lustre/ldlm/ldlm_extent.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2002, 2003 Cluster File Systems, Inc.
   5  *   Author: Peter Braam <braam@clusterfs.com>
   6  *   Author: Phil Schwan <phil@clusterfs.com>
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #define DEBUG_SUBSYSTEM S_LDLM
  25 #ifndef __KERNEL__
  26 # include <liblustre.h>
  27 #endif
  28
  29 #include <linux/lustre_dlm.h>
  30 #include <linux/obd_support.h>
  31 #include <linux/lustre_lib.h>
  32
  33 #include "ldlm_internal.h"
  34
  35 /* The purpose of this function is to return:
  36  * - the maximum extent
  37  * - containing the requested extent
  38  * - and not overlapping existing conflicting extents outside the requested one
  39  */
  40 static void
  41 ldlm_extent_internal_policy(struct list_head *queue, struct ldlm_lock *req,
  42                             struct ldlm_extent *new_ex)
  43 {
  44         struct list_head *tmp;
  45         ldlm_mode_t req_mode = req->l_req_mode;
  46         __u64 req_start = req->l_req_extent.start;
  47         __u64 req_end = req->l_req_extent.end;
  48         int conflicting = 0;
  49         ENTRY;
  50
  51         lockmode_verify(req_mode);
  52
  53         list_for_each(tmp, queue) {
  54                 struct ldlm_lock *lock;
  55                 struct ldlm_extent *l_extent;
  56
  57                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
  58                 l_extent = &lock->l_policy_data.l_extent;
  59
  60                 if (new_ex->start == req_start && new_ex->end == req_end) {
  61                         EXIT;
  62                         return;
  63                 }
  64
  65                 /* Don't conflict with ourselves */
  66                 if (req == lock)
  67                         continue;
  68
  69                 /* Locks are compatible, overlap doesn't matter */
  70                 /* Until bug 20 is fixed, try to avoid granting overlapping
  71                  * locks on one client (they take a long time to cancel) */
  72                 if (lockmode_compat(lock->l_req_mode, req_mode) &&
  73                     lock->l_export != req->l_export)
  74                         continue;
  75
  76                 /* If this is a high-traffic lock, don't grow downwards at all
  77                  * or grow upwards too much */
  78                 ++conflicting;
  79                 if (conflicting > 4)
  80                         new_ex->start = req_start;
  81
  82                 /* If lock doesn't overlap new_ex, skip it. */
  83                 if (l_extent->end < new_ex->start ||
  84                     l_extent->start > new_ex->end)
  85                         continue;
  86
  87                 /* Locks conflicting in requested extents and we can't satisfy
  88                  * both locks, so ignore it.  Either we will ping-pong this
  89                  * extent (we would regardless of what extent we granted) or
  90                  * lock is unused and it shouldn't limit our extent growth. */
  91                 if (lock->l_req_extent.end >= req_start &&
  92                     lock->l_req_extent.start <= req_end)
  93                         continue;
  94
  95                 /* We grow extents downwards only as far as they don't overlap
  96                  * with already-granted locks, on the assumtion that clients
  97                  * will be writing beyond the initial requested end and would
  98                  * then need to enqueue a new lock beyond previous request.
  99                  * l_req_extent->end strictly < req_start, checked above. */
 100                 if (l_extent->start < req_start && new_ex->start != req_start) {
 101                         if (l_extent->end >= req_start)
 102                                 new_ex->start = req_start;
 103                         else
 104                                 new_ex->start = min(l_extent->end+1, req_start);
 105                 }
 106
 107                 /* If we need to cancel this lock anyways because our request
 108                  * overlaps the granted lock, we grow up to its requested
 109                  * extent start instead of limiting this extent, assuming that
 110                  * clients are writing forwards and the lock had over grown
 111                  * its extent downwards before we enqueued our request. */
 112                 if (l_extent->end > req_end) {
 113                         if (l_extent->start <= req_end)
 114                                 new_ex->end = max(lock->l_req_extent.start - 1,
 115                                                   req_end);
 116                         else
 117                                 new_ex->end = max(l_extent->start - 1, req_end);
 118                 }
 119         }
 120
 121 #define LDLM_MAX_GROWN_EXTENT (32 * 1024 * 1024 - 1)
 122         if (conflicting > 32 && (req_mode == LCK_PW || req_mode == LCK_CW)) {
 123                 if (req_end < req_start + LDLM_MAX_GROWN_EXTENT)
 124                         new_ex->end = min(req_start + LDLM_MAX_GROWN_EXTENT,
 125                                           new_ex->end);
 126         }
 127         EXIT;
 128 }
 129
 130 /* In order to determine the largest possible extent we can grant, we need
 131  * to scan all of the queues. */
 132 static void ldlm_extent_policy(struct ldlm_resource *res,
 133                                struct ldlm_lock *lock, int *flags)
 134 {
 135         struct ldlm_extent new_ex = { .start = 0, .end = ~0};
 136
 137         if (lock->l_req_mode == LCK_GROUP)
 138                 return;
 139
 140         ldlm_extent_internal_policy(&res->lr_granted, lock, &new_ex);
 141         ldlm_extent_internal_policy(&res->lr_waiting, lock, &new_ex);
 142
 143         if (new_ex.start != lock->l_policy_data.l_extent.start ||
 144             new_ex.end != lock->l_policy_data.l_extent.end) {
 145                 *flags |= LDLM_FL_LOCK_CHANGED;
 146                 lock->l_policy_data.l_extent.start = new_ex.start;
 147                 lock->l_policy_data.l_extent.end = new_ex.end;
 148         }
 149 }
 150
 151 /* Determine if the lock is compatible with all locks on the queue.
