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 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 (!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         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(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                         struct lov_lock_link *link;
 221                         /*
 222                          * we could race with lock deletion which temporarily
 223                          * put the lock in freeing state, bug 19080.
 224                          */
 225                         LASSERT(!(lls->sub_flags & LSF_HELD));
 226
 227                         link = lov_lock_link_find(env, lck, sublock);
 228                         LASSERT(link != NULL);
 229                         lov_lock_unlink(env, link, sublock);
 230                         lov_sublock_unlock(env, sublock, closure, NULL);
 231                         lck->lls_cancel_race = 1;
 232                         result = CLO_REPEAT;
 233                 } else if (lsep) {
 234                         struct lov_sublock_env *subenv;
 235                         subenv = lov_sublock_env_get(env, parent, lls);
 236                         if (IS_ERR(subenv)) {
 237                                 lov_sublock_unlock(env, sublock,
 238                                                    closure, NULL);
 239                                 result = PTR_ERR(subenv);
 240                         } else {
 241                                 *lsep = subenv;
 242                         }
 243                 }
 244         }
 245         RETURN(result);
 246 }
 247
 248 /**
 249  * Updates the result of a top-lock operation from a result of sub-lock
 250  * sub-operations. Top-operations like lov_lock_{enqueue,use,unuse}() iterate
 251  * over sub-locks and lov_subresult() is used to calculate return value of a
 252  * top-operation. To this end, possible return values of sub-operations are
 253  * ordered as
 254  *
 255  *     - 0                  success
 256  *     - CLO_WAIT           wait for event
 257  *     - CLO_REPEAT         repeat top-operation
 258  *     - -ne                fundamental error
 259  *
 260  * Top-level return code can only go down through this list. CLO_REPEAT
 261  * overwrites CLO_WAIT, because lock mutex was released and sleeping condition
 262  * has to be rechecked by the upper layer.
 263  */
 264 static int lov_subresult(int result, int rc)
 265 {
 266         int result_rank;
 267         int rc_rank;
 268
 269         LASSERT(result <= 0 || result == CLO_REPEAT || result == CLO_WAIT);
 270         LASSERT(rc <= 0 || rc == CLO_REPEAT || rc == CLO_WAIT);
 271         CLASSERT(CLO_WAIT < CLO_REPEAT);
 272
 273         ENTRY;
 274
 275         /* calculate ranks in the ordering above */
 276         result_rank = result < 0 ? 1 + CLO_REPEAT : result;
 277         rc_rank = rc < 0 ? 1 + CLO_REPEAT : rc;
 278
 279         if (result_rank < rc_rank)
 280                 result = rc;
 281         RETURN(result);
 282 }
 283
 284 /**
 285  * Creates sub-locks for a given lov_lock for the first time.
 286  *
 287  * Goes through all sub-objects of top-object, and creates sub-locks on every
 288  * sub-object intersecting with top-lock extent. This is complicated by the
 289  * fact that top-lock (that is being created) can be accessed concurrently
 290  * through already created sub-locks (possibly shared with other top-locks).
 291  */
 292 static int lov_lock_sub_init(const struct lu_env *env,
 293                              struct lov_lock *lck, const struct cl_io *io)
 294 {
 295         int result = 0;
 296         int i;
 297         int nr;
 298         obd_off start;
 299         obd_off end;
 300         obd_off file_start;
 301         obd_off file_end;
 302
 303         struct lov_object       *loo    = cl2lov(lck->lls_cl.cls_obj);
 304         struct lov_layout_raid0 *r0     = lov_r0(loo);
 305         struct cl_lock          *parent = lck->lls_cl.cls_lock;
 306
 307         ENTRY;
 308
 309         lck->lls_orig = parent->cll_descr;
 310         file_start = cl_offset(lov2cl(loo), parent->cll_descr.cld_start);
 311         file_end   = cl_offset(lov2cl(loo), parent->cll_descr.cld_end + 1) - 1;
 312
 313         for (i = 0, nr = 0; i < r0->lo_nr; i++) {
 314                 /*
 315                  * XXX for wide striping smarter algorithm is desirable,
 316                  * breaking out of the loop, early.
 317                  */
 318                 if (lov_stripe_intersects(r0->lo_lsm, i,
 319                                           file_start, file_end, &start, &end))
 320                         nr++;
 321         }
 322         LASSERT(nr > 0);
 323         OBD_ALLOC(lck->lls_sub, nr * sizeof lck->lls_sub[0]);
 324         if (lck->lls_sub == NULL)
 325                 RETURN(-ENOMEM);
 326
 327         lck->lls_nr = nr;
 328         /*
 329          * First, fill in sub-lock descriptions in
 330          * lck->lls_sub[].sub_descr. They are used by lov_sublock_alloc()
 331          * (called below in this function, and by lov_lock_enqueue()) to
 332          * create sub-locks. At this moment, no other thread can access
 333          * top-lock.
