1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
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);
112 /* We use the same basis for both server side and client side functions
113 from a single node. */
114 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
116 int timeout = at_get(ldlm_lock_to_ns_at(lock));
118 return obd_timeout / 2;
119 /* Since these are non-updating timeouts, we should be conservative.
120 It would be nice to have some kind of "early reply" mechanism for
121 lock callbacks too... */
122 timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */
123 return max(timeout, ldlm_enqueue_min);
125 EXPORT_SYMBOL(ldlm_get_enq_timeout);
128 * Helper function for ldlm_completion_ast(), updating timings when lock is
131 static int ldlm_completion_tail(struct ldlm_lock *lock)
136 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
137 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
140 delay = cfs_time_sub(cfs_time_current_sec(),
141 lock->l_last_activity);
142 LDLM_DEBUG(lock, "client-side enqueue: granted after "
143 CFS_DURATION_T"s", delay);
145 /* Update our time estimate */
146 at_measured(ldlm_lock_to_ns_at(lock),
154 * Implementation of ->l_completion_ast() for a client, that doesn't wait
155 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
156 * other threads that cannot block for long.
158 int ldlm_completion_ast_async(struct ldlm_lock *lock, int flags, void *data)
162 if (flags == LDLM_FL_WAIT_NOREPROC) {
163 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
167 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
168 LDLM_FL_BLOCK_CONV))) {
169 cfs_waitq_signal(&lock->l_waitq);
170 RETURN(ldlm_completion_tail(lock));
173 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
175 ldlm_lock_dump(D_OTHER, lock, 0);
176 ldlm_reprocess_all(lock->l_resource);
181 * Client side LDLM "completion" AST. This is called in several cases:
183 * - when a reply to an ENQUEUE rpc is received from the server
184 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
185 * this point (determined by flags);
187 * - when LDLM_CP_CALLBACK rpc comes to client to notify it that lock has
190 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
193 * - to force all locks when resource is destroyed (cleanup_resource());
195 * - during lock conversion (not used currently).
197 * If lock is not granted in the first case, this function waits until second
198 * or penultimate cases happen in some other thread.
201 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
203 /* XXX ALLOCATE - 160 bytes */
204 struct lock_wait_data lwd;
205 struct obd_device *obd;
206 struct obd_import *imp = NULL;
207 struct l_wait_info lwi;
212 if (flags == LDLM_FL_WAIT_NOREPROC) {
213 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
217 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
218 LDLM_FL_BLOCK_CONV))) {
219 cfs_waitq_signal(&lock->l_waitq);
223 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
225 ldlm_lock_dump(D_OTHER, lock, 0);
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 cfs_spin_lock(&imp->imp_lock);
254 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
255 cfs_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));
279 * A helper to build a blocking ast function
281 * Perform a common operation for blocking asts:
282 * defferred lock cancellation.
284 * \param lock the lock blocking or canceling ast was called on
286 * \see mdt_blocking_ast
287 * \see ldlm_blocking_ast
289 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
294 lock->l_flags |= LDLM_FL_CBPENDING;
295 do_ast = (!lock->l_readers && !lock->l_writers);
296 unlock_res_and_lock(lock);
299 struct lustre_handle lockh;
302 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
303 ldlm_lock2handle(lock, &lockh);
304 rc = ldlm_cli_cancel(&lockh);
306 CERROR("ldlm_cli_cancel: %d\n", rc);
308 LDLM_DEBUG(lock, "Lock still has references, will be "
315 * Server blocking AST
317 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
320 * \param lock the lock which blocks a request or cancelling lock
323 * \param flag indicates whether this cancelling or blocking callback
325 * \see ldlm_blocking_ast_nocheck
327 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
328 void *data, int flag)
332 if (flag == LDLM_CB_CANCELING) {
333 /* Don't need to do anything here. */
337 lock_res_and_lock(lock);
338 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
339 * that ldlm_blocking_ast is called just before intent_policy method
340 * takes the ns_lock, then by the time we get the lock, we might not
341 * be the correct blocking function anymore. So check, and return
343 if (lock->l_blocking_ast != ldlm_blocking_ast) {
344 unlock_res_and_lock(lock);
347 RETURN(ldlm_blocking_ast_nocheck(lock));
351 * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
352 * comment in filter_intent_policy() on why you may need this.
354 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
357 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
358 * that is rather subtle: with OST-side locking, it may so happen that
359 * _all_ extent locks are held by the OST. If client wants to obtain
360 * current file size it calls ll{,u}_glimpse_size(), and (as locks are
361 * on the server), dummy glimpse callback fires and does
362 * nothing. Client still receives correct file size due to the
363 * following fragment in filter_intent_policy():
365 * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
366 * if (rc != 0 && res->lr_namespace->ns_lvbo &&
367 * res->lr_namespace->ns_lvbo->lvbo_update) {
368 * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
371 * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
372 * returns correct file size to the client.
374 return -ELDLM_NO_LOCK_DATA;
377 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
378 const struct ldlm_res_id *res_id,
379 ldlm_type_t type, ldlm_policy_data_t *policy,
380 ldlm_mode_t mode, int *flags,
381 ldlm_blocking_callback blocking,
382 ldlm_completion_callback completion,
383 ldlm_glimpse_callback glimpse,
384 void *data, __u32 lvb_len,
385 const __u64 *client_cookie,
386 struct lustre_handle *lockh)
388 struct ldlm_lock *lock;
390 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
391 .lcs_blocking = blocking,
392 .lcs_glimpse = glimpse,
396 LASSERT(!(*flags & LDLM_FL_REPLAY));
397 if (unlikely(ns_is_client(ns))) {
398 CERROR("Trying to enqueue local lock in a shadow namespace\n");
402 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len);
404 GOTO(out_nolock, err = -ENOMEM);
406 ldlm_lock_addref_internal(lock, mode);
407 ldlm_lock2handle(lock, lockh);
408 lock_res_and_lock(lock);
409 lock->l_flags |= LDLM_FL_LOCAL;
410 if (*flags & LDLM_FL_ATOMIC_CB)
411 lock->l_flags |= LDLM_FL_ATOMIC_CB;
412 unlock_res_and_lock(lock);
414 lock->l_policy_data = *policy;
415 if (client_cookie != NULL)
416 lock->l_client_cookie = *client_cookie;
417 if (type == LDLM_EXTENT)
418 lock->l_req_extent = policy->l_extent;
420 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
421 if (unlikely(err != ELDLM_OK))
425 *policy = lock->l_policy_data;
427 if (lock->l_completion_ast)
428 lock->l_completion_ast(lock, *flags, NULL);
430 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
433 LDLM_LOCK_RELEASE(lock);
438 static void failed_lock_cleanup(struct ldlm_namespace *ns,
439 struct ldlm_lock *lock, int mode)
443 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
444 lock_res_and_lock(lock);
445 /* Check that lock is not granted or failed, we might race. */
446 if ((lock->l_req_mode != lock->l_granted_mode) &&
447 !(lock->l_flags & LDLM_FL_FAILED)) {
448 /* Make sure that this lock will not be found by raced
449 * bl_ast and -EINVAL reply is sent to server anyways.