 152  * We stop walking the queue if we hit ourselves so we don't take
 153  * conflicting locks enqueued after us into accound, or we'd wait forever.
 154  *
 155  * 0 if the lock is not compatible
 156  * 1 if the lock is compatible
 157  * 2 if this group lock is compatible and requires no further checking
 158  * negative error, such as EWOULDBLOCK for group locks
 159  */
 160 static int
 161 ldlm_extent_compat_queue(struct list_head *queue, struct ldlm_lock *req,
 162                          int *flags, struct list_head *work_list,
 163                          struct list_head **insertp)
 164 {
 165         struct list_head *tmp;
 166         struct list_head *save = NULL;
 167         struct ldlm_lock *lock = NULL;
 168         ldlm_mode_t req_mode = req->l_req_mode;
 169         int compat = 1;
 170         int found = 0;
 171         ENTRY;
 172
 173         lockmode_verify(req_mode);
 174
 175         /* Extent locks are only queued once. We can get back here with
 176          * insertp != NULL if the blocking ASTs returned -ERESTART. */
 177         if (!list_empty(&req->l_res_link))
 178                 insertp = NULL;
 179
 180         if (req->l_req_mode != LCK_GROUP) {
 181                 __u64 req_start = req->l_req_extent.start;
 182                 __u64 req_end = req->l_req_extent.end;
 183
 184                 list_for_each(tmp, queue) {
 185                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 186                         if (req == lock)
 187                                 break;
 188
 189                         if (lock->l_req_mode == LCK_GROUP) {
 190                                 if (*flags & LDLM_FL_BLOCK_NOWAIT)
 191                                         RETURN(-EWOULDBLOCK);
 192
 193                                 /* No blocking ASTs are sent for group locks. */
 194                                 compat = 0;
 195
 196                                 /* there's a blocking group lock in front
 197                                  * of us on the queue.  It can be held
 198                                  * indefinitely, so don't timeout. */
 199                                 if (insertp) {
 200                                         *flags |= LDLM_FL_NO_TIMEOUT;
 201                                         /* lock_bitlock(req) is held here. */
 202                                         req->l_flags |= LDLM_FL_NO_TIMEOUT;
 203                                 }
 204
 205                                 if (work_list)
 206                                         continue;
 207                                 else
 208                                         break;
 209                         }
 210
 211                         /* locks are compatible, overlap doesn't matter */
 212                         if (lockmode_compat(lock->l_req_mode, req_mode))
 213                                 continue;
 214
 215                         if (lock->l_policy_data.l_extent.end < req_start ||
 216                             lock->l_policy_data.l_extent.start > req_end)
 217                                 continue;
 218
 219                         compat = 0;
 220
 221                         if (!work_list)
 222                                 break;
 223
 224                         if (lock->l_blocking_ast)
 225                                 ldlm_add_ast_work_item(lock, req, work_list);
 226                 }
 227
 228                 if (insertp)
 229                         *insertp = queue;
 230
 231                 RETURN(compat);
 232         }
 233
 234         list_for_each(tmp, queue) {
 235                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 236                 if (req == lock)
 237                         break;
 238
 239                 if (lock->l_req_mode != LCK_GROUP) {
 240                         if (lock->l_req_mode != lock->l_granted_mode) {
 241                                 /* we must be traversing the waitq. */
 242
 243                                 /* If a group lock was already found then
 244                                  * req can be queued before any extent locks
 245                                  * that come after the found group lock. */
 246                                 if (found)
 247                                         break;
 248
 249                                 if (!insertp) {
 250                                         /* We've hit a conflicting extent lock
 251                                          * on the waitq before hitting the req
 252                                          * group lock. See comments below. */
 253                                         compat = 0;
 254                                         break;
 255                                 }
 256
 257                                 /* Group locks are not normally blocked by
 258                                  * waiting PR|PW locks. */
 259
 260                                 /* If NO_TIMEOUT was sent back to the client
 261                                  * we can queue the group lock in front of
 262                                  * this extent lock. */
 263                                 if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
 264                                         if (save == NULL)
