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 2008 Sun Microsystems, Inc.  All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Implementation of cl_lock for OSC layer.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  */
  40
  41 /** \addtogroup osc osc @{ */
  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 /*****************************************************************************
  56  *
  57  * Type conversions.
  58  *
  59  */
  60
  61 static const struct cl_lock_operations osc_lock_ops;
  62 static const struct cl_lock_operations osc_lock_lockless_ops;
  63 static void osc_lock_to_lockless(const struct lu_env *env,
  64                                  struct osc_lock *ols, int force);
  65
  66 int osc_lock_is_lockless(const struct osc_lock *olck)
  67 {
  68         return (olck->ols_cl.cls_ops == &osc_lock_lockless_ops);
  69 }
  70
  71 /**
  72  * Returns a weak pointer to the ldlm lock identified by a handle. Returned
  73  * pointer cannot be dereferenced, as lock is not protected from concurrent
  74  * reclaim. This function is a helper for osc_lock_invariant().
  75  */
  76 static struct ldlm_lock *osc_handle_ptr(struct lustre_handle *handle)
  77 {
  78         struct ldlm_lock *lock;
  79
  80         lock = ldlm_handle2lock(handle);
  81         if (lock != NULL)
  82                 LDLM_LOCK_PUT(lock);
  83         return lock;
  84 }
  85
  86 /**
  87  * Invariant that has to be true all of the time.
  88  */
  89 static int osc_lock_invariant(struct osc_lock *ols)
  90 {
  91         struct ldlm_lock *lock        = osc_handle_ptr(&ols->ols_handle);
  92         struct ldlm_lock *olock       = ols->ols_lock;
  93         int               handle_used = lustre_handle_is_used(&ols->ols_handle);
  94
  95         return
  96                 ergo(osc_lock_is_lockless(ols),
  97                      ols->ols_locklessable && ols->ols_lock == NULL)  ||
  98                 (ergo(olock != NULL, handle_used) &&
  99                  ergo(olock != NULL,
 100                       olock->l_handle.h_cookie == ols->ols_handle.cookie) &&
 101                  /*
 102                   * Check that ->ols_handle and ->ols_lock are consistent, but
 103                   * take into account that they are set at the different time.
 104                   */
 105                  ergo(handle_used,
 106                       ergo(lock != NULL && olock != NULL, lock == olock) &&
 107                       ergo(lock == NULL, olock == NULL)) &&
 108                  ergo(ols->ols_state == OLS_CANCELLED,
 109                       olock == NULL && !handle_used) &&
 110                  /*
 111                   * DLM lock is destroyed only after we have seen cancellation
 112                   * ast.
 113                   */
 114                  ergo(olock != NULL && ols->ols_state < OLS_CANCELLED,
 115                       !olock->l_destroyed) &&
 116                  ergo(ols->ols_state == OLS_GRANTED,
 117                       olock != NULL &&
 118                       olock->l_req_mode == olock->l_granted_mode &&
 119                       ols->ols_hold));
 120 }
 121
 122 /*****************************************************************************
 123  *
 124  * Lock operations.
 125  *
 126  */
 127
 128 /**
 129  * Breaks a link between osc_lock and dlm_lock.
 130  */
 131 static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck)
 132 {
 133         struct ldlm_lock *dlmlock;
 134
 135         spin_lock(&osc_ast_guard);
 136         dlmlock = olck->ols_lock;
 137         if (dlmlock == NULL) {
 138                 spin_unlock(&osc_ast_guard);
 139                 return;
 140         }
 141
 142         olck->ols_lock = NULL;
 143         /* wb(); --- for all who checks (ols->ols_lock != NULL) before
 144          * call to osc_lock_detach() */
 145         dlmlock->l_ast_data = NULL;
 146         olck->ols_handle.cookie = 0ULL;
 147         spin_unlock(&osc_ast_guard);
 148
 149         lock_res_and_lock(dlmlock);
 150         if (dlmlock->l_granted_mode == dlmlock->l_req_mode) {
 151                 struct cl_object *obj = olck->ols_cl.cls_obj;
 152                 struct cl_attr *attr  = &osc_env_info(env)->oti_attr;
 153                 __u64 old_kms = cl2osc(obj)->oo_oinfo->loi_kms;
 154
 155                 /* Update the kms. Need to loop all granted locks.
 156                  * Not a problem for the client */
 157                 attr->cat_kms = ldlm_extent_shift_kms(dlmlock, old_kms);
 158                 unlock_res_and_lock(dlmlock);
 159
 160                 cl_object_attr_lock(obj);
 161                 cl_object_attr_set(env, obj, attr, CAT_KMS);
 162                 cl_object_attr_unlock(obj);
 163         } else
 164                 unlock_res_and_lock(dlmlock);
 165
 166         /* release a reference taken in osc_lock_upcall0(). */
 167         lu_ref_del(&dlmlock->l_reference, "osc_lock", olck);
 168         LDLM_LOCK_RELEASE(dlmlock);
 169 }
 170
 171 static int osc_lock_unuse(const struct lu_env *env,
 172                           const struct cl_lock_slice *slice)
 173 {
 174         struct osc_lock *ols = cl2osc_lock(slice);
 175         int result;
 176
 177         LASSERT(ols->ols_state == OLS_GRANTED ||
 178                 ols->ols_state == OLS_UPCALL_RECEIVED);
 179         LINVRNT(osc_lock_invariant(ols));
 180
 181         if (ols->ols_glimpse) {
 182                 LASSERT(ols->ols_hold == 0);
 183                 return 0;
 184         }
 185         LASSERT(ols->ols_hold);
 186
 187         /*
 188          * Move lock into OLS_RELEASED state before calling osc_cancel_base()
 189          * so that possible synchronous cancellation (that always happens
 190          * e.g., for liblustre) sees that lock is released.
 191          */
 192         ols->ols_state = OLS_RELEASED;
 193         ols->ols_hold = 0;
 194         result = osc_cancel_base(&ols->ols_handle, ols->ols_einfo.ei_mode);
 195         ols->ols_has_ref = 0;
 196         return result;
 197 }
 198
 199 static void osc_lock_fini(const struct lu_env *env,
 200                           struct cl_lock_slice *slice)
 201 {
 202         struct osc_lock  *ols = cl2osc_lock(slice);
 203
 204         LINVRNT(osc_lock_invariant(ols));
 205         /*
 206          * ->ols_hold can still be true at this point if, for example, a
 207          * thread that requested a lock was killed (and released a reference
 208          * to the lock), before reply from a server was received. In this case
 209          * lock is destroyed immediately after upcall.
 210          */
 211         if (ols->ols_hold)
 212                 osc_lock_unuse(env, slice);
 213         LASSERT(ols->ols_lock == NULL);
 214
 215         OBD_SLAB_FREE_PTR(ols, osc_lock_kmem);
 216 }
 217
 218 void osc_lock_build_res(const struct lu_env *env, const struct osc_object *obj,
 219                         struct ldlm_res_id *resname)
 220 {
 221         const struct lu_fid *fid = lu_object_fid(&obj->oo_cl.co_lu);
 222         if (0) {
 223                 /*
 224                  * In the perfect world of the future, where ost servers talk
 225                  * idif-fids...
 226                  */
 227                 fid_build_reg_res_name(fid, resname);
 228         } else {
 229                 /*
 230                  * In reality, where ost server expects ->lsm_object_id and
 231                  * ->lsm_object_gr in rename.
 232                  */
 233                 osc_build_res_name(obj->oo_oinfo->loi_id, obj->oo_oinfo->loi_gr,
 234                                    resname);
 235         }
 236 }
 237
 238 static void osc_lock_build_policy(const struct lu_env *env,
 239                                   const struct cl_lock *lock,
 240                                   ldlm_policy_data_t *policy)
 241 {
 242         const struct cl_lock_descr *d = &lock->cll_descr;
 243
 244         osc_index2policy(policy, d->cld_obj, d->cld_start, d->cld_end);
 245 }
 246
 247 static int osc_enq2ldlm_flags(__u32 enqflags)
 248 {
 249         int result = 0;
 250
 251         LASSERT((enqflags & ~CEF_MASK) == 0);
 252
 253         if (enqflags & CEF_NONBLOCK)
 254                 result |= LDLM_FL_BLOCK_NOWAIT;
 255         if (enqflags & CEF_ASYNC)
 256                 result |= LDLM_FL_HAS_INTENT;
 257         if (enqflags & CEF_DISCARD_DATA)
 258                 result |= LDLM_AST_DISCARD_DATA;
 259         return result;
 260 }
 261
 262 /**
 263  * Global spin-lock protecting consistency of ldlm_lock::l_ast_data
 264  * pointers. Initialized in osc_init().
