lustre/lov/lov_lock.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  * Implementation of cl_lock for LOV layer.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_LOV
  42
  43 #include "lov_cl_internal.h"
  44
  45 /** \addtogroup lov lov @{ */
  46
  47 static struct cl_lock_closure *lov_closure_get(const struct lu_env *env,
  48                                                struct cl_lock *parent);
  49
  50 /*****************************************************************************
  51  *
  52  * Lov lock operations.
  53  *
  54  */
  55
  56 static struct lov_sublock_env *lov_sublock_env_get(const struct lu_env *env,
  57                                                    struct cl_lock *parent,
  58                                                    struct lov_lock_sub *lls)
  59 {
  60         struct lov_sublock_env *subenv;
  61         struct lov_io          *lio    = lov_env_io(env);
  62         struct cl_io           *io     = lio->lis_cl.cis_io;
  63         struct lov_io_sub      *sub;
  64
  65         subenv = &lov_env_session(env)->ls_subenv;
  66
  67         /*
  68          * FIXME: We tend to use the subio's env & io to call the sublock
  69          * lock operations because osc lock sometimes stores some control
  70          * variables in thread's IO infomation(Now only lockless information).
  71          * However, if the lock's host(object) is different from the object
  72          * for current IO, we have no way to get the subenv and subio because
  73          * they are not initialized at all. As a temp fix, in this case,
  74          * we still borrow the parent's env to call sublock operations.
  75          */
  76         if (!cl_object_same(io->ci_obj, parent->cll_descr.cld_obj)) {
  77                 subenv->lse_env = env;
  78                 subenv->lse_io  = io;
  79                 subenv->lse_sub = NULL;
  80         } else {
  81                 LASSERT(io != NULL);
  82                 sub = lov_sub_get(env, lio, lls->sub_stripe);
  83                 if (!IS_ERR(sub)) {
  84                         subenv->lse_env = sub->sub_env;
  85                         subenv->lse_io  = sub->sub_io;
  86                         subenv->lse_sub = sub;
  87                 } else {
  88                         subenv = (void*)sub;
  89                 }
  90         }
  91         return subenv;
  92 }
  93
  94 static void lov_sublock_env_put(struct lov_sublock_env *subenv)
  95 {
  96         if (subenv && subenv->lse_sub)
  97                 lov_sub_put(subenv->lse_sub);
  98 }
  99
 100 static void lov_sublock_adopt(const struct lu_env *env, struct lov_lock *lck,
 101                               struct cl_lock *sublock, int idx,
 102                               struct lov_lock_link *link)
 103 {
 104         struct lovsub_lock *lsl;
 105         struct cl_lock     *parent = lck->lls_cl.cls_lock;
 106         int                 rc;
 107
 108         LASSERT(cl_lock_is_mutexed(parent));
 109         LASSERT(cl_lock_is_mutexed(sublock));
 110         ENTRY;
 111
 112         lsl = cl2sub_lock(sublock);
 113         /*
 114          * check that sub-lock doesn't have lock link to this top-lock.
 115          */
 116         LASSERT(lov_lock_link_find(env, lck, lsl) == NULL);
 117         LASSERT(idx < lck->lls_nr);
 118
 119         lck->lls_sub[idx].sub_lock = lsl;
 120         lck->lls_nr_filled++;
 121         LASSERT(lck->lls_nr_filled <= lck->lls_nr);
 122         list_add_tail(&link->lll_list, &lsl->lss_parents);
 123         link->lll_idx = idx;
 124         link->lll_super = lck;
 125         cl_lock_get(parent);
 126         lu_ref_add(&parent->cll_reference, "lov-child", sublock);
 127         lck->lls_sub[idx].sub_flags |= LSF_HELD;
 128         cl_lock_user_add(env, sublock);
 129
 130         rc = lov_sublock_modify(env, lck, lsl, &sublock->cll_descr, idx);
 131         LASSERT(rc == 0); /* there is no way this can fail, currently */
 132         EXIT;
 133 }
 134
 135 static struct cl_lock *lov_sublock_alloc(const struct lu_env *env,
 136                                          const struct cl_io *io,
 137                                          struct lov_lock *lck,
 138                                          int idx, struct lov_lock_link **out)
 139 {
 140         struct cl_lock       *sublock;
 141         struct cl_lock       *parent;
 142         struct lov_lock_link *link;
 143
 144         LASSERT(idx < lck->lls_nr);
 145         ENTRY;
 146
 147         OBD_SLAB_ALLOC_PTR_GFP(link, lov_lock_link_kmem, CFS_ALLOC_IO);
 148         if (link != NULL) {
 149                 struct lov_sublock_env *subenv;
 150                 struct lov_lock_sub  *lls;
 151                 struct cl_lock_descr *descr;
 152
 153                 parent = lck->lls_cl.cls_lock;
 154                 lls    = &lck->lls_sub[idx];
 155                 descr  = &lls->sub_descr;
 156
 157                 subenv = lov_sublock_env_get(env, parent, lls);
 158                 if (!IS_ERR(subenv)) {
 159                         /* CAVEAT: Don't try to add a field in lov_lock_sub
 160                          * to remember the subio. This is because lock is able
 161                          * to be cached, but this is not true for IO. This
 162                          * further means a sublock might be referenced in
 163                          * different io context. -jay */
 164
 165                         sublock = cl_lock_hold(subenv->lse_env, subenv->lse_io,
 166                                                descr, "lov-parent", parent);
 167                         lov_sublock_env_put(subenv);
 168                 } else {
 169                         /* error occurs. */
 170                         sublock = (void*)subenv;
 171                 }
 172
 173                 if (!IS_ERR(sublock))
