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         sema_init(&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                         CERROR("%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 (!ll_sb_any_quota_active(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 (!ll_sb_any_quota_active(obt->obt_sb))
 190                 RETURN(0);
 191
 192         OBD_ALLOC_PTR(oqctl);
 193         if (!oqctl)
 194                 RETURN(-ENOMEM);
 195
 196         /* set over quota flags for a uid/gid */
 197         oa->o_valid |= OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA;
 198         oa->o_flags &= ~(OBD_FL_NO_USRQUOTA | OBD_FL_NO_GRPQUOTA);
 199
 200         for (cnt = 0; cnt < MAXQUOTAS; cnt++) {
 201                 struct lustre_qunit_size *lqs = NULL;
 202
 203                 lqs = quota_search_lqs(LQS_KEY(cnt, GET_OA_ID(cnt, oa)),
 204                                        qctxt, 0);
 205                 if (lqs == NULL || IS_ERR(lqs)) {
 206                         rc = PTR_ERR(lqs);
 207                         if (rc)
 208                                 CDEBUG(D_QUOTA, "search lqs for %s %d failed, "
 209                                        "(rc = %d)\n",
 210                                        cnt == USRQUOTA ? "user" : "group",
 211                                        cnt == USRQUOTA ? oa->o_uid : oa->o_gid,
 212                                        rc);
 213                         break;
 214                 } else {
 215                         spin_lock(&lqs->lqs_lock);
 216                         if (lqs->lqs_bunit_sz <= qctxt->lqc_sync_blk) {
 217                                 oa->o_flags |= (cnt == USRQUOTA) ?
 218                                         OBD_FL_NO_USRQUOTA : OBD_FL_NO_GRPQUOTA;
 219                                 spin_unlock(&lqs->lqs_lock);
 220                                 CDEBUG(D_QUOTA, "set sync flag: bunit(%lu), "
 221                                        "sync_blk(%d)\n", lqs->lqs_bunit_sz,
 222                                        qctxt->lqc_sync_blk);
 223                                 /* this is for quota_search_lqs */
 224                                 lqs_putref(lqs);
 225                                 continue;
 226                         }
 227                         spin_unlock(&lqs->lqs_lock);
 228                         /* this is for quota_search_lqs */
 229                         lqs_putref(lqs);
 230                 }
 231
 232                 memset(oqctl, 0, sizeof(*oqctl));
 233
 234                 oqctl->qc_cmd = Q_GETQUOTA;
 235                 oqctl->qc_type = cnt;
 236                 oqctl->qc_id = (cnt == USRQUOTA) ? oa->o_uid : oa->o_gid;
 237                 err = fsfilt_quotactl(obd, obt->obt_sb, oqctl);
 238                 if (err) {
 239                         if (!rc)
 240                                 rc = err;
 241                         oa->o_valid &= ~((cnt == USRQUOTA) ? OBD_MD_FLUSRQUOTA :
 242                                                              OBD_MD_FLGRPQUOTA);
 243                         CDEBUG(D_QUOTA, "fsfilt getquota for %s %d failed, "
 244                                "(rc = %d)\n",
 245                                cnt == USRQUOTA ? "user" : "group",
 246                                cnt == USRQUOTA ? oa->o_uid : oa->o_gid, err);
 247                         continue;
 248                 }
 249
 250                 if (oqctl->qc_dqblk.dqb_bhardlimit &&
 251                    (toqb(oqctl->qc_dqblk.dqb_curspace) >=
 252                     oqctl->qc_dqblk.dqb_bhardlimit)) {
 253                         oa->o_flags |= (cnt == USRQUOTA) ?
