lustre/osc/osc_lock.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, Intel Corporation.
  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 OSC layer.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_OSC
  42
  43 #ifdef __KERNEL__
  44 # include <libcfs/libcfs.h>
  45 #else
  46 # include <liblustre.h>
  47 #endif
  48 /* fid_build_reg_res_name() */
  49 #include <lustre_fid.h>
  50
  51 #include "osc_cl_internal.h"
  52
  53 /** \addtogroup osc
  54  *  @{
  55  */
  56
  57 #define _PAGEREF_MAGIC  (-10000000)
  58
  59 /*****************************************************************************
  60  *
  61  * Type conversions.
  62  *
  63  */
  64
  65 static const struct cl_lock_operations osc_lock_ops;
  66 static const struct cl_lock_operations osc_lock_lockless_ops;
  67 static void osc_lock_to_lockless(const struct lu_env *env,
  68                                  struct osc_lock *ols, int force);
  69 static int osc_lock_has_pages(struct osc_lock *olck);
  70
  71 int osc_lock_is_lockless(const struct osc_lock *olck)
  72 {
  73         return (olck->ols_cl.cls_ops == &osc_lock_lockless_ops);
  74 }
  75
  76 /**
  77  * Returns a weak pointer to the ldlm lock identified by a handle. Returned
  78  * pointer cannot be dereferenced, as lock is not protected from concurrent
  79  * reclaim. This function is a helper for osc_lock_invariant().
  80  */
  81 static struct ldlm_lock *osc_handle_ptr(struct lustre_handle *handle)
  82 {
  83         struct ldlm_lock *lock;
  84
  85         lock = ldlm_handle2lock(handle);
  86         if (lock != NULL)
  87                 LDLM_LOCK_PUT(lock);
  88         return lock;
  89 }
  90
  91 /**
  92  * Invariant that has to be true all of the time.
  93  */
  94 static int osc_lock_invariant(struct osc_lock *ols)
  95 {
  96         struct ldlm_lock *lock        = osc_handle_ptr(&ols->ols_handle);
  97         struct ldlm_lock *olock       = ols->ols_lock;
  98         int               handle_used = lustre_handle_is_used(&ols->ols_handle);
  99
 100         if (ergo(osc_lock_is_lockless(ols),
 101                  ols->ols_locklessable && ols->ols_lock == NULL))
 102                 return 1;
 103
 104         /*
 105          * If all the following "ergo"s are true, return 1, otherwise 0
 106          */
 107         if (! ergo(olock != NULL, handle_used))
 108                 return 0;
 109
 110         if (! ergo(olock != NULL,
 111                    olock->l_handle.h_cookie == ols->ols_handle.cookie))
 112                 return 0;
 113
 114         if (! ergo(handle_used,
 115                    ergo(lock != NULL && olock != NULL, lock == olock) &&
 116                    ergo(lock == NULL, olock == NULL)))
 117                 return 0;
 118         /*
 119          * Check that ->ols_handle and ->ols_lock are consistent, but
 120          * take into account that they are set at the different time.
 121          */
 122         if (! ergo(ols->ols_state == OLS_CANCELLED,
 123                    olock == NULL && !handle_used))
 124                 return 0;
 125         /*
 126          * DLM lock is destroyed only after we have seen cancellation
 127          * ast.
 128          */
 129         if (! ergo(olock != NULL && ols->ols_state < OLS_CANCELLED,
 130                    ((olock->l_flags & LDLM_FL_DESTROYED) == 0)))
 131                 return 0;
 132
 133         if (! ergo(ols->ols_state == OLS_GRANTED,
 134                    olock != NULL &&
 135                    olock->l_req_mode == olock->l_granted_mode &&
 136                    ols->ols_hold))
 137                 return 0;
 138         return 1;
 139 }
 140
 141 /*****************************************************************************
 142  *
 143  * Lock operations.
 144  *
 145  */
 146
 147 /**
 148  * Breaks a link between osc_lock and dlm_lock.
 149  */
 150 static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck)
 151 {
 152         struct ldlm_lock *dlmlock;
 153
 154         spin_lock(&osc_ast_guard);
 155         dlmlock = olck->ols_lock;
 156         if (dlmlock == NULL) {
 157                 spin_unlock(&osc_ast_guard);
 158                 return;
 159         }
 160
 161         olck->ols_lock = NULL;
 162         /* wb(); --- for all who checks (ols->ols_lock != NULL) before
 163          * call to osc_lock_detach() */
 164         dlmlock->l_ast_data = NULL;
 165         olck->ols_handle.cookie = 0ULL;
 166         spin_unlock(&osc_ast_guard);
 167
 168         lock_res_and_lock(dlmlock);
 169         if (dlmlock->l_granted_mode == dlmlock->l_req_mode) {
 170                 struct cl_object *obj = olck->ols_cl.cls_obj;
 171                 struct cl_attr *attr  = &osc_env_info(env)->oti_attr;
 172                 __u64 old_kms;
 173
 174                 cl_object_attr_lock(obj);
 175                 /* Must get the value under the lock to avoid possible races. */
 176                 old_kms = cl2osc(obj)->oo_oinfo->loi_kms;
 177                 /* Update the kms. Need to loop all granted locks.
 178                  * Not a problem for the client */
 179                 attr->cat_kms = ldlm_extent_shift_kms(dlmlock, old_kms);
 180
 181                 cl_object_attr_set(env, obj, attr, CAT_KMS);
 182                 cl_object_attr_unlock(obj);
 183         }
 184         unlock_res_and_lock(dlmlock);
 185
 186         /* release a reference taken in osc_lock_upcall0(). */
 187         LASSERT(olck->ols_has_ref);
 188         lu_ref_del(&dlmlock->l_reference, "osc_lock", olck);
 189         LDLM_LOCK_RELEASE(dlmlock);
 190         olck->ols_has_ref = 0;
 191 }
 192
 193 static int osc_lock_unhold(struct osc_lock *ols)
 194 {
 195         int result = 0;
 196
 197         if (ols->ols_hold) {
 198                 ols->ols_hold = 0;
 199                 result = osc_cancel_base(&ols->ols_handle,
 200                                          ols->ols_einfo.ei_mode);
 201         }
 202         return result;
 203 }
 204
 205 static int osc_lock_unuse(const struct lu_env *env,
 206                           const struct cl_lock_slice *slice)
 207 {
 208         struct osc_lock *ols = cl2osc_lock(slice);
 209
 210         LINVRNT(osc_lock_invariant(ols));
 211
 212         switch (ols->ols_state) {
 213         case OLS_NEW:
 214                 LASSERT(!ols->ols_hold);
 215                 LASSERT(ols->ols_agl);
 216                 return 0;
 217         case OLS_UPCALL_RECEIVED:
 218                 osc_lock_unhold(ols);
 219         case OLS_ENQUEUED:
 220                 LASSERT(!ols->ols_hold);
 221                 osc_lock_detach(env, ols);
 222                 ols->ols_state = OLS_NEW;
 223                 return 0;
 224         case OLS_GRANTED:
 225                 LASSERT(!ols->ols_glimpse);
 226                 LASSERT(ols->ols_hold);
 227                 /*
 228                  * Move lock into OLS_RELEASED state before calling
 229                  * osc_cancel_base() so that possible synchronous cancellation
 230                  * (that always happens e.g., for liblustre) sees that lock is
 231                  * released.
 232                  */
 233                 ols->ols_state = OLS_RELEASED;
 234                 return osc_lock_unhold(ols);
 235         default:
 236                 CERROR("Impossible state: %d\n", ols->ols_state);
 237                 LBUG();
 238         }
 239 }
 240
 241 static void osc_lock_fini(const struct lu_env *env,
 242                           struct cl_lock_slice *slice)
 243 {
 244         struct osc_lock  *ols = cl2osc_lock(slice);
 245
 246         LINVRNT(osc_lock_invariant(ols));
 247         /*
 248          * ->ols_hold can still be true at this point if, for example, a
 249          * thread that requested a lock was killed (and released a reference
 250          * to the lock), before reply from a server was received. In this case
 251          * lock is destroyed immediately after upcall.
