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, 2012, Whamcloud, Inc.
  33  */
  34 /*
  35  * This file is part of Lustre, http://www.lustre.org/
  36  * Lustre is a trademark of Sun Microsystems, Inc.
  37  *
  38  * Implementation of cl_lock for OSC layer.
  39  *
  40  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_OSC
  44
  45 #ifdef __KERNEL__
  46 # include <libcfs/libcfs.h>
  47 #else
  48 # include <liblustre.h>
  49 #endif
  50 /* fid_build_reg_res_name() */
  51 #include <lustre_fid.h>
  52
  53 #include "osc_cl_internal.h"
  54
  55 /** \addtogroup osc
  56  *  @{
  57  */
  58
  59 #define _PAGEREF_MAGIC  (-10000000)
  60
  61 /*****************************************************************************
  62  *
  63  * Type conversions.
  64  *
  65  */
  66
  67 static const struct cl_lock_operations osc_lock_ops;
  68 static const struct cl_lock_operations osc_lock_lockless_ops;
  69 static void osc_lock_to_lockless(const struct lu_env *env,
  70                                  struct osc_lock *ols, int force);
  71 static int osc_lock_has_pages(struct osc_lock *olck);
  72
  73 int osc_lock_is_lockless(const struct osc_lock *olck)
  74 {
  75         return (olck->ols_cl.cls_ops == &osc_lock_lockless_ops);
  76 }
  77
  78 /**
  79  * Returns a weak pointer to the ldlm lock identified by a handle. Returned
  80  * pointer cannot be dereferenced, as lock is not protected from concurrent
  81  * reclaim. This function is a helper for osc_lock_invariant().
  82  */
  83 static struct ldlm_lock *osc_handle_ptr(struct lustre_handle *handle)
  84 {
  85         struct ldlm_lock *lock;
  86
  87         lock = ldlm_handle2lock(handle);
  88         if (lock != NULL)
  89                 LDLM_LOCK_PUT(lock);
  90         return lock;
  91 }
  92
  93 /**
  94  * Invariant that has to be true all of the time.
  95  */
  96 static int osc_lock_invariant(struct osc_lock *ols)
  97 {
  98         struct ldlm_lock *lock        = osc_handle_ptr(&ols->ols_handle);
  99         struct ldlm_lock *olock       = ols->ols_lock;
 100         int               handle_used = lustre_handle_is_used(&ols->ols_handle);
 101
 102         return
 103                 ergo(osc_lock_is_lockless(ols),
 104                      ols->ols_locklessable && ols->ols_lock == NULL)  ||
 105                 (ergo(olock != NULL, handle_used) &&
 106                  ergo(olock != NULL,
 107                       olock->l_handle.h_cookie == ols->ols_handle.cookie) &&
 108                  /*
 109                   * Check that ->ols_handle and ->ols_lock are consistent, but
 110                   * take into account that they are set at the different time.
 111                   */
 112                  ergo(handle_used,
 113                       ergo(lock != NULL && olock != NULL, lock == olock) &&
 114                       ergo(lock == NULL, olock == NULL)) &&
 115                  ergo(ols->ols_state == OLS_CANCELLED,
 116                       olock == NULL && !handle_used) &&
 117                  /*
 118                   * DLM lock is destroyed only after we have seen cancellation
 119                   * ast.
 120                   */
 121                  ergo(olock != NULL && ols->ols_state < OLS_CANCELLED,
 122                       !olock->l_destroyed) &&
 123                  ergo(ols->ols_state == OLS_GRANTED,
 124                       olock != NULL &&
 125                       olock->l_req_mode == olock->l_granted_mode &&
 126                       ols->ols_hold));
 127 }
 128
 129 /*****************************************************************************
 130  *
 131  * Lock operations.
 132  *
 133  */
 134
 135 /**
 136  * Breaks a link between osc_lock and dlm_lock.
 137  */
 138 static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck)
 139 {
 140         struct ldlm_lock *dlmlock;
 141
 142         cfs_spin_lock(&osc_ast_guard);
 143         dlmlock = olck->ols_lock;
 144         if (dlmlock == NULL) {
 145                 cfs_spin_unlock(&osc_ast_guard);
 146                 return;
 147         }
 148
 149         olck->ols_lock = NULL;
 150         /* wb(); --- for all who checks (ols->ols_lock != NULL) before
 151          * call to osc_lock_detach() */
 152         dlmlock->l_ast_data = NULL;
 153         olck->ols_handle.cookie = 0ULL;
 154         cfs_spin_unlock(&osc_ast_guard);
 155
 156         lock_res_and_lock(dlmlock);
 157         if (dlmlock->l_granted_mode == dlmlock->l_req_mode) {
 158                 struct cl_object *obj = olck->ols_cl.cls_obj;
 159                 struct cl_attr *attr  = &osc_env_info(env)->oti_attr;
 160                 __u64 old_kms;
 161
 162                 cl_object_attr_lock(obj);
 163                 /* Must get the value under the lock to avoid possible races. */
 164                 old_kms = cl2osc(obj)->oo_oinfo->loi_kms;
 165                 /* Update the kms. Need to loop all granted locks.
 166                  * Not a problem for the client */
 167                 attr->cat_kms = ldlm_extent_shift_kms(dlmlock, old_kms);
 168
 169                 cl_object_attr_set(env, obj, attr, CAT_KMS);
 170                 cl_object_attr_unlock(obj);
 171         }
 172         unlock_res_and_lock(dlmlock);
 173
 174         /* release a reference taken in osc_lock_upcall0(). */
 175         LASSERT(olck->ols_has_ref);
 176         lu_ref_del(&dlmlock->l_reference, "osc_lock", olck);
 177         LDLM_LOCK_RELEASE(dlmlock);
 178         olck->ols_has_ref = 0;
 179 }
 180
 181 static int osc_lock_unhold(struct osc_lock *ols)
 182 {
 183         int result = 0;
 184
 185         if (ols->ols_hold) {
 186                 ols->ols_hold = 0;
 187                 result = osc_cancel_base(&ols->ols_handle,
 188                                          ols->ols_einfo.ei_mode);
 189         }
 190         return result;
 191 }
 192
 193 static int osc_lock_unuse(const struct lu_env *env,
 194                           const struct cl_lock_slice *slice)
 195 {
 196         struct osc_lock *ols = cl2osc_lock(slice);
 197
 198         LINVRNT(osc_lock_invariant(ols));
 199
 200         switch (ols->ols_state) {
 201         case OLS_NEW:
 202                 LASSERT(!ols->ols_hold);
 203                 LASSERT(ols->ols_agl);
 204                 return 0;
 205         case OLS_UPCALL_RECEIVED:
 206                 LASSERT(!ols->ols_hold);
 207                 ols->ols_state = OLS_NEW;
 208                 return 0;
 209         case OLS_GRANTED:
 210                 LASSERT(!ols->ols_glimpse);
 211                 LASSERT(ols->ols_hold);
 212                 /*
 213                  * Move lock into OLS_RELEASED state before calling
 214                  * osc_cancel_base() so that possible synchronous cancellation
 215                  * (that always happens e.g., for liblustre) sees that lock is
 216                  * released.
 217                  */
 218                 ols->ols_state = OLS_RELEASED;
 219                 return osc_lock_unhold(ols);
 220         default:
 221                 CERROR("Impossible state: %d\n", ols->ols_state);
 222                 LBUG();
 223         }
 224 }
 225
 226 static void osc_lock_fini(const struct lu_env *env,
 227                           struct cl_lock_slice *slice)
 228 {
 229         struct osc_lock  *ols = cl2osc_lock(slice);
 230
 231         LINVRNT(osc_lock_invariant(ols));
 232         /*
 233          * ->ols_hold can still be true at this point if, for example, a
 234          * thread that requested a lock was killed (and released a reference
 235          * to the lock), before reply from a server was received. In this case
 236          * lock is destroyed immediately after upcall.
 237          */
 238         osc_lock_unhold(ols);
 239         LASSERT(ols->ols_lock == NULL);
 240         LASSERT(cfs_atomic_read(&ols->ols_pageref) == 0 ||
 241                 cfs_atomic_read(&ols->ols_pageref) == _PAGEREF_MAGIC);
 242
 243         OBD_SLAB_FREE_PTR(ols, osc_lock_kmem);
 244 }
 245
 246 void osc_lock_build_res(const struct lu_env *env, const struct osc_object *obj,
 247                         struct ldlm_res_id *resname)
 248 {
 249         const struct lu_fid *fid = lu_object_fid(&obj->oo_cl.co_lu);
 250         if (0) {
 251                 /*
 252                  * In the perfect world of the future, where ost servers talk
 253                  * idif-fids...
