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 2008 Sun Microsystems, Inc. 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 static void interrupted_completion_wait(void *data)
57 struct lock_wait_data {
58 struct ldlm_lock *lwd_lock;
62 struct ldlm_async_args {
63 struct lustre_handle lock_handle;
66 int ldlm_expired_completion_wait(void *data)
68 struct lock_wait_data *lwd = data;
69 struct ldlm_lock *lock = lwd->lwd_lock;
70 struct obd_import *imp;
71 struct obd_device *obd;
74 if (lock->l_conn_export == NULL) {
75 static cfs_time_t next_dump = 0, last_dump = 0;
77 if (ptlrpc_check_suspend())
80 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
81 CFS_DURATION_T"s ago); not entering recovery in "
82 "server code, just going back to sleep",
83 lock->l_enqueued_time.tv_sec,
84 cfs_time_sub(cfs_time_current_sec(),
85 lock->l_enqueued_time.tv_sec));
86 if (cfs_time_after(cfs_time_current(), next_dump)) {
87 last_dump = next_dump;
88 next_dump = cfs_time_shift(300);
89 ldlm_namespace_dump(D_DLMTRACE,
90 lock->l_resource->lr_namespace);
92 libcfs_debug_dumplog();
97 obd = lock->l_conn_export->exp_obd;
98 imp = obd->u.cli.cl_import;
99 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
100 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
101 CFS_DURATION_T"s ago), entering recovery for %s@%s",
102 lock->l_enqueued_time.tv_sec,
103 cfs_time_sub(cfs_time_current_sec(),
104 lock->l_enqueued_time.tv_sec),
105 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
110 /* We use the same basis for both server side and client side functions
111 from a single node. */
112 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
114 int timeout = at_get(&lock->l_resource->lr_namespace->ns_at_estimate);
116 return obd_timeout / 2;
117 /* Since these are non-updating timeouts, we should be conservative.
118 It would be nice to have some kind of "early reply" mechanism for
119 lock callbacks too... */
120 timeout = timeout + (timeout >> 1); /* 150% */
121 return max(timeout, ldlm_enqueue_min);
123 EXPORT_SYMBOL(ldlm_get_enq_timeout);
125 static int is_granted_or_cancelled(struct ldlm_lock *lock)
129 lock_res_and_lock(lock);
130 if (((lock->l_req_mode == lock->l_granted_mode) &&
131 !(lock->l_flags & LDLM_FL_CP_REQD)) ||
132 (lock->l_flags & LDLM_FL_FAILED))
134 unlock_res_and_lock(lock);
140 * Helper function for ldlm_completion_ast(), updating timings when lock is
143 static int ldlm_completion_tail(struct ldlm_lock *lock)
148 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
149 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
152 delay = cfs_time_sub(cfs_time_current_sec(),
153 lock->l_enqueued_time.tv_sec);
154 LDLM_DEBUG(lock, "client-side enqueue: granted after "
155 CFS_DURATION_T"s", delay);
157 /* Update our time estimate */
158 at_add(&lock->l_resource->lr_namespace->ns_at_estimate, delay);
165 * Implementation of ->l_completion_ast() for a client, that doesn't wait
166 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
167 * other threads that cannot block for long.
169 int ldlm_completion_ast_async(struct ldlm_lock *lock, int flags, void *data)
173 if (flags == LDLM_FL_WAIT_NOREPROC) {
174 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
178 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
179 LDLM_FL_BLOCK_CONV))) {
180 cfs_waitq_signal(&lock->l_waitq);
181 RETURN(ldlm_completion_tail(lock));
184 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
186 ldlm_lock_dump(D_OTHER, lock, 0);
187 ldlm_reprocess_all(lock->l_resource);
192 * Client side LDLM "completion" AST. This is called in several cases:
194 * - when a reply to an ENQUEUE rpc is received from the server
195 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
196 * this point (determined by flags);
198 * - when LDLM_CP_CALLBACK rpc comes to client to notify it that lock has
201 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
204 * - to force all locks when resource is destroyed (cleanup_resource());
206 * - during lock conversion (not used currently).
208 * If lock is not granted in the first case, this function waits until second
209 * or penultimate cases happen in some other thread.
212 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
214 /* XXX ALLOCATE - 160 bytes */
215 struct lock_wait_data lwd;
216 struct obd_device *obd;
217 struct obd_import *imp = NULL;
218 struct l_wait_info lwi;
223 if (flags == LDLM_FL_WAIT_NOREPROC) {
224 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
228 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
229 LDLM_FL_BLOCK_CONV))) {
230 cfs_waitq_signal(&lock->l_waitq);
234 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
236 ldlm_lock_dump(D_OTHER, lock, 0);
240 obd = class_exp2obd(lock->l_conn_export);
242 /* if this is a local lock, then there is no import */
244 imp = obd->u.cli.cl_import;
247 /* Wait a long time for enqueue - server may have to callback a
248 lock from another client. Server will evict the other client if it
249 doesn't respond reasonably, and then give us the lock. */
250 timeout = ldlm_get_enq_timeout(lock) * 2;
254 if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
255 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
256 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
258 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
259 ldlm_expired_completion_wait,
260 interrupted_completion_wait, &lwd);
264 spin_lock(&imp->imp_lock);
265 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
266 spin_unlock(&imp->imp_lock);
269 if (ns_is_client(lock->l_resource->lr_namespace) &&
270 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
271 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
272 lock->l_flags |= LDLM_FL_FAIL_LOC;
275 /* Go to sleep until the lock is granted or cancelled. */
276 rc = l_wait_event(lock->l_waitq,
277 is_granted_or_cancelled(lock), &lwi);
281 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
286 RETURN(ldlm_completion_tail(lock));
290 * A helper to build a blocking ast function
292 * Perform a common operation for blocking asts:
293 * defferred lock cancellation.
295 * \param lock the lock blocking or canceling ast was called on
297 * \see mdt_blocking_ast
298 * \see ldlm_blocking_ast
300 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
305 lock->l_flags |= LDLM_FL_CBPENDING;
306 do_ast = (!lock->l_readers && !lock->l_writers);
307 unlock_res_and_lock(lock);
310 struct lustre_handle lockh;
313 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
314 ldlm_lock2handle(lock, &lockh);
315 rc = ldlm_cli_cancel(&lockh);
317 CERROR("ldlm_cli_cancel: %d\n", rc);
319 LDLM_DEBUG(lock, "Lock still has references, will be "
326 * Server blocking AST
328 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
331 * \param lock the lock which blocks a request or cancelling lock
334 * \param flag indicates whether this cancelling or blocking callback
336 * \see ldlm_blocking_ast_nocheck
338 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
339 void *data, int flag)
343 if (flag == LDLM_CB_CANCELING) {
344 /* Don't need to do anything here. */
348 lock_res_and_lock(lock);
349 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
350 * that ldlm_blocking_ast is called just before intent_policy method
351 * takes the ns_lock, then by the time we get the lock, we might not
352 * be the correct blocking function anymore. So check, and return
354 if (lock->l_blocking_ast != ldlm_blocking_ast) {
355 unlock_res_and_lock(lock);
358 RETURN(ldlm_blocking_ast_nocheck(lock));
362 * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
363 * comment in filter_intent_policy() on why you may need this.
