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 (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * Copyright (c) 2011 Whamcloud, Inc.
  34  */
  35 /*
  36  * This file is part of Lustre, http://www.lustre.org/
  37  * Lustre is a trademark of Sun Microsystems, Inc.
  38  *
  39  * lustre/quota/quota_context.c
  40  *
  41  * Lustre Quota Context
  42  *
  43  * Author: Niu YaWei <niu@clusterfs.com>
  44  */
  45
  46 #ifndef EXPORT_SYMTAB
  47 # define EXPORT_SYMTAB
  48 #endif
  49
  50 #define DEBUG_SUBSYSTEM S_LQUOTA
  51
  52 #include <linux/version.h>
  53 #include <linux/fs.h>
  54 #include <asm/unistd.h>
  55 #include <linux/slab.h>
  56 #include <linux/quotaops.h>
  57 #include <linux/module.h>
  58 #include <linux/init.h>
  59
  60 #include <obd_class.h>
  61 #include <lustre_quota.h>
  62 #include <lustre_fsfilt.h>
  63 #include <lprocfs_status.h>
  64 #include "quota_internal.h"
  65
  66 static int hash_lqs_cur_bits = HASH_LQS_CUR_BITS;
  67 CFS_MODULE_PARM(hash_lqs_cur_bits, "i", int, 0444,
  68                 "the current bits of lqs hash");
  69
  70 #ifdef HAVE_QUOTA_SUPPORT
  71
  72 static cfs_hash_ops_t lqs_hash_ops;
  73
  74 unsigned long default_bunit_sz = 128 * 1024 * 1024; /* 128M bytes */
  75 unsigned long default_btune_ratio = 50;             /* 50 percentage */
  76 unsigned long default_iunit_sz = 5120;              /* 5120 inodes */
  77 unsigned long default_itune_ratio = 50;             /* 50 percentage */
  78
  79 cfs_mem_cache_t *qunit_cachep = NULL;
  80 cfs_list_t qunit_hash[NR_DQHASH];
  81 cfs_spinlock_t qunit_hash_lock = CFS_SPIN_LOCK_UNLOCKED;
  82
  83 /* please sync qunit_state with qunit_state_names */
  84 enum qunit_state {
  85         /**
  86          * a qunit is created
  87          */
  88         QUNIT_CREATED      = 0,
  89         /**
  90          * a qunit is added into qunit hash, that means
  91          * a quota req will be sent or is flying
  92          */
  93         QUNIT_IN_HASH      = 1,
  94         /**
  95          * a qunit is removed from qunit hash, that
  96          * means a quota req is handled and comes back
  97          */
  98         QUNIT_RM_FROM_HASH = 2,
  99         /**
 100          * qunit can wake up all threads waiting for it
 101          */
 102         QUNIT_FINISHED     = 3,
 103 };
 104
 105 static const char *qunit_state_names[] = {
 106         [QUNIT_CREATED]      = "CREATED",
 107         [QUNIT_IN_HASH]      = "IN_HASH",
 108         [QUNIT_RM_FROM_HASH] = "RM_FROM_HASH",
 109         [QUNIT_FINISHED]     = "FINISHED",
 110 };
 111
 112 struct lustre_qunit {
 113         cfs_list_t lq_hash;      /** Hash list in memory */
 114         cfs_atomic_t lq_refcnt;            /** Use count */
 115         struct lustre_quota_ctxt *lq_ctxt; /** Quota context this applies to */
 116         struct qunit_data lq_data;         /** See qunit_data */
 117         unsigned int lq_opc;               /** QUOTA_DQACQ, QUOTA_DQREL */
 118         cfs_waitq_t lq_waitq;              /** Threads waiting for this qunit */
 119         cfs_spinlock_t lq_lock;            /** Protect the whole structure */
 120         enum qunit_state lq_state;         /** Present the status of qunit */
 121         int lq_rc;                         /** The rc of lq_data */
 122         pid_t lq_owner;
 123 };
 124
 125 #define QUNIT_SET_STATE(qunit, state)                                   \
 126 do {                                                                    \
 127         cfs_spin_lock(&qunit->lq_lock);                                 \
 128         QDATA_DEBUG((&qunit->lq_data), "qunit(%p) lq_state(%s->%s), "   \
 129                     "lq_rc(%d), lq_owner(%d)\n",                        \
 130                     qunit, qunit_state_names[qunit->lq_state],          \
 131                     qunit_state_names[state], qunit->lq_rc,             \
 132                     qunit->lq_owner);                                   \
 133         qunit->lq_state = state;                                        \
 134         cfs_spin_unlock(&qunit->lq_lock);                               \
 135 } while(0)
 136
 137 #define QUNIT_SET_STATE_AND_RC(qunit, state, rc)                        \
 138 do {                                                                    \
 139         cfs_spin_lock(&qunit->lq_lock);                                 \
 140         qunit->lq_rc = rc;                                              \
 141         QDATA_DEBUG((&qunit->lq_data), "qunit(%p) lq_state(%s->%s), "   \
 142                     "lq_rc(%d), lq_owner(%d)\n",                        \
 143                     qunit, qunit_state_names[qunit->lq_state],          \
 144                     qunit_state_names[state], qunit->lq_rc,             \
 145                     qunit->lq_owner);                                   \
 146         qunit->lq_state = state;                                        \
 147         cfs_spin_unlock(&qunit->lq_lock);                               \
 148 } while(0)
 149
 150 int should_translate_quota (struct obd_import *imp)
 151 {
 152         ENTRY;
 153
 154         LASSERT(imp);
 155         if (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_QUOTA64)
 156                 RETURN(0);
 157         else
 158                 RETURN(1);
 159 }
 160
 161 void qunit_cache_cleanup(void)
 162 {
 163         int i;
 164         ENTRY;
 165
 166         cfs_spin_lock(&qunit_hash_lock);
 167         for (i = 0; i < NR_DQHASH; i++)
 168                 LASSERT(cfs_list_empty(qunit_hash + i));
 169         cfs_spin_unlock(&qunit_hash_lock);
 170
 171         if (qunit_cachep) {
 172                 int rc;
 173                 rc = cfs_mem_cache_destroy(qunit_cachep);
 174                 LASSERTF(rc == 0, "couldn't destroy qunit_cache slab\n");
 175                 qunit_cachep = NULL;
 176         }
 177         EXIT;
 178 }
 179
 180 int qunit_cache_init(void)
 181 {
 182         int i;
 183         ENTRY;
 184
 185         LASSERT(qunit_cachep == NULL);
 186         qunit_cachep = cfs_mem_cache_create("ll_qunit_cache",
 187                                             sizeof(struct lustre_qunit),
 188                                             0, 0);
 189         if (!qunit_cachep)
 190                 RETURN(-ENOMEM);
 191
 192         cfs_spin_lock(&qunit_hash_lock);
 193         for (i = 0; i < NR_DQHASH; i++)
 194                 CFS_INIT_LIST_HEAD(qunit_hash + i);
 195         cfs_spin_unlock(&qunit_hash_lock);
 196         RETURN(0);
 197 }
 198
 199 static inline int
 200 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
 201              __attribute__((__const__));
 202
 203 static inline int
 204 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
 205 {
 206         unsigned int id = qdata->qd_id;
 207         unsigned int type = QDATA_IS_GRP(qdata);
 208
 209         unsigned long tmp = ((unsigned long)qctxt >> L1_CACHE_SHIFT) ^ id;
 210         tmp = (tmp * (MAXQUOTAS - type)) % NR_DQHASH;
 211         return tmp;
 212 }
 213
 214 /* caller must hold qunit_hash_lock */
 215 static inline struct lustre_qunit *find_qunit(unsigned int hashent,
 216                                               struct lustre_quota_ctxt *qctxt,
 217                                               struct qunit_data *qdata)
 218 {
 219         struct lustre_qunit *qunit = NULL;
 220         struct qunit_data *tmp;
 221
 222         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
 223         cfs_list_for_each_entry(qunit, qunit_hash + hashent, lq_hash) {
 224                 tmp = &qunit->lq_data;
 225                 if (qunit->lq_ctxt == qctxt &&
 226                     qdata->qd_id == tmp->qd_id &&
 227                     (qdata->qd_flags & LQUOTA_QUNIT_FLAGS) ==
 228                     (tmp->qd_flags & LQUOTA_QUNIT_FLAGS))
 229                         return qunit;
 230         }
 231         return NULL;
 232 }
 233
 234 /* check_cur_qunit - check the current usage of qunit.
 235  * @qctxt: quota context
 236  * @qdata: the type of quota unit to be checked
 237  *
 238  * return: 1 - need acquire qunit;
 239  *         2 - need release qunit;
 240  *         0 - need do nothing.
