4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 * This file contains Asynchronous System Trap (AST) handlers and related
38 * LDLM request-processing routines.
40 * An AST is a callback issued on a lock when its state is changed. There are
41 * several different types of ASTs (callbacks) registered for each lock:
43 * - completion AST: when a lock is enqueued by some process, but cannot be
44 * granted immediately due to other conflicting locks on the same resource,
45 * the completion AST is sent to notify the caller when the lock is
48 * - blocking AST: when a lock is granted to some process, if another process
49 * enqueues a conflicting (blocking) lock on a resource, a blocking AST is
50 * sent to notify the holder(s) of the lock(s) of the conflicting lock
51 * request. The lock holder(s) must release their lock(s) on that resource in
52 * a timely manner or be evicted by the server.
54 * - glimpse AST: this is used when a process wants information about a lock
55 * (i.e. the lock value block (LVB)) but does not necessarily require holding
56 * the lock. If the resource is locked, the lock holder(s) are sent glimpse
57 * ASTs and the LVB is returned to the caller, and lock holder(s) may CANCEL
58 * their lock(s) if they are idle. If the resource is not locked, the server
62 #define DEBUG_SUBSYSTEM S_LDLM
64 #include <lustre_dlm.h>
65 #include <obd_class.h>
68 #include "ldlm_internal.h"
70 int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
71 CFS_MODULE_PARM(ldlm_enqueue_min, "i", int, 0644,
72 "lock enqueue timeout minimum");
74 /* in client side, whether the cached locks will be canceled before replay */
75 unsigned int ldlm_cancel_unused_locks_before_replay = 1;
77 static void interrupted_completion_wait(void *data)
81 struct lock_wait_data {
82 struct ldlm_lock *lwd_lock;
86 struct ldlm_async_args {
87 struct lustre_handle lock_handle;
90 int ldlm_expired_completion_wait(void *data)
92 struct lock_wait_data *lwd = data;
93 struct ldlm_lock *lock = lwd->lwd_lock;
94 struct obd_import *imp;
95 struct obd_device *obd;
97 if (lock->l_conn_export == NULL) {
98 static cfs_time_t next_dump = 0, last_dump = 0;
100 if (ptlrpc_check_suspend())
103 LCONSOLE_WARN("lock timed out (enqueued at "CFS_TIME_T", "
104 CFS_DURATION_T"s ago)\n",
105 lock->l_last_activity,
106 cfs_time_sub(cfs_time_current_sec(),
107 lock->l_last_activity));
108 LDLM_DEBUG(lock, "lock timed out (enqueued at "CFS_TIME_T", "
109 CFS_DURATION_T"s ago); not entering recovery in "
110 "server code, just going back to sleep",
111 lock->l_last_activity,
112 cfs_time_sub(cfs_time_current_sec(),
113 lock->l_last_activity));
114 if (cfs_time_after(cfs_time_current(), next_dump)) {
115 last_dump = next_dump;
116 next_dump = cfs_time_shift(300);
117 ldlm_namespace_dump(D_DLMTRACE,
118 ldlm_lock_to_ns(lock));
120 libcfs_debug_dumplog();
125 obd = lock->l_conn_export->exp_obd;
126 imp = obd->u.cli.cl_import;
127 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
128 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
129 CFS_DURATION_T"s ago), entering recovery for %s@%s",
130 lock->l_last_activity,
131 cfs_time_sub(cfs_time_current_sec(), lock->l_last_activity),
132 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
136 EXPORT_SYMBOL(ldlm_expired_completion_wait);
138 /* We use the same basis for both server side and client side functions
139 from a single node. */
140 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
142 int timeout = at_get(ldlm_lock_to_ns_at(lock));
144 return obd_timeout / 2;
145 /* Since these are non-updating timeouts, we should be conservative.
146 It would be nice to have some kind of "early reply" mechanism for
147 lock callbacks too... */
148 timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */
149 return max(timeout, ldlm_enqueue_min);
151 EXPORT_SYMBOL(ldlm_get_enq_timeout);
154 * Helper function for ldlm_completion_ast(), updating timings when lock is
157 static int ldlm_completion_tail(struct ldlm_lock *lock)
162 if (lock->l_flags & (LDLM_FL_DESTROYED | LDLM_FL_FAILED)) {
163 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
166 delay = cfs_time_sub(cfs_time_current_sec(),
167 lock->l_last_activity);
168 LDLM_DEBUG(lock, "client-side enqueue: granted after "
169 CFS_DURATION_T"s", delay);
171 /* Update our time estimate */
172 at_measured(ldlm_lock_to_ns_at(lock),
180 * Implementation of ->l_completion_ast() for a client, that doesn't wait
181 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
182 * other threads that cannot block for long.
184 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
186 if (flags == LDLM_FL_WAIT_NOREPROC) {
187 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
191 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
192 LDLM_FL_BLOCK_CONV))) {
193 wake_up(&lock->l_waitq);
194 return ldlm_completion_tail(lock);
197 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
199 ldlm_reprocess_all(lock->l_resource);
202 EXPORT_SYMBOL(ldlm_completion_ast_async);
205 * Generic LDLM "completion" AST. This is called in several cases:
207 * - when a reply to an ENQUEUE RPC is received from the server
208 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
209 * this point (determined by flags);
211 * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
214 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
217 * - to force all locks when resource is destroyed (cleanup_resource());
219 * - during lock conversion (not used currently).
221 * If lock is not granted in the first case, this function waits until second
222 * or penultimate cases happen in some other thread.
225 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
227 /* XXX ALLOCATE - 160 bytes */
228 struct lock_wait_data lwd;
229 struct obd_device *obd;
230 struct obd_import *imp = NULL;
231 struct l_wait_info lwi;
235 if (flags == LDLM_FL_WAIT_NOREPROC) {
236 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
240 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
241 LDLM_FL_BLOCK_CONV))) {
242 wake_up(&lock->l_waitq);
246 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
251 obd = class_exp2obd(lock->l_conn_export);
253 /* if this is a local lock, then there is no import */
255 imp = obd->u.cli.cl_import;
258 /* Wait a long time for enqueue - server may have to callback a
259 lock from another client. Server will evict the other client if it
260 doesn't respond reasonably, and then give us the lock. */
261 timeout = ldlm_get_enq_timeout(lock) * 2;
265 if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
266 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
267 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
269 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
270 ldlm_expired_completion_wait,
271 interrupted_completion_wait, &lwd);
275 spin_lock(&imp->imp_lock);
276 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
277 spin_unlock(&imp->imp_lock);
280 if (ns_is_client(ldlm_lock_to_ns(lock)) &&
281 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
282 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
283 lock->l_flags |= LDLM_FL_FAIL_LOC;
286 /* Go to sleep until the lock is granted or cancelled. */
287 rc = l_wait_event(lock->l_waitq,
288 is_granted_or_cancelled(lock), &lwi);
292 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
297 return ldlm_completion_tail(lock);
299 EXPORT_SYMBOL(ldlm_completion_ast);
302 * A helper to build a blocking AST function
304 * Perform a common operation for blocking ASTs:
305 * defferred lock cancellation.