 252          */
 253         osc_lock_unhold(ols);
 254         LASSERT(ols->ols_lock == NULL);
 255         LASSERT(cfs_atomic_read(&ols->ols_pageref) == 0 ||
 256                 cfs_atomic_read(&ols->ols_pageref) == _PAGEREF_MAGIC);
 257
 258         OBD_SLAB_FREE_PTR(ols, osc_lock_kmem);
 259 }
 260
 261 static void osc_lock_build_policy(const struct lu_env *env,
 262                                   const struct cl_lock *lock,
 263                                   ldlm_policy_data_t *policy)
 264 {
 265         const struct cl_lock_descr *d = &lock->cll_descr;
 266
 267         osc_index2policy(policy, d->cld_obj, d->cld_start, d->cld_end);
 268         policy->l_extent.gid = d->cld_gid;
 269 }
 270
 271 static __u64 osc_enq2ldlm_flags(__u32 enqflags)
 272 {
 273         __u64 result = 0;
 274
 275         LASSERT((enqflags & ~CEF_MASK) == 0);
 276
 277         if (enqflags & CEF_NONBLOCK)
 278                 result |= LDLM_FL_BLOCK_NOWAIT;
 279         if (enqflags & CEF_ASYNC)
 280                 result |= LDLM_FL_HAS_INTENT;
 281         if (enqflags & CEF_DISCARD_DATA)
 282                 result |= LDLM_FL_AST_DISCARD_DATA;
 283         return result;
 284 }
 285
 286 /**
 287  * Global spin-lock protecting consistency of ldlm_lock::l_ast_data
 288  * pointers. Initialized in osc_init().
 289  */
 290 spinlock_t osc_ast_guard;
 291
 292 static struct osc_lock *osc_ast_data_get(struct ldlm_lock *dlm_lock)
 293 {
 294         struct osc_lock *olck;
 295
 296         lock_res_and_lock(dlm_lock);
 297         spin_lock(&osc_ast_guard);
 298         olck = dlm_lock->l_ast_data;
 299         if (olck != NULL) {
 300                 struct cl_lock *lock = olck->ols_cl.cls_lock;
 301                 /*
 302                  * If osc_lock holds a reference on ldlm lock, return it even
 303                  * when cl_lock is in CLS_FREEING state. This way
 304                  *
 305                  *         osc_ast_data_get(dlmlock) == NULL
 306                  *
 307                  * guarantees that all osc references on dlmlock were
 308                  * released. osc_dlm_blocking_ast0() relies on that.
 309                  */
 310                 if (lock->cll_state < CLS_FREEING || olck->ols_has_ref) {
 311                         cl_lock_get_trust(lock);
 312                         lu_ref_add_atomic(&lock->cll_reference,
 313                                           "ast", cfs_current());
 314                 } else
 315                         olck = NULL;
 316         }
 317         spin_unlock(&osc_ast_guard);
 318         unlock_res_and_lock(dlm_lock);
 319         return olck;
 320 }
 321
 322 static void osc_ast_data_put(const struct lu_env *env, struct osc_lock *olck)
 323 {
 324         struct cl_lock *lock;
 325
 326         lock = olck->ols_cl.cls_lock;
 327         lu_ref_del(&lock->cll_reference, "ast", cfs_current());
 328         cl_lock_put(env, lock);
 329 }
 330
 331 /**
 332  * Updates object attributes from a lock value block (lvb) received together
 333  * with the DLM lock reply from the server. Copy of osc_update_enqueue()
 334  * logic.
 335  *
 336  * This can be optimized to not update attributes when lock is a result of a
 337  * local match.
 338  *
 339  * Called under lock and resource spin-locks.
 340  */
 341 static void osc_lock_lvb_update(const struct lu_env *env, struct osc_lock *olck,
 342                                 int rc)
 343 {
 344         struct ost_lvb    *lvb;
 345         struct cl_object  *obj;
 346         struct lov_oinfo  *oinfo;
 347         struct cl_attr    *attr;
 348         unsigned           valid;
 349
 350         ENTRY;
 351
 352         if (!(olck->ols_flags & LDLM_FL_LVB_READY))
 353                 RETURN_EXIT;
 354
 355         lvb   = &olck->ols_lvb;
 356         obj   = olck->ols_cl.cls_obj;
 357         oinfo = cl2osc(obj)->oo_oinfo;
 358         attr  = &osc_env_info(env)->oti_attr;
 359         valid = CAT_BLOCKS | CAT_ATIME | CAT_CTIME | CAT_MTIME | CAT_SIZE;
 360         cl_lvb2attr(attr, lvb);
 361
 362         cl_object_attr_lock(obj);
 363         if (rc == 0) {
 364                 struct ldlm_lock  *dlmlock;
 365                 __u64 size;
 366
 367                 dlmlock = olck->ols_lock;
 368                 LASSERT(dlmlock != NULL);
 369
 370                 /* re-grab LVB from a dlm lock under DLM spin-locks. */
 371                 *lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
 372                 size = lvb->lvb_size;
 373                 /* Extend KMS up to the end of this lock and no further
 374                  * A lock on [x,y] means a KMS of up to y + 1 bytes! */
 375                 if (size > dlmlock->l_policy_data.l_extent.end)
 376                         size = dlmlock->l_policy_data.l_extent.end + 1;
 377                 if (size >= oinfo->loi_kms) {
 378                         LDLM_DEBUG(dlmlock, "lock acquired, setting rss="LPU64
 379                                    ", kms="LPU64, lvb->lvb_size, size);
 380                         valid |= CAT_KMS;
 381                         attr->cat_kms = size;
 382                 } else {
 383                         LDLM_DEBUG(dlmlock, "lock acquired, setting rss="
 384                                    LPU64"; leaving kms="LPU64", end="LPU64,
 385                                    lvb->lvb_size, oinfo->loi_kms,
 386                                    dlmlock->l_policy_data.l_extent.end);
 387                 }
 388                 ldlm_lock_allow_match_locked(dlmlock);
 389         } else if (rc == -ENAVAIL && olck->ols_glimpse) {
 390                 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
 391                        " kms="LPU64"\n", lvb->lvb_size, oinfo->loi_kms);
 392         } else
 393                 valid = 0;
 394
 395         if (valid != 0)
 396                 cl_object_attr_set(env, obj, attr, valid);
 397
 398         cl_object_attr_unlock(obj);
 399
 400         EXIT;
 401 }
 402
 403 /**
 404  * Called when a lock is granted, from an upcall (when server returned a
 405  * granted lock), or from completion AST, when server returned a blocked lock.
 406  *
 407  * Called under lock and resource spin-locks, that are released temporarily
 408  * here.
 409  */
 410 static void osc_lock_granted(const struct lu_env *env, struct osc_lock *olck,
 411                              struct ldlm_lock *dlmlock, int rc)
 412 {
 413         struct ldlm_extent   *ext;
 414         struct cl_lock       *lock;
 415         struct cl_lock_descr *descr;
 416
 417         LASSERT(dlmlock->l_granted_mode == dlmlock->l_req_mode);
 418
 419         ENTRY;
 420         if (olck->ols_state < OLS_GRANTED) {
 421                 lock  = olck->ols_cl.cls_lock;
 422                 ext   = &dlmlock->l_policy_data.l_extent;
 423                 descr = &osc_env_info(env)->oti_descr;
 424                 descr->cld_obj = lock->cll_descr.cld_obj;
 425
 426                 /* XXX check that ->l_granted_mode is valid. */
 427                 descr->cld_mode  = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
 428                 descr->cld_start = cl_index(descr->cld_obj, ext->start);
 429                 descr->cld_end   = cl_index(descr->cld_obj, ext->end);
 430                 descr->cld_gid   = ext->gid;
 431                 /*
 432                  * tell upper layers the extent of the lock that was actually
 433                  * granted
 434                  */
 435                 olck->ols_state = OLS_GRANTED;
 436                 osc_lock_lvb_update(env, olck, rc);
 437
 438                 /* release DLM spin-locks to allow cl_lock_{modify,signal}()
 439                  * to take a semaphore on a parent lock. This is safe, because
 440                  * spin-locks are needed to protect consistency of
 441                  * dlmlock->l_*_mode and LVB, and we have finished processing
 442                  * them. */
 443                 unlock_res_and_lock(dlmlock);
 444                 cl_lock_modify(env, lock, descr);
 445                 cl_lock_signal(env, lock);
 446                 LINVRNT(osc_lock_invariant(olck));
 447                 lock_res_and_lock(dlmlock);
 448         }
 449         EXIT;
 450 }
 451
 452 static void osc_lock_upcall0(const struct lu_env *env, struct osc_lock *olck)
 453
 454 {
 455         struct ldlm_lock *dlmlock;
 456
 457         ENTRY;
 458
 459         dlmlock = ldlm_handle2lock_long(&olck->ols_handle, 0);
 460         LASSERT(dlmlock != NULL);
 461
 462         lock_res_and_lock(dlmlock);
 463         spin_lock(&osc_ast_guard);
 464         LASSERT(dlmlock->l_ast_data == olck);
 465         LASSERT(olck->ols_lock == NULL);
 466         olck->ols_lock = dlmlock;
 467         spin_unlock(&osc_ast_guard);
 468
 469         /*
 470          * Lock might be not yet granted. In this case, completion ast
 471          * (osc_ldlm_completion_ast()) comes later and finishes lock
 472          * granting.
