lustre/lov/lov_lock.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Implementation of cl_lock for LOV layer.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_LOV
  42
  43 #include "lov_cl_internal.h"
  44
  45 /** \addtogroup lov lov @{ */
  46
  47 static struct cl_lock_closure *lov_closure_get(const struct lu_env *env,
  48                                                struct cl_lock *parent);
  49
  50 /*****************************************************************************
  51  *
  52  * Lov lock operations.
  53  *
  54  */
  55
  56 static struct lov_sublock_env *lov_sublock_env_get(const struct lu_env *env,
  57                                                    struct cl_lock *parent,
  58                                                    struct lov_lock_sub *lls)
  59 {
  60         struct lov_sublock_env *subenv;
  61         struct lov_io          *lio    = lov_env_io(env);
  62         struct cl_io           *io     = lio->lis_cl.cis_io;
  63         struct lov_io_sub      *sub;
  64
  65         subenv = &lov_env_session(env)->ls_subenv;
  66
  67         /*
  68          * FIXME: We tend to use the subio's env & io to call the sublock
  69          * lock operations because osc lock sometimes stores some control
  70          * variables in thread's IO infomation(Now only lockless information).
  71          * However, if the lock's host(object) is different from the object
  72          * for current IO, we have no way to get the subenv and subio because
  73          * they are not initialized at all. As a temp fix, in this case,
  74          * we still borrow the parent's env to call sublock operations.
  75          */
  76         if (!cl_object_same(io->ci_obj, parent->cll_descr.cld_obj)) {
  77                 subenv->lse_env = env;
  78                 subenv->lse_io  = io;
  79                 subenv->lse_sub = NULL;
  80         } else {
  81                 LASSERT(io != NULL);
  82                 sub = lov_sub_get(env, lio, lls->sub_stripe);
  83                 if (!IS_ERR(sub)) {
  84                         subenv->lse_env = sub->sub_env;
  85                         subenv->lse_io  = sub->sub_io;
  86                         subenv->lse_sub = sub;
  87                 } else {
  88                         subenv = (void*)sub;
  89                 }
  90         }
  91         return subenv;
  92 }
  93
  94 static void lov_sublock_env_put(struct lov_sublock_env *subenv)
  95 {
  96         if (subenv && subenv->lse_sub)
  97                 lov_sub_put(subenv->lse_sub);
  98 }
  99
 100 static void lov_sublock_adopt(const struct lu_env *env, struct lov_lock *lck,
 101                               struct cl_lock *sublock, int idx,
 102                               struct lov_lock_link *link)
 103 {
 104         struct lovsub_lock *lsl;
 105         struct cl_lock     *parent = lck->lls_cl.cls_lock;
 106         int                 rc;
 107
 108         LASSERT(cl_lock_is_mutexed(parent));
 109         LASSERT(cl_lock_is_mutexed(sublock));
 110         ENTRY;
 111
 112         lsl = cl2sub_lock(sublock);
 113         /*
 114          * check that sub-lock doesn't have lock link to this top-lock.
 115          */
 116         LASSERT(lov_lock_link_find(env, lck, lsl) == NULL);
 117         LASSERT(idx < lck->lls_nr);
 118
 119         lck->lls_sub[idx].sub_lock = lsl;
 120         lck->lls_nr_filled++;
 121         LASSERT(lck->lls_nr_filled <= lck->lls_nr);
 122         list_add_tail(&link->lll_list, &lsl->lss_parents);
 123         link->lll_idx = idx;
 124         link->lll_super = lck;
 125         cl_lock_get(parent);
 126         lu_ref_add(&parent->cll_reference, "lov-child", sublock);
 127         lck->lls_sub[idx].sub_flags |= LSF_HELD;
 128         cl_lock_user_add(env, sublock);
 129
 130         rc = lov_sublock_modify(env, lck, lsl, &sublock->cll_descr, idx);
 131         LASSERT(rc == 0); /* there is no way this can fail, currently */
 132         EXIT;
 133 }
 134
 135 static struct cl_lock *lov_sublock_alloc(const struct lu_env *env,
 136                                          const struct cl_io *io,
 137                                          struct lov_lock *lck,
 138                                          int idx, struct lov_lock_link **out)
 139 {
 140         struct cl_lock       *sublock;
 141         struct cl_lock       *parent;
 142         struct lov_lock_link *link;
 143
 144         LASSERT(idx < lck->lls_nr);
 145         ENTRY;
 146
 147         OBD_SLAB_ALLOC_PTR_GFP(link, lov_lock_link_kmem, CFS_ALLOC_IO);
 148         if (link != NULL) {
 149                 struct lov_sublock_env *subenv;
 150                 struct lov_lock_sub  *lls;
 151                 struct cl_lock_descr *descr;
 152
 153                 parent = lck->lls_cl.cls_lock;
 154                 lls    = &lck->lls_sub[idx];
 155                 descr  = &lls->sub_descr;
 156
 157                 subenv = lov_sublock_env_get(env, parent, lls);
 158                 if (!IS_ERR(subenv)) {
 159                         /* CAVEAT: Don't try to add a field in lov_lock_sub
 160                          * to remember the subio. This is because lock is able
 161                          * to be cached, but this is not true for IO. This
 162                          * further means a sublock might be referenced in
 163                          * different io context. -jay */
 164
 165                         sublock = cl_lock_hold(subenv->lse_env, subenv->lse_io,
 166                                                descr, "lov-parent", parent);
 167                         lov_sublock_env_put(subenv);
 168                 } else {
 169                         /* error occurs. */
 170                         sublock = (void*)subenv;
 171                 }
 172
 173                 if (!IS_ERR(sublock))
 174                         *out = link;