 254                  */
 255                 fid_build_reg_res_name(fid, resname);
 256         } else {
 257                 /*
 258                  * In reality, where ost server expects ->lsm_object_id and
 259                  * ->lsm_object_seq in rename.
 260                  */
 261                 osc_build_res_name(obj->oo_oinfo->loi_id, obj->oo_oinfo->loi_seq,
 262                                    resname);
 263         }
 264 }
 265
 266 static void osc_lock_build_policy(const struct lu_env *env,
 267                                   const struct cl_lock *lock,
 268                                   ldlm_policy_data_t *policy)
 269 {
 270         const struct cl_lock_descr *d = &lock->cll_descr;
 271
 272         osc_index2policy(policy, d->cld_obj, d->cld_start, d->cld_end);
 273         policy->l_extent.gid = d->cld_gid;
 274 }
 275
 276 static int osc_enq2ldlm_flags(__u32 enqflags)
 277 {
 278         int result = 0;
 279
 280         LASSERT((enqflags & ~CEF_MASK) == 0);
 281
 282         if (enqflags & CEF_NONBLOCK)
 283                 result |= LDLM_FL_BLOCK_NOWAIT;
 284         if (enqflags & CEF_ASYNC)
 285                 result |= LDLM_FL_HAS_INTENT;
 286         if (enqflags & CEF_DISCARD_DATA)
 287                 result |= LDLM_AST_DISCARD_DATA;
 288         return result;
 289 }
 290
 291 /**
 292  * Global spin-lock protecting consistency of ldlm_lock::l_ast_data
 293  * pointers. Initialized in osc_init().
 294  */
 295 cfs_spinlock_t osc_ast_guard;
 296
 297 static struct osc_lock *osc_ast_data_get(struct ldlm_lock *dlm_lock)
 298 {
 299         struct osc_lock *olck;
 300
 301         lock_res_and_lock(dlm_lock);
 302         cfs_spin_lock(&osc_ast_guard);
 303         olck = dlm_lock->l_ast_data;
 304         if (olck != NULL) {
 305                 struct cl_lock *lock = olck->ols_cl.cls_lock;
 306                 /*
 307                  * If osc_lock holds a reference on ldlm lock, return it even
 308                  * when cl_lock is in CLS_FREEING state. This way
 309                  *
 310                  *         osc_ast_data_get(dlmlock) == NULL
 311                  *
 312                  * guarantees that all osc references on dlmlock were
 313                  * released. osc_dlm_blocking_ast0() relies on that.
 314                  */
 315                 if (lock->cll_state < CLS_FREEING || olck->ols_has_ref) {
 316                         cl_lock_get_trust(lock);
 317                         lu_ref_add_atomic(&lock->cll_reference,
 318                                           "ast", cfs_current());
 319                 } else
 320                         olck = NULL;
 321         }
 322         cfs_spin_unlock(&osc_ast_guard);
 323         unlock_res_and_lock(dlm_lock);
 324         return olck;
 325 }
 326
 327 static void osc_ast_data_put(const struct lu_env *env, struct osc_lock *olck)
 328 {
 329         struct cl_lock *lock;
 330
 331         lock = olck->ols_cl.cls_lock;
 332         lu_ref_del(&lock->cll_reference, "ast", cfs_current());
 333         cl_lock_put(env, lock);
 334 }
 335
 336 /**
 337  * Updates object attributes from a lock value block (lvb) received together
 338  * with the DLM lock reply from the server. Copy of osc_update_enqueue()
 339  * logic.
 340  *
 341  * This can be optimized to not update attributes when lock is a result of a
 342  * local match.
 343  *
 344  * Called under lock and resource spin-locks.
 345  */
 346 static void osc_lock_lvb_update(const struct lu_env *env, struct osc_lock *olck,
 347                                 int rc)
 348 {
 349         struct ost_lvb    *lvb;
 350         struct cl_object  *obj;
 351         struct lov_oinfo  *oinfo;
 352         struct cl_attr    *attr;
 353         unsigned           valid;
 354
 355         ENTRY;
 356
 357         if (!(olck->ols_flags & LDLM_FL_LVB_READY))
 358                 RETURN_EXIT;
 359
 360         lvb   = &olck->ols_lvb;
 361         obj   = olck->ols_cl.cls_obj;
 362         oinfo = cl2osc(obj)->oo_oinfo;
 363         attr  = &osc_env_info(env)->oti_attr;
 364         valid = CAT_BLOCKS | CAT_ATIME | CAT_CTIME | CAT_MTIME | CAT_SIZE;
 365         cl_lvb2attr(attr, lvb);
 366
 367         cl_object_attr_lock(obj);
 368         if (rc == 0) {
 369                 struct ldlm_lock  *dlmlock;
 370                 __u64 size;
 371
 372                 dlmlock = olck->ols_lock;
 373                 LASSERT(dlmlock != NULL);
 374
 375                 /* re-grab LVB from a dlm lock under DLM spin-locks. */
 376                 *lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
 377                 size = lvb->lvb_size;
 378                 /* Extend KMS up to the end of this lock and no further
 379                  * A lock on [x,y] means a KMS of up to y + 1 bytes! */
 380                 if (size > dlmlock->l_policy_data.l_extent.end)
 381                         size = dlmlock->l_policy_data.l_extent.end + 1;
 382                 if (size >= oinfo->loi_kms) {
 383                         LDLM_DEBUG(dlmlock, "lock acquired, setting rss="LPU64
 384                                    ", kms="LPU64, lvb->lvb_size, size);
 385                         valid |= CAT_KMS;
 386                         attr->cat_kms = size;
 387                 } else {
 388                         LDLM_DEBUG(dlmlock, "lock acquired, setting rss="
 389                                    LPU64"; leaving kms="LPU64", end="LPU64,
 390                                    lvb->lvb_size, oinfo->loi_kms,
 391                                    dlmlock->l_policy_data.l_extent.end);
 392                 }
 393                 ldlm_lock_allow_match_locked(dlmlock);
 394         } else if (rc == -ENAVAIL && olck->ols_glimpse) {
 395                 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
 396                        " kms="LPU64"\n", lvb->lvb_size, oinfo->loi_kms);
 397         } else
 398                 valid = 0;
 399
 400         if (valid != 0)
 401                 cl_object_attr_set(env, obj, attr, valid);
 402
 403         cl_object_attr_unlock(obj);
 404
 405         EXIT;
 406 }
 407
 408 /**
 409  * Called when a lock is granted, from an upcall (when server returned a
 410  * granted lock), or from completion AST, when server returned a blocked lock.
 411  *
 412  * Called under lock and resource spin-locks, that are released temporarily
 413  * here.
 414  */
 415 static void osc_lock_granted(const struct lu_env *env, struct osc_lock *olck,
 416                              struct ldlm_lock *dlmlock, int rc)
 417 {
 418         struct ldlm_extent   *ext;
 419         struct cl_lock       *lock;
 420         struct cl_lock_descr *descr;
 421
 422         LASSERT(dlmlock->l_granted_mode == dlmlock->l_req_mode);
 423
 424         ENTRY;
 425         if (olck->ols_state < OLS_GRANTED) {
 426                 lock  = olck->ols_cl.cls_lock;
 427                 ext   = &dlmlock->l_policy_data.l_extent;
 428                 descr = &osc_env_info(env)->oti_descr;
 429                 descr->cld_obj = lock->cll_descr.cld_obj;
 430
 431                 /* XXX check that ->l_granted_mode is valid. */
 432                 descr->cld_mode  = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
 433                 descr->cld_start = cl_index(descr->cld_obj, ext->start);
 434                 descr->cld_end   = cl_index(descr->cld_obj, ext->end);
 435                 descr->cld_gid   = ext->gid;
 436                 /*
 437                  * tell upper layers the extent of the lock that was actually
 438                  * granted
 439                  */
 440                 olck->ols_state = OLS_GRANTED;
 441                 osc_lock_lvb_update(env, olck, rc);
 442
 443                 /* release DLM spin-locks to allow cl_lock_{modify,signal}()
 444                  * to take a semaphore on a parent lock. This is safe, because
 445                  * spin-locks are needed to protect consistency of
 446                  * dlmlock->l_*_mode and LVB, and we have finished processing
 447                  * them. */
 448                 unlock_res_and_lock(dlmlock);
 449                 cl_lock_modify(env, lock, descr);
 450                 cl_lock_signal(env, lock);
 451                 LINVRNT(osc_lock_invariant(olck));
 452                 lock_res_and_lock(dlmlock);
 453         }
 454         EXIT;
 455 }
 456
 457 static void osc_lock_upcall0(const struct lu_env *env, struct osc_lock *olck)
 458
 459 {
 460         struct ldlm_lock *dlmlock;
 461
 462         ENTRY;
 463
 464         dlmlock = ldlm_handle2lock_long(&olck->ols_handle, 0);
 465         LASSERT(dlmlock != NULL);
 466
 467         lock_res_and_lock(dlmlock);
 468         cfs_spin_lock(&osc_ast_guard);
 469         LASSERT(dlmlock->l_ast_data == olck);
 470         LASSERT(olck->ols_lock == NULL);
 471         olck->ols_lock = dlmlock;
 472         cfs_spin_unlock(&osc_ast_guard);
 473
 474         /*
 475          * Lock might be not yet granted. In this case, completion ast
 476          * (osc_ldlm_completion_ast()) comes later and finishes lock
 477          * granting.
