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                         descr->cld_gid   = parent->cll_descr.cld_gid;
 334                         /* XXX has no effect */
 335                         lck->lls_sub[nr].sub_got = *descr;
 336                         lck->lls_sub[nr].sub_stripe = stripe;
 337                         nr++;
 338                 }
 339         }
 340         LASSERT(nr == lck->lls_nr);
 341         /*
 342          * Then, create sub-locks. Once at least one sub-lock was created,
 343          * top-lock can be reached by other threads.
 344          */
 345         for (i = 0; i < lck->lls_nr; ++i) {
 346                 struct cl_lock       *sublock;
 347                 struct lov_lock_link *link;
 348
 349                 if (lck->lls_sub[i].sub_lock == NULL) {
 350                         sublock = lov_sublock_alloc(env, io, lck, i, &link);
 351                         if (IS_ERR(sublock)) {
 352                                 result = PTR_ERR(sublock);
 353                                 break;
 354                         }
 355                         cl_lock_mutex_get(env, sublock);
 356                         cl_lock_mutex_get(env, parent);
 357                         /*
 358                          * recheck under mutex that sub-lock wasn't created
 359                          * concurrently, and that top-lock is still alive.
 360                          */
 361                         if (lck->lls_sub[i].sub_lock == NULL &&
 362                             parent->cll_state < CLS_FREEING) {
 363                                 lov_sublock_adopt(env, lck, sublock, i, link);
 364                                 cl_lock_mutex_put(env, parent);
 365                         } else {
 366                                 cl_lock_mutex_put(env, parent);
 367                                 cl_lock_unhold(env, sublock,
 368                                                "lov-parent", parent);
 369                         }
 370                         cl_lock_mutex_put(env, sublock);
 371                 }
 372         }
 373         /*
 374          * Some sub-locks can be missing at this point. This is not a problem,
 375          * because enqueue will create them anyway. Main duty of this function
 376          * is to fill in sub-lock descriptions in a race free manner.
 377          */
 378         RETURN(result);
 379 }
 380
 381 static int lov_sublock_release(const struct lu_env *env, struct lov_lock *lck,
 382                                int i, int deluser, int rc)
 383 {
 384         struct cl_lock *parent = lck->lls_cl.cls_lock;
 385
 386         LASSERT(cl_lock_is_mutexed(parent));
 387         ENTRY;
 388
 389         if (lck->lls_sub[i].sub_flags & LSF_HELD) {
 390                 struct cl_lock    *sublock;
 391                 int dying;
 392
 393                 LASSERT(lck->lls_sub[i].sub_lock != NULL);
 394                 sublock = lck->lls_sub[i].sub_lock->lss_cl.cls_lock;
 395                 LASSERT(cl_lock_is_mutexed(sublock));
 396
 397                 lck->lls_sub[i].sub_flags &= ~LSF_HELD;
 398                 if (deluser)
 399                         cl_lock_user_del(env, sublock);
 400                 /*
 401                  * If the last hold is released, and cancellation is pending
 402                  * for a sub-lock, release parent mutex, to avoid keeping it
 403                  * while sub-lock is being paged out.
 404                  */
 405                 dying = (sublock->cll_descr.cld_mode == CLM_PHANTOM ||
 406                          sublock->cll_descr.cld_mode == CLM_GROUP ||
 407                          (sublock->cll_flags & (CLF_CANCELPEND|CLF_DOOMED))) &&
 408                         sublock->cll_holds == 1;
 409                 if (dying)
 410                         cl_lock_mutex_put(env, parent);
 411                 cl_lock_unhold(env, sublock, "lov-parent", parent);
 412                 if (dying) {
 413                         cl_lock_mutex_get(env, parent);
 414                         rc = lov_subresult(rc, CLO_REPEAT);
 415                 }
 416                 /*
 417                  * From now on lck->lls_sub[i].sub_lock is a "weak" pointer,
 418                  * not backed by a reference on a
 419                  * sub-lock. lovsub_lock_delete() will clear
 420                  * lck->lls_sub[i].sub_lock under semaphores, just before
 421                  * sub-lock is destroyed.
