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 (ret == 2 && pending_write) {
 341                         CDEBUG(D_QUOTA, "delay quota release\n");
 342                         ret = 0;
 343                 }
 344         }
 345         CDEBUG(D_QUOTA, "type: %c, limit: "LPU64", usage: "LPU64
 346                ", pending_write: "LPU64", record: "LPD64
 347                ", qunit_sz: %lu, tune_sz: %lu, ret: %d.\n",
 348                QDATA_IS_BLK(qdata) ? 'b' : 'i', limit, usage, pending_write,
 349                record, qunit_sz, tune_sz, ret);
 350         LASSERT(ret == 0 || qdata->qd_count);
 351
 352         spin_unlock(&lqs->lqs_lock);
 353         lqs_putref(lqs);
 354
 355         EXIT;
 356  out:
 357         OBD_FREE_PTR(qctl);
 358         return ret;
 359 }
 360
 361 /**
 362  * Compute the remaining quota for certain gid or uid b=11693
 363  */
 364 int compute_remquota(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 365                      struct qunit_data *qdata, int isblk)
 366 {
 367         struct super_block *sb = qctxt->lqc_sb;
 368         __u64 usage, limit;
 369         struct obd_quotactl *qctl;
 370         int ret = QUOTA_RET_OK;
 371         ENTRY;
 372
 373         if (!ll_sb_any_quota_active(sb))
 374                 RETURN(QUOTA_RET_NOQUOTA);
 375
 376         /* ignore root user */
 377         if (qdata->qd_id == 0 && QDATA_IS_GRP(qdata) == USRQUOTA)
 378                 RETURN(QUOTA_RET_NOLIMIT);
 379
 380         OBD_ALLOC_PTR(qctl);
 381         if (qctl == NULL)
 382                 RETURN(-ENOMEM);
 383
 384         /* get fs quota usage & limit */
 385         qctl->qc_cmd = Q_GETQUOTA;
 386         qctl->qc_id = qdata->qd_id;
 387         qctl->qc_type = QDATA_IS_GRP(qdata);
 388         ret = fsfilt_quotactl(obd, sb, qctl);
 389         if (ret) {
 390                 if (ret == -ESRCH)      /* no limit */
 391                         ret = QUOTA_RET_NOLIMIT;
 392                 else
 393                         CDEBUG(D_QUOTA, "can't get fs quota usage! (rc:%d)",
 394                                ret);
 395                 GOTO(out, ret);
 396         }
 397
 398         usage = isblk ? qctl->qc_dqblk.dqb_curspace :
 399                 qctl->qc_dqblk.dqb_curinodes;
 400         limit = isblk ? qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS :
 401                 qctl->qc_dqblk.dqb_ihardlimit;
 402         if (!limit){            /* no limit */
 403                 ret = QUOTA_RET_NOLIMIT;
 404                 GOTO(out, ret);
 405         }
 406
 407         if (limit >= usage)
 408                 qdata->qd_count = limit - usage;
 409         else
 410                 qdata->qd_count = 0;
 411         EXIT;
 412 out:
 413         OBD_FREE_PTR(qctl);
 414         return ret;
 415 }
 416
 417 static struct lustre_qunit *alloc_qunit(struct lustre_quota_ctxt *qctxt,
 418                                         struct qunit_data *qdata, int opc)
 419 {
 420         struct lustre_qunit *qunit = NULL;
 421         ENTRY;
 422
 423         OBD_SLAB_ALLOC_PTR_GFP(qunit, qunit_cachep, CFS_ALLOC_IO);
 424         if (qunit == NULL)
 425                 RETURN(NULL);
 426
 427         CFS_INIT_LIST_HEAD(&qunit->lq_hash);
 428         init_waitqueue_head(&qunit->lq_waitq);
 429         atomic_set(&qunit->lq_refcnt, 1);
 430         qunit->lq_ctxt = qctxt;
 431         memcpy(&qunit->lq_data, qdata, sizeof(*qdata));
 432         qunit->lq_opc = opc;
 433         qunit->lq_lock = SPIN_LOCK_UNLOCKED;
 434         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_CREATED, 0);
 435         qunit->lq_owner = cfs_curproc_pid();
 436         RETURN(qunit);
 437 }
 438
 439 static inline void free_qunit(struct lustre_qunit *qunit)
 440 {
 441         OBD_SLAB_FREE(qunit, qunit_cachep, sizeof(*qunit));
 442 }
 443
 444 static inline void qunit_get(struct lustre_qunit *qunit)
 445 {
 446         atomic_inc(&qunit->lq_refcnt);
 447 }
 448
 449 static void qunit_put(struct lustre_qunit *qunit)
 450 {
 451         LASSERT(atomic_read(&qunit->lq_refcnt));
 452         if (atomic_dec_and_test(&qunit->lq_refcnt))
 453                 free_qunit(qunit);
 454 }
 455
 456 /* caller must hold qunit_hash_lock and release ref of qunit after using it */
 457 static struct lustre_qunit *dqacq_in_flight(struct lustre_quota_ctxt *qctxt,
 458                                             struct qunit_data *qdata)
 459 {
 460         unsigned int hashent = qunit_hashfn(qctxt, qdata);
 461         struct lustre_qunit *qunit;
 462         ENTRY;
 463
 464         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
 465         qunit = find_qunit(hashent, qctxt, qdata);
 466         if (qunit)
 467                 qunit_get(qunit);
 468         RETURN(qunit);
 469 }
 470
 471 static void
 472 insert_qunit_nolock(struct lustre_quota_ctxt *qctxt, struct lustre_qunit *qunit)
 473 {
 474         struct list_head *head;
 475
 476         LASSERT(list_empty(&qunit->lq_hash));
 477         qunit_get(qunit);
 478         head = qunit_hash + qunit_hashfn(qctxt, &qunit->lq_data);
 479         list_add(&qunit->lq_hash, head);
 480         QUNIT_SET_STATE(qunit, QUNIT_IN_HASH);
 481 }
 482
 483 static void compute_lqs_after_removing_qunit(struct lustre_qunit *qunit)
 484 {
 485         struct lustre_qunit_size *lqs;
 486
 487         lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(&qunit->lq_data),
 488                                        qunit->lq_data.qd_id),
 489                                qunit->lq_ctxt, 0);
 490         if (lqs && !IS_ERR(lqs)) {
 491                 spin_lock(&lqs->lqs_lock);
 492                 if (qunit->lq_opc == QUOTA_DQACQ)
 493                         quota_compute_lqs(&qunit->lq_data, lqs, 0, 1);
 494                 if (qunit->lq_opc == QUOTA_DQREL)
 495                         quota_compute_lqs(&qunit->lq_data, lqs, 0, 0);
 496                 spin_unlock(&lqs->lqs_lock);
 497                 /* this is for quota_search_lqs */
 498                 lqs_putref(lqs);
 499                 /* this is for schedule_dqacq */
 500                 lqs_putref(lqs);
 501         }
 502 }
 503
 504 static void remove_qunit_nolock(struct lustre_qunit *qunit)
 505 {
