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
= 0, last_dump
= 0;
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
", "
106 CFS_DURATION_T
"s ago); not entering recovery in "
107 "server code, just going back to sleep",
108 lock
->l_last_activity
,
109 cfs_time_sub(get_seconds(),
110 lock
->l_last_activity
));
111 if (cfs_time_after(cfs_time_current(), next_dump
)) {
112 last_dump
= next_dump
;
113 next_dump
= cfs_time_shift(300);
114 ldlm_namespace_dump(D_DLMTRACE
,
115 ldlm_lock_to_ns(lock
));
117 libcfs_debug_dumplog();
122 obd
= lock
->l_conn_export
->exp_obd
;
123 imp
= obd
->u
.cli
.cl_import
;
124 ptlrpc_fail_import(imp
, lwd
->lwd_conn_cnt
);
125 LDLM_ERROR(lock
, "lock timed out (enqueued at "CFS_TIME_T
", "
126 CFS_DURATION_T
"s ago), entering recovery for %s@%s",
127 lock
->l_last_activity
,
128 cfs_time_sub(get_seconds(), lock
->l_last_activity
),
129 obd2cli_tgt(obd
), imp
->imp_connection
->c_remote_uuid
.uuid
);
133 EXPORT_SYMBOL(ldlm_expired_completion_wait
);
135 /* We use the same basis for both server side and client side functions
136 from a single node. */
137 int ldlm_get_enq_timeout(struct ldlm_lock
*lock
)
139 int timeout
= at_get(ldlm_lock_to_ns_at(lock
));
141 return obd_timeout
/ 2;
142 /* Since these are non-updating timeouts, we should be conservative.
143 It would be nice to have some kind of "early reply" mechanism for
144 lock callbacks too... */
145 timeout
= min_t(int, at_max
, timeout
+ (timeout
>> 1)); /* 150% */
146 return max(timeout
, ldlm_enqueue_min
);
148 EXPORT_SYMBOL(ldlm_get_enq_timeout
);
151 * Helper function for ldlm_completion_ast(), updating timings when lock is
154 static int ldlm_completion_tail(struct ldlm_lock
*lock
)
159 if (lock
->l_flags
& (LDLM_FL_DESTROYED
| LDLM_FL_FAILED
)) {
160 LDLM_DEBUG(lock
, "client-side enqueue: destroyed");
163 delay
= cfs_time_sub(get_seconds(),
164 lock
->l_last_activity
);
165 LDLM_DEBUG(lock
, "client-side enqueue: granted after "
166 CFS_DURATION_T
"s", delay
);
168 /* Update our time estimate */
169 at_measured(ldlm_lock_to_ns_at(lock
),
177 * Implementation of ->l_completion_ast() for a client, that doesn't wait
178 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
179 * other threads that cannot block for long.
181 int ldlm_completion_ast_async(struct ldlm_lock
*lock
, __u64 flags
, void *data
)
183 if (flags
== LDLM_FL_WAIT_NOREPROC
) {
184 LDLM_DEBUG(lock
, "client-side enqueue waiting on pending lock");
188 if (!(flags
& (LDLM_FL_BLOCK_WAIT
| LDLM_FL_BLOCK_GRANTED
|
189 LDLM_FL_BLOCK_CONV
))) {
190 wake_up(&lock
->l_waitq
);
191 return ldlm_completion_tail(lock
);
194 LDLM_DEBUG(lock
, "client-side enqueue returned a blocked lock, "
196 ldlm_reprocess_all(lock
->l_resource
);
199 EXPORT_SYMBOL(ldlm_completion_ast_async
);
202 * Generic LDLM "completion" AST. This is called in several cases:
204 * - when a reply to an ENQUEUE RPC is received from the server
205 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
206 * this point (determined by flags);
208 * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
211 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
214 * - to force all locks when resource is destroyed (cleanup_resource());
216 * - during lock conversion (not used currently).
218 * If lock is not granted in the first case, this function waits until second
219 * or penultimate cases happen in some other thread.
222 int ldlm_completion_ast(struct ldlm_lock
*lock
, __u64 flags
, void *data
)
224 /* XXX ALLOCATE - 160 bytes */
225 struct lock_wait_data lwd
;
226 struct obd_device
*obd
;
227 struct obd_import
*imp
= NULL
;
228 struct l_wait_info lwi
;
232 if (flags
== LDLM_FL_WAIT_NOREPROC
) {
233 LDLM_DEBUG(lock
, "client-side enqueue waiting on pending lock");
237 if (!(flags
& (LDLM_FL_BLOCK_WAIT
| LDLM_FL_BLOCK_GRANTED
|
238 LDLM_FL_BLOCK_CONV
))) {
239 wake_up(&lock
->l_waitq
);
243 LDLM_DEBUG(lock
, "client-side enqueue returned a blocked lock, "
248 obd
= class_exp2obd(lock
->l_conn_export
);
250 /* if this is a local lock, then there is no import */
252 imp
= obd
->u
.cli
.cl_import
;
255 /* Wait a long time for enqueue - server may have to callback a
256 lock from another client. Server will evict the other client if it
257 doesn't respond reasonably, and then give us the lock. */
258 timeout
= ldlm_get_enq_timeout(lock
) * 2;
262 if (lock
->l_flags
& LDLM_FL_NO_TIMEOUT
) {
263 LDLM_DEBUG(lock
, "waiting indefinitely because of NO_TIMEOUT");
264 lwi
= LWI_INTR(interrupted_completion_wait
, &lwd
);
266 lwi
= LWI_TIMEOUT_INTR(cfs_time_seconds(timeout
),
267 ldlm_expired_completion_wait
,
268 interrupted_completion_wait
, &lwd
);
272 spin_lock(&imp
->imp_lock
);
273 lwd
.lwd_conn_cnt
= imp
->imp_conn_cnt
;
274 spin_unlock(&imp
->imp_lock
);
277 if (ns_is_client(ldlm_lock_to_ns(lock
)) &&
278 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST
,
279 OBD_FAIL_LDLM_CP_BL_RACE
| OBD_FAIL_ONCE
)) {
280 lock
->l_flags
|= LDLM_FL_FAIL_LOC
;
283 /* Go to sleep until the lock is granted or cancelled. */
284 rc
= l_wait_event(lock
->l_waitq
,
285 is_granted_or_cancelled(lock
), &lwi
);
289 LDLM_DEBUG(lock
, "client-side enqueue waking up: failed (%d)",
294 return ldlm_completion_tail(lock
);
296 EXPORT_SYMBOL(ldlm_completion_ast
);
299 * A helper to build a blocking AST function
301 * Perform a common operation for blocking ASTs:
302 * deferred lock cancellation.