 254                                 OBD_FL_NO_USRQUOTA : OBD_FL_NO_GRPQUOTA;
 255                         CDEBUG(D_QUOTA, "out of quota for %s %d\n",
 256                                cnt == USRQUOTA ? "user" : "group",
 257                                cnt == USRQUOTA ? oa->o_uid : oa->o_gid);
 258                 }
 259         }
 260         OBD_FREE_PTR(oqctl);
 261         RETURN(rc);
 262 }
 263
 264 /**
 265  * check whether the left quota of certain uid and gid can satisfy a block_write
 266  * or inode_create rpc. When need to acquire quota, return QUOTA_RET_ACQUOTA
 267  */
 268 static int quota_check_common(struct obd_device *obd, const unsigned int id[],
 269                               int pending[], int count, int cycle, int isblk,
 270                               struct inode *inode, int frags)
 271 {
 272         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
 273         int i;
 274         struct qunit_data qdata[MAXQUOTAS];
 275         int mb = 0;
 276         int rc = 0, rc2[2] = { 0, 0 };
 277         ENTRY;
 278
 279         spin_lock(&qctxt->lqc_lock);
 280         if (!qctxt->lqc_valid){
 281                 spin_unlock(&qctxt->lqc_lock);
 282                 RETURN(rc);
 283         }
 284         spin_unlock(&qctxt->lqc_lock);
 285
 286         for (i = 0; i < MAXQUOTAS; i++) {
 287                 struct lustre_qunit_size *lqs = NULL;
 288
 289                 qdata[i].qd_id = id[i];
 290                 qdata[i].qd_flags = i;
 291                 if (isblk)
 292                         QDATA_SET_BLK(&qdata[i]);
 293                 qdata[i].qd_count = 0;
 294
 295                 /* ignore root user */
 296                 if (qdata[i].qd_id == 0 && !QDATA_IS_GRP(&qdata[i]))
 297                         continue;
 298
 299                 lqs = quota_search_lqs(LQS_KEY(i, id[i]), qctxt, 0);
 300                 if (lqs == NULL || IS_ERR(lqs))
 301                         continue;
 302
 303                 if (IS_ERR(lqs)) {
 304                         CERROR("can not find lqs for check_common: "
 305                                "[id %u] [%c] [isblk %d] [count %d] [rc %ld]\n",
 306                                id[i], i % 2 ? 'g': 'u', isblk, count,
 307                                PTR_ERR(lqs));
 308                         RETURN(PTR_ERR(lqs));
 309                 }
 310
 311                 rc2[i] = compute_remquota(obd, qctxt, &qdata[i], isblk);
 312                 spin_lock(&lqs->lqs_lock);
 313                 if (!cycle) {
 314                         if (isblk) {
 315                                 pending[i] = count * CFS_PAGE_SIZE;
 316                                 /* in order to complete this write, we need extra
 317                                  * meta blocks. This function can get it through
 318                                  * data needed to be written b=16542 */
 319                                 if (inode) {
 320                                         mb = pending[i];
 321                                         rc = fsfilt_get_mblk(obd, qctxt->lqc_sb,
 322                                                              &mb, inode,frags);
 323                                         if (rc)
 324                                                 CERROR("%s: can't get extra "
 325                                                        "meta blocks\n",
 326                                                        obd->obd_name);
 327                                         else
 328                                                 pending[i] += mb;
 329                                 }
 330                                 lqs->lqs_bwrite_pending += pending[i];
 331                         } else {
 332                                 pending[i] = count;
 333                                 lqs->lqs_iwrite_pending += pending[i];
 334                         }
 335                 }
 336
 337                 /* if xx_rec < 0, that means quota are releasing,
 338                  * and it may return before we use quota. So if
 339                  * we find this situation, we assuming it has
 340                  * returned b=18491 */
 341                 if (isblk && lqs->lqs_blk_rec < 0) {
 342                         if (qdata[i].qd_count < -lqs->lqs_blk_rec)
 343                                 qdata[i].qd_count = 0;
 344                         else
 345                                 qdata[i].qd_count += lqs->lqs_blk_rec;
 346                 }
 347                 if (!isblk && lqs->lqs_ino_rec < 0) {
 348                         if (qdata[i].qd_count < -lqs->lqs_ino_rec)
 349                                 qdata[i].qd_count = 0;
 350                         else
 351                                 qdata[i].qd_count += lqs->lqs_ino_rec;
 352                 }
 353
 354                 CDEBUG(D_QUOTA, "[id %u] [%c] [isblk %d] [count %d]"
 355                        " [lqs pending: %lu] [qd_count: "LPU64"] [metablocks: %d]"
 356                        " [pending: %d]\n", id[i], i % 2 ? 'g': 'u', isblk, count,