 506         LASSERT(!list_empty(&qunit->lq_hash));
 507         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
 508
 509         list_del_init(&qunit->lq_hash);
 510         QUNIT_SET_STATE(qunit, QUNIT_RM_FROM_HASH);
 511         qunit_put(qunit);
 512 }
 513
 514 void* quota_barrier(struct lustre_quota_ctxt *qctxt,
 515                     struct obd_quotactl *oqctl, int isblk)
 516 {
 517         struct lustre_qunit *qunit, *find_qunit;
 518         int cycle = 1;
 519
 520         OBD_SLAB_ALLOC(qunit, qunit_cachep, CFS_ALLOC_IO, sizeof(*qunit));
 521         if (qunit == NULL) {
 522                 CERROR("locating %sunit failed for %sid %u\n",
 523                        isblk ? "b" : "i", oqctl->qc_type ? "g" : "u",
 524                        oqctl->qc_id);
 525                 qctxt_wait_pending_dqacq(qctxt, oqctl->qc_id,
 526                                          oqctl->qc_type, isblk);
 527                 return NULL;
 528         }
 529
 530         INIT_LIST_HEAD(&qunit->lq_hash);
 531         qunit->lq_lock = SPIN_LOCK_UNLOCKED;
 532         init_waitqueue_head(&qunit->lq_waitq);
 533         atomic_set(&qunit->lq_refcnt, 1);
 534         qunit->lq_ctxt = qctxt;
 535         qunit->lq_data.qd_id = oqctl->qc_id;
 536         qunit->lq_data.qd_flags =  oqctl->qc_type;
 537         if (isblk)
 538                 QDATA_SET_BLK(&qunit->lq_data);
 539         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_CREATED, 0);
 540         /* it means it is only an invalid qunit for barrier */
 541         qunit->lq_opc = QUOTA_LAST_OPC;
 542
 543         while (1) {
 544                 spin_lock(&qunit_hash_lock);
 545                 find_qunit = dqacq_in_flight(qctxt, &qunit->lq_data);
 546                 if (find_qunit) {
 547                         spin_unlock(&qunit_hash_lock);
 548                         qunit_put(find_qunit);
 549                         qctxt_wait_pending_dqacq(qctxt, oqctl->qc_id,
 550                                                  oqctl->qc_type, isblk);
 551                         CDEBUG(D_QUOTA, "cycle=%d\n", cycle++);
 552                         continue;
 553                 }
 554                 break;
 555         }
 556         insert_qunit_nolock(qctxt, qunit);
 557         spin_unlock(&qunit_hash_lock);
 558         return qunit;
 559 }
 560
 561 void quota_unbarrier(void *handle)
 562 {
 563         struct lustre_qunit *qunit = (struct lustre_qunit *)handle;
 564
 565         if (qunit == NULL) {
 566                 CERROR("handle is NULL\n");
 567                 return;
 568         }
 569
 570         LASSERT(qunit->lq_opc == QUOTA_LAST_OPC);
 571         spin_lock(&qunit_hash_lock);
 572         remove_qunit_nolock(qunit);
 573         spin_unlock(&qunit_hash_lock);
 574         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, QUOTA_REQ_RETURNED);
 575         wake_up(&qunit->lq_waitq);
 576         qunit_put(qunit);
 577 }
 578
 579 #define INC_QLIMIT(limit, count) (limit == MIN_QLIMIT) ? \
 580                                  (limit = count) : (limit += count)
 581
 582
 583 static inline int is_master(struct lustre_quota_ctxt *qctxt)
 584 {
 585         return qctxt->lqc_handler ? 1 : 0;
 586 }
 587
 588 static int
 589 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 590                struct qunit_data *qdata, int opc, int wait,
 591                struct obd_trans_info *oti);
 592
 593 static inline void qdata_to_oqaq(struct qunit_data *qdata,
 594                                  struct quota_adjust_qunit *oqaq)
 595 {
 596         LASSERT(qdata);
 597         LASSERT(oqaq);
 598
 599         oqaq->qaq_flags = qdata->qd_flags;
 600         oqaq->qaq_id    = qdata->qd_id;
 601         if (QDATA_IS_ADJBLK(qdata))
 602                 oqaq->qaq_bunit_sz = qdata->qd_qunit;
 603         if (QDATA_IS_ADJINO(qdata))
 604                 oqaq->qaq_iunit_sz = qdata->qd_qunit;
 605 }
 606
 607 static int
 608 dqacq_completion(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 609                  struct qunit_data *qdata, int rc, int opc)
 610 {
 611         struct lustre_qunit *qunit = NULL;
 612         struct super_block *sb = qctxt->lqc_sb;
 613         int err = 0;
 614         struct quota_adjust_qunit *oqaq = NULL;
 615         int rc1 = 0;
 616         ENTRY;
 617
 618         LASSERT(qdata);
 619         QDATA_DEBUG(qdata, "obd(%s): complete %s quota req\n",
 620                     obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
 621
 622         /* do it only when a releasing quota req more than 5MB b=18491 */
 623         if (opc == QUOTA_DQREL && qdata->qd_count >= 5242880)
 624                 OBD_FAIL_TIMEOUT(OBD_FAIL_QUOTA_DELAY_REL, 5);
 625
 626         /* update local operational quota file */
 627         if (rc == 0) {
 628                 __u64 count = QUSG(qdata->qd_count, QDATA_IS_BLK(qdata));
 629                 struct obd_quotactl *qctl;
 630                 __u64 *hardlimit;
 631
 632                 OBD_ALLOC_PTR(qctl);
 633                 if (qctl == NULL)
 634                         GOTO(out, err = -ENOMEM);
 635
 636                 /* acq/rel qunit for specified uid/gid is serialized,
 637                  * so there is no race between get fs quota limit and
 638                  * set fs quota limit */
 639                 qctl->qc_cmd = Q_GETQUOTA;
 640                 qctl->qc_id = qdata->qd_id;
 641                 qctl->qc_type = QDATA_IS_GRP(qdata);
 642                 err = fsfilt_quotactl(obd, sb, qctl);
 643                 if (err) {
 644                         CERROR("error get quota fs limit! (rc:%d)\n", err);
 645                         GOTO(out_mem, err);
 646                 }
 647
 648                 if (QDATA_IS_BLK(qdata)) {
 649                         qctl->qc_dqblk.dqb_valid = QIF_BLIMITS;
 650                         hardlimit = &qctl->qc_dqblk.dqb_bhardlimit;
 651                 } else {
 652                         qctl->qc_dqblk.dqb_valid = QIF_ILIMITS;
 653                         hardlimit = &qctl->qc_dqblk.dqb_ihardlimit;
 654                 }
 655
 656                 CDEBUG(D_QUOTA, "hardlimt: "LPU64"\n", *hardlimit);
 657
 658                 if (*hardlimit == 0)
