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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2011, Whamcloud, Inc.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_LDLM
40 #include <liblustre.h>
43 #include <lustre_dlm.h>
44 #include <obd_class.h>
47 #include "ldlm_internal.h"
49 int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
50 CFS_MODULE_PARM(ldlm_enqueue_min, "i", int, 0644,
51 "lock enqueue timeout minimum");
53 /* in client side, whether the cached locks will be canceled before replay */
54 unsigned int ldlm_cancel_unused_locks_before_replay = 1;
56 static void interrupted_completion_wait(void *data)
60 struct lock_wait_data {
61 struct ldlm_lock *lwd_lock;
65 struct ldlm_async_args {
66 struct lustre_handle lock_handle;
69 int ldlm_expired_completion_wait(void *data)
71 struct lock_wait_data *lwd = data;
72 struct ldlm_lock *lock = lwd->lwd_lock;
73 struct obd_import *imp;
74 struct obd_device *obd;
77 if (lock->l_conn_export == NULL) {
78 static cfs_time_t next_dump = 0, last_dump = 0;
80 if (ptlrpc_check_suspend())
83 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
84 CFS_DURATION_T"s ago); not entering recovery in "
85 "server code, just going back to sleep",
86 lock->l_last_activity,
87 cfs_time_sub(cfs_time_current_sec(),
88 lock->l_last_activity));
89 if (cfs_time_after(cfs_time_current(), next_dump)) {
90 last_dump = next_dump;
91 next_dump = cfs_time_shift(300);
92 ldlm_namespace_dump(D_DLMTRACE,
93 ldlm_lock_to_ns(lock));
95 libcfs_debug_dumplog();
100 obd = lock->l_conn_export->exp_obd;
101 imp = obd->u.cli.cl_import;
102 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
103 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
104 CFS_DURATION_T"s ago), entering recovery for %s@%s",
105 lock->l_last_activity,
106 cfs_time_sub(cfs_time_current_sec(), lock->l_last_activity),
107 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
111 EXPORT_SYMBOL(ldlm_expired_completion_wait);
113 /* We use the same basis for both server side and client side functions
114 from a single node. */
115 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
117 int timeout = at_get(ldlm_lock_to_ns_at(lock));
119 return obd_timeout / 2;
120 /* Since these are non-updating timeouts, we should be conservative.
121 It would be nice to have some kind of "early reply" mechanism for
122 lock callbacks too... */
123 timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */
124 return max(timeout, ldlm_enqueue_min);
126 EXPORT_SYMBOL(ldlm_get_enq_timeout);
129 * Helper function for ldlm_completion_ast(), updating timings when lock is
132 static int ldlm_completion_tail(struct ldlm_lock *lock)
137 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
138 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
141 delay = cfs_time_sub(cfs_time_current_sec(),
142 lock->l_last_activity);
143 LDLM_DEBUG(lock, "client-side enqueue: granted after "
144 CFS_DURATION_T"s", delay);
146 /* Update our time estimate */
147 at_measured(ldlm_lock_to_ns_at(lock),
155 * Implementation of ->l_completion_ast() for a client, that doesn't wait
156 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
157 * other threads that cannot block for long.
159 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
163 if (flags == LDLM_FL_WAIT_NOREPROC) {
164 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
168 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
169 LDLM_FL_BLOCK_CONV))) {
170 cfs_waitq_signal(&lock->l_waitq);
171 RETURN(ldlm_completion_tail(lock));
174 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
176 ldlm_reprocess_all(lock->l_resource);
179 EXPORT_SYMBOL(ldlm_completion_ast_async);
182 * Client side LDLM "completion" AST. This is called in several cases:
184 * - when a reply to an ENQUEUE rpc is received from the server
185 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
186 * this point (determined by flags);
188 * - when LDLM_CP_CALLBACK rpc comes to client to notify it that lock has
191 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
194 * - to force all locks when resource is destroyed (cleanup_resource());
196 * - during lock conversion (not used currently).
198 * If lock is not granted in the first case, this function waits until second
199 * or penultimate cases happen in some other thread.
202 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
204 /* XXX ALLOCATE - 160 bytes */
205 struct lock_wait_data lwd;
206 struct obd_device *obd;
207 struct obd_import *imp = NULL;
208 struct l_wait_info lwi;
213 if (flags == LDLM_FL_WAIT_NOREPROC) {
214 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
218 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
219 LDLM_FL_BLOCK_CONV))) {
220 cfs_waitq_signal(&lock->l_waitq);
224 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
229 obd = class_exp2obd(lock->l_conn_export);
231 /* if this is a local lock, then there is no import */
233 imp = obd->u.cli.cl_import;
236 /* Wait a long time for enqueue - server may have to callback a
237 lock from another client. Server will evict the other client if it
238 doesn't respond reasonably, and then give us the lock. */
239 timeout = ldlm_get_enq_timeout(lock) * 2;
243 if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
244 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
245 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
247 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
248 ldlm_expired_completion_wait,
249 interrupted_completion_wait, &lwd);
253 spin_lock(&imp->imp_lock);
254 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
255 spin_unlock(&imp->imp_lock);
258 if (ns_is_client(ldlm_lock_to_ns(lock)) &&
259 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
260 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
261 lock->l_flags |= LDLM_FL_FAIL_LOC;
264 /* Go to sleep until the lock is granted or cancelled. */
265 rc = l_wait_event(lock->l_waitq,
266 is_granted_or_cancelled(lock), &lwi);
270 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
275 RETURN(ldlm_completion_tail(lock));
277 EXPORT_SYMBOL(ldlm_completion_ast);
280 * A helper to build a blocking ast function
282 * Perform a common operation for blocking asts:
283 * defferred lock cancellation.
285 * \param lock the lock blocking or canceling ast was called on
287 * \see mdt_blocking_ast
288 * \see ldlm_blocking_ast
290 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
295 lock->l_flags |= LDLM_FL_CBPENDING;
296 do_ast = (!lock->l_readers && !lock->l_writers);
297 unlock_res_and_lock(lock);
300 struct lustre_handle lockh;
303 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
304 ldlm_lock2handle(lock, &lockh);
305 rc = ldlm_cli_cancel(&lockh);
307 CERROR("ldlm_cli_cancel: %d\n", rc);
309 LDLM_DEBUG(lock, "Lock still has references, will be "
314 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
317 * Server blocking AST
319 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
322 * \param lock the lock which blocks a request or cancelling lock
325 * \param flag indicates whether this cancelling or blocking callback
327 * \see ldlm_blocking_ast_nocheck
329 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
330 void *data, int flag)
334 if (flag == LDLM_CB_CANCELING) {
335 /* Don't need to do anything here. */
339 lock_res_and_lock(lock);
340 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
341 * that ldlm_blocking_ast is called just before intent_policy method
342 * takes the lr_lock, then by the time we get the lock, we might not
343 * be the correct blocking function anymore. So check, and return
345 if (lock->l_blocking_ast != ldlm_blocking_ast) {
346 unlock_res_and_lock(lock);
349 RETURN(ldlm_blocking_ast_nocheck(lock));
351 EXPORT_SYMBOL(ldlm_blocking_ast);
354 * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
355 * comment in filter_intent_policy() on why you may need this.
