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 "../include/lustre_dlm.h"
65 #include "../include/obd_class.h"
66 #include "../include/obd.h"
68 #include "ldlm_internal.h"
70 int ldlm_enqueue_min
= OBD_TIMEOUT_DEFAULT
;
71 module_param(ldlm_enqueue_min
, int, 0644);
72 MODULE_PARM_DESC(ldlm_enqueue_min
, "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 unsigned long next_dump
, last_dump
;
100 LCONSOLE_WARN("lock timed out (enqueued at "CFS_TIME_T
", "
101 CFS_DURATION_T
"s ago)\n",
102 lock
->l_last_activity
,
103 cfs_time_sub(get_seconds(),
104 lock
->l_last_activity
));
105 LDLM_DEBUG(lock
, "lock timed out (enqueued at " CFS_TIME_T
", " CFS_DURATION_T
"s ago); not entering recovery in server code, just going back to sleep",
106 lock
->l_last_activity
,
107 cfs_time_sub(get_seconds(),
108 lock
->l_last_activity
));
109 if (cfs_time_after(cfs_time_current(), next_dump
)) {
110 last_dump
= next_dump
;
111 next_dump
= cfs_time_shift(300);
112 ldlm_namespace_dump(D_DLMTRACE
,
113 ldlm_lock_to_ns(lock
));
115 libcfs_debug_dumplog();
120 obd
= lock
->l_conn_export
->exp_obd
;
121 imp
= obd
->u
.cli
.cl_import
;
122 ptlrpc_fail_import(imp
, lwd
->lwd_conn_cnt
);
123 LDLM_ERROR(lock
, "lock timed out (enqueued at "CFS_TIME_T
", "
124 CFS_DURATION_T
"s ago), entering recovery for %s@%s",
125 lock
->l_last_activity
,
126 cfs_time_sub(get_seconds(), lock
->l_last_activity
),
127 obd2cli_tgt(obd
), imp
->imp_connection
->c_remote_uuid
.uuid
);
131 EXPORT_SYMBOL(ldlm_expired_completion_wait
);
133 /* We use the same basis for both server side and client side functions
134 from a single node. */
135 int ldlm_get_enq_timeout(struct ldlm_lock
*lock
)
137 int timeout
= at_get(ldlm_lock_to_ns_at(lock
));
140 return obd_timeout
/ 2;
141 /* Since these are non-updating timeouts, we should be conservative.
142 It would be nice to have some kind of "early reply" mechanism for
143 lock callbacks too... */
144 timeout
= min_t(int, at_max
, timeout
+ (timeout
>> 1)); /* 150% */
145 return max(timeout
, ldlm_enqueue_min
);
147 EXPORT_SYMBOL(ldlm_get_enq_timeout
);
150 * Helper function for ldlm_completion_ast(), updating timings when lock is
153 static int ldlm_completion_tail(struct ldlm_lock
*lock
)
158 if (lock
->l_flags
& (LDLM_FL_DESTROYED
| LDLM_FL_FAILED
)) {
159 LDLM_DEBUG(lock
, "client-side enqueue: destroyed");
162 delay
= cfs_time_sub(get_seconds(),
163 lock
->l_last_activity
);
164 LDLM_DEBUG(lock
, "client-side enqueue: granted after "
165 CFS_DURATION_T
"s", delay
);
167 /* Update our time estimate */
168 at_measured(ldlm_lock_to_ns_at(lock
),
176 * Implementation of ->l_completion_ast() for a client, that doesn't wait
177 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
178 * other threads that cannot block for long.
180 int ldlm_completion_ast_async(struct ldlm_lock
*lock
, __u64 flags
, void *data
)
182 if (flags
== LDLM_FL_WAIT_NOREPROC
) {
183 LDLM_DEBUG(lock
, "client-side enqueue waiting on pending lock");
187 if (!(flags
& (LDLM_FL_BLOCK_WAIT
| LDLM_FL_BLOCK_GRANTED
|
188 LDLM_FL_BLOCK_CONV
))) {
189 wake_up(&lock
->l_waitq
);
190 return ldlm_completion_tail(lock
);
193 LDLM_DEBUG(lock
, "client-side enqueue returned a blocked lock, going forward");
194 ldlm_reprocess_all(lock
->l_resource
);
197 EXPORT_SYMBOL(ldlm_completion_ast_async
);
200 * Generic LDLM "completion" AST. This is called in several cases:
202 * - when a reply to an ENQUEUE RPC is received from the server
203 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
204 * this point (determined by flags);
206 * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
209 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
212 * - to force all locks when resource is destroyed (cleanup_resource());
214 * - during lock conversion (not used currently).
216 * If lock is not granted in the first case, this function waits until second
217 * or penultimate cases happen in some other thread.
220 int ldlm_completion_ast(struct ldlm_lock
*lock
, __u64 flags
, void *data
)
222 /* XXX ALLOCATE - 160 bytes */
223 struct lock_wait_data lwd
;
224 struct obd_device
*obd
;
225 struct obd_import
*imp
= NULL
;
226 struct l_wait_info lwi
;
230 if (flags
== LDLM_FL_WAIT_NOREPROC
) {
231 LDLM_DEBUG(lock
, "client-side enqueue waiting on pending lock");
235 if (!(flags
& (LDLM_FL_BLOCK_WAIT
| LDLM_FL_BLOCK_GRANTED
|
236 LDLM_FL_BLOCK_CONV
))) {
237 wake_up(&lock
->l_waitq
);
241 LDLM_DEBUG(lock
, "client-side enqueue returned a blocked lock, sleeping");
245 obd
= class_exp2obd(lock
->l_conn_export
);
247 /* if this is a local lock, then there is no import */
249 imp
= obd
->u
.cli
.cl_import
;
251 /* Wait a long time for enqueue - server may have to callback a
252 lock from another client. Server will evict the other client if it
253 doesn't respond reasonably, and then give us the lock. */
254 timeout
= ldlm_get_enq_timeout(lock
) * 2;
258 if (lock
->l_flags
& LDLM_FL_NO_TIMEOUT
) {
259 LDLM_DEBUG(lock
, "waiting indefinitely because of NO_TIMEOUT");
260 lwi
= LWI_INTR(interrupted_completion_wait
, &lwd
);
262 lwi
= LWI_TIMEOUT_INTR(cfs_time_seconds(timeout
),
263 ldlm_expired_completion_wait
,
264 interrupted_completion_wait
, &lwd
);
268 spin_lock(&imp
->imp_lock
);
269 lwd
.lwd_conn_cnt
= imp
->imp_conn_cnt
;
270 spin_unlock(&imp
->imp_lock
);
273 if (ns_is_client(ldlm_lock_to_ns(lock
)) &&
274 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST
,
275 OBD_FAIL_LDLM_CP_BL_RACE
| OBD_FAIL_ONCE
)) {
276 lock
->l_flags
|= LDLM_FL_FAIL_LOC
;
279 /* Go to sleep until the lock is granted or cancelled. */
280 rc
= l_wait_event(lock
->l_waitq
,
281 is_granted_or_cancelled(lock
), &lwi
);
285 LDLM_DEBUG(lock
, "client-side enqueue waking up: failed (%d)",
290 return ldlm_completion_tail(lock
);
292 EXPORT_SYMBOL(ldlm_completion_ast
);
295 * A helper to build a blocking AST function
297 * Perform a common operation for blocking ASTs:
298 * deferred lock cancellation.
300 * \param lock the lock blocking or canceling AST was called on
302 * \see mdt_blocking_ast
303 * \see ldlm_blocking_ast
305 int ldlm_blocking_ast_nocheck(struct ldlm_lock
*lock
)
309 lock
->l_flags
|= LDLM_FL_CBPENDING
;
310 do_ast
= !lock
->l_readers
&& !lock
->l_writers
;
311 unlock_res_and_lock(lock
);
314 struct lustre_handle lockh
;
317 LDLM_DEBUG(lock
, "already unused, calling ldlm_cli_cancel");
318 ldlm_lock2handle(lock
, &lockh
);
319 rc
= ldlm_cli_cancel(&lockh
, LCF_ASYNC
);
321 CERROR("ldlm_cli_cancel: %d\n", rc
);
323 LDLM_DEBUG(lock
, "Lock still has references, will be cancelled later");
327 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck
);
330 * Server blocking AST
332 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
335 * \param lock the lock which blocks a request or cancelling lock
338 * \param flag indicates whether this cancelling or blocking callback
340 * \see ldlm_blocking_ast_nocheck
342 int ldlm_blocking_ast(struct ldlm_lock
*lock
, struct ldlm_lock_desc
*desc
,
343 void *data
, int flag
)
345 if (flag
== LDLM_CB_CANCELING
) {
346 /* Don't need to do anything here. */
350 lock_res_and_lock(lock
);
351 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
352 * that ldlm_blocking_ast is called just before intent_policy method
353 * takes the lr_lock, then by the time we get the lock, we might not
354 * be the correct blocking function anymore. So check, and return
356 if (lock
->l_blocking_ast
!= ldlm_blocking_ast
) {
357 unlock_res_and_lock(lock
);
360 return ldlm_blocking_ast_nocheck(lock
);
362 EXPORT_SYMBOL(ldlm_blocking_ast
);
365 * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
366 * comment in filter_intent_policy() on why you may need this.