 357                        isblk ? lqs->lqs_bwrite_pending : lqs->lqs_iwrite_pending,
 358                        qdata[i].qd_count, mb, pending[i]);
 359                 if (rc2[i] == QUOTA_RET_OK) {
 360                         if (isblk && qdata[i].qd_count < lqs->lqs_bwrite_pending)
 361                                 rc2[i] = QUOTA_RET_ACQUOTA;
 362                         if (!isblk && qdata[i].qd_count <
 363                             lqs->lqs_iwrite_pending)
 364                                 rc2[i] = QUOTA_RET_ACQUOTA;
 365                 }
 366
 367                 spin_unlock(&lqs->lqs_lock);
 368
 369                 if (lqs->lqs_blk_rec  < 0 &&
 370                     qdata[i].qd_count <
 371                     lqs->lqs_bwrite_pending - lqs->lqs_blk_rec - mb)
 372                         OBD_FAIL_TIMEOUT(OBD_FAIL_QUOTA_DELAY_REL, 5);
 373
 374                 /* When cycle is zero, lqs_*_pending will be changed. We will
 375                  * get reference of the lqs here and put reference of lqs in
 376                  * quota_pending_commit b=14784 */
 377                 if (!cycle)
 378                         lqs_getref(lqs);
 379
 380                 /* this is for quota_search_lqs */
 381                 lqs_putref(lqs);
 382         }
 383
 384         if (rc2[0] == QUOTA_RET_ACQUOTA || rc2[1] == QUOTA_RET_ACQUOTA)
 385                 RETURN(QUOTA_RET_ACQUOTA);
 386         else
 387                 RETURN(rc);
 388 }
 389
 390 int quota_is_set(struct obd_device *obd, const unsigned int id[], int flag)
 391 {
 392         struct lustre_qunit_size *lqs;
 393         int i, q_set = 0;
 394
 395         if (!ll_sb_any_quota_active(obd->u.obt.obt_qctxt.lqc_sb))
 396                 RETURN(0);
 397
 398         for (i = 0; i < MAXQUOTAS; i++) {
 399                 lqs = quota_search_lqs(LQS_KEY(i, id[i]),
 400                                        &obd->u.obt.obt_qctxt, 0);
 401                 if (lqs && !IS_ERR(lqs)) {
 402                         if (lqs->lqs_flags & flag)
 403                                 q_set = 1;
 404                         lqs_putref(lqs);
 405                 }
 406         }
 407
 408         return q_set;
 409 }
 410
 411 static int quota_chk_acq_common(struct obd_device *obd, struct obd_export *exp,
 412                                 const unsigned int id[], int pending[],
 413                                 int count, quota_acquire acquire,
 414                                 struct obd_trans_info *oti, int isblk,
 415                                 struct inode *inode, int frags)
 416 {
 417         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
 418         struct timeval work_start;
 419         struct timeval work_end;
 420         long timediff;
 421         struct l_wait_info lwi = { 0 };
 422         int rc = 0, cycle = 0, count_err = 1;
 423         ENTRY;
 424
 425         if (!quota_is_set(obd, id, isblk ? QB_SET : QI_SET))
 426                 RETURN(0);
 427
 428         if (isblk && (exp->exp_failed || exp->exp_abort_active_req))
 429                 /* If the client has been evicted or if it
 430                  * timed out and tried to reconnect already,
 431                  * abort the request immediately */
 432                 RETURN(-ENOTCONN);
 433
 434         CDEBUG(D_QUOTA, "check quota for %s\n", obd->obd_name);
 435         pending[USRQUOTA] = pending[GRPQUOTA] = 0;
 436         /* Unfortunately, if quota master is too busy to handle the
 437          * pre-dqacq in time and quota hash on ost is used up, we
 438          * have to wait for the completion of in flight dqacq/dqrel,
 439          * in order to get enough quota for write b=12588 */
 440         do_gettimeofday(&work_start);
 441         while ((rc = quota_check_common(obd, id, pending, count, cycle, isblk,
 442                                         inode, frags)) &
 443                QUOTA_RET_ACQUOTA) {
 444
 445                 spin_lock(&qctxt->lqc_lock);
 446                 if (!qctxt->lqc_import && oti) {
 447                         spin_unlock(&qctxt->lqc_lock);
 448
 449                         LASSERT(oti && oti->oti_thread &&
 450                                 oti->oti_thread->t_watchdog);
 451
 452                         lc_watchdog_disable(oti->oti_thread->t_watchdog);
 453                         CDEBUG(D_QUOTA, "sleep for quota master\n");
 454                         l_wait_event(qctxt->lqc_wait_for_qmaster, check_qm(qctxt),
 455                                      &lwi);
 456                         CDEBUG(D_QUOTA, "wake up when quota master is back\n");
 457                         lc_watchdog_touch(oti->oti_thread->t_watchdog,
 458                                  GET_TIMEOUT(oti->oti_thread->t_svc));
 459                 } else {
 460                         spin_unlock(&qctxt->lqc_lock);
 461                 }
 462
 463                 cycle++;
 464                 if (isblk)
