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 (c) 2008, 2010, Oracle and/or its affiliates. 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 static int lov_lock_unuse(const struct lu_env *env,
  53                           const struct cl_lock_slice *slice);
  54 /*****************************************************************************
  55  *
  56  * Lov lock operations.
  57  *
  58  */
  59
  60 static struct lov_sublock_env *lov_sublock_env_get(const struct lu_env *env,
  61                                                    struct cl_lock *parent,
  62                                                    struct lov_lock_sub *lls)
  63 {
  64         struct lov_sublock_env *subenv;
  65         struct lov_io          *lio    = lov_env_io(env);
  66         struct cl_io           *io     = lio->lis_cl.cis_io;
  67         struct lov_io_sub      *sub;
  68
  69         subenv = &lov_env_session(env)->ls_subenv;
  70
  71         /*
  72          * FIXME: We tend to use the subio's env & io to call the sublock
  73          * lock operations because osc lock sometimes stores some control
  74          * variables in thread's IO infomation(Now only lockless information).
  75          * However, if the lock's host(object) is different from the object
  76          * for current IO, we have no way to get the subenv and subio because
  77          * they are not initialized at all. As a temp fix, in this case,
  78          * we still borrow the parent's env to call sublock operations.
  79          */
  80         if (!io || !cl_object_same(io->ci_obj, parent->cll_descr.cld_obj)) {
  81                 subenv->lse_env = env;
  82                 subenv->lse_io  = io;
  83                 subenv->lse_sub = NULL;
  84         } else {
  85                 sub = lov_sub_get(env, lio, lls->sub_stripe);
  86                 if (!IS_ERR(sub)) {
  87                         subenv->lse_env = sub->sub_env;
  88                         subenv->lse_io  = sub->sub_io;
  89                         subenv->lse_sub = sub;
  90                 } else {
  91                         subenv = (void*)sub;
  92                 }
  93         }
  94         return subenv;
  95 }
  96
  97 static void lov_sublock_env_put(struct lov_sublock_env *subenv)
  98 {
  99         if (subenv && subenv->lse_sub)
 100                 lov_sub_put(subenv->lse_sub);
 101 }
 102
 103 static void lov_sublock_adopt(const struct lu_env *env, struct lov_lock *lck,
 104                               struct cl_lock *sublock, int idx,
 105                               struct lov_lock_link *link)
 106 {
 107         struct lovsub_lock *lsl;
 108         struct cl_lock     *parent = lck->lls_cl.cls_lock;
 109         int                 rc;
 110
 111         LASSERT(cl_lock_is_mutexed(parent));
 112         LASSERT(cl_lock_is_mutexed(sublock));
 113         ENTRY;
 114
 115         lsl = cl2sub_lock(sublock);
 116         /*
 117          * check that sub-lock doesn't have lock link to this top-lock.
 118          */
 119         LASSERT(lov_lock_link_find(env, lck, lsl) == NULL);
 120         LASSERT(idx < lck->lls_nr);
 121
 122         lck->lls_sub[idx].sub_lock = lsl;
 123         lck->lls_nr_filled++;
 124         LASSERT(lck->lls_nr_filled <= lck->lls_nr);
 125         cfs_list_add_tail(&link->lll_list, &lsl->lss_parents);
 126         link->lll_idx = idx;
 127         link->lll_super = lck;
 128         cl_lock_get(parent);
 129         lu_ref_add(&parent->cll_reference, "lov-child", sublock);
 130         lck->lls_sub[idx].sub_flags |= LSF_HELD;
 131         cl_lock_user_add(env, sublock);
 132
 133         rc = lov_sublock_modify(env, lck, lsl, &sublock->cll_descr, idx);
 134         LASSERT(rc == 0); /* there is no way this can fail, currently */
 135         EXIT;
 136 }
 137
 138 static struct cl_lock *lov_sublock_alloc(const struct lu_env *env,
 139                                          const struct cl_io *io,
 140                                          struct lov_lock *lck,
 141                                          int idx, struct lov_lock_link **out)
 142 {
 143         struct cl_lock       *sublock;
 144         struct cl_lock       *parent;
 145         struct lov_lock_link *link;
 146
 147         LASSERT(idx < lck->lls_nr);
 148         ENTRY;
 149
 150         OBD_SLAB_ALLOC_PTR_GFP(link, lov_lock_link_kmem, CFS_ALLOC_IO);
 151         if (link != NULL) {
 152                 struct lov_sublock_env *subenv;
 153                 struct lov_lock_sub  *lls;
 154                 struct cl_lock_descr *descr;
 155
 156                 parent = lck->lls_cl.cls_lock;
 157                 lls    = &lck->lls_sub[idx];
 158                 descr  = &lls->sub_descr;
 159
 160                 subenv = lov_sublock_env_get(env, parent, lls);
 161                 if (!IS_ERR(subenv)) {
 162                         /* CAVEAT: Don't try to add a field in lov_lock_sub
 163                          * to remember the subio. This is because lock is able
 164                          * to be cached, but this is not true for IO. This
 165                          * further means a sublock might be referenced in
 166                          * different io context. -jay */
 167
 168                         sublock = cl_lock_hold(subenv->lse_env, subenv->lse_io,
 169                                                descr, "lov-parent", parent);
 170                         lov_sublock_env_put(subenv);
 171                 } else {
 172                         /* error occurs. */
 173                         sublock = (void*)subenv;
 174                 }
 175
 176                 if (!IS_ERR(sublock))
 177                         *out = link;
 178                 else
 179                         OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
 180         } else
 181                 sublock = ERR_PTR(-ENOMEM);
 182         RETURN(sublock);
 183 }
 184
 185 static void lov_sublock_unlock(const struct lu_env *env,
 186                                struct lovsub_lock *lsl,
 187                                struct cl_lock_closure *closure,
 188                                struct lov_sublock_env *subenv)
 189 {
 190         ENTRY;
 191         lov_sublock_env_put(subenv);
 192         lsl->lss_active = NULL;
 193         cl_lock_disclosure(env, closure);
 194         EXIT;
 195 }
 196
 197 static int lov_sublock_lock(const struct lu_env *env,
 198                             struct lov_lock *lck,
 199                             struct lov_lock_sub *lls,
 200                             struct cl_lock_closure *closure,
 201                             struct lov_sublock_env **lsep)
 202 {
 203         struct lovsub_lock *sublock;
 204         struct cl_lock     *child;
 205         int                 result = 0;
 206         ENTRY;
 207
 208         LASSERT(cfs_list_empty(&closure->clc_list));
 209
 210         sublock = lls->sub_lock;
 211         child = sublock->lss_cl.cls_lock;
 212         result = cl_lock_closure_build(env, child, closure);
 213         if (result == 0) {
 214                 struct cl_lock *parent = closure->clc_origin;
 215
 216                 LASSERT(cl_lock_is_mutexed(child));
 217                 sublock->lss_active = parent;
 218
 219                 if (unlikely((child->cll_state == CLS_FREEING) ||
 220                              (child->cll_flags & CLF_CANCELLED))) {
 221                         struct lov_lock_link *link;
 222                         /*
 223                          * we could race with lock deletion which temporarily
 224                          * put the lock in freeing state, bug 19080.
