lustre/osc/osc_lock.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  *
  32  * Copyright (c) 2011 Whamcloud, Inc.
  33  *
  34  */
  35 /*
  36  * This file is part of Lustre, http://www.lustre.org/
  37  * Lustre is a trademark of Sun Microsystems, Inc.
  38  *
  39  * Implementation of cl_lock for OSC layer.
  40  *
  41  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  42  */
  43
  44 #define DEBUG_SUBSYSTEM S_OSC
  45
  46 #ifdef __KERNEL__
  47 # include <libcfs/libcfs.h>
  48 #else
  49 # include <liblustre.h>
  50 #endif
  51 /* fid_build_reg_res_name() */
  52 #include <lustre_fid.h>
  53
  54 #include "osc_cl_internal.h"
  55
  56 /** \addtogroup osc
  57  *  @{
  58  */
  59
  60 #define _PAGEREF_MAGIC  (-10000000)
  61
  62 /*****************************************************************************
  63  *
  64  * Type conversions.
  65  *
  66  */
  67
  68 static const struct cl_lock_operations osc_lock_ops;
  69 static const struct cl_lock_operations osc_lock_lockless_ops;
  70 static void osc_lock_to_lockless(const struct lu_env *env,
  71                                  struct osc_lock *ols, int force);
  72 static int osc_lock_has_pages(struct osc_lock *olck);
  73
  74 int osc_lock_is_lockless(const struct osc_lock *olck)
  75 {
  76         return (olck->ols_cl.cls_ops == &osc_lock_lockless_ops);
  77 }
  78
  79 /**
  80  * Returns a weak pointer to the ldlm lock identified by a handle. Returned
  81  * pointer cannot be dereferenced, as lock is not protected from concurrent
  82  * reclaim. This function is a helper for osc_lock_invariant().
  83  */
  84 static struct ldlm_lock *osc_handle_ptr(struct lustre_handle *handle)
  85 {
  86         struct ldlm_lock *lock;
  87
  88         lock = ldlm_handle2lock(handle);
  89         if (lock != NULL)
  90                 LDLM_LOCK_PUT(lock);
  91         return lock;
  92 }
  93
  94 /**
  95  * Invariant that has to be true all of the time.
  96  */
  97 static int osc_lock_invariant(struct osc_lock *ols)
  98 {
  99         struct ldlm_lock *lock        = osc_handle_ptr(&ols->ols_handle);
 100         struct ldlm_lock *olock       = ols->ols_lock;
 101         int               handle_used = lustre_handle_is_used(&ols->ols_handle);
 102
 103         return
 104                 ergo(osc_lock_is_lockless(ols),
 105                      ols->ols_locklessable && ols->ols_lock == NULL)  ||
 106                 (ergo(olock != NULL, handle_used) &&
 107                  ergo(olock != NULL,
 108                       olock->l_handle.h_cookie == ols->ols_handle.cookie) &&
 109                  /*
 110                   * Check that ->ols_handle and ->ols_lock are consistent, but
 111                   * take into account that they are set at the different time.
 112                   */
 113                  ergo(handle_used,
 114                       ergo(lock != NULL && olock != NULL, lock == olock) &&
 115                       ergo(lock == NULL, olock == NULL)) &&
 116                  ergo(ols->ols_state == OLS_CANCELLED,
 117                       olock == NULL && !handle_used) &&
 118                  /*
 119                   * DLM lock is destroyed only after we have seen cancellation
 120                   * ast.
 121                   */
 122                  ergo(olock != NULL && ols->ols_state < OLS_CANCELLED,
 123                       !olock->l_destroyed) &&
 124                  ergo(ols->ols_state == OLS_GRANTED,
 125                       olock != NULL &&
 126                       olock->l_req_mode == olock->l_granted_mode &&
 127                       ols->ols_hold));
 128 }
 129
 130 /*****************************************************************************
 131  *
 132  * Lock operations.
 133  *
 134  */
 135
 136 /**
 137  * Breaks a link between osc_lock and dlm_lock.
 138  */
 139 static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck)
 140 {
 141         struct ldlm_lock *dlmlock;
 142
 143         cfs_spin_lock(&osc_ast_guard);
 144         dlmlock = olck->ols_lock;
 145         if (dlmlock == NULL) {
 146                 cfs_spin_unlock(&osc_ast_guard);
 147                 return;
 148         }
 149
 150         olck->ols_lock = NULL;
 151         /* wb(); --- for all who checks (ols->ols_lock != NULL) before
 152          * call to osc_lock_detach() */
 153         dlmlock->l_ast_data = NULL;
 154         olck->ols_handle.cookie = 0ULL;
 155         cfs_spin_unlock(&osc_ast_guard);
 156
 157         lock_res_and_lock(dlmlock);
 158         if (dlmlock->l_granted_mode == dlmlock->l_req_mode) {
 159                 struct cl_object *obj = olck->ols_cl.cls_obj;
 160                 struct cl_attr *attr  = &osc_env_info(env)->oti_attr;
 161                 __u64 old_kms;
 162
 163                 cl_object_attr_lock(obj);
 164                 /* Must get the value under the lock to avoid possible races. */
 165                 old_kms = cl2osc(obj)->oo_oinfo->loi_kms;
 166                 /* Update the kms. Need to loop all granted locks.
 167                  * Not a problem for the client */
 168                 attr->cat_kms = ldlm_extent_shift_kms(dlmlock, old_kms);
 169
 170                 cl_object_attr_set(env, obj, attr, CAT_KMS);
 171                 cl_object_attr_unlock(obj);
 172         }
 173         unlock_res_and_lock(dlmlock);
 174
 175         /* release a reference taken in osc_lock_upcall0(). */
 176         LASSERT(olck->ols_has_ref);
 177         lu_ref_del(&dlmlock->l_reference, "osc_lock", olck);
 178         LDLM_LOCK_RELEASE(dlmlock);
 179         olck->ols_has_ref = 0;
 180 }
 181
 182 static int osc_lock_unhold(struct osc_lock *ols)
 183 {
 184         int result = 0;
 185
 186         if (ols->ols_hold) {
 187                 ols->ols_hold = 0;
 188                 result = osc_cancel_base(&ols->ols_handle,
 189                                          ols->ols_einfo.ei_mode);
 190         }
 191         return result;
 192 }
 193
 194 static int osc_lock_unuse(const struct lu_env *env,
 195                           const struct cl_lock_slice *slice)
 196 {
 197         struct osc_lock *ols = cl2osc_lock(slice);
 198
 199         LINVRNT(osc_lock_invariant(ols));
 200
 201         switch (ols->ols_state) {
 202         case OLS_NEW:
 203                 LASSERT(!ols->ols_hold);
 204                 LASSERT(ols->ols_agl);
 205                 return 0;
 206         case OLS_UPCALL_RECEIVED:
 207                 LASSERT(!ols->ols_hold);
 208                 ols->ols_state = OLS_NEW;
 209                 return 0;
 210         case OLS_GRANTED:
 211                 LASSERT(!ols->ols_glimpse);
 212                 LASSERT(ols->ols_hold);
 213                 /*
 214                  * Move lock into OLS_RELEASED state before calling
 215                  * osc_cancel_base() so that possible synchronous cancellation
 216                  * (that always happens e.g., for liblustre) sees that lock is
 217                  * released.
 218                  */
 219                 ols->ols_state = OLS_RELEASED;
 220                 return osc_lock_unhold(ols);
 221         default:
 222                 CERROR("Impossible state: %d\n", ols->ols_state);
 223                 LBUG();
 224         }
 225 }
 226
 227 static void osc_lock_fini(const struct lu_env *env,
 228                           struct cl_lock_slice *slice)
 229 {
 230         struct osc_lock  *ols = cl2osc_lock(slice);
 231
 232         LINVRNT(osc_lock_invariant(ols));
 233         /*
 234          * ->ols_hold can still be true at this point if, for example, a
 235          * thread that requested a lock was killed (and released a reference
 236          * to the lock), before reply from a server was received. In this case
 237          * lock is destroyed immediately after upcall.
 238          */
 239         osc_lock_unhold(ols);
 240         LASSERT(ols->ols_lock == NULL);
 241         LASSERT(cfs_atomic_read(&ols->ols_pageref) == 0 ||
 242                 cfs_atomic_read(&ols->ols_pageref) == _PAGEREF_MAGIC);
 243
 244         OBD_SLAB_FREE_PTR(ols, osc_lock_kmem);
 245 }
 246
 247 void osc_lock_build_res(const struct lu_env *env, const struct osc_object *obj,
 248                         struct ldlm_res_id *resname)
 249 {
 250         const struct lu_fid *fid = lu_object_fid(&obj->oo_cl.co_lu);
 251         if (0) {
 252                 /*
 253                  * In the perfect world of the future, where ost servers talk
 254                  * idif-fids...