 465                         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_HOLD_WRITE_RPC, 90);
 466                 /* after acquire(), we should run quota_check_common again
 467                  * so that we confirm there are enough quota to finish write */
 468                 rc = acquire(obd, id, oti, isblk);
 469
 470                 /* please reference to dqacq_completion for the below */
 471                 /* a new request is finished, try again */
 472                 if (rc == QUOTA_REQ_RETURNED) {
 473                         CDEBUG(D_QUOTA, "finish a quota req, try again\n");
 474                         continue;
 475                 }
 476
 477                 /* it is out of quota already */
 478                 if (rc == -EDQUOT) {
 479                         CDEBUG(D_QUOTA, "out of quota,  return -EDQUOT\n");
 480                         break;
 481                 }
 482
 483                 /* Related quota has been disabled by master, but enabled by
 484                  * slave, do not try again. */
 485                 if (unlikely(rc == -ESRCH)) {
 486                         CERROR("mismatched quota configuration, stop try.\n");
 487                         break;
 488                 }
 489
 490                 if (isblk && (exp->exp_failed || exp->exp_abort_active_req))
 491                         /* The client has been evicted or tried to
 492                          * to reconnect already, abort the request */
 493                         RETURN(-ENOTCONN);
 494
 495                 /* -EBUSY and others, wait a second and try again */
 496                 if (rc < 0) {
 497                         cfs_waitq_t        waitq;
 498                         struct l_wait_info lwi;
 499
 500                         if (oti && oti->oti_thread && oti->oti_thread->t_watchdog)
 501                                 lc_watchdog_touch(oti->oti_thread->t_watchdog,
 502                                          GET_TIMEOUT(oti->oti_thread->t_svc));
 503                         CDEBUG(D_QUOTA, "rc: %d, count_err: %d\n", rc,
 504                                count_err++);
 505
 506                         init_waitqueue_head(&waitq);
 507                         lwi = LWI_TIMEOUT(cfs_time_seconds(min(cycle, 10)), NULL,
 508                                           NULL);
 509                         l_wait_event(waitq, 0, &lwi);
 510                 }
 511
 512                 if (rc < 0 || cycle % 10 == 0) {
 513                         spin_lock(&last_print_lock);
 514                         if (last_print == 0 ||
 515                             cfs_time_before((last_print + cfs_time_seconds(30)),
 516                                             cfs_time_current())) {
 517                                 last_print = cfs_time_current();
 518                                 spin_unlock(&last_print_lock);
 519                                 CWARN("still haven't managed to acquire quota "
 520                                       "space from the quota master after %d "
 521                                       "retries (err=%d, rc=%d)\n",
 522                                       cycle, count_err - 1, rc);
 523                         } else {
 524                                 spin_unlock(&last_print_lock);
 525                         }
 526                 }
 527
 528                 CDEBUG(D_QUOTA, "recheck quota with rc: %d, cycle: %d\n", rc,
 529                        cycle);
 530         }
 531         do_gettimeofday(&work_end);
 532         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
 533         lprocfs_counter_add(qctxt->lqc_stats,
 534                             isblk ? LQUOTA_WAIT_FOR_CHK_BLK :
 535                                     LQUOTA_WAIT_FOR_CHK_INO,
 536                             timediff);
 537
 538         if (rc > 0)
 539                 rc = 0;
 540         RETURN(rc);
 541 }
 542
 543 /**
 544  * when a block_write or inode_create rpc is finished, adjust the record for
 545  * pending blocks and inodes
 546  */
 547 static int quota_pending_commit(struct obd_device *obd, const unsigned int id[],
 548                                 int pending[], int isblk)
 549 {
 550         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
 551         struct timeval work_start;
 552         struct timeval work_end;
 553         long timediff;
 554         int i;
 555         struct qunit_data qdata[MAXQUOTAS];
 556         ENTRY;
 557
 558         CDEBUG(D_QUOTA, "commit pending quota for  %s\n", obd->obd_name);
 559         CLASSERT(MAXQUOTAS < 4);
 560         if (!ll_sb_any_quota_active(qctxt->lqc_sb))
 561                 RETURN(0);
 562
 563         do_gettimeofday(&work_start);
 564         for (i = 0; i < MAXQUOTAS; i++) {
 565                 struct lustre_qunit_size *lqs = NULL;
 566
 567                 LASSERT(pending[i] >= 0);
 568                 if (pending[i] == 0)
 569                         continue;
 570
 571                 qdata[i].qd_id = id[i];
 572                 qdata[i].qd_flags = i;
 573                 if (isblk)