357 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
360 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
361 * that is rather subtle: with OST-side locking, it may so happen that
362 * _all_ extent locks are held by the OST. If client wants to obtain
363 * current file size it calls ll{,u}_glimpse_size(), and (as locks are
364 * on the server), dummy glimpse callback fires and does
365 * nothing. Client still receives correct file size due to the
366 * following fragment in filter_intent_policy():
368 * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
369 * if (rc != 0 && res->lr_namespace->ns_lvbo &&
370 * res->lr_namespace->ns_lvbo->lvbo_update) {
371 * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
374 * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
375 * returns correct file size to the client.
377 return -ELDLM_NO_LOCK_DATA;
379 EXPORT_SYMBOL(ldlm_glimpse_ast);
381 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
382 const struct ldlm_res_id *res_id,
383 ldlm_type_t type, ldlm_policy_data_t *policy,
384 ldlm_mode_t mode, __u64 *flags,
385 ldlm_blocking_callback blocking,
386 ldlm_completion_callback completion,
387 ldlm_glimpse_callback glimpse,
388 void *data, __u32 lvb_len,
389 const __u64 *client_cookie,
390 struct lustre_handle *lockh)
392 struct ldlm_lock *lock;
394 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
395 .lcs_blocking = blocking,
396 .lcs_glimpse = glimpse,
400 LASSERT(!(*flags & LDLM_FL_REPLAY));
401 if (unlikely(ns_is_client(ns))) {
402 CERROR("Trying to enqueue local lock in a shadow namespace\n");
406 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len);
408 GOTO(out_nolock, err = -ENOMEM);
410 ldlm_lock2handle(lock, lockh);
412 /* NB: we don't have any lock now (lock_res_and_lock)
413 * because it's a new lock */
414 ldlm_lock_addref_internal_nolock(lock, mode);
415 lock->l_flags |= LDLM_FL_LOCAL;
416 if (*flags & LDLM_FL_ATOMIC_CB)
417 lock->l_flags |= LDLM_FL_ATOMIC_CB;
420 lock->l_policy_data = *policy;
421 if (client_cookie != NULL)
422 lock->l_client_cookie = *client_cookie;
423 if (type == LDLM_EXTENT)
424 lock->l_req_extent = policy->l_extent;
426 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
427 if (unlikely(err != ELDLM_OK))
431 *policy = lock->l_policy_data;
433 if (lock->l_completion_ast)
434 lock->l_completion_ast(lock, *flags, NULL);
436 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
439 LDLM_LOCK_RELEASE(lock);
443 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
445 static void failed_lock_cleanup(struct ldlm_namespace *ns,
446 struct ldlm_lock *lock, int mode)
450 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
451 lock_res_and_lock(lock);
452 /* Check that lock is not granted or failed, we might race. */
453 if ((lock->l_req_mode != lock->l_granted_mode) &&
454 !(lock->l_flags & LDLM_FL_FAILED)) {
455 /* Make sure that this lock will not be found by raced
456 * bl_ast and -EINVAL reply is sent to server anyways.
458 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
459 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
462 unlock_res_and_lock(lock);
466 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
467 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
469 LDLM_DEBUG(lock, "lock was granted or failed in race");
471 ldlm_lock_decref_internal(lock, mode);
473 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
474 * from llite/file.c/ll_file_flock(). */
475 /* This code makes for the fact that we do not have blocking handler on
476 * a client for flock locks. As such this is the place where we must
477 * completely kill failed locks. (interrupted and those that
478 * were waiting to be granted when server evicted us. */
479 if (lock->l_resource->lr_type == LDLM_FLOCK) {
480 lock_res_and_lock(lock);
481 ldlm_resource_unlink_lock(lock);
482 ldlm_lock_destroy_nolock(lock);
483 unlock_res_and_lock(lock);
487 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
488 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
489 __u64 *flags, void *lvb, __u32 lvb_len,
490 struct lustre_handle *lockh,int rc)
492 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
493 int is_replay = *flags & LDLM_FL_REPLAY;
494 struct ldlm_lock *lock;
495 struct ldlm_reply *reply;
496 struct ost_lvb *tmplvb;
497 int cleanup_phase = 1;
500 lock = ldlm_handle2lock(lockh);
501 /* ldlm_cli_enqueue is holding a reference on this lock. */
503 LASSERT(type == LDLM_FLOCK);
507 if (rc != ELDLM_OK) {
509 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
510 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
511 if (rc == ELDLM_LOCK_ABORTED) {
512 /* Before we return, swab the reply */
513 reply = req_capsule_server_get(&req->rq_pill,
519 req_capsule_set_size(&req->rq_pill,
520 &RMF_DLM_LVB, RCL_SERVER,
522 tmplvb = req_capsule_server_get(&req->rq_pill,
525 GOTO(cleanup, rc = -EPROTO);
527 memcpy(lvb, tmplvb, lvb_len);
533 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
535 GOTO(cleanup, rc = -EPROTO);
537 /* lock enqueued on the server */
540 lock_res_and_lock(lock);
541 /* Key change rehash lock in per-export hash with new key */
542 if (exp->exp_lock_hash) {
543 /* In the function below, .hs_keycmp resolves to
544 * ldlm_export_lock_keycmp() */
545 /* coverity[overrun-buffer-val] */
546 cfs_hash_rehash_key(exp->exp_lock_hash,
547 &lock->l_remote_handle,
551 lock->l_remote_handle = reply->lock_handle;
554 *flags = ldlm_flags_from_wire(reply->lock_flags);
555 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
557 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
558 * to wait with no timeout as well */
559 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
561 unlock_res_and_lock(lock);
563 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%llx\n",
564 lock, reply->lock_handle.cookie, *flags);
566 /* If enqueue returned a blocked lock but the completion handler has
567 * already run, then it fixed up the resource and we don't need to do it
569 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
570 int newmode = reply->lock_desc.l_req_mode;
572 if (newmode && newmode != lock->l_req_mode) {
573 LDLM_DEBUG(lock, "server returned different mode %s",
574 ldlm_lockname[newmode]);
575 lock->l_req_mode = newmode;
578 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
579 lock->l_resource->lr_name.name,
580 sizeof(struct ldlm_res_id))) {
581 CDEBUG(D_INFO, "remote intent success, locking "
582 "(%ld,%ld,%ld) instead of "
584 (long)reply->lock_desc.l_resource.lr_name.name[0],
585 (long)reply->lock_desc.l_resource.lr_name.name[1],
586 (long)reply->lock_desc.l_resource.lr_name.name[2],
587 (long)lock->l_resource->lr_name.name[0],
588 (long)lock->l_resource->lr_name.name[1],
589 (long)lock->l_resource->lr_name.name[2]);
591 rc = ldlm_lock_change_resource(ns, lock,
592 &reply->lock_desc.l_resource.lr_name);
593 if (rc || lock->l_resource == NULL)
594 GOTO(cleanup, rc = -ENOMEM);
595 LDLM_DEBUG(lock, "client-side enqueue, new resource");
598 if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
600 /* We assume lock type cannot change on server*/
601 ldlm_convert_policy_to_local(exp,
602 lock->l_resource->lr_type,
603 &reply->lock_desc.l_policy_data,
604 &lock->l_policy_data);
605 if (type != LDLM_PLAIN)
606 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
609 if ((*flags) & LDLM_FL_AST_SENT ||
610 /* Cancel extent locks as soon as possible on a liblustre client,
611 * because it cannot handle asynchronous ASTs robustly (see
613 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
614 lock_res_and_lock(lock);
615 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
616 unlock_res_and_lock(lock);
617 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
620 /* If the lock has already been granted by a completion AST, don't
621 * clobber the LVB with an older one. */
623 /* We must lock or a racing completion might update lvb
624 without letting us know and we'll clobber the correct value.