451 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
452 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
455 unlock_res_and_lock(lock);
459 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
460 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
462 LDLM_DEBUG(lock, "lock was granted or failed in race");
464 ldlm_lock_decref_internal(lock, mode);
466 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
467 * from llite/file.c/ll_file_flock(). */
468 /* This code makes for the fact that we do not have blocking handler on
469 * a client for flock locks. As such this is the place where we must
470 * completely kill failed locks. (interrupted and those that
471 * were waiting to be granted when server evicted us. */
472 if (lock->l_resource->lr_type == LDLM_FLOCK) {
473 lock_res_and_lock(lock);
474 ldlm_resource_unlink_lock(lock);
475 ldlm_lock_destroy_nolock(lock);
476 unlock_res_and_lock(lock);
480 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
481 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
482 int *flags, void *lvb, __u32 lvb_len,
483 struct lustre_handle *lockh,int rc)
485 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
486 int is_replay = *flags & LDLM_FL_REPLAY;
487 struct ldlm_lock *lock;
488 struct ldlm_reply *reply;
489 int cleanup_phase = 1;
492 lock = ldlm_handle2lock(lockh);
493 /* ldlm_cli_enqueue is holding a reference on this lock. */
495 LASSERT(type == LDLM_FLOCK);
499 if (rc != ELDLM_OK) {
501 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
502 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
503 if (rc == ELDLM_LOCK_ABORTED) {
504 /* Before we return, swab the reply */
505 reply = req_capsule_server_get(&req->rq_pill,
510 struct ost_lvb *tmplvb;
512 req_capsule_set_size(&req->rq_pill,
513 &RMF_DLM_LVB, RCL_SERVER,
515 tmplvb = req_capsule_server_get(&req->rq_pill,
518 GOTO(cleanup, rc = -EPROTO);
520 memcpy(lvb, tmplvb, lvb_len);
526 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
528 GOTO(cleanup, rc = -EPROTO);
530 /* lock enqueued on the server */
533 lock_res_and_lock(lock);
534 /* Key change rehash lock in per-export hash with new key */
535 if (exp->exp_lock_hash) {
536 cfs_hash_rehash_key(exp->exp_lock_hash,
537 &lock->l_remote_handle,
541 lock->l_remote_handle = reply->lock_handle;
544 *flags = reply->lock_flags;
545 lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
546 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
547 * to wait with no timeout as well */
548 lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
549 unlock_res_and_lock(lock);
551 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
552 lock, reply->lock_handle.cookie, *flags);
554 /* If enqueue returned a blocked lock but the completion handler has
555 * already run, then it fixed up the resource and we don't need to do it
557 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
558 int newmode = reply->lock_desc.l_req_mode;
560 if (newmode && newmode != lock->l_req_mode) {
561 LDLM_DEBUG(lock, "server returned different mode %s",
562 ldlm_lockname[newmode]);
563 lock->l_req_mode = newmode;
566 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
567 lock->l_resource->lr_name.name,
568 sizeof(struct ldlm_res_id))) {
569 CDEBUG(D_INFO, "remote intent success, locking "
570 "(%ld,%ld,%ld) instead of "
572 (long)reply->lock_desc.l_resource.lr_name.name[0],
573 (long)reply->lock_desc.l_resource.lr_name.name[1],
574 (long)reply->lock_desc.l_resource.lr_name.name[2],
575 (long)lock->l_resource->lr_name.name[0],
576 (long)lock->l_resource->lr_name.name[1],
577 (long)lock->l_resource->lr_name.name[2]);
579 rc = ldlm_lock_change_resource(ns, lock,
580 &reply->lock_desc.l_resource.lr_name);
581 if (rc || lock->l_resource == NULL)
582 GOTO(cleanup, rc = -ENOMEM);
583 LDLM_DEBUG(lock, "client-side enqueue, new resource");
586 if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
588 lock->l_policy_data =
589 reply->lock_desc.l_policy_data;
590 if (type != LDLM_PLAIN)
591 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
594 if ((*flags) & LDLM_FL_AST_SENT ||
595 /* Cancel extent locks as soon as possible on a liblustre client,
596 * because it cannot handle asynchronous ASTs robustly (see
598 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
599 lock_res_and_lock(lock);
600 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
601 unlock_res_and_lock(lock);
602 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
605 /* If the lock has already been granted by a completion AST, don't
606 * clobber the LVB with an older one. */
607 if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
610 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
612 tmplvb = req_capsule_server_get(&req->rq_pill,
615 GOTO(cleanup, rc = -EPROTO);
616 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
620 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
621 if (lock->l_completion_ast != NULL) {
622 int err = lock->l_completion_ast(lock, *flags, NULL);
630 if (lvb_len && lvb != NULL) {
631 /* Copy the LVB here, and not earlier, because the completion
632 * AST (if any) can override what we got in the reply */
633 memcpy(lvb, lock->l_lvb_data, lvb_len);
636 LDLM_DEBUG(lock, "client-side enqueue END");
639 if (cleanup_phase == 1 && rc)
640 failed_lock_cleanup(ns, lock, mode);
641 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
643 LDLM_LOCK_RELEASE(lock);
647 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
648 * a single page on the send/receive side. XXX: 512 should be changed
649 * to more adequate value. */
650 static inline int ldlm_req_handles_avail(int req_size, int off)
654 avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size;
655 if (likely(avail >= 0))
656 avail /= (int)sizeof(struct lustre_handle);
659 avail += LDLM_LOCKREQ_HANDLES - off;
664 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
665 enum req_location loc,
668 int size = req_capsule_msg_size(pill, loc);
669 return ldlm_req_handles_avail(size, off);
672 static inline int ldlm_format_handles_avail(struct obd_import *imp,
673 const struct req_format *fmt,
674 enum req_location loc, int off)
676 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
677 return ldlm_req_handles_avail(size, off);
680 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
681 * @count locks in @cancels. */
682 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
683 int version, int opc, int canceloff,
684 cfs_list_t *cancels, int count)
686 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
687 struct req_capsule *pill = &req->rq_pill;
688 struct ldlm_request *dlm = NULL;
689 int flags, avail, to_free, bufcount, pack = 0;
696 if (exp_connect_cancelset(exp)) {
697 /* Estimate the amount of available space in the request. */
698 bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT);
699 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
701 flags = ns_connect_lru_resize(ns) ?
