lustre/quota/quota_context.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  * lustre/quota/quota_context.c
  37  *
  38  * Lustre Quota Context
  39  *
  40  * Author: Niu YaWei <niu@clusterfs.com>
  41  */
  42
  43 #ifndef EXPORT_SYMTAB
  44 # define EXPORT_SYMTAB
  45 #endif
  46
  47 #define DEBUG_SUBSYSTEM S_LQUOTA
  48
  49 #include <linux/version.h>
  50 #include <linux/fs.h>
  51 #include <asm/unistd.h>
  52 #include <linux/slab.h>
  53 #include <linux/quotaops.h>
  54 #include <linux/module.h>
  55 #include <linux/init.h>
  56
  57 #include <obd_class.h>
  58 #include <lustre_quota.h>
  59 #include <lustre_fsfilt.h>
  60 #include <class_hash.h>
  61 #include <lprocfs_status.h>
  62 #include "quota_internal.h"
  63
  64 #ifdef HAVE_QUOTA_SUPPORT
  65
  66 static lustre_hash_ops_t lqs_hash_ops;
  67
  68 unsigned long default_bunit_sz = 128 * 1024 * 1024; /* 128M bytes */
  69 unsigned long default_btune_ratio = 50;             /* 50 percentage */
  70 unsigned long default_iunit_sz = 5120;              /* 5120 inodes */
  71 unsigned long default_itune_ratio = 50;             /* 50 percentage */
  72
  73 cfs_mem_cache_t *qunit_cachep = NULL;
  74 struct list_head qunit_hash[NR_DQHASH];
  75 spinlock_t qunit_hash_lock = SPIN_LOCK_UNLOCKED;
  76
  77 /* please sync qunit_state with qunit_state_names */
  78 enum qunit_state {
  79         /**
  80          * a qunit is created
  81          */
  82         QUNIT_CREATED      = 0,
  83         /**
  84          * a qunit is added into qunit hash, that means
  85          * a quota req will be sent or is flying
  86          */
  87         QUNIT_IN_HASH      = 1,
  88         /**
  89          * a qunit is removed from qunit hash, that
  90          * means a quota req is handled and comes back
  91          */
  92         QUNIT_RM_FROM_HASH = 2,
  93         /**
  94          * qunit can wake up all threads waiting for it
  95          */
  96         QUNIT_FINISHED     = 3,
  97 };
  98
  99 static const char *qunit_state_names[] = {
 100         [QUNIT_CREATED]      = "CREATED",
 101         [QUNIT_IN_HASH]      = "IN_HASH",
 102         [QUNIT_RM_FROM_HASH] = "RM_FROM_HASH",
 103         [QUNIT_FINISHED]     = "FINISHED",
 104 };
 105
 106 struct lustre_qunit {
 107         struct list_head lq_hash;          /** Hash list in memory */
 108         atomic_t lq_refcnt;                /** Use count */
 109         struct lustre_quota_ctxt *lq_ctxt; /** Quota context this applies to */
 110         struct qunit_data lq_data;         /** See qunit_data */
 111         unsigned int lq_opc;               /** QUOTA_DQACQ, QUOTA_DQREL */
 112         cfs_waitq_t lq_waitq;              /** Threads waiting for this qunit */
 113         spinlock_t lq_lock;                /** Protect the whole structure */
 114         enum qunit_state lq_state;         /** Present the status of qunit */
 115         int lq_rc;                         /** The rc of lq_data */
 116         pid_t lq_owner;
 117 };
 118
 119 #define QUNIT_SET_STATE(qunit, state)                                   \
 120 do {                                                                    \
 121         spin_lock(&qunit->lq_lock);                                     \
 122         QDATA_DEBUG((&qunit->lq_data), "qunit(%p) lq_state(%s->%s), "   \
 123                     "lq_rc(%d), lq_owner(%d)\n",                        \
 124                     qunit, qunit_state_names[qunit->lq_state],          \
 125                     qunit_state_names[state], qunit->lq_rc,             \
 126                     qunit->lq_owner);                                   \
 127         qunit->lq_state = state;                                        \
 128         spin_unlock(&qunit->lq_lock);                                   \
 129 } while(0)
 130
 131 #define QUNIT_SET_STATE_AND_RC(qunit, state, rc)                        \
 132 do {                                                                    \
 133         spin_lock(&qunit->lq_lock);                                     \
 134         qunit->lq_rc = rc;                                              \
 135         QDATA_DEBUG((&qunit->lq_data), "qunit(%p) lq_state(%s->%s), "   \
 136                     "lq_rc(%d), lq_owner(%d)\n",                        \
 137                     qunit, qunit_state_names[qunit->lq_state],          \
 138                     qunit_state_names[state], qunit->lq_rc,             \
 139                     qunit->lq_owner);                                   \
 140         qunit->lq_state = state;                                        \
 141         spin_unlock(&qunit->lq_lock);                                   \
 142 } while(0)
 143
 144 int should_translate_quota (struct obd_import *imp)
 145 {
 146         ENTRY;
 147
 148         LASSERT(imp);
 149         if (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_QUOTA64)
 150                 RETURN(0);
 151         else
 152                 RETURN(1);
 153 }
 154
 155 void qunit_cache_cleanup(void)
 156 {
 157         int i;
 158         ENTRY;
 159
 160         spin_lock(&qunit_hash_lock);
 161         for (i = 0; i < NR_DQHASH; i++)
 162                 LASSERT(list_empty(qunit_hash + i));
 163         spin_unlock(&qunit_hash_lock);
 164
 165         if (qunit_cachep) {
 166                 int rc;
 167                 rc = cfs_mem_cache_destroy(qunit_cachep);
 168                 LASSERTF(rc == 0, "couldn't destroy qunit_cache slab\n");
 169                 qunit_cachep = NULL;
 170         }
 171         EXIT;
 172 }
 173
 174 int qunit_cache_init(void)
 175 {
 176         int i;
 177         ENTRY;
 178
 179         LASSERT(qunit_cachep == NULL);
 180         qunit_cachep = cfs_mem_cache_create("ll_qunit_cache",
 181                                             sizeof(struct lustre_qunit),
 182                                             0, 0);
 183         if (!qunit_cachep)
 184                 RETURN(-ENOMEM);
 185
 186         spin_lock(&qunit_hash_lock);
 187         for (i = 0; i < NR_DQHASH; i++)
 188                 CFS_INIT_LIST_HEAD(qunit_hash + i);
 189         spin_unlock(&qunit_hash_lock);
 190         RETURN(0);
 191 }
 192
 193 static inline int
 194 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
 195              __attribute__((__const__));
 196
 197 static inline int
 198 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
 199 {
 200         unsigned int id = qdata->qd_id;
 201         unsigned int type = QDATA_IS_GRP(qdata);
 202
 203         unsigned long tmp = ((unsigned long)qctxt >> L1_CACHE_SHIFT) ^ id;
 204         tmp = (tmp * (MAXQUOTAS - type)) % NR_DQHASH;
 205         return tmp;
 206 }
 207
 208 /* caller must hold qunit_hash_lock */
 209 static inline struct lustre_qunit *find_qunit(unsigned int hashent,
 210                                               struct lustre_quota_ctxt *qctxt,
 211                                               struct qunit_data *qdata)
 212 {
 213         struct lustre_qunit *qunit = NULL;
 214         struct qunit_data *tmp;
 215
 216         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
 217         list_for_each_entry(qunit, qunit_hash + hashent, lq_hash) {
 218                 tmp = &qunit->lq_data;
 219                 if (qunit->lq_ctxt == qctxt &&
 220                     qdata->qd_id == tmp->qd_id &&
 221                     (qdata->qd_flags & LQUOTA_QUNIT_FLAGS) ==
 222                     (tmp->qd_flags & LQUOTA_QUNIT_FLAGS))
 223                         return qunit;
 224         }
 225         return NULL;
 226 }
 227
 228 /* check_cur_qunit - check the current usage of qunit.
 229  * @qctxt: quota context
 230  * @qdata: the type of quota unit to be checked
 231  *
 232  * return: 1 - need acquire qunit;
 233  *         2 - need release qunit;
 234  *         0 - need do nothing.
 235  *       < 0 - error.