368 int ldlm_glimpse_ast(struct ldlm_lock
*lock
, void *reqp
)
371 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
372 * that is rather subtle: with OST-side locking, it may so happen that
373 * _all_ extent locks are held by the OST. If client wants to obtain
374 * current file size it calls ll{,u}_glimpse_size(), and (as locks are
375 * on the server), dummy glimpse callback fires and does
376 * nothing. Client still receives correct file size due to the
377 * following fragment in filter_intent_policy():
379 * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
380 * if (rc != 0 && res->lr_namespace->ns_lvbo &&
381 * res->lr_namespace->ns_lvbo->lvbo_update) {
382 * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
385 * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
386 * returns correct file size to the client.
388 return -ELDLM_NO_LOCK_DATA
;
390 EXPORT_SYMBOL(ldlm_glimpse_ast
);
393 * Enqueue a local lock (typically on a server).
395 int ldlm_cli_enqueue_local(struct ldlm_namespace
*ns
,
396 const struct ldlm_res_id
*res_id
,
397 ldlm_type_t type
, ldlm_policy_data_t
*policy
,
398 ldlm_mode_t mode
, __u64
*flags
,
399 ldlm_blocking_callback blocking
,
400 ldlm_completion_callback completion
,
401 ldlm_glimpse_callback glimpse
,
402 void *data
, __u32 lvb_len
, enum lvb_type lvb_type
,
403 const __u64
*client_cookie
,
404 struct lustre_handle
*lockh
)
406 struct ldlm_lock
*lock
;
408 const struct ldlm_callback_suite cbs
= { .lcs_completion
= completion
,
409 .lcs_blocking
= blocking
,
410 .lcs_glimpse
= glimpse
,
413 LASSERT(!(*flags
& LDLM_FL_REPLAY
));
414 if (unlikely(ns_is_client(ns
))) {
415 CERROR("Trying to enqueue local lock in a shadow namespace\n");
419 lock
= ldlm_lock_create(ns
, res_id
, type
, mode
, &cbs
, data
, lvb_len
,
421 if (unlikely(!lock
)) {
426 ldlm_lock2handle(lock
, lockh
);
428 /* NB: we don't have any lock now (lock_res_and_lock)
429 * because it's a new lock */
430 ldlm_lock_addref_internal_nolock(lock
, mode
);
431 lock
->l_flags
|= LDLM_FL_LOCAL
;
432 if (*flags
& LDLM_FL_ATOMIC_CB
)
433 lock
->l_flags
|= LDLM_FL_ATOMIC_CB
;
436 lock
->l_policy_data
= *policy
;
437 if (client_cookie
!= NULL
)
438 lock
->l_client_cookie
= *client_cookie
;
439 if (type
== LDLM_EXTENT
)
440 lock
->l_req_extent
= policy
->l_extent
;
442 err
= ldlm_lock_enqueue(ns
, &lock
, policy
, flags
);
443 if (unlikely(err
!= ELDLM_OK
))
447 *policy
= lock
->l_policy_data
;
449 if (lock
->l_completion_ast
)
450 lock
->l_completion_ast(lock
, *flags
, NULL
);
452 LDLM_DEBUG(lock
, "client-side local enqueue handler, new lock created");
454 LDLM_LOCK_RELEASE(lock
);
458 EXPORT_SYMBOL(ldlm_cli_enqueue_local
);
460 static void failed_lock_cleanup(struct ldlm_namespace
*ns
,
461 struct ldlm_lock
*lock
, int mode
)
465 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
466 lock_res_and_lock(lock
);
467 /* Check that lock is not granted or failed, we might race. */
468 if ((lock
->l_req_mode
!= lock
->l_granted_mode
) &&
469 !(lock
->l_flags
& LDLM_FL_FAILED
)) {
470 /* Make sure that this lock will not be found by raced
471 * bl_ast and -EINVAL reply is sent to server anyways.
473 lock
->l_flags
|= LDLM_FL_LOCAL_ONLY
| LDLM_FL_FAILED
|
474 LDLM_FL_ATOMIC_CB
| LDLM_FL_CBPENDING
;
477 unlock_res_and_lock(lock
);
481 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
483 LDLM_DEBUG(lock
, "lock was granted or failed in race");
485 ldlm_lock_decref_internal(lock
, mode
);
487 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
488 * from llite/file.c/ll_file_flock(). */
489 /* This code makes for the fact that we do not have blocking handler on
490 * a client for flock locks. As such this is the place where we must
491 * completely kill failed locks. (interrupted and those that
492 * were waiting to be granted when server evicted us. */
493 if (lock
->l_resource
->lr_type
== LDLM_FLOCK
) {
494 lock_res_and_lock(lock
);
495 ldlm_resource_unlink_lock(lock
);
496 ldlm_lock_destroy_nolock(lock
);
497 unlock_res_and_lock(lock
);
502 * Finishing portion of client lock enqueue code.
504 * Called after receiving reply from server.
506 int ldlm_cli_enqueue_fini(struct obd_export
*exp
, struct ptlrpc_request
*req
,
507 ldlm_type_t type
, __u8 with_policy
, ldlm_mode_t mode
,
508 __u64
*flags
, void *lvb
, __u32 lvb_len
,
509 struct lustre_handle
*lockh
, int rc
)
511 struct ldlm_namespace
*ns
= exp
->exp_obd
->obd_namespace
;
512 int is_replay
= *flags
& LDLM_FL_REPLAY
;
513 struct ldlm_lock
*lock
;
514 struct ldlm_reply
*reply
;
515 int cleanup_phase
= 1;
518 lock
= ldlm_handle2lock(lockh
);
519 /* ldlm_cli_enqueue is holding a reference on this lock. */
521 LASSERT(type
== LDLM_FLOCK
);
525 LASSERTF(ergo(lvb_len
!= 0, lvb_len
== lock
->l_lvb_len
),
526 "lvb_len = %d, l_lvb_len = %d\n", lvb_len
, lock
->l_lvb_len
);
528 if (rc
!= ELDLM_OK
) {
530 LDLM_DEBUG(lock
, "client-side enqueue END (%s)",
531 rc
== ELDLM_LOCK_ABORTED
? "ABORTED" : "FAILED");
533 if (rc
!= ELDLM_LOCK_ABORTED
)
537 /* Before we return, swab the reply */
538 reply
= req_capsule_server_get(&req
->rq_pill
, &RMF_DLM_REP
);
545 LASSERT(lvb
!= NULL
);
547 size
= req_capsule_get_size(&req
->rq_pill
, &RMF_DLM_LVB
,
550 LDLM_ERROR(lock
, "Fail to get lvb_len, rc = %d", size
);
553 } else if (unlikely(size
> lvb_len
)) {
554 LDLM_ERROR(lock
, "Replied LVB is larger than expectation, expected = %d, replied = %d",
561 if (rc
== ELDLM_LOCK_ABORTED
) {
563 rc
= ldlm_fill_lvb(lock
, &req
->rq_pill
, RCL_SERVER
,
566 rc
= ELDLM_LOCK_ABORTED
;
570 /* lock enqueued on the server */
573 lock_res_and_lock(lock
);
574 /* Key change rehash lock in per-export hash with new key */
575 if (exp
->exp_lock_hash
) {
576 /* In the function below, .hs_keycmp resolves to
577 * ldlm_export_lock_keycmp() */
578 /* coverity[overrun-buffer-val] */
579 cfs_hash_rehash_key(exp
->exp_lock_hash
,
580 &lock
->l_remote_handle
,
584 lock
->l_remote_handle
= reply
->lock_handle
;
587 *flags
= ldlm_flags_from_wire(reply
->lock_flags
);
588 lock
->l_flags
|= ldlm_flags_from_wire(reply
->lock_flags
&
590 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
591 * to wait with no timeout as well */
592 lock
->l_flags
|= ldlm_flags_from_wire(reply
->lock_flags
&
594 unlock_res_and_lock(lock
);
596 CDEBUG(D_INFO
, "local: %p, remote cookie: %#llx, flags: 0x%llx\n",
597 lock
, reply
->lock_handle
.cookie
, *flags
);
599 /* If enqueue returned a blocked lock but the completion handler has
600 * already run, then it fixed up the resource and we don't need to do it
602 if ((*flags
) & LDLM_FL_LOCK_CHANGED
) {
603 int newmode
= reply
->lock_desc
.l_req_mode
;
606 if (newmode
&& newmode
!= lock
->l_req_mode
) {
607 LDLM_DEBUG(lock
, "server returned different mode %s",
608 ldlm_lockname
[newmode
]);
609 lock
->l_req_mode
= newmode
;
612 if (!ldlm_res_eq(&reply
->lock_desc
.l_resource
.lr_name
,
613 &lock
->l_resource
->lr_name
)) {
614 CDEBUG(D_INFO
, "remote intent success, locking "DLDLMRES
615 " instead of "DLDLMRES
"\n",
616 PLDLMRES(&reply
->lock_desc
.l_resource
),
617 PLDLMRES(lock
->l_resource
));
619 rc
= ldlm_lock_change_resource(ns
, lock
,
620 &reply
->lock_desc
.l_resource
.lr_name
);
621 if (rc
|| lock
->l_resource
== NULL
) {
625 LDLM_DEBUG(lock
, "client-side enqueue, new resource");
628 if (!(type
== LDLM_IBITS
&&
629 !(exp_connect_flags(exp
) & OBD_CONNECT_IBITS
)))
630 /* We assume lock type cannot change on server*/
631 ldlm_convert_policy_to_local(exp
,
632 lock
->l_resource
->lr_type
,
633 &reply
->lock_desc
.l_policy_data
,
634 &lock
->l_policy_data
);
635 if (type
!= LDLM_PLAIN
)
637 "client-side enqueue, new policy data");
640 if ((*flags
) & LDLM_FL_AST_SENT
||
641 /* Cancel extent locks as soon as possible on a liblustre client,
642 * because it cannot handle asynchronous ASTs robustly (see
644 (LIBLUSTRE_CLIENT
&& type
== LDLM_EXTENT
)) {
645 lock_res_and_lock(lock
);
646 lock
->l_flags
|= LDLM_FL_CBPENDING
| LDLM_FL_BL_AST
;
647 unlock_res_and_lock(lock
);
648 LDLM_DEBUG(lock
, "enqueue reply includes blocking AST");
651 /* If the lock has already been granted by a completion AST, don't
652 * clobber the LVB with an older one. */
654 /* We must lock or a racing completion might update lvb without
655 * letting us know and we'll clobber the correct value.
