lustre/quota/quota_context.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/quota/quota_context.c
  37  *
  38  * Lustre Quota Context
  39  *
  40  * Author: Niu YaWei <niu@clusterfs.com>
  41  */
  42
  43 #ifndef EXPORT_SYMTAB
  44 # define EXPORT_SYMTAB
  45 #endif
  46
  47 #define DEBUG_SUBSYSTEM S_LQUOTA
  48
  49 #include <linux/version.h>
  50 #include <linux/fs.h>
  51 #include <asm/unistd.h>
  52 #include <linux/slab.h>
  53 #include <linux/quotaops.h>
  54 #include <linux/module.h>
  55 #include <linux/init.h>
  56
  57 #include <obd_class.h>
  58 #include <lustre_quota.h>
  59 #include <lustre_fsfilt.h>
  60 #include <class_hash.h>
  61 #include <lprocfs_status.h>
  62 #include "quota_internal.h"
  63
  64 #ifdef HAVE_QUOTA_SUPPORT
  65
  66 static lustre_hash_ops_t lqs_hash_ops;
  67
  68 unsigned long default_bunit_sz = 128 * 1024 * 1024; /* 128M bytes */
  69 unsigned long default_btune_ratio = 50;             /* 50 percentage */
  70 unsigned long default_iunit_sz = 5120;              /* 5120 inodes */
  71 unsigned long default_itune_ratio = 50;             /* 50 percentage */
  72
  73 cfs_mem_cache_t *qunit_cachep = NULL;
  74 struct list_head qunit_hash[NR_DQHASH];
  75 spinlock_t qunit_hash_lock = SPIN_LOCK_UNLOCKED;
  76
  77 /* please sync qunit_state with qunit_state_names */
  78 enum qunit_state {
  79         QUNIT_CREATED      = 0,   /* a qunit is created */
  80         QUNIT_IN_HASH      = 1,   /* a qunit is added into qunit hash, that means
  81                                    * a quota req will be sent or is flying */
  82         QUNIT_RM_FROM_HASH = 2,   /* a qunit is removed from qunit hash, that
  83                                    * means a quota req is handled and comes
  84                                    * back */
  85         QUNIT_FINISHED     = 3,   /* qunit can wake up all threads waiting
  86                                    * for it */
  87 };
  88
  89 static const char *qunit_state_names[] = {
  90         [QUNIT_CREATED]      = "CREATED",
  91         [QUNIT_IN_HASH]      = "IN_HASH",
  92         [QUNIT_RM_FROM_HASH] = "RM_FROM_HASH",
  93         [QUNIT_FINISHED]     = "FINISHED",
  94 };
  95
  96 struct lustre_qunit {
  97         struct list_head lq_hash;          /* Hash list in memory */
  98         atomic_t lq_refcnt;                /* Use count */
  99         struct lustre_quota_ctxt *lq_ctxt; /* Quota context this applies to */
 100         struct qunit_data lq_data;         /* See qunit_data */
 101         unsigned int lq_opc;               /* QUOTA_DQACQ, QUOTA_DQREL */
 102         cfs_waitq_t lq_waitq;              /* Threads waiting for this qunit */
 103         spinlock_t lq_lock;                /* Protect the whole structure */
 104         enum qunit_state lq_state;         /* Present the status of qunit */
 105         int lq_rc;                         /* The rc of lq_data */
 106 };
 107
 108 #define QUNIT_SET_STATE(qunit, state)                                   \
 109 do {                                                                    \
 110         spin_lock(&qunit->lq_lock);                                     \
 111         QDATA_DEBUG((&qunit->lq_data), "qunit(%p) lq_state(%s->%s), "   \
 112                     "lq_rc(%d)\n",                                      \
 113                     qunit, qunit_state_names[qunit->lq_state],          \
 114                     qunit_state_names[state], qunit->lq_rc);            \
 115         qunit->lq_state = state;                                        \
 116         spin_unlock(&qunit->lq_lock);                                   \
 117 } while(0)
 118
 119 #define QUNIT_SET_STATE_AND_RC(qunit, state, rc)                        \
 120 do {                                                                    \
 121         spin_lock(&qunit->lq_lock);                                     \
 122         qunit->lq_rc = rc;                                              \
 123         QDATA_DEBUG((&qunit->lq_data), "qunit(%p) lq_state(%s->%s), "   \
 124                     "lq_rc(%d)\n",                                      \
 125                     qunit, qunit_state_names[qunit->lq_state],          \
 126                     qunit_state_names[state], qunit->lq_rc);            \
 127         qunit->lq_state = state;                                        \
 128         spin_unlock(&qunit->lq_lock);                                   \
 129 } while(0)
 130
 131 int should_translate_quota (struct obd_import *imp)
 132 {
 133         ENTRY;
 134
 135         LASSERT(imp);
 136         if (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_QUOTA64)
 137                 RETURN(0);
 138         else
 139                 RETURN(1);
 140 }
 141
 142 void qunit_cache_cleanup(void)
 143 {
 144         int i;
 145         ENTRY;
 146
 147         spin_lock(&qunit_hash_lock);
 148         for (i = 0; i < NR_DQHASH; i++)
 149                 LASSERT(list_empty(qunit_hash + i));
 150         spin_unlock(&qunit_hash_lock);
 151
 152         if (qunit_cachep) {
 153                 int rc;
 154                 rc = cfs_mem_cache_destroy(qunit_cachep);
 155                 LASSERTF(rc == 0, "couldn't destory qunit_cache slab\n");
 156                 qunit_cachep = NULL;
 157         }
 158         EXIT;
 159 }
 160
 161 int qunit_cache_init(void)
 162 {
 163         int i;
 164         ENTRY;
 165
 166         LASSERT(qunit_cachep == NULL);
 167         qunit_cachep = cfs_mem_cache_create("ll_qunit_cache",
 168                                             sizeof(struct lustre_qunit),
 169                                             0, 0);
 170         if (!qunit_cachep)
 171                 RETURN(-ENOMEM);
 172
 173         spin_lock(&qunit_hash_lock);
 174         for (i = 0; i < NR_DQHASH; i++)
 175                 INIT_LIST_HEAD(qunit_hash + i);
 176         spin_unlock(&qunit_hash_lock);
 177         RETURN(0);
 178 }
 179
 180 static inline int
 181 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
 182              __attribute__((__const__));
 183
 184 static inline int
 185 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
 186 {
 187         unsigned int id = qdata->qd_id;
 188         unsigned int type = QDATA_IS_GRP(qdata);
 189
 190         unsigned long tmp = ((unsigned long)qctxt >> L1_CACHE_SHIFT) ^ id;
 191         tmp = (tmp * (MAXQUOTAS - type)) % NR_DQHASH;
 192         return tmp;
 193 }
 194
 195 /* caller must hold qunit_hash_lock */
 196 static inline struct lustre_qunit *find_qunit(unsigned int hashent,
 197                                               struct lustre_quota_ctxt *qctxt,
 198                                               struct qunit_data *qdata)
 199 {
 200         struct lustre_qunit *qunit = NULL;
 201         struct qunit_data *tmp;
 202
 203         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
 204         list_for_each_entry(qunit, qunit_hash + hashent, lq_hash) {
 205                 tmp = &qunit->lq_data;
 206                 if (qunit->lq_ctxt == qctxt &&
 207                     qdata->qd_id == tmp->qd_id &&
 208                     (qdata->qd_flags & LQUOTA_QUNIT_FLAGS) ==
 209                     (tmp->qd_flags & LQUOTA_QUNIT_FLAGS))
 210                         return qunit;
 211         }
 212         return NULL;
 213 }
 214
 215 /* check_cur_qunit - check the current usage of qunit.