702 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
703 to_free = !ns_connect_lru_resize(ns) &&
704 opc == LDLM_ENQUEUE ? 1 : 0;
706 /* Cancel lru locks here _only_ if the server supports
707 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
708 * rpc, what will make us slower. */
710 count += ldlm_cancel_lru_local(ns, cancels, to_free,
711 avail - count, 0, flags);
716 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
717 ldlm_request_bufsize(pack, opc));
720 rc = ptlrpc_request_pack(req, version, opc);
722 ldlm_lock_list_put(cancels, l_bl_ast, count);
726 if (exp_connect_cancelset(exp)) {
728 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
730 /* Skip first lock handler in ldlm_request_pack(),
731 * this method will incrment @lock_count according
732 * to the lock handle amount actually written to
734 dlm->lock_count = canceloff;
736 /* Pack into the request @pack lock handles. */
737 ldlm_cli_cancel_list(cancels, pack, req, 0);
738 /* Prepare and send separate cancel rpc for others. */
739 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
741 ldlm_lock_list_put(cancels, l_bl_ast, count);
746 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
747 cfs_list_t *cancels, int count)
749 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
750 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
753 /* If a request has some specific initialisation it is passed in @reqp,
754 * otherwise it is created in ldlm_cli_enqueue.
756 * Supports sync and async requests, pass @async flag accordingly. If a
757 * request was created in ldlm_cli_enqueue and it is the async request,
758 * pass it to the caller in @reqp. */
759 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
760 struct ldlm_enqueue_info *einfo,
761 const struct ldlm_res_id *res_id,
762 ldlm_policy_data_t const *policy, int *flags,
763 void *lvb, __u32 lvb_len, struct lustre_handle *lockh,
766 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
767 struct ldlm_lock *lock;
768 struct ldlm_request *body;
769 int is_replay = *flags & LDLM_FL_REPLAY;
770 int req_passed_in = 1;
772 struct ptlrpc_request *req;
775 LASSERT(exp != NULL);
777 /* If we're replaying this lock, just check some invariants.
778 * If we're creating a new lock, get everything all setup nice. */
780 lock = ldlm_handle2lock_long(lockh, 0);
781 LASSERT(lock != NULL);
782 LDLM_DEBUG(lock, "client-side enqueue START");
783 LASSERT(exp == lock->l_conn_export);
785 const struct ldlm_callback_suite cbs = {
786 .lcs_completion = einfo->ei_cb_cp,
787 .lcs_blocking = einfo->ei_cb_bl,
788 .lcs_glimpse = einfo->ei_cb_gl,
789 .lcs_weigh = einfo->ei_cb_wg
791 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
792 einfo->ei_mode, &cbs, einfo->ei_cbdata,
796 /* for the local lock, add the reference */
797 ldlm_lock_addref_internal(lock, einfo->ei_mode);
798 ldlm_lock2handle(lock, lockh);
799 if (policy != NULL) {
800 /* INODEBITS_INTEROP: If the server does not support
801 * inodebits, we will request a plain lock in the
802 * descriptor (ldlm_lock2desc() below) but use an
803 * inodebits lock internally with both bits set.
805 if (einfo->ei_type == LDLM_IBITS &&
806 !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
807 lock->l_policy_data.l_inodebits.bits =
808 MDS_INODELOCK_LOOKUP |
809 MDS_INODELOCK_UPDATE;
811 lock->l_policy_data = *policy;
814 if (einfo->ei_type == LDLM_EXTENT)
815 lock->l_req_extent = policy->l_extent;
816 LDLM_DEBUG(lock, "client-side enqueue START");
819 /* lock not sent to server yet */
821 if (reqp == NULL || *reqp == NULL) {
822 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
827 failed_lock_cleanup(ns, lock, einfo->ei_mode);
828 LDLM_LOCK_RELEASE(lock);
838 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
840 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
841 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
844 lock->l_conn_export = exp;
845 lock->l_export = NULL;
846 lock->l_blocking_ast = einfo->ei_cb_bl;
848 /* Dump lock data into the request buffer */
849 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
850 ldlm_lock2desc(lock, &body->lock_desc);
851 body->lock_flags = *flags;
852 body->lock_handle[0] = *lockh;
854 /* Continue as normal. */
855 if (!req_passed_in) {
857 req_capsule_extend(&req->rq_pill,
858 &RQF_LDLM_ENQUEUE_LVB);
859 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
860 RCL_SERVER, lvb_len);
862 ptlrpc_request_set_replen(req);
866 * Liblustre client doesn't get extent locks, except for O_APPEND case
867 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
868 * [i_size, OBD_OBJECT_EOF] lock is taken.
