device *obd)
 820 {
 821         struct client_obd *cli = &obd->u.cli;
 822         struct osc_quota_info *oqi, *n;
 823         int i;
 824         ENTRY;
 825
 826         spin_lock(&qinfo_list_lock);
 827         for (i = 0; i < NR_DQHASH; i++) {
 828                 list_for_each_entry_safe(oqi, n, &qinfo_hash[i], oqi_hash) {
 829                         if (oqi->oqi_cli != cli)
 830                                 continue;
 831                         remove_qinfo_hash(oqi);
 832                         free_qinfo(oqi);
 833                 }
 834         }
 835         spin_unlock(&qinfo_list_lock);
 836
 837         RETURN(0);
 838 }
 839
 840 int osc_quota_init(void)
 841 {
 842         int i;
 843         ENTRY;
 844
 845         LASSERT(qinfo_cachep == NULL);
 846         qinfo_cachep = cfs_mem_cache_create("osc_quota_info",
 847                                             sizeof(struct osc_quota_info),
 848                                             0, 0);
 849         if (!qinfo_cachep)
 850                 RETURN(-ENOMEM);
 851
 852         for (i = 0; i < NR_DQHASH; i++)
 853                 CFS_INIT_LIST_HEAD(qinfo_hash + i);
 854
 855         RETURN(0);
 856 }
 857
 858 int osc_quota_exit(void)
 859 {
 860         struct osc_quota_info *oqi, *n;
 861         int i, rc;
 862         ENTRY;
 863
 864         spin_lock(&qinfo_list_lock);
 865         for (i = 0; i < NR_DQHASH; i++) {
 866                 list_for_each_entry_safe(oqi, n, &qinfo_hash[i], oqi_hash) {
 867                         remove_qinfo_hash(oqi);
 868                         free_qinfo(oqi);
 869                 }
 870         }
 871         spin_unlock(&qinfo_list_lock);
 872
 873         rc = cfs_mem_cache_destroy(qinfo_cachep);
 874         LASSERTF(rc == 0, "couldn't destory qinfo_cachep slab\n");
 875         qinfo_cachep = NULL;
 876
 877         RETURN(0);
 878 }
 879
 880 #ifdef __KERNEL__
 881 #ifdef HAVE_QUOTA_SUPPORT
 882 quota_interface_t mds_quota_interface = {
 883         .quota_init     = mds_quota_init,
 884         .quota_exit     = mds_quota_exit,
 885         .quota_setup    = mds_quota_setup,
 886         .quota_cleanup  = mds_quota_cleanup,
 887         .quota_check    = target_quota_check,
 888         .quota_ctl      = mds_quota_ctl,
 889         .quota_setinfo  = mds_quota_setinfo,
 890         .quota_fs_cleanup = mds_quota_fs_cleanup,
 891         .quota_recovery = mds_quota_recovery,
 892         .quota_adjust   = mds_quota_adjust,
 893         .quota_chkquota = quota_chk_acq_common,
 894         .quota_acquire  = quota_acquire_common,
 895         .quota_pending_commit = quota_pending_commit,
 896 };
 897
 898 quota_interface_t filter_quota_interface = {
 899         .quota_setup    = filter_quota_setup,
 900         .quota_cleanup  = filter_quota_cleanup,
 901         .quota_check    = target_quota_check,
 902         .quota_ctl      = filter_quota_ctl,
 903         .quota_setinfo  = filter_quota_setinfo,
 904         .quota_clearinfo = filter_quota_clearinfo,
 905         .quota_enforce  = filter_quota_enforce,
 906         .quota_getflag  = filter_quota_getflag,
 907         .quota_acquire  = quota_acquire_common,
 908         .quota_adjust   = filter_quota_adjust,
 909         .quota_chkquota = quota_chk_acq_common,
 910         .quota_adjust_qunit   = filter_quota_adjust_qunit,
 911         .quota_pending_commit = quota_pending_commit,
 912 };
