Whamcloud - gitweb
b=11270
[fs/lustre-release.git] / 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 the Lustre file system, http://www.lustre.org
9  *   Lustre is a trademark of Cluster File Systems, Inc.
10  *
11  *   You may have signed or agreed to another license before downloading
12  *   this software.  If so, you are bound by the terms and conditions
13  *   of that agreement, and the following does not apply to you.  See the
14  *   LICENSE file included with this distribution for more information.
15  *
16  *   If you did not agree to a different license, then this copy of Lustre
17  *   is open source software; you can redistribute it and/or modify it
18  *   under the terms of version 2 of the GNU General Public License as
19  *   published by the Free Software Foundation.
20  *
21  *   In either case, Lustre is distributed in the hope that it will be
22  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
23  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
24  *   license text for more details.
25  */
26
27 #define DEBUG_SUBSYSTEM S_LDLM
28 #ifndef __KERNEL__
29 # include <liblustre.h>
30 #else
31 # include <libcfs/libcfs.h>
32 # include <libcfs/kp30.h>
33 #endif
34
35 #include <lustre_dlm.h>
36 #include <obd_support.h>
37 #include <obd.h>
38 #include <lustre_lib.h>
39
40 #include "ldlm_internal.h"
41
42 #define LDLM_MAX_GROWN_EXTENT (32 * 1024 * 1024 - 1)
43
44 /* fixup the ldlm_extent after expanding */
45 static void ldlm_extent_internal_policy_fixup(struct ldlm_lock *req,
46                                               struct ldlm_extent *new_ex,
47                                               int conflicting)
48 {
49         ldlm_mode_t req_mode = req->l_req_mode;
50         __u64 req_start = req->l_req_extent.start;
51         __u64 req_end = req->l_req_extent.end;
52         __u64 req_align, mask;
53  
54         if (conflicting > 32 && (req_mode == LCK_PW || req_mode == LCK_CW)) {
55                 if (req_end < req_start + LDLM_MAX_GROWN_EXTENT)
56                         new_ex->end = min(req_start + LDLM_MAX_GROWN_EXTENT,
57                                           new_ex->end);
58         }
59
60         if (new_ex->start == 0 && new_ex->end == OBD_OBJECT_EOF) {
61                 EXIT;
62                 return;
63         }
64
65         /* we need to ensure that the lock extent is properly aligned to what
66          * the client requested.  We align it to the lowest-common denominator
67          * of the clients requested lock start and end alignment. */
68         mask = 0x1000ULL;
69         req_align = (req_end + 1) | req_start;
70         if (req_align != 0) {
71                 while ((req_align & mask) == 0)
72                         mask <<= 1;
73         }
74         mask -= 1;
75         /* We can only shrink the lock, not grow it.
76          * This should never cause lock to be smaller than requested,
77          * since requested lock was already aligned on these boundaries. */
78         new_ex->start = ((new_ex->start - 1) | mask) + 1;
79         new_ex->end = ((new_ex->end + 1) & ~mask) - 1;
80         LASSERTF(new_ex->start <= req_start,
81                  "mask "LPX64" grant start "LPU64" req start "LPU64"\n",
82                  mask, new_ex->start, req_start);
83         LASSERTF(new_ex->end >= req_end,
84                  "mask "LPX64" grant end "LPU64" req end "LPU64"\n",
85                  mask, new_ex->end, req_end);
86 }
87
88 /* The purpose of this function is to return:
89  * - the maximum extent
90  * - containing the requested extent
91  * - and not overlapping existing conflicting extents outside the requested one
92  *
93  * Use interval tree to expand the lock extent for granted lock.