 225                          */
 226                         LASSERT(!(lls->sub_flags & LSF_HELD));
 227
 228                         link = lov_lock_link_find(env, lck, sublock);
 229                         LASSERT(link != NULL);
 230                         lov_lock_unlink(env, link, sublock);
 231                         lov_sublock_unlock(env, sublock, closure, NULL);
 232                         lck->lls_cancel_race = 1;
 233                         result = CLO_REPEAT;
 234                 } else if (lsep) {
 235                         struct lov_sublock_env *subenv;
 236                         subenv = lov_sublock_env_get(env, parent, lls);
 237                         if (IS_ERR(subenv)) {
 238                                 lov_sublock_unlock(env, sublock,
 239                                                    closure, NULL);
 240                                 result = PTR_ERR(subenv);
 241                         } else {
 242                                 *lsep = subenv;
 243                         }
 244                 }
 245         }
 246         RETURN(result);
 247 }
 248
 249 /**
 250  * Updates the result of a top-lock operation from a result of sub-lock
 251  * sub-operations. Top-operations like lov_lock_{enqueue,use,unuse}() iterate
 252  * over sub-locks and lov_subresult() is used to calculate return value of a
 253  * top-operation. To this end, possible return values of sub-operations are
 254  * ordered as
 255  *
 256  *     - 0                  success
 257  *     - CLO_WAIT           wait for event
 258  *     - CLO_REPEAT         repeat top-operation
 259  *     - -ne                fundamental error
 260  *
 261  * Top-level return code can only go down through this list. CLO_REPEAT
 262  * overwrites CLO_WAIT, because lock mutex was released and sleeping condition
 263  * has to be rechecked by the upper layer.
 264  */
 265 static int lov_subresult(int result, int rc)
 266 {
 267         int result_rank;
 268         int rc_rank;
 269
 270         LASSERT(result <= 0 || result == CLO_REPEAT || result == CLO_WAIT);
 271         LASSERT(rc <= 0 || rc == CLO_REPEAT || rc == CLO_WAIT);
 272         CLASSERT(CLO_WAIT < CLO_REPEAT);
 273
 274         ENTRY;
 275
 276         /* calculate ranks in the ordering above */
 277         result_rank = result < 0 ? 1 + CLO_REPEAT : result;
 278         rc_rank = rc < 0 ? 1 + CLO_REPEAT : rc;
 279
 280         if (result_rank < rc_rank)
 281                 result = rc;
 282         RETURN(result);
 283 }
 284
 285 /**
 286  * Creates sub-locks for a given lov_lock for the first time.
 287  *
 288  * Goes through all sub-objects of top-object, and creates sub-locks on every
 289  * sub-object intersecting with top-lock extent. This is complicated by the
 290  * fact that top-lock (that is being created) can be accessed concurrently
 291  * through already created sub-locks (possibly shared with other top-locks).
 292  */
 293 static int lov_lock_sub_init(const struct lu_env *env,
 294                              struct lov_lock *lck, const struct cl_io *io)
 295 {
 296         int result = 0;
 297         int i;
 298         int nr;
 299         obd_off start;
 300         obd_off end;
 301         obd_off file_start;
 302         obd_off file_end;
 303
 304         struct lov_object       *loo    = cl2lov(lck->lls_cl.cls_obj);
 305         struct lov_layout_raid0 *r0     = lov_r0(loo);
 306         struct cl_lock          *parent = lck->lls_cl.cls_lock;
 307
 308         ENTRY;
 309
 310         lck->lls_orig = parent->cll_descr;
 311         file_start = cl_offset(lov2cl(loo), parent->cll_descr.cld_start);
 312         file_end   = cl_offset(lov2cl(loo), parent->cll_descr.cld_end + 1) - 1;
 313
 314         for (i = 0, nr = 0; i < r0->lo_nr; i++) {
 315                 /*
 316                  * XXX for wide striping smarter algorithm is desirable,
 317                  * breaking out of the loop, early.