 659                         goto out_mem;
 660
 661                 switch (opc) {
 662                 case QUOTA_DQACQ:
 663                         INC_QLIMIT(*hardlimit, count);
 664                         break;
 665                 case QUOTA_DQREL:
 666                         LASSERTF(count < *hardlimit,
 667                                  "id(%u) flag(%u) type(%c) isblk(%c) "
 668                                  "count("LPU64") qd_qunit("LPU64") "
 669                                  "hardlimit("LPU64").\n",
 670                                  qdata->qd_id, qdata->qd_flags,
 671                                  QDATA_IS_GRP(qdata) ? 'g' : 'u',
 672                                  QDATA_IS_BLK(qdata) ? 'b': 'i',
 673                                  qdata->qd_count, qdata->qd_qunit, *hardlimit);
 674                         *hardlimit -= count;
 675                         break;
 676                 default:
 677                         LBUG();
 678                 }
 679
 680                 /* clear quota limit */
 681                 if (count == 0)
 682                         *hardlimit = 0;
 683
 684                 qctl->qc_cmd = Q_SETQUOTA;
 685                 err = fsfilt_quotactl(obd, sb, qctl);
 686                 if (err)
 687                         CERROR("error set quota fs limit! (rc:%d)\n", err);
 688
 689                 QDATA_DEBUG(qdata, "%s completion\n",
 690                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 691 out_mem:
 692                 OBD_FREE_PTR(qctl);
 693         } else if (rc == -EDQUOT) {
 694                 QDATA_DEBUG(qdata, "acquire qunit got EDQUOT.\n");
 695         } else if (rc == -EBUSY) {
 696                 QDATA_DEBUG(qdata, "it's is recovering, got EBUSY.\n");
 697         } else {
 698                 CERROR("acquire qunit got error! (rc:%d)\n", rc);
 699         }
 700 out:
 701         /* remove the qunit from hash */
 702         spin_lock(&qunit_hash_lock);
 703
 704         qunit = dqacq_in_flight(qctxt, qdata);
 705         /* this qunit has been removed by qctxt_cleanup() */
 706         if (!qunit) {
 707                 spin_unlock(&qunit_hash_lock);
 708                 QDATA_DEBUG(qdata, "%s is discarded because qunit isn't found\n",
 709                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 710                 RETURN(err);
 711         }
 712
 713         LASSERT(opc == qunit->lq_opc);
 714         /* remove this qunit from lq_hash so that new processes cannot be added
 715          * to qunit->lq_waiters */
 716         remove_qunit_nolock(qunit);
 717         spin_unlock(&qunit_hash_lock);
 718
 719         compute_lqs_after_removing_qunit(qunit);
 720
 721         if (rc == 0)
 722                 rc = QUOTA_REQ_RETURNED;
 723         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, rc);
 724         /* wake up all waiters */
 725         wake_up_all(&qunit->lq_waitq);
 726
 727         /* this is for dqacq_in_flight() */
 728         qunit_put(qunit);
 729         if (rc < 0 && rc != -EDQUOT)
 730                 GOTO(out1, err);
 731
 732         /* don't reschedule in such cases:
 733          *   - acq/rel failure and qunit isn't changed,
 734          *     but not for quota recovery.
 735          *   - local dqacq/dqrel.
 736          *   - local disk io failure.
 737          */
 738          OBD_ALLOC_PTR(oqaq);
 739          if (!oqaq)
 740                  GOTO(out1, err = -ENOMEM);
 741          qdata_to_oqaq(qdata, oqaq);
 742          /* adjust the qunit size in slaves */
 743          rc1 = quota_adjust_slave_lqs(oqaq, qctxt);
 744          OBD_FREE_PTR(oqaq);
 745          if (rc1 < 0) {
 746                  CERROR("adjust slave's qunit size failed!(rc:%d)\n", rc1);
 747                  GOTO(out1, err = rc1);
 748          }
 749          if (err || (rc < 0 && rc != -EBUSY && rc1 == 0) || is_master(qctxt))
 750                  GOTO(out1, err);
 751
 752          if (opc == QUOTA_DQREL && qdata->qd_count >= 5242880 &&
 753              OBD_FAIL_CHECK(OBD_FAIL_QUOTA_DELAY_REL))
 754                  GOTO(out1, err);
 755
 756         /* reschedule another dqacq/dqrel if needed */
 757         qdata->qd_count = 0;
 758         qdata->qd_flags &= LQUOTA_QUNIT_FLAGS;
 759         rc1 = check_cur_qunit(obd, qctxt, qdata);
 760         if (rc1 > 0) {
 761                 int opc;
 762                 opc = rc1 == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
 763                 rc1 = schedule_dqacq(obd, qctxt, qdata, opc, 0, NULL);
 764                 QDATA_DEBUG(qdata, "reschedudle opc(%d) rc(%d)\n", opc, rc1);
 765         }
 766  out1:
 767         /* this is for alloc_qunit() */
 768         qunit_put(qunit);
 769         RETURN(err);
 770 }
 771
 772 struct dqacq_async_args {
 773         struct lustre_quota_ctxt *aa_ctxt;
 774         struct lustre_qunit *aa_qunit;
 775 };
 776
 777 static int dqacq_interpret(const struct lu_env *env,
 778                            struct ptlrpc_request *req, void *data, int rc)
 779 {
 780         struct dqacq_async_args *aa = (struct dqacq_async_args *)data;
 781         struct lustre_quota_ctxt *qctxt = aa->aa_ctxt;
 782         struct obd_device_target *obt = qctxt->lqc_obt;
 783         struct lustre_qunit *qunit = aa->aa_qunit;
 784         struct obd_device *obd = req->rq_import->imp_obd;
 785         struct qunit_data *qdata = NULL;
 786         ENTRY;
 787
 788         LASSERT(req);
 789         LASSERT(req->rq_import);
 790
 791         down_read(&obt->obt_rwsem);
 792         /* if a quota req timeouts or is dropped, we should update quota
 793          * statistics which will be handled in dqacq_completion. And in
 794          * this situation we should get qdata from request instead of
 795          * reply */
 796         qdata = quota_get_qdata(req, (rc != 0) ? QUOTA_REQUEST : QUOTA_REPLY,
 797                                 QUOTA_IMPORT);
 798         if (IS_ERR(qdata)) {
 799                 rc = PTR_ERR(qdata);
 800                 DEBUG_REQ(D_ERROR, req,
 801                           "error unpacking qunit_data(rc: %ld)\n",
 802                           PTR_ERR(qdata));
 803                 qdata = &qunit->lq_data;
