lustre/ldlm/ldlm_extent.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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2010, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/ldlm/ldlm_extent.c
  33  *
  34  * Author: Peter Braam <braam@clusterfs.com>
  35  * Author: Phil Schwan <phil@clusterfs.com>
  36  */
  37
  38 /**
  39  * This file contains implementation of EXTENT lock type
  40  *
  41  * EXTENT lock type is for locking a contiguous range of values, represented
  42  * by 64-bit starting and ending offsets (inclusive). There are several extent
  43  * lock modes, some of which may be mutually incompatible. Extent locks are
  44  * considered incompatible if their modes are incompatible and their extents
  45  * intersect.  See the lock mode compatibility matrix in lustre_dlm.h.
  46  */
  47
  48 #define DEBUG_SUBSYSTEM S_LDLM
  49
  50 #include <libcfs/libcfs.h>
  51 #include <lustre_dlm.h>
  52 #include <obd_support.h>
  53 #include <obd.h>
  54 #include <obd_class.h>
  55 #include <lustre_lib.h>
  56
  57 #include "ldlm_internal.h"
  58
  59 #ifdef HAVE_SERVER_SUPPORT
  60 # define LDLM_MAX_GROWN_EXTENT (32 * 1024 * 1024 - 1)
  61
  62 /**
  63  * Fix up the ldlm_extent after expanding it.
  64  *
  65  * After expansion has been done, we might still want to do certain adjusting
  66  * based on overall contention of the resource and the like to avoid granting
  67  * overly wide locks.
  68  */
  69 static void ldlm_extent_internal_policy_fixup(struct ldlm_lock *req,
  70                                               struct ldlm_extent *new_ex,
  71                                               int conflicting)
  72 {
  73         enum ldlm_mode req_mode = req->l_req_mode;
  74         __u64 req_start = req->l_req_extent.start;
  75         __u64 req_end = req->l_req_extent.end;
  76         __u64 req_align, mask;
  77
  78         if (conflicting > 32 && (req_mode == LCK_PW || req_mode == LCK_CW)) {
  79                 if (req_end < req_start + LDLM_MAX_GROWN_EXTENT)
  80                         new_ex->end = min(req_start + LDLM_MAX_GROWN_EXTENT,
  81                                           new_ex->end);
  82         }
  83
  84         if (new_ex->start == 0 && new_ex->end == OBD_OBJECT_EOF) {
  85                 EXIT;
  86                 return;
  87         }
  88
  89         /* we need to ensure that the lock extent is properly aligned to what
  90          * the client requested. Also we need to make sure it's also server
  91          * page size aligned otherwise a server page can be covered by two
  92          * write locks. */
  93         mask = PAGE_SIZE;
  94         req_align = (req_end + 1) | req_start;
  95         if (req_align != 0 && (req_align & (mask - 1)) == 0) {
  96                 while ((req_align & mask) == 0)
  97                         mask <<= 1;
  98         }
  99         mask -= 1;
 100         /* We can only shrink the lock, not grow it.
 101          * This should never cause lock to be smaller than requested,
 102          * since requested lock was already aligned on these boundaries. */
 103         new_ex->start = ((new_ex->start - 1) | mask) + 1;
 104         new_ex->end = ((new_ex->end + 1) & ~mask) - 1;
 105         LASSERTF(new_ex->start <= req_start,
 106                  "mask %#llx grant start %llu req start %llu\n",
 107                  mask, new_ex->start, req_start);
 108         LASSERTF(new_ex->end >= req_end,
 109                  "mask %#llx grant end %llu req end %llu\n",
 110                  mask, new_ex->end, req_end);
 111 }
 112
 113 /**
 114  * Return the maximum extent that:
 115  * - contains the requested extent
 116  * - does not overlap existing conflicting extents outside the requested one
 117  *
 118  * This allows clients to request a small required extent range, but if there
 119  * is no contention on the lock the full lock can be granted to the client.
 120  * This avoids the need for many smaller lock requests to be granted in the
 121  * common (uncontended) case.
 122  *
 123  * Use interval tree to expand the lock extent for granted lock.