870 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
871 policy->l_extent.end == OBD_OBJECT_EOF));
874 LASSERT(reqp != NULL);
878 LDLM_DEBUG(lock, "sending request");
880 rc = ptlrpc_queue_wait(req);
882 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
883 einfo->ei_mode, flags, lvb, lvb_len,
886 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
887 * one reference that we took */
889 LDLM_LOCK_RELEASE(lock);
893 if (!req_passed_in && req != NULL) {
894 ptlrpc_req_finished(req);
902 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
905 struct ldlm_resource *res;
908 if (ns_is_client(ldlm_lock_to_ns(lock))) {
909 CERROR("Trying to cancel local lock\n");
912 LDLM_DEBUG(lock, "client-side local convert");
914 res = ldlm_lock_convert(lock, new_mode, flags);
916 ldlm_reprocess_all(res);
921 LDLM_DEBUG(lock, "client-side local convert handler END");
926 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
927 * conversion of locks which are on the waiting or converting queue */
928 /* Caller of this code is supposed to take care of lock readers/writers
930 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
932 struct ldlm_request *body;
933 struct ldlm_reply *reply;
934 struct ldlm_lock *lock;
935 struct ldlm_resource *res;
936 struct ptlrpc_request *req;
940 lock = ldlm_handle2lock(lockh);
947 if (lock->l_conn_export == NULL)
948 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
950 LDLM_DEBUG(lock, "client-side convert");
952 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
953 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
960 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
961 body->lock_handle[0] = lock->l_remote_handle;
963 body->lock_desc.l_req_mode = new_mode;
964 body->lock_flags = *flags;
967 ptlrpc_request_set_replen(req);
968 rc = ptlrpc_queue_wait(req);
972 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
974 GOTO(out, rc = -EPROTO);
977 GOTO(out, rc = req->rq_status);
979 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
981 ldlm_reprocess_all(res);
982 /* Go to sleep until the lock is granted. */
983 /* FIXME: or cancelled. */
984 if (lock->l_completion_ast) {
985 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
996 ptlrpc_req_finished(req);
1000 /* Cancel locks locally.
1002 * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
1003 * LDLM_FL_CANCELING otherwise;
1004 * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
1005 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
1007 int rc = LDLM_FL_LOCAL_ONLY;
1010 if (lock->l_conn_export) {
1013 LDLM_DEBUG(lock, "client-side cancel");
1014 /* Set this flag to prevent others from getting new references*/
1015 lock_res_and_lock(lock);
1016 lock->l_flags |= LDLM_FL_CBPENDING;
1017 local_only = (lock->l_flags &
1018 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1019 ldlm_cancel_callback(lock);
1020 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
1021 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1022 unlock_res_and_lock(lock);
1025 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1027 rc = LDLM_FL_LOCAL_ONLY;
1029 ldlm_lock_cancel(lock);
1031 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1032 LDLM_ERROR(lock, "Trying to cancel local lock");
1035 LDLM_DEBUG(lock, "server-side local cancel");
1036 ldlm_lock_cancel(lock);
1037 ldlm_reprocess_all(lock->l_resource);
1043 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
1044 of the request @req. */
1045 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1046 cfs_list_t *head, int count)
1048 struct ldlm_request *dlm;
1049 struct ldlm_lock *lock;
1050 int max, packed = 0;
1053 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1054 LASSERT(dlm != NULL);
1056 /* Check the room in the request buffer. */
1057 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1058 sizeof(struct ldlm_request);
1059 max /= sizeof(struct lustre_handle);
1060 max += LDLM_LOCKREQ_HANDLES;
1061 LASSERT(max >= dlm->lock_count + count);
1063 /* XXX: it would be better to pack lock handles grouped by resource.
1064 * so that the server cancel would call filter_lvbo_update() less
1066 cfs_list_for_each_entry(lock, head, l_bl_ast) {
1069 LASSERT(lock->l_conn_export);
1070 /* Pack the lock handle to the given request buffer. */
1071 LDLM_DEBUG(lock, "packing");
1072 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1075 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1079 /* Prepare and send a batched cancel rpc, it will include count lock handles
1080 * of locks given in @head. */
1081 int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels,
1082 int count, ldlm_cancel_flags_t flags)
1084 struct ptlrpc_request *req = NULL;
1085 struct obd_import *imp;
1090 LASSERT(exp != NULL);
1093 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, obd_fail_val);
1095 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1098 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1099 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1106 imp = class_exp2cliimp(exp);
1107 if (imp == NULL || imp->imp_invalid) {
1109 "skipping cancel on invalid import %p\n", imp);
1113 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1115 GOTO(out, rc = -ENOMEM);
1117 bufcount = req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1118 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1119 ldlm_request_bufsize(count, LDLM_CANCEL));
1121 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1123 ptlrpc_request_free(req);
1126 req->rq_no_resend = 1;
1127 req->rq_no_delay = 1;
1129 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1130 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1131 ptlrpc_at_set_req_timeout(req);
1133 ldlm_cancel_pack(req, cancels, count);
1135 ptlrpc_request_set_replen(req);
1136 if (flags & LCF_ASYNC) {
1137 ptlrpcd_add_req(req, PSCOPE_OTHER);
1141 rc = ptlrpc_queue_wait(req);
1144 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1145 "out of sync -- not fatal\n",
1146 libcfs_nid2str(req->rq_import->
1147 imp_connection->c_peer.nid));
1149 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1150 req->rq_import_generation == imp->imp_generation) {
1151 ptlrpc_req_finished(req);
1153 } else if (rc != ELDLM_OK) {
1154 CERROR("Got rc %d from cancel RPC: canceling "
1162 ptlrpc_req_finished(req);
1165 return sent ? sent : rc;
1168 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1170 LASSERT(imp != NULL);
1171 return &imp->imp_obd->obd_namespace->ns_pool;
1175 * Update client's obd pool related fields with new SLV and Limit from \a req.
1177 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1179 struct obd_device *obd;
1180 __u64 old_slv, new_slv;
1183 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1184 !imp_connect_lru_resize(req->rq_import)))
1187 * Do nothing for corner cases.
1193 * In some cases RPC may contain slv and limit zeroed out. This is
1194 * the case when server does not support lru resize feature. This is
1195 * also possible in some recovery cases when server side reqs have no
1196 * ref to obd export and thus access to server side namespace is no
1199 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1200 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1201 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1202 "(SLV: "LPU64", Limit: %u)",
1203 lustre_msg_get_slv(req->rq_repmsg),
1204 lustre_msg_get_limit(req->rq_repmsg));
1208 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1209 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1210 obd = req->rq_import->imp_obd;
1213 * Set new SLV and Limit to obd fields to make accessible for pool
1214 * thread. We do not access obd_namespace and pool directly here
1215 * as there is no reliable way to make sure that they are still
1216 * alive in cleanup time. Evil races are possible which may cause
1217 * oops in that time.
