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
62 * It should iterate through all waiting locks on a given resource queue and
63 * attempt to grant them. An optimization is to check only heads waitintg
64 * locks for each inodebit type.
66 * Must be called with resource lock held.
68 int ldlm_reprocess_inodebits_queue(struct ldlm_resource *res,
69 struct list_head *queue,
70 struct list_head *work_list,
71 enum ldlm_process_intention intention,
72 struct ldlm_lock *hint)
75 int rc = LDLM_ITER_CONTINUE;
77 LIST_HEAD(bl_ast_list);
78 struct ldlm_ibits_queues *queues = res->lr_ibits_queues;
83 check_res_locked(res);
85 LASSERT(res->lr_type == LDLM_IBITS);
86 LASSERT(intention == LDLM_PROCESS_RESCAN ||
87 intention == LDLM_PROCESS_RECOVERY);
89 if (intention == LDLM_PROCESS_RECOVERY)
90 return ldlm_reprocess_queue(res, queue, work_list, intention,
94 CDEBUG(D_DLMTRACE, "--- Reprocess resource "DLDLMRES" (%p)\n",
97 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++) {
99 struct list_head *head = &queues->liq_waiting[i];
100 struct ldlm_lock *pending;
101 struct ldlm_ibits_node *node;
103 if (list_empty(head))
105 if (hint && !(hint->l_policy_data.l_inodebits.bits & BIT(i)))
108 node = list_entry(head->next, struct ldlm_ibits_node,
111 pending = node->lock;
112 LDLM_DEBUG(pending, "Reprocessing lock from queue %d", i);
115 rc = ldlm_process_inodebits_lock(pending, &flags, intention,
117 if (ldlm_is_granted(pending)) {
118 list_splice(&rpc_list, work_list);
119 /* Try to grant more locks from current queue */
122 list_splice(&rpc_list, &bl_ast_list);
126 if (!list_empty(&bl_ast_list)) {
129 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &bl_ast_list,
137 if (!list_empty(&bl_ast_list))
138 ldlm_discard_bl_list(&bl_ast_list);
144 * Determine if the lock is compatible with all locks on the queue.
146 * If \a work_list is provided, conflicting locks are linked there.
147 * If \a work_list is not provided, we exit this function on first conflict.
149 * \retval 0 if there are conflicting locks in the \a queue
150 * \retval 1 if the lock is compatible to all locks in \a queue
152 * IBITS locks in granted queue are organized in bunches of
153 * same-mode/same-bits locks called "skip lists". The First lock in the
154 * bunch contains a pointer to the end of the bunch. This allows us to
155 * skip an entire bunch when iterating the list in search for conflicting
156 * locks if first lock of the bunch is not conflicting with us.
159 ldlm_inodebits_compat_queue(struct list_head *queue, struct ldlm_lock *req,
160 struct list_head *work_list)
162 struct list_head *tmp;
163 struct ldlm_lock *lock;
164 __u64 req_bits = req->l_policy_data.l_inodebits.bits;
165 __u64 *try_bits = &req->l_policy_data.l_inodebits.try_bits;
170 /* There is no sense in lock with no bits set. Also such a lock
171 * would be compatible with any other bit lock.
172 * Meanwhile that can be true if there were just try_bits and all
173 * are failed, so just exit gracefully and let the caller to care.
175 if ((req_bits | *try_bits) == 0)
178 list_for_each(tmp, queue) {
179 struct list_head *mode_tail;
181 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
183 /* We stop walking the queue if we hit ourselves so we don't
184 * take conflicting locks enqueued after us into account,
185 * or we'd wait forever. */
189 /* last lock in mode group */
190 LASSERT(lock->l_sl_mode.prev != NULL);
191 mode_tail = &list_entry(lock->l_sl_mode.prev, struct ldlm_lock,
192 l_sl_mode)->l_res_link;
194 /* if request lock is not COS_INCOMPAT and COS is disabled,
195 * they are compatible, IOW this request is from a local
196 * transaction on a DNE system. */
197 if (lock->l_req_mode == LCK_COS && !ldlm_is_cos_incompat(req) &&
198 !ldlm_is_cos_enabled(req)) {
199 /* jump to last lock in mode group */
204 /* locks' mode are compatible, bits don't matter */
205 if (lockmode_compat(lock->l_req_mode, req->l_req_mode)) {
206 /* jump to last lock in mode group */
212 struct list_head *head;
214 /* Advance loop cursor to last lock in policy group. */
215 tmp = &list_entry(lock->l_sl_policy.prev,
217 l_sl_policy)->l_res_link;
219 /* New lock's try_bits are filtered out by ibits
220 * of all locks in both granted and waiting queues.