 334          */
 335         for (i = 0, nr = 0; i < r0->lo_nr; ++i) {
 336                 if (lov_stripe_intersects(r0->lo_lsm, i,
 337                                           file_start, file_end, &start, &end)) {
 338                         struct cl_lock_descr *descr;
 339
 340                         descr = &lck->lls_sub[nr].sub_descr;
 341
 342                         LASSERT(descr->cld_obj == NULL);
 343                         descr->cld_obj   = lovsub2cl(r0->lo_sub[i]);
 344                         descr->cld_start = cl_index(descr->cld_obj, start);
 345                         descr->cld_end   = cl_index(descr->cld_obj, end);
 346                         descr->cld_mode  = parent->cll_descr.cld_mode;
 347                         descr->cld_gid   = parent->cll_descr.cld_gid;
 348                         descr->cld_enq_flags   = parent->cll_descr.cld_enq_flags;
 349                         /* XXX has no effect */
 350                         lck->lls_sub[nr].sub_got = *descr;
 351                         lck->lls_sub[nr].sub_stripe = i;
 352                         nr++;
 353                 }
 354         }
 355         LASSERT(nr == lck->lls_nr);
 356         /*
 357          * Then, create sub-locks. Once at least one sub-lock was created,
 358          * top-lock can be reached by other threads.
 359          */
 360         for (i = 0; i < lck->lls_nr; ++i) {
 361                 struct cl_lock       *sublock;
 362                 struct lov_lock_link *link;
 363
 364                 if (lck->lls_sub[i].sub_lock == NULL) {
 365                         sublock = lov_sublock_alloc(env, io, lck, i, &link);
 366                         if (IS_ERR(sublock)) {
 367                                 result = PTR_ERR(sublock);
 368                                 break;
 369                         }
 370                         cl_lock_get_trust(sublock);
 371                         cl_lock_mutex_get(env, sublock);
 372                         cl_lock_mutex_get(env, parent);
 373                         /*
 374                          * recheck under mutex that sub-lock wasn't created
 375                          * concurrently, and that top-lock is still alive.
 376                          */
 377                         if (lck->lls_sub[i].sub_lock == NULL &&
 378                             parent->cll_state < CLS_FREEING) {
 379                                 lov_sublock_adopt(env, lck, sublock, i, link);
 380                                 cl_lock_mutex_put(env, parent);
 381                         } else {
 382                                 OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
 383                                 cl_lock_mutex_put(env, parent);
 384                                 cl_lock_unhold(env, sublock,
 385                                                "lov-parent", parent);
 386                         }
 387                         cl_lock_mutex_put(env, sublock);
 388                         cl_lock_put(env, sublock);
 389                 }
 390         }
 391         /*
 392          * Some sub-locks can be missing at this point. This is not a problem,
 393          * because enqueue will create them anyway. Main duty of this function
 394          * is to fill in sub-lock descriptions in a race free manner.
 395          */
 396         RETURN(result);
 397 }
 398
 399 static int lov_sublock_release(const struct lu_env *env, struct lov_lock *lck,
 400                                int i, int deluser, int rc)
 401 {
 402         struct cl_lock *parent = lck->lls_cl.cls_lock;
 403
 404         LASSERT(cl_lock_is_mutexed(parent));
 405         ENTRY;
 406
 407         if (lck->lls_sub[i].sub_flags & LSF_HELD) {
 408                 struct cl_lock    *sublock;
 409                 int dying;
 410
 411                 LASSERT(lck->lls_sub[i].sub_lock != NULL);
 412                 sublock = lck->lls_sub[i].sub_lock->lss_cl.cls_lock;
 413                 LASSERT(cl_lock_is_mutexed(sublock));
 414
 415                 lck->lls_sub[i].sub_flags &= ~LSF_HELD;
 416                 if (deluser)
 417                         cl_lock_user_del(env, sublock);
 418                 /*
 419                  * If the last hold is released, and cancellation is pending
 420                  * for a sub-lock, release parent mutex, to avoid keeping it
 421                  * while sub-lock is being paged out.
 422                  */
 423                 dying = (sublock->cll_descr.cld_mode == CLM_PHANTOM ||
 424                          sublock->cll_descr.cld_mode == CLM_GROUP ||
 425                          (sublock->cll_flags & (CLF_CANCELPEND|CLF_DOOMED))) &&
 426                         sublock->cll_holds == 1;
 427                 if (dying)
 428                         cl_lock_mutex_put(env, parent);
 429                 cl_lock_unhold(env, sublock, "lov-parent", parent);
 430                 if (dying) {
 431                         cl_lock_mutex_get(env, parent);
 432                         rc = lov_subresult(rc, CLO_REPEAT);
 433                 }
 434                 /*
 435                  * From now on lck->lls_sub[i].sub_lock is a "weak" pointer,
 436                  * not backed by a reference on a
 437                  * sub-lock. lovsub_lock_delete() will clear
 438                  * lck->lls_sub[i].sub_lock under semaphores, just before
 439                  * sub-lock is destroyed.