 804         }
 805
 806         QDATA_DEBUG(qdata, "qdata: interpret rc(%d).\n", rc);
 807         QDATA_DEBUG((&qunit->lq_data), "lq_data: \n");
 808
 809         if (qdata->qd_id != qunit->lq_data.qd_id ||
 810             OBD_FAIL_CHECK(OBD_FAIL_QUOTA_RET_QDATA)) {
 811                 CERROR("the returned qd_id isn't expected!"
 812                        "(qdata: %u, lq_data: %u)\n", qdata->qd_id,
 813                        qunit->lq_data.qd_id);
 814                 qdata->qd_id = qunit->lq_data.qd_id;
 815                 rc = -EPROTO;
 816         }
 817         if (QDATA_IS_GRP(qdata) != QDATA_IS_GRP(&qunit->lq_data)) {
 818                 CERROR("the returned grp/usr isn't expected!"
 819                        "(qdata: %u, lq_data: %u)\n", qdata->qd_flags,
 820                        qunit->lq_data.qd_flags);
 821                 if (QDATA_IS_GRP(&qunit->lq_data))
 822                         QDATA_SET_GRP(qdata);
 823                 else
 824                         QDATA_CLR_GRP(qdata);
 825                 rc = -EPROTO;
 826         }
 827         if (qdata->qd_count > qunit->lq_data.qd_count) {
 828                 CERROR("the returned qd_count isn't expected!"
 829                        "(qdata: "LPU64", lq_data: "LPU64")\n", qdata->qd_count,
 830                        qunit->lq_data.qd_count);
 831                 rc = -EPROTO;
 832         }
 833
 834         if (unlikely(rc == -ESRCH))
 835                 CERROR("quota for %s has been enabled by master, but disabled "
 836                        "by slave.\n", QDATA_IS_GRP(qdata) ? "group" : "user");
 837
 838         rc = dqacq_completion(obd, qctxt, qdata, rc,
 839                               lustre_msg_get_opc(req->rq_reqmsg));
 840
 841         up_read(&obt->obt_rwsem);
 842         RETURN(rc);
 843 }
 844
 845 /**
 846  * check if quota master is online
 847  */
 848 int check_qm(struct lustre_quota_ctxt *qctxt)
 849 {
 850         int rc;
 851         ENTRY;
 852
 853         spin_lock(&qctxt->lqc_lock);
 854         /* quit waiting when mds is back or qctxt is cleaned up */
 855         rc = qctxt->lqc_import || !qctxt->lqc_valid;
 856         spin_unlock(&qctxt->lqc_lock);
 857
 858         RETURN(rc);
 859 }
 860
 861 /* wake up all waiting threads when lqc_import is NULL */
 862 void dqacq_interrupt(struct lustre_quota_ctxt *qctxt)
 863 {
 864         struct lustre_qunit *qunit, *tmp;
 865         int i;
 866         ENTRY;
 867
 868         spin_lock(&qunit_hash_lock);
 869         for (i = 0; i < NR_DQHASH; i++) {
 870                 list_for_each_entry_safe(qunit, tmp, &qunit_hash[i], lq_hash) {
 871                         if (qunit->lq_ctxt != qctxt)
 872                                 continue;
 873
 874                         /* Wake up all waiters. Do not change lq_state.
 875                          * The waiters will check lq_rc which is kept as 0
 876                          * if no others change it, then the waiters will return
 877                          * -EAGAIN to caller who can perform related quota
 878                          * acq/rel if necessary. */
 879                         wake_up_all(&qunit->lq_waitq);
 880                 }
 881         }
 882         spin_unlock(&qunit_hash_lock);
 883         EXIT;
 884 }
 885
 886 static int got_qunit(struct lustre_qunit *qunit, int is_master)
 887 {
 888         struct lustre_quota_ctxt *qctxt = qunit->lq_ctxt;
 889         int rc = 0;
 890         ENTRY;
 891
 892         spin_lock(&qunit->lq_lock);
 893         switch (qunit->lq_state) {
 894         case QUNIT_IN_HASH:
 895         case QUNIT_RM_FROM_HASH:
 896                 break;
 897         case QUNIT_FINISHED:
 898                 rc = 1;
 899                 break;
 900         default:
 901                 CERROR("invalid qunit state %d\n", qunit->lq_state);
 902         }
 903         spin_unlock(&qunit->lq_lock);
 904
 905         if (!rc) {
 906                 spin_lock(&qctxt->lqc_lock);
 907                 rc = !qctxt->lqc_valid;
 908                 if (!is_master)
 909                         rc |= !qctxt->lqc_import;
 910                 spin_unlock(&qctxt->lqc_lock);
 911         }
 912
 913         RETURN(rc);
 914 }
 915
 916 static int
 917 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 918                struct qunit_data *qdata, int opc, int wait,
 919                struct obd_trans_info *oti)
 920 {
 921         struct lustre_qunit *qunit, *empty;
 922         struct l_wait_info lwi = { 0 };
 923         struct ptlrpc_request *req;
 924         struct dqacq_async_args *aa;
 925         struct obd_import *imp = NULL;
 926         struct lustre_qunit_size *lqs = NULL;
 927         struct timeval work_start;
 928         struct timeval work_end;
 929         long timediff;
 930         int rc = 0;
 931         ENTRY;
 932
 933         LASSERT(opc == QUOTA_DQACQ || opc == QUOTA_DQREL);
 934         do_gettimeofday(&work_start);
 935         if ((empty = alloc_qunit(qctxt, qdata, opc)) == NULL)
 936                 RETURN(-ENOMEM);
 937
 938         spin_lock(&qunit_hash_lock);
 939         qunit = dqacq_in_flight(qctxt, qdata);
 940         if (qunit) {
 941                 spin_unlock(&qunit_hash_lock);
 942                 qunit_put(empty);
 943
 944                 goto wait_completion;
 945         }
 946         qunit = empty;
 947         qunit_get(qunit);
 948         insert_qunit_nolock(qctxt, qunit);
 949         spin_unlock(&qunit_hash_lock);
 950
 951         lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(qdata), qdata->qd_id),
 952                                qctxt, 0);
 953         if (lqs && !IS_ERR(lqs)) {
 954                 spin_lock(&lqs->lqs_lock);
 955                 quota_compute_lqs(qdata, lqs, 1, (opc == QUOTA_DQACQ) ? 1 : 0);
 956                 /* when this qdata returned from mds, it will call lqs_putref */
 957                 lqs_getref(lqs);
 958                 spin_unlock(&lqs->lqs_lock);
 959                 /* this is for quota_search_lqs */
 960                 lqs_putref(lqs);
 961         } else {
 962                 CERROR("Can't find the lustre qunit size!\n");
 963         }
 964
 965         QDATA_DEBUG(qdata, "obd(%s): send %s quota req\n",
 966                     obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
 967         /* master is going to dqacq/dqrel from itself */
 968         if (is_master(qctxt)) {
 969                 int rc2;
 970                 QDATA_DEBUG(qdata, "local %s.\n",
 971                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 972                 QDATA_SET_CHANGE_QS(qdata);
 973                 rc = qctxt->lqc_handler(obd, qdata, opc);
 974                 rc2 = dqacq_completion(obd, qctxt, qdata, rc, opc);
 975                 /* this is for qunit_get() */
 976                 qunit_put(qunit);
 977
 978                 do_gettimeofday(&work_end);
 979                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