625 Cannot unlock after the check either, a that still leaves
626 a tiny window for completion to get in */
627 lock_res_and_lock(lock);
628 if (lock->l_req_mode != lock->l_granted_mode) {
630 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
631 RCL_SERVER, lvb_len);
632 tmplvb = req_capsule_server_get(&req->rq_pill,
634 if (tmplvb == NULL) {
635 unlock_res_and_lock(lock);
636 GOTO(cleanup, rc = -EPROTO);
638 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
640 unlock_res_and_lock(lock);
644 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
645 if (lock->l_completion_ast != NULL) {
646 int err = lock->l_completion_ast(lock, *flags, NULL);
654 if (lvb_len && lvb != NULL) {
655 /* Copy the LVB here, and not earlier, because the completion
656 * AST (if any) can override what we got in the reply */
657 memcpy(lvb, lock->l_lvb_data, lvb_len);
660 LDLM_DEBUG(lock, "client-side enqueue END");
663 if (cleanup_phase == 1 && rc)
664 failed_lock_cleanup(ns, lock, mode);
665 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
667 LDLM_LOCK_RELEASE(lock);
670 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
672 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
673 * a single page on the send/receive side. XXX: 512 should be changed
674 * to more adequate value. */
675 static inline int ldlm_req_handles_avail(int req_size, int off)
679 avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size;
680 if (likely(avail >= 0))
681 avail /= (int)sizeof(struct lustre_handle);
684 avail += LDLM_LOCKREQ_HANDLES - off;
689 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
690 enum req_location loc,
693 int size = req_capsule_msg_size(pill, loc);
694 return ldlm_req_handles_avail(size, off);
697 static inline int ldlm_format_handles_avail(struct obd_import *imp,
698 const struct req_format *fmt,
699 enum req_location loc, int off)
701 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
702 return ldlm_req_handles_avail(size, off);
705 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
706 * @count locks in @cancels. */
707 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
708 int version, int opc, int canceloff,
709 cfs_list_t *cancels, int count)
711 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
712 struct req_capsule *pill = &req->rq_pill;
713 struct ldlm_request *dlm = NULL;
714 int flags, avail, to_free, pack = 0;
721 if (ns_connect_cancelset(ns)) {
722 /* Estimate the amount of available space in the request. */
723 req_capsule_filled_sizes(pill, RCL_CLIENT);
724 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
726 flags = ns_connect_lru_resize(ns) ?
727 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
728 to_free = !ns_connect_lru_resize(ns) &&
729 opc == LDLM_ENQUEUE ? 1 : 0;
731 /* Cancel lru locks here _only_ if the server supports
732 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
733 * rpc, what will make us slower. */
735 count += ldlm_cancel_lru_local(ns, cancels, to_free,
736 avail - count, 0, flags);
741 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
742 ldlm_request_bufsize(pack, opc));
745 rc = ptlrpc_request_pack(req, version, opc);
747 ldlm_lock_list_put(cancels, l_bl_ast, count);
751 if (ns_connect_cancelset(ns)) {
753 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
755 /* Skip first lock handler in ldlm_request_pack(),
756 * this method will incrment @lock_count according
757 * to the lock handle amount actually written to
759 dlm->lock_count = canceloff;
761 /* Pack into the request @pack lock handles. */
762 ldlm_cli_cancel_list(cancels, pack, req, 0);
763 /* Prepare and send separate cancel rpc for others. */
764 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
766 ldlm_lock_list_put(cancels, l_bl_ast, count);
770 EXPORT_SYMBOL(ldlm_prep_elc_req);
772 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
773 cfs_list_t *cancels, int count)
775 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
776 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
778 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
780 /* If a request has some specific initialisation it is passed in @reqp,
781 * otherwise it is created in ldlm_cli_enqueue.
783 * Supports sync and async requests, pass @async flag accordingly. If a
784 * request was created in ldlm_cli_enqueue and it is the async request,
785 * pass it to the caller in @reqp. */
786 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
787 struct ldlm_enqueue_info *einfo,
788 const struct ldlm_res_id *res_id,
789 ldlm_policy_data_t const *policy, __u64 *flags,
790 void *lvb, __u32 lvb_len, struct lustre_handle *lockh,
793 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
794 struct ldlm_lock *lock;
795 struct ldlm_request *body;
796 int is_replay = *flags & LDLM_FL_REPLAY;
797 int req_passed_in = 1;
799 struct ptlrpc_request *req;
802 LASSERT(exp != NULL);
804 /* If we're replaying this lock, just check some invariants.
805 * If we're creating a new lock, get everything all setup nice. */
807 lock = ldlm_handle2lock_long(lockh, 0);
808 LASSERT(lock != NULL);
809 LDLM_DEBUG(lock, "client-side enqueue START");
810 LASSERT(exp == lock->l_conn_export);
812 const struct ldlm_callback_suite cbs = {
813 .lcs_completion = einfo->ei_cb_cp,
814 .lcs_blocking = einfo->ei_cb_bl,
815 .lcs_glimpse = einfo->ei_cb_gl,
816 .lcs_weigh = einfo->ei_cb_wg
818 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
819 einfo->ei_mode, &cbs, einfo->ei_cbdata,
823 /* for the local lock, add the reference */
824 ldlm_lock_addref_internal(lock, einfo->ei_mode);
825 ldlm_lock2handle(lock, lockh);
826 if (policy != NULL) {
827 /* INODEBITS_INTEROP: If the server does not support
828 * inodebits, we will request a plain lock in the
829 * descriptor (ldlm_lock2desc() below) but use an
830 * inodebits lock internally with both bits set.