304 * \param lock the lock blocking or canceling AST was called on
306 * \see mdt_blocking_ast
307 * \see ldlm_blocking_ast
309 int ldlm_blocking_ast_nocheck(struct ldlm_lock
*lock
)
313 lock
->l_flags
|= LDLM_FL_CBPENDING
;
314 do_ast
= (!lock
->l_readers
&& !lock
->l_writers
);
315 unlock_res_and_lock(lock
);
318 struct lustre_handle lockh
;
321 LDLM_DEBUG(lock
, "already unused, calling ldlm_cli_cancel");
322 ldlm_lock2handle(lock
, &lockh
);
323 rc
= ldlm_cli_cancel(&lockh
, LCF_ASYNC
);
325 CERROR("ldlm_cli_cancel: %d\n", rc
);
327 LDLM_DEBUG(lock
, "Lock still has references, will be "
332 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck
);
335 * Server blocking AST
337 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
340 * \param lock the lock which blocks a request or cancelling lock
343 * \param flag indicates whether this cancelling or blocking callback
345 * \see ldlm_blocking_ast_nocheck
347 int ldlm_blocking_ast(struct ldlm_lock
*lock
, struct ldlm_lock_desc
*desc
,
348 void *data
, int flag
)
350 if (flag
== LDLM_CB_CANCELING
) {
351 /* Don't need to do anything here. */
355 lock_res_and_lock(lock
);
356 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
357 * that ldlm_blocking_ast is called just before intent_policy method
358 * takes the lr_lock, then by the time we get the lock, we might not
359 * be the correct blocking function anymore. So check, and return
361 if (lock
->l_blocking_ast
!= ldlm_blocking_ast
) {
362 unlock_res_and_lock(lock
);
365 return ldlm_blocking_ast_nocheck(lock
);
367 EXPORT_SYMBOL(ldlm_blocking_ast
);
370 * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
371 * comment in filter_intent_policy() on why you may need this.
373 int ldlm_glimpse_ast(struct ldlm_lock
*lock
, void *reqp
)
376 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
377 * that is rather subtle: with OST-side locking, it may so happen that
378 * _all_ extent locks are held by the OST. If client wants to obtain
379 * current file size it calls ll{,u}_glimpse_size(), and (as locks are
380 * on the server), dummy glimpse callback fires and does
381 * nothing. Client still receives correct file size due to the
382 * following fragment in filter_intent_policy():
384 * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
385 * if (rc != 0 && res->lr_namespace->ns_lvbo &&
386 * res->lr_namespace->ns_lvbo->lvbo_update) {
387 * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
390 * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
391 * returns correct file size to the client.
393 return -ELDLM_NO_LOCK_DATA
;
395 EXPORT_SYMBOL(ldlm_glimpse_ast
);
398 * Enqueue a local lock (typically on a server).
400 int ldlm_cli_enqueue_local(struct ldlm_namespace
*ns
,
401 const struct ldlm_res_id
*res_id
,
402 ldlm_type_t type
, ldlm_policy_data_t
*policy
,
403 ldlm_mode_t mode
, __u64
*flags
,
404 ldlm_blocking_callback blocking
,
405 ldlm_completion_callback completion
,
406 ldlm_glimpse_callback glimpse
,
407 void *data
, __u32 lvb_len
, enum lvb_type lvb_type
,
408 const __u64
*client_cookie
,
409 struct lustre_handle
*lockh
)
411 struct ldlm_lock
*lock
;
413 const struct ldlm_callback_suite cbs
= { .lcs_completion
= completion
,
414 .lcs_blocking
= blocking
,
415 .lcs_glimpse
= glimpse
,
418 LASSERT(!(*flags
& LDLM_FL_REPLAY
));
419 if (unlikely(ns_is_client(ns
))) {
420 CERROR("Trying to enqueue local lock in a shadow namespace\n");
424 lock
= ldlm_lock_create(ns
, res_id
, type
, mode
, &cbs
, data
, lvb_len
,
427 GOTO(out_nolock
, err
= -ENOMEM
);
429 ldlm_lock2handle(lock
, lockh
);
431 /* NB: we don't have any lock now (lock_res_and_lock)
432 * because it's a new lock */
433 ldlm_lock_addref_internal_nolock(lock
, mode
);
434 lock
->l_flags
|= LDLM_FL_LOCAL
;
435 if (*flags
& LDLM_FL_ATOMIC_CB
)
436 lock
->l_flags
|= LDLM_FL_ATOMIC_CB
;
439 lock
->l_policy_data
= *policy
;
440 if (client_cookie
!= NULL
)
441 lock
->l_client_cookie
= *client_cookie
;
442 if (type
== LDLM_EXTENT
)
443 lock
->l_req_extent
= policy
->l_extent
;
445 err
= ldlm_lock_enqueue(ns
, &lock
, policy
, flags
);
446 if (unlikely(err
!= ELDLM_OK
))
450 *policy
= lock
->l_policy_data
;
452 if (lock
->l_completion_ast
)
453 lock
->l_completion_ast(lock
, *flags
, NULL
);
455 LDLM_DEBUG(lock
, "client-side local enqueue handler, new lock created");
457 LDLM_LOCK_RELEASE(lock
);
461 EXPORT_SYMBOL(ldlm_cli_enqueue_local
);
463 static void failed_lock_cleanup(struct ldlm_namespace
*ns
,
464 struct ldlm_lock
*lock
, int mode
)
468 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
469 lock_res_and_lock(lock
);
470 /* Check that lock is not granted or failed, we might race. */
471 if ((lock
->l_req_mode
!= lock
->l_granted_mode
) &&
472 !(lock
->l_flags
& LDLM_FL_FAILED
)) {
473 /* Make sure that this lock will not be found by raced
474 * bl_ast and -EINVAL reply is sent to server anyways.
476 lock
->l_flags
|= LDLM_FL_LOCAL_ONLY
| LDLM_FL_FAILED
|
477 LDLM_FL_ATOMIC_CB
| LDLM_FL_CBPENDING
;
480 unlock_res_and_lock(lock
);
484 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
485 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
487 LDLM_DEBUG(lock
, "lock was granted or failed in race");
489 ldlm_lock_decref_internal(lock
, mode
);
491 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
492 * from llite/file.c/ll_file_flock(). */
493 /* This code makes for the fact that we do not have blocking handler on
494 * a client for flock locks. As such this is the place where we must
495 * completely kill failed locks. (interrupted and those that
496 * were waiting to be granted when server evicted us. */
497 if (lock
->l_resource
->lr_type
== LDLM_FLOCK
) {
498 lock_res_and_lock(lock
);
499 ldlm_resource_unlink_lock(lock
);
500 ldlm_lock_destroy_nolock(lock
);
501 unlock_res_and_lock(lock
);
506 * Finishing portion of client lock enqueue code.
508 * Called after receiving reply from server.