 174                         *out = link;
 175                 else
 176                         OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
 177         } else
 178                 sublock = ERR_PTR(-ENOMEM);
 179         RETURN(sublock);
 180 }
 181
 182 static void lov_sublock_unlock(const struct lu_env *env,
 183                                struct lovsub_lock *lsl,
 184                                struct cl_lock_closure *closure,
 185                                struct lov_sublock_env *subenv)
 186 {
 187         ENTRY;
 188         lov_sublock_env_put(subenv);
 189         lsl->lss_active = NULL;
 190         cl_lock_disclosure(env, closure);
 191         EXIT;
 192 }
 193
 194 static int lov_sublock_lock(const struct lu_env *env,
 195                             struct lov_lock *lck,
 196                             struct lov_lock_sub *lls,
 197                             struct cl_lock_closure *closure,
 198                             struct lov_sublock_env **lsep)
 199 {
 200         struct lovsub_lock *sublock;
 201         struct cl_lock     *child;
 202         int                 result = 0;
 203         ENTRY;
 204
 205         LASSERT(list_empty(&closure->clc_list));
 206
 207         sublock = lls->sub_lock;
 208         child = sublock->lss_cl.cls_lock;
 209         result = cl_lock_closure_build(env, child, closure);
 210         if (result == 0) {
 211                 struct cl_lock *parent = closure->clc_origin;
 212
 213                 LASSERT(cl_lock_is_mutexed(child));
 214                 sublock->lss_active = parent;
 215
 216                 if (unlikely(child->cll_state == CLS_FREEING)) {
 217                         struct lov_lock_link *link;
 218                         /*
 219                          * we could race with lock deletion which temporarily
 220                          * put the lock in freeing state, bug 19080.
 221                          */
 222                         LASSERT(!(lls->sub_flags & LSF_HELD));
 223
 224                         link = lov_lock_link_find(env, lck, sublock);
 225                         LASSERT(link != NULL);
 226                         lov_lock_unlink(env, link, sublock);
 227                         lov_sublock_unlock(env, sublock, closure, NULL);
 228                         result = CLO_REPEAT;
 229                 } else if (lsep) {
 230                         struct lov_sublock_env *subenv;
 231                         subenv = lov_sublock_env_get(env, parent, lls);
 232                         if (IS_ERR(subenv)) {
 233                                 lov_sublock_unlock(env, sublock,
 234                                                    closure, NULL);
 235                                 result = PTR_ERR(subenv);
 236                         } else {
 237                                 *lsep = subenv;
 238                         }
 239                 }
 240         }
 241         RETURN(result);
 242 }
 243
 244 /**
 245  * Updates the result of a top-lock operation from a result of sub-lock
 246  * sub-operations. Top-operations like lov_lock_{enqueue,use,unuse}() iterate
 247  * over sub-locks and lov_subresult() is used to calculate return value of a
 248  * top-operation. To this end, possible return values of sub-operations are
 249  * ordered as
 250  *
 251  *     - 0                  success
 252  *     - CLO_WAIT           wait for event
 253  *     - CLO_REPEAT         repeat top-operation
 254  *     - -ne                fundamental error
 255  *
 256  * Top-level return code can only go down through this list. CLO_REPEAT
 257  * overwrites CLO_WAIT, because lock mutex was released and sleeping condition
 258  * has to be rechecked by the upper layer.
 259  */
 260 static int lov_subresult(int result, int rc)
 261 {
 262         int result_rank;
 263         int rc_rank;
 264
 265         LASSERT(result <= 0 || result == CLO_REPEAT || result == CLO_WAIT);
 266         LASSERT(rc <= 0 || rc == CLO_REPEAT || rc == CLO_WAIT);
 267         CLASSERT(CLO_WAIT < CLO_REPEAT);
 268
 269         ENTRY;
 270
 271         /* calculate ranks in the ordering above */
 272         result_rank = result < 0 ? 1 + CLO_REPEAT : result;
 273         rc_rank = rc < 0 ? 1 + CLO_REPEAT : rc;
 274
 275         if (result_rank < rc_rank)
 276                 result = rc;
 277         RETURN(result);
 278 }
 279
 280 /**
 281  * Creates sub-locks for a given lov_lock for the first time.
 282  *
 283  * Goes through all sub-objects of top-object, and creates sub-locks on every
 284  * sub-object intersecting with top-lock extent. This is complicated by the
 285  * fact that top-lock (that is being created) can be accessed concurrently
 286  * through already created sub-locks (possibly shared with other top-locks).
 287  */
 288 static int lov_lock_sub_init(const struct lu_env *env,
 289                              struct lov_lock *lck, const struct cl_io *io)
 290 {
 291         int result = 0;
 292         int i;
 293         int nr;
 294         obd_off start;
 295         obd_off end;
 296         obd_off file_start;
 297         obd_off file_end;
 298
 299         struct lov_object       *loo    = cl2lov(lck->lls_cl.cls_obj);
 300         struct lov_layout_raid0 *r0     = lov_r0(loo);
 301         struct cl_lock          *parent = lck->lls_cl.cls_lock;
 302
 303         ENTRY;
 304
 305         lck->lls_orig = parent->cll_descr;
 306         file_start = cl_offset(lov2cl(loo), parent->cll_descr.cld_start);
 307         file_end   = cl_offset(lov2cl(loo), parent->cll_descr.cld_end + 1) - 1;
 308
 309         for (i = 0, nr = 0; i < r0->lo_nr; i++) {
 310                 /*
 311                  * XXX for wide striping smarter algorithm is desirable,
 312                  * breaking out of the loop, early.
