4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
23 * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ldlm/ldlm_inodebits.c
34 * Author: Peter Braam <braam@clusterfs.com>
35 * Author: Phil Schwan <phil@clusterfs.com>
39 * This file contains implementation of IBITS lock type
41 * IBITS lock type contains a bit mask determining various properties of an
42 * object. The meanings of specific bits are specific to the caller and are
43 * opaque to LDLM code.
45 * Locks with intersecting bitmasks and conflicting lock modes (e.g. LCK_PW)
46 * are considered conflicting. See the lock mode compatibility matrix
50 #define DEBUG_SUBSYSTEM S_LDLM
52 #include <lustre_dlm.h>
53 #include <obd_support.h>
54 #include <lustre_lib.h>
55 #include <obd_class.h>
57 #include "ldlm_internal.h"
59 #ifdef HAVE_SERVER_SUPPORT
61 * local lock will be canceled after use, and it should run blocking ast only
62 * when it should trigger Commit-on-Sharing, otherwise if the blocking ast
63 * is run and does nothing, it will prevent subsequent operations to trigger
64 * Commit-on-Sharing, see LU-11102.
66 static bool ldlm_should_run_bl_ast(const struct ldlm_lock *lock,
67 const struct ldlm_lock *req)
70 if (!lock->l_blocking_ast)
74 if (!ldlm_is_local(lock))
77 /* should trigger Commit-on-Sharing */
78 if ((lock->l_req_mode & LCK_COS))
81 /* local read lock will be canceld after use */
82 if (!(lock->l_req_mode & (LCK_PW | LCK_EX)))
85 /* if CoS enabled, check if @req is from different client */
86 if (ldlm_is_cos_enabled(req))
87 return lock->l_client_cookie != req->l_client_cookie;
89 /* check if @req is COS incompatible */
90 if (ldlm_is_cos_incompat(req))
97 * Determine if the lock is compatible with all locks on the queue.
99 * If \a work_list is provided, conflicting locks are linked there.
100 * If \a work_list is not provided, we exit this function on first conflict.
102 * \retval 0 if there are conflicting locks in the \a queue
103 * \retval 1 if the lock is compatible to all locks in \a queue
105 * IBITS locks in granted queue are organized in bunches of
106 * same-mode/same-bits locks called "skip lists". The First lock in the
107 * bunch contains a pointer to the end of the bunch. This allows us to
108 * skip an entire bunch when iterating the list in search for conflicting
109 * locks if first lock of the bunch is not conflicting with us.
112 ldlm_inodebits_compat_queue(struct list_head *queue, struct ldlm_lock *req,
113 struct list_head *work_list)
115 struct list_head *tmp;
116 struct ldlm_lock *lock;
117 __u64 req_bits = req->l_policy_data.l_inodebits.bits;
118 __u64 *try_bits = &req->l_policy_data.l_inodebits.try_bits;
123 /* There is no sense in lock with no bits set. Also such a lock
124 * would be compatible with any other bit lock.
125 * Meanwhile that can be true if there were just try_bits and all
126 * are failed, so just exit gracefully and let the caller to care.
128 if ((req_bits | *try_bits) == 0)
131 list_for_each(tmp, queue) {
132 struct list_head *mode_tail;
134 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
136 /* We stop walking the queue if we hit ourselves so we don't
137 * take conflicting locks enqueued after us into account,
138 * or we'd wait forever. */
142 /* last lock in mode group */
143 LASSERT(lock->l_sl_mode.prev != NULL);
144 mode_tail = &list_entry(lock->l_sl_mode.prev, struct ldlm_lock,
145 l_sl_mode)->l_res_link;
147 /* if request lock is not COS_INCOMPAT and COS is disabled,
148 * they are compatible, IOW this request is from a local
149 * transaction on a DNE system. */
150 if (lock->l_req_mode == LCK_COS && !ldlm_is_cos_incompat(req) &&
151 !ldlm_is_cos_enabled(req)) {
152 /* jump to last lock in mode group */
157 /* locks' mode are compatible, bits don't matter */
158 if (lockmode_compat(lock->l_req_mode, req->l_req_mode)) {
159 /* jump to last lock in mode group */
165 struct list_head *head;
167 /* Advance loop cursor to last lock in policy group. */
168 tmp = &list_entry(lock->l_sl_policy.prev,
170 l_sl_policy)->l_res_link;
172 /* New lock's try_bits are filtered out by ibits
173 * of all locks in both granted and waiting queues.
175 *try_bits &= ~(lock->l_policy_data.l_inodebits.bits |
176 lock->l_policy_data.l_inodebits.try_bits);
178 if ((req_bits | *try_bits) == 0)
181 /* The new lock ibits is more preferable than try_bits
182 * of waiting locks so drop conflicting try_bits in
184 * Notice that try_bits of granted locks must be zero.