 241  *       < 0 - error.
 242  */
 243 static int
 244 check_cur_qunit(struct obd_device *obd,
 245                 struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
 246 {
 247         struct super_block *sb = qctxt->lqc_sb;
 248         unsigned long qunit_sz, tune_sz;
 249         __u64 usage, limit, limit_org, pending_write = 0;
 250         long long record = 0;
 251         struct obd_quotactl *qctl;
 252         struct lustre_qunit_size *lqs = NULL;
 253         int ret = 0;
 254         ENTRY;
 255
 256         if (!ll_sb_any_quota_active(sb))
 257                 RETURN(0);
 258
 259         cfs_spin_lock(&qctxt->lqc_lock);
 260         if (!qctxt->lqc_valid){
 261                 cfs_spin_unlock(&qctxt->lqc_lock);
 262                 RETURN(0);
 263         }
 264         cfs_spin_unlock(&qctxt->lqc_lock);
 265
 266         OBD_ALLOC_PTR(qctl);
 267         if (qctl == NULL)
 268                 RETURN(-ENOMEM);
 269
 270         /* get fs quota usage & limit */
 271         qctl->qc_cmd = Q_GETQUOTA;
 272         qctl->qc_id = qdata->qd_id;
 273         qctl->qc_type = QDATA_IS_GRP(qdata);
 274         ret = fsfilt_quotactl(obd, sb, qctl);
 275         if (ret) {
 276                 if (ret == -ESRCH)      /* no limit */
 277                         ret = 0;
 278                 else
 279                         CERROR("can't get fs quota usage! (rc:%d)\n", ret);
 280                 GOTO(out, ret);
 281         }
 282
 283         if (QDATA_IS_BLK(qdata)) {
 284                 usage = qctl->qc_dqblk.dqb_curspace;
 285                 limit = qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS;
 286         } else {
 287                 usage = qctl->qc_dqblk.dqb_curinodes;
 288                 limit = qctl->qc_dqblk.dqb_ihardlimit;
 289         }
 290
 291         /* ignore the no quota limit case; and it can avoid creating
 292          * unnecessary lqs for uid/gid */
 293         if (!limit)
 294                 GOTO(out, ret = 0);
 295
 296         lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(qdata), qdata->qd_id),
 297                                qctxt, 0);
 298         if (IS_ERR(lqs) || lqs == NULL) {
 299                 CERROR("fail to find a lqs for %sid: %u)!\n",
 300                        QDATA_IS_GRP(qdata) ? "g" : "u", qdata->qd_id);
 301                 GOTO (out, ret = 0);
 302         }
 303         cfs_spin_lock(&lqs->lqs_lock);
 304
 305         if (QDATA_IS_BLK(qdata)) {
 306                 qunit_sz = lqs->lqs_bunit_sz;
 307                 tune_sz  = lqs->lqs_btune_sz;
 308                 pending_write = lqs->lqs_bwrite_pending;
 309                 record   = lqs->lqs_blk_rec;
 310                 LASSERT(!(qunit_sz % QUOTABLOCK_SIZE));
 311         } else {
 312                 /* we didn't need change inode qunit size now */
 313                 qunit_sz = lqs->lqs_iunit_sz;
 314                 tune_sz  = lqs->lqs_itune_sz;
 315                 pending_write = lqs->lqs_iwrite_pending;
 316                 record   = lqs->lqs_ino_rec;
 317         }
 318
 319         /* we don't count the MIN_QLIMIT */
 320         if ((limit == MIN_QLIMIT && !QDATA_IS_BLK(qdata)) ||
 321             (toqb(limit) == MIN_QLIMIT && QDATA_IS_BLK(qdata)))
 322                 limit = 0;
 323
 324         usage += pending_write;
 325         limit_org = limit;
 326         /* when a releasing quota req is sent, before it returned
 327            limit is assigned a small value. limit will overflow */
 328         if (limit + record < 0)
 329                 usage -= record;
 330         else
 331                 limit += record;
 332
 333         LASSERT(qdata->qd_count == 0);
 334         if (limit <= usage + tune_sz) {
 335                 while (qdata->qd_count + limit <=
 336                        usage + tune_sz)
 337                         qdata->qd_count += qunit_sz;
 338                 ret = 1;
 339         } else if (limit > usage + qunit_sz + tune_sz &&
 340                    limit_org > qdata->qd_count + qunit_sz) {
 341                 while (limit - qdata->qd_count > usage + qunit_sz + tune_sz &&
 342                        limit_org > qdata->qd_count + qunit_sz)
 343                         qdata->qd_count += qunit_sz;
 344                 ret = 2;
 345                 /* if there are other pending writes for this uid/gid, releasing
 346                  * quota is put off until the last pending write b=16645 */
 347                 /* if there is an ongoing quota request, a releasing request is aborted.
 348                  * That ongoing quota request will call this function again when
 349                  * it returned b=18630 */
 350                 if (pending_write || record) {
 351                         CDEBUG(D_QUOTA, "delay quota release\n");
 352                         ret = 0;
 353                 }
 354         }
 355         if (ret > 0)
 356                 quota_compute_lqs(qdata, lqs, 1, (ret == 1) ? 1 : 0);
 357
 358         CDEBUG(D_QUOTA, "type: %c, limit: "LPU64", usage: "LPU64
 359                ", pending_write: "LPU64", record: %lld"
 360                ", qunit_sz: %lu, tune_sz: %lu, ret: %d.\n",
 361                QDATA_IS_BLK(qdata) ? 'b' : 'i', limit, usage, pending_write,
 362                record, qunit_sz, tune_sz, ret);
 363         LASSERT(ret == 0 || qdata->qd_count);
 364
 365         cfs_spin_unlock(&lqs->lqs_lock);
 366         lqs_putref(lqs);
 367
 368         EXIT;
 369  out:
 370         OBD_FREE_PTR(qctl);
 371         return ret;
 372 }
 373
 374 /**
 375  * Compute the remaining quota for certain gid or uid b=11693
 376  */
 377 int compute_remquota(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 378                      struct qunit_data *qdata, int isblk)
 379 {
 380         struct super_block *sb = qctxt->lqc_sb;
 381         __u64 usage, limit;
 382         struct obd_quotactl *qctl;
 383         int ret = QUOTA_RET_OK;
 384         ENTRY;
 385
 386         if (!ll_sb_any_quota_active(sb))
 387                 RETURN(QUOTA_RET_NOQUOTA);
 388
 389         /* ignore root user */
 390         if (qdata->qd_id == 0 && QDATA_IS_GRP(qdata) == USRQUOTA)
 391                 RETURN(QUOTA_RET_NOLIMIT);
 392
 393         OBD_ALLOC_PTR(qctl);
 394         if (qctl == NULL)
 395                 RETURN(-ENOMEM);
 396
 397         /* get fs quota usage & limit */
 398         qctl->qc_cmd = Q_GETQUOTA;
 399         qctl->qc_id = qdata->qd_id;
 400         qctl->qc_type = QDATA_IS_GRP(qdata);
 401         ret = fsfilt_quotactl(obd, sb, qctl);
 402         if (ret) {
 403                 if (ret == -ESRCH)      /* no limit */
 404                         ret = QUOTA_RET_NOLIMIT;
 405                 else
 406                         CDEBUG(D_QUOTA, "can't get fs quota usage! (rc:%d)",
 407                                ret);
 408                 GOTO(out, ret);
 409         }
 410
 411         usage = isblk ? qctl->qc_dqblk.dqb_curspace :
 412                 qctl->qc_dqblk.dqb_curinodes;
 413         limit = isblk ? qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS :
 414                 qctl->qc_dqblk.dqb_ihardlimit;
 415         if (!limit){            /* no limit */
 416                 ret = QUOTA_RET_NOLIMIT;
 417                 GOTO(out, ret);
 418         }
 419
 420         if (limit >= usage)
 421                 qdata->qd_count = limit - usage;
 422         else
 423                 qdata->qd_count = 0;
 424         EXIT;
 425 out:
 426         OBD_FREE_PTR(qctl);
 427         return ret;
 428 }
 429
 430 static struct lustre_qunit *alloc_qunit(struct lustre_quota_ctxt *qctxt,
 431                                         struct qunit_data *qdata, int opc)
 432 {
 433         struct lustre_qunit *qunit = NULL;
 434         ENTRY;
 435
 436         OBD_SLAB_ALLOC_PTR_GFP(qunit, qunit_cachep, CFS_ALLOC_IO);