 980                 if (opc == QUOTA_DQACQ)
 981                         lprocfs_counter_add(qctxt->lqc_stats,
 982                                             wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
 983                                             timediff);
 984                 else
 985                         lprocfs_counter_add(qctxt->lqc_stats,
 986                                             wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
 987                                             timediff);
 988                 RETURN(rc ? rc : rc2);
 989         }
 990
 991         spin_lock(&qctxt->lqc_lock);
 992         if (!qctxt->lqc_import) {
 993                 spin_unlock(&qctxt->lqc_lock);
 994                 QDATA_DEBUG(qdata, "lqc_import is invalid.\n");
 995
 996                 spin_lock(&qunit_hash_lock);
 997                 remove_qunit_nolock(qunit);
 998                 spin_unlock(&qunit_hash_lock);
 999
1000                 compute_lqs_after_removing_qunit(qunit);
1001
1002                 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, -EAGAIN);
1003                 wake_up_all(&qunit->lq_waitq);
1004
1005                 /* this is for qunit_get() */
1006                 qunit_put(qunit);
1007                 /* this for alloc_qunit() */
1008                 qunit_put(qunit);
1009                 spin_lock(&qctxt->lqc_lock);
1010                 if (wait && !qctxt->lqc_import) {
1011                         spin_unlock(&qctxt->lqc_lock);
1012
1013                         LASSERT(oti && oti->oti_thread &&
1014                                 oti->oti_thread->t_watchdog);
1015
1016                         lc_watchdog_disable(oti->oti_thread->t_watchdog);
1017                         CDEBUG(D_QUOTA, "sleep for quota master\n");
1018                         l_wait_event(qctxt->lqc_wait_for_qmaster,
1019                                      check_qm(qctxt), &lwi);
1020                         CDEBUG(D_QUOTA, "wake up when quota master is back\n");
1021                         lc_watchdog_touch(oti->oti_thread->t_watchdog,
1022                                  GET_TIMEOUT(oti->oti_thread->t_svc));
1023                 } else {
1024                         spin_unlock(&qctxt->lqc_lock);
1025                 }
1026
1027                 RETURN(-EAGAIN);
1028         }
1029         imp = class_import_get(qctxt->lqc_import);
1030         spin_unlock(&qctxt->lqc_lock);
1031
1032         /* build dqacq/dqrel request */
1033         LASSERT(imp);
1034
1035         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_QUOTA_DQACQ,
1036                                         LUSTRE_MDS_VERSION, opc);
1037         class_import_put(imp);
1038         if (req == NULL) {
1039                 CERROR("Can't alloc request\n");
1040                 dqacq_completion(obd, qctxt, qdata, -ENOMEM, opc);
1041                 /* this is for qunit_get() */
1042                 qunit_put(qunit);
1043                 RETURN(-ENOMEM);
1044         }
1045
1046         ptlrpc_request_set_replen(req);
1047         req->rq_no_resend = req->rq_no_delay = 1;
1048         rc = quota_copy_qdata(req, qdata, QUOTA_REQUEST, QUOTA_IMPORT);
1049         if (rc < 0) {
1050                 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
1051                 ptlrpc_req_finished(req);
1052                 dqacq_completion(obd, qctxt, qdata, -EPROTO, opc);
1053                 /* this is for qunit_get() */
1054                 qunit_put(qunit);
1055                 RETURN(rc);
1056         }
1057
1058         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1059         aa = ptlrpc_req_async_args(req);
1060         aa->aa_ctxt = qctxt;
1061         aa->aa_qunit = qunit;
1062
1063         req->rq_interpret_reply = dqacq_interpret;
1064         ptlrpcd_add_req(req, PSCOPE_OTHER);
1065
1066         QDATA_DEBUG(qdata, "%s scheduled.\n",
1067                     opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
1068 wait_completion:
1069         if (wait && qunit) {
1070                 struct qunit_data *p = &qunit->lq_data;
1071
1072                 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
1073                 l_wait_event(qunit->lq_waitq, got_qunit(qunit, is_master(qctxt)),
1074                              &lwi);
1075                 /* rc = -EAGAIN, it means the quota master isn't ready yet
1076                  * rc = QUOTA_REQ_RETURNED, it means a quota req is finished;
1077                  * rc = -EDQUOT, it means out of quota
1078                  * rc = -EBUSY, it means recovery is happening
1079                  * other rc < 0, it means real errors, functions who call
1080                  * schedule_dqacq should take care of this */
1081                 spin_lock(&qunit->lq_lock);
1082                 rc = qunit->lq_rc;
1083                 spin_unlock(&qunit->lq_lock);
1084                 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting: id(%u) flag(%u) "
1085                        "rc(%d) owner(%d)\n", qunit, qunit->lq_data.qd_id,
1086                        qunit->lq_data.qd_flags, rc, qunit->lq_owner);
1087         }
1088
1089         qunit_put(qunit);
1090         do_gettimeofday(&work_end);
1091         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1092         if (opc == QUOTA_DQACQ)
1093                 lprocfs_counter_add(qctxt->lqc_stats,
1094                                     wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
1095                                     timediff);
1096         else
1097                 lprocfs_counter_add(qctxt->lqc_stats,
1098                                     wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
1099                                     timediff);
1100
1101         RETURN(rc);
1102 }
1103
1104 int
1105 qctxt_adjust_qunit(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
1106                    const unsigned int id[], __u32 isblk, int wait,
1107                    struct obd_trans_info *oti)
1108 {
1109         int rc = 0, i = USRQUOTA;
1110         struct qunit_data qdata[MAXQUOTAS];
1111         ENTRY;
1112
1113         if (quota_is_set(obd, id, isblk ? QB_SET : QI_SET) == 0)
1114                 RETURN(0);
1115
1116         for (i = 0; i < MAXQUOTAS; i++) {
1117                 qdata[i].qd_id = id[i];
1118                 qdata[i].qd_flags = i;
1119                 if (isblk)
1120                         QDATA_SET_BLK(&qdata[i]);
1121                 qdata[i].qd_count = 0;
1122
1123                 rc = check_cur_qunit(obd, qctxt, &qdata[i]);
1124                 if (rc > 0) {
1125                         int opc;
1126                         /* need acquire or release */
1127                         opc = rc == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1128                         rc = schedule_dqacq(obd, qctxt, &qdata[i], opc,
1129                                             wait,oti);