 440                  */
 441         }
 442         RETURN(rc);
 443 }
 444
 445 static void lov_sublock_hold(const struct lu_env *env, struct lov_lock *lck,
 446                              int i)
 447 {
 448         struct cl_lock *parent = lck->lls_cl.cls_lock;
 449
 450         LASSERT(cl_lock_is_mutexed(parent));
 451         ENTRY;
 452
 453         if (!(lck->lls_sub[i].sub_flags & LSF_HELD)) {
 454                 struct cl_lock *sublock;
 455
 456                 LASSERT(lck->lls_sub[i].sub_lock != NULL);
 457                 sublock = lck->lls_sub[i].sub_lock->lss_cl.cls_lock;
 458                 LASSERT(cl_lock_is_mutexed(sublock));
 459                 LASSERT(sublock->cll_state != CLS_FREEING);
 460
 461                 lck->lls_sub[i].sub_flags |= LSF_HELD;
 462
 463                 cl_lock_get_trust(sublock);
 464                 cl_lock_hold_add(env, sublock, "lov-parent", parent);
 465                 cl_lock_user_add(env, sublock);
 466                 cl_lock_put(env, sublock);
 467         }
 468         EXIT;
 469 }
 470
 471 static void lov_lock_fini(const struct lu_env *env,
 472                           struct cl_lock_slice *slice)
 473 {
 474         struct lov_lock *lck;
 475         int i;
 476
 477         ENTRY;
 478         lck = cl2lov_lock(slice);
 479         LASSERT(lck->lls_nr_filled == 0);
 480         if (lck->lls_sub != NULL) {
 481                 for (i = 0; i < lck->lls_nr; ++i)
 482                         /*
 483                          * No sub-locks exists at this point, as sub-lock has
 484                          * a reference on its parent.
 485                          */
 486                         LASSERT(lck->lls_sub[i].sub_lock == NULL);
 487                 OBD_FREE(lck->lls_sub, lck->lls_nr * sizeof lck->lls_sub[0]);
 488         }
 489         OBD_SLAB_FREE_PTR(lck, lov_lock_kmem);
 490         EXIT;
 491 }
 492
 493 /**
 494  * Tries to advance a state machine of a given sub-lock toward enqueuing of
 495  * the top-lock.
 496  *
 497  * \retval 0 if state-transition can proceed
 498  * \retval -ve otherwise.
 499  */
 500 static int lov_lock_enqueue_one(const struct lu_env *env, struct lov_lock *lck,
 501                                 struct cl_lock *sublock,
 502                                 struct cl_io *io, __u32 enqflags, int last)
 503 {
 504         int result;
 505         ENTRY;
 506
 507         /* first, try to enqueue a sub-lock ... */
 508         result = cl_enqueue_try(env, sublock, io, enqflags);
 509         if (sublock->cll_state == CLS_ENQUEUED)
 510                 /* if it is enqueued, try to `wait' on it---maybe it's already
 511                  * granted */
 512                 result = cl_wait_try(env, sublock);
 513         /*
 514          * If CEF_ASYNC flag is set, then all sub-locks can be enqueued in
 515          * parallel, otherwise---enqueue has to wait until sub-lock is granted
 516          * before proceeding to the next one.
 517          */
 518         if (result == CLO_WAIT && sublock->cll_state <= CLS_HELD &&
 519             enqflags & CEF_ASYNC && !last)
 520                 result = 0;
 521         RETURN(result);
 522 }
 523
 524 /**
 525  * Helper function for lov_lock_enqueue() that creates missing sub-lock.
 526  */
 527 static int lov_sublock_fill(const struct lu_env *env, struct cl_lock *parent,
 528                             struct cl_io *io, struct lov_lock *lck, int idx)
 529 {
 530         struct lov_lock_link *link;
 531         struct cl_lock       *sublock;
 532         int                   result;
 533
 534         LASSERT(parent->cll_depth == 1);
 535         cl_lock_mutex_put(env, parent);
 536         sublock = lov_sublock_alloc(env, io, lck, idx, &link);
 537         if (!IS_ERR(sublock))
 538                 cl_lock_mutex_get(env, sublock);
 539         cl_lock_mutex_get(env, parent);
 540
 541         if (!IS_ERR(sublock)) {
 542                 cl_lock_get_trust(sublock);
 543                 if (parent->cll_state == CLS_QUEUING &&
 544                     lck->lls_sub[idx].sub_lock == NULL) {
 545                         lov_sublock_adopt(env, lck, sublock, idx, link);
 546                 } else {
 547                         OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
 548                         /* other thread allocated sub-lock, or enqueue is no
 549                          * longer going on */
 550                         cl_lock_mutex_put(env, parent);
 551                         cl_lock_unhold(env, sublock, "lov-parent", parent);
 552                         cl_lock_mutex_get(env, parent);
 553                 }
 554                 cl_lock_mutex_put(env, sublock);
 555                 cl_lock_put(env, sublock);
 556                 result = CLO_REPEAT;
 557         } else
 558                 result = PTR_ERR(sublock);
 559         return result;
 560 }
 561
 562 /**
 563  * Implementation of cl_lock_operations::clo_enqueue() for lov layer. This
 564  * function is rather subtle, as it enqueues top-lock (i.e., advances top-lock
 565  * state machine from CLS_QUEUING to CLS_ENQUEUED states) by juggling sub-lock
 566  * state machines in the face of sub-locks sharing (by multiple top-locks),
 567  * and concurrent sub-lock cancellations.