 265  */
 266 spinlock_t osc_ast_guard;
 267
 268 static struct osc_lock *osc_ast_data_get(struct ldlm_lock *dlm_lock)
 269 {
 270         struct osc_lock *olck;
 271
 272         lock_res_and_lock(dlm_lock);
 273         spin_lock(&osc_ast_guard);
 274         olck = dlm_lock->l_ast_data;
 275         if (olck != NULL) {
 276                 struct cl_lock *lock = olck->ols_cl.cls_lock;
 277                 /*
 278                  * If osc_lock holds a reference on ldlm lock, return it even
 279                  * when cl_lock is in CLS_FREEING state. This way
 280                  *
 281                  *         osc_ast_data_get(dlmlock) == NULL
 282                  *
 283                  * guarantees that all osc references on dlmlock were
 284                  * released. osc_dlm_blocking_ast0() relies on that.
 285                  */
 286                 if (lock->cll_state < CLS_FREEING || olck->ols_has_ref) {
 287                         cl_lock_get_trust(lock);
 288                         lu_ref_add_atomic(&lock->cll_reference,
 289                                           "ast", cfs_current());
 290                 } else
 291                         olck = NULL;
 292         }
 293         spin_unlock(&osc_ast_guard);
 294         unlock_res_and_lock(dlm_lock);
 295         return olck;
 296 }
 297
 298 static void osc_ast_data_put(const struct lu_env *env, struct osc_lock *olck)
 299 {
 300         struct cl_lock *lock;
 301
 302         lock = olck->ols_cl.cls_lock;
 303         lu_ref_del(&lock->cll_reference, "ast", cfs_current());
 304         cl_lock_put(env, lock);
 305 }
 306
 307 /**
 308  * Updates object attributes from a lock value block (lvb) received together
 309  * with the DLM lock reply from the server. Copy of osc_update_enqueue()
 310  * logic.
 311  *
 312  * This can be optimized to not update attributes when lock is a result of a
 313  * local match.
 314  *
 315  * Called under lock and resource spin-locks.
 316  */
 317 static void osc_lock_lvb_update(const struct lu_env *env, struct osc_lock *olck,
 318                                 int rc)
 319 {
 320         struct ost_lvb    *lvb;
 321         struct cl_object  *obj;
 322         struct lov_oinfo  *oinfo;
 323         struct cl_attr    *attr;
 324         unsigned           valid;
 325
 326         ENTRY;
 327
 328         if (!(olck->ols_flags & LDLM_FL_LVB_READY)) {
 329                 EXIT;
 330                 return;
 331         }
 332
 333         lvb   = &olck->ols_lvb;
 334         obj   = olck->ols_cl.cls_obj;
 335         oinfo = cl2osc(obj)->oo_oinfo;
 336         attr  = &osc_env_info(env)->oti_attr;
 337         valid = CAT_BLOCKS | CAT_ATIME | CAT_CTIME | CAT_MTIME | CAT_SIZE;
 338         cl_lvb2attr(attr, lvb);
 339
 340         cl_object_attr_lock(obj);
 341         if (rc == 0) {
 342                 struct ldlm_lock  *dlmlock;
 343                 __u64 size;
 344
 345                 dlmlock = olck->ols_lock;
 346                 LASSERT(dlmlock != NULL);
 347
 348                 /* re-grab LVB from a dlm lock under DLM spin-locks. */
 349                 *lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
 350                 size = lvb->lvb_size;
 351                 /* Extend KMS up to the end of this lock and no further
 352                  * A lock on [x,y] means a KMS of up to y + 1 bytes! */
 353                 if (size > dlmlock->l_policy_data.l_extent.end)
 354                         size = dlmlock->l_policy_data.l_extent.end + 1;
 355                 if (size >= oinfo->loi_kms) {
 356                         LDLM_DEBUG(dlmlock, "lock acquired, setting rss="LPU64
 357                                    ", kms="LPU64, lvb->lvb_size, size);
 358                         valid |= CAT_KMS;
 359                         attr->cat_kms = size;
 360                 } else {
 361                         LDLM_DEBUG(dlmlock, "lock acquired, setting rss="
 362                                    LPU64"; leaving kms="LPU64", end="LPU64,
 363                                    lvb->lvb_size, oinfo->loi_kms,
 364                                    dlmlock->l_policy_data.l_extent.end);
 365                 }
 366                 ldlm_lock_allow_match_locked(dlmlock);
 367         } else if (rc == -ENAVAIL && olck->ols_glimpse) {
 368                 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
 369                        " kms="LPU64"\n", lvb->lvb_size, oinfo->loi_kms);
 370         } else
 371                 valid = 0;
 372
 373         if (valid != 0)
 374                 cl_object_attr_set(env, obj, attr, valid);
 375
 376         cl_object_attr_unlock(obj);
 377
 378         EXIT;
 379 }
 380
 381 /**
 382  * Called when a lock is granted, from an upcall (when server returned a
 383  * granted lock), or from completion AST, when server returned a blocked lock.
 384  *
 385  * Called under lock and resource spin-locks, that are released temporarily
 386  * here.
 387  */
 388 static void osc_lock_granted(const struct lu_env *env, struct osc_lock *olck,
 389                              struct ldlm_lock *dlmlock, int rc)
 390 {
 391         struct ldlm_extent   *ext;
 392         struct cl_lock       *lock;
 393         struct cl_lock_descr *descr;
 394
 395         LASSERT(dlmlock->l_granted_mode == dlmlock->l_req_mode);
 396
 397         ENTRY;
 398         if (olck->ols_state != OLS_GRANTED) {
 399                 lock  = olck->ols_cl.cls_lock;
 400                 ext   = &dlmlock->l_policy_data.l_extent;
 401                 descr = &osc_env_info(env)->oti_descr;
 402                 descr->cld_obj = lock->cll_descr.cld_obj;
 403
 404                 /* XXX check that ->l_granted_mode is valid. */
 405                 descr->cld_mode  = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
 406                 descr->cld_start = cl_index(descr->cld_obj, ext->start);
 407                 descr->cld_end   = cl_index(descr->cld_obj, ext->end);
 408                 /*
 409                  * tell upper layers the extent of the lock that was actually
 410                  * granted
 411                  */
 412                 olck->ols_state = OLS_GRANTED;
 413                 osc_lock_lvb_update(env, olck, rc);
 414
 415                 /* release DLM spin-locks to allow cl_lock_{modify,signal}()
 416                  * to take a semaphore on a parent lock. This is safe, because
 417                  * spin-locks are needed to protect consistency of
 418                  * dlmlock->l_*_mode and LVB, and we have finished processing
 419                  * them. */
 420                 unlock_res_and_lock(dlmlock);
 421                 cl_lock_modify(env, lock, descr);
 422                 cl_lock_signal(env, lock);
 423                 LINVRNT(osc_lock_invariant(olck));
 424                 lock_res_and_lock(dlmlock);
 425         }
 426         EXIT;
 427 }
 428
 429 static void osc_lock_upcall0(const struct lu_env *env, struct osc_lock *olck)
 430
 431 {
 432         struct ldlm_lock *dlmlock;
 433
 434         ENTRY;
 435
 436         dlmlock = ldlm_handle2lock_long(&olck->ols_handle, 0);
 437         LASSERT(dlmlock != NULL);
 438
 439         lock_res_and_lock(dlmlock);
 440         spin_lock(&osc_ast_guard);
 441         LASSERT(dlmlock->l_ast_data == olck);
 442         LASSERT(olck->ols_lock == NULL);
 443         olck->ols_lock = dlmlock;
 444         spin_unlock(&osc_ast_guard);
 445
 446         /*
 447          * Lock might be not yet granted. In this case, completion ast
 448          * (osc_ldlm_completion_ast()) comes later and finishes lock
 449          * granting.
 450          */
 451         if (dlmlock->l_granted_mode == dlmlock->l_req_mode)
 452                 osc_lock_granted(env, olck, dlmlock, 0);
 453         unlock_res_and_lock(dlmlock);
 454
 455         /*
 456          * osc_enqueue_interpret() decrefs asynchronous locks, counter
 457          * this.