307 * \param lock the lock blocking or canceling AST was called on
309 * \see mdt_blocking_ast
310 * \see ldlm_blocking_ast
312 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
316 lock->l_flags |= LDLM_FL_CBPENDING;
317 do_ast = (!lock->l_readers && !lock->l_writers);
318 unlock_res_and_lock(lock);
321 struct lustre_handle lockh;
324 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
325 ldlm_lock2handle(lock, &lockh);
326 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
328 CERROR("ldlm_cli_cancel: %d\n", rc);
330 LDLM_DEBUG(lock, "Lock still has references, will be "
335 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
338 * Server blocking AST
340 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
343 * \param lock the lock which blocks a request or cancelling lock
346 * \param flag indicates whether this cancelling or blocking callback
348 * \see ldlm_blocking_ast_nocheck
350 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
351 void *data, int flag)
353 if (flag == LDLM_CB_CANCELING) {
354 /* Don't need to do anything here. */
358 lock_res_and_lock(lock);
359 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
360 * that ldlm_blocking_ast is called just before intent_policy method
361 * takes the lr_lock, then by the time we get the lock, we might not
362 * be the correct blocking function anymore. So check, and return
364 if (lock->l_blocking_ast != ldlm_blocking_ast) {
365 unlock_res_and_lock(lock);
368 return ldlm_blocking_ast_nocheck(lock);
370 EXPORT_SYMBOL(ldlm_blocking_ast);
373 * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
374 * comment in filter_intent_policy() on why you may need this.
376 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
379 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
380 * that is rather subtle: with OST-side locking, it may so happen that
381 * _all_ extent locks are held by the OST. If client wants to obtain
382 * current file size it calls ll{,u}_glimpse_size(), and (as locks are
383 * on the server), dummy glimpse callback fires and does
384 * nothing. Client still receives correct file size due to the
385 * following fragment in filter_intent_policy():
387 * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
388 * if (rc != 0 && res->lr_namespace->ns_lvbo &&
389 * res->lr_namespace->ns_lvbo->lvbo_update) {
390 * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
393 * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
394 * returns correct file size to the client.
396 return -ELDLM_NO_LOCK_DATA;
398 EXPORT_SYMBOL(ldlm_glimpse_ast);
401 * Enqueue a local lock (typically on a server).
403 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
404 const struct ldlm_res_id *res_id,
405 ldlm_type_t type, ldlm_policy_data_t *policy,
406 ldlm_mode_t mode, __u64 *flags,
407 ldlm_blocking_callback blocking,
408 ldlm_completion_callback completion,
409 ldlm_glimpse_callback glimpse,
410 void *data, __u32 lvb_len, enum lvb_type lvb_type,
411 const __u64 *client_cookie,
412 struct lustre_handle *lockh)
414 struct ldlm_lock *lock;
416 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
417 .lcs_blocking = blocking,
418 .lcs_glimpse = glimpse,
421 LASSERT(!(*flags & LDLM_FL_REPLAY));
422 if (unlikely(ns_is_client(ns))) {
423 CERROR("Trying to enqueue local lock in a shadow namespace\n");
427 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
430 GOTO(out_nolock, err = -ENOMEM);
432 ldlm_lock2handle(lock, lockh);
434 /* NB: we don't have any lock now (lock_res_and_lock)
435 * because it's a new lock */
436 ldlm_lock_addref_internal_nolock(lock, mode);
437 lock->l_flags |= LDLM_FL_LOCAL;
438 if (*flags & LDLM_FL_ATOMIC_CB)
439 lock->l_flags |= LDLM_FL_ATOMIC_CB;
442 lock->l_policy_data = *policy;
443 if (client_cookie != NULL)
444 lock->l_client_cookie = *client_cookie;
445 if (type == LDLM_EXTENT)
446 lock->l_req_extent = policy->l_extent;
448 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
449 if (unlikely(err != ELDLM_OK))
453 *policy = lock->l_policy_data;
455 if (lock->l_completion_ast)
456 lock->l_completion_ast(lock, *flags, NULL);
458 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
460 LDLM_LOCK_RELEASE(lock);
464 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
466 static void failed_lock_cleanup(struct ldlm_namespace *ns,
467 struct ldlm_lock *lock, int mode)
471 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
472 lock_res_and_lock(lock);
473 /* Check that lock is not granted or failed, we might race. */
474 if ((lock->l_req_mode != lock->l_granted_mode) &&
475 !(lock->l_flags & LDLM_FL_FAILED)) {
476 /* Make sure that this lock will not be found by raced
477 * bl_ast and -EINVAL reply is sent to server anyways.
479 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
480 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
483 unlock_res_and_lock(lock);
487 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
488 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
490 LDLM_DEBUG(lock, "lock was granted or failed in race");
492 ldlm_lock_decref_internal(lock, mode);
494 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
495 * from llite/file.c/ll_file_flock(). */
496 /* This code makes for the fact that we do not have blocking handler on
497 * a client for flock locks. As such this is the place where we must
498 * completely kill failed locks. (interrupted and those that
499 * were waiting to be granted when server evicted us. */
500 if (lock->l_resource->lr_type == LDLM_FLOCK) {
501 lock_res_and_lock(lock);
502 ldlm_resource_unlink_lock(lock);
503 ldlm_lock_destroy_nolock(lock);
504 unlock_res_and_lock(lock);
509 * Finishing portion of client lock enqueue code.
511 * Called after receiving reply from server.