832 if (einfo->ei_type == LDLM_IBITS &&
833 !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
834 lock->l_policy_data.l_inodebits.bits =
835 MDS_INODELOCK_LOOKUP |
836 MDS_INODELOCK_UPDATE;
838 lock->l_policy_data = *policy;
841 if (einfo->ei_type == LDLM_EXTENT)
842 lock->l_req_extent = policy->l_extent;
843 LDLM_DEBUG(lock, "client-side enqueue START, flags %llx\n",
847 lock->l_conn_export = exp;
848 lock->l_export = NULL;
849 lock->l_blocking_ast = einfo->ei_cb_bl;
850 lock->l_flags |= (*flags & LDLM_FL_NO_LRU);
852 /* lock not sent to server yet */
854 if (reqp == NULL || *reqp == NULL) {
855 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
860 failed_lock_cleanup(ns, lock, einfo->ei_mode);
861 LDLM_LOCK_RELEASE(lock);
871 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
873 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
874 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
877 /* Dump lock data into the request buffer */
878 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
879 ldlm_lock2desc(lock, &body->lock_desc);
880 body->lock_flags = ldlm_flags_to_wire(*flags);
881 body->lock_handle[0] = *lockh;
883 /* Continue as normal. */
884 if (!req_passed_in) {
886 req_capsule_extend(&req->rq_pill,
887 &RQF_LDLM_ENQUEUE_LVB);
888 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
889 RCL_SERVER, lvb_len);
891 ptlrpc_request_set_replen(req);
895 * Liblustre client doesn't get extent locks, except for O_APPEND case
896 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
897 * [i_size, OBD_OBJECT_EOF] lock is taken.
899 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
900 policy->l_extent.end == OBD_OBJECT_EOF));
903 LASSERT(reqp != NULL);
907 LDLM_DEBUG(lock, "sending request");
909 rc = ptlrpc_queue_wait(req);
911 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
912 einfo->ei_mode, flags, lvb, lvb_len,
915 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
916 * one reference that we took */
918 LDLM_LOCK_RELEASE(lock);
922 if (!req_passed_in && req != NULL) {
923 ptlrpc_req_finished(req);
930 EXPORT_SYMBOL(ldlm_cli_enqueue);
932 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
935 struct ldlm_resource *res;
938 if (ns_is_client(ldlm_lock_to_ns(lock))) {
939 CERROR("Trying to cancel local lock\n");
942 LDLM_DEBUG(lock, "client-side local convert");
944 res = ldlm_lock_convert(lock, new_mode, flags);
946 ldlm_reprocess_all(res);
951 LDLM_DEBUG(lock, "client-side local convert handler END");
956 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
957 * conversion of locks which are on the waiting or converting queue */
958 /* Caller of this code is supposed to take care of lock readers/writers
960 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
962 struct ldlm_request *body;
963 struct ldlm_reply *reply;
964 struct ldlm_lock *lock;
965 struct ldlm_resource *res;
966 struct ptlrpc_request *req;
970 lock = ldlm_handle2lock(lockh);
977 if (lock->l_conn_export == NULL)
978 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
980 LDLM_DEBUG(lock, "client-side convert");
982 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
983 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
990 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
991 body->lock_handle[0] = lock->l_remote_handle;
993 body->lock_desc.l_req_mode = new_mode;
994 body->lock_flags = ldlm_flags_to_wire(*flags);
997 ptlrpc_request_set_replen(req);
998 rc = ptlrpc_queue_wait(req);
1002 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1004 GOTO(out, rc = -EPROTO);
1007 GOTO(out, rc = req->rq_status);
1009 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
1011 ldlm_reprocess_all(res);
1012 /* Go to sleep until the lock is granted. */
1013 /* FIXME: or cancelled. */
1014 if (lock->l_completion_ast) {
1015 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1025 LDLM_LOCK_PUT(lock);
1026 ptlrpc_req_finished(req);
1029 EXPORT_SYMBOL(ldlm_cli_convert);
1031 /* Cancel locks locally.
1033 * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
1034 * LDLM_FL_CANCELING otherwise;
1035 * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
1036 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1038 __u64 rc = LDLM_FL_LOCAL_ONLY;
1041 if (lock->l_conn_export) {
1044 LDLM_DEBUG(lock, "client-side cancel");
1045 /* Set this flag to prevent others from getting new references*/
1046 lock_res_and_lock(lock);
1047 lock->l_flags |= LDLM_FL_CBPENDING;
1048 local_only = !!(lock->l_flags &
1049 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1050 ldlm_cancel_callback(lock);
1051 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
1052 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1053 unlock_res_and_lock(lock);
1056 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1058 rc = LDLM_FL_LOCAL_ONLY;
1060 ldlm_lock_cancel(lock);
1062 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1063 LDLM_ERROR(lock, "Trying to cancel local lock");
1066 LDLM_DEBUG(lock, "server-side local cancel");
1067 ldlm_lock_cancel(lock);
1068 ldlm_reprocess_all(lock->l_resource);
1074 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
1075 of the request @req. */
1076 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1077 cfs_list_t *head, int count)
1079 struct ldlm_request *dlm;
1080 struct ldlm_lock *lock;
1081 int max, packed = 0;
1084 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1085 LASSERT(dlm != NULL);
1087 /* Check the room in the request buffer. */
1088 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1089 sizeof(struct ldlm_request);
1090 max /= sizeof(struct lustre_handle);
1091 max += LDLM_LOCKREQ_HANDLES;
1092 LASSERT(max >= dlm->lock_count + count);
1094 /* XXX: it would be better to pack lock handles grouped by resource.