 458          */
 459         ldlm_lock_addref(&olck->ols_handle, olck->ols_einfo.ei_mode);
 460         olck->ols_hold = olck->ols_has_ref = 1;
 461
 462         /* lock reference taken by ldlm_handle2lock_long() is owned by
 463          * osc_lock and released in osc_lock_detach() */
 464         lu_ref_add(&dlmlock->l_reference, "osc_lock", olck);
 465 }
 466
 467 /**
 468  * Lock upcall function that is executed either when a reply to ENQUEUE rpc is
 469  * received from a server, or after osc_enqueue_base() matched a local DLM
 470  * lock.
 471  */
 472 static int osc_lock_upcall(void *cookie, int errcode)
 473 {
 474         struct osc_lock      *olck  = cookie;
 475         struct cl_lock_slice *slice = &olck->ols_cl;
 476         struct cl_lock       *lock  = slice->cls_lock;
 477         struct lu_env        *env;
 478
 479         int refcheck;
 480
 481         ENTRY;
 482         /*
 483          * XXX environment should be created in ptlrpcd.
 484          */
 485         env = cl_env_get(&refcheck);
 486         if (!IS_ERR(env)) {
 487                 int rc;
 488
 489                 cl_lock_mutex_get(env, lock);
 490
 491                 LASSERT(lock->cll_state >= CLS_QUEUING);
 492                 if (olck->ols_state == OLS_ENQUEUED) {
 493                         olck->ols_state = OLS_UPCALL_RECEIVED;
 494                         rc = ldlm_error2errno(errcode);
 495                 } else if (olck->ols_state == OLS_CANCELLED) {
 496                         rc = -EIO;
 497                 } else {
 498                         CERROR("Impossible state: %i\n", olck->ols_state);
 499                         LBUG();
 500                 }
 501                 if (rc) {
 502                         struct ldlm_lock *dlmlock;
 503
 504                         dlmlock = ldlm_handle2lock(&olck->ols_handle);
 505                         if (dlmlock != NULL) {
 506                                 lock_res_and_lock(dlmlock);
 507                                 spin_lock(&osc_ast_guard);
 508                                 LASSERT(olck->ols_lock == NULL);
 509                                 dlmlock->l_ast_data = NULL;
 510                                 olck->ols_handle.cookie = 0ULL;
 511                                 spin_unlock(&osc_ast_guard);
 512                                 unlock_res_and_lock(dlmlock);
 513                                 LDLM_LOCK_PUT(dlmlock);
 514                         }
 515                 } else {
 516                         if (olck->ols_glimpse)
 517                                 olck->ols_glimpse = 0;
 518                         osc_lock_upcall0(env, olck);
 519                 }
 520
 521                 /* Error handling, some errors are tolerable. */
 522                 if (olck->ols_locklessable && rc == -EUSERS) {
 523                         /* This is a tolerable error, turn this lock into
 524                          * lockless lock.
 525                          */
 526                         osc_object_set_contended(cl2osc(slice->cls_obj));
 527                         LASSERT(slice->cls_ops == &osc_lock_ops);
 528
 529                         /* Change this lock to ldlmlock-less lock. */
 530                         osc_lock_to_lockless(env, olck, 1);
 531                         olck->ols_state = OLS_GRANTED;
 532                         rc = 0;
 533                 } else if (olck->ols_glimpse && rc == -ENAVAIL) {
 534                         osc_lock_lvb_update(env, olck, rc);
 535                         cl_lock_delete(env, lock);
 536                         /* Hide the error. */
 537                         rc = 0;
 538                 }
 539
 540                 if (rc == 0)
 541                         /* on error, lock was signaled by cl_lock_error() */
 542                         cl_lock_signal(env, lock);
 543                 else
 544                         cl_lock_error(env, lock, rc);
 545
 546                 cl_lock_mutex_put(env, lock);
 547
 548                 /* release cookie reference, acquired by osc_lock_enqueue() */
 549                 lu_ref_del(&lock->cll_reference, "upcall", lock);
 550                 cl_lock_put(env, lock);
 551                 cl_env_put(env, &refcheck);
 552         } else
 553                 /* should never happen, similar to osc_ldlm_blocking_ast(). */
 554                 LBUG();
 555         RETURN(errcode);
 556 }
 557
 558 /**
 559  * Core of osc_dlm_blocking_ast() logic.
 560  */
 561 static void osc_lock_blocking(const struct lu_env *env,
 562                               struct ldlm_lock *dlmlock,
 563                               struct osc_lock *olck, int blocking)
 564 {
 565         struct cl_lock *lock = olck->ols_cl.cls_lock;
 566
 567         LASSERT(olck->ols_lock == dlmlock);
 568         CLASSERT(OLS_BLOCKED < OLS_CANCELLED);
 569         LASSERT(!osc_lock_is_lockless(olck));
 570
 571         if (olck->ols_hold)
 572                 /*
 573                  * Lock might be still addref-ed here, if e.g., blocking ast
 574                  * is sent for a failed lock.
 575                  */
 576                 osc_lock_unuse(env, &olck->ols_cl);
 577
 578         if (blocking && olck->ols_state < OLS_BLOCKED)
 579                 /*
 580                  * Move osc_lock into OLS_BLOCKED before canceling the lock,
 581                  * because it recursively re-enters osc_lock_blocking(), with
 582                  * the state set to OLS_CANCELLED.
 583                  */
 584                 olck->ols_state = OLS_BLOCKED;
 585         /*
 586          * cancel and destroy lock at least once no matter how blocking ast is
 587          * entered (see comment above osc_ldlm_blocking_ast() for use
 588          * cases). cl_lock_cancel() and cl_lock_delete() are idempotent.
 589          */
 590         cl_lock_cancel(env, lock);
 591         cl_lock_delete(env, lock);
 592 }
 593
 594 /**
 595  * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
 596  * and ldlm_lock caches.
 597  */
 598 static int osc_dlm_blocking_ast0(const struct lu_env *env,
 599                                  struct ldlm_lock *dlmlock,
 600                                  void *data, int flag)
 601 {
 602         struct osc_lock *olck;
 603         struct cl_lock  *lock;
 604         int result;
 605         int cancel;
 606
 607         LASSERT(flag == LDLM_CB_BLOCKING || flag == LDLM_CB_CANCELING);
 608
 609         cancel = 0;
 610         olck = osc_ast_data_get(dlmlock);
 611         if (olck != NULL) {
 612                 lock = olck->ols_cl.cls_lock;
 613                 cl_lock_mutex_get(env, lock);
 614                 LINVRNT(osc_lock_invariant(olck));
 615                 if (olck->ols_ast_wait) {
 616                         /* wake up osc_lock_use() */
 617                         cl_lock_signal(env, lock);
 618                         olck->ols_ast_wait = 0;
 619                 }
 620                 /*
 621                  * Lock might have been canceled while this thread was
 622                  * sleeping for lock mutex, but olck is pinned in memory.
 623                  */
 624                 if (olck == dlmlock->l_ast_data) {
 625                         /*
 626                          * NOTE: DLM sends blocking AST's for failed locks
 627                          *       (that are still in pre-OLS_GRANTED state)
 628                          *       too, and they have to be canceled otherwise
 629                          *       DLM lock is never destroyed and stuck in
 630                          *       the memory.
 631                          *
 632                          *       Alternatively, ldlm_cli_cancel() can be
 633                          *       called here directly for osc_locks with
 634                          *       ols_state < OLS_GRANTED to maintain an
 635                          *       invariant that ->clo_cancel() is only called
 636                          *       for locks that were granted.
 637                          */
 638                         LASSERT(data == olck);
 639                         osc_lock_blocking(env, dlmlock,
 640                                           olck, flag == LDLM_CB_BLOCKING);
 641                 } else
 642                         cancel = 1;
 643                 cl_lock_mutex_put(env, lock);
 644                 osc_ast_data_put(env, olck);
 645         } else
 646                 /*
 647                  * DLM lock exists, but there is no cl_lock attached to it.
 648                  * This is a `normal' race. cl_object and its cl_lock's can be
 649                  * removed by memory pressure, together with all pages.
 650                  */
 651                 cancel = (flag == LDLM_CB_BLOCKING);
 652
 653         if (cancel) {
 654                 struct lustre_handle *lockh;
 655
 656                 lockh = &osc_env_info(env)->oti_handle;
 657                 ldlm_lock2handle(dlmlock, lockh);
 658                 result = ldlm_cli_cancel(lockh);
 659         } else
 660                 result = 0;
 661         return result;
 662 }
 663
 664 /**
 665  * Blocking ast invoked by ldlm when dlm lock is either blocking progress of
 666  * some other lock, or is canceled. This function is installed as a
 667  * ldlm_lock::l_blocking_ast() for client extent locks.