186 lock->l_policy_data.l_inodebits.try_bits &= ~req_bits;
188 /* Locks with overlapping bits conflict. */
189 if (lock->l_policy_data.l_inodebits.bits & req_bits) {
190 /* COS lock mode has a special compatibility
191 * requirement: it is only compatible with
192 * locks from the same client. */
193 if (lock->l_req_mode == LCK_COS &&
194 !ldlm_is_cos_incompat(req) &&
195 ldlm_is_cos_enabled(req) &&
196 lock->l_client_cookie == req->l_client_cookie)
197 goto not_conflicting;
199 /* Found a conflicting policy group. */
205 /* Add locks of the policy group to @work_list
206 * as blocking locks for @req */
207 if (ldlm_should_run_bl_ast(lock, req))
208 ldlm_add_ast_work_item(lock, req,
210 head = &lock->l_sl_policy;
211 list_for_each_entry(lock, head, l_sl_policy)
212 if (ldlm_should_run_bl_ast(lock, req))
213 ldlm_add_ast_work_item(lock,
217 if (tmp == mode_tail)
221 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
222 } /* Loop over policy groups within one mode group. */
223 } /* Loop over mode groups within @queue. */
229 * Process a granting attempt for IBITS lock.
230 * Must be called with ns lock held
232 * This function looks for any conflicts for \a lock in the granted or
233 * waiting queues. The lock is granted if no conflicts are found in
236 int ldlm_process_inodebits_lock(struct ldlm_lock *lock, __u64 *flags,
237 enum ldlm_process_intention intention,
238 enum ldlm_error *err,
239 struct list_head *work_list)
241 struct ldlm_resource *res = lock->l_resource;
242 struct list_head *grant_work = intention == LDLM_PROCESS_ENQUEUE ?
248 LASSERT(lock->l_granted_mode != lock->l_req_mode);
249 check_res_locked(res);
251 if (intention == LDLM_PROCESS_RESCAN) {
252 struct list_head *bl_list;
254 if (*flags & LDLM_FL_BLOCK_NOWAIT) {
256 *err = ELDLM_LOCK_WOULDBLOCK;
259 *err = ELDLM_LOCK_ABORTED;
262 LASSERT(lock->l_policy_data.l_inodebits.bits != 0);
264 /* It is possible that some of granted locks was not canceled
265 * but converted and is kept in granted queue. So there is
266 * a window where lock with 'ast_sent' might become granted
267 * again. Meanwhile a new lock may appear in that window and
268 * conflicts with the converted lock so the following scenario
271 * 1) lock1 conflicts with lock2
272 * 2) bl_ast was sent for lock2
273 * 3) lock3 comes and conflicts with lock2 too
274 * 4) no bl_ast sent because lock2->l_bl_ast_sent is 1
275 * 5) lock2 was converted for lock1 but not for lock3
276 * 6) lock1 granted, lock3 still is waiting for lock2, but
277 * there will never be another bl_ast for that
279 * To avoid this scenario the work_list is used below to collect
280 * any blocked locks from granted queue during every reprocess
281 * and bl_ast will be sent if needed.
283 rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock,
286 RETURN(LDLM_ITER_STOP);
287 rc = ldlm_inodebits_compat_queue(&res->lr_waiting, lock, NULL);
289 RETURN(LDLM_ITER_STOP);
291 /* grant also try_bits if any */
292 if (lock->l_policy_data.l_inodebits.try_bits != 0) {
293 lock->l_policy_data.l_inodebits.bits |=
294 lock->l_policy_data.l_inodebits.try_bits;
295 lock->l_policy_data.l_inodebits.try_bits = 0;
296 *flags |= LDLM_FL_LOCK_CHANGED;
298 ldlm_resource_unlink_lock(lock);
299 ldlm_grant_lock(lock, grant_work);
302 RETURN(LDLM_ITER_CONTINUE);
305 rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock, work_list);
306 rc += ldlm_inodebits_compat_queue(&res->lr_waiting, lock, work_list);
309 /* if there were only bits to try and all are conflicting */
310 if ((lock->l_policy_data.l_inodebits.bits |
311 lock->l_policy_data.l_inodebits.try_bits) == 0) {
312 *err = ELDLM_LOCK_WOULDBLOCK;
317 /* grant also all remaining try_bits */
318 if (lock->l_policy_data.l_inodebits.try_bits != 0) {
319 lock->l_policy_data.l_inodebits.bits |=
320 lock->l_policy_data.l_inodebits.try_bits;
321 lock->l_policy_data.l_inodebits.try_bits = 0;
322 *flags |= LDLM_FL_LOCK_CHANGED;
324 LASSERT(lock->l_policy_data.l_inodebits.bits);
325 ldlm_resource_unlink_lock(lock);
326 ldlm_grant_lock(lock, grant_work);
330 RETURN(LDLM_ITER_CONTINUE);
332 #endif /* HAVE_SERVER_SUPPORT */
334 void ldlm_ibits_policy_wire_to_local(const union ldlm_wire_policy_data *wpolicy,
335 union ldlm_policy_data *lpolicy)
337 lpolicy->l_inodebits.bits = wpolicy->l_inodebits.bits;
338 lpolicy->l_inodebits.try_bits = wpolicy->l_inodebits.try_bits;
341 void ldlm_ibits_policy_local_to_wire(const union ldlm_policy_data *lpolicy,
342 union ldlm_wire_policy_data *wpolicy)
344 memset(wpolicy, 0, sizeof(*wpolicy));
345 wpolicy->l_inodebits.bits = lpolicy->l_inodebits.bits;
346 wpolicy->l_inodebits.try_bits = lpolicy->l_inodebits.try_bits;
350 * Attempt to convert already granted IBITS lock with several bits set to
351 * a lock with less bits (downgrade).