1095 * so that the server cancel would call filter_lvbo_update() less
1097 cfs_list_for_each_entry(lock, head, l_bl_ast) {
1100 LASSERT(lock->l_conn_export);
1101 /* Pack the lock handle to the given request buffer. */
1102 LDLM_DEBUG(lock, "packing");
1103 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1106 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1110 /* Prepare and send a batched cancel rpc, it will include count lock handles
1111 * of locks given in @head. */
1112 int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels,
1113 int count, ldlm_cancel_flags_t flags)
1115 struct ptlrpc_request *req = NULL;
1116 struct obd_import *imp;
1121 LASSERT(exp != NULL);
1124 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1126 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1129 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1130 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1135 imp = class_exp2cliimp(exp);
1136 if (imp == NULL || imp->imp_invalid) {
1138 "skipping cancel on invalid import %p\n", imp);
1142 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1144 GOTO(out, rc = -ENOMEM);
1146 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1147 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1148 ldlm_request_bufsize(count, LDLM_CANCEL));
1150 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1152 ptlrpc_request_free(req);
1156 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1157 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1158 ptlrpc_at_set_req_timeout(req);
1160 ldlm_cancel_pack(req, cancels, count);
1162 ptlrpc_request_set_replen(req);
1163 if (flags & LCF_ASYNC) {
1164 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1168 rc = ptlrpc_queue_wait(req);
1171 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1172 "out of sync -- not fatal\n",
1173 libcfs_nid2str(req->rq_import->
1174 imp_connection->c_peer.nid));
1176 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1177 req->rq_import_generation == imp->imp_generation) {
1178 ptlrpc_req_finished(req);
1180 } else if (rc != ELDLM_OK) {
1181 if (rc != -ESHUTDOWN)
1182 CERROR("Got rc %d from cancel RPC: canceling "
1190 ptlrpc_req_finished(req);
1193 return sent ? sent : rc;
1195 EXPORT_SYMBOL(ldlm_cli_cancel_req);
1197 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1199 LASSERT(imp != NULL);
1200 return &imp->imp_obd->obd_namespace->ns_pool;
1204 * Update client's obd pool related fields with new SLV and Limit from \a req.
1206 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1208 struct obd_device *obd;
1212 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1213 !imp_connect_lru_resize(req->rq_import)))
1216 * Do nothing for corner cases.
1222 * In some cases RPC may contain slv and limit zeroed out. This is
1223 * the case when server does not support lru resize feature. This is
1224 * also possible in some recovery cases when server side reqs have no
1225 * ref to obd export and thus access to server side namespace is no
1228 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1229 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1230 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1231 "(SLV: "LPU64", Limit: %u)",
1232 lustre_msg_get_slv(req->rq_repmsg),
1233 lustre_msg_get_limit(req->rq_repmsg));
1237 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1238 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1239 obd = req->rq_import->imp_obd;
1242 * Set new SLV and Limit to obd fields to make accessible for pool
1243 * thread. We do not access obd_namespace and pool directly here
1244 * as there is no reliable way to make sure that they are still
1245 * alive in cleanup time. Evil races are possible which may cause
1246 * oops in that time.
1248 write_lock(&obd->obd_pool_lock);
1249 obd->obd_pool_slv = new_slv;
1250 obd->obd_pool_limit = new_limit;
1251 write_unlock(&obd->obd_pool_lock);
1255 EXPORT_SYMBOL(ldlm_cli_update_pool);
1257 int ldlm_cli_cancel(struct lustre_handle *lockh)
1259 struct obd_export *exp;
1260 int avail, flags, count = 1;
1262 struct ldlm_namespace *ns;
1263 struct ldlm_lock *lock;
1264 CFS_LIST_HEAD(cancels);
1267 /* concurrent cancels on the same handle can happen */
1268 lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1270 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1274 rc = ldlm_cli_cancel_local(lock);
1275 if (rc == LDLM_FL_LOCAL_ONLY) {
1276 LDLM_LOCK_RELEASE(lock);
1279 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1280 * rpc which goes to canceld portal, so we can cancel other lru locks
1281 * here and send them all as one LDLM_CANCEL rpc. */
1282 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1283 cfs_list_add(&lock->l_bl_ast, &cancels);
1285 exp = lock->l_conn_export;
1286 if (exp_connect_cancelset(exp)) {
1287 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1292 ns = ldlm_lock_to_ns(lock);
1293 flags = ns_connect_lru_resize(ns) ?
1294 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1295 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1298 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1301 EXPORT_SYMBOL(ldlm_cli_cancel);
1303 /* XXX until we will have compound requests and can cut cancels from generic rpc
1304 * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */
1305 int ldlm_cli_cancel_list_local(cfs_list_t *cancels, int count,
1306 ldlm_cancel_flags_t flags)
1308 CFS_LIST_HEAD(head);
1309 struct ldlm_lock *lock, *next;
1310 int left = 0, bl_ast = 0;
1314 cfs_list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1318 if (flags & LCF_LOCAL) {
1319 rc = LDLM_FL_LOCAL_ONLY;
1320 ldlm_lock_cancel(lock);
1322 rc = ldlm_cli_cancel_local(lock);
1324 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1325 LDLM_DEBUG(lock, "Cancel lock separately");
1326 cfs_list_del_init(&lock->l_bl_ast);
1327 cfs_list_add(&lock->l_bl_ast, &head);
1331 if (rc == LDLM_FL_LOCAL_ONLY) {
1332 /* CANCEL RPC should not be sent to server. */
1333 cfs_list_del_init(&lock->l_bl_ast);
1334 LDLM_LOCK_RELEASE(lock);
1341 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1346 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
1349 * Cancel as many locks as possible w/o sending any rpcs (e.g. to write back
1350 * dirty data, to close a file, ...) or waiting for any rpcs in-flight (e.g.
1351 * readahead requests, ...)
1353 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1354 struct ldlm_lock *lock,
1355 int unused, int added,
1358 ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1359 ldlm_cancel_for_recovery cb = ns->ns_cancel_for_recovery;
1360 lock_res_and_lock(lock);
1362 /* don't check added & count since we want to process all locks
1363 * from unused list */
1364 switch (lock->l_resource->lr_type) {
1370 result = LDLM_POLICY_SKIP_LOCK;
1371 lock->l_flags |= LDLM_FL_SKIPPED;
1375 unlock_res_and_lock(lock);
1380 * Callback function for lru-resize policy. Makes decision whether to keep
1381 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1382 * \a added and number of locks to be preferably canceled \a count.
1384 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1386 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1388 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1389 struct ldlm_lock *lock,
1390 int unused, int added,
1393 cfs_time_t cur = cfs_time_current();
1394 struct ldlm_pool *pl = &ns->ns_pool;
1399 * Stop lru processing when we reached passed @count or checked all
1402 if (count && added >= count)
1403 return LDLM_POLICY_KEEP_LOCK;
1405 slv = ldlm_pool_get_slv(pl);
1406 lvf = ldlm_pool_get_lvf(pl);
1407 la = cfs_duration_sec(cfs_time_sub(cur,
1408 lock->l_last_used));
1411 * Stop when slv is not yet come from server or lv is smaller than
1414 lv = lvf * la * unused;
1417 * Inform pool about current CLV to see it via proc.
1419 ldlm_pool_set_clv(pl, lv);
1420 return (slv == 0 || lv < slv) ?
1421 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1425 * Callback function for proc used policy. Makes decision whether to keep
1426 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1427 * \a added and number of locks to be preferably canceled \a count.
1429 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1431 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1433 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1434 struct ldlm_lock *lock,
1435 int unused, int added,
1439 * Stop lru processing when we reached passed @count or checked all
1442 return (added >= count) ?
1443 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1447 * Callback function for aged policy. Makes decision whether to keep
1448 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1449 * \a added and number of locks to be preferably canceled \a count.