 422                  */
 423         }
 424         RETURN(rc);
 425 }
 426
 427 static void lov_sublock_hold(const struct lu_env *env, struct lov_lock *lck,
 428                              int i)
 429 {
 430         struct cl_lock *parent = lck->lls_cl.cls_lock;
 431
 432         LASSERT(cl_lock_is_mutexed(parent));
 433         ENTRY;
 434
 435         if (!(lck->lls_sub[i].sub_flags & LSF_HELD)) {
 436                 struct cl_lock *sublock;
 437
 438                 LASSERT(lck->lls_sub[i].sub_lock != NULL);
 439                 sublock = lck->lls_sub[i].sub_lock->lss_cl.cls_lock;
 440                 LASSERT(cl_lock_is_mutexed(sublock));
 441                 LASSERT(sublock->cll_state != CLS_FREEING);
 442
 443                 lck->lls_sub[i].sub_flags |= LSF_HELD;
 444
 445                 cl_lock_get_trust(sublock);
 446                 cl_lock_hold_add(env, sublock, "lov-parent", parent);
 447                 cl_lock_user_add(env, sublock);
 448                 cl_lock_put(env, sublock);
 449         }
 450         EXIT;
 451 }
 452
 453 static void lov_lock_fini(const struct lu_env *env,
 454                           struct cl_lock_slice *slice)
 455 {
 456         struct lov_lock *lck;
 457         int i;
 458
 459         ENTRY;
 460         lck = cl2lov_lock(slice);
 461         LASSERT(lck->lls_nr_filled == 0);
 462         if (lck->lls_sub != NULL) {
 463                 for (i = 0; i < lck->lls_nr; ++i)
 464                         /*
 465                          * No sub-locks exists at this point, as sub-lock has
 466                          * a reference on its parent.
 467                          */
 468                         LASSERT(lck->lls_sub[i].sub_lock == NULL);
 469                 OBD_FREE(lck->lls_sub, lck->lls_nr * sizeof lck->lls_sub[0]);
 470         }
 471         OBD_SLAB_FREE_PTR(lck, lov_lock_kmem);
 472         EXIT;
 473 }
 474
 475 /**
 476  * Tries to advance a state machine of a given sub-lock toward enqueuing of
 477  * the top-lock.
 478  *
 479  * \retval 0 if state-transition can proceed
 480  * \retval -ve otherwise.
 481  */
 482 static int lov_lock_enqueue_one(const struct lu_env *env, struct lov_lock *lck,
 483                                 struct cl_lock *sublock,
 484                                 struct cl_io *io, __u32 enqflags, int last)
 485 {
 486         int result;
 487         ENTRY;
 488
 489         /* first, try to enqueue a sub-lock ... */
 490         result = cl_enqueue_try(env, sublock, io, enqflags);
 491         if (sublock->cll_state == CLS_ENQUEUED)
 492                 /* if it is enqueued, try to `wait' on it---maybe it's already
 493                  * granted */
 494                 result = cl_wait_try(env, sublock);
 495         /*
 496          * If CEF_ASYNC flag is set, then all sub-locks can be enqueued in
 497          * parallel, otherwise---enqueue has to wait until sub-lock is granted
 498          * before proceeding to the next one.
 499          */
 500         if (result == CLO_WAIT && sublock->cll_state <= CLS_HELD &&
 501             enqflags & CEF_ASYNC && !last)
 502                 result = 0;
 503         RETURN(result);
 504 }
 505
 506 /**
 507  * Helper function for lov_lock_enqueue() that creates missing sub-lock.
