lustre/quota/quota_context.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/quota/quota_context.c
  37  *
  38  * Lustre Quota Context
  39  *
  40  * Author: Niu YaWei <niu@clusterfs.com>
  41  */
  42
  43 #ifndef EXPORT_SYMTAB
  44 # define EXPORT_SYMTAB
  45 #endif
  46
  47 #define DEBUG_SUBSYSTEM S_LQUOTA
  48
  49 #include <linux/version.h>
  50 #include <linux/fs.h>
  51 #include <asm/unistd.h>
  52 #include <linux/slab.h>
  53 #include <linux/quotaops.h>
  54 #include <linux/module.h>
  55 #include <linux/init.h>
  56
  57 #include <obd_class.h>
  58 #include <lustre_quota.h>
  59 #include <lustre_fsfilt.h>
  60 #include <lprocfs_status.h>
  61 #include "quota_internal.h"
  62
  63 #ifdef HAVE_QUOTA_SUPPORT
  64
  65 static cfs_hash_ops_t lqs_hash_ops;
  66
  67 unsigned long default_bunit_sz = 128 * 1024 * 1024; /* 128M bytes */
  68 unsigned long default_btune_ratio = 50;             /* 50 percentage */
  69 unsigned long default_iunit_sz = 5120;              /* 5120 inodes */
  70 unsigned long default_itune_ratio = 50;             /* 50 percentage */
  71
  72 cfs_mem_cache_t *qunit_cachep = NULL;
  73 struct list_head qunit_hash[NR_DQHASH];
  74 spinlock_t qunit_hash_lock = SPIN_LOCK_UNLOCKED;
  75
  76 /* please sync qunit_state with qunit_state_names */
  77 enum qunit_state {
  78         /**
  79          * a qunit is created
  80          */
  81         QUNIT_CREATED      = 0,
  82         /**
  83          * a qunit is added into qunit hash, that means
  84          * a quota req will be sent or is flying
  85          */
  86         QUNIT_IN_HASH      = 1,
  87         /**
  88          * a qunit is removed from qunit hash, that
  89          * means a quota req is handled and comes back
  90          */
  91         QUNIT_RM_FROM_HASH = 2,
  92         /**
  93          * qunit can wake up all threads waiting for it
  94          */
  95         QUNIT_FINISHED     = 3,
  96 };
  97
  98 static const char *qunit_state_names[] = {
  99         [QUNIT_CREATED]      = "CREATED",
 100         [QUNIT_IN_HASH]      = "IN_HASH",
 101         [QUNIT_RM_FROM_HASH] = "RM_FROM_HASH",
 102         [QUNIT_FINISHED]     = "FINISHED",
 103 };
 104
 105 struct lustre_qunit {
 106         struct list_head lq_hash;          /** Hash list in memory */
 107         atomic_t lq_refcnt;                /** Use count */
 108         struct lustre_quota_ctxt *lq_ctxt; /** Quota context this applies to */
 109         struct qunit_data lq_data;         /** See qunit_data */
 110         unsigned int lq_opc;               /** QUOTA_DQACQ, QUOTA_DQREL */
 111         cfs_waitq_t lq_waitq;              /** Threads waiting for this qunit */
 112         spinlock_t lq_lock;                /** Protect the whole structure */
 113         enum qunit_state lq_state;         /** Present the status of qunit */
 114         int lq_rc;                         /** The rc of lq_data */
 115         pid_t lq_owner;
 116 };
 117
 118 #define QUNIT_SET_STATE(qunit, state)                                   \
 119 do {                                                                    \
 120         spin_lock(&qunit->lq_lock);                                     \
 121         QDATA_DEBUG((&qunit->lq_data), "qunit(%p) lq_state(%s->%s), "   \
 122                     "lq_rc(%d), lq_owner(%d)\n",                        \
 123                     qunit, qunit_state_names[qunit->lq_state],          \
 124                     qunit_state_names[state], qunit->lq_rc,             \
 125                     qunit->lq_owner);                                   \
 126         qunit->lq_state = state;                                        \
 127         spin_unlock(&qunit->lq_lock);                                   \
 128 } while(0)
 129
 130 #define QUNIT_SET_STATE_AND_RC(qunit, state, rc)                        \
 131 do {                                                                    \
 132         spin_lock(&qunit->lq_lock);                                     \
 133         qunit->lq_rc = rc;                                              \
 134         QDATA_DEBUG((&qunit->lq_data), "qunit(%p) lq_state(%s->%s), "   \
 135                     "lq_rc(%d), lq_owner(%d)\n",                        \
 136                     qunit, qunit_state_names[qunit->lq_state],          \
 137                     qunit_state_names[state], qunit->lq_rc,             \
 138                     qunit->lq_owner);                                   \
 139         qunit->lq_state = state;                                        \
 140         spin_unlock(&qunit->lq_lock);                                   \
 141 } while(0)
 142
 143 int should_translate_quota (struct obd_import *imp)
 144 {
 145         ENTRY;
 146
 147         LASSERT(imp);
 148         if (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_QUOTA64)
 149                 RETURN(0);
 150         else
 151                 RETURN(1);
 152 }
 153
 154 void qunit_cache_cleanup(void)
 155 {
 156         int i;
 157         ENTRY;
 158
 159         spin_lock(&qunit_hash_lock);
 160         for (i = 0; i < NR_DQHASH; i++)
 161                 LASSERT(list_empty(qunit_hash + i));
 162         spin_unlock(&qunit_hash_lock);
 163
 164         if (qunit_cachep) {
 165                 int rc;
 166                 rc = cfs_mem_cache_destroy(qunit_cachep);
 167                 LASSERTF(rc == 0, "couldn't destroy qunit_cache slab\n");
 168                 qunit_cachep = NULL;
 169         }
 170         EXIT;
 171 }
 172
 173 int qunit_cache_init(void)
 174 {
 175         int i;
 176         ENTRY;
 177
 178         LASSERT(qunit_cachep == NULL);
 179         qunit_cachep = cfs_mem_cache_create("ll_qunit_cache",
 180                                             sizeof(struct lustre_qunit),
 181                                             0, 0);
 182         if (!qunit_cachep)
 183                 RETURN(-ENOMEM);
 184
 185         spin_lock(&qunit_hash_lock);
 186         for (i = 0; i < NR_DQHASH; i++)
 187                 CFS_INIT_LIST_HEAD(qunit_hash + i);
 188         spin_unlock(&qunit_hash_lock);
 189         RETURN(0);
 190 }
 191
 192 static inline int
 193 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
 194              __attribute__((__const__));
 195
 196 static inline int
 197 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
 198 {
 199         unsigned int id = qdata->qd_id;
 200         unsigned int type = QDATA_IS_GRP(qdata);
 201
 202         unsigned long tmp = ((unsigned long)qctxt >> L1_CACHE_SHIFT) ^ id;
 203         tmp = (tmp * (MAXQUOTAS - type)) % NR_DQHASH;
 204         return tmp;
 205 }
 206
 207 /* caller must hold qunit_hash_lock */
 208 static inline struct lustre_qunit *find_qunit(unsigned int hashent,
 209                                               struct lustre_quota_ctxt *qctxt,
 210                                               struct qunit_data *qdata)
 211 {
 212         struct lustre_qunit *qunit = NULL;
 213         struct qunit_data *tmp;
 214
 215         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
 216         list_for_each_entry(qunit, qunit_hash + hashent, lq_hash) {
 217                 tmp = &qunit->lq_data;
 218                 if (qunit->lq_ctxt == qctxt &&
 219                     qdata->qd_id == tmp->qd_id &&
 220                     (qdata->qd_flags & LQUOTA_QUNIT_FLAGS) ==
 221                     (tmp->qd_flags & LQUOTA_QUNIT_FLAGS))
 222                         return qunit;
 223         }
 224         return NULL;
 225 }
 226
 227 /* check_cur_qunit - check the current usage of qunit.
 228  * @qctxt: quota context
 229  * @qdata: the type of quota unit to be checked
 230  *
 231  * return: 1 - need acquire qunit;
 232  *         2 - need release qunit;
 233  *         0 - need do nothing.
 234  *       < 0 - error.