 236  */
 237 static int
 238 check_cur_qunit(struct obd_device *obd,
 239                 struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
 240 {
 241         struct super_block *sb = qctxt->lqc_sb;
 242         unsigned long qunit_sz, tune_sz;
 243         __u64 usage, limit, limit_org, pending_write = 0;
 244         long long record = 0;
 245         struct obd_quotactl *qctl;
 246         struct lustre_qunit_size *lqs = NULL;
 247         int ret = 0;
 248         ENTRY;
 249
 250         if (!ll_sb_any_quota_active(sb))
 251                 RETURN(0);
 252
 253         spin_lock(&qctxt->lqc_lock);
 254         if (!qctxt->lqc_valid){
 255                 spin_unlock(&qctxt->lqc_lock);
 256                 RETURN(0);
 257         }
 258         spin_unlock(&qctxt->lqc_lock);
 259
 260         OBD_ALLOC_PTR(qctl);
 261         if (qctl == NULL)
 262                 RETURN(-ENOMEM);
 263
 264         /* get fs quota usage & limit */
 265         qctl->qc_cmd = Q_GETQUOTA;
 266         qctl->qc_id = qdata->qd_id;
 267         qctl->qc_type = QDATA_IS_GRP(qdata);
 268         ret = fsfilt_quotactl(obd, sb, qctl);
 269         if (ret) {
 270                 if (ret == -ESRCH)      /* no limit */
 271                         ret = 0;
 272                 else
 273                         CERROR("can't get fs quota usage! (rc:%d)\n", ret);
 274                 GOTO(out, ret);
 275         }
 276
 277         if (QDATA_IS_BLK(qdata)) {
 278                 usage = qctl->qc_dqblk.dqb_curspace;
 279                 limit = qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS;
 280         } else {
 281                 usage = qctl->qc_dqblk.dqb_curinodes;
 282                 limit = qctl->qc_dqblk.dqb_ihardlimit;
 283         }
 284
 285         /* ignore the no quota limit case; and it can avoid creating
 286          * unnecessary lqs for uid/gid */
 287         if (!limit)
 288                 GOTO(out, ret = 0);
 289
 290         lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(qdata), qdata->qd_id),
 291                                qctxt, 0);
 292         if (IS_ERR(lqs) || lqs == NULL) {
 293                 CDEBUG(D_ERROR, "fail to find a lqs(%s id: %u)!\n",
 294                        QDATA_IS_GRP(qdata) ? "group" : "user", qdata->qd_id);
 295                 GOTO (out, ret = 0);
 296         }
 297         spin_lock(&lqs->lqs_lock);
 298
 299         if (QDATA_IS_BLK(qdata)) {
 300                 qunit_sz = lqs->lqs_bunit_sz;
 301                 tune_sz  = lqs->lqs_btune_sz;
 302                 pending_write = lqs->lqs_bwrite_pending;
 303                 record   = lqs->lqs_blk_rec;
 304                 LASSERT(!(qunit_sz % QUOTABLOCK_SIZE));
 305         } else {
 306                 /* we didn't need change inode qunit size now */
 307                 qunit_sz = lqs->lqs_iunit_sz;
 308                 tune_sz  = lqs->lqs_itune_sz;
 309                 pending_write = lqs->lqs_iwrite_pending;
 310                 record   = lqs->lqs_ino_rec;
 311         }
 312
 313         /* we don't count the MIN_QLIMIT */
 314         if ((limit == MIN_QLIMIT && !QDATA_IS_BLK(qdata)) ||
 315             (toqb(limit) == MIN_QLIMIT && QDATA_IS_BLK(qdata)))
 316                 limit = 0;
 317
 318         usage += pending_write;
 319         limit_org = limit;
 320         /* when a releasing quota req is sent, before it returned
 321            limit is assigned a small value. limit will overflow */
 322         if (limit + record < 0)
 323                 usage -= record;
 324         else
 325                 limit += record;
 326
 327         LASSERT(qdata->qd_count == 0);
 328         if (limit <= usage + tune_sz) {
 329                 while (qdata->qd_count + limit <=
 330                        usage + tune_sz)
 331                         qdata->qd_count += qunit_sz;
 332                 ret = 1;
 333         } else if (limit > usage + qunit_sz + tune_sz &&
 334                    limit_org > qdata->qd_count + qunit_sz) {
 335                 while (limit - qdata->qd_count > usage + qunit_sz + tune_sz &&
 336                        limit_org > qdata->qd_count + qunit_sz)
 337                         qdata->qd_count += qunit_sz;
 338                 ret = 2;
 339                 /* if there are other pending writes for this uid/gid, releasing
 340                  * quota is put off until the last pending write b=16645 */
 341                 if (ret == 2 && pending_write) {
 342                         CDEBUG(D_QUOTA, "delay quota release\n");
 343                         ret = 0;
 344                 }
 345         }
 346         CDEBUG(D_QUOTA, "type: %c, limit: "LPU64", usage: "LPU64
 347                ", pending_write: "LPU64", record: "LPD64
 348                ", qunit_sz: %lu, tune_sz: %lu, ret: %d.\n",
 349                QDATA_IS_BLK(qdata) ? 'b' : 'i', limit, usage, pending_write,
 350                record, qunit_sz, tune_sz, ret);
 351         LASSERT(ret == 0 || qdata->qd_count);
 352
 353         spin_unlock(&lqs->lqs_lock);
 354         lqs_putref(lqs);
 355
 356         EXIT;
 357  out:
 358         OBD_FREE_PTR(qctl);
 359         return ret;
 360 }
 361
 362 /**
 363  * Compute the remaining quota for certain gid or uid b=11693
 364  */
 365 int compute_remquota(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 366                      struct qunit_data *qdata, int isblk)
 367 {
 368         struct super_block *sb = qctxt->lqc_sb;
 369         __u64 usage, limit;
 370         struct obd_quotactl *qctl;
 371         int ret = QUOTA_RET_OK;
 372         ENTRY;
 373
 374         if (!ll_sb_any_quota_active(sb))
 375                 RETURN(QUOTA_RET_NOQUOTA);
 376
 377         /* ignore root user */
 378         if (qdata->qd_id == 0 && QDATA_IS_GRP(qdata) == USRQUOTA)
 379                 RETURN(QUOTA_RET_NOLIMIT);
 380
 381         OBD_ALLOC_PTR(qctl);
 382         if (qctl == NULL)
 383                 RETURN(-ENOMEM);
 384
 385         /* get fs quota usage & limit */
 386         qctl->qc_cmd = Q_GETQUOTA;
 387         qctl->qc_id = qdata->qd_id;
 388         qctl->qc_type = QDATA_IS_GRP(qdata);
 389         ret = fsfilt_quotactl(obd, sb, qctl);
 390         if (ret) {
 391                 if (ret == -ESRCH)      /* no limit */
 392                         ret = QUOTA_RET_NOLIMIT;
 393                 else
 394                         CDEBUG(D_QUOTA, "can't get fs quota usage! (rc:%d)",
 395                                ret);
 396                 GOTO(out, ret);
 397         }
 398
 399         usage = isblk ? qctl->qc_dqblk.dqb_curspace :
 400                 qctl->qc_dqblk.dqb_curinodes;
 401         limit = isblk ? qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS :
 402                 qctl->qc_dqblk.dqb_ihardlimit;
 403         if (!limit){            /* no limit */
 404                 ret = QUOTA_RET_NOLIMIT;
 405                 GOTO(out, ret);
 406         }
 407
 408         if (limit >= usage)
 409                 qdata->qd_count = limit - usage;
 410         else
 411                 qdata->qd_count = 0;
 412         EXIT;
 413 out:
 414         OBD_FREE_PTR(qctl);
 415         return ret;
 416 }
 417
 418 static struct lustre_qunit *alloc_qunit(struct lustre_quota_ctxt *qctxt,
 419                                         struct qunit_data *qdata, int opc)
 420 {
 421         struct lustre_qunit *qunit = NULL;
 422         ENTRY;
 423
 424         OBD_SLAB_ALLOC_PTR_GFP(qunit, qunit_cachep, CFS_ALLOC_IO);
 425         if (qunit == NULL)
 426                 RETURN(NULL);
 427