 255                  */
 256                 fid_build_reg_res_name(fid, resname);
 257         } else {
 258                 /*
 259                  * In reality, where ost server expects ->lsm_object_id and
 260                  * ->lsm_object_seq in rename.
 261                  */
 262                 osc_build_res_name(obj->oo_oinfo->loi_id, obj->oo_oinfo->loi_seq,
 263                                    resname);
 264         }
 265 }
 266
 267 static void osc_lock_build_policy(const struct lu_env *env,
 268                                   const struct cl_lock *lock,
 269                                   ldlm_policy_data_t *policy)
 270 {
 271         const struct cl_lock_descr *d = &lock->cll_descr;
 272
 273         osc_index2policy(policy, d->cld_obj, d->cld_start, d->cld_end);
 274         policy->l_extent.gid = d->cld_gid;
 275 }
 276
 277 static int osc_enq2ldlm_flags(__u32 enqflags)
 278 {
 279         int result = 0;
 280
 281         LASSERT((enqflags & ~CEF_MASK) == 0);
 282
 283         if (enqflags & CEF_NONBLOCK)
 284                 result |= LDLM_FL_BLOCK_NOWAIT;
 285         if (enqflags & CEF_ASYNC)
 286                 result |= LDLM_FL_HAS_INTENT;
 287         if (enqflags & CEF_DISCARD_DATA)
 288                 result |= LDLM_AST_DISCARD_DATA;
 289         return result;
 290 }
 291
 292 /**
 293  * Global spin-lock protecting consistency of ldlm_lock::l_ast_data
 294  * pointers. Initialized in osc_init().
 295  */
 296 cfs_spinlock_t osc_ast_guard;
 297
 298 static struct osc_lock *osc_ast_data_get(struct ldlm_lock *dlm_lock)
 299 {
 300         struct osc_lock *olck;
 301
 302         lock_res_and_lock(dlm_lock);
 303         cfs_spin_lock(&osc_ast_guard);
 304         olck = dlm_lock->l_ast_data;
 305         if (olck != NULL) {
 306                 struct cl_lock *lock = olck->ols_cl.cls_lock;
 307                 /*
 308                  * If osc_lock holds a reference on ldlm lock, return it even
 309                  * when cl_lock is in CLS_FREEING state. This way
 310                  *
 311                  *         osc_ast_data_get(dlmlock) == NULL
 312                  *
 313                  * guarantees that all osc references on dlmlock were
 314                  * released. osc_dlm_blocking_ast0() relies on that.
 315                  */
 316                 if (lock->cll_state < CLS_FREEING || olck->ols_has_ref) {
 317                         cl_lock_get_trust(lock);
 318                         lu_ref_add_atomic(&lock->cll_reference,
 319                                           "ast", cfs_current());
 320                 } else
 321                         olck = NULL;
 322         }
 323         cfs_spin_unlock(&osc_ast_guard);
 324         unlock_res_and_lock(dlm_lock);
 325         return olck;
 326 }
 327
 328 static void osc_ast_data_put(const struct lu_env *env, struct osc_lock *olck)
 329 {
 330         struct cl_lock *lock;
 331
 332         lock = olck->ols_cl.cls_lock;
 333         lu_ref_del(&lock->cll_reference, "ast", cfs_current());
 334         cl_lock_put(env, lock);
 335 }
 336
 337 /**
 338  * Updates object attributes from a lock value block (lvb) received together
 339  * with the DLM lock reply from the server. Copy of osc_update_enqueue()
 340  * logic.
 341  *
 342  * This can be optimized to not update attributes when lock is a result of a
 343  * local match.
 344  *
 345  * Called under lock and resource spin-locks.
 346  */
 347 static void osc_lock_lvb_update(const struct lu_env *env, struct osc_lock *olck,
 348                                 int rc)
 349 {
 350         struct ost_lvb    *lvb;
 351         struct cl_object  *obj;
 352         struct lov_oinfo  *oinfo;
 353         struct cl_attr    *attr;
 354         unsigned           valid;
 355
 356         ENTRY;
 357
 358         if (!(olck->ols_flags & LDLM_FL_LVB_READY))
 359                 RETURN_EXIT;
 360
 361         lvb   = &olck->ols_lvb;
 362         obj   = olck->ols_cl.cls_obj;
 363         oinfo = cl2osc(obj)->oo_oinfo;
 364         attr  = &osc_env_info(env)->oti_attr;
 365         valid = CAT_BLOCKS | CAT_ATIME | CAT_CTIME | CAT_MTIME | CAT_SIZE;
 366         cl_lvb2attr(attr, lvb);
 367
 368         cl_object_attr_lock(obj);
 369         if (rc == 0) {
 370                 struct ldlm_lock  *dlmlock;
 371                 __u64 size;
 372
 373                 dlmlock = olck->ols_lock;
 374                 LASSERT(dlmlock != NULL);
 375
 376                 /* re-grab LVB from a dlm lock under DLM spin-locks. */
 377                 *lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
 378                 size = lvb->lvb_size;
 379                 /* Extend KMS up to the end of this lock and no further
 380                  * A lock on [x,y] means a KMS of up to y + 1 bytes! */
 381                 if (size > dlmlock->l_policy_data.l_extent.end)
 382                         size = dlmlock->l_policy_data.l_extent.end + 1;
 383                 if (size >= oinfo->loi_kms) {
 384                         LDLM_DEBUG(dlmlock, "lock acquired, setting rss="LPU64
 385                                    ", kms="LPU64, lvb->lvb_size, size);
 386                         valid |= CAT_KMS;
 387                         attr->cat_kms = size;
 388                 } else {
 389                         LDLM_DEBUG(dlmlock, "lock acquired, setting rss="
 390                                    LPU64"; leaving kms="LPU64", end="LPU64,
 391                                    lvb->lvb_size, oinfo->loi_kms,
 392                                    dlmlock->l_policy_data.l_extent.end);
 393                 }
 394                 ldlm_lock_allow_match_locked(dlmlock);
 395         } else if (rc == -ENAVAIL && olck->ols_glimpse) {
 396                 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
 397                        " kms="LPU64"\n", lvb->lvb_size, oinfo->loi_kms);
 398         } else
 399                 valid = 0;
 400
 401         if (valid != 0)
 402                 cl_object_attr_set(env, obj, attr, valid);
 403
 404         cl_object_attr_unlock(obj);
 405
 406         EXIT;
 407 }
 408
 409 /**
 410  * Called when a lock is granted, from an upcall (when server returned a
 411  * granted lock), or from completion AST, when server returned a blocked lock.
 412  *
 413  * Called under lock and resource spin-locks, that are released temporarily
 414  * here.
 415  */
 416 static void osc_lock_granted(const struct lu_env *env, struct osc_lock *olck,
 417                              struct ldlm_lock *dlmlock, int rc)
 418 {
 419         struct ldlm_extent   *ext;
 420         struct cl_lock       *lock;
 421         struct cl_lock_descr *descr;
 422
 423         LASSERT(dlmlock->l_granted_mode == dlmlock->l_req_mode);
 424
 425         ENTRY;
 426         if (olck->ols_state < OLS_GRANTED) {
 427                 lock  = olck->ols_cl.cls_lock;
 428                 ext   = &dlmlock->l_policy_data.l_extent;
 429                 descr = &osc_env_info(env)->oti_descr;
 430                 descr->cld_obj = lock->cll_descr.cld_obj;
 431
 432                 /* XXX check that ->l_granted_mode is valid. */
 433                 descr->cld_mode  = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
 434                 descr->cld_start = cl_index(descr->cld_obj, ext->start);
 435                 descr->cld_end   = cl_index(descr->cld_obj, ext->end);
 436                 descr->cld_gid   = ext->gid;
 437                 /*
 438                  * tell upper layers the extent of the lock that was actually
 439                  * granted
 440                  */
 441                 olck->ols_state = OLS_GRANTED;
 442                 osc_lock_lvb_update(env, olck, rc);
 443
 444                 /* release DLM spin-locks to allow cl_lock_{modify,signal}()
 445                  * to take a semaphore on a parent lock. This is safe, because
 446                  * spin-locks are needed to protect consistency of
 447                  * dlmlock->l_*_mode and LVB, and we have finished processing
 448                  * them. */
 449                 unlock_res_and_lock(dlmlock);
 450                 cl_lock_modify(env, lock, descr);
 451                 cl_lock_signal(env, lock);
 452                 LINVRNT(osc_lock_invariant(olck));
 453                 lock_res_and_lock(dlmlock);
 454         }
 455         EXIT;
 456 }
 457
 458 static void osc_lock_upcall0(const struct lu_env *env, struct osc_lock *olck)
 459
 460 {
 461         struct ldlm_lock *dlmlock;
 462
 463         ENTRY;
 464
 465         dlmlock = ldlm_handle2lock_long(&olck->ols_handle, 0);
 466         LASSERT(dlmlock != NULL);
 467
 468         lock_res_and_lock(dlmlock);
 469         cfs_spin_lock(&osc_ast_guard);
 470         LASSERT(dlmlock->l_ast_data == olck);
 471         LASSERT(olck->ols_lock == NULL);
 472         olck->ols_lock = dlmlock;
 473         cfs_spin_unlock(&osc_ast_guard);
 474
 475         /*
 476          * Lock might be not yet granted. In this case, completion ast
 477          * (osc_ldlm_completion_ast()) comes later and finishes lock
 478          * granting.