656 * Cannot unlock after the check either, a that still leaves
657 * a tiny window for completion to get in */
658 lock_res_and_lock(lock
);
659 if (lock
->l_req_mode
!= lock
->l_granted_mode
)
660 rc
= ldlm_fill_lvb(lock
, &req
->rq_pill
, RCL_SERVER
,
661 lock
->l_lvb_data
, size
);
662 unlock_res_and_lock(lock
);
670 rc
= ldlm_lock_enqueue(ns
, &lock
, NULL
, flags
);
671 if (lock
->l_completion_ast
!= NULL
) {
672 int err
= lock
->l_completion_ast(lock
, *flags
, NULL
);
681 if (lvb_len
&& lvb
!= NULL
) {
682 /* Copy the LVB here, and not earlier, because the completion
683 * AST (if any) can override what we got in the reply */
684 memcpy(lvb
, lock
->l_lvb_data
, lvb_len
);
687 LDLM_DEBUG(lock
, "client-side enqueue END");
689 if (cleanup_phase
== 1 && rc
)
690 failed_lock_cleanup(ns
, lock
, mode
);
691 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
693 LDLM_LOCK_RELEASE(lock
);
696 EXPORT_SYMBOL(ldlm_cli_enqueue_fini
);
699 * Estimate number of lock handles that would fit into request of given
700 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
701 * a single page on the send/receive side. XXX: 512 should be changed to
702 * more adequate value.
704 static inline int ldlm_req_handles_avail(int req_size
, int off
)
708 avail
= min_t(int, LDLM_MAXREQSIZE
, PAGE_CACHE_SIZE
- 512) - req_size
;
709 if (likely(avail
>= 0))
710 avail
/= (int)sizeof(struct lustre_handle
);
713 avail
+= LDLM_LOCKREQ_HANDLES
- off
;
718 static inline int ldlm_capsule_handles_avail(struct req_capsule
*pill
,
719 enum req_location loc
,
722 int size
= req_capsule_msg_size(pill
, loc
);
724 return ldlm_req_handles_avail(size
, off
);
727 static inline int ldlm_format_handles_avail(struct obd_import
*imp
,
728 const struct req_format
*fmt
,
729 enum req_location loc
, int off
)
731 int size
= req_capsule_fmt_size(imp
->imp_msg_magic
, fmt
, loc
);
733 return ldlm_req_handles_avail(size
, off
);
737 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
738 * \a count locks in \a cancels.
740 * This is to be called by functions preparing their own requests that
741 * might contain lists of locks to cancel in addition to actual operation
742 * that needs to be performed.
744 int ldlm_prep_elc_req(struct obd_export
*exp
, struct ptlrpc_request
*req
,
745 int version
, int opc
, int canceloff
,
746 struct list_head
*cancels
, int count
)
748 struct ldlm_namespace
*ns
= exp
->exp_obd
->obd_namespace
;
749 struct req_capsule
*pill
= &req
->rq_pill
;
750 struct ldlm_request
*dlm
= NULL
;
751 int flags
, avail
, to_free
, pack
= 0;
757 if (ns_connect_cancelset(ns
)) {
758 /* Estimate the amount of available space in the request. */
759 req_capsule_filled_sizes(pill
, RCL_CLIENT
);
760 avail
= ldlm_capsule_handles_avail(pill
, RCL_CLIENT
, canceloff
);
762 flags
= ns_connect_lru_resize(ns
) ?
763 LDLM_CANCEL_LRUR
: LDLM_CANCEL_AGED
;
764 to_free
= !ns_connect_lru_resize(ns
) &&
765 opc
== LDLM_ENQUEUE
? 1 : 0;
767 /* Cancel LRU locks here _only_ if the server supports
768 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
769 * RPC, which will make us slower. */
771 count
+= ldlm_cancel_lru_local(ns
, cancels
, to_free
,
772 avail
- count
, 0, flags
);
777 req_capsule_set_size(pill
, &RMF_DLM_REQ
, RCL_CLIENT
,
778 ldlm_request_bufsize(pack
, opc
));
781 rc
= ptlrpc_request_pack(req
, version
, opc
);
783 ldlm_lock_list_put(cancels
, l_bl_ast
, count
);
787 if (ns_connect_cancelset(ns
)) {
789 dlm
= req_capsule_client_get(pill
, &RMF_DLM_REQ
);
791 /* Skip first lock handler in ldlm_request_pack(),
792 * this method will increment @lock_count according
793 * to the lock handle amount actually written to
795 dlm
->lock_count
= canceloff
;
797 /* Pack into the request @pack lock handles. */
798 ldlm_cli_cancel_list(cancels
, pack
, req
, 0);
799 /* Prepare and send separate cancel RPC for others. */
800 ldlm_cli_cancel_list(cancels
, count
- pack
, NULL
, 0);
802 ldlm_lock_list_put(cancels
, l_bl_ast
, count
);
806 EXPORT_SYMBOL(ldlm_prep_elc_req
);
808 int ldlm_prep_enqueue_req(struct obd_export
*exp
, struct ptlrpc_request
*req
,
809 struct list_head
*cancels
, int count
)
811 return ldlm_prep_elc_req(exp
, req
, LUSTRE_DLM_VERSION
, LDLM_ENQUEUE
,
812 LDLM_ENQUEUE_CANCEL_OFF
, cancels
, count
);
814 EXPORT_SYMBOL(ldlm_prep_enqueue_req
);
816 struct ptlrpc_request
*ldlm_enqueue_pack(struct obd_export
*exp
, int lvb_len
)
818 struct ptlrpc_request
*req
;
821 req
= ptlrpc_request_alloc(class_exp2cliimp(exp
), &RQF_LDLM_ENQUEUE
);
823 return ERR_PTR(-ENOMEM
);
825 rc
= ldlm_prep_enqueue_req(exp
, req
, NULL
, 0);
827 ptlrpc_request_free(req
);
831 req_capsule_set_size(&req
->rq_pill
, &RMF_DLM_LVB
, RCL_SERVER
, lvb_len
);
832 ptlrpc_request_set_replen(req
);
835 EXPORT_SYMBOL(ldlm_enqueue_pack
);
838 * Client-side lock enqueue.
840 * If a request has some specific initialisation it is passed in \a reqp,
841 * otherwise it is created in ldlm_cli_enqueue.
843 * Supports sync and async requests, pass \a async flag accordingly. If a
844 * request was created in ldlm_cli_enqueue and it is the async request,
845 * pass it to the caller in \a reqp.
847 int ldlm_cli_enqueue(struct obd_export
*exp
, struct ptlrpc_request
**reqp
,
848 struct ldlm_enqueue_info
*einfo
,
849 const struct ldlm_res_id
*res_id
,
850 ldlm_policy_data_t
const *policy
, __u64
*flags
,
851 void *lvb
, __u32 lvb_len
, enum lvb_type lvb_type
,
852 struct lustre_handle
*lockh
, int async
)
854 struct ldlm_namespace
*ns
;
855 struct ldlm_lock
*lock
;
856 struct ldlm_request
*body
;
857 int is_replay
= *flags
& LDLM_FL_REPLAY
;
858 int req_passed_in
= 1;
860 struct ptlrpc_request
*req
;
862 LASSERT(exp
!= NULL
);
864 ns
= exp
->exp_obd
->obd_namespace
;
866 /* If we're replaying this lock, just check some invariants.