94  */
95 static void ldlm_extent_internal_policy_granted(struct ldlm_lock *req,
96                                                 struct ldlm_extent *new_ex)
97 {
98         struct ldlm_resource *res = req->l_resource;
99         ldlm_mode_t req_mode = req->l_req_mode;
100         __u64 req_start = req->l_req_extent.start;
101         __u64 req_end = req->l_req_extent.end;
102         struct ldlm_interval_tree *tree;
103         struct interval_node_extent limiter = { new_ex->start, new_ex->end };
104         int conflicting = 0;
105         int idx;
106         ENTRY;
107
108         lockmode_verify(req_mode);
109
110         /* using interval tree to handle the ldlm extent granted locks */
111         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
112                 struct interval_node_extent ext = { req_start, req_end };
113
114                 tree = &res->lr_itree[idx];
115                 if (lockmode_compat(tree->lit_mode, req_mode))
116                         continue;
117
118                 conflicting += tree->lit_size;
119                 if (conflicting > 4)
120                         limiter.start = req_start;
121
122                 if (interval_is_overlapped(tree->lit_root, &ext))
123                         printk("req_mode = %d, tree->lit_mode = %d, tree->lit_size = %d\n",
124                                req_mode, tree->lit_mode, tree->lit_size);
125                 interval_expand(tree->lit_root, &ext, &limiter);
126                 limiter.start = max(limiter.start, ext.start);
127                 limiter.end = min(limiter.end, ext.end);
128                 if (limiter.start == req_start && limiter.end == req_end)
129                         break;
130         }
131
132         new_ex->start = limiter.start;
133         new_ex->end = limiter.end;
134         LASSERT(new_ex->start <= req_start);
135         LASSERT(new_ex->end >= req_end);
136
137         ldlm_extent_internal_policy_fixup(req, new_ex, conflicting);
138         EXIT;
139 }
140
141 /* The purpose of this function is to return:
142  * - the maximum extent
143  * - containing the requested extent
144  * - and not overlapping existing conflicting extents outside the requested one
145  */
146 static void
147 ldlm_extent_internal_policy_waiting(struct ldlm_lock *req,
148                                     struct ldlm_extent *new_ex)
149 {
150         struct list_head *tmp;
151         struct ldlm_resource *res = req->l_resource;
152         ldlm_mode_t req_mode = req->l_req_mode;
153         __u64 req_start = req->l_req_extent.start;
154         __u64 req_end = req->l_req_extent.end;
155         int conflicting = 0;
156         ENTRY;
157
158         lockmode_verify(req_mode);
159
160         /* for waiting locks */
161         list_for_each(tmp, &res->lr_waiting) {
162                 struct ldlm_lock *lock;
163                 struct ldlm_extent *l_extent;
164
165                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
166                 l_extent = &lock->l_policy_data.l_extent;
167
168                 /* We already hit the minimum requested size, search no more */
169                 if (new_ex->start == req_start && new_ex->end == req_end) {
170                         EXIT;
171                         return;
172                 }
173
174                 /* Don't conflict with ourselves */
175                 if (req == lock)
176                         continue;
177
178                 /* Locks are compatible, overlap doesn't matter */
179                 /* Until bug 20 is fixed, try to avoid granting overlapping
180                  * locks on one client (they take a long time to cancel) */
181                 if (lockmode_compat(lock->l_req_mode, req_mode) &&
182                     lock->l_export != req->l_export)
183                         continue;
184
185                 /* If this is a high-traffic lock, don't grow downwards at all
186                  * or grow upwards too much */
187                 ++conflicting;
188                 if (conflicting > 4)
189                         new_ex->start = req_start;
190
191                 /* If lock doesn't overlap new_ex, skip it. */
192                 if (!ldlm_extent_overlap(l_extent, new_ex))
193                         continue;
194
195                 /* Locks conflicting in requested extents and we can't satisfy
196                  * both locks, so ignore it.  Either we will ping-pong this
197                  * extent (we would regardless of what extent we granted) or
198                  * lock is unused and it shouldn't limit our extent growth. */
199                 if (ldlm_extent_overlap(&lock->l_req_extent,&req->l_req_extent))
200                         continue;
201
202                 /* We grow extents downwards only as far as they don't overlap
203                  * with already-granted locks, on the assumtion that clients
204                  * will be writing beyond the initial requested end and would
205                  * then need to enqueue a new lock beyond previous request.