 428         CFS_INIT_LIST_HEAD(&qunit->lq_hash);
 429         init_waitqueue_head(&qunit->lq_waitq);
 430         atomic_set(&qunit->lq_refcnt, 1);
 431         qunit->lq_ctxt = qctxt;
 432         memcpy(&qunit->lq_data, qdata, sizeof(*qdata));
 433         qunit->lq_opc = opc;
 434         qunit->lq_lock = SPIN_LOCK_UNLOCKED;
 435         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_CREATED, 0);
 436         qunit->lq_owner = cfs_curproc_pid();
 437         RETURN(qunit);
 438 }
 439
 440 static inline void free_qunit(struct lustre_qunit *qunit)
 441 {
 442         OBD_SLAB_FREE(qunit, qunit_cachep, sizeof(*qunit));
 443 }
 444
 445 static inline void qunit_get(struct lustre_qunit *qunit)
 446 {
 447         atomic_inc(&qunit->lq_refcnt);
 448 }
 449
 450 static void qunit_put(struct lustre_qunit *qunit)
 451 {
 452         LASSERT(atomic_read(&qunit->lq_refcnt));
 453         if (atomic_dec_and_test(&qunit->lq_refcnt))
 454                 free_qunit(qunit);
 455 }
 456
 457 /* caller must hold qunit_hash_lock and release ref of qunit after using it */
 458 static struct lustre_qunit *dqacq_in_flight(struct lustre_quota_ctxt *qctxt,
 459                                             struct qunit_data *qdata)
 460 {
 461         unsigned int hashent = qunit_hashfn(qctxt, qdata);
 462         struct lustre_qunit *qunit;
 463         ENTRY;
 464
 465         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
 466         qunit = find_qunit(hashent, qctxt, qdata);
 467         if (qunit)
 468                 qunit_get(qunit);
 469         RETURN(qunit);
 470 }
 471
 472 static void
 473 insert_qunit_nolock(struct lustre_quota_ctxt *qctxt, struct lustre_qunit *qunit)
 474 {
 475         struct list_head *head;
 476
 477         LASSERT(list_empty(&qunit->lq_hash));
 478         qunit_get(qunit);
 479         head = qunit_hash + qunit_hashfn(qctxt, &qunit->lq_data);
 480         list_add(&qunit->lq_hash, head);
 481         QUNIT_SET_STATE(qunit, QUNIT_IN_HASH);
 482 }
 483
 484 static void compute_lqs_after_removing_qunit(struct lustre_qunit *qunit)
 485 {
 486         struct lustre_qunit_size *lqs;
 487
 488         lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(&qunit->lq_data),
 489                                        qunit->lq_data.qd_id),
 490                                qunit->lq_ctxt, 0);
 491         if (lqs && !IS_ERR(lqs)) {
 492                 spin_lock(&lqs->lqs_lock);
 493                 if (qunit->lq_opc == QUOTA_DQACQ)
 494                         quota_compute_lqs(&qunit->lq_data, lqs, 0, 1);
 495                 if (qunit->lq_opc == QUOTA_DQREL)
 496                         quota_compute_lqs(&qunit->lq_data, lqs, 0, 0);
 497                 spin_unlock(&lqs->lqs_lock);
 498                 /* this is for quota_search_lqs */
 499                 lqs_putref(lqs);
 500                 /* this is for schedule_dqacq */
 501                 lqs_putref(lqs);
 502         }
 503 }
 504
 505 static void remove_qunit_nolock(struct lustre_qunit *qunit)
 506 {
 507         LASSERT(!list_empty(&qunit->lq_hash));
 508         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
 509
 510         list_del_init(&qunit->lq_hash);
 511         QUNIT_SET_STATE(qunit, QUNIT_RM_FROM_HASH);
 512         qunit_put(qunit);
 513 }
 514
 515 void* quota_barrier(struct lustre_quota_ctxt *qctxt,
 516                     struct obd_quotactl *oqctl, int isblk)
 517 {
 518         struct lustre_qunit *qunit, *find_qunit;
 519         int cycle = 1;
 520
 521         OBD_SLAB_ALLOC(qunit, qunit_cachep, CFS_ALLOC_IO, sizeof(*qunit));
 522         if (qunit == NULL) {
 523                 CERROR("locating qunit failed.(id=%u isblk=%d %s)\n",
 524                        oqctl->qc_id, isblk, oqctl->qc_type ? "grp" : "usr");
 525                 qctxt_wait_pending_dqacq(qctxt, oqctl->qc_id,
 526                                          oqctl->qc_type, isblk);
 527                 return NULL;
 528         }
 529
 530         INIT_LIST_HEAD(&qunit->lq_hash);
 531         qunit->lq_lock = SPIN_LOCK_UNLOCKED;
 532         init_waitqueue_head(&qunit->lq_waitq);
 533         atomic_set(&qunit->lq_refcnt, 1);
 534         qunit->lq_ctxt = qctxt;
 535         qunit->lq_data.qd_id = oqctl->qc_id;
 536         qunit->lq_data.qd_flags =  oqctl->qc_type;
 537         if (isblk)
 538                 QDATA_SET_BLK(&qunit->lq_data);
 539         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_CREATED, 0);
 540         /* it means it is only an invalid qunit for barrier */
 541         qunit->lq_opc = QUOTA_LAST_OPC;
 542
 543         while (1) {
 544                 spin_lock(&qunit_hash_lock);
 545                 find_qunit = dqacq_in_flight(qctxt, &qunit->lq_data);
 546                 if (find_qunit) {
 547                         spin_unlock(&qunit_hash_lock);
 548                         qunit_put(find_qunit);
 549                         qctxt_wait_pending_dqacq(qctxt, oqctl->qc_id,
 550                                                  oqctl->qc_type, isblk);
 551                         CDEBUG(D_QUOTA, "cycle=%d\n", cycle++);
 552                         continue;
 553                 }
 554                 break;
 555         }
 556         insert_qunit_nolock(qctxt, qunit);
 557         spin_unlock(&qunit_hash_lock);
 558         return qunit;
 559 }
 560
 561 void quota_unbarrier(void *handle)
 562 {
 563         struct lustre_qunit *qunit = (struct lustre_qunit *)handle;
 564
 565         if (qunit == NULL) {
 566                 CERROR("handle is NULL\n");
 567                 return;
 568         }
 569
 570         LASSERT(qunit->lq_opc == QUOTA_LAST_OPC);
 571         spin_lock(&qunit_hash_lock);
 572         remove_qunit_nolock(qunit);
 573         spin_unlock(&qunit_hash_lock);
 574         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, QUOTA_REQ_RETURNED);
 575         wake_up(&qunit->lq_waitq);
 576         qunit_put(qunit);
 577 }
 578
 579 #define INC_QLIMIT(limit, count) (limit == MIN_QLIMIT) ? \
 580                                  (limit = count) : (limit += count)
 581
 582
 583 static inline int is_master(struct lustre_quota_ctxt *qctxt)
 584 {
 585         return qctxt->lqc_handler ? 1 : 0;
 586 }
 587
 588 static int
 589 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 590                struct qunit_data *qdata, int opc, int wait,
 591                struct obd_trans_info *oti);
 592
 593 static inline void qdata_to_oqaq(struct qunit_data *qdata,
 594                                  struct quota_adjust_qunit *oqaq)
 595 {
 596         LASSERT(qdata);
 597         LASSERT(oqaq);
 598
 599         oqaq->qaq_flags = qdata->qd_flags;
 600         oqaq->qaq_id    = qdata->qd_id;
 601         if (QDATA_IS_ADJBLK(qdata))
 602                 oqaq->qaq_bunit_sz = qdata->qd_qunit;
 603         if (QDATA_IS_ADJINO(qdata))
 604                 oqaq->qaq_iunit_sz = qdata->qd_qunit;
 605 }
 606
 607 static int
 608 dqacq_completion(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 609                  struct qunit_data *qdata, int rc, int opc)
 610 {
 611         struct lustre_qunit *qunit = NULL;
 612         struct super_block *sb = qctxt->lqc_sb;
 613         int err = 0;
 614         struct quota_adjust_qunit *oqaq = NULL;
 615         int rc1 = 0;
 616         ENTRY;
 617
 618         LASSERT(qdata);
 619         QDATA_DEBUG(qdata, "obd(%s): complete %s quota req\n",
 620                     obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
 621
 622         /* do it only when a releasing quota req more than 5MB b=18491 */
