lustre/quota/quota_context.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  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 <lprocfs_status.h>
  61 #include "quota_internal.h"
  62
  63 static int hash_lqs_cur_bits = HASH_LQS_CUR_BITS;
  64 CFS_MODULE_PARM(hash_lqs_cur_bits, "i", int, 0444,
  65                 "the current bits of lqs hash");
  66
  67 static cfs_hash_ops_t lqs_hash_ops;
  68
  69 unsigned long default_bunit_sz = 128 * 1024 * 1024; /* 128M bytes */
  70 unsigned long default_btune_ratio = 50;             /* 50 percentage */
  71 unsigned long default_iunit_sz = 5120;              /* 5120 inodes */
  72 unsigned long default_itune_ratio = 50;             /* 50 percentage */
  73
  74 cfs_mem_cache_t *qunit_cachep = NULL;
  75 cfs_list_t qunit_hash[NR_DQHASH];
  76 cfs_spinlock_t qunit_hash_lock = CFS_SPIN_LOCK_UNLOCKED;
  77
  78 /* please sync qunit_state with qunit_state_names */
  79 enum qunit_state {
  80         /**
  81          * a qunit is created
  82          */
  83         QUNIT_CREATED      = 0,
  84         /**
  85          * a qunit is added into qunit hash, that means
  86          * a quota req will be sent or is flying
  87          */
  88         QUNIT_IN_HASH      = 1,
  89         /**
  90          * a qunit is removed from qunit hash, that
  91          * means a quota req is handled and comes back
  92          */
  93         QUNIT_RM_FROM_HASH = 2,
  94         /**
  95          * qunit can wake up all threads waiting for it
  96          */
  97         QUNIT_FINISHED     = 3,
  98 };
  99
 100 static const char *qunit_state_names[] = {
 101         [QUNIT_CREATED]      = "CREATED",
 102         [QUNIT_IN_HASH]      = "IN_HASH",
 103         [QUNIT_RM_FROM_HASH] = "RM_FROM_HASH",
 104         [QUNIT_FINISHED]     = "FINISHED",
 105 };
 106
 107 struct lustre_qunit {
 108         cfs_list_t lq_hash;      /** Hash list in memory */
 109         cfs_atomic_t lq_refcnt;            /** Use count */
 110         struct lustre_quota_ctxt *lq_ctxt; /** Quota context this applies to */
 111         struct qunit_data lq_data;         /** See qunit_data */
 112         unsigned int lq_opc;               /** QUOTA_DQACQ, QUOTA_DQREL */
 113         cfs_waitq_t lq_waitq;              /** Threads waiting for this qunit */
 114         cfs_spinlock_t lq_lock;            /** Protect the whole structure */
 115         enum qunit_state lq_state;         /** Present the status of qunit */
 116         int lq_rc;                         /** The rc of lq_data */
 117         pid_t lq_owner;
 118 };
 119
 120 #define QUNIT_SET_STATE(qunit, state)                                   \
 121 do {                                                                    \
 122         cfs_spin_lock(&qunit->lq_lock);                                 \
 123         QDATA_DEBUG((&qunit->lq_data), "qunit(%p) lq_state(%s->%s), "   \
 124                     "lq_rc(%d), lq_owner(%d)\n",                        \
 125                     qunit, qunit_state_names[qunit->lq_state],          \
 126                     qunit_state_names[state], qunit->lq_rc,             \
 127                     qunit->lq_owner);                                   \
 128         qunit->lq_state = state;                                        \
 129         cfs_spin_unlock(&qunit->lq_lock);                               \
 130 } while(0)
 131
 132 #define QUNIT_SET_STATE_AND_RC(qunit, state, rc)                        \
 133 do {                                                                    \
 134         cfs_spin_lock(&qunit->lq_lock);                                 \
 135         qunit->lq_rc = rc;                                              \
 136         QDATA_DEBUG((&qunit->lq_data), "qunit(%p) lq_state(%s->%s), "   \
 137                     "lq_rc(%d), lq_owner(%d)\n",                        \
 138                     qunit, qunit_state_names[qunit->lq_state],          \
 139                     qunit_state_names[state], qunit->lq_rc,             \
 140                     qunit->lq_owner);                                   \
 141         qunit->lq_state = state;                                        \
 142         cfs_spin_unlock(&qunit->lq_lock);                               \
 143 } while(0)
 144
 145 int should_translate_quota (struct obd_import *imp)
 146 {
 147         ENTRY;
 148
 149         LASSERT(imp);
 150         if (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_QUOTA64)
 151                 RETURN(0);
 152         else
 153                 RETURN(1);
 154 }
 155
 156 void qunit_cache_cleanup(void)
 157 {
 158         int i;
 159         ENTRY;
 160
 161         cfs_spin_lock(&qunit_hash_lock);
 162         for (i = 0; i < NR_DQHASH; i++)
 163                 LASSERT(cfs_list_empty(qunit_hash + i));
 164         cfs_spin_unlock(&qunit_hash_lock);
 165
 166         if (qunit_cachep) {
 167                 int rc;
 168                 rc = cfs_mem_cache_destroy(qunit_cachep);
 169                 LASSERTF(rc == 0, "couldn't destroy qunit_cache slab\n");
 170                 qunit_cachep = NULL;
 171         }
 172         EXIT;
 173 }
 174
 175 int qunit_cache_init(void)
 176 {
 177         int i;
 178         ENTRY;
 179
 180         LASSERT(qunit_cachep == NULL);
 181         qunit_cachep = cfs_mem_cache_create("ll_qunit_cache",
 182                                             sizeof(struct lustre_qunit),
 183                                             0, 0);
 184         if (!qunit_cachep)
 185                 RETURN(-ENOMEM);
 186
 187         cfs_spin_lock(&qunit_hash_lock);
 188         for (i = 0; i < NR_DQHASH; i++)
 189                 CFS_INIT_LIST_HEAD(qunit_hash + i);
 190         cfs_spin_unlock(&qunit_hash_lock);
 191         RETURN(0);
 192 }
 193
 194 static inline int
 195 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
 196              __attribute__((__const__));
 197
 198 static inline int
 199 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
 200 {
 201         unsigned int id = qdata->qd_id;
 202         unsigned int type = QDATA_IS_GRP(qdata);
 203
 204         unsigned long tmp = ((unsigned long)qctxt >> L1_CACHE_SHIFT) ^ id;
 205         tmp = (tmp * (MAXQUOTAS - type)) % NR_DQHASH;
 206         return tmp;
 207 }
 208
 209 /* caller must hold qunit_hash_lock */
 210 static inline struct lustre_qunit *find_qunit(unsigned int hashent,
 211                                               struct lustre_quota_ctxt *qctxt,
 212                                               struct qunit_data *qdata)
 213 {
 214         struct lustre_qunit *qunit = NULL;
 215         struct qunit_data *tmp;
 216
 217         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
 218         cfs_list_for_each_entry(qunit, qunit_hash + hashent, lq_hash) {
 219                 tmp = &qunit->lq_data;
 220                 if (qunit->lq_ctxt == qctxt &&
 221                     qdata->qd_id == tmp->qd_id &&
 222                     (qdata->qd_flags & LQUOTA_QUNIT_FLAGS) ==
 223                     (tmp->qd_flags & LQUOTA_QUNIT_FLAGS))
 224                         return qunit;
 225         }
 226         return NULL;
 227 }
 228
 229 /* check_cur_qunit - check the current usage of qunit.
 230  * @qctxt: quota context
 231  * @qdata: the type of quota unit to be checked
 232  *
 233  * return: 1 - need acquire qunit;
 234  *         2 - need release qunit;
 235  *         0 - need do nothing.
 236  *       < 0 - error.