867 * If we're creating a new lock, get everything all setup nice. */
869 lock
= ldlm_handle2lock_long(lockh
, 0);
870 LASSERT(lock
!= NULL
);
871 LDLM_DEBUG(lock
, "client-side enqueue START");
872 LASSERT(exp
== lock
->l_conn_export
);
874 const struct ldlm_callback_suite cbs
= {
875 .lcs_completion
= einfo
->ei_cb_cp
,
876 .lcs_blocking
= einfo
->ei_cb_bl
,
877 .lcs_glimpse
= einfo
->ei_cb_gl
879 lock
= ldlm_lock_create(ns
, res_id
, einfo
->ei_type
,
880 einfo
->ei_mode
, &cbs
, einfo
->ei_cbdata
,
884 /* for the local lock, add the reference */
885 ldlm_lock_addref_internal(lock
, einfo
->ei_mode
);
886 ldlm_lock2handle(lock
, lockh
);
888 lock
->l_policy_data
= *policy
;
890 if (einfo
->ei_type
== LDLM_EXTENT
)
891 lock
->l_req_extent
= policy
->l_extent
;
892 LDLM_DEBUG(lock
, "client-side enqueue START, flags %llx\n",
896 lock
->l_conn_export
= exp
;
897 lock
->l_export
= NULL
;
898 lock
->l_blocking_ast
= einfo
->ei_cb_bl
;
899 lock
->l_flags
|= (*flags
& (LDLM_FL_NO_LRU
| LDLM_FL_EXCL
));
901 /* lock not sent to server yet */
903 if (reqp
== NULL
|| *reqp
== NULL
) {
904 req
= ptlrpc_request_alloc_pack(class_exp2cliimp(exp
),
909 failed_lock_cleanup(ns
, lock
, einfo
->ei_mode
);
910 LDLM_LOCK_RELEASE(lock
);
920 len
= req_capsule_get_size(&req
->rq_pill
, &RMF_DLM_REQ
,
922 LASSERTF(len
>= sizeof(*body
), "buflen[%d] = %d, not %d\n",
923 DLM_LOCKREQ_OFF
, len
, (int)sizeof(*body
));
926 /* Dump lock data into the request buffer */
927 body
= req_capsule_client_get(&req
->rq_pill
, &RMF_DLM_REQ
);
928 ldlm_lock2desc(lock
, &body
->lock_desc
);
929 body
->lock_flags
= ldlm_flags_to_wire(*flags
);
930 body
->lock_handle
[0] = *lockh
;
932 /* Continue as normal. */
933 if (!req_passed_in
) {
935 req_capsule_extend(&req
->rq_pill
,
936 &RQF_LDLM_ENQUEUE_LVB
);
937 req_capsule_set_size(&req
->rq_pill
, &RMF_DLM_LVB
, RCL_SERVER
,
939 ptlrpc_request_set_replen(req
);
943 * Liblustre client doesn't get extent locks, except for O_APPEND case
944 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
945 * [i_size, OBD_OBJECT_EOF] lock is taken.
947 LASSERT(ergo(LIBLUSTRE_CLIENT
, einfo
->ei_type
!= LDLM_EXTENT
||
948 policy
->l_extent
.end
== OBD_OBJECT_EOF
));
951 LASSERT(reqp
!= NULL
);
955 LDLM_DEBUG(lock
, "sending request");
957 rc
= ptlrpc_queue_wait(req
);
959 err
= ldlm_cli_enqueue_fini(exp
, req
, einfo
->ei_type
, policy
? 1 : 0,
960 einfo
->ei_mode
, flags
, lvb
, lvb_len
,
963 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
964 * one reference that we took */
966 LDLM_LOCK_RELEASE(lock
);
970 if (!req_passed_in
&& req
!= NULL
) {
971 ptlrpc_req_finished(req
);
978 EXPORT_SYMBOL(ldlm_cli_enqueue
);
980 static int ldlm_cli_convert_local(struct ldlm_lock
*lock
, int new_mode
,
983 struct ldlm_resource
*res
;
986 if (ns_is_client(ldlm_lock_to_ns(lock
))) {
987 CERROR("Trying to cancel local lock\n");
990 LDLM_DEBUG(lock
, "client-side local convert");
992 res
= ldlm_lock_convert(lock
, new_mode
, flags
);
994 ldlm_reprocess_all(res
);
999 LDLM_DEBUG(lock
, "client-side local convert handler END");
1000 LDLM_LOCK_PUT(lock
);
1004 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
1005 * conversion of locks which are on the waiting or converting queue */
1006 /* Caller of this code is supposed to take care of lock readers/writers
1008 int ldlm_cli_convert(struct lustre_handle
*lockh
, int new_mode
, __u32
*flags
)
1010 struct ldlm_request
*body
;
1011 struct ldlm_reply
*reply
;
1012 struct ldlm_lock
*lock
;
1013 struct ldlm_resource
*res
;
1014 struct ptlrpc_request
*req
;
1017 lock
= ldlm_handle2lock(lockh
);
1024 if (lock
->l_conn_export
== NULL
)
1025 return ldlm_cli_convert_local(lock
, new_mode
, flags
);
1027 LDLM_DEBUG(lock
, "client-side convert");
1029 req
= ptlrpc_request_alloc_pack(class_exp2cliimp(lock
->l_conn_export
),
1030 &RQF_LDLM_CONVERT
, LUSTRE_DLM_VERSION
,
1033 LDLM_LOCK_PUT(lock
);
1037 body
= req_capsule_client_get(&req
->rq_pill
, &RMF_DLM_REQ
);
1038 body
->lock_handle
[0] = lock
->l_remote_handle
;
1040 body
->lock_desc
.l_req_mode
= new_mode
;
1041 body
->lock_flags
= ldlm_flags_to_wire(*flags
);
1044 ptlrpc_request_set_replen(req
);
1045 rc
= ptlrpc_queue_wait(req
);
1049 reply
= req_capsule_server_get(&req
->rq_pill
, &RMF_DLM_REP
);
1050 if (reply
== NULL
) {
1055 if (req
->rq_status
) {
1056 rc
= req
->rq_status
;
1060 res
= ldlm_lock_convert(lock
, new_mode
, &reply
->lock_flags
);
1062 ldlm_reprocess_all(res
);
1063 /* Go to sleep until the lock is granted. */
1064 /* FIXME: or cancelled. */
1065 if (lock
->l_completion_ast
) {
1066 rc
= lock
->l_completion_ast(lock
, LDLM_FL_WAIT_NOREPROC
,
1072 rc
= LUSTRE_EDEADLK
;
1075 LDLM_LOCK_PUT(lock
);
1076 ptlrpc_req_finished(req
);
1079 EXPORT_SYMBOL(ldlm_cli_convert
);
1082 * Cancel locks locally.
1084 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1085 * \retval LDLM_FL_CANCELING otherwise;
1086 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1088 static __u64
ldlm_cli_cancel_local(struct ldlm_lock
*lock
)
1090 __u64 rc
= LDLM_FL_LOCAL_ONLY
;
1092 if (lock
->l_conn_export
) {
1095 LDLM_DEBUG(lock
, "client-side cancel");
1096 /* Set this flag to prevent others from getting new references*/
1097 lock_res_and_lock(lock
);
1098 lock
->l_flags
|= LDLM_FL_CBPENDING
;
1099 local_only
= !!(lock
->l_flags
&
1100 (LDLM_FL_LOCAL_ONLY
|LDLM_FL_CANCEL_ON_BLOCK
));
1101 ldlm_cancel_callback(lock
);
1102 rc
= (lock
->l_flags
& LDLM_FL_BL_AST
) ?
1103 LDLM_FL_BL_AST
: LDLM_FL_CANCELING
;
1104 unlock_res_and_lock(lock
);
1107 CDEBUG(D_DLMTRACE
, "not sending request (at caller's instruction)\n");
1108 rc
= LDLM_FL_LOCAL_ONLY
;
1110 ldlm_lock_cancel(lock
);
1112 if (ns_is_client(ldlm_lock_to_ns(lock
))) {
1113 LDLM_ERROR(lock
, "Trying to cancel local lock");
1116 LDLM_DEBUG(lock
, "server-side local cancel");
1117 ldlm_lock_cancel(lock
);
1118 ldlm_reprocess_all(lock
->l_resource
);
1125 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1127 static void ldlm_cancel_pack(struct ptlrpc_request
*req
,
1128 struct list_head
*head
, int count
)
1130 struct ldlm_request
*dlm
;
1131 struct ldlm_lock
*lock
;
1132 int max
, packed
= 0;
1134 dlm
= req_capsule_client_get(&req
->rq_pill
, &RMF_DLM_REQ
);
1135 LASSERT(dlm
!= NULL
);
1137 /* Check the room in the request buffer. */
1138 max
= req_capsule_get_size(&req
->rq_pill
, &RMF_DLM_REQ
, RCL_CLIENT
) -
1139 sizeof(struct ldlm_request
);
1140 max
/= sizeof(struct lustre_handle
);
1141 max
+= LDLM_LOCKREQ_HANDLES
;
1142 LASSERT(max
>= dlm
->lock_count
+ count
);
1144 /* XXX: it would be better to pack lock handles grouped by resource.