365 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
368 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
369 * that is rather subtle: with OST-side locking, it may so happen that
370 * _all_ extent locks are held by the OST. If client wants to obtain
371 * current file size it calls ll{,u}_glimpse_size(), and (as locks are
372 * on the server), dummy glimpse callback fires and does
373 * nothing. Client still receives correct file size due to the
374 * following fragment in filter_intent_policy():
376 * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
377 * if (rc != 0 && res->lr_namespace->ns_lvbo &&
378 * res->lr_namespace->ns_lvbo->lvbo_update) {
379 * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
382 * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
383 * returns correct file size to the client.
385 return -ELDLM_NO_LOCK_DATA;
388 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
389 const struct ldlm_res_id *res_id,
390 ldlm_type_t type, ldlm_policy_data_t *policy,
391 ldlm_mode_t mode, int *flags,
392 ldlm_blocking_callback blocking,
393 ldlm_completion_callback completion,
394 ldlm_glimpse_callback glimpse,
395 void *data, __u32 lvb_len, void *lvb_swabber,
396 const __u64 *client_cookie,
397 struct lustre_handle *lockh)
399 struct ldlm_lock *lock;
401 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
402 .lcs_blocking = blocking,
403 .lcs_glimpse = glimpse,
407 LASSERT(!(*flags & LDLM_FL_REPLAY));
408 if (unlikely(ns_is_client(ns))) {
409 CERROR("Trying to enqueue local lock in a shadow namespace\n");
413 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len);
415 GOTO(out_nolock, err = -ENOMEM);
416 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
418 ldlm_lock_addref_internal(lock, mode);
419 ldlm_lock2handle(lock, lockh);
420 lock_res_and_lock(lock);
421 lock->l_flags |= LDLM_FL_LOCAL;
422 if (*flags & LDLM_FL_ATOMIC_CB)
423 lock->l_flags |= LDLM_FL_ATOMIC_CB;
424 lock->l_lvb_swabber = lvb_swabber;
425 unlock_res_and_lock(lock);
427 lock->l_policy_data = *policy;
428 if (client_cookie != NULL)
429 lock->l_client_cookie = *client_cookie;
430 if (type == LDLM_EXTENT)
431 lock->l_req_extent = policy->l_extent;
433 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
434 if (unlikely(err != ELDLM_OK))
438 *policy = lock->l_policy_data;
440 LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
443 if (lock->l_completion_ast)
444 lock->l_completion_ast(lock, *flags, NULL);
446 LDLM_DEBUG(lock, "client-side local enqueue END");
449 LDLM_LOCK_RELEASE(lock);
454 static void failed_lock_cleanup(struct ldlm_namespace *ns,
455 struct ldlm_lock *lock,
456 struct lustre_handle *lockh, int mode)
460 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
461 lock_res_and_lock(lock);
462 /* Check that lock is not granted or failed, we might race. */
463 if ((lock->l_req_mode != lock->l_granted_mode) &&
464 !(lock->l_flags & LDLM_FL_FAILED)) {
465 /* Make sure that this lock will not be found by raced
466 * bl_ast and -EINVAL reply is sent to server anyways.
468 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
472 unlock_res_and_lock(lock);
476 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
477 "LDLM_FL_ATOMIC_CB");
478 ldlm_lock_decref_and_cancel(lockh, mode);
480 LDLM_DEBUG(lock, "lock was granted or failed in race");
481 ldlm_lock_decref(lockh, mode);
484 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
485 * from llite/file.c/ll_file_flock(). */
486 if (lock->l_resource->lr_type == LDLM_FLOCK) {
487 ldlm_lock_destroy(lock);
491 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
492 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
493 int *flags, void *lvb, __u32 lvb_len,
494 void *lvb_swabber, struct lustre_handle *lockh,int rc)
496 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
497 int is_replay = *flags & LDLM_FL_REPLAY;
498 struct lustre_handle old_hash_key;
499 struct ldlm_lock *lock;
500 struct ldlm_reply *reply;
501 int cleanup_phase = 1;
504 lock = ldlm_handle2lock(lockh);
505 /* ldlm_cli_enqueue is holding a reference on this lock. */
507 LASSERT(type == LDLM_FLOCK);
511 if (rc != ELDLM_OK) {
513 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
514 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
515 if (rc == ELDLM_LOCK_ABORTED) {
516 /* Before we return, swab the reply */
517 reply = req_capsule_server_get(&req->rq_pill,
522 struct ost_lvb *tmplvb;
524 req_capsule_set_size(&req->rq_pill,
525 &RMF_DLM_LVB, RCL_SERVER,
527 tmplvb = req_capsule_server_swab_get(&req->rq_pill,
531 GOTO(cleanup, rc = -EPROTO);
533 memcpy(lvb, tmplvb, lvb_len);
539 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
541 GOTO(cleanup, rc = -EPROTO);
543 /* lock enqueued on the server */
546 lock_res_and_lock(lock);
547 old_hash_key = lock->l_remote_handle;
548 lock->l_remote_handle = reply->lock_handle;
550 /* Key change rehash lock in per-export hash with new key */
551 if (exp->exp_lock_hash)
552 lustre_hash_rehash_key(exp->exp_lock_hash, &old_hash_key,
553 &lock->l_remote_handle,
556 *flags = reply->lock_flags;
557 lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
558 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
559 * to wait with no timeout as well */
560 lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
561 unlock_res_and_lock(lock);
563 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\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 lock->l_policy_data =
601 reply->lock_desc.l_policy_data;
602 if (type != LDLM_PLAIN)
603 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
606 if ((*flags) & LDLM_FL_AST_SENT ||
607 /* Cancel extent locks as soon as possible on a liblustre client,
608 * because it cannot handle asynchronous ASTs robustly (see
610 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
611 lock_res_and_lock(lock);
612 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
613 unlock_res_and_lock(lock);
614 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
617 /* If the lock has already been granted by a completion AST, don't
618 * clobber the LVB with an older one. */
619 if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
622 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
624 tmplvb = req_capsule_server_swab_get(&req->rq_pill,
628 GOTO(cleanup, rc = -EPROTO);
629 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
633 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
634 if (lock->l_completion_ast != NULL) {
635 int err = lock->l_completion_ast(lock, *flags, NULL);
638 if (rc && type != LDLM_FLOCK) /* bug 9425, bug 10250 */
643 if (lvb_len && lvb != NULL) {
644 /* Copy the LVB here, and not earlier, because the completion
645 * AST (if any) can override what we got in the reply */
646 memcpy(lvb, lock->l_lvb_data, lvb_len);
649 LDLM_DEBUG(lock, "client-side enqueue END");
652 if (cleanup_phase == 1 && rc)
653 failed_lock_cleanup(ns, lock, lockh, mode);
654 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
656 LDLM_LOCK_RELEASE(lock);
660 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
661 * a single page on the send/receive side. XXX: 512 should be changed
662 * to more adequate value. */
663 static inline int ldlm_req_handles_avail(int req_size, int off)
667 avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size;
668 if (likely(avail >= 0))
669 avail /= (int)sizeof(struct lustre_handle);
672 avail += LDLM_LOCKREQ_HANDLES - off;
677 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
678 enum req_location loc,
681 int size = req_capsule_msg_size(pill, loc);
682 return ldlm_req_handles_avail(size, off);
685 static inline int ldlm_format_handles_avail(struct obd_import *imp,
686 const struct req_format *fmt,
687 enum req_location loc, int off)
689 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
690 return ldlm_req_handles_avail(size, off);
693 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
694 * @count locks in @cancels. */
695 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
696 int version, int opc, int canceloff,
697 struct list_head *cancels, int count)
699 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
700 struct req_capsule *pill = &req->rq_pill;
701 struct ldlm_request *dlm = NULL;
702 int flags, avail, to_free, bufcount, pack = 0;
709 if (exp_connect_cancelset(exp)) {
710 /* Estimate the amount of available space in the request. */
711 bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT);
712 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
714 flags = ns_connect_lru_resize(ns) ?