 473          */
 474         if (dlmlock->l_granted_mode == dlmlock->l_req_mode)
 475                 osc_lock_granted(env, olck, dlmlock, 0);
 476         unlock_res_and_lock(dlmlock);
 477
 478         /*
 479          * osc_enqueue_interpret() decrefs asynchronous locks, counter
 480          * this.
 481          */
 482         ldlm_lock_addref(&olck->ols_handle, olck->ols_einfo.ei_mode);
 483         olck->ols_hold = 1;
 484
 485         /* lock reference taken by ldlm_handle2lock_long() is owned by
 486          * osc_lock and released in osc_lock_detach() */
 487         lu_ref_add(&dlmlock->l_reference, "osc_lock", olck);
 488         olck->ols_has_ref = 1;
 489 }
 490
 491 /**
 492  * Lock upcall function that is executed either when a reply to ENQUEUE rpc is
 493  * received from a server, or after osc_enqueue_base() matched a local DLM
 494  * lock.
 495  */
 496 static int osc_lock_upcall(void *cookie, int errcode)
 497 {
 498         struct osc_lock         *olck  = cookie;
 499         struct cl_lock_slice    *slice = &olck->ols_cl;
 500         struct cl_lock          *lock  = slice->cls_lock;
 501         struct lu_env           *env;
 502         struct cl_env_nest       nest;
 503
 504         ENTRY;
 505         env = cl_env_nested_get(&nest);
 506         if (!IS_ERR(env)) {
 507                 int rc;
 508
 509                 cl_lock_mutex_get(env, lock);
 510
 511                 LASSERT(lock->cll_state >= CLS_QUEUING);
 512                 if (olck->ols_state == OLS_ENQUEUED) {
 513                         olck->ols_state = OLS_UPCALL_RECEIVED;
 514                         rc = ldlm_error2errno(errcode);
 515                 } else if (olck->ols_state == OLS_CANCELLED) {
 516                         rc = -EIO;
 517                 } else {
 518                         CERROR("Impossible state: %d\n", olck->ols_state);
 519                         LBUG();
 520                 }
 521                 if (rc) {
 522                         struct ldlm_lock *dlmlock;
 523
 524                         dlmlock = ldlm_handle2lock(&olck->ols_handle);
 525                         if (dlmlock != NULL) {
 526                                 lock_res_and_lock(dlmlock);
 527                                 spin_lock(&osc_ast_guard);
 528                                 LASSERT(olck->ols_lock == NULL);
 529                                 dlmlock->l_ast_data = NULL;
 530                                 olck->ols_handle.cookie = 0ULL;
 531                                 spin_unlock(&osc_ast_guard);
 532                                 ldlm_lock_fail_match_locked(dlmlock);
 533                                 unlock_res_and_lock(dlmlock);
 534                                 LDLM_LOCK_PUT(dlmlock);
 535                         }
 536                 } else {
 537                         if (olck->ols_glimpse)
 538                                 olck->ols_glimpse = 0;
 539                         osc_lock_upcall0(env, olck);
 540                 }
 541
 542                 /* Error handling, some errors are tolerable. */
 543                 if (olck->ols_locklessable && rc == -EUSERS) {
 544                         /* This is a tolerable error, turn this lock into
 545                          * lockless lock.
 546                          */
 547                         osc_object_set_contended(cl2osc(slice->cls_obj));
 548                         LASSERT(slice->cls_ops == &osc_lock_ops);
 549
 550                         /* Change this lock to ldlmlock-less lock. */
 551                         osc_lock_to_lockless(env, olck, 1);
 552                         olck->ols_state = OLS_GRANTED;
 553                         rc = 0;
 554                 } else if (olck->ols_glimpse && rc == -ENAVAIL) {
 555                         osc_lock_lvb_update(env, olck, rc);
 556                         cl_lock_delete(env, lock);
 557                         /* Hide the error. */
 558                         rc = 0;
 559                 }
 560
 561                 if (rc == 0) {
 562                         /* For AGL case, the RPC sponsor may exits the cl_lock
 563                         *  processing without wait() called before related OSC
 564                         *  lock upcall(). So update the lock status according
 565                         *  to the enqueue result inside AGL upcall(). */
 566                         if (olck->ols_agl) {
 567                                 lock->cll_flags |= CLF_FROM_UPCALL;
 568                                 cl_wait_try(env, lock);
 569                                 lock->cll_flags &= ~CLF_FROM_UPCALL;
 570                                 if (!olck->ols_glimpse)
 571                                         olck->ols_agl = 0;
 572                         }
 573                         cl_lock_signal(env, lock);
 574                         /* del user for lock upcall cookie */
 575                         cl_unuse_try(env, lock);
 576                 } else {
 577                         /* del user for lock upcall cookie */
 578                         cl_lock_user_del(env, lock);
 579                         cl_lock_error(env, lock, rc);
 580                 }
 581
 582                 /* release cookie reference, acquired by osc_lock_enqueue() */
 583                 cl_lock_hold_release(env, lock, "upcall", lock);
 584                 cl_lock_mutex_put(env, lock);
 585
 586                 lu_ref_del(&lock->cll_reference, "upcall", lock);
 587                 /* This maybe the last reference, so must be called after
 588                  * cl_lock_mutex_put(). */
 589                 cl_lock_put(env, lock);
 590
 591                 cl_env_nested_put(&nest, env);
 592         } else {
 593                 /* should never happen, similar to osc_ldlm_blocking_ast(). */
 594                 LBUG();
 595         }
 596         RETURN(errcode);
 597 }
 598
 599 /**
 600  * Core of osc_dlm_blocking_ast() logic.
 601  */
 602 static void osc_lock_blocking(const struct lu_env *env,
 603                               struct ldlm_lock *dlmlock,
 604                               struct osc_lock *olck, int blocking)
 605 {
 606         struct cl_lock *lock = olck->ols_cl.cls_lock;
 607
 608         LASSERT(olck->ols_lock == dlmlock);
 609         CLASSERT(OLS_BLOCKED < OLS_CANCELLED);
 610         LASSERT(!osc_lock_is_lockless(olck));
 611
 612         /*
 613          * Lock might be still addref-ed here, if e.g., blocking ast
 614          * is sent for a failed lock.
 615          */
 616         osc_lock_unhold(olck);
 617
 618         if (blocking && olck->ols_state < OLS_BLOCKED)
 619                 /*
 620                  * Move osc_lock into OLS_BLOCKED before canceling the lock,
 621                  * because it recursively re-enters osc_lock_blocking(), with
 622                  * the state set to OLS_CANCELLED.
 623                  */
 624                 olck->ols_state = OLS_BLOCKED;
 625         /*
 626          * cancel and destroy lock at least once no matter how blocking ast is
 627          * entered (see comment above osc_ldlm_blocking_ast() for use
 628          * cases). cl_lock_cancel() and cl_lock_delete() are idempotent.
 629          */
 630         cl_lock_cancel(env, lock);
 631         cl_lock_delete(env, lock);
 632 }
 633
 634 /**
 635  * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
 636  * and ldlm_lock caches.
 637  */
 638 static int osc_dlm_blocking_ast0(const struct lu_env *env,
 639                                  struct ldlm_lock *dlmlock,
 640                                  void *data, int flag)
 641 {
 642         struct osc_lock *olck;
 643         struct cl_lock  *lock;
 644         int result;
 645         int cancel;
 646
 647         LASSERT(flag == LDLM_CB_BLOCKING || flag == LDLM_CB_CANCELING);
 648
 649         cancel = 0;
 650         olck = osc_ast_data_get(dlmlock);
 651         if (olck != NULL) {
 652                 lock = olck->ols_cl.cls_lock;
 653                 cl_lock_mutex_get(env, lock);
 654                 LINVRNT(osc_lock_invariant(olck));
 655                 if (olck->ols_ast_wait) {
 656                         /* wake up osc_lock_use() */
 657                         cl_lock_signal(env, lock);
 658                         olck->ols_ast_wait = 0;
 659                 }
 660                 /*
 661                  * Lock might have been canceled while this thread was
 662                  * sleeping for lock mutex, but olck is pinned in memory.
 663                  */
 664                 if (olck == dlmlock->l_ast_data) {
 665                         /*
 666                          * NOTE: DLM sends blocking AST's for failed locks
 667                          *       (that are still in pre-OLS_GRANTED state)
 668                          *       too, and they have to be canceled otherwise
 669                          *       DLM lock is never destroyed and stuck in
 670                          *       the memory.