 124  */
 125 static void ldlm_extent_internal_policy_granted(struct ldlm_lock *req,
 126                                                 struct ldlm_extent *new_ex)
 127 {
 128         struct ldlm_resource *res = req->l_resource;
 129         enum ldlm_mode req_mode = req->l_req_mode;
 130         __u64 req_start = req->l_req_extent.start;
 131         __u64 req_end = req->l_req_extent.end;
 132         struct ldlm_interval_tree *tree;
 133         struct interval_node_extent limiter = {
 134                 .start  = new_ex->start,
 135                 .end    = new_ex->end,
 136         };
 137         int conflicting = 0;
 138         int idx;
 139         ENTRY;
 140
 141         lockmode_verify(req_mode);
 142
 143         /* Using interval tree to handle the LDLM extent granted locks. */
 144         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
 145                 struct interval_node_extent ext = {
 146                         .start  = req_start,
 147                         .end    = req_end,
 148                 };
 149
 150                 tree = &res->lr_itree[idx];
 151                 if (lockmode_compat(tree->lit_mode, req_mode))
 152                         continue;
 153
 154                 conflicting += tree->lit_size;
 155                 if (conflicting > 4)
 156                         limiter.start = req_start;
 157
 158                 if (interval_is_overlapped(tree->lit_root, &ext))
 159                         CDEBUG(D_INFO,
 160                                "req_mode = %d, tree->lit_mode = %d, "
 161                                "tree->lit_size = %d\n",
 162                                req_mode, tree->lit_mode, tree->lit_size);
 163                 interval_expand(tree->lit_root, &ext, &limiter);
 164                 limiter.start = max(limiter.start, ext.start);
 165                 limiter.end = min(limiter.end, ext.end);
 166                 if (limiter.start == req_start && limiter.end == req_end)
 167                         break;
 168         }
 169
 170         new_ex->start = limiter.start;
 171         new_ex->end = limiter.end;
 172         LASSERT(new_ex->start <= req_start);
 173         LASSERT(new_ex->end >= req_end);
 174
 175         ldlm_extent_internal_policy_fixup(req, new_ex, conflicting);
 176         EXIT;
 177 }
 178
 179 /* The purpose of this function is to return:
 180  * - the maximum extent
 181  * - containing the requested extent
 182  * - and not overlapping existing conflicting extents outside the requested one
 183  */
 184 static void
 185 ldlm_extent_internal_policy_waiting(struct ldlm_lock *req,
 186                                     struct ldlm_extent *new_ex)
 187 {
 188         struct ldlm_resource *res = req->l_resource;
 189         enum ldlm_mode req_mode = req->l_req_mode;
 190         __u64 req_start = req->l_req_extent.start;
 191         __u64 req_end = req->l_req_extent.end;
 192         struct ldlm_lock *lock;
 193         int conflicting = 0;
 194         ENTRY;
 195
 196         lockmode_verify(req_mode);
 197
 198         /* for waiting locks */
 199         list_for_each_entry(lock, &res->lr_waiting, l_res_link) {
 200                 struct ldlm_extent *l_extent = &lock->l_policy_data.l_extent;
 201
 202                 /* We already hit the minimum requested size, search no more */
 203                 if (new_ex->start == req_start && new_ex->end == req_end) {
 204                         EXIT;
 205                         return;
 206                 }
 207
 208                 /* Don't conflict with ourselves */
 209                 if (req == lock)
 210                         continue;
 211
 212                 /* Locks are compatible, overlap doesn't matter */
 213                 /* Until bug 20 is fixed, try to avoid granting overlapping
 214                  * locks on one client (they take a long time to cancel) */
 215                 if (lockmode_compat(lock->l_req_mode, req_mode) &&
 216                     lock->l_export != req->l_export)
 217                         continue;
 218
 219                 /* If this is a high-traffic lock, don't grow downwards at all
 220                  * or grow upwards too much */
 221                 ++conflicting;
 222                 if (conflicting > 4)
 223                         new_ex->start = req_start;
 224
 225                 /* If lock doesn't overlap new_ex, skip it. */
 226                 if (!ldlm_extent_overlap(l_extent, new_ex))
 227                         continue;
 228
 229                 /* Locks conflicting in requested extents and we can't satisfy
 230                  * both locks, so ignore it.  Either we will ping-pong this
 231                  * extent (we would regardless of what extent we granted) or
 232                  * lock is unused and it shouldn't limit our extent growth. */
 233                 if (ldlm_extent_overlap(&lock->l_req_extent,&req->l_req_extent))
 234                         continue;
 235
 236                 /* We grow extents downwards only as far as they don't overlap
 237                  * with already-granted locks, on the assumption that clients
 238                  * will be writing beyond the initial requested end and would
 239                  * then need to enqueue a new lock beyond previous request.
 240                  * l_req_extent->end strictly < req_start, checked above. */
 241                 if (l_extent->start < req_start && new_ex->start != req_start) {
 242                         if (l_extent->end >= req_start)
 243                                 new_ex->start = req_start;
 244                         else
 245                                 new_ex->start = min(l_extent->end+1, req_start);
 246                 }
 247
 248                 /* If we need to cancel this lock anyways because our request
 249                  * overlaps the granted lock, we grow up to its requested
 250                  * extent start instead of limiting this extent, assuming that
 251                  * clients are writing forwards and the lock had over grown
 252                  * its extent downwards before we enqueued our request. */
 253                 if (l_extent->end > req_end) {
 254                         if (l_extent->start <= req_end)
 255                                 new_ex->end = max(lock->l_req_extent.start - 1,
 256                                                   req_end);
 257                         else
 258                                 new_ex->end = max(l_extent->start - 1, req_end);
 259                 }
 260         }
 261
 262         ldlm_extent_internal_policy_fixup(req, new_ex, conflicting);
 263         EXIT;
 264 }
 265
 266
 267 /* In order to determine the largest possible extent we can grant, we need
 268  * to scan all of the queues. */
 269 static void ldlm_extent_policy(struct ldlm_resource *res,
 270                                struct ldlm_lock *lock, __u64 *flags)
 271 {
 272         struct ldlm_extent new_ex = { .start = 0, .end = OBD_OBJECT_EOF };
 273
 274         if (lock->l_export == NULL)
 275                 /*
 276                  * this is a local lock taken by server (e.g., as a part of
 277                  * OST-side locking, or unlink handling). Expansion doesn't
 278                  * make a lot of sense for local locks, because they are
 279                  * dropped immediately on operation completion and would only
 280                  * conflict with other threads.