 237  */
 238 static int
 239 check_cur_qunit(struct obd_device *obd,
 240                 struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
 241 {
 242         struct super_block *sb = qctxt->lqc_sb;
 243         unsigned long qunit_sz, tune_sz;
 244         __u64 usage, limit, limit_org, pending_write = 0;
 245         long long record = 0;
 246         struct obd_quotactl *qctl;
 247         struct lustre_qunit_size *lqs = NULL;
 248         int ret = 0;
 249         ENTRY;
 250
 251         if (!ll_sb_any_quota_active(sb))
 252                 RETURN(0);
 253
 254         cfs_spin_lock(&qctxt->lqc_lock);
 255         if (!qctxt->lqc_valid){
 256                 cfs_spin_unlock(&qctxt->lqc_lock);
 257                 RETURN(0);
 258         }
 259         cfs_spin_unlock(&qctxt->lqc_lock);
 260
 261         OBD_ALLOC_PTR(qctl);
 262         if (qctl == NULL)
 263                 RETURN(-ENOMEM);
 264
 265         /* get fs quota usage & limit */
 266         qctl->qc_cmd = Q_GETQUOTA;
 267         qctl->qc_id = qdata->qd_id;
 268         qctl->qc_type = QDATA_IS_GRP(qdata);
 269         ret = fsfilt_quotactl(obd, sb, qctl);
 270         if (ret) {
 271                 if (ret == -ESRCH)      /* no limit */
 272                         ret = 0;
 273                 else
 274                         CERROR("can't get fs quota usage! (rc:%d)\n", ret);
 275                 GOTO(out, ret);
 276         }
 277
 278         if (QDATA_IS_BLK(qdata)) {
 279                 usage = qctl->qc_dqblk.dqb_curspace;
 280                 limit = qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS;
 281         } else {
 282                 usage = qctl->qc_dqblk.dqb_curinodes;
 283                 limit = qctl->qc_dqblk.dqb_ihardlimit;
 284         }
 285
 286         /* ignore the no quota limit case; and it can avoid creating
 287          * unnecessary lqs for uid/gid */
 288         if (!limit)
 289                 GOTO(out, ret = 0);
 290
 291         lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(qdata), qdata->qd_id),
 292                                qctxt, 0);
 293         if (IS_ERR(lqs) || lqs == NULL) {
 294                 CERROR("fail to find a lqs for %sid: %u)!\n",
 295                        QDATA_IS_GRP(qdata) ? "g" : "u", qdata->qd_id);
 296                 GOTO (out, ret = 0);
 297         }
 298         cfs_spin_lock(&lqs->lqs_lock);
 299
 300         if (QDATA_IS_BLK(qdata)) {
 301                 qunit_sz = lqs->lqs_bunit_sz;
 302                 tune_sz  = lqs->lqs_btune_sz;
 303                 pending_write = lqs->lqs_bwrite_pending;
 304                 record   = lqs->lqs_blk_rec;
 305                 LASSERT(!(qunit_sz % QUOTABLOCK_SIZE));
 306         } else {
 307                 /* we didn't need change inode qunit size now */
 308                 qunit_sz = lqs->lqs_iunit_sz;
 309                 tune_sz  = lqs->lqs_itune_sz;
 310                 pending_write = lqs->lqs_iwrite_pending;
 311                 record   = lqs->lqs_ino_rec;
 312         }
 313
 314         /* we don't count the MIN_QLIMIT */
 315         if ((limit == MIN_QLIMIT && !QDATA_IS_BLK(qdata)) ||
 316             (toqb(limit) == MIN_QLIMIT && QDATA_IS_BLK(qdata)))
 317                 limit = 0;
 318
 319         usage += pending_write;
 320         limit_org = limit;
 321         /* when a releasing quota req is sent, before it returned
 322            limit is assigned a small value. limit will overflow */
 323         if (record < 0)
 324                 usage -= record;
 325         else
 326                 limit += record;
 327
 328         LASSERT(qdata->qd_count == 0);
 329         if (limit <= usage + tune_sz) {
 330                 while (qdata->qd_count + limit <=
 331                        usage + tune_sz)
 332                         qdata->qd_count += qunit_sz;
 333                 ret = 1;
 334         } else if (limit > usage + qunit_sz + tune_sz &&
 335                    limit_org > qdata->qd_count + qunit_sz) {
 336                 while (limit - qdata->qd_count > usage + qunit_sz + tune_sz &&
 337                        limit_org > qdata->qd_count + qunit_sz)
 338                         qdata->qd_count += qunit_sz;
 339                 ret = 2;
 340                 /* if there are other pending writes for this uid/gid, releasing
 341                  * quota is put off until the last pending write b=16645 */
 342                 /* if there is an ongoing quota request, a releasing request is aborted.
 343                  * That ongoing quota request will call this function again when
 344                  * it returned b=18630 */
 345                 if (pending_write || record) {
 346                         CDEBUG(D_QUOTA, "delay quota release\n");
 347                         ret = 0;
 348                 }
 349         }
 350         if (ret > 0)
 351                 quota_compute_lqs(qdata, lqs, 1, (ret == 1) ? 1 : 0);
 352
 353         CDEBUG(D_QUOTA, "type: %c, limit: "LPU64", usage: "LPU64
 354                ", pending_write: "LPU64", record: %lld"
 355                ", qunit_sz: %lu, tune_sz: %lu, ret: %d.\n",
 356                QDATA_IS_BLK(qdata) ? 'b' : 'i', limit, usage, pending_write,
 357                record, qunit_sz, tune_sz, ret);
 358         LASSERT(ret == 0 || qdata->qd_count);
 359
 360         cfs_spin_unlock(&lqs->lqs_lock);
 361         lqs_putref(lqs);
 362
 363         EXIT;
 364  out:
 365         OBD_FREE_PTR(qctl);
 366         return ret;
 367 }
 368
 369 /**
 370  * Compute the remaining quota for certain gid or uid b=11693
 371  */
 372 int compute_remquota(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 373                      struct qunit_data *qdata, int isblk)
 374 {
 375         struct super_block *sb = qctxt->lqc_sb;
 376         __u64 usage, limit;
 377         struct obd_quotactl *qctl;
 378         int ret = QUOTA_RET_OK;
 379         ENTRY;
 380
 381         if (!ll_sb_any_quota_active(sb))
 382                 RETURN(QUOTA_RET_NOQUOTA);
 383
 384         /* ignore root user */
 385         if (qdata->qd_id == 0 && QDATA_IS_GRP(qdata) == USRQUOTA)
 386                 RETURN(QUOTA_RET_NOLIMIT);
 387
 388         OBD_ALLOC_PTR(qctl);
 389         if (qctl == NULL)
 390                 RETURN(-ENOMEM);
 391
 392         /* get fs quota usage & limit */
 393         qctl->qc_cmd = Q_GETQUOTA;
 394         qctl->qc_id = qdata->qd_id;
 395         qctl->qc_type = QDATA_IS_GRP(qdata);
 396         ret = fsfilt_quotactl(obd, sb, qctl);
 397         if (ret) {
 398                 if (ret == -ESRCH)      /* no limit */
 399                         ret = QUOTA_RET_NOLIMIT;
 400                 else
 401                         CDEBUG(D_QUOTA, "can't get fs quota usage! (rc:%d)",
 402                                ret);
 403                 GOTO(out, ret);
 404         }
 405
 406         usage = isblk ? qctl->qc_dqblk.dqb_curspace :
 407                 qctl->qc_dqblk.dqb_curinodes;
 408         limit = isblk ? qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS :
 409                 qctl->qc_dqblk.dqb_ihardlimit;
 410         if (!limit){            /* no limit */
 411                 ret = QUOTA_RET_NOLIMIT;
 412                 GOTO(out, ret);
 413         }
 414
 415         if (limit >= usage)
 416                 qdata->qd_count = limit - usage;
 417         else
 418                 qdata->qd_count = 0;
 419         EXIT;
 420 out:
 421         OBD_FREE_PTR(qctl);
 422         return ret;
 423 }
 424
 425 static struct lustre_qunit *alloc_qunit(struct lustre_quota_ctxt *qctxt,
 426                                         struct qunit_data *qdata, int opc)
 427 {
 428         struct lustre_qunit *qunit = NULL;
 429         ENTRY;
 430
 431         OBD_SLAB_ALLOC_PTR_GFP(qunit, qunit_cachep, CFS_ALLOC_IO);
 432         if (qunit == NULL)
 433                 RETURN(NULL);
 434
 435         CFS_INIT_LIST_HEAD(&qunit->lq_hash);
 436         cfs_waitq_init(&qunit->lq_waitq);