513 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
514 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
515 __u64 *flags, void *lvb, __u32 lvb_len,
516 struct lustre_handle *lockh,int rc)
518 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
519 int is_replay = *flags & LDLM_FL_REPLAY;
520 struct ldlm_lock *lock;
521 struct ldlm_reply *reply;
522 int cleanup_phase = 1;
525 lock = ldlm_handle2lock(lockh);
526 /* ldlm_cli_enqueue is holding a reference on this lock. */
528 LASSERT(type == LDLM_FLOCK);
532 LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
533 "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
535 if (rc != ELDLM_OK) {
537 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
538 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
540 if (rc != ELDLM_LOCK_ABORTED)
544 /* Before we return, swab the reply */
545 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
547 GOTO(cleanup, rc = -EPROTO);
550 LASSERT(lvb != NULL);
552 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
555 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
556 GOTO(cleanup, rc = size);
557 } else if (unlikely(size > lvb_len)) {
558 LDLM_ERROR(lock, "Replied LVB is larger than "
559 "expectation, expected = %d, replied = %d",
561 GOTO(cleanup, rc = -EINVAL);
565 if (rc == ELDLM_LOCK_ABORTED) {
567 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
569 GOTO(cleanup, rc = (rc != 0 ? rc : ELDLM_LOCK_ABORTED));
572 /* lock enqueued on the server */
575 lock_res_and_lock(lock);
576 /* Key change rehash lock in per-export hash with new key */
577 if (exp->exp_lock_hash) {
578 /* In the function below, .hs_keycmp resolves to
579 * ldlm_export_lock_keycmp() */
580 /* coverity[overrun-buffer-val] */
581 cfs_hash_rehash_key(exp->exp_lock_hash,
582 &lock->l_remote_handle,
586 lock->l_remote_handle = reply->lock_handle;
589 *flags = ldlm_flags_from_wire(reply->lock_flags);
590 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
592 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
593 * to wait with no timeout as well */
594 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
596 unlock_res_and_lock(lock);
598 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%llx\n",
599 lock, reply->lock_handle.cookie, *flags);
601 /* If enqueue returned a blocked lock but the completion handler has
602 * already run, then it fixed up the resource and we don't need to do it
604 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
605 int newmode = reply->lock_desc.l_req_mode;
607 if (newmode && newmode != lock->l_req_mode) {
608 LDLM_DEBUG(lock, "server returned different mode %s",
609 ldlm_lockname[newmode]);
610 lock->l_req_mode = newmode;
613 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
614 &lock->l_resource->lr_name)) {
615 CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES
616 " instead of "DLDLMRES"\n",
617 PLDLMRES(&reply->lock_desc.l_resource),
618 PLDLMRES(lock->l_resource));
620 rc = ldlm_lock_change_resource(ns, lock,
621 &reply->lock_desc.l_resource.lr_name);
622 if (rc || lock->l_resource == NULL)
623 GOTO(cleanup, rc = -ENOMEM);
624 LDLM_DEBUG(lock, "client-side enqueue, new resource");
627 if (!(type == LDLM_IBITS &&
628 !(exp_connect_flags(exp) & OBD_CONNECT_IBITS)))
629 /* We assume lock type cannot change on server*/
630 ldlm_convert_policy_to_local(exp,
631 lock->l_resource->lr_type,
632 &reply->lock_desc.l_policy_data,
633 &lock->l_policy_data);
634 if (type != LDLM_PLAIN)
635 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
638 if ((*flags) & LDLM_FL_AST_SENT ||
639 /* Cancel extent locks as soon as possible on a liblustre client,
640 * because it cannot handle asynchronous ASTs robustly (see
642 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
643 lock_res_and_lock(lock);
644 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
645 unlock_res_and_lock(lock);
646 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
649 /* If the lock has already been granted by a completion AST, don't
650 * clobber the LVB with an older one. */
652 /* We must lock or a racing completion might update lvb without
653 * letting us know and we'll clobber the correct value.
654 * Cannot unlock after the check either, a that still leaves
655 * a tiny window for completion to get in */
656 lock_res_and_lock(lock);
657 if (lock->l_req_mode != lock->l_granted_mode)
658 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
659 lock->l_lvb_data, size);
660 unlock_res_and_lock(lock);
668 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
669 if (lock->l_completion_ast != NULL) {
670 int err = lock->l_completion_ast(lock, *flags, NULL);
678 if (lvb_len && lvb != NULL) {
679 /* Copy the LVB here, and not earlier, because the completion
680 * AST (if any) can override what we got in the reply */
681 memcpy(lvb, lock->l_lvb_data, lvb_len);
684 LDLM_DEBUG(lock, "client-side enqueue END");
686 if (cleanup_phase == 1 && rc)
687 failed_lock_cleanup(ns, lock, mode);
688 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
690 LDLM_LOCK_RELEASE(lock);
693 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
696 * Estimate number of lock handles that would fit into request of given
697 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
698 * a single page on the send/receive side. XXX: 512 should be changed to
699 * more adequate value.
701 static inline int ldlm_req_handles_avail(int req_size, int off)
705 avail = min_t(int, LDLM_MAXREQSIZE, PAGE_CACHE_SIZE - 512) - req_size;
706 if (likely(avail >= 0))
707 avail /= (int)sizeof(struct lustre_handle);
710 avail += LDLM_LOCKREQ_HANDLES - off;
715 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
716 enum req_location loc,
719 int size = req_capsule_msg_size(pill, loc);
720 return ldlm_req_handles_avail(size, off);
723 static inline int ldlm_format_handles_avail(struct obd_import *imp,
724 const struct req_format *fmt,
725 enum req_location loc, int off)
727 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
728 return ldlm_req_handles_avail(size, off);
732 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
733 * \a count locks in \a cancels.
735 * This is to be called by functions preparing their own requests that
736 * might contain lists of locks to cancel in addition to actual operation
737 * that needs to be performed.
739 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
740 int version, int opc, int canceloff,
741 struct list_head *cancels, int count)
743 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
744 struct req_capsule *pill = &req->rq_pill;
745 struct ldlm_request *dlm = NULL;
746 int flags, avail, to_free, pack = 0;
752 if (ns_connect_cancelset(ns)) {
753 /* Estimate the amount of available space in the request. */
754 req_capsule_filled_sizes(pill, RCL_CLIENT);
755 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
757 flags = ns_connect_lru_resize(ns) ?
758 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
759 to_free = !ns_connect_lru_resize(ns) &&
760 opc == LDLM_ENQUEUE ? 1 : 0;
762 /* Cancel LRU locks here _only_ if the server supports
763 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
764 * RPC, which will make us slower. */
766 count += ldlm_cancel_lru_local(ns, cancels, to_free,
767 avail - count, 0, flags);
772 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
773 ldlm_request_bufsize(pack, opc));
776 rc = ptlrpc_request_pack(req, version, opc);
778 ldlm_lock_list_put(cancels, l_bl_ast, count);
782 if (ns_connect_cancelset(ns)) {
784 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
786 /* Skip first lock handler in ldlm_request_pack(),
787 * this method will incrment @lock_count according
788 * to the lock handle amount actually written to
790 dlm->lock_count = canceloff;
792 /* Pack into the request @pack lock handles. */
793 ldlm_cli_cancel_list(cancels, pack, req, 0);
794 /* Prepare and send separate cancel RPC for others. */
795 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
797 ldlm_lock_list_put(cancels, l_bl_ast, count);
801 EXPORT_SYMBOL(ldlm_prep_elc_req);
803 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
804 struct list_head *cancels, int count)
806 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
807 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
809 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
811 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
813 struct ptlrpc_request *req;
816 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
818 return ERR_PTR(-ENOMEM);
820 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
822 ptlrpc_request_free(req);
826 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
827 ptlrpc_request_set_replen(req);
830 EXPORT_SYMBOL(ldlm_enqueue_pack);
833 * Client-side lock enqueue.
835 * If a request has some specific initialisation it is passed in \a reqp,
836 * otherwise it is created in ldlm_cli_enqueue.
838 * Supports sync and async requests, pass \a async flag accordingly. If a
839 * request was created in ldlm_cli_enqueue and it is the async request,
840 * pass it to the caller in \a reqp.