 235  */
 236 static int
 237 check_cur_qunit(struct obd_device *obd,
 238                 struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
 239 {
 240         struct super_block *sb = qctxt->lqc_sb;
 241         unsigned long qunit_sz, tune_sz;
 242         __u64 usage, limit, limit_org, pending_write = 0;
 243         long long record = 0;
 244         struct obd_quotactl *qctl;
 245         struct lustre_qunit_size *lqs = NULL;
 246         int ret = 0;
 247         ENTRY;
 248
 249         if (!ll_sb_any_quota_active(sb))
 250                 RETURN(0);
 251
 252         spin_lock(&qctxt->lqc_lock);
 253         if (!qctxt->lqc_valid){
 254                 spin_unlock(&qctxt->lqc_lock);
 255                 RETURN(0);
 256         }
 257         spin_unlock(&qctxt->lqc_lock);
 258
 259         OBD_ALLOC_PTR(qctl);
 260         if (qctl == NULL)
 261                 RETURN(-ENOMEM);
 262
 263         /* get fs quota usage & limit */
 264         qctl->qc_cmd = Q_GETQUOTA;
 265         qctl->qc_id = qdata->qd_id;
 266         qctl->qc_type = QDATA_IS_GRP(qdata);
 267         ret = fsfilt_quotactl(obd, sb, qctl);
 268         if (ret) {
 269                 if (ret == -ESRCH)      /* no limit */
 270                         ret = 0;
 271                 else
 272                         CERROR("can't get fs quota usage! (rc:%d)\n", ret);
 273                 GOTO(out, ret);
 274         }
 275
 276         if (QDATA_IS_BLK(qdata)) {
 277                 usage = qctl->qc_dqblk.dqb_curspace;
 278                 limit = qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS;
 279         } else {
 280                 usage = qctl->qc_dqblk.dqb_curinodes;
 281                 limit = qctl->qc_dqblk.dqb_ihardlimit;
 282         }
 283
 284         /* ignore the no quota limit case; and it can avoid creating
 285          * unnecessary lqs for uid/gid */
 286         if (!limit)
 287                 GOTO(out, ret = 0);
 288
 289         lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(qdata), qdata->qd_id),
 290                                qctxt, 0);
 291         if (IS_ERR(lqs) || lqs == NULL) {
 292                 CERROR("fail to find a lqs for %sid: %u)!\n",
 293                        QDATA_IS_GRP(qdata) ? "g" : "u", qdata->qd_id);
 294                 GOTO (out, ret = 0);
 295         }
 296         spin_lock(&lqs->lqs_lock);
 297
 298         if (QDATA_IS_BLK(qdata)) {
 299                 qunit_sz = lqs->lqs_bunit_sz;
 300                 tune_sz  = lqs->lqs_btune_sz;
 301                 pending_write = lqs->lqs_bwrite_pending;
 302                 record   = lqs->lqs_blk_rec;
 303                 LASSERT(!(qunit_sz % QUOTABLOCK_SIZE));
 304         } else {
 305                 /* we didn't need change inode qunit size now */
 306                 qunit_sz = lqs->lqs_iunit_sz;
 307                 tune_sz  = lqs->lqs_itune_sz;
 308                 pending_write = lqs->lqs_iwrite_pending;
 309                 record   = lqs->lqs_ino_rec;
 310         }
 311
 312         /* we don't count the MIN_QLIMIT */
 313         if ((limit == MIN_QLIMIT && !QDATA_IS_BLK(qdata)) ||
 314             (toqb(limit) == MIN_QLIMIT && QDATA_IS_BLK(qdata)))
 315                 limit = 0;
 316
 317         usage += pending_write;
 318         limit_org = limit;
 319         /* when a releasing quota req is sent, before it returned
 320            limit is assigned a small value. limit will overflow */
 321         if (limit + record < 0)
 322                 usage -= record;
 323         else
 324                 limit += record;
 325
 326         LASSERT(qdata->qd_count == 0);
 327         if (limit <= usage + tune_sz) {
 328                 while (qdata->qd_count + limit <=
 329                        usage + tune_sz)
 330                         qdata->qd_count += qunit_sz;
 331                 ret = 1;
 332         } else if (limit > usage + qunit_sz + tune_sz &&
 333                    limit_org > qdata->qd_count + qunit_sz) {
 334                 while (limit - qdata->qd_count > usage + qunit_sz + tune_sz &&
 335                        limit_org > qdata->qd_count + qunit_sz)
 336                         qdata->qd_count += qunit_sz;
 337                 ret = 2;
 338                 /* if there are other pending writes for this uid/gid, releasing
 339                  * quota is put off until the last pending write b=16645 */
 340                 /* if there is an ongoing quota request, a releasing request is aborted.
 341                  * That ongoing quota request will call this function again when
 342                  * it returned b=18630 */
 343                 if (pending_write || record) {
 344                         CDEBUG(D_QUOTA, "delay quota release\n");
 345                         ret = 0;
 346                 }
 347         }
 348         if (ret > 0)
 349                 quota_compute_lqs(qdata, lqs, 1, (ret == 1) ? 1 : 0);
 350
 351         CDEBUG(D_QUOTA, "type: %c, limit: "LPU64", usage: "LPU64
 352                ", pending_write: "LPU64", record: %lld"
 353                ", qunit_sz: %lu, tune_sz: %lu, ret: %d.\n",
 354                QDATA_IS_BLK(qdata) ? 'b' : 'i', limit, usage, pending_write,
 355                record, qunit_sz, tune_sz, ret);
 356         LASSERT(ret == 0 || qdata->qd_count);
 357
 358         spin_unlock(&lqs->lqs_lock);
 359         lqs_putref(lqs);
 360
 361         EXIT;
 362  out:
 363         OBD_FREE_PTR(qctl);
 364         return ret;
 365 }
 366
 367 /**
 368  * Compute the remaining quota for certain gid or uid b=11693
 369  */
 370 int compute_remquota(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 371                      struct qunit_data *qdata, int isblk)
 372 {
 373         struct super_block *sb = qctxt->lqc_sb;
 374         __u64 usage, limit;
 375         struct obd_quotactl *qctl;
 376         int ret = QUOTA_RET_OK;
 377         ENTRY;
 378
 379         if (!ll_sb_any_quota_active(sb))
 380                 RETURN(QUOTA_RET_NOQUOTA);
 381
 382         /* ignore root user */
 383         if (qdata->qd_id == 0 && QDATA_IS_GRP(qdata) == USRQUOTA)
 384                 RETURN(QUOTA_RET_NOLIMIT);
 385
 386         OBD_ALLOC_PTR(qctl);
 387         if (qctl == NULL)
 388                 RETURN(-ENOMEM);
 389
 390         /* get fs quota usage & limit */
 391         qctl->qc_cmd = Q_GETQUOTA;
 392         qctl->qc_id = qdata->qd_id;
 393         qctl->qc_type = QDATA_IS_GRP(qdata);
 394         ret = fsfilt_quotactl(obd, sb, qctl);
 395         if (ret) {
 396                 if (ret == -ESRCH)      /* no limit */
 397                         ret = QUOTA_RET_NOLIMIT;
 398                 else
 399                         CDEBUG(D_QUOTA, "can't get fs quota usage! (rc:%d)",
 400                                ret);
 401                 GOTO(out, ret);
 402         }
 403
 404         usage = isblk ? qctl->qc_dqblk.dqb_curspace :
 405                 qctl->qc_dqblk.dqb_curinodes;
 406         limit = isblk ? qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS :
 407                 qctl->qc_dqblk.dqb_ihardlimit;
 408         if (!limit){            /* no limit */
 409                 ret = QUOTA_RET_NOLIMIT;
 410                 GOTO(out, ret);
 411         }
 412
 413         if (limit >= usage)
 414                 qdata->qd_count = limit - usage;
 415         else
 416                 qdata->qd_count = 0;
 417         EXIT;
 418 out:
 419         OBD_FREE_PTR(qctl);
 420         return ret;
 421 }
 422
 423 static struct lustre_qunit *alloc_qunit(struct lustre_quota_ctxt *qctxt,
 424                                         struct qunit_data *qdata, int opc)
 425 {
 426         struct lustre_qunit *qunit = NULL;
 427         ENTRY;
 428
 429         OBD_SLAB_ALLOC_PTR_GFP(qunit, qunit_cachep, CFS_ALLOC_IO);