 175                 else
 176                         OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
 177         } else
 178                 sublock = ERR_PTR(-ENOMEM);
 179         RETURN(sublock);
 180 }
 181
 182 static void lov_sublock_unlock(const struct lu_env *env,
 183                                struct lovsub_lock *lsl,
 184                                struct cl_lock_closure *closure,
 185                                struct lov_sublock_env *subenv)
 186 {
 187         ENTRY;
 188         lov_sublock_env_put(subenv);
 189         lsl->lss_active = NULL;
 190         cl_lock_disclosure(env, closure);
 191         EXIT;
 192 }
 193
 194 static int lov_sublock_lock(const struct lu_env *env,
 195                             struct lov_lock *lck,
 196                             struct lov_lock_sub *lls,
 197                             struct cl_lock_closure *closure,
 198                             struct lov_sublock_env **lsep)
 199 {
 200         struct lovsub_lock *sublock;
 201         struct cl_lock     *child;
 202         int                 result = 0;
 203         ENTRY;
 204
 205         LASSERT(list_empty(&closure->clc_list));
 206
 207         sublock = lls->sub_lock;
 208         child = sublock->lss_cl.cls_lock;
 209         result = cl_lock_closure_build(env, child, closure);
 210         if (result == 0) {
 211                 struct cl_lock *parent = closure->clc_origin;
 212
 213                 LASSERT(cl_lock_is_mutexed(child));
 214                 sublock->lss_active = parent;
 215
 216                 if (unlikely(child->cll_state == CLS_FREEING)) {
 217                         struct lov_lock_link *link;
 218                         /*
 219                          * we could race with lock deletion which temporarily
 220                          * put the lock in freeing state, bug 19080.
 221                          */
 222                         LASSERT(!(lls->sub_flags & LSF_HELD));
 223
 224                         link = lov_lock_link_find(env, lck, sublock);
 225                         LASSERT(link != NULL);
 226                         lov_lock_unlink(env, link, sublock);
 227                         lov_sublock_unlock(env, sublock, closure, NULL);
 228                         result = CLO_REPEAT;
 229                 } else if (lsep) {
 230                         struct lov_sublock_env *subenv;
 231                         subenv = lov_sublock_env_get(env, parent, lls);
 232                         if (IS_ERR(subenv)) {
 233                                 lov_sublock_unlock(env, sublock,
 234                                                    closure, NULL);
 235                                 result = PTR_ERR(subenv);
 236                         } else {
 237                                 *lsep = subenv;
 238                         }
 239                 }
 240         }
 241         RETURN(result);
 242 }
 243
 244 /**
 245  * Updates the result of a top-lock operation from a result of sub-lock
 246  * sub-operations. Top-operations like lov_lock_{enqueue,use,unuse}() iterate
 247  * over sub-locks and lov_subresult() is used to calculate return value of a
 248  * top-operation. To this end, possible return values of sub-operations are
 249  * ordered as
 250  *
 251  *     - 0                  success
 252  *     - CLO_WAIT           wait for event
 253  *     - CLO_REPEAT         repeat top-operation
 254  *     - -ne                fundamental error
 255  *
 256  * Top-level return code can only go down through this list. CLO_REPEAT
 257  * overwrites CLO_WAIT, because lock mutex was released and sleeping condition
 258  * has to be rechecked by the upper layer.
 259  */
 260 static int lov_subresult(int result, int rc)
 261 {
 262         int result_rank;
 263         int rc_rank;
 264
 265         LASSERT(result <= 0 || result == CLO_REPEAT || result == CLO_WAIT);
 266         LASSERT(rc <= 0 || rc == CLO_REPEAT || rc == CLO_WAIT);
 267         CLASSERT(CLO_WAIT < CLO_REPEAT);
 268
 269         ENTRY;
 270
 271         /* calculate ranks in the ordering above */
 272         result_rank = result < 0 ? 1 + CLO_REPEAT : result;
 273         rc_rank = rc < 0 ? 1 + CLO_REPEAT : rc;
 274
 275         if (result_rank < rc_rank)
 276                 result = rc;
 277         RETURN(result);
 278 }
 279
 280 /**
 281  * Creates sub-locks for a given lov_lock for the first time.
 282  *
 283  * Goes through all sub-objects of top-object, and creates sub-locks on every
 284  * sub-object intersecting with top-lock extent. This is complicated by the
 285  * fact that top-lock (that is being created) can be accessed concurrently
 286  * through already created sub-locks (possibly shared with other top-locks).
 287  */
 288 static int lov_lock_sub_init(const struct lu_env *env,
 289                              struct lov_lock *lck, const struct cl_io *io)
 290 {
 291         int result = 0;
 292         int i;
 293         int j;
 294         int nr;
 295         int stripe;
 296         int start_stripe;
 297         obd_off start;
 298         obd_off end;
 299         obd_off file_start;
 300         obd_off file_end;
 301
 302         struct lov_object       *loo    = cl2lov(lck->lls_cl.cls_obj);
 303         struct lov_layout_raid0 *r0     = lov_r0(loo);
 304         struct cl_lock          *parent = lck->lls_cl.cls_lock;
 305
 306         ENTRY;
 307
 308         lck->lls_orig = parent->cll_descr;
 309         file_start = cl_offset(lov2cl(loo), parent->cll_descr.cld_start);
 310         file_end   = cl_offset(lov2cl(loo), parent->cll_descr.cld_end + 1) - 1;
 311
 312         start_stripe = lov_stripe_number(r0->lo_lsm, file_start);
 313         for (i = 0, nr = 0; i < r0->lo_nr; i++) {
 314                 /*
 315                  * XXX for wide striping smarter algorithm is desirable,
 316                  * breaking out of the loop, early.
