lustre/osc/osc_lock.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Implementation of cl_lock for OSC layer.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_OSC
  43
  44 #ifdef __KERNEL__
  45 # include <libcfs/libcfs.h>
  46 #else
  47 # include <liblustre.h>
  48 #endif
  49 /* fid_build_reg_res_name() */
  50 #include <lustre_fid.h>
  51
  52 #include "osc_cl_internal.h"
  53
  54 /** \addtogroup osc
  55  *  @{
  56  */
  57
  58 /*****************************************************************************
  59  *
  60  * Type conversions.
  61  *
  62  */
  63
  64 static const struct cl_lock_operations osc_lock_ops;
  65 static const struct cl_lock_operations osc_lock_lockless_ops;
  66 static void osc_lock_to_lockless(const struct lu_env *env,
  67                                  struct osc_lock *ols, int force);
  68 static bool osc_lock_has_pages(struct osc_lock *olck);
  69
  70 int osc_lock_is_lockless(const struct osc_lock *olck)
  71 {
  72         return (olck->ols_cl.cls_ops == &osc_lock_lockless_ops);
  73 }
  74
  75 /**
  76  * Returns a weak pointer to the ldlm lock identified by a handle. Returned
  77  * pointer cannot be dereferenced, as lock is not protected from concurrent
  78  * reclaim. This function is a helper for osc_lock_invariant().
  79  */
  80 static struct ldlm_lock *osc_handle_ptr(struct lustre_handle *handle)
  81 {
  82         struct ldlm_lock *lock;
  83
  84         lock = ldlm_handle2lock(handle);
  85         if (lock != NULL)
  86                 LDLM_LOCK_PUT(lock);
  87         return lock;
  88 }
  89
  90 /**
  91  * Invariant that has to be true all of the time.
  92  */
  93 static int osc_lock_invariant(struct osc_lock *ols)
  94 {
  95         struct ldlm_lock *lock        = osc_handle_ptr(&ols->ols_handle);
  96         struct ldlm_lock *olock       = ols->ols_lock;
  97         int               handle_used = lustre_handle_is_used(&ols->ols_handle);
  98
  99         if (ergo(osc_lock_is_lockless(ols),
 100                  ols->ols_locklessable && ols->ols_lock == NULL))
 101                 return 1;
 102
 103         /*
 104          * If all the following "ergo"s are true, return 1, otherwise 0
 105          */
 106         if (! ergo(olock != NULL, handle_used))
 107                 return 0;
 108
 109         if (! ergo(olock != NULL,
 110                    olock->l_handle.h_cookie == ols->ols_handle.cookie))
 111                 return 0;
 112
 113         if (! ergo(handle_used,
 114                    ergo(lock != NULL && olock != NULL, lock == olock) &&
 115                    ergo(lock == NULL, olock == NULL)))
 116                 return 0;
 117         /*
 118          * Check that ->ols_handle and ->ols_lock are consistent, but
 119          * take into account that they are set at the different time.
 120          */
 121         if (! ergo(ols->ols_state == OLS_CANCELLED,
 122                    olock == NULL && !handle_used))
 123                 return 0;
 124         /*
 125          * DLM lock is destroyed only after we have seen cancellation
 126          * ast.
 127          */
 128         if (! ergo(olock != NULL && ols->ols_state < OLS_CANCELLED,
 129                    !ldlm_is_destroyed(olock)))
 130                 return 0;
 131
 132         if (! ergo(ols->ols_state == OLS_GRANTED,
 133                    olock != NULL &&
 134                    olock->l_req_mode == olock->l_granted_mode &&
 135                    ols->ols_hold))
 136                 return 0;
 137         return 1;
 138 }
 139
 140 /*****************************************************************************
 141  *
 142  * Lock operations.
 143  *
 144  */
 145
 146 /**
 147  * Breaks a link between osc_lock and dlm_lock.
 148  */
 149 static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck)
 150 {
 151         struct ldlm_lock *dlmlock;
 152
 153         spin_lock(&osc_ast_guard);
 154         dlmlock = olck->ols_lock;
 155         if (dlmlock == NULL) {
 156                 spin_unlock(&osc_ast_guard);
 157                 return;
 158         }
 159
 160         olck->ols_lock = NULL;
 161         /* wb(); --- for all who checks (ols->ols_lock != NULL) before
 162          * call to osc_lock_detach() */
 163         dlmlock->l_ast_data = NULL;
 164         olck->ols_handle.cookie = 0ULL;
 165         spin_unlock(&osc_ast_guard);
 166
 167         lock_res_and_lock(dlmlock);
 168         if (dlmlock->l_granted_mode == dlmlock->l_req_mode) {
 169                 struct cl_object *obj = olck->ols_cl.cls_obj;
 170                 struct cl_attr *attr  = &osc_env_info(env)->oti_attr;
 171                 __u64 old_kms;
 172
 173                 cl_object_attr_lock(obj);
 174                 /* Must get the value under the lock to avoid possible races. */
 175                 old_kms = cl2osc(obj)->oo_oinfo->loi_kms;
 176                 /* Update the kms. Need to loop all granted locks.
 177                  * Not a problem for the client */
 178                 attr->cat_kms = ldlm_extent_shift_kms(dlmlock, old_kms);
 179
 180                 cl_object_attr_set(env, obj, attr, CAT_KMS);
 181                 cl_object_attr_unlock(obj);
 182         }
 183         unlock_res_and_lock(dlmlock);
 184
 185         /* release a reference taken in osc_lock_upcall0(). */
 186         LASSERT(olck->ols_has_ref);
 187         lu_ref_del(&dlmlock->l_reference, "osc_lock", olck);
 188         LDLM_LOCK_RELEASE(dlmlock);
 189         olck->ols_has_ref = 0;
 190 }
 191
 192 static int osc_lock_unhold(struct osc_lock *ols)
 193 {
 194         int result = 0;
 195
 196         if (ols->ols_hold) {
 197                 ols->ols_hold = 0;
 198                 result = osc_cancel_base(&ols->ols_handle,
 199                                          ols->ols_einfo.ei_mode);
 200         }
 201         return result;
 202 }
 203
 204 static int osc_lock_unuse(const struct lu_env *env,
 205                           const struct cl_lock_slice *slice)
 206 {
 207         struct osc_lock *ols = cl2osc_lock(slice);
 208
 209         LINVRNT(osc_lock_invariant(ols));
 210
 211         switch (ols->ols_state) {
 212         case OLS_NEW:
 213                 LASSERT(!ols->ols_hold);
 214                 LASSERT(ols->ols_agl);
 215                 return 0;
 216         case OLS_UPCALL_RECEIVED:
 217                 osc_lock_unhold(ols);
 218         case OLS_ENQUEUED:
 219                 LASSERT(!ols->ols_hold);
 220                 osc_lock_detach(env, ols);
 221                 ols->ols_state = OLS_NEW;
 222                 return 0;
 223         case OLS_GRANTED:
 224                 LASSERT(!ols->ols_glimpse);
 225                 LASSERT(ols->ols_hold);
 226                 /*
 227                  * Move lock into OLS_RELEASED state before calling
 228                  * osc_cancel_base() so that possible synchronous cancellation
 229                  * (that always happens e.g., for liblustre) sees that lock is
 230                  * released.
 231                  */
 232                 ols->ols_state = OLS_RELEASED;
 233                 return osc_lock_unhold(ols);
 234         default:
 235                 CERROR("Impossible state: %d\n", ols->ols_state);
 236                 LBUG();
 237         }
 238 }
 239
 240 static void osc_lock_fini(const struct lu_env *env,
 241                           struct cl_lock_slice *slice)
 242 {
 243         struct osc_lock  *ols = cl2osc_lock(slice);
 244
 245         LINVRNT(osc_lock_invariant(ols));
 246         /*
 247          * ->ols_hold can still be true at this point if, for example, a
 248          * thread that requested a lock was killed (and released a reference
 249          * to the lock), before reply from a server was received. In this case
 250          * lock is destroyed immediately after upcall.
 251          */
 252         osc_lock_unhold(ols);
 253         LASSERT(ols->ols_lock == NULL);
 254         OBD_SLAB_FREE_PTR(ols, osc_lock_kmem);
 255 }
 256
 257 static void osc_lock_build_policy(const struct lu_env *env,
 258                                   const struct cl_lock *lock,
 259                                   ldlm_policy_data_t *policy)
 260 {
 261         const struct cl_lock_descr *d = &lock->cll_descr;
 262
 263         osc_index2policy(policy, d->cld_obj, d->cld_start, d->cld_end);
 264         policy->l_extent.gid = d->cld_gid;
 265 }
 266
 267 static __u64 osc_enq2ldlm_flags(__u32 enqflags)
 268 {
 269         __u64 result = 0;
 270
 271         LASSERT((enqflags & ~CEF_MASK) == 0);
 272
 273         if (enqflags & CEF_NONBLOCK)
 274                 result |= LDLM_FL_BLOCK_NOWAIT;
 275         if (enqflags & CEF_ASYNC)
 276                 result |= LDLM_FL_HAS_INTENT;
 277         if (enqflags & CEF_DISCARD_DATA)
 278                 result |= LDLM_FL_AST_DISCARD_DATA;
 279         return result;
 280 }
 281
 282 /**
 283  * Global spin-lock protecting consistency of ldlm_lock::l_ast_data
 284  * pointers. Initialized in osc_init().