1130                         if (rc < 0)
1131                                 RETURN(rc);
1132                 } else if (wait == 1) {
1133                         /* when wait equates 1, that means mds_quota_acquire
1134                          * or filter_quota_acquire is calling it. */
1135                         rc = qctxt_wait_pending_dqacq(qctxt, id[i], i, isblk);
1136                         if (rc < 0)
1137                                 RETURN(rc);
1138                 }
1139         }
1140
1141         RETURN(rc);
1142 }
1143
1144 int
1145 qctxt_wait_pending_dqacq(struct lustre_quota_ctxt *qctxt, unsigned int id,
1146                          unsigned short type, int isblk)
1147 {
1148         struct lustre_qunit *qunit = NULL;
1149         struct qunit_data qdata;
1150         struct timeval work_start;
1151         struct timeval work_end;
1152         long timediff;
1153         struct l_wait_info lwi = { 0 };
1154         int rc = 0;
1155         ENTRY;
1156
1157         do_gettimeofday(&work_start);
1158         qdata.qd_id = id;
1159         qdata.qd_flags = type;
1160         if (isblk)
1161                 QDATA_SET_BLK(&qdata);
1162         qdata.qd_count = 0;
1163
1164         spin_lock(&qunit_hash_lock);
1165         qunit = dqacq_in_flight(qctxt, &qdata);
1166         spin_unlock(&qunit_hash_lock);
1167
1168         if (qunit) {
1169                 struct qunit_data *p = &qunit->lq_data;
1170
1171                 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
1172                 l_wait_event(qunit->lq_waitq, got_qunit(qunit, is_master(qctxt)),
1173                              &lwi);
1174                 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting: rc(%d) "
1175                        "owner(%d)\n", qunit, qunit->lq_rc, qunit->lq_owner);
1176                 /* keep same as schedule_dqacq() b=17030 */
1177                 spin_lock(&qunit->lq_lock);
1178                 rc = qunit->lq_rc;
1179                 spin_unlock(&qunit->lq_lock);
1180                 /* this is for dqacq_in_flight() */
1181                 qunit_put(qunit);
1182                 do_gettimeofday(&work_end);
1183                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1184                 lprocfs_counter_add(qctxt->lqc_stats,
1185                                     isblk ? LQUOTA_WAIT_PENDING_BLK_QUOTA :
1186                                             LQUOTA_WAIT_PENDING_INO_QUOTA,
1187                                     timediff);
1188         } else {
1189                 do_gettimeofday(&work_end);
1190                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1191                 lprocfs_counter_add(qctxt->lqc_stats,
1192                                     isblk ? LQUOTA_NOWAIT_PENDING_BLK_QUOTA :
1193                                             LQUOTA_NOWAIT_PENDING_INO_QUOTA,
1194                                     timediff);
1195         }
1196
1197         RETURN(rc);
1198 }
1199
1200 int
1201 qctxt_init(struct obd_device *obd, dqacq_handler_t handler)
1202 {
1203         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
1204         struct obd_device_target *obt = &obd->u.obt;
1205         struct super_block *sb = obt->obt_sb;
1206         int rc = 0;
1207         ENTRY;
1208
1209         LASSERT(qctxt);
1210
1211         rc = ptlrpcd_addref();
1212         if (rc)
1213                 RETURN(rc);
1214
1215         cfs_waitq_init(&qctxt->lqc_wait_for_qmaster);
1216         cfs_waitq_init(&qctxt->lqc_lqs_waitq);
1217         atomic_set(&qctxt->lqc_lqs, 0);
1218         spin_lock_init(&qctxt->lqc_lock);
1219         spin_lock(&qctxt->lqc_lock);
1220         qctxt->lqc_handler = handler;
1221         qctxt->lqc_sb = sb;
1222         qctxt->lqc_obt = obt;
1223         qctxt->lqc_import = NULL;
1224         qctxt->lqc_recovery = 0;
1225         qctxt->lqc_switch_qs = 1; /* Change qunit size in default setting */
1226         qctxt->lqc_valid = 1;
1227         qctxt->lqc_cqs_boundary_factor = 4;
1228         qctxt->lqc_cqs_least_bunit = PTLRPC_MAX_BRW_SIZE;
1229         qctxt->lqc_cqs_least_iunit = 2;
1230         qctxt->lqc_cqs_qs_factor = 2;
1231         qctxt->lqc_flags = 0;
1232         QUOTA_MASTER_UNREADY(qctxt);
1233         qctxt->lqc_bunit_sz = default_bunit_sz;
1234         qctxt->lqc_btune_sz = default_bunit_sz / 100 * default_btune_ratio;
1235         qctxt->lqc_iunit_sz = default_iunit_sz;
1236         qctxt->lqc_itune_sz = default_iunit_sz * default_itune_ratio / 100;
1237         qctxt->lqc_switch_seconds = 300; /* enlarging will wait 5 minutes
1238                                           * after the last shrinking */
1239         qctxt->lqc_sync_blk = 0;
1240         spin_unlock(&qctxt->lqc_lock);
1241
1242         qctxt->lqc_lqs_hash = cfs_hash_create("LQS_HASH",
1243                                               HASH_LQS_CUR_BITS,
1244                                               HASH_LQS_MAX_BITS,
1245                                               &lqs_hash_ops, CFS_HASH_REHASH);
1246         if (!qctxt->lqc_lqs_hash) {
1247                 CERROR("initialize hash lqs for %s error!\n", obd->obd_name);
1248                 RETURN(-ENOMEM);
1249         }
1250
1251 #ifdef LPROCFS
1252         rc = lquota_proc_setup(obd, is_master(qctxt));
1253         if (rc)
1254                 CERROR("initialize proc for %s error!\n", obd->obd_name);
1255 #endif
1256
1257         RETURN(rc);
1258 }
1259
1260 static int check_lqs(struct lustre_quota_ctxt *qctxt)
1261 {
1262         int rc;
1263         ENTRY;
1264
1265         rc = !atomic_read(&qctxt->lqc_lqs);
1266
1267         RETURN(rc);
1268 }
1269
1270
1271 void hash_put_lqs(void *obj, void *data)
1272 {
1273         lqs_putref((struct lustre_qunit_size *)obj);
1274 }
1275
1276 void qctxt_cleanup(struct lustre_quota_ctxt *qctxt, int force)
1277 {
1278         struct lustre_qunit *qunit, *tmp;
1279         struct list_head tmp_list;
1280         struct l_wait_info lwi = { 0 };
1281         struct obd_device_target *obt = qctxt->lqc_obt;
1282         int i;
1283         ENTRY;
1284
1285         CFS_INIT_LIST_HEAD(&tmp_list);
1286
1287         spin_lock(&qctxt->lqc_lock);
1288         qctxt->lqc_valid = 0;
1289         spin_unlock(&qctxt->lqc_lock);
1290
1291         spin_lock(&qunit_hash_lock);
1292         for (i = 0; i < NR_DQHASH; i++) {
1293                 list_for_each_entry_safe(qunit, tmp, &qunit_hash[i], lq_hash) {
1294                         if (qunit->lq_ctxt != qctxt)
1295                                 continue;
1296                         remove_qunit_nolock(qunit);
1297                         list_add(&qunit->lq_hash, &tmp_list);
1298                 }
1299         }
1300         spin_unlock(&qunit_hash_lock);
1301
1302         list_for_each_entry_safe(qunit, tmp, &tmp_list, lq_hash) {
1303                 list_del_init(&qunit->lq_hash);
1304                 compute_lqs_after_removing_qunit(qunit);
1305