 668  *
 669  * Control flow is tricky, because ldlm uses the same call-back
 670  * (ldlm_lock::l_blocking_ast()) for both blocking and cancellation ast's.
 671  *
 672  * \param dlmlock lock for which ast occurred.
 673  *
 674  * \param new description of a conflicting lock in case of blocking ast.
 675  *
 676  * \param data value of dlmlock->l_ast_data
 677  *
 678  * \param flag LDLM_CB_BLOCKING or LDLM_CB_CANCELING. Used to distinguish
 679  *             cancellation and blocking ast's.
 680  *
 681  * Possible use cases:
 682  *
 683  *     - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) to cancel
 684  *       lock due to lock lru pressure, or explicit user request to purge
 685  *       locks.
 686  *
 687  *     - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_BLOCKING) to notify
 688  *       us that dlmlock conflicts with another lock that some client is
 689  *       enqueing. Lock is canceled.
 690  *
 691  *           - cl_lock_cancel() is called. osc_lock_cancel() calls
 692  *             ldlm_cli_cancel() that calls
 693  *
 694  *                  dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
 695  *
 696  *             recursively entering osc_ldlm_blocking_ast().
 697  *
 698  *     - client cancels lock voluntary (e.g., as a part of early cancellation):
 699  *
 700  *           cl_lock_cancel()->
 701  *             osc_lock_cancel()->
 702  *               ldlm_cli_cancel()->
 703  *                 dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
 704  *
 705  */
 706 static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
 707                                  struct ldlm_lock_desc *new, void *data,
 708                                  int flag)
 709 {
 710         struct lu_env     *env;
 711         struct cl_env_nest nest;
 712         int                result;
 713
 714         /*
 715          * This can be called in the context of outer IO, e.g.,
 716          *
 717          *     cl_enqueue()->...
 718          *       ->osc_enqueue_base()->...
 719          *         ->ldlm_prep_elc_req()->...
 720          *           ->ldlm_cancel_callback()->...
 721          *             ->osc_ldlm_blocking_ast()
 722          *
 723          * new environment has to be created to not corrupt outer context.
 724          */
 725         env = cl_env_nested_get(&nest);
 726         if (!IS_ERR(env))
 727                 result = osc_dlm_blocking_ast0(env, dlmlock, data, flag);
 728         else {
 729                 result = PTR_ERR(env);
 730                 /*
 731                  * XXX This should never happen, as cl_lock is
 732                  * stuck. Pre-allocated environment a la vvp_inode_fini_env
 733                  * should be used.
 734                  */
 735                 LBUG();
 736         }
 737         if (result != 0) {
 738                 if (result == -ENODATA)
 739                         result = 0;
 740                 else
 741                         CERROR("BAST failed: %d\n", result);
 742         }
 743         cl_env_nested_put(&nest, env);
 744         return result;
 745 }
 746
 747 static int osc_ldlm_completion_ast(struct ldlm_lock *dlmlock,
 748                                    int flags, void *data)
 749 {
 750         struct lu_env   *env;
 751         void            *env_cookie;
 752         struct osc_lock *olck;
 753         struct cl_lock  *lock;
 754         int refcheck;
 755         int result;
 756         int dlmrc;
 757
 758         /* first, do dlm part of the work */
 759         dlmrc = ldlm_completion_ast_async(dlmlock, flags, data);
 760         /* then, notify cl_lock */
 761         env_cookie = cl_env_reenter();
 762         env = cl_env_get(&refcheck);
 763         if (!IS_ERR(env)) {
 764                 olck = osc_ast_data_get(dlmlock);
 765                 if (olck != NULL) {
 766                         lock = olck->ols_cl.cls_lock;
 767                         cl_lock_mutex_get(env, lock);
 768                         /*
 769                          * ldlm_handle_cp_callback() copied LVB from request
 770                          * to lock->l_lvb_data, store it in osc_lock.
 771                          */
 772                         LASSERT(dlmlock->l_lvb_data != NULL);
 773                         lock_res_and_lock(dlmlock);
 774                         olck->ols_lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
 775                         if (olck->ols_lock == NULL)
 776                                 /*
 777                                  * upcall (osc_lock_upcall()) hasn't yet been
 778                                  * called. Do nothing now, upcall will bind
 779                                  * olck to dlmlock and signal the waiters.
 780                                  *
 781                                  * This maintains an invariant that osc_lock
 782                                  * and ldlm_lock are always bound when
 783                                  * osc_lock is in OLS_GRANTED state.
 784                                  */
 785                                 ;
 786                         else if (dlmlock->l_granted_mode != LCK_MINMODE)
 787                                 osc_lock_granted(env, olck, dlmlock, dlmrc);
 788                         unlock_res_and_lock(dlmlock);
 789                         if (dlmrc != 0)
 790                                 cl_lock_error(env, lock, dlmrc);
 791                         cl_lock_mutex_put(env, lock);
 792                         osc_ast_data_put(env, olck);
 793                         result = 0;
 794                 } else
 795                         result = -ELDLM_NO_LOCK_DATA;
 796                 cl_env_put(env, &refcheck);
 797         } else
 798                 result = PTR_ERR(env);
 799         cl_env_reexit(env_cookie);
 800         return dlmrc ?: result;
 801 }
 802
 803 static int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
 804 {
 805         struct ptlrpc_request  *req  = data;
 806         struct osc_lock        *olck;
 807         struct cl_lock         *lock;
 808         struct cl_object       *obj;
 809         struct lu_env          *env;
 810         struct ost_lvb         *lvb;
 811         struct req_capsule     *cap;
 812         int                     result;
 813         int                     refcheck;
 814
 815         LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK);
 816
 817         env = cl_env_get(&refcheck);
 818         if (!IS_ERR(env)) {
 819                 /*
 820                  * osc_ast_data_get() has to go after environment is
 821                  * allocated, because osc_ast_data() acquires a
 822                  * reference to a lock, and it can only be released in
 823                  * environment.
 824                  */
 825                 olck = osc_ast_data_get(dlmlock);
 826                 if (olck != NULL) {
 827                         lock = olck->ols_cl.cls_lock;
 828                         cl_lock_mutex_get(env, lock);
 829                         cap = &req->rq_pill;
 830                         req_capsule_extend(cap, &RQF_LDLM_GL_CALLBACK);
 831                         req_capsule_set_size(cap, &RMF_DLM_LVB, RCL_SERVER,
 832                                              sizeof *lvb);
 833                         result = req_capsule_server_pack(cap);
 834                         if (result == 0) {
 835                                 lvb = req_capsule_server_get(cap, &RMF_DLM_LVB);
 836                                 obj = lock->cll_descr.cld_obj;
 837                                 result = cl_object_glimpse(env, obj, lvb);
 838                         }
 839                         cl_lock_mutex_put(env, lock);
 840                         osc_ast_data_put(env, olck);
 841                 } else {
 842                         /*
 843                          * These errors are normal races, so we don't want to
 844                          * fill the console with messages by calling
 845                          * ptlrpc_error()
 846                          */
 847                         lustre_pack_reply(req, 1, NULL, NULL);
 848                         result = -ELDLM_NO_LOCK_DATA;
 849                 }
 850                 cl_env_put(env, &refcheck);
 851         } else
 852                 result = PTR_ERR(env);
 853         req->rq_status = result;
 854         return result;
 855 }
 856
 857 static unsigned long osc_lock_weigh(const struct lu_env *env,
 858                                     const struct cl_lock_slice *slice)
 859 {
 860         /*
 861          * don't need to grab coh_page_guard since we don't care the exact #
 862          * of pages..
 863          */
 864         return cl_object_header(slice->cls_obj)->coh_pages;
 865 }
 866
 867 /**
 868  * Get the weight of dlm lock for early cancellation.
 869  *
 870  * XXX: it should return the pages covered by this \a dlmlock.
 871  */
 872 static unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock)
 873 {
 874         struct lu_env           *env;
 875         int                      refcheck;
 876         void                    *cookie;
 877         struct osc_lock         *lock;
 878         struct cl_lock          *cll;
 879         unsigned long            weight;
 880         ENTRY;
 881
 882         might_sleep();
 883         cookie = cl_env_reenter();
 884         /*
 885          * osc_ldlm_weigh_ast has a complex context since it might be called
 886          * because of lock canceling, or from user's input. We have to make
 887          * a new environment for it. Probably it is implementation safe to use
 888          * the upper context because cl_lock_put don't modify environment
 889          * variables. But in case of ..