 437         cfs_atomic_set(&qunit->lq_refcnt, 1);
 438         qunit->lq_ctxt = qctxt;
 439         memcpy(&qunit->lq_data, qdata, sizeof(*qdata));
 440         qunit->lq_opc = opc;
 441         qunit->lq_lock = CFS_SPIN_LOCK_UNLOCKED;
 442         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_CREATED, 0);
 443         qunit->lq_owner = cfs_curproc_pid();
 444         RETURN(qunit);
 445 }
 446
 447 static inline void free_qunit(struct lustre_qunit *qunit)
 448 {
 449         OBD_SLAB_FREE(qunit, qunit_cachep, sizeof(*qunit));
 450 }
 451
 452 static inline void qunit_get(struct lustre_qunit *qunit)
 453 {
 454         cfs_atomic_inc(&qunit->lq_refcnt);
 455 }
 456
 457 static void qunit_put(struct lustre_qunit *qunit)
 458 {
 459         LASSERT(cfs_atomic_read(&qunit->lq_refcnt));
 460         if (cfs_atomic_dec_and_test(&qunit->lq_refcnt))
 461                 free_qunit(qunit);
 462 }
 463
 464 /* caller must hold qunit_hash_lock and release ref of qunit after using it */
 465 static struct lustre_qunit *dqacq_in_flight(struct lustre_quota_ctxt *qctxt,
 466                                             struct qunit_data *qdata)
 467 {
 468         unsigned int hashent = qunit_hashfn(qctxt, qdata);
 469         struct lustre_qunit *qunit;
 470         ENTRY;
 471
 472         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
 473         qunit = find_qunit(hashent, qctxt, qdata);
 474         if (qunit)
 475                 qunit_get(qunit);
 476         RETURN(qunit);
 477 }
 478
 479 static void
 480 insert_qunit_nolock(struct lustre_quota_ctxt *qctxt, struct lustre_qunit *qunit)
 481 {
 482         cfs_list_t *head;
 483
 484         LASSERT(cfs_list_empty(&qunit->lq_hash));
 485         qunit_get(qunit);
 486         head = qunit_hash + qunit_hashfn(qctxt, &qunit->lq_data);
 487         cfs_list_add(&qunit->lq_hash, head);
 488         QUNIT_SET_STATE(qunit, QUNIT_IN_HASH);
 489 }
 490
 491 static void compute_lqs_after_removing_qunit(struct lustre_qunit *qunit)
 492 {
 493         struct lustre_qunit_size *lqs;
 494
 495         lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(&qunit->lq_data),
 496                                        qunit->lq_data.qd_id),
 497                                qunit->lq_ctxt, 0);
 498         if (lqs && !IS_ERR(lqs)) {
 499                 cfs_spin_lock(&lqs->lqs_lock);
 500                 if (qunit->lq_opc == QUOTA_DQACQ)
 501                         quota_compute_lqs(&qunit->lq_data, lqs, 0, 1);
 502                 if (qunit->lq_opc == QUOTA_DQREL)
 503                         quota_compute_lqs(&qunit->lq_data, lqs, 0, 0);
 504                 cfs_spin_unlock(&lqs->lqs_lock);
 505                 /* this is for quota_search_lqs */
 506                 lqs_putref(lqs);
 507                 /* this is for schedule_dqacq */
 508                 lqs_putref(lqs);
 509         }
 510 }
 511
 512 static void remove_qunit_nolock(struct lustre_qunit *qunit)
 513 {
 514         LASSERT(!cfs_list_empty(&qunit->lq_hash));
 515         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
 516
 517         cfs_list_del_init(&qunit->lq_hash);
 518         QUNIT_SET_STATE(qunit, QUNIT_RM_FROM_HASH);
 519         qunit_put(qunit);
 520 }
 521
 522 void* quota_barrier(struct lustre_quota_ctxt *qctxt,
 523                     struct obd_quotactl *oqctl, int isblk)
 524 {
 525         struct lustre_qunit *qunit, *find_qunit;
 526         int cycle = 1;
 527
 528         OBD_SLAB_ALLOC(qunit, qunit_cachep, CFS_ALLOC_IO, sizeof(*qunit));
 529         if (qunit == NULL) {
 530                 CERROR("locating %sunit failed for %sid %u\n",
 531                        isblk ? "b" : "i", oqctl->qc_type ? "g" : "u",
 532                        oqctl->qc_id);
 533                 qctxt_wait_pending_dqacq(qctxt, oqctl->qc_id,
 534                                          oqctl->qc_type, isblk);
 535                 return NULL;
 536         }
 537
 538         CFS_INIT_LIST_HEAD(&qunit->lq_hash);
 539         qunit->lq_lock = CFS_SPIN_LOCK_UNLOCKED;
 540         cfs_waitq_init(&qunit->lq_waitq);
 541         cfs_atomic_set(&qunit->lq_refcnt, 1);
 542         qunit->lq_ctxt = qctxt;
 543         qunit->lq_data.qd_id = oqctl->qc_id;
 544         qunit->lq_data.qd_flags =  oqctl->qc_type;
 545         if (isblk)
 546                 QDATA_SET_BLK(&qunit->lq_data);
 547         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_CREATED, 0);
 548         /* it means it is only an invalid qunit for barrier */
 549         qunit->lq_opc = QUOTA_LAST_OPC;
 550
 551         while (1) {
 552                 cfs_spin_lock(&qunit_hash_lock);
 553                 find_qunit = dqacq_in_flight(qctxt, &qunit->lq_data);
 554                 if (find_qunit) {
 555                         cfs_spin_unlock(&qunit_hash_lock);
 556                         qunit_put(find_qunit);
 557                         qctxt_wait_pending_dqacq(qctxt, oqctl->qc_id,
 558                                                  oqctl->qc_type, isblk);
 559                         CDEBUG(D_QUOTA, "cycle=%d\n", cycle++);
 560                         continue;
 561                 }
 562                 break;
 563         }
 564         insert_qunit_nolock(qctxt, qunit);
 565         cfs_spin_unlock(&qunit_hash_lock);
 566         return qunit;
 567 }
 568
 569 void quota_unbarrier(void *handle)
 570 {
 571         struct lustre_qunit *qunit = (struct lustre_qunit *)handle;
 572
 573         if (qunit == NULL) {
 574                 CERROR("handle is NULL\n");
 575                 return;
 576         }
 577
 578         LASSERT(qunit->lq_opc == QUOTA_LAST_OPC);
 579         cfs_spin_lock(&qunit_hash_lock);
 580         remove_qunit_nolock(qunit);
 581         cfs_spin_unlock(&qunit_hash_lock);
 582         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, QUOTA_REQ_RETURNED);
 583         cfs_waitq_signal(&qunit->lq_waitq);
 584         qunit_put(qunit);
 585 }
 586
 587 #define INC_QLIMIT(limit, count) (limit == MIN_QLIMIT) ? \
 588                                  (limit = count) : (limit += count)
 589
 590
 591 static inline int is_master(struct lustre_quota_ctxt *qctxt)
 592 {
 593         return qctxt->lqc_handler ? 1 : 0;
 594 }
 595
 596 static int
 597 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 598                struct qunit_data *qdata, int opc, int wait,
 599                struct obd_trans_info *oti);
 600
 601 static inline void qdata_to_oqaq(struct qunit_data *qdata,
 602                                  struct quota_adjust_qunit *oqaq)
 603 {
 604         LASSERT(qdata);
 605         LASSERT(oqaq);
 606
 607         oqaq->qaq_flags = qdata->qd_flags;
 608         oqaq->qaq_id    = qdata->qd_id;
 609         if (QDATA_IS_ADJBLK(qdata))
 610                 oqaq->qaq_bunit_sz = qdata->qd_qunit;
 611         if (QDATA_IS_ADJINO(qdata))
 612                 oqaq->qaq_iunit_sz = qdata->qd_qunit;
 613 }
 614
 615 static int
 616 dqacq_completion(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 617                  struct qunit_data *qdata, int rc, int opc)
 618 {
 619         struct lustre_qunit *qunit = NULL;
 620         struct super_block *sb = qctxt->lqc_sb;
 621         int err = 0;
 622         struct quota_adjust_qunit *oqaq = NULL;
 623         int rc1 = 0;
 624         ENTRY;
 625
 626         LASSERT(qdata);
 627         QDATA_DEBUG(qdata, "obd(%s): complete %s quota req\n",
 628                     obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
 629
 630         /* do it only when a releasing quota req more than 5MB b=18491 */
 631         if (opc == QUOTA_DQREL && qdata->qd_count >= 5242880)
 632                 OBD_FAIL_TIMEOUT(OBD_FAIL_QUOTA_DELAY_REL, 5);
 633
 634         /* update local operational quota file */
 635         if (rc == 0) {
 636                 __u64 count = QUSG(qdata->qd_count, QDATA_IS_BLK(qdata));
 637                 struct obd_quotactl *qctl;
 638                 __u64 *hardlimit;
 639
 640                 OBD_ALLOC_PTR(qctl);