 430         if (qunit == NULL)
 431                 RETURN(NULL);
 432
 433         CFS_INIT_LIST_HEAD(&qunit->lq_hash);
 434         init_waitqueue_head(&qunit->lq_waitq);
 435         atomic_set(&qunit->lq_refcnt, 1);
 436         qunit->lq_ctxt = qctxt;
 437         memcpy(&qunit->lq_data, qdata, sizeof(*qdata));
 438         qunit->lq_opc = opc;
 439         qunit->lq_lock = SPIN_LOCK_UNLOCKED;
 440         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_CREATED, 0);
 441         qunit->lq_owner = cfs_curproc_pid();
 442         RETURN(qunit);
 443 }
 444
 445 static inline void free_qunit(struct lustre_qunit *qunit)
 446 {
 447         OBD_SLAB_FREE(qunit, qunit_cachep, sizeof(*qunit));
 448 }
 449
 450 static inline void qunit_get(struct lustre_qunit *qunit)
 451 {
 452         atomic_inc(&qunit->lq_refcnt);
 453 }
 454
 455 static void qunit_put(struct lustre_qunit *qunit)
 456 {
 457         LASSERT(atomic_read(&qunit->lq_refcnt));
 458         if (atomic_dec_and_test(&qunit->lq_refcnt))
 459                 free_qunit(qunit);
 460 }
 461
 462 /* caller must hold qunit_hash_lock and release ref of qunit after using it */
 463 static struct lustre_qunit *dqacq_in_flight(struct lustre_quota_ctxt *qctxt,
 464                                             struct qunit_data *qdata)
 465 {
 466         unsigned int hashent = qunit_hashfn(qctxt, qdata);
 467         struct lustre_qunit *qunit;
 468         ENTRY;
 469
 470         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
 471         qunit = find_qunit(hashent, qctxt, qdata);
 472         if (qunit)
 473                 qunit_get(qunit);
 474         RETURN(qunit);
 475 }
 476
 477 static void
 478 insert_qunit_nolock(struct lustre_quota_ctxt *qctxt, struct lustre_qunit *qunit)
 479 {
 480         struct list_head *head;
 481
 482         LASSERT(list_empty(&qunit->lq_hash));
 483         qunit_get(qunit);
 484         head = qunit_hash + qunit_hashfn(qctxt, &qunit->lq_data);
 485         list_add(&qunit->lq_hash, head);
 486         QUNIT_SET_STATE(qunit, QUNIT_IN_HASH);
 487 }
 488
 489 static void compute_lqs_after_removing_qunit(struct lustre_qunit *qunit)
 490 {
 491         struct lustre_qunit_size *lqs;
 492
 493         lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(&qunit->lq_data),
 494                                        qunit->lq_data.qd_id),
 495                                qunit->lq_ctxt, 0);
 496         if (lqs && !IS_ERR(lqs)) {
 497                 spin_lock(&lqs->lqs_lock);
 498                 if (qunit->lq_opc == QUOTA_DQACQ)
 499                         quota_compute_lqs(&qunit->lq_data, lqs, 0, 1);
 500                 if (qunit->lq_opc == QUOTA_DQREL)
 501                         quota_compute_lqs(&qunit->lq_data, lqs, 0, 0);
 502                 spin_unlock(&lqs->lqs_lock);
 503                 /* this is for quota_search_lqs */
 504                 lqs_putref(lqs);
 505                 /* this is for schedule_dqacq */
 506                 lqs_putref(lqs);
 507         }
 508 }
 509
 510 static void remove_qunit_nolock(struct lustre_qunit *qunit)
 511 {
 512         LASSERT(!list_empty(&qunit->lq_hash));
 513         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
 514
 515         list_del_init(&qunit->lq_hash);
 516         QUNIT_SET_STATE(qunit, QUNIT_RM_FROM_HASH);
 517         qunit_put(qunit);
 518 }
 519
 520 void* quota_barrier(struct lustre_quota_ctxt *qctxt,
 521                     struct obd_quotactl *oqctl, int isblk)
 522 {
 523         struct lustre_qunit *qunit, *find_qunit;
 524         int cycle = 1;
 525
 526         OBD_SLAB_ALLOC(qunit, qunit_cachep, CFS_ALLOC_IO, sizeof(*qunit));
 527         if (qunit == NULL) {
 528                 CERROR("locating %sunit failed for %sid %u\n",
 529                        isblk ? "b" : "i", oqctl->qc_type ? "g" : "u",
 530                        oqctl->qc_id);
 531                 qctxt_wait_pending_dqacq(qctxt, oqctl->qc_id,
 532                                          oqctl->qc_type, isblk);
 533                 return NULL;
 534         }
 535
 536         INIT_LIST_HEAD(&qunit->lq_hash);
 537         qunit->lq_lock = SPIN_LOCK_UNLOCKED;
 538         init_waitqueue_head(&qunit->lq_waitq);
 539         atomic_set(&qunit->lq_refcnt, 1);
 540         qunit->lq_ctxt = qctxt;
 541         qunit->lq_data.qd_id = oqctl->qc_id;
 542         qunit->lq_data.qd_flags =  oqctl->qc_type;
 543         if (isblk)
 544                 QDATA_SET_BLK(&qunit->lq_data);
 545         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_CREATED, 0);
 546         /* it means it is only an invalid qunit for barrier */
 547         qunit->lq_opc = QUOTA_LAST_OPC;
 548
 549         while (1) {
 550                 spin_lock(&qunit_hash_lock);
 551                 find_qunit = dqacq_in_flight(qctxt, &qunit->lq_data);
 552                 if (find_qunit) {
 553                         spin_unlock(&qunit_hash_lock);
 554                         qunit_put(find_qunit);
 555                         qctxt_wait_pending_dqacq(qctxt, oqctl->qc_id,
 556                                                  oqctl->qc_type, isblk);
 557                         CDEBUG(D_QUOTA, "cycle=%d\n", cycle++);
 558                         continue;
 559                 }
 560                 break;
 561         }
 562         insert_qunit_nolock(qctxt, qunit);
 563         spin_unlock(&qunit_hash_lock);
 564         return qunit;
 565 }
 566
 567 void quota_unbarrier(void *handle)
 568 {
 569         struct lustre_qunit *qunit = (struct lustre_qunit *)handle;
 570
 571         if (qunit == NULL) {
 572                 CERROR("handle is NULL\n");
 573                 return;
 574         }
 575
 576         LASSERT(qunit->lq_opc == QUOTA_LAST_OPC);
 577         spin_lock(&qunit_hash_lock);
 578         remove_qunit_nolock(qunit);
 579         spin_unlock(&qunit_hash_lock);
 580         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, QUOTA_REQ_RETURNED);
 581         wake_up(&qunit->lq_waitq);
 582         qunit_put(qunit);
 583 }
 584
 585 #define INC_QLIMIT(limit, count) (limit == MIN_QLIMIT) ? \
 586                                  (limit = count) : (limit += count)
 587
 588
 589 static inline int is_master(struct lustre_quota_ctxt *qctxt)
 590 {
 591         return qctxt->lqc_handler ? 1 : 0;
 592 }
 593
 594 static int
 595 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 596                struct qunit_data *qdata, int opc, int wait,
 597                struct obd_trans_info *oti);
 598
 599 static inline void qdata_to_oqaq(struct qunit_data *qdata,
 600                                  struct quota_adjust_qunit *oqaq)
 601 {
 602         LASSERT(qdata);
 603         LASSERT(oqaq);
 604
 605         oqaq->qaq_flags = qdata->qd_flags;
 606         oqaq->qaq_id    = qdata->qd_id;
 607         if (QDATA_IS_ADJBLK(qdata))
 608                 oqaq->qaq_bunit_sz = qdata->qd_qunit;
 609         if (QDATA_IS_ADJINO(qdata))
 610                 oqaq->qaq_iunit_sz = qdata->qd_qunit;
 611 }
 612
 613 static int
 614 dqacq_completion(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 615                  struct qunit_data *qdata, int rc, int opc)
 616 {
 617         struct lustre_qunit *qunit = NULL;
 618         struct super_block *sb = qctxt->lqc_sb;
 619         int err = 0;
 620         struct quota_adjust_qunit *oqaq = NULL;
 621         int rc1 = 0;
 622         ENTRY;
 623
 624         LASSERT(qdata);
 625         QDATA_DEBUG(qdata, "obd(%s): complete %s quota req\n",
 626                     obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
 627
 628         /* do it only when a releasing quota req more than 5MB b=18491 */
 629         if (opc == QUOTA_DQREL && qdata->qd_count >= 5242880)