 508  */
 509 static int lov_sublock_fill(const struct lu_env *env, struct cl_lock *parent,
 510                             struct cl_io *io, struct lov_lock *lck, int idx)
 511 {
 512         struct lov_lock_link *link;
 513         struct cl_lock       *sublock;
 514         int                   result;
 515
 516         LASSERT(parent->cll_depth == 1);
 517         cl_lock_mutex_put(env, parent);
 518         sublock = lov_sublock_alloc(env, io, lck, idx, &link);
 519         if (!IS_ERR(sublock))
 520                 cl_lock_mutex_get(env, sublock);
 521         cl_lock_mutex_get(env, parent);
 522
 523         if (!IS_ERR(sublock)) {
 524                 if (parent->cll_state == CLS_QUEUING &&
 525                     lck->lls_sub[idx].sub_lock == NULL)
 526                         lov_sublock_adopt(env, lck, sublock, idx, link);
 527                 else {
 528                         /* other thread allocated sub-lock, or enqueue is no
 529                          * longer going on */
 530                         cl_lock_mutex_put(env, parent);
 531                         cl_lock_unhold(env, sublock, "lov-parent", parent);
 532                         cl_lock_mutex_get(env, parent);
 533                 }
 534                 cl_lock_mutex_put(env, sublock);
 535                 result = CLO_REPEAT;
 536         } else
 537                 result = PTR_ERR(sublock);
 538         return result;
 539 }
 540
 541 /**
 542  * Implementation of cl_lock_operations::clo_enqueue() for lov layer. This
 543  * function is rather subtle, as it enqueues top-lock (i.e., advances top-lock
 544  * state machine from CLS_QUEUING to CLS_ENQUEUED states) by juggling sub-lock
 545  * state machines in the face of sub-locks sharing (by multiple top-locks),
 546  * and concurrent sub-lock cancellations.
 547  */
 548 static int lov_lock_enqueue(const struct lu_env *env,
 549                             const struct cl_lock_slice *slice,
 550                             struct cl_io *io, __u32 enqflags)
 551 {
 552         struct cl_lock         *lock    = slice->cls_lock;
 553         struct lov_lock        *lck     = cl2lov_lock(slice);
 554         struct cl_lock_closure *closure = lov_closure_get(env, lock);
 555         int i;
 556         int result;
 557         enum cl_lock_state minstate;
 558
 559         ENTRY;
 560
 561         for (result = 0, minstate = CLS_FREEING, i = 0; i < lck->lls_nr; ++i) {
 562                 int rc;
 563                 struct lovsub_lock     *sub;
 564                 struct lov_lock_sub    *lls;
 565                 struct cl_lock         *sublock;
 566                 struct lov_sublock_env *subenv;
 567
 568                 if (lock->cll_state != CLS_QUEUING) {
 569                         /*
 570                          * Lock might have left QUEUING state if previous
 571                          * iteration released its mutex. Stop enqueing in this
 572                          * case and let the upper layer to decide what to do.
 573                          */
 574                         LASSERT(i > 0 && result != 0);
 575                         break;
 576                 }
 577
 578                 lls = &lck->lls_sub[i];
 579                 sub = lls->sub_lock;
 580                 /*
 581                  * Sub-lock might have been canceled, while top-lock was
 582                  * cached.
 583                  */
 584                 if (sub == NULL) {
 585                         result = lov_sublock_fill(env, lock, io, lck, i);
 586                         /* lov_sublock_fill() released @lock mutex,
 587                          * restart. */
 588                         break;
 589                 }
 590                 sublock = sub->lss_cl.cls_lock;
 591                 rc = lov_sublock_lock(env, lls, closure, &subenv);
 592                 if (rc == 0) {
 593                         lov_sublock_hold(env, lck, i);
 594                         rc = lov_lock_enqueue_one(subenv->lse_env, lck, sublock,
 595                                                   subenv->lse_io, enqflags,
 596                                                   i == lck->lls_nr - 1);
 597                         minstate = min(minstate, sublock->cll_state);
 598                         /*
 599                          * Don't hold a sub-lock in CLS_CACHED state, see
 600                          * description for lov_lock::lls_sub.