 641                 if (qctl == NULL)
 642                         GOTO(out, err = -ENOMEM);
 643
 644                 /* acq/rel qunit for specified uid/gid is serialized,
 645                  * so there is no race between get fs quota limit and
 646                  * set fs quota limit */
 647                 qctl->qc_cmd = Q_GETQUOTA;
 648                 qctl->qc_id = qdata->qd_id;
 649                 qctl->qc_type = QDATA_IS_GRP(qdata);
 650                 err = fsfilt_quotactl(obd, sb, qctl);
 651                 if (err) {
 652                         CERROR("error get quota fs limit! (rc:%d)\n", err);
 653                         GOTO(out_mem, err);
 654                 }
 655
 656                 if (QDATA_IS_BLK(qdata)) {
 657                         qctl->qc_dqblk.dqb_valid = QIF_BLIMITS;
 658                         hardlimit = &qctl->qc_dqblk.dqb_bhardlimit;
 659                 } else {
 660                         qctl->qc_dqblk.dqb_valid = QIF_ILIMITS;
 661                         hardlimit = &qctl->qc_dqblk.dqb_ihardlimit;
 662                 }
 663
 664                 CDEBUG(D_QUOTA, "hardlimt: "LPU64"\n", *hardlimit);
 665
 666                 if (*hardlimit == 0)
 667                         goto out_mem;
 668
 669                 switch (opc) {
 670                 case QUOTA_DQACQ:
 671                         INC_QLIMIT(*hardlimit, count);
 672                         break;
 673                 case QUOTA_DQREL:
 674                         LASSERTF(count < *hardlimit,
 675                                  "id(%u) flag(%u) type(%c) isblk(%c) "
 676                                  "count("LPU64") qd_qunit("LPU64") "
 677                                  "hardlimit("LPU64").\n",
 678                                  qdata->qd_id, qdata->qd_flags,
 679                                  QDATA_IS_GRP(qdata) ? 'g' : 'u',
 680                                  QDATA_IS_BLK(qdata) ? 'b': 'i',
 681                                  qdata->qd_count, qdata->qd_qunit, *hardlimit);
 682                         *hardlimit -= count;
 683                         break;
 684                 default:
 685                         LBUG();
 686                 }
 687
 688                 /* clear quota limit */
 689                 if (count == 0)
 690                         *hardlimit = 0;
 691
 692                 qctl->qc_cmd = Q_SETQUOTA;
 693                 err = fsfilt_quotactl(obd, sb, qctl);
 694                 if (err)
 695                         CERROR("error set quota fs limit! (rc:%d)\n", err);
 696
 697                 QDATA_DEBUG(qdata, "%s completion\n",
 698                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 699 out_mem:
 700                 OBD_FREE_PTR(qctl);
 701         } else if (rc == -EDQUOT) {
 702                 QDATA_DEBUG(qdata, "acquire qunit got EDQUOT.\n");
 703         } else if (rc == -EBUSY) {
 704                 QDATA_DEBUG(qdata, "it's is recovering, got EBUSY.\n");
 705         } else {
 706                 CERROR("acquire qunit got error! (rc:%d)\n", rc);
 707         }
 708 out:
 709         /* remove the qunit from hash */
 710         cfs_spin_lock(&qunit_hash_lock);
 711
 712         qunit = dqacq_in_flight(qctxt, qdata);
 713         /* this qunit has been removed by qctxt_cleanup() */
 714         if (!qunit) {
 715                 cfs_spin_unlock(&qunit_hash_lock);
 716                 QDATA_DEBUG(qdata, "%s is discarded because qunit isn't found\n",
 717                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 718                 RETURN(err);
 719         }
 720
 721         LASSERT(opc == qunit->lq_opc);
 722         /* remove this qunit from lq_hash so that new processes cannot be added
 723          * to qunit->lq_waiters */
 724         remove_qunit_nolock(qunit);
 725         cfs_spin_unlock(&qunit_hash_lock);
 726
 727         compute_lqs_after_removing_qunit(qunit);
 728
 729         if (rc == 0)
 730                 rc = QUOTA_REQ_RETURNED;
 731         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, rc);
 732         /* wake up all waiters */
 733         cfs_waitq_broadcast(&qunit->lq_waitq);
 734
 735         /* this is for dqacq_in_flight() */
 736         qunit_put(qunit);
 737         if (rc < 0 && rc != -EDQUOT)
 738                 GOTO(out1, err);
 739
 740         /* don't reschedule in such cases:
 741          *   - acq/rel failure and qunit isn't changed,
 742          *     but not for quota recovery.
 743          *   - local dqacq/dqrel.
 744          *   - local disk io failure.
 745          */
 746          OBD_ALLOC_PTR(oqaq);
 747          if (!oqaq)
 748                  GOTO(out1, err = -ENOMEM);
 749          qdata_to_oqaq(qdata, oqaq);
 750          /* adjust the qunit size in slaves */
 751          rc1 = quota_adjust_slave_lqs(oqaq, qctxt);
 752          OBD_FREE_PTR(oqaq);
 753          if (rc1 < 0) {
 754                  CERROR("adjust slave's qunit size failed!(rc:%d)\n", rc1);
 755                  GOTO(out1, err = rc1);
 756          }
 757          if (err || (rc < 0 && rc != -EBUSY && rc1 == 0) || is_master(qctxt))
 758                  GOTO(out1, err);
 759
 760          if (opc == QUOTA_DQREL && qdata->qd_count >= 5242880 &&
 761              OBD_FAIL_CHECK(OBD_FAIL_QUOTA_DELAY_REL))
 762                  GOTO(out1, err);
 763
 764         /* reschedule another dqacq/dqrel if needed */
 765         qdata->qd_count = 0;
 766         qdata->qd_flags &= LQUOTA_QUNIT_FLAGS;
 767         rc1 = check_cur_qunit(obd, qctxt, qdata);
 768         if (rc1 > 0) {
 769                 int opc;
 770                 opc = rc1 == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
 771                 rc1 = schedule_dqacq(obd, qctxt, qdata, opc, 0, NULL);
 772                 QDATA_DEBUG(qdata, "reschedudle opc(%d) rc(%d)\n", opc, rc1);
 773         }
 774  out1:
 775         /* this is for alloc_qunit() */
 776         qunit_put(qunit);
 777         RETURN(err);
 778 }
 779
 780 struct dqacq_async_args {
 781         struct lustre_quota_ctxt *aa_ctxt;
 782         struct lustre_qunit *aa_qunit;
 783 };
 784
 785 static int dqacq_interpret(const struct lu_env *env,
 786                            struct ptlrpc_request *req, void *data, int rc)
 787 {
 788         struct dqacq_async_args *aa = (struct dqacq_async_args *)data;
 789         struct lustre_quota_ctxt *qctxt = aa->aa_ctxt;
 790         struct obd_device_target *obt = qctxt->lqc_obt;
 791         struct lustre_qunit *qunit = aa->aa_qunit;
 792         struct obd_device *obd = req->rq_import->imp_obd;
 793         struct qunit_data *qdata = NULL;
 794         ENTRY;
 795
 796         LASSERT(req);
 797         LASSERT(req->rq_import);
 798
 799         cfs_down_read(&obt->obt_rwsem);
 800         /* if a quota req timeouts or is dropped, we should update quota
 801          * statistics which will be handled in dqacq_completion. And in
 802          * this situation we should get qdata from request instead of
 803          * reply */
 804         qdata = quota_get_qdata(req, (rc != 0) ? QUOTA_REQUEST : QUOTA_REPLY,
 805                                 QUOTA_IMPORT);
 806         if (IS_ERR(qdata)) {
 807                 rc = PTR_ERR(qdata);
 808                 DEBUG_REQ(D_ERROR, req,
 809                           "error unpacking qunit_data(rc: %ld)\n",
 810                           PTR_ERR(qdata));
 811                 qdata = &qunit->lq_data;
 812         }
 813
 814         QDATA_DEBUG(qdata, "qdata: interpret rc(%d).\n", rc);
 815         QDATA_DEBUG((&qunit->lq_data), "lq_data: \n");
 816
 817         if (qdata->qd_id != qunit->lq_data.qd_id ||
 818             OBD_FAIL_CHECK(OBD_FAIL_QUOTA_RET_QDATA)) {
 819                 CERROR("the returned qd_id isn't expected!"
 820                        "(qdata: %u, lq_data: %u)\n", qdata->qd_id,
 821                        qunit->lq_data.qd_id);
 822                 qdata->qd_id = qunit->lq_data.qd_id;
 823                 rc = -EPROTO;
