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 struct list_head bl_ast_list = LIST_HEAD_INIT(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++) {
98 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
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 & (1 << 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 ?
295 LASSERT(!ldlm_is_granted(lock));
296 check_res_locked(res);
298 if (intention == LDLM_PROCESS_RESCAN) {
299 struct list_head *bl_list;
301 if (*flags & LDLM_FL_BLOCK_NOWAIT) {
303 *err = ELDLM_LOCK_WOULDBLOCK;
306 *err = ELDLM_LOCK_ABORTED;
309 LASSERT(lock->l_policy_data.l_inodebits.bits != 0);
311 /* It is possible that some of granted locks was not canceled
312 * but converted and is kept in granted queue. So there is
313 * a window where lock with 'ast_sent' might become granted
314 * again. Meanwhile a new lock may appear in that window and
315 * conflicts with the converted lock so the following scenario
318 * 1) lock1 conflicts with lock2
319 * 2) bl_ast was sent for lock2
320 * 3) lock3 comes and conflicts with lock2 too
321 * 4) no bl_ast sent because lock2->l_bl_ast_sent is 1
322 * 5) lock2 was converted for lock1 but not for lock3
323 * 6) lock1 granted, lock3 still is waiting for lock2, but
324 * there will never be another bl_ast for that
326 * To avoid this scenario the work_list is used below to collect
327 * any blocked locks from granted queue during every reprocess
328 * and bl_ast will be sent if needed.
330 rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock,
333 RETURN(LDLM_ITER_STOP);
334 rc = ldlm_inodebits_compat_queue(&res->lr_waiting, lock, NULL);
336 RETURN(LDLM_ITER_STOP);
338 /* grant also try_bits if any */
339 if (lock->l_policy_data.l_inodebits.try_bits != 0) {
340 lock->l_policy_data.l_inodebits.bits |=
341 lock->l_policy_data.l_inodebits.try_bits;
342 lock->l_policy_data.l_inodebits.try_bits = 0;
343 *flags |= LDLM_FL_LOCK_CHANGED;
345 ldlm_resource_unlink_lock(lock);
346 ldlm_grant_lock(lock, grant_work);
349 RETURN(LDLM_ITER_CONTINUE);
352 rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock, work_list);
353 rc += ldlm_inodebits_compat_queue(&res->lr_waiting, lock, work_list);
356 /* if there were only bits to try and all are conflicting */
357 if ((lock->l_policy_data.l_inodebits.bits |
358 lock->l_policy_data.l_inodebits.try_bits) == 0) {
359 *err = ELDLM_LOCK_WOULDBLOCK;
364 /* grant also all remaining try_bits */
365 if (lock->l_policy_data.l_inodebits.try_bits != 0) {
366 lock->l_policy_data.l_inodebits.bits |=
367 lock->l_policy_data.l_inodebits.try_bits;
368 lock->l_policy_data.l_inodebits.try_bits = 0;
369 *flags |= LDLM_FL_LOCK_CHANGED;
371 LASSERT(lock->l_policy_data.l_inodebits.bits);
372 ldlm_resource_unlink_lock(lock);
373 ldlm_grant_lock(lock, grant_work);
377 RETURN(LDLM_ITER_CONTINUE);
379 #endif /* HAVE_SERVER_SUPPORT */
381 void ldlm_ibits_policy_wire_to_local(const union ldlm_wire_policy_data *wpolicy,
382 union ldlm_policy_data *lpolicy)
384 lpolicy->l_inodebits.bits = wpolicy->l_inodebits.bits;
385 lpolicy->l_inodebits.try_bits = wpolicy->l_inodebits.try_bits;
388 void ldlm_ibits_policy_local_to_wire(const union ldlm_policy_data *lpolicy,
389 union ldlm_wire_policy_data *wpolicy)
391 memset(wpolicy, 0, sizeof(*wpolicy));
392 wpolicy->l_inodebits.bits = lpolicy->l_inodebits.bits;
393 wpolicy->l_inodebits.try_bits = lpolicy->l_inodebits.try_bits;
397 * Attempt to convert already granted IBITS lock with several bits set to
398 * a lock with less bits (downgrade).
400 * Such lock conversion is used to keep lock with non-blocking bits instead of
401 * cancelling it, introduced for better support of DoM files.