715 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
716 to_free = !ns_connect_lru_resize(ns) &&
717 opc == LDLM_ENQUEUE ? 1 : 0;
719 /* Cancel lru locks here _only_ if the server supports
720 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
721 * rpc, what will make us slower. */
723 count += ldlm_cancel_lru_local(ns, cancels, to_free,
724 avail - count, 0, flags);
729 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
730 ldlm_request_bufsize(pack, opc));
733 rc = ptlrpc_request_pack(req, version, opc);
735 ldlm_lock_list_put(cancels, l_bl_ast, count);
739 if (exp_connect_cancelset(exp)) {
741 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
743 /* Skip first lock handler in ldlm_request_pack(),
744 * this method will incrment @lock_count according
745 * to the lock handle amount actually written to
747 dlm->lock_count = canceloff;
749 /* Pack into the request @pack lock handles. */
750 ldlm_cli_cancel_list(cancels, pack, req, 0);
751 /* Prepare and send separate cancel rpc for others. */
752 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
754 ldlm_lock_list_put(cancels, l_bl_ast, count);
759 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
760 struct list_head *cancels, int count)
762 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
763 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
766 /* If a request has some specific initialisation it is passed in @reqp,
767 * otherwise it is created in ldlm_cli_enqueue.
769 * Supports sync and async requests, pass @async flag accordingly. If a
770 * request was created in ldlm_cli_enqueue and it is the async request,
771 * pass it to the caller in @reqp. */
772 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
773 struct ldlm_enqueue_info *einfo,
774 const struct ldlm_res_id *res_id,
775 ldlm_policy_data_t *policy, int *flags,
776 void *lvb, __u32 lvb_len, void *lvb_swabber,
777 struct lustre_handle *lockh, int async)
779 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
780 struct ldlm_lock *lock;
781 struct ldlm_request *body;
782 int is_replay = *flags & LDLM_FL_REPLAY;
783 int req_passed_in = 1;
785 struct ptlrpc_request *req;
788 LASSERT(exp != NULL);
790 /* If we're replaying this lock, just check some invariants.
791 * If we're creating a new lock, get everything all setup nice. */
793 lock = ldlm_handle2lock_long(lockh, 0);
794 LASSERT(lock != NULL);
795 LDLM_DEBUG(lock, "client-side enqueue START");
796 LASSERT(exp == lock->l_conn_export);
798 const struct ldlm_callback_suite cbs = {
799 .lcs_completion = einfo->ei_cb_cp,
800 .lcs_blocking = einfo->ei_cb_bl,
801 .lcs_glimpse = einfo->ei_cb_gl,
802 .lcs_weigh = einfo->ei_cb_wg
804 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
805 einfo->ei_mode, &cbs, einfo->ei_cbdata,
809 /* for the local lock, add the reference */
810 ldlm_lock_addref_internal(lock, einfo->ei_mode);
811 ldlm_lock2handle(lock, lockh);
812 lock->l_lvb_swabber = lvb_swabber;
813 if (policy != NULL) {
814 /* INODEBITS_INTEROP: If the server does not support
815 * inodebits, we will request a plain lock in the
816 * descriptor (ldlm_lock2desc() below) but use an
817 * inodebits lock internally with both bits set.
819 if (einfo->ei_type == LDLM_IBITS &&
820 !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
821 lock->l_policy_data.l_inodebits.bits =
822 MDS_INODELOCK_LOOKUP |
823 MDS_INODELOCK_UPDATE;
825 lock->l_policy_data = *policy;
828 if (einfo->ei_type == LDLM_EXTENT)
829 lock->l_req_extent = policy->l_extent;
830 LDLM_DEBUG(lock, "client-side enqueue START");
833 /* lock not sent to server yet */
835 if (reqp == NULL || *reqp == NULL) {
836 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
841 failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode);
842 LDLM_LOCK_RELEASE(lock);
852 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
854 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
855 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
858 lock->l_conn_export = exp;
859 lock->l_export = NULL;
860 lock->l_blocking_ast = einfo->ei_cb_bl;
862 /* Dump lock data into the request buffer */
863 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
864 ldlm_lock2desc(lock, &body->lock_desc);
865 body->lock_flags = *flags;
866 body->lock_handle[0] = *lockh;
868 /* Continue as normal. */
869 if (!req_passed_in) {
871 req_capsule_extend(&req->rq_pill,
872 &RQF_LDLM_ENQUEUE_LVB);
873 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
874 RCL_SERVER, lvb_len);
876 ptlrpc_request_set_replen(req);
880 * Liblustre client doesn't get extent locks, except for O_APPEND case
881 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
882 * [i_size, OBD_OBJECT_EOF] lock is taken.