1451 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1453 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1455 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1456 struct ldlm_lock *lock,
1457 int unused, int added,
1461 * Stop lru processing if young lock is found and we reached passed
1464 return ((added >= count) &&
1465 cfs_time_before(cfs_time_current(),
1466 cfs_time_add(lock->l_last_used,
1468 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1472 * Callback function for default policy. Makes decision whether to keep
1473 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1474 * \a added and number of locks to be preferably canceled \a count.
1476 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1478 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1480 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1481 struct ldlm_lock *lock,
1482 int unused, int added,
1486 * Stop lru processing when we reached passed @count or checked all
1489 return (added >= count) ?
1490 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1493 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1494 struct ldlm_lock *, int,
1497 static ldlm_cancel_lru_policy_t
1498 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1500 if (flags & LDLM_CANCEL_NO_WAIT)
1501 return ldlm_cancel_no_wait_policy;
1503 if (ns_connect_lru_resize(ns)) {
1504 if (flags & LDLM_CANCEL_SHRINK)
1505 /* We kill passed number of old locks. */
1506 return ldlm_cancel_passed_policy;
1507 else if (flags & LDLM_CANCEL_LRUR)
1508 return ldlm_cancel_lrur_policy;
1509 else if (flags & LDLM_CANCEL_PASSED)
1510 return ldlm_cancel_passed_policy;
1512 if (flags & LDLM_CANCEL_AGED)
1513 return ldlm_cancel_aged_policy;
1516 return ldlm_cancel_default_policy;
1519 /* - Free space in lru for @count new locks,
1520 * redundant unused locks are canceled locally;
1521 * - also cancel locally unused aged locks;
1522 * - do not cancel more than @max locks;
1523 * - GET the found locks and add them into the @cancels list.
1525 * A client lock can be added to the l_bl_ast list only when it is
1526 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
1527 * There are the following use cases: ldlm_cancel_resource_local(),
1528 * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
1529 * flag properly. As any attempt to cancel a lock rely on this flag,
1530 * l_bl_ast list is accessed later without any special locking.
1532 * Calling policies for enabled lru resize:
1533 * ----------------------------------------
1534 * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to
1535 * cancel not more than @count locks;
1537 * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at
1538 * the beginning of lru list);
1540 * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
1541 * memory pressre policy function;
1543 * flags & LDLM_CANCEL_AGED - cancel alocks according to "aged policy".
1545 * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1546 * (typically before replaying locks) w/o
1547 * sending any rpcs or waiting for any
1548 * outstanding rpc to complete.
1550 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels,
1551 int count, int max, int flags)
1553 ldlm_cancel_lru_policy_t pf;
1554 struct ldlm_lock *lock, *next;
1555 int added = 0, unused, remained;
1558 spin_lock(&ns->ns_lock);
1559 unused = ns->ns_nr_unused;
1562 if (!ns_connect_lru_resize(ns))
1563 count += unused - ns->ns_max_unused;
1565 pf = ldlm_cancel_lru_policy(ns, flags);
1566 LASSERT(pf != NULL);
1568 while (!cfs_list_empty(&ns->ns_unused_list)) {
1569 ldlm_policy_res_t result;
1571 /* all unused locks */
1572 if (remained-- <= 0)
1575 /* For any flags, stop scanning if @max is reached. */
1576 if (max && added >= max)
1579 cfs_list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1581 /* No locks which got blocking requests. */
1582 LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1584 if (flags & LDLM_CANCEL_NO_WAIT &&
1585 lock->l_flags & LDLM_FL_SKIPPED)
1586 /* already processed */
1589 /* Somebody is already doing CANCEL. No need in this
1590 * lock in lru, do not traverse it again. */
1591 if (!(lock->l_flags & LDLM_FL_CANCELING))
1594 ldlm_lock_remove_from_lru_nolock(lock);
1596 if (&lock->l_lru == &ns->ns_unused_list)
1599 LDLM_LOCK_GET(lock);
1600 spin_unlock(&ns->ns_lock);
1601 lu_ref_add(&lock->l_reference, __FUNCTION__, cfs_current());
1603 /* Pass the lock through the policy filter and see if it
1604 * should stay in lru.
1606 * Even for shrinker policy we stop scanning if
1607 * we find a lock that should stay in the cache.
1608 * We should take into account lock age anyway
1609 * as new lock even if it is small of weight is
1610 * valuable resource.
1612 * That is, for shrinker policy we drop only
1613 * old locks, but additionally chose them by
1614 * their weight. Big extent locks will stay in
1616 result = pf(ns, lock, unused, added, count);
1617 if (result == LDLM_POLICY_KEEP_LOCK) {
1618 lu_ref_del(&lock->l_reference,
1619 __FUNCTION__, cfs_current());
1620 LDLM_LOCK_RELEASE(lock);
1621 spin_lock(&ns->ns_lock);
1624 if (result == LDLM_POLICY_SKIP_LOCK) {
1625 lu_ref_del(&lock->l_reference,
1626 __func__, cfs_current());
1627 LDLM_LOCK_RELEASE(lock);
1628 spin_lock(&ns->ns_lock);
1632 lock_res_and_lock(lock);
1633 /* Check flags again under the lock. */
1634 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1635 (ldlm_lock_remove_from_lru(lock) == 0)) {
1636 /* other thread is removing lock from lru or
1637 * somebody is already doing CANCEL or
1638 * there is a blocking request which will send
1639 * cancel by itseft or the lock is matched
1640 * is already not unused. */
1641 unlock_res_and_lock(lock);
1642 lu_ref_del(&lock->l_reference,
1643 __FUNCTION__, cfs_current());
1644 LDLM_LOCK_RELEASE(lock);
1645 spin_lock(&ns->ns_lock);
1648 LASSERT(!lock->l_readers && !lock->l_writers);
1650 /* If we have chosen to cancel this lock voluntarily, we
1651 * better send cancel notification to server, so that it
1652 * frees appropriate state. This might lead to a race
1653 * where while we are doing cancel here, server is also
1654 * silently cancelling this lock. */
1655 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1657 /* Setting the CBPENDING flag is a little misleading,
1658 * but prevents an important race; namely, once
1659 * CBPENDING is set, the lock can accumulate no more
1660 * readers/writers. Since readers and writers are
1661 * already zero here, ldlm_lock_decref() won't see
1662 * this flag and call l_blocking_ast */
1663 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1665 /* We can't re-add to l_lru as it confuses the
1666 * refcounting in ldlm_lock_remove_from_lru() if an AST
1667 * arrives after we drop lr_lock below. We use l_bl_ast
1668 * and can't use l_pending_chain as it is used both on
1669 * server and client nevertheless bug 5666 says it is
1670 * used only on server */
1671 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1672 cfs_list_add(&lock->l_bl_ast, cancels);
1673 unlock_res_and_lock(lock);
1674 lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current());
1675 spin_lock(&ns->ns_lock);
1679 spin_unlock(&ns->ns_lock);
1683 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, cfs_list_t *cancels,
1684 int count, int max, ldlm_cancel_flags_t cancel_flags,
1688 added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1691 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1694 /* when called with LDLM_ASYNC the blocking callback will be handled
1695 * in a thread and this function will return after the thread has been
1696 * asked to call the callback. when called with LDLM_SYNC the blocking
1697 * callback will be performed in this function. */
1698 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t mode,
1701 CFS_LIST_HEAD(cancels);
1706 mode = LDLM_SYNC; /* force to be sync in user space */
1708 /* Just prepare the list of locks, do not actually cancel them yet.