 630                 OBD_FAIL_TIMEOUT(OBD_FAIL_QUOTA_DELAY_REL, 5);
 631
 632         /* update local operational quota file */
 633         if (rc == 0) {
 634                 __u64 count = QUSG(qdata->qd_count, QDATA_IS_BLK(qdata));
 635                 struct obd_quotactl *qctl;
 636                 __u64 *hardlimit;
 637
 638                 OBD_ALLOC_PTR(qctl);
 639                 if (qctl == NULL)
 640                         GOTO(out, err = -ENOMEM);
 641
 642                 /* acq/rel qunit for specified uid/gid is serialized,
 643                  * so there is no race between get fs quota limit and
 644                  * set fs quota limit */
 645                 qctl->qc_cmd = Q_GETQUOTA;
 646                 qctl->qc_id = qdata->qd_id;
 647                 qctl->qc_type = QDATA_IS_GRP(qdata);
 648                 err = fsfilt_quotactl(obd, sb, qctl);
 649                 if (err) {
 650                         CERROR("error get quota fs limit! (rc:%d)\n", err);
 651                         GOTO(out_mem, err);
 652                 }
 653
 654                 if (QDATA_IS_BLK(qdata)) {
 655                         qctl->qc_dqblk.dqb_valid = QIF_BLIMITS;
 656                         hardlimit = &qctl->qc_dqblk.dqb_bhardlimit;
 657                 } else {
 658                         qctl->qc_dqblk.dqb_valid = QIF_ILIMITS;
 659                         hardlimit = &qctl->qc_dqblk.dqb_ihardlimit;
 660                 }
 661
 662                 CDEBUG(D_QUOTA, "hardlimt: "LPU64"\n", *hardlimit);
 663
 664                 if (*hardlimit == 0)
 665                         goto out_mem;
 666
 667                 switch (opc) {
 668                 case QUOTA_DQACQ:
 669                         INC_QLIMIT(*hardlimit, count);
 670                         break;
 671                 case QUOTA_DQREL:
 672                         LASSERTF(count < *hardlimit,
 673                                  "id(%u) flag(%u) type(%c) isblk(%c) "
 674                                  "count("LPU64") qd_qunit("LPU64") "
 675                                  "hardlimit("LPU64").\n",
 676                                  qdata->qd_id, qdata->qd_flags,
 677                                  QDATA_IS_GRP(qdata) ? 'g' : 'u',
 678                                  QDATA_IS_BLK(qdata) ? 'b': 'i',
 679                                  qdata->qd_count, qdata->qd_qunit, *hardlimit);
 680                         *hardlimit -= count;
 681                         break;
 682                 default:
 683                         LBUG();
 684                 }
 685
 686                 /* clear quota limit */
 687                 if (count == 0)
 688                         *hardlimit = 0;
 689
 690                 qctl->qc_cmd = Q_SETQUOTA;
 691                 err = fsfilt_quotactl(obd, sb, qctl);
 692                 if (err)
 693                         CERROR("error set quota fs limit! (rc:%d)\n", err);
 694
 695                 QDATA_DEBUG(qdata, "%s completion\n",
 696                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 697 out_mem:
 698                 OBD_FREE_PTR(qctl);
 699         } else if (rc == -EDQUOT) {
 700                 QDATA_DEBUG(qdata, "acquire qunit got EDQUOT.\n");
 701         } else if (rc == -EBUSY) {
 702                 QDATA_DEBUG(qdata, "it's is recovering, got EBUSY.\n");
 703         } else {
 704                 CERROR("acquire qunit got error! (rc:%d)\n", rc);
 705         }
 706 out:
 707         /* remove the qunit from hash */
 708         spin_lock(&qunit_hash_lock);
 709
 710         qunit = dqacq_in_flight(qctxt, qdata);
 711         /* this qunit has been removed by qctxt_cleanup() */
 712         if (!qunit) {
 713                 spin_unlock(&qunit_hash_lock);
 714                 QDATA_DEBUG(qdata, "%s is discarded because qunit isn't found\n",
 715                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 716                 RETURN(err);
 717         }
 718
 719         LASSERT(opc == qunit->lq_opc);
 720         /* remove this qunit from lq_hash so that new processes cannot be added
 721          * to qunit->lq_waiters */
 722         remove_qunit_nolock(qunit);
 723         spin_unlock(&qunit_hash_lock);
 724
 725         compute_lqs_after_removing_qunit(qunit);
 726
 727         if (rc == 0)
 728                 rc = QUOTA_REQ_RETURNED;
 729         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, rc);
 730         /* wake up all waiters */
 731         wake_up_all(&qunit->lq_waitq);
 732
 733         /* this is for dqacq_in_flight() */
 734         qunit_put(qunit);
 735         if (rc < 0 && rc != -EDQUOT)
 736                 GOTO(out1, err);
 737
 738         /* don't reschedule in such cases:
 739          *   - acq/rel failure and qunit isn't changed,
 740          *     but not for quota recovery.
 741          *   - local dqacq/dqrel.
 742          *   - local disk io failure.
 743          */
 744          OBD_ALLOC_PTR(oqaq);
 745          if (!oqaq)
 746                  GOTO(out1, err = -ENOMEM);
 747          qdata_to_oqaq(qdata, oqaq);
 748          /* adjust the qunit size in slaves */
 749          rc1 = quota_adjust_slave_lqs(oqaq, qctxt);
 750          OBD_FREE_PTR(oqaq);
 751          if (rc1 < 0) {
 752                  CERROR("adjust slave's qunit size failed!(rc:%d)\n", rc1);
 753                  GOTO(out1, err = rc1);
 754          }
 755          if (err || (rc < 0 && rc != -EBUSY && rc1 == 0) || is_master(qctxt))
 756                  GOTO(out1, err);
 757
 758          if (opc == QUOTA_DQREL && qdata->qd_count >= 5242880 &&
 759              OBD_FAIL_CHECK(OBD_FAIL_QUOTA_DELAY_REL))
 760                  GOTO(out1, err);
 761
 762         /* reschedule another dqacq/dqrel if needed */
 763         qdata->qd_count = 0;
 764         qdata->qd_flags &= LQUOTA_QUNIT_FLAGS;
 765         rc1 = check_cur_qunit(obd, qctxt, qdata);
 766         if (rc1 > 0) {
 767                 int opc;
 768                 opc = rc1 == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
 769                 rc1 = schedule_dqacq(obd, qctxt, qdata, opc, 0, NULL);
 770                 QDATA_DEBUG(qdata, "reschedudle opc(%d) rc(%d)\n", opc, rc1);
 771         }
 772  out1:
 773         /* this is for alloc_qunit() */
 774         qunit_put(qunit);
 775         RETURN(err);
 776 }
 777
 778 struct dqacq_async_args {
 779         struct lustre_quota_ctxt *aa_ctxt;
 780         struct lustre_qunit *aa_qunit;
 781 };
 782
 783 static int dqacq_interpret(const struct lu_env *env,
 784                            struct ptlrpc_request *req, void *data, int rc)
 785 {
 786         struct dqacq_async_args *aa = (struct dqacq_async_args *)data;
 787         struct lustre_quota_ctxt *qctxt = aa->aa_ctxt;
 788         struct obd_device_target *obt = qctxt->lqc_obt;
 789         struct lustre_qunit *qunit = aa->aa_qunit;
 790         struct obd_device *obd = req->rq_import->imp_obd;
 791         struct qunit_data *qdata = NULL;
 792         ENTRY;
 793
 794         LASSERT(req);
 795         LASSERT(req->rq_import);
 796
 797         down_read(&obt->obt_rwsem);
 798         /* if a quota req timeouts or is dropped, we should update quota
 799          * statistics which will be handled in dqacq_completion. And in
 800          * this situation we should get qdata from request instead of
 801          * reply */
 802         qdata = quota_get_qdata(req, (rc != 0) ? QUOTA_REQUEST : QUOTA_REPLY,
 803                                 QUOTA_IMPORT);
 804         if (IS_ERR(qdata)) {
 805                 rc = PTR_ERR(qdata);
 806                 DEBUG_REQ(D_ERROR, req,
 807                           "error unpacking qunit_data(rc: %ld)\n",
 808                           PTR_ERR(qdata));
 809                 qdata = &qunit->lq_data;