 913 #endif
 914 #endif /* __KERNEL__ */
 915
 916 quota_interface_t mdc_quota_interface = {
 917         .quota_ctl      = client_quota_ctl,
 918         .quota_check    = client_quota_check,
 919         .quota_poll_check = client_quota_poll_check,
 920 };
 921
 922 quota_interface_t lmv_quota_interface = {
 923         .quota_ctl      = lmv_quota_ctl,
 924         .quota_check    = lmv_quota_check,
 925 };
 926
 927 quota_interface_t osc_quota_interface = {
 928         .quota_ctl      = client_quota_ctl,
 929         .quota_check    = client_quota_check,
 930         .quota_poll_check = client_quota_poll_check,
 931         .quota_init     = osc_quota_init,
 932         .quota_exit     = osc_quota_exit,
 933         .quota_chkdq    = osc_quota_chkdq,
 934         .quota_setdq    = osc_quota_setdq,
 935         .quota_cleanup  = osc_quota_cleanup,
 936         .quota_adjust_qunit = client_quota_adjust_qunit,
 937 };
 938
 939 quota_interface_t lov_quota_interface = {
 940         .quota_ctl      = lov_quota_ctl,
 941         .quota_check    = lov_quota_check,
 942         .quota_adjust_qunit = lov_quota_adjust_qunit,
 943 };
 944
 945 #ifdef __KERNEL__
 946
 947 cfs_proc_dir_entry_t *lquota_type_proc_dir = NULL;
 948
 949 static int __init init_lustre_quota(void)
 950 {
 951 #ifdef HAVE_QUOTA_SUPPORT
 952         int rc = 0;
 953
 954         lquota_type_proc_dir = lprocfs_register(OBD_LQUOTA_DEVICENAME,
 955                                                 proc_lustre_root,
 956                                                 NULL, NULL);
 957         if (IS_ERR(lquota_type_proc_dir)) {
 958                 CERROR("LProcFS failed in lquota-init\n");
 959                 rc = PTR_ERR(lquota_type_proc_dir);
 960                 return rc;
 961         }
 962
 963         rc = qunit_cache_init();
 964         if (rc)
 965                 return rc;
 966
 967         PORTAL_SYMBOL_REGISTER(filter_quota_interface);
 968         PORTAL_SYMBOL_REGISTER(mds_quota_interface);
 969 #endif
 970         PORTAL_SYMBOL_REGISTER(mdc_quota_interface);
 971         PORTAL_SYMBOL_REGISTER(lmv_quota_interface);
 972         PORTAL_SYMBOL_REGISTER(osc_quota_interface);
 973         PORTAL_SYMBOL_REGISTER(lov_quota_interface);
 974         return 0;
 975 }
 976
 977 static void /*__exit*/ exit_lustre_quota(void)
 978 {
 979         PORTAL_SYMBOL_UNREGISTER(mdc_quota_interface);
 980         PORTAL_SYMBOL_UNREGISTER(lmv_quota_interface);
 981         PORTAL_SYMBOL_UNREGISTER(osc_quota_interface);
 982         PORTAL_SYMBOL_UNREGISTER(lov_quota_interface);
 983 #ifdef HAVE_QUOTA_SUPPORT
 984         PORTAL_SYMBOL_UNREGISTER(filter_quota_interface);
 985         PORTAL_SYMBOL_UNREGISTER(mds_quota_interface);
 986
 987         qunit_cache_cleanup();
 988
 989         if (lquota_type_proc_dir)
 990                 lprocfs_remove(&lquota_type_proc_dir);
 991 #endif
 992 }
 993
 994 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
 995 MODULE_DESCRIPTION("Lustre Quota");
 996 MODULE_LICENSE("GPL");
 997
 998 cfs_module(lquota, "1.0.0", init_lustre_quota, exit_lustre_quota);
 999
1000 #ifdef HAVE_QUOTA_SUPPORT
1001 EXPORT_SYMBOL(mds_quota_interface);
1002 EXPORT_SYMBOL(filter_quota_interface);
1003 #endif
1004 EXPORT_SYMBOL(mdc_quota_interface);
1005 EXPORT_SYMBOL(lmv_quota_interface);
1006 EXPORT_SYMBOL(osc_quota_interface);
1007 EXPORT_SYMBOL(lov_quota_interface);
1008 #endif /* __KERNEL */