 318                  */
 319                 if (lov_stripe_intersects(r0->lo_lsm, i,
 320                                           file_start, file_end, &start, &end))
 321                         nr++;
 322         }
 323         LASSERT(nr > 0);
 324         OBD_ALLOC_LARGE(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 (i = 0, nr = 0; i < r0->lo_nr; ++i) {
 337                 if (lov_stripe_intersects(r0->lo_lsm, i,
 338                                           file_start, file_end, &start, &end)) {
 339                         struct cl_lock_descr *descr;
 340
 341                         descr = &lck->lls_sub[nr].sub_descr;
 342
 343                         LASSERT(descr->cld_obj == NULL);
 344                         descr->cld_obj   = lovsub2cl(r0->lo_sub[i]);
 345                         descr->cld_start = cl_index(descr->cld_obj, start);
 346                         descr->cld_end   = cl_index(descr->cld_obj, end);
 347                         descr->cld_mode  = parent->cll_descr.cld_mode;
 348                         descr->cld_gid   = parent->cll_descr.cld_gid;
 349                         descr->cld_enq_flags   = parent->cll_descr.cld_enq_flags;
 350                         /* XXX has no effect */
 351                         lck->lls_sub[nr].sub_got = *descr;
 352                         lck->lls_sub[nr].sub_stripe = i;
 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_get_trust(sublock);
 372                         cl_lock_mutex_get(env, sublock);
 373                         cl_lock_mutex_get(env, parent);
 374                         /*
 375                          * recheck under mutex that sub-lock wasn't created
 376                          * concurrently, and that top-lock is still alive.
 377                          */
 378                         if (lck->lls_sub[i].sub_lock == NULL &&
 379                             parent->cll_state < CLS_FREEING) {
 380                                 lov_sublock_adopt(env, lck, sublock, i, link);
 381                                 cl_lock_mutex_put(env, parent);
 382                         } else {
 383                                 OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
 384                                 cl_lock_mutex_put(env, parent);
 385                                 cl_lock_unhold(env, sublock,
 386                                                "lov-parent", parent);
 387                         }
 388                         cl_lock_mutex_put(env, sublock);
 389                         cl_lock_put(env, sublock);
 390                 }
 391         }
 392         /*
 393          * Some sub-locks can be missing at this point. This is not a problem,
 394          * because enqueue will create them anyway. Main duty of this function
 395          * is to fill in sub-lock descriptions in a race free manner.
 396          */
 397         RETURN(result);
 398 }
 399
 400 static int lov_sublock_release(const struct lu_env *env, struct lov_lock *lck,
 401                                int i, int deluser, int rc)
 402 {
 403         struct cl_lock *parent = lck->lls_cl.cls_lock;
 404
 405         LASSERT(cl_lock_is_mutexed(parent));
 406         ENTRY;
 407
 408         if (lck->lls_sub[i].sub_flags & LSF_HELD) {
 409                 struct cl_lock    *sublock;
 410                 int dying;
 411
 412                 LASSERT(lck->lls_sub[i].sub_lock != NULL);
 413                 sublock = lck->lls_sub[i].sub_lock->lss_cl.cls_lock;
 414                 LASSERT(cl_lock_is_mutexed(sublock));
 415
 416                 lck->lls_sub[i].sub_flags &= ~LSF_HELD;
 417                 if (deluser)
 418                         cl_lock_user_del(env, sublock);
 419                 /*
 420                  * If the last hold is released, and cancellation is pending
 421                  * for a sub-lock, release parent mutex, to avoid keeping it
 422                  * while sub-lock is being paged out.
 423                  */
 424                 dying = (sublock->cll_descr.cld_mode == CLM_PHANTOM ||
 425                          sublock->cll_descr.cld_mode == CLM_GROUP ||
 426                          (sublock->cll_flags & (CLF_CANCELPEND|CLF_DOOMED))) &&
 427                         sublock->cll_holds == 1;
 428                 if (dying)
 429                         cl_lock_mutex_put(env, parent);
 430                 cl_lock_unhold(env, sublock, "lov-parent", parent);
 431                 if (dying) {
 432                         cl_lock_mutex_get(env, parent);
 433                         rc = lov_subresult(rc, CLO_REPEAT);
 434                 }
 435                 /*
 436                  * From now on lck->lls_sub[i].sub_lock is a "weak" pointer,
 437                  * not backed by a reference on a
 438                  * sub-lock. lovsub_lock_delete() will clear
 439                  * lck->lls_sub[i].sub_lock under semaphores, just before
 440                  * sub-lock is destroyed.
 441                  */
 442         }
 443         RETURN(rc);
 444 }
 445
 446 static void lov_sublock_hold(const struct lu_env *env, struct lov_lock *lck,
 447                              int i)
 448 {
 449         struct cl_lock *parent = lck->lls_cl.cls_lock;
 450
 451         LASSERT(cl_lock_is_mutexed(parent));
 452         ENTRY;
 453
 454         if (!(lck->lls_sub[i].sub_flags & LSF_HELD)) {
 455                 struct cl_lock *sublock;
 456
 457                 LASSERT(lck->lls_sub[i].sub_lock != NULL);
 458                 sublock = lck->lls_sub[i].sub_lock->lss_cl.cls_lock;
 459                 LASSERT(cl_lock_is_mutexed(sublock));
 460                 LASSERT(sublock->cll_state != CLS_FREEING);
 461
 462                 lck->lls_sub[i].sub_flags |= LSF_HELD;
 463
 464                 cl_lock_get_trust(sublock);
 465                 cl_lock_hold_add(env, sublock, "lov-parent", parent);
 466                 cl_lock_user_add(env, sublock);
 467                 cl_lock_put(env, sublock);
 468         }
 469         EXIT;
 470 }
 471
 472 static void lov_lock_fini(const struct lu_env *env,
 473                           struct cl_lock_slice *slice)
 474 {
 475         struct lov_lock *lck;
 476         int i;
 477
 478         ENTRY;
 479         lck = cl2lov_lock(slice);
 480         LASSERT(lck->lls_nr_filled == 0);
 481         if (lck->lls_sub != NULL) {
 482                 for (i = 0; i < lck->lls_nr; ++i)
 483                         /*
 484                          * No sub-locks exists at this point, as sub-lock has
 485                          * a reference on its parent.