 317                  */
 318                 stripe = (start_stripe + i) % r0->lo_nr;
 319                 if (lov_stripe_intersects(r0->lo_lsm, stripe,
 320                                           file_start, file_end, &start, &end))
 321                         nr++;
 322         }
 323         LASSERT(nr > 0);
 324         OBD_ALLOC(lck->lls_sub, nr * sizeof lck->lls_sub[0]);
 325         if (lck->lls_sub == NULL)
 326                 RETURN(-ENOMEM);
 327
 328         lck->lls_nr = nr;
 329         /*
 330          * First, fill in sub-lock descriptions in
 331          * lck->lls_sub[].sub_descr. They are used by lov_sublock_alloc()
 332          * (called below in this function, and by lov_lock_enqueue()) to
 333          * create sub-locks. At this moment, no other thread can access
 334          * top-lock.
 335          */
 336         for (j = 0, nr = 0; j < i; ++j) {
 337                 stripe = (start_stripe + j) % r0->lo_nr;
 338                 if (lov_stripe_intersects(r0->lo_lsm, stripe,
 339                                           file_start, file_end, &start, &end)) {
 340                         struct cl_lock_descr *descr;
 341
 342                         descr = &lck->lls_sub[nr].sub_descr;
 343
 344                         LASSERT(descr->cld_obj == NULL);
 345                         descr->cld_obj   = lovsub2cl(r0->lo_sub[stripe]);
 346                         descr->cld_start = cl_index(descr->cld_obj, start);
 347                         descr->cld_end   = cl_index(descr->cld_obj, end);
 348                         descr->cld_mode  = parent->cll_descr.cld_mode;
 349                         descr->cld_gid   = parent->cll_descr.cld_gid;
 350                         /* XXX has no effect */
 351                         lck->lls_sub[nr].sub_got = *descr;
 352                         lck->lls_sub[nr].sub_stripe = stripe;
 353                         nr++;
 354                 }
 355         }
 356         LASSERT(nr == lck->lls_nr);
 357         /*
 358          * Then, create sub-locks. Once at least one sub-lock was created,
 359          * top-lock can be reached by other threads.
 360          */
 361         for (i = 0; i < lck->lls_nr; ++i) {
 362                 struct cl_lock       *sublock;
 363                 struct lov_lock_link *link;
 364
 365                 if (lck->lls_sub[i].sub_lock == NULL) {
 366                         sublock = lov_sublock_alloc(env, io, lck, i, &link);
 367                         if (IS_ERR(sublock)) {
 368                                 result = PTR_ERR(sublock);
 369                                 break;
 370                         }
 371                         cl_lock_mutex_get(env, sublock);
 372                         cl_lock_mutex_get(env, parent);
 373                         /*
 374                          * recheck under mutex that sub-lock wasn't created
 375                          * concurrently, and that top-lock is still alive.
 376                          */
 377                         if (lck->lls_sub[i].sub_lock == NULL &&
 378                             parent->cll_state < CLS_FREEING) {
 379                                 lov_sublock_adopt(env, lck, sublock, i, link);
 380                                 cl_lock_mutex_put(env, parent);
 381                         } else {
 382                                 OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
 383                                 cl_lock_mutex_put(env, parent);
 384                                 cl_lock_unhold(env, sublock,
 385                                                "lov-parent", parent);
 386                         }
 387                         cl_lock_mutex_put(env, sublock);
 388                 }
 389         }
 390         /*
 391          * Some sub-locks can be missing at this point. This is not a problem,
 392          * because enqueue will create them anyway. Main duty of this function
 393          * is to fill in sub-lock descriptions in a race free manner.
 394          */
 395         RETURN(result);
 396 }
 397
 398 static int lov_sublock_release(const struct lu_env *env, struct lov_lock *lck,
 399                                int i, int deluser, int rc)
 400 {
 401         struct cl_lock *parent = lck->lls_cl.cls_lock;
 402
 403         LASSERT(cl_lock_is_mutexed(parent));
 404         ENTRY;
 405
 406         if (lck->lls_sub[i].sub_flags & LSF_HELD) {
 407                 struct cl_lock    *sublock;
 408                 int dying;
 409
 410                 LASSERT(lck->lls_sub[i].sub_lock != NULL);
 411                 sublock = lck->lls_sub[i].sub_lock->lss_cl.cls_lock;
 412                 LASSERT(cl_lock_is_mutexed(sublock));
 413
 414                 lck->lls_sub[i].sub_flags &= ~LSF_HELD;
 415                 if (deluser)
 416                         cl_lock_user_del(env, sublock);
 417                 /*
 418                  * If the last hold is released, and cancellation is pending
 419                  * for a sub-lock, release parent mutex, to avoid keeping it
 420                  * while sub-lock is being paged out.