353 * Such lock conversion is used to keep lock with non-blocking bits instead of
354 * cancelling it, introduced for better support of DoM files.
356 int ldlm_inodebits_drop(struct ldlm_lock *lock, __u64 to_drop)
360 check_res_locked(lock->l_resource);
362 /* Just return if there are no conflicting bits */
363 if ((lock->l_policy_data.l_inodebits.bits & to_drop) == 0) {
364 LDLM_WARN(lock, "try to drop unset bits %#llx/%#llx",
365 lock->l_policy_data.l_inodebits.bits, to_drop);
370 /* remove lock from a skiplist and put in the new place
371 * according with new inodebits */
372 ldlm_resource_unlink_lock(lock);
373 lock->l_policy_data.l_inodebits.bits &= ~to_drop;
374 ldlm_grant_lock_with_skiplist(lock);
377 EXPORT_SYMBOL(ldlm_inodebits_drop);
379 /* convert single lock */
380 int ldlm_cli_dropbits(struct ldlm_lock *lock, __u64 drop_bits)
382 struct lustre_handle lockh;
389 LASSERT(!lock->l_readers && !lock->l_writers);
391 LDLM_DEBUG(lock, "client lock convert START");
393 ldlm_lock2handle(lock, &lockh);
394 lock_res_and_lock(lock);
395 /* check if all bits are blocked */
396 if (!(lock->l_policy_data.l_inodebits.bits & ~drop_bits)) {
397 unlock_res_and_lock(lock);
398 /* return error to continue with cancel */
399 GOTO(exit, rc = -EINVAL);
402 /* check if no common bits, consider this as successful convert */
403 if (!(lock->l_policy_data.l_inodebits.bits & drop_bits)) {
404 unlock_res_and_lock(lock);
408 /* check if there is race with cancel */
409 if (ldlm_is_canceling(lock) || ldlm_is_cancel(lock)) {
410 unlock_res_and_lock(lock);
411 GOTO(exit, rc = -EINVAL);
414 /* clear cbpending flag early, it is safe to match lock right after
415 * client convert because it is downgrade always.
417 ldlm_clear_cbpending(lock);
418 ldlm_clear_bl_ast(lock);
420 /* If lock is being converted already, check drop bits first */
421 if (ldlm_is_converting(lock)) {
422 /* raced lock convert, lock inodebits are remaining bits
423 * so check if they are conflicting with new convert or not.
425 if (!(lock->l_policy_data.l_inodebits.bits & drop_bits)) {
426 unlock_res_and_lock(lock);
429 /* Otherwise drop new conflicting bits in new convert */
431 ldlm_set_converting(lock);
432 /* from all bits of blocking lock leave only conflicting */
433 drop_bits &= lock->l_policy_data.l_inodebits.bits;
434 /* save them in cancel_bits, so l_blocking_ast will know
435 * which bits from the current lock were dropped. */
436 lock->l_policy_data.l_inodebits.cancel_bits = drop_bits;
437 /* Finally clear these bits in lock ibits */
438 ldlm_inodebits_drop(lock, drop_bits);
439 unlock_res_and_lock(lock);
440 /* Finally call cancel callback for remaining bits only.
441 * It is important to have converting flag during that
442 * so blocking_ast callback can distinguish convert from
445 if (lock->l_blocking_ast)
446 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
449 /* now notify server about convert */
450 rc = ldlm_cli_convert(lock, &flags);
452 lock_res_and_lock(lock);
453 if (ldlm_is_converting(lock)) {
454 ldlm_clear_converting(lock);
455 ldlm_set_cbpending(lock);
456 ldlm_set_bl_ast(lock);
458 unlock_res_and_lock(lock);
463 LDLM_DEBUG(lock, "client lock convert END");