 486                          */
 487                         LASSERT(lck->lls_sub[i].sub_lock == NULL);
 488                 OBD_FREE_LARGE(lck->lls_sub,
 489                                lck->lls_nr * sizeof lck->lls_sub[0]);
 490         }
 491         OBD_SLAB_FREE_PTR(lck, lov_lock_kmem);
 492         EXIT;
 493 }
 494
 495 static int lov_lock_enqueue_wait(const struct lu_env *env,
 496                                  struct lov_lock *lck,
 497                                  struct cl_lock *sublock)
 498 {
 499         struct cl_lock *lock = lck->lls_cl.cls_lock;
 500         int             result;
 501         ENTRY;
 502
 503         LASSERT(cl_lock_is_mutexed(lock));
 504
 505         cl_lock_mutex_put(env, lock);
 506         result = cl_lock_enqueue_wait(env, sublock, 0);
 507         cl_lock_mutex_get(env, lock);
 508         RETURN(result ?: CLO_REPEAT);
 509 }
 510
 511 /**
 512  * Tries to advance a state machine of a given sub-lock toward enqueuing of
 513  * the top-lock.
 514  *
 515  * \retval 0 if state-transition can proceed
 516  * \retval -ve otherwise.
 517  */
 518 static int lov_lock_enqueue_one(const struct lu_env *env, struct lov_lock *lck,
 519                                 struct cl_lock *sublock,
 520                                 struct cl_io *io, __u32 enqflags, int last)
 521 {
 522         int result;
 523         ENTRY;
 524
 525         /* first, try to enqueue a sub-lock ... */
 526         result = cl_enqueue_try(env, sublock, io, enqflags);
 527         if (sublock->cll_state == CLS_ENQUEUED)
 528                 /* if it is enqueued, try to `wait' on it---maybe it's already
 529                  * granted */
 530                 result = cl_wait_try(env, sublock);
 531         /*
 532          * If CEF_ASYNC flag is set, then all sub-locks can be enqueued in
 533          * parallel, otherwise---enqueue has to wait until sub-lock is granted
 534          * before proceeding to the next one.
 535          */
 536         if (result == CLO_WAIT && sublock->cll_state <= CLS_HELD &&
 537             enqflags & CEF_ASYNC && !last)
 538                 result = 0;
 539         RETURN(result);
 540 }
 541
 542 /**
 543  * Helper function for lov_lock_enqueue() that creates missing sub-lock.
 544  */
 545 static int lov_sublock_fill(const struct lu_env *env, struct cl_lock *parent,
 546                             struct cl_io *io, struct lov_lock *lck, int idx)
 547 {
 548         struct lov_lock_link *link;
 549         struct cl_lock       *sublock;
 550         int                   result;
 551
 552         LASSERT(parent->cll_depth == 1);
 553         cl_lock_mutex_put(env, parent);
 554         sublock = lov_sublock_alloc(env, io, lck, idx, &link);
 555         if (!IS_ERR(sublock))
 556                 cl_lock_mutex_get(env, sublock);
 557         cl_lock_mutex_get(env, parent);
 558
 559         if (!IS_ERR(sublock)) {
 560                 cl_lock_get_trust(sublock);
 561                 if (parent->cll_state == CLS_QUEUING &&
 562                     lck->lls_sub[idx].sub_lock == NULL) {
 563                         lov_sublock_adopt(env, lck, sublock, idx, link);
 564                 } else {
 565                         OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
 566                         /* other thread allocated sub-lock, or enqueue is no
 567                          * longer going on */
 568                         cl_lock_mutex_put(env, parent);
 569                         cl_lock_unhold(env, sublock, "lov-parent", parent);
 570                         cl_lock_mutex_get(env, parent);
 571                 }
 572                 cl_lock_mutex_put(env, sublock);
 573                 cl_lock_put(env, sublock);
 574                 result = CLO_REPEAT;
 575         } else
 576                 result = PTR_ERR(sublock);
 577         return result;
 578 }
 579
 580 /**
 581  * Implementation of cl_lock_operations::clo_enqueue() for lov layer. This
 582  * function is rather subtle, as it enqueues top-lock (i.e., advances top-lock
 583  * state machine from CLS_QUEUING to CLS_ENQUEUED states) by juggling sub-lock
 584  * state machines in the face of sub-locks sharing (by multiple top-locks),
 585  * and concurrent sub-lock cancellations.
 586  */
 587 static int lov_lock_enqueue(const struct lu_env *env,
 588                             const struct cl_lock_slice *slice,
 589                             struct cl_io *io, __u32 enqflags)
 590 {
 591         struct cl_lock         *lock    = slice->cls_lock;
 592         struct lov_lock        *lck     = cl2lov_lock(slice);
 593         struct cl_lock_closure *closure = lov_closure_get(env, lock);
 594         int i;
 595         int result;
 596         enum cl_lock_state minstate;
 597
 598         ENTRY;
 599
 600         for (result = 0, minstate = CLS_FREEING, i = 0; i < lck->lls_nr; ++i) {
 601                 int rc;
 602                 struct lovsub_lock     *sub;
 603                 struct lov_lock_sub    *lls;
 604                 struct cl_lock         *sublock;
 605                 struct lov_sublock_env *subenv;
 606
 607                 if (lock->cll_state != CLS_QUEUING) {
 608                         /*
 609                          * Lock might have left QUEUING state if previous
 610                          * iteration released its mutex. Stop enqueing in this
 611                          * case and let the upper layer to decide what to do.