 568  */
 569 static int lov_lock_enqueue(const struct lu_env *env,
 570                             const struct cl_lock_slice *slice,
 571                             struct cl_io *io, __u32 enqflags)
 572 {
 573         struct cl_lock         *lock    = slice->cls_lock;
 574         struct lov_lock        *lck     = cl2lov_lock(slice);
 575         struct cl_lock_closure *closure = lov_closure_get(env, lock);
 576         int i;
 577         int result;
 578         enum cl_lock_state minstate;
 579
 580         ENTRY;
 581
 582         for (result = 0, minstate = CLS_FREEING, i = 0; i < lck->lls_nr; ++i) {
 583                 int rc;
 584                 struct lovsub_lock     *sub;
 585                 struct lov_lock_sub    *lls;
 586                 struct cl_lock         *sublock;
 587                 struct lov_sublock_env *subenv;
 588
 589                 if (lock->cll_state != CLS_QUEUING) {
 590                         /*
 591                          * Lock might have left QUEUING state if previous
 592                          * iteration released its mutex. Stop enqueing in this
 593                          * case and let the upper layer to decide what to do.
 594                          */
 595                         LASSERT(i > 0 && result != 0);
 596                         break;
 597                 }
 598
 599                 lls = &lck->lls_sub[i];
 600                 sub = lls->sub_lock;
 601                 /*
 602                  * Sub-lock might have been canceled, while top-lock was
 603                  * cached.
 604                  */
 605                 if (sub == NULL) {
 606                         result = lov_sublock_fill(env, lock, io, lck, i);
 607                         /* lov_sublock_fill() released @lock mutex,
 608                          * restart. */
 609                         break;
 610                 }
 611                 sublock = sub->lss_cl.cls_lock;
 612                 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
 613                 if (rc == 0) {
 614                         lov_sublock_hold(env, lck, i);
 615                         rc = lov_lock_enqueue_one(subenv->lse_env, lck, sublock,
 616                                                   subenv->lse_io, enqflags,
 617                                                   i == lck->lls_nr - 1);
 618                         minstate = min(minstate, sublock->cll_state);
 619                         /*
 620                          * Don't hold a sub-lock in CLS_CACHED state, see
 621                          * description for lov_lock::lls_sub.
 622                          */
 623                         if (sublock->cll_state > CLS_HELD)
 624                                 rc = lov_sublock_release(env, lck, i, 1, rc);
 625                         lov_sublock_unlock(env, sub, closure, subenv);
 626                 }
 627                 result = lov_subresult(result, rc);
 628                 if (result != 0)
 629                         break;
 630         }
 631         cl_lock_closure_fini(closure);
 632         RETURN(result ?: minstate >= CLS_ENQUEUED ? 0 : CLO_WAIT);
 633 }
 634
 635 static int lov_lock_unuse(const struct lu_env *env,
 636                           const struct cl_lock_slice *slice)
 637 {
 638         struct lov_lock        *lck     = cl2lov_lock(slice);
 639         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
 640         int i;
 641         int result;
 642
 643         ENTRY;
 644
 645         for (result = 0, i = 0; i < lck->lls_nr; ++i) {
 646                 int rc;
 647                 struct lovsub_lock     *sub;
 648                 struct cl_lock         *sublock;
 649                 struct lov_lock_sub    *lls;
 650                 struct lov_sublock_env *subenv;
 651
 652                 /* top-lock state cannot change concurrently, because single
 653                  * thread (one that released the last hold) carries unlocking
 654                  * to the completion. */
 655                 LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT);
 656                 lls = &lck->lls_sub[i];
 657                 sub = lls->sub_lock;
 658                 if (sub == NULL)
 659                         continue;
 660
 661                 sublock = sub->lss_cl.cls_lock;
 662                 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
 663                 if (rc == 0) {
 664                         if (lls->sub_flags & LSF_HELD) {
 665                                 LASSERT(sublock->cll_state == CLS_HELD);
 666                                 rc = cl_unuse_try(subenv->lse_env, sublock);
 667                                 if (rc != CLO_WAIT)
 668                                         rc = lov_sublock_release(env, lck,
 669                                                                  i, 0, rc);
 670                         }
 671                         lov_sublock_unlock(env, sub, closure, subenv);
 672                 }
 673                 result = lov_subresult(result, rc);
 674                 if (result < 0)
 675                         break;
 676         }
 677