 479          */
 480         if (dlmlock->l_granted_mode == dlmlock->l_req_mode)
 481                 osc_lock_granted(env, olck, dlmlock, 0);
 482         unlock_res_and_lock(dlmlock);
 483
 484         /*
 485          * osc_enqueue_interpret() decrefs asynchronous locks, counter
 486          * this.
 487          */
 488         ldlm_lock_addref(&olck->ols_handle, olck->ols_einfo.ei_mode);
 489         olck->ols_hold = 1;
 490
 491         /* lock reference taken by ldlm_handle2lock_long() is owned by
 492          * osc_lock and released in osc_lock_detach() */
 493         lu_ref_add(&dlmlock->l_reference, "osc_lock", olck);
 494         olck->ols_has_ref = 1;
 495 }
 496
 497 /**
 498  * Lock upcall function that is executed either when a reply to ENQUEUE rpc is
 499  * received from a server, or after osc_enqueue_base() matched a local DLM
 500  * lock.
 501  */
 502 static int osc_lock_upcall(void *cookie, int errcode)
 503 {
 504         struct osc_lock         *olck  = cookie;
 505         struct cl_lock_slice    *slice = &olck->ols_cl;
 506         struct cl_lock          *lock  = slice->cls_lock;
 507         struct lu_env           *env;
 508         struct cl_env_nest       nest;
 509
 510         ENTRY;
 511         env = cl_env_nested_get(&nest);
 512         if (!IS_ERR(env)) {
 513                 int rc;
 514
 515                 cl_lock_mutex_get(env, lock);
 516
 517                 LASSERT(lock->cll_state >= CLS_QUEUING);
 518                 if (olck->ols_state == OLS_ENQUEUED) {
 519                         olck->ols_state = OLS_UPCALL_RECEIVED;
 520                         rc = ldlm_error2errno(errcode);
 521                 } else if (olck->ols_state == OLS_CANCELLED) {
 522                         rc = -EIO;
 523                 } else {
 524                         CERROR("Impossible state: %d\n", olck->ols_state);
 525                         LBUG();
 526                 }
 527                 if (rc) {
 528                         struct ldlm_lock *dlmlock;
 529
 530                         dlmlock = ldlm_handle2lock(&olck->ols_handle);
 531                         if (dlmlock != NULL) {
 532                                 lock_res_and_lock(dlmlock);
 533                                 cfs_spin_lock(&osc_ast_guard);
 534                                 LASSERT(olck->ols_lock == NULL);
 535                                 dlmlock->l_ast_data = NULL;
 536                                 olck->ols_handle.cookie = 0ULL;
 537                                 cfs_spin_unlock(&osc_ast_guard);
 538                                 ldlm_lock_fail_match_locked(dlmlock);
 539                                 unlock_res_and_lock(dlmlock);
 540                                 LDLM_LOCK_PUT(dlmlock);
 541                         }
 542                 } else {
 543                         if (olck->ols_glimpse) {
 544                                 olck->ols_glimpse = 0;
 545                                 olck->ols_agl = 0 ;
 546                         }
 547                         osc_lock_upcall0(env, olck);
 548                 }
 549
 550                 /* Error handling, some errors are tolerable. */
 551                 if (olck->ols_locklessable && rc == -EUSERS) {
 552                         /* This is a tolerable error, turn this lock into
 553                          * lockless lock.
 554                          */
 555                         osc_object_set_contended(cl2osc(slice->cls_obj));
 556                         LASSERT(slice->cls_ops == &osc_lock_ops);
 557
 558                         /* Change this lock to ldlmlock-less lock. */
 559                         osc_lock_to_lockless(env, olck, 1);
 560                         olck->ols_state = OLS_GRANTED;
 561                         rc = 0;
 562                 } else if (olck->ols_glimpse && rc == -ENAVAIL) {
 563                         osc_lock_lvb_update(env, olck, rc);
 564                         cl_lock_delete(env, lock);
 565                         /* Hide the error. */
 566                         rc = 0;
 567                 }
 568
 569                 if (rc == 0) {
 570                         cl_lock_signal(env, lock);
 571                         /* del user for lock upcall cookie */
 572                         cl_unuse_try(env, lock);
 573                 } else {
 574                         /* del user for lock upcall cookie */
 575                         cl_lock_user_del(env, lock);
 576                         cl_lock_error(env, lock, rc);
 577                 }
 578
 579                 cl_lock_mutex_put(env, lock);
 580
 581                 /* release cookie reference, acquired by osc_lock_enqueue() */
 582                 lu_ref_del(&lock->cll_reference, "upcall", lock);
 583                 cl_lock_put(env, lock);
 584
 585                 cl_env_nested_put(&nest, env);
 586         } else
 587                 /* should never happen, similar to osc_ldlm_blocking_ast(). */
 588                 LBUG();
 589         RETURN(errcode);
 590 }
 591
 592 /**
 593  * Core of osc_dlm_blocking_ast() logic.
 594  */
 595 static void osc_lock_blocking(const struct lu_env *env,
 596                               struct ldlm_lock *dlmlock,
 597                               struct osc_lock *olck, int blocking)
 598 {
 599         struct cl_lock *lock = olck->ols_cl.cls_lock;
 600
 601         LASSERT(olck->ols_lock == dlmlock);
 602         CLASSERT(OLS_BLOCKED < OLS_CANCELLED);
 603         LASSERT(!osc_lock_is_lockless(olck));
 604
 605         /*
 606          * Lock might be still addref-ed here, if e.g., blocking ast
 607          * is sent for a failed lock.
 608          */
 609         osc_lock_unhold(olck);
 610
 611         if (blocking && olck->ols_state < OLS_BLOCKED)
 612                 /*
 613                  * Move osc_lock into OLS_BLOCKED before canceling the lock,
 614                  * because it recursively re-enters osc_lock_blocking(), with
 615                  * the state set to OLS_CANCELLED.
 616                  */
 617                 olck->ols_state = OLS_BLOCKED;
 618         /*
 619          * cancel and destroy lock at least once no matter how blocking ast is
 620          * entered (see comment above osc_ldlm_blocking_ast() for use
 621          * cases). cl_lock_cancel() and cl_lock_delete() are idempotent.
 622          */
 623         cl_lock_cancel(env, lock);
 624         cl_lock_delete(env, lock);
 625 }
 626
 627 /**
 628  * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
 629  * and ldlm_lock caches.
 630  */
 631 static int osc_dlm_blocking_ast0(const struct lu_env *env,
 632                                  struct ldlm_lock *dlmlock,
 633                                  void *data, int flag)
 634 {
 635         struct osc_lock *olck;
 636         struct cl_lock  *lock;
 637         int result;
 638         int cancel;
 639
 640         LASSERT(flag == LDLM_CB_BLOCKING || flag == LDLM_CB_CANCELING);
 641
 642         cancel = 0;
 643         olck = osc_ast_data_get(dlmlock);
 644         if (olck != NULL) {
 645                 lock = olck->ols_cl.cls_lock;
 646                 cl_lock_mutex_get(env, lock);
 647                 LINVRNT(osc_lock_invariant(olck));
 648                 if (olck->ols_ast_wait) {
 649                         /* wake up osc_lock_use() */
 650                         cl_lock_signal(env, lock);
 651                         olck->ols_ast_wait = 0;
 652                 }
 653                 /*
 654                  * Lock might have been canceled while this thread was
 655                  * sleeping for lock mutex, but olck is pinned in memory.
 656                  */
 657                 if (olck == dlmlock->l_ast_data) {
 658                         /*
 659                          * NOTE: DLM sends blocking AST's for failed locks
 660                          *       (that are still in pre-OLS_GRANTED state)
 661                          *       too, and they have to be canceled otherwise
 662                          *       DLM lock is never destroyed and stuck in
 663                          *       the memory.