 478          */
 479         if (dlmlock->l_granted_mode == dlmlock->l_req_mode)
 480                 osc_lock_granted(env, olck, dlmlock, 0);
 481         unlock_res_and_lock(dlmlock);
 482
 483         /*
 484          * osc_enqueue_interpret() decrefs asynchronous locks, counter
 485          * this.
 486          */
 487         ldlm_lock_addref(&olck->ols_handle, olck->ols_einfo.ei_mode);
 488         olck->ols_hold = 1;
 489
 490         /* lock reference taken by ldlm_handle2lock_long() is owned by
 491          * osc_lock and released in osc_lock_detach() */
 492         lu_ref_add(&dlmlock->l_reference, "osc_lock", olck);
 493         olck->ols_has_ref = 1;
 494 }
 495
 496 /**
 497  * Lock upcall function that is executed either when a reply to ENQUEUE rpc is
 498  * received from a server, or after osc_enqueue_base() matched a local DLM
 499  * lock.
 500  */
 501 static int osc_lock_upcall(void *cookie, int errcode)
 502 {
 503         struct osc_lock         *olck  = cookie;
 504         struct cl_lock_slice    *slice = &olck->ols_cl;
 505         struct cl_lock          *lock  = slice->cls_lock;
 506         struct lu_env           *env;
 507         struct cl_env_nest       nest;
 508
 509         ENTRY;
 510         env = cl_env_nested_get(&nest);
 511         if (!IS_ERR(env)) {
 512                 int rc;
 513
 514                 cl_lock_mutex_get(env, lock);
 515
 516                 LASSERT(lock->cll_state >= CLS_QUEUING);
 517                 if (olck->ols_state == OLS_ENQUEUED) {
 518                         olck->ols_state = OLS_UPCALL_RECEIVED;
 519                         rc = ldlm_error2errno(errcode);
 520                 } else if (olck->ols_state == OLS_CANCELLED) {
 521                         rc = -EIO;
 522                 } else {
 523                         CERROR("Impossible state: %d\n", olck->ols_state);
 524                         LBUG();
 525                 }
 526                 if (rc) {
 527                         struct ldlm_lock *dlmlock;
 528
 529                         dlmlock = ldlm_handle2lock(&olck->ols_handle);
 530                         if (dlmlock != NULL) {
 531                                 lock_res_and_lock(dlmlock);
 532                                 cfs_spin_lock(&osc_ast_guard);
 533                                 LASSERT(olck->ols_lock == NULL);
 534                                 dlmlock->l_ast_data = NULL;
 535                                 olck->ols_handle.cookie = 0ULL;
 536                                 cfs_spin_unlock(&osc_ast_guard);
 537                                 ldlm_lock_fail_match_locked(dlmlock);
 538                                 unlock_res_and_lock(dlmlock);
 539                                 LDLM_LOCK_PUT(dlmlock);
 540                         }
 541                 } else {
 542                         if (olck->ols_glimpse) {
 543                                 olck->ols_glimpse = 0;
 544                                 olck->ols_agl = 0 ;
 545                         }
 546                         osc_lock_upcall0(env, olck);
 547                 }
 548
 549                 /* Error handling, some errors are tolerable. */
 550                 if (olck->ols_locklessable && rc == -EUSERS) {
 551                         /* This is a tolerable error, turn this lock into
 552                          * lockless lock.
 553                          */
 554                         osc_object_set_contended(cl2osc(slice->cls_obj));
 555                         LASSERT(slice->cls_ops == &osc_lock_ops);
 556
 557                         /* Change this lock to ldlmlock-less lock. */
 558                         osc_lock_to_lockless(env, olck, 1);
 559                         olck->ols_state = OLS_GRANTED;
 560                         rc = 0;
 561                 } else if (olck->ols_glimpse && rc == -ENAVAIL) {
 562                         osc_lock_lvb_update(env, olck, rc);
 563                         cl_lock_delete(env, lock);
 564                         /* Hide the error. */
 565                         rc = 0;
 566                 }
 567
 568                 if (rc == 0) {
 569                         cl_lock_signal(env, lock);
 570                         /* del user for lock upcall cookie */
 571                         cl_unuse_try(env, lock);
 572                 } else {
 573                         /* del user for lock upcall cookie */
 574                         cl_lock_user_del(env, lock);
 575                         cl_lock_error(env, lock, rc);
 576                 }
 577
 578                 cl_lock_mutex_put(env, lock);
 579
 580                 /* release cookie reference, acquired by osc_lock_enqueue() */
 581                 lu_ref_del(&lock->cll_reference, "upcall", lock);
 582                 cl_lock_put(env, lock);
 583
 584                 cl_env_nested_put(&nest, env);
 585         } else
 586                 /* should never happen, similar to osc_ldlm_blocking_ast(). */
 587                 LBUG();
 588         RETURN(errcode);
 589 }
 590
 591 /**
 592  * Core of osc_dlm_blocking_ast() logic.
 593  */
 594 static void osc_lock_blocking(const struct lu_env *env,
 595                               struct ldlm_lock *dlmlock,
 596                               struct osc_lock *olck, int blocking)
 597 {
 598         struct cl_lock *lock = olck->ols_cl.cls_lock;
 599
 600         LASSERT(olck->ols_lock == dlmlock);
 601         CLASSERT(OLS_BLOCKED < OLS_CANCELLED);
 602         LASSERT(!osc_lock_is_lockless(olck));
 603
 604         /*
 605          * Lock might be still addref-ed here, if e.g., blocking ast
 606          * is sent for a failed lock.
 607          */
 608         osc_lock_unhold(olck);
 609
 610         if (blocking && olck->ols_state < OLS_BLOCKED)
 611                 /*
 612                  * Move osc_lock into OLS_BLOCKED before canceling the lock,
 613                  * because it recursively re-enters osc_lock_blocking(), with
 614                  * the state set to OLS_CANCELLED.
 615                  */
 616                 olck->ols_state = OLS_BLOCKED;
 617         /*
 618          * cancel and destroy lock at least once no matter how blocking ast is
 619          * entered (see comment above osc_ldlm_blocking_ast() for use
 620          * cases). cl_lock_cancel() and cl_lock_delete() are idempotent.
 621          */
 622         cl_lock_cancel(env, lock);
 623         cl_lock_delete(env, lock);
 624 }
 625
 626 /**
 627  * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
 628  * and ldlm_lock caches.
 629  */
 630 static int osc_dlm_blocking_ast0(const struct lu_env *env,
 631                                  struct ldlm_lock *dlmlock,
 632                                  void *data, int flag)
 633 {
 634         struct osc_lock *olck;
 635         struct cl_lock  *lock;
 636         int result;
 637         int cancel;
 638
 639         LASSERT(flag == LDLM_CB_BLOCKING || flag == LDLM_CB_CANCELING);
 640
 641         cancel = 0;
 642         olck = osc_ast_data_get(dlmlock);
 643         if (olck != NULL) {
 644                 lock = olck->ols_cl.cls_lock;
 645                 cl_lock_mutex_get(env, lock);
 646                 LINVRNT(osc_lock_invariant(olck));
 647                 if (olck->ols_ast_wait) {
 648                         /* wake up osc_lock_use() */
 649                         cl_lock_signal(env, lock);
 650                         olck->ols_ast_wait = 0;
 651                 }
 652                 /*
 653                  * Lock might have been canceled while this thread was
 654                  * sleeping for lock mutex, but olck is pinned in memory.
 655                  */
 656                 if (olck == dlmlock->l_ast_data) {
 657                         /*
 658                          * NOTE: DLM sends blocking AST's for failed locks
 659                          *       (that are still in pre-OLS_GRANTED state)
 660                          *       too, and they have to be canceled otherwise
 661                          *       DLM lock is never destroyed and stuck in
 662                          *       the memory.