206                  * l_req_extent->end strictly < req_start, checked above. */
207                 if (l_extent->start < req_start && new_ex->start != req_start) {
208                         if (l_extent->end >= req_start)
209                                 new_ex->start = req_start;
210                         else
211                                 new_ex->start = min(l_extent->end+1, req_start);
212                 }
213
214                 /* If we need to cancel this lock anyways because our request
215                  * overlaps the granted lock, we grow up to its requested
216                  * extent start instead of limiting this extent, assuming that
217                  * clients are writing forwards and the lock had over grown
218                  * its extent downwards before we enqueued our request. */
219                 if (l_extent->end > req_end) {
220                         if (l_extent->start <= req_end)
221                                 new_ex->end = max(lock->l_req_extent.start - 1,
222                                                   req_end);
223                         else
224                                 new_ex->end = max(l_extent->start - 1, req_end);
225                 }
226         }
227
228         ldlm_extent_internal_policy_fixup(req, new_ex, conflicting);
229         EXIT;
230 }
231
232
233 /* In order to determine the largest possible extent we can grant, we need
234  * to scan all of the queues. */
235 static void ldlm_extent_policy(struct ldlm_resource *res,
236                                struct ldlm_lock *lock, int *flags)
237 {
238         struct ldlm_extent new_ex = { .start = 0, .end = OBD_OBJECT_EOF };
239
240         if (lock->l_export == NULL)
241                 /*
242                  * this is local lock taken by server (e.g., as a part of
243                  * OST-side locking, or unlink handling). Expansion doesn't
244                  * make a lot of sense for local locks, because they are
245                  * dropped immediately on operation completion and would only
246                  * conflict with other threads.
247                  */
248                 return;
249
250         if (lock->l_policy_data.l_extent.start == 0 &&
251             lock->l_policy_data.l_extent.end == OBD_OBJECT_EOF)
252                 /* fast-path whole file locks */
253                 return;
254
255         ldlm_extent_internal_policy_granted(lock, &new_ex);
256         ldlm_extent_internal_policy_waiting(lock, &new_ex);
257
258         if (new_ex.start != lock->l_policy_data.l_extent.start ||
259             new_ex.end != lock->l_policy_data.l_extent.end) {
260                 *flags |= LDLM_FL_LOCK_CHANGED;
261                 lock->l_policy_data.l_extent.start = new_ex.start;
262                 lock->l_policy_data.l_extent.end = new_ex.end;
263         }
264 }
265
266 static int ldlm_check_contention(struct ldlm_lock *lock, int contended_locks)
267 {
268         struct ldlm_resource *res = lock->l_resource;
269         cfs_time_t now = cfs_time_current();
270
271         CDEBUG(D_DLMTRACE, "contended locks = %d\n", contended_locks);
272         if (contended_locks > res->lr_namespace->ns_contended_locks)
273                 res->lr_contention_time = now;
274         return cfs_time_before(now, cfs_time_add(res->lr_contention_time,
275                 cfs_time_seconds(res->lr_namespace->ns_contention_time)));
276 }
277
278 struct ldlm_extent_compat_args {
279         struct list_head *work_list;
280         struct ldlm_lock *lock;
281         ldlm_mode_t mode;
282         int *locks;
283         int *compat;
284 };
285
286 static enum interval_iter ldlm_extent_compat_cb(struct interval_node *n,
287                                                 void *data)
288 {
289         struct ldlm_extent_compat_args *priv = data;
290         struct ldlm_interval *node = to_ldlm_interval(n);
291         struct ldlm_extent *extent;
292         struct list_head *work_list = priv->work_list;
293         struct ldlm_lock *lock, *enq = priv->lock;
294         ldlm_mode_t mode = priv->mode;
295         int count = 0;
296         ENTRY;
297
298         LASSERT(!list_empty(&node->li_group));
299
300         list_for_each_entry(lock, &node->li_group, l_sl_policy) {
301                 /* interval tree is for granted lock */
302                 LASSERTF(mode == lock->l_granted_mode,
303                          "mode = %s, lock->l_granted_mode = %s\n",
304                          ldlm_lockname[mode],
305                          ldlm_lockname[lock->l_granted_mode]);
306                 count++;
307                 if (lock->l_blocking_ast)
308                         ldlm_add_ast_work_item(lock, enq, work_list);
309         }
310
311         /* don't count conflicting glimpse locks */
312         extent = ldlm_interval_extent(node);
313         if (!(mode == LCK_PR &&
314             extent->start == 0 && extent->end == OBD_OBJECT_EOF))
315                 *priv->locks += count;
316
317         if (priv->compat)
318                 *priv->compat = 0;
319
320         RETURN(INTERVAL_ITER_CONT);
321 }
322
323 /* Determine if the lock is compatible with all locks on the queue.
324  * We stop walking the queue if we hit ourselves so we don't take
325  * conflicting locks enqueued after us into accound, or we'd wait forever.