 623         if (opc == QUOTA_DQREL && qdata->qd_count >= 5242880)
 624                 OBD_FAIL_TIMEOUT(OBD_FAIL_QUOTA_DELAY_REL, 5);
 625
 626         /* update local operational quota file */
 627         if (rc == 0) {
 628                 __u64 count = QUSG(qdata->qd_count, QDATA_IS_BLK(qdata));
 629                 struct obd_quotactl *qctl;
 630                 __u64 *hardlimit;
 631
 632                 OBD_ALLOC_PTR(qctl);
 633                 if (qctl == NULL)
 634                         GOTO(out, err = -ENOMEM);
 635
 636                 /* acq/rel qunit for specified uid/gid is serialized,
 637                  * so there is no race between get fs quota limit and
 638                  * set fs quota limit */
 639                 qctl->qc_cmd = Q_GETQUOTA;
 640                 qctl->qc_id = qdata->qd_id;
 641                 qctl->qc_type = QDATA_IS_GRP(qdata);
 642                 err = fsfilt_quotactl(obd, sb, qctl);
 643                 if (err) {
 644                         CERROR("error get quota fs limit! (rc:%d)\n", err);
 645                         GOTO(out_mem, err);
 646                 }
 647
 648                 if (QDATA_IS_BLK(qdata)) {
 649                         qctl->qc_dqblk.dqb_valid = QIF_BLIMITS;
 650                         hardlimit = &qctl->qc_dqblk.dqb_bhardlimit;
 651                 } else {
 652                         qctl->qc_dqblk.dqb_valid = QIF_ILIMITS;
 653                         hardlimit = &qctl->qc_dqblk.dqb_ihardlimit;
 654                 }
 655
 656                 CDEBUG(D_QUOTA, "hardlimt: "LPU64"\n", *hardlimit);
 657
 658                 if (*hardlimit == 0)
 659                         goto out_mem;
 660
 661                 switch (opc) {
 662                 case QUOTA_DQACQ:
 663                         INC_QLIMIT(*hardlimit, count);
 664                         break;
 665                 case QUOTA_DQREL:
 666                         LASSERTF(count < *hardlimit,
 667                                  "id(%u) flag(%u) type(%c) isblk(%c) "
 668                                  "count("LPU64") qd_qunit("LPU64") "
 669                                  "hardlimit("LPU64").\n",
 670                                  qdata->qd_id, qdata->qd_flags,
 671                                  QDATA_IS_GRP(qdata) ? 'g' : 'u',
 672                                  QDATA_IS_BLK(qdata) ? 'b': 'i',
 673                                  qdata->qd_count, qdata->qd_qunit, *hardlimit);
 674                         *hardlimit -= count;
 675                         break;
 676                 default:
 677                         LBUG();
 678                 }
 679
 680                 /* clear quota limit */
 681                 if (count == 0)
 682                         *hardlimit = 0;
 683
 684                 qctl->qc_cmd = Q_SETQUOTA;
 685                 err = fsfilt_quotactl(obd, sb, qctl);
 686                 if (err)
 687                         CERROR("error set quota fs limit! (rc:%d)\n", err);
 688
 689                 QDATA_DEBUG(qdata, "%s completion\n",
 690                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 691 out_mem:
 692                 OBD_FREE_PTR(qctl);
 693         } else if (rc == -EDQUOT) {
 694                 QDATA_DEBUG(qdata, "acquire qunit got EDQUOT.\n");
 695         } else if (rc == -EBUSY) {
 696                 QDATA_DEBUG(qdata, "it's is recovering, got EBUSY.\n");
 697         } else {
 698                 CERROR("acquire qunit got error! (rc:%d)\n", rc);
 699         }
 700 out:
 701         /* remove the qunit from hash */
 702         spin_lock(&qunit_hash_lock);
 703
 704         qunit = dqacq_in_flight(qctxt, qdata);
 705         /* this qunit has been removed by qctxt_cleanup() */
 706         if (!qunit) {
 707                 spin_unlock(&qunit_hash_lock);
 708                 QDATA_DEBUG(qdata, "%s is discarded because qunit isn't found\n",
 709                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 710                 RETURN(err);
 711         }
 712
 713         LASSERT(opc == qunit->lq_opc);
 714         /* remove this qunit from lq_hash so that new processes cannot be added
 715          * to qunit->lq_waiters */
 716         remove_qunit_nolock(qunit);
 717         spin_unlock(&qunit_hash_lock);
 718
 719         compute_lqs_after_removing_qunit(qunit);
 720
 721         if (rc == 0)
 722                 rc = QUOTA_REQ_RETURNED;
 723         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, rc);
 724         /* wake up all waiters */
 725         wake_up_all(&qunit->lq_waitq);
 726
 727         /* this is for dqacq_in_flight() */
 728         qunit_put(qunit);
 729         /* this is for alloc_qunit() */
 730         qunit_put(qunit);
 731         if (rc < 0 && rc != -EDQUOT)
 732                  RETURN(err);
 733
 734         /* don't reschedule in such cases:
 735          *   - acq/rel failure and qunit isn't changed,
 736          *     but not for quota recovery.
 737          *   - local dqacq/dqrel.
 738          *   - local disk io failure.
 739          */
 740          OBD_ALLOC_PTR(oqaq);
 741          if (!oqaq)
 742                  RETURN(-ENOMEM);
 743          qdata_to_oqaq(qdata, oqaq);
 744          /* adjust the qunit size in slaves */
 745          rc1 = quota_adjust_slave_lqs(oqaq, qctxt);
 746          OBD_FREE_PTR(oqaq);
 747          if (rc1 < 0) {
 748                  CERROR("adjust slave's qunit size failed!(rc:%d)\n", rc1);
 749                  RETURN(rc1);
 750          }
 751          if (err || (rc < 0 && rc != -EBUSY && rc1 == 0) || is_master(qctxt))
 752                  RETURN(err);
 753
 754          if (opc == QUOTA_DQREL && qdata->qd_count >= 5242880 &&
 755              OBD_FAIL_CHECK(OBD_FAIL_QUOTA_DELAY_REL))
 756                  RETURN(err);
 757
 758         /* reschedule another dqacq/dqrel if needed */
 759         qdata->qd_count = 0;
 760         qdata->qd_flags &= LQUOTA_QUNIT_FLAGS;
 761         rc1 = check_cur_qunit(obd, qctxt, qdata);
 762         if (rc1 > 0) {
 763                 int opc;
 764                 opc = rc1 == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
 765                 rc1 = schedule_dqacq(obd, qctxt, qdata, opc, 0, NULL);
 766                 QDATA_DEBUG(qdata, "reschedudle opc(%d) rc(%d)\n", opc, rc1);
 767         }
 768         RETURN(err);
 769 }
 770
 771 struct dqacq_async_args {
 772         struct lustre_quota_ctxt *aa_ctxt;
 773         struct lustre_qunit *aa_qunit;
 774 };
 775
 776 static int dqacq_interpret(const struct lu_env *env,
 777                            struct ptlrpc_request *req, void *data, int rc)
 778 {
 779         struct dqacq_async_args *aa = (struct dqacq_async_args *)data;
 780         struct lustre_quota_ctxt *qctxt = aa->aa_ctxt;
 781         struct obd_device_target *obt = qctxt->lqc_obt;
 782         struct lustre_qunit *qunit = aa->aa_qunit;
 783         struct obd_device *obd = req->rq_import->imp_obd;
 784         struct qunit_data *qdata = NULL;
 785         ENTRY;
 786
 787         LASSERT(req);
 788         LASSERT(req->rq_import);
 789
 790         down_read(&obt->obt_rwsem);
 791         /* if a quota req timeouts or is dropped, we should update quota
 792          * statistics which will be handled in dqacq_completion. And in
 793          * this situation we should get qdata from request instead of
 794          * reply */
 795         qdata = quota_get_qdata(req, (rc != 0) ? QUOTA_REQUEST : QUOTA_REPLY,
 796                                 QUOTA_IMPORT);
 797         if (IS_ERR(qdata)) {
 798                 rc = PTR_ERR(qdata);