 437         if (qunit == NULL)
 438                 RETURN(NULL);
 439
 440         CFS_INIT_LIST_HEAD(&qunit->lq_hash);
 441         cfs_waitq_init(&qunit->lq_waitq);
 442         cfs_atomic_set(&qunit->lq_refcnt, 1);
 443         qunit->lq_ctxt = qctxt;
 444         memcpy(&qunit->lq_data, qdata, sizeof(*qdata));
 445         qunit->lq_opc = opc;
 446         qunit->lq_lock = CFS_SPIN_LOCK_UNLOCKED;
 447         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_CREATED, 0);
 448         qunit->lq_owner = cfs_curproc_pid();
 449         RETURN(qunit);
 450 }
 451
 452 static inline void free_qunit(struct lustre_qunit *qunit)
 453 {
 454         OBD_SLAB_FREE(qunit, qunit_cachep, sizeof(*qunit));
 455 }
 456
 457 static inline void qunit_get(struct lustre_qunit *qunit)
 458 {
 459         cfs_atomic_inc(&qunit->lq_refcnt);
 460 }
 461
 462 static void qunit_put(struct lustre_qunit *qunit)
 463 {
 464         LASSERT(cfs_atomic_read(&qunit->lq_refcnt));
 465         if (cfs_atomic_dec_and_test(&qunit->lq_refcnt))
 466                 free_qunit(qunit);
 467 }
 468
 469 /* caller must hold qunit_hash_lock and release ref of qunit after using it */
 470 static struct lustre_qunit *dqacq_in_flight(struct lustre_quota_ctxt *qctxt,
 471                                             struct qunit_data *qdata)
 472 {
 473         unsigned int hashent = qunit_hashfn(qctxt, qdata);
 474         struct lustre_qunit *qunit;
 475         ENTRY;
 476
 477         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
 478         qunit = find_qunit(hashent, qctxt, qdata);
 479         if (qunit)
 480                 qunit_get(qunit);
 481         RETURN(qunit);
 482 }
 483
 484 static void
 485 insert_qunit_nolock(struct lustre_quota_ctxt *qctxt, struct lustre_qunit *qunit)
 486 {
 487         cfs_list_t *head;
 488
 489         LASSERT(cfs_list_empty(&qunit->lq_hash));
 490         qunit_get(qunit);
 491         head = qunit_hash + qunit_hashfn(qctxt, &qunit->lq_data);
 492         cfs_list_add(&qunit->lq_hash, head);
 493         QUNIT_SET_STATE(qunit, QUNIT_IN_HASH);
 494 }
 495
 496 static void compute_lqs_after_removing_qunit(struct lustre_qunit *qunit)
 497 {
 498         struct lustre_qunit_size *lqs;
 499
 500         lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(&qunit->lq_data),
 501                                        qunit->lq_data.qd_id),
 502                                qunit->lq_ctxt, 0);
 503         if (lqs && !IS_ERR(lqs)) {
 504                 cfs_spin_lock(&lqs->lqs_lock);
 505                 if (qunit->lq_opc == QUOTA_DQACQ)
 506                         quota_compute_lqs(&qunit->lq_data, lqs, 0, 1);
 507                 if (qunit->lq_opc == QUOTA_DQREL)
 508                         quota_compute_lqs(&qunit->lq_data, lqs, 0, 0);
 509                 cfs_spin_unlock(&lqs->lqs_lock);
 510                 /* this is for quota_search_lqs */
 511                 lqs_putref(lqs);
 512                 /* this is for schedule_dqacq */
 513                 lqs_putref(lqs);
 514         }
 515 }
 516
 517 static void remove_qunit_nolock(struct lustre_qunit *qunit)
 518 {
 519         LASSERT(!cfs_list_empty(&qunit->lq_hash));
 520         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
 521
 522         cfs_list_del_init(&qunit->lq_hash);
 523         QUNIT_SET_STATE(qunit, QUNIT_RM_FROM_HASH);
 524         qunit_put(qunit);
 525 }
 526
 527 void* quota_barrier(struct lustre_quota_ctxt *qctxt,
 528                     struct obd_quotactl *oqctl, int isblk)
 529 {
 530         struct lustre_qunit *qunit, *find_qunit;
 531         int cycle = 1;
 532
 533         OBD_SLAB_ALLOC(qunit, qunit_cachep, CFS_ALLOC_IO, sizeof(*qunit));
 534         if (qunit == NULL) {
 535                 CERROR("locating %sunit failed for %sid %u\n",
 536                        isblk ? "b" : "i", oqctl->qc_type ? "g" : "u",
 537                        oqctl->qc_id);
 538                 qctxt_wait_pending_dqacq(qctxt, oqctl->qc_id,
 539                                          oqctl->qc_type, isblk);
 540                 return NULL;
 541         }
 542
 543         CFS_INIT_LIST_HEAD(&qunit->lq_hash);
 544         qunit->lq_lock = CFS_SPIN_LOCK_UNLOCKED;
 545         cfs_waitq_init(&qunit->lq_waitq);
 546         cfs_atomic_set(&qunit->lq_refcnt, 1);
 547         qunit->lq_ctxt = qctxt;
 548         qunit->lq_data.qd_id = oqctl->qc_id;
 549         qunit->lq_data.qd_flags =  oqctl->qc_type;
 550         if (isblk)
 551                 QDATA_SET_BLK(&qunit->lq_data);
 552         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_CREATED, 0);
 553         /* it means it is only an invalid qunit for barrier */
 554         qunit->lq_opc = QUOTA_LAST_OPC;
 555
 556         while (1) {
 557                 cfs_spin_lock(&qunit_hash_lock);
 558                 find_qunit = dqacq_in_flight(qctxt, &qunit->lq_data);
 559                 if (find_qunit) {
 560                         cfs_spin_unlock(&qunit_hash_lock);
 561                         qunit_put(find_qunit);
 562                         qctxt_wait_pending_dqacq(qctxt, oqctl->qc_id,
 563                                                  oqctl->qc_type, isblk);
 564                         CDEBUG(D_QUOTA, "cycle=%d\n", cycle++);
 565                         continue;
 566                 }
 567                 break;
 568         }
 569         insert_qunit_nolock(qctxt, qunit);
 570         cfs_spin_unlock(&qunit_hash_lock);
 571         return qunit;
 572 }
 573
 574 void quota_unbarrier(void *handle)
 575 {
 576         struct lustre_qunit *qunit = (struct lustre_qunit *)handle;
 577
 578         if (qunit == NULL) {
 579                 CERROR("handle is NULL\n");
 580                 return;
 581         }
 582
 583         LASSERT(qunit->lq_opc == QUOTA_LAST_OPC);
 584         cfs_spin_lock(&qunit_hash_lock);
 585         remove_qunit_nolock(qunit);
 586         cfs_spin_unlock(&qunit_hash_lock);
 587         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, QUOTA_REQ_RETURNED);
 588         cfs_waitq_signal(&qunit->lq_waitq);
 589         qunit_put(qunit);
 590 }
 591
 592 #define INC_QLIMIT(limit, count) (limit == MIN_QLIMIT) ? \
 593                                  (limit = count) : (limit += count)
 594
 595
 596 static inline int is_master(struct lustre_quota_ctxt *qctxt)
 597 {
 598         return qctxt->lqc_handler ? 1 : 0;
 599 }
 600
 601 static int
 602 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 603                struct qunit_data *qdata, int opc, int wait,
 604                struct obd_trans_info *oti);
 605
 606 static inline void qdata_to_oqaq(struct qunit_data *qdata,
 607                                  struct quota_adjust_qunit *oqaq)
 608 {
 609         LASSERT(qdata);
 610         LASSERT(oqaq);
 611
 612         oqaq->qaq_flags = qdata->qd_flags;
 613         oqaq->qaq_id    = qdata->qd_id;
 614         if (QDATA_IS_ADJBLK(qdata))
 615                 oqaq->qaq_bunit_sz = qdata->qd_qunit;
 616         if (QDATA_IS_ADJINO(qdata))
 617                 oqaq->qaq_iunit_sz = qdata->qd_qunit;
 618 }
 619
 620 static int
 621 dqacq_completion(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 622                  struct qunit_data *qdata, int rc, int opc)
 623 {
 624         struct lustre_qunit *qunit = NULL;
 625         struct super_block *sb = qctxt->lqc_sb;
 626         int err = 0;
 627         struct quota_adjust_qunit *oqaq = NULL;
 628         int rc1 = 0;
 629         ENTRY;
 630
 631         LASSERT(qdata);
 632         QDATA_DEBUG(qdata, "obd(%s): complete %s quota req\n",
 633                     obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
 634
 635         /* do it only when a releasing quota req more than 5MB b=18491 */
 636         if (opc == QUOTA_DQREL && qdata->qd_count >= 5242880)