 678         if (result == 0 && lck->lls_cancel_race) {
 679                 lck->lls_cancel_race = 0;
 680                 result = -ESTALE;
 681         }
 682         cl_lock_closure_fini(closure);
 683         RETURN(result);
 684 }
 685
 686
 687 static void lov_lock_cancel(const struct lu_env *env,
 688                            const struct cl_lock_slice *slice)
 689 {
 690         struct lov_lock        *lck     = cl2lov_lock(slice);
 691         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
 692         int i;
 693         int result;
 694
 695         ENTRY;
 696
 697         for (result = 0, i = 0; i < lck->lls_nr; ++i) {
 698                 int rc;
 699                 struct lovsub_lock     *sub;
 700                 struct cl_lock         *sublock;
 701                 struct lov_lock_sub    *lls;
 702                 struct lov_sublock_env *subenv;
 703
 704                 /* top-lock state cannot change concurrently, because single
 705                  * thread (one that released the last hold) carries unlocking
 706                  * to the completion. */
 707                 lls = &lck->lls_sub[i];
 708                 sub = lls->sub_lock;
 709                 if (sub == NULL)
 710                         continue;
 711
 712                 sublock = sub->lss_cl.cls_lock;
 713                 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
 714                 if (rc == 0) {
 715                         if (lls->sub_flags & LSF_HELD) {
 716                                 if (sublock->cll_state == CLS_HELD) {
 717                                         rc = cl_unuse_try(subenv->lse_env,
 718                                                           sublock);
 719                                         lov_sublock_release(env, lck, i, 0, 0);
 720                                 } else {
 721                                         lov_sublock_release(env, lck, i, 1, 0);
 722                                 }
 723                         }
 724                         lov_sublock_unlock(env, sub, closure, subenv);
 725                 }
 726                 result = lov_subresult(result, rc);
 727                 if (result < 0)
 728                         break;
 729         }
 730
 731         cl_lock_closure_fini(closure);
 732
 733         return;
 734 }
 735
 736 static int lov_lock_wait(const struct lu_env *env,
 737                          const struct cl_lock_slice *slice)
 738 {
 739         struct lov_lock        *lck     = cl2lov_lock(slice);
 740         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
 741         enum cl_lock_state      minstate;
 742         int                     result;
 743         int                     i;
 744
 745         ENTRY;
 746
 747         for (result = 0, minstate = CLS_FREEING, i = 0; i < lck->lls_nr; ++i) {
 748                 int rc;
 749                 struct lovsub_lock     *sub;
 750                 struct cl_lock         *sublock;
 751                 struct lov_lock_sub    *lls;
 752                 struct lov_sublock_env *subenv;
 753
 754                 lls = &lck->lls_sub[i];
 755                 sub = lls->sub_lock;
 756                 LASSERT(sub != NULL);
 757                 sublock = sub->lss_cl.cls_lock;
 758                 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
 759                 if (rc == 0) {
 760                         LASSERT(sublock->cll_state >= CLS_ENQUEUED);
 761                         if (sublock->cll_state < CLS_HELD)
 762                                 rc = cl_wait_try(env, sublock);
 763
 764                         minstate = min(minstate, sublock->cll_state);
 765                         lov_sublock_unlock(env, sub, closure, subenv);
 766                 }
 767                 result = lov_subresult(result, rc);
 768                 if (result != 0)
 769                         break;
 770         }
 771         cl_lock_closure_fini(closure);
 772         RETURN(result ?: minstate >= CLS_HELD ? 0 : CLO_WAIT);
 773 }
 774
 775 static int lov_lock_use(const struct lu_env *env,
 776                         const struct cl_lock_slice *slice)
 777 {
 778         struct lov_lock        *lck     = cl2lov_lock(slice);
 779         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
 780         int                     result;
 781         int                     i;
 782
 783         LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT);
 784         ENTRY;
 785
 786         for (result = 0, i = 0; i < lck->lls_nr; ++i) {
 787                 int rc;
 788                 struct lovsub_lock     *sub;
 789                 struct cl_lock         *sublock;
 790                 struct lov_lock_sub    *lls;
 791                 struct lov_sublock_env *subenv;
 792
 793                 LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT);
 794
 795                 lls = &lck->lls_sub[i];
 796                 sub = lls->sub_lock;
 797                 if (sub == NULL) {
 798                         /*
 799                          * Sub-lock might have been canceled, while top-lock was
 800                          * cached.