 612                          */
 613                         LASSERT(i > 0 && result != 0);
 614                         break;
 615                 }
 616
 617                 lls = &lck->lls_sub[i];
 618                 sub = lls->sub_lock;
 619                 /*
 620                  * Sub-lock might have been canceled, while top-lock was
 621                  * cached.
 622                  */
 623                 if (sub == NULL) {
 624                         result = lov_sublock_fill(env, lock, io, lck, i);
 625                         /* lov_sublock_fill() released @lock mutex,
 626                          * restart. */
 627                         break;
 628                 }
 629                 sublock = sub->lss_cl.cls_lock;
 630                 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
 631                 if (rc == 0) {
 632                         lov_sublock_hold(env, lck, i);
 633                         rc = lov_lock_enqueue_one(subenv->lse_env, lck, sublock,
 634                                                   subenv->lse_io, enqflags,
 635                                                   i == lck->lls_nr - 1);
 636                         minstate = min(minstate, sublock->cll_state);
 637                         if (rc == CLO_WAIT) {
 638                                 switch (sublock->cll_state) {
 639                                 case CLS_QUEUING:
 640                                         /* take recursive mutex, the lock is
 641                                          * released in lov_lock_enqueue_wait.
 642                                          */
 643                                         cl_lock_mutex_get(env, sublock);
 644                                         lov_sublock_unlock(env, sub, closure,
 645                                                            subenv);
 646                                         rc = lov_lock_enqueue_wait(env, lck,
 647                                                                    sublock);
 648                                         break;
 649                                 case CLS_CACHED:
 650                                         rc = lov_sublock_release(env, lck, i,
 651                                                                  1, rc);
 652                                 default:
 653                                         lov_sublock_unlock(env, sub, closure,
 654                                                            subenv);
 655                                         break;
 656                                 }
 657                         } else {
 658                                 LASSERT(sublock->cll_conflict == NULL);
 659                                 lov_sublock_unlock(env, sub, closure, subenv);
 660                         }
 661                 }
 662                 result = lov_subresult(result, rc);
 663                 if (result != 0)
 664                         break;
 665         }
 666         cl_lock_closure_fini(closure);
 667         RETURN(result ?: minstate >= CLS_ENQUEUED ? 0 : CLO_WAIT);
 668 }
 669
 670 static int lov_lock_unuse(const struct lu_env *env,
 671                           const struct cl_lock_slice *slice)
 672 {
 673         struct lov_lock        *lck     = cl2lov_lock(slice);
 674         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
 675         int i;
 676         int result;
 677
 678         ENTRY;
 679
 680         for (result = 0, i = 0; i < lck->lls_nr; ++i) {
 681                 int rc;
 682                 struct lovsub_lock     *sub;
 683                 struct cl_lock         *sublock;
 684                 struct lov_lock_sub    *lls;
 685                 struct lov_sublock_env *subenv;
 686
 687                 /* top-lock state cannot change concurrently, because single
 688                  * thread (one that released the last hold) carries unlocking
 689                  * to the completion. */
 690                 LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT);
 691                 lls = &lck->lls_sub[i];
 692                 sub = lls->sub_lock;
 693                 if (sub == NULL)
 694                         continue;
 695
 696                 sublock = sub->lss_cl.cls_lock;
 697                 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
 698                 if (rc == 0) {
 699                         if (lls->sub_flags & LSF_HELD) {
 700                                 LASSERT(sublock->cll_state == CLS_HELD);
 701                                 rc = cl_unuse_try(subenv->lse_env, sublock);
 702                                 rc = lov_sublock_release(env, lck, i, 0, rc);
 703                         }
 704                         lov_sublock_unlock(env, sub, closure, subenv);
 705                 }
 706                 result = lov_subresult(result, rc);
 707         }
 708
 709         if (result == 0 && lck->lls_cancel_race) {
 710                 lck->lls_cancel_race = 0;
 711                 result = -ESTALE;
 712         }
 713         cl_lock_closure_fini(closure);
 714         RETURN(result);
 715 }
 716
 717
 718 static void lov_lock_cancel(const struct lu_env *env,
 719                            const struct cl_lock_slice *slice)
 720 {
 721         struct lov_lock        *lck     = cl2lov_lock(slice);
 722         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
 723         int i;
 724         int result;
 725
 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                 /* top-lock state cannot change concurrently, because single
 736                  * thread (one that released the last hold) carries unlocking
 737                  * to the completion. */
 738                 lls = &lck->lls_sub[i];
 739                 sub = lls->sub_lock;
 740                 if (sub == NULL)
 741                         continue;
 742
 743                 sublock = sub->lss_cl.cls_lock;
 744                 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
 745                 if (rc == 0) {
 746                         if (!(lls->sub_flags & LSF_HELD)) {
 747                                 lov_sublock_unlock(env, sub, closure, subenv);
 748                                 continue;
 749                         }
 750
 751                         switch(sublock->cll_state) {
 752                         case CLS_HELD:
 753                                 rc = cl_unuse_try(subenv->lse_env,
 754                                                   sublock);
 755                                 lov_sublock_release(env, lck, i, 0, 0);
 756                                 break;
 757                         case CLS_ENQUEUED:
 758                                 /* TODO: it's not a good idea to cancel this
 759                                  * lock because it's innocent. But it's
 760                                  * acceptable. The better way would be to
 761                                  * define a new lock method to unhold the
 762                                  * dlm lock. */
 763                                 cl_lock_cancel(env, sublock);