 671                          *
 672                          *       Alternatively, ldlm_cli_cancel() can be
 673                          *       called here directly for osc_locks with
 674                          *       ols_state < OLS_GRANTED to maintain an
 675                          *       invariant that ->clo_cancel() is only called
 676                          *       for locks that were granted.
 677                          */
 678                         LASSERT(data == olck);
 679                         osc_lock_blocking(env, dlmlock,
 680                                           olck, flag == LDLM_CB_BLOCKING);
 681                 } else
 682                         cancel = 1;
 683                 cl_lock_mutex_put(env, lock);
 684                 osc_ast_data_put(env, olck);
 685         } else
 686                 /*
 687                  * DLM lock exists, but there is no cl_lock attached to it.
 688                  * This is a `normal' race. cl_object and its cl_lock's can be
 689                  * removed by memory pressure, together with all pages.
 690                  */
 691                 cancel = (flag == LDLM_CB_BLOCKING);
 692
 693         if (cancel) {
 694                 struct lustre_handle *lockh;
 695
 696                 lockh = &osc_env_info(env)->oti_handle;
 697                 ldlm_lock2handle(dlmlock, lockh);
 698                 result = ldlm_cli_cancel(lockh, LCF_ASYNC);
 699         } else
 700                 result = 0;
 701         return result;
 702 }
 703
 704 /**
 705  * Blocking ast invoked by ldlm when dlm lock is either blocking progress of
 706  * some other lock, or is canceled. This function is installed as a
 707  * ldlm_lock::l_blocking_ast() for client extent locks.
 708  *
 709  * Control flow is tricky, because ldlm uses the same call-back
 710  * (ldlm_lock::l_blocking_ast()) for both blocking and cancellation ast's.
 711  *
 712  * \param dlmlock lock for which ast occurred.
 713  *
 714  * \param new description of a conflicting lock in case of blocking ast.
 715  *
 716  * \param data value of dlmlock->l_ast_data
 717  *
 718  * \param flag LDLM_CB_BLOCKING or LDLM_CB_CANCELING. Used to distinguish
 719  *             cancellation and blocking ast's.
 720  *
 721  * Possible use cases:
 722  *
 723  *     - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) to cancel
 724  *       lock due to lock lru pressure, or explicit user request to purge
 725  *       locks.
 726  *
 727  *     - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_BLOCKING) to notify
 728  *       us that dlmlock conflicts with another lock that some client is
 729  *       enqueing. Lock is canceled.
 730  *
 731  *           - cl_lock_cancel() is called. osc_lock_cancel() calls
 732  *             ldlm_cli_cancel() that calls
 733  *
 734  *                  dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
 735  *
 736  *             recursively entering osc_ldlm_blocking_ast().
 737  *
 738  *     - client cancels lock voluntary (e.g., as a part of early cancellation):
 739  *
 740  *           cl_lock_cancel()->
 741  *             osc_lock_cancel()->
 742  *               ldlm_cli_cancel()->
 743  *                 dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
 744  *
 745  */
 746 static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
 747                                  struct ldlm_lock_desc *new, void *data,
 748                                  int flag)
 749 {
 750         struct lu_env     *env;
 751         struct cl_env_nest nest;
 752         int                result;
 753
 754         /*
 755          * This can be called in the context of outer IO, e.g.,
 756          *
 757          *     cl_enqueue()->...
 758          *       ->osc_enqueue_base()->...
 759          *         ->ldlm_prep_elc_req()->...
 760          *           ->ldlm_cancel_callback()->...
 761          *             ->osc_ldlm_blocking_ast()
 762          *
 763          * new environment has to be created to not corrupt outer context.
 764          */
 765         env = cl_env_nested_get(&nest);
 766         if (!IS_ERR(env)) {
 767                 result = osc_dlm_blocking_ast0(env, dlmlock, data, flag);
 768                 cl_env_nested_put(&nest, env);
 769         } else {
 770                 result = PTR_ERR(env);
 771                 /*
 772                  * XXX This should never happen, as cl_lock is
 773                  * stuck. Pre-allocated environment a la vvp_inode_fini_env
 774                  * should be used.
 775                  */
 776                 LBUG();
 777         }
 778         if (result != 0) {
 779                 if (result == -ENODATA)
 780                         result = 0;
 781                 else
 782                         CERROR("BAST failed: %d\n", result);
 783         }
 784         return result;
 785 }
 786
 787 static int osc_ldlm_completion_ast(struct ldlm_lock *dlmlock,
 788                                    __u64 flags, void *data)
 789 {
 790         struct cl_env_nest nest;
 791         struct lu_env     *env;
 792         struct osc_lock   *olck;
 793         struct cl_lock    *lock;
 794         int result;
 795         int dlmrc;
 796
 797         /* first, do dlm part of the work */
 798         dlmrc = ldlm_completion_ast_async(dlmlock, flags, data);
 799         /* then, notify cl_lock */
 800         env = cl_env_nested_get(&nest);
 801         if (!IS_ERR(env)) {
 802                 olck = osc_ast_data_get(dlmlock);
 803                 if (olck != NULL) {
 804                         lock = olck->ols_cl.cls_lock;
 805                         cl_lock_mutex_get(env, lock);
 806                         /*
 807                          * ldlm_handle_cp_callback() copied LVB from request
 808                          * to lock->l_lvb_data, store it in osc_lock.
 809                          */
 810                         LASSERT(dlmlock->l_lvb_data != NULL);
 811                         lock_res_and_lock(dlmlock);
 812                         olck->ols_lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
 813                         if (olck->ols_lock == NULL) {
 814                                 /*
 815                                  * upcall (osc_lock_upcall()) hasn't yet been
 816                                  * called. Do nothing now, upcall will bind
 817                                  * olck to dlmlock and signal the waiters.
 818                                  *
 819                                  * This maintains an invariant that osc_lock
 820                                  * and ldlm_lock are always bound when
 821                                  * osc_lock is in OLS_GRANTED state.
 822                                  */
 823                         } else if (dlmlock->l_granted_mode ==
 824                                    dlmlock->l_req_mode) {
 825                                 osc_lock_granted(env, olck, dlmlock, dlmrc);
 826                         }
 827                         unlock_res_and_lock(dlmlock);
 828
 829                         if (dlmrc != 0) {
 830                                 CL_LOCK_DEBUG(D_ERROR, env, lock,
 831                                               "dlmlock returned %d\n", dlmrc);
 832                                 cl_lock_error(env, lock, dlmrc);
 833                         }
 834                         cl_lock_mutex_put(env, lock);
 835                         osc_ast_data_put(env, olck);
 836                         result = 0;
 837                 } else
 838                         result = -ELDLM_NO_LOCK_DATA;
 839                 cl_env_nested_put(&nest, env);
 840         } else
 841                 result = PTR_ERR(env);
 842         return dlmrc ?: result;
 843 }
 844
 845 static int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
 846 {
 847         struct ptlrpc_request  *req  = data;
 848         struct osc_lock        *olck;
 849         struct cl_lock         *lock;
 850         struct cl_object       *obj;
 851         struct cl_env_nest      nest;
 852         struct lu_env          *env;
 853         struct ost_lvb         *lvb;
 854         struct req_capsule     *cap;
 855         int                     result;
 856
 857         LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK);
 858
 859         env = cl_env_nested_get(&nest);
 860         if (!IS_ERR(env)) {
 861                 /* osc_ast_data_get() has to go after environment is
 862                  * allocated, because osc_ast_data() acquires a
 863                  * reference to a lock, and it can only be released in
 864                  * environment.
 865                  */
 866                 olck = osc_ast_data_get(dlmlock);
 867                 if (olck != NULL) {
 868                         lock = olck->ols_cl.cls_lock;
 869                         /* Do not grab the mutex of cl_lock for glimpse.
 870                          * See LU-1274 for details.
 871                          * BTW, it's okay for cl_lock to be cancelled during
 872                          * this period because server can handle this race.
 873                          * See ldlm_server_glimpse_ast() for details.