 574                         QDATA_SET_BLK(&qdata[i]);
 575                 qdata[i].qd_count = 0;
 576
 577                 if (qdata[i].qd_id == 0 && !QDATA_IS_GRP(&qdata[i]))
 578                         continue;
 579
 580                 lqs = quota_search_lqs(LQS_KEY(i, qdata[i].qd_id), qctxt, 0);
 581                 if (lqs == NULL || IS_ERR(lqs)) {
 582                         CERROR("can not find lqs for pending_commit: "
 583                                "[id %u] [%c] [pending %u] [isblk %d] (rc %ld), "
 584                                "maybe cause unexpected lqs refcount error!\n",
 585                                id[i], i ? 'g': 'u', pending[i], isblk,
 586                                lqs ? PTR_ERR(lqs) : -1);
 587                         continue;
 588                 }
 589
 590                 spin_lock(&lqs->lqs_lock);
 591                 if (isblk) {
 592                         LASSERTF(lqs->lqs_bwrite_pending >= pending[i],
 593                                  "there are too many blocks! [id %u] [%c] "
 594                                  "[bwrite_pending %lu] [pending %u]\n",
 595                                  id[i], i % 2 ? 'g' : 'u',
 596                                  lqs->lqs_bwrite_pending, pending[i]);
 597
 598                         lqs->lqs_bwrite_pending -= pending[i];
 599                 } else {
 600                         LASSERTF(lqs->lqs_iwrite_pending >= pending[i],
 601                                 "there are too many files! [id %u] [%c] "
 602                                 "[iwrite_pending %lu] [pending %u]\n",
 603                                 id[i], i % 2 ? 'g' : 'u',
 604                                 lqs->lqs_iwrite_pending, pending[i]);
 605
 606                         lqs->lqs_iwrite_pending -= pending[i];
 607                 }
 608                 CDEBUG(D_QUOTA, "%s: lqs_pending=%lu pending[%d]=%d isblk=%d\n",
 609                        obd->obd_name,
 610                        isblk ? lqs->lqs_bwrite_pending : lqs->lqs_iwrite_pending,
 611                        i, pending[i], isblk);
 612                 spin_unlock(&lqs->lqs_lock);
 613
 614                 /* for quota_search_lqs in pending_commit */
 615                 lqs_putref(lqs);
 616                 /* for quota_search_lqs in quota_check */
 617                 lqs_putref(lqs);
 618         }
 619         do_gettimeofday(&work_end);
 620         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
 621         lprocfs_counter_add(qctxt->lqc_stats,
 622                             isblk ? LQUOTA_WAIT_FOR_COMMIT_BLK :
 623                                     LQUOTA_WAIT_FOR_COMMIT_INO,
 624                             timediff);
 625
 626         RETURN(0);
 627 }
 628
 629 static int mds_quota_init(void)
 630 {
 631         return lustre_dquot_init();
 632 }
 633
 634 static int mds_quota_exit(void)
 635 {
 636         lustre_dquot_exit();
 637         return 0;
 638 }
 639
 640 static int mds_quota_setup(struct obd_device *obd)
 641 {
 642         struct obd_device_target *obt = &obd->u.obt;
 643         struct mds_obd *mds = &obd->u.mds;
 644         int rc;
 645         ENTRY;
 646
 647         if (unlikely(mds->mds_quota)) {
 648                 CWARN("try to reinitialize quota context!\n");
 649                 RETURN(0);
 650         }
 651
 652         init_rwsem(&obt->obt_rwsem);
 653         obt->obt_qfmt = LUSTRE_QUOTA_V2;
 654         mds->mds_quota_info.qi_version = LUSTRE_QUOTA_V2;
 655         sema_init(&obt->obt_quotachecking, 1);
 656         /* initialize quota master and quota context */
 657         sema_init(&mds->mds_qonoff_sem, 1);
 658         rc = qctxt_init(obd, dqacq_handler);
 659         if (rc) {
 660                 CERROR("%s: initialize quota context failed! (rc:%d)\n",
 661                        obd->obd_name, rc);
 662                 RETURN(rc);
 663         }
 664         mds->mds_quota = 1;
 665         RETURN(rc);
 666 }
 667
 668 static int mds_quota_cleanup(struct obd_device *obd)
 669 {
 670         ENTRY;
 671         if (unlikely(!obd->u.mds.mds_quota))
 672                 RETURN(0);
 673
 674         qctxt_cleanup(&obd->u.obt.obt_qctxt, 0);
 675         RETURN(0);
 676 }
 677
 678 static int mds_quota_setinfo(struct obd_device *obd, void *data)
 679 {
 680         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
 681         ENTRY;
 682
 683         if (unlikely(!obd->u.mds.mds_quota))
 684                 RETURN(0);
 685
 686         if (data != NULL)
 687                 QUOTA_MASTER_READY(qctxt);
 688         else
 689                 QUOTA_MASTER_UNREADY(qctxt);
 690         RETURN(0);
 691 }
 692
 693 static int mds_quota_fs_cleanup(struct obd_device *obd)
 694 {
 695         struct mds_obd *mds = &obd->u.mds;
 696         struct obd_quotactl oqctl;
 697         ENTRY;
 698
 699         if (unlikely(!mds->mds_quota))
 700                 RETURN(0);
 701
 702         mds->mds_quota = 0;
 703         memset(&oqctl, 0, sizeof(oqctl));