 799                 DEBUG_REQ(D_ERROR, req,
 800                           "error unpacking qunit_data(rc: %ld)\n",
 801                           PTR_ERR(qdata));
 802                 RETURN(PTR_ERR(qdata));
 803         }
 804
 805         QDATA_DEBUG(qdata, "qdata: interpret rc(%d).\n", rc);
 806         QDATA_DEBUG((&qunit->lq_data), "lq_data: \n");
 807
 808         if (qdata->qd_id != qunit->lq_data.qd_id ||
 809             OBD_FAIL_CHECK(OBD_FAIL_QUOTA_RET_QDATA)) {
 810                 CDEBUG(D_ERROR, "the returned qd_id isn't expected!"
 811                        "(qdata: %u, lq_data: %u)\n", qdata->qd_id,
 812                        qunit->lq_data.qd_id);
 813                 qdata->qd_id = qunit->lq_data.qd_id;
 814                 rc = -EPROTO;
 815         }
 816         if (QDATA_IS_GRP(qdata) != QDATA_IS_GRP(&qunit->lq_data)) {
 817                 CDEBUG(D_ERROR, "the returned grp/usr isn't expected!"
 818                        "(qdata: %u, lq_data: %u)\n", qdata->qd_flags,
 819                        qunit->lq_data.qd_flags);
 820                 if (QDATA_IS_GRP(&qunit->lq_data))
 821                         QDATA_SET_GRP(qdata);
 822                 else
 823                         QDATA_CLR_GRP(qdata);
 824                 rc = -EPROTO;
 825         }
 826         if (qdata->qd_count > qunit->lq_data.qd_count) {
 827                 CDEBUG(D_ERROR, "the returned qd_count isn't expected!"
 828                        "(qdata: "LPU64", lq_data: "LPU64")\n", qdata->qd_count,
 829                        qunit->lq_data.qd_count);
 830                 rc = -EPROTO;
 831         }
 832
 833         if (unlikely(rc == -ESRCH))
 834                 CERROR("quota for %s has been enabled by master, but disabled "
 835                        "by slave.\n", QDATA_IS_GRP(qdata) ? "group" : "user");
 836
 837         rc = dqacq_completion(obd, qctxt, qdata, rc,
 838                               lustre_msg_get_opc(req->rq_reqmsg));
 839
 840         up_read(&obt->obt_rwsem);
 841         RETURN(rc);
 842 }
 843
 844 /**
 845  * check if quota master is online
 846  */
 847 int check_qm(struct lustre_quota_ctxt *qctxt)
 848 {
 849         int rc;
 850         ENTRY;
 851
 852         spin_lock(&qctxt->lqc_lock);
 853         /* quit waiting when mds is back or qctxt is cleaned up */
 854         rc = qctxt->lqc_import || !qctxt->lqc_valid;
 855         spin_unlock(&qctxt->lqc_lock);
 856
 857         RETURN(rc);
 858 }
 859
 860 /* wake up all waiting threads when lqc_import is NULL */
 861 void dqacq_interrupt(struct lustre_quota_ctxt *qctxt)
 862 {
 863         struct lustre_qunit *qunit, *tmp;
 864         int i;
 865         ENTRY;
 866
 867         spin_lock(&qunit_hash_lock);
 868         for (i = 0; i < NR_DQHASH; i++) {
 869                 list_for_each_entry_safe(qunit, tmp, &qunit_hash[i], lq_hash) {
 870                         if (qunit->lq_ctxt != qctxt)
 871                                 continue;
 872
 873                         /* Wake up all waiters. Do not change lq_state.
 874                          * The waiters will check lq_rc which is kept as 0
 875                          * if no others change it, then the waiters will return
 876                          * -EAGAIN to caller who can perform related quota
 877                          * acq/rel if necessary. */
 878                         wake_up_all(&qunit->lq_waitq);
 879                 }
 880         }
 881         spin_unlock(&qunit_hash_lock);
 882         EXIT;
 883 }
 884
 885 static int got_qunit(struct lustre_qunit *qunit, int is_master)
 886 {
 887         struct lustre_quota_ctxt *qctxt = qunit->lq_ctxt;
 888         int rc = 0;
 889         ENTRY;
 890
 891         spin_lock(&qunit->lq_lock);
 892         switch (qunit->lq_state) {
 893         case QUNIT_IN_HASH:
 894         case QUNIT_RM_FROM_HASH:
 895                 break;
 896         case QUNIT_FINISHED:
 897                 rc = 1;
 898                 break;
 899         default:
 900                 CERROR("invalid qunit state %d\n", qunit->lq_state);
 901         }
 902         spin_unlock(&qunit->lq_lock);
 903
 904         if (!rc) {
 905                 spin_lock(&qctxt->lqc_lock);
 906                 rc = !qctxt->lqc_valid;
 907                 if (!is_master)
 908                         rc |= !qctxt->lqc_import;
 909                 spin_unlock(&qctxt->lqc_lock);
 910         }
 911
 912         RETURN(rc);
 913 }
 914
 915 static int
 916 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 917                struct qunit_data *qdata, int opc, int wait,
 918                struct obd_trans_info *oti)
 919 {
 920         struct lustre_qunit *qunit, *empty;
 921         struct l_wait_info lwi = { 0 };
 922         struct ptlrpc_request *req;
 923         struct dqacq_async_args *aa;
 924         struct obd_import *imp = NULL;
 925         struct lustre_qunit_size *lqs = NULL;
 926         struct timeval work_start;
 927         struct timeval work_end;
 928         long timediff;
 929         int rc = 0;
 930         ENTRY;
 931
 932         LASSERT(opc == QUOTA_DQACQ || opc == QUOTA_DQREL);
 933         do_gettimeofday(&work_start);
 934         if ((empty = alloc_qunit(qctxt, qdata, opc)) == NULL)
 935                 RETURN(-ENOMEM);
 936
 937         spin_lock(&qunit_hash_lock);
 938         qunit = dqacq_in_flight(qctxt, qdata);
 939         if (qunit) {
 940                 spin_unlock(&qunit_hash_lock);
 941                 qunit_put(empty);
 942
 943                 goto wait_completion;
 944         }
 945         qunit = empty;
 946         qunit_get(qunit);
 947         insert_qunit_nolock(qctxt, qunit);
 948         spin_unlock(&qunit_hash_lock);
 949
 950         lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(qdata), qdata->qd_id),
 951                                qctxt, 0);
 952         if (lqs && !IS_ERR(lqs)) {
 953                 spin_lock(&lqs->lqs_lock);
 954                 quota_compute_lqs(qdata, lqs, 1, (opc == QUOTA_DQACQ) ? 1 : 0);
 955                 /* when this qdata returned from mds, it will call lqs_putref */
 956                 lqs_getref(lqs);
 957                 spin_unlock(&lqs->lqs_lock);
 958                 /* this is for quota_search_lqs */
 959                 lqs_putref(lqs);
 960         } else {
 961                 CDEBUG(D_ERROR, "Can't find the lustre qunit size!\n");
 962         }
 963
 964         QDATA_DEBUG(qdata, "obd(%s): send %s quota req\n",
 965                     obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
 966         /* master is going to dqacq/dqrel from itself */
 967         if (is_master(qctxt)) {
 968                 int rc2;
 969                 QDATA_DEBUG(qdata, "local %s.\n",
 970                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 971                 QDATA_SET_CHANGE_QS(qdata);
 972                 rc = qctxt->lqc_handler(obd, qdata, opc);
 973                 rc2 = dqacq_completion(obd, qctxt, qdata, rc, opc);
 974                 /* this is for qunit_get() */
 975                 qunit_put(qunit);
 976
 977                 do_gettimeofday(&work_end);
 978                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
 979                 if (opc == QUOTA_DQACQ)
 980                         lprocfs_counter_add(qctxt->lqc_stats,
 981                                             wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
 982                                             timediff);
 983                 else
 984                         lprocfs_counter_add(qctxt->lqc_stats,
 985                                             wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
 986                                             timediff);
 987                 RETURN(rc ? rc : rc2);
 988         }
 989
 990         spin_lock(&qctxt->lqc_lock);
 991         if (!qctxt->lqc_import) {
 992                 spin_unlock(&qctxt->lqc_lock);
 993                 QDATA_DEBUG(qdata, "lqc_import is invalid.\n");
 994
 995                 spin_lock(&qunit_hash_lock);
 996                 remove_qunit_nolock(qunit);
 997                 spin_unlock(&qunit_hash_lock);
 998