1709 * Locks are cancelled later in a separate thread. */
1710 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1711 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, mode);
1718 /* Find and cancel locally unused locks found on resource, matched to the
1719 * given policy, mode. GET the found locks and add them into the @cancels
1721 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1722 cfs_list_t *cancels,
1723 ldlm_policy_data_t *policy,
1724 ldlm_mode_t mode, int lock_flags,
1725 ldlm_cancel_flags_t cancel_flags, void *opaque)
1727 struct ldlm_lock *lock;
1732 cfs_list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1733 if (opaque != NULL && lock->l_ast_data != opaque) {
1734 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1735 lock->l_ast_data, opaque);
1740 if (lock->l_readers || lock->l_writers)
1743 /* If somebody is already doing CANCEL, or blocking ast came,
1744 * skip this lock. */
1745 if (lock->l_flags & LDLM_FL_BL_AST ||
1746 lock->l_flags & LDLM_FL_CANCELING)
1749 if (lockmode_compat(lock->l_granted_mode, mode))
1752 /* If policy is given and this is IBITS lock, add to list only
1753 * those locks that match by policy. */
1754 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1755 !(lock->l_policy_data.l_inodebits.bits &
1756 policy->l_inodebits.bits))
1759 /* See CBPENDING comment in ldlm_cancel_lru */
1760 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1763 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1764 cfs_list_add(&lock->l_bl_ast, cancels);
1765 LDLM_LOCK_GET(lock);
1770 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
1772 EXPORT_SYMBOL(ldlm_cancel_resource_local);
1774 /* If @req is NULL, send CANCEL request to server with handles of locks
1775 * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1776 * separately per lock.
1777 * If @req is not NULL, put handles of locks in @cancels into the request
1778 * buffer at the offset @off.
1779 * Destroy @cancels at the end. */
1780 int ldlm_cli_cancel_list(cfs_list_t *cancels, int count,
1781 struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1783 struct ldlm_lock *lock;
1787 if (cfs_list_empty(cancels) || count == 0)
1790 /* XXX: requests (both batched and not) could be sent in parallel.
1791 * Usually it is enough to have just 1 RPC, but it is possible that
1792 * there are to many locks to be cancelled in LRU or on a resource.
1793 * It would also speed up the case when the server does not support
1796 LASSERT(!cfs_list_empty(cancels));
1797 lock = cfs_list_entry(cancels->next, struct ldlm_lock,
1799 LASSERT(lock->l_conn_export);
1801 if (exp_connect_cancelset(lock->l_conn_export)) {
1804 ldlm_cancel_pack(req, cancels, count);
1806 res = ldlm_cli_cancel_req(lock->l_conn_export,
1810 res = ldlm_cli_cancel_req(lock->l_conn_export,
1815 if (res != -ESHUTDOWN)
1816 CERROR("ldlm_cli_cancel_list: %d\n", res);
1821 ldlm_lock_list_put(cancels, l_bl_ast, res);
1823 LASSERT(count == 0);
1826 EXPORT_SYMBOL(ldlm_cli_cancel_list);
1828 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1829 const struct ldlm_res_id *res_id,
1830 ldlm_policy_data_t *policy,
1832 ldlm_cancel_flags_t flags,
1835 struct ldlm_resource *res;
1836 CFS_LIST_HEAD(cancels);
1841 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1843 /* This is not a problem. */
1844 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1848 LDLM_RESOURCE_ADDREF(res);
1849 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1850 0, flags | LCF_BL_AST, opaque);
1851 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1853 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1855 LDLM_RESOURCE_DELREF(res);
1856 ldlm_resource_putref(res);
1859 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
1861 struct ldlm_cli_cancel_arg {
1866 static int ldlm_cli_hash_cancel_unused(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1867 cfs_hlist_node_t *hnode, void *arg)
1869 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1870 struct ldlm_cli_cancel_arg *lc = arg;
1873 rc = ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1875 lc->lc_flags, lc->lc_opaque);
1877 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1878 res->lr_name.name[0], rc);
1880 /* must return 0 for hash iteration */
1884 /* Cancel all locks on a namespace (or a specific resource, if given)
1885 * that have 0 readers/writers.
1887 * If flags & LCF_LOCAL, throw the locks away without trying
1888 * to notify the server. */
1889 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1890 const struct ldlm_res_id *res_id,
1891 ldlm_cancel_flags_t flags, void *opaque)
1893 struct ldlm_cli_cancel_arg arg = {
1895 .lc_opaque = opaque,
1903 if (res_id != NULL) {
1904 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1908 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1909 ldlm_cli_hash_cancel_unused, &arg);
1913 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1915 /* Lock iterators. */
1917 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1920 cfs_list_t *tmp, *next;
1921 struct ldlm_lock *lock;
1922 int rc = LDLM_ITER_CONTINUE;
1927 RETURN(LDLM_ITER_CONTINUE);
1930 cfs_list_for_each_safe(tmp, next, &res->lr_granted) {
1931 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1933 if (iter(lock, closure) == LDLM_ITER_STOP)
1934 GOTO(out, rc = LDLM_ITER_STOP);
1937 cfs_list_for_each_safe(tmp, next, &res->lr_converting) {
1938 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1940 if (iter(lock, closure) == LDLM_ITER_STOP)
1941 GOTO(out, rc = LDLM_ITER_STOP);
1944 cfs_list_for_each_safe(tmp, next, &res->lr_waiting) {
1945 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1947 if (iter(lock, closure) == LDLM_ITER_STOP)
1948 GOTO(out, rc = LDLM_ITER_STOP);
1954 EXPORT_SYMBOL(ldlm_resource_foreach);
1956 struct iter_helper_data {
1957 ldlm_iterator_t iter;
1961 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1963 struct iter_helper_data *helper = closure;
1964 return helper->iter(lock, helper->closure);
1967 static int ldlm_res_iter_helper(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1968 cfs_hlist_node_t *hnode, void *arg)
1971 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1973 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
1977 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
1978 ldlm_iterator_t iter, void *closure)
1981 struct iter_helper_data helper = { iter: iter, closure: closure };
1983 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1984 ldlm_res_iter_helper, &helper);
1987 EXPORT_SYMBOL(ldlm_namespace_foreach);
1989 /* non-blocking function to manipulate a lock whose cb_data is being put away.