 764                         default:
 765                                 lov_sublock_release(env, lck, i, 1, 0);
 766                                 break;
 767                         }
 768                         lov_sublock_unlock(env, sub, closure, subenv);
 769                 }
 770
 771                 if (rc == CLO_REPEAT) {
 772                         --i;
 773                         continue;
 774                 }
 775
 776                 result = lov_subresult(result, rc);
 777         }
 778
 779         if (result)
 780                 CL_LOCK_DEBUG(D_ERROR, env, slice->cls_lock,
 781                               "lov_lock_cancel fails with %d.\n", result);
 782
 783         cl_lock_closure_fini(closure);
 784 }
 785
 786 static int lov_lock_wait(const struct lu_env *env,
 787                          const struct cl_lock_slice *slice)
 788 {
 789         struct lov_lock        *lck     = cl2lov_lock(slice);
 790         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
 791         enum cl_lock_state      minstate;
 792         int                     result;
 793         int                     i;
 794
 795         ENTRY;
 796
 797         for (result = 0, minstate = CLS_FREEING, i = 0; i < lck->lls_nr; ++i) {
 798                 int rc;
 799                 struct lovsub_lock     *sub;
 800                 struct cl_lock         *sublock;
 801                 struct lov_lock_sub    *lls;
 802                 struct lov_sublock_env *subenv;
 803
 804                 lls = &lck->lls_sub[i];
 805                 sub = lls->sub_lock;
 806                 LASSERT(sub != NULL);
 807                 sublock = sub->lss_cl.cls_lock;
 808                 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
 809                 if (rc == 0) {
 810                         LASSERT(sublock->cll_state >= CLS_ENQUEUED);
 811                         if (sublock->cll_state < CLS_HELD)
 812                                 rc = cl_wait_try(env, sublock);
 813
 814                         minstate = min(minstate, sublock->cll_state);
 815                         lov_sublock_unlock(env, sub, closure, subenv);
 816                 }
 817                 result = lov_subresult(result, rc);
 818                 if (result != 0)
 819                         break;
 820         }
 821         cl_lock_closure_fini(closure);
 822         RETURN(result ?: minstate >= CLS_HELD ? 0 : CLO_WAIT);
 823 }
 824
 825 static int lov_lock_use(const struct lu_env *env,
 826                         const struct cl_lock_slice *slice)
 827 {
 828         struct lov_lock        *lck     = cl2lov_lock(slice);
 829         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
 830         int                     result;
 831         int                     i;
 832
 833         LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT);
 834         ENTRY;
 835
 836         for (result = 0, i = 0; i < lck->lls_nr; ++i) {
 837                 int rc;
 838                 struct lovsub_lock     *sub;
 839                 struct cl_lock         *sublock;
 840                 struct lov_lock_sub    *lls;
 841                 struct lov_sublock_env *subenv;
 842
 843                 LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT);
 844
 845                 lls = &lck->lls_sub[i];
 846                 sub = lls->sub_lock;
 847                 if (sub == NULL) {
 848                         /*
 849                          * Sub-lock might have been canceled, while top-lock was
 850                          * cached.
 851                          */
 852                         result = -ESTALE;
 853                         break;
 854                 }
 855
 856                 sublock = sub->lss_cl.cls_lock;
 857                 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
 858                 if (rc == 0) {
 859                         LASSERT(sublock->cll_state != CLS_FREEING);
 860                         lov_sublock_hold(env, lck, i);
 861                         if (sublock->cll_state == CLS_CACHED) {
 862                                 rc = cl_use_try(subenv->lse_env, sublock, 0);
 863                                 if (rc != 0)
 864                                         rc = lov_sublock_release(env, lck,
 865                                                                  i, 1, rc);
 866                         }
 867                         lov_sublock_unlock(env, sub, closure, subenv);
 868                 }
 869                 result = lov_subresult(result, rc);
 870                 if (result != 0)
 871                         break;
 872         }
 873
 874         if (lck->lls_cancel_race) {
 875                 /*
 876                  * If there is unlocking happened at the same time, then
 877                  * sublock_lock state should be FREEING, and lov_sublock_lock
 878                  * should return CLO_REPEAT. In this case, it should return
 879                  * ESTALE, and up layer should reset the lock state to be NEW.
 880                  */
 881                 lck->lls_cancel_race = 0;
 882                 LASSERT(result != 0);
 883                 result = -ESTALE;
 884         }
 885         cl_lock_closure_fini(closure);
 886         RETURN(result);
 887 }
 888
 889 #if 0
 890 static int lock_lock_multi_match()
 891 {
 892         struct cl_lock          *lock    = slice->cls_lock;
 893         struct cl_lock_descr    *subneed = &lov_env_info(env)->lti_ldescr;
 894         struct lov_object       *loo     = cl2lov(lov->lls_cl.cls_obj);
 895         struct lov_layout_raid0 *r0      = lov_r0(loo);
 896         struct lov_lock_sub     *sub;
 897         struct cl_object        *subobj;
 898         obd_off  fstart;
 899         obd_off  fend;
 900         obd_off  start;
 901         obd_off  end;
 902         int i;
 903
 904         fstart = cl_offset(need->cld_obj, need->cld_start);
 905         fend   = cl_offset(need->cld_obj, need->cld_end + 1) - 1;
 906         subneed->cld_mode = need->cld_mode;
 907         cl_lock_mutex_get(env, lock);
 908         for (i = 0; i < lov->lls_nr; ++i) {
 909                 sub = &lov->lls_sub[i];
 910                 if (sub->sub_lock == NULL)
 911                         continue;
 912                 subobj = sub->sub_descr.cld_obj;
 913                 if (!lov_stripe_intersects(r0->lo_lsm, sub->sub_stripe,
 914                                            fstart, fend, &start, &end))
 915                         continue;
 916                 subneed->cld_start = cl_index(subobj, start);
 917                 subneed->cld_end   = cl_index(subobj, end);
 918                 subneed->cld_obj   = subobj;
 919                 if (!cl_lock_ext_match(&sub->sub_got, subneed)) {
 920                         result = 0;
 921                         break;
 922                 }
 923         }
 924         cl_lock_mutex_put(env, lock);
 925 }
 926 #endif
 927
 928 /**
 929  * Check if the extent region \a descr is covered by \a child against the
 930  * specific \a stripe.