326  *
327  * 0 if the lock is not compatible
328  * 1 if the lock is compatible
329  * 2 if this group lock is compatible and requires no further checking
330  * negative error, such as EWOULDBLOCK for group locks
331  */
332 static int
333 ldlm_extent_compat_queue(struct list_head *queue, struct ldlm_lock *req,
334                          int *flags, ldlm_error_t *err,
335                          struct list_head *work_list, int *contended_locks)
336 {
337         struct list_head *tmp;
338         struct ldlm_lock *lock;
339         struct ldlm_resource *res = req->l_resource;
340         ldlm_mode_t req_mode = req->l_req_mode;
341         __u64 req_start = req->l_req_extent.start;
342         __u64 req_end = req->l_req_extent.end;
343         int compat = 1;
344         int scan = 0;
345         int check_contention;
346         ENTRY;
347
348         lockmode_verify(req_mode);
349
350         /* Using interval tree for granted lock */
351         if (queue == &res->lr_granted) {
352                 struct ldlm_interval_tree *tree;
353                 struct ldlm_extent_compat_args data = {.work_list = work_list,
354                                                .lock = req,
355                                                .locks = contended_locks,
356                                                .compat = &compat };
357                 struct interval_node_extent ex = { .start = req_start,
358                                                    .end = req_end };
359                 int idx, rc;
360
361                 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
362                         tree = &res->lr_itree[idx];
363                         if (tree->lit_root == NULL) /* empty tree, skipped */
364                                 continue;
365
366                         data.mode = tree->lit_mode;
367                         if (lockmode_compat(req_mode, tree->lit_mode)) {
368                                 struct ldlm_interval *node;
369                                 struct ldlm_extent *extent;
370
371                                 if (req_mode != LCK_GROUP)
372                                         continue;
373
374                                 /* group lock, grant it immediately if
375                                  * compatible */
376                                 node = to_ldlm_interval(tree->lit_root);
377                                 extent = ldlm_interval_extent(node);
378                                 if (req->l_policy_data.l_extent.gid ==
379                                     extent->gid)
380                                         RETURN(2);
381                         }
382
383                         if (tree->lit_mode == LCK_GROUP) {
384                                 if (*flags & LDLM_FL_BLOCK_NOWAIT) {
385                                         compat = -EWOULDBLOCK;
386                                         goto destroylock;
387                                 }
388
389                                 *flags |= LDLM_FL_NO_TIMEOUT;
390                                 if (!work_list)
391                                         RETURN(0);
392
393                                 /* if work list is not NULL,add all
394                                    locks in the tree to work list */
395                                 compat = 0;
396                                 interval_iterate(tree->lit_root,
397                                                  ldlm_extent_compat_cb, &data);
398                                 continue;
399                         }
400
401                         if (!work_list) {
402                                 rc = interval_is_overlapped(tree->lit_root,&ex);
403                                 if (rc)
404                                         RETURN(0);
405                         } else {
406                                 interval_search(tree->lit_root, &ex,
407                                                 ldlm_extent_compat_cb, &data);
408                                 if (!list_empty(work_list) && compat)
409                                         compat = 0;
410                         }
411                 }
412         } else { /* for waiting queue */
413                 list_for_each(tmp, queue) {
414                         check_contention = 1;
415
416                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
417
418                         if (req == lock)
419                                 break;
420
421                         if (unlikely(scan)) {
422                                 /* We only get here if we are queuing GROUP lock
423                                    and met some incompatible one. The main idea of this
424                                    code is to insert GROUP lock past compatible GROUP
425                                    lock in the waiting queue or if there is not any,
426                                    then in front of first non-GROUP lock */
427                                 if (lock->l_req_mode != LCK_GROUP) {
428                                         /* Ok, we hit non-GROUP lock, there should
429                                          * be no more GROUP locks later on, queue in
430                                          * front of first non-GROUP lock */
431
432                                         ldlm_resource_insert_lock_after(lock, req);
433                                         list_del_init(&lock->l_res_link);
434                                         ldlm_resource_insert_lock_after(req, lock);
435                                         compat = 0;
436                                         break;
437                                 }
438                                 if (req->l_policy_data.l_extent.gid ==
439                                     lock->l_policy_data.l_extent.gid) {
440                                         /* found it */
441                                         ldlm_resource_insert_lock_after(lock, req);
442                                         compat = 0;
443                                         break;
444                                 }
445                                 continue;
446                         }
447
448                         /* locks are compatible, overlap doesn't matter */
449                         if (lockmode_compat(lock->l_req_mode, req_mode)) {
450                                 if (req_mode == LCK_PR &&
451                                     ((lock->l_policy_data.l_extent.start <=
452                                       req->l_policy_data.l_extent.start) &&
453                                      (lock->l_policy_data.l_extent.end >=
454                                       req->l_policy_data.l_extent.end))) {
455                                         /* If we met a PR lock just like us or wider,
456                                            and nobody down the list conflicted with
457                                            it, that means we can skip processing of
458                                            the rest of the list and safely place
459                                            ourselves at the end of the list, or grant
460                                            (dependent if we met an conflicting locks
461                                            before in the list).