 810         }
 811
 812         QDATA_DEBUG(qdata, "qdata: interpret rc(%d).\n", rc);
 813         QDATA_DEBUG((&qunit->lq_data), "lq_data: \n");
 814
 815         if (qdata->qd_id != qunit->lq_data.qd_id ||
 816             OBD_FAIL_CHECK(OBD_FAIL_QUOTA_RET_QDATA)) {
 817                 CERROR("the returned qd_id isn't expected!"
 818                        "(qdata: %u, lq_data: %u)\n", qdata->qd_id,
 819                        qunit->lq_data.qd_id);
 820                 qdata->qd_id = qunit->lq_data.qd_id;
 821                 rc = -EPROTO;
 822         }
 823         if (QDATA_IS_GRP(qdata) != QDATA_IS_GRP(&qunit->lq_data)) {
 824                 CERROR("the returned grp/usr isn't expected!"
 825                        "(qdata: %u, lq_data: %u)\n", qdata->qd_flags,
 826                        qunit->lq_data.qd_flags);
 827                 if (QDATA_IS_GRP(&qunit->lq_data))
 828                         QDATA_SET_GRP(qdata);
 829                 else
 830                         QDATA_CLR_GRP(qdata);
 831                 rc = -EPROTO;
 832         }
 833         if (qdata->qd_count > qunit->lq_data.qd_count) {
 834                 CERROR("the returned qd_count isn't expected!"
 835                        "(qdata: "LPU64", lq_data: "LPU64")\n", qdata->qd_count,
 836                        qunit->lq_data.qd_count);
 837                 rc = -EPROTO;
 838         }
 839
 840         if (unlikely(rc == -ESRCH))
 841                 CERROR("quota for %s has been enabled by master, but disabled "
 842                        "by slave.\n", QDATA_IS_GRP(qdata) ? "group" : "user");
 843
 844         rc = dqacq_completion(obd, qctxt, qdata, rc,
 845                               lustre_msg_get_opc(req->rq_reqmsg));
 846
 847         up_read(&obt->obt_rwsem);
 848         RETURN(rc);
 849 }
 850
 851 /**
 852  * check if quota master is online
 853  */
 854 int check_qm(struct lustre_quota_ctxt *qctxt)
 855 {
 856         int rc;
 857         ENTRY;
 858
 859         spin_lock(&qctxt->lqc_lock);
 860         /* quit waiting when mds is back or qctxt is cleaned up */
 861         rc = qctxt->lqc_import || !qctxt->lqc_valid;
 862         spin_unlock(&qctxt->lqc_lock);
 863
 864         RETURN(rc);
 865 }
 866
 867 /* wake up all waiting threads when lqc_import is NULL */
 868 void dqacq_interrupt(struct lustre_quota_ctxt *qctxt)
 869 {
 870         struct lustre_qunit *qunit, *tmp;
 871         int i;
 872         ENTRY;
 873
 874         spin_lock(&qunit_hash_lock);
 875         for (i = 0; i < NR_DQHASH; i++) {
 876                 list_for_each_entry_safe(qunit, tmp, &qunit_hash[i], lq_hash) {
 877                         if (qunit->lq_ctxt != qctxt)
 878                                 continue;
 879
 880                         /* Wake up all waiters. Do not change lq_state.
 881                          * The waiters will check lq_rc which is kept as 0
 882                          * if no others change it, then the waiters will return
 883                          * -EAGAIN to caller who can perform related quota
 884                          * acq/rel if necessary. */
 885                         wake_up_all(&qunit->lq_waitq);
 886                 }
 887         }
 888         spin_unlock(&qunit_hash_lock);
 889         EXIT;
 890 }
 891
 892 static int got_qunit(struct lustre_qunit *qunit, int is_master)
 893 {
 894         struct lustre_quota_ctxt *qctxt = qunit->lq_ctxt;
 895         int rc = 0;
 896         ENTRY;
 897
 898         spin_lock(&qunit->lq_lock);
 899         switch (qunit->lq_state) {
 900         case QUNIT_IN_HASH:
 901         case QUNIT_RM_FROM_HASH:
 902                 break;
 903         case QUNIT_FINISHED:
 904                 rc = 1;
 905                 break;
 906         default:
 907                 CERROR("invalid qunit state %d\n", qunit->lq_state);
 908         }
 909         spin_unlock(&qunit->lq_lock);
 910
 911         if (!rc) {
 912                 spin_lock(&qctxt->lqc_lock);
 913                 rc = !qctxt->lqc_valid;
 914                 if (!is_master)
 915                         rc |= !qctxt->lqc_import;
 916                 spin_unlock(&qctxt->lqc_lock);
 917         }
 918
 919         RETURN(rc);
 920 }
 921
 922 static inline void
 923 revoke_lqs_rec(struct lustre_qunit_size *lqs, struct qunit_data *qdata, int opc)
 924 {
 925         /* revoke lqs_xxx_rec which is computed in check_cur_qunit
 926          * b=18630 */
 927         spin_lock(&lqs->lqs_lock);
 928         quota_compute_lqs(qdata, lqs, 0, (opc == QUOTA_DQACQ) ? 1 : 0);
 929         spin_unlock(&lqs->lqs_lock);
 930 }
 931
 932 static int
 933 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 934                struct qunit_data *qdata, int opc, int wait,
 935                struct obd_trans_info *oti)
 936 {
 937         struct lustre_qunit *qunit, *empty;
 938         struct l_wait_info lwi = { 0 };
 939         struct ptlrpc_request *req;
 940         struct dqacq_async_args *aa;
 941         struct obd_import *imp = NULL;
 942         struct lustre_qunit_size *lqs = NULL;
 943         struct timeval work_start;
 944         struct timeval work_end;
 945         long timediff;
 946         int rc = 0;
 947         ENTRY;
 948
 949         LASSERT(opc == QUOTA_DQACQ || opc == QUOTA_DQREL);
 950         do_gettimeofday(&work_start);
 951
 952         lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(qdata), qdata->qd_id),
 953                                qctxt, 0);
 954         if (lqs == NULL || IS_ERR(lqs)) {
 955                 CERROR("Can't find the lustre qunit size!\n");
 956                 RETURN(-EPERM);
 957         }
 958
 959         if ((empty = alloc_qunit(qctxt, qdata, opc)) == NULL) {
 960                 revoke_lqs_rec(lqs, qdata, opc);
 961                 /* this is for quota_search_lqs */
 962                 lqs_putref(lqs);
 963                 RETURN(-ENOMEM);
 964         }
 965
 966         OBD_FAIL_TIMEOUT(OBD_FAIL_QUOTA_DELAY_SD, 5);
 967
 968         spin_lock(&qunit_hash_lock);
 969         qunit = dqacq_in_flight(qctxt, qdata);
 970         if (qunit) {
 971                 spin_unlock(&qunit_hash_lock);
 972                 qunit_put(empty);
 973
 974                 revoke_lqs_rec(lqs, qdata, opc);
 975                 /* this is for quota_search_lqs */
 976                 lqs_putref(lqs);
 977                 goto wait_completion;
 978         }
 979         qunit = empty;
 980         qunit_get(qunit);
 981         insert_qunit_nolock(qctxt, qunit);
 982         spin_unlock(&qunit_hash_lock);
 983
 984         /* From here, the quota request will be sent anyway.