 890          */
 891         env = cl_env_get(&refcheck);
 892         if (IS_ERR(env)) {
 893                 /* Mostly because lack of memory, tend to eliminate this lock*/
 894                 cl_env_reexit(cookie);
 895                 RETURN(0);
 896         }
 897
 898         LASSERT(dlmlock->l_resource->lr_type == LDLM_EXTENT);
 899         lock = osc_ast_data_get(dlmlock);
 900         if (lock == NULL) {
 901                 /* cl_lock was destroyed because of memory pressure.
 902                  * It is much reasonable to assign this type of lock
 903                  * a lower cost.
 904                  */
 905                 GOTO(out, weight = 0);
 906         }
 907
 908         cll = lock->ols_cl.cls_lock;
 909         cl_lock_mutex_get(env, cll);
 910         weight = cl_lock_weigh(env, cll);
 911         cl_lock_mutex_put(env, cll);
 912         osc_ast_data_put(env, lock);
 913         EXIT;
 914
 915 out:
 916         cl_env_put(env, &refcheck);
 917         cl_env_reexit(cookie);
 918         return weight;
 919 }
 920
 921 static void osc_lock_build_einfo(const struct lu_env *env,
 922                                  const struct cl_lock *clock,
 923                                  struct osc_lock *lock,
 924                                  struct ldlm_enqueue_info *einfo)
 925 {
 926         enum cl_lock_mode mode;
 927
 928         mode = clock->cll_descr.cld_mode;
 929         if (mode == CLM_PHANTOM)
 930                 /*
 931                  * For now, enqueue all glimpse locks in read mode. In the
 932                  * future, client might choose to enqueue LCK_PW lock for
 933                  * glimpse on a file opened for write.
 934                  */
 935                 mode = CLM_READ;
 936
 937         einfo->ei_type   = LDLM_EXTENT;
 938         einfo->ei_mode   = osc_cl_lock2ldlm(mode);
 939         einfo->ei_cb_bl  = osc_ldlm_blocking_ast;
 940         einfo->ei_cb_cp  = osc_ldlm_completion_ast;
 941         einfo->ei_cb_gl  = osc_ldlm_glimpse_ast;
 942         einfo->ei_cb_wg  = osc_ldlm_weigh_ast;
 943         einfo->ei_cbdata = lock; /* value to be put into ->l_ast_data */
 944 }
 945
 946 static int osc_lock_delete0(struct cl_lock *conflict)
 947 {
 948         struct cl_env_nest    nest;
 949         struct lu_env        *env;
 950         int    rc = 0;
 951
 952         env = cl_env_nested_get(&nest);
 953         if (!IS_ERR(env)) {
 954                 cl_lock_delete(env, conflict);
 955                 cl_env_nested_put(&nest, env);
 956         } else
 957                 rc = PTR_ERR(env);
 958         return rc;
 959 }
 960 /**
 961  * Cancels \a conflict lock and waits until it reached CLS_FREEING state. This
 962  * is called as a part of enqueuing to cancel conflicting locks early.
 963  *
 964  * \retval            0: success, \a conflict was cancelled and destroyed.
 965  *
 966  * \retval   CLO_REPEAT: \a conflict was cancelled, but \a lock mutex was
 967  *                       released in the process. Repeat enqueing.
 968  *
 969  * \retval -EWOULDBLOCK: \a conflict cannot be cancelled immediately, and
 970  *                       either \a lock is non-blocking, or current thread
 971  *                       holds other locks, that prevent it from waiting
 972  *                       for cancel to complete.
 973  *
 974  * \retval          -ve: other error, including -EINTR.
 975  *
 976  */
 977 static int osc_lock_cancel_wait(const struct lu_env *env, struct cl_lock *lock,
 978                                 struct cl_lock *conflict, int canwait)
 979 {
 980         int rc;
 981
 982         LASSERT(cl_lock_is_mutexed(lock));
 983         LASSERT(cl_lock_is_mutexed(conflict));
 984
 985         rc = 0;
 986         if (conflict->cll_state != CLS_FREEING) {
 987                 cl_lock_cancel(env, conflict);
 988                 rc = osc_lock_delete0(conflict);
 989                 if (rc)
 990                         return rc;
 991                 if (conflict->cll_flags & (CLF_CANCELPEND|CLF_DOOMED)) {
 992                         rc = -EWOULDBLOCK;
 993                         if (cl_lock_nr_mutexed(env) > 2)
 994                                 /*
 995                                  * If mutices of locks other than @lock and
 996                                  * @scan are held by the current thread, it
 997                                  * cannot wait on @scan state change in a
 998                                  * dead-lock safe matter, so simply skip early
 999                                  * cancellation in this case.
1000                                  *
1001                                  * This means that early cancellation doesn't
1002                                  * work when there is even slight mutex
1003                                  * contention, as top-lock's mutex is usually
1004                                  * held at this time.
1005                                  */
1006                                 ;
1007                         else if (canwait) {
1008                                 /* Waiting for @scan to be destroyed */
1009                                 cl_lock_mutex_put(env, lock);
1010                                 do {
1011                                         rc = cl_lock_state_wait(env, conflict);
1012                                 } while (!rc &&
1013                                          conflict->cll_state < CLS_FREEING);
1014                                 /* mutex was released, repeat enqueue. */
1015                                 rc = rc ?: CLO_REPEAT;
1016                                 cl_lock_mutex_get(env, lock);
1017                         }
1018                 }
1019                 LASSERT(ergo(!rc, conflict->cll_state == CLS_FREEING));
1020                 CDEBUG(D_INFO, "lock %p was %s freed now, rc (%d)\n",
1021                        conflict, rc ? "not":"", rc);
1022         }
1023         return rc;
1024 }
1025
1026 /**
1027  * Determine if the lock should be converted into a lockless lock.
1028  *
1029  * Steps to check:
1030  * - if the lock has an explicite requirment for a non-lockless lock;
1031  * - if the io lock request type ci_lockreq;
1032  * - send the enqueue rpc to ost to make the further decision;
1033  * - special treat to truncate lockless lock
1034  *
1035  *  Additional policy can be implemented here, e.g., never do lockless-io
1036  *  for large extents.
1037  */
1038 static void osc_lock_to_lockless(const struct lu_env *env,
1039                                  struct osc_lock *ols, int force)
1040 {
1041         struct cl_lock_slice *slice = &ols->ols_cl;
1042         struct cl_lock *lock        = slice->cls_lock;
1043
1044         LASSERT(ols->ols_state == OLS_NEW ||
1045                 ols->ols_state == OLS_UPCALL_RECEIVED);
1046
1047         if (force) {
1048                 ols->ols_locklessable = 1;
1049                 LASSERT(cl_lock_is_mutexed(lock));
1050                 slice->cls_ops = &osc_lock_lockless_ops;
1051         } else {
1052                 struct osc_io *oio     = osc_env_io(env);
1053                 struct cl_io  *io      = oio->oi_cl.cis_io;
1054                 struct cl_object *obj  = slice->cls_obj;
1055                 struct osc_object *oob = cl2osc(obj);
1056                 const struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
1057                 struct obd_connect_data *ocd;
1058
1059                 LASSERT(io->ci_lockreq == CILR_MANDATORY ||
1060                         io->ci_lockreq == CILR_MAYBE ||
1061                         io->ci_lockreq == CILR_NEVER);
1062
1063                 ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data;
1064                 ols->ols_locklessable = (io->ci_type != CIT_TRUNC) &&
1065                                 (io->ci_lockreq == CILR_MAYBE) &&
1066                                 (ocd->ocd_connect_flags & OBD_CONNECT_SRVLOCK);
1067                 if (io->ci_lockreq == CILR_NEVER ||
1068                         /* lockless IO */
1069                     (ols->ols_locklessable && osc_object_is_contended(oob)) ||
1070                         /* lockless truncate */
1071                     (io->ci_type == CIT_TRUNC &&
1072                      (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK) &&
1073                       osd->od_lockless_truncate)) {
1074                         ols->ols_locklessable = 1;
1075                         slice->cls_ops = &osc_lock_lockless_ops;
1076                 }
1077         }
1078         LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
1079 }
1080
1081 /**
1082  * Cancel all conflicting locks and wait for them to be destroyed.
1083  *
1084  * This function is used for two purposes:
1085  *
1086  *     - early cancel all conflicting locks before starting IO, and
1087  *
1088  *     - guarantee that pages added to the page cache by lockless IO are never
1089  *       covered by locks other than lockless IO lock, and, hence, are not
1090  *       visible to other threads.