 801                          */
 802                         result = -ESTALE;
 803                         break;
 804                 }
 805
 806                 sublock = sub->lss_cl.cls_lock;
 807                 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
 808                 if (rc == 0) {
 809                         LASSERT(sublock->cll_state != CLS_FREEING);
 810                         lov_sublock_hold(env, lck, i);
 811                         if (sublock->cll_state == CLS_CACHED) {
 812                                 rc = cl_use_try(subenv->lse_env, sublock, 0);
 813                                 if (rc != 0)
 814                                         rc = lov_sublock_release(env, lck,
 815                                                                  i, 1, rc);
 816                         }
 817                         lov_sublock_unlock(env, sub, closure, subenv);
 818                 }
 819                 result = lov_subresult(result, rc);
 820                 if (result != 0)
 821                         break;
 822         }
 823
 824         if (lck->lls_cancel_race) {
 825                 /*
 826                  * If there is unlocking happened at the same time, then
 827                  * sublock_lock state should be FREEING, and lov_sublock_lock
 828                  * should return CLO_REPEAT. In this case, it should return
 829                  * ESTALE, and up layer should reset the lock state to be NEW.
 830                  */
 831                 lck->lls_cancel_race = 0;
 832                 LASSERT(result != 0);
 833                 result = -ESTALE;
 834         }
 835         cl_lock_closure_fini(closure);
 836         RETURN(result);
 837 }
 838
 839 #if 0
 840 static int lock_lock_multi_match()
 841 {
 842         struct cl_lock          *lock    = slice->cls_lock;
 843         struct cl_lock_descr    *subneed = &lov_env_info(env)->lti_ldescr;
 844         struct lov_object       *loo     = cl2lov(lov->lls_cl.cls_obj);
 845         struct lov_layout_raid0 *r0      = lov_r0(loo);
 846         struct lov_lock_sub     *sub;
 847         struct cl_object        *subobj;
 848         obd_off  fstart;
 849         obd_off  fend;
 850         obd_off  start;
 851         obd_off  end;
 852         int i;
 853
 854         fstart = cl_offset(need->cld_obj, need->cld_start);
 855         fend   = cl_offset(need->cld_obj, need->cld_end + 1) - 1;
 856         subneed->cld_mode = need->cld_mode;
 857         cl_lock_mutex_get(env, lock);
 858         for (i = 0; i < lov->lls_nr; ++i) {
 859                 sub = &lov->lls_sub[i];
 860                 if (sub->sub_lock == NULL)
 861                         continue;
 862                 subobj = sub->sub_descr.cld_obj;
 863                 if (!lov_stripe_intersects(r0->lo_lsm, sub->sub_stripe,
 864                                            fstart, fend, &start, &end))
 865                         continue;
 866                 subneed->cld_start = cl_index(subobj, start);
 867                 subneed->cld_end   = cl_index(subobj, end);
 868                 subneed->cld_obj   = subobj;
 869                 if (!cl_lock_ext_match(&sub->sub_got, subneed)) {
 870                         result = 0;
 871                         break;
 872                 }
 873         }
 874         cl_lock_mutex_put(env, lock);
 875 }
 876 #endif
 877
 878 /**
 879  * Check if the extent region \a descr is covered by \a child against the
 880  * specific \a stripe.
 881  */
 882 static int lov_lock_stripe_is_matching(const struct lu_env *env,
 883                                        struct lov_object *lov, int stripe,
 884                                        const struct cl_lock_descr *child,
 885                                        const struct cl_lock_descr *descr)
 886 {
 887         struct lov_stripe_md *lsm = lov_r0(lov)->lo_lsm;
 888         obd_off start;
 889         obd_off end;
 890         int result;
 891
 892         if (lov_r0(lov)->lo_nr == 1)
 893                 return cl_lock_ext_match(child, descr);
 894
 895         /*
 896          * For a multi-stripes object:
 897          * - make sure the descr only covers child's stripe, and
 898          * - check if extent is matching.
 899          */
 900         start = cl_offset(&lov->lo_cl, descr->cld_start);
 901         end   = cl_offset(&lov->lo_cl, descr->cld_end + 1) - 1;
 902         result = end - start <= lsm->lsm_stripe_size &&
 903                  stripe == lov_stripe_number(lsm, start) &&
 904                  stripe == lov_stripe_number(lsm, end);
 905         if (result) {
 906                 struct cl_lock_descr *subd = &lov_env_info(env)->lti_ldescr;
 907                 obd_off sub_start;
 908                 obd_off sub_end;
 909
 910                 subd->cld_obj  = NULL;   /* don't need sub object at all */
 911                 subd->cld_mode = descr->cld_mode;
 912                 subd->cld_gid  = descr->cld_gid;
 913                 result = lov_stripe_intersects(lsm, stripe, start, end,
 914                                                &sub_start, &sub_end);
 915                 LASSERT(result);
 916                 subd->cld_start = cl_index(child->cld_obj, sub_start);
 917                 subd->cld_end   = cl_index(child->cld_obj, sub_end);
 918                 result = cl_lock_ext_match(child, subd);
 919         }
 920         return result;
 921 }
 922
 923 /**
 924  * An implementation of cl_lock_operations::clo_fits_into() method.