 285  */
 286 spinlock_t osc_ast_guard;
 287
 288 static struct osc_lock *osc_ast_data_get(struct ldlm_lock *dlm_lock)
 289 {
 290         struct osc_lock *olck;
 291
 292         lock_res_and_lock(dlm_lock);
 293         spin_lock(&osc_ast_guard);
 294         olck = dlm_lock->l_ast_data;
 295         if (olck != NULL) {
 296                 struct cl_lock *lock = olck->ols_cl.cls_lock;
 297                 /*
 298                  * If osc_lock holds a reference on ldlm lock, return it even
 299                  * when cl_lock is in CLS_FREEING state. This way
 300                  *
 301                  *         osc_ast_data_get(dlmlock) == NULL
 302                  *
 303                  * guarantees that all osc references on dlmlock were
 304                  * released. osc_dlm_blocking_ast0() relies on that.
 305                  */
 306                 if (lock->cll_state < CLS_FREEING || olck->ols_has_ref) {
 307                         cl_lock_get_trust(lock);
 308                         lu_ref_add_atomic(&lock->cll_reference,
 309                                           "ast", current);
 310                 } else
 311                         olck = NULL;
 312         }
 313         spin_unlock(&osc_ast_guard);
 314         unlock_res_and_lock(dlm_lock);
 315         return olck;
 316 }
 317
 318 static void osc_ast_data_put(const struct lu_env *env, struct osc_lock *olck)
 319 {
 320         struct cl_lock *lock;
 321
 322         lock = olck->ols_cl.cls_lock;
 323         lu_ref_del(&lock->cll_reference, "ast", current);
 324         cl_lock_put(env, lock);
 325 }
 326
 327 /**
 328  * Updates object attributes from a lock value block (lvb) received together
 329  * with the DLM lock reply from the server. Copy of osc_update_enqueue()
 330  * logic.
 331  *
 332  * This can be optimized to not update attributes when lock is a result of a
 333  * local match.
 334  *
 335  * Called under lock and resource spin-locks.
 336  */
 337 static void osc_lock_lvb_update(const struct lu_env *env, struct osc_lock *olck,
 338                                 int rc)
 339 {
 340         struct ost_lvb    *lvb;
 341         struct cl_object  *obj;
 342         struct lov_oinfo  *oinfo;
 343         struct cl_attr    *attr;
 344         unsigned           valid;
 345
 346         ENTRY;
 347
 348         if (!(olck->ols_flags & LDLM_FL_LVB_READY))
 349                 RETURN_EXIT;
 350
 351         lvb   = &olck->ols_lvb;
 352         obj   = olck->ols_cl.cls_obj;
 353         oinfo = cl2osc(obj)->oo_oinfo;
 354         attr  = &osc_env_info(env)->oti_attr;
 355         valid = CAT_BLOCKS | CAT_ATIME | CAT_CTIME | CAT_MTIME | CAT_SIZE;
 356         cl_lvb2attr(attr, lvb);
 357
 358         cl_object_attr_lock(obj);
 359         if (rc == 0) {
 360                 struct ldlm_lock  *dlmlock;
 361                 __u64 size;
 362
 363                 dlmlock = olck->ols_lock;
 364                 LASSERT(dlmlock != NULL);
 365
 366                 /* re-grab LVB from a dlm lock under DLM spin-locks. */
 367                 *lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
 368                 size = lvb->lvb_size;
 369                 /* Extend KMS up to the end of this lock and no further
 370                  * A lock on [x,y] means a KMS of up to y + 1 bytes! */
 371                 if (size > dlmlock->l_policy_data.l_extent.end)
 372                         size = dlmlock->l_policy_data.l_extent.end + 1;
 373                 if (size >= oinfo->loi_kms) {
 374                         LDLM_DEBUG(dlmlock, "lock acquired, setting rss="LPU64
 375                                    ", kms="LPU64, lvb->lvb_size, size);
 376                         valid |= CAT_KMS;
 377                         attr->cat_kms = size;
 378                 } else {
 379                         LDLM_DEBUG(dlmlock, "lock acquired, setting rss="
 380                                    LPU64"; leaving kms="LPU64", end="LPU64,
 381                                    lvb->lvb_size, oinfo->loi_kms,
 382                                    dlmlock->l_policy_data.l_extent.end);
 383                 }
 384                 ldlm_lock_allow_match_locked(dlmlock);
 385         } else if (rc == -ENAVAIL && olck->ols_glimpse) {
 386                 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
 387                        " kms="LPU64"\n", lvb->lvb_size, oinfo->loi_kms);
 388         } else
 389                 valid = 0;
 390
 391         if (valid != 0)
 392                 cl_object_attr_set(env, obj, attr, valid);
 393
 394         cl_object_attr_unlock(obj);
 395
 396         EXIT;
 397 }
 398
 399 /**
 400  * Called when a lock is granted, from an upcall (when server returned a
 401  * granted lock), or from completion AST, when server returned a blocked lock.
 402  *
 403  * Called under lock and resource spin-locks, that are released temporarily
 404  * here.
 405  */
 406 static void osc_lock_granted(const struct lu_env *env, struct osc_lock *olck,
 407                              struct ldlm_lock *dlmlock, int rc)
 408 {
 409         struct ldlm_extent   *ext;
 410         struct cl_lock       *lock;
 411         struct cl_lock_descr *descr;
 412
 413         LASSERT(dlmlock->l_granted_mode == dlmlock->l_req_mode);
 414
 415         ENTRY;
 416         if (olck->ols_state < OLS_GRANTED) {
 417                 lock  = olck->ols_cl.cls_lock;
 418                 ext   = &dlmlock->l_policy_data.l_extent;
 419                 descr = &osc_env_info(env)->oti_descr;
 420                 descr->cld_obj = lock->cll_descr.cld_obj;
 421
 422                 /* XXX check that ->l_granted_mode is valid. */
 423                 descr->cld_mode  = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
 424                 descr->cld_start = cl_index(descr->cld_obj, ext->start);
 425                 descr->cld_end   = cl_index(descr->cld_obj, ext->end);
 426                 descr->cld_gid   = ext->gid;
 427                 /*
 428                  * tell upper layers the extent of the lock that was actually
 429                  * granted
 430                  */
 431                 olck->ols_state = OLS_GRANTED;
 432                 osc_lock_lvb_update(env, olck, rc);
 433
 434                 /* release DLM spin-locks to allow cl_lock_{modify,signal}()
 435                  * to take a semaphore on a parent lock. This is safe, because
 436                  * spin-locks are needed to protect consistency of
 437                  * dlmlock->l_*_mode and LVB, and we have finished processing
 438                  * them. */
 439                 unlock_res_and_lock(dlmlock);
 440                 cl_lock_modify(env, lock, descr);
 441                 cl_lock_signal(env, lock);
 442                 LINVRNT(osc_lock_invariant(olck));
 443                 lock_res_and_lock(dlmlock);
 444         }
 445         EXIT;
 446 }
 447
 448 static void osc_lock_upcall0(const struct lu_env *env, struct osc_lock *olck)
 449
 450 {
 451         struct ldlm_lock *dlmlock;
 452
 453         ENTRY;
 454
 455         dlmlock = ldlm_handle2lock_long(&olck->ols_handle, 0);
 456         LASSERT(dlmlock != NULL);
 457
 458         lock_res_and_lock(dlmlock);
 459         spin_lock(&osc_ast_guard);
 460         LASSERT(dlmlock->l_ast_data == olck);
 461         LASSERT(olck->ols_lock == NULL);
 462         olck->ols_lock = dlmlock;
 463         spin_unlock(&osc_ast_guard);
 464
 465         /*
 466          * Lock might be not yet granted. In this case, completion ast
 467          * (osc_ldlm_completion_ast()) comes later and finishes lock
 468          * granting.
 469          */
 470         if (dlmlock->l_granted_mode == dlmlock->l_req_mode)
 471                 osc_lock_granted(env, olck, dlmlock, 0);
 472         unlock_res_and_lock(dlmlock);
 473
 474         /*
 475          * osc_enqueue_interpret() decrefs asynchronous locks, counter
 476          * this.