 216  * @qctxt: quota context
 217  * @qdata: the type of quota unit to be checked
 218  *
 219  * return: 1 - need acquire qunit;
 220  *         2 - need release qunit;
 221  *         0 - need do nothing.
 222  *       < 0 - error.
 223  */
 224 static int
 225 check_cur_qunit(struct obd_device *obd,
 226                 struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
 227 {
 228         struct super_block *sb = qctxt->lqc_sb;
 229         unsigned long qunit_sz, tune_sz;
 230         __u64 usage, limit, limit_org, pending_write = 0;
 231         long long record = 0;
 232         struct obd_quotactl *qctl;
 233         struct lustre_qunit_size *lqs = NULL;
 234         int ret = 0;
 235         ENTRY;
 236
 237         if (!ll_sb_any_quota_active(sb))
 238                 RETURN(0);
 239
 240         spin_lock(&qctxt->lqc_lock);
 241         if (!qctxt->lqc_valid){
 242                 spin_unlock(&qctxt->lqc_lock);
 243                 RETURN(0);
 244         }
 245         spin_unlock(&qctxt->lqc_lock);
 246
 247         OBD_ALLOC_PTR(qctl);
 248         if (qctl == NULL)
 249                 RETURN(-ENOMEM);
 250
 251         /* get fs quota usage & limit */
 252         qctl->qc_cmd = Q_GETQUOTA;
 253         qctl->qc_id = qdata->qd_id;
 254         qctl->qc_type = QDATA_IS_GRP(qdata);
 255         ret = fsfilt_quotactl(obd, sb, qctl);
 256         if (ret) {
 257                 if (ret == -ESRCH)      /* no limit */
 258                         ret = 0;
 259                 else
 260                         CERROR("can't get fs quota usage! (rc:%d)\n", ret);
 261                 GOTO(out, ret);
 262         }
 263
 264         if (QDATA_IS_BLK(qdata)) {
 265                 usage = qctl->qc_dqblk.dqb_curspace;
 266                 limit = qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS;
 267         } else {
 268                 usage = qctl->qc_dqblk.dqb_curinodes;
 269                 limit = qctl->qc_dqblk.dqb_ihardlimit;
 270         }
 271
 272         /* ignore the no quota limit case; and it can avoid creating
 273          * unnecessary lqs for uid/gid */
 274         if (!limit)
 275                 GOTO(out, ret = 0);
 276
 277         lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(qdata), qdata->qd_id),
 278                                qctxt, 0);
 279         if (IS_ERR(lqs) || lqs == NULL) {
 280                 CERROR("fail to find a lqs for %sid: %u)!\n",
 281                        QDATA_IS_GRP(qdata) ? "g" : "u", qdata->qd_id);
 282                 GOTO (out, ret = 0);
 283         }
 284         spin_lock(&lqs->lqs_lock);
 285
 286         if (QDATA_IS_BLK(qdata)) {
 287                 qunit_sz = lqs->lqs_bunit_sz;
 288                 tune_sz  = lqs->lqs_btune_sz;
 289                 pending_write = lqs->lqs_bwrite_pending;
 290                 record   = lqs->lqs_blk_rec;
 291                 LASSERT(!(qunit_sz % QUOTABLOCK_SIZE));
 292         } else {
 293                 /* we didn't need change inode qunit size now */
 294                 qunit_sz = lqs->lqs_iunit_sz;
 295                 tune_sz  = lqs->lqs_itune_sz;
 296                 pending_write = lqs->lqs_iwrite_pending;
 297                 record   = lqs->lqs_ino_rec;
 298         }
 299
 300         /* we don't count the MIN_QLIMIT */
 301         if ((limit == MIN_QLIMIT && !QDATA_IS_BLK(qdata)) ||
 302             (toqb(limit) == MIN_QLIMIT && QDATA_IS_BLK(qdata)))
 303                 limit = 0;
 304
 305         usage += pending_write;
 306         limit_org = limit;
 307         /* when a releasing quota req is sent, before it returned
 308            limit is assigned a small value. limit will overflow */
 309         if (limit + record < 0)
 310                 usage -= record;
 311         else
 312                 limit += record;
 313
 314         LASSERT(qdata->qd_count == 0);
 315         if (limit <= usage + tune_sz) {
 316                 while (qdata->qd_count + limit <=
 317                        usage + tune_sz)
 318                         qdata->qd_count += qunit_sz;
 319                 ret = 1;
 320         } else if (limit > usage + qunit_sz + tune_sz &&
 321                    limit_org > qdata->qd_count + qunit_sz) {
 322                 while (limit - qdata->qd_count > usage + qunit_sz + tune_sz &&
 323                        limit_org > qdata->qd_count + qunit_sz)
 324                         qdata->qd_count += qunit_sz;
 325                 ret = 2;
 326                 /* if there are other pending writes for this uid/gid, releasing
 327                  * quota is put off until the last pending write b=16645 */
 328                 if (ret == 2 && pending_write) {
 329                         CDEBUG(D_QUOTA, "delay quota release\n");
 330                         ret = 0;
 331                 }
 332         }
 333         CDEBUG(D_QUOTA, "type: %c, limit: "LPU64", usage: "LPU64
 334                ", pending_write: "LPU64", record: %llu"
 335                ", qunit_sz: %lu, tune_sz: %lu, ret: %d.\n",
 336                QDATA_IS_BLK(qdata) ? 'b' : 'i', limit, usage, pending_write,
 337                (__s64)record, qunit_sz, tune_sz, ret);
 338         LASSERT(ret == 0 || qdata->qd_count);
 339
 340         spin_unlock(&lqs->lqs_lock);
 341         lqs_putref(lqs);
 342
 343         EXIT;
 344  out:
 345         OBD_FREE_PTR(qctl);
 346         return ret;
 347 }
 348
 349 /* compute the remaining quota for certain gid or uid b=11693 */
 350 int compute_remquota(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 351                      struct qunit_data *qdata, int isblk)
 352 {
 353         struct super_block *sb = qctxt->lqc_sb;
 354         __u64 usage, limit;
 355         struct obd_quotactl *qctl;
 356         int ret = QUOTA_RET_OK;
 357         ENTRY;
 358
 359         if (!ll_sb_any_quota_active(sb))
 360                 RETURN(QUOTA_RET_NOQUOTA);
 361
 362         /* ignore root user */
 363         if (qdata->qd_id == 0 && QDATA_IS_GRP(qdata) == USRQUOTA)
 364                 RETURN(QUOTA_RET_NOLIMIT);
 365
 366         OBD_ALLOC_PTR(qctl);
 367         if (qctl == NULL)
 368                 RETURN(-ENOMEM);
 369
 370         /* get fs quota usage & limit */
 371         qctl->qc_cmd = Q_GETQUOTA;
 372         qctl->qc_id = qdata->qd_id;
 373         qctl->qc_type = QDATA_IS_GRP(qdata);
 374         ret = fsfilt_quotactl(obd, sb, qctl);
 375         if (ret) {
 376                 if (ret == -ESRCH)      /* no limit */
 377                         ret = QUOTA_RET_NOLIMIT;
 378                 else
 379                         CDEBUG(D_QUOTA, "can't get fs quota usage! (rc:%d)",
 380                                ret);
 381                 GOTO(out, ret);
 382         }
 383
 384         usage = isblk ? qctl->qc_dqblk.dqb_curspace :
 385                 qctl->qc_dqblk.dqb_curinodes;
 386         limit = isblk ? qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS :
 387                 qctl->qc_dqblk.dqb_ihardlimit;
 388         if (!limit){            /* no limit */
 389                 ret = QUOTA_RET_NOLIMIT;
 390                 GOTO(out, ret);
 391         }
 392
 393         if (limit >= usage)
 394                 qdata->qd_count = limit - usage;
 395         else
 396                 qdata->qd_count = 0;
 397         EXIT;
 398 out:
 399         OBD_FREE_PTR(qctl);
 400         return ret;
 401 }
 402
 403 static struct lustre_qunit *alloc_qunit(struct lustre_quota_ctxt *qctxt,
 404                                         struct qunit_data *qdata, int opc)
 405 {
 406         struct lustre_qunit *qunit = NULL;
 407         ENTRY;
 408
 409         OBD_SLAB_ALLOC(qunit, qunit_cachep, CFS_ALLOC_IO, sizeof(*qunit));
 410         if (qunit == NULL)
 411                 RETURN(NULL);
 412
 413         INIT_LIST_HEAD(&qunit->lq_hash);
 414         init_waitqueue_head(&qunit->lq_waitq);
 415         atomic_set(&qunit->lq_refcnt, 1);
 416         qunit->lq_ctxt = qctxt;
 417         memcpy(&qunit->lq_data, qdata, sizeof(*qdata));
 418         qunit->lq_opc = opc;
 419         qunit->lq_lock = SPIN_LOCK_UNLOCKED;
 420         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_CREATED, 0);
 421         RETURN(qunit);
 422 }
 423
 424 static inline void free_qunit(struct lustre_qunit *qunit)
 425 {
 426         OBD_SLAB_FREE(qunit, qunit_cachep, sizeof(*qunit));
 427 }
 428
 429 static inline void qunit_get(struct lustre_qunit *qunit)
 430 {
 431         atomic_inc(&qunit->lq_refcnt);
 432 }
 433
 434 static void qunit_put(struct lustre_qunit *qunit)
 435 {
 436         LASSERT(atomic_read(&qunit->lq_refcnt));
 437         if (atomic_dec_and_test(&qunit->lq_refcnt))
 438                 free_qunit(qunit);
 439 }
 440
 441 /* caller must hold qunit_hash_lock and release ref of qunit after using it */
 442 static struct lustre_qunit *dqacq_in_flight(struct lustre_quota_ctxt *qctxt,
 443                                             struct qunit_data *qdata)
 444 {
 445         unsigned int hashent = qunit_hashfn(qctxt, qdata);
 446         struct lustre_qunit *qunit;
 447         ENTRY;
 448
 449         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
 450         qunit = find_qunit(hashent, qctxt, qdata);
 451         if (qunit)
 452                 qunit_get(qunit);
 453         RETURN(qunit);
 454 }
 455
 456 static void
 457 insert_qunit_nolock(struct lustre_quota_ctxt *qctxt, struct lustre_qunit *qunit)
 458 {
 459         struct list_head *head;
 460
 461         LASSERT(list_empty(&qunit->lq_hash));
 462         qunit_get(qunit);
 463         head = qunit_hash + qunit_hashfn(qctxt, &qunit->lq_data);
 464         list_add(&qunit->lq_hash, head);
 465         QUNIT_SET_STATE(qunit, QUNIT_IN_HASH);
 466 }
 467
 468 static void compute_lqs_after_removing_qunit(struct lustre_qunit *qunit)
 469 {
 470         struct lustre_qunit_size *lqs;
 471
 472         lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(&qunit->lq_data),
 473                                        qunit->lq_data.qd_id),
 474                                qunit->lq_ctxt, 0);
 475         if (lqs && !IS_ERR(lqs)) {
 476                 spin_lock(&lqs->lqs_lock);
 477                 if (qunit->lq_opc == QUOTA_DQACQ)
 478                         quota_compute_lqs(&qunit->lq_data, lqs, 0, 1);
 479                 if (qunit->lq_opc == QUOTA_DQREL)
 480                         quota_compute_lqs(&qunit->lq_data, lqs, 0, 0);
 481                 spin_unlock(&lqs->lqs_lock);
 482                 /* this is for quota_search_lqs */
 483                 lqs_putref(lqs);
 484                 /* this is for schedule_dqacq */
 485                 lqs_putref(lqs);
 486         }
 487 }
 488
 489 static void remove_qunit_nolock(struct lustre_qunit *qunit)
 490 {
 491         LASSERT(!list_empty(&qunit->lq_hash));
 492         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
 493
 494         list_del_init(&qunit->lq_hash);
 495         QUNIT_SET_STATE(qunit, QUNIT_RM_FROM_HASH);
 496         qunit_put(qunit);
 497 }
 498
 499 void* quota_barrier(struct lustre_quota_ctxt *qctxt,
 500                     struct obd_quotactl *oqctl, int isblk)
 501 {
 502         struct lustre_qunit *qunit, *find_qunit;
 503         int cycle = 1;
 504
 505         OBD_SLAB_ALLOC(qunit, qunit_cachep, CFS_ALLOC_IO, sizeof(*qunit));
 506         if (qunit == NULL) {
 507                 CERROR("locating %sunit failed for %sid %u\n",
 508                        isblk ? "b" : "i", oqctl->qc_type ? "g" : "u",
 509                        oqctl->qc_id);
 510                 qctxt_wait_pending_dqacq(qctxt, oqctl->qc_id,
 511                                          oqctl->qc_type, isblk);
 512                 return NULL;
 513         }
 514
 515         INIT_LIST_HEAD(&qunit->lq_hash);
 516         qunit->lq_lock = SPIN_LOCK_UNLOCKED;
 517         init_waitqueue_head(&qunit->lq_waitq);
 518         atomic_set(&qunit->lq_refcnt, 1);
 519         qunit->lq_ctxt = qctxt;
 520         qunit->lq_data.qd_id = oqctl->qc_id;
 521         qunit->lq_data.qd_flags =  oqctl->qc_type;
 522         if (isblk)
 523                 QDATA_SET_BLK(&qunit->lq_data);
 524         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_CREATED, 0);
 525         /* it means it is only an invalid qunit for barrier */
 526         qunit->lq_opc = QUOTA_LAST_OPC;
 527
 528         while (1) {
 529                 spin_lock(&qunit_hash_lock);
 530                 find_qunit = dqacq_in_flight(qctxt, &qunit->lq_data);
 531                 if (find_qunit) {
 532                         spin_unlock(&qunit_hash_lock);
 533                         qunit_put(find_qunit);
 534                         qctxt_wait_pending_dqacq(qctxt, oqctl->qc_id,
 535                                                  oqctl->qc_type, isblk);
 536                         CDEBUG(D_QUOTA, "cycle=%d\n", cycle++);
 537                         continue;
 538                 }
 539                 break;
 540         }
 541         insert_qunit_nolock(qctxt, qunit);
 542         spin_unlock(&qunit_hash_lock);
 543         return qunit;
 544 }
 545
 546 void quota_unbarrier(void *handle)
 547 {
 548         struct lustre_qunit *qunit = (struct lustre_qunit *)handle;
 549
 550         if (qunit == NULL) {
 551                 CERROR("handle is NULL\n");
 552                 return;
 553         }
 554
 555         LASSERT(qunit->lq_opc == QUOTA_LAST_OPC);
 556         spin_lock(&qunit_hash_lock);
 557         remove_qunit_nolock(qunit);
 558         spin_unlock(&qunit_hash_lock);
 559         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, QUOTA_REQ_RETURNED);
 560         wake_up(&qunit->lq_waitq);
 561         qunit_put(qunit);
 562 }
 563
 564 #define INC_QLIMIT(limit, count) (limit == MIN_QLIMIT) ? \
 565                                  (limit = count) : (limit += count)