 637                 OBD_FAIL_TIMEOUT(OBD_FAIL_QUOTA_DELAY_REL, 5);
 638
 639         /* update local operational quota file */
 640         if (rc == 0) {
 641                 __u64 count = QUSG(qdata->qd_count, QDATA_IS_BLK(qdata));
 642                 struct obd_quotactl *qctl;
 643                 __u64 *hardlimit;
 644
 645                 OBD_ALLOC_PTR(qctl);
 646                 if (qctl == NULL)
 647                         GOTO(out, err = -ENOMEM);
 648
 649                 /* acq/rel qunit for specified uid/gid is serialized,
 650                  * so there is no race between get fs quota limit and
 651                  * set fs quota limit */
 652                 qctl->qc_cmd = Q_GETQUOTA;
 653                 qctl->qc_id = qdata->qd_id;
 654                 qctl->qc_type = QDATA_IS_GRP(qdata);
 655                 err = fsfilt_quotactl(obd, sb, qctl);
 656                 if (err) {
 657                         CERROR("error get quota fs limit! (rc:%d)\n", err);
 658                         GOTO(out_mem, err);
 659                 }
 660
 661                 if (QDATA_IS_BLK(qdata)) {
 662                         qctl->qc_dqblk.dqb_valid = QIF_BLIMITS;
 663                         hardlimit = &qctl->qc_dqblk.dqb_bhardlimit;
 664                 } else {
 665                         qctl->qc_dqblk.dqb_valid = QIF_ILIMITS;
 666                         hardlimit = &qctl->qc_dqblk.dqb_ihardlimit;
 667                 }
 668
 669                 CDEBUG(D_QUOTA, "hardlimt: "LPU64"\n", *hardlimit);
 670
 671                 if (*hardlimit == 0)
 672                         goto out_mem;
 673
 674                 switch (opc) {
 675                 case QUOTA_DQACQ:
 676                         INC_QLIMIT(*hardlimit, count);
 677                         break;
 678                 case QUOTA_DQREL:
 679                         LASSERTF(count < *hardlimit,
 680                                  "id(%u) flag(%u) type(%c) isblk(%c) "
 681                                  "count("LPU64") qd_qunit("LPU64") "
 682                                  "hardlimit("LPU64").\n",
 683                                  qdata->qd_id, qdata->qd_flags,
 684                                  QDATA_IS_GRP(qdata) ? 'g' : 'u',
 685                                  QDATA_IS_BLK(qdata) ? 'b': 'i',
 686                                  qdata->qd_count, qdata->qd_qunit, *hardlimit);
 687                         *hardlimit -= count;
 688                         break;
 689                 default:
 690                         LBUG();
 691                 }
 692
 693                 /* clear quota limit */
 694                 if (count == 0)
 695                         *hardlimit = 0;
 696
 697                 qctl->qc_cmd = Q_SETQUOTA;
 698                 err = fsfilt_quotactl(obd, sb, qctl);
 699                 if (err)
 700                         CERROR("error set quota fs limit! (rc:%d)\n", err);
 701
 702                 QDATA_DEBUG(qdata, "%s completion\n",
 703                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 704 out_mem:
 705                 OBD_FREE_PTR(qctl);
 706         } else if (rc == -EDQUOT) {
 707                 QDATA_DEBUG(qdata, "acquire qunit got EDQUOT.\n");
 708         } else if (rc == -EBUSY) {
 709                 QDATA_DEBUG(qdata, "it's is recovering, got EBUSY.\n");
 710         } else {
 711                 CERROR("acquire qunit got error! (rc:%d)\n", rc);
 712         }
 713 out:
 714         /* remove the qunit from hash */
 715         cfs_spin_lock(&qunit_hash_lock);
 716
 717         qunit = dqacq_in_flight(qctxt, qdata);
 718         /* this qunit has been removed by qctxt_cleanup() */
 719         if (!qunit) {
 720                 cfs_spin_unlock(&qunit_hash_lock);
 721                 QDATA_DEBUG(qdata, "%s is discarded because qunit isn't found\n",
 722                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 723                 RETURN(err);
 724         }
 725
 726         LASSERT(opc == qunit->lq_opc);
 727         /* remove this qunit from lq_hash so that new processes cannot be added
 728          * to qunit->lq_waiters */
 729         remove_qunit_nolock(qunit);
 730         cfs_spin_unlock(&qunit_hash_lock);
 731
 732         compute_lqs_after_removing_qunit(qunit);
 733
 734         if (rc == 0)
 735                 rc = QUOTA_REQ_RETURNED;
 736         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, rc);
 737         /* wake up all waiters */
 738         cfs_waitq_broadcast(&qunit->lq_waitq);
 739
 740         /* this is for dqacq_in_flight() */
 741         qunit_put(qunit);
 742         if (rc < 0 && rc != -EDQUOT)
 743                 GOTO(out1, err);
 744
 745         /* don't reschedule in such cases:
 746          *   - acq/rel failure and qunit isn't changed,
 747          *     but not for quota recovery.
 748          *   - local dqacq/dqrel.
 749          *   - local disk io failure.
 750          */
 751          OBD_ALLOC_PTR(oqaq);
 752          if (!oqaq)
 753                  GOTO(out1, err = -ENOMEM);
 754          qdata_to_oqaq(qdata, oqaq);
 755          /* adjust the qunit size in slaves */
 756          rc1 = quota_adjust_slave_lqs(oqaq, qctxt);
 757          OBD_FREE_PTR(oqaq);
 758          if (rc1 < 0) {
 759                  CERROR("adjust slave's qunit size failed!(rc:%d)\n", rc1);
 760                  GOTO(out1, err = rc1);
 761          }
 762          if (err || (rc < 0 && rc != -EBUSY && rc1 == 0) || is_master(qctxt))
 763                  GOTO(out1, err);
 764
 765          if (opc == QUOTA_DQREL && qdata->qd_count >= 5242880 &&
 766              OBD_FAIL_CHECK(OBD_FAIL_QUOTA_DELAY_REL))
 767                  GOTO(out1, err);
 768
 769         /* reschedule another dqacq/dqrel if needed */
 770         qdata->qd_count = 0;
 771         qdata->qd_flags &= LQUOTA_QUNIT_FLAGS;
 772         rc1 = check_cur_qunit(obd, qctxt, qdata);
 773         if (rc1 > 0) {
 774                 int opc;
 775                 opc = rc1 == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
 776                 rc1 = schedule_dqacq(obd, qctxt, qdata, opc, 0, NULL);
 777                 QDATA_DEBUG(qdata, "reschedudle opc(%d) rc(%d)\n", opc, rc1);
 778         }
 779  out1:
 780         /* this is for alloc_qunit() */
 781         qunit_put(qunit);
 782         RETURN(err);
 783 }
 784
 785 struct dqacq_async_args {
 786         struct lustre_quota_ctxt *aa_ctxt;
 787         struct lustre_qunit *aa_qunit;
 788 };
 789
 790 static int dqacq_interpret(const struct lu_env *env,
 791                            struct ptlrpc_request *req, void *data, int rc)
 792 {
 793         struct dqacq_async_args *aa = (struct dqacq_async_args *)data;
 794         struct lustre_quota_ctxt *qctxt = aa->aa_ctxt;
 795         struct obd_device_target *obt = qctxt->lqc_obt;
 796         struct lustre_qunit *qunit = aa->aa_qunit;
 797         struct obd_device *obd = req->rq_import->imp_obd;
 798         struct qunit_data *qdata = NULL;
 799         ENTRY;
 800
 801         LASSERT(req);
 802         LASSERT(req->rq_import);
 803
 804         cfs_down_read(&obt->obt_rwsem);
 805         /* if a quota req timeouts or is dropped, we should update quota
 806          * statistics which will be handled in dqacq_completion. And in
 807          * this situation we should get qdata from request instead of
 808          * reply */
 809         qdata = quota_get_qdata(req, (rc != 0) ? QUOTA_REQUEST : QUOTA_REPLY,
 810                                 QUOTA_IMPORT);
 811         if (IS_ERR(qdata)) {
 812                 rc = PTR_ERR(qdata);
 813                 DEBUG_REQ(D_ERROR, req,
 814                           "error unpacking qunit_data(rc: %ld)\n",
 815                           PTR_ERR(qdata));
 816                 qdata = &qunit->lq_data;
