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/
31 * lustre/ldlm/ldlm_inodebits.c
33 * Author: Peter Braam <braam@clusterfs.com>
34 * Author: Phil Schwan <phil@clusterfs.com>
38 * This file contains implementation of IBITS lock type
40 * IBITS lock type contains a bit mask determining various properties of an
41 * object. The meanings of specific bits are specific to the caller and are
42 * opaque to LDLM code.
44 * Locks with intersecting bitmasks and conflicting lock modes (e.g. LCK_PW)
45 * are considered conflicting. See the lock mode compatibility matrix
49 #define DEBUG_SUBSYSTEM S_LDLM
51 #include <lustre_dlm.h>
52 #include <obd_support.h>
53 #include <lustre_lib.h>
54 #include <obd_class.h>
56 #include "ldlm_internal.h"
58 #ifdef HAVE_SERVER_SUPPORT
61 * It should iterate through all waiting locks on a given resource queue and
62 * attempt to grant them. An optimization is to check only heads waitintg
63 * locks for each inodebit type.
65 * Must be called with resource lock held.
67 int ldlm_reprocess_inodebits_queue(struct ldlm_resource *res,
68 struct list_head *queue,
69 struct list_head *work_list,
70 enum ldlm_process_intention intention,
74 int rc = LDLM_ITER_CONTINUE;
76 LIST_HEAD(bl_ast_list);
77 struct ldlm_ibits_queues *queues = res->lr_ibits_queues;
82 check_res_locked(res);
84 LASSERT(res->lr_type == LDLM_IBITS);
85 LASSERT(intention == LDLM_PROCESS_RESCAN ||
86 intention == LDLM_PROCESS_RECOVERY);
88 if (intention == LDLM_PROCESS_RECOVERY)
89 return ldlm_reprocess_queue(res, queue, work_list, intention,
93 CDEBUG(D_DLMTRACE, "--- Reprocess resource "DLDLMRES" (%p)\n",
96 CDEBUG(D_DLMTRACE, "Hint %llx\n", mask);
98 mask = MDS_INODELOCK_FULL;
100 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++) {
102 struct list_head *head = &queues->liq_waiting[i];
103 struct ldlm_lock *pending;
104 struct ldlm_ibits_node *node;
106 if (list_empty(head) || !(mask & (1 << i)))
109 node = list_first_entry(head, struct ldlm_ibits_node,
112 pending = node->lock;
113 LDLM_DEBUG(pending, "Reprocessing lock from queue %d", i);
116 rc = ldlm_process_inodebits_lock(pending, &flags, intention,
118 if (ldlm_is_granted(pending)) {
119 list_splice(&rpc_list, work_list);
120 mask |= pending->l_policy_data.l_inodebits.bits;
121 i = ffs(pending->l_policy_data.l_inodebits.bits) - 2;
123 list_splice(&rpc_list, &bl_ast_list);
127 if (!list_empty(&bl_ast_list)) {
130 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &bl_ast_list,
134 if (rc == -ERESTART) {
140 if (!list_empty(&bl_ast_list))
141 ldlm_discard_bl_list(&bl_ast_list);
146 /* lock of COS mode is compatible with locks from the same client. */
147 static inline bool ldlm_cos_same_client(const struct ldlm_lock *req,
148 const struct ldlm_lock *lock)
150 return lock->l_req_mode == LCK_COS &&
151 lock->l_client_cookie == req->l_client_cookie;
154 /* lock of TXN mode is compatible with locks from the same MDT. */
155 static inline bool ldlm_txn_same_server(const struct ldlm_lock *req,
156 const struct ldlm_lock *lock)
158 return lock->l_req_mode == LCK_TXN &&
159 lock->l_policy_data.l_inodebits.li_initiator_id ==
160 req->l_policy_data.l_inodebits.li_initiator_id;
164 * Determine if the lock is compatible with all locks on the queue.
166 * If \a work_list is provided, conflicting locks are linked there.
167 * If \a work_list is not provided, we exit this function on first conflict.
169 * \retval 0 if there are conflicting locks in the \a queue
170 * \retval 1 if the lock is compatible to all locks in \a queue
172 * IBITS locks in granted queue are organized in bunches of
173 * same-mode/same-bits locks called "skip lists". The First lock in the
174 * bunch contains a pointer to the end of the bunch. This allows us to
175 * skip an entire bunch when iterating the list in search for conflicting
176 * locks if first lock of the bunch is not conflicting with us.