 874                          * cl_lock_mutex_get(env, lock); */
 875                         cap = &req->rq_pill;
 876                         req_capsule_extend(cap, &RQF_LDLM_GL_CALLBACK);
 877                         req_capsule_set_size(cap, &RMF_DLM_LVB, RCL_SERVER,
 878                                              sizeof *lvb);
 879                         result = req_capsule_server_pack(cap);
 880                         if (result == 0) {
 881                                 lvb = req_capsule_server_get(cap, &RMF_DLM_LVB);
 882                                 obj = lock->cll_descr.cld_obj;
 883                                 result = cl_object_glimpse(env, obj, lvb);
 884                         }
 885                         if (!exp_connect_lvb_type(req->rq_export))
 886                                 req_capsule_shrink(&req->rq_pill,
 887                                                    &RMF_DLM_LVB,
 888                                                    sizeof(struct ost_lvb_v1),
 889                                                    RCL_SERVER);
 890                         osc_ast_data_put(env, olck);
 891                 } else {
 892                         /*
 893                          * These errors are normal races, so we don't want to
 894                          * fill the console with messages by calling
 895                          * ptlrpc_error()
 896                          */
 897                         lustre_pack_reply(req, 1, NULL, NULL);
 898                         result = -ELDLM_NO_LOCK_DATA;
 899                 }
 900                 cl_env_nested_put(&nest, env);
 901         } else
 902                 result = PTR_ERR(env);
 903         req->rq_status = result;
 904         return result;
 905 }
 906
 907 static unsigned long osc_lock_weigh(const struct lu_env *env,
 908                                     const struct cl_lock_slice *slice)
 909 {
 910         /*
 911          * don't need to grab coh_page_guard since we don't care the exact #
 912          * of pages..
 913          */
 914         return cl_object_header(slice->cls_obj)->coh_pages;
 915 }
 916
 917 static void osc_lock_build_einfo(const struct lu_env *env,
 918                                  const struct cl_lock *clock,
 919                                  struct osc_lock *lock,
 920                                  struct ldlm_enqueue_info *einfo)
 921 {
 922         enum cl_lock_mode mode;
 923
 924         mode = clock->cll_descr.cld_mode;
 925         if (mode == CLM_PHANTOM)
 926                 /*
 927                  * For now, enqueue all glimpse locks in read mode. In the
 928                  * future, client might choose to enqueue LCK_PW lock for
 929                  * glimpse on a file opened for write.
 930                  */
 931                 mode = CLM_READ;
 932
 933         einfo->ei_type   = LDLM_EXTENT;
 934         einfo->ei_mode   = osc_cl_lock2ldlm(mode);
 935         einfo->ei_cb_bl  = osc_ldlm_blocking_ast;
 936         einfo->ei_cb_cp  = osc_ldlm_completion_ast;
 937         einfo->ei_cb_gl  = osc_ldlm_glimpse_ast;
 938         einfo->ei_cbdata = lock; /* value to be put into ->l_ast_data */
 939 }
 940
 941 /**
 942  * Determine if the lock should be converted into a lockless lock.
 943  *
 944  * Steps to check:
 945  * - if the lock has an explicite requirment for a non-lockless lock;
 946  * - if the io lock request type ci_lockreq;
 947  * - send the enqueue rpc to ost to make the further decision;
 948  * - special treat to truncate lockless lock
 949  *
 950  *  Additional policy can be implemented here, e.g., never do lockless-io
 951  *  for large extents.
 952  */
 953 static void osc_lock_to_lockless(const struct lu_env *env,
 954                                  struct osc_lock *ols, int force)
 955 {
 956         struct cl_lock_slice *slice = &ols->ols_cl;
 957
 958         LASSERT(ols->ols_state == OLS_NEW ||
 959                 ols->ols_state == OLS_UPCALL_RECEIVED);
 960
 961         if (force) {
 962                 ols->ols_locklessable = 1;
 963                 slice->cls_ops = &osc_lock_lockless_ops;
 964         } else {
 965                 struct osc_io *oio     = osc_env_io(env);
 966                 struct cl_io  *io      = oio->oi_cl.cis_io;
 967                 struct cl_object *obj  = slice->cls_obj;
 968                 struct osc_object *oob = cl2osc(obj);
 969                 const struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
 970                 struct obd_connect_data *ocd;
 971
 972                 LASSERT(io->ci_lockreq == CILR_MANDATORY ||
 973                         io->ci_lockreq == CILR_MAYBE ||
 974                         io->ci_lockreq == CILR_NEVER);
 975
 976                 ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data;
 977                 ols->ols_locklessable = (io->ci_type != CIT_SETATTR) &&
 978                                 (io->ci_lockreq == CILR_MAYBE) &&
 979                                 (ocd->ocd_connect_flags & OBD_CONNECT_SRVLOCK);
 980                 if (io->ci_lockreq == CILR_NEVER ||
 981                         /* lockless IO */
 982                     (ols->ols_locklessable && osc_object_is_contended(oob)) ||
 983                         /* lockless truncate */
 984                     (cl_io_is_trunc(io) &&
 985                      (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK) &&
 986                       osd->od_lockless_truncate)) {
 987                         ols->ols_locklessable = 1;
 988                         slice->cls_ops = &osc_lock_lockless_ops;
 989                 }
 990         }
 991         LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
 992 }
 993
 994 static int osc_lock_compatible(const struct osc_lock *qing,
 995                                const struct osc_lock *qed)
 996 {
 997         enum cl_lock_mode qing_mode;
 998         enum cl_lock_mode qed_mode;
 999
1000         qing_mode = qing->ols_cl.cls_lock->cll_descr.cld_mode;
1001         if (qed->ols_glimpse &&
1002             (qed->ols_state >= OLS_UPCALL_RECEIVED || qing_mode == CLM_READ))
1003                 return 1;
1004
1005         qed_mode = qed->ols_cl.cls_lock->cll_descr.cld_mode;
1006         return ((qing_mode == CLM_READ) && (qed_mode == CLM_READ));
1007 }
1008
1009 /**
1010  * Cancel all conflicting locks and wait for them to be destroyed.
1011  *
1012  * This function is used for two purposes:
1013  *
1014  *     - early cancel all conflicting locks before starting IO, and
1015  *
1016  *     - guarantee that pages added to the page cache by lockless IO are never
1017  *       covered by locks other than lockless IO lock, and, hence, are not
1018  *       visible to other threads.
1019  */
1020 static int osc_lock_enqueue_wait(const struct lu_env *env,
1021                                  const struct osc_lock *olck)
1022 {
1023         struct cl_lock          *lock    = olck->ols_cl.cls_lock;
1024         struct cl_lock_descr    *descr   = &lock->cll_descr;
1025         struct cl_object_header *hdr     = cl_object_header(descr->cld_obj);
1026         struct cl_lock          *scan;
1027         struct cl_lock          *conflict= NULL;
1028         int lockless                     = osc_lock_is_lockless(olck);
1029         int rc                           = 0;
1030         ENTRY;
1031
1032         LASSERT(cl_lock_is_mutexed(lock));
1033
1034         /* make it enqueue anyway for glimpse lock, because we actually
1035          * don't need to cancel any conflicting locks. */
1036         if (olck->ols_glimpse)
1037                 return 0;
1038
1039         spin_lock(&hdr->coh_lock_guard);
1040         cfs_list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) {
1041                 struct cl_lock_descr *cld = &scan->cll_descr;
1042                 const struct osc_lock *scan_ols;
1043
1044                 if (scan == lock)
1045                         break;
1046
1047                 if (scan->cll_state < CLS_QUEUING ||
1048                     scan->cll_state == CLS_FREEING ||
1049                     cld->cld_start > descr->cld_end ||
1050                     cld->cld_end < descr->cld_start)
1051                         continue;
1052
1053                 /* overlapped and living locks. */
1054
1055                 /* We're not supposed to give up group lock. */
1056                 if (scan->cll_descr.cld_mode == CLM_GROUP) {
1057                         LASSERT(descr->cld_mode != CLM_GROUP ||
1058                                 descr->cld_gid != scan->cll_descr.cld_gid);
1059                         continue;
1060                 }
1061
1062                 scan_ols = osc_lock_at(scan);
1063
1064                 /* We need to cancel the compatible locks if we're enqueuing
1065                  * a lockless lock, for example:
1066                  * imagine that client has PR lock on [0, 1000], and thread T0
1067                  * is doing lockless IO in [500, 1500] region. Concurrent
1068                  * thread T1 can see lockless data in [500, 1000], which is
1069                  * wrong, because these data are possibly stale. */
1070                 if (!lockless && osc_lock_compatible(olck, scan_ols))
1071                         continue;
1072
1073                 cl_lock_get_trust(scan);
1074                 conflict = scan;
1075                 break;
1076         }
1077         spin_unlock(&hdr->coh_lock_guard);
1078
1079         if (conflict) {
1080                 if (lock->cll_descr.cld_mode == CLM_GROUP) {
1081                         /* we want a group lock but a previous lock request
1082                          * conflicts, we do not wait but return 0 so the
1083                          * request is send to the server
1084                          */
1085                         CDEBUG(D_DLMTRACE, "group lock %p is conflicted "
1086                                            "with %p, no wait, send to server\n",
1087                                lock, conflict);
1088                         cl_lock_put(env, conflict);
1089                         rc = 0;
1090                 } else {
1091                         CDEBUG(D_DLMTRACE, "lock %p is conflicted with %p, "
1092                                            "will wait\n",
1093                                lock, conflict);
1094                         LASSERT(lock->cll_conflict == NULL);
1095                         lu_ref_add(&conflict->cll_reference, "cancel-wait",
1096                                    lock);
1097                         lock->cll_conflict = conflict;
1098                         rc = CLO_WAIT;
1099                 }
1100         }
1101         RETURN(rc);
1102 }
1103
1104 /**
1105  * Implementation of cl_lock_operations::clo_enqueue() method for osc
1106  * layer. This initiates ldlm enqueue:
1107  *
1108  *     - cancels conflicting locks early (osc_lock_enqueue_wait());
1109  *
1110  *     - calls osc_enqueue_base() to do actual enqueue.