 421                  */
 422                 dying = (sublock->cll_descr.cld_mode == CLM_PHANTOM ||
 423                          sublock->cll_descr.cld_mode == CLM_GROUP ||
 424                          (sublock->cll_flags & (CLF_CANCELPEND|CLF_DOOMED))) &&
 425                         sublock->cll_holds == 1;
 426                 if (dying)
 427                         cl_lock_mutex_put(env, parent);
 428                 cl_lock_unhold(env, sublock, "lov-parent", parent);
 429                 if (dying) {
 430                         cl_lock_mutex_get(env, parent);
 431                         rc = lov_subresult(rc, CLO_REPEAT);
 432                 }
 433                 /*
 434                  * From now on lck->lls_sub[i].sub_lock is a "weak" pointer,
 435                  * not backed by a reference on a
 436                  * sub-lock. lovsub_lock_delete() will clear
 437                  * lck->lls_sub[i].sub_lock under semaphores, just before
 438                  * sub-lock is destroyed.
 439                  */
 440         }
 441         RETURN(rc);
 442 }
 443
 444 static void lov_sublock_hold(const struct lu_env *env, struct lov_lock *lck,
 445                              int i)
 446 {
 447         struct cl_lock *parent = lck->lls_cl.cls_lock;
 448
 449         LASSERT(cl_lock_is_mutexed(parent));
 450         ENTRY;
 451
 452         if (!(lck->lls_sub[i].sub_flags & LSF_HELD)) {
 453                 struct cl_lock *sublock;
 454
 455                 LASSERT(lck->lls_sub[i].sub_lock != NULL);
 456                 sublock = lck->lls_sub[i].sub_lock->lss_cl.cls_lock;
 457                 LASSERT(cl_lock_is_mutexed(sublock));
 458                 LASSERT(sublock->cll_state != CLS_FREEING);
 459
 460                 lck->lls_sub[i].sub_flags |= LSF_HELD;
 461
 462                 cl_lock_get_trust(sublock);
 463                 cl_lock_hold_add(env, sublock, "lov-parent", parent);
 464                 cl_lock_user_add(env, sublock);
 465                 cl_lock_put(env, sublock);
 466         }
 467         EXIT;
 468 }
 469
 470 static void lov_lock_fini(const struct lu_env *env,
 471                           struct cl_lock_slice *slice)
 472 {
 473         struct lov_lock *lck;
 474         int i;
 475
 476         ENTRY;
 477         lck = cl2lov_lock(slice);
 478         LASSERT(lck->lls_nr_filled == 0);
 479         if (lck->lls_sub != NULL) {
 480                 for (i = 0; i < lck->lls_nr; ++i)
 481                         /*
 482                          * No sub-locks exists at this point, as sub-lock has
 483                          * a reference on its parent.
 484                          */
 485                         LASSERT(lck->lls_sub[i].sub_lock == NULL);
 486                 OBD_FREE(lck->lls_sub, lck->lls_nr * sizeof lck->lls_sub[0]);
 487         }
 488         OBD_SLAB_FREE_PTR(lck, lov_lock_kmem);
 489         EXIT;
 490 }
 491
 492 /**
 493  * Tries to advance a state machine of a given sub-lock toward enqueuing of
 494  * the top-lock.
 495  *
 496  * \retval 0 if state-transition can proceed
 497  * \retval -ve otherwise.
 498  */
 499 static int lov_lock_enqueue_one(const struct lu_env *env, struct lov_lock *lck,
 500                                 struct cl_lock *sublock,
 501                                 struct cl_io *io, __u32 enqflags, int last)
 502 {
 503         int result;
 504         ENTRY;
 505
 506         /* first, try to enqueue a sub-lock ... */
 507         result = cl_enqueue_try(env, sublock, io, enqflags);
 508         if (sublock->cll_state == CLS_ENQUEUED)
 509                 /* if it is enqueued, try to `wait' on it---maybe it's already
 510                  * granted */
 511                 result = cl_wait_try(env, sublock);
 512         /*
 513          * If CEF_ASYNC flag is set, then all sub-locks can be enqueued in
 514          * parallel, otherwise---enqueue has to wait until sub-lock is granted
 515          * before proceeding to the next one.
 516          */
 517         if (result == CLO_WAIT && sublock->cll_state <= CLS_HELD &&
 518             enqflags & CEF_ASYNC && !last)
 519                 result = 0;
 520         RETURN(result);
 521 }
 522
 523 /**
 524  * Helper function for lov_lock_enqueue() that creates missing sub-lock.
 525  */
 526 static int lov_sublock_fill(const struct lu_env *env, struct cl_lock *parent,
 527                             struct cl_io *io, struct lov_lock *lck, int idx)
 528 {
 529         struct lov_lock_link *link;
 530         struct cl_lock       *sublock;
 531         int                   result;
 532
 533         LASSERT(parent->cll_depth == 1);
 534         cl_lock_mutex_put(env, parent);
 535         sublock = lov_sublock_alloc(env, io, lck, idx, &link);
 536         if (!IS_ERR(sublock))
 537                 cl_lock_mutex_get(env, sublock);
 538         cl_lock_mutex_get(env, parent);
 539
 540         if (!IS_ERR(sublock)) {
 541                 if (parent->cll_state == CLS_QUEUING &&
 542                     lck->lls_sub[idx].sub_lock == NULL)
 543                         lov_sublock_adopt(env, lck, sublock, idx, link);
 544                 else {
 545                         OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
 546                         /* other thread allocated sub-lock, or enqueue is no
 547                          * longer going on */
 548                         cl_lock_mutex_put(env, parent);
 549                         cl_lock_unhold(env, sublock, "lov-parent", parent);
 550                         cl_lock_mutex_get(env, parent);
 551                 }
 552                 cl_lock_mutex_put(env, sublock);
 553                 result = CLO_REPEAT;
 554         } else
 555                 result = PTR_ERR(sublock);
 556         return result;
 557 }
 558
 559 /**
 560  * Implementation of cl_lock_operations::clo_enqueue() for lov layer. This
 561  * function is rather subtle, as it enqueues top-lock (i.e., advances top-lock
 562  * state machine from CLS_QUEUING to CLS_ENQUEUED states) by juggling sub-lock
 563  * state machines in the face of sub-locks sharing (by multiple top-locks),
 564  * and concurrent sub-lock cancellations.