1306                 /* wake up all waiters */
1307                 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, 0);
1308                 wake_up_all(&qunit->lq_waitq);
1309                 qunit_put(qunit);
1310         }
1311
1312         /* after qctxt_cleanup, qctxt might be freed, then check_qm() is
1313          * unpredicted. So we must wait until lqc_wait_for_qmaster is empty */
1314         while (cfs_waitq_active(&qctxt->lqc_wait_for_qmaster)) {
1315                 cfs_waitq_signal(&qctxt->lqc_wait_for_qmaster);
1316                 cfs_schedule_timeout(CFS_TASK_INTERRUPTIBLE,
1317                                      cfs_time_seconds(1));
1318         }
1319
1320         cfs_hash_for_each_safe(qctxt->lqc_lqs_hash, hash_put_lqs, NULL);
1321         l_wait_event(qctxt->lqc_lqs_waitq, check_lqs(qctxt), &lwi);
1322         down_write(&obt->obt_rwsem);
1323         cfs_hash_destroy(qctxt->lqc_lqs_hash);
1324         qctxt->lqc_lqs_hash = NULL;
1325         up_write(&obt->obt_rwsem);
1326
1327         ptlrpcd_decref();
1328
1329 #ifdef LPROCFS
1330         if (lquota_proc_cleanup(qctxt))
1331                 CERROR("cleanup proc error!\n");
1332 #endif
1333
1334         EXIT;
1335 }
1336
1337 struct qslave_recov_thread_data {
1338         struct obd_device *obd;
1339         struct lustre_quota_ctxt *qctxt;
1340         struct completion comp;
1341 };
1342
1343 /* FIXME only recovery block quota by now */
1344 static int qslave_recovery_main(void *arg)
1345 {
1346         struct qslave_recov_thread_data *data = arg;
1347         struct obd_device *obd = data->obd;
1348         struct lustre_quota_ctxt *qctxt = data->qctxt;
1349         unsigned int type;
1350         int rc = 0;
1351         ENTRY;
1352
1353         cfs_daemonize_ctxt("qslave_recovd");
1354
1355         /* for obdfilter */
1356         class_incref(obd, "qslave_recovd_filter", obd);
1357
1358         complete(&data->comp);
1359
1360         spin_lock(&qctxt->lqc_lock);
1361         if (qctxt->lqc_recovery) {
1362                 spin_unlock(&qctxt->lqc_lock);
1363                 class_decref(obd, "qslave_recovd_filter", obd);
1364                 RETURN(0);
1365         } else {
1366                 qctxt->lqc_recovery = 1;
1367                 spin_unlock(&qctxt->lqc_lock);
1368         }
1369
1370         for (type = USRQUOTA; type < MAXQUOTAS; type++) {
1371                 struct qunit_data qdata;
1372                 struct quota_info *dqopt = sb_dqopt(qctxt->lqc_sb);
1373                 struct list_head id_list;
1374                 struct dquot_id *dqid, *tmp;
1375                 int ret;
1376
1377                 LOCK_DQONOFF_MUTEX(dqopt);
1378                 if (!ll_sb_has_quota_active(qctxt->lqc_sb, type)) {
1379                         UNLOCK_DQONOFF_MUTEX(dqopt);
1380                         break;
1381                 }
1382
1383                 LASSERT(dqopt->files[type] != NULL);
1384                 CFS_INIT_LIST_HEAD(&id_list);
1385 #ifndef KERNEL_SUPPORTS_QUOTA_READ
1386                 rc = fsfilt_qids(obd, dqopt->files[type], NULL, type, &id_list);
1387 #else
1388                 rc = fsfilt_qids(obd, NULL, dqopt->files[type], type, &id_list);
1389 #endif
1390                 UNLOCK_DQONOFF_MUTEX(dqopt);
1391                 if (rc)
1392                         CERROR("Get ids from quota file failed. (rc:%d)\n", rc);
1393
1394                 list_for_each_entry_safe(dqid, tmp, &id_list, di_link) {
1395                         list_del_init(&dqid->di_link);
1396                         /* skip slave recovery on itself */
1397                         if (is_master(qctxt))
1398                                 goto free;
1399                         if (rc && rc != -EBUSY)
1400                                 goto free;
1401
1402                         qdata.qd_id = dqid->di_id;
1403                         qdata.qd_flags = type;
1404                         QDATA_SET_BLK(&qdata);
1405                         qdata.qd_count = 0;
1406
1407                         ret = check_cur_qunit(obd, qctxt, &qdata);
1408                         if (ret > 0) {
1409                                 int opc;
1410                                 opc = ret == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1411                                 rc = schedule_dqacq(obd, qctxt, &qdata, opc,
1412                                                     0, NULL);
1413                                 if (rc == -EDQUOT)
1414                                         rc = 0;
1415                         } else {
1416                                 rc = 0;
1417                         }
1418
1419                         if (rc && rc != -EBUSY)
1420                                 CERROR("qslave recovery failed! (id:%d type:%d "
1421                                        " rc:%d)\n", dqid->di_id, type, rc);
1422 free:
1423                         OBD_FREE_PTR(dqid);
1424                 }
1425         }
1426
1427         spin_lock(&qctxt->lqc_lock);
1428         qctxt->lqc_recovery = 0;
1429         spin_unlock(&qctxt->lqc_lock);
1430         class_decref(obd, "qslave_recovd_filter", obd);
1431         RETURN(rc);
1432 }
1433
1434 void
1435 qslave_start_recovery(struct obd_device *obd, struct lustre_quota_ctxt *qctxt)
1436 {
1437         struct qslave_recov_thread_data data;
1438         int rc;
1439         ENTRY;
1440
1441         if (!ll_sb_any_quota_active(qctxt->lqc_sb))
1442                 goto exit;
1443
1444         data.obd = obd;
1445         data.qctxt = qctxt;
1446         init_completion(&data.comp);
1447
1448         rc = kernel_thread(qslave_recovery_main, &data, CLONE_VM|CLONE_FILES);
1449         if (rc < 0) {
1450                 CERROR("Cannot start quota recovery thread: rc %d\n", rc);
1451                 goto exit;
1452         }
1453         wait_for_completion(&data.comp);
1454 exit:
1455         EXIT;
1456 }
1457
1458 int quota_is_on(struct lustre_quota_ctxt *qctxt, struct obd_quotactl *oqctl)
1459 {
1460         unsigned int type;
1461
1462         for (type = USRQUOTA; type < MAXQUOTAS; type++) {
1463                 if (!Q_TYPESET(oqctl, type))
1464                         continue;
1465                 if (!(qctxt->lqc_flags & UGQUOTA2LQC(oqctl->qc_type)))
1466                         return 0;
1467         }
1468         return 1;
1469 }
1470
1471 int quota_is_off(struct lustre_quota_ctxt *qctxt, struct obd_quotactl *oqctl)
1472 {
1473         unsigned int type;
1474
1475         for (type = USRQUOTA; type < MAXQUOTAS; type++) {
1476                 if (!Q_TYPESET(oqctl, type))
1477                         continue;
1478                 if (qctxt->lqc_flags & UGQUOTA2LQC(oqctl->qc_type))
1479                         return 0;
1480         }
1481         return 1;
1482 }
1483
1484 /**
1485  * When quotaon, build a lqs for every uid/gid who has been set limitation
1486  * for quota. After quota_search_lqs, it will hold one ref for the lqs.