 281                  */
 282                 return;
 283
 284         if (lock->l_policy_data.l_extent.start == 0 &&
 285             lock->l_policy_data.l_extent.end == OBD_OBJECT_EOF)
 286                 /* fast-path whole file locks */
 287                 return;
 288
 289         /* Because reprocess_queue zeroes flags and uses it to return
 290          * LDLM_FL_LOCK_CHANGED, we must check for the NO_EXPANSION flag
 291          * in the lock flags rather than the 'flags' argument */
 292         if (likely(!(lock->l_flags & LDLM_FL_NO_EXPANSION))) {
 293                 ldlm_extent_internal_policy_granted(lock, &new_ex);
 294                 ldlm_extent_internal_policy_waiting(lock, &new_ex);
 295         } else {
 296                 LDLM_DEBUG(lock, "Not expanding manually requested lock.\n");
 297                 new_ex.start = lock->l_policy_data.l_extent.start;
 298                 new_ex.end = lock->l_policy_data.l_extent.end;
 299                 /* In case the request is not on correct boundaries, we call
 300                  * fixup. (normally called in ldlm_extent_internal_policy_*) */
 301                 ldlm_extent_internal_policy_fixup(lock, &new_ex, 0);
 302         }
 303
 304         if (!ldlm_extent_equal(&new_ex, &lock->l_policy_data.l_extent)) {
 305                 *flags |= LDLM_FL_LOCK_CHANGED;
 306                 lock->l_policy_data.l_extent.start = new_ex.start;
 307                 lock->l_policy_data.l_extent.end = new_ex.end;
 308         }
 309 }
 310
 311 static bool ldlm_check_contention(struct ldlm_lock *lock, int contended_locks)
 312 {
 313         struct ldlm_resource *res = lock->l_resource;
 314         time64_t now = ktime_get_seconds();
 315
 316         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_SET_CONTENTION))
 317                 return true;
 318
 319         CDEBUG(D_DLMTRACE, "contended locks = %d\n", contended_locks);
 320         if (contended_locks > ldlm_res_to_ns(res)->ns_contended_locks)
 321                 res->lr_contention_time = now;
 322
 323         return now < res->lr_contention_time +
 324                      ldlm_res_to_ns(res)->ns_contention_time;
 325 }
 326
 327 struct ldlm_extent_compat_args {
 328         struct list_head *work_list;
 329         struct ldlm_lock *lock;
 330         enum ldlm_mode mode;
 331         int *locks;
 332         int *compat;
 333 };
 334
 335 static enum interval_iter ldlm_extent_compat_cb(struct interval_node *n,
 336                                                 void *data)
 337 {
 338         struct ldlm_extent_compat_args *priv = data;
 339         struct ldlm_interval *node = to_ldlm_interval(n);
 340         struct ldlm_extent *extent;
 341         struct list_head *work_list = priv->work_list;
 342         struct ldlm_lock *lock, *enq = priv->lock;
 343         enum ldlm_mode mode = priv->mode;
 344         int count = 0;
 345         ENTRY;
 346
 347         LASSERT(!list_empty(&node->li_group));
 348
 349         list_for_each_entry(lock, &node->li_group, l_sl_policy) {
 350                 /* interval tree is for granted lock */
 351                 LASSERTF(mode == lock->l_granted_mode,
 352                          "mode = %s, lock->l_granted_mode = %s\n",
 353                          ldlm_lockname[mode],
 354                          ldlm_lockname[lock->l_granted_mode]);
 355                 count++;
 356                 if (lock->l_blocking_ast &&
 357                     lock->l_granted_mode != LCK_GROUP)
 358                         ldlm_add_ast_work_item(lock, enq, work_list);
 359         }
 360
 361         /* don't count conflicting glimpse locks */
 362         extent = ldlm_interval_extent(node);
 363         if (!(mode == LCK_PR &&
 364             extent->start == 0 && extent->end == OBD_OBJECT_EOF))
 365                 *priv->locks += count;
 366
 367         if (priv->compat)
 368                 *priv->compat = 0;
 369
 370         RETURN(INTERVAL_ITER_CONT);
 371 }
 372
 373 /**
 374  * Determine if the lock is compatible with all locks on the queue.
 375  *
 376  * If \a work_list is provided, conflicting locks are linked there.
 377  * If \a work_list is not provided, we exit this function on first conflict.
 378  *
 379  * \retval 0 if the lock is not compatible
 380  * \retval 1 if the lock is compatible
 381  * \retval 2 if \a req is a group lock and it is compatible and requires
 382  *           no further checking
 383  * \retval negative error, such as EAGAIN for group locks
 384  */
 385 static int
 386 ldlm_extent_compat_queue(struct list_head *queue, struct ldlm_lock *req,
 387                          __u64 *flags, struct list_head *work_list,
 388                          int *contended_locks)
 389 {
 390         struct ldlm_resource *res = req->l_resource;
 391         enum ldlm_mode req_mode = req->l_req_mode;
 392         __u64 req_start = req->l_req_extent.start;
 393         __u64 req_end = req->l_req_extent.end;
 394         struct ldlm_lock *lock;
 395         int check_contention;
 396         int compat = 1;
 397         ENTRY;
 398
 399         lockmode_verify(req_mode);
 400
 401         /* Using interval tree for granted lock */
 402         if (queue == &res->lr_granted) {
 403                 struct ldlm_interval_tree *tree;
 404                 struct ldlm_extent_compat_args data = {.work_list = work_list,
 405                                                .lock = req,
 406                                                .locks = contended_locks,
 407                                                .compat = &compat };
 408                 struct interval_node_extent ex = { .start = req_start,
 409                                                    .end = req_end };
 410                 int idx, rc;
 411
 412                 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
 413                         tree = &res->lr_itree[idx];
 414                         if (tree->lit_root == NULL) /* empty tree, skipped */
 415                                 continue;
 416
 417                         data.mode = tree->lit_mode;
 418                         if (lockmode_compat(req_mode, tree->lit_mode)) {
 419                                 struct ldlm_interval *node;
 420                                 struct ldlm_extent *extent;
 421
 422                                 if (req_mode != LCK_GROUP)
 423                                         continue;
 424
 425                                 /* group lock, grant it immediately if
 426                                  * compatible */
 427                                 node = to_ldlm_interval(tree->lit_root);
 428                                 extent = ldlm_interval_extent(node);
 429                                 if (req->l_policy_data.l_extent.gid ==
 430                                     extent->gid)
 431                                         RETURN(2);
 432                         }
 433
 434                         if (tree->lit_mode == LCK_GROUP) {
 435                                 if (*flags & (LDLM_FL_BLOCK_NOWAIT |
 436                                               LDLM_FL_SPECULATIVE)) {
 437                                         compat = -EAGAIN;
 438                                         goto destroylock;
 439                                 }
 440
 441                                 if (!work_list)
 442                                         RETURN(0);
 443
 444                                 /* if work list is not NULL,add all
 445                                    locks in the tree to work list */
 446                                 compat = 0;
 447                                 interval_iterate(tree->lit_root,
 448                                                  ldlm_extent_compat_cb, &data);
 449                                 continue;
 450                         }
 451
 452                         /* We've found a potentially blocking lock, check
 453                          * compatibility.  This handles locks other than GROUP
 454                          * locks, which are handled separately above.