 817         }
 818
 819         QDATA_DEBUG(qdata, "qdata: interpret rc(%d).\n", rc);
 820         QDATA_DEBUG((&qunit->lq_data), "lq_data: \n");
 821
 822         if (qdata->qd_id != qunit->lq_data.qd_id ||
 823             OBD_FAIL_CHECK(OBD_FAIL_QUOTA_RET_QDATA)) {
 824                 CERROR("the returned qd_id isn't expected!"
 825                        "(qdata: %u, lq_data: %u)\n", qdata->qd_id,
 826                        qunit->lq_data.qd_id);
 827                 qdata->qd_id = qunit->lq_data.qd_id;
 828                 rc = -EPROTO;
 829         }
 830         if (QDATA_IS_GRP(qdata) != QDATA_IS_GRP(&qunit->lq_data)) {
 831                 CERROR("the returned grp/usr isn't expected!"
 832                        "(qdata: %u, lq_data: %u)\n", qdata->qd_flags,
 833                        qunit->lq_data.qd_flags);
 834                 if (QDATA_IS_GRP(&qunit->lq_data))
 835                         QDATA_SET_GRP(qdata);
 836                 else
 837                         QDATA_CLR_GRP(qdata);
 838                 rc = -EPROTO;
 839         }
 840         if (qdata->qd_count > qunit->lq_data.qd_count) {
 841                 CERROR("the returned qd_count isn't expected!"
 842                        "(qdata: "LPU64", lq_data: "LPU64")\n", qdata->qd_count,
 843                        qunit->lq_data.qd_count);
 844                 rc = -EPROTO;
 845         }
 846
 847         if (unlikely(rc == -ESRCH))
 848                 CERROR("quota for %s has been enabled by master, but disabled "
 849                        "by slave.\n", QDATA_IS_GRP(qdata) ? "group" : "user");
 850
 851         rc = dqacq_completion(obd, qctxt, qdata, rc,
 852                               lustre_msg_get_opc(req->rq_reqmsg));
 853
 854         cfs_up_read(&obt->obt_rwsem);
 855         RETURN(rc);
 856 }
 857
 858 /**
 859  * check if quota master is online
 860  */
 861 int check_qm(struct lustre_quota_ctxt *qctxt)
 862 {
 863         int rc;
 864         ENTRY;
 865
 866         cfs_spin_lock(&qctxt->lqc_lock);
 867         /* quit waiting when mds is back or qctxt is cleaned up */
 868         rc = qctxt->lqc_import || !qctxt->lqc_valid;
 869         cfs_spin_unlock(&qctxt->lqc_lock);
 870
 871         RETURN(rc);
 872 }
 873
 874 /* wake up all waiting threads when lqc_import is NULL */
 875 void dqacq_interrupt(struct lustre_quota_ctxt *qctxt)
 876 {
 877         struct lustre_qunit *qunit, *tmp;
 878         int i;
 879         ENTRY;
 880
 881         cfs_spin_lock(&qunit_hash_lock);
 882         for (i = 0; i < NR_DQHASH; i++) {
 883                 cfs_list_for_each_entry_safe(qunit, tmp, &qunit_hash[i],
 884                                              lq_hash) {
 885                         if (qunit->lq_ctxt != qctxt)
 886                                 continue;
 887
 888                         /* Wake up all waiters. Do not change lq_state.
 889                          * The waiters will check lq_rc which is kept as 0
 890                          * if no others change it, then the waiters will return
 891                          * -EAGAIN to caller who can perform related quota
 892                          * acq/rel if necessary. */
 893                         cfs_waitq_broadcast(&qunit->lq_waitq);
 894                 }
 895         }
 896         cfs_spin_unlock(&qunit_hash_lock);
 897         EXIT;
 898 }
 899
 900 static int got_qunit(struct lustre_qunit *qunit, int is_master)
 901 {
 902         struct lustre_quota_ctxt *qctxt = qunit->lq_ctxt;
 903         int rc = 0;
 904         ENTRY;
 905
 906         cfs_spin_lock(&qunit->lq_lock);
 907         switch (qunit->lq_state) {
 908         case QUNIT_IN_HASH:
 909         case QUNIT_RM_FROM_HASH:
 910                 break;
 911         case QUNIT_FINISHED:
 912                 rc = 1;
 913                 break;
 914         default:
 915                 CERROR("invalid qunit state %d\n", qunit->lq_state);
 916         }
 917         cfs_spin_unlock(&qunit->lq_lock);
 918
 919         if (!rc) {
 920                 cfs_spin_lock(&qctxt->lqc_lock);
 921                 rc = !qctxt->lqc_valid;
 922                 if (!is_master)
 923                         rc |= !qctxt->lqc_import;
 924                 cfs_spin_unlock(&qctxt->lqc_lock);
 925         }
 926
 927         RETURN(rc);
 928 }
 929
 930 static inline void
 931 revoke_lqs_rec(struct lustre_qunit_size *lqs, struct qunit_data *qdata, int opc)
 932 {
 933         /* revoke lqs_xxx_rec which is computed in check_cur_qunit
 934          * b=18630 */
 935         cfs_spin_lock(&lqs->lqs_lock);
 936         quota_compute_lqs(qdata, lqs, 0, (opc == QUOTA_DQACQ) ? 1 : 0);
 937         cfs_spin_unlock(&lqs->lqs_lock);
 938 }
 939
 940 static int
 941 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 942                struct qunit_data *qdata, int opc, int wait,
 943                struct obd_trans_info *oti)
 944 {
 945         struct lustre_qunit *qunit, *empty;
 946         struct l_wait_info lwi = { 0 };
 947         struct ptlrpc_request *req;
 948         struct dqacq_async_args *aa;
 949         struct obd_import *imp = NULL;
 950         struct lustre_qunit_size *lqs = NULL;
 951         struct timeval work_start;
 952         struct timeval work_end;
 953         long timediff;
 954         int rc = 0;
 955         ENTRY;
 956
 957         LASSERT(opc == QUOTA_DQACQ || opc == QUOTA_DQREL);
 958         cfs_gettimeofday(&work_start);
 959
 960         lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(qdata), qdata->qd_id),
 961                                qctxt, 0);
 962         if (lqs == NULL || IS_ERR(lqs)) {
 963                 CERROR("Can't find the lustre qunit size!\n");
 964                 RETURN(-EPERM);
 965         }
 966
 967         if ((empty = alloc_qunit(qctxt, qdata, opc)) == NULL) {
 968                 revoke_lqs_rec(lqs, qdata, opc);
 969                 /* this is for quota_search_lqs */
 970                 lqs_putref(lqs);
 971                 RETURN(-ENOMEM);
 972         }
 973
 974         OBD_FAIL_TIMEOUT(OBD_FAIL_QUOTA_DELAY_SD, 5);
 975
 976         cfs_spin_lock(&qunit_hash_lock);
 977         qunit = dqacq_in_flight(qctxt, qdata);
 978         if (qunit) {
 979                 cfs_spin_unlock(&qunit_hash_lock);
 980                 qunit_put(empty);
 981
 982                 revoke_lqs_rec(lqs, qdata, opc);
 983                 /* this is for quota_search_lqs */
 984                 lqs_putref(lqs);
 985                 goto wait_completion;
 986         }
 987         qunit = empty;
 988         qunit_get(qunit);
 989         insert_qunit_nolock(qctxt, qunit);
 990         cfs_spin_unlock(&qunit_hash_lock);
 991
 992         /* From here, the quota request will be sent anyway.