222 *try_bits &= ~(lock->l_policy_data.l_inodebits.bits |
223 lock->l_policy_data.l_inodebits.try_bits);
225 if ((req_bits | *try_bits) == 0)
228 /* The new lock ibits is more preferable than try_bits
229 * of waiting locks so drop conflicting try_bits in
231 * Notice that try_bits of granted locks must be zero.
233 lock->l_policy_data.l_inodebits.try_bits &= ~req_bits;
235 /* Locks with overlapping bits conflict. */
236 if (lock->l_policy_data.l_inodebits.bits & req_bits) {
237 /* COS lock mode has a special compatibility
238 * requirement: it is only compatible with
239 * locks from the same client. */
240 if (lock->l_req_mode == LCK_COS &&
241 !ldlm_is_cos_incompat(req) &&
242 ldlm_is_cos_enabled(req) &&
243 lock->l_client_cookie == req->l_client_cookie)
244 goto not_conflicting;
246 /* Found a conflicting policy group. */
252 /* Add locks of the policy group to @work_list
253 * as blocking locks for @req */
254 if (lock->l_blocking_ast)
255 ldlm_add_ast_work_item(lock, req,
257 head = &lock->l_sl_policy;
258 list_for_each_entry(lock, head, l_sl_policy)
259 if (lock->l_blocking_ast)
260 ldlm_add_ast_work_item(lock,
264 if (tmp == mode_tail)
268 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
269 } /* Loop over policy groups within one mode group. */
270 } /* Loop over mode groups within @queue. */
276 * Process a granting attempt for IBITS lock.
277 * Must be called with ns lock held
279 * This function looks for any conflicts for \a lock in the granted or
280 * waiting queues. The lock is granted if no conflicts are found in
283 int ldlm_process_inodebits_lock(struct ldlm_lock *lock, __u64 *flags,
284 enum ldlm_process_intention intention,
285 enum ldlm_error *err,
286 struct list_head *work_list)
288 struct ldlm_resource *res = lock->l_resource;
289 struct list_head *grant_work = intention == LDLM_PROCESS_ENQUEUE ?
294 *err = ELDLM_LOCK_ABORTED;
295 LASSERT(!ldlm_is_granted(lock));
296 check_res_locked(res);
298 if (intention == LDLM_PROCESS_RESCAN) {
299 struct list_head *bl_list =
300 *flags & LDLM_FL_BLOCK_NOWAIT ? NULL : work_list;
302 LASSERT(lock->l_policy_data.l_inodebits.bits != 0);
304 /* It is possible that some of granted locks was not canceled
305 * but converted and is kept in granted queue. So there is
306 * a window where lock with 'ast_sent' might become granted
307 * again. Meanwhile a new lock may appear in that window and
308 * conflicts with the converted lock so the following scenario
311 * 1) lock1 conflicts with lock2
312 * 2) bl_ast was sent for lock2
313 * 3) lock3 comes and conflicts with lock2 too
314 * 4) no bl_ast sent because lock2->l_bl_ast_sent is 1
315 * 5) lock2 was converted for lock1 but not for lock3
316 * 6) lock1 granted, lock3 still is waiting for lock2, but
317 * there will never be another bl_ast for that
319 * To avoid this scenario the work_list is used below to collect
320 * any blocked locks from granted queue during every reprocess
321 * and bl_ast will be sent if needed.