 601                          */
 602                         if (sublock->cll_state > CLS_HELD)
 603                                 rc = lov_sublock_release(env, lck, i, 1, rc);
 604                         lov_sublock_unlock(env, sub, closure, subenv);
 605                 }
 606                 result = lov_subresult(result, rc);
 607                 if (result < 0)
 608                         break;
 609         }
 610         cl_lock_closure_fini(closure);
 611         RETURN(result ?: minstate >= CLS_ENQUEUED ? 0 : CLO_WAIT);
 612 }
 613
 614 static int lov_lock_unuse(const struct lu_env *env,
 615                           const struct cl_lock_slice *slice)
 616 {
 617         struct lov_lock        *lck     = cl2lov_lock(slice);
 618         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
 619         int i;
 620         int result;
 621
 622         ENTRY;
 623
 624         for (result = 0, i = 0; i < lck->lls_nr; ++i) {
 625                 int rc;
 626                 struct lovsub_lock     *sub;
 627                 struct cl_lock         *sublock;
 628                 struct lov_lock_sub    *lls;
 629                 struct lov_sublock_env *subenv;
 630
 631                 /* top-lock state cannot change concurrently, because single
 632                  * thread (one that released the last hold) carries unlocking
 633                  * to the completion. */
 634                 LASSERT(slice->cls_lock->cll_state == CLS_UNLOCKING);
 635                 lls = &lck->lls_sub[i];
 636                 sub = lls->sub_lock;
 637                 if (sub == NULL)
 638                         continue;
 639
 640                 sublock = sub->lss_cl.cls_lock;
 641                 rc = lov_sublock_lock(env, lls, closure, &subenv);
 642                 if (rc == 0) {
 643                         if (lck->lls_sub[i].sub_flags & LSF_HELD) {
 644                                 LASSERT(sublock->cll_state == CLS_HELD);
 645                                 rc = cl_unuse_try(subenv->lse_env, sublock);
 646                                 if (rc != CLO_WAIT)
 647                                         rc = lov_sublock_release(env, lck,
 648                                                                  i, 0, rc);
 649                         }
 650                         lov_sublock_unlock(env, sub, closure, subenv);
 651                 }
 652                 result = lov_subresult(result, rc);
 653                 if (result < 0)
 654                         break;
 655         }
 656         if (result == 0 && lck->lls_unuse_race) {
 657                 lck->lls_unuse_race = 0;
 658                 result = -ESTALE;
 659         }
 660         cl_lock_closure_fini(closure);
 661         RETURN(result);
 662 }
 663
 664 static int lov_lock_wait(const struct lu_env *env,
 665                          const struct cl_lock_slice *slice)
 666 {
 667         struct lov_lock        *lck     = cl2lov_lock(slice);
 668         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
 669         enum cl_lock_state      minstate;
 670         int                     result;
 671         int                     i;
 672
 673         ENTRY;
 674
 675         for (result = 0, minstate = CLS_FREEING, i = 0; i < lck->lls_nr; ++i) {
 676                 int rc;
 677                 struct lovsub_lock     *sub;
 678                 struct cl_lock         *sublock;
 679                 struct lov_lock_sub    *lls;
 680                 struct lov_sublock_env *subenv;
 681
 682                 lls = &lck->lls_sub[i];
 683                 sub = lls->sub_lock;
 684                 LASSERT(sub != NULL);
 685                 sublock = sub->lss_cl.cls_lock;
 686                 rc = lov_sublock_lock(env, lls, closure, &subenv);
 687                 if (rc == 0) {
 688                         LASSERT(sublock->cll_state >= CLS_ENQUEUED);
 689                         if (sublock->cll_state < CLS_HELD)
 690                                 rc = cl_wait_try(env, sublock);
 691
 692                         minstate = min(minstate, sublock->cll_state);
 693                         lov_sublock_unlock(env, sub, closure, subenv);
 694                 }
 695                 result = lov_subresult(result, rc);
 696                 if (result < 0)
 697                         break;
 698         }