 664                          *
 665                          *       Alternatively, ldlm_cli_cancel() can be
 666                          *       called here directly for osc_locks with
 667                          *       ols_state < OLS_GRANTED to maintain an
 668                          *       invariant that ->clo_cancel() is only called
 669                          *       for locks that were granted.
 670                          */
 671                         LASSERT(data == olck);
 672                         osc_lock_blocking(env, dlmlock,
 673                                           olck, flag == LDLM_CB_BLOCKING);
 674                 } else
 675                         cancel = 1;
 676                 cl_lock_mutex_put(env, lock);
 677                 osc_ast_data_put(env, olck);
 678         } else
 679                 /*
 680                  * DLM lock exists, but there is no cl_lock attached to it.
 681                  * This is a `normal' race. cl_object and its cl_lock's can be
 682                  * removed by memory pressure, together with all pages.
 683                  */
 684                 cancel = (flag == LDLM_CB_BLOCKING);
 685
 686         if (cancel) {
 687                 struct lustre_handle *lockh;
 688
 689                 lockh = &osc_env_info(env)->oti_handle;
 690                 ldlm_lock2handle(dlmlock, lockh);
 691                 result = ldlm_cli_cancel(lockh);
 692         } else
 693                 result = 0;
 694         return result;
 695 }
 696
 697 /**
 698  * Blocking ast invoked by ldlm when dlm lock is either blocking progress of
 699  * some other lock, or is canceled. This function is installed as a
 700  * ldlm_lock::l_blocking_ast() for client extent locks.
 701  *
 702  * Control flow is tricky, because ldlm uses the same call-back
 703  * (ldlm_lock::l_blocking_ast()) for both blocking and cancellation ast's.
 704  *
 705  * \param dlmlock lock for which ast occurred.
 706  *
 707  * \param new description of a conflicting lock in case of blocking ast.
 708  *
 709  * \param data value of dlmlock->l_ast_data
 710  *
 711  * \param flag LDLM_CB_BLOCKING or LDLM_CB_CANCELING. Used to distinguish
 712  *             cancellation and blocking ast's.
 713  *
 714  * Possible use cases:
 715  *
 716  *     - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) to cancel
 717  *       lock due to lock lru pressure, or explicit user request to purge
 718  *       locks.
 719  *
 720  *     - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_BLOCKING) to notify
 721  *       us that dlmlock conflicts with another lock that some client is
 722  *       enqueing. Lock is canceled.
 723  *
 724  *           - cl_lock_cancel() is called. osc_lock_cancel() calls
 725  *             ldlm_cli_cancel() that calls
 726  *
 727  *                  dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
 728  *
 729  *             recursively entering osc_ldlm_blocking_ast().
 730  *
 731  *     - client cancels lock voluntary (e.g., as a part of early cancellation):
 732  *
 733  *           cl_lock_cancel()->
 734  *             osc_lock_cancel()->
 735  *               ldlm_cli_cancel()->
 736  *                 dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
 737  *
 738  */
 739 static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
 740                                  struct ldlm_lock_desc *new, void *data,
 741                                  int flag)
 742 {
 743         struct lu_env     *env;
 744         struct cl_env_nest nest;
 745         int                result;
 746
 747         /*
 748          * This can be called in the context of outer IO, e.g.,
 749          *
 750          *     cl_enqueue()->...
 751          *       ->osc_enqueue_base()->...
 752          *         ->ldlm_prep_elc_req()->...
 753          *           ->ldlm_cancel_callback()->...
 754          *             ->osc_ldlm_blocking_ast()
 755          *
 756          * new environment has to be created to not corrupt outer context.
 757          */
 758         env = cl_env_nested_get(&nest);
 759         if (!IS_ERR(env)) {
 760                 result = osc_dlm_blocking_ast0(env, dlmlock, data, flag);
 761                 cl_env_nested_put(&nest, env);
 762         } else {
 763                 result = PTR_ERR(env);
 764                 /*
 765                  * XXX This should never happen, as cl_lock is
 766                  * stuck. Pre-allocated environment a la vvp_inode_fini_env
 767                  * should be used.
 768                  */
 769                 LBUG();
 770         }
 771         if (result != 0) {
 772                 if (result == -ENODATA)
 773                         result = 0;
 774                 else
 775                         CERROR("BAST failed: %d\n", result);
 776         }
 777         return result;
 778 }
 779
 780 static int osc_ldlm_completion_ast(struct ldlm_lock *dlmlock,
 781                                    int flags, void *data)
 782 {
 783         struct cl_env_nest nest;
 784         struct lu_env     *env;
 785         struct osc_lock   *olck;
 786         struct cl_lock    *lock;
 787         int result;
 788         int dlmrc;
 789
 790         /* first, do dlm part of the work */
 791         dlmrc = ldlm_completion_ast_async(dlmlock, flags, data);
 792         /* then, notify cl_lock */
 793         env = cl_env_nested_get(&nest);
 794         if (!IS_ERR(env)) {
 795                 olck = osc_ast_data_get(dlmlock);
 796                 if (olck != NULL) {
 797                         lock = olck->ols_cl.cls_lock;
 798                         cl_lock_mutex_get(env, lock);
 799                         /*
 800                          * ldlm_handle_cp_callback() copied LVB from request
 801                          * to lock->l_lvb_data, store it in osc_lock.
 802                          */
 803                         LASSERT(dlmlock->l_lvb_data != NULL);
 804                         lock_res_and_lock(dlmlock);
 805                         olck->ols_lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
 806                         if (olck->ols_lock == NULL) {
 807                                 /*
 808                                  * upcall (osc_lock_upcall()) hasn't yet been
 809                                  * called. Do nothing now, upcall will bind
 810                                  * olck to dlmlock and signal the waiters.
 811                                  *
 812                                  * This maintains an invariant that osc_lock
 813                                  * and ldlm_lock are always bound when
 814                                  * osc_lock is in OLS_GRANTED state.
 815                                  */
 816                         } else if (dlmlock->l_granted_mode ==
 817                                    dlmlock->l_req_mode) {
 818                                 osc_lock_granted(env, olck, dlmlock, dlmrc);
 819                         }
 820                         unlock_res_and_lock(dlmlock);
 821
 822                         if (dlmrc != 0) {
 823                                 CL_LOCK_DEBUG(D_ERROR, env, lock,
 824                                               "dlmlock returned %d\n", dlmrc);
 825                                 cl_lock_error(env, lock, dlmrc);
 826                         }
 827                         cl_lock_mutex_put(env, lock);
 828                         osc_ast_data_put(env, olck);
 829                         result = 0;
 830                 } else
 831                         result = -ELDLM_NO_LOCK_DATA;
 832                 cl_env_nested_put(&nest, env);
 833         } else
 834                 result = PTR_ERR(env);
 835         return dlmrc ?: result;
 836 }
 837
 838 static int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
 839 {
 840         struct ptlrpc_request  *req  = data;
 841         struct osc_lock        *olck;
 842         struct cl_lock         *lock;
 843         struct cl_object       *obj;
 844         struct cl_env_nest      nest;
 845         struct lu_env          *env;
 846         struct ost_lvb         *lvb;
 847         struct req_capsule     *cap;
 848         int                     result;
 849
 850         LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK);
 851
 852         env = cl_env_nested_get(&nest);
 853         if (!IS_ERR(env)) {
 854                 /*
 855                  * osc_ast_data_get() has to go after environment is
 856                  * allocated, because osc_ast_data() acquires a
 857                  * reference to a lock, and it can only be released in
 858                  * environment.
 859                  */
 860                 olck = osc_ast_data_get(dlmlock);
 861                 if (olck != NULL) {
 862                         lock = olck->ols_cl.cls_lock;
 863                         cl_lock_mutex_get(env, lock);
 864                         cap = &req->rq_pill;
 865                         req_capsule_extend(cap, &RQF_LDLM_GL_CALLBACK);
 866                         req_capsule_set_size(cap, &RMF_DLM_LVB, RCL_SERVER,
 867                                              sizeof *lvb);
 868                         result = req_capsule_server_pack(cap);
 869                         if (result == 0) {
 870                                 lvb = req_capsule_server_get(cap, &RMF_DLM_LVB);
 871                                 obj = lock->cll_descr.cld_obj;
 872                                 result = cl_object_glimpse(env, obj, lvb);
 873                         }
 874                         cl_lock_mutex_put(env, lock);
 875                         osc_ast_data_put(env, olck);
 876                 } else {
 877                         /*
 878                          * These errors are normal races, so we don't want to
 879                          * fill the console with messages by calling
 880                          * ptlrpc_error()
 881                          */
 882                         lustre_pack_reply(req, 1, NULL, NULL);
 883                         result = -ELDLM_NO_LOCK_DATA;
 884                 }
 885                 cl_env_nested_put(&nest, env);
 886         } else
 887                 result = PTR_ERR(env);
 888         req->rq_status = result;
 889         return result;
 890 }
 891
 892 static unsigned long osc_lock_weigh(const struct lu_env *env,
 893                                     const struct cl_lock_slice *slice)
 894 {
 895         /*
 896          * don't need to grab coh_page_guard since we don't care the exact #
 897          * of pages..