 455                          *
 456                          * Locks with FL_SPECULATIVE are asynchronous requests
 457                          * which must never wait behind another lock, so they
 458                          * fail if any conflicting lock is found. */
 459                         if (!work_list || (*flags & LDLM_FL_SPECULATIVE)) {
 460                                 rc = interval_is_overlapped(tree->lit_root,
 461                                                             &ex);
 462                                 if (rc) {
 463                                         if (!work_list) {
 464                                                 RETURN(0);
 465                                         } else {
 466                                                 compat = -EAGAIN;
 467                                                 goto destroylock;
 468                                         }
 469                                 }
 470                         } else {
 471                                 interval_search(tree->lit_root, &ex,
 472                                                 ldlm_extent_compat_cb, &data);
 473                                 if (!list_empty(work_list) && compat)
 474                                         compat = 0;
 475                         }
 476                 }
 477         } else { /* for waiting queue */
 478                 list_for_each_entry(lock, queue, l_res_link) {
 479                         check_contention = 1;
 480
 481                         /* We stop walking the queue if we hit ourselves so
 482                          * we don't take conflicting locks enqueued after us
 483                          * into account, or we'd wait forever. */
 484                         if (req == lock)
 485                                 break;
 486
 487                         /* locks are compatible, overlap doesn't matter */
 488                         if (lockmode_compat(lock->l_req_mode, req_mode)) {
 489                                 if (req_mode == LCK_PR &&
 490                                     ((lock->l_policy_data.l_extent.start <=
 491                                       req->l_policy_data.l_extent.start) &&
 492                                      (lock->l_policy_data.l_extent.end >=
 493                                       req->l_policy_data.l_extent.end))) {
 494                                         /* If we met a PR lock just like us or
 495                                            wider, and nobody down the list
 496                                            conflicted with it, that means we
 497                                            can skip processing of the rest of
 498                                            the list and safely place ourselves
 499                                            at the end of the list, or grant
 500                                            (dependent if we met an conflicting
 501                                            locks before in the list).  In case
 502                                            of 1st enqueue only we continue
 503                                            traversing if there is something
 504                                            conflicting down the list because
 505                                            we need to make sure that something
 506                                            is marked as AST_SENT as well, in
 507                                            cse of empy worklist we would exit
 508                                            on first conflict met. */
 509                                         /* There IS a case where such flag is
 510                                            not set for a lock, yet it blocks
 511                                            something. Luckily for us this is
 512                                            only during destroy, so lock is
 513                                            exclusive. So here we are safe */
 514                                         if (!ldlm_is_ast_sent(lock))
 515                                                 RETURN(compat);
 516                                 }
 517
 518                                 /* non-group locks are compatible, overlap doesn't
 519                                    matter */
 520                                 if (likely(req_mode != LCK_GROUP))
 521                                         continue;
 522
 523                                 /* If we are trying to get a GROUP lock and there is
 524                                    another one of this kind, we need to compare gid */
 525                                 if (req->l_policy_data.l_extent.gid ==
 526                                     lock->l_policy_data.l_extent.gid) {
 527                                         /* If existing lock with matched gid is granted,
 528                                            we grant new one too. */
 529                                         if (ldlm_is_granted(lock))
 530                                                 RETURN(2);
 531
 532                                         /* Otherwise we are scanning queue of waiting
 533                                          * locks and it means current request would
 534                                          * block along with existing lock (that is
 535                                          * already blocked.
 536                                          * If we are in nonblocking mode - return
 537                                          * immediately */
 538                                         if (*flags & (LDLM_FL_BLOCK_NOWAIT
 539                                                       | LDLM_FL_SPECULATIVE)) {
 540                                                 compat = -EAGAIN;
 541                                                 goto destroylock;
 542                                         }
 543                                         /* If this group lock is compatible with another
 544                                          * group lock on the waiting list, they must be
 545                                          * together in the list, so they can be granted
 546                                          * at the same time.  Otherwise the later lock
 547                                          * can get stuck behind another, incompatible,
 548                                          * lock. */
 549                                         ldlm_resource_insert_lock_after(lock, req);
 550                                         /* Because 'lock' is not granted, we can stop
 551                                          * processing this queue and return immediately.