842 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
843 struct ldlm_enqueue_info *einfo,
844 const struct ldlm_res_id *res_id,
845 ldlm_policy_data_t const *policy, __u64 *flags,
846 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
847 struct lustre_handle *lockh, int async)
849 struct ldlm_namespace *ns;
850 struct ldlm_lock *lock;
851 struct ldlm_request *body;
852 int is_replay = *flags & LDLM_FL_REPLAY;
853 int req_passed_in = 1;
855 struct ptlrpc_request *req;
857 LASSERT(exp != NULL);
859 ns = exp->exp_obd->obd_namespace;
861 /* If we're replaying this lock, just check some invariants.
862 * If we're creating a new lock, get everything all setup nice. */
864 lock = ldlm_handle2lock_long(lockh, 0);
865 LASSERT(lock != NULL);
866 LDLM_DEBUG(lock, "client-side enqueue START");
867 LASSERT(exp == lock->l_conn_export);
869 const struct ldlm_callback_suite cbs = {
870 .lcs_completion = einfo->ei_cb_cp,
871 .lcs_blocking = einfo->ei_cb_bl,
872 .lcs_glimpse = einfo->ei_cb_gl
874 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
875 einfo->ei_mode, &cbs, einfo->ei_cbdata,
879 /* for the local lock, add the reference */
880 ldlm_lock_addref_internal(lock, einfo->ei_mode);
881 ldlm_lock2handle(lock, lockh);
882 if (policy != NULL) {
883 /* INODEBITS_INTEROP: If the server does not support
884 * inodebits, we will request a plain lock in the
885 * descriptor (ldlm_lock2desc() below) but use an
886 * inodebits lock internally with both bits set.
888 if (einfo->ei_type == LDLM_IBITS &&
889 !(exp_connect_flags(exp) &
891 lock->l_policy_data.l_inodebits.bits =
892 MDS_INODELOCK_LOOKUP |
893 MDS_INODELOCK_UPDATE;
895 lock->l_policy_data = *policy;
898 if (einfo->ei_type == LDLM_EXTENT)
899 lock->l_req_extent = policy->l_extent;
900 LDLM_DEBUG(lock, "client-side enqueue START, flags %llx\n",
904 lock->l_conn_export = exp;
905 lock->l_export = NULL;
906 lock->l_blocking_ast = einfo->ei_cb_bl;
907 lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
909 /* lock not sent to server yet */
911 if (reqp == NULL || *reqp == NULL) {
912 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
917 failed_lock_cleanup(ns, lock, einfo->ei_mode);
918 LDLM_LOCK_RELEASE(lock);
928 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
930 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
931 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
934 /* Dump lock data into the request buffer */
935 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
936 ldlm_lock2desc(lock, &body->lock_desc);
937 body->lock_flags = ldlm_flags_to_wire(*flags);
938 body->lock_handle[0] = *lockh;
940 /* Continue as normal. */
941 if (!req_passed_in) {
943 req_capsule_extend(&req->rq_pill,
944 &RQF_LDLM_ENQUEUE_LVB);
945 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
947 ptlrpc_request_set_replen(req);
951 * Liblustre client doesn't get extent locks, except for O_APPEND case
952 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
953 * [i_size, OBD_OBJECT_EOF] lock is taken.
955 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
956 policy->l_extent.end == OBD_OBJECT_EOF));
959 LASSERT(reqp != NULL);
963 LDLM_DEBUG(lock, "sending request");
965 rc = ptlrpc_queue_wait(req);
967 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
968 einfo->ei_mode, flags, lvb, lvb_len,
971 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
972 * one reference that we took */
974 LDLM_LOCK_RELEASE(lock);
978 if (!req_passed_in && req != NULL) {
979 ptlrpc_req_finished(req);
986 EXPORT_SYMBOL(ldlm_cli_enqueue);
988 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
991 struct ldlm_resource *res;
994 if (ns_is_client(ldlm_lock_to_ns(lock))) {
995 CERROR("Trying to cancel local lock\n");
998 LDLM_DEBUG(lock, "client-side local convert");
1000 res = ldlm_lock_convert(lock, new_mode, flags);
1002 ldlm_reprocess_all(res);
1005 rc = LUSTRE_EDEADLK;
1007 LDLM_DEBUG(lock, "client-side local convert handler END");
1008 LDLM_LOCK_PUT(lock);
1012 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
1013 * conversion of locks which are on the waiting or converting queue */
1014 /* Caller of this code is supposed to take care of lock readers/writers
1016 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
1018 struct ldlm_request *body;
1019 struct ldlm_reply *reply;
1020 struct ldlm_lock *lock;
1021 struct ldlm_resource *res;
1022 struct ptlrpc_request *req;
1025 lock = ldlm_handle2lock(lockh);
1032 if (lock->l_conn_export == NULL)
1033 return ldlm_cli_convert_local(lock, new_mode, flags);
1035 LDLM_DEBUG(lock, "client-side convert");
1037 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
1038 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1041 LDLM_LOCK_PUT(lock);
1045 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1046 body->lock_handle[0] = lock->l_remote_handle;
1048 body->lock_desc.l_req_mode = new_mode;
1049 body->lock_flags = ldlm_flags_to_wire(*flags);
1052 ptlrpc_request_set_replen(req);
1053 rc = ptlrpc_queue_wait(req);
1057 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1059 GOTO(out, rc = -EPROTO);
1062 GOTO(out, rc = req->rq_status);
1064 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
1066 ldlm_reprocess_all(res);
1067 /* Go to sleep until the lock is granted. */
1068 /* FIXME: or cancelled. */
1069 if (lock->l_completion_ast) {
1070 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1076 rc = LUSTRE_EDEADLK;
1079 LDLM_LOCK_PUT(lock);
1080 ptlrpc_req_finished(req);
1083 EXPORT_SYMBOL(ldlm_cli_convert);
1086 * Cancel locks locally.
1088 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1089 * \retval LDLM_FL_CANCELING otherwise;
1090 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1092 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1094 __u64 rc = LDLM_FL_LOCAL_ONLY;
1096 if (lock->l_conn_export) {
1099 LDLM_DEBUG(lock, "client-side cancel");
1100 /* Set this flag to prevent others from getting new references*/
1101 lock_res_and_lock(lock);
1102 lock->l_flags |= LDLM_FL_CBPENDING;
1103 local_only = !!(lock->l_flags &
1104 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1105 ldlm_cancel_callback(lock);
1106 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
1107 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1108 unlock_res_and_lock(lock);
1111 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1113 rc = LDLM_FL_LOCAL_ONLY;
1115 ldlm_lock_cancel(lock);
1117 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1118 LDLM_ERROR(lock, "Trying to cancel local lock");
1121 LDLM_DEBUG(lock, "server-side local cancel");
1122 ldlm_lock_cancel(lock);
1123 ldlm_reprocess_all(lock->l_resource);
1130 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1132 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1133 struct list_head *head, int count)
1135 struct ldlm_request *dlm;
1136 struct ldlm_lock *lock;
1137 int max, packed = 0;
1139 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1140 LASSERT(dlm != NULL);
1142 /* Check the room in the request buffer. */
1143 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1144 sizeof(struct ldlm_request);
1145 max /= sizeof(struct lustre_handle);
1146 max += LDLM_LOCKREQ_HANDLES;
1147 LASSERT(max >= dlm->lock_count + count);
1149 /* XXX: it would be better to pack lock handles grouped by resource.