1219 cfs_write_lock(&obd->obd_pool_lock);
1220 old_slv = obd->obd_pool_slv;
1221 obd->obd_pool_slv = new_slv;
1222 obd->obd_pool_limit = new_limit;
1223 cfs_write_unlock(&obd->obd_pool_lock);
1227 EXPORT_SYMBOL(ldlm_cli_update_pool);
1229 int ldlm_cli_cancel(struct lustre_handle *lockh)
1231 struct obd_export *exp;
1232 int avail, flags, count = 1, rc = 0;
1233 struct ldlm_namespace *ns;
1234 struct ldlm_lock *lock;
1235 CFS_LIST_HEAD(cancels);
1238 /* concurrent cancels on the same handle can happen */
1239 lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1241 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1245 rc = ldlm_cli_cancel_local(lock);
1246 if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) {
1247 LDLM_LOCK_RELEASE(lock);
1248 RETURN(rc < 0 ? rc : 0);
1250 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1251 * rpc which goes to canceld portal, so we can cancel other lru locks
1252 * here and send them all as one LDLM_CANCEL rpc. */
1253 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1254 cfs_list_add(&lock->l_bl_ast, &cancels);
1256 exp = lock->l_conn_export;
1257 if (exp_connect_cancelset(exp)) {
1258 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1263 ns = ldlm_lock_to_ns(lock);
1264 flags = ns_connect_lru_resize(ns) ?
1265 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1266 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1269 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1273 /* XXX until we will have compound requests and can cut cancels from generic rpc
1274 * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */
1275 int ldlm_cli_cancel_list_local(cfs_list_t *cancels, int count,
1276 ldlm_cancel_flags_t flags)
1278 CFS_LIST_HEAD(head);
1279 struct ldlm_lock *lock, *next;
1280 int left = 0, bl_ast = 0, rc;
1283 cfs_list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1287 if (flags & LCF_LOCAL) {
1288 rc = LDLM_FL_LOCAL_ONLY;
1289 ldlm_lock_cancel(lock);
1291 rc = ldlm_cli_cancel_local(lock);
1293 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1294 LDLM_DEBUG(lock, "Cancel lock separately");
1295 cfs_list_del_init(&lock->l_bl_ast);
1296 cfs_list_add(&lock->l_bl_ast, &head);
1300 if (rc == LDLM_FL_LOCAL_ONLY) {
1301 /* CANCEL RPC should not be sent to server. */
1302 cfs_list_del_init(&lock->l_bl_ast);
1303 LDLM_LOCK_RELEASE(lock);
1310 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1317 * Cancel as many locks as possible w/o sending any rpcs (e.g. to write back
1318 * dirty data, to close a file, ...) or waiting for any rpcs in-flight (e.g.
1319 * readahead requests, ...)
1321 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1322 struct ldlm_lock *lock,
1323 int unused, int added,
1326 ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1327 ldlm_cancel_for_recovery cb = ns->ns_cancel_for_recovery;
1328 lock_res_and_lock(lock);
1330 /* don't check added & count since we want to process all locks
1331 * from unused list */
1332 switch (lock->l_resource->lr_type) {
1338 result = LDLM_POLICY_SKIP_LOCK;
1339 lock->l_flags |= LDLM_FL_SKIPPED;
1343 unlock_res_and_lock(lock);
1348 * Callback function for lru-resize policy. Makes decision whether to keep
1349 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1350 * \a added and number of locks to be preferably canceled \a count.
1352 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1354 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1356 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1357 struct ldlm_lock *lock,
1358 int unused, int added,
1361 cfs_time_t cur = cfs_time_current();
1362 struct ldlm_pool *pl = &ns->ns_pool;
1367 * Stop lru processing when we reached passed @count or checked all
1370 if (count && added >= count)
1371 return LDLM_POLICY_KEEP_LOCK;
1373 slv = ldlm_pool_get_slv(pl);
1374 lvf = ldlm_pool_get_lvf(pl);
1375 la = cfs_duration_sec(cfs_time_sub(cur,
1376 lock->l_last_used));
1379 * Stop when slv is not yet come from server or lv is smaller than
1382 lv = lvf * la * unused;
1385 * Inform pool about current CLV to see it via proc.
1387 ldlm_pool_set_clv(pl, lv);
1388 return (slv == 0 || lv < slv) ?
1389 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1393 * Callback function for proc used policy. Makes decision whether to keep
1394 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1395 * \a added and number of locks to be preferably canceled \a count.
1397 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1399 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1401 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1402 struct ldlm_lock *lock,
1403 int unused, int added,
1407 * Stop lru processing when we reached passed @count or checked all
1410 return (added >= count) ?
1411 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1415 * Callback function for aged policy. Makes decision whether to keep
1416 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1417 * \a added and number of locks to be preferably canceled \a count.
1419 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1421 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1423 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1424 struct ldlm_lock *lock,
1425 int unused, int added,
1429 * Stop lru processing if young lock is found and we reached passed
1432 return ((added >= count) &&
1433 cfs_time_before(cfs_time_current(),
1434 cfs_time_add(lock->l_last_used,
1436 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1440 * Callback function for default policy. Makes decision whether to keep
1441 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1442 * \a added and number of locks to be preferably canceled \a count.
1444 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1446 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1448 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1449 struct ldlm_lock *lock,
1450 int unused, int added,
1454 * Stop lru processing when we reached passed @count or checked all
1457 return (added >= count) ?
1458 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1461 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1462 struct ldlm_lock *, int,
1465 static ldlm_cancel_lru_policy_t
1466 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1468 if (flags & LDLM_CANCEL_NO_WAIT)
1469 return ldlm_cancel_no_wait_policy;
1471 if (ns_connect_lru_resize(ns)) {
1472 if (flags & LDLM_CANCEL_SHRINK)
1473 /* We kill passed number of old locks. */
1474 return ldlm_cancel_passed_policy;
1475 else if (flags & LDLM_CANCEL_LRUR)
1476 return ldlm_cancel_lrur_policy;
1477 else if (flags & LDLM_CANCEL_PASSED)
1478 return ldlm_cancel_passed_policy;
1480 if (flags & LDLM_CANCEL_AGED)
1481 return ldlm_cancel_aged_policy;
1484 return ldlm_cancel_default_policy;
1487 /* - Free space in lru for @count new locks,
1488 * redundant unused locks are canceled locally;
1489 * - also cancel locally unused aged locks;
1490 * - do not cancel more than @max locks;
1491 * - GET the found locks and add them into the @cancels list.
1493 * A client lock can be added to the l_bl_ast list only when it is
1494 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
1495 * There are the following use cases: ldlm_cancel_resource_local(),
1496 * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
1497 * flag properly. As any attempt to cancel a lock rely on this flag,
1498 * l_bl_ast list is accessed later without any special locking.