 313                  */
 314                 if (lov_stripe_intersects(r0->lo_lsm, i,
 315                                           file_start, file_end, &start, &end))
 316                         nr++;
 317         }
 318         LASSERT(nr > 0);
 319         OBD_ALLOC(lck->lls_sub, nr * sizeof lck->lls_sub[0]);
 320         if (lck->lls_sub == NULL)
 321                 RETURN(-ENOMEM);
 322
 323         lck->lls_nr = nr;
 324         /*
 325          * First, fill in sub-lock descriptions in
 326          * lck->lls_sub[].sub_descr. They are used by lov_sublock_alloc()
 327          * (called below in this function, and by lov_lock_enqueue()) to
 328          * create sub-locks. At this moment, no other thread can access
 329          * top-lock.
 330          */
 331         for (i = 0, nr = 0; i < r0->lo_nr; ++i) {
 332                 if (lov_stripe_intersects(r0->lo_lsm, i,
 333                                           file_start, file_end, &start, &end)) {
 334                         struct cl_lock_descr *descr;
 335
 336                         descr = &lck->lls_sub[nr].sub_descr;
 337
 338                         LASSERT(descr->cld_obj == NULL);
 339                         descr->cld_obj   = lovsub2cl(r0->lo_sub[i]);
 340                         descr->cld_start = cl_index(descr->cld_obj, start);
 341                         descr->cld_end   = cl_index(descr->cld_obj, end);
 342                         descr->cld_mode  = parent->cll_descr.cld_mode;
 343                         descr->cld_gid   = parent->cll_descr.cld_gid;
 344                         /* XXX has no effect */
 345                         lck->lls_sub[nr].sub_got = *descr;
 346                         lck->lls_sub[nr].sub_stripe = i;
 347                         nr++;
 348                 }
 349         }
 350         LASSERT(nr == lck->lls_nr);
 351         /*
 352          * Then, create sub-locks. Once at least one sub-lock was created,
 353          * top-lock can be reached by other threads.
 354          */
 355         for (i = 0; i < lck->lls_nr; ++i) {
 356                 struct cl_lock       *sublock;
 357                 struct lov_lock_link *link;
 358
 359                 if (lck->lls_sub[i].sub_lock == NULL) {
 360                         sublock = lov_sublock_alloc(env, io, lck, i, &link);
 361                         if (IS_ERR(sublock)) {
 362                                 result = PTR_ERR(sublock);
 363                                 break;
 364                         }
 365                         cl_lock_mutex_get(env, sublock);
 366                         cl_lock_mutex_get(env, parent);
 367                         /*
 368                          * recheck under mutex that sub-lock wasn't created
 369                          * concurrently, and that top-lock is still alive.
 370                          */
 371                         if (lck->lls_sub[i].sub_lock == NULL &&
 372                             parent->cll_state < CLS_FREEING) {
 373                                 lov_sublock_adopt(env, lck, sublock, i, link);
 374                                 cl_lock_mutex_put(env, parent);
 375                         } else {
 376                                 OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
 377                                 cl_lock_mutex_put(env, parent);
 378                                 cl_lock_unhold(env, sublock,
 379                                                "lov-parent", parent);
 380                         }
 381                         cl_lock_mutex_put(env, sublock);
 382                 }
 383         }
 384         /*
 385          * Some sub-locks can be missing at this point. This is not a problem,
 386          * because enqueue will create them anyway. Main duty of this function
 387          * is to fill in sub-lock descriptions in a race free manner.
 388          */
 389         RETURN(result);
 390 }
 391
 392 static int lov_sublock_release(const struct lu_env *env, struct lov_lock *lck,
 393                                int i, int deluser, int rc)
 394 {
 395         struct cl_lock *parent = lck->lls_cl.cls_lock;
 396
 397         LASSERT(cl_lock_is_mutexed(parent));
 398         ENTRY;
 399
 400         if (lck->lls_sub[i].sub_flags & LSF_HELD) {
 401                 struct cl_lock    *sublock;
 402                 int dying;
 403
 404                 LASSERT(lck->lls_sub[i].sub_lock != NULL);
 405                 sublock = lck->lls_sub[i].sub_lock->lss_cl.cls_lock;
 406                 LASSERT(cl_lock_is_mutexed(sublock));
 407
 408                 lck->lls_sub[i].sub_flags &= ~LSF_HELD;
 409                 if (deluser)
 410                         cl_lock_user_del(env, sublock);
 411                 /*
 412                  * If the last hold is released, and cancellation is pending
 413                  * for a sub-lock, release parent mutex, to avoid keeping it
 414                  * while sub-lock is being paged out.
 415                  */
 416                 dying = (sublock->cll_descr.cld_mode == CLM_PHANTOM ||
 417                          sublock->cll_descr.cld_mode == CLM_GROUP ||
 418                          (sublock->cll_flags & (CLF_CANCELPEND|CLF_DOOMED))) &&
 419                         sublock->cll_holds == 1;
 420                 if (dying)
 421                         cl_lock_mutex_put(env, parent);
 422                 cl_lock_unhold(env, sublock, "lov-parent", parent);
 423                 if (dying) {
 424                         cl_lock_mutex_get(env, parent);
 425                         rc = lov_subresult(rc, CLO_REPEAT);
 426                 }
 427                 /*
 428                  * From now on lck->lls_sub[i].sub_lock is a "weak" pointer,
 429                  * not backed by a reference on a
 430                  * sub-lock. lovsub_lock_delete() will clear
 431                  * lck->lls_sub[i].sub_lock under semaphores, just before
 432                  * sub-lock is destroyed.