 704         oqctl.qc_type = UGQUOTA;
 705
 706         down(&mds->mds_qonoff_sem);
 707         mds_admin_quota_off(obd, &oqctl);
 708         up(&mds->mds_qonoff_sem);
 709         RETURN(0);
 710 }
 711
 712 static int quota_acquire_common(struct obd_device *obd, const unsigned int id[],
 713                                 struct obd_trans_info *oti, int isblk)
 714 {
 715         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
 716         int rc;
 717         ENTRY;
 718
 719         rc = qctxt_adjust_qunit(obd, qctxt, id, isblk, 1, oti);
 720         RETURN(rc);
 721 }
 722
 723 #endif /* HAVE_QUOTA_SUPPORT */
 724 #endif /* __KERNEL__ */
 725
 726 struct osc_quota_info {
 727         struct list_head        oqi_hash;       /* hash list */
 728         struct client_obd      *oqi_cli;        /* osc obd */
 729         unsigned int            oqi_id;         /* uid/gid of a file */
 730         short                   oqi_type;       /* quota type */
 731 };
 732
 733 spinlock_t qinfo_list_lock = SPIN_LOCK_UNLOCKED;
 734
 735 static struct list_head qinfo_hash[NR_DQHASH];
 736 /* SLAB cache for client quota context */
 737 cfs_mem_cache_t *qinfo_cachep = NULL;
 738
 739 static inline int hashfn(struct client_obd *cli, unsigned long id, int type)
 740                          __attribute__((__const__));
 741
 742 static inline int hashfn(struct client_obd *cli, unsigned long id, int type)
 743 {
 744         unsigned long tmp = ((unsigned long)cli>>6) ^ id;
 745         tmp = (tmp * (MAXQUOTAS - type)) % NR_DQHASH;
 746         return tmp;
 747 }
 748
 749 /* caller must hold qinfo_list_lock */
 750 static inline void insert_qinfo_hash(struct osc_quota_info *oqi)
 751 {
 752         struct list_head *head = qinfo_hash +
 753                 hashfn(oqi->oqi_cli, oqi->oqi_id, oqi->oqi_type);
 754
 755         LASSERT_SPIN_LOCKED(&qinfo_list_lock);
 756         list_add(&oqi->oqi_hash, head);
 757 }
 758
 759 /* caller must hold qinfo_list_lock */
 760 static inline void remove_qinfo_hash(struct osc_quota_info *oqi)
 761 {
 762         LASSERT_SPIN_LOCKED(&qinfo_list_lock);
 763         list_del_init(&oqi->oqi_hash);
 764 }
 765
 766 /* caller must hold qinfo_list_lock */
 767 static inline struct osc_quota_info *find_qinfo(struct client_obd *cli,
 768                                                 unsigned int id, int type)
 769 {
 770         unsigned int hashent = hashfn(cli, id, type);
 771         struct osc_quota_info *oqi;
 772         ENTRY;
 773
 774         LASSERT_SPIN_LOCKED(&qinfo_list_lock);
 775         list_for_each_entry(oqi, &qinfo_hash[hashent], oqi_hash) {
 776                 if (oqi->oqi_cli == cli &&
 777                     oqi->oqi_id == id && oqi->oqi_type == type)
 778                         return oqi;
 779         }
 780         RETURN(NULL);
 781 }
 782
 783 static struct osc_quota_info *alloc_qinfo(struct client_obd *cli,
 784                                           unsigned int id, int type)
 785 {
 786         struct osc_quota_info *oqi;
 787         ENTRY;
 788
 789         OBD_SLAB_ALLOC(oqi, qinfo_cachep, CFS_ALLOC_STD, sizeof(*oqi));
 790         if(!oqi)
 791                 RETURN(NULL);
 792
 793         CFS_INIT_LIST_HEAD(&oqi->oqi_hash);
 794         oqi->oqi_cli = cli;
 795         oqi->oqi_id = id;
 796         oqi->oqi_type = type;
 797
 798         RETURN(oqi);
 799 }
 800
 801 static void free_qinfo(struct osc_quota_info *oqi)
 802 {
 803         OBD_SLAB_FREE(oqi, qinfo_cachep, sizeof(*oqi));
 804 }
 805
 806 int osc_quota_chkdq(struct client_obd *cli, const unsigned int qid[])
 807 {
 808         unsigned int id;
 809         int cnt, rc = QUOTA_OK;
 810         ENTRY;
 811
 812         spin_lock(&qinfo_list_lock);
 813         for (cnt = 0; cnt < MAXQUOTAS; cnt++) {
 814                 struct osc_quota_info *oqi = NULL;
 815
 816                 id = (cnt == USRQUOTA) ? qid[USRQUOTA] : qid[GRPQUOTA];
 817                 oqi = find_qinfo(cli, id, cnt);
 818                 if (oqi) {
 819                         rc = NO_QUOTA;
 820                         break;
 821                 }
 822         }
 823         spin_unlock(&qinfo_list_lock);
 824
 825         if (rc == NO_QUOTA)
 826                 CDEBUG(D_QUOTA, "chkdq found noquota for %s %d\n",
 827                        cnt == USRQUOTA ? "user" : "group", id);
 828         RETURN(rc);
 829 }
 830
 831 int osc_quota_setdq(struct client_obd *cli, const unsigned int qid[],
 832                     obd_flag valid, obd_flag flags)
 833 {
 834         unsigned int id;
 835         obd_flag noquota;
 836         int cnt, rc = 0;
 837         ENTRY;
 838
 839
 840         for (cnt = 0; cnt < MAXQUOTAS; cnt++) {
 841                 struct osc_quota_info *oqi, *old;
 842
 843                 if (!(valid & ((cnt == USRQUOTA) ?
 844                     OBD_MD_FLUSRQUOTA : OBD_MD_FLGRPQUOTA)))
 845                         continue;
 846
 847                 id = (cnt == USRQUOTA) ? qid[USRQUOTA] : qid[GRPQUOTA];
 848                 noquota = (cnt == USRQUOTA) ?
