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 /* Local combined DOM lock came across
294 * GROUP DOM lock, it makes the thread
295 * to be blocked for a long time, not
296 * allowed, the trybits to be used
299 if (!req->l_export &&
300 (req_bits & MDS_INODELOCK_DOM) &&
301 (req_bits & ~MDS_INODELOCK_DOM))
307 /* Found a conflicting policy group. */
311 /* Add locks of the policy group to @work_list
312 * as blocking locks for @req */
313 if (lock->l_blocking_ast)
314 ldlm_add_ast_work_item(lock, req,
316 head = &lock->l_sl_policy;
317 list_for_each_entry(lock, head, l_sl_policy)
318 if (lock->l_blocking_ast)
319 ldlm_add_ast_work_item(lock,
323 if (tmp == mode_tail)
327 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
328 } /* Loop over policy groups within one mode group. */
329 } /* Loop over mode groups within @queue. */
335 * Process a granting attempt for IBITS lock.
336 * Must be called with ns lock held
338 * This function looks for any conflicts for \a lock in the granted or
339 * waiting queues. The lock is granted if no conflicts are found in
342 int ldlm_process_inodebits_lock(struct ldlm_lock *lock, __u64 *ldlm_flags,
343 enum ldlm_process_intention intention,
344 enum ldlm_error *err,
345 struct list_head *work_list)
347 struct ldlm_resource *res = lock->l_resource;
348 struct list_head *grant_work = intention == LDLM_PROCESS_ENQUEUE ?
353 *err = ELDLM_LOCK_ABORTED;
354 LASSERT(!ldlm_is_granted(lock));
355 check_res_locked(res);
357 if (intention == LDLM_PROCESS_RESCAN) {
358 struct list_head *bl_list =
359 *ldlm_flags & LDLM_FL_BLOCK_NOWAIT ? NULL : work_list;
361 LASSERT(lock->l_policy_data.l_inodebits.bits != 0);
363 /* It is possible that some of granted locks was not canceled
364 * but converted and is kept in granted queue. So there is
365 * a window where lock with 'ast_sent' might become granted
366 * again. Meanwhile a new lock may appear in that window and
367 * conflicts with the converted lock so the following scenario
370 * 1) lock1 conflicts with lock2
371 * 2) bl_ast was sent for lock2
372 * 3) lock3 comes and conflicts with lock2 too
373 * 4) no bl_ast sent because lock2->l_bl_ast_sent is 1
374 * 5) lock2 was converted for lock1 but not for lock3
375 * 6) lock1 granted, lock3 still is waiting for lock2, but
376 * there will never be another bl_ast for that
378 * To avoid this scenario the work_list is used below to collect
379 * any blocked locks from granted queue during every reprocess
380 * and bl_ast will be sent if needed.
383 rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock,
384 ldlm_flags, bl_list);
386 RETURN(LDLM_ITER_STOP);
387 rc = ldlm_inodebits_compat_queue(&res->lr_waiting, lock,
390 RETURN(LDLM_ITER_STOP);
392 /* grant also try_bits if any */
393 if (lock->l_policy_data.l_inodebits.try_bits != 0) {
394 lock->l_policy_data.l_inodebits.bits |=
395 lock->l_policy_data.l_inodebits.try_bits;
396 lock->l_policy_data.l_inodebits.try_bits = 0;
397 *ldlm_flags |= LDLM_FL_LOCK_CHANGED;
399 ldlm_resource_unlink_lock(lock);
400 ldlm_grant_lock(lock, grant_work);
403 RETURN(LDLM_ITER_CONTINUE);
406 rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock,
407 ldlm_flags, work_list);
409 GOTO(out, *err = rc);
412 rc2 = ldlm_inodebits_compat_queue(&res->lr_waiting, lock,
413 ldlm_flags, work_list);
415 GOTO(out, *err = rc = rc2);
419 /* if there were only bits to try and all are conflicting */
420 if ((lock->l_policy_data.l_inodebits.bits |
421 lock->l_policy_data.l_inodebits.try_bits)) {
422 /* There is no sense to set LDLM_FL_NO_TIMEOUT to
423 * @ldlm_flags for DOM lock while they are enqueued
424 * through intents, i.e. @lock here is local which does
429 /* grant also all remaining try_bits */
430 if (lock->l_policy_data.l_inodebits.try_bits != 0) {
431 lock->l_policy_data.l_inodebits.bits |=
432 lock->l_policy_data.l_inodebits.try_bits;
433 lock->l_policy_data.l_inodebits.try_bits = 0;
434 *ldlm_flags |= LDLM_FL_LOCK_CHANGED;
436 LASSERT(lock->l_policy_data.l_inodebits.bits);
437 ldlm_resource_unlink_lock(lock);
438 ldlm_grant_lock(lock, grant_work);
442 RETURN(LDLM_ITER_CONTINUE);
446 #endif /* HAVE_SERVER_SUPPORT */
448 void ldlm_ibits_policy_wire_to_local(const union ldlm_wire_policy_data *wpolicy,
449 union ldlm_policy_data *lpolicy)
451 lpolicy->l_inodebits.bits = wpolicy->l_inodebits.bits;
453 * try_bits and li_gid are to be handled outside of generic
454 * write_to_local due to different behavior on a server and client.