510 int ldlm_cli_enqueue_fini(struct obd_export
*exp
, struct ptlrpc_request
*req
,
511 ldlm_type_t type
, __u8 with_policy
, ldlm_mode_t mode
,
512 __u64
*flags
, void *lvb
, __u32 lvb_len
,
513 struct lustre_handle
*lockh
,int rc
)
515 struct ldlm_namespace
*ns
= exp
->exp_obd
->obd_namespace
;
516 int is_replay
= *flags
& LDLM_FL_REPLAY
;
517 struct ldlm_lock
*lock
;
518 struct ldlm_reply
*reply
;
519 int cleanup_phase
= 1;
522 lock
= ldlm_handle2lock(lockh
);
523 /* ldlm_cli_enqueue is holding a reference on this lock. */
525 LASSERT(type
== LDLM_FLOCK
);
529 LASSERTF(ergo(lvb_len
!= 0, lvb_len
== lock
->l_lvb_len
),
530 "lvb_len = %d, l_lvb_len = %d\n", lvb_len
, lock
->l_lvb_len
);
532 if (rc
!= ELDLM_OK
) {
534 LDLM_DEBUG(lock
, "client-side enqueue END (%s)",
535 rc
== ELDLM_LOCK_ABORTED
? "ABORTED" : "FAILED");
537 if (rc
!= ELDLM_LOCK_ABORTED
)
541 /* Before we return, swab the reply */
542 reply
= req_capsule_server_get(&req
->rq_pill
, &RMF_DLM_REP
);
544 GOTO(cleanup
, rc
= -EPROTO
);
547 LASSERT(lvb
!= NULL
);
549 size
= req_capsule_get_size(&req
->rq_pill
, &RMF_DLM_LVB
,
552 LDLM_ERROR(lock
, "Fail to get lvb_len, rc = %d", size
);
553 GOTO(cleanup
, rc
= size
);
554 } else if (unlikely(size
> lvb_len
)) {
555 LDLM_ERROR(lock
, "Replied LVB is larger than "
556 "expectation, expected = %d, replied = %d",
558 GOTO(cleanup
, rc
= -EINVAL
);
562 if (rc
== ELDLM_LOCK_ABORTED
) {
564 rc
= ldlm_fill_lvb(lock
, &req
->rq_pill
, RCL_SERVER
,
566 GOTO(cleanup
, rc
= (rc
!= 0 ? rc
: ELDLM_LOCK_ABORTED
));
569 /* lock enqueued on the server */
572 lock_res_and_lock(lock
);
573 /* Key change rehash lock in per-export hash with new key */
574 if (exp
->exp_lock_hash
) {
575 /* In the function below, .hs_keycmp resolves to
576 * ldlm_export_lock_keycmp() */
577 /* coverity[overrun-buffer-val] */
578 cfs_hash_rehash_key(exp
->exp_lock_hash
,
579 &lock
->l_remote_handle
,
583 lock
->l_remote_handle
= reply
->lock_handle
;
586 *flags
= ldlm_flags_from_wire(reply
->lock_flags
);
587 lock
->l_flags
|= ldlm_flags_from_wire(reply
->lock_flags
&
589 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
590 * to wait with no timeout as well */
591 lock
->l_flags
|= ldlm_flags_from_wire(reply
->lock_flags
&
593 unlock_res_and_lock(lock
);
595 CDEBUG(D_INFO
, "local: %p, remote cookie: %#llx, flags: 0x%llx\n",
596 lock
, reply
->lock_handle
.cookie
, *flags
);
598 /* If enqueue returned a blocked lock but the completion handler has
599 * already run, then it fixed up the resource and we don't need to do it
601 if ((*flags
) & LDLM_FL_LOCK_CHANGED
) {
602 int newmode
= reply
->lock_desc
.l_req_mode
;
604 if (newmode
&& newmode
!= lock
->l_req_mode
) {
605 LDLM_DEBUG(lock
, "server returned different mode %s",
606 ldlm_lockname
[newmode
]);
607 lock
->l_req_mode
= newmode
;
610 if (!ldlm_res_eq(&reply
->lock_desc
.l_resource
.lr_name
,
611 &lock
->l_resource
->lr_name
)) {
612 CDEBUG(D_INFO
, "remote intent success, locking "DLDLMRES
613 " instead of "DLDLMRES
"\n",
614 PLDLMRES(&reply
->lock_desc
.l_resource
),
615 PLDLMRES(lock
->l_resource
));
617 rc
= ldlm_lock_change_resource(ns
, lock
,
618 &reply
->lock_desc
.l_resource
.lr_name
);
619 if (rc
|| lock
->l_resource
== NULL
)
620 GOTO(cleanup
, rc
= -ENOMEM
);
621 LDLM_DEBUG(lock
, "client-side enqueue, new resource");
624 if (!(type
== LDLM_IBITS
&&
625 !(exp_connect_flags(exp
) & OBD_CONNECT_IBITS
)))
626 /* We assume lock type cannot change on server*/
627 ldlm_convert_policy_to_local(exp
,
628 lock
->l_resource
->lr_type
,
629 &reply
->lock_desc
.l_policy_data
,
630 &lock
->l_policy_data
);
631 if (type
!= LDLM_PLAIN
)
632 LDLM_DEBUG(lock
,"client-side enqueue, new policy data");
635 if ((*flags
) & LDLM_FL_AST_SENT
||
636 /* Cancel extent locks as soon as possible on a liblustre client,
637 * because it cannot handle asynchronous ASTs robustly (see
639 (LIBLUSTRE_CLIENT
&& type
== LDLM_EXTENT
)) {
640 lock_res_and_lock(lock
);
641 lock
->l_flags
|= LDLM_FL_CBPENDING
| LDLM_FL_BL_AST
;
642 unlock_res_and_lock(lock
);
643 LDLM_DEBUG(lock
, "enqueue reply includes blocking AST");
646 /* If the lock has already been granted by a completion AST, don't
647 * clobber the LVB with an older one. */
649 /* We must lock or a racing completion might update lvb without
650 * letting us know and we'll clobber the correct value.
651 * Cannot unlock after the check either, a that still leaves
652 * a tiny window for completion to get in */
653 lock_res_and_lock(lock
);
654 if (lock
->l_req_mode
!= lock
->l_granted_mode
)
655 rc
= ldlm_fill_lvb(lock
, &req
->rq_pill
, RCL_SERVER
,
656 lock
->l_lvb_data
, size
);
657 unlock_res_and_lock(lock
);
665 rc
= ldlm_lock_enqueue(ns
, &lock
, NULL
, flags
);
666 if (lock
->l_completion_ast
!= NULL
) {
667 int err
= lock
->l_completion_ast(lock
, *flags
, NULL
);
675 if (lvb_len
&& lvb
!= NULL
) {
676 /* Copy the LVB here, and not earlier, because the completion
677 * AST (if any) can override what we got in the reply */
678 memcpy(lvb
, lock
->l_lvb_data
, lvb_len
);
681 LDLM_DEBUG(lock
, "client-side enqueue END");
683 if (cleanup_phase
== 1 && rc
)
684 failed_lock_cleanup(ns
, lock
, mode
);
685 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
687 LDLM_LOCK_RELEASE(lock
);
690 EXPORT_SYMBOL(ldlm_cli_enqueue_fini
);
693 * Estimate number of lock handles that would fit into request of given
694 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
695 * a single page on the send/receive side. XXX: 512 should be changed to
696 * more adequate value.
698 static inline int ldlm_req_handles_avail(int req_size
, int off
)
702 avail
= min_t(int, LDLM_MAXREQSIZE
, PAGE_CACHE_SIZE
- 512) - req_size
;
703 if (likely(avail
>= 0))
704 avail
/= (int)sizeof(struct lustre_handle
);
707 avail
+= LDLM_LOCKREQ_HANDLES
- off
;
712 static inline int ldlm_capsule_handles_avail(struct req_capsule
*pill
,
713 enum req_location loc
,
716 int size
= req_capsule_msg_size(pill
, loc
);
717 return ldlm_req_handles_avail(size
, off
);
720 static inline int ldlm_format_handles_avail(struct obd_import
*imp
,
721 const struct req_format
*fmt
,
722 enum req_location loc
, int off
)
724 int size
= req_capsule_fmt_size(imp
->imp_msg_magic
, fmt
, loc
);
725 return ldlm_req_handles_avail(size
, off
);
729 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
730 * \a count locks in \a cancels.
732 * This is to be called by functions preparing their own requests that
733 * might contain lists of locks to cancel in addition to actual operation
734 * that needs to be performed.
