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