884 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
885 policy->l_extent.end == OBD_OBJECT_EOF));
888 LASSERT(reqp != NULL);
892 LDLM_DEBUG(lock, "sending request");
894 rc = ptlrpc_queue_wait(req);
896 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
897 einfo->ei_mode, flags, lvb, lvb_len,
898 lvb_swabber, lockh, rc);
900 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
901 * one reference that we took */
903 LDLM_LOCK_RELEASE(lock);
907 if (!req_passed_in && req != NULL) {
908 ptlrpc_req_finished(req);
916 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
919 struct ldlm_resource *res;
922 if (ns_is_client(lock->l_resource->lr_namespace)) {
923 CERROR("Trying to cancel local lock\n");
926 LDLM_DEBUG(lock, "client-side local convert");
928 res = ldlm_lock_convert(lock, new_mode, flags);
930 ldlm_reprocess_all(res);
935 LDLM_DEBUG(lock, "client-side local convert handler END");
940 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
941 * conversion of locks which are on the waiting or converting queue */
942 /* Caller of this code is supposed to take care of lock readers/writers
944 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
946 struct ldlm_request *body;
947 struct ldlm_reply *reply;
948 struct ldlm_lock *lock;
949 struct ldlm_resource *res;
950 struct ptlrpc_request *req;
954 lock = ldlm_handle2lock(lockh);
961 if (lock->l_conn_export == NULL)
962 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
964 LDLM_DEBUG(lock, "client-side convert");
966 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
967 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
974 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
975 body->lock_handle[0] = lock->l_remote_handle;
977 body->lock_desc.l_req_mode = new_mode;
978 body->lock_flags = *flags;
981 ptlrpc_request_set_replen(req);
982 rc = ptlrpc_queue_wait(req);
986 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
988 GOTO(out, rc = -EPROTO);
991 GOTO(out, rc = req->rq_status);
993 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
995 ldlm_reprocess_all(res);
996 /* Go to sleep until the lock is granted. */
997 /* FIXME: or cancelled. */
998 if (lock->l_completion_ast) {
999 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1009 LDLM_LOCK_PUT(lock);
1010 ptlrpc_req_finished(req);
1014 /* Cancel locks locally.
1016 * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
1017 * LDLM_FL_CANCELING otherwise;
1018 * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
1019 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
1021 int rc = LDLM_FL_LOCAL_ONLY;
1024 if (lock->l_conn_export) {
1027 LDLM_DEBUG(lock, "client-side cancel");
1028 /* Set this flag to prevent others from getting new references*/
1029 lock_res_and_lock(lock);
1030 lock->l_flags |= LDLM_FL_CBPENDING;
1031 local_only = (lock->l_flags &
1032 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1033 ldlm_cancel_callback(lock);
1034 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
1035 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1036 unlock_res_and_lock(lock);
1039 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1041 rc = LDLM_FL_LOCAL_ONLY;
1043 ldlm_lock_cancel(lock);
1045 if (ns_is_client(lock->l_resource->lr_namespace)) {
1046 LDLM_ERROR(lock, "Trying to cancel local lock");
1049 LDLM_DEBUG(lock, "server-side local cancel");
1050 ldlm_lock_cancel(lock);
1051 ldlm_reprocess_all(lock->l_resource);
1052 LDLM_DEBUG(lock, "server-side local cancel handler END");
1058 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
1059 of the request @req. */
1060 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1061 struct list_head *head, int count)
1063 struct ldlm_request *dlm;
1064 struct ldlm_lock *lock;
1065 int max, packed = 0;
1068 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1069 LASSERT(dlm != NULL);
1071 /* Check the room in the request buffer. */
1072 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1073 sizeof(struct ldlm_request);
1074 max /= sizeof(struct lustre_handle);
1075 max += LDLM_LOCKREQ_HANDLES;
1076 LASSERT(max >= dlm->lock_count + count);
1078 /* XXX: it would be better to pack lock handles grouped by resource.
1079 * so that the server cancel would call filter_lvbo_update() less
1081 list_for_each_entry(lock, head, l_bl_ast) {
1084 LASSERT(lock->l_conn_export);
1085 /* Pack the lock handle to the given request buffer. */
1086 LDLM_DEBUG(lock, "packing");
1087 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1090 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1094 /* Prepare and send a batched cancel rpc, it will include count lock handles
1095 * of locks given in @head. */
1096 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1097 int count, int flags)
1099 struct ptlrpc_request *req = NULL;
1100 struct obd_import *imp;
1105 LASSERT(exp != NULL);
1108 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, obd_fail_val);
1110 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1113 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1114 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1121 imp = class_exp2cliimp(exp);
1122 if (imp == NULL || imp->imp_invalid) {
1124 "skipping cancel on invalid import %p\n", imp);
1128 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1130 GOTO(out, rc = -ENOMEM);
1132 bufcount = req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1133 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1134 ldlm_request_bufsize(count, LDLM_CANCEL));
1136 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1138 ptlrpc_request_free(req);
1141 req->rq_no_resend = 1;
1142 req->rq_no_delay = 1;
1144 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1145 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1146 ptlrpc_at_set_req_timeout(req);
1148 ldlm_cancel_pack(req, cancels, count);
1150 ptlrpc_request_set_replen(req);
1151 if (flags & LDLM_FL_ASYNC) {
1152 ptlrpcd_add_req(req, PSCOPE_OTHER);
1156 rc = ptlrpc_queue_wait(req);
1159 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1160 "out of sync -- not fatal\n",
1161 libcfs_nid2str(req->rq_import->
1162 imp_connection->c_peer.nid));
1164 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1165 req->rq_import_generation == imp->imp_generation) {
1166 ptlrpc_req_finished(req);
1168 } else if (rc != ELDLM_OK) {
1169 CERROR("Got rc %d from cancel RPC: canceling "
1177 ptlrpc_req_finished(req);
1180 return sent ? sent : rc;
1183 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1185 LASSERT(imp != NULL);
1186 return &imp->imp_obd->obd_namespace->ns_pool;
1190 * Update client's obd pool related fields with new SLV and Limit from \a req.
1192 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1194 struct obd_device *obd;
1195 __u64 old_slv, new_slv;
1198 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1199 !imp_connect_lru_resize(req->rq_import)))
1202 * Do nothing for corner cases.