 824         }
 825         if (QDATA_IS_GRP(qdata) != QDATA_IS_GRP(&qunit->lq_data)) {
 826                 CERROR("the returned grp/usr isn't expected!"
 827                        "(qdata: %u, lq_data: %u)\n", qdata->qd_flags,
 828                        qunit->lq_data.qd_flags);
 829                 if (QDATA_IS_GRP(&qunit->lq_data))
 830                         QDATA_SET_GRP(qdata);
 831                 else
 832                         QDATA_CLR_GRP(qdata);
 833                 rc = -EPROTO;
 834         }
 835         if (qdata->qd_count > qunit->lq_data.qd_count) {
 836                 CERROR("the returned qd_count isn't expected!"
 837                        "(qdata: "LPU64", lq_data: "LPU64")\n", qdata->qd_count,
 838                        qunit->lq_data.qd_count);
 839                 rc = -EPROTO;
 840         }
 841
 842         if (unlikely(rc == -ESRCH))
 843                 CERROR("quota for %s has been enabled by master, but disabled "
 844                        "by slave.\n", QDATA_IS_GRP(qdata) ? "group" : "user");
 845
 846         rc = dqacq_completion(obd, qctxt, qdata, rc,
 847                               lustre_msg_get_opc(req->rq_reqmsg));
 848
 849         cfs_up_read(&obt->obt_rwsem);
 850         RETURN(rc);
 851 }
 852
 853 /**
 854  * check if quota master is online
 855  */
 856 int check_qm(struct lustre_quota_ctxt *qctxt)
 857 {
 858         int rc;
 859         ENTRY;
 860
 861         cfs_spin_lock(&qctxt->lqc_lock);
 862         /* quit waiting when mds is back or qctxt is cleaned up */
 863         rc = qctxt->lqc_import || !qctxt->lqc_valid;
 864         cfs_spin_unlock(&qctxt->lqc_lock);
 865
 866         RETURN(rc);
 867 }
 868
 869 /* wake up all waiting threads when lqc_import is NULL */
 870 void dqacq_interrupt(struct lustre_quota_ctxt *qctxt)
 871 {
 872         struct lustre_qunit *qunit, *tmp;
 873         int i;
 874         ENTRY;
 875
 876         cfs_spin_lock(&qunit_hash_lock);
 877         for (i = 0; i < NR_DQHASH; i++) {
 878                 cfs_list_for_each_entry_safe(qunit, tmp, &qunit_hash[i],
 879                                              lq_hash) {
 880                         if (qunit->lq_ctxt != qctxt)
 881                                 continue;
 882
 883                         /* Wake up all waiters. Do not change lq_state.
 884                          * The waiters will check lq_rc which is kept as 0
 885                          * if no others change it, then the waiters will return
 886                          * -EAGAIN to caller who can perform related quota
 887                          * acq/rel if necessary. */
 888                         cfs_waitq_broadcast(&qunit->lq_waitq);
 889                 }
 890         }
 891         cfs_spin_unlock(&qunit_hash_lock);
 892         EXIT;
 893 }
 894
 895 static int got_qunit(struct lustre_qunit *qunit, int is_master)
 896 {
 897         struct lustre_quota_ctxt *qctxt = qunit->lq_ctxt;
 898         int rc = 0;
 899         ENTRY;
 900
 901         cfs_spin_lock(&qunit->lq_lock);
 902         switch (qunit->lq_state) {
 903         case QUNIT_IN_HASH:
 904         case QUNIT_RM_FROM_HASH:
 905                 break;
 906         case QUNIT_FINISHED:
 907                 rc = 1;
 908                 break;
 909         default:
 910                 CERROR("invalid qunit state %d\n", qunit->lq_state);
 911         }
 912         cfs_spin_unlock(&qunit->lq_lock);
 913
 914         if (!rc) {
 915                 cfs_spin_lock(&qctxt->lqc_lock);
 916                 rc = !qctxt->lqc_valid;
 917                 if (!is_master)
 918                         rc |= !qctxt->lqc_import;
 919                 cfs_spin_unlock(&qctxt->lqc_lock);
 920         }
 921
 922         RETURN(rc);
 923 }
 924
 925 static inline void
 926 revoke_lqs_rec(struct lustre_qunit_size *lqs, struct qunit_data *qdata, int opc)
 927 {
 928         /* revoke lqs_xxx_rec which is computed in check_cur_qunit
 929          * b=18630 */
 930         cfs_spin_lock(&lqs->lqs_lock);
 931         quota_compute_lqs(qdata, lqs, 0, (opc == QUOTA_DQACQ) ? 1 : 0);
 932         cfs_spin_unlock(&lqs->lqs_lock);
 933 }
 934
 935 static int
 936 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 937                struct qunit_data *qdata, int opc, int wait,
 938                struct obd_trans_info *oti)
 939 {
 940         struct lustre_qunit *qunit, *empty;
 941         struct l_wait_info lwi = { 0 };
 942         struct ptlrpc_request *req;
 943         struct dqacq_async_args *aa;
 944         struct obd_import *imp = NULL;
 945         struct lustre_qunit_size *lqs = NULL;
 946         struct timeval work_start;
 947         struct timeval work_end;
 948         long timediff;
 949         int rc = 0;
 950         ENTRY;
 951
 952         LASSERT(opc == QUOTA_DQACQ || opc == QUOTA_DQREL);
 953         cfs_gettimeofday(&work_start);
 954
 955         lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(qdata), qdata->qd_id),
 956                                qctxt, 0);
 957         if (lqs == NULL || IS_ERR(lqs)) {
 958                 CERROR("Can't find the lustre qunit size!\n");
 959                 RETURN(-EPERM);
 960         }
 961
 962         if ((empty = alloc_qunit(qctxt, qdata, opc)) == NULL) {
 963                 revoke_lqs_rec(lqs, qdata, opc);
 964                 /* this is for quota_search_lqs */
 965                 lqs_putref(lqs);
 966                 RETURN(-ENOMEM);
 967         }
 968
 969         OBD_FAIL_TIMEOUT(OBD_FAIL_QUOTA_DELAY_SD, 5);
 970
 971         cfs_spin_lock(&qunit_hash_lock);
 972         qunit = dqacq_in_flight(qctxt, qdata);
 973         if (qunit) {
 974                 cfs_spin_unlock(&qunit_hash_lock);
 975                 qunit_put(empty);
 976
 977                 revoke_lqs_rec(lqs, qdata, opc);
 978                 /* this is for quota_search_lqs */
 979                 lqs_putref(lqs);
 980                 goto wait_completion;
 981         }
 982         qunit = empty;
 983         qunit_get(qunit);
 984         insert_qunit_nolock(qctxt, qunit);
 985         cfs_spin_unlock(&qunit_hash_lock);
 986
 987         /* From here, the quota request will be sent anyway.