 663                          *
 664                          *       Alternatively, ldlm_cli_cancel() can be
 665                          *       called here directly for osc_locks with
 666                          *       ols_state < OLS_GRANTED to maintain an
 667                          *       invariant that ->clo_cancel() is only called
 668                          *       for locks that were granted.
 669                          */
 670                         LASSERT(data == olck);
 671                         osc_lock_blocking(env, dlmlock,
 672                                           olck, flag == LDLM_CB_BLOCKING);
 673                 } else
 674                         cancel = 1;
 675                 cl_lock_mutex_put(env, lock);
 676                 osc_ast_data_put(env, olck);
 677         } else
 678                 /*
 679                  * DLM lock exists, but there is no cl_lock attached to it.
 680                  * This is a `normal' race. cl_object and its cl_lock's can be
 681                  * removed by memory pressure, together with all pages.
 682                  */
 683                 cancel = (flag == LDLM_CB_BLOCKING);
 684
 685         if (cancel) {
 686                 struct lustre_handle *lockh;
 687
 688                 lockh = &osc_env_info(env)->oti_handle;
 689                 ldlm_lock2handle(dlmlock, lockh);
 690                 result = ldlm_cli_cancel(lockh);
 691         } else
 692                 result = 0;
 693         return result;
 694 }
 695
 696 /**
 697  * Blocking ast invoked by ldlm when dlm lock is either blocking progress of
 698  * some other lock, or is canceled. This function is installed as a
 699  * ldlm_lock::l_blocking_ast() for client extent locks.
 700  *
 701  * Control flow is tricky, because ldlm uses the same call-back
 702  * (ldlm_lock::l_blocking_ast()) for both blocking and cancellation ast's.
 703  *
 704  * \param dlmlock lock for which ast occurred.
 705  *
 706  * \param new description of a conflicting lock in case of blocking ast.
 707  *
 708  * \param data value of dlmlock->l_ast_data
 709  *
 710  * \param flag LDLM_CB_BLOCKING or LDLM_CB_CANCELING. Used to distinguish
 711  *             cancellation and blocking ast's.
 712  *
 713  * Possible use cases:
 714  *
 715  *     - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) to cancel
 716  *       lock due to lock lru pressure, or explicit user request to purge
 717  *       locks.
 718  *
 719  *     - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_BLOCKING) to notify
 720  *       us that dlmlock conflicts with another lock that some client is
 721  *       enqueing. Lock is canceled.
 722  *
 723  *           - cl_lock_cancel() is called. osc_lock_cancel() calls
 724  *             ldlm_cli_cancel() that calls
 725  *
 726  *                  dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
 727  *
 728  *             recursively entering osc_ldlm_blocking_ast().
 729  *
 730  *     - client cancels lock voluntary (e.g., as a part of early cancellation):
 731  *
 732  *           cl_lock_cancel()->
 733  *             osc_lock_cancel()->
 734  *               ldlm_cli_cancel()->
 735  *                 dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
 736  *
 737  */
 738 static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
 739                                  struct ldlm_lock_desc *new, void *data,
 740                                  int flag)
 741 {
 742         struct lu_env     *env;
 743         struct cl_env_nest nest;
 744         int                result;
 745
 746         /*
 747          * This can be called in the context of outer IO, e.g.,
 748          *
 749          *     cl_enqueue()->...
 750          *       ->osc_enqueue_base()->...
 751          *         ->ldlm_prep_elc_req()->...
 752          *           ->ldlm_cancel_callback()->...
 753          *             ->osc_ldlm_blocking_ast()
 754          *
 755          * new environment has to be created to not corrupt outer context.
 756          */
 757         env = cl_env_nested_get(&nest);
 758         if (!IS_ERR(env)) {
 759                 result = osc_dlm_blocking_ast0(env, dlmlock, data, flag);
 760                 cl_env_nested_put(&nest, env);
 761         } else {
 762                 result = PTR_ERR(env);
 763                 /*
 764                  * XXX This should never happen, as cl_lock is
 765                  * stuck. Pre-allocated environment a la vvp_inode_fini_env
 766                  * should be used.
 767                  */
 768                 LBUG();
 769         }
 770         if (result != 0) {
 771                 if (result == -ENODATA)
 772                         result = 0;
 773                 else
 774                         CERROR("BAST failed: %d\n", result);
 775         }
 776         return result;
 777 }
 778
 779 static int osc_ldlm_completion_ast(struct ldlm_lock *dlmlock,
 780                                    int flags, void *data)
 781 {
 782         struct cl_env_nest nest;
 783         struct lu_env     *env;
 784         struct osc_lock   *olck;
 785         struct cl_lock    *lock;
 786         int result;
 787         int dlmrc;
 788
 789         /* first, do dlm part of the work */
 790         dlmrc = ldlm_completion_ast_async(dlmlock, flags, data);
 791         /* then, notify cl_lock */
 792         env = cl_env_nested_get(&nest);
 793         if (!IS_ERR(env)) {
 794                 olck = osc_ast_data_get(dlmlock);
 795                 if (olck != NULL) {
 796                         lock = olck->ols_cl.cls_lock;
 797                         cl_lock_mutex_get(env, lock);
 798                         /*
 799                          * ldlm_handle_cp_callback() copied LVB from request
 800                          * to lock->l_lvb_data, store it in osc_lock.
 801                          */
 802                         LASSERT(dlmlock->l_lvb_data != NULL);
 803                         lock_res_and_lock(dlmlock);
 804                         olck->ols_lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
 805                         if (olck->ols_lock == NULL) {
 806                                 /*
 807                                  * upcall (osc_lock_upcall()) hasn't yet been
 808                                  * called. Do nothing now, upcall will bind
 809                                  * olck to dlmlock and signal the waiters.
 810                                  *
 811                                  * This maintains an invariant that osc_lock
 812                                  * and ldlm_lock are always bound when
 813                                  * osc_lock is in OLS_GRANTED state.
 814                                  */
 815                         } else if (dlmlock->l_granted_mode ==
 816                                    dlmlock->l_req_mode) {
 817                                 osc_lock_granted(env, olck, dlmlock, dlmrc);
 818                         }
 819                         unlock_res_and_lock(dlmlock);
 820
 821                         if (dlmrc != 0) {
 822                                 CL_LOCK_DEBUG(D_ERROR, env, lock,
 823                                               "dlmlock returned %d\n", dlmrc);
 824                                 cl_lock_error(env, lock, dlmrc);
 825                         }
 826                         cl_lock_mutex_put(env, lock);
 827                         osc_ast_data_put(env, olck);
 828                         result = 0;
 829                 } else
 830                         result = -ELDLM_NO_LOCK_DATA;
 831                 cl_env_nested_put(&nest, env);
 832         } else
 833                 result = PTR_ERR(env);
 834         return dlmrc ?: result;
 835 }
 836
 837 static int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
 838 {
 839         struct ptlrpc_request  *req  = data;
 840         struct osc_lock        *olck;
 841         struct cl_lock         *lock;
 842         struct cl_object       *obj;
 843         struct cl_env_nest      nest;
 844         struct lu_env          *env;
 845         struct ost_lvb         *lvb;
 846         struct req_capsule     *cap;
 847         int                     result;
 848
 849         LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK);
 850
 851         env = cl_env_nested_get(&nest);
 852         if (!IS_ERR(env)) {
 853                 /*
 854                  * osc_ast_data_get() has to go after environment is
 855                  * allocated, because osc_ast_data() acquires a
 856                  * reference to a lock, and it can only be released in
 857                  * environment.
 858                  */
 859                 olck = osc_ast_data_get(dlmlock);
 860                 if (olck != NULL) {
 861                         lock = olck->ols_cl.cls_lock;
 862                         cl_lock_mutex_get(env, lock);
 863                         cap = &req->rq_pill;
 864                         req_capsule_extend(cap, &RQF_LDLM_GL_CALLBACK);
 865                         req_capsule_set_size(cap, &RMF_DLM_LVB, RCL_SERVER,
 866                                              sizeof *lvb);
 867                         result = req_capsule_server_pack(cap);
 868                         if (result == 0) {
 869                                 lvb = req_capsule_server_get(cap, &RMF_DLM_LVB);
 870                                 obj = lock->cll_descr.cld_obj;
 871                                 result = cl_object_glimpse(env, obj, lvb);
 872                         }
 873                         cl_lock_mutex_put(env, lock);
 874                         osc_ast_data_put(env, olck);
 875                 } else {
 876                         /*
 877                          * These errors are normal races, so we don't want to
 878                          * fill the console with messages by calling
 879                          * ptlrpc_error()
 880                          */
 881                         lustre_pack_reply(req, 1, NULL, NULL);
 882                         result = -ELDLM_NO_LOCK_DATA;
 883                 }
 884                 cl_env_nested_put(&nest, env);
 885         } else
 886                 result = PTR_ERR(env);
 887         req->rq_status = result;
 888         return result;
 889 }
 890
 891 static unsigned long osc_lock_weigh(const struct lu_env *env,
 892                                     const struct cl_lock_slice *slice)
 893 {
 894         /*
 895          * don't need to grab coh_page_guard since we don't care the exact #
 896          * of pages..