1145 * so that the server cancel would call filter_lvbo_update() less
1147 list_for_each_entry(lock
, head
, l_bl_ast
) {
1150 LASSERT(lock
->l_conn_export
);
1151 /* Pack the lock handle to the given request buffer. */
1152 LDLM_DEBUG(lock
, "packing");
1153 dlm
->lock_handle
[dlm
->lock_count
++] = lock
->l_remote_handle
;
1156 CDEBUG(D_DLMTRACE
, "%d locks packed\n", packed
);
1160 * Prepare and send a batched cancel RPC. It will include \a count lock
1161 * handles of locks given in \a cancels list. */
1162 int ldlm_cli_cancel_req(struct obd_export
*exp
, struct list_head
*cancels
,
1163 int count
, ldlm_cancel_flags_t flags
)
1165 struct ptlrpc_request
*req
= NULL
;
1166 struct obd_import
*imp
;
1170 LASSERT(exp
!= NULL
);
1173 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL
, cfs_fail_val
);
1175 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE
))
1178 free
= ldlm_format_handles_avail(class_exp2cliimp(exp
),
1179 &RQF_LDLM_CANCEL
, RCL_CLIENT
, 0);
1184 imp
= class_exp2cliimp(exp
);
1185 if (imp
== NULL
|| imp
->imp_invalid
) {
1187 "skipping cancel on invalid import %p\n", imp
);
1191 req
= ptlrpc_request_alloc(imp
, &RQF_LDLM_CANCEL
);
1197 req_capsule_filled_sizes(&req
->rq_pill
, RCL_CLIENT
);
1198 req_capsule_set_size(&req
->rq_pill
, &RMF_DLM_REQ
, RCL_CLIENT
,
1199 ldlm_request_bufsize(count
, LDLM_CANCEL
));
1201 rc
= ptlrpc_request_pack(req
, LUSTRE_DLM_VERSION
, LDLM_CANCEL
);
1203 ptlrpc_request_free(req
);
1207 req
->rq_request_portal
= LDLM_CANCEL_REQUEST_PORTAL
;
1208 req
->rq_reply_portal
= LDLM_CANCEL_REPLY_PORTAL
;
1209 ptlrpc_at_set_req_timeout(req
);
1211 ldlm_cancel_pack(req
, cancels
, count
);
1213 ptlrpc_request_set_replen(req
);
1214 if (flags
& LCF_ASYNC
) {
1215 ptlrpcd_add_req(req
);
1220 rc
= ptlrpc_queue_wait(req
);
1221 if (rc
== LUSTRE_ESTALE
) {
1222 CDEBUG(D_DLMTRACE
, "client/server (nid %s) out of sync -- not fatal\n",
1223 libcfs_nid2str(req
->rq_import
->
1224 imp_connection
->c_peer
.nid
));
1226 } else if (rc
== -ETIMEDOUT
&& /* check there was no reconnect*/
1227 req
->rq_import_generation
== imp
->imp_generation
) {
1228 ptlrpc_req_finished(req
);
1230 } else if (rc
!= ELDLM_OK
) {
1231 /* -ESHUTDOWN is common on umount */
1232 CDEBUG_LIMIT(rc
== -ESHUTDOWN
? D_DLMTRACE
: D_ERROR
,
1233 "Got rc %d from cancel RPC: canceling anyway\n",
1241 ptlrpc_req_finished(req
);
1243 return sent
? sent
: rc
;
1245 EXPORT_SYMBOL(ldlm_cli_cancel_req
);
1247 static inline struct ldlm_pool
*ldlm_imp2pl(struct obd_import
*imp
)
1249 LASSERT(imp
!= NULL
);
1250 return &imp
->imp_obd
->obd_namespace
->ns_pool
;
1254 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1256 int ldlm_cli_update_pool(struct ptlrpc_request
*req
)
1258 struct obd_device
*obd
;
1262 if (unlikely(!req
->rq_import
|| !req
->rq_import
->imp_obd
||
1263 !imp_connect_lru_resize(req
->rq_import
))) {
1265 * Do nothing for corner cases.
1270 /* In some cases RPC may contain SLV and limit zeroed out. This
1271 * is the case when server does not support LRU resize feature.
1272 * This is also possible in some recovery cases when server-side
1273 * reqs have no reference to the OBD export and thus access to
1274 * server-side namespace is not possible. */
1275 if (lustre_msg_get_slv(req
->rq_repmsg
) == 0 ||
1276 lustre_msg_get_limit(req
->rq_repmsg
) == 0) {
1277 DEBUG_REQ(D_HA
, req
,
1278 "Zero SLV or Limit found (SLV: %llu, Limit: %u)",
1279 lustre_msg_get_slv(req
->rq_repmsg
),
1280 lustre_msg_get_limit(req
->rq_repmsg
));
1284 new_limit
= lustre_msg_get_limit(req
->rq_repmsg
);
1285 new_slv
= lustre_msg_get_slv(req
->rq_repmsg
);
1286 obd
= req
->rq_import
->imp_obd
;
1288 /* Set new SLV and limit in OBD fields to make them accessible
1289 * to the pool thread. We do not access obd_namespace and pool
1290 * directly here as there is no reliable way to make sure that
1291 * they are still alive at cleanup time. Evil races are possible
1292 * which may cause Oops at that time. */
1293 write_lock(&obd
->obd_pool_lock
);
1294 obd
->obd_pool_slv
= new_slv
;
1295 obd
->obd_pool_limit
= new_limit
;
1296 write_unlock(&obd
->obd_pool_lock
);
1300 EXPORT_SYMBOL(ldlm_cli_update_pool
);
1303 * Client side lock cancel.
1305 * Lock must not have any readers or writers by this time.
1307 int ldlm_cli_cancel(struct lustre_handle
*lockh
,
1308 ldlm_cancel_flags_t cancel_flags
)
1310 struct obd_export
*exp
;
1311 int avail
, flags
, count
= 1;
1313 struct ldlm_namespace
*ns
;
1314 struct ldlm_lock
*lock
;
1317 /* concurrent cancels on the same handle can happen */
1318 lock
= ldlm_handle2lock_long(lockh
, LDLM_FL_CANCELING
);
1320 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1324 rc
= ldlm_cli_cancel_local(lock
);
1325 if (rc
== LDLM_FL_LOCAL_ONLY
|| cancel_flags
& LCF_LOCAL
) {
1326 LDLM_LOCK_RELEASE(lock
);
1329 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1330 * RPC which goes to canceld portal, so we can cancel other LRU locks
1331 * here and send them all as one LDLM_CANCEL RPC. */
1332 LASSERT(list_empty(&lock
->l_bl_ast
));
1333 list_add(&lock
->l_bl_ast
, &cancels
);
1335 exp
= lock
->l_conn_export
;
1336 if (exp_connect_cancelset(exp
)) {
1337 avail
= ldlm_format_handles_avail(class_exp2cliimp(exp
),
1342 ns
= ldlm_lock_to_ns(lock
);
1343 flags
= ns_connect_lru_resize(ns
) ?
1344 LDLM_CANCEL_LRUR
: LDLM_CANCEL_AGED
;
1345 count
+= ldlm_cancel_lru_local(ns
, &cancels
, 0, avail
- 1,
1348 ldlm_cli_cancel_list(&cancels
, count
, NULL
, cancel_flags
);
1351 EXPORT_SYMBOL(ldlm_cli_cancel
);
1354 * Locally cancel up to \a count locks in list \a cancels.
1355 * Return the number of cancelled locks.
1357 int ldlm_cli_cancel_list_local(struct list_head
*cancels
, int count
,
1358 ldlm_cancel_flags_t flags
)
1361 struct ldlm_lock
*lock
, *next
;
1362 int left
= 0, bl_ast
= 0;
1366 list_for_each_entry_safe(lock
, next
, cancels
, l_bl_ast
) {
1370 if (flags
& LCF_LOCAL
) {
1371 rc
= LDLM_FL_LOCAL_ONLY
;
1372 ldlm_lock_cancel(lock
);
1374 rc
= ldlm_cli_cancel_local(lock
);
1376 /* Until we have compound requests and can send LDLM_CANCEL
1377 * requests batched with generic RPCs, we need to send cancels
1378 * with the LDLM_FL_BL_AST flag in a separate RPC from
1379 * the one being generated now. */
1380 if (!(flags
& LCF_BL_AST
) && (rc
== LDLM_FL_BL_AST
)) {
1381 LDLM_DEBUG(lock
, "Cancel lock separately");
1382 list_del_init(&lock
->l_bl_ast
);
1383 list_add(&lock
->l_bl_ast
, &head
);
1387 if (rc
== LDLM_FL_LOCAL_ONLY
) {
1388 /* CANCEL RPC should not be sent to server. */
1389 list_del_init(&lock
->l_bl_ast
);
1390 LDLM_LOCK_RELEASE(lock
);
1396 ldlm_cli_cancel_list(&head
, bl_ast
, NULL
, 0);
1401 EXPORT_SYMBOL(ldlm_cli_cancel_list_local
);
1404 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1405 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1406 * readahead requests, ...)
1408 static ldlm_policy_res_t
ldlm_cancel_no_wait_policy(struct ldlm_namespace
*ns
,
1409 struct ldlm_lock
*lock
,
1410 int unused
, int added
,
1413 ldlm_policy_res_t result
= LDLM_POLICY_CANCEL_LOCK
;
1414 ldlm_cancel_for_recovery cb
= ns
->ns_cancel_for_recovery
;
1416 lock_res_and_lock(lock
);
1418 /* don't check added & count since we want to process all locks
1419 * from unused list */
1420 switch (lock
->l_resource
->lr_type
) {
1426 result
= LDLM_POLICY_SKIP_LOCK
;
1427 lock
->l_flags
|= LDLM_FL_SKIPPED
;
1431 unlock_res_and_lock(lock
);
1436 * Callback function for LRU-resize policy. Decides whether to keep
1437 * \a lock in LRU for current \a LRU size \a unused, added in current
1438 * scan \a added and number of locks to be preferably canceled \a count.
1440 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1442 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1444 static ldlm_policy_res_t
ldlm_cancel_lrur_policy(struct ldlm_namespace
*ns
,
1445 struct ldlm_lock
*lock
,
1446 int unused
, int added
,
1449 unsigned long cur
= cfs_time_current();
1450 struct ldlm_pool
*pl
= &ns
->ns_pool
;
1454 /* Stop LRU processing when we reach past @count or have checked all
1456 if (count
&& added
>= count
)
1457 return LDLM_POLICY_KEEP_LOCK
;
1459 slv
= ldlm_pool_get_slv(pl
);
1460 lvf
= ldlm_pool_get_lvf(pl
);
1461 la
= cfs_duration_sec(cfs_time_sub(cur
,
1462 lock
->l_last_used
));
1463 lv
= lvf
* la
* unused
;
1465 /* Inform pool about current CLV to see it via debugfs. */
1466 ldlm_pool_set_clv(pl
, lv
);
1468 /* Stop when SLV is not yet come from server or lv is smaller than
1470 return (slv
== 0 || lv
< slv
) ?