323 rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock,
326 RETURN(LDLM_ITER_STOP);
327 rc = ldlm_inodebits_compat_queue(&res->lr_waiting, lock, NULL);
329 RETURN(LDLM_ITER_STOP);
331 /* grant also try_bits if any */
332 if (lock->l_policy_data.l_inodebits.try_bits != 0) {
333 lock->l_policy_data.l_inodebits.bits |=
334 lock->l_policy_data.l_inodebits.try_bits;
335 lock->l_policy_data.l_inodebits.try_bits = 0;
336 *flags |= LDLM_FL_LOCK_CHANGED;
338 ldlm_resource_unlink_lock(lock);
339 ldlm_grant_lock(lock, grant_work);
342 RETURN(LDLM_ITER_CONTINUE);
345 rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock, work_list);
346 rc += ldlm_inodebits_compat_queue(&res->lr_waiting, lock, work_list);
349 /* if there were only bits to try and all are conflicting */
350 if ((lock->l_policy_data.l_inodebits.bits |
351 lock->l_policy_data.l_inodebits.try_bits)) {
352 /* There is no sense to set LDLM_FL_NO_TIMEOUT to @flags
353 * for DOM lock while they are enqueued through intents,
354 * i.e. @lock here is local which does not timeout. */
358 /* grant also all remaining try_bits */
359 if (lock->l_policy_data.l_inodebits.try_bits != 0) {
360 lock->l_policy_data.l_inodebits.bits |=
361 lock->l_policy_data.l_inodebits.try_bits;
362 lock->l_policy_data.l_inodebits.try_bits = 0;
363 *flags |= LDLM_FL_LOCK_CHANGED;
365 LASSERT(lock->l_policy_data.l_inodebits.bits);
366 ldlm_resource_unlink_lock(lock);
367 ldlm_grant_lock(lock, grant_work);
371 RETURN(LDLM_ITER_CONTINUE);
373 #endif /* HAVE_SERVER_SUPPORT */
375 void ldlm_ibits_policy_wire_to_local(const union ldlm_wire_policy_data *wpolicy,
376 union ldlm_policy_data *lpolicy)
378 lpolicy->l_inodebits.bits = wpolicy->l_inodebits.bits;
380 * try_bits are to be handled outside of generic write_to_local due
381 * to different behavior on a server and client.
385 void ldlm_ibits_policy_local_to_wire(const union ldlm_policy_data *lpolicy,
386 union ldlm_wire_policy_data *wpolicy)
388 memset(wpolicy, 0, sizeof(*wpolicy));
389 wpolicy->l_inodebits.bits = lpolicy->l_inodebits.bits;
390 wpolicy->l_inodebits.try_bits = lpolicy->l_inodebits.try_bits;
394 * Attempt to convert already granted IBITS lock with several bits set to
395 * a lock with less bits (downgrade).
397 * Such lock conversion is used to keep lock with non-blocking bits instead of
398 * cancelling it, introduced for better support of DoM files.
400 int ldlm_inodebits_drop(struct ldlm_lock *lock, __u64 to_drop)
404 check_res_locked(lock->l_resource);
406 /* Just return if there are no conflicting bits */
407 if ((lock->l_policy_data.l_inodebits.bits & to_drop) == 0) {
408 LDLM_WARN(lock, "try to drop unset bits %#llx/%#llx",
409 lock->l_policy_data.l_inodebits.bits, to_drop);
414 /* remove lock from a skiplist and put in the new place
415 * according with new inodebits */
416 ldlm_resource_unlink_lock(lock);
417 lock->l_policy_data.l_inodebits.bits &= ~to_drop;
418 ldlm_grant_lock_with_skiplist(lock);
421 EXPORT_SYMBOL(ldlm_inodebits_drop);
423 /* convert single lock */
424 int ldlm_cli_inodebits_convert(struct ldlm_lock *lock,
425 enum ldlm_cancel_flags cancel_flags)
427 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
428 struct ldlm_lock_desc ld = { { 0 } };
429 __u64 drop_bits, new_bits;
435 check_res_locked(lock->l_resource);
437 /* Lock is being converted already */
438 if (ldlm_is_converting(lock)) {
439 if (!(cancel_flags & LCF_ASYNC)) {
440 unlock_res_and_lock(lock);
441 wait_event_idle(lock->l_waitq,
442 is_lock_converted(lock));
443 lock_res_and_lock(lock);
448 /* lru_cancel may happen in parallel and call ldlm_cli_cancel_list()
451 if (ldlm_is_canceling(lock))
454 /* no need in only local convert */
455 if (lock->l_flags & (LDLM_FL_LOCAL_ONLY | LDLM_FL_CANCEL_ON_BLOCK))
458 drop_bits = lock->l_policy_data.l_inodebits.cancel_bits;
459 /* no cancel bits - means that caller needs full cancel */
463 new_bits = lock->l_policy_data.l_inodebits.bits & ~drop_bits;
464 /* check if all lock bits are dropped, proceed with cancel */
468 /* check if no dropped bits, consider this as successful convert */
469 if (lock->l_policy_data.l_inodebits.bits == new_bits)
472 ldlm_set_converting(lock);
473 /* Finally call cancel callback for remaining bits only.