 999                 compute_lqs_after_removing_qunit(qunit);
1000
1001                 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, -EAGAIN);
1002                 wake_up_all(&qunit->lq_waitq);
1003
1004                 /* this is for qunit_get() */
1005                 qunit_put(qunit);
1006                 /* this for alloc_qunit() */
1007                 qunit_put(qunit);
1008                 spin_lock(&qctxt->lqc_lock);
1009                 if (wait && !qctxt->lqc_import) {
1010                         spin_unlock(&qctxt->lqc_lock);
1011
1012                         LASSERT(oti && oti->oti_thread &&
1013                                 oti->oti_thread->t_watchdog);
1014
1015                         lc_watchdog_disable(oti->oti_thread->t_watchdog);
1016                         CDEBUG(D_QUOTA, "sleep for quota master\n");
1017                         l_wait_event(qctxt->lqc_wait_for_qmaster,
1018                                      check_qm(qctxt), &lwi);
1019                         CDEBUG(D_QUOTA, "wake up when quota master is back\n");
1020                         lc_watchdog_touch(oti->oti_thread->t_watchdog,
1021                                  GET_TIMEOUT(oti->oti_thread->t_svc));
1022                 } else {
1023                         spin_unlock(&qctxt->lqc_lock);
1024                 }
1025
1026                 RETURN(-EAGAIN);
1027         }
1028         imp = class_import_get(qctxt->lqc_import);
1029         spin_unlock(&qctxt->lqc_lock);
1030
1031         /* build dqacq/dqrel request */
1032         LASSERT(imp);
1033
1034         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_QUOTA_DQACQ,
1035                                         LUSTRE_MDS_VERSION, opc);
1036         class_import_put(imp);
1037         if (req == NULL) {
1038                 CDEBUG(D_ERROR, "Can't alloc request\n");
1039                 dqacq_completion(obd, qctxt, qdata, -ENOMEM, opc);
1040                 /* this is for qunit_get() */
1041                 qunit_put(qunit);
1042                 RETURN(-ENOMEM);
1043         }
1044
1045         ptlrpc_request_set_replen(req);
1046         req->rq_no_resend = req->rq_no_delay = 1;
1047         rc = quota_copy_qdata(req, qdata, QUOTA_REQUEST, QUOTA_IMPORT);
1048         if (rc < 0) {
1049                 CDEBUG(D_ERROR, "Can't pack qunit_data(rc: %d)\n", rc);
1050                 ptlrpc_req_finished(req);
1051                 dqacq_completion(obd, qctxt, qdata, -EPROTO, opc);
1052                 /* this is for qunit_get() */
1053                 qunit_put(qunit);
1054                 RETURN(rc);
1055         }
1056
1057         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1058         aa = ptlrpc_req_async_args(req);
1059         aa->aa_ctxt = qctxt;
1060         aa->aa_qunit = qunit;
1061
1062         req->rq_interpret_reply = dqacq_interpret;
1063         ptlrpcd_add_req(req, PSCOPE_OTHER);
1064
1065         QDATA_DEBUG(qdata, "%s scheduled.\n",
1066                     opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
1067 wait_completion:
1068         if (wait && qunit) {
1069                 struct qunit_data *p = &qunit->lq_data;
1070
1071                 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
1072                 l_wait_event(qunit->lq_waitq, got_qunit(qunit, is_master(qctxt)),
1073                              &lwi);
1074                 /* rc = -EAGAIN, it means the quota master isn't ready yet
1075                  * rc = QUOTA_REQ_RETURNED, it means a quota req is finished;
1076                  * rc = -EDQUOT, it means out of quota
1077                  * rc = -EBUSY, it means recovery is happening
1078                  * other rc < 0, it means real errors, functions who call
1079                  * schedule_dqacq should take care of this */
1080                 spin_lock(&qunit->lq_lock);
1081                 rc = qunit->lq_rc;
1082                 spin_unlock(&qunit->lq_lock);
1083                 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting: id(%u) flag(%u) "
1084                        "rc(%d) owner(%d)\n", qunit, qunit->lq_data.qd_id,
1085                        qunit->lq_data.qd_flags, rc, qunit->lq_owner);
1086         }
1087
1088         qunit_put(qunit);
1089         do_gettimeofday(&work_end);
1090         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1091         if (opc == QUOTA_DQACQ)
1092                 lprocfs_counter_add(qctxt->lqc_stats,
1093                                     wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
1094                                     timediff);
1095         else
1096                 lprocfs_counter_add(qctxt->lqc_stats,
1097                                     wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
1098                                     timediff);
1099
1100         RETURN(rc);
1101 }
1102
1103 int
1104 qctxt_adjust_qunit(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
1105                    const unsigned int id[], __u32 isblk, int wait,
1106                    struct obd_trans_info *oti)
1107 {
1108         int rc = 0, i = USRQUOTA;
1109         struct qunit_data qdata[MAXQUOTAS];
1110         ENTRY;
1111
1112         if (quota_is_set(obd, id, isblk ? QB_SET : QI_SET) == 0)
1113                 RETURN(0);
1114
1115         for (i = 0; i < MAXQUOTAS; i++) {
1116                 qdata[i].qd_id = id[i];
1117                 qdata[i].qd_flags = i;
1118                 if (isblk)
1119                         QDATA_SET_BLK(&qdata[i]);
1120                 qdata[i].qd_count = 0;
1121
1122                 rc = check_cur_qunit(obd, qctxt, &qdata[i]);
1123                 if (rc > 0) {
1124                         int opc;
1125                         /* need acquire or release */
1126                         opc = rc == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1127                         rc = schedule_dqacq(obd, qctxt, &qdata[i], opc,
1128                                             wait,oti);
1129                         if (rc < 0)
1130                                 RETURN(rc);
1131                 } else if (wait == 1) {
1132                         /* when wait equates 1, that means mds_quota_acquire
1133                          * or filter_quota_acquire is calling it. */
1134                         rc = qctxt_wait_pending_dqacq(qctxt, id[i], i, isblk);
1135                         if (rc < 0)
1136                                 RETURN(rc);
1137                 }
1138         }
1139
1140         RETURN(rc);
1141 }
1142
1143 int
1144 qctxt_wait_pending_dqacq(struct lustre_quota_ctxt *qctxt, unsigned int id,
1145                          unsigned short type, int isblk)
1146 {
1147         struct lustre_qunit *qunit = NULL;
1148         struct qunit_data qdata;
1149         struct timeval work_start;
1150         struct timeval work_end;
1151         long timediff;
1152         struct l_wait_info lwi = { 0 };
1153         int rc = 0;
1154         ENTRY;
1155
1156         do_gettimeofday(&work_start);
1157         qdata.qd_id = id;
1158         qdata.qd_flags = type;
1159         if (isblk)
1160                 QDATA_SET_BLK(&qdata);
1161         qdata.qd_count = 0;
1162
1163         spin_lock(&qunit_hash_lock);
1164         qunit = dqacq_in_flight(qctxt, &qdata);
1165         spin_unlock(&qunit_hash_lock);
1166
1167         if (qunit) {
1168                 struct qunit_data *p = &qunit->lq_data;
1169
1170                 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
1171                 l_wait_event(qunit->lq_waitq, got_qunit(qunit, is_master(qctxt)),
1172                              &lwi);
1173                 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting: rc(%d) "
1174                        "owner(%d)\n", qunit, qunit->lq_rc, qunit->lq_owner);
1175                 /* keep same as schedule_dqacq() b=17030 */
1176                 spin_lock(&qunit->lq_lock);
1177                 rc = qunit->lq_rc;
1178                 spin_unlock(&qunit->lq_lock);
1179                 /* this is for dqacq_in_flight() */
1180                 qunit_put(qunit);
1181                 do_gettimeofday(&work_end);
1182                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1183                 lprocfs_counter_add(qctxt->lqc_stats,