179 ldlm_inodebits_compat_queue(struct list_head *queue, struct ldlm_lock *req,
180 __u64 *ldlm_flags, struct list_head *work_list)
182 enum ldlm_mode req_mode = req->l_req_mode;
183 struct list_head *tmp;
184 struct ldlm_lock *lock;
185 __u64 req_bits = req->l_policy_data.l_inodebits.bits;
186 __u64 *try_bits = &req->l_policy_data.l_inodebits.try_bits;
191 lockmode_verify(req_mode);
193 /* There is no sense in lock with no bits set. Also such a lock
194 * would be compatible with any other bit lock.
195 * Meanwhile that can be true if there were just try_bits and all
196 * are failed, so just exit gracefully and let the caller to care.
198 if ((req_bits | *try_bits) == 0)
201 /* Group lock could be only DOM */
202 if (unlikely(req_mode == LCK_GROUP &&
203 (req_bits | *try_bits) != MDS_INODELOCK_DOM))
206 list_for_each(tmp, queue) {
207 struct list_head *mode_tail;
209 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
211 /* We stop walking the queue if we hit ourselves so we don't
212 * take conflicting locks enqueued after us into account,
213 * or we'd wait forever.
218 /* last lock in mode group */
219 LASSERT(lock->l_sl_mode.prev != NULL);
220 mode_tail = &list_entry(lock->l_sl_mode.prev, struct ldlm_lock,
221 l_sl_mode)->l_res_link;
223 if (lockmode_compat(lock->l_req_mode, req_mode)) {
224 /* non group locks are compatible, bits don't matter */
225 if (likely(req_mode != LCK_GROUP)) {
226 /* jump to last lock in mode group */
231 if (req->l_policy_data.l_inodebits.li_gid ==
232 lock->l_policy_data.l_inodebits.li_gid) {
233 if (ldlm_is_granted(lock))
236 if (*ldlm_flags & LDLM_FL_BLOCK_NOWAIT)
237 RETURN(-EWOULDBLOCK);
239 /* Place the same group together */
240 ldlm_resource_insert_lock_after(lock, req);
243 } else if (ldlm_cos_same_client(req, lock) ||
244 ldlm_txn_same_server(req, lock)) {
245 /* COS/TXN locks need to be checked one by one,
246 * because client cookie or initiator id may be
247 * different for locks in mode/policy skiplist.
253 /* GROUP(by gid) locks placed to a head of the waiting list */
254 if (unlikely(req_mode == LCK_GROUP && !ldlm_is_granted(lock))) {
256 if (lock->l_req_mode != LCK_GROUP) {
257 /* Already not a GROUP lock, insert before. */
258 ldlm_resource_insert_lock_before(lock, req);
261 /* Still GROUP but a different gid(the same gid would
262 * be handled above). Keep searching for the same gid
264 LASSERT(req->l_policy_data.l_inodebits.li_gid !=
265 lock->l_policy_data.l_inodebits.li_gid);
270 struct list_head *head;
272 /* Advance loop cursor to last lock in policy group. */
273 tmp = &list_entry(lock->l_sl_policy.prev,
275 l_sl_policy)->l_res_link;
277 /* New lock's try_bits are filtered out by ibits
278 * of all locks in both granted and waiting queues.
280 *try_bits &= ~(lock->l_policy_data.l_inodebits.bits |
281 lock->l_policy_data.l_inodebits.try_bits);
283 if ((req_bits | *try_bits) == 0)
286 /* The new lock ibits is more preferable than try_bits
287 * of waiting locks so drop conflicting try_bits in
289 * Notice that try_bits of granted locks must be zero.