 552                                          * There is no need to check the rest of the
 553                                          * list. */
 554                                         RETURN(0);
 555                                 }
 556                         }
 557
 558                         if (unlikely(req_mode == LCK_GROUP &&
 559                                      !ldlm_is_granted(lock))) {
 560                                 compat = 0;
 561                                 if (lock->l_req_mode != LCK_GROUP) {
 562                                         /* Ok, we hit non-GROUP lock, there should be no
 563                                            more GROUP locks later on, queue in front of
 564                                            first non-GROUP lock */
 565
 566                                         ldlm_resource_insert_lock_before(lock, req);
 567                                         break;
 568                                 }
 569                                 LASSERT(req->l_policy_data.l_extent.gid !=
 570                                         lock->l_policy_data.l_extent.gid);
 571                                 continue;
 572                         }
 573
 574                         if (unlikely(lock->l_req_mode == LCK_GROUP)) {
 575                                 /* If compared lock is GROUP, then requested is
 576                                  * PR/PW so this is not compatible; extent
 577                                  * range does not matter */
 578                                 if (*flags & (LDLM_FL_BLOCK_NOWAIT
 579                                               | LDLM_FL_SPECULATIVE)) {
 580                                         compat = -EAGAIN;
 581                                         goto destroylock;
 582                                 }
 583                         } else if (lock->l_policy_data.l_extent.end < req_start ||
 584                                    lock->l_policy_data.l_extent.start > req_end) {
 585                                 /* if a non group lock doesn't overlap skip it */
 586                                 continue;
 587                         } else if (lock->l_req_extent.end < req_start ||
 588                                    lock->l_req_extent.start > req_end) {
 589                                 /* false contention, the requests doesn't really overlap */
 590                                 check_contention = 0;
 591                         }
 592
 593                         if (!work_list)
 594                                 RETURN(0);
 595
 596                         if (*flags & LDLM_FL_SPECULATIVE) {
 597                                 compat = -EAGAIN;
 598                                 goto destroylock;
 599                         }
 600
 601                         /* don't count conflicting glimpse locks */
 602                         if (lock->l_req_mode == LCK_PR &&
 603                             lock->l_policy_data.l_extent.start == 0 &&
 604                             lock->l_policy_data.l_extent.end == OBD_OBJECT_EOF)
 605                                 check_contention = 0;
 606
 607                         *contended_locks += check_contention;
 608
 609                         compat = 0;
 610                         if (lock->l_blocking_ast &&
 611                             lock->l_req_mode != LCK_GROUP)
 612                                 ldlm_add_ast_work_item(lock, req, work_list);
 613                 }
 614         }
 615
 616         if (ldlm_check_contention(req, *contended_locks) &&
 617             compat == 0 &&
 618             (*flags & LDLM_FL_DENY_ON_CONTENTION) &&
 619             req->l_req_mode != LCK_GROUP &&
 620             req_end - req_start <=
 621             ldlm_res_to_ns(req->l_resource)->ns_max_nolock_size)
 622                 GOTO(destroylock, compat = -EUSERS);
 623
 624         RETURN(compat);
 625 destroylock:
 626         list_del_init(&req->l_res_link);
 627         ldlm_lock_destroy_nolock(req);
 628         RETURN(compat);
 629 }
 630
 631 /**
 632  * This function refresh eviction timer for cancelled lock.
 633  * \param[in] lock              ldlm lock for refresh
 634  * \param[in] arg               ldlm prolong arguments, timeout, export, extent
 635  *                              and counter are used
 636  */
 637 void ldlm_lock_prolong_one(struct ldlm_lock *lock,
 638                            struct ldlm_prolong_args *arg)
 639 {
 640         timeout_t timeout;
 641
 642         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PROLONG_PAUSE, 3);
 643
 644         if (arg->lpa_export != lock->l_export ||
 645             lock->l_flags & LDLM_FL_DESTROYED)
 646                 /* ignore unrelated locks */
 647                 return;
 648
 649         arg->lpa_locks_cnt++;
 650
 651         if (!(lock->l_flags & LDLM_FL_AST_SENT))
 652                 /* ignore locks not being cancelled */
 653                 return;
 654
 655         /* We are in the middle of the process - BL AST is sent, CANCEL
 656          * is ahead. Take half of BL AT + IO AT process time.
 657          */
 658         timeout = arg->lpa_timeout + (ldlm_bl_timeout(lock) >> 1);
 659
 660         arg->lpa_blocks_cnt++;
 661
 662         /* OK. this is a possible lock the user holds doing I/O
 663          * let's refresh eviction timer for it.
 664          */
 665         ldlm_refresh_waiting_lock(lock, timeout);
 666 }
 667 EXPORT_SYMBOL(ldlm_lock_prolong_one);
 668
 669 static enum interval_iter ldlm_resource_prolong_cb(struct interval_node *n,
 670                                                    void *data)
 671 {
 672         struct ldlm_prolong_args *arg = data;
 673         struct ldlm_interval *node = to_ldlm_interval(n);
 674         struct ldlm_lock *lock;
 675
 676         ENTRY;
 677
 678         LASSERT(!list_empty(&node->li_group));
 679
 680         list_for_each_entry(lock, &node->li_group, l_sl_policy) {
 681                 ldlm_lock_prolong_one(lock, arg);
 682         }
 683
 684         RETURN(INTERVAL_ITER_CONT);
 685 }
 686
 687 /**
 688  * Walk through granted tree and prolong locks if they overlaps extent.