 898          */
 899         return cl_object_header(slice->cls_obj)->coh_pages;
 900 }
 901
 902 /**
 903  * Get the weight of dlm lock for early cancellation.
 904  *
 905  * XXX: it should return the pages covered by this \a dlmlock.
 906  */
 907 static unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock)
 908 {
 909         struct cl_env_nest       nest;
 910         struct lu_env           *env;
 911         struct osc_lock         *lock;
 912         struct cl_lock          *cll;
 913         unsigned long            weight;
 914         ENTRY;
 915
 916         cfs_might_sleep();
 917         /*
 918          * osc_ldlm_weigh_ast has a complex context since it might be called
 919          * because of lock canceling, or from user's input. We have to make
 920          * a new environment for it. Probably it is implementation safe to use
 921          * the upper context because cl_lock_put don't modify environment
 922          * variables. But in case of ..
 923          */
 924         env = cl_env_nested_get(&nest);
 925         if (IS_ERR(env))
 926                 /* Mostly because lack of memory, tend to eliminate this lock*/
 927                 RETURN(0);
 928
 929         LASSERT(dlmlock->l_resource->lr_type == LDLM_EXTENT);
 930         lock = osc_ast_data_get(dlmlock);
 931         if (lock == NULL) {
 932                 /* cl_lock was destroyed because of memory pressure.
 933                  * It is much reasonable to assign this type of lock
 934                  * a lower cost.
 935                  */
 936                 GOTO(out, weight = 0);
 937         }
 938
 939         cll = lock->ols_cl.cls_lock;
 940         cl_lock_mutex_get(env, cll);
 941         weight = cl_lock_weigh(env, cll);
 942         cl_lock_mutex_put(env, cll);
 943         osc_ast_data_put(env, lock);
 944         EXIT;
 945
 946 out:
 947         cl_env_nested_put(&nest, env);
 948         return weight;
 949 }
 950
 951 static void osc_lock_build_einfo(const struct lu_env *env,
 952                                  const struct cl_lock *clock,
 953                                  struct osc_lock *lock,
 954                                  struct ldlm_enqueue_info *einfo)
 955 {
 956         enum cl_lock_mode mode;
 957
 958         mode = clock->cll_descr.cld_mode;
 959         if (mode == CLM_PHANTOM)
 960                 /*
 961                  * For now, enqueue all glimpse locks in read mode. In the
 962                  * future, client might choose to enqueue LCK_PW lock for
 963                  * glimpse on a file opened for write.
 964                  */
 965                 mode = CLM_READ;
 966
 967         einfo->ei_type   = LDLM_EXTENT;
 968         einfo->ei_mode   = osc_cl_lock2ldlm(mode);
 969         einfo->ei_cb_bl  = osc_ldlm_blocking_ast;
 970         einfo->ei_cb_cp  = osc_ldlm_completion_ast;
 971         einfo->ei_cb_gl  = osc_ldlm_glimpse_ast;
 972         einfo->ei_cb_wg  = osc_ldlm_weigh_ast;
 973         einfo->ei_cbdata = lock; /* value to be put into ->l_ast_data */
 974 }
 975
 976 /**
 977  * Determine if the lock should be converted into a lockless lock.
 978  *
 979  * Steps to check:
 980  * - if the lock has an explicite requirment for a non-lockless lock;
 981  * - if the io lock request type ci_lockreq;
 982  * - send the enqueue rpc to ost to make the further decision;
 983  * - special treat to truncate lockless lock
 984  *
 985  *  Additional policy can be implemented here, e.g., never do lockless-io
 986  *  for large extents.
 987  */
 988 static void osc_lock_to_lockless(const struct lu_env *env,
 989                                  struct osc_lock *ols, int force)
 990 {
 991         struct cl_lock_slice *slice = &ols->ols_cl;
 992         struct cl_lock *lock        = slice->cls_lock;
 993
 994         LASSERT(ols->ols_state == OLS_NEW ||
 995                 ols->ols_state == OLS_UPCALL_RECEIVED);
 996
 997         if (force) {
 998                 ols->ols_locklessable = 1;
 999                 LASSERT(cl_lock_is_mutexed(lock));
1000                 slice->cls_ops = &osc_lock_lockless_ops;
1001         } else {
1002                 struct osc_io *oio     = osc_env_io(env);
1003                 struct cl_io  *io      = oio->oi_cl.cis_io;
1004                 struct cl_object *obj  = slice->cls_obj;
1005                 struct osc_object *oob = cl2osc(obj);
1006                 const struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
1007                 struct obd_connect_data *ocd;
1008
1009                 LASSERT(io->ci_lockreq == CILR_MANDATORY ||
1010                         io->ci_lockreq == CILR_MAYBE ||
1011                         io->ci_lockreq == CILR_NEVER);
1012
1013                 ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data;
1014                 ols->ols_locklessable = (io->ci_type != CIT_SETATTR) &&
1015                                 (io->ci_lockreq == CILR_MAYBE) &&
1016                                 (ocd->ocd_connect_flags & OBD_CONNECT_SRVLOCK);
1017                 if (io->ci_lockreq == CILR_NEVER ||
1018                         /* lockless IO */
1019                     (ols->ols_locklessable && osc_object_is_contended(oob)) ||
1020                         /* lockless truncate */
1021                     (cl_io_is_trunc(io) &&
1022                      (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK) &&
1023                       osd->od_lockless_truncate)) {
1024                         ols->ols_locklessable = 1;
1025                         slice->cls_ops = &osc_lock_lockless_ops;
1026                 }
1027         }
1028         LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
1029 }
1030
1031 static int osc_lock_compatible(const struct osc_lock *qing,
1032                                const struct osc_lock *qed)
1033 {
1034         enum cl_lock_mode qing_mode;
1035         enum cl_lock_mode qed_mode;
1036
1037         qing_mode = qing->ols_cl.cls_lock->cll_descr.cld_mode;
1038         if (qed->ols_glimpse &&
1039             (qed->ols_state >= OLS_UPCALL_RECEIVED || qing_mode == CLM_READ))
1040                 return 1;
1041
1042         qed_mode = qed->ols_cl.cls_lock->cll_descr.cld_mode;
1043         return ((qing_mode == CLM_READ) && (qed_mode == CLM_READ));
1044 }
1045
1046 /**
1047  * Cancel all conflicting locks and wait for them to be destroyed.
1048  *
1049  * This function is used for two purposes:
1050  *
1051  *     - early cancel all conflicting locks before starting IO, and
1052  *
1053  *     - guarantee that pages added to the page cache by lockless IO are never
1054  *       covered by locks other than lockless IO lock, and, hence, are not
1055  *       visible to other threads.