1208 * In some cases RPC may contain slv and limit zeroed out. This is
1209 * the case when server does not support lru resize feature. This is
1210 * also possible in some recovery cases when server side reqs have no
1211 * ref to obd export and thus access to server side namespace is no
1214 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1215 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1216 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1217 "(SLV: "LPU64", Limit: %u)",
1218 lustre_msg_get_slv(req->rq_repmsg),
1219 lustre_msg_get_limit(req->rq_repmsg));
1223 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1224 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1225 obd = req->rq_import->imp_obd;
1228 * Set new SLV and Limit to obd fields to make accessible for pool
1229 * thread. We do not access obd_namespace and pool directly here
1230 * as there is no reliable way to make sure that they are still
1231 * alive in cleanup time. Evil races are possible which may cause
1232 * oops in that time.
1234 write_lock(&obd->obd_pool_lock);
1235 old_slv = obd->obd_pool_slv;
1236 obd->obd_pool_slv = new_slv;
1237 obd->obd_pool_limit = new_limit;
1238 write_unlock(&obd->obd_pool_lock);
1242 EXPORT_SYMBOL(ldlm_cli_update_pool);
1244 int ldlm_cli_cancel(struct lustre_handle *lockh)
1246 struct obd_export *exp;
1247 int avail, flags, count = 1, rc = 0;
1248 struct ldlm_namespace *ns;
1249 struct ldlm_lock *lock;
1250 CFS_LIST_HEAD(cancels);
1253 /* concurrent cancels on the same handle can happen */
1254 lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1256 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1260 rc = ldlm_cli_cancel_local(lock);
1261 if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) {
1262 LDLM_LOCK_RELEASE(lock);
1263 RETURN(rc < 0 ? rc : 0);
1265 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1266 * rpc which goes to canceld portal, so we can cancel other lru locks
1267 * here and send them all as one LDLM_CANCEL rpc. */
1268 LASSERT(list_empty(&lock->l_bl_ast));
1269 list_add(&lock->l_bl_ast, &cancels);
1271 exp = lock->l_conn_export;
1272 if (exp_connect_cancelset(exp)) {
1273 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1278 ns = lock->l_resource->lr_namespace;
1279 flags = ns_connect_lru_resize(ns) ?
1280 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1281 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1282 LDLM_FL_BL_AST, flags);
1284 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1288 /* XXX until we will have compound requests and can cut cancels from generic rpc
1289 * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */
1290 static int ldlm_cancel_list(struct list_head *cancels, int count, int flags)
1292 CFS_LIST_HEAD(head);
1293 struct ldlm_lock *lock, *next;
1294 int left = 0, bl_ast = 0, rc;
1297 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1301 if (flags & LDLM_FL_LOCAL_ONLY) {
1302 rc = LDLM_FL_LOCAL_ONLY;
1303 ldlm_lock_cancel(lock);
1305 rc = ldlm_cli_cancel_local(lock);
1307 if (!(flags & LDLM_FL_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1308 LDLM_DEBUG(lock, "Cancel lock separately");
1309 list_del_init(&lock->l_bl_ast);
1310 list_add(&lock->l_bl_ast, &head);
1314 if (rc == LDLM_FL_LOCAL_ONLY) {
1315 /* CANCEL RPC should not be sent to server. */
1316 list_del_init(&lock->l_bl_ast);
1317 LDLM_LOCK_RELEASE(lock);
1324 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1331 * Callback function for shrink policy. Makes decision whether to keep
1332 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1333 * \a added and number of locks to be preferably canceled \a count.
1335 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1337 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1339 static ldlm_policy_res_t ldlm_cancel_shrink_policy(struct ldlm_namespace *ns,
1340 struct ldlm_lock *lock,
1341 int unused, int added,
1348 * Stop lru processing when we reached passed @count or checked all
1351 if (count && added >= count)
1352 return LDLM_POLICY_KEEP_LOCK;
1354 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1355 if (lock->l_weigh_ast) {
1357 * For liblustre, l_weigh_ast should return 0 since it
1360 page_nr = lock->l_weigh_ast(lock);
1362 struct ldlm_extent *l_extent;
1365 * For all extent locks cost is 1 + number of pages in
1368 l_extent = &lock->l_policy_data.l_extent;
1369 page_nr = l_extent->end - l_extent->start;
1370 do_div(page_nr, CFS_PAGE_SIZE);
1372 lock_cost = 1 + page_nr;
1375 * For all locks which are not extent ones cost is 1
1381 * Keep all expensive locks in lru for the memory pressure time
1382 * cancel policy. They anyways may be canceled by lru resize
1383 * pplicy if they have not small enough CLV.
1385 return lock_cost > ns->ns_shrink_thumb ?
1386 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1390 * Callback function for lru-resize policy. Makes decision whether to keep
1391 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1392 * \a added and number of locks to be preferably canceled \a count.
1394 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1396 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1398 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1399 struct ldlm_lock *lock,
1400 int unused, int added,
1403 cfs_time_t cur = cfs_time_current();
1404 struct ldlm_pool *pl = &ns->ns_pool;
1409 * Stop lru processing when we reached passed @count or checked all
1412 if (count && added >= count)
1413 return LDLM_POLICY_KEEP_LOCK;
1415 slv = ldlm_pool_get_slv(pl);
1416 lvf = ldlm_pool_get_lvf(pl);
1417 la = cfs_duration_sec(cfs_time_sub(cur,
1418 lock->l_last_used));
1421 * Stop when slv is not yet come from server or lv is smaller than
1424 lv = lvf * la * unused;
1427 * Inform pool about current CLV to see it via proc.
1429 ldlm_pool_set_clv(pl, lv);
1430 return (slv == 1 || lv < slv) ?
1431 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1435 * Callback function for proc used policy. Makes decision whether to keep
1436 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1437 * \a added and number of locks to be preferably canceled \a count.
1439 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1441 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1443 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1444 struct ldlm_lock *lock,
1445 int unused, int added,
1449 * Stop lru processing when we reached passed @count or checked all
1452 return (added >= count) ?
1453 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1457 * Callback function for aged policy. Makes decision whether to keep
1458 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1459 * \a added and number of locks to be preferably canceled \a count.
1461 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1463 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1465 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1466 struct ldlm_lock *lock,
1467 int unused, int added,
1471 * Stop lru processing if young lock is found and we reached passed
1474 return ((added >= count) &&
1475 cfs_time_before(cfs_time_current(),
1476 cfs_time_add(lock->l_last_used,
1478 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1482 * Callback function for default policy. Makes decision whether to keep
1483 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1484 * \a added and number of locks to be preferably canceled \a count.
1486 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1488 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1490 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1491 struct ldlm_lock *lock,
1492 int unused, int added,
1496 * Stop lru processing when we reached passed @count or checked all
1499 return (added >= count) ?