 477          */
 478         ldlm_lock_addref(&olck->ols_handle, olck->ols_einfo.ei_mode);
 479         olck->ols_hold = 1;
 480
 481         /* lock reference taken by ldlm_handle2lock_long() is owned by
 482          * osc_lock and released in osc_lock_detach() */
 483         lu_ref_add(&dlmlock->l_reference, "osc_lock", olck);
 484         olck->ols_has_ref = 1;
 485 }
 486
 487 /**
 488  * Lock upcall function that is executed either when a reply to ENQUEUE rpc is
 489  * received from a server, or after osc_enqueue_base() matched a local DLM
 490  * lock.
 491  */
 492 static int osc_lock_upcall(void *cookie, int errcode)
 493 {
 494         struct osc_lock         *olck  = cookie;
 495         struct cl_lock_slice    *slice = &olck->ols_cl;
 496         struct cl_lock          *lock  = slice->cls_lock;
 497         struct lu_env           *env;
 498         struct cl_env_nest       nest;
 499
 500         ENTRY;
 501         env = cl_env_nested_get(&nest);
 502         if (!IS_ERR(env)) {
 503                 int rc;
 504
 505                 cl_lock_mutex_get(env, lock);
 506
 507                 LASSERT(lock->cll_state >= CLS_QUEUING);
 508                 if (olck->ols_state == OLS_ENQUEUED) {
 509                         olck->ols_state = OLS_UPCALL_RECEIVED;
 510                         rc = ldlm_error2errno(errcode);
 511                 } else if (olck->ols_state == OLS_CANCELLED) {
 512                         rc = -EIO;
 513                 } else {
 514                         CERROR("Impossible state: %d\n", olck->ols_state);
 515                         LBUG();
 516                 }
 517                 if (rc) {
 518                         struct ldlm_lock *dlmlock;
 519
 520                         dlmlock = ldlm_handle2lock(&olck->ols_handle);
 521                         if (dlmlock != NULL) {
 522                                 lock_res_and_lock(dlmlock);
 523                                 spin_lock(&osc_ast_guard);
 524                                 LASSERT(olck->ols_lock == NULL);
 525                                 dlmlock->l_ast_data = NULL;
 526                                 olck->ols_handle.cookie = 0ULL;
 527                                 spin_unlock(&osc_ast_guard);
 528                                 ldlm_lock_fail_match_locked(dlmlock);
 529                                 unlock_res_and_lock(dlmlock);
 530                                 LDLM_LOCK_PUT(dlmlock);
 531                         }
 532                 } else {
 533                         if (olck->ols_glimpse)
 534                                 olck->ols_glimpse = 0;
 535                         osc_lock_upcall0(env, olck);
 536                 }
 537
 538                 /* Error handling, some errors are tolerable. */
 539                 if (olck->ols_locklessable && rc == -EUSERS) {
 540                         /* This is a tolerable error, turn this lock into
 541                          * lockless lock.
 542                          */
 543                         osc_object_set_contended(cl2osc(slice->cls_obj));
 544                         LASSERT(slice->cls_ops == &osc_lock_ops);
 545
 546                         /* Change this lock to ldlmlock-less lock. */
 547                         osc_lock_to_lockless(env, olck, 1);
 548                         olck->ols_state = OLS_GRANTED;
 549                         rc = 0;
 550                 } else if (olck->ols_glimpse && rc == -ENAVAIL) {
 551                         osc_lock_lvb_update(env, olck, rc);
 552                         cl_lock_delete(env, lock);
 553                         /* Hide the error. */
 554                         rc = 0;
 555                 }
 556
 557                 if (rc == 0) {
 558                         /* For AGL case, the RPC sponsor may exits the cl_lock
 559                         *  processing without wait() called before related OSC
 560                         *  lock upcall(). So update the lock status according
 561                         *  to the enqueue result inside AGL upcall(). */
 562                         if (olck->ols_agl) {
 563                                 lock->cll_flags |= CLF_FROM_UPCALL;
 564                                 cl_wait_try(env, lock);
 565                                 lock->cll_flags &= ~CLF_FROM_UPCALL;
 566                         }
 567                         cl_lock_signal(env, lock);
 568                         /* del user for lock upcall cookie */
 569                         if (olck->ols_agl) {
 570                                 if (!olck->ols_glimpse)
 571                                         olck->ols_agl = 0;
 572                                 cl_unuse_try(env, lock);
 573                         }
 574                 } else {
 575                         /* del user for lock upcall cookie */
 576                         if (olck->ols_agl)
 577                                 cl_lock_user_del(env, lock);
 578                         cl_lock_error(env, lock, rc);
 579                 }
 580
 581                 /* release cookie reference, acquired by osc_lock_enqueue() */
 582                 cl_lock_hold_release(env, lock, "upcall", lock);
 583                 cl_lock_mutex_put(env, lock);
 584
 585                 lu_ref_del(&lock->cll_reference, "upcall", lock);
 586                 /* This maybe the last reference, so must be called after
 587                  * cl_lock_mutex_put(). */
 588                 cl_lock_put(env, lock);
 589
 590                 cl_env_nested_put(&nest, env);
 591         } else {
 592                 /* should never happen, similar to osc_ldlm_blocking_ast(). */
 593                 LBUG();
 594         }
 595         RETURN(errcode);
 596 }
 597
 598 /**
 599  * Core of osc_dlm_blocking_ast() logic.
 600  */
 601 static void osc_lock_blocking(const struct lu_env *env,
 602                               struct ldlm_lock *dlmlock,
 603                               struct osc_lock *olck, int blocking)
 604 {
 605         struct cl_lock *lock = olck->ols_cl.cls_lock;
 606
 607         LASSERT(olck->ols_lock == dlmlock);
 608         CLASSERT(OLS_BLOCKED < OLS_CANCELLED);
 609         LASSERT(!osc_lock_is_lockless(olck));
 610
 611         /*
 612          * Lock might be still addref-ed here, if e.g., blocking ast
 613          * is sent for a failed lock.
 614          */
 615         osc_lock_unhold(olck);
 616
 617         if (blocking && olck->ols_state < OLS_BLOCKED)
 618                 /*
 619                  * Move osc_lock into OLS_BLOCKED before canceling the lock,
 620                  * because it recursively re-enters osc_lock_blocking(), with
 621                  * the state set to OLS_CANCELLED.
 622                  */
 623                 olck->ols_state = OLS_BLOCKED;
 624         /*
 625          * cancel and destroy lock at least once no matter how blocking ast is
 626          * entered (see comment above osc_ldlm_blocking_ast() for use
 627          * cases). cl_lock_cancel() and cl_lock_delete() are idempotent.
 628          */
 629         cl_lock_cancel(env, lock);
 630         cl_lock_delete(env, lock);
 631 }
 632
 633 /**
 634  * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
 635  * and ldlm_lock caches.
 636  */
 637 static int osc_dlm_blocking_ast0(const struct lu_env *env,
 638                                  struct ldlm_lock *dlmlock,
 639                                  void *data, int flag)
 640 {
 641         struct osc_lock *olck;
 642         struct cl_lock  *lock;
 643         int result;
 644         int cancel;
 645
 646         LASSERT(flag == LDLM_CB_BLOCKING || flag == LDLM_CB_CANCELING);
 647
 648         cancel = 0;
 649         olck = osc_ast_data_get(dlmlock);
 650         if (olck != NULL) {
 651                 lock = olck->ols_cl.cls_lock;
 652                 cl_lock_mutex_get(env, lock);
 653                 LINVRNT(osc_lock_invariant(olck));
 654                 if (olck->ols_ast_wait) {
 655                         /* wake up osc_lock_use() */
 656                         cl_lock_signal(env, lock);
 657                         olck->ols_ast_wait = 0;
 658                 }
 659                 /*
 660                  * Lock might have been canceled while this thread was
 661                  * sleeping for lock mutex, but olck is pinned in memory.
 662                  */
 663                 if (olck == dlmlock->l_ast_data) {
 664                         /*
 665                          * NOTE: DLM sends blocking AST's for failed locks
 666                          *       (that are still in pre-OLS_GRANTED state)
 667                          *       too, and they have to be canceled otherwise
 668                          *       DLM lock is never destroyed and stuck in
 669                          *       the memory.
 670                          *
 671                          *       Alternatively, ldlm_cli_cancel() can be
 672                          *       called here directly for osc_locks with
 673                          *       ols_state < OLS_GRANTED to maintain an
 674                          *       invariant that ->clo_cancel() is only called
 675                          *       for locks that were granted.
 676                          */
 677                         LASSERT(data == olck);
 678                         osc_lock_blocking(env, dlmlock,
 679                                           olck, flag == LDLM_CB_BLOCKING);
 680                 } else
 681                         cancel = 1;
 682                 cl_lock_mutex_put(env, lock);
 683                 osc_ast_data_put(env, olck);
 684         } else
 685                 /*
 686                  * DLM lock exists, but there is no cl_lock attached to it.