 433                  */
 434         }
 435         RETURN(rc);
 436 }
 437
 438 static void lov_sublock_hold(const struct lu_env *env, struct lov_lock *lck,
 439                              int i)
 440 {
 441         struct cl_lock *parent = lck->lls_cl.cls_lock;
 442
 443         LASSERT(cl_lock_is_mutexed(parent));
 444         ENTRY;
 445
 446         if (!(lck->lls_sub[i].sub_flags & LSF_HELD)) {
 447                 struct cl_lock *sublock;
 448
 449                 LASSERT(lck->lls_sub[i].sub_lock != NULL);
 450                 sublock = lck->lls_sub[i].sub_lock->lss_cl.cls_lock;
 451                 LASSERT(cl_lock_is_mutexed(sublock));
 452                 LASSERT(sublock->cll_state != CLS_FREEING);
 453
 454                 lck->lls_sub[i].sub_flags |= LSF_HELD;
 455
 456                 cl_lock_get_trust(sublock);
 457                 cl_lock_hold_add(env, sublock, "lov-parent", parent);
 458                 cl_lock_user_add(env, sublock);
 459                 cl_lock_put(env, sublock);
 460         }
 461         EXIT;
 462 }
 463
 464 static void lov_lock_fini(const struct lu_env *env,
 465                           struct cl_lock_slice *slice)
 466 {
 467         struct lov_lock *lck;
 468         int i;
 469
 470         ENTRY;
 471         lck = cl2lov_lock(slice);
 472         LASSERT(lck->lls_nr_filled == 0);
 473         if (lck->lls_sub != NULL) {
 474                 for (i = 0; i < lck->lls_nr; ++i)
 475                         /*
 476                          * No sub-locks exists at this point, as sub-lock has
 477                          * a reference on its parent.
 478                          */
 479                         LASSERT(lck->lls_sub[i].sub_lock == NULL);
 480                 OBD_FREE(lck->lls_sub, lck->lls_nr * sizeof lck->lls_sub[0]);
 481         }
 482         OBD_SLAB_FREE_PTR(lck, lov_lock_kmem);
 483         EXIT;
 484 }
 485
 486 /**
 487  * Tries to advance a state machine of a given sub-lock toward enqueuing of
 488  * the top-lock.
 489  *
 490  * \retval 0 if state-transition can proceed
 491  * \retval -ve otherwise.
 492  */
 493 static int lov_lock_enqueue_one(const struct lu_env *env, struct lov_lock *lck,
 494                                 struct cl_lock *sublock,
 495                                 struct cl_io *io, __u32 enqflags, int last)
 496 {
 497         int result;
 498         ENTRY;
 499
 500         /* first, try to enqueue a sub-lock ... */
 501         result = cl_enqueue_try(env, sublock, io, enqflags);
 502         if (sublock->cll_state == CLS_ENQUEUED)
 503                 /* if it is enqueued, try to `wait' on it---maybe it's already
 504                  * granted */
 505                 result = cl_wait_try(env, sublock);
 506         /*
 507          * If CEF_ASYNC flag is set, then all sub-locks can be enqueued in
 508          * parallel, otherwise---enqueue has to wait until sub-lock is granted
 509          * before proceeding to the next one.
 510          */
 511         if (result == CLO_WAIT && sublock->cll_state <= CLS_HELD &&
 512             enqflags & CEF_ASYNC && !last)
 513                 result = 0;
 514         RETURN(result);
 515 }
 516
 517 /**
 518  * Helper function for lov_lock_enqueue() that creates missing sub-lock.
 519  */
 520 static int lov_sublock_fill(const struct lu_env *env, struct cl_lock *parent,
 521                             struct cl_io *io, struct lov_lock *lck, int idx)
 522 {
 523         struct lov_lock_link *link;
 524         struct cl_lock       *sublock;
 525         int                   result;
 526
 527         LASSERT(parent->cll_depth == 1);
 528         cl_lock_mutex_put(env, parent);
 529         sublock = lov_sublock_alloc(env, io, lck, idx, &link);
 530         if (!IS_ERR(sublock))
 531                 cl_lock_mutex_get(env, sublock);
 532         cl_lock_mutex_get(env, parent);
 533
 534         if (!IS_ERR(sublock)) {
 535                 if (parent->cll_state == CLS_QUEUING &&
 536                     lck->lls_sub[idx].sub_lock == NULL)
 537                         lov_sublock_adopt(env, lck, sublock, idx, link);
 538                 else {
 539                         OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
 540                         /* other thread allocated sub-lock, or enqueue is no
 541                          * longer going on */
 542                         cl_lock_mutex_put(env, parent);
 543                         cl_lock_unhold(env, sublock, "lov-parent", parent);
 544                         cl_lock_mutex_get(env, parent);
 545                 }
 546                 cl_lock_mutex_put(env, sublock);
 547                 result = CLO_REPEAT;
 548         } else
 549                 result = PTR_ERR(sublock);
 550         return result;
 551 }
 552
 553 /**
 554  * Implementation of cl_lock_operations::clo_enqueue() for lov layer. This
 555  * function is rather subtle, as it enqueues top-lock (i.e., advances top-lock
 556  * state machine from CLS_QUEUING to CLS_ENQUEUED states) by juggling sub-lock
 557  * state machines in the face of sub-locks sharing (by multiple top-locks),
 558  * and concurrent sub-lock cancellations.