 566
 567
 568 /* FIXME check if this mds is the master of specified id */
 569 static int
 570 is_master(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 571           unsigned int id, int type)
 572 {
 573         return qctxt->lqc_handler ? 1 : 0;
 574 }
 575
 576 static int
 577 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 578                struct qunit_data *qdata, int opc, int wait,
 579                struct obd_trans_info *oti);
 580
 581 static inline void qdata_to_oqaq(struct qunit_data *qdata,
 582                                  struct quota_adjust_qunit *oqaq)
 583 {
 584         LASSERT(qdata);
 585         LASSERT(oqaq);
 586
 587         oqaq->qaq_flags = qdata->qd_flags;
 588         oqaq->qaq_id    = qdata->qd_id;
 589         if (QDATA_IS_ADJBLK(qdata))
 590                 oqaq->qaq_bunit_sz = qdata->qd_qunit;
 591         if (QDATA_IS_ADJINO(qdata))
 592                 oqaq->qaq_iunit_sz = qdata->qd_qunit;
 593 }
 594
 595 static int
 596 dqacq_completion(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 597                  struct qunit_data *qdata, int rc, int opc)
 598 {
 599         struct lustre_qunit *qunit = NULL;
 600         struct super_block *sb = qctxt->lqc_sb;
 601         int err = 0;
 602         struct quota_adjust_qunit *oqaq = NULL;
 603         int rc1 = 0;
 604         ENTRY;
 605
 606         LASSERT(qdata);
 607         QDATA_DEBUG(qdata, "obd(%s): complete %s quota req\n",
 608                     obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
 609
 610         /* do it only when a releasing quota req more than 5MB b=18491 */
 611         if (opc == QUOTA_DQREL && qdata->qd_count >= 5242880)
 612                 OBD_FAIL_TIMEOUT(OBD_FAIL_QUOTA_DELAY_REL, 5);
 613
 614         /* update local operational quota file */
 615         if (rc == 0) {
 616                 __u64 count = QUSG(qdata->qd_count, QDATA_IS_BLK(qdata));
 617                 struct obd_quotactl *qctl;
 618                 __u64 *hardlimit;
 619
 620                 OBD_ALLOC_PTR(qctl);
 621                 if (qctl == NULL)
 622                         GOTO(out, err = -ENOMEM);
 623
 624                 /* acq/rel qunit for specified uid/gid is serialized,
 625                  * so there is no race between get fs quota limit and
 626                  * set fs quota limit */
 627                 qctl->qc_cmd = Q_GETQUOTA;
 628                 qctl->qc_id = qdata->qd_id;
 629                 qctl->qc_type = QDATA_IS_GRP(qdata);
 630                 err = fsfilt_quotactl(obd, sb, qctl);
 631                 if (err) {
 632                         CERROR("error get quota fs limit! (rc:%d)\n", err);
 633                         GOTO(out_mem, err);
 634                 }
 635
 636                 if (QDATA_IS_BLK(qdata)) {
 637                         qctl->qc_dqblk.dqb_valid = QIF_BLIMITS;
 638                         hardlimit = &qctl->qc_dqblk.dqb_bhardlimit;
 639                 } else {
 640                         qctl->qc_dqblk.dqb_valid = QIF_ILIMITS;
 641                         hardlimit = &qctl->qc_dqblk.dqb_ihardlimit;
 642                 }
 643
 644                 CDEBUG(D_QUOTA, "hardlimt: "LPU64"\n", *hardlimit);
 645
 646                 if (*hardlimit == 0)
 647                         goto out_mem;
 648
 649                 switch (opc) {
 650                 case QUOTA_DQACQ:
 651                         INC_QLIMIT(*hardlimit, count);
 652                         break;
 653                 case QUOTA_DQREL:
 654                         if (count >= *hardlimit)
 655                                 CERROR("release quota error: id(%u) flag(%u) "
 656                                        "type(%c) isblk(%c) count("LPU64") "
 657                                        "qd_qunit("LPU64") hardlimit("LPU64") "
 658                                        "qdata(%p)\n",
 659                                        qdata->qd_id, qdata->qd_flags,
 660                                        QDATA_IS_GRP(qdata) ? 'g' : 'u',
 661                                        QDATA_IS_BLK(qdata) ? 'b': 'i',
 662                                        qdata->qd_count, qdata->qd_qunit, *hardlimit,
 663                                        qdata);
 664                         else
 665                                 *hardlimit -= count;
 666                         break;
 667                 default:
 668                         LBUG();
 669                 }
 670
 671                 /* clear quota limit */
 672                 if (count == 0)
 673                         *hardlimit = 0;
 674
 675                 qctl->qc_cmd = Q_SETQUOTA;
 676                 err = fsfilt_quotactl(obd, sb, qctl);
 677                 if (err)
 678                         CERROR("error set quota fs limit! (rc:%d)\n", err);
 679
 680                 QDATA_DEBUG(qdata, "%s completion\n",
 681                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 682 out_mem:
 683                 OBD_FREE_PTR(qctl);
 684         } else if (rc == -EDQUOT) {
 685                 QDATA_DEBUG(qdata, "acquire qunit got EDQUOT.\n");
 686         } else if (rc == -EBUSY) {
 687                 QDATA_DEBUG(qdata, "it's is recovering, got EBUSY.\n");
 688         } else {
 689                 CERROR("acquire qunit got error! (rc:%d)\n", rc);
 690         }
 691 out:
 692         /* remove the qunit from hash */
 693         spin_lock(&qunit_hash_lock);
 694
 695         qunit = dqacq_in_flight(qctxt, qdata);
 696         /* this qunit has been removed by qctxt_cleanup() */
 697         if (!qunit) {
 698                 spin_unlock(&qunit_hash_lock);
 699                 QDATA_DEBUG(qdata, "%s is discarded because qunit isn't found\n",
 700                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 701                 RETURN(err);
 702         }
 703
 704         LASSERT(opc == qunit->lq_opc);
 705         /* remove this qunit from lq_hash so that new processes cannot be added
 706          * to qunit->lq_waiters */
 707         remove_qunit_nolock(qunit);
 708         spin_unlock(&qunit_hash_lock);
 709
 710         compute_lqs_after_removing_qunit(qunit);
 711
 712
 713         if (rc == 0)
 714                 rc = QUOTA_REQ_RETURNED;
 715         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, rc);
 716         /* wake up all waiters */
 717         wake_up(&qunit->lq_waitq);
 718
 719         /* this is for dqacq_in_flight() */
 720         qunit_put(qunit);
 721         /* this is for alloc_qunit() */
 722         qunit_put(qunit);
 723         if (rc < 0 && rc != -EDQUOT)
 724                  RETURN(err);
 725
 726         /* don't reschedule in such cases:
 727          *   - acq/rel failure and qunit isn't changed,
 728          *     but not for quota recovery.
 729          *   - local dqacq/dqrel.
 730          *   - local disk io failure.