 897          */
 898         return cl_object_header(slice->cls_obj)->coh_pages;
 899 }
 900
 901 /**
 902  * Get the weight of dlm lock for early cancellation.
 903  *
 904  * XXX: it should return the pages covered by this \a dlmlock.
 905  */
 906 static unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock)
 907 {
 908         struct cl_env_nest       nest;
 909         struct lu_env           *env;
 910         struct osc_lock         *lock;
 911         struct cl_lock          *cll;
 912         unsigned long            weight;
 913         ENTRY;
 914
 915         cfs_might_sleep();
 916         /*
 917          * osc_ldlm_weigh_ast has a complex context since it might be called
 918          * because of lock canceling, or from user's input. We have to make
 919          * a new environment for it. Probably it is implementation safe to use
 920          * the upper context because cl_lock_put don't modify environment
 921          * variables. But in case of ..
 922          */
 923         env = cl_env_nested_get(&nest);
 924         if (IS_ERR(env))
 925                 /* Mostly because lack of memory, tend to eliminate this lock*/
 926                 RETURN(0);
 927
 928         LASSERT(dlmlock->l_resource->lr_type == LDLM_EXTENT);
 929         lock = osc_ast_data_get(dlmlock);
 930         if (lock == NULL) {
 931                 /* cl_lock was destroyed because of memory pressure.
 932                  * It is much reasonable to assign this type of lock
 933                  * a lower cost.
 934                  */
 935                 GOTO(out, weight = 0);
 936         }
 937
 938         cll = lock->ols_cl.cls_lock;
 939         cl_lock_mutex_get(env, cll);
 940         weight = cl_lock_weigh(env, cll);
 941         cl_lock_mutex_put(env, cll);
 942         osc_ast_data_put(env, lock);
 943         EXIT;
 944
 945 out:
 946         cl_env_nested_put(&nest, env);
 947         return weight;
 948 }
 949
 950 static void osc_lock_build_einfo(const struct lu_env *env,
 951                                  const struct cl_lock *clock,
 952                                  struct osc_lock *lock,
 953                                  struct ldlm_enqueue_info *einfo)
 954 {
 955         enum cl_lock_mode mode;
 956
 957         mode = clock->cll_descr.cld_mode;
 958         if (mode == CLM_PHANTOM)
 959                 /*
 960                  * For now, enqueue all glimpse locks in read mode. In the
 961                  * future, client might choose to enqueue LCK_PW lock for
 962                  * glimpse on a file opened for write.
 963                  */
 964                 mode = CLM_READ;
 965
 966         einfo->ei_type   = LDLM_EXTENT;
 967         einfo->ei_mode   = osc_cl_lock2ldlm(mode);
 968         einfo->ei_cb_bl  = osc_ldlm_blocking_ast;
 969         einfo->ei_cb_cp  = osc_ldlm_completion_ast;
 970         einfo->ei_cb_gl  = osc_ldlm_glimpse_ast;
 971         einfo->ei_cb_wg  = osc_ldlm_weigh_ast;
 972         einfo->ei_cbdata = lock; /* value to be put into ->l_ast_data */
 973 }
 974
 975 /**
 976  * Determine if the lock should be converted into a lockless lock.
 977  *
 978  * Steps to check:
 979  * - if the lock has an explicite requirment for a non-lockless lock;
 980  * - if the io lock request type ci_lockreq;
 981  * - send the enqueue rpc to ost to make the further decision;
 982  * - special treat to truncate lockless lock
 983  *
 984  *  Additional policy can be implemented here, e.g., never do lockless-io
 985  *  for large extents.
 986  */
 987 static void osc_lock_to_lockless(const struct lu_env *env,
 988                                  struct osc_lock *ols, int force)
 989 {
 990         struct cl_lock_slice *slice = &ols->ols_cl;
 991         struct cl_lock *lock        = slice->cls_lock;
 992
 993         LASSERT(ols->ols_state == OLS_NEW ||
 994                 ols->ols_state == OLS_UPCALL_RECEIVED);
 995
 996         if (force) {
 997                 ols->ols_locklessable = 1;
 998                 LASSERT(cl_lock_is_mutexed(lock));
 999                 slice->cls_ops = &osc_lock_lockless_ops;
1000         } else {
1001                 struct osc_io *oio     = osc_env_io(env);
1002                 struct cl_io  *io      = oio->oi_cl.cis_io;
1003                 struct cl_object *obj  = slice->cls_obj;
1004                 struct osc_object *oob = cl2osc(obj);
1005                 const struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
1006                 struct obd_connect_data *ocd;
1007
1008                 LASSERT(io->ci_lockreq == CILR_MANDATORY ||
1009                         io->ci_lockreq == CILR_MAYBE ||
1010                         io->ci_lockreq == CILR_NEVER);
1011
1012                 ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data;
1013                 ols->ols_locklessable = (io->ci_type != CIT_SETATTR) &&
1014                                 (io->ci_lockreq == CILR_MAYBE) &&
1015                                 (ocd->ocd_connect_flags & OBD_CONNECT_SRVLOCK);
1016                 if (io->ci_lockreq == CILR_NEVER ||
1017                         /* lockless IO */
1018                     (ols->ols_locklessable && osc_object_is_contended(oob)) ||
1019                         /* lockless truncate */
1020                     (cl_io_is_trunc(io) &&
1021                      (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK) &&
1022                       osd->od_lockless_truncate)) {
1023                         ols->ols_locklessable = 1;
1024                         slice->cls_ops = &osc_lock_lockless_ops;
1025                 }
1026         }
1027         LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
1028 }
1029
1030 static int osc_lock_compatible(const struct osc_lock *qing,
1031                                const struct osc_lock *qed)
1032 {
1033         enum cl_lock_mode qing_mode;
1034         enum cl_lock_mode qed_mode;
1035
1036         qing_mode = qing->ols_cl.cls_lock->cll_descr.cld_mode;
1037         if (qed->ols_glimpse &&
1038             (qed->ols_state >= OLS_UPCALL_RECEIVED || qing_mode == CLM_READ))
1039                 return 1;
1040
1041         qed_mode = qed->ols_cl.cls_lock->cll_descr.cld_mode;
1042         return ((qing_mode == CLM_READ) && (qed_mode == CLM_READ));
1043 }
1044
1045 /**
1046  * Cancel all conflicting locks and wait for them to be destroyed.
1047  *
1048  * This function is used for two purposes:
1049  *
1050  *     - early cancel all conflicting locks before starting IO, and
1051  *
1052  *     - guarantee that pages added to the page cache by lockless IO are never
1053  *       covered by locks other than lockless IO lock, and, hence, are not
1054  *       visible to other threads.