 565  */
 566 static int lov_lock_enqueue(const struct lu_env *env,
 567                             const struct cl_lock_slice *slice,
 568                             struct cl_io *io, __u32 enqflags)
 569 {
 570         struct cl_lock         *lock    = slice->cls_lock;
 571         struct lov_lock        *lck     = cl2lov_lock(slice);
 572         struct cl_lock_closure *closure = lov_closure_get(env, lock);
 573         int i;
 574         int result;
 575         enum cl_lock_state minstate;
 576
 577         ENTRY;
 578
 579         for (result = 0, minstate = CLS_FREEING, i = 0; i < lck->lls_nr; ++i) {
 580                 int rc;
 581                 struct lovsub_lock     *sub;
 582                 struct lov_lock_sub    *lls;
 583                 struct cl_lock         *sublock;
 584                 struct lov_sublock_env *subenv;
 585
 586                 if (lock->cll_state != CLS_QUEUING) {
 587                         /*
 588                          * Lock might have left QUEUING state if previous
 589                          * iteration released its mutex. Stop enqueing in this
 590                          * case and let the upper layer to decide what to do.
 591                          */
 592                         LASSERT(i > 0 && result != 0);
 593                         break;
 594                 }
 595
 596                 lls = &lck->lls_sub[i];
 597                 sub = lls->sub_lock;
 598                 /*
 599                  * Sub-lock might have been canceled, while top-lock was
 600                  * cached.
 601                  */
 602                 if (sub == NULL) {
 603                         result = lov_sublock_fill(env, lock, io, lck, i);
 604                         /* lov_sublock_fill() released @lock mutex,
 605                          * restart. */
 606                         break;
 607                 }
 608                 sublock = sub->lss_cl.cls_lock;
 609                 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
 610                 if (rc == 0) {
 611                         lov_sublock_hold(env, lck, i);
 612                         rc = lov_lock_enqueue_one(subenv->lse_env, lck, sublock,
 613                                                   subenv->lse_io, enqflags,
 614                                                   i == lck->lls_nr - 1);
 615                         minstate = min(minstate, sublock->cll_state);
 616                         /*
 617                          * Don't hold a sub-lock in CLS_CACHED state, see
 618                          * description for lov_lock::lls_sub.
 619                          */
 620                         if (sublock->cll_state > CLS_HELD)
 621                                 rc = lov_sublock_release(env, lck, i, 1, rc);
 622                         lov_sublock_unlock(env, sub, closure, subenv);
 623                 }
 624                 result = lov_subresult(result, rc);
 625                 if (result < 0)
 626                         break;
 627         }
 628         cl_lock_closure_fini(closure);
 629         RETURN(result ?: minstate >= CLS_ENQUEUED ? 0 : CLO_WAIT);
 630 }
 631
 632 static int lov_lock_unuse(const struct lu_env *env,
 633                           const struct cl_lock_slice *slice)
 634 {
 635         struct lov_lock        *lck     = cl2lov_lock(slice);
 636         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
 637         int i;
 638         int result;
 639
 640         ENTRY;
 641
 642         for (result = 0, i = 0; i < lck->lls_nr; ++i) {
 643                 int rc;
 644                 struct lovsub_lock     *sub;
 645                 struct cl_lock         *sublock;
 646                 struct lov_lock_sub    *lls;
 647                 struct lov_sublock_env *subenv;
 648
 649                 /* top-lock state cannot change concurrently, because single
 650                  * thread (one that released the last hold) carries unlocking
 651                  * to the completion. */
 652                 LASSERT(slice->cls_lock->cll_state == CLS_UNLOCKING);
 653                 lls = &lck->lls_sub[i];
 654                 sub = lls->sub_lock;
 655                 if (sub == NULL)
 656                         continue;
 657
 658                 sublock = sub->lss_cl.cls_lock;
 659                 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
 660                 if (rc == 0) {
 661                         if (lck->lls_sub[i].sub_flags & LSF_HELD) {
 662                                 LASSERT(sublock->cll_state == CLS_HELD);
 663                                 rc = cl_unuse_try(subenv->lse_env, sublock);
 664                                 if (rc != CLO_WAIT)
 665                                         rc = lov_sublock_release(env, lck,
 666                                                                  i, 0, rc);
 667                         }
 668                         lov_sublock_unlock(env, sub, closure, subenv);
 669                 }
 670                 result = lov_subresult(result, rc);
 671                 if (result < 0)
 672                         break;
 673         }
 674         if (result == 0 && lck->lls_unuse_race) {
 675                 lck->lls_unuse_race = 0;
 676                 result = -ESTALE;
 677         }
 678         cl_lock_closure_fini(closure);
 679         RETURN(result);
 680 }
 681
 682 static int lov_lock_wait(const struct lu_env *env,
 683                          const struct cl_lock_slice *slice)
 684 {
 685         struct lov_lock        *lck     = cl2lov_lock(slice);
 686         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