1056  */
1057 static int osc_lock_enqueue_wait(const struct lu_env *env,
1058                                  const struct osc_lock *olck)
1059 {
1060         struct cl_lock          *lock    = olck->ols_cl.cls_lock;
1061         struct cl_lock_descr    *descr   = &lock->cll_descr;
1062         struct cl_object_header *hdr     = cl_object_header(descr->cld_obj);
1063         struct cl_lock          *scan;
1064         struct cl_lock          *conflict= NULL;
1065         int lockless                     = osc_lock_is_lockless(olck);
1066         int rc                           = 0;
1067         ENTRY;
1068
1069         LASSERT(cl_lock_is_mutexed(lock));
1070
1071         /* make it enqueue anyway for glimpse lock, because we actually
1072          * don't need to cancel any conflicting locks. */
1073         if (olck->ols_glimpse)
1074                 return 0;
1075
1076         cfs_spin_lock(&hdr->coh_lock_guard);
1077         cfs_list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) {
1078                 struct cl_lock_descr *cld = &scan->cll_descr;
1079                 const struct osc_lock *scan_ols;
1080
1081                 if (scan == lock)
1082                         break;
1083
1084                 if (scan->cll_state < CLS_QUEUING ||
1085                     scan->cll_state == CLS_FREEING ||
1086                     cld->cld_start > descr->cld_end ||
1087                     cld->cld_end < descr->cld_start)
1088                         continue;
1089
1090                 /* overlapped and living locks. */
1091
1092                 /* We're not supposed to give up group lock. */
1093                 if (scan->cll_descr.cld_mode == CLM_GROUP) {
1094                         LASSERT(descr->cld_mode != CLM_GROUP ||
1095                                 descr->cld_gid != scan->cll_descr.cld_gid);
1096                         continue;
1097                 }
1098
1099                 scan_ols = osc_lock_at(scan);
1100
1101                 /* We need to cancel the compatible locks if we're enqueuing
1102                  * a lockless lock, for example:
1103                  * imagine that client has PR lock on [0, 1000], and thread T0
1104                  * is doing lockless IO in [500, 1500] region. Concurrent
1105                  * thread T1 can see lockless data in [500, 1000], which is
1106                  * wrong, because these data are possibly stale. */
1107                 if (!lockless && osc_lock_compatible(olck, scan_ols))
1108                         continue;
1109
1110                 /* Now @scan is conflicting with @lock, this means current
1111                  * thread have to sleep for @scan being destroyed. */
1112                 if (scan_ols->ols_owner == osc_env_io(env)) {
1113                         CERROR("DEADLOCK POSSIBLE!\n");
1114                         CL_LOCK_DEBUG(D_ERROR, env, scan, "queued.\n");
1115                         CL_LOCK_DEBUG(D_ERROR, env, lock, "queuing.\n");
1116                         libcfs_debug_dumpstack(NULL);
1117                 }
1118                 cl_lock_get_trust(scan);
1119                 conflict = scan;
1120                 break;
1121         }
1122         cfs_spin_unlock(&hdr->coh_lock_guard);
1123
1124         if (conflict) {
1125                 if (lock->cll_descr.cld_mode == CLM_GROUP) {
1126                         /* we want a group lock but a previous lock request
1127                          * conflicts, we do not wait but return 0 so the
1128                          * request is send to the server
1129                          */
1130                         CDEBUG(D_DLMTRACE, "group lock %p is conflicted "
1131                                            "with %p, no wait, send to server\n",
1132                                lock, conflict);
1133                         cl_lock_put(env, conflict);
1134                         rc = 0;
1135                 } else {
1136                         CDEBUG(D_DLMTRACE, "lock %p is conflicted with %p, "
1137                                            "will wait\n",
1138                                lock, conflict);
1139                         LASSERT(lock->cll_conflict == NULL);
1140                         lu_ref_add(&conflict->cll_reference, "cancel-wait",
1141                                    lock);
1142                         lock->cll_conflict = conflict;
1143                         rc = CLO_WAIT;
1144                 }
1145         }
1146         RETURN(rc);
1147 }
1148
1149 /**
1150  * Implementation of cl_lock_operations::clo_enqueue() method for osc
1151  * layer. This initiates ldlm enqueue:
1152  *
1153  *     - cancels conflicting locks early (osc_lock_enqueue_wait());
1154  *
1155  *     - calls osc_enqueue_base() to do actual enqueue.
1156  *
1157  * osc_enqueue_base() is supplied with an upcall function that is executed
1158  * when lock is received either after a local cached ldlm lock is matched, or
1159  * when a reply from the server is received.
1160  *
1161  * This function does not wait for the network communication to complete.
1162  */
1163 static int osc_lock_enqueue(const struct lu_env *env,
1164                             const struct cl_lock_slice *slice,
1165                             struct cl_io *unused, __u32 enqflags)
1166 {
1167         struct osc_lock          *ols     = cl2osc_lock(slice);
1168         struct cl_lock           *lock    = ols->ols_cl.cls_lock;
1169         int result;
1170         ENTRY;
1171
1172         LASSERT(cl_lock_is_mutexed(lock));
1173         LASSERTF(ols->ols_state == OLS_NEW,
1174                  "Impossible state: %d\n", ols->ols_state);
1175
1176         ols->ols_flags = osc_enq2ldlm_flags(enqflags);
1177         if (enqflags & CEF_AGL) {
1178                 ols->ols_flags |= LDLM_FL_BLOCK_NOWAIT;
1179                 ols->ols_agl = 1;
1180         }
1181         if (ols->ols_flags & LDLM_FL_HAS_INTENT)
1182                 ols->ols_glimpse = 1;
1183         if (!osc_lock_is_lockless(ols) && !(enqflags & CEF_MUST))
1184                 /* try to convert this lock to a lockless lock */
1185                 osc_lock_to_lockless(env, ols, (enqflags & CEF_NEVER));
1186
1187         result = osc_lock_enqueue_wait(env, ols);
1188         if (result == 0) {
1189                 if (!osc_lock_is_lockless(ols)) {
1190                         struct osc_object        *obj = cl2osc(slice->cls_obj);
1191                         struct osc_thread_info   *info = osc_env_info(env);
1192                         struct ldlm_res_id       *resname = &info->oti_resname;
1193                         ldlm_policy_data_t       *policy = &info->oti_policy;
1194                         struct ldlm_enqueue_info *einfo = &ols->ols_einfo;
1195
1196                         if (ols->ols_locklessable)
1197                                 ols->ols_flags |= LDLM_FL_DENY_ON_CONTENTION;
1198
1199                         /* a reference for lock, passed as an upcall cookie */
1200                         cl_lock_get(lock);
1201                         lu_ref_add(&lock->cll_reference, "upcall", lock);
1202                         /* a user for lock also */
1203                         cl_lock_user_add(env, lock);
1204                         ols->ols_state = OLS_ENQUEUED;
1205
1206                         /*
1207                          * XXX: this is possible blocking point as
1208                          * ldlm_lock_match(LDLM_FL_LVB_READY) waits for
1209                          * LDLM_CP_CALLBACK.
1210                          */
1211                         osc_lock_build_res(env, obj, resname);
1212                         osc_lock_build_policy(env, lock, policy);
1213                         result = osc_enqueue_base(osc_export(obj), resname,
1214                                           &ols->ols_flags, policy,
1215                                           &ols->ols_lvb,
1216                                           obj->oo_oinfo->loi_kms_valid,
1217                                           osc_lock_upcall,
1218                                           ols, einfo, &ols->ols_handle,
1219                                           PTLRPCD_SET, 1, ols->ols_agl);
1220                         if (result != 0) {
1221                                 cl_lock_user_del(env, lock);
1222                                 lu_ref_del(&lock->cll_reference,
1223                                            "upcall", lock);
1224                                 cl_lock_put(env, lock);
1225                                 if (unlikely(result == -ECANCELED)) {
1226                                         ols->ols_state = OLS_NEW;
1227                                         result = 0;
1228                                 }
1229                         }
1230                 } else {
1231                         ols->ols_state = OLS_GRANTED;
1232                         ols->ols_owner = osc_env_io(env);
1233                 }
1234         }
1235         LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
1236         RETURN(result);
1237 }
1238
1239 static int osc_lock_wait(const struct lu_env *env,
1240                          const struct cl_lock_slice *slice)
1241 {
1242         struct osc_lock *olck = cl2osc_lock(slice);
1243         struct cl_lock  *lock = olck->ols_cl.cls_lock;
1244
1245         LINVRNT(osc_lock_invariant(olck));
1246
1247         if (olck->ols_glimpse && olck->ols_state >= OLS_UPCALL_RECEIVED) {
1248                 if (olck->ols_flags & LDLM_FL_LVB_READY) {
1249                         return 0;
1250                 } else if (olck->ols_agl) {
1251                         olck->ols_state = OLS_NEW;
1252                 } else {
1253                         LASSERT(lock->cll_error);
1254                         return lock->cll_error;
1255                 }
1256         }
1257
1258         if (olck->ols_state == OLS_NEW) {
1259                 if (lock->cll_descr.cld_enq_flags & CEF_NO_REENQUEUE) {
1260                         return -ENAVAIL;
1261                 } else {
1262                         int rc;
1263
1264                         LASSERT(olck->ols_agl);
1265
1266                         rc = osc_lock_enqueue(env, slice, NULL, CEF_ASYNC |
1267                                                                 CEF_MUST);
1268                         if (rc != 0)
1269                                 return rc;
1270                         else
1271                                 return CLO_REENQUEUED;
1272                 }
1273         }
1274
1275         LASSERT(equi(olck->ols_state >= OLS_UPCALL_RECEIVED &&
1276                      lock->cll_error == 0, olck->ols_lock != NULL));
1277
1278         return lock->cll_error ?: olck->ols_state >= OLS_GRANTED ? 0 : CLO_WAIT;
1279 }
1280
1281 /**
1282  * An implementation of cl_lock_operations::clo_use() method that pins cached
1283  * lock.
1284  */
1285 static int osc_lock_use(const struct lu_env *env,
1286                         const struct cl_lock_slice *slice)
1287 {
1288         struct osc_lock *olck = cl2osc_lock(slice);
1289         int rc;
1290
1291         LASSERT(!olck->ols_hold);
1292
1293         /*
1294          * Atomically check for LDLM_FL_CBPENDING and addref a lock if this
1295          * flag is not set. This protects us from a concurrent blocking ast.
