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