1990 * return 0: find no resource
1991 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
1994 int ldlm_resource_iterate(struct ldlm_namespace *ns,
1995 const struct ldlm_res_id *res_id,
1996 ldlm_iterator_t iter, void *data)
1998 struct ldlm_resource *res;
2003 CERROR("must pass in namespace\n");
2007 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2011 LDLM_RESOURCE_ADDREF(res);
2012 rc = ldlm_resource_foreach(res, iter, data);
2013 LDLM_RESOURCE_DELREF(res);
2014 ldlm_resource_putref(res);
2017 EXPORT_SYMBOL(ldlm_resource_iterate);
2021 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2023 cfs_list_t *list = closure;
2025 /* we use l_pending_chain here, because it's unused on clients. */
2026 LASSERTF(cfs_list_empty(&lock->l_pending_chain),
2027 "lock %p next %p prev %p\n",
2028 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2029 /* bug 9573: don't replay locks left after eviction, or
2030 * bug 17614: locks being actively cancelled. Get a reference
2031 * on a lock so that it does not disapear under us (e.g. due to cancel)
2033 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
2034 cfs_list_add(&lock->l_pending_chain, list);
2035 LDLM_LOCK_GET(lock);
2038 return LDLM_ITER_CONTINUE;
2041 static int replay_lock_interpret(const struct lu_env *env,
2042 struct ptlrpc_request *req,
2043 struct ldlm_async_args *aa, int rc)
2045 struct ldlm_lock *lock;
2046 struct ldlm_reply *reply;
2047 struct obd_export *exp;
2050 cfs_atomic_dec(&req->rq_import->imp_replay_inflight);
2055 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2057 GOTO(out, rc = -EPROTO);
2059 lock = ldlm_handle2lock(&aa->lock_handle);
2061 CERROR("received replay ack for unknown local cookie "LPX64
2062 " remote cookie "LPX64 " from server %s id %s\n",
2063 aa->lock_handle.cookie, reply->lock_handle.cookie,
2064 req->rq_export->exp_client_uuid.uuid,
2065 libcfs_id2str(req->rq_peer));
2066 GOTO(out, rc = -ESTALE);
2069 /* Key change rehash lock in per-export hash with new key */
2070 exp = req->rq_export;
2071 if (exp && exp->exp_lock_hash) {
2072 /* In the function below, .hs_keycmp resolves to
2073 * ldlm_export_lock_keycmp() */
2074 /* coverity[overrun-buffer-val] */
2075 cfs_hash_rehash_key(exp->exp_lock_hash,
2076 &lock->l_remote_handle,
2077 &reply->lock_handle,
2080 lock->l_remote_handle = reply->lock_handle;
2083 LDLM_DEBUG(lock, "replayed lock:");
2084 ptlrpc_import_recovery_state_machine(req->rq_import);
2085 LDLM_LOCK_PUT(lock);
2088 ptlrpc_connect_import(req->rq_import);
2093 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2095 struct ptlrpc_request *req;
2096 struct ldlm_async_args *aa;
2097 struct ldlm_request *body;
2102 /* Bug 11974: Do not replay a lock which is actively being canceled */
2103 if (lock->l_flags & LDLM_FL_CANCELING) {
2104 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2108 /* If this is reply-less callback lock, we cannot replay it, since
2109 * server might have long dropped it, but notification of that event was
2110 * lost by network. (and server granted conflicting lock already) */
2111 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2112 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2113 ldlm_lock_cancel(lock);
2117 * If granted mode matches the requested mode, this lock is granted.
2119 * If they differ, but we have a granted mode, then we were granted
2120 * one mode and now want another: ergo, converting.
2122 * If we haven't been granted anything and are on a resource list,
2123 * then we're blocked/waiting.
2125 * If we haven't been granted anything and we're NOT on a resource list,
2126 * then we haven't got a reply yet and don't have a known disposition.
2127 * This happens whenever a lock enqueue is the request that triggers
2130 if (lock->l_granted_mode == lock->l_req_mode)
2131 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2132 else if (lock->l_granted_mode)
2133 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2134 else if (!cfs_list_empty(&lock->l_res_link))
2135 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2137 flags = LDLM_FL_REPLAY;
2139 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2140 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2144 /* We're part of recovery, so don't wait for it. */
2145 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2147 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2148 ldlm_lock2desc(lock, &body->lock_desc);
2149 body->lock_flags = ldlm_flags_to_wire(flags);
2151 ldlm_lock2handle(lock, &body->lock_handle[0]);
2152 if (lock->l_lvb_len != 0) {
2153 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2154 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2157 ptlrpc_request_set_replen(req);
2158 /* notify the server we've replayed all requests.
2159 * also, we mark the request to be put on a dedicated
2160 * queue to be processed after all request replayes.
2162 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2164 LDLM_DEBUG(lock, "replaying lock:");
2166 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2167 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2168 aa = ptlrpc_req_async_args(req);
2169 aa->lock_handle = body->lock_handle[0];
2170 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2171 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2177 * Cancel as many unused locks as possible before replay. since we are
2178 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2181 * Called only in recovery before replaying locks. there is no need to
2182 * replay locks that are unused. since the clients may hold thousands of
2183 * cached unused locks, dropping the unused locks can greatly reduce the
2184 * load on the servers at recovery time.
2186 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2189 CFS_LIST_HEAD(cancels);
2191 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2192 "replay for namespace %s (%d)\n",
2193 ldlm_ns_name(ns), ns->ns_nr_unused);
2195 /* We don't need to care whether or not LRU resize is enabled
2196 * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2197 * count parameter */
2198 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2199 LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2201 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2202 canceled, ldlm_ns_name(ns));
2205 int ldlm_replay_locks(struct obd_import *imp)
2207 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2208 CFS_LIST_HEAD(list);
2209 struct ldlm_lock *lock, *next;
2214 LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0);
2216 /* don't replay locks if import failed recovery */
2217 if (imp->imp_vbr_failed)
2220 /* ensure this doesn't fall to 0 before all have been queued */
2221 cfs_atomic_inc(&imp->imp_replay_inflight);
2223 if (ldlm_cancel_unused_locks_before_replay)
2224 ldlm_cancel_unused_locks_for_replay(ns);
2226 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2228 cfs_list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2229 cfs_list_del_init(&lock->l_pending_chain);
2231 LDLM_LOCK_RELEASE(lock);
2232 continue; /* or try to do the rest? */
2234 rc = replay_one_lock(imp, lock);
2235 LDLM_LOCK_RELEASE(lock);
2238 cfs_atomic_dec(&imp->imp_replay_inflight);
2242 EXPORT_SYMBOL(ldlm_replay_locks);