1184                                     isblk ? LQUOTA_WAIT_PENDING_BLK_QUOTA :
1185                                             LQUOTA_WAIT_PENDING_INO_QUOTA,
1186                                     timediff);
1187         } else {
1188                 do_gettimeofday(&work_end);
1189                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1190                 lprocfs_counter_add(qctxt->lqc_stats,
1191                                     isblk ? LQUOTA_NOWAIT_PENDING_BLK_QUOTA :
1192                                             LQUOTA_NOWAIT_PENDING_INO_QUOTA,
1193                                     timediff);
1194         }
1195
1196         RETURN(rc);
1197 }
1198
1199 int
1200 qctxt_init(struct obd_device *obd, dqacq_handler_t handler)
1201 {
1202         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
1203         struct obd_device_target *obt = &obd->u.obt;
1204         struct super_block *sb = obt->obt_sb;
1205         int rc = 0;
1206         ENTRY;
1207
1208         LASSERT(qctxt);
1209
1210         rc = ptlrpcd_addref();
1211         if (rc)
1212                 RETURN(rc);
1213
1214         cfs_waitq_init(&qctxt->lqc_wait_for_qmaster);
1215         cfs_waitq_init(&qctxt->lqc_lqs_waitq);
1216         atomic_set(&qctxt->lqc_lqs, 0);
1217         spin_lock_init(&qctxt->lqc_lock);
1218         spin_lock(&qctxt->lqc_lock);
1219         qctxt->lqc_handler = handler;
1220         qctxt->lqc_sb = sb;
1221         qctxt->lqc_obt = obt;
1222         qctxt->lqc_import = NULL;
1223         qctxt->lqc_recovery = 0;
1224         qctxt->lqc_switch_qs = 1; /* Change qunit size in default setting */
1225         qctxt->lqc_valid = 1;
1226         qctxt->lqc_cqs_boundary_factor = 4;
1227         qctxt->lqc_cqs_least_bunit = PTLRPC_MAX_BRW_SIZE;
1228         qctxt->lqc_cqs_least_iunit = 2;
1229         qctxt->lqc_cqs_qs_factor = 2;
1230         qctxt->lqc_flags = 0;
1231         QUOTA_MASTER_UNREADY(qctxt);
1232         qctxt->lqc_bunit_sz = default_bunit_sz;
1233         qctxt->lqc_btune_sz = default_bunit_sz / 100 * default_btune_ratio;
1234         qctxt->lqc_iunit_sz = default_iunit_sz;
1235         qctxt->lqc_itune_sz = default_iunit_sz * default_itune_ratio / 100;
1236         qctxt->lqc_switch_seconds = 300; /* enlarging will wait 5 minutes
1237                                           * after the last shrinking */
1238         qctxt->lqc_sync_blk = 0;
1239         spin_unlock(&qctxt->lqc_lock);
1240
1241         qctxt->lqc_lqs_hash = lustre_hash_init("LQS_HASH",
1242                                                HASH_LQS_CUR_BITS,
1243                                                HASH_LQS_MAX_BITS,
1244                                                &lqs_hash_ops, 0);
1245         if (!qctxt->lqc_lqs_hash) {
1246                 CERROR("initialize hash lqs for %s error!\n", obd->obd_name);
1247                 RETURN(-ENOMEM);
1248         }
1249
1250 #ifdef LPROCFS
1251         rc = lquota_proc_setup(obd, is_master(qctxt));
1252         if (rc)
1253                 CERROR("initialize proc for %s error!\n", obd->obd_name);
1254 #endif
1255
1256         RETURN(rc);
1257 }
1258
1259 static int check_lqs(struct lustre_quota_ctxt *qctxt)
1260 {
1261         int rc;
1262         ENTRY;
1263
1264         rc = !atomic_read(&qctxt->lqc_lqs);
1265
1266         RETURN(rc);
1267 }
1268
1269
1270 void hash_put_lqs(void *obj, void *data)
1271 {
1272         lqs_putref((struct lustre_qunit_size *)obj);
1273 }
1274
1275 void qctxt_cleanup(struct lustre_quota_ctxt *qctxt, int force)
1276 {
1277         struct lustre_qunit *qunit, *tmp;
1278         struct list_head tmp_list;
1279         struct l_wait_info lwi = { 0 };
1280         struct obd_device_target *obt = qctxt->lqc_obt;
1281         int i;
1282         ENTRY;
1283
1284         CFS_INIT_LIST_HEAD(&tmp_list);
1285
1286         spin_lock(&qctxt->lqc_lock);
1287         qctxt->lqc_valid = 0;
1288         spin_unlock(&qctxt->lqc_lock);
1289
1290         spin_lock(&qunit_hash_lock);
1291         for (i = 0; i < NR_DQHASH; i++) {
1292                 list_for_each_entry_safe(qunit, tmp, &qunit_hash[i], lq_hash) {
1293                         if (qunit->lq_ctxt != qctxt)
1294                                 continue;
1295                         remove_qunit_nolock(qunit);
1296                         list_add(&qunit->lq_hash, &tmp_list);
1297                 }
1298         }
1299         spin_unlock(&qunit_hash_lock);
1300
1301         list_for_each_entry_safe(qunit, tmp, &tmp_list, lq_hash) {
1302                 list_del_init(&qunit->lq_hash);
1303                 compute_lqs_after_removing_qunit(qunit);
1304
1305                 /* wake up all waiters */
1306                 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, 0);
1307                 wake_up_all(&qunit->lq_waitq);
1308                 qunit_put(qunit);
1309         }
1310
1311         /* after qctxt_cleanup, qctxt might be freed, then check_qm() is
1312          * unpredicted. So we must wait until lqc_wait_for_qmaster is empty */
1313         while (cfs_waitq_active(&qctxt->lqc_wait_for_qmaster)) {
1314                 cfs_waitq_signal(&qctxt->lqc_wait_for_qmaster);
1315                 cfs_schedule_timeout(CFS_TASK_INTERRUPTIBLE,
1316                                      cfs_time_seconds(1));
1317         }
1318
1319         lustre_hash_for_each_safe(qctxt->lqc_lqs_hash, hash_put_lqs, NULL);
1320         l_wait_event(qctxt->lqc_lqs_waitq, check_lqs(qctxt), &lwi);
1321         down_write(&obt->obt_rwsem);
1322         lustre_hash_exit(qctxt->lqc_lqs_hash);
1323         qctxt->lqc_lqs_hash = NULL;
1324         up_write(&obt->obt_rwsem);
1325
1326         ptlrpcd_decref();
1327
1328 #ifdef LPROCFS
1329         if (lquota_proc_cleanup(qctxt))
1330                 CERROR("cleanup proc error!\n");
1331 #endif
1332
1333         EXIT;
1334 }
1335
1336 struct qslave_recov_thread_data {
1337         struct obd_device *obd;
1338         struct lustre_quota_ctxt *qctxt;
1339         struct completion comp;
1340 };
1341
1342 /* FIXME only recovery block quota by now */
1343 static int qslave_recovery_main(void *arg)
1344 {
1345         struct qslave_recov_thread_data *data = arg;
1346         struct obd_device *obd = data->obd;
1347         struct lustre_quota_ctxt *qctxt = data->qctxt;
1348         unsigned int type;
1349         int rc = 0;
1350         ENTRY;
1351
1352         cfs_daemonize_ctxt("qslave_recovd");
1353
1354         /* for obdfilter */
1355         class_incref(obd, "qslave_recovd_filter", obd);
1356
1357         complete(&data->comp);
1358
1359         spin_lock(&qctxt->lqc_lock);
1360         if (qctxt->lqc_recovery) {
1361                 spin_unlock(&qctxt->lqc_lock);
1362                 class_decref(obd, "qslave_recovd_filter", obd);
1363                 RETURN(0);
1364         } else {
1365                 qctxt->lqc_recovery = 1;
1366                 spin_unlock(&qctxt->lqc_lock);
1367         }
1368
1369         for (type = USRQUOTA; type < MAXQUOTAS; type++) {
1370                 struct qunit_data qdata;
1371                 struct quota_info *dqopt = sb_dqopt(qctxt->lqc_sb);
1372                 struct list_head id_list;
1373                 struct dquot_id *dqid, *tmp;
1374                 int ret;
1375
1376                 LOCK_DQONOFF_MUTEX(dqopt);
1377                 if (!ll_sb_has_quota_active(qctxt->lqc_sb, type)) {
1378                         UNLOCK_DQONOFF_MUTEX(dqopt);
1379                         break;
1380                 }
1381
1382                 LASSERT(dqopt->files[type] != NULL);
1383                 CFS_INIT_LIST_HEAD(&id_list);
1384 #ifndef KERNEL_SUPPORTS_QUOTA_READ
1385                 rc = fsfilt_qids(obd, dqopt->files[type], NULL, type, &id_list);
1386 #else
1387                 rc = fsfilt_qids(obd, NULL, dqopt->files[type], type, &id_list);
1388 #endif
1389                 UNLOCK_DQONOFF_MUTEX(dqopt);
1390                 if (rc)
1391                         CERROR("Get ids from quota file failed. (rc:%d)\n", rc);
1392
1393                 list_for_each_entry_safe(dqid, tmp, &id_list, di_link) {
1394                         list_del_init(&dqid->di_link);