 687         enum cl_lock_state      minstate;
 688         int                     result;
 689         int                     i;
 690
 691         ENTRY;
 692
 693         for (result = 0, minstate = CLS_FREEING, i = 0; i < lck->lls_nr; ++i) {
 694                 int rc;
 695                 struct lovsub_lock     *sub;
 696                 struct cl_lock         *sublock;
 697                 struct lov_lock_sub    *lls;
 698                 struct lov_sublock_env *subenv;
 699
 700                 lls = &lck->lls_sub[i];
 701                 sub = lls->sub_lock;
 702                 LASSERT(sub != NULL);
 703                 sublock = sub->lss_cl.cls_lock;
 704                 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
 705                 if (rc == 0) {
 706                         LASSERT(sublock->cll_state >= CLS_ENQUEUED);
 707                         if (sublock->cll_state < CLS_HELD)
 708                                 rc = cl_wait_try(env, sublock);
 709
 710                         minstate = min(minstate, sublock->cll_state);
 711                         lov_sublock_unlock(env, sub, closure, subenv);
 712                 }
 713                 result = lov_subresult(result, rc);
 714                 if (result < 0)
 715                         break;
 716         }
 717         cl_lock_closure_fini(closure);
 718         RETURN(result ?: minstate >= CLS_HELD ? 0 : CLO_WAIT);
 719 }
 720
 721 static int lov_lock_use(const struct lu_env *env,
 722                         const struct cl_lock_slice *slice)
 723 {
 724         struct lov_lock        *lck     = cl2lov_lock(slice);
 725         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
 726         int                     result;
 727         int                     i;
 728
 729         LASSERT(slice->cls_lock->cll_state == CLS_CACHED);
 730         ENTRY;
 731
 732         for (result = 0, i = 0; i < lck->lls_nr; ++i) {
 733                 int rc;
 734                 struct lovsub_lock     *sub;
 735                 struct cl_lock         *sublock;
 736                 struct lov_lock_sub    *lls;
 737                 struct lov_sublock_env *subenv;
 738
 739                 if (slice->cls_lock->cll_state != CLS_CACHED) {
 740                         /* see comment in lov_lock_enqueue(). */
 741                         LASSERT(i > 0 && result != 0);
 742                         break;
 743                 }
 744                 /*
 745                  * if a sub-lock was destroyed while top-lock was in
 746                  * CLS_CACHED state, top-lock would have been moved into
 747                  * CLS_NEW state, so all sub-locks have to be in place.
 748                  */
 749                 lls = &lck->lls_sub[i];
 750                 sub = lls->sub_lock;
 751                 LASSERT(sub != NULL);
 752                 sublock = sub->lss_cl.cls_lock;
 753                 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
 754                 if (rc == 0) {
 755                         LASSERT(sublock->cll_state != CLS_FREEING);
 756                         lov_sublock_hold(env, lck, i);
 757                         if (sublock->cll_state == CLS_CACHED) {
 758                                 rc = cl_use_try(subenv->lse_env, sublock);
 759                                 if (rc != 0)
 760                                         rc = lov_sublock_release(env, lck,
 761                                                                  i, 1, rc);
 762                         } else
 763                                 rc = 0;
 764                         lov_sublock_unlock(env, sub, closure, subenv);
 765                 }
 766                 result = lov_subresult(result, rc);
 767                 if (result < 0)
 768                         break;
 769         }
 770         cl_lock_closure_fini(closure);
 771         RETURN(result);
 772 }
 773
 774 #if 0
 775 static int lock_lock_multi_match()
 776 {
 777         struct cl_lock          *lock    = slice->cls_lock;
 778         struct cl_lock_descr    *subneed = &lov_env_info(env)->lti_ldescr;
 779         struct lov_object       *loo     = cl2lov(lov->lls_cl.cls_obj);
 780         struct lov_layout_raid0 *r0      = lov_r0(loo);
 781         struct lov_lock_sub     *sub;
 782         struct cl_object        *subobj;
 783         obd_off  fstart;
 784         obd_off  fend;
 785         obd_off  start;
 786         obd_off  end;
 787         int i;
 788
 789         fstart = cl_offset(need->cld_obj, need->cld_start);
 790         fend   = cl_offset(need->cld_obj, need->cld_end + 1) - 1;
 791         subneed->cld_mode = need->cld_mode;
 792         cl_lock_mutex_get(env, lock);
 793         for (i = 0; i < lov->lls_nr; ++i) {
 794                 sub = &lov->lls_sub[i];
 795                 if (sub->sub_lock == NULL)
 796                         continue;
 797                 subobj = sub->sub_descr.cld_obj;
 798                 if (!lov_stripe_intersects(r0->lo_lsm, sub->sub_stripe,
 799                                            fstart, fend, &start, &end))
 800                         continue;
 801                 subneed->cld_start = cl_index(subobj, start);
 802                 subneed->cld_end   = cl_index(subobj, end);
 803                 subneed->cld_obj   = subobj;
 804                 if (!cl_lock_ext_match(&sub->sub_got, subneed)) {
 805                         result = 0;
 806                         break;
 807                 }
 808         }
 809         cl_lock_mutex_put(env, lock);
 810 }
 811 #endif
 812
 813 /**
 814  * Check if the extent region \a descr is covered by \a child against the
 815  * specific \a stripe.
