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 __u64 *ldlm_flags, struct list_head *work_list)
162 enum ldlm_mode req_mode = req->l_req_mode;
163 struct list_head *tmp;
164 struct ldlm_lock *lock;
165 __u64 req_bits = req->l_policy_data.l_inodebits.bits;
166 __u64 *try_bits = &req->l_policy_data.l_inodebits.try_bits;
171 lockmode_verify(req_mode);
173 /* There is no sense in lock with no bits set. Also such a lock
174 * would be compatible with any other bit lock.
175 * Meanwhile that can be true if there were just try_bits and all
176 * are failed, so just exit gracefully and let the caller to care.
178 if ((req_bits | *try_bits) == 0)
181 /* Group lock could be only DOM */
182 if (unlikely(req_mode == LCK_GROUP &&
183 (req_bits | *try_bits) != MDS_INODELOCK_DOM))
186 list_for_each(tmp, queue) {
187 struct list_head *mode_tail;
189 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
191 /* We stop walking the queue if we hit ourselves so we don't
192 * take conflicting locks enqueued after us into account,
193 * or we'd wait forever. */
197 /* last lock in mode group */
198 LASSERT(lock->l_sl_mode.prev != NULL);
199 mode_tail = &list_entry(lock->l_sl_mode.prev, struct ldlm_lock,
200 l_sl_mode)->l_res_link;
202 /* if request lock is not COS_INCOMPAT and COS is disabled,
203 * they are compatible, IOW this request is from a local
204 * transaction on a DNE system. */
205 if (lock->l_req_mode == LCK_COS && !ldlm_is_cos_incompat(req) &&
206 !ldlm_is_cos_enabled(req)) {
207 /* jump to last lock in mode group */
212 if (lockmode_compat(lock->l_req_mode, req_mode)) {
213 /* non group locks are compatible, bits don't matter */
214 if (likely(req_mode != LCK_GROUP)) {
215 /* jump to last lock in mode group */
220 if (req->l_policy_data.l_inodebits.li_gid ==
221 lock->l_policy_data.l_inodebits.li_gid) {
222 if (ldlm_is_granted(lock))
225 if (*ldlm_flags & LDLM_FL_BLOCK_NOWAIT)
226 RETURN(-EWOULDBLOCK);
228 /* Place the same group together */
229 ldlm_resource_insert_lock_after(lock, req);
234 /* GROUP locks are placed to a head of the waiting list, but
236 if (unlikely(req_mode == LCK_GROUP && !ldlm_is_granted(lock))) {
238 if (lock->l_req_mode != LCK_GROUP) {
239 /* Already not a GROUP lock, insert before. */
240 ldlm_resource_insert_lock_before(lock, req);
243 /* Still GROUP but a different gid(the same gid would
244 * be handled above). Keep searching for the same gid */
245 LASSERT(req->l_policy_data.l_inodebits.li_gid !=
246 lock->l_policy_data.l_inodebits.li_gid);
251 struct list_head *head;
253 /* Advance loop cursor to last lock in policy group. */
254 tmp = &list_entry(lock->l_sl_policy.prev,
256 l_sl_policy)->l_res_link;
258 /* New lock's try_bits are filtered out by ibits
259 * of all locks in both granted and waiting queues.
261 *try_bits &= ~(lock->l_policy_data.l_inodebits.bits |
262 lock->l_policy_data.l_inodebits.try_bits);
264 if ((req_bits | *try_bits) == 0)
267 /* The new lock ibits is more preferable than try_bits
268 * of waiting locks so drop conflicting try_bits in
270 * Notice that try_bits of granted locks must be zero.
272 lock->l_policy_data.l_inodebits.try_bits &= ~req_bits;
274 /* Locks with overlapping bits conflict. */
275 if (lock->l_policy_data.l_inodebits.bits & req_bits) {
276 /* COS lock mode has a special compatibility
277 * requirement: it is only compatible with
278 * locks from the same client. */
279 if (lock->l_req_mode == LCK_COS &&
280 !ldlm_is_cos_incompat(req) &&
281 ldlm_is_cos_enabled(req) &&
282 lock->l_client_cookie == req->l_client_cookie)
287 if (unlikely(lock->l_req_mode == LCK_GROUP)) {
288 LASSERT(ldlm_has_dom(lock));
290 if (*ldlm_flags & LDLM_FL_BLOCK_NOWAIT)
291 RETURN(-EWOULDBLOCK);
293 /* Combined DOM lock came across GROUP
294 * DOM lock, it makes the thread to be
295 * blocked for a long time, not allowed,
296 * the trybits to be used instead.