 265                                                 save = tmp;
 266                                         continue;
 267                                 }
 268
 269                                 /* If we did NOT send NO_TIMEOUT back to the
 270                                  * client for this extent lock then the client
 271                                  * could possibly timeout if we queue this
 272                                  * group lock before it, so don't. This is the
 273                                  * only way to get a conflicting extent lock
 274                                  * in front of a group lock on the waitq. */
 275                         }
 276
 277                         compat = 0;
 278                         if (!work_list) {
 279                                 LASSERT(save == NULL);
 280                                 break;
 281                         }
 282
 283                         /* If we previously skipped over some extent locks
 284                          * because we thought we were going to queue the
 285                          * group lock in front of them then we need to go back
 286                          * and send blocking ASTs for the locks we skipped. */
 287                         if (save != NULL) {
 288                                 struct ldlm_lock *lck2;
 289
 290                                 for (; save != tmp; save = save->next) {
 291                                         lck2 = list_entry(save,
 292                                                           struct ldlm_lock,
 293                                                           l_res_link);
 294
 295                                         /* If there was a group lock after save
 296                                          * then we would have exited this loop
 297                                          * above. */
 298                                         LASSERT(lck2->l_req_mode!=LCK_GROUP);
 299
 300                                         if (lck2->l_blocking_ast) {
 301                                                 ldlm_add_ast_work_item(lck2,req,
 302                                                                      work_list);
 303                                         }
 304                                 }
 305                                 save = NULL;
 306                         }
 307
 308                         if (lock->l_blocking_ast)
 309                                 ldlm_add_ast_work_item(lock, req, work_list);
 310                         continue;
 311                 }
 312
 313                 /* If it was safe to insert a group lock at save,
 314                  * i.e. save != NULL, then this group lock already
 315                  * on the queue would have been inserted before save. */
 316                 LASSERT(save == NULL);
 317
 318                 /* Note: no blocking ASTs are sent for group locks. */
 319
 320                 if (lock->l_policy_data.l_extent.gid ==
 321                     req->l_policy_data.l_extent.gid) {
 322                         /* group locks with this gid already on the waitq. */
 323                         found = 2;
 324
 325                         if (lock->l_req_mode == lock->l_granted_mode) {
 326                                 /* if a group lock with this gid has already
 327                                  * been granted then grant this one. */
 328                                 compat = 2;
 329                                 break;
 330                         }
 331                 } else {
 332                         if (found == 2)
 333                                 break;
 334
 335                         /* group locks already exist on the queue. */
 336                         found = 1;
 337
 338                         if (*flags & LDLM_FL_BLOCK_NOWAIT)
 339                                 RETURN(-EWOULDBLOCK);
 340
 341                         compat = 0;
 342
 343                         /* there's a blocking group lock in front
 344                          * of us on the queue.  It can be held
 345                          * indefinitely, so don't timeout. */
 346                         *flags |= LDLM_FL_NO_TIMEOUT;
 347
 348                         /* the only reason to continue traversing the
 349                          * list at this point is to find the proper
 350                          * place to insert the lock in the waitq. */
 351                         if (!insertp)
 352                                 break;
 353                 }
 354         }
 355
 356         if (insertp != NULL) {
 357                 if (save != NULL)
 358                         *insertp = save;
 359                 else
 360                         *insertp = tmp;
 361         }
 362
 363         RETURN(compat);
 364 }
 365
 366 /* If first_enq is 0 (ie, called from ldlm_reprocess_queue):
 367   *   - blocking ASTs have already been sent
 368   *   - the caller has already initialized req->lr_tmp
 369   *   - must call this function with the ns lock held
 370   *
 371   * If first_enq is 1 (ie, called from ldlm_lock_enqueue):
 372   *   - blocking ASTs have not been sent
 373   *   - the caller has NOT initialized req->lr_tmp, so we must
 374   *   - must call this function with the ns lock held once */