 993          * When this qdata request returned or is cancelled,
 994          * lqs_putref will be called at that time */
 995         lqs_getref(lqs);
 996         /* this is for quota_search_lqs */
 997         lqs_putref(lqs);
 998
 999         QDATA_DEBUG(qdata, "obd(%s): send %s quota req\n",
1000                     obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
1001         /* master is going to dqacq/dqrel from itself */
1002         if (is_master(qctxt)) {
1003                 int rc2;
1004                 QDATA_DEBUG(qdata, "local %s.\n",
1005                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
1006                 QDATA_SET_CHANGE_QS(qdata);
1007                 rc = qctxt->lqc_handler(obd, qdata, opc);
1008                 rc2 = dqacq_completion(obd, qctxt, qdata, rc, opc);
1009                 /* this is for qunit_get() */
1010                 qunit_put(qunit);
1011
1012                 cfs_gettimeofday(&work_end);
1013                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1014                 if (opc == QUOTA_DQACQ)
1015                         lprocfs_counter_add(qctxt->lqc_stats,
1016                                             wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
1017                                             timediff);
1018                 else
1019                         lprocfs_counter_add(qctxt->lqc_stats,
1020                                             wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
1021                                             timediff);
1022                 RETURN(rc ? rc : rc2);
1023         }
1024
1025         cfs_spin_lock(&qctxt->lqc_lock);
1026         if (!qctxt->lqc_import) {
1027                 cfs_spin_unlock(&qctxt->lqc_lock);
1028                 QDATA_DEBUG(qdata, "lqc_import is invalid.\n");
1029
1030                 cfs_spin_lock(&qunit_hash_lock);
1031                 remove_qunit_nolock(qunit);
1032                 cfs_spin_unlock(&qunit_hash_lock);
1033
1034                 compute_lqs_after_removing_qunit(qunit);
1035
1036                 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, -EAGAIN);
1037                 cfs_waitq_broadcast(&qunit->lq_waitq);
1038
1039                 /* this is for qunit_get() */
1040                 qunit_put(qunit);
1041                 /* this for alloc_qunit() */
1042                 qunit_put(qunit);
1043                 cfs_spin_lock(&qctxt->lqc_lock);
1044                 if (wait && !qctxt->lqc_import) {
1045                         cfs_spin_unlock(&qctxt->lqc_lock);
1046                         LASSERT(oti && oti->oti_thread);
1047                         /* The recovery thread doesn't have watchdog
1048                          * attached. LU-369 */
1049                         if (oti->oti_thread->t_watchdog)
1050                                 lc_watchdog_disable(oti->oti_thread->\
1051                                                 t_watchdog);
1052                         CDEBUG(D_QUOTA, "sleep for quota master\n");
1053                         l_wait_event(qctxt->lqc_wait_for_qmaster,
1054                                      check_qm(qctxt), &lwi);
1055                         CDEBUG(D_QUOTA, "wake up when quota master is back\n");
1056                         if (oti->oti_thread->t_watchdog)
1057                                 lc_watchdog_touch(oti->oti_thread->t_watchdog,
1058                                       CFS_GET_TIMEOUT(oti->oti_thread->t_svc));
1059                 } else {
1060                         cfs_spin_unlock(&qctxt->lqc_lock);
1061                 }
1062
1063                 RETURN(-EAGAIN);
1064         }
1065         imp = class_import_get(qctxt->lqc_import);
1066         cfs_spin_unlock(&qctxt->lqc_lock);
1067
1068         /* build dqacq/dqrel request */
1069         LASSERT(imp);
1070
1071         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_QUOTA_DQACQ,
1072                                         LUSTRE_MDS_VERSION, opc);
1073         class_import_put(imp);
1074         if (req == NULL) {
1075                 CERROR("Can't alloc request\n");
1076                 dqacq_completion(obd, qctxt, qdata, -ENOMEM, opc);
1077                 /* this is for qunit_get() */
1078                 qunit_put(qunit);
1079                 RETURN(-ENOMEM);
1080         }
1081
1082         ptlrpc_request_set_replen(req);
1083         req->rq_no_resend = req->rq_no_delay = 1;
1084         rc = quota_copy_qdata(req, qdata, QUOTA_REQUEST, QUOTA_IMPORT);
1085         if (rc < 0) {
1086                 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
1087                 ptlrpc_req_finished(req);
1088                 dqacq_completion(obd, qctxt, qdata, -EPROTO, opc);
1089                 /* this is for qunit_get() */
1090                 qunit_put(qunit);
1091                 RETURN(rc);
1092         }
1093
1094         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1095         aa = ptlrpc_req_async_args(req);
1096         aa->aa_ctxt = qctxt;
1097         aa->aa_qunit = qunit;
1098
1099         req->rq_interpret_reply = dqacq_interpret;
1100         ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1101
1102         QDATA_DEBUG(qdata, "%s scheduled.\n",
1103                     opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
1104 wait_completion:
1105         if (wait && qunit) {
1106                 struct qunit_data *p = &qunit->lq_data;
1107
1108                 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
1109                 l_wait_event(qunit->lq_waitq, got_qunit(qunit, is_master(qctxt)),
1110                              &lwi);
1111                 /* rc = -EAGAIN, it means the quota master isn't ready yet
1112                  * rc = QUOTA_REQ_RETURNED, it means a quota req is finished;
1113                  * rc = -EDQUOT, it means out of quota
1114                  * rc = -EBUSY, it means recovery is happening
1115                  * other rc < 0, it means real errors, functions who call
1116                  * schedule_dqacq should take care of this */
1117                 cfs_spin_lock(&qunit->lq_lock);
1118                 rc = qunit->lq_rc;
1119                 cfs_spin_unlock(&qunit->lq_lock);
1120                 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting: id(%u) flag(%u) "
1121                        "rc(%d) owner(%d)\n", qunit, qunit->lq_data.qd_id,
1122                        qunit->lq_data.qd_flags, rc, qunit->lq_owner);
1123         }
1124
1125         qunit_put(qunit);
1126         cfs_gettimeofday(&work_end);
1127         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1128         if (opc == QUOTA_DQACQ)
1129                 lprocfs_counter_add(qctxt->lqc_stats,
1130                                     wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
1131                                     timediff);
1132         else
1133                 lprocfs_counter_add(qctxt->lqc_stats,
1134                                     wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
1135                                     timediff);
1136
1137         RETURN(rc);
1138 }
1139
1140 int
1141 qctxt_adjust_qunit(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
1142                    const unsigned int id[], __u32 isblk, int wait,
1143                    struct obd_trans_info *oti)
1144 {
1145         int rc = 0, i = USRQUOTA;
1146         struct qunit_data qdata[MAXQUOTAS];
1147         ENTRY;
1148
1149         if (quota_is_set(obd, id, isblk ? QB_SET : QI_SET) == 0)
1150                 RETURN(0);
1151
1152         for (i = 0; i < MAXQUOTAS; i++) {
1153                 qdata[i].qd_id = id[i];
1154                 qdata[i].qd_flags = i;
1155                 if (isblk)
1156                         QDATA_SET_BLK(&qdata[i]);
1157                 qdata[i].qd_count = 0;
1158
1159                 rc = check_cur_qunit(obd, qctxt, &qdata[i]);
1160                 if (rc > 0) {
1161                         int opc;
1162                         /* need acquire or release */
1163                         opc = rc == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1164                         rc = schedule_dqacq(obd, qctxt, &qdata[i], opc,
1165                                             wait,oti);
1166                         if (rc < 0)
1167                                 RETURN(rc);
1168                 } else if (wait == 1) {
1169                         /* when wait equates 1, that means mds_quota_acquire
1170                          * or filter_quota_acquire is calling it. */
1171                         rc = qctxt_wait_pending_dqacq(qctxt, id[i], i, isblk);
1172                         if (rc < 0)
1173                                 RETURN(rc);
1174                 }
1175         }
1176
1177         RETURN(rc);
1178 }
1179
1180 int
1181 qctxt_wait_pending_dqacq(struct lustre_quota_ctxt *qctxt, unsigned int id,
1182                          unsigned short type, int isblk)
1183 {
1184         struct lustre_qunit *qunit = NULL;