1111  *
1112  * osc_enqueue_base() is supplied with an upcall function that is executed
1113  * when lock is received either after a local cached ldlm lock is matched, or
1114  * when a reply from the server is received.
1115  *
1116  * This function does not wait for the network communication to complete.
1117  */
1118 static int osc_lock_enqueue(const struct lu_env *env,
1119                             const struct cl_lock_slice *slice,
1120                             struct cl_io *unused, __u32 enqflags)
1121 {
1122         struct osc_lock          *ols     = cl2osc_lock(slice);
1123         struct cl_lock           *lock    = ols->ols_cl.cls_lock;
1124         int result;
1125         ENTRY;
1126
1127         LASSERT(cl_lock_is_mutexed(lock));
1128         LASSERTF(ols->ols_state == OLS_NEW,
1129                  "Impossible state: %d\n", ols->ols_state);
1130
1131         LASSERTF(ergo(ols->ols_glimpse, lock->cll_descr.cld_mode <= CLM_READ),
1132                 "lock = %p, ols = %p\n", lock, ols);
1133
1134         result = osc_lock_enqueue_wait(env, ols);
1135         if (result == 0) {
1136                 if (!osc_lock_is_lockless(ols)) {
1137                         struct osc_object        *obj = cl2osc(slice->cls_obj);
1138                         struct osc_thread_info   *info = osc_env_info(env);
1139                         struct ldlm_res_id       *resname = &info->oti_resname;
1140                         ldlm_policy_data_t       *policy = &info->oti_policy;
1141                         struct ldlm_enqueue_info *einfo = &ols->ols_einfo;
1142
1143                         /* lock will be passed as upcall cookie,
1144                          * hold ref to prevent to be released. */
1145                         cl_lock_hold_add(env, lock, "upcall", lock);
1146                         /* a user for lock also */
1147                         cl_lock_user_add(env, lock);
1148                         ols->ols_state = OLS_ENQUEUED;
1149
1150                         /*
1151                          * XXX: this is possible blocking point as
1152                          * ldlm_lock_match(LDLM_FL_LVB_READY) waits for
1153                          * LDLM_CP_CALLBACK.
1154                          */
1155                         ostid_build_res_name(&obj->oo_oinfo->loi_oi, resname);
1156                         osc_lock_build_policy(env, lock, policy);
1157                         result = osc_enqueue_base(osc_export(obj), resname,
1158                                           &ols->ols_flags, policy,
1159                                           &ols->ols_lvb,
1160                                           obj->oo_oinfo->loi_kms_valid,
1161                                           osc_lock_upcall,
1162                                           ols, einfo, &ols->ols_handle,
1163                                           PTLRPCD_SET, 1, ols->ols_agl);
1164                         if (result != 0) {
1165                                 cl_lock_user_del(env, lock);
1166                                 cl_lock_unhold(env, lock, "upcall", lock);
1167                                 if (unlikely(result == -ECANCELED)) {
1168                                         ols->ols_state = OLS_NEW;
1169                                         result = 0;
1170                                 }
1171                         }
1172                 } else {
1173                         ols->ols_state = OLS_GRANTED;
1174                         ols->ols_owner = osc_env_io(env);
1175                 }
1176         }
1177         LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
1178         RETURN(result);
1179 }
1180
1181 static int osc_lock_wait(const struct lu_env *env,
1182                          const struct cl_lock_slice *slice)
1183 {
1184         struct osc_lock *olck = cl2osc_lock(slice);
1185         struct cl_lock  *lock = olck->ols_cl.cls_lock;
1186
1187         LINVRNT(osc_lock_invariant(olck));
1188
1189         if (olck->ols_glimpse && olck->ols_state >= OLS_UPCALL_RECEIVED) {
1190                 if (olck->ols_flags & LDLM_FL_LVB_READY) {
1191                         return 0;
1192                 } else if (olck->ols_agl) {
1193                         if (lock->cll_flags & CLF_FROM_UPCALL)
1194                                 /* It is from enqueue RPC reply upcall for
1195                                  * updating state. Do not re-enqueue. */
1196                                 return -ENAVAIL;
1197                         else
1198                                 olck->ols_state = OLS_NEW;
1199                 } else {
1200                         LASSERT(lock->cll_error);
1201                         return lock->cll_error;
1202                 }
1203         }
1204
1205         if (olck->ols_state == OLS_NEW) {
1206                 int rc;
1207
1208                 LASSERT(olck->ols_agl);
1209                 olck->ols_agl = 0;
1210                 rc = osc_lock_enqueue(env, slice, NULL, CEF_ASYNC | CEF_MUST);
1211                 if (rc != 0)
1212                         return rc;
1213                 else
1214                         return CLO_REENQUEUED;
1215         }
1216
1217         LASSERT(equi(olck->ols_state >= OLS_UPCALL_RECEIVED &&
1218                      lock->cll_error == 0, olck->ols_lock != NULL));
1219
1220         return lock->cll_error ?: olck->ols_state >= OLS_GRANTED ? 0 : CLO_WAIT;
1221 }
1222
1223 /**
1224  * An implementation of cl_lock_operations::clo_use() method that pins cached
1225  * lock.
1226  */
1227 static int osc_lock_use(const struct lu_env *env,
1228                         const struct cl_lock_slice *slice)
1229 {
1230         struct osc_lock *olck = cl2osc_lock(slice);
1231         int rc;
1232
1233         LASSERT(!olck->ols_hold);
1234
1235         /*
1236          * Atomically check for LDLM_FL_CBPENDING and addref a lock if this
1237          * flag is not set. This protects us from a concurrent blocking ast.
1238          */
1239         rc = ldlm_lock_addref_try(&olck->ols_handle, olck->ols_einfo.ei_mode);
1240         if (rc == 0) {
1241                 olck->ols_hold = 1;
1242                 olck->ols_state = OLS_GRANTED;
1243         } else {
1244                 struct cl_lock *lock;
1245
1246                 /*
1247                  * Lock is being cancelled somewhere within
1248                  * ldlm_handle_bl_callback(): LDLM_FL_CBPENDING is already
1249                  * set, but osc_ldlm_blocking_ast() hasn't yet acquired
1250                  * cl_lock mutex.
1251                  */
1252                 lock = slice->cls_lock;
1253                 LASSERT(lock->cll_state == CLS_INTRANSIT);
1254                 LASSERT(lock->cll_users > 0);
1255                 /* set a flag for osc_dlm_blocking_ast0() to signal the
1256                  * lock.*/
1257                 olck->ols_ast_wait = 1;
1258                 rc = CLO_WAIT;
1259         }
1260         return rc;
1261 }
1262
1263 static int osc_lock_flush(struct osc_lock *ols, int discard)
1264 {
1265         struct cl_lock       *lock  = ols->ols_cl.cls_lock;
1266         struct cl_env_nest    nest;
1267         struct lu_env        *env;
1268         int result = 0;
1269         ENTRY;
1270
1271         env = cl_env_nested_get(&nest);
1272         if (!IS_ERR(env)) {
1273                 struct osc_object    *obj   = cl2osc(ols->ols_cl.cls_obj);
1274                 struct cl_lock_descr *descr = &lock->cll_descr;
1275                 int rc = 0;
1276
1277                 if (descr->cld_mode >= CLM_WRITE) {
1278                         result = osc_cache_writeback_range(env, obj,
1279                                         descr->cld_start, descr->cld_end,
1280                                         1, discard);
1281                         LDLM_DEBUG(ols->ols_lock,
1282                                 "lock %p: %d pages were %s.\n", lock, result,
1283                                 discard ? "discarded" : "written");
1284                         if (result > 0)
1285                                 result = 0;
1286                 }
1287
1288                 rc = cl_lock_discard_pages(env, lock);
1289                 if (result == 0 && rc < 0)
1290                         result = rc;
1291
1292                 cl_env_nested_put(&nest, env);
1293         } else
1294                 result = PTR_ERR(env);
1295         if (result == 0) {
1296                 ols->ols_flush = 1;
1297                 LINVRNT(!osc_lock_has_pages(ols));
1298         }
1299         RETURN(result);
1300 }
1301
1302 /**
1303  * Implements cl_lock_operations::clo_cancel() method for osc layer. This is
1304  * called (as part of cl_lock_cancel()) when lock is canceled either voluntary
1305  * (LRU pressure, early cancellation, umount, etc.) or due to the conflict
1306  * with some other lock some where in the cluster. This function does the
1307  * following:
1308  *
1309  *     - invalidates all pages protected by this lock (after sending dirty
1310  *       ones to the server, as necessary);
1311  *
1312  *     - decref's underlying ldlm lock;
1313  *
1314  *     - cancels ldlm lock (ldlm_cli_cancel()).