1395                         /* skip slave recovery on itself */
1396                         if (is_master(qctxt))
1397                                 goto free;
1398                         if (rc && rc != -EBUSY)
1399                                 goto free;
1400
1401                         qdata.qd_id = dqid->di_id;
1402                         qdata.qd_flags = type;
1403                         QDATA_SET_BLK(&qdata);
1404                         qdata.qd_count = 0;
1405
1406                         ret = check_cur_qunit(obd, qctxt, &qdata);
1407                         if (ret > 0) {
1408                                 int opc;
1409                                 opc = ret == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1410                                 rc = schedule_dqacq(obd, qctxt, &qdata, opc,
1411                                                     0, NULL);
1412                                 if (rc == -EDQUOT)
1413                                         rc = 0;
1414                         } else {
1415                                 rc = 0;
1416                         }
1417
1418                         if (rc)
1419                                 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
1420                                        "qslave recovery failed! (id:%d type:%d "
1421                                        " rc:%d)\n", dqid->di_id, type, rc);
1422 free:
1423                         OBD_FREE_PTR(dqid);
1424                 }
1425         }
1426
1427         spin_lock(&qctxt->lqc_lock);
1428         qctxt->lqc_recovery = 0;
1429         spin_unlock(&qctxt->lqc_lock);
1430         class_decref(obd, "qslave_recovd_filter", obd);
1431         RETURN(rc);
1432 }
1433
1434 void
1435 qslave_start_recovery(struct obd_device *obd, struct lustre_quota_ctxt *qctxt)
1436 {
1437         struct qslave_recov_thread_data data;
1438         int rc;
1439         ENTRY;
1440
1441         if (!ll_sb_any_quota_active(qctxt->lqc_sb))
1442                 goto exit;
1443
1444         data.obd = obd;
1445         data.qctxt = qctxt;
1446         init_completion(&data.comp);
1447
1448         rc = kernel_thread(qslave_recovery_main, &data, CLONE_VM|CLONE_FILES);
1449         if (rc < 0) {
1450                 CERROR("Cannot start quota recovery thread: rc %d\n", rc);
1451                 goto exit;
1452         }
1453         wait_for_completion(&data.comp);
1454 exit:
1455         EXIT;
1456 }
1457
1458 int quota_is_on(struct lustre_quota_ctxt *qctxt, struct obd_quotactl *oqctl)
1459 {
1460         unsigned int type;
1461
1462         for (type = USRQUOTA; type < MAXQUOTAS; type++) {
1463                 if (!Q_TYPESET(oqctl, type))
1464                         continue;
1465                 if (!(qctxt->lqc_flags & UGQUOTA2LQC(oqctl->qc_type)))
1466                         return 0;
1467         }
1468         return 1;
1469 }
1470
1471 int quota_is_off(struct lustre_quota_ctxt *qctxt, struct obd_quotactl *oqctl)
1472 {
1473         unsigned int type;
1474
1475         for (type = USRQUOTA; type < MAXQUOTAS; type++) {
1476                 if (!Q_TYPESET(oqctl, type))
1477                         continue;
1478                 if (qctxt->lqc_flags & UGQUOTA2LQC(oqctl->qc_type))
1479                         return 0;
1480         }
1481         return 1;
1482 }
1483
1484 /**
1485  * When quotaon, build a lqs for every uid/gid who has been set limitation
1486  * for quota. After quota_search_lqs, it will hold one ref for the lqs.
1487  * It will be released when qctxt_cleanup() is executed b=18574
1488  *
1489  * Should be called with obt->obt_quotachecking held. b=20152
1490  */
1491 void build_lqs(struct obd_device *obd)
1492 {
1493         struct obd_device_target *obt = &obd->u.obt;
1494         struct lustre_quota_ctxt *qctxt = &obt->obt_qctxt;
1495         struct list_head id_list;
1496         int i, rc;
1497
1498         LASSERT_SEM_LOCKED(&obt->obt_quotachecking);
1499         INIT_LIST_HEAD(&id_list);
1500         for (i = 0; i < MAXQUOTAS; i++) {
1501                 struct dquot_id *dqid, *tmp;
1502
1503                 if (sb_dqopt(qctxt->lqc_sb)->files[i] == NULL)
1504                         continue;
1505
1506 #ifndef KERNEL_SUPPORTS_QUOTA_READ
1507                 rc = fsfilt_qids(obd, sb_dqopt(qctxt->lqc_sb)->files[i], NULL,
1508                                  i, &id_list);
1509 #else
1510                 rc = fsfilt_qids(obd, NULL, sb_dqopt(qctxt->lqc_sb)->files[i],
1511                                  i, &id_list);
1512 #endif
1513                 if (rc) {
1514                         CDEBUG(D_ERROR, "fail to get %s qids!\n",
1515                                i ? "group" : "user");
1516                         continue;
1517                 }
1518
1519                 list_for_each_entry_safe(dqid, tmp, &id_list,
1520                                          di_link) {
1521                         struct lustre_qunit_size *lqs;
1522
1523                         list_del_init(&dqid->di_link);
1524                         lqs = quota_search_lqs(LQS_KEY(i, dqid->di_id),
1525                                                qctxt, 1);
1526                         if (lqs && !IS_ERR(lqs)) {
1527                                 lqs->lqs_flags |= dqid->di_flag;
1528                                 lqs_putref(lqs);
1529                         } else {
1530                                 CDEBUG(D_ERROR, "fail to create a lqs"
1531                                        "(%s id: %u)!\n", i ? "group" : "user",
1532                                        dqid->di_id);
1533                         }
1534
1535                         OBD_FREE_PTR(dqid);
1536                 }
1537         }
1538 }
1539
1540 /**
1541  * lqs<->qctxt hash operations
1542  */
1543
1544 /**
1545  * string hashing using djb2 hash algorithm
1546  */
1547 static unsigned
1548 lqs_hash(lustre_hash_t *lh, void *key, unsigned mask)
1549 {
1550         struct quota_adjust_qunit *lqs_key;
1551         unsigned hash;
1552         ENTRY;
1553
1554         LASSERT(key);
1555         lqs_key = (struct quota_adjust_qunit *)key;
1556         hash = (QAQ_IS_GRP(lqs_key) ? 5381 : 5387) * lqs_key->qaq_id;
1557
1558         RETURN(hash & mask);
1559 }
1560
1561 static int
1562 lqs_compare(void *key, struct hlist_node *hnode)
1563 {
1564         struct lustre_qunit_size *q;
1565         int rc;
1566         ENTRY;
1567
1568         LASSERT(key);
1569         q = hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1570
1571         spin_lock(&q->lqs_lock);
1572         rc = (q->lqs_key == *((unsigned long long *)key));
1573         spin_unlock(&q->lqs_lock);
1574
1575         RETURN(rc);
1576 }
1577
1578 static void *
1579 lqs_get(struct hlist_node *hnode)
1580 {
1581         struct lustre_qunit_size *q =
1582                 hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1583         ENTRY;
1584
1585         __lqs_getref(q);
1586
1587         RETURN(q);
1588 }
1589
1590 static void *
1591 lqs_put(struct hlist_node *hnode)
1592 {
1593         struct lustre_qunit_size *q =
1594                 hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1595         ENTRY;
1596
1597         __lqs_putref(q);
1598
1599         RETURN(q);
1600 }
1601
1602 static void
1603 lqs_exit(struct hlist_node *hnode)
1604 {
1605         struct lustre_qunit_size *q =
1606                 hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1607         ENTRY;
1608
1609         /*
1610          * Nothing should be left. User of lqs put it and
1611          * lqs also was deleted from table by this time
1612          * so we should have 0 refs.
1613          */
1614         LASSERTF(atomic_read(&q->lqs_refcount) == 0,
1615                  "Busy lqs %p with %d refs\n", q,
1616                  atomic_read(&q->lqs_refcount));
1617         OBD_FREE_PTR(q);
1618         EXIT;
1619 }
1620
1621 static lustre_hash_ops_t lqs_hash_ops = {
1622         .lh_hash    = lqs_hash,
1623         .lh_compare = lqs_compare,
1624         .lh_get     = lqs_get,
1625         .lh_put     = lqs_put,
1626         .lh_exit    = lqs_exit
1627 };
1628 #endif /* HAVE_QUOTA_SUPPORT */