1471 LDLM_POLICY_KEEP_LOCK
: LDLM_POLICY_CANCEL_LOCK
;
1475 * Callback function for debugfs used policy. Makes decision whether to keep
1476 * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1477 * added and number of locks to be preferably canceled \a count.
1479 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1481 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1483 static ldlm_policy_res_t
ldlm_cancel_passed_policy(struct ldlm_namespace
*ns
,
1484 struct ldlm_lock
*lock
,
1485 int unused
, int added
,
1488 /* Stop LRU processing when we reach past @count or have checked all
1490 return (added
>= count
) ?
1491 LDLM_POLICY_KEEP_LOCK
: LDLM_POLICY_CANCEL_LOCK
;
1495 * Callback function for aged policy. Makes decision whether to keep \a lock in
1496 * LRU for current LRU size \a unused, added in current scan \a added and
1497 * number of locks to be preferably canceled \a count.
1499 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1501 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1503 static ldlm_policy_res_t
ldlm_cancel_aged_policy(struct ldlm_namespace
*ns
,
1504 struct ldlm_lock
*lock
,
1505 int unused
, int added
,
1508 /* Stop LRU processing if young lock is found and we reach past count */
1509 return ((added
>= count
) &&
1510 time_before(cfs_time_current(),
1511 cfs_time_add(lock
->l_last_used
, ns
->ns_max_age
))) ?
1512 LDLM_POLICY_KEEP_LOCK
: LDLM_POLICY_CANCEL_LOCK
;
1516 * Callback function for default policy. Makes decision whether to keep \a lock
1517 * in LRU for current LRU size \a unused, added in current scan \a added and
1518 * number of locks to be preferably canceled \a count.
1520 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1522 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1524 static ldlm_policy_res_t
ldlm_cancel_default_policy(struct ldlm_namespace
*ns
,
1525 struct ldlm_lock
*lock
,
1526 int unused
, int added
,
1529 /* Stop LRU processing when we reach past count or have checked all
1531 return (added
>= count
) ?
1532 LDLM_POLICY_KEEP_LOCK
: LDLM_POLICY_CANCEL_LOCK
;
1535 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t
)(struct ldlm_namespace
*,
1536 struct ldlm_lock
*, int,
1539 static ldlm_cancel_lru_policy_t
1540 ldlm_cancel_lru_policy(struct ldlm_namespace
*ns
, int flags
)
1542 if (flags
& LDLM_CANCEL_NO_WAIT
)
1543 return ldlm_cancel_no_wait_policy
;
1545 if (ns_connect_lru_resize(ns
)) {
1546 if (flags
& LDLM_CANCEL_SHRINK
)
1547 /* We kill passed number of old locks. */
1548 return ldlm_cancel_passed_policy
;
1549 else if (flags
& LDLM_CANCEL_LRUR
)
1550 return ldlm_cancel_lrur_policy
;
1551 else if (flags
& LDLM_CANCEL_PASSED
)
1552 return ldlm_cancel_passed_policy
;
1554 if (flags
& LDLM_CANCEL_AGED
)
1555 return ldlm_cancel_aged_policy
;
1558 return ldlm_cancel_default_policy
;
1562 * - Free space in LRU for \a count new locks,
1563 * redundant unused locks are canceled locally;
1564 * - also cancel locally unused aged locks;
1565 * - do not cancel more than \a max locks;
1566 * - GET the found locks and add them into the \a cancels list.
1568 * A client lock can be added to the l_bl_ast list only when it is
1569 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1570 * CANCEL. There are the following use cases:
1571 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1572 * ldlm_cli_cancel(), which check and set this flag properly. As any
1573 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1574 * later without any special locking.
1576 * Calling policies for enabled LRU resize:
1577 * ----------------------------------------
1578 * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to
1579 * cancel not more than \a count locks;
1581 * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at
1582 * the beginning of LRU list);
1584 * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to
1585 * memory pressure policy function;
1587 * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy".
1589 * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1590 * (typically before replaying locks) w/o
1591 * sending any RPCs or waiting for any
1592 * outstanding RPC to complete.
1594 static int ldlm_prepare_lru_list(struct ldlm_namespace
*ns
,
1595 struct list_head
*cancels
, int count
, int max
,
1598 ldlm_cancel_lru_policy_t pf
;
1599 struct ldlm_lock
*lock
, *next
;
1600 int added
= 0, unused
, remained
;
1602 spin_lock(&ns
->ns_lock
);
1603 unused
= ns
->ns_nr_unused
;
1606 if (!ns_connect_lru_resize(ns
))
1607 count
+= unused
- ns
->ns_max_unused
;
1609 pf
= ldlm_cancel_lru_policy(ns
, flags
);
1610 LASSERT(pf
!= NULL
);
1612 while (!list_empty(&ns
->ns_unused_list
)) {
1613 ldlm_policy_res_t result
;
1615 /* all unused locks */
1616 if (remained
-- <= 0)
1619 /* For any flags, stop scanning if @max is reached. */
1620 if (max
&& added
>= max
)
1623 list_for_each_entry_safe(lock
, next
, &ns
->ns_unused_list
,
1625 /* No locks which got blocking requests. */
1626 LASSERT(!(lock
->l_flags
& LDLM_FL_BL_AST
));
1628 if (flags
& LDLM_CANCEL_NO_WAIT
&&
1629 lock
->l_flags
& LDLM_FL_SKIPPED
)
1630 /* already processed */
1633 /* Somebody is already doing CANCEL. No need for this
1634 * lock in LRU, do not traverse it again. */
1635 if (!(lock
->l_flags
& LDLM_FL_CANCELING
))
1638 ldlm_lock_remove_from_lru_nolock(lock
);
1640 if (&lock
->l_lru
== &ns
->ns_unused_list
)
1643 LDLM_LOCK_GET(lock
);
1644 spin_unlock(&ns
->ns_lock
);
1645 lu_ref_add(&lock
->l_reference
, __func__
, current
);
1647 /* Pass the lock through the policy filter and see if it
1648 * should stay in LRU.
1650 * Even for shrinker policy we stop scanning if
1651 * we find a lock that should stay in the cache.
1652 * We should take into account lock age anyway
1653 * as a new lock is a valuable resource even if
1654 * it has a low weight.
1656 * That is, for shrinker policy we drop only
1657 * old locks, but additionally choose them by
1658 * their weight. Big extent locks will stay in
1660 result
= pf(ns
, lock
, unused
, added
, count
);
1661 if (result
== LDLM_POLICY_KEEP_LOCK
) {
1662 lu_ref_del(&lock
->l_reference
,
1664 LDLM_LOCK_RELEASE(lock
);
1665 spin_lock(&ns
->ns_lock
);
1668 if (result
== LDLM_POLICY_SKIP_LOCK
) {
1669 lu_ref_del(&lock
->l_reference
,
1671 LDLM_LOCK_RELEASE(lock
);
1672 spin_lock(&ns
->ns_lock
);
1676 lock_res_and_lock(lock
);
1677 /* Check flags again under the lock. */
1678 if ((lock
->l_flags
& LDLM_FL_CANCELING
) ||
1679 (ldlm_lock_remove_from_lru(lock
) == 0)) {
1680 /* Another thread is removing lock from LRU, or
1681 * somebody is already doing CANCEL, or there
1682 * is a blocking request which will send cancel
1683 * by itself, or the lock is no longer unused. */
1684 unlock_res_and_lock(lock
);
1685 lu_ref_del(&lock
->l_reference
,
1687 LDLM_LOCK_RELEASE(lock
);
1688 spin_lock(&ns
->ns_lock
);
1691 LASSERT(!lock
->l_readers
&& !lock
->l_writers
);
1693 /* If we have chosen to cancel this lock voluntarily, we
1694 * better send cancel notification to server, so that it
1695 * frees appropriate state. This might lead to a race
1696 * where while we are doing cancel here, server is also
1697 * silently cancelling this lock. */
1698 lock
->l_flags
&= ~LDLM_FL_CANCEL_ON_BLOCK
;
1700 /* Setting the CBPENDING flag is a little misleading,
1701 * but prevents an important race; namely, once
1702 * CBPENDING is set, the lock can accumulate no more
1703 * readers/writers. Since readers and writers are
1704 * already zero here, ldlm_lock_decref() won't see
1705 * this flag and call l_blocking_ast */
1706 lock
->l_flags
|= LDLM_FL_CBPENDING
| LDLM_FL_CANCELING
;
1708 /* We can't re-add to l_lru as it confuses the
1709 * refcounting in ldlm_lock_remove_from_lru() if an AST
1710 * arrives after we drop lr_lock below. We use l_bl_ast
1711 * and can't use l_pending_chain as it is used both on
1712 * server and client nevertheless bug 5666 says it is
1713 * used only on server */
1714 LASSERT(list_empty(&lock
->l_bl_ast
));
1715 list_add(&lock
->l_bl_ast
, cancels
);
1716 unlock_res_and_lock(lock
);
1717 lu_ref_del(&lock
->l_reference
, __func__
, current
);
1718 spin_lock(&ns
->ns_lock
);
1722 spin_unlock(&ns
->ns_lock
);
1726 int ldlm_cancel_lru_local(struct ldlm_namespace
*ns
, struct list_head
*cancels
,
1727 int count
, int max
, ldlm_cancel_flags_t cancel_flags
,
1732 added
= ldlm_prepare_lru_list(ns
, cancels
, count
, max
, flags
);
1735 return ldlm_cli_cancel_list_local(cancels
, added
, cancel_flags
);
1739 * Cancel at least \a nr locks from given namespace LRU.