1296          */
1297         rc = ldlm_lock_addref_try(&olck->ols_handle, olck->ols_einfo.ei_mode);
1298         if (rc == 0) {
1299                 olck->ols_hold = 1;
1300                 olck->ols_state = OLS_GRANTED;
1301         } else {
1302                 struct cl_lock *lock;
1303
1304                 /*
1305                  * Lock is being cancelled somewhere within
1306                  * ldlm_handle_bl_callback(): LDLM_FL_CBPENDING is already
1307                  * set, but osc_ldlm_blocking_ast() hasn't yet acquired
1308                  * cl_lock mutex.
1309                  */
1310                 lock = slice->cls_lock;
1311                 LASSERT(lock->cll_state == CLS_INTRANSIT);
1312                 LASSERT(lock->cll_users > 0);
1313                 /* set a flag for osc_dlm_blocking_ast0() to signal the
1314                  * lock.*/
1315                 olck->ols_ast_wait = 1;
1316                 rc = CLO_WAIT;
1317         }
1318         return rc;
1319 }
1320
1321 static int osc_lock_flush(struct osc_lock *ols, int discard)
1322 {
1323         struct cl_lock       *lock  = ols->ols_cl.cls_lock;
1324         struct cl_env_nest    nest;
1325         struct lu_env        *env;
1326         int result = 0;
1327
1328         env = cl_env_nested_get(&nest);
1329         if (!IS_ERR(env)) {
1330                 result = cl_lock_page_out(env, lock, discard);
1331                 cl_env_nested_put(&nest, env);
1332         } else
1333                 result = PTR_ERR(env);
1334         if (result == 0) {
1335                 ols->ols_flush = 1;
1336                 LINVRNT(!osc_lock_has_pages(ols));
1337         }
1338         return result;
1339 }
1340
1341 /**
1342  * Implements cl_lock_operations::clo_cancel() method for osc layer. This is
1343  * called (as part of cl_lock_cancel()) when lock is canceled either voluntary
1344  * (LRU pressure, early cancellation, umount, etc.) or due to the conflict
1345  * with some other lock some where in the cluster. This function does the
1346  * following:
1347  *
1348  *     - invalidates all pages protected by this lock (after sending dirty
1349  *       ones to the server, as necessary);
1350  *
1351  *     - decref's underlying ldlm lock;
1352  *
1353  *     - cancels ldlm lock (ldlm_cli_cancel()).
1354  */
1355 static void osc_lock_cancel(const struct lu_env *env,
1356                             const struct cl_lock_slice *slice)
1357 {
1358         struct cl_lock   *lock    = slice->cls_lock;
1359         struct osc_lock  *olck    = cl2osc_lock(slice);
1360         struct ldlm_lock *dlmlock = olck->ols_lock;
1361         int               result  = 0;
1362         int               discard;
1363
1364         LASSERT(cl_lock_is_mutexed(lock));
1365         LINVRNT(osc_lock_invariant(olck));
1366
1367         if (dlmlock != NULL) {
1368                 int do_cancel;
1369
1370                 discard = !!(dlmlock->l_flags & LDLM_FL_DISCARD_DATA);
1371                 result = osc_lock_flush(olck, discard);
1372                 osc_lock_unhold(olck);
1373
1374                 lock_res_and_lock(dlmlock);
1375                 /* Now that we're the only user of dlm read/write reference,
1376                  * mostly the ->l_readers + ->l_writers should be zero.
1377                  * However, there is a corner case.
1378                  * See bug 18829 for details.*/
1379                 do_cancel = (dlmlock->l_readers == 0 &&
1380                              dlmlock->l_writers == 0);
1381                 dlmlock->l_flags |= LDLM_FL_CBPENDING;
1382                 unlock_res_and_lock(dlmlock);
1383                 if (do_cancel)
1384                         result = ldlm_cli_cancel(&olck->ols_handle);
1385                 if (result < 0)
1386                         CL_LOCK_DEBUG(D_ERROR, env, lock,
1387                                       "lock %p cancel failure with error(%d)\n",
1388                                       lock, result);
1389         }
1390         olck->ols_state = OLS_CANCELLED;
1391         olck->ols_flags &= ~LDLM_FL_LVB_READY;
1392         osc_lock_detach(env, olck);
1393 }
1394
1395 #ifdef INVARIANT_CHECK
1396 static int check_cb(const struct lu_env *env, struct cl_io *io,
1397                     struct cl_page *page, void *cbdata)
1398 {
1399         struct cl_lock *lock = cbdata;
1400
1401         if (lock->cll_descr.cld_mode == CLM_READ) {
1402                 struct cl_lock *tmp;
1403                 tmp = cl_lock_at_page(env, lock->cll_descr.cld_obj,
1404                                      page, lock, 1, 0);
1405                 if (tmp != NULL) {
1406                         cl_lock_put(env, tmp);
1407                         return CLP_GANG_OKAY;
1408                 }
1409         }
1410
1411         CL_LOCK_DEBUG(D_ERROR, env, lock, "still has pages\n");
1412         CL_PAGE_DEBUG(D_ERROR, env, page, "\n");
1413         return CLP_GANG_ABORT;
1414 }
1415
1416 /**
1417  * Returns true iff there are pages under \a olck not protected by other
1418  * locks.
1419  */
1420 static int osc_lock_has_pages(struct osc_lock *olck)
1421 {
1422         struct cl_lock       *lock;
1423         struct cl_lock_descr *descr;
1424         struct cl_object     *obj;
1425         struct osc_object    *oob;
1426         struct cl_env_nest    nest;
1427         struct cl_io         *io;
1428         struct lu_env        *env;
1429         int                   result;
1430
1431         env = cl_env_nested_get(&nest);
1432         if (IS_ERR(env))
1433                 return 0;
1434
1435         obj   = olck->ols_cl.cls_obj;
1436         oob   = cl2osc(obj);
1437         io    = &oob->oo_debug_io;
1438         lock  = olck->ols_cl.cls_lock;
1439         descr = &lock->cll_descr;
1440
1441         cfs_mutex_lock(&oob->oo_debug_mutex);
1442
1443         io->ci_obj = cl_object_top(obj);
1444         cl_io_init(env, io, CIT_MISC, io->ci_obj);
1445         do {
1446                 result = cl_page_gang_lookup(env, obj, io,
1447                                              descr->cld_start, descr->cld_end,
1448                                              check_cb, (void *)lock);
1449                 if (result == CLP_GANG_ABORT)
1450                         break;
1451                 if (result == CLP_GANG_RESCHED)
1452                         cfs_cond_resched();
1453         } while (result != CLP_GANG_OKAY);
1454         cl_io_fini(env, io);
1455         cfs_mutex_unlock(&oob->oo_debug_mutex);
1456         cl_env_nested_put(&nest, env);
1457
1458         return (result == CLP_GANG_ABORT);
1459 }
1460 #else
1461 static int osc_lock_has_pages(struct osc_lock *olck)
1462 {
1463         return 0;
1464 }
1465 #endif /* INVARIANT_CHECK */
1466
1467 static void osc_lock_delete(const struct lu_env *env,
1468                             const struct cl_lock_slice *slice)
1469 {
1470         struct osc_lock *olck;
1471
1472         olck = cl2osc_lock(slice);
1473         if (olck->ols_glimpse) {
1474                 LASSERT(!olck->ols_hold);
1475                 LASSERT(!olck->ols_lock);
1476                 return;
1477         }
1478
1479         LINVRNT(osc_lock_invariant(olck));
1480         LINVRNT(!osc_lock_has_pages(olck));
1481
1482         osc_lock_unhold(olck);
1483         osc_lock_detach(env, olck);
1484 }
1485
1486 /**
1487  * Implements cl_lock_operations::clo_state() method for osc layer.
1488  *
1489  * Maintains osc_lock::ols_owner field.
1490  *
1491  * This assumes that lock always enters CLS_HELD (from some other state) in
1492  * the same IO context as one that requested the lock. This should not be a
1493  * problem, because context is by definition shared by all activity pertaining
1494  * to the same high-level IO.
1495  */
1496 static void osc_lock_state(const struct lu_env *env,
1497                            const struct cl_lock_slice *slice,
1498                            enum cl_lock_state state)
1499 {
1500         struct osc_lock *lock = cl2osc_lock(slice);
1501
1502         /*
1503          * XXX multiple io contexts can use the lock at the same time.