 849                     (flags & OBD_FL_NO_USRQUOTA) : (flags & OBD_FL_NO_GRPQUOTA);
 850
 851                 oqi = alloc_qinfo(cli, id, cnt);
 852                 if (!oqi) {
 853                         rc = -ENOMEM;
 854                         CDEBUG(D_QUOTA, "setdq for %s %d failed, (rc = %d)\n",
 855                                cnt == USRQUOTA ? "user" : "group", id, rc);
 856                         break;
 857                 }
 858
 859                 spin_lock(&qinfo_list_lock);
 860                 old = find_qinfo(cli, id, cnt);
 861                 if (old && !noquota)
 862                         remove_qinfo_hash(old);
 863                 else if (!old && noquota)
 864                         insert_qinfo_hash(oqi);
 865                 spin_unlock(&qinfo_list_lock);
 866
 867                 if (old || !noquota)
 868                         free_qinfo(oqi);
 869                 if (old && !noquota)
 870                         free_qinfo(old);
 871
 872                 if (old && !noquota)
 873                         CDEBUG(D_QUOTA, "setdq to remove for %s %d\n",
 874                                cnt == USRQUOTA ? "user" : "group", id);
 875                 else if (!old && noquota)
 876                         CDEBUG(D_QUOTA, "setdq to insert for %s %d\n",
 877                                cnt == USRQUOTA ? "user" : "group", id);
 878         }
 879
 880         RETURN(rc);
 881 }
 882
 883 int osc_quota_cleanup(struct obd_device *obd)
 884 {
 885         struct client_obd *cli = &obd->u.cli;
 886         struct osc_quota_info *oqi, *n;
 887         int i;
 888         ENTRY;
 889
 890         spin_lock(&qinfo_list_lock);
 891         for (i = 0; i < NR_DQHASH; i++) {
 892                 list_for_each_entry_safe(oqi, n, &qinfo_hash[i], oqi_hash) {
 893                         if (oqi->oqi_cli != cli)
 894                                 continue;
 895                         remove_qinfo_hash(oqi);
 896                         free_qinfo(oqi);
 897                 }
 898         }
 899         spin_unlock(&qinfo_list_lock);
 900
 901         RETURN(0);
 902 }
 903
 904 int osc_quota_init(void)
 905 {
 906         int i;
 907         ENTRY;
 908
 909         LASSERT(qinfo_cachep == NULL);
 910         qinfo_cachep = cfs_mem_cache_create("osc_quota_info",
 911                                             sizeof(struct osc_quota_info),
 912                                             0, 0);
 913         if (!qinfo_cachep)
 914                 RETURN(-ENOMEM);
 915
 916         for (i = 0; i < NR_DQHASH; i++)
 917                 CFS_INIT_LIST_HEAD(qinfo_hash + i);
 918
 919         RETURN(0);
 920 }
 921
 922 int osc_quota_exit(void)
 923 {
 924         struct osc_quota_info *oqi, *n;
 925         int i, rc;
 926         ENTRY;
 927
 928         spin_lock(&qinfo_list_lock);
 929         for (i = 0; i < NR_DQHASH; i++) {
 930                 list_for_each_entry_safe(oqi, n, &qinfo_hash[i], oqi_hash) {
 931                         remove_qinfo_hash(oqi);
 932                         free_qinfo(oqi);
 933                 }
 934         }
 935         spin_unlock(&qinfo_list_lock);
 936
 937         rc = cfs_mem_cache_destroy(qinfo_cachep);
 938         LASSERTF(rc == 0, "couldn't destory qinfo_cachep slab\n");
 939         qinfo_cachep = NULL;
 940
 941         RETURN(0);
 942 }
 943
 944 #ifdef __KERNEL__
 945 #ifdef HAVE_QUOTA_SUPPORT
 946 quota_interface_t mds_quota_interface = {
 947         .quota_init     = mds_quota_init,
 948         .quota_exit     = mds_quota_exit,
 949         .quota_setup    = mds_quota_setup,
 950         .quota_cleanup  = mds_quota_cleanup,
 951         .quota_check    = target_quota_check,
 952         .quota_ctl      = mds_quota_ctl,
 953         .quota_setinfo  = mds_quota_setinfo,
 954         .quota_fs_cleanup = mds_quota_fs_cleanup,
 955         .quota_recovery = mds_quota_recovery,
 956         .quota_adjust   = mds_quota_adjust,
 957         .quota_chkquota = quota_chk_acq_common,
 958         .quota_acquire  = quota_acquire_common,
 959         .quota_pending_commit = quota_pending_commit,
 960 };
 961