1091  */
1092 static int osc_lock_enqueue_wait(const struct lu_env *env,
1093                                  const struct osc_lock *olck)
1094 {
1095         struct cl_lock          *lock    = olck->ols_cl.cls_lock;
1096         struct cl_lock_descr    *descr   = &lock->cll_descr;
1097         struct cl_object_header *hdr     = cl_object_header(descr->cld_obj);
1098         struct cl_lock_closure  *closure = &osc_env_info(env)->oti_closure;
1099         struct cl_lock          *scan;
1100         struct cl_lock          *temp;
1101         int lockless                     = osc_lock_is_lockless(olck);
1102         int rc                           = 0;
1103         int canwait;
1104         int stop;
1105         ENTRY;
1106
1107         LASSERT(cl_lock_is_mutexed(lock));
1108         LASSERT(lock->cll_state == CLS_QUEUING);
1109
1110         /*
1111          * XXX This function could be sped up if we had asynchronous
1112          * cancellation.
1113          */
1114
1115         canwait =
1116                 !(olck->ols_flags & LDLM_FL_BLOCK_NOWAIT) &&
1117                 cl_lock_nr_mutexed(env) == 1;
1118         cl_lock_closure_init(env, closure, lock, canwait);
1119         spin_lock(&hdr->coh_lock_guard);
1120         list_for_each_entry_safe(scan, temp, &hdr->coh_locks, cll_linkage) {
1121                 if (scan == lock)
1122                         continue;
1123
1124                 if (scan->cll_state < CLS_QUEUING ||
1125                     scan->cll_state == CLS_FREEING ||
1126                     scan->cll_descr.cld_start > descr->cld_end ||
1127                     scan->cll_descr.cld_end < descr->cld_start)
1128                         continue;
1129
1130                 /* overlapped and living locks. */
1131                 /* A tricky case for lockless pages:
1132                  * We need to cancel the compatible locks if we're enqueuing
1133                  * a lockless lock, for example:
1134                  * imagine that client has PR lock on [0, 1000], and thread T0
1135                  * is doing lockless IO in [500, 1500] region. Concurrent
1136                  * thread T1 can see lockless data in [500, 1000], which is
1137                  * wrong, because these data are possibly stale.
1138                  */
1139                 if (!lockless && cl_lock_compatible(scan, lock))
1140                         continue;
1141
1142                 /* Now @scan is conflicting with @lock, this means current
1143                  * thread have to sleep for @scan being destroyed. */
1144                 cl_lock_get_trust(scan);
1145                 if (&temp->cll_linkage != &hdr->coh_locks)
1146                         cl_lock_get_trust(temp);
1147                 spin_unlock(&hdr->coh_lock_guard);
1148                 lu_ref_add(&scan->cll_reference, "cancel-wait", lock);
1149
1150                 LASSERT(list_empty(&closure->clc_list));
1151                 rc = cl_lock_closure_build(env, scan, closure);
1152                 if (rc == 0) {
1153                         rc = osc_lock_cancel_wait(env, lock, scan, canwait);
1154                         cl_lock_disclosure(env, closure);
1155                         if (rc == -EWOULDBLOCK)
1156                                 rc = 0;
1157                 }
1158                 if (rc == CLO_REPEAT && !canwait)
1159                         /* cannot wait... no early cancellation. */
1160                         rc = 0;
1161
1162                 lu_ref_del(&scan->cll_reference, "cancel-wait", lock);
1163                 cl_lock_put(env, scan);
1164                 spin_lock(&hdr->coh_lock_guard);
1165                 /*
1166                  * Lock list could have been modified, while spin-lock was
1167                  * released. Check that it is safe to continue.
1168                  */
1169                 stop = list_empty(&temp->cll_linkage);
1170                 if (&temp->cll_linkage != &hdr->coh_locks)
1171                         cl_lock_put(env, temp);
1172                 if (stop || rc != 0)
1173                         break;
1174         }
1175         spin_unlock(&hdr->coh_lock_guard);
1176         cl_lock_closure_fini(closure);
1177         RETURN(rc);
1178 }
1179
1180 /**
1181  * Deadlock avoidance for osc_lock_enqueue(). Consider following scenario:
1182  *
1183  *     - Thread0: obtains PR:[0, 10]. Lock is busy.
1184  *
1185  *     - Thread1: enqueues PW:[5, 50]. Blocking ast is sent to
1186  *       PR:[0, 10], but cancellation of busy lock is postponed.
1187  *
1188  *     - Thread0: enqueue PR:[30, 40]. Lock is locally matched to
1189  *       PW:[5, 50], and thread0 waits for the lock completion never
1190  *       releasing PR:[0, 10]---deadlock.
1191  *
1192  * The second PR lock can be glimpse (it is to deal with that situation that
1193  * ll_glimpse_size() has second argument, preventing local match of
1194  * not-yet-granted locks, see bug 10295). Similar situation is possible in the
1195  * case of memory mapped user level buffer.
1196  *
1197  * To prevent this we can detect a situation when current "thread" or "io"
1198  * already holds a lock on this object and either add LDLM_FL_BLOCK_GRANTED to
1199  * the ols->ols_flags, or prevent local match with PW locks.
1200  */
1201 static int osc_deadlock_is_possible(const struct lu_env *env,
1202                                     struct cl_lock *lock)
1203 {
1204         struct cl_object        *obj;
1205         struct cl_object_header *head;
1206         struct cl_lock          *scan;
1207         struct osc_io           *oio;
1208
1209         int result;
1210
1211         ENTRY;
1212
1213         LASSERT(cl_lock_is_mutexed(lock));
1214
1215         oio  = osc_env_io(env);
1216         obj  = lock->cll_descr.cld_obj;
1217         head = cl_object_header(obj);
1218
1219         result = 0;
1220         spin_lock(&head->coh_lock_guard);
1221         list_for_each_entry(scan, &head->coh_locks, cll_linkage) {
1222                 if (scan != lock) {
1223                         struct osc_lock *oscan;
1224
1225                         oscan = osc_lock_at(scan);
1226                         LASSERT(oscan != NULL);
1227                         if (oscan->ols_owner == oio) {
1228                                 result = 1;
1229                                 break;
1230                         }
1231                 }
1232         }
1233         spin_unlock(&head->coh_lock_guard);
1234         RETURN(result);
1235 }
1236
1237 /**
1238  * Implementation of cl_lock_operations::clo_enqueue() method for osc
1239  * layer. This initiates ldlm enqueue:
1240  *
1241  *     - checks for possible dead-lock conditions (osc_deadlock_is_possible());
1242  *
1243  *     - cancels conflicting locks early (osc_lock_enqueue_wait());
1244  *
1245  *     - calls osc_enqueue_base() to do actual enqueue.
1246  *
1247  * osc_enqueue_base() is supplied with an upcall function that is executed
1248  * when lock is received either after a local cached ldlm lock is matched, or
1249  * when a reply from the server is received.
1250  *
1251  * This function does not wait for the network communication to complete.
1252  */
1253 static int osc_lock_enqueue(const struct lu_env *env,
1254                             const struct cl_lock_slice *slice,
1255                             struct cl_io *_, __u32 enqflags)
1256 {
1257         struct osc_lock          *ols     = cl2osc_lock(slice);
1258         struct cl_lock           *lock    = ols->ols_cl.cls_lock;
1259         struct osc_object        *obj     = cl2osc(slice->cls_obj);
1260         struct osc_thread_info   *info    = osc_env_info(env);
1261         struct ldlm_res_id       *resname = &info->oti_resname;
1262         ldlm_policy_data_t       *policy  = &info->oti_policy;
1263         struct ldlm_enqueue_info *einfo   = &ols->ols_einfo;
1264         int result;
1265         ENTRY;
1266
1267         LASSERT(cl_lock_is_mutexed(lock));
1268         LASSERT(lock->cll_state == CLS_QUEUING);
1269         LASSERT(ols->ols_state == OLS_NEW);
1270
1271         osc_lock_build_res(env, obj, resname);
1272         osc_lock_build_policy(env, lock, policy);
1273         ols->ols_flags = osc_enq2ldlm_flags(enqflags);
1274         if (osc_deadlock_is_possible(env, lock))
1275                 ols->ols_flags |= LDLM_FL_BLOCK_GRANTED;
1276         if (ols->ols_flags & LDLM_FL_HAS_INTENT)
1277                 ols->ols_glimpse = 1;
1278
1279         result = osc_lock_enqueue_wait(env, ols);
1280         if (result == 0) {
1281                 if (!(enqflags & CEF_MUST))
1282                         /* try to convert this lock to a lockless lock */
1283                         osc_lock_to_lockless(env, ols, (enqflags & CEF_NEVER));
1284                 if (!osc_lock_is_lockless(ols)) {
1285                         if (ols->ols_locklessable)
1286                                 ols->ols_flags |= LDLM_FL_DENY_ON_CONTENTION;
1287
1288                         /* a reference for lock, passed as an upcall cookie */
1289                         cl_lock_get(lock);
1290                         lu_ref_add(&lock->cll_reference, "upcall", lock);
1291                         ols->ols_state = OLS_ENQUEUED;
1292
1293                         /*
1294                          * XXX: this is possible blocking point as
1295                          * ldlm_lock_match(LDLM_FL_LVB_READY) waits for
1296                          * LDLM_CP_CALLBACK.