 687                  * This is a `normal' race. cl_object and its cl_lock's can be
 688                  * removed by memory pressure, together with all pages.
 689                  */
 690                 cancel = (flag == LDLM_CB_BLOCKING);
 691
 692         if (cancel) {
 693                 struct lustre_handle *lockh;
 694
 695                 lockh = &osc_env_info(env)->oti_handle;
 696                 ldlm_lock2handle(dlmlock, lockh);
 697                 result = ldlm_cli_cancel(lockh, LCF_ASYNC);
 698         } else
 699                 result = 0;
 700         return result;
 701 }
 702
 703 /**
 704  * Blocking ast invoked by ldlm when dlm lock is either blocking progress of
 705  * some other lock, or is canceled. This function is installed as a
 706  * ldlm_lock::l_blocking_ast() for client extent locks.
 707  *
 708  * Control flow is tricky, because ldlm uses the same call-back
 709  * (ldlm_lock::l_blocking_ast()) for both blocking and cancellation ast's.
 710  *
 711  * \param dlmlock lock for which ast occurred.
 712  *
 713  * \param new description of a conflicting lock in case of blocking ast.
 714  *
 715  * \param data value of dlmlock->l_ast_data
 716  *
 717  * \param flag LDLM_CB_BLOCKING or LDLM_CB_CANCELING. Used to distinguish
 718  *             cancellation and blocking ast's.
 719  *
 720  * Possible use cases:
 721  *
 722  *     - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) to cancel
 723  *       lock due to lock lru pressure, or explicit user request to purge
 724  *       locks.
 725  *
 726  *     - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_BLOCKING) to notify
 727  *       us that dlmlock conflicts with another lock that some client is
 728  *       enqueing. Lock is canceled.
 729  *
 730  *           - cl_lock_cancel() is called. osc_lock_cancel() calls
 731  *             ldlm_cli_cancel() that calls
 732  *
 733  *                  dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
 734  *
 735  *             recursively entering osc_ldlm_blocking_ast().
 736  *
 737  *     - client cancels lock voluntary (e.g., as a part of early cancellation):
 738  *
 739  *           cl_lock_cancel()->
 740  *             osc_lock_cancel()->
 741  *               ldlm_cli_cancel()->
 742  *                 dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
 743  *
 744  */
 745 static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
 746                                  struct ldlm_lock_desc *new, void *data,
 747                                  int flag)
 748 {
 749         struct lu_env     *env;
 750         struct cl_env_nest nest;
 751         int                result;
 752
 753         /*
 754          * This can be called in the context of outer IO, e.g.,
 755          *
 756          *     cl_enqueue()->...
 757          *       ->osc_enqueue_base()->...
 758          *         ->ldlm_prep_elc_req()->...
 759          *           ->ldlm_cancel_callback()->...
 760          *             ->osc_ldlm_blocking_ast()
 761          *
 762          * new environment has to be created to not corrupt outer context.
 763          */
 764         env = cl_env_nested_get(&nest);
 765         if (!IS_ERR(env)) {
 766                 result = osc_dlm_blocking_ast0(env, dlmlock, data, flag);
 767                 cl_env_nested_put(&nest, env);
 768         } else {
 769                 result = PTR_ERR(env);
 770                 /*
 771                  * XXX This should never happen, as cl_lock is
 772                  * stuck. Pre-allocated environment a la vvp_inode_fini_env
 773                  * should be used.
 774                  */
 775                 LBUG();
 776         }
 777         if (result != 0) {
 778                 if (result == -ENODATA)
 779                         result = 0;
 780                 else
 781                         CERROR("BAST failed: %d\n", result);
 782         }
 783         return result;
 784 }
 785
 786 static int osc_ldlm_completion_ast(struct ldlm_lock *dlmlock,
 787                                    __u64 flags, void *data)
 788 {
 789         struct cl_env_nest nest;
 790         struct lu_env     *env;
 791         struct osc_lock   *olck;
 792         struct cl_lock    *lock;
 793         int result;
 794         int dlmrc;
 795
 796         /* first, do dlm part of the work */
 797         dlmrc = ldlm_completion_ast_async(dlmlock, flags, data);
 798         /* then, notify cl_lock */
 799         env = cl_env_nested_get(&nest);
 800         if (!IS_ERR(env)) {
 801                 olck = osc_ast_data_get(dlmlock);
 802                 if (olck != NULL) {
 803                         lock = olck->ols_cl.cls_lock;
 804                         cl_lock_mutex_get(env, lock);
 805                         /*
 806                          * ldlm_handle_cp_callback() copied LVB from request
 807                          * to lock->l_lvb_data, store it in osc_lock.
 808                          */
 809                         LASSERT(dlmlock->l_lvb_data != NULL);
 810                         lock_res_and_lock(dlmlock);
 811                         olck->ols_lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
 812                         if (olck->ols_lock == NULL) {
 813                                 /*
 814                                  * upcall (osc_lock_upcall()) hasn't yet been
 815                                  * called. Do nothing now, upcall will bind
 816                                  * olck to dlmlock and signal the waiters.
 817                                  *
 818                                  * This maintains an invariant that osc_lock
 819                                  * and ldlm_lock are always bound when
 820                                  * osc_lock is in OLS_GRANTED state.
 821                                  */
 822                         } else if (dlmlock->l_granted_mode ==
 823                                    dlmlock->l_req_mode) {
 824                                 osc_lock_granted(env, olck, dlmlock, dlmrc);
 825                         }
 826                         unlock_res_and_lock(dlmlock);
 827
 828                         if (dlmrc != 0) {
 829                                 CL_LOCK_DEBUG(D_ERROR, env, lock,
 830                                               "dlmlock returned %d\n", dlmrc);
 831                                 cl_lock_error(env, lock, dlmrc);
 832                         }
 833                         cl_lock_mutex_put(env, lock);
 834                         osc_ast_data_put(env, olck);
 835                         result = 0;
 836                 } else
 837                         result = -ELDLM_NO_LOCK_DATA;
 838                 cl_env_nested_put(&nest, env);
 839         } else
 840                 result = PTR_ERR(env);
 841         return dlmrc ?: result;
 842 }
 843
 844 static int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
 845 {
 846         struct ptlrpc_request  *req  = data;
 847         struct osc_lock        *olck;
 848         struct cl_lock         *lock;
 849         struct cl_object       *obj;
 850         struct cl_env_nest      nest;
 851         struct lu_env          *env;
 852         struct ost_lvb         *lvb;
 853         struct req_capsule     *cap;
 854         int                     result;
 855
 856         LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK);
 857
 858         env = cl_env_nested_get(&nest);
 859         if (!IS_ERR(env)) {
 860                 /* osc_ast_data_get() has to go after environment is
 861                  * allocated, because osc_ast_data() acquires a
 862                  * reference to a lock, and it can only be released in
 863                  * environment.
 864                  */
 865                 olck = osc_ast_data_get(dlmlock);
 866                 if (olck != NULL) {
 867                         lock = olck->ols_cl.cls_lock;
 868                         /* Do not grab the mutex of cl_lock for glimpse.
 869                          * See LU-1274 for details.
 870                          * BTW, it's okay for cl_lock to be cancelled during
 871                          * this period because server can handle this race.
 872                          * See ldlm_server_glimpse_ast() for details.