291 lock->l_policy_data.l_inodebits.try_bits &= ~req_bits;
293 /* Locks with overlapping bits conflict. */
294 if (lock->l_policy_data.l_inodebits.bits & req_bits) {
297 if (unlikely(lock->l_req_mode == LCK_GROUP)) {
298 LASSERT(ldlm_has_dom(lock));
300 if (*ldlm_flags & LDLM_FL_BLOCK_NOWAIT)
301 RETURN(-EWOULDBLOCK);
303 /* Local combined DOM lock came across
304 * GROUP DOM lock, it makes the thread
305 * to be blocked for a long time, not
306 * allowed, the trybits to be used
309 if (!req->l_export &&
310 (req_bits & MDS_INODELOCK_DOM) &&
311 (req_bits & ~MDS_INODELOCK_DOM))
317 /* Found a conflicting policy group. */
321 /* Add locks of the policy group to @work_list
322 * as blocking locks for @req
324 if (lock->l_blocking_ast)
325 ldlm_add_ast_work_item(lock, req,
327 head = &lock->l_sl_policy;
328 list_for_each_entry(lock, head, l_sl_policy)
329 if (lock->l_blocking_ast)
330 ldlm_add_ast_work_item(lock,
334 if (tmp == mode_tail)
338 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
339 } /* Loop over policy groups within one mode group. */
340 } /* Loop over mode groups within @queue. */
346 * Process a granting attempt for IBITS lock.
347 * Must be called with ns lock held
349 * This function looks for any conflicts for \a lock in the granted or
350 * waiting queues. The lock is granted if no conflicts are found in
353 int ldlm_process_inodebits_lock(struct ldlm_lock *lock, __u64 *ldlm_flags,
354 enum ldlm_process_intention intention,
355 enum ldlm_error *err,
356 struct list_head *work_list)
358 struct ldlm_resource *res = lock->l_resource;
359 struct list_head *grant_work = intention == LDLM_PROCESS_ENQUEUE ?
365 *err = ELDLM_LOCK_ABORTED;
366 LASSERT(!ldlm_is_granted(lock));
367 check_res_locked(res);
369 if (intention == LDLM_PROCESS_RESCAN) {
370 struct list_head *bl_list =
371 *ldlm_flags & LDLM_FL_BLOCK_NOWAIT ? NULL : work_list;
373 LASSERT(lock->l_policy_data.l_inodebits.bits != 0);
375 /* It is possible that some of granted locks was not canceled
376 * but converted and is kept in granted queue. So there is
377 * a window where lock with 'ast_sent' might become granted
378 * again. Meanwhile a new lock may appear in that window and
379 * conflicts with the converted lock so the following scenario
382 * 1) lock1 conflicts with lock2
383 * 2) bl_ast was sent for lock2
384 * 3) lock3 comes and conflicts with lock2 too
385 * 4) no bl_ast sent because lock2->l_bl_ast_sent is 1
386 * 5) lock2 was converted for lock1 but not for lock3
387 * 6) lock1 granted, lock3 still is waiting for lock2, but
388 * there will never be another bl_ast for that
390 * To avoid this scenario the work_list is used below to collect
391 * any blocked locks from granted queue during every reprocess
392 * and bl_ast will be sent if needed.
395 rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock,
396 ldlm_flags, bl_list);
398 RETURN(LDLM_ITER_STOP);
399 rc = ldlm_inodebits_compat_queue(&res->lr_waiting, lock,
402 RETURN(LDLM_ITER_STOP);
404 /* grant also try_bits if any */
405 if (lock->l_policy_data.l_inodebits.try_bits != 0) {
406 lock->l_policy_data.l_inodebits.bits |=
407 lock->l_policy_data.l_inodebits.try_bits;
408 lock->l_policy_data.l_inodebits.try_bits = 0;
409 *ldlm_flags |= LDLM_FL_LOCK_CHANGED;
411 ldlm_resource_unlink_lock(lock);
412 ldlm_grant_lock(lock, grant_work);
415 RETURN(LDLM_ITER_CONTINUE);
418 rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock,
419 ldlm_flags, work_list);
421 GOTO(out, *err = rc);
424 rc2 = ldlm_inodebits_compat_queue(&res->lr_waiting, lock,
425 ldlm_flags, work_list);
427 GOTO(out, *err = rc = rc2);
431 /* if there were only bits to try and all are conflicting */
432 if ((lock->l_policy_data.l_inodebits.bits |
433 lock->l_policy_data.l_inodebits.try_bits)) {
434 /* There is no sense to set LDLM_FL_NO_TIMEOUT to
435 * @ldlm_flags for DOM lock while they are enqueued
436 * through intents, i.e. @lock here is local which does
442 /* grant also all remaining try_bits */
443 if (lock->l_policy_data.l_inodebits.try_bits != 0) {
444 lock->l_policy_data.l_inodebits.bits |=
445 lock->l_policy_data.l_inodebits.try_bits;
446 lock->l_policy_data.l_inodebits.try_bits = 0;
447 *ldlm_flags |= LDLM_FL_LOCK_CHANGED;
449 LASSERT(lock->l_policy_data.l_inodebits.bits);
450 ldlm_resource_unlink_lock(lock);
451 ldlm_grant_lock(lock, grant_work);
455 RETURN(LDLM_ITER_CONTINUE);
459 #endif /* HAVE_SERVER_SUPPORT */
461 void ldlm_ibits_policy_wire_to_local(const union ldlm_wire_policy_data *wpolicy,
462 union ldlm_policy_data *lpolicy)
464 lpolicy->l_inodebits.bits = wpolicy->l_inodebits.bits;
465 lpolicy->l_inodebits.li_initiator_id =
466 wpolicy->l_inodebits.li_initiator_id;
468 * try_bits and li_gid are to be handled outside of generic
469 * write_to_local due to different behavior on a server and client.