 699         cl_lock_closure_fini(closure);
 700         RETURN(result ?: minstate >= CLS_HELD ? 0 : CLO_WAIT);
 701 }
 702
 703 static int lov_lock_use(const struct lu_env *env,
 704                         const struct cl_lock_slice *slice)
 705 {
 706         struct lov_lock        *lck     = cl2lov_lock(slice);
 707         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
 708         int                     result;
 709         int                     i;
 710
 711         LASSERT(slice->cls_lock->cll_state == CLS_CACHED);
 712         ENTRY;
 713
 714         for (result = 0, i = 0; i < lck->lls_nr; ++i) {
 715                 int rc;
 716                 struct lovsub_lock     *sub;
 717                 struct cl_lock         *sublock;
 718                 struct lov_lock_sub    *lls;
 719                 struct lov_sublock_env *subenv;
 720
 721                 if (slice->cls_lock->cll_state != CLS_CACHED) {
 722                         /* see comment in lov_lock_enqueue(). */
 723                         LASSERT(i > 0 && result != 0);
 724                         break;
 725                 }
 726                 /*
 727                  * if a sub-lock was destroyed while top-lock was in
 728                  * CLS_CACHED state, top-lock would have been moved into
 729                  * CLS_NEW state, so all sub-locks have to be in place.
 730                  */
 731                 lls = &lck->lls_sub[i];
 732                 sub = lls->sub_lock;
 733                 LASSERT(sub != NULL);
 734                 sublock = sub->lss_cl.cls_lock;
 735                 rc = lov_sublock_lock(env, lls, closure, &subenv);
 736                 if (rc == 0) {
 737                         LASSERT(sublock->cll_state != CLS_FREEING);
 738                         lov_sublock_hold(env, lck, i);
 739                         if (sublock->cll_state == CLS_CACHED) {
 740                                 rc = cl_use_try(subenv->lse_env, sublock);
 741                                 if (rc != 0)
 742                                         rc = lov_sublock_release(env, lck,
 743                                                                  i, 1, rc);
 744                         } else
 745                                 rc = 0;
 746                         lov_sublock_unlock(env, sub, closure, subenv);
 747                 }
 748                 result = lov_subresult(result, rc);
 749                 if (result < 0)
 750                         break;
 751         }
 752         cl_lock_closure_fini(closure);
 753         RETURN(result);
 754 }
 755
 756 #if 0
 757 static int lock_lock_multi_match()
 758 {
 759         struct cl_lock          *lock    = slice->cls_lock;
 760         struct cl_lock_descr    *subneed = &lov_env_info(env)->lti_ldescr;
 761         struct lov_object       *loo     = cl2lov(lov->lls_cl.cls_obj);
 762         struct lov_layout_raid0 *r0      = lov_r0(loo);
 763         struct lov_lock_sub     *sub;
 764         struct cl_object        *subobj;
 765         obd_off  fstart;
 766         obd_off  fend;
 767         obd_off  start;
 768         obd_off  end;
 769         int i;
 770
 771         fstart = cl_offset(need->cld_obj, need->cld_start);
 772         fend   = cl_offset(need->cld_obj, need->cld_end + 1) - 1;
 773         subneed->cld_mode = need->cld_mode;
 774         cl_lock_mutex_get(env, lock);
 775         for (i = 0; i < lov->lls_nr; ++i) {
 776                 sub = &lov->lls_sub[i];
 777                 if (sub->sub_lock == NULL)
 778                         continue;
 779                 subobj = sub->sub_descr.cld_obj;
 780                 if (!lov_stripe_intersects(r0->lo_lsm, sub->sub_stripe,
 781                                            fstart, fend, &start, &end))
 782                         continue;
 783                 subneed->cld_start = cl_index(subobj, start);
 784                 subneed->cld_end   = cl_index(subobj, end);
 785                 subneed->cld_obj   = subobj;
 786                 if (!cl_lock_ext_match(&sub->sub_got, subneed)) {
 787                         result = 0;
 788                         break;
 789                 }
 790         }
 791         cl_lock_mutex_put(env, lock);
 792 }
 793 #endif
 794
 795 /**
 796  * Check if the extent region \a descr is covered by \a child against the
 797  * specific \a stripe.