 873                          * cl_lock_mutex_get(env, lock); */
 874                         cap = &req->rq_pill;
 875                         req_capsule_extend(cap, &RQF_LDLM_GL_CALLBACK);
 876                         req_capsule_set_size(cap, &RMF_DLM_LVB, RCL_SERVER,
 877                                              sizeof *lvb);
 878                         result = req_capsule_server_pack(cap);
 879                         if (result == 0) {
 880                                 lvb = req_capsule_server_get(cap, &RMF_DLM_LVB);
 881                                 obj = lock->cll_descr.cld_obj;
 882                                 result = cl_object_glimpse(env, obj, lvb);
 883                         }
 884                         if (!exp_connect_lvb_type(req->rq_export))
 885                                 req_capsule_shrink(&req->rq_pill,
 886                                                    &RMF_DLM_LVB,
 887                                                    sizeof(struct ost_lvb_v1),
 888                                                    RCL_SERVER);
 889                         osc_ast_data_put(env, olck);
 890                 } else {
 891                         /*
 892                          * These errors are normal races, so we don't want to
 893                          * fill the console with messages by calling
 894                          * ptlrpc_error()
 895                          */
 896                         lustre_pack_reply(req, 1, NULL, NULL);
 897                         result = -ELDLM_NO_LOCK_DATA;
 898                 }
 899                 cl_env_nested_put(&nest, env);
 900         } else
 901                 result = PTR_ERR(env);
 902         req->rq_status = result;
 903         return result;
 904 }
 905
 906 static int weigh_cb(const struct lu_env *env, struct cl_io *io,
 907                     struct osc_page *ops, void *cbdata)
 908 {
 909         struct cl_page *page = ops->ops_cl.cpl_page;
 910
 911         if (cl_page_is_vmlocked(env, page)) {
 912                 (*(unsigned long *)cbdata)++;
 913                 return CLP_GANG_ABORT;
 914         }
 915
 916         return CLP_GANG_OKAY;
 917 }
 918
 919 static unsigned long osc_lock_weight(const struct lu_env *env,
 920                                      const struct osc_lock *ols)
 921 {
 922         struct cl_io *io = &osc_env_info(env)->oti_io;
 923         struct cl_lock_descr *descr = &ols->ols_cl.cls_lock->cll_descr;
 924         struct cl_object *obj = ols->ols_cl.cls_obj;
 925         unsigned long npages = 0;
 926         int result;
 927         ENTRY;
 928
 929         io->ci_obj = cl_object_top(obj);
 930         io->ci_ignore_layout = 1;
 931         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
 932         if (result != 0)
 933                 RETURN(result);
 934
 935         do {
 936                 result = osc_page_gang_lookup(env, io, cl2osc(obj),
 937                                               descr->cld_start, descr->cld_end,
 938                                               weigh_cb, (void *)&npages);
 939                 if (result == CLP_GANG_ABORT)
 940                         break;
 941                 if (result == CLP_GANG_RESCHED)
 942                         cond_resched();
 943         } while (result != CLP_GANG_OKAY);
 944         cl_io_fini(env, io);
 945
 946         return npages;
 947 }
 948
 949 /**
 950  * Get the weight of dlm lock for early cancellation.
 951  */
 952 unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock)
 953 {
 954         struct cl_env_nest       nest;
 955         struct lu_env           *env;
 956         struct osc_lock         *lock;
 957         unsigned long            weight;
 958         ENTRY;
 959
 960         might_sleep();
 961         /*
 962          * osc_ldlm_weigh_ast has a complex context since it might be called
 963          * because of lock canceling, or from user's input. We have to make
 964          * a new environment for it. Probably it is implementation safe to use
 965          * the upper context because cl_lock_put don't modify environment
 966          * variables. But just in case ..
 967          */
 968         env = cl_env_nested_get(&nest);
 969         if (IS_ERR(env))
 970                 /* Mostly because lack of memory, do not eliminate this lock */
 971                 RETURN(1);
 972
 973         LASSERT(dlmlock->l_resource->lr_type == LDLM_EXTENT);
 974         lock = osc_ast_data_get(dlmlock);
 975         if (lock == NULL) {
 976                 /* cl_lock was destroyed because of memory pressure.
 977                  * It is much reasonable to assign this type of lock
 978                  * a lower cost.
 979                  */
 980                 GOTO(out, weight = 0);
 981         }
 982
 983         weight = osc_lock_weight(env, lock);
 984         osc_ast_data_put(env, lock);
 985         EXIT;
 986
 987 out:
 988         cl_env_nested_put(&nest, env);
 989         return weight;
 990 }
 991
 992 static void osc_lock_build_einfo(const struct lu_env *env,
 993                                  const struct cl_lock *clock,
 994                                  struct osc_lock *lock,
 995                                  struct ldlm_enqueue_info *einfo)
 996 {
 997         enum cl_lock_mode mode;
 998
 999         mode = clock->cll_descr.cld_mode;
1000         if (mode == CLM_PHANTOM)
1001                 /*
1002                  * For now, enqueue all glimpse locks in read mode. In the
1003                  * future, client might choose to enqueue LCK_PW lock for
1004                  * glimpse on a file opened for write.
1005                  */
1006                 mode = CLM_READ;
1007
1008         einfo->ei_type   = LDLM_EXTENT;
1009         einfo->ei_mode   = osc_cl_lock2ldlm(mode);
1010         einfo->ei_cb_bl  = osc_ldlm_blocking_ast;
1011         einfo->ei_cb_cp  = osc_ldlm_completion_ast;
1012         einfo->ei_cb_gl  = osc_ldlm_glimpse_ast;
1013         einfo->ei_cbdata = lock; /* value to be put into ->l_ast_data */
1014 }
1015
1016 /**
1017  * Determine if the lock should be converted into a lockless lock.
1018  *
1019  * Steps to check:
1020  * - if the lock has an explicite requirment for a non-lockless lock;
1021  * - if the io lock request type ci_lockreq;
1022  * - send the enqueue rpc to ost to make the further decision;
1023  * - special treat to truncate lockless lock
1024  *
1025  *  Additional policy can be implemented here, e.g., never do lockless-io
1026  *  for large extents.
1027  */
1028 static void osc_lock_to_lockless(const struct lu_env *env,
1029                                  struct osc_lock *ols, int force)
1030 {
1031         struct cl_lock_slice *slice = &ols->ols_cl;
1032
1033         LASSERT(ols->ols_state == OLS_NEW ||
1034                 ols->ols_state == OLS_UPCALL_RECEIVED);
1035
1036         if (force) {
1037                 ols->ols_locklessable = 1;
1038                 slice->cls_ops = &osc_lock_lockless_ops;
1039         } else {
1040                 struct osc_io *oio     = osc_env_io(env);
1041                 struct cl_io  *io      = oio->oi_cl.cis_io;
1042                 struct cl_object *obj  = slice->cls_obj;
1043                 struct osc_object *oob = cl2osc(obj);
1044                 const struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
1045                 struct obd_connect_data *ocd;
1046
1047                 LASSERT(io->ci_lockreq == CILR_MANDATORY ||
1048                         io->ci_lockreq == CILR_MAYBE ||
1049                         io->ci_lockreq == CILR_NEVER);
1050
1051                 ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data;
1052                 ols->ols_locklessable = (io->ci_type != CIT_SETATTR) &&
1053                                 (io->ci_lockreq == CILR_MAYBE) &&
1054                                 (ocd->ocd_connect_flags & OBD_CONNECT_SRVLOCK);
1055                 if (io->ci_lockreq == CILR_NEVER ||
1056                         /* lockless IO */
1057                     (ols->ols_locklessable && osc_object_is_contended(oob)) ||
1058                         /* lockless truncate */
1059                     (cl_io_is_trunc(io) &&
1060                      (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK) &&
1061                       osd->od_lockless_truncate)) {
1062                         ols->ols_locklessable = 1;
1063                         slice->cls_ops = &osc_lock_lockless_ops;
1064                 }
1065         }
1066         LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
1067 }
1068
1069 static int osc_lock_compatible(const struct osc_lock *qing,
1070                                const struct osc_lock *qed)
1071 {
1072         enum cl_lock_mode qing_mode;
1073         enum cl_lock_mode qed_mode;
1074
1075         qing_mode = qing->ols_cl.cls_lock->cll_descr.cld_mode;
1076         if (qed->ols_glimpse &&
1077             (qed->ols_state >= OLS_UPCALL_RECEIVED || qing_mode == CLM_READ))
1078                 return 1;
1079
1080         qed_mode = qed->ols_cl.cls_lock->cll_descr.cld_mode;
1081         return ((qing_mode == CLM_READ) && (qed_mode == CLM_READ));
1082 }
1083
1084 /**
1085  * Cancel all conflicting locks and wait for them to be destroyed.
1086  *
1087  * This function is used for two purposes:
1088  *
1089  *     - early cancel all conflicting locks before starting IO, and
1090  *
1091  *     - guarantee that pages added to the page cache by lockless IO are never
1092  *       covered by locks other than lockless IO lock, and, hence, are not
1093  *       visible to other threads.