1055  */
1056 static int osc_lock_enqueue_wait(const struct lu_env *env,
1057                                  const struct osc_lock *olck)
1058 {
1059         struct cl_lock          *lock    = olck->ols_cl.cls_lock;
1060         struct cl_lock_descr    *descr   = &lock->cll_descr;
1061         struct cl_object_header *hdr     = cl_object_header(descr->cld_obj);
1062         struct cl_lock          *scan;
1063         struct cl_lock          *conflict= NULL;
1064         int lockless                     = osc_lock_is_lockless(olck);
1065         int rc                           = 0;
1066         ENTRY;
1067
1068         LASSERT(cl_lock_is_mutexed(lock));
1069
1070         /* make it enqueue anyway for glimpse lock, because we actually
1071          * don't need to cancel any conflicting locks. */
1072         if (olck->ols_glimpse)
1073                 return 0;
1074
1075         cfs_spin_lock(&hdr->coh_lock_guard);
1076         cfs_list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) {
1077                 struct cl_lock_descr *cld = &scan->cll_descr;
1078                 const struct osc_lock *scan_ols;
1079
1080                 if (scan == lock)
1081                         break;
1082
1083                 if (scan->cll_state < CLS_QUEUING ||
1084                     scan->cll_state == CLS_FREEING ||
1085                     cld->cld_start > descr->cld_end ||
1086                     cld->cld_end < descr->cld_start)
1087                         continue;
1088
1089                 /* overlapped and living locks. */
1090
1091                 /* We're not supposed to give up group lock. */
1092                 if (scan->cll_descr.cld_mode == CLM_GROUP) {
1093                         LASSERT(descr->cld_mode != CLM_GROUP ||
1094                                 descr->cld_gid != scan->cll_descr.cld_gid);
1095                         continue;
1096                 }
1097
1098                 scan_ols = osc_lock_at(scan);
1099
1100                 /* We need to cancel the compatible locks if we're enqueuing
1101                  * a lockless lock, for example:
1102                  * imagine that client has PR lock on [0, 1000], and thread T0
1103                  * is doing lockless IO in [500, 1500] region. Concurrent
1104                  * thread T1 can see lockless data in [500, 1000], which is
1105                  * wrong, because these data are possibly stale. */
1106                 if (!lockless && osc_lock_compatible(olck, scan_ols))
1107                         continue;
1108
1109                 /* Now @scan is conflicting with @lock, this means current
1110                  * thread have to sleep for @scan being destroyed. */
1111                 if (scan_ols->ols_owner == osc_env_io(env)) {
1112                         CERROR("DEADLOCK POSSIBLE!\n");
1113                         CL_LOCK_DEBUG(D_ERROR, env, scan, "queued.\n");
1114                         CL_LOCK_DEBUG(D_ERROR, env, lock, "queuing.\n");
1115                         libcfs_debug_dumpstack(NULL);
1116                 }
1117                 cl_lock_get_trust(scan);
1118                 conflict = scan;
1119                 break;
1120         }
1121         cfs_spin_unlock(&hdr->coh_lock_guard);
1122
1123         if (conflict) {
1124                 if (lock->cll_descr.cld_mode == CLM_GROUP) {
1125                         /* we want a group lock but a previous lock request
1126                          * conflicts, we do not wait but return 0 so the
1127                          * request is send to the server
1128                          */
1129                         CDEBUG(D_DLMTRACE, "group lock %p is conflicted "
1130                                            "with %p, no wait, send to server\n",
1131                                lock, conflict);
1132                         cl_lock_put(env, conflict);
1133                         rc = 0;
1134                 } else {
1135                         CDEBUG(D_DLMTRACE, "lock %p is conflicted with %p, "
1136                                            "will wait\n",
1137                                lock, conflict);
1138                         LASSERT(lock->cll_conflict == NULL);
1139                         lu_ref_add(&conflict->cll_reference, "cancel-wait",
1140                                    lock);
1141                         lock->cll_conflict = conflict;
1142                         rc = CLO_WAIT;
1143                 }
1144         }
1145         RETURN(rc);
1146 }
1147
1148 /**
1149  * Implementation of cl_lock_operations::clo_enqueue() method for osc
1150  * layer. This initiates ldlm enqueue:
1151  *
1152  *     - cancels conflicting locks early (osc_lock_enqueue_wait());
1153  *
1154  *     - calls osc_enqueue_base() to do actual enqueue.
1155  *
1156  * osc_enqueue_base() is supplied with an upcall function that is executed
1157  * when lock is received either after a local cached ldlm lock is matched, or
1158  * when a reply from the server is received.
1159  *
1160  * This function does not wait for the network communication to complete.
1161  */
1162 static int osc_lock_enqueue(const struct lu_env *env,
1163                             const struct cl_lock_slice *slice,
1164                             struct cl_io *unused, __u32 enqflags)
1165 {
1166         struct osc_lock          *ols     = cl2osc_lock(slice);
1167         struct cl_lock           *lock    = ols->ols_cl.cls_lock;
1168         int result;
1169         ENTRY;
1170
1171         LASSERT(cl_lock_is_mutexed(lock));
1172         LASSERTF(ols->ols_state == OLS_NEW,
1173                  "Impossible state: %d\n", ols->ols_state);
1174
1175         ols->ols_flags = osc_enq2ldlm_flags(enqflags);
1176         if (enqflags & CEF_AGL) {
1177                 ols->ols_flags |= LDLM_FL_BLOCK_NOWAIT;
1178                 ols->ols_agl = 1;
1179         }
1180         if (ols->ols_flags & LDLM_FL_HAS_INTENT)
1181                 ols->ols_glimpse = 1;
1182         if (!osc_lock_is_lockless(ols) && !(enqflags & CEF_MUST))
1183                 /* try to convert this lock to a lockless lock */
1184                 osc_lock_to_lockless(env, ols, (enqflags & CEF_NEVER));
1185
1186         result = osc_lock_enqueue_wait(env, ols);
1187         if (result == 0) {
1188                 if (!osc_lock_is_lockless(ols)) {
1189                         struct osc_object        *obj = cl2osc(slice->cls_obj);
1190                         struct osc_thread_info   *info = osc_env_info(env);
1191                         struct ldlm_res_id       *resname = &info->oti_resname;
1192                         ldlm_policy_data_t       *policy = &info->oti_policy;
1193                         struct ldlm_enqueue_info *einfo = &ols->ols_einfo;
1194
1195                         if (ols->ols_locklessable)
1196                                 ols->ols_flags |= LDLM_FL_DENY_ON_CONTENTION;
1197
1198                         /* a reference for lock, passed as an upcall cookie */
1199                         cl_lock_get(lock);
1200                         lu_ref_add(&lock->cll_reference, "upcall", lock);
1201                         /* a user for lock also */
1202                         cl_lock_user_add(env, lock);
1203                         ols->ols_state = OLS_ENQUEUED;
1204
1205                         /*
1206                          * XXX: this is possible blocking point as
1207                          * ldlm_lock_match(LDLM_FL_LVB_READY) waits for
1208                          * LDLM_CP_CALLBACK.
1209                          */
1210                         osc_lock_build_res(env, obj, resname);
1211                         osc_lock_build_policy(env, lock, policy);
1212                         result = osc_enqueue_base(osc_export(obj), resname,
1213                                           &ols->ols_flags, policy,
1214                                           &ols->ols_lvb,
1215                                           obj->oo_oinfo->loi_kms_valid,
1216                                           osc_lock_upcall,
1217                                           ols, einfo, &ols->ols_handle,
1218                                           PTLRPCD_SET, 1, ols->ols_agl);
1219                         if (result != 0) {
1220                                 cl_lock_user_del(env, lock);
1221                                 lu_ref_del(&lock->cll_reference,
1222                                            "upcall", lock);
1223                                 cl_lock_put(env, lock);
1224                                 if (unlikely(result == -ECANCELED)) {
1225                                         ols->ols_state = OLS_NEW;
1226                                         result = 0;
1227                                 }
1228                         }
1229                 } else {
1230                         ols->ols_state = OLS_GRANTED;
1231                         ols->ols_owner = osc_env_io(env);
1232                 }
1233         }
1234         LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
1235         RETURN(result);
1236 }
1237
1238 static int osc_lock_wait(const struct lu_env *env,
1239                          const struct cl_lock_slice *slice)
1240 {
1241         struct osc_lock *olck = cl2osc_lock(slice);
1242         struct cl_lock  *lock = olck->ols_cl.cls_lock;
1243
1244         LINVRNT(osc_lock_invariant(olck));
1245
1246         if (olck->ols_glimpse && olck->ols_state >= OLS_UPCALL_RECEIVED) {
1247                 if (olck->ols_flags & LDLM_FL_LVB_READY) {
1248                         return 0;
1249                 } else if (olck->ols_agl) {
1250                         olck->ols_state = OLS_NEW;
1251                 } else {
1252                         LASSERT(lock->cll_error);
1253                         return lock->cll_error;
1254                 }
1255         }
1256
1257         if (olck->ols_state == OLS_NEW) {
1258                 if (lock->cll_descr.cld_enq_flags & CEF_NO_REENQUEUE) {
1259                         return -ENAVAIL;
1260                 } else {
1261                         int rc;
1262
1263                         LASSERT(olck->ols_agl);
1264
1265                         rc = osc_lock_enqueue(env, slice, NULL, CEF_ASYNC |
1266                                                                 CEF_MUST);
1267                         if (rc != 0)
1268                                 return rc;
1269                         else
1270                                 return CLO_REENQUEUED;
1271                 }
1272         }
1273
1274         LASSERT(equi(olck->ols_state >= OLS_UPCALL_RECEIVED &&
1275                      lock->cll_error == 0, olck->ols_lock != NULL));
1276
1277         return lock->cll_error ?: olck->ols_state >= OLS_GRANTED ? 0 : CLO_WAIT;
1278 }
1279
1280 /**
1281  * An implementation of cl_lock_operations::clo_use() method that pins cached
1282  * lock.
1283  */
1284 static int osc_lock_use(const struct lu_env *env,
1285                         const struct cl_lock_slice *slice)
1286 {
1287         struct osc_lock *olck = cl2osc_lock(slice);
1288         int rc;
1289
1290         LASSERT(!olck->ols_hold);
1291
1292         /*
1293          * Atomically check for LDLM_FL_CBPENDING and addref a lock if this
1294          * flag is not set. This protects us from a concurrent blocking ast.