 798  */
 799 static int lov_lock_stripe_is_matching(const struct lu_env *env,
 800                                        struct lov_object *lov, int stripe,
 801                                        const struct cl_lock_descr *child,
 802                                        const struct cl_lock_descr *descr)
 803 {
 804         struct lov_stripe_md *lsm = lov_r0(lov)->lo_lsm;
 805         obd_off start;
 806         obd_off end;
 807         int result;
 808
 809         if (lov_r0(lov)->lo_nr == 1)
 810                 return cl_lock_ext_match(child, descr);
 811
 812         /*
 813          * For a multi-stripes object:
 814          * - make sure the descr only covers child's stripe, and
 815          * - check if extent is matching.
 816          */
 817         start = cl_offset(&lov->lo_cl, descr->cld_start);
 818         end   = cl_offset(&lov->lo_cl, descr->cld_end + 1) - 1;
 819         result = end - start <= lsm->lsm_stripe_size &&
 820                  stripe == lov_stripe_number(lsm, start) &&
 821                  stripe == lov_stripe_number(lsm, end);
 822         if (result) {
 823                 struct cl_lock_descr *subd = &lov_env_info(env)->lti_ldescr;
 824                 obd_off sub_start;
 825                 obd_off sub_end;
 826
 827                 subd->cld_obj  = NULL;   /* don't need sub object at all */
 828                 subd->cld_mode = descr->cld_mode;
 829                 subd->cld_gid  = descr->cld_gid;
 830                 result = lov_stripe_intersects(lsm, stripe, start, end,
 831                                                &sub_start, &sub_end);
 832                 LASSERT(result);
 833                 subd->cld_start = cl_index(child->cld_obj, sub_start);
 834                 subd->cld_end   = cl_index(child->cld_obj, sub_end);
 835                 result = cl_lock_ext_match(child, subd);
 836         }
 837         return result;
 838 }
 839
 840 /**
 841  * An implementation of cl_lock_operations::clo_fits_into() method.
 842  *
 843  * Checks whether a lock (given by \a slice) is suitable for \a
 844  * io. Multi-stripe locks can be used only for "quick" io, like truncate, or
 845  * O_APPEND write.
 846  *
 847  * \see ccc_lock_fits_into().
 848  */
 849 static int lov_lock_fits_into(const struct lu_env *env,
 850                               const struct cl_lock_slice *slice,
 851                               const struct cl_lock_descr *need,
 852                               const struct cl_io *io)
 853 {
 854         struct lov_lock   *lov = cl2lov_lock(slice);
 855         struct lov_object *obj = cl2lov(slice->cls_obj);
 856         int result;
 857
 858         LASSERT(cl_object_same(need->cld_obj, slice->cls_obj));
 859         LASSERT(lov->lls_nr > 0);
 860
 861         ENTRY;
 862
 863         if (need->cld_mode == CLM_GROUP)
 864                 /*
 865                  * always allow to match group lock.
 866                  */
 867                 result = cl_lock_ext_match(&lov->lls_orig, need);
 868         else if (lov->lls_nr == 1) {
 869                 struct cl_lock_descr *got = &lov->lls_sub[0].sub_got;
 870                 result = lov_lock_stripe_is_matching(env,
 871                                                      cl2lov(slice->cls_obj),
 872                                                      lov->lls_sub[0].sub_stripe,
 873                                                      got, need);
 874         } else if (io->ci_type != CIT_TRUNC && io->ci_type != CIT_MISC &&
 875                    !cl_io_is_append(io) && need->cld_mode != CLM_PHANTOM)
 876                 /*
 877                  * Multi-stripe locks are only suitable for `quick' IO and for
 878                  * glimpse.