1094  */
1095 static int osc_lock_enqueue_wait(const struct lu_env *env,
1096                                  const struct osc_lock *olck)
1097 {
1098         struct cl_lock          *lock    = olck->ols_cl.cls_lock;
1099         struct cl_lock_descr    *descr   = &lock->cll_descr;
1100         struct cl_object_header *hdr     = cl_object_header(descr->cld_obj);
1101         struct cl_lock          *scan;
1102         struct cl_lock          *conflict= NULL;
1103         int lockless                     = osc_lock_is_lockless(olck);
1104         int rc                           = 0;
1105         ENTRY;
1106
1107         LASSERT(cl_lock_is_mutexed(lock));
1108
1109         /* make it enqueue anyway for glimpse lock, because we actually
1110          * don't need to cancel any conflicting locks. */
1111         if (olck->ols_glimpse)
1112                 return 0;
1113
1114         spin_lock(&hdr->coh_lock_guard);
1115         cfs_list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) {
1116                 struct cl_lock_descr *cld = &scan->cll_descr;
1117                 const struct osc_lock *scan_ols;
1118
1119                 if (scan == lock)
1120                         break;
1121
1122                 if (scan->cll_state < CLS_QUEUING ||
1123                     scan->cll_state == CLS_FREEING ||
1124                     cld->cld_start > descr->cld_end ||
1125                     cld->cld_end < descr->cld_start)
1126                         continue;
1127
1128                 /* overlapped and living locks. */
1129
1130                 /* We're not supposed to give up group lock. */
1131                 if (scan->cll_descr.cld_mode == CLM_GROUP) {
1132                         LASSERT(descr->cld_mode != CLM_GROUP ||
1133                                 descr->cld_gid != scan->cll_descr.cld_gid);
1134                         continue;
1135                 }
1136
1137                 scan_ols = osc_lock_at(scan);
1138
1139                 /* We need to cancel the compatible locks if we're enqueuing
1140                  * a lockless lock, for example:
1141                  * imagine that client has PR lock on [0, 1000], and thread T0
1142                  * is doing lockless IO in [500, 1500] region. Concurrent
1143                  * thread T1 can see lockless data in [500, 1000], which is
1144                  * wrong, because these data are possibly stale. */
1145                 if (!lockless && osc_lock_compatible(olck, scan_ols))
1146                         continue;
1147
1148                 cl_lock_get_trust(scan);
1149                 conflict = scan;
1150                 break;
1151         }
1152         spin_unlock(&hdr->coh_lock_guard);
1153
1154         if (conflict) {
1155                 if (lock->cll_descr.cld_mode == CLM_GROUP) {
1156                         /* we want a group lock but a previous lock request
1157                          * conflicts, we do not wait but return 0 so the
1158                          * request is send to the server
1159                          */
1160                         CDEBUG(D_DLMTRACE, "group lock %p is conflicted "
1161                                            "with %p, no wait, send to server\n",
1162                                lock, conflict);
1163                         cl_lock_put(env, conflict);
1164                         rc = 0;
1165                 } else {
1166                         CDEBUG(D_DLMTRACE, "lock %p is conflicted with %p, "
1167                                            "will wait\n",
1168                                lock, conflict);
1169                         LASSERT(lock->cll_conflict == NULL);
1170                         lu_ref_add(&conflict->cll_reference, "cancel-wait",
1171                                    lock);
1172                         lock->cll_conflict = conflict;
1173                         rc = CLO_WAIT;
1174                 }
1175         }
1176         RETURN(rc);
1177 }
1178
1179 /**
1180  * Implementation of cl_lock_operations::clo_enqueue() method for osc
1181  * layer. This initiates ldlm enqueue:
1182  *
1183  *     - cancels conflicting locks early (osc_lock_enqueue_wait());
1184  *
1185  *     - calls osc_enqueue_base() to do actual enqueue.
1186  *
1187  * osc_enqueue_base() is supplied with an upcall function that is executed
1188  * when lock is received either after a local cached ldlm lock is matched, or
1189  * when a reply from the server is received.
1190  *
1191  * This function does not wait for the network communication to complete.
1192  */
1193 static int osc_lock_enqueue(const struct lu_env *env,
1194                             const struct cl_lock_slice *slice,
1195                             struct cl_io *unused, __u32 enqflags)
1196 {
1197         struct osc_lock          *ols     = cl2osc_lock(slice);
1198         struct cl_lock           *lock    = ols->ols_cl.cls_lock;
1199         int result;
1200         ENTRY;
1201
1202         LASSERT(cl_lock_is_mutexed(lock));
1203         LASSERTF(ols->ols_state == OLS_NEW,
1204                  "Impossible state: %d\n", ols->ols_state);
1205
1206         LASSERTF(ergo(ols->ols_glimpse, lock->cll_descr.cld_mode <= CLM_READ),
1207                 "lock = %p, ols = %p\n", lock, ols);
1208
1209         result = osc_lock_enqueue_wait(env, ols);
1210         if (result == 0) {
1211                 if (!osc_lock_is_lockless(ols)) {
1212                         struct osc_object        *obj = cl2osc(slice->cls_obj);
1213                         struct osc_thread_info   *info = osc_env_info(env);
1214                         struct ldlm_res_id       *resname = &info->oti_resname;
1215                         ldlm_policy_data_t       *policy = &info->oti_policy;
1216                         struct ldlm_enqueue_info *einfo = &ols->ols_einfo;
1217
1218                         /* lock will be passed as upcall cookie,
1219                          * hold ref to prevent to be released. */
1220                         cl_lock_hold_add(env, lock, "upcall", lock);
1221                         /* a user for agl lock also */
1222                         if (ols->ols_agl)
1223                                 cl_lock_user_add(env, lock);
1224                         ols->ols_state = OLS_ENQUEUED;
1225
1226                         /*
1227                          * XXX: this is possible blocking point as
1228                          * ldlm_lock_match(LDLM_FL_LVB_READY) waits for
1229                          * LDLM_CP_CALLBACK.
1230                          */
1231                         ostid_build_res_name(&obj->oo_oinfo->loi_oi, resname);
1232                         osc_lock_build_policy(env, lock, policy);
1233                         result = osc_enqueue_base(osc_export(obj), resname,
1234                                           &ols->ols_flags, policy,
1235                                           &ols->ols_lvb,
1236                                           obj->oo_oinfo->loi_kms_valid,
1237                                           osc_lock_upcall,
1238                                           ols, einfo, &ols->ols_handle,
1239                                           PTLRPCD_SET, 1, ols->ols_agl);
1240                         if (result != 0) {
1241                                 if (ols->ols_agl)
1242                                         cl_lock_user_del(env, lock);
1243                                 cl_lock_unhold(env, lock, "upcall", lock);
1244                                 if (unlikely(result == -ECANCELED)) {
1245                                         ols->ols_state = OLS_NEW;
1246                                         result = 0;
1247                                 }
1248                         }
1249                 } else {
1250                         ols->ols_state = OLS_GRANTED;
1251                         ols->ols_owner = osc_env_io(env);
1252                 }
1253         }
1254         LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
1255         RETURN(result);
1256 }
1257
1258 static int osc_lock_wait(const struct lu_env *env,
1259                          const struct cl_lock_slice *slice)
1260 {
1261         struct osc_lock *olck = cl2osc_lock(slice);
1262         struct cl_lock  *lock = olck->ols_cl.cls_lock;
1263
1264         LINVRNT(osc_lock_invariant(olck));
1265
1266         if (olck->ols_glimpse && olck->ols_state >= OLS_UPCALL_RECEIVED) {
1267                 if (olck->ols_flags & LDLM_FL_LVB_READY) {
1268                         return 0;
1269                 } else if (olck->ols_agl) {
1270                         if (lock->cll_flags & CLF_FROM_UPCALL)
1271                                 /* It is from enqueue RPC reply upcall for
1272                                  * updating state. Do not re-enqueue. */
1273                                 return -ENAVAIL;
1274                         else
1275                                 olck->ols_state = OLS_NEW;
1276                 } else {
1277                         LASSERT(lock->cll_error);
1278                         return lock->cll_error;
1279                 }
1280         }
1281
1282         if (olck->ols_state == OLS_NEW) {
1283                 int rc;
1284
1285                 LASSERT(olck->ols_agl);
1286                 olck->ols_agl = 0;
1287                 olck->ols_flags &= ~LDLM_FL_BLOCK_NOWAIT;
1288                 rc = osc_lock_enqueue(env, slice, NULL, CEF_ASYNC | CEF_MUST);
1289                 if (rc != 0)
1290                         return rc;
1291                 else
1292                         return CLO_REENQUEUED;
1293         }
1294
1295         LASSERT(equi(olck->ols_state >= OLS_UPCALL_RECEIVED &&
1296                      lock->cll_error == 0, olck->ols_lock != NULL));
1297
1298         return lock->cll_error ?: olck->ols_state >= OLS_GRANTED ? 0 : CLO_WAIT;
1299 }
1300
1301 /**
1302  * An implementation of cl_lock_operations::clo_use() method that pins cached
1303  * lock.
1304  */
1305 static int osc_lock_use(const struct lu_env *env,
1306                         const struct cl_lock_slice *slice)
1307 {
1308         struct osc_lock *olck = cl2osc_lock(slice);
1309         int rc;
1310
1311         LASSERT(!olck->ols_hold);
1312
1313         /*
1314          * Atomically check for LDLM_FL_CBPENDING and addref a lock if this
1315          * flag is not set. This protects us from a concurrent blocking ast.
1316          */
1317         rc = ldlm_lock_addref_try(&olck->ols_handle, olck->ols_einfo.ei_mode);
1318         if (rc == 0) {
1319                 olck->ols_hold = 1;
1320                 olck->ols_state = OLS_GRANTED;
1321         } else {
1322                 struct cl_lock *lock;
1323
1324                 /*
1325                  * Lock is being cancelled somewhere within
1326                  * ldlm_handle_bl_callback(): LDLM_FL_CBPENDING is already
1327                  * set, but osc_ldlm_blocking_ast() hasn't yet acquired
1328                  * cl_lock mutex.