462                                            In case of 1st enqueue only we continue
463                                            traversing if there is something conflicting
464                                            down the list because we need to make sure
465                                            that something is marked as AST_SENT as well,
466                                            in cse of empy worklist we would exit on
467                                            first conflict met. */
468                                         /* There IS a case where such flag is
469                                            not set for a lock, yet it blocks
470                                            something. Luckily for us this is
471                                            only during destroy, so lock is
472                                            exclusive. So here we are safe */
473                                         if (!(lock->l_flags & LDLM_FL_AST_SENT)) {
474                                                 RETURN(compat);
475                                         }
476                                 }
477
478                                 /* non-group locks are compatible, overlap doesn't
479                                    matter */
480                                 if (likely(req_mode != LCK_GROUP))
481                                         continue;
482
483                                 /* If we are trying to get a GROUP lock and there is
484                                    another one of this kind, we need to compare gid */
485                                 if (req->l_policy_data.l_extent.gid ==
486                                     lock->l_policy_data.l_extent.gid) {
487                                         /* If existing lock with matched gid is granted,
488                                            we grant new one too. */
489                                         if (lock->l_req_mode == lock->l_granted_mode)
490                                                 RETURN(2);
491
492                                         /* Otherwise we are scanning queue of waiting
493                                          * locks and it means current request would
494                                          * block along with existing lock (that is
495                                          * already blocked.
496                                          * If we are in nonblocking mode - return
497                                          * immediately */
498                                         if (*flags & LDLM_FL_BLOCK_NOWAIT) {
499                                                 compat = -EWOULDBLOCK;
500                                                 goto destroylock;
501                                         }
502                                         /* If this group lock is compatible with another
503                                          * group lock on the waiting list, they must be
504                                          * together in the list, so they can be granted
505                                          * at the same time.  Otherwise the later lock
506                                          * can get stuck behind another, incompatible,
507                                          * lock. */
508                                         ldlm_resource_insert_lock_after(lock, req);
509                                         /* Because 'lock' is not granted, we can stop
510                                          * processing this queue and return immediately.
511                                          * There is no need to check the rest of the
512                                          * list. */
513                                         RETURN(0);
514                                 }
515                         }
516
517                         if (unlikely(req_mode == LCK_GROUP &&
518                                      (lock->l_req_mode != lock->l_granted_mode))) {
519                                 scan = 1;
520                                 compat = 0;
521                                 if (lock->l_req_mode != LCK_GROUP) {
522                                         /* Ok, we hit non-GROUP lock, there should be no
523                                            more GROUP locks later on, queue in front of
524                                            first non-GROUP lock */
525
526                                         ldlm_resource_insert_lock_after(lock, req);
527                                         list_del_init(&lock->l_res_link);
528                                         ldlm_resource_insert_lock_after(req, lock);
529                                         break;
530                                 }
531                                 if (req->l_policy_data.l_extent.gid ==
532                                     lock->l_policy_data.l_extent.gid) {