736 int ldlm_prep_elc_req(struct obd_export
*exp
, struct ptlrpc_request
*req
,
737 int version
, int opc
, int canceloff
,
738 struct list_head
*cancels
, int count
)
740 struct ldlm_namespace
*ns
= exp
->exp_obd
->obd_namespace
;
741 struct req_capsule
*pill
= &req
->rq_pill
;
742 struct ldlm_request
*dlm
= NULL
;
743 int flags
, avail
, to_free
, pack
= 0;
749 if (ns_connect_cancelset(ns
)) {
750 /* Estimate the amount of available space in the request. */
751 req_capsule_filled_sizes(pill
, RCL_CLIENT
);
752 avail
= ldlm_capsule_handles_avail(pill
, RCL_CLIENT
, canceloff
);
754 flags
= ns_connect_lru_resize(ns
) ?
755 LDLM_CANCEL_LRUR
: LDLM_CANCEL_AGED
;
756 to_free
= !ns_connect_lru_resize(ns
) &&
757 opc
== LDLM_ENQUEUE
? 1 : 0;
759 /* Cancel LRU locks here _only_ if the server supports
760 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
761 * RPC, which will make us slower. */
763 count
+= ldlm_cancel_lru_local(ns
, cancels
, to_free
,
764 avail
- count
, 0, flags
);
769 req_capsule_set_size(pill
, &RMF_DLM_REQ
, RCL_CLIENT
,
770 ldlm_request_bufsize(pack
, opc
));
773 rc
= ptlrpc_request_pack(req
, version
, opc
);
775 ldlm_lock_list_put(cancels
, l_bl_ast
, count
);
779 if (ns_connect_cancelset(ns
)) {
781 dlm
= req_capsule_client_get(pill
, &RMF_DLM_REQ
);
783 /* Skip first lock handler in ldlm_request_pack(),
784 * this method will increment @lock_count according
785 * to the lock handle amount actually written to
787 dlm
->lock_count
= canceloff
;
789 /* Pack into the request @pack lock handles. */
790 ldlm_cli_cancel_list(cancels
, pack
, req
, 0);
791 /* Prepare and send separate cancel RPC for others. */
792 ldlm_cli_cancel_list(cancels
, count
- pack
, NULL
, 0);
794 ldlm_lock_list_put(cancels
, l_bl_ast
, count
);
798 EXPORT_SYMBOL(ldlm_prep_elc_req
);
800 int ldlm_prep_enqueue_req(struct obd_export
*exp
, struct ptlrpc_request
*req
,
801 struct list_head
*cancels
, int count
)
803 return ldlm_prep_elc_req(exp
, req
, LUSTRE_DLM_VERSION
, LDLM_ENQUEUE
,
804 LDLM_ENQUEUE_CANCEL_OFF
, cancels
, count
);
806 EXPORT_SYMBOL(ldlm_prep_enqueue_req
);
808 struct ptlrpc_request
*ldlm_enqueue_pack(struct obd_export
*exp
, int lvb_len
)
810 struct ptlrpc_request
*req
;
813 req
= ptlrpc_request_alloc(class_exp2cliimp(exp
), &RQF_LDLM_ENQUEUE
);
815 return ERR_PTR(-ENOMEM
);
817 rc
= ldlm_prep_enqueue_req(exp
, req
, NULL
, 0);
819 ptlrpc_request_free(req
);
823 req_capsule_set_size(&req
->rq_pill
, &RMF_DLM_LVB
, RCL_SERVER
, lvb_len
);
824 ptlrpc_request_set_replen(req
);
827 EXPORT_SYMBOL(ldlm_enqueue_pack
);
830 * Client-side lock enqueue.
832 * If a request has some specific initialisation it is passed in \a reqp,
833 * otherwise it is created in ldlm_cli_enqueue.
835 * Supports sync and async requests, pass \a async flag accordingly. If a
836 * request was created in ldlm_cli_enqueue and it is the async request,
837 * pass it to the caller in \a reqp.
839 int ldlm_cli_enqueue(struct obd_export
*exp
, struct ptlrpc_request
**reqp
,
840 struct ldlm_enqueue_info
*einfo
,
841 const struct ldlm_res_id
*res_id
,
842 ldlm_policy_data_t
const *policy
, __u64
*flags
,
843 void *lvb
, __u32 lvb_len
, enum lvb_type lvb_type
,
844 struct lustre_handle
*lockh
, int async
)
846 struct ldlm_namespace
*ns
;
847 struct ldlm_lock
*lock
;
848 struct ldlm_request
*body
;
849 int is_replay
= *flags
& LDLM_FL_REPLAY
;
850 int req_passed_in
= 1;
852 struct ptlrpc_request
*req
;
854 LASSERT(exp
!= NULL
);
856 ns
= exp
->exp_obd
->obd_namespace
;
858 /* If we're replaying this lock, just check some invariants.
859 * If we're creating a new lock, get everything all setup nice. */
861 lock
= ldlm_handle2lock_long(lockh
, 0);
862 LASSERT(lock
!= NULL
);
863 LDLM_DEBUG(lock
, "client-side enqueue START");
864 LASSERT(exp
== lock
->l_conn_export
);
866 const struct ldlm_callback_suite cbs
= {
867 .lcs_completion
= einfo
->ei_cb_cp
,
868 .lcs_blocking
= einfo
->ei_cb_bl
,
869 .lcs_glimpse
= einfo
->ei_cb_gl
871 lock
= ldlm_lock_create(ns
, res_id
, einfo
->ei_type
,
872 einfo
->ei_mode
, &cbs
, einfo
->ei_cbdata
,
876 /* for the local lock, add the reference */
877 ldlm_lock_addref_internal(lock
, einfo
->ei_mode
);
878 ldlm_lock2handle(lock
, lockh
);
879 if (policy
!= NULL
) {
880 /* INODEBITS_INTEROP: If the server does not support
881 * inodebits, we will request a plain lock in the
882 * descriptor (ldlm_lock2desc() below) but use an
883 * inodebits lock internally with both bits set.
885 if (einfo
->ei_type
== LDLM_IBITS
&&
886 !(exp_connect_flags(exp
) &
888 lock
->l_policy_data
.l_inodebits
.bits
=
889 MDS_INODELOCK_LOOKUP
|
890 MDS_INODELOCK_UPDATE
;
892 lock
->l_policy_data
= *policy
;
895 if (einfo
->ei_type
== LDLM_EXTENT
)
896 lock
->l_req_extent
= policy
->l_extent
;
897 LDLM_DEBUG(lock
, "client-side enqueue START, flags %llx\n",
901 lock
->l_conn_export
= exp
;
902 lock
->l_export
= NULL
;
903 lock
->l_blocking_ast
= einfo
->ei_cb_bl
;
904 lock
->l_flags
|= (*flags
& (LDLM_FL_NO_LRU
| LDLM_FL_EXCL
));
906 /* lock not sent to server yet */
908 if (reqp
== NULL
|| *reqp
== NULL
) {
909 req
= ptlrpc_request_alloc_pack(class_exp2cliimp(exp
),
914 failed_lock_cleanup(ns
, lock
, einfo
->ei_mode
);
915 LDLM_LOCK_RELEASE(lock
);
925 len
= req_capsule_get_size(&req
->rq_pill
, &RMF_DLM_REQ
,
927 LASSERTF(len
>= sizeof(*body
), "buflen[%d] = %d, not %d\n",
928 DLM_LOCKREQ_OFF
, len
, (int)sizeof(*body
));
931 /* Dump lock data into the request buffer */
932 body
= req_capsule_client_get(&req
->rq_pill
, &RMF_DLM_REQ
);
933 ldlm_lock2desc(lock
, &body
->lock_desc
);
934 body
->lock_flags
= ldlm_flags_to_wire(*flags
);
935 body
->lock_handle
[0] = *lockh
;
937 /* Continue as normal. */
938 if (!req_passed_in
) {
940 req_capsule_extend(&req
->rq_pill
,
941 &RQF_LDLM_ENQUEUE_LVB
);
942 req_capsule_set_size(&req
->rq_pill
, &RMF_DLM_LVB
, RCL_SERVER
,
944 ptlrpc_request_set_replen(req
);
948 * Liblustre client doesn't get extent locks, except for O_APPEND case
949 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
950 * [i_size, OBD_OBJECT_EOF] lock is taken.