 731          */
 732          OBD_ALLOC_PTR(oqaq);
 733          if (!oqaq)
 734                  RETURN(-ENOMEM);
 735          qdata_to_oqaq(qdata, oqaq);
 736          /* adjust the qunit size in slaves */
 737          rc1 = quota_adjust_slave_lqs(oqaq, qctxt);
 738          OBD_FREE_PTR(oqaq);
 739          if (rc1 < 0) {
 740                  CERROR("adjust slave's qunit size failed!(rc:%d)\n", rc1);
 741                  RETURN(rc1);
 742          }
 743          if (err || (rc < 0 && rc != -EBUSY && rc1 == 0) ||
 744              is_master(obd, qctxt, qdata->qd_id, QDATA_IS_GRP(qdata)))
 745                 RETURN(err);
 746
 747          if (opc == QUOTA_DQREL && qdata->qd_count >= 5242880)
 748                  OBD_FAIL_RETURN(OBD_FAIL_QUOTA_DELAY_REL, err);
 749
 750         /* reschedule another dqacq/dqrel if needed */
 751         qdata->qd_count = 0;
 752         qdata->qd_flags &= LQUOTA_QUNIT_FLAGS;
 753         rc1 = check_cur_qunit(obd, qctxt, qdata);
 754         if (rc1 > 0) {
 755                 int opc;
 756                 opc = rc1 == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
 757                 rc1 = schedule_dqacq(obd, qctxt, qdata, opc, 0, NULL);
 758                 QDATA_DEBUG(qdata, "reschedudle opc(%d) rc(%d)\n", opc, rc1);
 759         }
 760         RETURN(err);
 761 }
 762
 763 struct dqacq_async_args {
 764         struct lustre_quota_ctxt *aa_ctxt;
 765         struct lustre_qunit *aa_qunit;
 766 };
 767
 768 static int dqacq_interpret(struct ptlrpc_request *req, void *data, int rc)
 769 {
 770         struct dqacq_async_args *aa = (struct dqacq_async_args *)data;
 771         struct lustre_quota_ctxt *qctxt = aa->aa_ctxt;
 772         struct lustre_qunit *qunit = aa->aa_qunit;
 773         struct obd_device *obd = req->rq_import->imp_obd;
 774         struct qunit_data *qdata = NULL;
 775         int rc1 = 0;
 776         ENTRY;
 777
 778         LASSERT(req);
 779         LASSERT(req->rq_import);
 780
 781         /* there are several forms of qunit(historic causes), so we need to
 782          * adjust qunit from slaves to the same form here */
 783         OBD_ALLOC(qdata, sizeof(struct qunit_data));
 784         if (!qdata)
 785                 RETURN(-ENOMEM);
 786
 787         /* if a quota req timeouts or is dropped, we should update quota
 788          * statistics which will be handled in dqacq_completion. And in
 789          * this situation we should get qdata from request instead of
 790          * reply */
 791         rc1 = quota_get_qdata(req, qdata,
 792                               (rc != 0) ? QUOTA_REQUEST : QUOTA_REPLY,
 793                               QUOTA_IMPORT);
 794         if (rc1 < 0) {
 795                 DEBUG_REQ(D_ERROR, req,
 796                           "error unpacking qunit_data(rc: %d)\n", rc1);
 797                 *qdata = qunit->lq_data;
 798         }
 799
 800         QDATA_DEBUG(qdata, "qdata: interpret rc(%d).\n", rc);
 801         QDATA_DEBUG((&qunit->lq_data), "lq_data: \n");
 802
 803         if (qdata->qd_id != qunit->lq_data.qd_id ||
 804             OBD_FAIL_CHECK_ONCE(OBD_FAIL_QUOTA_RET_QDATA)) {
 805                 CERROR("the returned qd_id isn't expected!"
 806                        "(qdata: %u, lq_data: %u)\n", qdata->qd_id,
 807                        qunit->lq_data.qd_id);
 808                 qdata->qd_id = qunit->lq_data.qd_id;
 809                 rc = -EPROTO;
 810         }
 811         if (QDATA_IS_GRP(qdata) != QDATA_IS_GRP(&qunit->lq_data)) {
 812                 CERROR("the returned grp/usr isn't expected!"
 813                        "(qdata: %u, lq_data: %u)\n", qdata->qd_flags,
 814                        qunit->lq_data.qd_flags);
 815                 if (QDATA_IS_GRP(&qunit->lq_data))
 816                         QDATA_SET_GRP(qdata);
 817                 else
 818                         QDATA_CLR_GRP(qdata);
 819                 rc = -EPROTO;
 820         }
 821         if (qdata->qd_count > qunit->lq_data.qd_count) {
 822                 CERROR("the returned qd_count isn't expected!"
 823                        "(qdata: "LPU64", lq_data: "LPU64")\n", qdata->qd_count,
 824                        qunit->lq_data.qd_count);
 825                 rc = -EPROTO;
 826         }
 827
 828         rc = dqacq_completion(obd, qctxt, qdata, rc,
 829                               lustre_msg_get_opc(req->rq_reqmsg));
 830
 831         OBD_FREE(qdata, sizeof(struct qunit_data));
 832
 833         RETURN(rc);
 834 }
 835
 836 /* check if quota master is online */
 837 int check_qm(struct lustre_quota_ctxt *qctxt)
 838 {
 839         int rc;
 840         ENTRY;
 841
 842         spin_lock(&qctxt->lqc_lock);
 843         /* quit waiting when mds is back or qctxt is cleaned up */
 844         rc = qctxt->lqc_import || !qctxt->lqc_valid;
 845         spin_unlock(&qctxt->lqc_lock);
 846
 847         RETURN(rc);
 848 }
 849
 850 /* wake up all waiting threads when lqc_import is NULL */
 851 void dqacq_interrupt(struct lustre_quota_ctxt *qctxt)
 852 {
 853         struct lustre_qunit *qunit, *tmp;
 854         int i;
 855         ENTRY;
 856
 857         spin_lock(&qunit_hash_lock);
 858         for (i = 0; i < NR_DQHASH; i++) {
 859                 list_for_each_entry_safe(qunit, tmp, &qunit_hash[i], lq_hash) {
 860                         if (qunit->lq_ctxt != qctxt)
 861                                 continue;
 862
 863                         /* Wake up all waiters. Do not change lq_state.
 864                          * The waiters will check lq_rc which is kept as 0
 865                          * if no others change it, then the waiters will return
 866                          * -EAGAIN to caller who can perform related quota
 867                          * acq/rel if necessary. */
 868                         wake_up_all(&qunit->lq_waitq);
 869                 }
 870         }
 871         spin_unlock(&qunit_hash_lock);
 872         EXIT;
 873 }
 874
 875 static int got_qunit(struct lustre_qunit *qunit, int is_master)
 876 {
 877         struct lustre_quota_ctxt *qctxt = qunit->lq_ctxt;
 878         int rc = 0;
 879         ENTRY;
 880
 881         spin_lock(&qunit->lq_lock);
 882         switch (qunit->lq_state) {
 883         case QUNIT_IN_HASH:
 884         case QUNIT_RM_FROM_HASH:
 885                 break;
 886         case QUNIT_FINISHED:
 887                 rc = 1;
 888                 break;
 889         default:
 890                 CERROR("invalid qunit state %d\n", qunit->lq_state);
 891         }
 892         spin_unlock(&qunit->lq_lock);
 893
 894         if (!rc) {
 895                 spin_lock(&qctxt->lqc_lock);
 896                 rc = !qctxt->lqc_valid;
 897                 if (!is_master)
 898                         rc |= !qctxt->lqc_import;
 899                 spin_unlock(&qctxt->lqc_lock);
 900         }
 901
 902         RETURN(rc);
 903 }
 904
 905 static int
 906 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 907                struct qunit_data *qdata, int opc, int wait,
 908                struct obd_trans_info *oti)
 909 {
 910         struct lustre_qunit *qunit, *empty;
 911         struct l_wait_info lwi = { 0 };
 912         struct ptlrpc_request *req;
 913         struct dqacq_async_args *aa;
 914         int size[2] = { sizeof(struct ptlrpc_body), 0 };
 915         struct obd_import *imp = NULL;
 916         struct lustre_qunit_size *lqs = NULL;
 917         struct timeval work_start;
 918         struct timeval work_end;
 919         long timediff;
 920         int ismaster = is_master(obd, qctxt, qdata->qd_id, QDATA_IS_GRP(qdata));
 921         int rc = 0;
 922         ENTRY;
 923
 924         LASSERT(opc == QUOTA_DQACQ || opc == QUOTA_DQREL);
 925         do_gettimeofday(&work_start);