473 void ldlm_ibits_policy_local_to_wire(const union ldlm_policy_data *lpolicy,
474 union ldlm_wire_policy_data *wpolicy)
476 memset(wpolicy, 0, sizeof(*wpolicy));
477 wpolicy->l_inodebits.bits = lpolicy->l_inodebits.bits;
478 wpolicy->l_inodebits.try_bits = lpolicy->l_inodebits.try_bits;
479 wpolicy->l_inodebits.li_gid = lpolicy->l_inodebits.li_gid;
480 wpolicy->l_inodebits.li_initiator_id =
481 lpolicy->l_inodebits.li_initiator_id;
485 * Attempt to convert already granted IBITS lock with several bits set to
486 * a lock with less bits (downgrade).
488 * Such lock conversion is used to keep lock with non-blocking bits instead of
489 * cancelling it, introduced for better support of DoM files.
491 int ldlm_inodebits_drop(struct ldlm_lock *lock, __u64 to_drop)
495 check_res_locked(lock->l_resource);
497 /* Just return if there are no conflicting bits */
498 if ((lock->l_policy_data.l_inodebits.bits & to_drop) == 0) {
499 LDLM_WARN(lock, "try to drop unset bits %#llx/%#llx",
500 lock->l_policy_data.l_inodebits.bits, to_drop);
505 /* remove lock from a skiplist and put in the new place
506 * according with new inodebits
508 ldlm_resource_unlink_lock(lock);
509 lock->l_policy_data.l_inodebits.bits &= ~to_drop;
510 ldlm_grant_lock_with_skiplist(lock);
513 EXPORT_SYMBOL(ldlm_inodebits_drop);
515 /* convert single lock */
516 int ldlm_cli_inodebits_convert(struct ldlm_lock *lock,
517 enum ldlm_cancel_flags cancel_flags)
519 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
520 struct ldlm_lock_desc ld = { { 0 } };
521 __u64 drop_bits, new_bits;
527 check_res_locked(lock->l_resource);
529 /* Lock is being converted already */
530 if (ldlm_is_converting(lock)) {
531 if (!(cancel_flags & LCF_ASYNC)) {
532 unlock_res_and_lock(lock);
533 wait_event_idle(lock->l_waitq,
534 is_lock_converted(lock));
535 lock_res_and_lock(lock);
540 /* lru_cancel may happen in parallel and call ldlm_cli_cancel_list()
543 if (ldlm_is_canceling(lock))
546 /* no need in only local convert */
547 if (lock->l_flags & (LDLM_FL_LOCAL_ONLY | LDLM_FL_CANCEL_ON_BLOCK))
550 drop_bits = lock->l_policy_data.l_inodebits.cancel_bits;
551 /* no cancel bits - means that caller needs full cancel */
555 new_bits = lock->l_policy_data.l_inodebits.bits & ~drop_bits;
556 /* check if all lock bits are dropped, proceed with cancel */
560 /* check if no dropped bits, consider this as successful convert */
561 if (lock->l_policy_data.l_inodebits.bits == new_bits)
564 ldlm_set_converting(lock);
565 /* Finally call cancel callback for remaining bits only.