952 LASSERT(ergo(LIBLUSTRE_CLIENT
, einfo
->ei_type
!= LDLM_EXTENT
||
953 policy
->l_extent
.end
== OBD_OBJECT_EOF
));
956 LASSERT(reqp
!= NULL
);
960 LDLM_DEBUG(lock
, "sending request");
962 rc
= ptlrpc_queue_wait(req
);
964 err
= ldlm_cli_enqueue_fini(exp
, req
, einfo
->ei_type
, policy
? 1 : 0,
965 einfo
->ei_mode
, flags
, lvb
, lvb_len
,
968 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
969 * one reference that we took */
971 LDLM_LOCK_RELEASE(lock
);
975 if (!req_passed_in
&& req
!= NULL
) {
976 ptlrpc_req_finished(req
);
983 EXPORT_SYMBOL(ldlm_cli_enqueue
);
985 static int ldlm_cli_convert_local(struct ldlm_lock
*lock
, int new_mode
,
988 struct ldlm_resource
*res
;
991 if (ns_is_client(ldlm_lock_to_ns(lock
))) {
992 CERROR("Trying to cancel local lock\n");
995 LDLM_DEBUG(lock
, "client-side local convert");
997 res
= ldlm_lock_convert(lock
, new_mode
, flags
);
999 ldlm_reprocess_all(res
);
1002 rc
= LUSTRE_EDEADLK
;
1004 LDLM_DEBUG(lock
, "client-side local convert handler END");
1005 LDLM_LOCK_PUT(lock
);
1009 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
1010 * conversion of locks which are on the waiting or converting queue */
1011 /* Caller of this code is supposed to take care of lock readers/writers
1013 int ldlm_cli_convert(struct lustre_handle
*lockh
, int new_mode
, __u32
*flags
)
1015 struct ldlm_request
*body
;
1016 struct ldlm_reply
*reply
;
1017 struct ldlm_lock
*lock
;
1018 struct ldlm_resource
*res
;
1019 struct ptlrpc_request
*req
;
1022 lock
= ldlm_handle2lock(lockh
);
1029 if (lock
->l_conn_export
== NULL
)
1030 return ldlm_cli_convert_local(lock
, new_mode
, flags
);
1032 LDLM_DEBUG(lock
, "client-side convert");
1034 req
= ptlrpc_request_alloc_pack(class_exp2cliimp(lock
->l_conn_export
),
1035 &RQF_LDLM_CONVERT
, LUSTRE_DLM_VERSION
,
1038 LDLM_LOCK_PUT(lock
);
1042 body
= req_capsule_client_get(&req
->rq_pill
, &RMF_DLM_REQ
);
1043 body
->lock_handle
[0] = lock
->l_remote_handle
;
1045 body
->lock_desc
.l_req_mode
= new_mode
;
1046 body
->lock_flags
= ldlm_flags_to_wire(*flags
);
1049 ptlrpc_request_set_replen(req
);
1050 rc
= ptlrpc_queue_wait(req
);
1054 reply
= req_capsule_server_get(&req
->rq_pill
, &RMF_DLM_REP
);
1056 GOTO(out
, rc
= -EPROTO
);
1059 GOTO(out
, rc
= req
->rq_status
);
1061 res
= ldlm_lock_convert(lock
, new_mode
, &reply
->lock_flags
);
1063 ldlm_reprocess_all(res
);
1064 /* Go to sleep until the lock is granted. */
1065 /* FIXME: or cancelled. */
1066 if (lock
->l_completion_ast
) {
1067 rc
= lock
->l_completion_ast(lock
, LDLM_FL_WAIT_NOREPROC
,
1073 rc
= LUSTRE_EDEADLK
;
1076 LDLM_LOCK_PUT(lock
);
1077 ptlrpc_req_finished(req
);
1080 EXPORT_SYMBOL(ldlm_cli_convert
);
1083 * Cancel locks locally.
1085 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1086 * \retval LDLM_FL_CANCELING otherwise;
1087 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1089 static __u64
ldlm_cli_cancel_local(struct ldlm_lock
*lock
)
1091 __u64 rc
= LDLM_FL_LOCAL_ONLY
;
1093 if (lock
->l_conn_export
) {
1096 LDLM_DEBUG(lock
, "client-side cancel");
1097 /* Set this flag to prevent others from getting new references*/
1098 lock_res_and_lock(lock
);
1099 lock
->l_flags
|= LDLM_FL_CBPENDING
;
1100 local_only
= !!(lock
->l_flags
&
1101 (LDLM_FL_LOCAL_ONLY
|LDLM_FL_CANCEL_ON_BLOCK
));
1102 ldlm_cancel_callback(lock
);
1103 rc
= (lock
->l_flags
& LDLM_FL_BL_AST
) ?
1104 LDLM_FL_BL_AST
: LDLM_FL_CANCELING
;
1105 unlock_res_and_lock(lock
);
1108 CDEBUG(D_DLMTRACE
, "not sending request (at caller's "
1110 rc
= LDLM_FL_LOCAL_ONLY
;
1112 ldlm_lock_cancel(lock
);
1114 if (ns_is_client(ldlm_lock_to_ns(lock
))) {
1115 LDLM_ERROR(lock
, "Trying to cancel local lock");
1118 LDLM_DEBUG(lock
, "server-side local cancel");
1119 ldlm_lock_cancel(lock
);
1120 ldlm_reprocess_all(lock
->l_resource
);
1127 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1129 static void ldlm_cancel_pack(struct ptlrpc_request
*req
,
1130 struct list_head
*head
, int count
)
1132 struct ldlm_request
*dlm
;
1133 struct ldlm_lock
*lock
;
1134 int max
, packed
= 0;
1136 dlm
= req_capsule_client_get(&req
->rq_pill
, &RMF_DLM_REQ
);
1137 LASSERT(dlm
!= NULL
);
1139 /* Check the room in the request buffer. */
1140 max
= req_capsule_get_size(&req
->rq_pill
, &RMF_DLM_REQ
, RCL_CLIENT
) -
1141 sizeof(struct ldlm_request
);
1142 max
/= sizeof(struct lustre_handle
);
1143 max
+= LDLM_LOCKREQ_HANDLES
;
1144 LASSERT(max
>= dlm
->lock_count
+ count
);
1146 /* XXX: it would be better to pack lock handles grouped by resource.