1487  * It will be released when qctxt_cleanup() is executed b=18574
1488  *
1489  * Should be called with obt->obt_quotachecking held. b=20152
1490  */
1491 void build_lqs(struct obd_device *obd)
1492 {
1493         struct obd_device_target *obt = &obd->u.obt;
1494         struct lustre_quota_ctxt *qctxt = &obt->obt_qctxt;
1495         struct list_head id_list;
1496         int i, rc;
1497
1498         LASSERT_SEM_LOCKED(&obt->obt_quotachecking);
1499         INIT_LIST_HEAD(&id_list);
1500         for (i = 0; i < MAXQUOTAS; i++) {
1501                 struct dquot_id *dqid, *tmp;
1502
1503                 if (sb_dqopt(qctxt->lqc_sb)->files[i] == NULL)
1504                         continue;
1505
1506 #ifndef KERNEL_SUPPORTS_QUOTA_READ
1507                 rc = fsfilt_qids(obd, sb_dqopt(qctxt->lqc_sb)->files[i], NULL,
1508                                  i, &id_list);
1509 #else
1510                 rc = fsfilt_qids(obd, NULL, sb_dqopt(qctxt->lqc_sb)->files[i],
1511                                  i, &id_list);
1512 #endif
1513                 if (rc) {
1514                         CERROR("%s: failed to get %s qids!\n", obd->obd_name,
1515                                i ? "group" : "user");
1516                         continue;
1517                 }
1518
1519                 list_for_each_entry_safe(dqid, tmp, &id_list,
1520                                          di_link) {
1521                         struct lustre_qunit_size *lqs;
1522
1523                         list_del_init(&dqid->di_link);
1524                         lqs = quota_search_lqs(LQS_KEY(i, dqid->di_id),
1525                                                qctxt, 1);
1526                         if (lqs && !IS_ERR(lqs)) {
1527                                 lqs->lqs_flags |= dqid->di_flag;
1528                                 lqs_putref(lqs);
1529                         } else {
1530                                 CERROR("%s: failed to create a lqs for %sid %u"
1531                                        "\n", obd->obd_name, i ? "g" : "u",
1532                                        dqid->di_id);
1533                         }
1534
1535                         OBD_FREE_PTR(dqid);
1536                 }
1537         }
1538 }
1539
1540 /**
1541  * lqs<->qctxt hash operations
1542  */
1543
1544 /**
1545  * string hashing using djb2 hash algorithm
1546  */
1547 static unsigned
1548 lqs_hash(cfs_hash_t *hs, void *key, unsigned mask)
1549 {
1550         struct quota_adjust_qunit *lqs_key;
1551         unsigned hash;
1552         ENTRY;
1553
1554         LASSERT(key);
1555         lqs_key = (struct quota_adjust_qunit *)key;
1556         hash = (QAQ_IS_GRP(lqs_key) ? 5381 : 5387) * lqs_key->qaq_id;
1557
1558         RETURN(hash & mask);
1559 }
1560
1561 static int
1562 lqs_compare(void *key, struct hlist_node *hnode)
1563 {
1564         struct lustre_qunit_size *q;
1565         int rc;
1566         ENTRY;
1567
1568         LASSERT(key);
1569         q = hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1570
1571         spin_lock(&q->lqs_lock);
1572         rc = (q->lqs_key == *((unsigned long long *)key));
1573         spin_unlock(&q->lqs_lock);
1574
1575         RETURN(rc);
1576 }
1577
1578 static void *
1579 lqs_get(struct hlist_node *hnode)
1580 {
1581         struct lustre_qunit_size *q =
1582                 hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1583         ENTRY;
1584
1585         __lqs_getref(q);
1586
1587         RETURN(q);
1588 }
1589
1590 static void *
1591 lqs_put(struct hlist_node *hnode)
1592 {
1593         struct lustre_qunit_size *q =
1594                 hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1595         ENTRY;
1596
1597         __lqs_putref(q);
1598
1599         RETURN(q);
1600 }
1601
1602 static void
1603 lqs_exit(struct hlist_node *hnode)
1604 {
1605         struct lustre_qunit_size *q =
1606                 hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1607         ENTRY;
1608
1609         /*
1610          * Nothing should be left. User of lqs put it and
1611          * lqs also was deleted from table by this time
1612          * so we should have 0 refs.
1613          */
1614         LASSERTF(atomic_read(&q->lqs_refcount) == 0,
1615                  "Busy lqs %p with %d refs\n", q,
1616                  atomic_read(&q->lqs_refcount));
1617         OBD_FREE_PTR(q);
1618         EXIT;
1619 }
1620
1621 static cfs_hash_ops_t lqs_hash_ops = {
1622         .hs_hash    = lqs_hash,
1623         .hs_compare = lqs_compare,
1624         .hs_get     = lqs_get,
1625         .hs_put     = lqs_put,
1626         .hs_exit    = lqs_exit
1627 };
1628 #endif /* HAVE_QUOTA_SUPPORT */