 985          * When this qdata request returned or is cancelled,
 986          * lqs_putref will be called at that time */
 987         lqs_getref(lqs);
 988         /* this is for quota_search_lqs */
 989         lqs_putref(lqs);
 990
 991         QDATA_DEBUG(qdata, "obd(%s): send %s quota req\n",
 992                     obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
 993         /* master is going to dqacq/dqrel from itself */
 994         if (is_master(qctxt)) {
 995                 int rc2;
 996                 QDATA_DEBUG(qdata, "local %s.\n",
 997                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 998                 QDATA_SET_CHANGE_QS(qdata);
 999                 rc = qctxt->lqc_handler(obd, qdata, opc);
1000                 rc2 = dqacq_completion(obd, qctxt, qdata, rc, opc);
1001                 /* this is for qunit_get() */
1002                 qunit_put(qunit);
1003
1004                 do_gettimeofday(&work_end);
1005                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1006                 if (opc == QUOTA_DQACQ)
1007                         lprocfs_counter_add(qctxt->lqc_stats,
1008                                             wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
1009                                             timediff);
1010                 else
1011                         lprocfs_counter_add(qctxt->lqc_stats,
1012                                             wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
1013                                             timediff);
1014                 RETURN(rc ? rc : rc2);
1015         }
1016
1017         spin_lock(&qctxt->lqc_lock);
1018         if (!qctxt->lqc_import) {
1019                 spin_unlock(&qctxt->lqc_lock);
1020                 QDATA_DEBUG(qdata, "lqc_import is invalid.\n");
1021
1022                 spin_lock(&qunit_hash_lock);
1023                 remove_qunit_nolock(qunit);
1024                 spin_unlock(&qunit_hash_lock);
1025
1026                 compute_lqs_after_removing_qunit(qunit);
1027
1028                 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, -EAGAIN);
1029                 wake_up_all(&qunit->lq_waitq);
1030
1031                 /* this is for qunit_get() */
1032                 qunit_put(qunit);
1033                 /* this for alloc_qunit() */
1034                 qunit_put(qunit);
1035                 spin_lock(&qctxt->lqc_lock);
1036                 if (wait && !qctxt->lqc_import) {
1037                         spin_unlock(&qctxt->lqc_lock);
1038
1039                         LASSERT(oti && oti->oti_thread &&
1040                                 oti->oti_thread->t_watchdog);
1041
1042                         lc_watchdog_disable(oti->oti_thread->t_watchdog);
1043                         CDEBUG(D_QUOTA, "sleep for quota master\n");
1044                         l_wait_event(qctxt->lqc_wait_for_qmaster,
1045                                      check_qm(qctxt), &lwi);
1046                         CDEBUG(D_QUOTA, "wake up when quota master is back\n");
1047                         lc_watchdog_touch(oti->oti_thread->t_watchdog,
1048                                  GET_TIMEOUT(oti->oti_thread->t_svc));
1049                 } else {
1050                         spin_unlock(&qctxt->lqc_lock);
1051                 }
1052
1053                 RETURN(-EAGAIN);
1054         }
1055         imp = class_import_get(qctxt->lqc_import);
1056         spin_unlock(&qctxt->lqc_lock);
1057
1058         /* build dqacq/dqrel request */
1059         LASSERT(imp);
1060
1061         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_QUOTA_DQACQ,
1062                                         LUSTRE_MDS_VERSION, opc);
1063         class_import_put(imp);
1064         if (req == NULL) {
1065                 CERROR("Can't alloc request\n");
1066                 dqacq_completion(obd, qctxt, qdata, -ENOMEM, opc);
1067                 /* this is for qunit_get() */
1068                 qunit_put(qunit);
1069                 RETURN(-ENOMEM);
1070         }
1071
1072         ptlrpc_request_set_replen(req);
1073         req->rq_no_resend = req->rq_no_delay = 1;
1074         rc = quota_copy_qdata(req, qdata, QUOTA_REQUEST, QUOTA_IMPORT);
1075         if (rc < 0) {
1076                 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
1077                 ptlrpc_req_finished(req);
1078                 dqacq_completion(obd, qctxt, qdata, -EPROTO, opc);
1079                 /* this is for qunit_get() */
1080                 qunit_put(qunit);
1081                 RETURN(rc);
1082         }
1083
1084         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1085         aa = ptlrpc_req_async_args(req);
1086         aa->aa_ctxt = qctxt;
1087         aa->aa_qunit = qunit;
1088
1089         req->rq_interpret_reply = dqacq_interpret;
1090         ptlrpcd_add_req(req, PSCOPE_OTHER);
1091
1092         QDATA_DEBUG(qdata, "%s scheduled.\n",
1093                     opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
1094 wait_completion:
1095         if (wait && qunit) {
1096                 struct qunit_data *p = &qunit->lq_data;
1097
1098                 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
1099                 l_wait_event(qunit->lq_waitq, got_qunit(qunit, is_master(qctxt)),
1100                              &lwi);
1101                 /* rc = -EAGAIN, it means the quota master isn't ready yet
1102                  * rc = QUOTA_REQ_RETURNED, it means a quota req is finished;
1103                  * rc = -EDQUOT, it means out of quota
1104                  * rc = -EBUSY, it means recovery is happening
1105                  * other rc < 0, it means real errors, functions who call
1106                  * schedule_dqacq should take care of this */
1107                 spin_lock(&qunit->lq_lock);
1108                 rc = qunit->lq_rc;
1109                 spin_unlock(&qunit->lq_lock);
1110                 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting: id(%u) flag(%u) "
1111                        "rc(%d) owner(%d)\n", qunit, qunit->lq_data.qd_id,
1112                        qunit->lq_data.qd_flags, rc, qunit->lq_owner);
1113         }
1114
1115         qunit_put(qunit);
1116         do_gettimeofday(&work_end);
1117         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1118         if (opc == QUOTA_DQACQ)
1119                 lprocfs_counter_add(qctxt->lqc_stats,
1120                                     wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
1121                                     timediff);
1122         else
1123                 lprocfs_counter_add(qctxt->lqc_stats,
1124                                     wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
1125                                     timediff);
1126
1127         RETURN(rc);
1128 }
1129
1130 int
1131 qctxt_adjust_qunit(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
1132                    const unsigned int id[], __u32 isblk, int wait,
1133                    struct obd_trans_info *oti)
1134 {
1135         int rc = 0, i = USRQUOTA;
1136         struct qunit_data qdata[MAXQUOTAS];
1137         ENTRY;
1138
1139         if (quota_is_set(obd, id, isblk ? QB_SET : QI_SET) == 0)
1140                 RETURN(0);
1141
1142         for (i = 0; i < MAXQUOTAS; i++) {
1143                 qdata[i].qd_id = id[i];
1144                 qdata[i].qd_flags = i;
1145                 if (isblk)
1146                         QDATA_SET_BLK(&qdata[i]);
1147                 qdata[i].qd_count = 0;
1148
1149                 rc = check_cur_qunit(obd, qctxt, &qdata[i]);
1150                 if (rc > 0) {
1151                         int opc;
1152                         /* need acquire or release */
1153                         opc = rc == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1154                         rc = schedule_dqacq(obd, qctxt, &qdata[i], opc,
1155                                             wait,oti);
1156                         if (rc < 0)
1157                                 RETURN(rc);
1158                 } else if (wait == 1) {
1159                         /* when wait equates 1, that means mds_quota_acquire
1160                          * or filter_quota_acquire is calling it. */
1161                         rc = qctxt_wait_pending_dqacq(qctxt, id[i], i, isblk);
1162                         if (rc < 0)
1163                                 RETURN(rc);
1164                 }
1165         }
1166
1167         RETURN(rc);
1168 }
1169
1170 int
1171 qctxt_wait_pending_dqacq(struct lustre_quota_ctxt *qctxt, unsigned int id,
1172                          unsigned short type, int isblk)
1173 {
1174         struct lustre_qunit *qunit = NULL;
1175         struct qunit_data qdata;
1176         struct timeval work_start;
1177         struct timeval work_end;