 689  *
 690  * \param[in] arg               prolong args
 691  */
 692 void ldlm_resource_prolong(struct ldlm_prolong_args *arg)
 693 {
 694         struct ldlm_interval_tree *tree;
 695         struct ldlm_resource *res;
 696         struct interval_node_extent ex = { .start = arg->lpa_extent.start,
 697                                            .end = arg->lpa_extent.end };
 698         int idx;
 699
 700         ENTRY;
 701
 702         res = ldlm_resource_get(arg->lpa_export->exp_obd->obd_namespace, NULL,
 703                                 &arg->lpa_resid, LDLM_EXTENT, 0);
 704         if (IS_ERR(res)) {
 705                 CDEBUG(D_DLMTRACE, "Failed to get resource for resid %llu/%llu\n",
 706                        arg->lpa_resid.name[0], arg->lpa_resid.name[1]);
 707                 RETURN_EXIT;
 708         }
 709
 710         lock_res(res);
 711         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
 712                 tree = &res->lr_itree[idx];
 713                 if (tree->lit_root == NULL) /* empty tree, skipped */
 714                         continue;
 715
 716                 /* There is no possibility to check for the groupID
 717                  * so all the group locks are considered as valid
 718                  * here, especially because the client is supposed
 719                  * to check it has such a lock before sending an RPC.
 720                  */
 721                 if (!(tree->lit_mode & arg->lpa_mode))
 722                         continue;
 723
 724                 interval_search(tree->lit_root, &ex,
 725                                 ldlm_resource_prolong_cb, arg);
 726         }
 727
 728         unlock_res(res);
 729         ldlm_resource_putref(res);
 730
 731         EXIT;
 732 }
 733 EXPORT_SYMBOL(ldlm_resource_prolong);
 734
 735 /**
 736  * Process a granting attempt for extent lock.
 737  * Must be called with ns lock held.
 738  *
 739  * This function looks for any conflicts for \a lock in the granted or
 740  * waiting queues. The lock is granted if no conflicts are found in
 741  * either queue.
 742  */
 743 int ldlm_process_extent_lock(struct ldlm_lock *lock, __u64 *flags,
 744                              enum ldlm_process_intention intention,
 745                              enum ldlm_error *err, struct list_head *work_list)
 746 {
 747         struct ldlm_resource *res = lock->l_resource;
 748         int rc, rc2 = 0;
 749         int contended_locks = 0;
 750         struct list_head *grant_work = intention == LDLM_PROCESS_ENQUEUE ?
 751                                                         NULL : work_list;
 752         ENTRY;
 753
 754         LASSERT(!ldlm_is_granted(lock));
 755         LASSERT(!(*flags & LDLM_FL_DENY_ON_CONTENTION) ||
 756                 !ldlm_is_ast_discard_data(lock));
 757         check_res_locked(res);
 758         *err = ELDLM_OK;
 759
 760         if (intention == LDLM_PROCESS_RESCAN) {
 761                 /* Careful observers will note that we don't handle -EAGAIN
 762                  * here, but it's ok for a non-obvious reason -- compat_queue
 763                  * can only return -EAGAIN if (flags & BLOCK_NOWAIT |
 764                  * SPECULATIVE). flags should always be zero here, and if that
 765                  * ever stops being true, we want to find out. */
 766                 LASSERT(*flags == 0);
 767                 rc = ldlm_extent_compat_queue(&res->lr_granted, lock, flags,
 768                                               NULL, &contended_locks);
 769                 if (rc == 1) {
 770                         rc = ldlm_extent_compat_queue(&res->lr_waiting, lock,
 771                                                       flags, NULL,
 772                                                       &contended_locks);
 773                 }
 774                 if (rc == 0)
 775                         RETURN(LDLM_ITER_STOP);
 776
 777                 ldlm_resource_unlink_lock(lock);
 778
 779                 if (!OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_EVICT_RACE))
 780                         ldlm_extent_policy(res, lock, flags);
 781                 ldlm_grant_lock(lock, grant_work);
 782                 RETURN(LDLM_ITER_CONTINUE);
 783         }
 784
 785         contended_locks = 0;
 786         rc = ldlm_extent_compat_queue(&res->lr_granted, lock, flags,
 787                                       work_list, &contended_locks);
 788         if (rc < 0)
 789                 GOTO(out, *err = rc);
 790
 791         if (rc != 2) {
 792                 rc2 = ldlm_extent_compat_queue(&res->lr_waiting, lock,
 793                                                flags, work_list,
 794                                                &contended_locks);
 795                 if (rc2 < 0)
 796                         GOTO(out, *err = rc = rc2);
 797         }
 798
 799         if (rc + rc2 == 2) {
 800                 ldlm_extent_policy(res, lock, flags);
 801                 ldlm_resource_unlink_lock(lock);
 802                 ldlm_grant_lock(lock, grant_work);
 803         } else {
 804                 /* Adding LDLM_FL_NO_TIMEOUT flag to granted lock to
 805                  * force client to wait for the lock endlessly once
 806                  * the lock is enqueued -bzzz */
 807                 *flags |= LDLM_FL_NO_TIMEOUT;
 808         }
 809
 810         RETURN(LDLM_ITER_CONTINUE);
 811 out:
 812         return rc;
 813 }
 814 #endif /* HAVE_SERVER_SUPPORT */
 815
 816 struct ldlm_kms_shift_args {
 817         __u64   old_kms;
 818         __u64   kms;