 988          * When this qdata request returned or is cancelled,
 989          * lqs_putref will be called at that time */
 990         lqs_getref(lqs);
 991         /* this is for quota_search_lqs */
 992         lqs_putref(lqs);
 993
 994         QDATA_DEBUG(qdata, "obd(%s): send %s quota req\n",
 995                     obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
 996         /* master is going to dqacq/dqrel from itself */
 997         if (is_master(qctxt)) {
 998                 int rc2;
 999                 QDATA_DEBUG(qdata, "local %s.\n",
1000                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
1001                 QDATA_SET_CHANGE_QS(qdata);
1002                 rc = qctxt->lqc_handler(obd, qdata, opc);
1003                 rc2 = dqacq_completion(obd, qctxt, qdata, rc, opc);
1004                 /* this is for qunit_get() */
1005                 qunit_put(qunit);
1006
1007                 cfs_gettimeofday(&work_end);
1008                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1009                 if (opc == QUOTA_DQACQ)
1010                         lprocfs_counter_add(qctxt->lqc_stats,
1011                                             wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
1012                                             timediff);
1013                 else
1014                         lprocfs_counter_add(qctxt->lqc_stats,
1015                                             wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
1016                                             timediff);
1017                 RETURN(rc ? rc : rc2);
1018         }
1019
1020         cfs_spin_lock(&qctxt->lqc_lock);
1021         if (!qctxt->lqc_import) {
1022                 cfs_spin_unlock(&qctxt->lqc_lock);
1023                 QDATA_DEBUG(qdata, "lqc_import is invalid.\n");
1024
1025                 cfs_spin_lock(&qunit_hash_lock);
1026                 remove_qunit_nolock(qunit);
1027                 cfs_spin_unlock(&qunit_hash_lock);
1028
1029                 compute_lqs_after_removing_qunit(qunit);
1030
1031                 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, -EAGAIN);
1032                 cfs_waitq_broadcast(&qunit->lq_waitq);
1033
1034                 /* this is for qunit_get() */
1035                 qunit_put(qunit);
1036                 /* this for alloc_qunit() */
1037                 qunit_put(qunit);
1038                 cfs_spin_lock(&qctxt->lqc_lock);
1039                 if (wait && !qctxt->lqc_import) {
1040                         cfs_spin_unlock(&qctxt->lqc_lock);
1041                         LASSERT(oti && oti->oti_thread);
1042                         /* The recovery thread doesn't have watchdog
1043                          * attached. LU-369 */
1044                         if (oti->oti_thread->t_watchdog)
1045                                 lc_watchdog_disable(oti->oti_thread->\
1046                                                 t_watchdog);
1047                         CDEBUG(D_QUOTA, "sleep for quota master\n");
1048                         l_wait_event(qctxt->lqc_wait_for_qmaster,
1049                                      check_qm(qctxt), &lwi);
1050                         CDEBUG(D_QUOTA, "wake up when quota master is back\n");
1051                         if (oti->oti_thread->t_watchdog)
1052                                 lc_watchdog_touch(oti->oti_thread->t_watchdog,
1053                                       CFS_GET_TIMEOUT(oti->oti_thread->t_svc));
1054                 } else {
1055                         cfs_spin_unlock(&qctxt->lqc_lock);
1056                 }
1057
1058                 RETURN(-EAGAIN);
1059         }
1060         imp = class_import_get(qctxt->lqc_import);
1061         cfs_spin_unlock(&qctxt->lqc_lock);
1062
1063         /* build dqacq/dqrel request */
1064         LASSERT(imp);
1065
1066         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_QUOTA_DQACQ,
1067                                         LUSTRE_MDS_VERSION, opc);
1068         class_import_put(imp);
1069         if (req == NULL) {
1070                 CERROR("Can't alloc request\n");
1071                 dqacq_completion(obd, qctxt, qdata, -ENOMEM, opc);
1072                 /* this is for qunit_get() */
1073                 qunit_put(qunit);
1074                 RETURN(-ENOMEM);
1075         }
1076
1077         ptlrpc_request_set_replen(req);
1078         req->rq_no_resend = req->rq_no_delay = 1;
1079         rc = quota_copy_qdata(req, qdata, QUOTA_REQUEST, QUOTA_IMPORT);
1080         if (rc < 0) {
1081                 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
1082                 ptlrpc_req_finished(req);
1083                 dqacq_completion(obd, qctxt, qdata, -EPROTO, opc);
1084                 /* this is for qunit_get() */
1085                 qunit_put(qunit);
1086                 RETURN(rc);
1087         }
1088
1089         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1090         aa = ptlrpc_req_async_args(req);
1091         aa->aa_ctxt = qctxt;
1092         aa->aa_qunit = qunit;
1093
1094         req->rq_interpret_reply = dqacq_interpret;
1095         ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1096
1097         QDATA_DEBUG(qdata, "%s scheduled.\n",
1098                     opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
1099 wait_completion:
1100         if (wait && qunit) {
1101                 struct qunit_data *p = &qunit->lq_data;
1102
1103                 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
1104                 l_wait_event(qunit->lq_waitq, got_qunit(qunit, is_master(qctxt)),
1105                              &lwi);
1106                 /* rc = -EAGAIN, it means the quota master isn't ready yet
1107                  * rc = QUOTA_REQ_RETURNED, it means a quota req is finished;
1108                  * rc = -EDQUOT, it means out of quota
1109                  * rc = -EBUSY, it means recovery is happening
1110                  * other rc < 0, it means real errors, functions who call
1111                  * schedule_dqacq should take care of this */
1112                 cfs_spin_lock(&qunit->lq_lock);
1113                 rc = qunit->lq_rc;
1114                 cfs_spin_unlock(&qunit->lq_lock);
1115                 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting: id(%u) flag(%u) "
1116                        "rc(%d) owner(%d)\n", qunit, qunit->lq_data.qd_id,
1117                        qunit->lq_data.qd_flags, rc, qunit->lq_owner);
1118         }
1119
1120         qunit_put(qunit);
1121         cfs_gettimeofday(&work_end);
1122         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1123         if (opc == QUOTA_DQACQ)
1124                 lprocfs_counter_add(qctxt->lqc_stats,
1125                                     wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
1126                                     timediff);
1127         else
1128                 lprocfs_counter_add(qctxt->lqc_stats,
1129                                     wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
1130                                     timediff);
1131
1132         RETURN(rc);
1133 }
1134
1135 int
1136 qctxt_adjust_qunit(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
1137                    const unsigned int id[], __u32 isblk, int wait,
1138                    struct obd_trans_info *oti)
1139 {
1140         int rc = 0, i = USRQUOTA;
1141         struct qunit_data qdata[MAXQUOTAS];
1142         ENTRY;
1143
1144         if (quota_is_set(obd, id, isblk ? QB_SET : QI_SET) == 0)
1145                 RETURN(0);
1146
1147         for (i = 0; i < MAXQUOTAS; i++) {
1148                 qdata[i].qd_id = id[i];
1149                 qdata[i].qd_flags = i;
1150                 if (isblk)
1151                         QDATA_SET_BLK(&qdata[i]);
1152                 qdata[i].qd_count = 0;
1153
1154                 rc = check_cur_qunit(obd, qctxt, &qdata[i]);
1155                 if (rc > 0) {
1156                         int opc;
1157                         /* need acquire or release */
1158                         opc = rc == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1159                         rc = schedule_dqacq(obd, qctxt, &qdata[i], opc,
1160                                             wait,oti);
1161                         if (rc < 0)
1162                                 RETURN(rc);
1163                 } else if (wait == 1) {
1164                         /* when wait equates 1, that means mds_quota_acquire
1165                          * or filter_quota_acquire is calling it. */
1166                         rc = qctxt_wait_pending_dqacq(qctxt, id[i], i, isblk);
1167                         if (rc < 0)
1168                                 RETURN(rc);
1169                 }
1170         }
1171
1172         RETURN(rc);
1173 }
1174
1175 int
1176 qctxt_wait_pending_dqacq(struct lustre_quota_ctxt *qctxt, unsigned int id,
1177                          unsigned short type, int isblk)
1178 {