 962 quota_interface_t filter_quota_interface = {
 963         .quota_setup    = filter_quota_setup,
 964         .quota_cleanup  = filter_quota_cleanup,
 965         .quota_check    = target_quota_check,
 966         .quota_ctl      = filter_quota_ctl,
 967         .quota_setinfo  = filter_quota_setinfo,
 968         .quota_clearinfo = filter_quota_clearinfo,
 969         .quota_enforce  = filter_quota_enforce,
 970         .quota_getflag  = filter_quota_getflag,
 971         .quota_acquire  = quota_acquire_common,
 972         .quota_adjust   = filter_quota_adjust,
 973         .quota_chkquota = quota_chk_acq_common,
 974         .quota_adjust_qunit   = filter_quota_adjust_qunit,
 975         .quota_pending_commit = quota_pending_commit,
 976 };
 977 #endif
 978 #endif /* __KERNEL__ */
 979
 980 quota_interface_t mdc_quota_interface = {
 981         .quota_ctl      = client_quota_ctl,
 982         .quota_check    = client_quota_check,
 983         .quota_poll_check = client_quota_poll_check,
 984 };
 985
 986 quota_interface_t lmv_quota_interface = {
 987         .quota_ctl      = lmv_quota_ctl,
 988         .quota_check    = lmv_quota_check,
 989 };
 990
 991 quota_interface_t osc_quota_interface = {
 992         .quota_ctl      = client_quota_ctl,
 993         .quota_check    = client_quota_check,
 994         .quota_poll_check = client_quota_poll_check,
 995         .quota_init     = osc_quota_init,
 996         .quota_exit     = osc_quota_exit,
 997         .quota_chkdq    = osc_quota_chkdq,
 998         .quota_setdq    = osc_quota_setdq,
 999         .quota_cleanup  = osc_quota_cleanup,
1000         .quota_adjust_qunit = client_quota_adjust_qunit,
1001 };
1002
1003 quota_interface_t lov_quota_interface = {
1004         .quota_ctl      = lov_quota_ctl,
1005         .quota_check    = lov_quota_check,
1006         .quota_adjust_qunit = lov_quota_adjust_qunit,
1007 };
1008
1009 #ifdef __KERNEL__
1010
1011 cfs_proc_dir_entry_t *lquota_type_proc_dir = NULL;
1012
1013 static int __init init_lustre_quota(void)
1014 {
1015 #ifdef HAVE_QUOTA_SUPPORT
1016         int rc = 0;
1017
1018         lquota_type_proc_dir = lprocfs_register(OBD_LQUOTA_DEVICENAME,
1019                                                 proc_lustre_root,
1020                                                 NULL, NULL);
1021         if (IS_ERR(lquota_type_proc_dir)) {
1022                 CERROR("LProcFS failed in lquota-init\n");
1023                 rc = PTR_ERR(lquota_type_proc_dir);
1024                 return rc;
1025         }
1026
1027         rc = qunit_cache_init();
1028         if (rc)
1029                 return rc;
1030
1031         PORTAL_SYMBOL_REGISTER(filter_quota_interface);
1032         PORTAL_SYMBOL_REGISTER(mds_quota_interface);
1033 #endif
1034         PORTAL_SYMBOL_REGISTER(mdc_quota_interface);
1035         PORTAL_SYMBOL_REGISTER(lmv_quota_interface);
1036         PORTAL_SYMBOL_REGISTER(osc_quota_interface);
1037         PORTAL_SYMBOL_REGISTER(lov_quota_interface);
1038         return 0;
1039 }
1040
1041 static void /*__exit*/ exit_lustre_quota(void)
1042 {
1043         PORTAL_SYMBOL_UNREGISTER(mdc_quota_interface);
1044         PORTAL_SYMBOL_UNREGISTER(lmv_quota_interface);
1045         PORTAL_SYMBOL_UNREGISTER(osc_quota_interface);
1046         PORTAL_SYMBOL_UNREGISTER(lov_quota_interface);
1047 #ifdef HAVE_QUOTA_SUPPORT
1048         PORTAL_SYMBOL_UNREGISTER(filter_quota_interface);
1049         PORTAL_SYMBOL_UNREGISTER(mds_quota_interface);
1050
1051         qunit_cache_cleanup();
1052
1053         if (lquota_type_proc_dir)
1054                 lprocfs_remove(&lquota_type_proc_dir);
1055 #endif
1056 }
1057
1058 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1059 MODULE_DESCRIPTION("Lustre Quota");
1060 MODULE_LICENSE("GPL");
1061
1062 cfs_module(lquota, "1.0.0", init_lustre_quota, exit_lustre_quota);
1063
1064 #ifdef HAVE_QUOTA_SUPPORT
1065 EXPORT_SYMBOL(mds_quota_interface);
1066 EXPORT_SYMBOL(filter_quota_interface);
1067 #endif
1068 EXPORT_SYMBOL(mdc_quota_interface);
1069 EXPORT_SYMBOL(lmv_quota_interface);
1070 EXPORT_SYMBOL(osc_quota_interface);
1071 EXPORT_SYMBOL(lov_quota_interface);
1072 #endif /* __KERNEL */