 819         bool    complete;
 820 };
 821
 822 /* Callback for interval_iterate functions, used by ldlm_extent_shift_Kms */
 823 static enum interval_iter ldlm_kms_shift_cb(struct interval_node *n,
 824                                             void *args)
 825 {
 826         struct ldlm_kms_shift_args *arg = args;
 827         struct ldlm_interval *node = to_ldlm_interval(n);
 828         struct ldlm_lock *tmplock;
 829         struct ldlm_lock *lock = NULL;
 830
 831         ENTRY;
 832
 833         /* Since all locks in an interval have the same extent, we can just
 834          * use the first lock without kms_ignore set. */
 835         list_for_each_entry(tmplock, &node->li_group, l_sl_policy) {
 836                 if (ldlm_is_kms_ignore(tmplock))
 837                         continue;
 838
 839                 lock = tmplock;
 840
 841                 break;
 842         }
 843
 844         /* No locks in this interval without kms_ignore set */
 845         if (!lock)
 846                 RETURN(INTERVAL_ITER_CONT);
 847
 848         /* If we find a lock with a greater or equal kms, we are not the
 849          * highest lock (or we share that distinction with another lock), and
 850          * don't need to update KMS.  Return old_kms and stop looking. */
 851         if (lock->l_policy_data.l_extent.end == OBD_OBJECT_EOF ||
 852             lock->l_policy_data.l_extent.end + 1 >= arg->old_kms) {
 853                 arg->kms = arg->old_kms;
 854                 arg->complete = true;
 855                 RETURN(INTERVAL_ITER_STOP);
 856         }
 857
 858         if (lock->l_policy_data.l_extent.end + 1 > arg->kms)
 859                 arg->kms = lock->l_policy_data.l_extent.end + 1;
 860
 861         /* Since interval_iterate_reverse starts with the highest lock and
 862          * works down, for PW locks, we only need to check if we should update
 863          * the kms, then stop walking the tree.  PR locks are not exclusive, so
 864          * the highest start does not imply the highest end and we must
 865          * continue. (Only one group lock is allowed per resource, so this is
 866          * irrelevant for group locks.)*/
 867         if (lock->l_granted_mode == LCK_PW)
 868                 RETURN(INTERVAL_ITER_STOP);
 869         else
 870                 RETURN(INTERVAL_ITER_CONT);
 871 }
 872
 873 /* When a lock is cancelled by a client, the KMS may undergo change if this
 874  * is the "highest lock".  This function returns the new KMS value, updating
 875  * it only if we were the highest lock.
 876  *
 877  * Caller must hold lr_lock already.
 878  *
 879  * NB: A lock on [x,y] protects a KMS of up to y + 1 bytes! */
 880 __u64 ldlm_extent_shift_kms(struct ldlm_lock *lock, __u64 old_kms)
 881 {
 882         struct ldlm_resource *res = lock->l_resource;
 883         struct ldlm_interval_tree *tree;
 884         struct ldlm_kms_shift_args args;
 885         int idx = 0;
 886
 887         ENTRY;
 888
 889         args.old_kms = old_kms;
 890         args.kms = 0;
 891         args.complete = false;
 892
 893         /* don't let another thread in ldlm_extent_shift_kms race in
 894          * just after we finish and take our lock into account in its
 895          * calculation of the kms */
 896         ldlm_set_kms_ignore(lock);
 897
 898         /* We iterate over the lock trees, looking for the largest kms smaller
 899          * than the current one. */
 900         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
 901                 tree = &res->lr_itree[idx];
 902
 903                 /* If our already known kms is >= than the highest 'end' in
 904                  * this tree, we don't need to check this tree, because
 905                  * the kms from a tree can be lower than in_max_high (due to
 906                  * kms_ignore), but it can never be higher. */
 907                 if (!tree->lit_root || args.kms >= tree->lit_root->in_max_high)
 908                         continue;
 909
 910                 interval_iterate_reverse(tree->lit_root, ldlm_kms_shift_cb,
 911                                          &args);
 912
 913                 /* this tells us we're not the highest lock, so we don't need
 914                  * to check the remaining trees */
 915                 if (args.complete)
 916                         break;
 917         }
 918
 919         LASSERTF(args.kms <= args.old_kms, "kms %llu old_kms %llu\n", args.kms,
 920                  args.old_kms);
 921
 922         RETURN(args.kms);
 923 }
 924 EXPORT_SYMBOL(ldlm_extent_shift_kms);
 925
 926 struct kmem_cache *ldlm_interval_slab;
 927 static struct ldlm_interval *ldlm_interval_alloc(struct ldlm_lock *lock)
 928 {
 929         struct ldlm_interval *node;
 930         ENTRY;
 931
 932         LASSERT(lock->l_resource->lr_type == LDLM_EXTENT);
 933         OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
 934         if (node == NULL)
 935                 RETURN(NULL);
 936
 937         INIT_LIST_HEAD(&node->li_group);
 938         ldlm_interval_attach(node, lock);
 939         RETURN(node);
 940 }
 941
 942 void ldlm_interval_free(struct ldlm_interval *node)
 943 {
 944         if (node) {
 945                 LASSERT(list_empty(&node->li_group));
 946                 LASSERT(!interval_is_intree(&node->li_node));
 947                 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
 948         }
 949 }
 950
 951 /* interval tree, for LDLM_EXTENT. */
 952 void ldlm_interval_attach(struct ldlm_interval *n,
 953                           struct ldlm_lock *l)