 925  *
 926  * Checks whether a lock (given by \a slice) is suitable for \a
 927  * io. Multi-stripe locks can be used only for "quick" io, like truncate, or
 928  * O_APPEND write.
 929  *
 930  * \see ccc_lock_fits_into().
 931  */
 932 static int lov_lock_fits_into(const struct lu_env *env,
 933                               const struct cl_lock_slice *slice,
 934                               const struct cl_lock_descr *need,
 935                               const struct cl_io *io)
 936 {
 937         struct lov_lock   *lov = cl2lov_lock(slice);
 938         struct lov_object *obj = cl2lov(slice->cls_obj);
 939         int result;
 940
 941         LASSERT(cl_object_same(need->cld_obj, slice->cls_obj));
 942         LASSERT(lov->lls_nr > 0);
 943
 944         ENTRY;
 945
 946         if (need->cld_mode == CLM_GROUP)
 947                 /*
 948                  * always allow to match group lock.
 949                  */
 950                 result = cl_lock_ext_match(&lov->lls_orig, need);
 951         else if (lov->lls_nr == 1) {
 952                 struct cl_lock_descr *got = &lov->lls_sub[0].sub_got;
 953                 result = lov_lock_stripe_is_matching(env,
 954                                                      cl2lov(slice->cls_obj),
 955                                                      lov->lls_sub[0].sub_stripe,
 956                                                      got, need);
 957         } else if (io->ci_type != CIT_TRUNC && io->ci_type != CIT_MISC &&
 958                    !cl_io_is_append(io) && need->cld_mode != CLM_PHANTOM)
 959                 /*
 960                  * Multi-stripe locks are only suitable for `quick' IO and for
 961                  * glimpse.
 962                  */
 963                 result = 0;
 964         else
 965                 /*
 966                  * Most general case: multi-stripe existing lock, and
 967                  * (potentially) multi-stripe @need lock. Check that @need is
 968                  * covered by @lov's sub-locks.
 969                  *
 970                  * For now, ignore lock expansions made by the server, and
 971                  * match against original lock extent.
 972                  */
 973                 result = cl_lock_ext_match(&lov->lls_orig, need);
 974         CDEBUG(D_DLMTRACE, DDESCR"/"DDESCR" %i %i/%i: %i\n",
 975                PDESCR(&lov->lls_orig), PDESCR(&lov->lls_sub[0].sub_got),
 976                lov->lls_sub[0].sub_stripe, lov->lls_nr, lov_r0(obj)->lo_nr,
 977                result);
 978         RETURN(result);
 979 }
 980
 981 void lov_lock_unlink(const struct lu_env *env,
 982                      struct lov_lock_link *link, struct lovsub_lock *sub)
 983 {
 984         struct lov_lock *lck    = link->lll_super;
 985         struct cl_lock  *parent = lck->lls_cl.cls_lock;
 986
 987         LASSERT(cl_lock_is_mutexed(parent));
 988         LASSERT(cl_lock_is_mutexed(sub->lss_cl.cls_lock));
 989         ENTRY;
 990
 991         list_del_init(&link->lll_list);
 992         LASSERT(lck->lls_sub[link->lll_idx].sub_lock == sub);
 993         /* yank this sub-lock from parent's array */
 994         lck->lls_sub[link->lll_idx].sub_lock = NULL;
 995         LASSERT(lck->lls_nr_filled > 0);
 996         lck->lls_nr_filled--;
 997         lu_ref_del(&parent->cll_reference, "lov-child", sub->lss_cl.cls_lock);
 998         cl_lock_put(env, parent);
 999         OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
1000         EXIT;
1001 }
1002
1003 struct lov_lock_link *lov_lock_link_find(const struct lu_env *env,
1004                                          struct lov_lock *lck,
1005                                          struct lovsub_lock *sub)
1006 {
1007         struct lov_lock_link *scan;
1008
1009         LASSERT(cl_lock_is_mutexed(sub->lss_cl.cls_lock));
1010         ENTRY;
1011
1012         list_for_each_entry(scan, &sub->lss_parents, lll_list) {
1013                 if (scan->lll_super == lck)
1014                         RETURN(scan);
1015         }
1016         RETURN(NULL);
1017 }
1018
1019 /**
1020  * An implementation of cl_lock_operations::clo_delete() method. This is
1021  * invoked for "top-to-bottom" delete, when lock destruction starts from the
1022  * top-lock, e.g., as a result of inode destruction.