1295          */
1296         rc = ldlm_lock_addref_try(&olck->ols_handle, olck->ols_einfo.ei_mode);
1297         if (rc == 0) {
1298                 olck->ols_hold = 1;
1299                 olck->ols_state = OLS_GRANTED;
1300         } else {
1301                 struct cl_lock *lock;
1302
1303                 /*
1304                  * Lock is being cancelled somewhere within
1305                  * ldlm_handle_bl_callback(): LDLM_FL_CBPENDING is already
1306                  * set, but osc_ldlm_blocking_ast() hasn't yet acquired
1307                  * cl_lock mutex.
1308                  */
1309                 lock = slice->cls_lock;
1310                 LASSERT(lock->cll_state == CLS_INTRANSIT);
1311                 LASSERT(lock->cll_users > 0);
1312                 /* set a flag for osc_dlm_blocking_ast0() to signal the
1313                  * lock.*/
1314                 olck->ols_ast_wait = 1;
1315                 rc = CLO_WAIT;
1316         }
1317         return rc;
1318 }
1319
1320 static int osc_lock_flush(struct osc_lock *ols, int discard)
1321 {
1322         struct cl_lock       *lock  = ols->ols_cl.cls_lock;
1323         struct cl_env_nest    nest;
1324         struct lu_env        *env;
1325         int result = 0;
1326
1327         env = cl_env_nested_get(&nest);
1328         if (!IS_ERR(env)) {
1329                 result = cl_lock_page_out(env, lock, discard);
1330                 cl_env_nested_put(&nest, env);
1331         } else
1332                 result = PTR_ERR(env);
1333         if (result == 0) {
1334                 ols->ols_flush = 1;
1335                 LINVRNT(!osc_lock_has_pages(ols));
1336         }
1337         return result;
1338 }
1339
1340 /**
1341  * Implements cl_lock_operations::clo_cancel() method for osc layer. This is
1342  * called (as part of cl_lock_cancel()) when lock is canceled either voluntary
1343  * (LRU pressure, early cancellation, umount, etc.) or due to the conflict
1344  * with some other lock some where in the cluster. This function does the
1345  * following:
1346  *
1347  *     - invalidates all pages protected by this lock (after sending dirty
1348  *       ones to the server, as necessary);
1349  *
1350  *     - decref's underlying ldlm lock;
1351  *
1352  *     - cancels ldlm lock (ldlm_cli_cancel()).
1353  */
1354 static void osc_lock_cancel(const struct lu_env *env,
1355                             const struct cl_lock_slice *slice)
1356 {
1357         struct cl_lock   *lock    = slice->cls_lock;
1358         struct osc_lock  *olck    = cl2osc_lock(slice);
1359         struct ldlm_lock *dlmlock = olck->ols_lock;
1360         int               result  = 0;
1361         int               discard;
1362
1363         LASSERT(cl_lock_is_mutexed(lock));
1364         LINVRNT(osc_lock_invariant(olck));
1365
1366         if (dlmlock != NULL) {
1367                 int do_cancel;
1368
1369                 discard = !!(dlmlock->l_flags & LDLM_FL_DISCARD_DATA);
1370                 result = osc_lock_flush(olck, discard);
1371                 osc_lock_unhold(olck);
1372
1373                 lock_res_and_lock(dlmlock);
1374                 /* Now that we're the only user of dlm read/write reference,
1375                  * mostly the ->l_readers + ->l_writers should be zero.
1376                  * However, there is a corner case.
1377                  * See bug 18829 for details.*/
1378                 do_cancel = (dlmlock->l_readers == 0 &&
1379                              dlmlock->l_writers == 0);
1380                 dlmlock->l_flags |= LDLM_FL_CBPENDING;
1381                 unlock_res_and_lock(dlmlock);
1382                 if (do_cancel)
1383                         result = ldlm_cli_cancel(&olck->ols_handle);
1384                 if (result < 0)
1385                         CL_LOCK_DEBUG(D_ERROR, env, lock,
1386                                       "lock %p cancel failure with error(%d)\n",
1387                                       lock, result);
1388         }
1389         olck->ols_state = OLS_CANCELLED;
1390         olck->ols_flags &= ~LDLM_FL_LVB_READY;
1391         osc_lock_detach(env, olck);
1392 }
1393
1394 #ifdef INVARIANT_CHECK
1395 static int check_cb(const struct lu_env *env, struct cl_io *io,
1396                     struct cl_page *page, void *cbdata)
1397 {
1398         struct cl_lock *lock = cbdata;
1399
1400         if (lock->cll_descr.cld_mode == CLM_READ) {
1401                 struct cl_lock *tmp;
1402                 tmp = cl_lock_at_page(env, lock->cll_descr.cld_obj,
1403                                      page, lock, 1, 0);
1404                 if (tmp != NULL) {
1405                         cl_lock_put(env, tmp);
1406                         return CLP_GANG_OKAY;
1407                 }
1408         }
1409
1410         CL_LOCK_DEBUG(D_ERROR, env, lock, "still has pages\n");
1411         CL_PAGE_DEBUG(D_ERROR, env, page, "\n");
1412         return CLP_GANG_ABORT;
1413 }
1414
1415 /**
1416  * Returns true iff there are pages under \a olck not protected by other
1417  * locks.
1418  */
1419 static int osc_lock_has_pages(struct osc_lock *olck)
1420 {
1421         struct cl_lock       *lock;
1422         struct cl_lock_descr *descr;
1423         struct cl_object     *obj;
1424         struct osc_object    *oob;
1425         struct cl_env_nest    nest;
1426         struct cl_io         *io;
1427         struct lu_env        *env;
1428         int                   result;
1429
1430         env = cl_env_nested_get(&nest);
1431         if (IS_ERR(env))
1432                 return 0;
1433
1434         obj   = olck->ols_cl.cls_obj;
1435         oob   = cl2osc(obj);
1436         io    = &oob->oo_debug_io;
1437         lock  = olck->ols_cl.cls_lock;
1438         descr = &lock->cll_descr;
1439
1440         cfs_mutex_lock(&oob->oo_debug_mutex);
1441
1442         io->ci_obj = cl_object_top(obj);
1443         cl_io_init(env, io, CIT_MISC, io->ci_obj);
1444         do {
1445                 result = cl_page_gang_lookup(env, obj, io,
1446                                              descr->cld_start, descr->cld_end,
1447                                              check_cb, (void *)lock);
1448                 if (result == CLP_GANG_ABORT)
1449                         break;
1450                 if (result == CLP_GANG_RESCHED)
1451                         cfs_cond_resched();
1452         } while (result != CLP_GANG_OKAY);
1453         cl_io_fini(env, io);
1454         cfs_mutex_unlock(&oob->oo_debug_mutex);
1455         cl_env_nested_put(&nest, env);
1456
1457         return (result == CLP_GANG_ABORT);
1458 }
1459 #else
1460 static int osc_lock_has_pages(struct osc_lock *olck)
1461 {
1462         return 0;
1463 }
1464 #endif /* INVARIANT_CHECK */
1465
1466 static void osc_lock_delete(const struct lu_env *env,
1467                             const struct cl_lock_slice *slice)
1468 {
1469         struct osc_lock *olck;
1470
1471         olck = cl2osc_lock(slice);
1472         if (olck->ols_glimpse) {
1473                 LASSERT(!olck->ols_hold);
1474                 LASSERT(!olck->ols_lock);
1475                 return;
1476         }
1477
1478         LINVRNT(osc_lock_invariant(olck));
1479         LINVRNT(!osc_lock_has_pages(olck));
1480
1481         osc_lock_unhold(olck);
1482         osc_lock_detach(env, olck);
1483 }
1484
1485 /**
1486  * Implements cl_lock_operations::clo_state() method for osc layer.
1487  *
1488  * Maintains osc_lock::ols_owner field.
1489  *
1490  * This assumes that lock always enters CLS_HELD (from some other state) in
1491  * the same IO context as one that requested the lock. This should not be a
1492  * problem, because context is by definition shared by all activity pertaining
1493  * to the same high-level IO.
1494  */
1495 static void osc_lock_state(const struct lu_env *env,
1496                            const struct cl_lock_slice *slice,
1497                            enum cl_lock_state state)
1498 {
1499         struct osc_lock *lock = cl2osc_lock(slice);
1500
1501         /*
1502          * XXX multiple io contexts can use the lock at the same time.