1297                          */
1298                         result = osc_enqueue_base(osc_export(obj), resname,
1299                                           &ols->ols_flags, policy,
1300                                           &ols->ols_lvb,
1301                                           obj->oo_oinfo->loi_kms_valid,
1302                                           osc_lock_upcall,
1303                                           ols, einfo, &ols->ols_handle,
1304                                           PTLRPCD_SET, 1);
1305                         if (result != 0) {
1306                                 lu_ref_del(&lock->cll_reference,
1307                                            "upcall", lock);
1308                                 cl_lock_put(env, lock);
1309                         }
1310                 } else {
1311                         ols->ols_state = OLS_GRANTED;
1312                 }
1313         }
1314         LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
1315         RETURN(result);
1316 }
1317
1318 static int osc_lock_wait(const struct lu_env *env,
1319                          const struct cl_lock_slice *slice)
1320 {
1321         struct osc_lock *olck = cl2osc_lock(slice);
1322         struct cl_lock  *lock = olck->ols_cl.cls_lock;
1323
1324         LINVRNT(osc_lock_invariant(olck));
1325         if (olck->ols_glimpse && olck->ols_state >= OLS_UPCALL_RECEIVED)
1326                 return 0;
1327
1328         LASSERT(equi(olck->ols_state >= OLS_UPCALL_RECEIVED &&
1329                      lock->cll_error == 0, olck->ols_lock != NULL));
1330
1331         return lock->cll_error ?: olck->ols_state >= OLS_GRANTED ? 0 : CLO_WAIT;
1332 }
1333
1334 /**
1335  * An implementation of cl_lock_operations::clo_use() method that pins cached
1336  * lock.
1337  */
1338 static int osc_lock_use(const struct lu_env *env,
1339                         const struct cl_lock_slice *slice)
1340 {
1341         struct osc_lock *olck = cl2osc_lock(slice);
1342         int rc;
1343
1344         LASSERT(!olck->ols_hold);
1345         /*
1346          * Atomically check for LDLM_FL_CBPENDING and addref a lock if this
1347          * flag is not set. This protects us from a concurrent blocking ast.
1348          */
1349         rc = ldlm_lock_addref_try(&olck->ols_handle, olck->ols_einfo.ei_mode);
1350         if (rc == 0) {
1351                 olck->ols_hold = olck->ols_has_ref = 1;
1352                 olck->ols_state = OLS_GRANTED;
1353         } else {
1354                 struct cl_lock *lock;
1355
1356                 /*
1357                  * Lock is being cancelled somewhere within
1358                  * ldlm_handle_bl_callback(): LDLM_FL_CBPENDING is already
1359                  * set, but osc_ldlm_blocking_ast() hasn't yet acquired
1360                  * cl_lock mutex.
1361                  */
1362                 lock = slice->cls_lock;
1363                 LASSERT(lock->cll_state == CLS_CACHED);
1364                 LASSERT(lock->cll_users > 0);
1365                 LASSERT(olck->ols_lock->l_flags & LDLM_FL_CBPENDING);
1366                 /* set a flag for osc_dlm_blocking_ast0() to signal the
1367                  * lock.*/
1368                 olck->ols_ast_wait = 1;
1369                 rc = CLO_WAIT;
1370         }
1371         return rc;
1372 }
1373
1374 static int osc_lock_flush(struct osc_lock *ols, int discard)
1375 {
1376         struct cl_lock       *lock  = ols->ols_cl.cls_lock;
1377         struct cl_env_nest    nest;
1378         struct lu_env        *env;
1379         int result = 0;
1380
1381         env = cl_env_nested_get(&nest);
1382         if (!IS_ERR(env)) {
1383                 result = cl_lock_page_out(env, lock, discard);
1384                 cl_env_nested_put(&nest, env);
1385         } else
1386                 result = PTR_ERR(env);
1387         if (result == 0)
1388                 ols->ols_flush = 1;
1389         return result;
1390 }
1391
1392 /**
1393  * Implements cl_lock_operations::clo_cancel() method for osc layer. This is
1394  * called (as part of cl_lock_cancel()) when lock is canceled either voluntary
1395  * (LRU pressure, early cancellation, umount, etc.) or due to the conflict
1396  * with some other lock some where in the cluster. This function does the
1397  * following:
1398  *
1399  *     - invalidates all pages protected by this lock (after sending dirty
1400  *       ones to the server, as necessary);
1401  *
1402  *     - decref's underlying ldlm lock;
1403  *
1404  *     - cancels ldlm lock (ldlm_cli_cancel()).
1405  */
1406 static void osc_lock_cancel(const struct lu_env *env,
1407                             const struct cl_lock_slice *slice)
1408 {
1409         struct cl_lock   *lock    = slice->cls_lock;
1410         struct osc_lock  *olck    = cl2osc_lock(slice);
1411         struct ldlm_lock *dlmlock = olck->ols_lock;
1412         int               result  = 0;
1413         int               discard;
1414
1415         LASSERT(cl_lock_is_mutexed(lock));
1416         LINVRNT(osc_lock_invariant(olck));
1417
1418         if (dlmlock != NULL) {
1419                 int do_cancel;
1420
1421                 discard = dlmlock->l_flags & LDLM_FL_DISCARD_DATA;
1422                 result = osc_lock_flush(olck, discard);
1423                 if (olck->ols_hold)
1424                         osc_lock_unuse(env, slice);
1425
1426                 lock_res_and_lock(dlmlock);
1427                 /* Now that we're the only user of dlm read/write reference,
1428                  * mostly the ->l_readers + ->l_writers should be zero.
1429                  * However, there is a corner case.
1430                  * See bug 18829 for details.*/
1431                 do_cancel = (dlmlock->l_readers == 0 &&
1432                              dlmlock->l_writers == 0);
1433                 dlmlock->l_flags |= LDLM_FL_CBPENDING;
1434                 unlock_res_and_lock(dlmlock);
1435                 if (do_cancel)
1436                         result = ldlm_cli_cancel(&olck->ols_handle);
1437                 if (result < 0)
1438                         CL_LOCK_DEBUG(D_ERROR, env, lock,
1439                                       "lock %p cancel failure with error(%d)\n",
1440                                       lock, result);
1441         }
1442         olck->ols_state = OLS_CANCELLED;
1443         osc_lock_detach(env, olck);
1444 }
1445
1446 void cl_lock_page_list_fixup(const struct lu_env *env,
1447                              struct cl_io *io, struct cl_lock *lock,
1448                              struct cl_page_list *queue);
1449
1450 #ifdef INVARIANT_CHECK
1451 /**
1452  * Returns true iff there are pages under \a olck not protected by other
1453  * locks.