1023  *
1024  * Unlinks top-lock from all its sub-locks. Sub-locks are not deleted there:
1025  * this is done separately elsewhere:
1026  *
1027  *     - for inode destruction, lov_object_delete() calls cl_object_kill() for
1028  *       each sub-object, purging its locks;
1029  *
1030  *     - in other cases (e.g., a fatal error with a top-lock) sub-locks are
1031  *       left in the cache.
1032  */
1033 static void lov_lock_delete(const struct lu_env *env,
1034                             const struct cl_lock_slice *slice)
1035 {
1036         struct lov_lock        *lck     = cl2lov_lock(slice);
1037         struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
1038         int i;
1039
1040         LASSERT(slice->cls_lock->cll_state == CLS_FREEING);
1041         ENTRY;
1042
1043         for (i = 0; i < lck->lls_nr; ++i) {
1044                 struct lov_lock_sub *lls;
1045                 struct lovsub_lock  *lsl;
1046                 struct cl_lock      *sublock;
1047                 int rc;
1048
1049                 lls = &lck->lls_sub[i];
1050                 lsl = lls->sub_lock;
1051                 if (lsl == NULL)
1052                         continue;
1053
1054                 sublock = lsl->lss_cl.cls_lock;
1055                 rc = lov_sublock_lock(env, lck, lls, closure, NULL);
1056                 if (rc == 0) {
1057                         if (lls->sub_flags & LSF_HELD)
1058                                 lov_sublock_release(env, lck, i, 1, 0);
1059                         if (sublock->cll_state < CLS_FREEING) {
1060                                 struct lov_lock_link *link;
1061
1062                                 link = lov_lock_link_find(env, lck, lsl);
1063                                 LASSERT(link != NULL);
1064                                 lov_lock_unlink(env, link, lsl);
1065                                 LASSERT(lck->lls_sub[i].sub_lock == NULL);
1066                         }
1067                         lov_sublock_unlock(env, lsl, closure, NULL);
1068                 } else if (rc == CLO_REPEAT) {
1069                         --i; /* repeat with this lock */
1070                 } else {
1071                         CL_LOCK_DEBUG(D_ERROR, env, sublock,
1072                                       "Cannot get sub-lock for delete: %i\n",
1073                                       rc);
1074                 }
1075         }
1076         cl_lock_closure_fini(closure);
1077         EXIT;
1078 }
1079
1080 static int lov_lock_print(const struct lu_env *env, void *cookie,
1081                           lu_printer_t p, const struct cl_lock_slice *slice)
1082 {
1083         struct lov_lock *lck = cl2lov_lock(slice);
1084         int              i;
1085
1086         (*p)(env, cookie, "%d\n", lck->lls_nr);
1087         for (i = 0; i < lck->lls_nr; ++i) {
1088                 struct lov_lock_sub *sub;
1089
1090                 sub = &lck->lls_sub[i];
1091                 (*p)(env, cookie, "    %d %x: ", i, sub->sub_flags);
1092                 if (sub->sub_lock != NULL)
1093                         cl_lock_print(env, cookie, p,
1094                                       sub->sub_lock->lss_cl.cls_lock);
1095                 else
1096                         (*p)(env, cookie, "---\n");
1097         }
1098         return 0;
1099 }
1100
1101 static const struct cl_lock_operations lov_lock_ops = {
1102         .clo_fini      = lov_lock_fini,
1103         .clo_enqueue   = lov_lock_enqueue,
1104         .clo_wait      = lov_lock_wait,
1105         .clo_use       = lov_lock_use,
1106         .clo_unuse     = lov_lock_unuse,
1107         .clo_cancel    = lov_lock_cancel,
1108         .clo_fits_into = lov_lock_fits_into,
1109         .clo_delete    = lov_lock_delete,
1110         .clo_print     = lov_lock_print
1111 };
1112
1113 int lov_lock_init_raid0(const struct lu_env *env, struct cl_object *obj,
1114                         struct cl_lock *lock, const struct cl_io *io)
1115 {
1116         struct lov_lock *lck;
1117         int result;
1118
1119         ENTRY;
1120         OBD_SLAB_ALLOC_PTR_GFP(lck, lov_lock_kmem, CFS_ALLOC_IO);
1121         if (lck != NULL) {
1122                 cl_lock_slice_add(lock, &lck->lls_cl, obj, &lov_lock_ops);
1123                 result = lov_lock_sub_init(env, lck, io);
1124         } else
1125                 result = -ENOMEM;
1126         RETURN(result);
1127 }
1128
1129 static struct cl_lock_closure *lov_closure_get(const struct lu_env *env,
1130                                                struct cl_lock *parent)
1131 {
1132         struct cl_lock_closure *closure;
1133
1134         closure = &lov_env_info(env)->lti_closure;
1135         LASSERT(list_empty(&closure->clc_list));
1136         cl_lock_closure_init(env, closure, parent, 1);
1137         return closure;
1138 }
1139
1140
1141 /** @} lov */