 954 {
 955         LASSERT(l->l_tree_node == NULL);
 956         LASSERT(l->l_resource->lr_type == LDLM_EXTENT);
 957
 958         list_add_tail(&l->l_sl_policy, &n->li_group);
 959         l->l_tree_node = n;
 960 }
 961
 962 struct ldlm_interval *ldlm_interval_detach(struct ldlm_lock *l)
 963 {
 964         struct ldlm_interval *n = l->l_tree_node;
 965
 966         if (n == NULL)
 967                 return NULL;
 968
 969         LASSERT(!list_empty(&n->li_group));
 970         l->l_tree_node = NULL;
 971         list_del_init(&l->l_sl_policy);
 972
 973         return list_empty(&n->li_group) ? n : NULL;
 974 }
 975
 976 static inline int ldlm_mode_to_index(enum ldlm_mode mode)
 977 {
 978         int index;
 979
 980         LASSERT(mode != 0);
 981         LASSERT(is_power_of_2(mode));
 982         index = ilog2(mode);
 983         LASSERT(index < LCK_MODE_NUM);
 984         return index;
 985 }
 986
 987 int ldlm_extent_alloc_lock(struct ldlm_lock *lock)
 988 {
 989         lock->l_tree_node = NULL;
 990         if (ldlm_interval_alloc(lock) == NULL)
 991                 return -ENOMEM;
 992         return 0;
 993 }
 994
 995 /** Add newly granted lock into interval tree for the resource. */
 996 void ldlm_extent_add_lock(struct ldlm_resource *res,
 997                           struct ldlm_lock *lock)
 998 {
 999         struct interval_node *found, **root;
1000         struct ldlm_interval *node;
1001         struct ldlm_extent *extent;
1002         int idx, rc;
1003
1004         LASSERT(ldlm_is_granted(lock));
1005
1006         node = lock->l_tree_node;
1007         LASSERT(node != NULL);
1008         LASSERT(!interval_is_intree(&node->li_node));
1009
1010         idx = ldlm_mode_to_index(lock->l_granted_mode);
1011         LASSERT(lock->l_granted_mode == BIT(idx));
1012         LASSERT(lock->l_granted_mode == res->lr_itree[idx].lit_mode);
1013
1014         /* node extent initialize */
1015         extent = &lock->l_policy_data.l_extent;
1016
1017         rc = interval_set(&node->li_node, extent->start, extent->end);
1018         LASSERT(!rc);
1019
1020         root = &res->lr_itree[idx].lit_root;
1021         found = interval_insert(&node->li_node, root);
1022         if (found) { /* The policy group found. */
1023                 struct ldlm_interval *tmp = ldlm_interval_detach(lock);
1024                 LASSERT(tmp != NULL);
1025                 ldlm_interval_free(tmp);
1026                 ldlm_interval_attach(to_ldlm_interval(found), lock);
1027         }
1028         res->lr_itree[idx].lit_size++;
1029
1030         /* even though we use interval tree to manage the extent lock, we also
1031          * add the locks into grant list, for debug purpose, .. */
1032         ldlm_resource_add_lock(res, &res->lr_granted, lock);
1033
1034         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GRANT_CHECK)) {
1035                 struct ldlm_lock *lck;
1036
1037                 list_for_each_entry_reverse(lck, &res->lr_granted,
1038                                             l_res_link) {
1039                         if (lck == lock)
1040                                 continue;
1041                         if (lockmode_compat(lck->l_granted_mode,
1042                                             lock->l_granted_mode))
1043                                 continue;
1044                         if (ldlm_extent_overlap(&lck->l_req_extent,
1045                                                 &lock->l_req_extent)) {
1046                                 CDEBUG(D_ERROR, "granting conflicting lock %p "
1047                                                 "%p\n", lck, lock);
1048                                 ldlm_resource_dump(D_ERROR, res);
1049                                 LBUG();
1050                         }
1051                 }
1052         }
1053 }
1054
1055 /** Remove cancelled lock from resource interval tree. */
1056 void ldlm_extent_unlink_lock(struct ldlm_lock *lock)
1057 {
1058         struct ldlm_resource *res = lock->l_resource;
1059         struct ldlm_interval *node = lock->l_tree_node;
1060         struct ldlm_interval_tree *tree;
1061         int idx;
1062
1063         if (!node || !interval_is_intree(&node->li_node)) /* duplicate unlink */
1064                 return;
1065
1066         idx = ldlm_mode_to_index(lock->l_granted_mode);
1067         LASSERT(lock->l_granted_mode == BIT(idx));
1068         tree = &res->lr_itree[idx];
1069
1070         LASSERT(tree->lit_root != NULL); /* assure the tree is not null */
1071
1072         tree->lit_size--;
1073         node = ldlm_interval_detach(lock);
1074         if (node) {
1075                 interval_erase(&node->li_node, &tree->lit_root);
1076                 ldlm_interval_free(node);
1077         }
1078 }
1079
1080 void ldlm_extent_policy_wire_to_local(const union ldlm_wire_policy_data *wpolicy,
1081                                       union ldlm_policy_data *lpolicy)
1082 {
1083         lpolicy->l_extent.start = wpolicy->l_extent.start;
1084         lpolicy->l_extent.end = wpolicy->l_extent.end;
1085         lpolicy->l_extent.gid = wpolicy->l_extent.gid;
1086 }
1087
1088 void ldlm_extent_policy_local_to_wire(const union ldlm_policy_data *lpolicy,
1089                                       union ldlm_wire_policy_data *wpolicy)
1090 {
1091         memset(wpolicy, 0, sizeof(*wpolicy));
1092         wpolicy->l_extent.start = lpolicy->l_extent.start;
1093         wpolicy->l_extent.end = lpolicy->l_extent.end;
1094         wpolicy->l_extent.gid = lpolicy->l_extent.gid;
1095 }
1096