1178         long timediff;
1179         struct l_wait_info lwi = { 0 };
1180         int rc = 0;
1181         ENTRY;
1182
1183         do_gettimeofday(&work_start);
1184         qdata.qd_id = id;
1185         qdata.qd_flags = type;
1186         if (isblk)
1187                 QDATA_SET_BLK(&qdata);
1188         qdata.qd_count = 0;
1189
1190         spin_lock(&qunit_hash_lock);
1191         qunit = dqacq_in_flight(qctxt, &qdata);
1192         spin_unlock(&qunit_hash_lock);
1193
1194         if (qunit) {
1195                 struct qunit_data *p = &qunit->lq_data;
1196
1197                 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
1198                 l_wait_event(qunit->lq_waitq, got_qunit(qunit, is_master(qctxt)),
1199                              &lwi);
1200                 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting: rc(%d) "
1201                        "owner(%d)\n", qunit, qunit->lq_rc, qunit->lq_owner);
1202                 /* keep same as schedule_dqacq() b=17030 */
1203                 spin_lock(&qunit->lq_lock);
1204                 rc = qunit->lq_rc;
1205                 spin_unlock(&qunit->lq_lock);
1206                 /* this is for dqacq_in_flight() */
1207                 qunit_put(qunit);
1208                 do_gettimeofday(&work_end);
1209                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1210                 lprocfs_counter_add(qctxt->lqc_stats,
1211                                     isblk ? LQUOTA_WAIT_PENDING_BLK_QUOTA :
1212                                             LQUOTA_WAIT_PENDING_INO_QUOTA,
1213                                     timediff);
1214         } else {
1215                 do_gettimeofday(&work_end);
1216                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1217                 lprocfs_counter_add(qctxt->lqc_stats,
1218                                     isblk ? LQUOTA_NOWAIT_PENDING_BLK_QUOTA :
1219                                             LQUOTA_NOWAIT_PENDING_INO_QUOTA,
1220                                     timediff);
1221         }
1222
1223         RETURN(rc);
1224 }
1225
1226 int
1227 qctxt_init(struct obd_device *obd, dqacq_handler_t handler)
1228 {
1229         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
1230         struct obd_device_target *obt = &obd->u.obt;
1231         struct super_block *sb = obt->obt_sb;
1232         int rc = 0;
1233         ENTRY;
1234
1235         LASSERT(qctxt);
1236
1237         rc = ptlrpcd_addref();
1238         if (rc)
1239                 RETURN(rc);
1240
1241         cfs_waitq_init(&qctxt->lqc_wait_for_qmaster);
1242         cfs_waitq_init(&qctxt->lqc_lqs_waitq);
1243         atomic_set(&qctxt->lqc_lqs, 0);
1244         spin_lock_init(&qctxt->lqc_lock);
1245         spin_lock(&qctxt->lqc_lock);
1246         qctxt->lqc_handler = handler;
1247         qctxt->lqc_sb = sb;
1248         qctxt->lqc_obt = obt;
1249         qctxt->lqc_import = NULL;
1250         qctxt->lqc_recovery = 0;
1251         qctxt->lqc_switch_qs = 1; /* Change qunit size in default setting */
1252         qctxt->lqc_valid = 1;
1253         qctxt->lqc_cqs_boundary_factor = 4;
1254         qctxt->lqc_cqs_least_bunit = PTLRPC_MAX_BRW_SIZE;
1255         qctxt->lqc_cqs_least_iunit = 2;
1256         qctxt->lqc_cqs_qs_factor = 2;
1257         qctxt->lqc_flags = 0;
1258         QUOTA_MASTER_UNREADY(qctxt);
1259         qctxt->lqc_bunit_sz = default_bunit_sz;
1260         qctxt->lqc_btune_sz = default_bunit_sz / 100 * default_btune_ratio;
1261         qctxt->lqc_iunit_sz = default_iunit_sz;
1262         qctxt->lqc_itune_sz = default_iunit_sz * default_itune_ratio / 100;
1263         qctxt->lqc_switch_seconds = 300; /* enlarging will wait 5 minutes
1264                                           * after the last shrinking */
1265         qctxt->lqc_sync_blk = 0;
1266         spin_unlock(&qctxt->lqc_lock);
1267
1268         qctxt->lqc_lqs_hash = cfs_hash_create("LQS_HASH",
1269                                               HASH_LQS_CUR_BITS,
1270                                               HASH_LQS_MAX_BITS,
1271                                               &lqs_hash_ops, CFS_HASH_REHASH);
1272         if (!qctxt->lqc_lqs_hash) {
1273                 CERROR("initialize hash lqs for %s error!\n", obd->obd_name);
1274                 RETURN(-ENOMEM);
1275         }
1276
1277 #ifdef LPROCFS
1278         rc = lquota_proc_setup(obd, is_master(qctxt));
1279         if (rc)
1280                 CERROR("initialize proc for %s error!\n", obd->obd_name);
1281 #endif
1282
1283         RETURN(rc);
1284 }
1285
1286 static int check_lqs(struct lustre_quota_ctxt *qctxt)
1287 {
1288         int rc;
1289         ENTRY;
1290
1291         rc = !atomic_read(&qctxt->lqc_lqs);
1292
1293         RETURN(rc);
1294 }
1295
1296
1297 void hash_put_lqs(void *obj, void *data)
1298 {
1299         lqs_putref((struct lustre_qunit_size *)obj);
1300 }
1301
1302 void qctxt_cleanup(struct lustre_quota_ctxt *qctxt, int force)
1303 {
1304         struct lustre_qunit *qunit, *tmp;
1305         struct list_head tmp_list;
1306         struct l_wait_info lwi = { 0 };
1307         struct obd_device_target *obt = qctxt->lqc_obt;
1308         int i;
1309         ENTRY;
1310
1311         CFS_INIT_LIST_HEAD(&tmp_list);
1312
1313         spin_lock(&qctxt->lqc_lock);
1314         qctxt->lqc_valid = 0;
1315         spin_unlock(&qctxt->lqc_lock);
1316
1317         spin_lock(&qunit_hash_lock);
1318         for (i = 0; i < NR_DQHASH; i++) {
1319                 list_for_each_entry_safe(qunit, tmp, &qunit_hash[i], lq_hash) {
1320                         if (qunit->lq_ctxt != qctxt)
1321                                 continue;
1322                         remove_qunit_nolock(qunit);
1323                         list_add(&qunit->lq_hash, &tmp_list);
1324                 }
1325         }
1326         spin_unlock(&qunit_hash_lock);
1327
1328         list_for_each_entry_safe(qunit, tmp, &tmp_list, lq_hash) {
1329                 list_del_init(&qunit->lq_hash);
1330                 compute_lqs_after_removing_qunit(qunit);
1331
1332                 /* wake up all waiters */
1333                 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, 0);
1334                 wake_up_all(&qunit->lq_waitq);
1335                 qunit_put(qunit);
1336         }
1337
1338         /* after qctxt_cleanup, qctxt might be freed, then check_qm() is
1339          * unpredicted. So we must wait until lqc_wait_for_qmaster is empty */
1340         while (cfs_waitq_active(&qctxt->lqc_wait_for_qmaster)) {
1341                 cfs_waitq_signal(&qctxt->lqc_wait_for_qmaster);
1342                 cfs_schedule_timeout(CFS_TASK_INTERRUPTIBLE,
1343                                      cfs_time_seconds(1));
1344         }
1345
1346         cfs_hash_for_each_safe(qctxt->lqc_lqs_hash, hash_put_lqs, NULL);
1347         l_wait_event(qctxt->lqc_lqs_waitq, check_lqs(qctxt), &lwi);
1348         down_write(&obt->obt_rwsem);
1349         cfs_hash_destroy(qctxt->lqc_lqs_hash);
1350         qctxt->lqc_lqs_hash = NULL;
1351         up_write(&obt->obt_rwsem);
1352
1353         ptlrpcd_decref();
1354
1355 #ifdef LPROCFS
1356         if (lquota_proc_cleanup(qctxt))
1357                 CERROR("cleanup proc error!\n");
1358 #endif
1359
1360         EXIT;
1361 }
1362
1363 struct qslave_recov_thread_data {
1364         struct obd_device *obd;
1365         struct lustre_quota_ctxt *qctxt;
1366         struct completion comp;
1367 };
1368
1369 /* FIXME only recovery block quota by now */
1370 static int qslave_recovery_main(void *arg)
1371 {
1372         struct qslave_recov_thread_data *data = arg;
1373         struct obd_device *obd = data->obd;
1374         struct lustre_quota_ctxt *qctxt = data->qctxt;
1375         unsigned int type;
1376         int rc = 0;
1377         ENTRY;
1378
1379         cfs_daemonize_ctxt("qslave_recovd");
1380
1381         /* for obdfilter */
1382         class_incref(obd, "qslave_recovd_filter", obd);
1383
1384         complete(&data->comp);
1385
1386         spin_lock(&qctxt->lqc_lock);
1387         if (qctxt->lqc_recovery) {
1388                 spin_unlock(&qctxt->lqc_lock);
1389                 class_decref(obd, "qslave_recovd_filter", obd);
1390                 RETURN(0);
1391         } else {
1392                 qctxt->lqc_recovery = 1;
1393                 spin_unlock(&qctxt->lqc_lock);
1394         }
1395
1396         for (type = USRQUOTA; type < MAXQUOTAS; type++) {
1397                 struct qunit_data qdata;
1398                 struct quota_info *dqopt = sb_dqopt(qctxt->lqc_sb);
1399                 struct list_head id_list;
1400                 struct dquot_id *dqid, *tmp;
1401                 int ret;
1402
1403                 LOCK_DQONOFF_MUTEX(dqopt);
1404                 if (!ll_sb_has_quota_active(qctxt->lqc_sb, type)) {
1405                         UNLOCK_DQONOFF_MUTEX(dqopt);