 926         if ((empty = alloc_qunit(qctxt, qdata, opc)) == NULL)
 927                 RETURN(-ENOMEM);
 928
 929         spin_lock(&qunit_hash_lock);
 930         qunit = dqacq_in_flight(qctxt, qdata);
 931         if (qunit) {
 932                 spin_unlock(&qunit_hash_lock);
 933                 qunit_put(empty);
 934
 935                 goto wait_completion;
 936         }
 937         qunit = empty;
 938         qunit_get(qunit);
 939         insert_qunit_nolock(qctxt, qunit);
 940         spin_unlock(&qunit_hash_lock);
 941
 942         lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(qdata), qdata->qd_id),
 943                                qctxt, 0);
 944         if (lqs && !IS_ERR(lqs)) {
 945                 spin_lock(&lqs->lqs_lock);
 946                 quota_compute_lqs(qdata, lqs, 1, (opc == QUOTA_DQACQ) ? 1 : 0);
 947                 /* when this qdata returned from mds, it will call lqs_putref */
 948                 lqs_getref(lqs);
 949                 spin_unlock(&lqs->lqs_lock);
 950                 /* this is for quota_search_lqs */
 951                 lqs_putref(lqs);
 952         } else {
 953                 CERROR("Can't find the lustre qunit size!\n");
 954         }
 955
 956         QDATA_DEBUG(qdata, "obd(%s): send %s quota req\n",
 957                     obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
 958         /* master is going to dqacq/dqrel from itself */
 959         if (ismaster) {
 960                 int rc2;
 961                 QDATA_DEBUG(qdata, "local %s.\n",
 962                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 963                 QDATA_SET_CHANGE_QS(qdata);
 964                 rc = qctxt->lqc_handler(obd, qdata, opc);
 965                 rc2 = dqacq_completion(obd, qctxt, qdata, rc, opc);
 966                 /* this is for qunit_get() */
 967                 qunit_put(qunit);
 968
 969                 do_gettimeofday(&work_end);
 970                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
 971                 if (opc == QUOTA_DQACQ)
 972                         lprocfs_counter_add(qctxt->lqc_stats,
 973                                             wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
 974                                             timediff);
 975                 else
 976                         lprocfs_counter_add(qctxt->lqc_stats,
 977                                             wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
 978                                             timediff);
 979                 RETURN(rc ? rc : rc2);
 980         }
 981
 982         spin_lock(&qctxt->lqc_lock);
 983         if (!qctxt->lqc_import) {
 984                 spin_unlock(&qctxt->lqc_lock);
 985                 QDATA_DEBUG(qdata, "lqc_import is invalid.\n");
 986
 987                 spin_lock(&qunit_hash_lock);
 988                 remove_qunit_nolock(qunit);
 989                 spin_unlock(&qunit_hash_lock);
 990
 991                 compute_lqs_after_removing_qunit(qunit);
 992
 993                 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, -EAGAIN);
 994                 wake_up(&qunit->lq_waitq);
 995
 996                 /* this is for qunit_get() */
 997                 qunit_put(qunit);
 998                 /* this for alloc_qunit() */
 999                 qunit_put(qunit);
1000                 spin_lock(&qctxt->lqc_lock);
1001                 if (wait && !qctxt->lqc_import) {
1002                         spin_unlock(&qctxt->lqc_lock);
1003
1004                         LASSERT(oti && oti->oti_thread &&
1005                                 oti->oti_thread->t_watchdog);
1006
1007                         lc_watchdog_disable(oti->oti_thread->t_watchdog);
1008                         CDEBUG(D_QUOTA, "sleep for quota master\n");
1009                         l_wait_event(qctxt->lqc_wait_for_qmaster,
1010                                      check_qm(qctxt), &lwi);
1011                         CDEBUG(D_QUOTA, "wake up when quota master is back\n");
1012                         lc_watchdog_touch(oti->oti_thread->t_watchdog,
1013                                  GET_TIMEOUT(oti->oti_thread->t_svc));
1014                 } else {
1015                         spin_unlock(&qctxt->lqc_lock);
1016                 }
1017
1018                 RETURN(-EAGAIN);
1019         }
1020         imp = class_import_get(qctxt->lqc_import);
1021         spin_unlock(&qctxt->lqc_lock);
1022
1023         /* build dqacq/dqrel request */
1024         LASSERT(imp);
1025         size[1] = quota_get_qunit_data_size(imp->
1026                                             imp_connect_data.ocd_connect_flags);
1027
1028         req = ptlrpc_prep_req(imp, LUSTRE_MDS_VERSION, opc, 2,
1029                               size, NULL);
1030         if (!req) {
1031                 dqacq_completion(obd, qctxt, qdata, -ENOMEM, opc);
1032                 class_import_put(imp);
1033                 /* this is for qunit_get() */
1034                 qunit_put(qunit);
1035                 RETURN(-ENOMEM);
1036         }
1037
1038         rc = quota_copy_qdata(req, qdata, QUOTA_REQUEST, QUOTA_IMPORT);
1039         if (rc < 0) {
1040                 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
1041                 dqacq_completion(obd, qctxt, qdata, rc, opc);
1042                 class_import_put(imp);
1043                 /* this is for qunit_get() */
1044                 qunit_put(qunit);
1045                 RETURN(rc);
1046         }
1047         ptlrpc_req_set_repsize(req, 2, size);
1048         req->rq_no_resend = req->rq_no_delay = 1;
1049         class_import_put(imp);
1050
1051         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1052         aa = ptlrpc_req_async_args(req);
1053         aa->aa_ctxt = qctxt;
1054         aa->aa_qunit = qunit;
1055
1056         req->rq_interpret_reply = dqacq_interpret;
1057         ptlrpcd_add_req(req);
1058
1059         QDATA_DEBUG(qdata, "%s scheduled.\n",
1060                     opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
1061 wait_completion:
1062         if (wait && qunit) {
1063                 struct qunit_data *p = &qunit->lq_data;
1064
1065                 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
1066                 l_wait_event(qunit->lq_waitq, got_qunit(qunit, ismaster), &lwi);
1067                 /* rc = -EAGAIN, it means the quota master isn't ready yet
1068                  * rc = QUOTA_REQ_RETURNED, it means a quota req is finished;
1069                  * rc = -EDQUOT, it means out of quota
1070                  * rc = -EBUSY, it means recovery is happening
1071                  * other rc < 0, it means real errors, functions who call
1072                  * schedule_dqacq should take care of this */
1073                 spin_lock(&qunit->lq_lock);
1074                 rc = qunit->lq_rc;
1075                 spin_unlock(&qunit->lq_lock);
1076                 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting. (rc:%d)\n",
1077                        qunit, rc);
1078         }
1079
1080         qunit_put(qunit);
1081         do_gettimeofday(&work_end);
1082         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1083         if (opc == QUOTA_DQACQ)
1084                 lprocfs_counter_add(qctxt->lqc_stats,
1085                                     wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
1086                                     timediff);
1087         else
1088                 lprocfs_counter_add(qctxt->lqc_stats,
1089                                     wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
1090                                     timediff);
1091
1092         RETURN(rc);
1093 }
1094
1095 int