533                                         /* found it */
534                                         ldlm_resource_insert_lock_after(lock, req);
535                                         break;
536                                 }
537                                 continue;
538                         }
539
540                         if (unlikely(lock->l_req_mode == LCK_GROUP)) {
541                                 /* If compared lock is GROUP, then requested is PR/PW/
542                                  * so this is not compatible; extent range does not
543                                  * matter */
544                                 if (*flags & LDLM_FL_BLOCK_NOWAIT) {
545                                         compat = -EWOULDBLOCK;
546                                         goto destroylock;
547                                 } else {
548                                         *flags |= LDLM_FL_NO_TIMEOUT;
549                                 }
550                         } else if (lock->l_policy_data.l_extent.end < req_start ||
551                                    lock->l_policy_data.l_extent.start > req_end) {
552                                 /* if a non group lock doesn't overlap skip it */
553                                 continue;
554                         } else if (lock->l_req_extent.end < req_start ||
555                                    lock->l_req_extent.start > req_end) {
556                                 /* false contention, the requests doesn't really overlap */
557                                 check_contention = 0;
558                         }
559
560                         if (!work_list)
561                                 RETURN(0);
562
563                         /* don't count conflicting glimpse locks */
564                         if (lock->l_req_mode == LCK_PR &&
565                             lock->l_policy_data.l_extent.start == 0 &&
566                             lock->l_policy_data.l_extent.end == OBD_OBJECT_EOF)
567                                 check_contention = 0;
568
569                         *contended_locks += check_contention;
570
571                         compat = 0;
572                         if (lock->l_blocking_ast)
573                                 ldlm_add_ast_work_item(lock, req, work_list);
574                 }
575         }
576
577         if (ldlm_check_contention(req, *contended_locks) &&
578             compat == 0 &&
579             (*flags & LDLM_FL_DENY_ON_CONTENTION) &&
580             req->l_req_mode != LCK_GROUP &&
581             req_end - req_start <=
582             req->l_resource->lr_namespace->ns_max_nolock_size)
583                 GOTO(destroylock, compat = -EUSERS);
584
585         RETURN(compat);
586 destroylock:
587         list_del_init(&req->l_res_link);
588         ldlm_lock_destroy_nolock(req);
589         *err = compat;
590         RETURN(compat);
591 }
592
593 static void discard_bl_list(struct list_head *bl_list)
594 {
595         struct list_head *tmp, *pos;
596         ENTRY;
597
598         list_for_each_safe(pos, tmp, bl_list) {
599                 struct ldlm_lock *lock =
600                         list_entry(pos, struct ldlm_lock, l_bl_ast);
601
602                 list_del_init(&lock->l_bl_ast);
603                 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
604                 lock->l_flags &= ~LDLM_FL_AST_SENT;
605                 LASSERT(lock->l_bl_ast_run == 0);
606                 LASSERT(lock->l_blocking_lock);
607                 LDLM_LOCK_PUT(lock->l_blocking_lock);
608                 lock->l_blocking_lock = NULL;
609                 LDLM_LOCK_PUT(lock);
610         }
611         EXIT;
612 }
613
614 /* If first_enq is 0 (ie, called from ldlm_reprocess_queue):
615   *   - blocking ASTs have already been sent
616   *   - must call this function with the ns lock held
617   *
618   * If first_enq is 1 (ie, called from ldlm_lock_enqueue):
619   *   - blocking ASTs have not been sent
620   *   - must call this function with the ns lock held once */
621 int ldlm_process_extent_lock(struct ldlm_lock *lock, int *flags, int first_enq,
622                              ldlm_error_t *err, struct list_head *work_list)
623 {
624         struct ldlm_resource *res = lock->l_resource;
625         struct list_head rpc_list = CFS_LIST_HEAD_INIT(rpc_list);
626         int rc, rc2;
627         int contended_locks = 0;
628         ENTRY;
629
630         LASSERT(list_empty(&res->lr_converting));
631         LASSERT(!(*flags & LDLM_FL_DENY_ON_CONTENTION) ||
632                 !(lock->l_flags & LDLM_AST_DISCARD_DATA));
633         check_res_locked(res);
634         *err = ELDLM_OK;
635
636         if (!first_enq) {
637                 /* Careful observers will note that we don't handle -EWOULDBLOCK
638                  * here, but it's ok for a non-obvious reason -- compat_queue
639                  * can only return -EWOULDBLOCK if (flags & BLOCK_NOWAIT).
