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