1096 qctxt_adjust_qunit(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
1097                    uid_t uid, gid_t gid, __u32 isblk, int wait,
1098                    struct obd_trans_info *oti)
1099 {
1100         int rc = 0, i = USRQUOTA;
1101         __u32 id[MAXQUOTAS] = { uid, gid };
1102         struct qunit_data qdata[MAXQUOTAS];
1103         ENTRY;
1104
1105         if (quota_is_set(obd, uid, gid, isblk ? QB_SET : QI_SET) == 0)
1106                 RETURN(0);
1107
1108         for (i = 0; i < MAXQUOTAS; i++) {
1109                 qdata[i].qd_id = id[i];
1110                 qdata[i].qd_flags = i;
1111                 if (isblk)
1112                         QDATA_SET_BLK(&qdata[i]);
1113                 qdata[i].qd_count = 0;
1114
1115                 rc = check_cur_qunit(obd, qctxt, &qdata[i]);
1116                 if (rc > 0) {
1117                         int opc;
1118                         /* need acquire or release */
1119                         opc = rc == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1120                         rc = schedule_dqacq(obd, qctxt, &qdata[i], opc,
1121                                             wait,oti);
1122                         if (rc < 0)
1123                                 RETURN(rc);
1124                 } else if (wait == 1) {
1125                         /* when wait equates 1, that means mds_quota_acquire
1126                          * or filter_quota_acquire is calling it. */
1127                         rc = qctxt_wait_pending_dqacq(qctxt, id[i], i, isblk);
1128                         if (rc < 0)
1129                                 RETURN(rc);
1130                 }
1131         }
1132
1133         RETURN(rc);
1134 }
1135
1136 int
1137 qctxt_wait_pending_dqacq(struct lustre_quota_ctxt *qctxt, unsigned int id,
1138                          unsigned short type, int isblk)
1139 {
1140         struct lustre_qunit *qunit = NULL;
1141         struct qunit_data qdata;
1142         struct timeval work_start;
1143         struct timeval work_end;
1144         long timediff;
1145         struct l_wait_info lwi = { 0 };
1146         int ismaster = is_master(NULL, qctxt, id, type);
1147         int rc = 0;
1148         ENTRY;
1149
1150         do_gettimeofday(&work_start);
1151         qdata.qd_id = id;
1152         qdata.qd_flags = type;
1153         if (isblk)
1154                 QDATA_SET_BLK(&qdata);
1155         qdata.qd_count = 0;
1156
1157         spin_lock(&qunit_hash_lock);
1158         qunit = dqacq_in_flight(qctxt, &qdata);
1159         spin_unlock(&qunit_hash_lock);
1160
1161         if (qunit) {
1162                 struct qunit_data *p = &qunit->lq_data;
1163
1164                 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
1165                 l_wait_event(qunit->lq_waitq, got_qunit(qunit, ismaster), &lwi);
1166                 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting. (rc:%d)\n",
1167                        qunit, qunit->lq_rc);
1168                 /* keep same as schedule_dqacq() b=17030 */
1169                 spin_lock(&qunit->lq_lock);
1170                 rc = qunit->lq_rc;
1171                 spin_unlock(&qunit->lq_lock);
1172                 /* this is for dqacq_in_flight() */
1173                 qunit_put(qunit);
1174                 do_gettimeofday(&work_end);
1175                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1176                 lprocfs_counter_add(qctxt->lqc_stats,
1177                                     isblk ? LQUOTA_WAIT_PENDING_BLK_QUOTA :
1178                                             LQUOTA_WAIT_PENDING_INO_QUOTA,
1179                                     timediff);
1180         } else {
1181                 do_gettimeofday(&work_end);
1182                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1183                 lprocfs_counter_add(qctxt->lqc_stats,
1184                                     isblk ? LQUOTA_NOWAIT_PENDING_BLK_QUOTA :
1185                                             LQUOTA_NOWAIT_PENDING_INO_QUOTA,
1186                                     timediff);
1187         }
1188
1189         RETURN(rc);
1190 }
1191
1192 int
1193 qctxt_init(struct obd_device *obd, dqacq_handler_t handler)
1194 {
1195         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
1196         struct super_block *sb = obd->u.obt.obt_sb;
1197         int rc = 0;
1198         ENTRY;
1199
1200         LASSERT(qctxt);
1201
1202         rc = ptlrpcd_addref();
1203         if (rc)
1204                 RETURN(rc);
1205
1206         cfs_waitq_init(&qctxt->lqc_wait_for_qmaster);
1207         cfs_waitq_init(&qctxt->lqc_lqs_waitq);
1208         atomic_set(&qctxt->lqc_lqs, 0);
1209         spin_lock_init(&qctxt->lqc_lock);
1210         spin_lock(&qctxt->lqc_lock);
1211         qctxt->lqc_handler = handler;
1212         qctxt->lqc_sb = sb;
1213         qctxt->lqc_import = NULL;
1214         qctxt->lqc_recovery = 0;
1215         qctxt->lqc_switch_qs = 1; /* Change qunit size in default setting */
1216         qctxt->lqc_valid = 1;
1217         qctxt->lqc_cqs_boundary_factor = 4;
1218         qctxt->lqc_cqs_least_bunit = PTLRPC_MAX_BRW_SIZE;
1219         qctxt->lqc_cqs_least_iunit = 2;
1220         qctxt->lqc_cqs_qs_factor = 2;
1221         qctxt->lqc_flags = 0;
1222         QUOTA_MASTER_UNREADY(qctxt);
1223         qctxt->lqc_bunit_sz = default_bunit_sz;
1224         qctxt->lqc_btune_sz = default_bunit_sz / 100 * default_btune_ratio;
1225         qctxt->lqc_iunit_sz = default_iunit_sz;
1226         qctxt->lqc_itune_sz = default_iunit_sz * default_itune_ratio / 100;
1227         qctxt->lqc_switch_seconds = 300; /* enlarging will wait 5 minutes
1228                                           * after the last shrinking */
1229         qctxt->lqc_sync_blk = 0;
1230         spin_unlock(&qctxt->lqc_lock);
1231
1232         qctxt->lqc_lqs_hash = lustre_hash_init("LQS_HASH",
1233                                                HASH_LQS_CUR_BITS,
1234                                                HASH_LQS_MAX_BITS,
1235                                                &lqs_hash_ops, 0);
1236         if (!qctxt->lqc_lqs_hash)
1237                 CERROR("%s: initialize hash lqs failed\n", obd->obd_name);
1238
1239 #ifdef LPROCFS
1240         if (lquota_proc_setup(obd, is_master(obd, qctxt, 0, 0)))
1241                 CERROR("%s: initialize proc failed\n", obd->obd_name);
1242 #endif
1243
1244         RETURN(rc);
1245 }
1246
1247 static int check_lqs(struct lustre_quota_ctxt *qctxt)
1248 {
1249         int rc;
1250         ENTRY;
1251
1252         rc = !atomic_read(&qctxt->lqc_lqs);
1253
1254         RETURN(rc);
1255 }
1256
1257
1258 void hash_put_lqs(void *obj, void *data)
1259 {
1260         lqs_putref((struct lustre_qunit_size *)obj);
1261 }
1262
1263 void qctxt_cleanup(struct lustre_quota_ctxt *qctxt, int force)
1264 {
1265         struct lustre_qunit *qunit, *tmp;
1266         struct list_head tmp_list;
1267         struct l_wait_info lwi = { 0 };
1268         int i;
1269         ENTRY;
1270
1271         INIT_LIST_HEAD(&tmp_list);
1272
1273         spin_lock(&qctxt->lqc_lock);
1274         qctxt->lqc_valid = 0;
1275         spin_unlock(&qctxt->lqc_lock);
1276
1277         spin_lock(&qunit_hash_lock);
1278         for (i = 0; i < NR_DQHASH; i++) {
1279                 list_for_each_entry_safe(qunit, tmp, &qunit_hash[i], lq_hash) {
1280                         if (qunit->lq_ctxt != qctxt)
1281                                 continue;
1282                         remove_qunit_nolock(qunit);
1283                         list_add(&qunit->lq_hash, &tmp_list);
1284                 }
1285         }
1286         spin_unlock(&qunit_hash_lock);
1287
1288         list_for_each_entry_safe(qunit, tmp, &tmp_list, lq_hash) {
1289                 list_del_init(&qunit->lq_hash);