640                  * flags should always be zero here, and if that ever stops
641                  * being true, we want to find out. */
642                 LASSERT(*flags == 0);
643                 rc = ldlm_extent_compat_queue(&res->lr_granted, lock, flags,
644                                               err, NULL, &contended_locks);
645                 if (rc == 1) {
646                         rc = ldlm_extent_compat_queue(&res->lr_waiting, lock,
647                                                       flags, err, NULL,
648                                                       &contended_locks);
649                 }
650                 if (rc == 0)
651                         RETURN(LDLM_ITER_STOP);
652
653                 ldlm_resource_unlink_lock(lock);
654
655                 if (!OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_EVICT_RACE))
656                         ldlm_extent_policy(res, lock, flags);
657                 ldlm_grant_lock(lock, work_list);
658                 RETURN(LDLM_ITER_CONTINUE);
659         }
660
661  restart:
662         contended_locks = 0;
663         rc = ldlm_extent_compat_queue(&res->lr_granted, lock, flags, err,
664                                       &rpc_list, &contended_locks);
665         if (rc < 0)
666                 GOTO(out, rc); /* lock was destroyed */
667         if (rc == 2)
668                 goto grant;
669
670         rc2 = ldlm_extent_compat_queue(&res->lr_waiting, lock, flags, err,
671                                        &rpc_list, &contended_locks);
672         if (rc2 < 0)
673                 GOTO(out, rc = rc2); /* lock was destroyed */
674
675         if (rc + rc2 == 2) {
676         grant:
677                 ldlm_extent_policy(res, lock, flags);
678                 ldlm_resource_unlink_lock(lock);
679                 ldlm_grant_lock(lock, NULL);
680         } else {
681                 /* If either of the compat_queue()s returned failure, then we
682                  * have ASTs to send and must go onto the waiting list.
683                  *
684                  * bug 2322: we used to unlink and re-add here, which was a
685                  * terrible folly -- if we goto restart, we could get
686                  * re-ordered!  Causes deadlock, because ASTs aren't sent! */
687                 if (list_empty(&lock->l_res_link))
688                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
689                 unlock_res(res);
690                 rc = ldlm_run_ast_work(&rpc_list, LDLM_WORK_BL_AST);
691                 lock_res(res);
692
693                 if (rc == -ERESTART) {
694                         /* lock was granted while resource was unlocked. */
695                         if (lock->l_granted_mode == lock->l_req_mode) {
696                                 /* bug 11300: if the lock has been granted,
697                                  * break earlier because otherwise, we will go
698                                  * to restart and ldlm_resource_unlink will be
699                                  * called and it causes the interval node to be
700                                  * freed. Then we will fail at 
701                                  * ldlm_extent_add_lock() */
702                                 *flags &= ~(LDLM_FL_BLOCK_GRANTED | LDLM_FL_BLOCK_CONV |
703                                             LDLM_FL_BLOCK_WAIT);
704                                 GOTO(out, rc = 0);
705                         }
706
707                         GOTO(restart, -ERESTART);
708                 }
709
710                 *flags |= LDLM_FL_BLOCK_GRANTED;
711                 /* this way we force client to wait for the lock
712                  * endlessly once the lock is enqueued -bzzz */
713                 *flags |= LDLM_FL_NO_TIMEOUT;
714
715         }
716         RETURN(0);
717 out:
718         if (!list_empty(&rpc_list)) {
719                 LASSERT(!(lock->l_flags & LDLM_AST_DISCARD_DATA));
720                 discard_bl_list(&rpc_list);
721         }
722         RETURN(rc);
723 }
724
725 /* When a lock is cancelled by a client, the KMS may undergo change if this
726  * is the "highest lock".  This function returns the new KMS value.
727  * Caller must hold ns_lock already.
728  *
729  * NB: A lock on [x,y] protects a KMS of up to y + 1 bytes! */
730 __u64 ldlm_extent_shift_kms(struct ldlm_lock *lock, __u64 old_kms)
731 {
732         struct ldlm_resource *res = lock->l_resource;
733         struct list_head *tmp;
734         struct ldlm_lock *lck;
735         __u64 kms = 0;
736         ENTRY;
737
738         /* don't let another thread in ldlm_extent_shift_kms race in
739          * just after we finish and take our lock into account in its
740          * calculation of the kms */
741         lock->l_flags |= LDLM_FL_KMS_IGNORE;
742
743         list_for_each(tmp, &res->lr_granted) {