 559  */
 560 static int lov_lock_enqueue(const struct lu_env *env,
 561                             const struct cl_lock_slice *slice,
 562                             struct cl_io *io, __u32 enqflags)
 563 {
 564         struct cl_lock         *lock    = slice->cls_lock;
 565         struct lov_lock        *lck     = cl2lov_lock(slice);
 566         struct cl_lock_closure *closure = lov_closure_get(env, lock);
 567         int i;
 568         int result;
 569         enum cl_lock_state minstate;
 570
 571         ENTRY;
 572
 573         for (result = 0, minstate = CLS_FREEING, i = 0; i < lck->lls_nr; ++i) {
 574                 int rc;
 575                 struct lovsub_lock     *sub;
 576                 struct lov_lock_sub    *lls;
 577                 struct cl_lock         *sublock;
 578                 struct lov_sublock_env *subenv;
 579
 580                 if (lock->cll_state != CLS_QUEUING) {
 581                         /*
 582                          * Lock might have left QUEUING state if previous
 583                          * iteration released its mutex. Stop enqueing in this
 584                          * case and let the upper layer to decide what to do.
 585                          */
 586                         LASSERT(i > 0 && result != 0);
 587                         break;
 588                 }
 589
 590                 lls = &lck->lls_sub[i];
 591                 sub = lls->sub_lock;
 592                 /*
 593                  * Sub-lock might have been canceled, while top-lock was
 594                  * cached.
 595                  */
 596                 if (sub == NULL) {
 597                         result = lov_sublock_fill(env, lock, io, lck, i);
 598                         /* lov_sublock_fill() released @lock mutex,
 599                          * restart. */
 600                         break;
 601                 }
 602                 sublock = sub->lss_cl.cls_lock;
 603                 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
 604                 if (rc == 0) {
 605                         lov_sublock_hold(env, lck, i);
 606                         rc = lov_lock_enqueue_one(subenv->lse_env, lck, sublock,
 607                                                   subenv->lse_io, enqflags,
 608                                                   i == lck->lls_nr - 1);
 609                         minstate = min(minstate, sublock->cll_state);
 610                         /*
 611                          * Don't hold a sub-lock in CLS_CACHED state, see
 612                          * description for lov_lock::lls_sub.
 613                          */
 614                         if (sublock->cll_state > CLS_HELD)
 615                                 rc = lov_sublock_release(env, lck, i, 1, rc);
 616                         lov_sublock_unlock(env, sub, closure, subenv);
 617                 }
 618                 result = lov_subresult(result, rc);
 619                 if (result != 0)
 620                         break;
 621         }
 622         cl_lock_closure_fini(closure);
 623         RETURN(result ?: minstate >= CLS_ENQUEUED ? 0 : CLO_WAIT);
 624 }
 625
 626 static int lov_lock_unuse(const struct lu_env *env,
 627                           const struct cl_lock_slice *slice)
 628 {
 629         struct lov_lock        *lck     = cl2lov_lock(slice);
 630         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
 631         int i;
 632         int result;
 633
 634         ENTRY;
 635
 636         for (result = 0, i = 0; i < lck->lls_nr; ++i) {
 637                 int rc;
 638                 struct lovsub_lock     *sub;
 639                 struct cl_lock         *sublock;
 640                 struct lov_lock_sub    *lls;
 641                 struct lov_sublock_env *subenv;
 642
 643                 /* top-lock state cannot change concurrently, because single
 644                  * thread (one that released the last hold) carries unlocking
 645                  * to the completion. */
 646                 LASSERT(slice->cls_lock->cll_state == CLS_UNLOCKING);
 647                 lls = &lck->lls_sub[i];
 648                 sub = lls->sub_lock;
 649                 if (sub == NULL)
 650                         continue;
 651
 652                 sublock = sub->lss_cl.cls_lock;
 653                 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
 654                 if (rc == 0) {
 655                         if (lck->lls_sub[i].sub_flags & LSF_HELD) {
 656                                 LASSERT(sublock->cll_state == CLS_HELD);
 657                                 rc = cl_unuse_try(subenv->lse_env, sublock);
 658                                 if (rc != CLO_WAIT)
 659                                         rc = lov_sublock_release(env, lck,
 660                                                                  i, 0, rc);
 661                         }
 662                         lov_sublock_unlock(env, sub, closure, subenv);
 663                 }
 664                 result = lov_subresult(result, rc);
 665                 if (result < 0)
 666                         break;
 667         }
 668         if (result == 0 && lck->lls_unuse_race) {
 669                 lck->lls_unuse_race = 0;
 670                 result = -ESTALE;
 671         }
 672         cl_lock_closure_fini(closure);
 673         RETURN(result);
 674 }
 675
 676 static int lov_lock_wait(const struct lu_env *env,
 677                          const struct cl_lock_slice *slice)
 678 {
 679         struct lov_lock        *lck     = cl2lov_lock(slice);
 680         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
 681         enum cl_lock_state      minstate;
 682         int                     result;
 683         int                     i;
 684
 685         ENTRY;
 686
 687         for (result = 0, minstate = CLS_FREEING, i = 0; i < lck->lls_nr; ++i) {
 688                 int rc;
 689                 struct lovsub_lock     *sub;
 690                 struct cl_lock         *sublock;