1150 * so that the server cancel would call filter_lvbo_update() less
1152 list_for_each_entry(lock, head, l_bl_ast) {
1155 LASSERT(lock->l_conn_export);
1156 /* Pack the lock handle to the given request buffer. */
1157 LDLM_DEBUG(lock, "packing");
1158 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1161 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1165 * Prepare and send a batched cancel RPC. It will include \a count lock
1166 * handles of locks given in \a cancels list. */
1167 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1168 int count, ldlm_cancel_flags_t flags)
1170 struct ptlrpc_request *req = NULL;
1171 struct obd_import *imp;
1175 LASSERT(exp != NULL);
1178 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1180 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1183 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1184 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1189 imp = class_exp2cliimp(exp);
1190 if (imp == NULL || imp->imp_invalid) {
1192 "skipping cancel on invalid import %p\n", imp);
1196 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1198 GOTO(out, rc = -ENOMEM);
1200 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1201 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1202 ldlm_request_bufsize(count, LDLM_CANCEL));
1204 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1206 ptlrpc_request_free(req);
1210 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1211 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1212 ptlrpc_at_set_req_timeout(req);
1214 ldlm_cancel_pack(req, cancels, count);
1216 ptlrpc_request_set_replen(req);
1217 if (flags & LCF_ASYNC) {
1218 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1222 rc = ptlrpc_queue_wait(req);
1224 if (rc == LUSTRE_ESTALE) {
1225 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1226 "out of sync -- not fatal\n",
1227 libcfs_nid2str(req->rq_import->
1228 imp_connection->c_peer.nid));
1230 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1231 req->rq_import_generation == imp->imp_generation) {
1232 ptlrpc_req_finished(req);
1234 } else if (rc != ELDLM_OK) {
1235 /* -ESHUTDOWN is common on umount */
1236 CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1237 "Got rc %d from cancel RPC: "
1238 "canceling anyway\n", rc);
1245 ptlrpc_req_finished(req);
1247 return sent ? sent : rc;
1249 EXPORT_SYMBOL(ldlm_cli_cancel_req);
1251 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1253 LASSERT(imp != NULL);
1254 return &imp->imp_obd->obd_namespace->ns_pool;
1258 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1260 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1262 struct obd_device *obd;
1266 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1267 !imp_connect_lru_resize(req->rq_import)))
1270 * Do nothing for corner cases.
1275 /* In some cases RPC may contain SLV and limit zeroed out. This
1276 * is the case when server does not support LRU resize feature.
1277 * This is also possible in some recovery cases when server-side
1278 * reqs have no reference to the OBD export and thus access to
1279 * server-side namespace is not possible. */
1280 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1281 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1282 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1283 "(SLV: "LPU64", Limit: %u)",
1284 lustre_msg_get_slv(req->rq_repmsg),
1285 lustre_msg_get_limit(req->rq_repmsg));
1289 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1290 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1291 obd = req->rq_import->imp_obd;
1293 /* Set new SLV and limit in OBD fields to make them accessible
1294 * to the pool thread. We do not access obd_namespace and pool
1295 * directly here as there is no reliable way to make sure that
1296 * they are still alive at cleanup time. Evil races are possible
1297 * which may cause Oops at that time. */
1298 write_lock(&obd->obd_pool_lock);
1299 obd->obd_pool_slv = new_slv;
1300 obd->obd_pool_limit = new_limit;
1301 write_unlock(&obd->obd_pool_lock);
1305 EXPORT_SYMBOL(ldlm_cli_update_pool);
1308 * Client side lock cancel.
1310 * Lock must not have any readers or writers by this time.
1312 int ldlm_cli_cancel(struct lustre_handle *lockh,
1313 ldlm_cancel_flags_t cancel_flags)
1315 struct obd_export *exp;
1316 int avail, flags, count = 1;
1318 struct ldlm_namespace *ns;
1319 struct ldlm_lock *lock;
1322 /* concurrent cancels on the same handle can happen */
1323 lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1325 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1329 rc = ldlm_cli_cancel_local(lock);
1330 if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
1331 LDLM_LOCK_RELEASE(lock);
1334 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1335 * RPC which goes to canceld portal, so we can cancel other LRU locks
1336 * here and send them all as one LDLM_CANCEL RPC. */
1337 LASSERT(list_empty(&lock->l_bl_ast));
1338 list_add(&lock->l_bl_ast, &cancels);
1340 exp = lock->l_conn_export;
1341 if (exp_connect_cancelset(exp)) {
1342 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1347 ns = ldlm_lock_to_ns(lock);
1348 flags = ns_connect_lru_resize(ns) ?
1349 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1350 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1353 ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1356 EXPORT_SYMBOL(ldlm_cli_cancel);
1359 * Locally cancel up to \a count locks in list \a cancels.
1360 * Return the number of cancelled locks.
1362 int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
1363 ldlm_cancel_flags_t flags)
1366 struct ldlm_lock *lock, *next;
1367 int left = 0, bl_ast = 0;
1371 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1375 if (flags & LCF_LOCAL) {
1376 rc = LDLM_FL_LOCAL_ONLY;
1377 ldlm_lock_cancel(lock);
1379 rc = ldlm_cli_cancel_local(lock);
1381 /* Until we have compound requests and can send LDLM_CANCEL
1382 * requests batched with generic RPCs, we need to send cancels
1383 * with the LDLM_FL_BL_AST flag in a separate RPC from
1384 * the one being generated now. */
1385 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1386 LDLM_DEBUG(lock, "Cancel lock separately");
1387 list_del_init(&lock->l_bl_ast);
1388 list_add(&lock->l_bl_ast, &head);
1392 if (rc == LDLM_FL_LOCAL_ONLY) {
1393 /* CANCEL RPC should not be sent to server. */
1394 list_del_init(&lock->l_bl_ast);
1395 LDLM_LOCK_RELEASE(lock);
1401 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1406 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
1409 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1410 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1411 * readahead requests, ...)
1413 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1414 struct ldlm_lock *lock,
1415 int unused, int added,
1418 ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1419 ldlm_cancel_for_recovery cb = ns->ns_cancel_for_recovery;
1420 lock_res_and_lock(lock);
1422 /* don't check added & count since we want to process all locks
1423 * from unused list */
1424 switch (lock->l_resource->lr_type) {
1430 result = LDLM_POLICY_SKIP_LOCK;
1431 lock->l_flags |= LDLM_FL_SKIPPED;
1435 unlock_res_and_lock(lock);
1440 * Callback function for LRU-resize policy. Decides whether to keep
1441 * \a lock in LRU for current \a LRU size \a unused, added in current
1442 * scan \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_lrur_policy(struct ldlm_namespace *ns,
1449 struct ldlm_lock *lock,
1450 int unused, int added,
1453 cfs_time_t cur = cfs_time_current();
1454 struct ldlm_pool *pl = &ns->ns_pool;
1458 /* Stop LRU processing when we reach past @count or have checked all
1460 if (count && added >= count)
1461 return LDLM_POLICY_KEEP_LOCK;
1463 slv = ldlm_pool_get_slv(pl);
1464 lvf = ldlm_pool_get_lvf(pl);
1465 la = cfs_duration_sec(cfs_time_sub(cur,
1466 lock->l_last_used));
1467 lv = lvf * la * unused;
1469 /* Inform pool about current CLV to see it via proc. */
1470 ldlm_pool_set_clv(pl, lv);
1472 /* Stop when SLV is not yet come from server or lv is smaller than
1474 return (slv == 0 || lv < slv) ?