1329                  */
1330                 lock = slice->cls_lock;
1331                 LASSERT(lock->cll_state == CLS_INTRANSIT);
1332                 LASSERT(lock->cll_users > 0);
1333                 /* set a flag for osc_dlm_blocking_ast0() to signal the
1334                  * lock.*/
1335                 olck->ols_ast_wait = 1;
1336                 rc = CLO_WAIT;
1337         }
1338         return rc;
1339 }
1340
1341 static int osc_lock_flush(struct osc_lock *ols, int discard)
1342 {
1343         struct cl_lock       *lock  = ols->ols_cl.cls_lock;
1344         struct cl_env_nest    nest;
1345         struct lu_env        *env;
1346         int result = 0;
1347         ENTRY;
1348
1349         env = cl_env_nested_get(&nest);
1350         if (!IS_ERR(env)) {
1351                 struct osc_object    *obj   = cl2osc(ols->ols_cl.cls_obj);
1352                 struct cl_lock_descr *descr = &lock->cll_descr;
1353                 int rc = 0;
1354
1355                 if (descr->cld_mode >= CLM_WRITE) {
1356                         result = osc_cache_writeback_range(env, obj,
1357                                         descr->cld_start, descr->cld_end,
1358                                         1, discard);
1359                         LDLM_DEBUG(ols->ols_lock,
1360                                 "lock %p: %d pages were %s.\n", lock, result,
1361                                 discard ? "discarded" : "written");
1362                         if (result > 0)
1363                                 result = 0;
1364                 }
1365
1366                 rc = osc_lock_discard_pages(env, ols);
1367                 if (result == 0 && rc < 0)
1368                         result = rc;
1369
1370                 cl_env_nested_put(&nest, env);
1371         } else
1372                 result = PTR_ERR(env);
1373         if (result == 0) {
1374                 ols->ols_flush = 1;
1375                 LINVRNT(!osc_lock_has_pages(ols));
1376         }
1377         RETURN(result);
1378 }
1379
1380 /**
1381  * Implements cl_lock_operations::clo_cancel() method for osc layer. This is
1382  * called (as part of cl_lock_cancel()) when lock is canceled either voluntary
1383  * (LRU pressure, early cancellation, umount, etc.) or due to the conflict
1384  * with some other lock some where in the cluster. This function does the
1385  * following:
1386  *
1387  *     - invalidates all pages protected by this lock (after sending dirty
1388  *       ones to the server, as necessary);
1389  *
1390  *     - decref's underlying ldlm lock;
1391  *
1392  *     - cancels ldlm lock (ldlm_cli_cancel()).
1393  */
1394 static void osc_lock_cancel(const struct lu_env *env,
1395                             const struct cl_lock_slice *slice)
1396 {
1397         struct cl_lock   *lock    = slice->cls_lock;
1398         struct osc_lock  *olck    = cl2osc_lock(slice);
1399         struct ldlm_lock *dlmlock = olck->ols_lock;
1400
1401         LASSERT(cl_lock_is_mutexed(lock));
1402         LINVRNT(osc_lock_invariant(olck));
1403
1404         if (dlmlock != NULL) {
1405                 bool do_cancel;
1406                 int  result = 0;
1407
1408                 if (olck->ols_state >= OLS_GRANTED)
1409                         result = osc_lock_flush(olck,
1410                                 ldlm_is_discard_data(dlmlock));
1411                 osc_lock_unhold(olck);
1412
1413                 lock_res_and_lock(dlmlock);
1414                 /* Now that we're the only user of dlm read/write reference,
1415                  * mostly the ->l_readers + ->l_writers should be zero.
1416                  * However, there is a corner case.
1417                  * See b=18829 for details.*/
1418                 do_cancel = (dlmlock->l_readers == 0 &&
1419                              dlmlock->l_writers == 0);
1420                 ldlm_set_cbpending(dlmlock);
1421                 unlock_res_and_lock(dlmlock);
1422                 if (do_cancel)
1423                         result = ldlm_cli_cancel(&olck->ols_handle, LCF_ASYNC);
1424                 if (result < 0)
1425                         CL_LOCK_DEBUG(D_ERROR, env, lock,
1426                                       "lock %p cancel failure with error(%d)\n",
1427                                       lock, result);
1428         }
1429         olck->ols_state = OLS_CANCELLED;
1430         olck->ols_flags &= ~LDLM_FL_LVB_READY;
1431         osc_lock_detach(env, olck);
1432 }
1433
1434 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
1435 static int check_cb(const struct lu_env *env, struct cl_io *io,
1436                     struct osc_page *ops, void *cbdata)
1437 {
1438         struct cl_lock *lock = cbdata;
1439
1440         if (lock->cll_descr.cld_mode == CLM_READ) {
1441                 struct cl_lock *tmp;
1442                 tmp = cl_lock_at_pgoff(env, lock->cll_descr.cld_obj,
1443                                        osc_index(ops), lock, 1, 0);
1444                 if (tmp != NULL) {
1445                         cl_lock_put(env, tmp);
1446                         return CLP_GANG_OKAY;
1447                 }
1448         }
1449
1450         CL_LOCK_DEBUG(D_ERROR, env, lock, "still has pages\n");
1451         CL_PAGE_DEBUG(D_ERROR, env, ops->ops_cl.cpl_page, "\n");
1452         return CLP_GANG_ABORT;
1453 }
1454
1455 /**
1456  * Returns true iff there are pages under \a olck not protected by other
1457  * locks.
1458  */
1459 static bool osc_lock_has_pages(struct osc_lock *olck)
1460 {
1461         struct cl_lock       *lock;
1462         struct cl_lock_descr *descr;
1463         struct cl_object     *obj;
1464         struct osc_object    *oob;
1465         struct cl_env_nest    nest;
1466         struct cl_io         *io;
1467         struct lu_env        *env;
1468         bool                     has_pages;
1469         int                      rc;
1470
1471         env = cl_env_nested_get(&nest);
1472         if (IS_ERR(env))
1473                 return false;
1474
1475         obj   = olck->ols_cl.cls_obj;
1476         oob   = cl2osc(obj);
1477         io    = &oob->oo_debug_io;
1478         lock  = olck->ols_cl.cls_lock;
1479         descr = &lock->cll_descr;
1480
1481         mutex_lock(&oob->oo_debug_mutex);
1482         io->ci_obj = cl_object_top(obj);
1483         io->ci_ignore_layout = 1;
1484         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1485         if (rc != 0)
1486                 GOTO(out, has_pages = false);
1487
1488         do {
1489                 rc = osc_page_gang_lookup(env, io, oob,
1490                                           descr->cld_start, descr->cld_end,
1491                                           check_cb, (void *)lock);
1492                 if (rc == CLP_GANG_ABORT)
1493                         break;
1494                 if (rc == CLP_GANG_RESCHED)
1495                         cond_resched();
1496         } while (rc != CLP_GANG_OKAY);
1497         has_pages = (rc == CLP_GANG_ABORT);
1498 out:
1499         cl_io_fini(env, io);
1500         mutex_unlock(&oob->oo_debug_mutex);
1501         cl_env_nested_put(&nest, env);
1502
1503         return has_pages;
1504 }
1505 #else
1506 static bool osc_lock_has_pages(struct osc_lock *olck)
1507 {
1508         return false;
1509 }
1510 #endif /* CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK */
1511
1512 static void osc_lock_delete(const struct lu_env *env,
1513                             const struct cl_lock_slice *slice)
1514 {
1515         struct osc_lock *olck;
1516
1517         olck = cl2osc_lock(slice);
1518         if (olck->ols_glimpse) {
1519                 LASSERT(!olck->ols_hold);
1520                 LASSERT(!olck->ols_lock);
1521                 return;
1522         }
1523
1524         LINVRNT(osc_lock_invariant(olck));
1525         LINVRNT(!osc_lock_has_pages(olck));
1526
1527         osc_lock_unhold(olck);
1528         osc_lock_detach(env, olck);
1529 }
1530
1531 /**
1532  * Implements cl_lock_operations::clo_state() method for osc layer.
1533  *
1534  * Maintains osc_lock::ols_owner field.
1535  *
1536  * This assumes that lock always enters CLS_HELD (from some other state) in
1537  * the same IO context as one that requested the lock. This should not be a
1538  * problem, because context is by definition shared by all activity pertaining
1539  * to the same high-level IO.
1540  */
1541 static void osc_lock_state(const struct lu_env *env,
1542                            const struct cl_lock_slice *slice,
1543                            enum cl_lock_state state)
1544 {
1545         struct osc_lock *lock = cl2osc_lock(slice);
1546
1547         /*
1548          * XXX multiple io contexts can use the lock at the same time.