 816  */
 817 static int lov_lock_stripe_is_matching(const struct lu_env *env,
 818                                        struct lov_object *lov, int stripe,
 819                                        const struct cl_lock_descr *child,
 820                                        const struct cl_lock_descr *descr)
 821 {
 822         struct lov_stripe_md *lsm = lov_r0(lov)->lo_lsm;
 823         obd_off start;
 824         obd_off end;
 825         int result;
 826
 827         if (lov_r0(lov)->lo_nr == 1)
 828                 return cl_lock_ext_match(child, descr);
 829
 830         /*
 831          * For a multi-stripes object:
 832          * - make sure the descr only covers child's stripe, and
 833          * - check if extent is matching.
 834          */
 835         start = cl_offset(&lov->lo_cl, descr->cld_start);
 836         end   = cl_offset(&lov->lo_cl, descr->cld_end + 1) - 1;
 837         result = end - start <= lsm->lsm_stripe_size &&
 838                  stripe == lov_stripe_number(lsm, start) &&
 839                  stripe == lov_stripe_number(lsm, end);
 840         if (result) {
 841                 struct cl_lock_descr *subd = &lov_env_info(env)->lti_ldescr;
 842                 obd_off sub_start;
 843                 obd_off sub_end;
 844
 845                 subd->cld_obj  = NULL;   /* don't need sub object at all */
 846                 subd->cld_mode = descr->cld_mode;
 847                 subd->cld_gid  = descr->cld_gid;
 848                 result = lov_stripe_intersects(lsm, stripe, start, end,
 849                                                &sub_start, &sub_end);
 850                 LASSERT(result);
 851                 subd->cld_start = cl_index(child->cld_obj, sub_start);
 852                 subd->cld_end   = cl_index(child->cld_obj, sub_end);
 853                 result = cl_lock_ext_match(child, subd);
 854         }
 855         return result;
 856 }
 857
 858 /**
 859  * An implementation of cl_lock_operations::clo_fits_into() method.
 860  *
 861  * Checks whether a lock (given by \a slice) is suitable for \a
 862  * io. Multi-stripe locks can be used only for "quick" io, like truncate, or
 863  * O_APPEND write.
 864  *
 865  * \see ccc_lock_fits_into().
 866  */
 867 static int lov_lock_fits_into(const struct lu_env *env,
 868                               const struct cl_lock_slice *slice,
 869                               const struct cl_lock_descr *need,
 870                               const struct cl_io *io)
 871 {
 872         struct lov_lock   *lov = cl2lov_lock(slice);
 873         struct lov_object *obj = cl2lov(slice->cls_obj);
 874         int result;
 875
 876         LASSERT(cl_object_same(need->cld_obj, slice->cls_obj));
 877         LASSERT(lov->lls_nr > 0);
 878
 879         ENTRY;
 880
 881         if (need->cld_mode == CLM_GROUP)
 882                 /*
 883                  * always allow to match group lock.
 884                  */
 885                 result = cl_lock_ext_match(&lov->lls_orig, need);
 886         else if (lov->lls_nr == 1) {
 887                 struct cl_lock_descr *got = &lov->lls_sub[0].sub_got;
 888                 result = lov_lock_stripe_is_matching(env,
 889                                                      cl2lov(slice->cls_obj),
 890                                                      lov->lls_sub[0].sub_stripe,
 891                                                      got, need);
 892         } else if (io->ci_type != CIT_TRUNC && io->ci_type != CIT_MISC &&
 893                    !cl_io_is_append(io) && need->cld_mode != CLM_PHANTOM)
 894                 /*
 895                  * Multi-stripe locks are only suitable for `quick' IO and for
 896                  * glimpse.
 897                  */
 898                 result = 0;
 899         else
 900                 /*
 901                  * Most general case: multi-stripe existing lock, and
 902                  * (potentially) multi-stripe @need lock. Check that @need is
 903                  * covered by @lov's sub-locks.
 904                  *
 905                  * For now, ignore lock expansions made by the server, and
 906                  * match against original lock extent.
 907                  */
 908                 result = cl_lock_ext_match(&lov->lls_orig, need);
 909         CDEBUG(D_DLMTRACE, DDESCR"/"DDESCR" %i %i/%i: %i\n",
 910                PDESCR(&lov->lls_orig), PDESCR(&lov->lls_sub[0].sub_got),
 911                lov->lls_sub[0].sub_stripe, lov->lls_nr, lov_r0(obj)->lo_nr,
 912                result);
 913         RETURN(result);
 914 }
 915
 916 void lov_lock_unlink(const struct lu_env *env,
 917                      struct lov_lock_link *link, struct lovsub_lock *sub)
 918 {
 919         struct lov_lock *lck    = link->lll_super;
 920         struct cl_lock  *parent = lck->lls_cl.cls_lock;
 921
 922         LASSERT(cl_lock_is_mutexed(parent));
 923         LASSERT(cl_lock_is_mutexed(sub->lss_cl.cls_lock));
 924         ENTRY;
 925
 926         list_del_init(&link->lll_list);
 927         LASSERT(lck->lls_sub[link->lll_idx].sub_lock == sub);
 928         /* yank this sub-lock from parent's array */
 929         lck->lls_sub[link->lll_idx].sub_lock = NULL;
 930         LASSERT(lck->lls_nr_filled > 0);
 931         lck->lls_nr_filled--;
 932         lu_ref_del(&parent->cll_reference, "lov-child", sub->lss_cl.cls_lock);
 933         cl_lock_put(env, parent);
 934         OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
 935         EXIT;
 936 }
 937
 938 struct lov_lock_link *lov_lock_link_find(const struct lu_env *env,
 939                                          struct lov_lock *lck,
 940                                          struct lovsub_lock *sub)
 941 {
 942         struct lov_lock_link *scan;
 943
 944         LASSERT(cl_lock_is_mutexed(sub->lss_cl.cls_lock));
 945         ENTRY;
 946
 947         list_for_each_entry(scan, &sub->lss_parents, lll_list) {
 948                 if (scan->lll_super == lck)
 949                         RETURN(scan);
 950         }
 951         RETURN(NULL);
 952 }
 953
 954 /**
 955  * An implementation of cl_lock_operations::clo_delete() method. This is
 956  * invoked for "top-to-bottom" delete, when lock destruction starts from the
 957  * top-lock, e.g., as a result of inode destruction.