 879                  */
 880                 result = 0;
 881         else
 882                 /*
 883                  * Most general case: multi-stripe existing lock, and
 884                  * (potentially) multi-stripe @need lock. Check that @need is
 885                  * covered by @lov's sub-locks.
 886                  *
 887                  * For now, ignore lock expansions made by the server, and
 888                  * match against original lock extent.
 889                  */
 890                 result = cl_lock_ext_match(&lov->lls_orig, need);
 891         CDEBUG(D_DLMTRACE, DDESCR"/"DDESCR" %i %i/%i: %i\n",
 892                PDESCR(&lov->lls_orig), PDESCR(&lov->lls_sub[0].sub_got),
 893                lov->lls_sub[0].sub_stripe, lov->lls_nr, lov_r0(obj)->lo_nr,
 894                result);
 895         RETURN(result);
 896 }
 897
 898 void lov_lock_unlink(const struct lu_env *env,
 899                      struct lov_lock_link *link, struct lovsub_lock *sub)
 900 {
 901         struct lov_lock *lck    = link->lll_super;
 902         struct cl_lock  *parent = lck->lls_cl.cls_lock;
 903
 904         LASSERT(cl_lock_is_mutexed(parent));
 905         LASSERT(cl_lock_is_mutexed(sub->lss_cl.cls_lock));
 906         ENTRY;
 907
 908         list_del_init(&link->lll_list);
 909         LASSERT(lck->lls_sub[link->lll_idx].sub_lock == sub);
 910         /* yank this sub-lock from parent's array */
 911         lck->lls_sub[link->lll_idx].sub_lock = NULL;
 912         LASSERT(lck->lls_nr_filled > 0);
 913         lck->lls_nr_filled--;
 914         lu_ref_del(&parent->cll_reference, "lov-child", sub->lss_cl.cls_lock);
 915         cl_lock_put(env, parent);
 916         OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
 917         EXIT;
 918 }
 919
 920 struct lov_lock_link *lov_lock_link_find(const struct lu_env *env,
 921                                          struct lov_lock *lck,
 922                                          struct lovsub_lock *sub)
 923 {
 924         struct lov_lock_link *scan;
 925
 926         LASSERT(cl_lock_is_mutexed(sub->lss_cl.cls_lock));
 927         ENTRY;
 928
 929         list_for_each_entry(scan, &sub->lss_parents, lll_list) {
 930                 if (scan->lll_super == lck)
 931                         RETURN(scan);
 932         }
 933         RETURN(NULL);
 934 }
 935
 936 /**
 937  * An implementation of cl_lock_operations::clo_delete() method. This is
 938  * invoked for "top-to-bottom" delete, when lock destruction starts from the
 939  * top-lock, e.g., as a result of inode destruction.
 940  *
 941  * Unlinks top-lock from all its sub-locks. Sub-locks are not deleted there:
 942  * this is done separately elsewhere:
 943  *
 944  *     - for inode destruction, lov_object_delete() calls cl_object_kill() for
 945  *       each sub-object, purging its locks;
 946  *
 947  *     - in other cases (e.g., a fatal error with a top-lock) sub-locks are
 948  *       left in the cache.