1475 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1479 * Callback function for proc used policy. Makes decision whether to keep
1480 * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1481 * added and number of locks to be preferably canceled \a count.
1483 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1485 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1487 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1488 struct ldlm_lock *lock,
1489 int unused, int added,
1492 /* Stop LRU processing when we reach past @count or have checked all
1494 return (added >= count) ?
1495 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1499 * Callback function for aged policy. Makes decision whether to keep \a lock in
1500 * LRU for current LRU size \a unused, added in current scan \a added and
1501 * number of locks to be preferably canceled \a count.
1503 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1505 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1507 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1508 struct ldlm_lock *lock,
1509 int unused, int added,
1512 /* Stop LRU processing if young lock is found and we reach past count */
1513 return ((added >= count) &&
1514 cfs_time_before(cfs_time_current(),
1515 cfs_time_add(lock->l_last_used,
1517 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1521 * Callback function for default policy. Makes decision whether to keep \a lock
1522 * in LRU for current LRU size \a unused, added in current scan \a added and
1523 * number of locks to be preferably canceled \a count.
1525 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1527 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1529 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1530 struct ldlm_lock *lock,
1531 int unused, int added,
1534 /* Stop LRU processing when we reach past count or have checked all
1536 return (added >= count) ?
1537 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1540 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1541 struct ldlm_lock *, int,
1544 static ldlm_cancel_lru_policy_t
1545 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1547 if (flags & LDLM_CANCEL_NO_WAIT)
1548 return ldlm_cancel_no_wait_policy;
1550 if (ns_connect_lru_resize(ns)) {
1551 if (flags & LDLM_CANCEL_SHRINK)
1552 /* We kill passed number of old locks. */
1553 return ldlm_cancel_passed_policy;
1554 else if (flags & LDLM_CANCEL_LRUR)
1555 return ldlm_cancel_lrur_policy;
1556 else if (flags & LDLM_CANCEL_PASSED)
1557 return ldlm_cancel_passed_policy;
1559 if (flags & LDLM_CANCEL_AGED)
1560 return ldlm_cancel_aged_policy;
1563 return ldlm_cancel_default_policy;
1567 * - Free space in LRU for \a count new locks,
1568 * redundant unused locks are canceled locally;
1569 * - also cancel locally unused aged locks;
1570 * - do not cancel more than \a max locks;
1571 * - GET the found locks and add them into the \a cancels list.
1573 * A client lock can be added to the l_bl_ast list only when it is
1574 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1575 * CANCEL. There are the following use cases:
1576 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1577 * ldlm_cli_cancel(), which check and set this flag properly. As any
1578 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1579 * later without any special locking.
1581 * Calling policies for enabled LRU resize:
1582 * ----------------------------------------
1583 * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to
1584 * cancel not more than \a count locks;
1586 * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at
1587 * the beginning of LRU list);
1589 * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to
1590 * memory pressre policy function;
1592 * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy".
1594 * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1595 * (typically before replaying locks) w/o
1596 * sending any RPCs or waiting for any
1597 * outstanding RPC to complete.
1599 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, struct list_head *cancels,
1600 int count, int max, int flags)
1602 ldlm_cancel_lru_policy_t pf;
1603 struct ldlm_lock *lock, *next;
1604 int added = 0, unused, remained;
1606 spin_lock(&ns->ns_lock);
1607 unused = ns->ns_nr_unused;
1610 if (!ns_connect_lru_resize(ns))
1611 count += unused - ns->ns_max_unused;
1613 pf = ldlm_cancel_lru_policy(ns, flags);
1614 LASSERT(pf != NULL);
1616 while (!list_empty(&ns->ns_unused_list)) {
1617 ldlm_policy_res_t result;
1619 /* all unused locks */
1620 if (remained-- <= 0)
1623 /* For any flags, stop scanning if @max is reached. */
1624 if (max && added >= max)
1627 list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1629 /* No locks which got blocking requests. */
1630 LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1632 if (flags & LDLM_CANCEL_NO_WAIT &&
1633 lock->l_flags & LDLM_FL_SKIPPED)
1634 /* already processed */
1637 /* Somebody is already doing CANCEL. No need for this
1638 * lock in LRU, do not traverse it again. */
1639 if (!(lock->l_flags & LDLM_FL_CANCELING))
1642 ldlm_lock_remove_from_lru_nolock(lock);
1644 if (&lock->l_lru == &ns->ns_unused_list)
1647 LDLM_LOCK_GET(lock);
1648 spin_unlock(&ns->ns_lock);
1649 lu_ref_add(&lock->l_reference, __func__, current);
1651 /* Pass the lock through the policy filter and see if it
1652 * should stay in LRU.
1654 * Even for shrinker policy we stop scanning if
1655 * we find a lock that should stay in the cache.
1656 * We should take into account lock age anyway
1657 * as a new lock is a valuable resource even if
1658 * it has a low weight.
1660 * That is, for shrinker policy we drop only
1661 * old locks, but additionally choose them by
1662 * their weight. Big extent locks will stay in
1664 result = pf(ns, lock, unused, added, count);
1665 if (result == LDLM_POLICY_KEEP_LOCK) {
1666 lu_ref_del(&lock->l_reference,
1668 LDLM_LOCK_RELEASE(lock);
1669 spin_lock(&ns->ns_lock);
1672 if (result == LDLM_POLICY_SKIP_LOCK) {
1673 lu_ref_del(&lock->l_reference,
1675 LDLM_LOCK_RELEASE(lock);
1676 spin_lock(&ns->ns_lock);
1680 lock_res_and_lock(lock);
1681 /* Check flags again under the lock. */
1682 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1683 (ldlm_lock_remove_from_lru(lock) == 0)) {
1684 /* Another thread is removing lock from LRU, or
1685 * somebody is already doing CANCEL, or there
1686 * is a blocking request which will send cancel
1687 * by itself, or the lock is no longer unused. */
1688 unlock_res_and_lock(lock);
1689 lu_ref_del(&lock->l_reference,
1691 LDLM_LOCK_RELEASE(lock);
1692 spin_lock(&ns->ns_lock);
1695 LASSERT(!lock->l_readers && !lock->l_writers);
1697 /* If we have chosen to cancel this lock voluntarily, we
1698 * better send cancel notification to server, so that it
1699 * frees appropriate state. This might lead to a race
1700 * where while we are doing cancel here, server is also
1701 * silently cancelling this lock. */
1702 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1704 /* Setting the CBPENDING flag is a little misleading,
1705 * but prevents an important race; namely, once
1706 * CBPENDING is set, the lock can accumulate no more
1707 * readers/writers. Since readers and writers are
1708 * already zero here, ldlm_lock_decref() won't see
1709 * this flag and call l_blocking_ast */
1710 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1712 /* We can't re-add to l_lru as it confuses the
1713 * refcounting in ldlm_lock_remove_from_lru() if an AST
1714 * arrives after we drop lr_lock below. We use l_bl_ast
1715 * and can't use l_pending_chain as it is used both on
1716 * server and client nevertheless bug 5666 says it is
1717 * used only on server */
1718 LASSERT(list_empty(&lock->l_bl_ast));
1719 list_add(&lock->l_bl_ast, cancels);
1720 unlock_res_and_lock(lock);
1721 lu_ref_del(&lock->l_reference, __func__, current);
1722 spin_lock(&ns->ns_lock);
1726 spin_unlock(&ns->ns_lock);
1730 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1731 int count, int max, ldlm_cancel_flags_t cancel_flags,
1735 added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1738 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1742 * Cancel at least \a nr locks from given namespace LRU.