 691                 struct lov_lock_sub    *lls;
 692                 struct lov_sublock_env *subenv;
 693
 694                 lls = &lck->lls_sub[i];
 695                 sub = lls->sub_lock;
 696                 LASSERT(sub != NULL);
 697                 sublock = sub->lss_cl.cls_lock;
 698                 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
 699                 if (rc == 0) {
 700                         LASSERT(sublock->cll_state >= CLS_ENQUEUED);
 701                         if (sublock->cll_state < CLS_HELD)
 702                                 rc = cl_wait_try(env, sublock);
 703
 704                         minstate = min(minstate, sublock->cll_state);
 705                         lov_sublock_unlock(env, sub, closure, subenv);
 706                 }
 707                 result = lov_subresult(result, rc);
 708                 if (result < 0)
 709                         break;
 710         }
 711         cl_lock_closure_fini(closure);
 712         RETURN(result ?: minstate >= CLS_HELD ? 0 : CLO_WAIT);
 713 }
 714
 715 static int lov_lock_use(const struct lu_env *env,
 716                         const struct cl_lock_slice *slice)
 717 {
 718         struct lov_lock        *lck     = cl2lov_lock(slice);
 719         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
 720         int                     result;
 721         int                     i;
 722
 723         LASSERT(slice->cls_lock->cll_state == CLS_CACHED);
 724         ENTRY;
 725
 726         for (result = 0, i = 0; i < lck->lls_nr; ++i) {
 727                 int rc;
 728                 struct lovsub_lock     *sub;
 729                 struct cl_lock         *sublock;
 730                 struct lov_lock_sub    *lls;
 731                 struct lov_sublock_env *subenv;
 732
 733                 if (slice->cls_lock->cll_state != CLS_CACHED) {
 734                         /* see comment in lov_lock_enqueue(). */
 735                         LASSERT(i > 0 && result != 0);
 736                         break;
 737                 }
 738                 /*
 739                  * if a sub-lock was destroyed while top-lock was in
 740                  * CLS_CACHED state, top-lock would have been moved into
 741                  * CLS_NEW state, so all sub-locks have to be in place.
 742                  */
 743                 lls = &lck->lls_sub[i];
 744                 sub = lls->sub_lock;
 745                 LASSERT(sub != NULL);
 746                 sublock = sub->lss_cl.cls_lock;
 747                 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
 748                 if (rc == 0) {
 749                         LASSERT(sublock->cll_state != CLS_FREEING);
 750                         lov_sublock_hold(env, lck, i);
 751                         if (sublock->cll_state == CLS_CACHED) {
 752                                 rc = cl_use_try(subenv->lse_env, sublock);
 753                                 if (rc != 0)
 754                                         rc = lov_sublock_release(env, lck,
 755                                                                  i, 1, rc);
 756                         } else
 757                                 rc = 0;
 758                         lov_sublock_unlock(env, sub, closure, subenv);
 759                 }
 760                 result = lov_subresult(result, rc);
 761                 if (result < 0)
 762                         break;
 763         }
 764         cl_lock_closure_fini(closure);
 765         RETURN(result);
 766 }
 767
 768 #if 0
 769 static int lock_lock_multi_match()
 770 {
 771         struct cl_lock          *lock    = slice->cls_lock;
 772         struct cl_lock_descr    *subneed = &lov_env_info(env)->lti_ldescr;
 773         struct lov_object       *loo     = cl2lov(lov->lls_cl.cls_obj);
 774         struct lov_layout_raid0 *r0      = lov_r0(loo);
 775         struct lov_lock_sub     *sub;
 776         struct cl_object        *subobj;
 777         obd_off  fstart;
 778         obd_off  fend;
 779         obd_off  start;
 780         obd_off  end;
 781         int i;
 782
 783         fstart = cl_offset(need->cld_obj, need->cld_start);
 784         fend   = cl_offset(need->cld_obj, need->cld_end + 1) - 1;
 785         subneed->cld_mode = need->cld_mode;
 786         cl_lock_mutex_get(env, lock);
 787         for (i = 0; i < lov->lls_nr; ++i) {
 788                 sub = &lov->lls_sub[i];
 789                 if (sub->sub_lock == NULL)
 790                         continue;
 791                 subobj = sub->sub_descr.cld_obj;
 792                 if (!lov_stripe_intersects(r0->lo_lsm, sub->sub_stripe,
 793                                            fstart, fend, &start, &end))
 794                         continue;
 795                 subneed->cld_start = cl_index(subobj, start);
 796                 subneed->cld_end   = cl_index(subobj, end);
 797                 subneed->cld_obj   = subobj;
 798                 if (!cl_lock_ext_match(&sub->sub_got, subneed)) {
 799                         result = 0;
 800                         break;
 801                 }
 802         }
 803         cl_lock_mutex_put(env, lock);
 804 }
 805 #endif
 806
 807 /**
 808  * Check if the extent region \a descr is covered by \a child against the
 809  * specific \a stripe.
 810  */
 811 static int lov_lock_stripe_is_matching(const struct lu_env *env,
 812                                        struct lov_object *lov, int stripe,
 813                                        const struct cl_lock_descr *child,
 814                                        const struct cl_lock_descr *descr)
 815 {
 816         struct lov_stripe_md *lsm = lov_r0(lov)->lo_lsm;
 817         obd_off start;
 818         obd_off end;
 819         int result;
 820
 821         if (lov_r0(lov)->lo_nr == 1)
 822                 return cl_lock_ext_match(child, descr);
 823
 824         /*
 825          * For a multi-stripes object:
 826          * - make sure the descr only covers child's stripe, and
 827          * - check if extent is matching.