1290                 compute_lqs_after_removing_qunit(qunit);
1291
1292                 /* wake up all waiters */
1293                 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, 0);
1294                 wake_up(&qunit->lq_waitq);
1295                 qunit_put(qunit);
1296         }
1297
1298         /* after qctxt_cleanup, qctxt might be freed, then check_qm() is
1299          * unpredicted. So we must wait until lqc_wait_for_qmaster is empty */
1300         while (cfs_waitq_active(&qctxt->lqc_wait_for_qmaster)) {
1301                 cfs_waitq_signal(&qctxt->lqc_wait_for_qmaster);
1302                 cfs_schedule_timeout(CFS_TASK_INTERRUPTIBLE,
1303                                      cfs_time_seconds(1));
1304         }
1305
1306         lustre_hash_for_each_safe(qctxt->lqc_lqs_hash, hash_put_lqs, NULL);
1307         l_wait_event(qctxt->lqc_lqs_waitq, check_lqs(qctxt), &lwi);
1308         lustre_hash_exit(qctxt->lqc_lqs_hash);
1309
1310         ptlrpcd_decref();
1311
1312 #ifdef LPROCFS
1313         if (lquota_proc_cleanup(qctxt))
1314                 CERROR("cleanup proc error!\n");
1315 #endif
1316
1317         EXIT;
1318 }
1319
1320 struct qslave_recov_thread_data {
1321         struct obd_device *obd;
1322         struct lustre_quota_ctxt *qctxt;
1323         struct completion comp;
1324 };
1325
1326 /* FIXME only recovery block quota by now */
1327 static int qslave_recovery_main(void *arg)
1328 {
1329         struct qslave_recov_thread_data *data = arg;
1330         struct obd_device *obd = data->obd;
1331         struct lustre_quota_ctxt *qctxt = data->qctxt;
1332         unsigned int type;
1333         int rc = 0;
1334         ENTRY;
1335
1336         cfs_daemonize_ctxt("qslave_recovd");
1337
1338         complete(&data->comp);
1339
1340         if (qctxt->lqc_recovery)
1341                 RETURN(0);
1342         qctxt->lqc_recovery = 1;
1343
1344         for (type = USRQUOTA; type < MAXQUOTAS; type++) {
1345                 struct qunit_data qdata;
1346                 struct quota_info *dqopt = sb_dqopt(qctxt->lqc_sb);
1347                 struct list_head id_list;
1348                 struct dquot_id *dqid, *tmp;
1349                 int ret;
1350
1351                 LOCK_DQONOFF_MUTEX(dqopt);
1352                 if (!ll_sb_has_quota_active(qctxt->lqc_sb, type)) {
1353                         UNLOCK_DQONOFF_MUTEX(dqopt);
1354                         break;
1355                 }
1356
1357                 LASSERT(dqopt->files[type] != NULL);
1358                 INIT_LIST_HEAD(&id_list);
1359 #ifndef KERNEL_SUPPORTS_QUOTA_READ
1360                 rc = fsfilt_qids(obd, dqopt->files[type], NULL, type, &id_list);
1361 #else
1362                 rc = fsfilt_qids(obd, NULL, dqopt->files[type], type, &id_list);
1363 #endif
1364                 UNLOCK_DQONOFF_MUTEX(dqopt);
1365                 if (rc)
1366                         CERROR("Get ids from quota file failed. (rc:%d)\n", rc);
1367
1368                 list_for_each_entry_safe(dqid, tmp, &id_list, di_link) {
1369                         list_del_init(&dqid->di_link);
1370                         /* skip slave recovery on itself */
1371                         if (is_master(obd, qctxt, dqid->di_id, type))
1372                                 goto free;
1373                         if (rc && rc != -EBUSY)
1374                                 goto free;
1375
1376                         qdata.qd_id = dqid->di_id;
1377                         qdata.qd_flags = type;
1378                         QDATA_SET_BLK(&qdata);
1379                         qdata.qd_count = 0;
1380
1381                         ret = check_cur_qunit(obd, qctxt, &qdata);
1382                         if (ret > 0) {
1383                                 int opc;
1384                                 opc = ret == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1385                                 rc = schedule_dqacq(obd, qctxt, &qdata, opc,
1386                                                     0, NULL);
1387                                 if (rc == -EDQUOT)
1388                                         rc = 0;
1389                         } else {
1390                                 rc = 0;
1391                         }
1392
1393                         if (rc && rc != -EBUSY)
1394                                 CERROR("qslave recovery failed! (id:%d type:%d "
1395                                        " rc:%d)\n", dqid->di_id, type, rc);
1396 free:
1397                         OBD_FREE_PTR(dqid);
1398                 }
1399         }
1400
1401         qctxt->lqc_recovery = 0;
1402         RETURN(rc);
1403 }
1404
1405 void
1406 qslave_start_recovery(struct obd_device *obd, struct lustre_quota_ctxt *qctxt)
1407 {
1408         struct qslave_recov_thread_data data;
1409         int rc;
1410         ENTRY;
1411
1412         if (!ll_sb_any_quota_active(qctxt->lqc_sb))
1413                 goto exit;
1414
1415         data.obd = obd;
1416         data.qctxt = qctxt;
1417         init_completion(&data.comp);
1418
1419         rc = kernel_thread(qslave_recovery_main, &data, CLONE_VM|CLONE_FILES);
1420         if (rc < 0) {
1421                 CERROR("Cannot start quota recovery thread: rc %d\n", rc);
1422                 goto exit;
1423         }
1424         wait_for_completion(&data.comp);
1425 exit:
1426         EXIT;
1427 }
1428
1429
1430 /*
1431  * lqs<->qctxt hash operations
1432  */
1433
1434 /* string hashing using djb2 hash algorithm */
1435 static unsigned
1436 lqs_hash(lustre_hash_t *lh, void *key, unsigned mask)
1437 {
1438         struct quota_adjust_qunit *lqs_key;
1439         unsigned hash;
1440         ENTRY;
1441
1442         LASSERT(key);
1443         lqs_key = (struct quota_adjust_qunit *)key;
1444         hash = (QAQ_IS_GRP(lqs_key) ? 5381 : 5387) * lqs_key->qaq_id;
1445
1446         RETURN(hash & mask);
1447 }
1448
1449 static int
1450 lqs_compare(void *key, struct hlist_node *hnode)
1451 {
1452         struct lustre_qunit_size *q;
1453         int rc;
1454         ENTRY;
1455
1456         LASSERT(key);
1457         q = hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1458
1459         spin_lock(&q->lqs_lock);
1460         rc = (q->lqs_key == *((unsigned long long *)key));
1461         spin_unlock(&q->lqs_lock);
1462
1463         RETURN(rc);
1464 }
1465
1466 static void *
1467 lqs_get(struct hlist_node *hnode)
1468 {
1469         struct lustre_qunit_size *q =
1470             hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1471         ENTRY;
1472
1473         lqs_getref(q);
1474
1475         RETURN(q);
1476 }
1477
1478 static void *
1479 lqs_put(struct hlist_node *hnode)
1480 {
1481         struct lustre_qunit_size *q =
1482             hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1483         ENTRY;
1484
1485         lqs_putref(q);
1486
1487         RETURN(q);
1488 }
1489
1490 static void
1491 lqs_exit(struct hlist_node *hnode)
1492 {
1493         struct lustre_qunit_size *q;
1494         ENTRY;
1495
1496         q = hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1497         /*
1498          * Nothing should be left. User of lqs put it and
1499          * lqs also was deleted from table by this time
1500          * so we should have 0 refs.
1501          */
1502         LASSERTF(atomic_read(&q->lqs_refcount) == 0,
1503                  "Busy lqs %p with %d refs\n", q,
1504                  atomic_read(&q->lqs_refcount));
1505         OBD_FREE_PTR(q);
1506         EXIT;
1507 }
1508
1509 static lustre_hash_ops_t lqs_hash_ops = {
1510         .lh_hash    = lqs_hash,
1511         .lh_compare = lqs_compare,
1512         .lh_get     = lqs_get,
1513         .lh_put     = lqs_put,
1514         .lh_exit    = lqs_exit
1515 };
1516 #endif /* HAVE_QUOTA_SUPPORT */