1500 * Calling policies for enabled lru resize:
1501 * ----------------------------------------
1502 * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to
1503 * cancel not more than @count locks;
1505 * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at
1506 * the beginning of lru list);
1508 * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
1509 * memory pressre policy function;
1511 * flags & LDLM_CANCEL_AGED - cancel alocks according to "aged policy".
1513 * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1514 * (typically before replaying locks) w/o
1515 * sending any rpcs or waiting for any
1516 * outstanding rpc to complete.
1518 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels,
1519 int count, int max, int flags)
1521 ldlm_cancel_lru_policy_t pf;
1522 struct ldlm_lock *lock, *next;
1523 int added = 0, unused, remained;
1526 cfs_spin_lock(&ns->ns_unused_lock);
1527 unused = ns->ns_nr_unused;
1530 if (!ns_connect_lru_resize(ns))
1531 count += unused - ns->ns_max_unused;
1533 pf = ldlm_cancel_lru_policy(ns, flags);
1534 LASSERT(pf != NULL);
1536 while (!cfs_list_empty(&ns->ns_unused_list)) {
1537 ldlm_policy_res_t result;
1539 /* all unused locks */
1540 if (remained-- <= 0)
1543 /* For any flags, stop scanning if @max is reached. */
1544 if (max && added >= max)
1547 cfs_list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1549 /* No locks which got blocking requests. */
1550 LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1552 if (flags & LDLM_CANCEL_NO_WAIT &&
1553 lock->l_flags & LDLM_FL_SKIPPED)
1554 /* already processed */
1557 /* Somebody is already doing CANCEL. No need in this
1558 * lock in lru, do not traverse it again. */
1559 if (!(lock->l_flags & LDLM_FL_CANCELING))
1562 ldlm_lock_remove_from_lru_nolock(lock);
1564 if (&lock->l_lru == &ns->ns_unused_list)
1567 LDLM_LOCK_GET(lock);
1568 cfs_spin_unlock(&ns->ns_unused_lock);
1569 lu_ref_add(&lock->l_reference, __FUNCTION__, cfs_current());
1571 /* Pass the lock through the policy filter and see if it
1572 * should stay in lru.
1574 * Even for shrinker policy we stop scanning if
1575 * we find a lock that should stay in the cache.
1576 * We should take into account lock age anyway
1577 * as new lock even if it is small of weight is
1578 * valuable resource.
1580 * That is, for shrinker policy we drop only
1581 * old locks, but additionally chose them by
1582 * their weight. Big extent locks will stay in
1584 result = pf(ns, lock, unused, added, count);
1585 if (result == LDLM_POLICY_KEEP_LOCK) {
1586 lu_ref_del(&lock->l_reference,
1587 __FUNCTION__, cfs_current());
1588 LDLM_LOCK_RELEASE(lock);
1589 cfs_spin_lock(&ns->ns_unused_lock);
1592 if (result == LDLM_POLICY_SKIP_LOCK) {
1593 lu_ref_del(&lock->l_reference,
1594 __FUNCTION__, cfs_current());
1595 LDLM_LOCK_RELEASE(lock);
1596 cfs_spin_lock(&ns->ns_unused_lock);
1600 lock_res_and_lock(lock);
1601 /* Check flags again under the lock. */
1602 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1603 (ldlm_lock_remove_from_lru(lock) == 0)) {
1604 /* other thread is removing lock from lru or
1605 * somebody is already doing CANCEL or
1606 * there is a blocking request which will send
1607 * cancel by itseft or the lock is matched
1608 * is already not unused. */
1609 unlock_res_and_lock(lock);
1610 lu_ref_del(&lock->l_reference,
1611 __FUNCTION__, cfs_current());
1612 LDLM_LOCK_RELEASE(lock);
1613 cfs_spin_lock(&ns->ns_unused_lock);
1616 LASSERT(!lock->l_readers && !lock->l_writers);
1618 /* If we have chosen to cancel this lock voluntarily, we
1619 * better send cancel notification to server, so that it
1620 * frees appropriate state. This might lead to a race
1621 * where while we are doing cancel here, server is also
1622 * silently cancelling this lock. */
1623 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1625 /* Setting the CBPENDING flag is a little misleading,
1626 * but prevents an important race; namely, once
1627 * CBPENDING is set, the lock can accumulate no more
1628 * readers/writers. Since readers and writers are
1629 * already zero here, ldlm_lock_decref() won't see
1630 * this flag and call l_blocking_ast */
1631 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1633 /* We can't re-add to l_lru as it confuses the
1634 * refcounting in ldlm_lock_remove_from_lru() if an AST
1635 * arrives after we drop ns_lock below. We use l_bl_ast
1636 * and can't use l_pending_chain as it is used both on
1637 * server and client nevertheless bug 5666 says it is
1638 * used only on server */
1639 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1640 cfs_list_add(&lock->l_bl_ast, cancels);
1641 unlock_res_and_lock(lock);
1642 lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current());
1643 cfs_spin_lock(&ns->ns_unused_lock);
1647 cfs_spin_unlock(&ns->ns_unused_lock);
1651 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, cfs_list_t *cancels,
1652 int count, int max, ldlm_cancel_flags_t cancel_flags,
1656 added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1659 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1662 /* when called with LDLM_ASYNC the blocking callback will be handled
1663 * in a thread and this function will return after the thread has been
1664 * asked to call the callback. when called with LDLM_SYNC the blocking
1665 * callback will be performed in this function. */
1666 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t mode,
1669 CFS_LIST_HEAD(cancels);
1674 mode = LDLM_SYNC; /* force to be sync in user space */
1676 /* Just prepare the list of locks, do not actually cancel them yet.
1677 * Locks are cancelled later in a separate thread. */
1678 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1679 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, mode);
1686 /* Find and cancel locally unused locks found on resource, matched to the
1687 * given policy, mode. GET the found locks and add them into the @cancels
1689 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1690 cfs_list_t *cancels,
1691 ldlm_policy_data_t *policy,
1692 ldlm_mode_t mode, int lock_flags,
1693 ldlm_cancel_flags_t cancel_flags, void *opaque)
1695 struct ldlm_lock *lock;
1700 cfs_list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1701 if (opaque != NULL && lock->l_ast_data != opaque) {
1702 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1703 lock->l_ast_data, opaque);
1708 if (lock->l_readers || lock->l_writers)
1711 /* If somebody is already doing CANCEL, or blocking ast came,
1712 * skip this lock. */
1713 if (lock->l_flags & LDLM_FL_BL_AST ||
1714 lock->l_flags & LDLM_FL_CANCELING)
1717 if (lockmode_compat(lock->l_granted_mode, mode))
1720 /* If policy is given and this is IBITS lock, add to list only
1721 * those locks that match by policy. */
1722 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1723 !(lock->l_policy_data.l_inodebits.bits &
1724 policy->l_inodebits.bits))
1727 /* See CBPENDING comment in ldlm_cancel_lru */
1728 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1731 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1732 cfs_list_add(&lock->l_bl_ast, cancels);
1733 LDLM_LOCK_GET(lock);
1738 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
1741 /* If @req is NULL, send CANCEL request to server with handles of locks
1742 * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1743 * separately per lock.