1741 * When called with LCF_ASYNC the blocking callback will be handled
1742 * in a thread and this function will return after the thread has been
1743 * asked to call the callback. When called with LCF_ASYNC the blocking
1744 * callback will be performed in this function.
1746 int ldlm_cancel_lru(struct ldlm_namespace
*ns
, int nr
,
1747 ldlm_cancel_flags_t cancel_flags
,
1753 /* Just prepare the list of locks, do not actually cancel them yet.
1754 * Locks are cancelled later in a separate thread. */
1755 count
= ldlm_prepare_lru_list(ns
, &cancels
, nr
, 0, flags
);
1756 rc
= ldlm_bl_to_thread_list(ns
, NULL
, &cancels
, count
, cancel_flags
);
1764 * Find and cancel locally unused locks found on resource, matched to the
1765 * given policy, mode. GET the found locks and add them into the \a cancels
1768 int ldlm_cancel_resource_local(struct ldlm_resource
*res
,
1769 struct list_head
*cancels
,
1770 ldlm_policy_data_t
*policy
,
1771 ldlm_mode_t mode
, __u64 lock_flags
,
1772 ldlm_cancel_flags_t cancel_flags
, void *opaque
)
1774 struct ldlm_lock
*lock
;
1778 list_for_each_entry(lock
, &res
->lr_granted
, l_res_link
) {
1779 if (opaque
!= NULL
&& lock
->l_ast_data
!= opaque
) {
1780 LDLM_ERROR(lock
, "data %p doesn't match opaque %p",
1781 lock
->l_ast_data
, opaque
);
1785 if (lock
->l_readers
|| lock
->l_writers
)
1788 /* If somebody is already doing CANCEL, or blocking AST came,
1789 * skip this lock. */
1790 if (lock
->l_flags
& LDLM_FL_BL_AST
||
1791 lock
->l_flags
& LDLM_FL_CANCELING
)
1794 if (lockmode_compat(lock
->l_granted_mode
, mode
))
1797 /* If policy is given and this is IBITS lock, add to list only
1798 * those locks that match by policy. */
1799 if (policy
&& (lock
->l_resource
->lr_type
== LDLM_IBITS
) &&
1800 !(lock
->l_policy_data
.l_inodebits
.bits
&
1801 policy
->l_inodebits
.bits
))
1804 /* See CBPENDING comment in ldlm_cancel_lru */
1805 lock
->l_flags
|= LDLM_FL_CBPENDING
| LDLM_FL_CANCELING
|
1808 LASSERT(list_empty(&lock
->l_bl_ast
));
1809 list_add(&lock
->l_bl_ast
, cancels
);
1810 LDLM_LOCK_GET(lock
);
1815 return ldlm_cli_cancel_list_local(cancels
, count
, cancel_flags
);
1817 EXPORT_SYMBOL(ldlm_cancel_resource_local
);
1820 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1822 * If \a req is NULL, send CANCEL request to server with handles of locks
1823 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1824 * separately per lock.
1825 * If \a req is not NULL, put handles of locks in \a cancels into the request
1826 * buffer at the offset \a off.
1827 * Destroy \a cancels at the end.
1829 int ldlm_cli_cancel_list(struct list_head
*cancels
, int count
,
1830 struct ptlrpc_request
*req
, ldlm_cancel_flags_t flags
)
1832 struct ldlm_lock
*lock
;
1835 if (list_empty(cancels
) || count
== 0)
1838 /* XXX: requests (both batched and not) could be sent in parallel.
1839 * Usually it is enough to have just 1 RPC, but it is possible that
1840 * there are too many locks to be cancelled in LRU or on a resource.
1841 * It would also speed up the case when the server does not support
1844 LASSERT(!list_empty(cancels
));
1845 lock
= list_entry(cancels
->next
, struct ldlm_lock
,
1847 LASSERT(lock
->l_conn_export
);
1849 if (exp_connect_cancelset(lock
->l_conn_export
)) {
1852 ldlm_cancel_pack(req
, cancels
, count
);
1854 res
= ldlm_cli_cancel_req(lock
->l_conn_export
,
1858 res
= ldlm_cli_cancel_req(lock
->l_conn_export
,
1863 CDEBUG_LIMIT(res
== -ESHUTDOWN
? D_DLMTRACE
: D_ERROR
,
1864 "ldlm_cli_cancel_list: %d\n", res
);
1869 ldlm_lock_list_put(cancels
, l_bl_ast
, res
);
1871 LASSERT(count
== 0);
1874 EXPORT_SYMBOL(ldlm_cli_cancel_list
);
1877 * Cancel all locks on a resource that have 0 readers/writers.
1879 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1880 * to notify the server. */
1881 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace
*ns
,
1882 const struct ldlm_res_id
*res_id
,
1883 ldlm_policy_data_t
*policy
,
1885 ldlm_cancel_flags_t flags
,
1888 struct ldlm_resource
*res
;
1893 res
= ldlm_resource_get(ns
, NULL
, res_id
, 0, 0);
1895 /* This is not a problem. */
1896 CDEBUG(D_INFO
, "No resource %llu\n", res_id
->name
[0]);
1900 LDLM_RESOURCE_ADDREF(res
);
1901 count
= ldlm_cancel_resource_local(res
, &cancels
, policy
, mode
,
1902 0, flags
| LCF_BL_AST
, opaque
);
1903 rc
= ldlm_cli_cancel_list(&cancels
, count
, NULL
, flags
);
1905 CERROR("canceling unused lock "DLDLMRES
": rc = %d\n",
1908 LDLM_RESOURCE_DELREF(res
);
1909 ldlm_resource_putref(res
);
1912 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource
);
1914 struct ldlm_cli_cancel_arg
{
1919 static int ldlm_cli_hash_cancel_unused(struct cfs_hash
*hs
,
1920 struct cfs_hash_bd
*bd
,
1921 struct hlist_node
*hnode
, void *arg
)
1923 struct ldlm_resource
*res
= cfs_hash_object(hs
, hnode
);
1924 struct ldlm_cli_cancel_arg
*lc
= arg
;
1926 ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res
), &res
->lr_name
,
1928 lc
->lc_flags
, lc
->lc_opaque
);
1929 /* must return 0 for hash iteration */
1934 * Cancel all locks on a namespace (or a specific resource, if given)
1935 * that have 0 readers/writers.
1937 * If flags & LCF_LOCAL, throw the locks away without trying
1938 * to notify the server. */
1939 int ldlm_cli_cancel_unused(struct ldlm_namespace
*ns
,
1940 const struct ldlm_res_id
*res_id
,
1941 ldlm_cancel_flags_t flags
, void *opaque
)
1943 struct ldlm_cli_cancel_arg arg
= {
1945 .lc_opaque
= opaque
,
1951 if (res_id
!= NULL
) {
1952 return ldlm_cli_cancel_unused_resource(ns
, res_id
, NULL
,
1956 cfs_hash_for_each_nolock(ns
->ns_rs_hash
,
1957 ldlm_cli_hash_cancel_unused
, &arg
);
1961 EXPORT_SYMBOL(ldlm_cli_cancel_unused
);
1963 /* Lock iterators. */
1965 int ldlm_resource_foreach(struct ldlm_resource
*res
, ldlm_iterator_t iter
,
1968 struct list_head
*tmp
, *next
;
1969 struct ldlm_lock
*lock
;
1970 int rc
= LDLM_ITER_CONTINUE
;
1973 return LDLM_ITER_CONTINUE
;
1976 list_for_each_safe(tmp
, next
, &res
->lr_granted
) {
1977 lock
= list_entry(tmp
, struct ldlm_lock
, l_res_link
);
1979 if (iter(lock
, closure
) == LDLM_ITER_STOP
) {
1980 rc
= LDLM_ITER_STOP
;
1985 list_for_each_safe(tmp
, next
, &res
->lr_converting
) {
1986 lock
= list_entry(tmp
, struct ldlm_lock
, l_res_link
);
1988 if (iter(lock
, closure
) == LDLM_ITER_STOP
) {
1989 rc
= LDLM_ITER_STOP
;
1994 list_for_each_safe(tmp
, next
, &res
->lr_waiting
) {
1995 lock
= list_entry(tmp
, struct ldlm_lock
, l_res_link
);
1997 if (iter(lock
, closure
) == LDLM_ITER_STOP
) {
1998 rc
= LDLM_ITER_STOP
;
2006 EXPORT_SYMBOL(ldlm_resource_foreach
);
2008 struct iter_helper_data
{
2009 ldlm_iterator_t iter
;
2013 static int ldlm_iter_helper(struct ldlm_lock
*lock
, void *closure
)
2015 struct iter_helper_data
*helper
= closure
;
2017 return helper
->iter(lock
, helper
->closure
);
2020 static int ldlm_res_iter_helper(struct cfs_hash
*hs
, struct cfs_hash_bd
*bd
,
2021 struct hlist_node
*hnode
, void *arg
)
2024 struct ldlm_resource
*res
= cfs_hash_object(hs
, hnode
);
2026 return ldlm_resource_foreach(res
, ldlm_iter_helper
, arg
) ==
2030 void ldlm_namespace_foreach(struct ldlm_namespace
*ns
,
2031 ldlm_iterator_t iter
, void *closure
)
2034 struct iter_helper_data helper
= {
2039 cfs_hash_for_each_nolock(ns
->ns_rs_hash
,
2040 ldlm_res_iter_helper
, &helper
);
2043 EXPORT_SYMBOL(ldlm_namespace_foreach
);
2045 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2046 * return 0: find no resource