458 void ldlm_ibits_policy_local_to_wire(const union ldlm_policy_data *lpolicy,
459 union ldlm_wire_policy_data *wpolicy)
461 memset(wpolicy, 0, sizeof(*wpolicy));
462 wpolicy->l_inodebits.bits = lpolicy->l_inodebits.bits;
463 wpolicy->l_inodebits.try_bits = lpolicy->l_inodebits.try_bits;
464 wpolicy->l_inodebits.li_gid = lpolicy->l_inodebits.li_gid;
468 * Attempt to convert already granted IBITS lock with several bits set to
469 * a lock with less bits (downgrade).
471 * Such lock conversion is used to keep lock with non-blocking bits instead of
472 * cancelling it, introduced for better support of DoM files.
474 int ldlm_inodebits_drop(struct ldlm_lock *lock, __u64 to_drop)
478 check_res_locked(lock->l_resource);
480 /* Just return if there are no conflicting bits */
481 if ((lock->l_policy_data.l_inodebits.bits & to_drop) == 0) {
482 LDLM_WARN(lock, "try to drop unset bits %#llx/%#llx",
483 lock->l_policy_data.l_inodebits.bits, to_drop);
488 /* remove lock from a skiplist and put in the new place
489 * according with new inodebits */
490 ldlm_resource_unlink_lock(lock);
491 lock->l_policy_data.l_inodebits.bits &= ~to_drop;
492 ldlm_grant_lock_with_skiplist(lock);
495 EXPORT_SYMBOL(ldlm_inodebits_drop);
497 /* convert single lock */
498 int ldlm_cli_inodebits_convert(struct ldlm_lock *lock,
499 enum ldlm_cancel_flags cancel_flags)
501 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
502 struct ldlm_lock_desc ld = { { 0 } };
503 __u64 drop_bits, new_bits;
509 check_res_locked(lock->l_resource);
511 /* Lock is being converted already */
512 if (ldlm_is_converting(lock)) {
513 if (!(cancel_flags & LCF_ASYNC)) {
514 unlock_res_and_lock(lock);
515 wait_event_idle(lock->l_waitq,
516 is_lock_converted(lock));
517 lock_res_and_lock(lock);
522 /* lru_cancel may happen in parallel and call ldlm_cli_cancel_list()
525 if (ldlm_is_canceling(lock))
528 /* no need in only local convert */
529 if (lock->l_flags & (LDLM_FL_LOCAL_ONLY | LDLM_FL_CANCEL_ON_BLOCK))
532 drop_bits = lock->l_policy_data.l_inodebits.cancel_bits;
533 /* no cancel bits - means that caller needs full cancel */
537 new_bits = lock->l_policy_data.l_inodebits.bits & ~drop_bits;
538 /* check if all lock bits are dropped, proceed with cancel */
542 /* check if no dropped bits, consider this as successful convert */
543 if (lock->l_policy_data.l_inodebits.bits == new_bits)
546 ldlm_set_converting(lock);
547 /* Finally call cancel callback for remaining bits only.