1500 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1503 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1504 struct ldlm_lock *, int,
1507 static ldlm_cancel_lru_policy_t
1508 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1510 if (ns_connect_lru_resize(ns)) {
1511 if (flags & LDLM_CANCEL_SHRINK)
1512 return ldlm_cancel_shrink_policy;
1513 else if (flags & LDLM_CANCEL_LRUR)
1514 return ldlm_cancel_lrur_policy;
1515 else if (flags & LDLM_CANCEL_PASSED)
1516 return ldlm_cancel_passed_policy;
1518 if (flags & LDLM_CANCEL_AGED)
1519 return ldlm_cancel_aged_policy;
1522 return ldlm_cancel_default_policy;
1525 /* - Free space in lru for @count new locks,
1526 * redundant unused locks are canceled locally;
1527 * - also cancel locally unused aged locks;
1528 * - do not cancel more than @max locks;
1529 * - GET the found locks and add them into the @cancels list.
1531 * A client lock can be added to the l_bl_ast list only when it is
1532 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
1533 * There are the following use cases: ldlm_cancel_resource_local(),
1534 * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
1535 * flag properly. As any attempt to cancel a lock rely on this flag,
1536 * l_bl_ast list is accessed later without any special locking.
1538 * Calling policies for enabled lru resize:
1539 * ----------------------------------------
1540 * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to
1541 * cancel not more than @count locks;
1543 * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at
1544 * the beginning of lru list);
1546 * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
1547 * memory pressre policy function;
1549 * flags & LDLM_CANCEL_AGED - cancel alocks according to "aged policy".
1551 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1552 int count, int max, int cancel_flags, int flags)
1554 ldlm_cancel_lru_policy_t pf;
1555 struct ldlm_lock *lock, *next;
1556 int added = 0, unused;
1559 spin_lock(&ns->ns_unused_lock);
1560 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 (!list_empty(&ns->ns_unused_list)) {
1569 /* For any flags, stop scanning if @max is reached. */
1570 if (max && added >= max)
1573 list_for_each_entry_safe(lock, next, &ns->ns_unused_list, l_lru){
1574 /* No locks which got blocking requests. */
1575 LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1577 /* Somebody is already doing CANCEL. No need in this
1578 * lock in lru, do not traverse it again. */
1579 if (!(lock->l_flags & LDLM_FL_CANCELING))
1582 ldlm_lock_remove_from_lru_nolock(lock);
1584 if (&lock->l_lru == &ns->ns_unused_list)
1587 LDLM_LOCK_GET(lock);
1588 spin_unlock(&ns->ns_unused_lock);
1589 lu_ref_add(&lock->l_reference, __FUNCTION__, cfs_current());
1591 /* Pass the lock through the policy filter and see if it
1592 * should stay in lru.
1594 * Even for shrinker policy we stop scanning if
1595 * we find a lock that should stay in the cache.
1596 * We should take into account lock age anyway
1597 * as new lock even if it is small of weight is
1598 * valuable resource.
1600 * That is, for shrinker policy we drop only
1601 * old locks, but additionally chose them by
1602 * their weight. Big extent locks will stay in
1604 if (pf(ns, lock, unused, added, count) ==
1605 LDLM_POLICY_KEEP_LOCK) {
1606 lu_ref_del(&lock->l_reference,
1607 __FUNCTION__, cfs_current());
1608 LDLM_LOCK_RELEASE(lock);
1609 spin_lock(&ns->ns_unused_lock);
1613 lock_res_and_lock(lock);
1614 /* Check flags again under the lock. */
1615 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1616 (ldlm_lock_remove_from_lru(lock) == 0)) {
1617 /* other thread is removing lock from lru or
1618 * somebody is already doing CANCEL or
1619 * there is a blocking request which will send
1620 * cancel by itseft or the lock is matched
1621 * is already not unused. */
1622 unlock_res_and_lock(lock);
1623 lu_ref_del(&lock->l_reference,
1624 __FUNCTION__, cfs_current());
1625 LDLM_LOCK_RELEASE(lock);
1626 spin_lock(&ns->ns_unused_lock);
1629 LASSERT(!lock->l_readers && !lock->l_writers);
1631 /* If we have chosen to cancel this lock voluntarily, we
1632 * better send cancel notification to server, so that it
1633 * frees appropriate state. This might lead to a race
1634 * where while we are doing cancel here, server is also
1635 * silently cancelling this lock. */
1636 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1638 /* Setting the CBPENDING flag is a little misleading,
1639 * but prevents an important race; namely, once
1640 * CBPENDING is set, the lock can accumulate no more
1641 * readers/writers. Since readers and writers are
1642 * already zero here, ldlm_lock_decref() won't see
1643 * this flag and call l_blocking_ast */
1644 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1646 /* We can't re-add to l_lru as it confuses the
1647 * refcounting in ldlm_lock_remove_from_lru() if an AST
1648 * arrives after we drop ns_lock below. We use l_bl_ast
1649 * and can't use l_pending_chain as it is used both on
1650 * server and client nevertheless bug 5666 says it is
1651 * used only on server */
1652 LASSERT(list_empty(&lock->l_bl_ast));
1653 list_add(&lock->l_bl_ast, cancels);
1654 unlock_res_and_lock(lock);
1655 lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current());
1656 spin_lock(&ns->ns_unused_lock);
1660 spin_unlock(&ns->ns_unused_lock);
1661 RETURN(ldlm_cancel_list(cancels, added, cancel_flags));
1664 /* Returns number of locks which could be canceled next time when
1665 * ldlm_cancel_lru() is called. Used from locks pool shrinker. */
1666 int ldlm_cancel_lru_estimate(struct ldlm_namespace *ns,
1667 int count, int max, int flags)
1669 struct list_head disp = CFS_LIST_HEAD_INIT(disp);