1744 * When called with LCF_ASYNC the blocking callback will be handled
1745 * in a thread and this function will return after the thread has been
1746 * asked to call the callback. When called with LCF_ASYNC the blocking
1747 * callback will be performed in this function.
1749 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
1750 ldlm_cancel_flags_t cancel_flags,
1756 /* Just prepare the list of locks, do not actually cancel them yet.
1757 * Locks are cancelled later in a separate thread. */
1758 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1759 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
1767 * Find and cancel locally unused locks found on resource, matched to the
1768 * given policy, mode. GET the found locks and add them into the \a cancels
1771 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1772 struct list_head *cancels,
1773 ldlm_policy_data_t *policy,
1774 ldlm_mode_t mode, int lock_flags,
1775 ldlm_cancel_flags_t cancel_flags, void *opaque)
1777 struct ldlm_lock *lock;
1781 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1782 if (opaque != NULL && lock->l_ast_data != opaque) {
1783 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1784 lock->l_ast_data, opaque);
1789 if (lock->l_readers || lock->l_writers)
1792 /* If somebody is already doing CANCEL, or blocking AST came,
1793 * skip this lock. */
1794 if (lock->l_flags & LDLM_FL_BL_AST ||
1795 lock->l_flags & LDLM_FL_CANCELING)
1798 if (lockmode_compat(lock->l_granted_mode, mode))
1801 /* If policy is given and this is IBITS lock, add to list only
1802 * those locks that match by policy. */
1803 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1804 !(lock->l_policy_data.l_inodebits.bits &
1805 policy->l_inodebits.bits))
1808 /* See CBPENDING comment in ldlm_cancel_lru */
1809 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1812 LASSERT(list_empty(&lock->l_bl_ast));
1813 list_add(&lock->l_bl_ast, cancels);
1814 LDLM_LOCK_GET(lock);
1819 return ldlm_cli_cancel_list_local(cancels, count, cancel_flags);
1821 EXPORT_SYMBOL(ldlm_cancel_resource_local);
1824 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1826 * If \a req is NULL, send CANCEL request to server with handles of locks
1827 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1828 * separately per lock.
1829 * If \a req is not NULL, put handles of locks in \a cancels into the request
1830 * buffer at the offset \a off.
1831 * Destroy \a cancels at the end.
1833 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1834 struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1836 struct ldlm_lock *lock;
1839 if (list_empty(cancels) || count == 0)
1842 /* XXX: requests (both batched and not) could be sent in parallel.
1843 * Usually it is enough to have just 1 RPC, but it is possible that
1844 * there are too many locks to be cancelled in LRU or on a resource.
1845 * It would also speed up the case when the server does not support
1848 LASSERT(!list_empty(cancels));
1849 lock = list_entry(cancels->next, struct ldlm_lock,
1851 LASSERT(lock->l_conn_export);
1853 if (exp_connect_cancelset(lock->l_conn_export)) {
1856 ldlm_cancel_pack(req, cancels, count);
1858 res = ldlm_cli_cancel_req(lock->l_conn_export,
1862 res = ldlm_cli_cancel_req(lock->l_conn_export,
1867 CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1868 "ldlm_cli_cancel_list: %d\n", res);
1873 ldlm_lock_list_put(cancels, l_bl_ast, res);
1875 LASSERT(count == 0);
1878 EXPORT_SYMBOL(ldlm_cli_cancel_list);
1881 * Cancel all locks on a resource that have 0 readers/writers.
1883 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1884 * to notify the server. */
1885 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1886 const struct ldlm_res_id *res_id,
1887 ldlm_policy_data_t *policy,
1889 ldlm_cancel_flags_t flags,
1892 struct ldlm_resource *res;
1897 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1899 /* This is not a problem. */
1900 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1904 LDLM_RESOURCE_ADDREF(res);
1905 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1906 0, flags | LCF_BL_AST, opaque);
1907 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1909 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
1912 LDLM_RESOURCE_DELREF(res);
1913 ldlm_resource_putref(res);
1916 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
1918 struct ldlm_cli_cancel_arg {
1923 static int ldlm_cli_hash_cancel_unused(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1924 struct hlist_node *hnode, void *arg)
1926 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1927 struct ldlm_cli_cancel_arg *lc = arg;
1929 ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1931 lc->lc_flags, lc->lc_opaque);
1932 /* must return 0 for hash iteration */
1937 * Cancel all locks on a namespace (or a specific resource, if given)
1938 * that have 0 readers/writers.