566 * It is important to have converting flag during that
567 * so blocking_ast callback can distinguish convert from
570 ld.l_policy_data.l_inodebits.cancel_bits = drop_bits;
571 unlock_res_and_lock(lock);
572 lock->l_blocking_ast(lock, &ld, lock->l_ast_data, LDLM_CB_CANCELING);
573 /* now notify server about convert */
574 rc = ldlm_cli_convert_req(lock, &flags, new_bits);
575 lock_res_and_lock(lock);
577 GOTO(full_cancel, rc);
580 * check that the lock is still actual as it could get
581 * invalidated by an eviction being unproteced few
584 if (ldlm_is_failed(lock))
585 GOTO(full_cancel, rc = -EINVAL);
587 /* Being locked again check if lock was canceled, it is important
588 * to do and don't drop cbpending below
590 if (ldlm_is_canceling(lock))
591 GOTO(full_cancel, rc = -EINVAL);
593 /* Finally clear these bits in lock ibits */
594 ldlm_inodebits_drop(lock, drop_bits);
596 /* also check again if more bits to be cancelled appeared */
597 if (drop_bits != lock->l_policy_data.l_inodebits.cancel_bits)
598 GOTO(clear_converting, rc = -EAGAIN);
600 /* clear cbpending flag early, it is safe to match lock right after
601 * client convert because it is downgrade always.
603 ldlm_clear_cbpending(lock);
604 ldlm_clear_bl_ast(lock);
605 spin_lock(&ns->ns_lock);
606 if (list_empty(&lock->l_lru))
607 ldlm_lock_add_to_lru_nolock(lock);
608 spin_unlock(&ns->ns_lock);
610 /* the job is done, zero the cancel_bits. If more conflicts appear,
611 * it will result in another cycle of ldlm_cli_inodebits_convert().
614 lock->l_policy_data.l_inodebits.cancel_bits = 0;
616 ldlm_clear_converting(lock);
620 int ldlm_inodebits_alloc_lock(struct ldlm_lock *lock)
622 if (ldlm_is_ns_srv(lock)) {
625 OBD_SLAB_ALLOC_PTR(lock->l_ibits_node, ldlm_inodebits_slab);
626 if (lock->l_ibits_node == NULL)
628 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
629 INIT_LIST_HEAD(&lock->l_ibits_node->lin_link[i]);
630 lock->l_ibits_node->lock = lock;
632 lock->l_ibits_node = NULL;
637 void ldlm_inodebits_add_lock(struct ldlm_resource *res, struct list_head *head,
638 struct ldlm_lock *lock, bool tail)
642 if (!ldlm_is_ns_srv(lock))
645 if (head == &res->lr_waiting) {
646 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++) {
647 if (!(lock->l_policy_data.l_inodebits.bits & BIT(i)))
650 list_add_tail(&lock->l_ibits_node->lin_link[i],
651 &res->lr_ibits_queues->liq_waiting[i]);
653 list_add(&lock->l_ibits_node->lin_link[i],
654 &res->lr_ibits_queues->liq_waiting[i]);
656 } else if (head == &res->lr_granted && lock->l_ibits_node != NULL) {
657 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
658 LASSERT(list_empty(&lock->l_ibits_node->lin_link[i]));
659 OBD_SLAB_FREE_PTR(lock->l_ibits_node, ldlm_inodebits_slab);
660 lock->l_ibits_node = NULL;
661 } else if (head != &res->lr_granted) {
662 /* we are inserting in a middle of a list, after @head */
663 struct ldlm_lock *orig = list_entry(head, struct ldlm_lock,
665 LASSERT(orig->l_policy_data.l_inodebits.bits ==
666 lock->l_policy_data.l_inodebits.bits);
667 /* should not insert before with exactly matched set of bits */
668 LASSERT(tail == false);
670 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++) {
671 if (!(lock->l_policy_data.l_inodebits.bits & (1 << i)))
673 list_add(&lock->l_ibits_node->lin_link[i],
674 &orig->l_ibits_node->lin_link[i]);
679 void ldlm_inodebits_unlink_lock(struct ldlm_lock *lock)
683 ldlm_unlink_lock_skiplist(lock);
684 if (!ldlm_is_ns_srv(lock))
687 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
688 list_del_init(&lock->l_ibits_node->lin_link[i]);