1744 * If @req is not NULL, put handles of locks in @cancels into the request
1745 * buffer at the offset @off.
1746 * Destroy @cancels at the end. */
1747 int ldlm_cli_cancel_list(cfs_list_t *cancels, int count,
1748 struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1750 struct ldlm_lock *lock;
1754 if (cfs_list_empty(cancels) || count == 0)
1757 /* XXX: requests (both batched and not) could be sent in parallel.
1758 * Usually it is enough to have just 1 RPC, but it is possible that
1759 * there are to many locks to be cancelled in LRU or on a resource.
1760 * It would also speed up the case when the server does not support
1763 LASSERT(!cfs_list_empty(cancels));
1764 lock = cfs_list_entry(cancels->next, struct ldlm_lock,
1766 LASSERT(lock->l_conn_export);
1768 if (exp_connect_cancelset(lock->l_conn_export)) {
1771 ldlm_cancel_pack(req, cancels, count);
1773 res = ldlm_cli_cancel_req(lock->l_conn_export,
1777 res = ldlm_cli_cancel_req(lock->l_conn_export,
1782 CERROR("ldlm_cli_cancel_list: %d\n", res);
1787 ldlm_lock_list_put(cancels, l_bl_ast, res);
1789 LASSERT(count == 0);
1793 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1794 const struct ldlm_res_id *res_id,
1795 ldlm_policy_data_t *policy,
1797 ldlm_cancel_flags_t flags,
1800 struct ldlm_resource *res;
1801 CFS_LIST_HEAD(cancels);
1806 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1808 /* This is not a problem. */
1809 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1813 LDLM_RESOURCE_ADDREF(res);
1814 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1815 0, flags | LCF_BL_AST, opaque);
1816 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1818 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1820 LDLM_RESOURCE_DELREF(res);
1821 ldlm_resource_putref(res);
1825 static inline int have_no_nsresource(struct ldlm_namespace *ns)
1827 int no_resource = 0;
1829 cfs_spin_lock(&ns->ns_hash_lock);
1830 if (ns->ns_resources == 0)
1832 cfs_spin_unlock(&ns->ns_hash_lock);
1834 RETURN(no_resource);
1837 /* Cancel all locks on a namespace (or a specific resource, if given)
1838 * that have 0 readers/writers.
1840 * If flags & LCF_LOCAL, throw the locks away without trying
1841 * to notify the server. */
1842 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1843 const struct ldlm_res_id *res_id,
1844 ldlm_cancel_flags_t flags, void *opaque)
1853 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1857 cfs_spin_lock(&ns->ns_hash_lock);
1858 for (i = 0; i < RES_HASH_SIZE; i++) {
1860 tmp = ns->ns_hash[i].next;
1861 while (tmp != &(ns->ns_hash[i])) {
1862 struct ldlm_resource *res;
1865 res = cfs_list_entry(tmp, struct ldlm_resource,
1867 ldlm_resource_getref(res);
1868 cfs_spin_unlock(&ns->ns_hash_lock);
1870 LDLM_RESOURCE_ADDREF(res);
1871 rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name,
1876 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1877 res->lr_name.name[0], rc);
1879 LDLM_RESOURCE_DELREF(res);
1880 cfs_spin_lock(&ns->ns_hash_lock);
1882 ldlm_resource_putref_locked(res);
1885 cfs_spin_unlock(&ns->ns_hash_lock);
1890 /* Lock iterators. */
1892 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1895 cfs_list_t *tmp, *next;
1896 struct ldlm_lock *lock;
1897 int rc = LDLM_ITER_CONTINUE;
1902 RETURN(LDLM_ITER_CONTINUE);
1905 cfs_list_for_each_safe(tmp, next, &res->lr_granted) {
1906 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1908 if (iter(lock, closure) == LDLM_ITER_STOP)
1909 GOTO(out, rc = LDLM_ITER_STOP);
1912 cfs_list_for_each_safe(tmp, next, &res->lr_converting) {
1913 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1915 if (iter(lock, closure) == LDLM_ITER_STOP)
1916 GOTO(out, rc = LDLM_ITER_STOP);
1919 cfs_list_for_each_safe(tmp, next, &res->lr_waiting) {
1920 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1922 if (iter(lock, closure) == LDLM_ITER_STOP)
1923 GOTO(out, rc = LDLM_ITER_STOP);
1930 struct iter_helper_data {
1931 ldlm_iterator_t iter;
1935 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1937 struct iter_helper_data *helper = closure;
1938 return helper->iter(lock, helper->closure);
1941 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
1943 return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
1946 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
1949 struct iter_helper_data helper = { iter: iter, closure: closure };
1950 return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
1953 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
1954 ldlm_res_iterator_t iter, void *closure)
1956 int i, rc = LDLM_ITER_CONTINUE;
1957 struct ldlm_resource *res;
1961 cfs_spin_lock(&ns->ns_hash_lock);
1962 for (i = 0; i < RES_HASH_SIZE; i++) {
1963 tmp = ns->ns_hash[i].next;
1964 while (tmp != &(ns->ns_hash[i])) {
1965 res = cfs_list_entry(tmp, struct ldlm_resource,
1967 ldlm_resource_getref(res);
1968 cfs_spin_unlock(&ns->ns_hash_lock);
1969 LDLM_RESOURCE_ADDREF(res);
1971 rc = iter(res, closure);
1973 LDLM_RESOURCE_DELREF(res);
1974 cfs_spin_lock(&ns->ns_hash_lock);
1976 ldlm_resource_putref_locked(res);
1977 if (rc == LDLM_ITER_STOP)
1982 cfs_spin_unlock(&ns->ns_hash_lock);
1986 /* non-blocking function to manipulate a lock whose cb_data is being put away.