1179         struct lustre_qunit *qunit = NULL;
1180         struct qunit_data qdata;
1181         struct timeval work_start;
1182         struct timeval work_end;
1183         long timediff;
1184         struct l_wait_info lwi = { 0 };
1185         int rc = 0;
1186         ENTRY;
1187
1188         cfs_gettimeofday(&work_start);
1189         qdata.qd_id = id;
1190         qdata.qd_flags = type;
1191         if (isblk)
1192                 QDATA_SET_BLK(&qdata);
1193         qdata.qd_count = 0;
1194
1195         cfs_spin_lock(&qunit_hash_lock);
1196         qunit = dqacq_in_flight(qctxt, &qdata);
1197         cfs_spin_unlock(&qunit_hash_lock);
1198
1199         if (qunit) {
1200                 struct qunit_data *p = &qunit->lq_data;
1201
1202                 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
1203                 l_wait_event(qunit->lq_waitq, got_qunit(qunit, is_master(qctxt)),
1204                              &lwi);
1205                 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting: rc(%d) "
1206                        "owner(%d)\n", qunit, qunit->lq_rc, qunit->lq_owner);
1207                 /* keep same as schedule_dqacq() b=17030 */
1208                 cfs_spin_lock(&qunit->lq_lock);
1209                 rc = qunit->lq_rc;
1210                 cfs_spin_unlock(&qunit->lq_lock);
1211                 /* this is for dqacq_in_flight() */
1212                 qunit_put(qunit);
1213                 cfs_gettimeofday(&work_end);
1214                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1215                 lprocfs_counter_add(qctxt->lqc_stats,
1216                                     isblk ? LQUOTA_WAIT_PENDING_BLK_QUOTA :
1217                                             LQUOTA_WAIT_PENDING_INO_QUOTA,
1218                                     timediff);
1219         } else {
1220                 cfs_gettimeofday(&work_end);
1221                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1222                 lprocfs_counter_add(qctxt->lqc_stats,
1223                                     isblk ? LQUOTA_NOWAIT_PENDING_BLK_QUOTA :
1224                                             LQUOTA_NOWAIT_PENDING_INO_QUOTA,
1225                                     timediff);
1226         }
1227
1228         RETURN(rc);
1229 }
1230
1231 int
1232 qctxt_init(struct obd_device *obd, dqacq_handler_t handler)
1233 {
1234         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
1235         struct obd_device_target *obt = &obd->u.obt;
1236         struct super_block *sb = obt->obt_sb;
1237         int rc = 0;
1238         ENTRY;
1239
1240         LASSERT(qctxt);
1241
1242         rc = ptlrpcd_addref();
1243         if (rc)
1244                 RETURN(rc);
1245
1246         cfs_waitq_init(&qctxt->lqc_wait_for_qmaster);
1247         cfs_waitq_init(&qctxt->lqc_lqs_waitq);
1248         cfs_atomic_set(&qctxt->lqc_lqs, 0);
1249         cfs_spin_lock_init(&qctxt->lqc_lock);
1250         cfs_spin_lock(&qctxt->lqc_lock);
1251         qctxt->lqc_handler = handler;
1252         qctxt->lqc_sb = sb;
1253         qctxt->lqc_obt = obt;
1254         qctxt->lqc_import = NULL;
1255         qctxt->lqc_recovery = 0;
1256         qctxt->lqc_switch_qs = 1; /* Change qunit size in default setting */
1257         qctxt->lqc_valid = 1;
1258         qctxt->lqc_cqs_boundary_factor = 4;
1259         qctxt->lqc_cqs_least_bunit = PTLRPC_MAX_BRW_SIZE;
1260         qctxt->lqc_cqs_least_iunit = 2;
1261         qctxt->lqc_cqs_qs_factor = 2;
1262         qctxt->lqc_flags = 0;
1263         QUOTA_MASTER_UNREADY(qctxt);
1264         qctxt->lqc_bunit_sz = default_bunit_sz;
1265         qctxt->lqc_btune_sz = default_bunit_sz / 100 * default_btune_ratio;
1266         qctxt->lqc_iunit_sz = default_iunit_sz;
1267         qctxt->lqc_itune_sz = default_iunit_sz * default_itune_ratio / 100;
1268         qctxt->lqc_switch_seconds = 300; /* enlarging will wait 5 minutes
1269                                           * after the last shrinking */
1270         qctxt->lqc_sync_blk = 0;
1271         cfs_spin_unlock(&qctxt->lqc_lock);
1272
1273         qctxt->lqc_lqs_hash = cfs_hash_create("LQS_HASH",
1274                                               hash_lqs_cur_bits,
1275                                               HASH_LQS_MAX_BITS,
1276                                               min(hash_lqs_cur_bits,
1277                                                   HASH_LQS_BKT_BITS),
1278                                               0, CFS_HASH_MIN_THETA,
1279                                               CFS_HASH_MAX_THETA,
1280                                               &lqs_hash_ops, CFS_HASH_DEFAULT);
1281         if (!qctxt->lqc_lqs_hash) {
1282                 CERROR("initialize hash lqs for %s error!\n", obd->obd_name);
1283                 RETURN(-ENOMEM);
1284         }
1285
1286 #ifdef LPROCFS
1287         rc = lquota_proc_setup(obd, is_master(qctxt));
1288         if (rc)
1289                 CERROR("initialize proc for %s error!\n", obd->obd_name);
1290 #endif
1291
1292         RETURN(rc);
1293 }
1294
1295 static int check_lqs(struct lustre_quota_ctxt *qctxt)
1296 {
1297         int rc;
1298         ENTRY;
1299
1300         rc = !cfs_atomic_read(&qctxt->lqc_lqs);
1301
1302         RETURN(rc);
1303 }
1304
1305 int qctxt_del_lqs(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1306                  cfs_hlist_node_t *hnode, void *data)
1307 {
1308         /* remove from hash and -1 refcount */
1309         cfs_hash_bd_del_locked(hs, bd, hnode);
1310         return 0;
1311 }
1312
1313 void qctxt_cleanup(struct lustre_quota_ctxt *qctxt, int force)
1314 {
1315         struct lustre_qunit *qunit, *tmp;
1316         cfs_list_t tmp_list;
1317         struct l_wait_info lwi = { 0 };
1318         struct obd_device_target *obt = qctxt->lqc_obt;
1319         int i;
1320         ENTRY;
1321
1322         CFS_INIT_LIST_HEAD(&tmp_list);
1323
1324         cfs_spin_lock(&qctxt->lqc_lock);
1325         qctxt->lqc_valid = 0;
1326         cfs_spin_unlock(&qctxt->lqc_lock);
1327
1328         cfs_spin_lock(&qunit_hash_lock);
1329         for (i = 0; i < NR_DQHASH; i++) {
1330                 cfs_list_for_each_entry_safe(qunit, tmp, &qunit_hash[i],
1331                                              lq_hash) {
1332                         if (qunit->lq_ctxt != qctxt)
1333                                 continue;
1334                         remove_qunit_nolock(qunit);
1335                         cfs_list_add(&qunit->lq_hash, &tmp_list);
1336                 }
1337         }
1338         cfs_spin_unlock(&qunit_hash_lock);
1339
1340         cfs_list_for_each_entry_safe(qunit, tmp, &tmp_list, lq_hash) {
1341                 cfs_list_del_init(&qunit->lq_hash);
1342                 compute_lqs_after_removing_qunit(qunit);
1343
1344                 /* wake up all waiters */
1345                 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, 0);
1346                 cfs_waitq_broadcast(&qunit->lq_waitq);
1347                 qunit_put(qunit);
1348         }
1349
1350         /* after qctxt_cleanup, qctxt might be freed, then check_qm() is
1351          * unpredicted. So we must wait until lqc_wait_for_qmaster is empty */
1352         while (cfs_waitq_active(&qctxt->lqc_wait_for_qmaster)) {
1353                 cfs_waitq_signal(&qctxt->lqc_wait_for_qmaster);
1354                 cfs_schedule_timeout_and_set_state(CFS_TASK_INTERRUPTIBLE,
1355                                                    cfs_time_seconds(1));
1356         }
1357
1358         /* release refcount on lustre_qunit_size holding by lqs_hash */
1359         cfs_hash_for_each_safe(qctxt->lqc_lqs_hash, qctxt_del_lqs, NULL);
1360
1361         l_wait_event(qctxt->lqc_lqs_waitq, check_lqs(qctxt), &lwi);
1362         cfs_down_write(&obt->obt_rwsem);
1363         cfs_hash_putref(qctxt->lqc_lqs_hash);
1364         qctxt->lqc_lqs_hash = NULL;
1365         cfs_up_write(&obt->obt_rwsem);
1366
1367         ptlrpcd_decref();
1368
1369 #ifdef LPROCFS
1370         if (lquota_proc_cleanup(qctxt))
1371                 CERROR("cleanup proc error!\n");
1372 #endif
1373
1374         EXIT;
1375 }
1376
1377 struct qslave_recov_thread_data {
1378         struct obd_device *obd;
1379         struct lustre_quota_ctxt *qctxt;
1380         cfs_completion_t comp;
1381 };
1382
1383 /* FIXME only recovery block quota by now */
1384 static int qslave_recovery_main(void *arg)
1385 {
1386         struct qslave_recov_thread_data *data = arg;
1387         struct obd_device *obd = data->obd;
1388         struct lustre_quota_ctxt *qctxt = data->qctxt;
1389         unsigned int type;
1390         int rc = 0;
1391         ENTRY;
1392
1393         cfs_daemonize_ctxt("qslave_recovd");
1394
1395         /* for obdfilter */
1396         class_incref(obd, "qslave_recovd_filter", obd);
1397
1398         cfs_complete(&data->comp);
1399
1400         cfs_spin_lock(&qctxt->lqc_lock);
1401         if (qctxt->lqc_recovery) {
1402                 cfs_spin_unlock(&qctxt->lqc_lock);