2047 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2050 int ldlm_resource_iterate(struct ldlm_namespace
*ns
,
2051 const struct ldlm_res_id
*res_id
,
2052 ldlm_iterator_t iter
, void *data
)
2054 struct ldlm_resource
*res
;
2058 CERROR("must pass in namespace\n");
2062 res
= ldlm_resource_get(ns
, NULL
, res_id
, 0, 0);
2066 LDLM_RESOURCE_ADDREF(res
);
2067 rc
= ldlm_resource_foreach(res
, iter
, data
);
2068 LDLM_RESOURCE_DELREF(res
);
2069 ldlm_resource_putref(res
);
2072 EXPORT_SYMBOL(ldlm_resource_iterate
);
2076 static int ldlm_chain_lock_for_replay(struct ldlm_lock
*lock
, void *closure
)
2078 struct list_head
*list
= closure
;
2080 /* we use l_pending_chain here, because it's unused on clients. */
2081 LASSERTF(list_empty(&lock
->l_pending_chain
),
2082 "lock %p next %p prev %p\n",
2083 lock
, &lock
->l_pending_chain
.next
,
2084 &lock
->l_pending_chain
.prev
);
2085 /* bug 9573: don't replay locks left after eviction, or
2086 * bug 17614: locks being actively cancelled. Get a reference
2087 * on a lock so that it does not disappear under us (e.g. due to cancel)
2089 if (!(lock
->l_flags
& (LDLM_FL_FAILED
|LDLM_FL_CANCELING
))) {
2090 list_add(&lock
->l_pending_chain
, list
);
2091 LDLM_LOCK_GET(lock
);
2094 return LDLM_ITER_CONTINUE
;
2097 static int replay_lock_interpret(const struct lu_env
*env
,
2098 struct ptlrpc_request
*req
,
2099 struct ldlm_async_args
*aa
, int rc
)
2101 struct ldlm_lock
*lock
;
2102 struct ldlm_reply
*reply
;
2103 struct obd_export
*exp
;
2105 atomic_dec(&req
->rq_import
->imp_replay_inflight
);
2110 reply
= req_capsule_server_get(&req
->rq_pill
, &RMF_DLM_REP
);
2111 if (reply
== NULL
) {
2116 lock
= ldlm_handle2lock(&aa
->lock_handle
);
2118 CERROR("received replay ack for unknown local cookie %#llx remote cookie %#llx from server %s id %s\n",
2119 aa
->lock_handle
.cookie
, reply
->lock_handle
.cookie
,
2120 req
->rq_export
->exp_client_uuid
.uuid
,
2121 libcfs_id2str(req
->rq_peer
));
2126 /* Key change rehash lock in per-export hash with new key */
2127 exp
= req
->rq_export
;
2128 if (exp
&& exp
->exp_lock_hash
) {
2129 /* In the function below, .hs_keycmp resolves to
2130 * ldlm_export_lock_keycmp() */
2131 /* coverity[overrun-buffer-val] */
2132 cfs_hash_rehash_key(exp
->exp_lock_hash
,
2133 &lock
->l_remote_handle
,
2134 &reply
->lock_handle
,
2137 lock
->l_remote_handle
= reply
->lock_handle
;
2140 LDLM_DEBUG(lock
, "replayed lock:");
2141 ptlrpc_import_recovery_state_machine(req
->rq_import
);
2142 LDLM_LOCK_PUT(lock
);
2145 ptlrpc_connect_import(req
->rq_import
);
2150 static int replay_one_lock(struct obd_import
*imp
, struct ldlm_lock
*lock
)
2152 struct ptlrpc_request
*req
;
2153 struct ldlm_async_args
*aa
;
2154 struct ldlm_request
*body
;
2157 /* Bug 11974: Do not replay a lock which is actively being canceled */
2158 if (lock
->l_flags
& LDLM_FL_CANCELING
) {
2159 LDLM_DEBUG(lock
, "Not replaying canceled lock:");
2163 /* If this is reply-less callback lock, we cannot replay it, since
2164 * server might have long dropped it, but notification of that event was
2165 * lost by network. (and server granted conflicting lock already) */
2166 if (lock
->l_flags
& LDLM_FL_CANCEL_ON_BLOCK
) {
2167 LDLM_DEBUG(lock
, "Not replaying reply-less lock:");
2168 ldlm_lock_cancel(lock
);
2173 * If granted mode matches the requested mode, this lock is granted.
2175 * If they differ, but we have a granted mode, then we were granted
2176 * one mode and now want another: ergo, converting.
2178 * If we haven't been granted anything and are on a resource list,
2179 * then we're blocked/waiting.
2181 * If we haven't been granted anything and we're NOT on a resource list,
2182 * then we haven't got a reply yet and don't have a known disposition.
2183 * This happens whenever a lock enqueue is the request that triggers
2186 if (lock
->l_granted_mode
== lock
->l_req_mode
)
2187 flags
= LDLM_FL_REPLAY
| LDLM_FL_BLOCK_GRANTED
;
2188 else if (lock
->l_granted_mode
)
2189 flags
= LDLM_FL_REPLAY
| LDLM_FL_BLOCK_CONV
;
2190 else if (!list_empty(&lock
->l_res_link
))
2191 flags
= LDLM_FL_REPLAY
| LDLM_FL_BLOCK_WAIT
;
2193 flags
= LDLM_FL_REPLAY
;
2195 req
= ptlrpc_request_alloc_pack(imp
, &RQF_LDLM_ENQUEUE
,
2196 LUSTRE_DLM_VERSION
, LDLM_ENQUEUE
);
2200 /* We're part of recovery, so don't wait for it. */
2201 req
->rq_send_state
= LUSTRE_IMP_REPLAY_LOCKS
;
2203 body
= req_capsule_client_get(&req
->rq_pill
, &RMF_DLM_REQ
);
2204 ldlm_lock2desc(lock
, &body
->lock_desc
);
2205 body
->lock_flags
= ldlm_flags_to_wire(flags
);
2207 ldlm_lock2handle(lock
, &body
->lock_handle
[0]);
2208 if (lock
->l_lvb_len
> 0)
2209 req_capsule_extend(&req
->rq_pill
, &RQF_LDLM_ENQUEUE_LVB
);
2210 req_capsule_set_size(&req
->rq_pill
, &RMF_DLM_LVB
, RCL_SERVER
,
2212 ptlrpc_request_set_replen(req
);
2213 /* notify the server we've replayed all requests.
2214 * also, we mark the request to be put on a dedicated
2215 * queue to be processed after all request replayes.
2217 lustre_msg_set_flags(req
->rq_reqmsg
, MSG_REQ_REPLAY_DONE
);
2219 LDLM_DEBUG(lock
, "replaying lock:");
2221 atomic_inc(&req
->rq_import
->imp_replay_inflight
);
2222 CLASSERT(sizeof(*aa
) <= sizeof(req
->rq_async_args
));
2223 aa
= ptlrpc_req_async_args(req
);
2224 aa
->lock_handle
= body
->lock_handle
[0];
2225 req
->rq_interpret_reply
= (ptlrpc_interpterer_t
)replay_lock_interpret
;
2226 ptlrpcd_add_req(req
);
2232 * Cancel as many unused locks as possible before replay. since we are
2233 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2236 * Called only in recovery before replaying locks. there is no need to
2237 * replay locks that are unused. since the clients may hold thousands of
2238 * cached unused locks, dropping the unused locks can greatly reduce the
2239 * load on the servers at recovery time.
2241 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace
*ns
)
2246 CDEBUG(D_DLMTRACE
, "Dropping as many unused locks as possible before replay for namespace %s (%d)\n",
2247 ldlm_ns_name(ns
), ns
->ns_nr_unused
);
2249 /* We don't need to care whether or not LRU resize is enabled
2250 * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2251 * count parameter */
2252 canceled
= ldlm_cancel_lru_local(ns
, &cancels
, ns
->ns_nr_unused
, 0,
2253 LCF_LOCAL
, LDLM_CANCEL_NO_WAIT
);
2255 CDEBUG(D_DLMTRACE
, "Canceled %d unused locks from namespace %s\n",
2256 canceled
, ldlm_ns_name(ns
));
2259 int ldlm_replay_locks(struct obd_import
*imp
)
2261 struct ldlm_namespace
*ns
= imp
->imp_obd
->obd_namespace
;
2263 struct ldlm_lock
*lock
, *next
;
2266 LASSERT(atomic_read(&imp
->imp_replay_inflight
) == 0);
2268 /* don't replay locks if import failed recovery */
2269 if (imp
->imp_vbr_failed
)
2272 /* ensure this doesn't fall to 0 before all have been queued */
2273 atomic_inc(&imp
->imp_replay_inflight
);
2275 if (ldlm_cancel_unused_locks_before_replay
)
2276 ldlm_cancel_unused_locks_for_replay(ns
);
2278 ldlm_namespace_foreach(ns
, ldlm_chain_lock_for_replay
, &list
);
2280 list_for_each_entry_safe(lock
, next
, &list
, l_pending_chain
) {
2281 list_del_init(&lock
->l_pending_chain
);
2283 LDLM_LOCK_RELEASE(lock
);
2284 continue; /* or try to do the rest? */
2286 rc
= replay_one_lock(imp
, lock
);
2287 LDLM_LOCK_RELEASE(lock
);
2290 atomic_dec(&imp
->imp_replay_inflight
);
2294 EXPORT_SYMBOL(ldlm_replay_locks
);