 828          */
 829         start = cl_offset(&lov->lo_cl, descr->cld_start);
 830         end   = cl_offset(&lov->lo_cl, descr->cld_end + 1) - 1;
 831         result = end - start <= lsm->lsm_stripe_size &&
 832                  stripe == lov_stripe_number(lsm, start) &&
 833                  stripe == lov_stripe_number(lsm, end);
 834         if (result) {
 835                 struct cl_lock_descr *subd = &lov_env_info(env)->lti_ldescr;
 836                 obd_off sub_start;
 837                 obd_off sub_end;
 838
 839                 subd->cld_obj  = NULL;   /* don't need sub object at all */
 840                 subd->cld_mode = descr->cld_mode;
 841                 subd->cld_gid  = descr->cld_gid;
 842                 result = lov_stripe_intersects(lsm, stripe, start, end,
 843                                                &sub_start, &sub_end);
 844                 LASSERT(result);
 845                 subd->cld_start = cl_index(child->cld_obj, sub_start);
 846                 subd->cld_end   = cl_index(child->cld_obj, sub_end);
 847                 result = cl_lock_ext_match(child, subd);
 848         }
 849         return result;
 850 }
 851
 852 /**
 853  * An implementation of cl_lock_operations::clo_fits_into() method.
 854  *
 855  * Checks whether a lock (given by \a slice) is suitable for \a
 856  * io. Multi-stripe locks can be used only for "quick" io, like truncate, or
 857  * O_APPEND write.
 858  *
 859  * \see ccc_lock_fits_into().
 860  */
 861 static int lov_lock_fits_into(const struct lu_env *env,
 862                               const struct cl_lock_slice *slice,
 863                               const struct cl_lock_descr *need,
 864                               const struct cl_io *io)
 865 {
 866         struct lov_lock   *lov = cl2lov_lock(slice);
 867         struct lov_object *obj = cl2lov(slice->cls_obj);
 868         int result;
 869
 870         LASSERT(cl_object_same(need->cld_obj, slice->cls_obj));
 871         LASSERT(lov->lls_nr > 0);
 872
 873         ENTRY;
 874
 875         if (need->cld_mode == CLM_GROUP)
 876                 /*
 877                  * always allow to match group lock.
 878                  */
 879                 result = cl_lock_ext_match(&lov->lls_orig, need);
 880         else if (lov->lls_nr == 1) {
 881                 struct cl_lock_descr *got = &lov->lls_sub[0].sub_got;
 882                 result = lov_lock_stripe_is_matching(env,
 883                                                      cl2lov(slice->cls_obj),
 884                                                      lov->lls_sub[0].sub_stripe,
 885                                                      got, need);
 886         } else if (io->ci_type != CIT_TRUNC && io->ci_type != CIT_MISC &&
 887                    !cl_io_is_append(io) && need->cld_mode != CLM_PHANTOM)
 888                 /*
 889                  * Multi-stripe locks are only suitable for `quick' IO and for
 890                  * glimpse.
 891                  */
 892                 result = 0;
 893         else
 894                 /*
 895                  * Most general case: multi-stripe existing lock, and
 896                  * (potentially) multi-stripe @need lock. Check that @need is
 897                  * covered by @lov's sub-locks.
 898                  *
 899                  * For now, ignore lock expansions made by the server, and
 900                  * match against original lock extent.
 901                  */
 902                 result = cl_lock_ext_match(&lov->lls_orig, need);
 903         CDEBUG(D_DLMTRACE, DDESCR"/"DDESCR" %i %i/%i: %i\n",
 904                PDESCR(&lov->lls_orig), PDESCR(&lov->lls_sub[0].sub_got),
 905                lov->lls_sub[0].sub_stripe, lov->lls_nr, lov_r0(obj)->lo_nr,
 906                result);
 907         RETURN(result);
 908 }
 909
 910 void lov_lock_unlink(const struct lu_env *env,
 911                      struct lov_lock_link *link, struct lovsub_lock *sub)
 912 {
 913         struct lov_lock *lck    = link->lll_super;
 914         struct cl_lock  *parent = lck->lls_cl.cls_lock;
 915
 916         LASSERT(cl_lock_is_mutexed(parent));
 917         LASSERT(cl_lock_is_mutexed(sub->lss_cl.cls_lock));
 918         ENTRY;
 919
 920         list_del_init(&link->lll_list);
 921         LASSERT(lck->lls_sub[link->lll_idx].sub_lock == sub);
 922         /* yank this sub-lock from parent's array */
 923         lck->lls_sub[link->lll_idx].sub_lock = NULL;
 924         LASSERT(lck->lls_nr_filled > 0);
 925         lck->lls_nr_filled--;
 926         lu_ref_del(&parent->cll_reference, "lov-child", sub->lss_cl.cls_lock);
 927         cl_lock_put(env, parent);
 928         OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
 929         EXIT;
 930 }
 931
 932 struct lov_lock_link *lov_lock_link_find(const struct lu_env *env,
 933                                          struct lov_lock *lck,
 934                                          struct lovsub_lock *sub)
 935 {
 936         struct lov_lock_link *scan;
 937
 938         LASSERT(cl_lock_is_mutexed(sub->lss_cl.cls_lock));
 939         ENTRY;
 940
 941         list_for_each_entry(scan, &sub->lss_parents, lll_list) {
 942                 if (scan->lll_super == lck)
 943                         RETURN(scan);
 944         }
 945         RETURN(NULL);
 946 }
 947
 948 /**
 949  * An implementation of cl_lock_operations::clo_delete() method. This is
 950  * invoked for "top-to-bottom" delete, when lock destruction starts from the
 951  * top-lock, e.g., as a result of inode destruction.