 958  *
 959  * Unlinks top-lock from all its sub-locks. Sub-locks are not deleted there:
 960  * this is done separately elsewhere:
 961  *
 962  *     - for inode destruction, lov_object_delete() calls cl_object_kill() for
 963  *       each sub-object, purging its locks;
 964  *
 965  *     - in other cases (e.g., a fatal error with a top-lock) sub-locks are
 966  *       left in the cache.
 967  */
 968 static void lov_lock_delete(const struct lu_env *env,
 969                             const struct cl_lock_slice *slice)
 970 {
 971         struct lov_lock        *lck     = cl2lov_lock(slice);
 972         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
 973         int i;
 974
 975         LASSERT(slice->cls_lock->cll_state == CLS_FREEING);
 976         ENTRY;
 977
 978         for (i = 0; i < lck->lls_nr; ++i) {
 979                 struct lov_lock_sub *lls;
 980                 struct lovsub_lock  *lsl;
 981                 struct cl_lock      *sublock;
 982                 int rc;
 983
 984                 lls = &lck->lls_sub[i];
 985                 lsl = lls->sub_lock;
 986                 if (lsl == NULL)
 987                         continue;
 988
 989                 sublock = lsl->lss_cl.cls_lock;
 990                 rc = lov_sublock_lock(env, lck, lls, closure, NULL);
 991                 if (rc == 0) {
 992                         if (lck->lls_sub[i].sub_flags & LSF_HELD)
 993                                 lov_sublock_release(env, lck, i, 1, 0);
 994                         if (sublock->cll_state < CLS_FREEING) {
 995                                 struct lov_lock_link *link;
 996
 997                                 link = lov_lock_link_find(env, lck, lsl);
 998                                 LASSERT(link != NULL);
 999                                 lov_lock_unlink(env, link, lsl);
1000                                 LASSERT(lck->lls_sub[i].sub_lock == NULL);
1001                         }
1002                         lov_sublock_unlock(env, lsl, closure, NULL);
1003                 } else if (rc == CLO_REPEAT) {
1004                         --i; /* repeat with this lock */
1005                 } else {
1006                         CL_LOCK_DEBUG(D_ERROR, env, sublock,
1007                                       "Cannot get sub-lock for delete: %i\n",
1008                                       rc);
1009                 }
1010         }
1011         cl_lock_closure_fini(closure);
1012         EXIT;
1013 }
1014
1015 static int lov_lock_print(const struct lu_env *env, void *cookie,
1016                           lu_printer_t p, const struct cl_lock_slice *slice)
1017 {
1018         struct lov_lock *lck = cl2lov_lock(slice);
1019         int              i;
1020
1021         (*p)(env, cookie, "%d\n", lck->lls_nr);
1022         for (i = 0; i < lck->lls_nr; ++i) {
1023                 struct lov_lock_sub *sub;
1024
1025                 sub = &lck->lls_sub[i];
1026                 (*p)(env, cookie, "    %d %x: ", i, sub->sub_flags);
1027                 if (sub->sub_lock != NULL)
1028                         cl_lock_print(env, cookie, p,
1029                                       sub->sub_lock->lss_cl.cls_lock);
1030                 else
1031                         (*p)(env, cookie, "---\n");
1032         }
1033         return 0;
1034 }
1035
1036 static const struct cl_lock_operations lov_lock_ops = {
1037         .clo_fini      = lov_lock_fini,
1038         .clo_enqueue   = lov_lock_enqueue,
1039         .clo_wait      = lov_lock_wait,
1040         .clo_use       = lov_lock_use,
1041         .clo_unuse     = lov_lock_unuse,
1042         .clo_fits_into = lov_lock_fits_into,
1043         .clo_delete    = lov_lock_delete,
1044         .clo_print     = lov_lock_print
1045 };
1046
1047 int lov_lock_init_raid0(const struct lu_env *env, struct cl_object *obj,
1048                         struct cl_lock *lock, const struct cl_io *io)
1049 {
1050         struct lov_lock *lck;
1051         int result;
1052
1053         ENTRY;
1054         OBD_SLAB_ALLOC_PTR_GFP(lck, lov_lock_kmem, CFS_ALLOC_IO);
1055         if (lck != NULL) {
1056                 cl_lock_slice_add(lock, &lck->lls_cl, obj, &lov_lock_ops);
1057                 result = lov_lock_sub_init(env, lck, io);
1058         } else
1059                 result = -ENOMEM;
1060         RETURN(result);
1061 }
1062
1063 static struct cl_lock_closure *lov_closure_get(const struct lu_env *env,
1064                                                struct cl_lock *parent)
1065 {
1066         struct cl_lock_closure *closure;
1067
1068         closure = &lov_env_info(env)->lti_closure;
1069         LASSERT(list_empty(&closure->clc_list));
1070         cl_lock_closure_init(env, closure, parent, 1);
1071         return closure;
1072 }
1073
1074
1075 /** @} lov */