1670 ldlm_cancel_lru_policy_t pf;
1671 struct ldlm_lock *lock;
1672 int added = 0, unused;
1676 pf = ldlm_cancel_lru_policy(ns, flags);
1677 LASSERT(pf != NULL);
1678 spin_lock(&ns->ns_unused_lock);
1679 unused = ns->ns_nr_unused;
1680 list_splice_init(&ns->ns_unused_list, &disp);
1681 while (!list_empty(&disp)) {
1682 lock = list_entry(disp.next, struct ldlm_lock, l_lru);
1683 list_move_tail(&lock->l_lru, &ns->ns_unused_list);
1685 /* For any flags, stop scanning if @max is reached. */
1686 if (max && added >= max)
1689 /* Somebody is already doing CANCEL or there is a
1690 * blocking request will send cancel. Let's not count
1692 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1693 (lock->l_flags & LDLM_FL_BL_AST))
1696 LDLM_LOCK_GET(lock);
1697 spin_unlock(&ns->ns_unused_lock);
1698 lu_ref_add(&lock->l_reference, __FUNCTION__, cfs_current());
1700 /* Pass the lock through the policy filter and see if it
1701 * should stay in lru. */
1702 if (pf(ns, lock, unused, added, count) == LDLM_POLICY_KEEP_LOCK)
1705 lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current());
1706 LDLM_LOCK_RELEASE(lock);
1707 spin_lock(&ns->ns_unused_lock);
1714 list_splice(&disp, ns->ns_unused_list.prev);
1715 spin_unlock(&ns->ns_unused_lock);
1719 /* when called with LDLM_ASYNC the blocking callback will be handled
1720 * in a thread and this function will return after the thread has been
1721 * asked to call the callback. when called with LDLM_SYNC the blocking
1722 * callback will be performed in this function. */
1723 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync,
1726 CFS_LIST_HEAD(cancels);
1731 sync = LDLM_SYNC; /* force to be sync in user space */
1733 count = ldlm_cancel_lru_local(ns, &cancels, nr, 0, 0, flags);
1734 if (sync == LDLM_ASYNC) {
1735 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count);
1740 /* If an error occured in ASYNC mode, or
1741 * this is SYNC mode, cancel the list. */
1742 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1746 /* Find and cancel locally unused locks found on resource, matched to the
1747 * given policy, mode. GET the found locks and add them into the @cancels
1749 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1750 struct list_head *cancels,
1751 ldlm_policy_data_t *policy,
1752 ldlm_mode_t mode, int lock_flags,
1753 int cancel_flags, void *opaque)
1755 struct ldlm_lock *lock;
1760 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1761 if (opaque != NULL && lock->l_ast_data != opaque) {
1762 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1763 lock->l_ast_data, opaque);
1768 if (lock->l_readers || lock->l_writers) {
1769 if (cancel_flags & LDLM_FL_WARN) {
1770 LDLM_ERROR(lock, "lock in use");
1776 /* If somebody is already doing CANCEL, or blocking ast came,
1777 * skip this lock. */
1778 if (lock->l_flags & LDLM_FL_BL_AST ||
1779 lock->l_flags & LDLM_FL_CANCELING)
1782 if (lockmode_compat(lock->l_granted_mode, mode))
1785 /* If policy is given and this is IBITS lock, add to list only
1786 * those locks that match by policy. */
1787 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1788 !(lock->l_policy_data.l_inodebits.bits &
1789 policy->l_inodebits.bits))
1792 /* See CBPENDING comment in ldlm_cancel_lru */
1793 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1796 LASSERT(list_empty(&lock->l_bl_ast));
1797 list_add(&lock->l_bl_ast, cancels);
1798 LDLM_LOCK_GET(lock);
1803 RETURN(ldlm_cancel_list(cancels, count, cancel_flags));
1806 /* If @req is NULL, send CANCEL request to server with handles of locks
1807 * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1808 * separately per lock.
1809 * If @req is not NULL, put handles of locks in @cancels into the request
1810 * buffer at the offset @off.
1811 * Destroy @cancels at the end. */
1812 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1813 struct ptlrpc_request *req, int flags)
1815 struct ldlm_lock *lock;
1819 if (list_empty(cancels) || count == 0)
1822 /* XXX: requests (both batched and not) could be sent in parallel.
1823 * Usually it is enough to have just 1 RPC, but it is possible that
1824 * there are to many locks to be cancelled in LRU or on a resource.
1825 * It would also speed up the case when the server does not support
1828 LASSERT(!list_empty(cancels));
1829 lock = list_entry(cancels->next, struct ldlm_lock, l_bl_ast);
1830 LASSERT(lock->l_conn_export);
1832 if (exp_connect_cancelset(lock->l_conn_export)) {
1835 ldlm_cancel_pack(req, cancels, count);
1837 res = ldlm_cli_cancel_req(lock->l_conn_export,
1841 res = ldlm_cli_cancel_req(lock->l_conn_export,
1846 CERROR("ldlm_cli_cancel_list: %d\n", res);
1851 ldlm_lock_list_put(cancels, l_bl_ast, res);
1853 LASSERT(count == 0);
1857 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1858 const struct ldlm_res_id *res_id,
1859 ldlm_policy_data_t *policy,
1860 ldlm_mode_t mode, int flags, void *opaque)
1862 struct ldlm_resource *res;
1863 CFS_LIST_HEAD(cancels);
1868 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1870 /* This is not a problem. */
1871 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1875 LDLM_RESOURCE_ADDREF(res);
1876 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1878 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1880 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1882 LDLM_RESOURCE_DELREF(res);
1883 ldlm_resource_putref(res);
1887 static inline int have_no_nsresource(struct ldlm_namespace *ns)
1889 int no_resource = 0;
1891 spin_lock(&ns->ns_hash_lock);
1892 if (ns->ns_resources == 0)
1894 spin_unlock(&ns->ns_hash_lock);
1896 RETURN(no_resource);
1899 /* Cancel all locks on a namespace (or a specific resource, if given)
1900 * that have 0 readers/writers.