 931  */
 932 static int lov_lock_stripe_is_matching(const struct lu_env *env,
 933                                        struct lov_object *lov, int stripe,
 934                                        const struct cl_lock_descr *child,
 935                                        const struct cl_lock_descr *descr)
 936 {
 937         struct lov_stripe_md *lsm = lov_r0(lov)->lo_lsm;
 938         obd_off start;
 939         obd_off end;
 940         int result;
 941
 942         if (lov_r0(lov)->lo_nr == 1)
 943                 return cl_lock_ext_match(child, descr);
 944
 945         /*
 946          * For a multi-stripes object:
 947          * - make sure the descr only covers child's stripe, and
 948          * - check if extent is matching.
 949          */
 950         start = cl_offset(&lov->lo_cl, descr->cld_start);
 951         end   = cl_offset(&lov->lo_cl, descr->cld_end + 1) - 1;
 952         result = end - start <= lsm->lsm_stripe_size &&
 953                  stripe == lov_stripe_number(lsm, start) &&
 954                  stripe == lov_stripe_number(lsm, end);
 955         if (result) {
 956                 struct cl_lock_descr *subd = &lov_env_info(env)->lti_ldescr;
 957                 obd_off sub_start;
 958                 obd_off sub_end;
 959
 960                 subd->cld_obj  = NULL;   /* don't need sub object at all */
 961                 subd->cld_mode = descr->cld_mode;
 962                 subd->cld_gid  = descr->cld_gid;
 963                 result = lov_stripe_intersects(lsm, stripe, start, end,
 964                                                &sub_start, &sub_end);
 965                 LASSERT(result);
 966                 subd->cld_start = cl_index(child->cld_obj, sub_start);
 967                 subd->cld_end   = cl_index(child->cld_obj, sub_end);
 968                 result = cl_lock_ext_match(child, subd);
 969         }
 970         return result;
 971 }
 972
 973 /**
 974  * An implementation of cl_lock_operations::clo_fits_into() method.
 975  *
 976  * Checks whether a lock (given by \a slice) is suitable for \a
 977  * io. Multi-stripe locks can be used only for "quick" io, like truncate, or
 978  * O_APPEND write.
 979  *
 980  * \see ccc_lock_fits_into().
 981  */
 982 static int lov_lock_fits_into(const struct lu_env *env,
 983                               const struct cl_lock_slice *slice,
 984                               const struct cl_lock_descr *need,
 985                               const struct cl_io *io)
 986 {
 987         struct lov_lock   *lov = cl2lov_lock(slice);
 988         struct lov_object *obj = cl2lov(slice->cls_obj);
 989         int result;
 990
 991         LASSERT(cl_object_same(need->cld_obj, slice->cls_obj));
 992         LASSERT(lov->lls_nr > 0);
 993
 994         ENTRY;
 995
 996         if (need->cld_mode == CLM_GROUP)
 997                 /*
 998                  * always allow to match group lock.
 999                  */
1000                 result = cl_lock_ext_match(&lov->lls_orig, need);
1001         else if (lov->lls_nr == 1) {
1002                 struct cl_lock_descr *got = &lov->lls_sub[0].sub_got;
1003                 result = lov_lock_stripe_is_matching(env,
1004                                                      cl2lov(slice->cls_obj),
1005                                                      lov->lls_sub[0].sub_stripe,
1006                                                      got, need);
1007         } else if (io->ci_type != CIT_SETATTR && io->ci_type != CIT_MISC &&
1008                    !cl_io_is_append(io) && need->cld_mode != CLM_PHANTOM)
1009                 /*
1010                  * Multi-stripe locks are only suitable for `quick' IO and for
1011                  * glimpse.
1012                  */
1013                 result = 0;
1014         else
1015                 /*
1016                  * Most general case: multi-stripe existing lock, and
1017                  * (potentially) multi-stripe @need lock. Check that @need is
1018                  * covered by @lov's sub-locks.
1019                  *
1020                  * For now, ignore lock expansions made by the server, and
1021                  * match against original lock extent.
1022                  */
1023                 result = cl_lock_ext_match(&lov->lls_orig, need);
1024         CDEBUG(D_DLMTRACE, DDESCR"/"DDESCR" %d %d/%d: %d\n",
1025                PDESCR(&lov->lls_orig), PDESCR(&lov->lls_sub[0].sub_got),
1026                lov->lls_sub[0].sub_stripe, lov->lls_nr, lov_r0(obj)->lo_nr,
1027                result);
1028         RETURN(result);
1029 }
1030
1031 void lov_lock_unlink(const struct lu_env *env,
1032                      struct lov_lock_link *link, struct lovsub_lock *sub)
1033 {
1034         struct lov_lock *lck    = link->lll_super;
1035         struct cl_lock  *parent = lck->lls_cl.cls_lock;
1036
1037         LASSERT(cl_lock_is_mutexed(parent));
1038         LASSERT(cl_lock_is_mutexed(sub->lss_cl.cls_lock));
1039         ENTRY;
1040
1041         cfs_list_del_init(&link->lll_list);
1042         LASSERT(lck->lls_sub[link->lll_idx].sub_lock == sub);
1043         /* yank this sub-lock from parent's array */
1044         lck->lls_sub[link->lll_idx].sub_lock = NULL;
1045         LASSERT(lck->lls_nr_filled > 0);
1046         lck->lls_nr_filled--;
1047         lu_ref_del(&parent->cll_reference, "lov-child", sub->lss_cl.cls_lock);
1048         cl_lock_put(env, parent);
1049         OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
1050         EXIT;
1051 }
1052
1053 struct lov_lock_link *lov_lock_link_find(const struct lu_env *env,
1054                                          struct lov_lock *lck,
1055                                          struct lovsub_lock *sub)
1056 {
1057         struct lov_lock_link *scan;
1058
1059         LASSERT(cl_lock_is_mutexed(sub->lss_cl.cls_lock));
1060         ENTRY;
1061
1062         cfs_list_for_each_entry(scan, &sub->lss_parents, lll_list) {
1063                 if (scan->lll_super == lck)
1064                         RETURN(scan);
1065         }
1066         RETURN(NULL);
1067 }
1068
1069 /**
1070  * An implementation of cl_lock_operations::clo_delete() method. This is
1071  * invoked for "top-to-bottom" delete, when lock destruction starts from the
1072  * top-lock, e.g., as a result of inode destruction.