1185         struct qunit_data qdata;
1186         struct timeval work_start;
1187         struct timeval work_end;
1188         long timediff;
1189         struct l_wait_info lwi = { 0 };
1190         int rc = 0;
1191         ENTRY;
1192
1193         cfs_gettimeofday(&work_start);
1194         qdata.qd_id = id;
1195         qdata.qd_flags = type;
1196         if (isblk)
1197                 QDATA_SET_BLK(&qdata);
1198         qdata.qd_count = 0;
1199
1200         cfs_spin_lock(&qunit_hash_lock);
1201         qunit = dqacq_in_flight(qctxt, &qdata);
1202         cfs_spin_unlock(&qunit_hash_lock);
1203
1204         if (qunit) {
1205                 struct qunit_data *p = &qunit->lq_data;
1206
1207                 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
1208                 l_wait_event(qunit->lq_waitq, got_qunit(qunit, is_master(qctxt)),
1209                              &lwi);
1210                 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting: rc(%d) "
1211                        "owner(%d)\n", qunit, qunit->lq_rc, qunit->lq_owner);
1212                 /* keep same as schedule_dqacq() b=17030 */
1213                 cfs_spin_lock(&qunit->lq_lock);
1214                 rc = qunit->lq_rc;
1215                 cfs_spin_unlock(&qunit->lq_lock);
1216                 /* this is for dqacq_in_flight() */
1217                 qunit_put(qunit);
1218                 cfs_gettimeofday(&work_end);
1219                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1220                 lprocfs_counter_add(qctxt->lqc_stats,
1221                                     isblk ? LQUOTA_WAIT_PENDING_BLK_QUOTA :
1222                                             LQUOTA_WAIT_PENDING_INO_QUOTA,
1223                                     timediff);
1224         } else {
1225                 cfs_gettimeofday(&work_end);
1226                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1227                 lprocfs_counter_add(qctxt->lqc_stats,
1228                                     isblk ? LQUOTA_NOWAIT_PENDING_BLK_QUOTA :
1229                                             LQUOTA_NOWAIT_PENDING_INO_QUOTA,
1230                                     timediff);
1231         }
1232
1233         RETURN(rc);
1234 }
1235
1236 int
1237 qctxt_init(struct obd_device *obd, dqacq_handler_t handler)
1238 {
1239         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
1240         struct obd_device_target *obt = &obd->u.obt;
1241         struct super_block *sb = obt->obt_sb;
1242         int rc = 0;
1243         ENTRY;
1244
1245         LASSERT(qctxt);
1246
1247         rc = ptlrpcd_addref();
1248         if (rc)
1249                 RETURN(rc);
1250
1251         cfs_waitq_init(&qctxt->lqc_wait_for_qmaster);
1252         cfs_waitq_init(&qctxt->lqc_lqs_waitq);
1253         cfs_atomic_set(&qctxt->lqc_lqs, 0);
1254         cfs_spin_lock_init(&qctxt->lqc_lock);
1255         cfs_spin_lock(&qctxt->lqc_lock);
1256         qctxt->lqc_handler = handler;
1257         qctxt->lqc_sb = sb;
1258         qctxt->lqc_obt = obt;
1259         qctxt->lqc_import = NULL;
1260         qctxt->lqc_recovery = 0;
1261         qctxt->lqc_switch_qs = 1; /* Change qunit size in default setting */
1262         qctxt->lqc_valid = 1;
1263         qctxt->lqc_cqs_boundary_factor = 4;
1264         qctxt->lqc_cqs_least_bunit = PTLRPC_MAX_BRW_SIZE;
1265         qctxt->lqc_cqs_least_iunit = 2;
1266         qctxt->lqc_cqs_qs_factor = 2;
1267         qctxt->lqc_flags = 0;
1268         QUOTA_MASTER_UNREADY(qctxt);
1269         qctxt->lqc_bunit_sz = default_bunit_sz;
1270         qctxt->lqc_btune_sz = default_bunit_sz / 100 * default_btune_ratio;
1271         qctxt->lqc_iunit_sz = default_iunit_sz;
1272         qctxt->lqc_itune_sz = default_iunit_sz * default_itune_ratio / 100;
1273         qctxt->lqc_switch_seconds = 300; /* enlarging will wait 5 minutes
1274                                           * after the last shrinking */
1275         qctxt->lqc_sync_blk = 0;
1276         cfs_spin_unlock(&qctxt->lqc_lock);
1277
1278         qctxt->lqc_lqs_hash = cfs_hash_create("LQS_HASH",
1279                                               hash_lqs_cur_bits,
1280                                               HASH_LQS_MAX_BITS,
1281                                               min(hash_lqs_cur_bits,
1282                                                   HASH_LQS_BKT_BITS),
1283                                               0, CFS_HASH_MIN_THETA,
1284                                               CFS_HASH_MAX_THETA,
1285                                               &lqs_hash_ops, CFS_HASH_DEFAULT);
1286         if (!qctxt->lqc_lqs_hash) {
1287                 CERROR("initialize hash lqs for %s error!\n", obd->obd_name);
1288                 RETURN(-ENOMEM);
1289         }
1290
1291 #ifdef LPROCFS
1292         rc = lquota_proc_setup(obd, is_master(qctxt));
1293         if (rc)
1294                 CERROR("initialize proc for %s error!\n", obd->obd_name);
1295 #endif
1296
1297         RETURN(rc);
1298 }
1299
1300 static int check_lqs(struct lustre_quota_ctxt *qctxt)
1301 {
1302         int rc;
1303         ENTRY;
1304
1305         rc = !cfs_atomic_read(&qctxt->lqc_lqs);
1306
1307         RETURN(rc);
1308 }
1309
1310 int qctxt_del_lqs(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1311                  cfs_hlist_node_t *hnode, void *data)
1312 {
1313         /* remove from hash and -1 refcount */
1314         cfs_hash_bd_del_locked(hs, bd, hnode);
1315         return 0;
1316 }
1317
1318 void qctxt_cleanup(struct lustre_quota_ctxt *qctxt, int force)
1319 {
1320         struct lustre_qunit *qunit, *tmp;
1321         cfs_list_t tmp_list;
1322         struct l_wait_info lwi = { 0 };
1323         struct obd_device_target *obt = qctxt->lqc_obt;
1324         int i;
1325         ENTRY;
1326
1327         CFS_INIT_LIST_HEAD(&tmp_list);
1328
1329         cfs_spin_lock(&qctxt->lqc_lock);
1330         qctxt->lqc_valid = 0;
1331         cfs_spin_unlock(&qctxt->lqc_lock);
1332
1333         cfs_spin_lock(&qunit_hash_lock);
1334         for (i = 0; i < NR_DQHASH; i++) {
1335                 cfs_list_for_each_entry_safe(qunit, tmp, &qunit_hash[i],
1336                                              lq_hash) {
1337                         if (qunit->lq_ctxt != qctxt)
1338                                 continue;
1339                         remove_qunit_nolock(qunit);
1340                         cfs_list_add(&qunit->lq_hash, &tmp_list);
1341                 }
1342         }
1343         cfs_spin_unlock(&qunit_hash_lock);
1344
1345         cfs_list_for_each_entry_safe(qunit, tmp, &tmp_list, lq_hash) {
1346                 cfs_list_del_init(&qunit->lq_hash);
1347                 compute_lqs_after_removing_qunit(qunit);
1348
1349                 /* wake up all waiters */
1350                 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, 0);
1351                 cfs_waitq_broadcast(&qunit->lq_waitq);
1352                 qunit_put(qunit);
1353         }
1354
1355         /* after qctxt_cleanup, qctxt might be freed, then check_qm() is
1356          * unpredicted. So we must wait until lqc_wait_for_qmaster is empty */
1357         while (cfs_waitq_active(&qctxt->lqc_wait_for_qmaster)) {
1358                 cfs_waitq_signal(&qctxt->lqc_wait_for_qmaster);
1359                 cfs_schedule_timeout_and_set_state(CFS_TASK_INTERRUPTIBLE,
1360                                                    cfs_time_seconds(1));
1361         }
1362
1363         /* release refcount on lustre_qunit_size holding by lqs_hash */
1364         cfs_hash_for_each_safe(qctxt->lqc_lqs_hash, qctxt_del_lqs, NULL);
1365
1366         l_wait_event(qctxt->lqc_lqs_waitq, check_lqs(qctxt), &lwi);
1367         cfs_down_write(&obt->obt_rwsem);
1368         cfs_hash_putref(qctxt->lqc_lqs_hash);
1369         qctxt->lqc_lqs_hash = NULL;
1370         cfs_up_write(&obt->obt_rwsem);
1371
1372         ptlrpcd_decref();
1373
1374 #ifdef LPROCFS
1375         if (lquota_proc_cleanup(qctxt))
1376                 CERROR("cleanup proc error!\n");
1377 #endif
1378
1379         EXIT;
1380 }
1381
1382 struct qslave_recov_thread_data {
1383         struct obd_device *obd;
1384         struct lustre_quota_ctxt *qctxt;
1385         cfs_completion_t comp;
1386 };
1387
1388 /* FIXME only recovery block quota by now */
1389 static int qslave_recovery_main(void *arg)
1390 {
1391         struct qslave_recov_thread_data *data = arg;
1392         struct obd_device *obd = data->obd;
1393         struct lustre_quota_ctxt *qctxt = data->qctxt;
1394         unsigned int type;
1395         int rc = 0;
1396         ENTRY;
1397
1398         cfs_daemonize_ctxt("qslave_recovd");
1399
1400         /* for obdfilter */
1401         class_incref(obd, "qslave_recovd_filter", obd);
1402
1403         cfs_complete(&data->comp);
1404
1405         cfs_spin_lock(&qctxt->lqc_lock);
1406         if (qctxt->lqc_recovery) {
1407                 cfs_spin_unlock(&qctxt->lqc_lock);