1147 * so that the server cancel would call filter_lvbo_update() less
1149 list_for_each_entry(lock
, head
, l_bl_ast
) {
1152 LASSERT(lock
->l_conn_export
);
1153 /* Pack the lock handle to the given request buffer. */
1154 LDLM_DEBUG(lock
, "packing");
1155 dlm
->lock_handle
[dlm
->lock_count
++] = lock
->l_remote_handle
;
1158 CDEBUG(D_DLMTRACE
, "%d locks packed\n", packed
);
1162 * Prepare and send a batched cancel RPC. It will include \a count lock
1163 * handles of locks given in \a cancels list. */
1164 int ldlm_cli_cancel_req(struct obd_export
*exp
, struct list_head
*cancels
,
1165 int count
, ldlm_cancel_flags_t flags
)
1167 struct ptlrpc_request
*req
= NULL
;
1168 struct obd_import
*imp
;
1172 LASSERT(exp
!= NULL
);
1175 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL
, cfs_fail_val
);
1177 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE
))
1180 free
= ldlm_format_handles_avail(class_exp2cliimp(exp
),
1181 &RQF_LDLM_CANCEL
, RCL_CLIENT
, 0);
1186 imp
= class_exp2cliimp(exp
);
1187 if (imp
== NULL
|| imp
->imp_invalid
) {
1189 "skipping cancel on invalid import %p\n", imp
);
1193 req
= ptlrpc_request_alloc(imp
, &RQF_LDLM_CANCEL
);
1195 GOTO(out
, rc
= -ENOMEM
);
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
, PDL_POLICY_LOCAL
, -1);
1219 rc
= ptlrpc_queue_wait(req
);
1221 if (rc
== LUSTRE_ESTALE
) {
1222 CDEBUG(D_DLMTRACE
, "client/server (nid %s) "
1223 "out of sync -- not fatal\n",
1224 libcfs_nid2str(req
->rq_import
->
1225 imp_connection
->c_peer
.nid
));
1227 } else if (rc
== -ETIMEDOUT
&& /* check there was no reconnect*/
1228 req
->rq_import_generation
== imp
->imp_generation
) {
1229 ptlrpc_req_finished(req
);
1231 } else if (rc
!= ELDLM_OK
) {
1232 /* -ESHUTDOWN is common on umount */
1233 CDEBUG_LIMIT(rc
== -ESHUTDOWN
? D_DLMTRACE
: D_ERROR
,
1234 "Got rc %d from cancel RPC: "
1235 "canceling anyway\n", rc
);
1242 ptlrpc_req_finished(req
);
1244 return sent
? sent
: rc
;
1246 EXPORT_SYMBOL(ldlm_cli_cancel_req
);
1248 static inline struct ldlm_pool
*ldlm_imp2pl(struct obd_import
*imp
)
1250 LASSERT(imp
!= NULL
);
1251 return &imp
->imp_obd
->obd_namespace
->ns_pool
;
1255 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1257 int ldlm_cli_update_pool(struct ptlrpc_request
*req
)
1259 struct obd_device
*obd
;
1263 if (unlikely(!req
->rq_import
|| !req
->rq_import
->imp_obd
||
1264 !imp_connect_lru_resize(req
->rq_import
)))
1267 * Do nothing for corner cases.
1272 /* In some cases RPC may contain SLV and limit zeroed out. This
1273 * is the case when server does not support LRU resize feature.
1274 * This is also possible in some recovery cases when server-side
1275 * reqs have no reference to the OBD export and thus access to
1276 * server-side namespace is not possible. */
1277 if (lustre_msg_get_slv(req
->rq_repmsg
) == 0 ||
1278 lustre_msg_get_limit(req
->rq_repmsg
) == 0) {
1279 DEBUG_REQ(D_HA
, req
, "Zero SLV or Limit found (SLV: %llu, Limit: %u)",
1280 lustre_msg_get_slv(req
->rq_repmsg
),
1281 lustre_msg_get_limit(req
->rq_repmsg
));
1285 new_limit
= lustre_msg_get_limit(req
->rq_repmsg
);
1286 new_slv
= lustre_msg_get_slv(req
->rq_repmsg
);
1287 obd
= req
->rq_import
->imp_obd
;
1289 /* Set new SLV and limit in OBD fields to make them accessible
1290 * to the pool thread. We do not access obd_namespace and pool
1291 * directly here as there is no reliable way to make sure that
1292 * they are still alive at cleanup time. Evil races are possible
1293 * which may cause Oops at that time. */
1294 write_lock(&obd
->obd_pool_lock
);
1295 obd
->obd_pool_slv
= new_slv
;
1296 obd
->obd_pool_limit
= new_limit
;
1297 write_unlock(&obd
->obd_pool_lock
);
1301 EXPORT_SYMBOL(ldlm_cli_update_pool
);
1304 * Client side lock cancel.
1306 * Lock must not have any readers or writers by this time.
1308 int ldlm_cli_cancel(struct lustre_handle
*lockh
,
1309 ldlm_cancel_flags_t cancel_flags
)
1311 struct obd_export
*exp
;
1312 int avail
, flags
, count
= 1;
1314 struct ldlm_namespace
*ns
;
1315 struct ldlm_lock
*lock
;
1318 /* concurrent cancels on the same handle can happen */
1319 lock
= ldlm_handle2lock_long(lockh
, LDLM_FL_CANCELING
);
1321 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1325 rc
= ldlm_cli_cancel_local(lock
);
1326 if (rc
== LDLM_FL_LOCAL_ONLY
|| cancel_flags
& LCF_LOCAL
) {
1327 LDLM_LOCK_RELEASE(lock
);
1330 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1331 * RPC which goes to canceld portal, so we can cancel other LRU locks
1332 * here and send them all as one LDLM_CANCEL RPC. */
1333 LASSERT(list_empty(&lock
->l_bl_ast
));
1334 list_add(&lock
->l_bl_ast
, &cancels
);
1336 exp
= lock
->l_conn_export
;
1337 if (exp_connect_cancelset(exp
)) {
1338 avail
= ldlm_format_handles_avail(class_exp2cliimp(exp
),
1343 ns
= ldlm_lock_to_ns(lock
);
1344 flags
= ns_connect_lru_resize(ns
) ?
1345 LDLM_CANCEL_LRUR
: LDLM_CANCEL_AGED
;
1346 count
+= ldlm_cancel_lru_local(ns
, &cancels
, 0, avail
- 1,
1349 ldlm_cli_cancel_list(&cancels
, count
, NULL
, cancel_flags
);
1352 EXPORT_SYMBOL(ldlm_cli_cancel
);
1355 * Locally cancel up to \a count locks in list \a cancels.
1356 * Return the number of cancelled locks.