474 * It is important to have converting flag during that
475 * so blocking_ast callback can distinguish convert from
478 ld.l_policy_data.l_inodebits.cancel_bits = drop_bits;
479 unlock_res_and_lock(lock);
480 lock->l_blocking_ast(lock, &ld, lock->l_ast_data, LDLM_CB_CANCELING);
481 /* now notify server about convert */
482 rc = ldlm_cli_convert_req(lock, &flags, new_bits);
483 lock_res_and_lock(lock);
485 GOTO(full_cancel, rc);
487 /* Finally clear these bits in lock ibits */
488 ldlm_inodebits_drop(lock, drop_bits);
490 /* Being locked again check if lock was canceled, it is important
491 * to do and don't drop cbpending below
493 if (ldlm_is_canceling(lock))
494 GOTO(full_cancel, rc = -EINVAL);
496 /* also check again if more bits to be cancelled appeared */
497 if (drop_bits != lock->l_policy_data.l_inodebits.cancel_bits)
498 GOTO(clear_converting, rc = -EAGAIN);
500 /* clear cbpending flag early, it is safe to match lock right after
501 * client convert because it is downgrade always.
503 ldlm_clear_cbpending(lock);
504 ldlm_clear_bl_ast(lock);
505 spin_lock(&ns->ns_lock);
506 if (list_empty(&lock->l_lru))
507 ldlm_lock_add_to_lru_nolock(lock);
508 spin_unlock(&ns->ns_lock);
510 /* the job is done, zero the cancel_bits. If more conflicts appear,
511 * it will result in another cycle of ldlm_cli_inodebits_convert().
514 lock->l_policy_data.l_inodebits.cancel_bits = 0;
516 ldlm_clear_converting(lock);
520 int ldlm_inodebits_alloc_lock(struct ldlm_lock *lock)
522 if (ldlm_is_ns_srv(lock)) {
525 OBD_SLAB_ALLOC_PTR(lock->l_ibits_node, ldlm_inodebits_slab);
526 if (lock->l_ibits_node == NULL)
528 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
529 INIT_LIST_HEAD(&lock->l_ibits_node->lin_link[i]);
530 lock->l_ibits_node->lock = lock;
532 lock->l_ibits_node = NULL;
537 void ldlm_inodebits_add_lock(struct ldlm_resource *res, struct list_head *head,
538 struct ldlm_lock *lock)
542 if (!ldlm_is_ns_srv(lock))
545 if (head == &res->lr_waiting) {
546 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++) {
547 if (lock->l_policy_data.l_inodebits.bits & BIT(i))
548 list_add_tail(&lock->l_ibits_node->lin_link[i],
549 &res->lr_ibits_queues->liq_waiting[i]);
551 } else if (head == &res->lr_granted && lock->l_ibits_node != NULL) {
552 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
553 LASSERT(list_empty(&lock->l_ibits_node->lin_link[i]));
554 OBD_SLAB_FREE_PTR(lock->l_ibits_node, ldlm_inodebits_slab);
555 lock->l_ibits_node = NULL;
559 void ldlm_inodebits_unlink_lock(struct ldlm_lock *lock)
563 ldlm_unlink_lock_skiplist(lock);
564 if (!ldlm_is_ns_srv(lock))
567 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
568 list_del_init(&lock->l_ibits_node->lin_link[i]);