297 * Not combined DOM lock is requested by
298 * client, and have to wait for long
299 * until re-worked to a non-intent
301 if ((req_bits & MDS_INODELOCK_DOM) &&
302 (req_bits & ~MDS_INODELOCK_DOM))
308 /* Found a conflicting policy group. */
312 /* Add locks of the policy group to @work_list
313 * as blocking locks for @req */
314 if (lock->l_blocking_ast)
315 ldlm_add_ast_work_item(lock, req,
317 head = &lock->l_sl_policy;
318 list_for_each_entry(lock, head, l_sl_policy)
319 if (lock->l_blocking_ast)
320 ldlm_add_ast_work_item(lock,
324 if (tmp == mode_tail)
328 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
329 } /* Loop over policy groups within one mode group. */
330 } /* Loop over mode groups within @queue. */
336 * Process a granting attempt for IBITS lock.
337 * Must be called with ns lock held
339 * This function looks for any conflicts for \a lock in the granted or
340 * waiting queues. The lock is granted if no conflicts are found in
343 int ldlm_process_inodebits_lock(struct ldlm_lock *lock, __u64 *ldlm_flags,
344 enum ldlm_process_intention intention,
345 enum ldlm_error *err,
346 struct list_head *work_list)
348 struct ldlm_resource *res = lock->l_resource;
349 struct list_head *grant_work = intention == LDLM_PROCESS_ENQUEUE ?
354 *err = ELDLM_LOCK_ABORTED;
355 LASSERT(!ldlm_is_granted(lock));
356 check_res_locked(res);
358 if (intention == LDLM_PROCESS_RESCAN) {
359 struct list_head *bl_list =
360 *ldlm_flags & LDLM_FL_BLOCK_NOWAIT ? NULL : work_list;
362 LASSERT(lock->l_policy_data.l_inodebits.bits != 0);
364 /* It is possible that some of granted locks was not canceled
365 * but converted and is kept in granted queue. So there is
366 * a window where lock with 'ast_sent' might become granted
367 * again. Meanwhile a new lock may appear in that window and
368 * conflicts with the converted lock so the following scenario
371 * 1) lock1 conflicts with lock2
372 * 2) bl_ast was sent for lock2
373 * 3) lock3 comes and conflicts with lock2 too
374 * 4) no bl_ast sent because lock2->l_bl_ast_sent is 1
375 * 5) lock2 was converted for lock1 but not for lock3
376 * 6) lock1 granted, lock3 still is waiting for lock2, but
377 * there will never be another bl_ast for that
379 * To avoid this scenario the work_list is used below to collect
380 * any blocked locks from granted queue during every reprocess
381 * and bl_ast will be sent if needed.