1358 int ldlm_cli_cancel_list_local(struct list_head
*cancels
, int count
,
1359 ldlm_cancel_flags_t flags
)
1362 struct ldlm_lock
*lock
, *next
;
1363 int left
= 0, bl_ast
= 0;
1367 list_for_each_entry_safe(lock
, next
, cancels
, l_bl_ast
) {
1371 if (flags
& LCF_LOCAL
) {
1372 rc
= LDLM_FL_LOCAL_ONLY
;
1373 ldlm_lock_cancel(lock
);
1375 rc
= ldlm_cli_cancel_local(lock
);
1377 /* Until we have compound requests and can send LDLM_CANCEL
1378 * requests batched with generic RPCs, we need to send cancels
1379 * with the LDLM_FL_BL_AST flag in a separate RPC from
1380 * the one being generated now. */
1381 if (!(flags
& LCF_BL_AST
) && (rc
== LDLM_FL_BL_AST
)) {
1382 LDLM_DEBUG(lock
, "Cancel lock separately");
1383 list_del_init(&lock
->l_bl_ast
);
1384 list_add(&lock
->l_bl_ast
, &head
);
1388 if (rc
== LDLM_FL_LOCAL_ONLY
) {
1389 /* CANCEL RPC should not be sent to server. */
1390 list_del_init(&lock
->l_bl_ast
);
1391 LDLM_LOCK_RELEASE(lock
);
1397 ldlm_cli_cancel_list(&head
, bl_ast
, NULL
, 0);
1402 EXPORT_SYMBOL(ldlm_cli_cancel_list_local
);
1405 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1406 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1407 * readahead requests, ...)
1409 static ldlm_policy_res_t
ldlm_cancel_no_wait_policy(struct ldlm_namespace
*ns
,
1410 struct ldlm_lock
*lock
,
1411 int unused
, int added
,
1414 ldlm_policy_res_t result
= LDLM_POLICY_CANCEL_LOCK
;
1415 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 proc. */
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 proc 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
, struct list_head
*cancels
,
1595 int count
, int max
, int flags
)
1597 ldlm_cancel_lru_policy_t pf
;
1598 struct ldlm_lock
*lock
, *next
;
1599 int added
= 0, unused
, remained
;
1601 spin_lock(&ns
->ns_lock
);
1602 unused
= ns
->ns_nr_unused
;
1605 if (!ns_connect_lru_resize(ns
))
1606 count
+= unused
- ns
->ns_max_unused
;
1608 pf
= ldlm_cancel_lru_policy(ns
, flags
);
1609 LASSERT(pf
!= NULL
);
1611 while (!list_empty(&ns
->ns_unused_list
)) {
1612 ldlm_policy_res_t result
;
1614 /* all unused locks */
1615 if (remained
-- <= 0)
1618 /* For any flags, stop scanning if @max is reached. */
1619 if (max
&& added
>= max
)
1622 list_for_each_entry_safe(lock
, next
, &ns
->ns_unused_list
,
1624 /* No locks which got blocking requests. */
1625 LASSERT(!(lock
->l_flags
& LDLM_FL_BL_AST
));
1627 if (flags
& LDLM_CANCEL_NO_WAIT
&&
1628 lock
->l_flags
& LDLM_FL_SKIPPED
)
1629 /* already processed */
1632 /* Somebody is already doing CANCEL. No need for this
1633 * lock in LRU, do not traverse it again. */
1634 if (!(lock
->l_flags
& LDLM_FL_CANCELING
))
1637 ldlm_lock_remove_from_lru_nolock(lock
);
1639 if (&lock
->l_lru
== &ns
->ns_unused_list
)
1642 LDLM_LOCK_GET(lock
);
1643 spin_unlock(&ns
->ns_lock
);
1644 lu_ref_add(&lock
->l_reference
, __func__
, current
);
1646 /* Pass the lock through the policy filter and see if it
1647 * should stay in LRU.
1649 * Even for shrinker policy we stop scanning if
1650 * we find a lock that should stay in the cache.
1651 * We should take into account lock age anyway
1652 * as a new lock is a valuable resource even if
1653 * it has a low weight.
1655 * That is, for shrinker policy we drop only
1656 * old locks, but additionally choose them by
1657 * their weight. Big extent locks will stay in
1659 result
= pf(ns
, lock
, unused
, added
, count
);
1660 if (result
== LDLM_POLICY_KEEP_LOCK
) {
1661 lu_ref_del(&lock
->l_reference
,
1663 LDLM_LOCK_RELEASE(lock
);
1664 spin_lock(&ns
->ns_lock
);
1667 if (result
== LDLM_POLICY_SKIP_LOCK
) {
1668 lu_ref_del(&lock
->l_reference
,
1670 LDLM_LOCK_RELEASE(lock
);
1671 spin_lock(&ns
->ns_lock
);
1675 lock_res_and_lock(lock
);
1676 /* Check flags again under the lock. */
1677 if ((lock
->l_flags
& LDLM_FL_CANCELING
) ||
1678 (ldlm_lock_remove_from_lru(lock
) == 0)) {
1679 /* Another thread is removing lock from LRU, or
1680 * somebody is already doing CANCEL, or there
1681 * is a blocking request which will send cancel
1682 * by itself, or the lock is no longer unused. */
1683 unlock_res_and_lock(lock
);
1684 lu_ref_del(&lock
->l_reference
,
1686 LDLM_LOCK_RELEASE(lock
);
1687 spin_lock(&ns
->ns_lock
);
1690 LASSERT(!lock
->l_readers
&& !lock
->l_writers
);
1692 /* If we have chosen to cancel this lock voluntarily, we
1693 * better send cancel notification to server, so that it
1694 * frees appropriate state. This might lead to a race
1695 * where while we are doing cancel here, server is also
1696 * silently cancelling this lock. */
1697 lock
->l_flags
&= ~LDLM_FL_CANCEL_ON_BLOCK
;
1699 /* Setting the CBPENDING flag is a little misleading,
1700 * but prevents an important race; namely, once
1701 * CBPENDING is set, the lock can accumulate no more
1702 * readers/writers. Since readers and writers are
1703 * already zero here, ldlm_lock_decref() won't see
1704 * this flag and call l_blocking_ast */
1705 lock
->l_flags
|= LDLM_FL_CBPENDING
| LDLM_FL_CANCELING
;
1707 /* We can't re-add to l_lru as it confuses the
1708 * refcounting in ldlm_lock_remove_from_lru() if an AST
1709 * arrives after we drop lr_lock below. We use l_bl_ast
1710 * and can't use l_pending_chain as it is used both on
1711 * server and client nevertheless bug 5666 says it is
1712 * used only on server */
1713 LASSERT(list_empty(&lock
->l_bl_ast
));
1714 list_add(&lock
->l_bl_ast
, cancels
);
1715 unlock_res_and_lock(lock
);
1716 lu_ref_del(&lock
->l_reference
, __func__
, current
);
1717 spin_lock(&ns
->ns_lock
);
1721 spin_unlock(&ns
->ns_lock
);
1725 int ldlm_cancel_lru_local(struct ldlm_namespace
*ns
, struct list_head
*cancels
,
1726 int count
, int max
, ldlm_cancel_flags_t cancel_flags
,
1730 added
= ldlm_prepare_lru_list(ns
, cancels
, count
, max
, flags
);
1733 return ldlm_cli_cancel_list_local(cancels
, added
, cancel_flags
);
1737 * Cancel at least \a nr locks from given namespace LRU.
1739 * When called with LCF_ASYNC the blocking callback will be handled
1740 * in a thread and this function will return after the thread has been
1741 * asked to call the callback. When called with LCF_ASYNC the blocking
1742 * callback will be performed in this function.
1744 int ldlm_cancel_lru(struct ldlm_namespace
*ns
, int nr
,
1745 ldlm_cancel_flags_t cancel_flags
,
1751 /* Just prepare the list of locks, do not actually cancel them yet.