1403                 class_decref(obd, "qslave_recovd_filter", obd);
1404                 RETURN(0);
1405         } else {
1406                 qctxt->lqc_recovery = 1;
1407                 cfs_spin_unlock(&qctxt->lqc_lock);
1408         }
1409
1410         for (type = USRQUOTA; type < MAXQUOTAS; type++) {
1411                 struct qunit_data qdata;
1412                 struct quota_info *dqopt = sb_dqopt(qctxt->lqc_sb);
1413                 cfs_list_t id_list;
1414                 struct dquot_id *dqid, *tmp;
1415                 int ret;
1416
1417                 LOCK_DQONOFF_MUTEX(dqopt);
1418                 if (!ll_sb_has_quota_active(qctxt->lqc_sb, type)) {
1419                         UNLOCK_DQONOFF_MUTEX(dqopt);
1420                         break;
1421                 }
1422
1423                 LASSERT(dqopt->files[type] != NULL);
1424                 CFS_INIT_LIST_HEAD(&id_list);
1425 #ifndef KERNEL_SUPPORTS_QUOTA_READ
1426                 rc = fsfilt_qids(obd, dqopt->files[type], NULL, type, &id_list);
1427 #else
1428                 rc = fsfilt_qids(obd, NULL, dqopt->files[type], type, &id_list);
1429 #endif
1430                 UNLOCK_DQONOFF_MUTEX(dqopt);
1431                 if (rc)
1432                         CERROR("Get ids from quota file failed. (rc:%d)\n", rc);
1433
1434                 cfs_list_for_each_entry_safe(dqid, tmp, &id_list, di_link) {
1435                         cfs_list_del_init(&dqid->di_link);
1436                         /* skip slave recovery on itself */
1437                         if (is_master(qctxt))
1438                                 goto free;
1439                         if (rc && rc != -EBUSY)
1440                                 goto free;
1441
1442                         qdata.qd_id = dqid->di_id;
1443                         qdata.qd_flags = type;
1444                         QDATA_SET_BLK(&qdata);
1445                         qdata.qd_count = 0;
1446
1447                         ret = check_cur_qunit(obd, qctxt, &qdata);
1448                         if (ret > 0) {
1449                                 int opc;
1450                                 opc = ret == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1451                                 rc = schedule_dqacq(obd, qctxt, &qdata, opc,
1452                                                     0, NULL);
1453                                 if (rc == -EDQUOT)
1454                                         rc = 0;
1455                         } else {
1456                                 rc = 0;
1457                         }
1458
1459                         if (rc && rc != -EBUSY)
1460                                 CERROR("qslave recovery failed! (id:%d type:%d "
1461                                        " rc:%d)\n", dqid->di_id, type, rc);
1462 free:
1463                         OBD_FREE_PTR(dqid);
1464                 }
1465         }
1466
1467         cfs_spin_lock(&qctxt->lqc_lock);
1468         qctxt->lqc_recovery = 0;
1469         cfs_spin_unlock(&qctxt->lqc_lock);
1470         class_decref(obd, "qslave_recovd_filter", obd);
1471         RETURN(rc);
1472 }
1473
1474 void
1475 qslave_start_recovery(struct obd_device *obd, struct lustre_quota_ctxt *qctxt)
1476 {
1477         struct qslave_recov_thread_data data;
1478         int rc;
1479         ENTRY;
1480
1481         if (!ll_sb_any_quota_active(qctxt->lqc_sb))
1482                 goto exit;
1483
1484         data.obd = obd;
1485         data.qctxt = qctxt;
1486         cfs_init_completion(&data.comp);
1487
1488         rc = cfs_create_thread(qslave_recovery_main, &data,
1489                                CFS_DAEMON_FLAGS);
1490         if (rc < 0) {
1491                 CERROR("Cannot start quota recovery thread: rc %d\n", rc);
1492                 goto exit;
1493         }
1494         cfs_wait_for_completion(&data.comp);
1495 exit:
1496         EXIT;
1497 }
1498
1499 inline int quota_is_on(struct lustre_quota_ctxt *qctxt,
1500                        struct obd_quotactl *oqctl)
1501 {
1502         return ((qctxt->lqc_flags & UGQUOTA2LQC(oqctl->qc_type)) ==
1503                 UGQUOTA2LQC(oqctl->qc_type));
1504 }
1505
1506 inline int quota_is_off(struct lustre_quota_ctxt *qctxt,
1507                         struct obd_quotactl *oqctl)
1508 {
1509         return !(qctxt->lqc_flags & UGQUOTA2LQC(oqctl->qc_type));
1510 }
1511
1512 /**
1513  * When quotaon, build a lqs for every uid/gid who has been set limitation
1514  * for quota. After quota_search_lqs, it will hold one ref for the lqs.
1515  * It will be released when qctxt_cleanup() is executed b=18574
1516  *
1517  * Should be called with obt->obt_quotachecking held. b=20152
1518  */
1519 void build_lqs(struct obd_device *obd)
1520 {
1521         struct obd_device_target *obt = &obd->u.obt;
1522         struct lustre_quota_ctxt *qctxt = &obt->obt_qctxt;
1523         cfs_list_t id_list;
1524         int i, rc;
1525
1526         LASSERT_SEM_LOCKED(&obt->obt_quotachecking);
1527         CFS_INIT_LIST_HEAD(&id_list);
1528         for (i = 0; i < MAXQUOTAS; i++) {
1529                 struct dquot_id *dqid, *tmp;
1530
1531                 if (sb_dqopt(qctxt->lqc_sb)->files[i] == NULL)
1532                         continue;
1533
1534 #ifndef KERNEL_SUPPORTS_QUOTA_READ
1535                 rc = fsfilt_qids(obd, sb_dqopt(qctxt->lqc_sb)->files[i], NULL,
1536                                  i, &id_list);
1537 #else
1538                 rc = fsfilt_qids(obd, NULL, sb_dqopt(qctxt->lqc_sb)->files[i],
1539                                  i, &id_list);
1540 #endif
1541                 if (rc) {
1542                         CERROR("%s: failed to get %s qids!\n", obd->obd_name,
1543                                i ? "group" : "user");
1544                         continue;
1545                 }
1546
1547                 cfs_list_for_each_entry_safe(dqid, tmp, &id_list,
1548                                              di_link) {
1549                         struct lustre_qunit_size *lqs;
1550
1551                         cfs_list_del_init(&dqid->di_link);
1552                         lqs = quota_search_lqs(LQS_KEY(i, dqid->di_id),
1553                                                qctxt, 1);
1554                         if (lqs && !IS_ERR(lqs)) {
1555                                 lqs->lqs_flags |= dqid->di_flag;
1556                                 lqs_putref(lqs);
1557                         } else {
1558                                 CERROR("%s: failed to create a lqs for %sid %u"
1559                                        "\n", obd->obd_name, i ? "g" : "u",
1560                                        dqid->di_id);
1561                         }
1562
1563                         OBD_FREE_PTR(dqid);
1564                 }
1565         }
1566 }
1567
1568 /**
1569  * lqs<->qctxt hash operations
1570  */
1571
1572 /**
1573  * string hashing using djb2 hash algorithm
1574  */
1575 static unsigned
1576 lqs_hash(cfs_hash_t *hs, const void *key, unsigned mask)
1577 {
1578         unsigned long long id;
1579         unsigned hash;
1580         ENTRY;
1581
1582         LASSERT(key);
1583         id = *((unsigned long long *)key);
1584         hash = (LQS_KEY_GRP(id) ? 5381 : 5387) * (unsigned)LQS_KEY_ID(id);
1585
1586         RETURN(hash & mask);
1587 }
1588
1589 static void *
1590 lqs_key(cfs_hlist_node_t *hnode)
1591 {
1592         struct lustre_qunit_size *lqs;
1593         ENTRY;
1594
1595         lqs = cfs_hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1596         RETURN(&lqs->lqs_key);
1597 }
1598
1599 static int
1600 lqs_keycmp(const void *key, cfs_hlist_node_t *hnode)
1601 {
1602         struct lustre_qunit_size *q =
1603                 cfs_hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1604
1605         RETURN(q->lqs_key == *((unsigned long long *)key));
1606 }
1607
1608 static void *
1609 lqs_object(cfs_hlist_node_t *hnode)
1610 {
1611         return cfs_hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1612 }
1613
1614 static void
1615 lqs_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
1616 {
1617         struct lustre_qunit_size *q =
1618                 cfs_hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1619
1620         lqs_getref(q);
1621 }
1622
1623 static void
1624 lqs_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
1625 {
1626         struct lustre_qunit_size *q =
1627                 cfs_hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1628
1629         lqs_putref(q);
1630 }
1631
1632 static void
1633 lqs_exit(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
1634 {
1635         CERROR("It should not have any item left to be handled by this!");
1636 }
1637
1638 static cfs_hash_ops_t lqs_hash_ops = {
1639         .hs_hash        = lqs_hash,
1640         .hs_key         = lqs_key,
1641         .hs_keycmp      = lqs_keycmp,
1642         .hs_object      = lqs_object,
1643         .hs_get         = lqs_get,
1644         .hs_put_locked  = lqs_put_locked,
1645         .hs_exit        = lqs_exit
1646 };