1902 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1903 * to notify the server. */
1904 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1905 const struct ldlm_res_id *res_id,
1906 int flags, void *opaque)
1915 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1919 spin_lock(&ns->ns_hash_lock);
1920 for (i = 0; i < RES_HASH_SIZE; i++) {
1921 struct list_head *tmp;
1922 tmp = ns->ns_hash[i].next;
1923 while (tmp != &(ns->ns_hash[i])) {
1924 struct ldlm_resource *res;
1927 res = list_entry(tmp, struct ldlm_resource, lr_hash);
1928 ldlm_resource_getref(res);
1929 spin_unlock(&ns->ns_hash_lock);
1931 LDLM_RESOURCE_ADDREF(res);
1932 rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name,
1937 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1938 res->lr_name.name[0], rc);
1940 LDLM_RESOURCE_DELREF(res);
1941 spin_lock(&ns->ns_hash_lock);
1943 ldlm_resource_putref_locked(res);
1946 spin_unlock(&ns->ns_hash_lock);
1951 /* Lock iterators. */
1953 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1956 struct list_head *tmp, *next;
1957 struct ldlm_lock *lock;
1958 int rc = LDLM_ITER_CONTINUE;
1963 RETURN(LDLM_ITER_CONTINUE);
1966 list_for_each_safe(tmp, next, &res->lr_granted) {
1967 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1969 if (iter(lock, closure) == LDLM_ITER_STOP)
1970 GOTO(out, rc = LDLM_ITER_STOP);
1973 list_for_each_safe(tmp, next, &res->lr_converting) {
1974 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1976 if (iter(lock, closure) == LDLM_ITER_STOP)
1977 GOTO(out, rc = LDLM_ITER_STOP);
1980 list_for_each_safe(tmp, next, &res->lr_waiting) {
1981 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1983 if (iter(lock, closure) == LDLM_ITER_STOP)
1984 GOTO(out, rc = LDLM_ITER_STOP);
1991 struct iter_helper_data {
1992 ldlm_iterator_t iter;
1996 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1998 struct iter_helper_data *helper = closure;
1999 return helper->iter(lock, helper->closure);
2002 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
2004 return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
2007 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
2010 struct iter_helper_data helper = { iter: iter, closure: closure };
2011 return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
2014 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
2015 ldlm_res_iterator_t iter, void *closure)
2017 int i, rc = LDLM_ITER_CONTINUE;
2018 struct ldlm_resource *res;
2019 struct list_head *tmp;
2022 spin_lock(&ns->ns_hash_lock);
2023 for (i = 0; i < RES_HASH_SIZE; i++) {
2024 tmp = ns->ns_hash[i].next;
2025 while (tmp != &(ns->ns_hash[i])) {
2026 res = list_entry(tmp, struct ldlm_resource, lr_hash);
2027 ldlm_resource_getref(res);
2028 spin_unlock(&ns->ns_hash_lock);
2029 LDLM_RESOURCE_ADDREF(res);
2031 rc = iter(res, closure);
2033 LDLM_RESOURCE_DELREF(res);
2034 spin_lock(&ns->ns_hash_lock);
2036 ldlm_resource_putref_locked(res);
2037 if (rc == LDLM_ITER_STOP)
2042 spin_unlock(&ns->ns_hash_lock);
2046 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
2047 void ldlm_resource_iterate(struct ldlm_namespace *ns,
2048 const struct ldlm_res_id *res_id,
2049 ldlm_iterator_t iter, void *data)
2051 struct ldlm_resource *res;
2055 CERROR("must pass in namespace\n");
2059 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2065 LDLM_RESOURCE_ADDREF(res);
2066 ldlm_resource_foreach(res, iter, data);
2067 LDLM_RESOURCE_DELREF(res);
2068 ldlm_resource_putref(res);
2074 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2076 struct list_head *list = closure;
2078 /* we use l_pending_chain here, because it's unused on clients. */
2079 LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n",
2080 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2081 /* bug 9573: don't replay locks left after eviction */
2082 if (!(lock->l_flags & LDLM_FL_FAILED))
2083 list_add(&lock->l_pending_chain, list);
2084 return LDLM_ITER_CONTINUE;
2087 static int replay_lock_interpret(const struct lu_env *env,
2088 struct ptlrpc_request *req,
2089 struct ldlm_async_args *aa, int rc)
2091 struct lustre_handle old_hash_key;
2092 struct ldlm_lock *lock;
2093 struct ldlm_reply *reply;
2094 struct obd_export *exp;
2097 atomic_dec(&req->rq_import->imp_replay_inflight);
2102 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2104 GOTO(out, rc = -EPROTO);
2106 lock = ldlm_handle2lock(&aa->lock_handle);
2108 CERROR("received replay ack for unknown local cookie "LPX64
2109 " remote cookie "LPX64 " from server %s id %s\n",
2110 aa->lock_handle.cookie, reply->lock_handle.cookie,
2111 req->rq_export->exp_client_uuid.uuid,
2112 libcfs_id2str(req->rq_peer));
2113 GOTO(out, rc = -ESTALE);
2116 old_hash_key = lock->l_remote_handle;
2117 lock->l_remote_handle = reply->lock_handle;
2119 /* Key change rehash lock in per-export hash with new key */
2120 exp = req->rq_export;
2121 if (exp && exp->exp_lock_hash)
2122 lustre_hash_rehash_key(exp->exp_lock_hash, &old_hash_key,
2123 &lock->l_remote_handle,
2126 LDLM_DEBUG(lock, "replayed lock:");
2127 ptlrpc_import_recovery_state_machine(req->rq_import);
2128 LDLM_LOCK_PUT(lock);
2131 ptlrpc_connect_import(req->rq_import, NULL);
2136 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2138 struct ptlrpc_request *req;
2139 struct ldlm_async_args *aa;
2140 struct ldlm_request *body;
2145 /* Bug 11974: Do not replay a lock which is actively being canceled */
2146 if (lock->l_flags & LDLM_FL_CANCELING) {
2147 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2151 /* If this is reply-less callback lock, we cannot replay it, since
2152 * server might have long dropped it, but notification of that event was
2153 * lost by network. (and server granted conflicting lock already) */
2154 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2155 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2156 ldlm_lock_cancel(lock);
2160 * If granted mode matches the requested mode, this lock is granted.
2162 * If they differ, but we have a granted mode, then we were granted
2163 * one mode and now want another: ergo, converting.
2165 * If we haven't been granted anything and are on a resource list,
2166 * then we're blocked/waiting.
2168 * If we haven't been granted anything and we're NOT on a resource list,
2169 * then we haven't got a reply yet and don't have a known disposition.
2170 * This happens whenever a lock enqueue is the request that triggers
2173 if (lock->l_granted_mode == lock->l_req_mode)
2174 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2175 else if (lock->l_granted_mode)
2176 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2177 else if (!list_empty(&lock->l_res_link))
2178 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2180 flags = LDLM_FL_REPLAY;
2182 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2183 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2187 /* We're part of recovery, so don't wait for it. */
2188 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2190 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2191 ldlm_lock2desc(lock, &body->lock_desc);
2192 body->lock_flags = flags;
2194 ldlm_lock2handle(lock, &body->lock_handle[0]);
2195 if (lock->l_lvb_len != 0) {
2196 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2197 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2200 ptlrpc_request_set_replen(req);
2201 /* notify the server we've replayed all requests.
2202 * also, we mark the request to be put on a dedicated
2203 * queue to be processed after all request replayes.
2205 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2207 LDLM_DEBUG(lock, "replaying lock:");
2209 atomic_inc(&req->rq_import->imp_replay_inflight);
2210 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2211 aa = ptlrpc_req_async_args(req);
2212 aa->lock_handle = body->lock_handle[0];
2213 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2214 ptlrpcd_add_req(req, PSCOPE_OTHER);
2219 int ldlm_replay_locks(struct obd_import *imp)
2221 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2222 CFS_LIST_HEAD(list);
2223 struct ldlm_lock *lock, *next;
2228 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2230 /* ensure this doesn't fall to 0 before all have been queued */
2231 atomic_inc(&imp->imp_replay_inflight);
2233 (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2235 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2236 list_del_init(&lock->l_pending_chain);
2238 continue; /* or try to do the rest? */
2239 rc = replay_one_lock(imp, lock);
2242 atomic_dec(&imp->imp_replay_inflight);