403 int ldlm_inodebits_drop(struct ldlm_lock *lock, __u64 to_drop)
407 check_res_locked(lock->l_resource);
409 /* Just return if there are no conflicting bits */
410 if ((lock->l_policy_data.l_inodebits.bits & to_drop) == 0) {
411 LDLM_WARN(lock, "try to drop unset bits %#llx/%#llx",
412 lock->l_policy_data.l_inodebits.bits, to_drop);
417 /* remove lock from a skiplist and put in the new place
418 * according with new inodebits */
419 ldlm_resource_unlink_lock(lock);
420 lock->l_policy_data.l_inodebits.bits &= ~to_drop;
421 ldlm_grant_lock_with_skiplist(lock);
424 EXPORT_SYMBOL(ldlm_inodebits_drop);
426 /* convert single lock */
427 int ldlm_cli_dropbits(struct ldlm_lock *lock, __u64 drop_bits)
429 struct lustre_handle lockh;
436 LASSERT(!lock->l_readers && !lock->l_writers);
438 LDLM_DEBUG(lock, "client lock convert START");
440 ldlm_lock2handle(lock, &lockh);
441 lock_res_and_lock(lock);
442 /* check if all bits are blocked */
443 if (!(lock->l_policy_data.l_inodebits.bits & ~drop_bits)) {
444 unlock_res_and_lock(lock);
445 /* return error to continue with cancel */
446 GOTO(exit, rc = -EINVAL);
449 /* check if no common bits, consider this as successful convert */
450 if (!(lock->l_policy_data.l_inodebits.bits & drop_bits)) {
451 unlock_res_and_lock(lock);
455 /* check if there is race with cancel */
456 if (ldlm_is_canceling(lock) || ldlm_is_cancel(lock)) {
457 unlock_res_and_lock(lock);
458 GOTO(exit, rc = -EINVAL);
461 /* clear cbpending flag early, it is safe to match lock right after
462 * client convert because it is downgrade always.
464 ldlm_clear_cbpending(lock);
465 ldlm_clear_bl_ast(lock);
467 /* If lock is being converted already, check drop bits first */
468 if (ldlm_is_converting(lock)) {
469 /* raced lock convert, lock inodebits are remaining bits
470 * so check if they are conflicting with new convert or not.
472 if (!(lock->l_policy_data.l_inodebits.bits & drop_bits)) {
473 unlock_res_and_lock(lock);
476 /* Otherwise drop new conflicting bits in new convert */
478 ldlm_set_converting(lock);
479 /* from all bits of blocking lock leave only conflicting */
480 drop_bits &= lock->l_policy_data.l_inodebits.bits;
481 /* save them in cancel_bits, so l_blocking_ast will know
482 * which bits from the current lock were dropped. */
483 lock->l_policy_data.l_inodebits.cancel_bits = drop_bits;
484 /* Finally clear these bits in lock ibits */
485 ldlm_inodebits_drop(lock, drop_bits);
486 unlock_res_and_lock(lock);
487 /* Finally call cancel callback for remaining bits only.
488 * It is important to have converting flag during that
489 * so blocking_ast callback can distinguish convert from
492 if (lock->l_blocking_ast)
493 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
496 /* now notify server about convert */
497 rc = ldlm_cli_convert(lock, &flags);
499 lock_res_and_lock(lock);
500 if (ldlm_is_converting(lock)) {
501 ldlm_clear_converting(lock);
502 ldlm_set_cbpending(lock);
503 ldlm_set_bl_ast(lock);
505 unlock_res_and_lock(lock);
510 LDLM_DEBUG(lock, "client lock convert END");
515 int ldlm_inodebits_alloc_lock(struct ldlm_lock *lock)
517 if (ldlm_is_ns_srv(lock)) {
520 OBD_SLAB_ALLOC_PTR(lock->l_ibits_node, ldlm_inodebits_slab);
521 if (lock->l_ibits_node == NULL)
523 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
524 INIT_LIST_HEAD(&lock->l_ibits_node->lin_link[i]);
525 lock->l_ibits_node->lock = lock;
527 lock->l_ibits_node = NULL;
532 void ldlm_inodebits_add_lock(struct ldlm_resource *res, struct list_head *head,
533 struct ldlm_lock *lock)
537 if (!ldlm_is_ns_srv(lock))
540 if (head == &res->lr_waiting) {
541 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++) {
542 if (lock->l_policy_data.l_inodebits.bits & (1 << i))
543 list_add_tail(&lock->l_ibits_node->lin_link[i],
544 &res->lr_ibits_queues->liq_waiting[i]);
546 } else if (head == &res->lr_granted && lock->l_ibits_node != NULL) {
547 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
548 LASSERT(list_empty(&lock->l_ibits_node->lin_link[i]));
549 OBD_SLAB_FREE_PTR(lock->l_ibits_node, ldlm_inodebits_slab);
550 lock->l_ibits_node = NULL;
554 void ldlm_inodebits_unlink_lock(struct ldlm_lock *lock)
558 ldlm_unlink_lock_skiplist(lock);
559 if (!ldlm_is_ns_srv(lock))
562 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
563 list_del_init(&lock->l_ibits_node->lin_link[i]);