 949  */
 950 static void lov_lock_delete(const struct lu_env *env,
 951                             const struct cl_lock_slice *slice)
 952 {
 953         struct lov_lock        *lck     = cl2lov_lock(slice);
 954         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
 955         int i;
 956
 957         LASSERT(slice->cls_lock->cll_state == CLS_FREEING);
 958         ENTRY;
 959
 960         for (i = 0; i < lck->lls_nr; ++i) {
 961                 struct lov_lock_sub *lls;
 962                 struct lovsub_lock  *lsl;
 963                 struct cl_lock      *sublock;
 964                 int rc;
 965
 966                 lls = &lck->lls_sub[i];
 967                 lsl = lls->sub_lock;
 968                 if (lsl == NULL)
 969                         continue;
 970
 971                 sublock = lsl->lss_cl.cls_lock;
 972                 rc = lov_sublock_lock(env, lls, closure, NULL);
 973                 if (rc == 0) {
 974                         if (lck->lls_sub[i].sub_flags & LSF_HELD)
 975                                 lov_sublock_release(env, lck, i, 1, 0);
 976                         if (sublock->cll_state < CLS_FREEING) {
 977                                 struct lov_lock_link *link;
 978
 979                                 link = lov_lock_link_find(env, lck, lsl);
 980                                 LASSERT(link != NULL);
 981                                 lov_lock_unlink(env, link, lsl);
 982                                 LASSERT(lck->lls_sub[i].sub_lock == NULL);
 983                         }
 984                         lov_sublock_unlock(env, lsl, closure, NULL);
 985                 } else if (rc == CLO_REPEAT) {
 986                         --i; /* repeat with this lock */
 987                 } else {
 988                         CL_LOCK_DEBUG(D_ERROR, env, sublock,
 989                                       "Cannot get sub-lock for delete: %i\n",
 990                                       rc);
 991                 }
 992         }
 993         cl_lock_closure_fini(closure);
 994         EXIT;
 995 }
 996
 997 static int lov_lock_print(const struct lu_env *env, void *cookie,
 998                           lu_printer_t p, const struct cl_lock_slice *slice)
 999 {
1000         struct lov_lock *lck = cl2lov_lock(slice);
1001         int              i;
1002
1003         (*p)(env, cookie, "%d\n", lck->lls_nr);
1004         for (i = 0; i < lck->lls_nr; ++i) {
1005                 struct lov_lock_sub *sub;
1006
1007                 sub = &lck->lls_sub[i];
1008                 (*p)(env, cookie, "    %d %x: ", i, sub->sub_flags);
1009                 if (sub->sub_lock != NULL)
1010                         cl_lock_print(env, cookie, p,
1011                                       sub->sub_lock->lss_cl.cls_lock);
1012                 else
1013                         (*p)(env, cookie, "---\n");
1014         }
1015         return 0;
1016 }
1017
1018 static const struct cl_lock_operations lov_lock_ops = {
1019         .clo_fini      = lov_lock_fini,
1020         .clo_enqueue   = lov_lock_enqueue,
1021         .clo_wait      = lov_lock_wait,
1022         .clo_use       = lov_lock_use,
1023         .clo_unuse     = lov_lock_unuse,
1024         .clo_fits_into = lov_lock_fits_into,
1025         .clo_delete    = lov_lock_delete,
1026         .clo_print     = lov_lock_print
1027 };
1028
1029 int lov_lock_init_raid0(const struct lu_env *env, struct cl_object *obj,
1030                         struct cl_lock *lock, const struct cl_io *io)
1031 {
1032         struct lov_lock *lck;
1033         int result;
1034
1035         ENTRY;
1036         OBD_SLAB_ALLOC_PTR_GFP(lck, lov_lock_kmem, CFS_ALLOC_IO);
1037         if (lck != NULL) {
1038                 cl_lock_slice_add(lock, &lck->lls_cl, obj, &lov_lock_ops);
1039                 result = lov_lock_sub_init(env, lck, io);
1040         } else
1041                 result = -ENOMEM;
1042         RETURN(result);
1043 }
1044
1045 static struct cl_lock_closure *lov_closure_get(const struct lu_env *env,
1046                                                struct cl_lock *parent)
1047 {
1048         struct cl_lock_closure *closure;
1049
1050         closure = &lov_env_info(env)->lti_closure;
1051         LASSERT(list_empty(&closure->clc_list));
1052         cl_lock_closure_init(env, closure, parent, 1);
1053         return closure;
1054 }
1055
1056
1057 /** @} lov */