744                 lck = list_entry(tmp, struct ldlm_lock, l_res_link);
745
746                 if (lck->l_flags & LDLM_FL_KMS_IGNORE)
747                         continue;
748
749                 if (lck->l_policy_data.l_extent.end >= old_kms)
750                         RETURN(old_kms);
751
752                 /* This extent _has_ to be smaller than old_kms (checked above)
753                  * so kms can only ever be smaller or the same as old_kms. */
754                 if (lck->l_policy_data.l_extent.end + 1 > kms)
755                         kms = lck->l_policy_data.l_extent.end + 1;
756         }
757         LASSERTF(kms <= old_kms, "kms "LPU64" old_kms "LPU64"\n", kms, old_kms);
758
759         RETURN(kms);
760 }
761
762 cfs_mem_cache_t *ldlm_interval_slab;
763 struct ldlm_interval *ldlm_interval_alloc(struct ldlm_lock *lock)
764 {
765         struct ldlm_interval *node;
766         ENTRY;
767
768         LASSERT(lock->l_resource->lr_type == LDLM_EXTENT);
769         OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO, sizeof(*node));
770         if (node == NULL)
771                 RETURN(NULL);
772
773         CFS_INIT_LIST_HEAD(&node->li_group);
774         ldlm_interval_attach(node, lock);
775         RETURN(node);
776 }
777
778 void ldlm_interval_free(struct ldlm_interval *node)
779 {
780         if (node) {
781                 LASSERT(list_empty(&node->li_group));
782                 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
783         }
784 }
785
786 /* interval tree, for LDLM_EXTENT. */
787 void ldlm_interval_attach(struct ldlm_interval *n,
788                           struct ldlm_lock *l)
789 {
790         LASSERT(l->l_tree_node == NULL);
791         LASSERT(l->l_resource->lr_type == LDLM_EXTENT);
792
793         list_add_tail(&l->l_sl_policy, &n->li_group);
794         l->l_tree_node = n;
795 }
796
797 struct ldlm_interval *ldlm_interval_detach(struct ldlm_lock *l)
798 {
799         struct ldlm_interval *n = l->l_tree_node;
800
801         if (n == NULL)
802                 return NULL;
803
804         LASSERT(!list_empty(&n->li_group));
805         l->l_tree_node = NULL;
806         list_del_init(&l->l_sl_policy);
807
808         return (list_empty(&n->li_group) ? n : NULL);
809 }
810
811 static inline int lock_mode_to_index(ldlm_mode_t mode)
812 {
813         int index;
814
815         LASSERT(mode != 0);
816         LASSERT(IS_PO2(mode));
817         for (index = -1; mode; index++, mode >>= 1) ;
818         LASSERT(index < LCK_MODE_NUM);
819         return index;
820 }
821
822 void ldlm_extent_add_lock(struct ldlm_resource *res,
823                           struct ldlm_lock *lock)
824 {
825         struct interval_node *found, **root;
826         struct ldlm_interval *node;
827         struct ldlm_extent *extent;
828         int idx;
829
830         LASSERT(lock->l_granted_mode == lock->l_req_mode);
831
832         node = lock->l_tree_node;
833         LASSERT(node != NULL);
834
835         idx = lock_mode_to_index(lock->l_granted_mode);
836         LASSERT(lock->l_granted_mode == 1 << idx);
837         LASSERT(lock->l_granted_mode == res->lr_itree[idx].lit_mode);
838
839         /* node extent initialize */
840         extent = &lock->l_policy_data.l_extent;
841         interval_set(&node->li_node, extent->start, extent->end);
842
843         root = &res->lr_itree[idx].lit_root;
844         found = interval_insert(&node->li_node, root);
845         if (found) { /* The policy group found. */
846                 struct ldlm_interval *tmp = ldlm_interval_detach(lock);
847                 LASSERT(tmp != NULL);
848                 ldlm_interval_free(tmp);
849                 ldlm_interval_attach(to_ldlm_interval(found), lock);
850         }
851         res->lr_itree[idx].lit_size++;
852
853         /* even though we use interval tree to manage the extent lock, we also
854          * add the locks into grant list, for debug purpose, .. */
855         ldlm_resource_add_lock(res, &res->lr_granted, lock);
856 }
857
858 void ldlm_extent_unlink_lock(struct ldlm_lock *lock)
859 {
860         struct ldlm_resource *res = lock->l_resource;
861         struct ldlm_interval *node;
862         struct ldlm_interval_tree *tree;
863         int idx;
864
865         if (lock->l_granted_mode != lock->l_req_mode)
866                 return;
867
868         LASSERT(lock->l_tree_node != NULL);
869         idx = lock_mode_to_index(lock->l_granted_mode);
870         LASSERT(lock->l_granted_mode == 1 << idx);
871         tree = &res->lr_itree[idx];
872
873         LASSERT(tree->lit_root != NULL); /* assure the tree is not null */
874
875         tree->lit_size--;
876         node = ldlm_interval_detach(lock);
877         if (node) {
878                 interval_erase(&node->li_node, &tree->lit_root);
879                 ldlm_interval_free(node);
880         }
881 }