1940 * If flags & LCF_LOCAL, throw the locks away without trying
1941 * to notify the server. */
1942 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1943 const struct ldlm_res_id *res_id,
1944 ldlm_cancel_flags_t flags, void *opaque)
1946 struct ldlm_cli_cancel_arg arg = {
1948 .lc_opaque = opaque,
1954 if (res_id != NULL) {
1955 return ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1959 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1960 ldlm_cli_hash_cancel_unused, &arg);
1964 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1966 /* Lock iterators. */
1968 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1971 struct list_head *tmp, *next;
1972 struct ldlm_lock *lock;
1973 int rc = LDLM_ITER_CONTINUE;
1976 return LDLM_ITER_CONTINUE;
1979 list_for_each_safe(tmp, next, &res->lr_granted) {
1980 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1982 if (iter(lock, closure) == LDLM_ITER_STOP)
1983 GOTO(out, rc = LDLM_ITER_STOP);
1986 list_for_each_safe(tmp, next, &res->lr_converting) {
1987 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1989 if (iter(lock, closure) == LDLM_ITER_STOP)
1990 GOTO(out, rc = LDLM_ITER_STOP);
1993 list_for_each_safe(tmp, next, &res->lr_waiting) {
1994 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1996 if (iter(lock, closure) == LDLM_ITER_STOP)
1997 GOTO(out, rc = LDLM_ITER_STOP);
2003 EXPORT_SYMBOL(ldlm_resource_foreach);
2005 struct iter_helper_data {
2006 ldlm_iterator_t iter;
2010 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2012 struct iter_helper_data *helper = closure;
2013 return helper->iter(lock, helper->closure);
2016 static int ldlm_res_iter_helper(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2017 struct hlist_node *hnode, void *arg)
2020 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2022 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2026 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2027 ldlm_iterator_t iter, void *closure)
2030 struct iter_helper_data helper = {
2035 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2036 ldlm_res_iter_helper, &helper);
2039 EXPORT_SYMBOL(ldlm_namespace_foreach);
2041 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2042 * return 0: find no resource
2043 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2046 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2047 const struct ldlm_res_id *res_id,
2048 ldlm_iterator_t iter, void *data)
2050 struct ldlm_resource *res;
2054 CERROR("must pass in namespace\n");
2058 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2062 LDLM_RESOURCE_ADDREF(res);
2063 rc = ldlm_resource_foreach(res, iter, data);
2064 LDLM_RESOURCE_DELREF(res);
2065 ldlm_resource_putref(res);
2068 EXPORT_SYMBOL(ldlm_resource_iterate);
2072 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2074 struct list_head *list = closure;
2076 /* we use l_pending_chain here, because it's unused on clients. */
2077 LASSERTF(list_empty(&lock->l_pending_chain),
2078 "lock %p next %p prev %p\n",
2079 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2080 /* bug 9573: don't replay locks left after eviction, or
2081 * bug 17614: locks being actively cancelled. Get a reference
2082 * on a lock so that it does not disapear under us (e.g. due to cancel)
2084 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
2085 list_add(&lock->l_pending_chain, list);
2086 LDLM_LOCK_GET(lock);
2089 return LDLM_ITER_CONTINUE;
2092 static int replay_lock_interpret(const struct lu_env *env,
2093 struct ptlrpc_request *req,
2094 struct ldlm_async_args *aa, int rc)
2096 struct ldlm_lock *lock;
2097 struct ldlm_reply *reply;
2098 struct obd_export *exp;
2100 atomic_dec(&req->rq_import->imp_replay_inflight);
2105 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2107 GOTO(out, rc = -EPROTO);
2109 lock = ldlm_handle2lock(&aa->lock_handle);
2111 CERROR("received replay ack for unknown local cookie "LPX64
2112 " remote cookie "LPX64 " from server %s id %s\n",
2113 aa->lock_handle.cookie, reply->lock_handle.cookie,
2114 req->rq_export->exp_client_uuid.uuid,
2115 libcfs_id2str(req->rq_peer));
2116 GOTO(out, rc = -ESTALE);
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 /* In the function below, .hs_keycmp resolves to
2123 * ldlm_export_lock_keycmp() */
2124 /* coverity[overrun-buffer-val] */
2125 cfs_hash_rehash_key(exp->exp_lock_hash,
2126 &lock->l_remote_handle,
2127 &reply->lock_handle,
2130 lock->l_remote_handle = reply->lock_handle;
2133 LDLM_DEBUG(lock, "replayed lock:");
2134 ptlrpc_import_recovery_state_machine(req->rq_import);
2135 LDLM_LOCK_PUT(lock);
2138 ptlrpc_connect_import(req->rq_import);
2143 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2145 struct ptlrpc_request *req;
2146 struct ldlm_async_args *aa;
2147 struct ldlm_request *body;
2150 /* Bug 11974: Do not replay a lock which is actively being canceled */
2151 if (lock->l_flags & LDLM_FL_CANCELING) {
2152 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2156 /* If this is reply-less callback lock, we cannot replay it, since
2157 * server might have long dropped it, but notification of that event was
2158 * lost by network. (and server granted conflicting lock already) */
2159 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2160 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2161 ldlm_lock_cancel(lock);
2166 * If granted mode matches the requested mode, this lock is granted.
2168 * If they differ, but we have a granted mode, then we were granted
2169 * one mode and now want another: ergo, converting.
2171 * If we haven't been granted anything and are on a resource list,
2172 * then we're blocked/waiting.
2174 * If we haven't been granted anything and we're NOT on a resource list,
2175 * then we haven't got a reply yet and don't have a known disposition.
2176 * This happens whenever a lock enqueue is the request that triggers
2179 if (lock->l_granted_mode == lock->l_req_mode)
2180 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2181 else if (lock->l_granted_mode)
2182 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2183 else if (!list_empty(&lock->l_res_link))
2184 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2186 flags = LDLM_FL_REPLAY;
2188 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2189 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2193 /* We're part of recovery, so don't wait for it. */
2194 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2196 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2197 ldlm_lock2desc(lock, &body->lock_desc);
2198 body->lock_flags = ldlm_flags_to_wire(flags);
2200 ldlm_lock2handle(lock, &body->lock_handle[0]);
2201 if (lock->l_lvb_len > 0)
2202 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2203 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2205 ptlrpc_request_set_replen(req);
2206 /* notify the server we've replayed all requests.
2207 * also, we mark the request to be put on a dedicated
2208 * queue to be processed after all request replayes.
2210 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2212 LDLM_DEBUG(lock, "replaying lock:");
2214 atomic_inc(&req->rq_import->imp_replay_inflight);
2215 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2216 aa = ptlrpc_req_async_args(req);
2217 aa->lock_handle = body->lock_handle[0];
2218 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2219 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2225 * Cancel as many unused locks as possible before replay. since we are
2226 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2229 * Called only in recovery before replaying locks. there is no need to
2230 * replay locks that are unused. since the clients may hold thousands of
2231 * cached unused locks, dropping the unused locks can greatly reduce the
2232 * load on the servers at recovery time.
2234 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2239 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2240 "replay for namespace %s (%d)\n",
2241 ldlm_ns_name(ns), ns->ns_nr_unused);
2243 /* We don't need to care whether or not LRU resize is enabled
2244 * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2245 * count parameter */
2246 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2247 LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2249 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2250 canceled, ldlm_ns_name(ns));
2253 int ldlm_replay_locks(struct obd_import *imp)
2255 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2257 struct ldlm_lock *lock, *next;
2260 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2262 /* don't replay locks if import failed recovery */
2263 if (imp->imp_vbr_failed)
2266 /* ensure this doesn't fall to 0 before all have been queued */
2267 atomic_inc(&imp->imp_replay_inflight);
2269 if (ldlm_cancel_unused_locks_before_replay)
2270 ldlm_cancel_unused_locks_for_replay(ns);
2272 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2274 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2275 list_del_init(&lock->l_pending_chain);
2277 LDLM_LOCK_RELEASE(lock);
2278 continue; /* or try to do the rest? */
2280 rc = replay_one_lock(imp, lock);
2281 LDLM_LOCK_RELEASE(lock);
2284 atomic_dec(&imp->imp_replay_inflight);
2288 EXPORT_SYMBOL(ldlm_replay_locks);