 952  *
 953  * Unlinks top-lock from all its sub-locks. Sub-locks are not deleted there:
 954  * this is done separately elsewhere:
 955  *
 956  *     - for inode destruction, lov_object_delete() calls cl_object_kill() for
 957  *       each sub-object, purging its locks;
 958  *
 959  *     - in other cases (e.g., a fatal error with a top-lock) sub-locks are
 960  *       left in the cache.
 961  */
 962 static void lov_lock_delete(const struct lu_env *env,
 963                             const struct cl_lock_slice *slice)
 964 {
 965         struct lov_lock        *lck     = cl2lov_lock(slice);
 966         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
 967         int i;
 968
 969         LASSERT(slice->cls_lock->cll_state == CLS_FREEING);
 970         ENTRY;
 971
 972         for (i = 0; i < lck->lls_nr; ++i) {
 973                 struct lov_lock_sub *lls;
 974                 struct lovsub_lock  *lsl;
 975                 struct cl_lock      *sublock;
 976                 int rc;
 977
 978                 lls = &lck->lls_sub[i];
 979                 lsl = lls->sub_lock;
 980                 if (lsl == NULL)
 981                         continue;
 982
 983                 sublock = lsl->lss_cl.cls_lock;
 984                 rc = lov_sublock_lock(env, lck, lls, closure, NULL);
 985                 if (rc == 0) {
 986                         if (lck->lls_sub[i].sub_flags & LSF_HELD)
 987                                 lov_sublock_release(env, lck, i, 1, 0);
 988                         if (sublock->cll_state < CLS_FREEING) {
 989                                 struct lov_lock_link *link;
 990
 991                                 link = lov_lock_link_find(env, lck, lsl);
 992                                 LASSERT(link != NULL);
 993                                 lov_lock_unlink(env, link, lsl);
 994                                 LASSERT(lck->lls_sub[i].sub_lock == NULL);
 995                         }
 996                         lov_sublock_unlock(env, lsl, closure, NULL);
 997                 } else if (rc == CLO_REPEAT) {
 998                         --i; /* repeat with this lock */
 999                 } else {
1000                         CL_LOCK_DEBUG(D_ERROR, env, sublock,
1001                                       "Cannot get sub-lock for delete: %i\n",
1002                                       rc);
1003                 }
1004         }
1005         cl_lock_closure_fini(closure);
1006         EXIT;
1007 }
1008
1009 static int lov_lock_print(const struct lu_env *env, void *cookie,
1010                           lu_printer_t p, const struct cl_lock_slice *slice)
1011 {
1012         struct lov_lock *lck = cl2lov_lock(slice);
1013         int              i;
1014
1015         (*p)(env, cookie, "%d\n", lck->lls_nr);
1016         for (i = 0; i < lck->lls_nr; ++i) {
1017                 struct lov_lock_sub *sub;
1018
1019                 sub = &lck->lls_sub[i];
1020                 (*p)(env, cookie, "    %d %x: ", i, sub->sub_flags);
1021                 if (sub->sub_lock != NULL)
1022                         cl_lock_print(env, cookie, p,
1023                                       sub->sub_lock->lss_cl.cls_lock);
1024                 else
1025                         (*p)(env, cookie, "---\n");
1026         }
1027         return 0;
1028 }
1029
1030 static const struct cl_lock_operations lov_lock_ops = {
1031         .clo_fini      = lov_lock_fini,
1032         .clo_enqueue   = lov_lock_enqueue,
1033         .clo_wait      = lov_lock_wait,
1034         .clo_use       = lov_lock_use,
1035         .clo_unuse     = lov_lock_unuse,
1036         .clo_fits_into = lov_lock_fits_into,
1037         .clo_delete    = lov_lock_delete,
1038         .clo_print     = lov_lock_print
1039 };
1040
1041 int lov_lock_init_raid0(const struct lu_env *env, struct cl_object *obj,
1042                         struct cl_lock *lock, const struct cl_io *io)
1043 {
1044         struct lov_lock *lck;
1045         int result;
1046
1047         ENTRY;
1048         OBD_SLAB_ALLOC_PTR_GFP(lck, lov_lock_kmem, CFS_ALLOC_IO);
1049         if (lck != NULL) {
1050                 cl_lock_slice_add(lock, &lck->lls_cl, obj, &lov_lock_ops);
1051                 result = lov_lock_sub_init(env, lck, io);
1052         } else
1053                 result = -ENOMEM;
1054         RETURN(result);
1055 }
1056
1057 static struct cl_lock_closure *lov_closure_get(const struct lu_env *env,
1058                                                struct cl_lock *parent)
1059 {
1060         struct cl_lock_closure *closure;
1061
1062         closure = &lov_env_info(env)->lti_closure;
1063         LASSERT(list_empty(&closure->clc_list));
1064         cl_lock_closure_init(env, closure, parent, 1);
1065         return closure;
1066 }
1067
1068
1069 /** @} lov */