1454  */
1455 static int osc_lock_has_pages(struct osc_lock *olck)
1456 {
1457         struct cl_lock       *lock;
1458         struct cl_lock_descr *descr;
1459         struct cl_object     *obj;
1460         struct osc_object    *oob;
1461         struct cl_page_list  *plist;
1462         struct cl_page       *page;
1463         struct cl_env_nest    nest;
1464         struct cl_io         *io;
1465         struct lu_env        *env;
1466         int                   result;
1467
1468         env = cl_env_nested_get(&nest);
1469         if (!IS_ERR(env)) {
1470                 obj   = olck->ols_cl.cls_obj;
1471                 oob   = cl2osc(obj);
1472                 io    = &oob->oo_debug_io;
1473                 lock  = olck->ols_cl.cls_lock;
1474                 descr = &lock->cll_descr;
1475                 plist = &osc_env_info(env)->oti_plist;
1476                 cl_page_list_init(plist);
1477
1478                 mutex_lock(&oob->oo_debug_mutex);
1479
1480                 io->ci_obj = cl_object_top(obj);
1481                 cl_io_init(env, io, CIT_MISC, io->ci_obj);
1482                 cl_page_gang_lookup(env, obj, io,
1483                                     descr->cld_start, descr->cld_end, plist);
1484                 cl_lock_page_list_fixup(env, io, lock, plist);
1485                 if (plist->pl_nr > 0) {
1486                         CL_LOCK_DEBUG(D_ERROR, env, lock, "still has pages\n");
1487                         cl_page_list_for_each(page, plist)
1488                                 CL_PAGE_DEBUG(D_ERROR, env, page, "\n");
1489                 }
1490                 result = plist->pl_nr > 0;
1491                 cl_page_list_disown(env, io, plist);
1492                 cl_page_list_fini(env, plist);
1493                 cl_io_fini(env, io);
1494                 mutex_unlock(&oob->oo_debug_mutex);
1495                 cl_env_nested_put(&nest, env);
1496         } else
1497                 result = 0;
1498         return result;
1499 }
1500 #else
1501 # define osc_lock_has_pages(olck) (0)
1502 #endif /* INVARIANT_CHECK */
1503
1504 static void osc_lock_delete(const struct lu_env *env,
1505                             const struct cl_lock_slice *slice)
1506 {
1507         struct osc_lock *olck;
1508
1509         olck = cl2osc_lock(slice);
1510         LINVRNT(osc_lock_invariant(olck));
1511         LINVRNT(!osc_lock_has_pages(olck));
1512
1513         if (olck->ols_hold)
1514                 osc_lock_unuse(env, slice);
1515         osc_lock_detach(env, olck);
1516 }
1517
1518 /**
1519  * Implements cl_lock_operations::clo_state() method for osc layer.
1520  *
1521  * Maintains osc_lock::ols_owner field.
1522  *
1523  * This assumes that lock always enters CLS_HELD (from some other state) in
1524  * the same IO context as one that requested the lock. This should not be a
1525  * problem, because context is by definition shared by all activity pertaining
1526  * to the same high-level IO.
1527  */
1528 static void osc_lock_state(const struct lu_env *env,
1529                            const struct cl_lock_slice *slice,
1530                            enum cl_lock_state state)
1531 {
1532         struct osc_lock *lock = cl2osc_lock(slice);
1533         struct osc_io   *oio  = osc_env_io(env);
1534
1535         /*
1536          * XXX multiple io contexts can use the lock at the same time.
1537          */
1538         LINVRNT(osc_lock_invariant(lock));
1539         if (state == CLS_HELD && slice->cls_lock->cll_state != CLS_HELD) {
1540                 LASSERT(lock->ols_owner == NULL);
1541                 lock->ols_owner = oio;
1542         } else if (state != CLS_HELD)
1543                 lock->ols_owner = NULL;
1544 }
1545
1546 static int osc_lock_print(const struct lu_env *env, void *cookie,
1547                           lu_printer_t p, const struct cl_lock_slice *slice)
1548 {
1549         struct osc_lock *lock = cl2osc_lock(slice);
1550
1551         /*
1552          * XXX print ldlm lock and einfo properly.
1553          */
1554         (*p)(env, cookie, "%p %08x "LPU64" %d %p ",
1555              lock->ols_lock, lock->ols_flags, lock->ols_handle.cookie,
1556              lock->ols_state, lock->ols_owner);
1557         osc_lvb_print(env, cookie, p, &lock->ols_lvb);
1558         return 0;
1559 }
1560
1561 static const struct cl_lock_operations osc_lock_ops = {
1562         .clo_fini    = osc_lock_fini,
1563         .clo_enqueue = osc_lock_enqueue,
1564         .clo_wait    = osc_lock_wait,
1565         .clo_unuse   = osc_lock_unuse,
1566         .clo_use     = osc_lock_use,
1567         .clo_delete  = osc_lock_delete,
1568         .clo_state   = osc_lock_state,
1569         .clo_cancel  = osc_lock_cancel,
1570         .clo_weigh   = osc_lock_weigh,
1571         .clo_print   = osc_lock_print
1572 };
1573
1574 static int osc_lock_lockless_enqueue(const struct lu_env *env,
1575                                      const struct cl_lock_slice *slice,
1576                                      struct cl_io *_, __u32 enqflags)
1577 {
1578         LBUG();
1579         return 0;
1580 }
1581
1582 static int osc_lock_lockless_unuse(const struct lu_env *env,
1583                                    const struct cl_lock_slice *slice)
1584 {
1585         struct osc_lock *ols = cl2osc_lock(slice);
1586         struct cl_lock *lock = slice->cls_lock;
1587
1588         LASSERT(ols->ols_state == OLS_GRANTED);
1589         LINVRNT(osc_lock_invariant(ols));
1590
1591         cl_lock_cancel(env, lock);
1592         cl_lock_delete(env, lock);
1593         return 0;
1594 }
1595
1596 static void osc_lock_lockless_cancel(const struct lu_env *env,
1597                                      const struct cl_lock_slice *slice)
1598 {
1599         struct osc_lock   *ols  = cl2osc_lock(slice);
1600         int result;
1601
1602         result = osc_lock_flush(ols, 0);
1603         if (result)
1604                 CERROR("Pages for lockless lock %p were not purged(%d)\n",
1605                        ols, result);
1606         ols->ols_state = OLS_CANCELLED;
1607 }
1608
1609 static int osc_lock_lockless_wait(const struct lu_env *env,
1610                                   const struct cl_lock_slice *slice)
1611 {
1612         struct osc_lock *olck = cl2osc_lock(slice);
1613         struct cl_lock  *lock = olck->ols_cl.cls_lock;
1614
1615         LINVRNT(osc_lock_invariant(olck));
1616         LASSERT(olck->ols_state >= OLS_UPCALL_RECEIVED);
1617
1618         return lock->cll_error;
1619 }
1620
1621 static void osc_lock_lockless_state(const struct lu_env *env,
1622                                     const struct cl_lock_slice *slice,
1623                                     enum cl_lock_state state)
1624 {
1625         struct osc_lock *lock = cl2osc_lock(slice);
1626         struct osc_io   *oio  = osc_env_io(env);
1627
1628         LINVRNT(osc_lock_invariant(lock));
1629         if (state == CLS_HELD) {
1630                 LASSERT(lock->ols_owner == NULL);
1631                 lock->ols_owner = oio;
1632
1633                 /* set the io to be lockless if this lock is for io's
1634                  * host object */
1635                 if (cl_object_same(oio->oi_cl.cis_obj, slice->cls_obj))
1636                         oio->oi_lockless = 1;
1637         } else
1638                 lock->ols_owner = NULL;
1639 }
1640
1641 static int osc_lock_lockless_fits_into(const struct lu_env *env,
1642                                        const struct cl_lock_slice *slice,
1643                                        const struct cl_lock_descr *need,
1644                                        const struct cl_io *io)
1645 {
1646         return 0;
1647 }
1648
1649 static const struct cl_lock_operations osc_lock_lockless_ops = {
1650         .clo_fini      = osc_lock_fini,
1651         .clo_enqueue   = osc_lock_lockless_enqueue,
1652         .clo_wait      = osc_lock_lockless_wait,
1653         .clo_unuse     = osc_lock_lockless_unuse,
1654         .clo_state     = osc_lock_lockless_state,
1655         .clo_fits_into = osc_lock_lockless_fits_into,
1656         .clo_cancel    = osc_lock_lockless_cancel,
1657         .clo_print     = osc_lock_print
1658 };
1659
1660 int osc_lock_init(const struct lu_env *env,
1661                   struct cl_object *obj, struct cl_lock *lock,
1662                   const struct cl_io *_)
1663 {
1664         struct osc_lock *clk;
1665         int result;
1666
1667         OBD_SLAB_ALLOC_PTR_GFP(clk, osc_lock_kmem, CFS_ALLOC_IO);
1668         if (clk != NULL) {
1669                 osc_lock_build_einfo(env, lock, clk, &clk->ols_einfo);
1670                 clk->ols_state = OLS_NEW;
1671                 cl_lock_slice_add(lock, &clk->ols_cl, obj, &osc_lock_ops);
1672                 result = 0;
1673         } else
1674                 result = -ENOMEM;
1675         return result;
1676 }
1677
1678
1679 /** @} osc */