1987 * return 0: find no resource
1988 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
1991 int ldlm_resource_iterate(struct ldlm_namespace *ns,
1992 const struct ldlm_res_id *res_id,
1993 ldlm_iterator_t iter, void *data)
1995 struct ldlm_resource *res;
2000 CERROR("must pass in namespace\n");
2004 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2008 LDLM_RESOURCE_ADDREF(res);
2009 rc = ldlm_resource_foreach(res, iter, data);
2010 LDLM_RESOURCE_DELREF(res);
2011 ldlm_resource_putref(res);
2017 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2019 cfs_list_t *list = closure;
2021 /* we use l_pending_chain here, because it's unused on clients. */
2022 LASSERTF(cfs_list_empty(&lock->l_pending_chain),
2023 "lock %p next %p prev %p\n",
2024 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2025 /* bug 9573: don't replay locks left after eviction, or
2026 * bug 17614: locks being actively cancelled. Get a reference
2027 * on a lock so that it does not disapear under us (e.g. due to cancel)
2029 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
2030 cfs_list_add(&lock->l_pending_chain, list);
2031 LDLM_LOCK_GET(lock);
2034 return LDLM_ITER_CONTINUE;
2037 static int replay_lock_interpret(const struct lu_env *env,
2038 struct ptlrpc_request *req,
2039 struct ldlm_async_args *aa, int rc)
2041 struct ldlm_lock *lock;
2042 struct ldlm_reply *reply;
2043 struct obd_export *exp;
2046 cfs_atomic_dec(&req->rq_import->imp_replay_inflight);
2051 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2053 GOTO(out, rc = -EPROTO);
2055 lock = ldlm_handle2lock(&aa->lock_handle);
2057 CERROR("received replay ack for unknown local cookie "LPX64
2058 " remote cookie "LPX64 " from server %s id %s\n",
2059 aa->lock_handle.cookie, reply->lock_handle.cookie,
2060 req->rq_export->exp_client_uuid.uuid,
2061 libcfs_id2str(req->rq_peer));
2062 GOTO(out, rc = -ESTALE);
2065 /* Key change rehash lock in per-export hash with new key */
2066 exp = req->rq_export;
2067 if (exp && exp->exp_lock_hash) {
2068 cfs_hash_rehash_key(exp->exp_lock_hash,
2069 &lock->l_remote_handle,
2070 &reply->lock_handle,
2073 lock->l_remote_handle = reply->lock_handle;
2076 LDLM_DEBUG(lock, "replayed lock:");
2077 ptlrpc_import_recovery_state_machine(req->rq_import);
2078 LDLM_LOCK_PUT(lock);
2081 ptlrpc_connect_import(req->rq_import, NULL);
2086 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2088 struct ptlrpc_request *req;
2089 struct ldlm_async_args *aa;
2090 struct ldlm_request *body;
2095 /* Bug 11974: Do not replay a lock which is actively being canceled */
2096 if (lock->l_flags & LDLM_FL_CANCELING) {
2097 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2101 /* If this is reply-less callback lock, we cannot replay it, since
2102 * server might have long dropped it, but notification of that event was
2103 * lost by network. (and server granted conflicting lock already) */
2104 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2105 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2106 ldlm_lock_cancel(lock);
2110 * If granted mode matches the requested mode, this lock is granted.
2112 * If they differ, but we have a granted mode, then we were granted
2113 * one mode and now want another: ergo, converting.
2115 * If we haven't been granted anything and are on a resource list,
2116 * then we're blocked/waiting.
2118 * If we haven't been granted anything and we're NOT on a resource list,
2119 * then we haven't got a reply yet and don't have a known disposition.
2120 * This happens whenever a lock enqueue is the request that triggers
2123 if (lock->l_granted_mode == lock->l_req_mode)
2124 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2125 else if (lock->l_granted_mode)
2126 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2127 else if (!cfs_list_empty(&lock->l_res_link))
2128 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2130 flags = LDLM_FL_REPLAY;
2132 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2133 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2137 /* We're part of recovery, so don't wait for it. */
2138 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2140 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2141 ldlm_lock2desc(lock, &body->lock_desc);
2142 body->lock_flags = flags;
2144 ldlm_lock2handle(lock, &body->lock_handle[0]);
2145 if (lock->l_lvb_len != 0) {
2146 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2147 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2150 ptlrpc_request_set_replen(req);
2151 /* notify the server we've replayed all requests.
2152 * also, we mark the request to be put on a dedicated
2153 * queue to be processed after all request replayes.
2155 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2157 LDLM_DEBUG(lock, "replaying lock:");
2159 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2160 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2161 aa = ptlrpc_req_async_args(req);
2162 aa->lock_handle = body->lock_handle[0];
2163 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2164 ptlrpcd_add_req(req, PSCOPE_OTHER);
2170 * Cancel as many unused locks as possible before replay. since we are
2171 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2174 * Called only in recovery before replaying locks. there is no need to
2175 * replay locks that are unused. since the clients may hold thousands of
2176 * cached unused locks, dropping the unused locks can greatly reduce the
2177 * load on the servers at recovery time.
2179 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2182 CFS_LIST_HEAD(cancels);
2184 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2185 "replay for namespace %s (%d)\n", ns->ns_name,
2188 /* We don't need to care whether or not LRU resize is enabled
2189 * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2190 * count parameter */
2191 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2192 LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2194 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2195 canceled, ns->ns_name);
2198 int ldlm_replay_locks(struct obd_import *imp)
2200 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2201 CFS_LIST_HEAD(list);
2202 struct ldlm_lock *lock, *next;
2207 LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0);
2209 /* don't replay locks if import failed recovery */
2210 if (imp->imp_vbr_failed)
2213 /* ensure this doesn't fall to 0 before all have been queued */
2214 cfs_atomic_inc(&imp->imp_replay_inflight);
2216 if (ldlm_cancel_unused_locks_before_replay)
2217 ldlm_cancel_unused_locks_for_replay(ns);
2219 (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2221 cfs_list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2222 cfs_list_del_init(&lock->l_pending_chain);
2224 LDLM_LOCK_RELEASE(lock);
2225 continue; /* or try to do the rest? */
2227 rc = replay_one_lock(imp, lock);
2228 LDLM_LOCK_RELEASE(lock);
2231 cfs_atomic_dec(&imp->imp_replay_inflight);