548 * It is important to have converting flag during that
549 * so blocking_ast callback can distinguish convert from
552 ld.l_policy_data.l_inodebits.cancel_bits = drop_bits;
553 unlock_res_and_lock(lock);
554 lock->l_blocking_ast(lock, &ld, lock->l_ast_data, LDLM_CB_CANCELING);
555 /* now notify server about convert */
556 rc = ldlm_cli_convert_req(lock, &flags, new_bits);
557 lock_res_and_lock(lock);
559 GOTO(full_cancel, rc);
561 /* Finally clear these bits in lock ibits */
562 ldlm_inodebits_drop(lock, drop_bits);
564 /* Being locked again check if lock was canceled, it is important
565 * to do and don't drop cbpending below
567 if (ldlm_is_canceling(lock))
568 GOTO(full_cancel, rc = -EINVAL);
570 /* also check again if more bits to be cancelled appeared */
571 if (drop_bits != lock->l_policy_data.l_inodebits.cancel_bits)
572 GOTO(clear_converting, rc = -EAGAIN);
574 /* clear cbpending flag early, it is safe to match lock right after
575 * client convert because it is downgrade always.
577 ldlm_clear_cbpending(lock);
578 ldlm_clear_bl_ast(lock);
579 spin_lock(&ns->ns_lock);
580 if (list_empty(&lock->l_lru))
581 ldlm_lock_add_to_lru_nolock(lock);
582 spin_unlock(&ns->ns_lock);
584 /* the job is done, zero the cancel_bits. If more conflicts appear,
585 * it will result in another cycle of ldlm_cli_inodebits_convert().
588 lock->l_policy_data.l_inodebits.cancel_bits = 0;
590 ldlm_clear_converting(lock);
594 int ldlm_inodebits_alloc_lock(struct ldlm_lock *lock)
596 if (ldlm_is_ns_srv(lock)) {
599 OBD_SLAB_ALLOC_PTR(lock->l_ibits_node, ldlm_inodebits_slab);
600 if (lock->l_ibits_node == NULL)
602 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
603 INIT_LIST_HEAD(&lock->l_ibits_node->lin_link[i]);
604 lock->l_ibits_node->lock = lock;
606 lock->l_ibits_node = NULL;
611 void ldlm_inodebits_add_lock(struct ldlm_resource *res, struct list_head *head,
612 struct ldlm_lock *lock, bool tail)
616 if (!ldlm_is_ns_srv(lock))
619 if (head == &res->lr_waiting) {
620 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++) {
621 if (!(lock->l_policy_data.l_inodebits.bits & BIT(i)))
624 list_add_tail(&lock->l_ibits_node->lin_link[i],
625 &res->lr_ibits_queues->liq_waiting[i]);
627 list_add(&lock->l_ibits_node->lin_link[i],
628 &res->lr_ibits_queues->liq_waiting[i]);
630 } else if (head == &res->lr_granted && lock->l_ibits_node != NULL) {
631 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
632 LASSERT(list_empty(&lock->l_ibits_node->lin_link[i]));
633 OBD_SLAB_FREE_PTR(lock->l_ibits_node, ldlm_inodebits_slab);
634 lock->l_ibits_node = NULL;
635 } else if (head != &res->lr_granted) {
636 /* we are inserting in a middle of a list, after @head */
637 struct ldlm_lock *orig = list_entry(head, struct ldlm_lock,
639 LASSERT(orig->l_policy_data.l_inodebits.bits ==
640 lock->l_policy_data.l_inodebits.bits);
641 /* The is no a use case to insert before with exactly matched
643 LASSERT(tail == false);
645 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++) {
646 if (!(lock->l_policy_data.l_inodebits.bits & (1 << i)))
648 list_add(&lock->l_ibits_node->lin_link[i],
649 &orig->l_ibits_node->lin_link[i]);
654 void ldlm_inodebits_unlink_lock(struct ldlm_lock *lock)
658 ldlm_unlink_lock_skiplist(lock);
659 if (!ldlm_is_ns_srv(lock))
662 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
663 list_del_init(&lock->l_ibits_node->lin_link[i]);