1406                         break;
1407                 }
1408
1409                 LASSERT(dqopt->files[type] != NULL);
1410                 CFS_INIT_LIST_HEAD(&id_list);
1411 #ifndef KERNEL_SUPPORTS_QUOTA_READ
1412                 rc = fsfilt_qids(obd, dqopt->files[type], NULL, type, &id_list);
1413 #else
1414                 rc = fsfilt_qids(obd, NULL, dqopt->files[type], type, &id_list);
1415 #endif
1416                 UNLOCK_DQONOFF_MUTEX(dqopt);
1417                 if (rc)
1418                         CERROR("Get ids from quota file failed. (rc:%d)\n", rc);
1419
1420                 list_for_each_entry_safe(dqid, tmp, &id_list, di_link) {
1421                         list_del_init(&dqid->di_link);
1422                         /* skip slave recovery on itself */
1423                         if (is_master(qctxt))
1424                                 goto free;
1425                         if (rc && rc != -EBUSY)
1426                                 goto free;
1427
1428                         qdata.qd_id = dqid->di_id;
1429                         qdata.qd_flags = type;
1430                         QDATA_SET_BLK(&qdata);
1431                         qdata.qd_count = 0;
1432
1433                         ret = check_cur_qunit(obd, qctxt, &qdata);
1434                         if (ret > 0) {
1435                                 int opc;
1436                                 opc = ret == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1437                                 rc = schedule_dqacq(obd, qctxt, &qdata, opc,
1438                                                     0, NULL);
1439                                 if (rc == -EDQUOT)
1440                                         rc = 0;
1441                         } else {
1442                                 rc = 0;
1443                         }
1444
1445                         if (rc && rc != -EBUSY)
1446                                 CERROR("qslave recovery failed! (id:%d type:%d "
1447                                        " rc:%d)\n", dqid->di_id, type, rc);
1448 free:
1449                         OBD_FREE_PTR(dqid);
1450                 }
1451         }
1452
1453         spin_lock(&qctxt->lqc_lock);
1454         qctxt->lqc_recovery = 0;
1455         spin_unlock(&qctxt->lqc_lock);
1456         class_decref(obd, "qslave_recovd_filter", obd);
1457         RETURN(rc);
1458 }
1459
1460 void
1461 qslave_start_recovery(struct obd_device *obd, struct lustre_quota_ctxt *qctxt)
1462 {
1463         struct qslave_recov_thread_data data;
1464         int rc;
1465         ENTRY;
1466
1467         if (!ll_sb_any_quota_active(qctxt->lqc_sb))
1468                 goto exit;
1469
1470         data.obd = obd;
1471         data.qctxt = qctxt;
1472         init_completion(&data.comp);
1473
1474         rc = kernel_thread(qslave_recovery_main, &data, CLONE_VM|CLONE_FILES);
1475         if (rc < 0) {
1476                 CERROR("Cannot start quota recovery thread: rc %d\n", rc);
1477                 goto exit;
1478         }
1479         wait_for_completion(&data.comp);
1480 exit:
1481         EXIT;
1482 }
1483
1484 int quota_is_on(struct lustre_quota_ctxt *qctxt, struct obd_quotactl *oqctl)
1485 {
1486         unsigned int type;
1487
1488         for (type = USRQUOTA; type < MAXQUOTAS; type++) {
1489                 if (!Q_TYPESET(oqctl, type))
1490                         continue;
1491                 if (!(qctxt->lqc_flags & UGQUOTA2LQC(oqctl->qc_type)))
1492                         return 0;
1493         }
1494         return 1;
1495 }
1496
1497 int quota_is_off(struct lustre_quota_ctxt *qctxt, struct obd_quotactl *oqctl)
1498 {
1499         unsigned int type;
1500
1501         for (type = USRQUOTA; type < MAXQUOTAS; type++) {
1502                 if (!Q_TYPESET(oqctl, type))
1503                         continue;
1504                 if (qctxt->lqc_flags & UGQUOTA2LQC(oqctl->qc_type))
1505                         return 0;
1506         }
1507         return 1;
1508 }
1509
1510 /**
1511  * When quotaon, build a lqs for every uid/gid who has been set limitation
1512  * for quota. After quota_search_lqs, it will hold one ref for the lqs.
1513  * It will be released when qctxt_cleanup() is executed b=18574
1514  *
1515  * Should be called with obt->obt_quotachecking held. b=20152
1516  */
1517 void build_lqs(struct obd_device *obd)
1518 {
1519         struct obd_device_target *obt = &obd->u.obt;
1520         struct lustre_quota_ctxt *qctxt = &obt->obt_qctxt;
1521         struct list_head id_list;
1522         int i, rc;
1523
1524         LASSERT_SEM_LOCKED(&obt->obt_quotachecking);
1525         INIT_LIST_HEAD(&id_list);
1526         for (i = 0; i < MAXQUOTAS; i++) {
1527                 struct dquot_id *dqid, *tmp;
1528
1529                 if (sb_dqopt(qctxt->lqc_sb)->files[i] == NULL)
1530                         continue;
1531
1532 #ifndef KERNEL_SUPPORTS_QUOTA_READ
1533                 rc = fsfilt_qids(obd, sb_dqopt(qctxt->lqc_sb)->files[i], NULL,
1534                                  i, &id_list);
1535 #else
1536                 rc = fsfilt_qids(obd, NULL, sb_dqopt(qctxt->lqc_sb)->files[i],
1537                                  i, &id_list);
1538 #endif
1539                 if (rc) {
1540                         CERROR("%s: failed to get %s qids!\n", obd->obd_name,
1541                                i ? "group" : "user");
1542                         continue;
1543                 }
1544
1545                 list_for_each_entry_safe(dqid, tmp, &id_list,
1546                                          di_link) {
1547                         struct lustre_qunit_size *lqs;
1548
1549                         list_del_init(&dqid->di_link);
1550                         lqs = quota_search_lqs(LQS_KEY(i, dqid->di_id),
1551                                                qctxt, 1);
1552                         if (lqs && !IS_ERR(lqs)) {
1553                                 lqs->lqs_flags |= dqid->di_flag;
1554                                 lqs_putref(lqs);
1555                         } else {
1556                                 CERROR("%s: failed to create a lqs for %sid %u"
1557                                        "\n", obd->obd_name, i ? "g" : "u",
1558                                        dqid->di_id);
1559                         }
1560
1561                         OBD_FREE_PTR(dqid);
1562                 }
1563         }
1564 }
1565
1566 /**
1567  * lqs<->qctxt hash operations
1568  */
1569
1570 /**
1571  * string hashing using djb2 hash algorithm
1572  */
1573 static unsigned
1574 lqs_hash(cfs_hash_t *hs, void *key, unsigned mask)
1575 {
1576         struct quota_adjust_qunit *lqs_key;
1577         unsigned hash;
1578         ENTRY;
1579
1580         LASSERT(key);
1581         lqs_key = (struct quota_adjust_qunit *)key;
1582         hash = (QAQ_IS_GRP(lqs_key) ? 5381 : 5387) * lqs_key->qaq_id;
1583
1584         RETURN(hash & mask);
1585 }
1586
1587 static int
1588 lqs_compare(void *key, struct hlist_node *hnode)
1589 {
1590         struct lustre_qunit_size *q;
1591         int rc;
1592         ENTRY;
1593
1594         LASSERT(key);
1595         q = hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1596
1597         spin_lock(&q->lqs_lock);
1598         rc = (q->lqs_key == *((unsigned long long *)key));
1599         spin_unlock(&q->lqs_lock);
1600
1601         RETURN(rc);
1602 }
1603
1604 static void *
1605 lqs_get(struct hlist_node *hnode)
1606 {
1607         struct lustre_qunit_size *q =
1608                 hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1609         ENTRY;
1610
1611         __lqs_getref(q);
1612
1613         RETURN(q);
1614 }
1615
1616 static void *
1617 lqs_put(struct hlist_node *hnode)
1618 {
1619         struct lustre_qunit_size *q =
1620                 hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1621         ENTRY;
1622
1623         __lqs_putref(q);
1624
1625         RETURN(q);
1626 }
1627
1628 static void
1629 lqs_exit(struct hlist_node *hnode)
1630 {
1631         struct lustre_qunit_size *q =
1632                 hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1633         ENTRY;
1634
1635         /*
1636          * Nothing should be left. User of lqs put it and
1637          * lqs also was deleted from table by this time
1638          * so we should have 0 refs.
1639          */
1640         LASSERTF(atomic_read(&q->lqs_refcount) == 0,
1641                  "Busy lqs %p with %d refs\n", q,
1642                  atomic_read(&q->lqs_refcount));
1643         OBD_FREE_PTR(q);
1644         EXIT;
1645 }
1646
1647 static cfs_hash_ops_t lqs_hash_ops = {
1648         .hs_hash    = lqs_hash,
1649         .hs_compare = lqs_compare,
1650         .hs_get     = lqs_get,
1651         .hs_put     = lqs_put,
1652         .hs_exit    = lqs_exit
1653 };
1654 #endif /* HAVE_QUOTA_SUPPORT */