384 rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock,
385 ldlm_flags, bl_list);
387 RETURN(LDLM_ITER_STOP);
388 rc = ldlm_inodebits_compat_queue(&res->lr_waiting, lock,
391 RETURN(LDLM_ITER_STOP);
393 /* grant also try_bits if any */
394 if (lock->l_policy_data.l_inodebits.try_bits != 0) {
395 lock->l_policy_data.l_inodebits.bits |=
396 lock->l_policy_data.l_inodebits.try_bits;
397 lock->l_policy_data.l_inodebits.try_bits = 0;
398 *ldlm_flags |= LDLM_FL_LOCK_CHANGED;
400 ldlm_resource_unlink_lock(lock);
401 ldlm_grant_lock(lock, grant_work);
404 RETURN(LDLM_ITER_CONTINUE);
407 rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock,
408 ldlm_flags, work_list);
410 GOTO(out, *err = rc);
413 rc2 = ldlm_inodebits_compat_queue(&res->lr_waiting, lock,
414 ldlm_flags, work_list);
416 GOTO(out, *err = rc = rc2);
420 /* if there were only bits to try and all are conflicting */
421 if ((lock->l_policy_data.l_inodebits.bits |
422 lock->l_policy_data.l_inodebits.try_bits)) {
423 /* There is no sense to set LDLM_FL_NO_TIMEOUT to
424 * @ldlm_flags for DOM lock while they are enqueued
425 * through intents, i.e. @lock here is local which does
430 /* grant also all remaining try_bits */
431 if (lock->l_policy_data.l_inodebits.try_bits != 0) {
432 lock->l_policy_data.l_inodebits.bits |=
433 lock->l_policy_data.l_inodebits.try_bits;
434 lock->l_policy_data.l_inodebits.try_bits = 0;
435 *ldlm_flags |= LDLM_FL_LOCK_CHANGED;
437 LASSERT(lock->l_policy_data.l_inodebits.bits);
438 ldlm_resource_unlink_lock(lock);
439 ldlm_grant_lock(lock, grant_work);
443 RETURN(LDLM_ITER_CONTINUE);
447 #endif /* HAVE_SERVER_SUPPORT */
449 void ldlm_ibits_policy_wire_to_local(const union ldlm_wire_policy_data *wpolicy,
450 union ldlm_policy_data *lpolicy)
452 lpolicy->l_inodebits.bits = wpolicy->l_inodebits.bits;
453 lpolicy->l_inodebits.li_gid = wpolicy->l_inodebits.li_gid;
455 * try_bits are to be handled outside of generic write_to_local due
456 * to different behavior on a server and client.
460 void ldlm_ibits_policy_local_to_wire(const union ldlm_policy_data *lpolicy,
461 union ldlm_wire_policy_data *wpolicy)
463 memset(wpolicy, 0, sizeof(*wpolicy));
464 wpolicy->l_inodebits.bits = lpolicy->l_inodebits.bits;
465 wpolicy->l_inodebits.try_bits = lpolicy->l_inodebits.try_bits;
466 wpolicy->l_inodebits.li_gid = lpolicy->l_inodebits.li_gid;
470 * Attempt to convert already granted IBITS lock with several bits set to
471 * a lock with less bits (downgrade).
473 * Such lock conversion is used to keep lock with non-blocking bits instead of
474 * cancelling it, introduced for better support of DoM files.
476 int ldlm_inodebits_drop(struct ldlm_lock *lock, __u64 to_drop)
480 check_res_locked(lock->l_resource);
482 /* Just return if there are no conflicting bits */
483 if ((lock->l_policy_data.l_inodebits.bits & to_drop) == 0) {
484 LDLM_WARN(lock, "try to drop unset bits %#llx/%#llx",
485 lock->l_policy_data.l_inodebits.bits, to_drop);
490 /* remove lock from a skiplist and put in the new place
491 * according with new inodebits */
492 ldlm_resource_unlink_lock(lock);
493 lock->l_policy_data.l_inodebits.bits &= ~to_drop;
494 ldlm_grant_lock_with_skiplist(lock);
497 EXPORT_SYMBOL(ldlm_inodebits_drop);
499 /* convert single lock */
500 int ldlm_cli_inodebits_convert(struct ldlm_lock *lock,
501 enum ldlm_cancel_flags cancel_flags)
503 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
504 struct ldlm_lock_desc ld = { { 0 } };
505 __u64 drop_bits, new_bits;
511 check_res_locked(lock->l_resource);
513 /* Lock is being converted already */
514 if (ldlm_is_converting(lock)) {
515 if (!(cancel_flags & LCF_ASYNC)) {
516 unlock_res_and_lock(lock);
517 wait_event_idle(lock->l_waitq,
518 is_lock_converted(lock));
519 lock_res_and_lock(lock);
524 /* lru_cancel may happen in parallel and call ldlm_cli_cancel_list()
527 if (ldlm_is_canceling(lock))
530 /* no need in only local convert */
531 if (lock->l_flags & (LDLM_FL_LOCAL_ONLY | LDLM_FL_CANCEL_ON_BLOCK))
534 drop_bits = lock->l_policy_data.l_inodebits.cancel_bits;
535 /* no cancel bits - means that caller needs full cancel */
539 new_bits = lock->l_policy_data.l_inodebits.bits & ~drop_bits;
540 /* check if all lock bits are dropped, proceed with cancel */
544 /* check if no dropped bits, consider this as successful convert */
545 if (lock->l_policy_data.l_inodebits.bits == new_bits)
548 ldlm_set_converting(lock);
549 /* Finally call cancel callback for remaining bits only.