1503          */
1504         LINVRNT(osc_lock_invariant(lock));
1505         if (state == CLS_HELD && slice->cls_lock->cll_state != CLS_HELD) {
1506                 struct osc_io *oio = osc_env_io(env);
1507
1508                 LASSERT(lock->ols_owner == NULL);
1509                 lock->ols_owner = oio;
1510         } else if (state != CLS_HELD)
1511                 lock->ols_owner = NULL;
1512 }
1513
1514 static int osc_lock_print(const struct lu_env *env, void *cookie,
1515                           lu_printer_t p, const struct cl_lock_slice *slice)
1516 {
1517         struct osc_lock *lock = cl2osc_lock(slice);
1518
1519         /*
1520          * XXX print ldlm lock and einfo properly.
1521          */
1522         (*p)(env, cookie, "%p %08x "LPX64" %d %p ",
1523              lock->ols_lock, lock->ols_flags, lock->ols_handle.cookie,
1524              lock->ols_state, lock->ols_owner);
1525         osc_lvb_print(env, cookie, p, &lock->ols_lvb);
1526         return 0;
1527 }
1528
1529 static int osc_lock_fits_into(const struct lu_env *env,
1530                               const struct cl_lock_slice *slice,
1531                               const struct cl_lock_descr *need,
1532                               const struct cl_io *io)
1533 {
1534         struct osc_lock *ols = cl2osc_lock(slice);
1535
1536         if (need->cld_enq_flags & CEF_NEVER)
1537                 return 0;
1538
1539         if (need->cld_mode == CLM_PHANTOM) {
1540                 if (ols->ols_agl)
1541                         return !(ols->ols_state > OLS_RELEASED);
1542
1543                 /*
1544                  * Note: the QUEUED lock can't be matched here, otherwise
1545                  * it might cause the deadlocks.
1546                  * In read_process,
1547                  * P1: enqueued read lock, create sublock1
1548                  * P2: enqueued write lock, create sublock2(conflicted
1549                  *     with sublock1).
1550                  * P1: Grant read lock.
1551                  * P1: enqueued glimpse lock(with holding sublock1_read),
1552                  *     matched with sublock2, waiting sublock2 to be granted.
1553                  *     But sublock2 can not be granted, because P1
1554                  *     will not release sublock1. Bang!
1555                  */
1556                 if (ols->ols_state < OLS_GRANTED ||
1557                     ols->ols_state > OLS_RELEASED)
1558                         return 0;
1559         } else if (need->cld_enq_flags & CEF_MUST) {
1560                 /*
1561                  * If the lock hasn't ever enqueued, it can't be matched
1562                  * because enqueue process brings in many information
1563                  * which can be used to determine things such as lockless,
1564                  * CEF_MUST, etc.
1565                  */
1566                 if (ols->ols_state < OLS_UPCALL_RECEIVED &&
1567                     ols->ols_locklessable)
1568                         return 0;
1569         }
1570         return 1;
1571 }
1572
1573 static const struct cl_lock_operations osc_lock_ops = {
1574         .clo_fini    = osc_lock_fini,
1575         .clo_enqueue = osc_lock_enqueue,
1576         .clo_wait    = osc_lock_wait,
1577         .clo_unuse   = osc_lock_unuse,
1578         .clo_use     = osc_lock_use,
1579         .clo_delete  = osc_lock_delete,
1580         .clo_state   = osc_lock_state,
1581         .clo_cancel  = osc_lock_cancel,
1582         .clo_weigh   = osc_lock_weigh,
1583         .clo_print   = osc_lock_print,
1584         .clo_fits_into = osc_lock_fits_into,
1585 };
1586
1587 static int osc_lock_lockless_unuse(const struct lu_env *env,
1588                                    const struct cl_lock_slice *slice)
1589 {
1590         struct osc_lock *ols = cl2osc_lock(slice);
1591         struct cl_lock *lock = slice->cls_lock;
1592
1593         LASSERT(ols->ols_state == OLS_GRANTED);
1594         LINVRNT(osc_lock_invariant(ols));
1595
1596         cl_lock_cancel(env, lock);
1597         cl_lock_delete(env, lock);
1598         return 0;
1599 }
1600
1601 static void osc_lock_lockless_cancel(const struct lu_env *env,
1602                                      const struct cl_lock_slice *slice)
1603 {
1604         struct osc_lock   *ols  = cl2osc_lock(slice);
1605         int result;
1606
1607         result = osc_lock_flush(ols, 0);
1608         if (result)
1609                 CERROR("Pages for lockless lock %p were not purged(%d)\n",
1610                        ols, result);
1611         ols->ols_state = OLS_CANCELLED;
1612 }
1613
1614 static int osc_lock_lockless_wait(const struct lu_env *env,
1615                                   const struct cl_lock_slice *slice)
1616 {
1617         struct osc_lock *olck = cl2osc_lock(slice);
1618         struct cl_lock  *lock = olck->ols_cl.cls_lock;
1619
1620         LINVRNT(osc_lock_invariant(olck));
1621         LASSERT(olck->ols_state >= OLS_UPCALL_RECEIVED);
1622
1623         return lock->cll_error;
1624 }
1625
1626 static void osc_lock_lockless_state(const struct lu_env *env,
1627                                     const struct cl_lock_slice *slice,
1628                                     enum cl_lock_state state)
1629 {
1630         struct osc_lock *lock = cl2osc_lock(slice);
1631
1632         LINVRNT(osc_lock_invariant(lock));
1633         if (state == CLS_HELD) {
1634                 struct osc_io *oio  = osc_env_io(env);
1635
1636                 LASSERT(ergo(lock->ols_owner, lock->ols_owner == oio));
1637                 lock->ols_owner = oio;
1638
1639                 /* set the io to be lockless if this lock is for io's
1640                  * host object */
1641                 if (cl_object_same(oio->oi_cl.cis_obj, slice->cls_obj))
1642                         oio->oi_lockless = 1;
1643         }
1644 }
1645
1646 static int osc_lock_lockless_fits_into(const struct lu_env *env,
1647                                        const struct cl_lock_slice *slice,
1648                                        const struct cl_lock_descr *need,
1649                                        const struct cl_io *io)
1650 {
1651         struct osc_lock *lock = cl2osc_lock(slice);
1652
1653         if (!(need->cld_enq_flags & CEF_NEVER))
1654                 return 0;
1655
1656         /* lockless lock should only be used by its owning io. b22147 */
1657         return (lock->ols_owner == osc_env_io(env));
1658 }
1659
1660 static const struct cl_lock_operations osc_lock_lockless_ops = {
1661         .clo_fini      = osc_lock_fini,
1662         .clo_enqueue   = osc_lock_enqueue,
1663         .clo_wait      = osc_lock_lockless_wait,
1664         .clo_unuse     = osc_lock_lockless_unuse,
1665         .clo_state     = osc_lock_lockless_state,
1666         .clo_fits_into = osc_lock_lockless_fits_into,
1667         .clo_cancel    = osc_lock_lockless_cancel,
1668         .clo_print     = osc_lock_print
1669 };
1670
1671 int osc_lock_init(const struct lu_env *env,
1672                   struct cl_object *obj, struct cl_lock *lock,
1673                   const struct cl_io *unused)
1674 {
1675         struct osc_lock *clk;
1676         int result;
1677
1678         OBD_SLAB_ALLOC_PTR_GFP(clk, osc_lock_kmem, CFS_ALLOC_IO);
1679         if (clk != NULL) {
1680                 osc_lock_build_einfo(env, lock, clk, &clk->ols_einfo);
1681                 cfs_atomic_set(&clk->ols_pageref, 0);
1682                 clk->ols_state = OLS_NEW;
1683                 cl_lock_slice_add(lock, &clk->ols_cl, obj, &osc_lock_ops);
1684                 result = 0;
1685         } else
1686                 result = -ENOMEM;
1687         return result;
1688 }
1689
1690 int osc_dlm_lock_pageref(struct ldlm_lock *dlm)
1691 {
1692         struct osc_lock *olock;
1693         int              rc = 0;
1694
1695         cfs_spin_lock(&osc_ast_guard);
1696         olock = dlm->l_ast_data;
1697         /*
1698          * there's a very rare race with osc_page_addref_lock(), but that
1699          * doesn't matter because in the worst case we don't cancel a lock
1700          * which we actually can, that's no harm.
1701          */
1702         if (olock != NULL &&
1703             cfs_atomic_add_return(_PAGEREF_MAGIC,
1704                                   &olock->ols_pageref) != _PAGEREF_MAGIC) {
1705                 cfs_atomic_sub(_PAGEREF_MAGIC, &olock->ols_pageref);
1706                 rc = 1;
1707         }
1708         cfs_spin_unlock(&osc_ast_guard);
1709         return rc;
1710 }
1711
1712 /** @} osc */