1315  */
1316 static void osc_lock_cancel(const struct lu_env *env,
1317                             const struct cl_lock_slice *slice)
1318 {
1319         struct cl_lock   *lock    = slice->cls_lock;
1320         struct osc_lock  *olck    = cl2osc_lock(slice);
1321         struct ldlm_lock *dlmlock = olck->ols_lock;
1322         int               result  = 0;
1323         int               discard;
1324
1325         LASSERT(cl_lock_is_mutexed(lock));
1326         LINVRNT(osc_lock_invariant(olck));
1327
1328         if (dlmlock != NULL) {
1329                 int do_cancel;
1330
1331                 discard = !!(dlmlock->l_flags & LDLM_FL_DISCARD_DATA);
1332                 if (olck->ols_state >= OLS_GRANTED)
1333                         result = osc_lock_flush(olck, discard);
1334                 osc_lock_unhold(olck);
1335
1336                 lock_res_and_lock(dlmlock);
1337                 /* Now that we're the only user of dlm read/write reference,
1338                  * mostly the ->l_readers + ->l_writers should be zero.
1339                  * However, there is a corner case.
1340                  * See bug 18829 for details.*/
1341                 do_cancel = (dlmlock->l_readers == 0 &&
1342                              dlmlock->l_writers == 0);
1343                 dlmlock->l_flags |= LDLM_FL_CBPENDING;
1344                 unlock_res_and_lock(dlmlock);
1345                 if (do_cancel)
1346                         result = ldlm_cli_cancel(&olck->ols_handle, LCF_ASYNC);
1347                 if (result < 0)
1348                         CL_LOCK_DEBUG(D_ERROR, env, lock,
1349                                       "lock %p cancel failure with error(%d)\n",
1350                                       lock, result);
1351         }
1352         olck->ols_state = OLS_CANCELLED;
1353         olck->ols_flags &= ~LDLM_FL_LVB_READY;
1354         osc_lock_detach(env, olck);
1355 }
1356
1357 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
1358 static int check_cb(const struct lu_env *env, struct cl_io *io,
1359                     struct cl_page *page, void *cbdata)
1360 {
1361         struct cl_lock *lock = cbdata;
1362
1363         if (lock->cll_descr.cld_mode == CLM_READ) {
1364                 struct cl_lock *tmp;
1365                 tmp = cl_lock_at_page(env, lock->cll_descr.cld_obj,
1366                                      page, lock, 1, 0);
1367                 if (tmp != NULL) {
1368                         cl_lock_put(env, tmp);
1369                         return CLP_GANG_OKAY;
1370                 }
1371         }
1372
1373         CL_LOCK_DEBUG(D_ERROR, env, lock, "still has pages\n");
1374         CL_PAGE_DEBUG(D_ERROR, env, page, "\n");
1375         return CLP_GANG_ABORT;
1376 }
1377
1378 /**
1379  * Returns true iff there are pages under \a olck not protected by other
1380  * locks.
1381  */
1382 static int osc_lock_has_pages(struct osc_lock *olck)
1383 {
1384         struct cl_lock       *lock;
1385         struct cl_lock_descr *descr;
1386         struct cl_object     *obj;
1387         struct osc_object    *oob;
1388         struct cl_env_nest    nest;
1389         struct cl_io         *io;
1390         struct lu_env        *env;
1391         int                   result;
1392
1393         env = cl_env_nested_get(&nest);
1394         if (IS_ERR(env))
1395                 return 0;
1396
1397         obj   = olck->ols_cl.cls_obj;
1398         oob   = cl2osc(obj);
1399         io    = &oob->oo_debug_io;
1400         lock  = olck->ols_cl.cls_lock;
1401         descr = &lock->cll_descr;
1402
1403         mutex_lock(&oob->oo_debug_mutex);
1404
1405         io->ci_obj = cl_object_top(obj);
1406         io->ci_ignore_layout = 1;
1407         cl_io_init(env, io, CIT_MISC, io->ci_obj);
1408         do {
1409                 result = cl_page_gang_lookup(env, obj, io,
1410                                              descr->cld_start, descr->cld_end,
1411                                              check_cb, (void *)lock);
1412                 if (result == CLP_GANG_ABORT)
1413                         break;
1414                 if (result == CLP_GANG_RESCHED)
1415                         cfs_cond_resched();
1416         } while (result != CLP_GANG_OKAY);
1417         cl_io_fini(env, io);
1418         mutex_unlock(&oob->oo_debug_mutex);
1419         cl_env_nested_put(&nest, env);
1420
1421         return (result == CLP_GANG_ABORT);
1422 }
1423 #else
1424 static int osc_lock_has_pages(struct osc_lock *olck)
1425 {
1426         return 0;
1427 }
1428 #endif /* CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK */
1429
1430 static void osc_lock_delete(const struct lu_env *env,
1431                             const struct cl_lock_slice *slice)
1432 {
1433         struct osc_lock *olck;
1434
1435         olck = cl2osc_lock(slice);
1436         if (olck->ols_glimpse) {
1437                 LASSERT(!olck->ols_hold);
1438                 LASSERT(!olck->ols_lock);
1439                 return;
1440         }
1441
1442         LINVRNT(osc_lock_invariant(olck));
1443         LINVRNT(!osc_lock_has_pages(olck));
1444
1445         osc_lock_unhold(olck);
1446         osc_lock_detach(env, olck);
1447 }
1448
1449 /**
1450  * Implements cl_lock_operations::clo_state() method for osc layer.
1451  *
1452  * Maintains osc_lock::ols_owner field.
1453  *
1454  * This assumes that lock always enters CLS_HELD (from some other state) in
1455  * the same IO context as one that requested the lock. This should not be a
1456  * problem, because context is by definition shared by all activity pertaining
1457  * to the same high-level IO.
1458  */
1459 static void osc_lock_state(const struct lu_env *env,
1460                            const struct cl_lock_slice *slice,
1461                            enum cl_lock_state state)
1462 {
1463         struct osc_lock *lock = cl2osc_lock(slice);
1464
1465         /*
1466          * XXX multiple io contexts can use the lock at the same time.
1467          */
1468         LINVRNT(osc_lock_invariant(lock));
1469         if (state == CLS_HELD && slice->cls_lock->cll_state != CLS_HELD) {
1470                 struct osc_io *oio = osc_env_io(env);
1471
1472                 LASSERT(lock->ols_owner == NULL);
1473                 lock->ols_owner = oio;
1474         } else if (state != CLS_HELD)
1475                 lock->ols_owner = NULL;
1476 }
1477
1478 static int osc_lock_print(const struct lu_env *env, void *cookie,
1479                           lu_printer_t p, const struct cl_lock_slice *slice)
1480 {
1481         struct osc_lock *lock = cl2osc_lock(slice);
1482
1483         /*
1484          * XXX print ldlm lock and einfo properly.
1485          */
1486         (*p)(env, cookie, "%p %#16llx "LPX64" %d %p ",
1487              lock->ols_lock, lock->ols_flags, lock->ols_handle.cookie,
1488              lock->ols_state, lock->ols_owner);
1489         osc_lvb_print(env, cookie, p, &lock->ols_lvb);
1490         return 0;
1491 }
1492
1493 static int osc_lock_fits_into(const struct lu_env *env,
1494                               const struct cl_lock_slice *slice,
1495                               const struct cl_lock_descr *need,
1496                               const struct cl_io *io)
1497 {
1498         struct osc_lock *ols = cl2osc_lock(slice);
1499
1500         if (need->cld_enq_flags & CEF_NEVER)
1501                 return 0;
1502
1503         if (ols->ols_state >= OLS_CANCELLED)
1504                 return 0;
1505
1506         if (need->cld_mode == CLM_PHANTOM) {
1507                 if (ols->ols_agl)
1508                         return !(ols->ols_state > OLS_RELEASED);
1509
1510                 /*
1511                  * Note: the QUEUED lock can't be matched here, otherwise
1512                  * it might cause the deadlocks.