1549          */
1550         LINVRNT(osc_lock_invariant(lock));
1551         if (state == CLS_HELD && slice->cls_lock->cll_state != CLS_HELD) {
1552                 struct osc_io *oio = osc_env_io(env);
1553
1554                 LASSERT(lock->ols_owner == NULL);
1555                 lock->ols_owner = oio;
1556         } else if (state != CLS_HELD)
1557                 lock->ols_owner = NULL;
1558 }
1559
1560 static int osc_lock_print(const struct lu_env *env, void *cookie,
1561                           lu_printer_t p, const struct cl_lock_slice *slice)
1562 {
1563         struct osc_lock *lock = cl2osc_lock(slice);
1564
1565         /*
1566          * XXX print ldlm lock and einfo properly.
1567          */
1568         (*p)(env, cookie, "%p %#16llx "LPX64" %d %p ",
1569              lock->ols_lock, lock->ols_flags, lock->ols_handle.cookie,
1570              lock->ols_state, lock->ols_owner);
1571         osc_lvb_print(env, cookie, p, &lock->ols_lvb);
1572         return 0;
1573 }
1574
1575 static int osc_lock_fits_into(const struct lu_env *env,
1576                               const struct cl_lock_slice *slice,
1577                               const struct cl_lock_descr *need,
1578                               const struct cl_io *io)
1579 {
1580         struct osc_lock *ols = cl2osc_lock(slice);
1581
1582         if (need->cld_enq_flags & CEF_NEVER)
1583                 return 0;
1584
1585         if (ols->ols_state >= OLS_CANCELLED)
1586                 return 0;
1587
1588         if (need->cld_mode == CLM_PHANTOM) {
1589                 if (ols->ols_agl)
1590                         return !(ols->ols_state > OLS_RELEASED);
1591
1592                 /*
1593                  * Note: the QUEUED lock can't be matched here, otherwise
1594                  * it might cause the deadlocks.
1595                  * In read_process,
1596                  * P1: enqueued read lock, create sublock1
1597                  * P2: enqueued write lock, create sublock2(conflicted
1598                  *     with sublock1).
1599                  * P1: Grant read lock.
1600                  * P1: enqueued glimpse lock(with holding sublock1_read),
1601                  *     matched with sublock2, waiting sublock2 to be granted.
1602                  *     But sublock2 can not be granted, because P1
1603                  *     will not release sublock1. Bang!
1604                  */
1605                 if (ols->ols_state < OLS_GRANTED ||
1606                     ols->ols_state > OLS_RELEASED)
1607                         return 0;
1608         } else if (need->cld_enq_flags & CEF_MUST) {
1609                 /*
1610                  * If the lock hasn't ever enqueued, it can't be matched
1611                  * because enqueue process brings in many information
1612                  * which can be used to determine things such as lockless,
1613                  * CEF_MUST, etc.
1614                  */
1615                 if (ols->ols_state < OLS_UPCALL_RECEIVED &&
1616                     ols->ols_locklessable)
1617                         return 0;
1618         }
1619         return 1;
1620 }
1621
1622 static const struct cl_lock_operations osc_lock_ops = {
1623         .clo_fini    = osc_lock_fini,
1624         .clo_enqueue = osc_lock_enqueue,
1625         .clo_wait    = osc_lock_wait,
1626         .clo_unuse   = osc_lock_unuse,
1627         .clo_use     = osc_lock_use,
1628         .clo_delete  = osc_lock_delete,
1629         .clo_state   = osc_lock_state,
1630         .clo_cancel  = osc_lock_cancel,
1631         .clo_print   = osc_lock_print,
1632         .clo_fits_into = osc_lock_fits_into,
1633 };
1634
1635 static int osc_lock_lockless_unuse(const struct lu_env *env,
1636                                    const struct cl_lock_slice *slice)
1637 {
1638         struct osc_lock *ols = cl2osc_lock(slice);
1639         struct cl_lock *lock = slice->cls_lock;
1640
1641         LASSERT(ols->ols_state == OLS_GRANTED);
1642         LINVRNT(osc_lock_invariant(ols));
1643
1644         cl_lock_cancel(env, lock);
1645         cl_lock_delete(env, lock);
1646         return 0;
1647 }
1648
1649 static void osc_lock_lockless_cancel(const struct lu_env *env,
1650                                      const struct cl_lock_slice *slice)
1651 {
1652         struct osc_lock   *ols  = cl2osc_lock(slice);
1653         int result;
1654
1655         result = osc_lock_flush(ols, 0);
1656         if (result)
1657                 CERROR("Pages for lockless lock %p were not purged(%d)\n",
1658                        ols, result);
1659         ols->ols_state = OLS_CANCELLED;
1660 }
1661
1662 static int osc_lock_lockless_wait(const struct lu_env *env,
1663                                   const struct cl_lock_slice *slice)
1664 {
1665         struct osc_lock *olck = cl2osc_lock(slice);
1666         struct cl_lock  *lock = olck->ols_cl.cls_lock;
1667
1668         LINVRNT(osc_lock_invariant(olck));
1669         LASSERT(olck->ols_state >= OLS_UPCALL_RECEIVED);
1670
1671         return lock->cll_error;
1672 }
1673
1674 static void osc_lock_lockless_state(const struct lu_env *env,
1675                                     const struct cl_lock_slice *slice,
1676                                     enum cl_lock_state state)
1677 {
1678         struct osc_lock *lock = cl2osc_lock(slice);
1679
1680         LINVRNT(osc_lock_invariant(lock));
1681         if (state == CLS_HELD) {
1682                 struct osc_io *oio  = osc_env_io(env);
1683
1684                 LASSERT(ergo(lock->ols_owner, lock->ols_owner == oio));
1685                 lock->ols_owner = oio;
1686
1687                 /* set the io to be lockless if this lock is for io's
1688                  * host object */
1689                 if (cl_object_same(oio->oi_cl.cis_obj, slice->cls_obj))
1690                         oio->oi_lockless = 1;
1691         }
1692 }
1693
1694 static int osc_lock_lockless_fits_into(const struct lu_env *env,
1695                                        const struct cl_lock_slice *slice,
1696                                        const struct cl_lock_descr *need,
1697                                        const struct cl_io *io)
1698 {
1699         struct osc_lock *lock = cl2osc_lock(slice);
1700
1701         if (!(need->cld_enq_flags & CEF_NEVER))
1702                 return 0;
1703
1704         /* lockless lock should only be used by its owning io. b22147 */
1705         return (lock->ols_owner == osc_env_io(env));
1706 }
1707
1708 static const struct cl_lock_operations osc_lock_lockless_ops = {
1709         .clo_fini      = osc_lock_fini,
1710         .clo_enqueue   = osc_lock_enqueue,
1711         .clo_wait      = osc_lock_lockless_wait,
1712         .clo_unuse     = osc_lock_lockless_unuse,
1713         .clo_state     = osc_lock_lockless_state,
1714         .clo_fits_into = osc_lock_lockless_fits_into,
1715         .clo_cancel    = osc_lock_lockless_cancel,
1716         .clo_print     = osc_lock_print
1717 };
1718
1719 int osc_lock_init(const struct lu_env *env,
1720                   struct cl_object *obj, struct cl_lock *lock,
1721                   const struct cl_io *unused)
1722 {
1723         struct osc_lock *clk;
1724         int result;
1725
1726         OBD_SLAB_ALLOC_PTR_GFP(clk, osc_lock_kmem, GFP_NOFS);
1727         if (clk != NULL) {
1728                 __u32 enqflags = lock->cll_descr.cld_enq_flags;
1729
1730                 osc_lock_build_einfo(env, lock, clk, &clk->ols_einfo);
1731                 clk->ols_state = OLS_NEW;
1732
1733                 clk->ols_flags = osc_enq2ldlm_flags(enqflags);
1734                 clk->ols_agl = !!(enqflags & CEF_AGL);
1735                 if (clk->ols_agl)
1736                         clk->ols_flags |= LDLM_FL_BLOCK_NOWAIT;
1737                 if (clk->ols_flags & LDLM_FL_HAS_INTENT)
1738                         clk->ols_glimpse = 1;
1739
1740                 cl_lock_slice_add(lock, &clk->ols_cl, obj, &osc_lock_ops);
1741
1742                 if (!(enqflags & CEF_MUST))
1743                         /* try to convert this lock to a lockless lock */
1744                         osc_lock_to_lockless(env, clk, (enqflags & CEF_NEVER));
1745                 if (clk->ols_locklessable && !(enqflags & CEF_DISCARD_DATA))
1746                         clk->ols_flags |= LDLM_FL_DENY_ON_CONTENTION;
1747
1748                 LDLM_DEBUG_NOLOCK("lock %p, osc lock %p, flags %llx\n",
1749                                 lock, clk, clk->ols_flags);
1750
1751                 result = 0;
1752         } else
1753                 result = -ENOMEM;
1754         return result;
1755 }
1756
1757 /** @} osc */