1408                 class_decref(obd, "qslave_recovd_filter", obd);
1409                 RETURN(0);
1410         } else {
1411                 qctxt->lqc_recovery = 1;
1412                 cfs_spin_unlock(&qctxt->lqc_lock);
1413         }
1414
1415         for (type = USRQUOTA; type < MAXQUOTAS; type++) {
1416                 struct qunit_data qdata;
1417                 struct quota_info *dqopt = sb_dqopt(qctxt->lqc_sb);
1418                 cfs_list_t id_list;
1419                 struct dquot_id *dqid, *tmp;
1420                 int ret;
1421
1422                 LOCK_DQONOFF_MUTEX(dqopt);
1423                 if (!ll_sb_has_quota_active(qctxt->lqc_sb, type)) {
1424                         UNLOCK_DQONOFF_MUTEX(dqopt);
1425                         break;
1426                 }
1427
1428                 LASSERT(dqopt->files[type] != NULL);
1429                 CFS_INIT_LIST_HEAD(&id_list);
1430 #ifndef KERNEL_SUPPORTS_QUOTA_READ
1431                 rc = fsfilt_qids(obd, dqopt->files[type], NULL, type, &id_list);
1432 #else
1433                 rc = fsfilt_qids(obd, NULL, dqopt->files[type], type, &id_list);
1434 #endif
1435                 UNLOCK_DQONOFF_MUTEX(dqopt);
1436                 if (rc)
1437                         CERROR("Get ids from quota file failed. (rc:%d)\n", rc);
1438
1439                 cfs_list_for_each_entry_safe(dqid, tmp, &id_list, di_link) {
1440                         cfs_list_del_init(&dqid->di_link);
1441                         /* skip slave recovery on itself */
1442                         if (is_master(qctxt))
1443                                 goto free;
1444                         if (rc && rc != -EBUSY)
1445                                 goto free;
1446
1447                         qdata.qd_id = dqid->di_id;
1448                         qdata.qd_flags = type;
1449                         QDATA_SET_BLK(&qdata);
1450                         qdata.qd_count = 0;
1451
1452                         ret = check_cur_qunit(obd, qctxt, &qdata);
1453                         if (ret > 0) {
1454                                 int opc;
1455                                 opc = ret == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1456                                 rc = schedule_dqacq(obd, qctxt, &qdata, opc,
1457                                                     0, NULL);
1458                                 if (rc == -EDQUOT)
1459                                         rc = 0;
1460                         } else {
1461                                 rc = 0;
1462                         }
1463
1464                         if (rc && rc != -EBUSY)
1465                                 CERROR("qslave recovery failed! (id:%d type:%d "
1466                                        " rc:%d)\n", dqid->di_id, type, rc);
1467 free:
1468                         OBD_FREE_PTR(dqid);
1469                 }
1470         }
1471
1472         cfs_spin_lock(&qctxt->lqc_lock);
1473         qctxt->lqc_recovery = 0;
1474         cfs_spin_unlock(&qctxt->lqc_lock);
1475         class_decref(obd, "qslave_recovd_filter", obd);
1476         RETURN(rc);
1477 }
1478
1479 void
1480 qslave_start_recovery(struct obd_device *obd, struct lustre_quota_ctxt *qctxt)
1481 {
1482         struct qslave_recov_thread_data data;
1483         int rc;
1484         ENTRY;
1485
1486         if (!ll_sb_any_quota_active(qctxt->lqc_sb))
1487                 goto exit;
1488
1489         data.obd = obd;
1490         data.qctxt = qctxt;
1491         cfs_init_completion(&data.comp);
1492
1493         rc = cfs_create_thread(qslave_recovery_main, &data,
1494                                CFS_DAEMON_FLAGS);
1495         if (rc < 0) {
1496                 CERROR("Cannot start quota recovery thread: rc %d\n", rc);
1497                 goto exit;
1498         }
1499         cfs_wait_for_completion(&data.comp);
1500 exit:
1501         EXIT;
1502 }
1503
1504 inline int quota_is_on(struct lustre_quota_ctxt *qctxt,
1505                        struct obd_quotactl *oqctl)
1506 {
1507         return ((qctxt->lqc_flags & UGQUOTA2LQC(oqctl->qc_type)) ==
1508                 UGQUOTA2LQC(oqctl->qc_type));
1509 }
1510
1511 inline int quota_is_off(struct lustre_quota_ctxt *qctxt,
1512                         struct obd_quotactl *oqctl)
1513 {
1514         return !(qctxt->lqc_flags & UGQUOTA2LQC(oqctl->qc_type));
1515 }
1516
1517 /**
1518  * When quotaon, build a lqs for every uid/gid who has been set limitation
1519  * for quota. After quota_search_lqs, it will hold one ref for the lqs.
1520  * It will be released when qctxt_cleanup() is executed b=18574
1521  *
1522  * Should be called with obt->obt_quotachecking held. b=20152
1523  */
1524 void build_lqs(struct obd_device *obd)
1525 {
1526         struct obd_device_target *obt = &obd->u.obt;
1527         struct lustre_quota_ctxt *qctxt = &obt->obt_qctxt;
1528         cfs_list_t id_list;
1529         int i, rc;
1530
1531         LASSERT_SEM_LOCKED(&obt->obt_quotachecking);
1532         CFS_INIT_LIST_HEAD(&id_list);
1533         for (i = 0; i < MAXQUOTAS; i++) {
1534                 struct dquot_id *dqid, *tmp;
1535
1536                 if (sb_dqopt(qctxt->lqc_sb)->files[i] == NULL)
1537                         continue;
1538
1539 #ifndef KERNEL_SUPPORTS_QUOTA_READ
1540                 rc = fsfilt_qids(obd, sb_dqopt(qctxt->lqc_sb)->files[i], NULL,
1541                                  i, &id_list);
1542 #else
1543                 rc = fsfilt_qids(obd, NULL, sb_dqopt(qctxt->lqc_sb)->files[i],
1544                                  i, &id_list);
1545 #endif
1546                 if (rc) {
1547                         CERROR("%s: failed to get %s qids!\n", obd->obd_name,
1548                                i ? "group" : "user");
1549                         continue;
1550                 }
1551
1552                 cfs_list_for_each_entry_safe(dqid, tmp, &id_list,
1553                                              di_link) {
1554                         struct lustre_qunit_size *lqs;
1555
1556                         cfs_list_del_init(&dqid->di_link);
1557                         lqs = quota_search_lqs(LQS_KEY(i, dqid->di_id),
1558                                                qctxt, 1);
1559                         if (lqs && !IS_ERR(lqs)) {
1560                                 lqs->lqs_flags |= dqid->di_flag;
1561                                 lqs_putref(lqs);
1562                         } else {
1563                                 CERROR("%s: failed to create a lqs for %sid %u"
1564                                        "\n", obd->obd_name, i ? "g" : "u",
1565                                        dqid->di_id);
1566                         }
1567
1568                         OBD_FREE_PTR(dqid);
1569                 }
1570         }
1571 }
1572
1573 /**
1574  * lqs<->qctxt hash operations
1575  */
1576
1577 /**
1578  * string hashing using djb2 hash algorithm
1579  */
1580 static unsigned
1581 lqs_hash(cfs_hash_t *hs, const void *key, unsigned mask)
1582 {
1583         unsigned long long id;
1584         unsigned hash;
1585         ENTRY;
1586
1587         LASSERT(key);
1588         id = *((unsigned long long *)key);
1589         hash = (LQS_KEY_GRP(id) ? 5381 : 5387) * (unsigned)LQS_KEY_ID(id);
1590
1591         RETURN(hash & mask);
1592 }
1593
1594 static void *
1595 lqs_key(cfs_hlist_node_t *hnode)
1596 {
1597         struct lustre_qunit_size *lqs;
1598         ENTRY;
1599
1600         lqs = cfs_hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1601         RETURN(&lqs->lqs_key);
1602 }
1603
1604 static int
1605 lqs_keycmp(const void *key, cfs_hlist_node_t *hnode)
1606 {
1607         struct lustre_qunit_size *q =
1608                 cfs_hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1609
1610         RETURN(q->lqs_key == *((unsigned long long *)key));
1611 }
1612
1613 static void *
1614 lqs_object(cfs_hlist_node_t *hnode)
1615 {
1616         return cfs_hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1617 }
1618
1619 static void
1620 lqs_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
1621 {
1622         struct lustre_qunit_size *q =
1623                 cfs_hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1624
1625         lqs_getref(q);
1626 }
1627
1628 static void
1629 lqs_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
1630 {
1631         struct lustre_qunit_size *q =
1632                 cfs_hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1633
1634         lqs_putref(q);
1635 }
1636
1637 static void
1638 lqs_exit(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
1639 {
1640         CERROR("It should not have any item left to be handled by this!");
1641 }
1642
1643 static cfs_hash_ops_t lqs_hash_ops = {
1644         .hs_hash        = lqs_hash,
1645         .hs_key         = lqs_key,
1646         .hs_keycmp      = lqs_keycmp,
1647         .hs_object      = lqs_object,
1648         .hs_get         = lqs_get,
1649         .hs_put_locked  = lqs_put_locked,
1650         .hs_exit        = lqs_exit
1651 };
1652 #endif /* HAVE_QUOTA_SUPPORT */