1513                  * In read_process,
1514                  * P1: enqueued read lock, create sublock1
1515                  * P2: enqueued write lock, create sublock2(conflicted
1516                  *     with sublock1).
1517                  * P1: Grant read lock.
1518                  * P1: enqueued glimpse lock(with holding sublock1_read),
1519                  *     matched with sublock2, waiting sublock2 to be granted.
1520                  *     But sublock2 can not be granted, because P1
1521                  *     will not release sublock1. Bang!
1522                  */
1523                 if (ols->ols_state < OLS_GRANTED ||
1524                     ols->ols_state > OLS_RELEASED)
1525                         return 0;
1526         } else if (need->cld_enq_flags & CEF_MUST) {
1527                 /*
1528                  * If the lock hasn't ever enqueued, it can't be matched
1529                  * because enqueue process brings in many information
1530                  * which can be used to determine things such as lockless,
1531                  * CEF_MUST, etc.
1532                  */
1533                 if (ols->ols_state < OLS_UPCALL_RECEIVED &&
1534                     ols->ols_locklessable)
1535                         return 0;
1536         }
1537         return 1;
1538 }
1539
1540 static const struct cl_lock_operations osc_lock_ops = {
1541         .clo_fini    = osc_lock_fini,
1542         .clo_enqueue = osc_lock_enqueue,
1543         .clo_wait    = osc_lock_wait,
1544         .clo_unuse   = osc_lock_unuse,
1545         .clo_use     = osc_lock_use,
1546         .clo_delete  = osc_lock_delete,
1547         .clo_state   = osc_lock_state,
1548         .clo_cancel  = osc_lock_cancel,
1549         .clo_weigh   = osc_lock_weigh,
1550         .clo_print   = osc_lock_print,
1551         .clo_fits_into = osc_lock_fits_into,
1552 };
1553
1554 static int osc_lock_lockless_unuse(const struct lu_env *env,
1555                                    const struct cl_lock_slice *slice)
1556 {
1557         struct osc_lock *ols = cl2osc_lock(slice);
1558         struct cl_lock *lock = slice->cls_lock;
1559
1560         LASSERT(ols->ols_state == OLS_GRANTED);
1561         LINVRNT(osc_lock_invariant(ols));
1562
1563         cl_lock_cancel(env, lock);
1564         cl_lock_delete(env, lock);
1565         return 0;
1566 }
1567
1568 static void osc_lock_lockless_cancel(const struct lu_env *env,
1569                                      const struct cl_lock_slice *slice)
1570 {
1571         struct osc_lock   *ols  = cl2osc_lock(slice);
1572         int result;
1573
1574         result = osc_lock_flush(ols, 0);
1575         if (result)
1576                 CERROR("Pages for lockless lock %p were not purged(%d)\n",
1577                        ols, result);
1578         ols->ols_state = OLS_CANCELLED;
1579 }
1580
1581 static int osc_lock_lockless_wait(const struct lu_env *env,
1582                                   const struct cl_lock_slice *slice)
1583 {
1584         struct osc_lock *olck = cl2osc_lock(slice);
1585         struct cl_lock  *lock = olck->ols_cl.cls_lock;
1586
1587         LINVRNT(osc_lock_invariant(olck));
1588         LASSERT(olck->ols_state >= OLS_UPCALL_RECEIVED);
1589
1590         return lock->cll_error;
1591 }
1592
1593 static void osc_lock_lockless_state(const struct lu_env *env,
1594                                     const struct cl_lock_slice *slice,
1595                                     enum cl_lock_state state)
1596 {
1597         struct osc_lock *lock = cl2osc_lock(slice);
1598
1599         LINVRNT(osc_lock_invariant(lock));
1600         if (state == CLS_HELD) {
1601                 struct osc_io *oio  = osc_env_io(env);
1602
1603                 LASSERT(ergo(lock->ols_owner, lock->ols_owner == oio));
1604                 lock->ols_owner = oio;
1605
1606                 /* set the io to be lockless if this lock is for io's
1607                  * host object */
1608                 if (cl_object_same(oio->oi_cl.cis_obj, slice->cls_obj))
1609                         oio->oi_lockless = 1;
1610         }
1611 }
1612
1613 static int osc_lock_lockless_fits_into(const struct lu_env *env,
1614                                        const struct cl_lock_slice *slice,
1615                                        const struct cl_lock_descr *need,
1616                                        const struct cl_io *io)
1617 {
1618         struct osc_lock *lock = cl2osc_lock(slice);
1619
1620         if (!(need->cld_enq_flags & CEF_NEVER))
1621                 return 0;
1622
1623         /* lockless lock should only be used by its owning io. b22147 */
1624         return (lock->ols_owner == osc_env_io(env));
1625 }
1626
1627 static const struct cl_lock_operations osc_lock_lockless_ops = {
1628         .clo_fini      = osc_lock_fini,
1629         .clo_enqueue   = osc_lock_enqueue,
1630         .clo_wait      = osc_lock_lockless_wait,
1631         .clo_unuse     = osc_lock_lockless_unuse,
1632         .clo_state     = osc_lock_lockless_state,
1633         .clo_fits_into = osc_lock_lockless_fits_into,
1634         .clo_cancel    = osc_lock_lockless_cancel,
1635         .clo_print     = osc_lock_print
1636 };
1637
1638 int osc_lock_init(const struct lu_env *env,
1639                   struct cl_object *obj, struct cl_lock *lock,
1640                   const struct cl_io *unused)
1641 {
1642         struct osc_lock *clk;
1643         int result;
1644
1645         OBD_SLAB_ALLOC_PTR_GFP(clk, osc_lock_kmem, __GFP_IO);
1646         if (clk != NULL) {
1647                 __u32 enqflags = lock->cll_descr.cld_enq_flags;
1648
1649                 osc_lock_build_einfo(env, lock, clk, &clk->ols_einfo);
1650                 cfs_atomic_set(&clk->ols_pageref, 0);
1651                 clk->ols_state = OLS_NEW;
1652
1653                 clk->ols_flags = osc_enq2ldlm_flags(enqflags);
1654                 clk->ols_agl = !!(enqflags & CEF_AGL);
1655                 if (clk->ols_agl)
1656                         clk->ols_flags |= LDLM_FL_BLOCK_NOWAIT;
1657                 if (clk->ols_flags & LDLM_FL_HAS_INTENT)
1658                         clk->ols_glimpse = 1;
1659
1660                 cl_lock_slice_add(lock, &clk->ols_cl, obj, &osc_lock_ops);
1661
1662                 if (!(enqflags & CEF_MUST))
1663                         /* try to convert this lock to a lockless lock */
1664                         osc_lock_to_lockless(env, clk, (enqflags & CEF_NEVER));
1665                 if (clk->ols_locklessable && !(enqflags & CEF_DISCARD_DATA))
1666                         clk->ols_flags |= LDLM_FL_DENY_ON_CONTENTION;
1667
1668                 LDLM_DEBUG_NOLOCK("lock %p, osc lock %p, flags %llx\n",
1669                                 lock, clk, clk->ols_flags);
1670
1671                 result = 0;
1672         } else
1673                 result = -ENOMEM;
1674         return result;
1675 }
1676
1677 int osc_dlm_lock_pageref(struct ldlm_lock *dlm)
1678 {
1679         struct osc_lock *olock;
1680         int              rc = 0;
1681
1682         spin_lock(&osc_ast_guard);
1683         olock = dlm->l_ast_data;
1684         /*
1685          * there's a very rare race with osc_page_addref_lock(), but that
1686          * doesn't matter because in the worst case we don't cancel a lock
1687          * which we actually can, that's no harm.
1688          */
1689         if (olock != NULL &&
1690             cfs_atomic_add_return(_PAGEREF_MAGIC,
1691                                   &olock->ols_pageref) != _PAGEREF_MAGIC) {
1692                 cfs_atomic_sub(_PAGEREF_MAGIC, &olock->ols_pageref);
1693                 rc = 1;
1694         }
1695         spin_unlock(&osc_ast_guard);
1696         return rc;
1697 }
1698
1699 /** @} osc */