1073  *
1074  * Unlinks top-lock from all its sub-locks. Sub-locks are not deleted there:
1075  * this is done separately elsewhere:
1076  *
1077  *     - for inode destruction, lov_object_delete() calls cl_object_kill() for
1078  *       each sub-object, purging its locks;
1079  *
1080  *     - in other cases (e.g., a fatal error with a top-lock) sub-locks are
1081  *       left in the cache.
1082  */
1083 static void lov_lock_delete(const struct lu_env *env,
1084                             const struct cl_lock_slice *slice)
1085 {
1086         struct lov_lock        *lck     = cl2lov_lock(slice);
1087         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
1088         struct lov_lock_link   *link;
1089         int                     rc;
1090         int                     i;
1091
1092         LASSERT(slice->cls_lock->cll_state == CLS_FREEING);
1093         ENTRY;
1094
1095         for (i = 0; i < lck->lls_nr; ++i) {
1096                 struct lov_lock_sub *lls = &lck->lls_sub[i];
1097                 struct lovsub_lock  *lsl = lls->sub_lock;
1098
1099                 if (lsl == NULL) /* already removed */
1100                         continue;
1101
1102                 rc = lov_sublock_lock(env, lck, lls, closure, NULL);
1103                 if (rc == CLO_REPEAT) {
1104                         --i;
1105                         continue;
1106                 }
1107
1108                 LASSERT(rc == 0);
1109                 LASSERT(lsl->lss_cl.cls_lock->cll_state < CLS_FREEING);
1110
1111                 if (lls->sub_flags & LSF_HELD)
1112                         lov_sublock_release(env, lck, i, 1, 0);
1113
1114                 link = lov_lock_link_find(env, lck, lsl);
1115                 LASSERT(link != NULL);
1116                 lov_lock_unlink(env, link, lsl);
1117                 LASSERT(lck->lls_sub[i].sub_lock == NULL);
1118
1119                 lov_sublock_unlock(env, lsl, closure, NULL);
1120         }
1121
1122         cl_lock_closure_fini(closure);
1123         EXIT;
1124 }
1125
1126 static int lov_lock_print(const struct lu_env *env, void *cookie,
1127                           lu_printer_t p, const struct cl_lock_slice *slice)
1128 {
1129         struct lov_lock *lck = cl2lov_lock(slice);
1130         int              i;
1131
1132         (*p)(env, cookie, "%d\n", lck->lls_nr);
1133         for (i = 0; i < lck->lls_nr; ++i) {
1134                 struct lov_lock_sub *sub;
1135
1136                 sub = &lck->lls_sub[i];
1137                 (*p)(env, cookie, "    %d %x: ", i, sub->sub_flags);
1138                 if (sub->sub_lock != NULL)
1139                         cl_lock_print(env, cookie, p,
1140                                       sub->sub_lock->lss_cl.cls_lock);
1141                 else
1142                         (*p)(env, cookie, "---\n");
1143         }
1144         return 0;
1145 }
1146
1147 static const struct cl_lock_operations lov_lock_ops = {
1148         .clo_fini      = lov_lock_fini,
1149         .clo_enqueue   = lov_lock_enqueue,
1150         .clo_wait      = lov_lock_wait,
1151         .clo_use       = lov_lock_use,
1152         .clo_unuse     = lov_lock_unuse,
1153         .clo_cancel    = lov_lock_cancel,
1154         .clo_fits_into = lov_lock_fits_into,
1155         .clo_delete    = lov_lock_delete,
1156         .clo_print     = lov_lock_print
1157 };
1158
1159 int lov_lock_init_raid0(const struct lu_env *env, struct cl_object *obj,
1160                         struct cl_lock *lock, const struct cl_io *io)
1161 {
1162         struct lov_lock *lck;
1163         int result;
1164
1165         ENTRY;
1166         OBD_SLAB_ALLOC_PTR_GFP(lck, lov_lock_kmem, CFS_ALLOC_IO);
1167         if (lck != NULL) {
1168                 cl_lock_slice_add(lock, &lck->lls_cl, obj, &lov_lock_ops);
1169                 result = lov_lock_sub_init(env, lck, io);
1170         } else
1171                 result = -ENOMEM;
1172         RETURN(result);
1173 }
1174
1175 static struct cl_lock_closure *lov_closure_get(const struct lu_env *env,
1176                                                struct cl_lock *parent)
1177 {
1178         struct cl_lock_closure *closure;
1179
1180         closure = &lov_env_info(env)->lti_closure;
1181         LASSERT(cfs_list_empty(&closure->clc_list));
1182         cl_lock_closure_init(env, closure, parent, 1);
1183         return closure;
1184 }
1185
1186
1187 /** @} lov */