1504          */
1505         LINVRNT(osc_lock_invariant(lock));
1506         if (state == CLS_HELD && slice->cls_lock->cll_state != CLS_HELD) {
1507                 struct osc_io *oio = osc_env_io(env);
1508
1509                 LASSERT(lock->ols_owner == NULL);
1510                 lock->ols_owner = oio;
1511         } else if (state != CLS_HELD)
1512                 lock->ols_owner = NULL;
1513 }
1514
1515 static int osc_lock_print(const struct lu_env *env, void *cookie,
1516                           lu_printer_t p, const struct cl_lock_slice *slice)
1517 {
1518         struct osc_lock *lock = cl2osc_lock(slice);
1519
1520         /*
1521          * XXX print ldlm lock and einfo properly.
1522          */
1523         (*p)(env, cookie, "%p %08x "LPX64" %d %p ",
1524              lock->ols_lock, lock->ols_flags, lock->ols_handle.cookie,
1525              lock->ols_state, lock->ols_owner);
1526         osc_lvb_print(env, cookie, p, &lock->ols_lvb);
1527         return 0;
1528 }
1529
1530 static int osc_lock_fits_into(const struct lu_env *env,
1531                               const struct cl_lock_slice *slice,
1532                               const struct cl_lock_descr *need,
1533                               const struct cl_io *io)
1534 {
1535         struct osc_lock *ols = cl2osc_lock(slice);
1536
1537         if (need->cld_enq_flags & CEF_NEVER)
1538                 return 0;
1539
1540         if (need->cld_mode == CLM_PHANTOM) {
1541                 if (ols->ols_agl)
1542                         return !(ols->ols_state > OLS_RELEASED);
1543
1544                 /*
1545                  * Note: the QUEUED lock can't be matched here, otherwise
1546                  * it might cause the deadlocks.
1547                  * In read_process,
1548                  * P1: enqueued read lock, create sublock1
1549                  * P2: enqueued write lock, create sublock2(conflicted
1550                  *     with sublock1).
1551                  * P1: Grant read lock.
1552                  * P1: enqueued glimpse lock(with holding sublock1_read),
1553                  *     matched with sublock2, waiting sublock2 to be granted.
1554                  *     But sublock2 can not be granted, because P1
1555                  *     will not release sublock1. Bang!
1556                  */
1557                 if (ols->ols_state < OLS_GRANTED ||
1558                     ols->ols_state > OLS_RELEASED)
1559                         return 0;
1560         } else if (need->cld_enq_flags & CEF_MUST) {
1561                 /*
1562                  * If the lock hasn't ever enqueued, it can't be matched
1563                  * because enqueue process brings in many information
1564                  * which can be used to determine things such as lockless,
1565                  * CEF_MUST, etc.
1566                  */
1567                 if (ols->ols_state < OLS_UPCALL_RECEIVED &&
1568                     ols->ols_locklessable)
1569                         return 0;
1570         }
1571         return 1;
1572 }
1573
1574 static const struct cl_lock_operations osc_lock_ops = {
1575         .clo_fini    = osc_lock_fini,
1576         .clo_enqueue = osc_lock_enqueue,
1577         .clo_wait    = osc_lock_wait,
1578         .clo_unuse   = osc_lock_unuse,
1579         .clo_use     = osc_lock_use,
1580         .clo_delete  = osc_lock_delete,
1581         .clo_state   = osc_lock_state,
1582         .clo_cancel  = osc_lock_cancel,
1583         .clo_weigh   = osc_lock_weigh,
1584         .clo_print   = osc_lock_print,
1585         .clo_fits_into = osc_lock_fits_into,
1586 };
1587
1588 static int osc_lock_lockless_unuse(const struct lu_env *env,
1589                                    const struct cl_lock_slice *slice)
1590 {
1591         struct osc_lock *ols = cl2osc_lock(slice);
1592         struct cl_lock *lock = slice->cls_lock;
1593
1594         LASSERT(ols->ols_state == OLS_GRANTED);
1595         LINVRNT(osc_lock_invariant(ols));
1596
1597         cl_lock_cancel(env, lock);
1598         cl_lock_delete(env, lock);
1599         return 0;
1600 }
1601
1602 static void osc_lock_lockless_cancel(const struct lu_env *env,
1603                                      const struct cl_lock_slice *slice)
1604 {
1605         struct osc_lock   *ols  = cl2osc_lock(slice);
1606         int result;
1607
1608         result = osc_lock_flush(ols, 0);
1609         if (result)
1610                 CERROR("Pages for lockless lock %p were not purged(%d)\n",
1611                        ols, result);
1612         ols->ols_state = OLS_CANCELLED;
1613 }
1614
1615 static int osc_lock_lockless_wait(const struct lu_env *env,
1616                                   const struct cl_lock_slice *slice)
1617 {
1618         struct osc_lock *olck = cl2osc_lock(slice);
1619         struct cl_lock  *lock = olck->ols_cl.cls_lock;
1620
1621         LINVRNT(osc_lock_invariant(olck));
1622         LASSERT(olck->ols_state >= OLS_UPCALL_RECEIVED);
1623
1624         return lock->cll_error;
1625 }
1626
1627 static void osc_lock_lockless_state(const struct lu_env *env,
1628                                     const struct cl_lock_slice *slice,
1629                                     enum cl_lock_state state)
1630 {
1631         struct osc_lock *lock = cl2osc_lock(slice);
1632
1633         LINVRNT(osc_lock_invariant(lock));
1634         if (state == CLS_HELD) {
1635                 struct osc_io *oio  = osc_env_io(env);
1636
1637                 LASSERT(ergo(lock->ols_owner, lock->ols_owner == oio));
1638                 lock->ols_owner = oio;
1639
1640                 /* set the io to be lockless if this lock is for io's
1641                  * host object */
1642                 if (cl_object_same(oio->oi_cl.cis_obj, slice->cls_obj))
1643                         oio->oi_lockless = 1;
1644         }
1645 }
1646
1647 static int osc_lock_lockless_fits_into(const struct lu_env *env,
1648                                        const struct cl_lock_slice *slice,
1649                                        const struct cl_lock_descr *need,
1650                                        const struct cl_io *io)
1651 {
1652         struct osc_lock *lock = cl2osc_lock(slice);
1653
1654         if (!(need->cld_enq_flags & CEF_NEVER))
1655                 return 0;
1656
1657         /* lockless lock should only be used by its owning io. b22147 */
1658         return (lock->ols_owner == osc_env_io(env));
1659 }
1660
1661 static const struct cl_lock_operations osc_lock_lockless_ops = {
1662         .clo_fini      = osc_lock_fini,
1663         .clo_enqueue   = osc_lock_enqueue,
1664         .clo_wait      = osc_lock_lockless_wait,
1665         .clo_unuse     = osc_lock_lockless_unuse,
1666         .clo_state     = osc_lock_lockless_state,
1667         .clo_fits_into = osc_lock_lockless_fits_into,
1668         .clo_cancel    = osc_lock_lockless_cancel,
1669         .clo_print     = osc_lock_print
1670 };
1671
1672 int osc_lock_init(const struct lu_env *env,
1673                   struct cl_object *obj, struct cl_lock *lock,
1674                   const struct cl_io *unused)
1675 {
1676         struct osc_lock *clk;
1677         int result;
1678
1679         OBD_SLAB_ALLOC_PTR_GFP(clk, osc_lock_kmem, CFS_ALLOC_IO);
1680         if (clk != NULL) {
1681                 osc_lock_build_einfo(env, lock, clk, &clk->ols_einfo);
1682                 cfs_atomic_set(&clk->ols_pageref, 0);
1683                 clk->ols_state = OLS_NEW;
1684                 cl_lock_slice_add(lock, &clk->ols_cl, obj, &osc_lock_ops);
1685                 result = 0;
1686         } else
1687                 result = -ENOMEM;
1688         return result;
1689 }
1690
1691 int osc_dlm_lock_pageref(struct ldlm_lock *dlm)
1692 {
1693         struct osc_lock *olock;
1694         int              rc = 0;
1695
1696         cfs_spin_lock(&osc_ast_guard);
1697         olock = dlm->l_ast_data;
1698         /*
1699          * there's a very rare race with osc_page_addref_lock(), but that
1700          * doesn't matter because in the worst case we don't cancel a lock
1701          * which we actually can, that's no harm.
1702          */
1703         if (olock != NULL &&
1704             cfs_atomic_add_return(_PAGEREF_MAGIC,
1705                                   &olock->ols_pageref) != _PAGEREF_MAGIC) {
1706                 cfs_atomic_sub(_PAGEREF_MAGIC, &olock->ols_pageref);
1707                 rc = 1;
1708         }
1709         cfs_spin_unlock(&osc_ast_guard);
1710         return rc;
1711 }
1712
1713 /** @} osc */