 375 int ldlm_process_extent_lock(struct ldlm_lock *lock, int *flags, int first_enq,
 376                              ldlm_error_t *err, struct list_head *work_list)
 377 {
 378         struct ldlm_resource *res = lock->l_resource;
 379         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 380         struct list_head *insertp = NULL;
 381         int rc, rc2;
 382         ENTRY;
 383
 384         LASSERT(list_empty(&res->lr_converting));
 385         *err = ELDLM_OK;
 386
 387         if (!first_enq) {
 388                 /* -EWOULDBLOCK can't occur here since (flags & BLOCK_NOWAIT)
 389                  * lock requests would either be granted or fail on their
 390                  * first_enq. flags should always be zero here, and if that
 391                  * ever changes we want to find out. */
 392                 LASSERT(*flags == 0);
 393                 rc = ldlm_extent_compat_queue(&res->lr_granted, lock,
 394                                               flags, NULL, NULL);
 395                 if (rc == 1) {
 396                         rc = ldlm_extent_compat_queue(&res->lr_waiting, lock,
 397                                                       flags, NULL, NULL);
 398                 }
 399                 if (rc == 0)
 400                         RETURN(LDLM_ITER_STOP);
 401
 402                 ldlm_resource_unlink_lock(lock);
 403                 ldlm_extent_policy(res, lock, flags);
 404                 lock_bitlock(lock);
 405                 lock->l_flags &= ~LDLM_FL_NO_TIMEOUT;
 406                 unlock_bitlock(lock);
 407                 ldlm_grant_lock(lock, work_list);
 408                 RETURN(LDLM_ITER_CONTINUE);
 409         }
 410
 411  restart:
 412         rc = ldlm_extent_compat_queue(&res->lr_granted, lock, flags, &rpc_list,
 413                                       NULL);
 414         if (rc < 0)
 415                 GOTO(destroylock, rc);
 416         if (rc == 2)
 417                 goto grant;
 418
 419         /* Traverse the waiting list in case there are other conflicting
 420          * lock requests ahead of us in the queue and send blocking ASTs */
 421         rc2 = ldlm_extent_compat_queue(&res->lr_waiting, lock, flags, &rpc_list,
 422                                        &insertp);
 423         if (rc2 < 0)
 424                 GOTO(destroylock, rc);
 425         if (rc + rc2 == 2) {
 426  grant:
 427                 ldlm_extent_policy(res, lock, flags);
 428                 ldlm_resource_unlink_lock(lock);
 429                 lock->l_flags &= ~LDLM_FL_NO_TIMEOUT;
 430                 ldlm_grant_lock(lock, NULL);
 431         } else {
 432                 /* If either of the compat_queue()s returned failure, then we
 433                  * have ASTs to send and must go onto the waiting list.
 434                  *
 435                  * bug 2322: we used to unlink and re-add here, which was a
 436                  * terrible folly -- if we goto restart, we could get
 437                  * re-ordered!  Causes deadlock, because ASTs aren't sent! */
 438                 if (list_empty(&lock->l_res_link))
 439                         ldlm_resource_add_lock(res, insertp, lock);
 440                 unlock_res(res);
 441                 rc = ldlm_run_bl_ast_work(&rpc_list);
 442                 lock_res(res);
 443                 if (rc == -ERESTART)
 444                         GOTO(restart, -ERESTART);
 445                 *flags |= LDLM_FL_BLOCK_GRANTED;
 446         }
 447
 448         RETURN(0);
 449
 450  destroylock:
 451         list_del_init(&lock->l_res_link);
 452         unlock_res(res);
 453         ldlm_lock_destroy(lock);
 454         lock_res(res);
 455         *err = rc;
 456         RETURN(rc);
 457 }
 458
 459 /* When a lock is cancelled by a client, the KMS may undergo change if this
 460  * is the "highest lock".  This function returns the new KMS value.
 461  * Caller must hold ns_lock already.
 462  *
 463  * NB: A lock on [x,y] protects a KMS of up to y + 1 bytes! */
 464 __u64 ldlm_extent_shift_kms(struct ldlm_lock *lock, __u64 old_kms)
 465 {
 466         struct ldlm_resource *res = lock->l_resource;
 467         struct list_head *tmp;
 468         struct ldlm_lock *lck;
 469         __u64 kms = 0;
 470         ENTRY;
 471
 472         /* don't let another thread in ldlm_extent_shift_kms race in
 473          * just after we finish and take our lock into account in its
 474          * calculation of the kms */
 475
 476         lock->l_flags |= LDLM_FL_KMS_IGNORE;
 477
 478         list_for_each(tmp, &res->lr_granted) {
 479                 lck = list_entry(tmp, struct ldlm_lock, l_res_link);
 480
 481                 if (lck->l_flags & LDLM_FL_KMS_IGNORE)
 482                         continue;
 483
 484                 if (lck->l_policy_data.l_extent.end >= old_kms)
 485                         RETURN(old_kms);
 486
 487                 /* This extent _has_ to be smaller than old_kms (checked above)
 488                  * so kms can only ever be smaller or the same as old_kms. */
 489                 if (lck->l_policy_data.l_extent.end + 1 > kms)
 490                         kms = lck->l_policy_data.l_extent.end + 1;
 491         }
 492         LASSERTF(kms <= old_kms, "kms "LPU64" old_kms "LPU64"\n", kms, old_kms);
 493
 494         RETURN(kms);
 495 }