550 * It is important to have converting flag during that
551 * so blocking_ast callback can distinguish convert from
554 ld.l_policy_data.l_inodebits.cancel_bits = drop_bits;
555 unlock_res_and_lock(lock);
556 lock->l_blocking_ast(lock, &ld, lock->l_ast_data, LDLM_CB_CANCELING);
557 /* now notify server about convert */
558 rc = ldlm_cli_convert_req(lock, &flags, new_bits);
559 lock_res_and_lock(lock);
561 GOTO(full_cancel, rc);
563 /* Finally clear these bits in lock ibits */
564 ldlm_inodebits_drop(lock, drop_bits);
566 /* Being locked again check if lock was canceled, it is important
567 * to do and don't drop cbpending below
569 if (ldlm_is_canceling(lock))
570 GOTO(full_cancel, rc = -EINVAL);
572 /* also check again if more bits to be cancelled appeared */
573 if (drop_bits != lock->l_policy_data.l_inodebits.cancel_bits)
574 GOTO(clear_converting, rc = -EAGAIN);
576 /* clear cbpending flag early, it is safe to match lock right after
577 * client convert because it is downgrade always.
579 ldlm_clear_cbpending(lock);
580 ldlm_clear_bl_ast(lock);
581 spin_lock(&ns->ns_lock);
582 if (list_empty(&lock->l_lru))
583 ldlm_lock_add_to_lru_nolock(lock);
584 spin_unlock(&ns->ns_lock);
586 /* the job is done, zero the cancel_bits. If more conflicts appear,
587 * it will result in another cycle of ldlm_cli_inodebits_convert().
590 lock->l_policy_data.l_inodebits.cancel_bits = 0;
592 ldlm_clear_converting(lock);
596 int ldlm_inodebits_alloc_lock(struct ldlm_lock *lock)
598 if (ldlm_is_ns_srv(lock)) {
601 OBD_SLAB_ALLOC_PTR(lock->l_ibits_node, ldlm_inodebits_slab);
602 if (lock->l_ibits_node == NULL)
604 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
605 INIT_LIST_HEAD(&lock->l_ibits_node->lin_link[i]);
606 lock->l_ibits_node->lock = lock;
608 lock->l_ibits_node = NULL;
613 void ldlm_inodebits_add_lock(struct ldlm_resource *res, struct list_head *head,
614 struct ldlm_lock *lock, bool tail)
618 if (!ldlm_is_ns_srv(lock))
621 if (head == &res->lr_waiting) {
622 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++) {
623 if (!(lock->l_policy_data.l_inodebits.bits & BIT(i)))
626 list_add_tail(&lock->l_ibits_node->lin_link[i],
627 &res->lr_ibits_queues->liq_waiting[i]);
629 list_add(&lock->l_ibits_node->lin_link[i],
630 &res->lr_ibits_queues->liq_waiting[i]);
632 } else if (head == &res->lr_granted && lock->l_ibits_node != NULL) {
633 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
634 LASSERT(list_empty(&lock->l_ibits_node->lin_link[i]));
635 OBD_SLAB_FREE_PTR(lock->l_ibits_node, ldlm_inodebits_slab);
636 lock->l_ibits_node = NULL;
637 } else if (head != &res->lr_granted) {
638 /* we are inserting in a middle of a list, after @head */
639 struct ldlm_lock *orig = list_entry(head, struct ldlm_lock,
641 LASSERT(orig->l_policy_data.l_inodebits.bits ==
642 lock->l_policy_data.l_inodebits.bits);
643 /* The is no a use case to insert before with exactly matched
645 LASSERT(tail == false);
647 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++) {
648 if (!(lock->l_policy_data.l_inodebits.bits & (1 << i)))
650 list_add(&lock->l_ibits_node->lin_link[i],
651 &orig->l_ibits_node->lin_link[i]);
656 void ldlm_inodebits_unlink_lock(struct ldlm_lock *lock)
660 ldlm_unlink_lock_skiplist(lock);
661 if (!ldlm_is_ns_srv(lock))
664 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
665 list_del_init(&lock->l_ibits_node->lin_link[i]);