1752 * Locks are cancelled later in a separate thread. */
1753 count
= ldlm_prepare_lru_list(ns
, &cancels
, nr
, 0, flags
);
1754 rc
= ldlm_bl_to_thread_list(ns
, NULL
, &cancels
, count
, cancel_flags
);
1762 * Find and cancel locally unused locks found on resource, matched to the
1763 * given policy, mode. GET the found locks and add them into the \a cancels
1766 int ldlm_cancel_resource_local(struct ldlm_resource
*res
,
1767 struct list_head
*cancels
,
1768 ldlm_policy_data_t
*policy
,
1769 ldlm_mode_t mode
, __u64 lock_flags
,
1770 ldlm_cancel_flags_t cancel_flags
, void *opaque
)
1772 struct ldlm_lock
*lock
;
1776 list_for_each_entry(lock
, &res
->lr_granted
, l_res_link
) {
1777 if (opaque
!= NULL
&& lock
->l_ast_data
!= opaque
) {
1778 LDLM_ERROR(lock
, "data %p doesn't match opaque %p",
1779 lock
->l_ast_data
, opaque
);
1784 if (lock
->l_readers
|| lock
->l_writers
)
1787 /* If somebody is already doing CANCEL, or blocking AST came,
1788 * skip this lock. */
1789 if (lock
->l_flags
& LDLM_FL_BL_AST
||
1790 lock
->l_flags
& LDLM_FL_CANCELING
)
1793 if (lockmode_compat(lock
->l_granted_mode
, mode
))
1796 /* If policy is given and this is IBITS lock, add to list only
1797 * those locks that match by policy. */
1798 if (policy
&& (lock
->l_resource
->lr_type
== LDLM_IBITS
) &&
1799 !(lock
->l_policy_data
.l_inodebits
.bits
&
1800 policy
->l_inodebits
.bits
))
1803 /* See CBPENDING comment in ldlm_cancel_lru */
1804 lock
->l_flags
|= LDLM_FL_CBPENDING
| LDLM_FL_CANCELING
|
1807 LASSERT(list_empty(&lock
->l_bl_ast
));
1808 list_add(&lock
->l_bl_ast
, cancels
);
1809 LDLM_LOCK_GET(lock
);
1814 return ldlm_cli_cancel_list_local(cancels
, count
, cancel_flags
);
1816 EXPORT_SYMBOL(ldlm_cancel_resource_local
);
1819 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1821 * If \a req is NULL, send CANCEL request to server with handles of locks
1822 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1823 * separately per lock.
1824 * If \a req is not NULL, put handles of locks in \a cancels into the request
1825 * buffer at the offset \a off.
1826 * Destroy \a cancels at the end.
1828 int ldlm_cli_cancel_list(struct list_head
*cancels
, int count
,
1829 struct ptlrpc_request
*req
, ldlm_cancel_flags_t flags
)
1831 struct ldlm_lock
*lock
;
1834 if (list_empty(cancels
) || count
== 0)
1837 /* XXX: requests (both batched and not) could be sent in parallel.
1838 * Usually it is enough to have just 1 RPC, but it is possible that
1839 * there are too many locks to be cancelled in LRU or on a resource.
1840 * It would also speed up the case when the server does not support
1843 LASSERT(!list_empty(cancels
));
1844 lock
= list_entry(cancels
->next
, struct ldlm_lock
,
1846 LASSERT(lock
->l_conn_export
);
1848 if (exp_connect_cancelset(lock
->l_conn_export
)) {
1851 ldlm_cancel_pack(req
, cancels
, count
);
1853 res
= ldlm_cli_cancel_req(lock
->l_conn_export
,
1857 res
= ldlm_cli_cancel_req(lock
->l_conn_export
,
1862 CDEBUG_LIMIT(res
== -ESHUTDOWN
? D_DLMTRACE
: D_ERROR
,
1863 "ldlm_cli_cancel_list: %d\n", res
);
1868 ldlm_lock_list_put(cancels
, l_bl_ast
, res
);
1870 LASSERT(count
== 0);
1873 EXPORT_SYMBOL(ldlm_cli_cancel_list
);
1876 * Cancel all locks on a resource that have 0 readers/writers.
1878 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1879 * to notify the server. */
1880 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace
*ns
,
1881 const struct ldlm_res_id
*res_id
,
1882 ldlm_policy_data_t
*policy
,
1884 ldlm_cancel_flags_t flags
,
1887 struct ldlm_resource
*res
;
1892 res
= ldlm_resource_get(ns
, NULL
, res_id
, 0, 0);
1894 /* This is not a problem. */
1895 CDEBUG(D_INFO
, "No resource %llu\n", res_id
->name
[0]);
1899 LDLM_RESOURCE_ADDREF(res
);
1900 count
= ldlm_cancel_resource_local(res
, &cancels
, policy
, mode
,
1901 0, flags
| LCF_BL_AST
, opaque
);
1902 rc
= ldlm_cli_cancel_list(&cancels
, count
, NULL
, flags
);
1904 CERROR("canceling unused lock "DLDLMRES
": rc = %d\n",
1907 LDLM_RESOURCE_DELREF(res
);
1908 ldlm_resource_putref(res
);
1911 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource
);
1913 struct ldlm_cli_cancel_arg
{
1918 static int ldlm_cli_hash_cancel_unused(struct cfs_hash
*hs
, struct cfs_hash_bd
*bd
,
1919 struct hlist_node
*hnode
, void *arg
)
1921 struct ldlm_resource
*res
= cfs_hash_object(hs
, hnode
);
1922 struct ldlm_cli_cancel_arg
*lc
= arg
;
1924 ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res
), &res
->lr_name
,
1926 lc
->lc_flags
, lc
->lc_opaque
);
1927 /* must return 0 for hash iteration */
1932 * Cancel all locks on a namespace (or a specific resource, if given)
1933 * that have 0 readers/writers.
1935 * If flags & LCF_LOCAL, throw the locks away without trying
1936 * to notify the server. */
1937 int ldlm_cli_cancel_unused(struct ldlm_namespace
*ns
,
1938 const struct ldlm_res_id
*res_id
,
1939 ldlm_cancel_flags_t flags
, void *opaque
)
1941 struct ldlm_cli_cancel_arg arg
= {
1943 .lc_opaque
= opaque
,
1949 if (res_id
!= NULL
) {
1950 return ldlm_cli_cancel_unused_resource(ns
, res_id
, NULL
,
1954 cfs_hash_for_each_nolock(ns
->ns_rs_hash
,
1955 ldlm_cli_hash_cancel_unused
, &arg
);
1959 EXPORT_SYMBOL(ldlm_cli_cancel_unused
);
1961 /* Lock iterators. */
1963 int ldlm_resource_foreach(struct ldlm_resource
*res
, ldlm_iterator_t iter
,
1966 struct list_head
*tmp
, *next
;
1967 struct ldlm_lock
*lock
;
1968 int rc
= LDLM_ITER_CONTINUE
;
1971 return LDLM_ITER_CONTINUE
;
1974 list_for_each_safe(tmp
, next
, &res
->lr_granted
) {
1975 lock
= list_entry(tmp
, struct ldlm_lock
, l_res_link
);
1977 if (iter(lock
, closure
) == LDLM_ITER_STOP
)
1978 GOTO(out
, rc
= LDLM_ITER_STOP
);
1981 list_for_each_safe(tmp
, next
, &res
->lr_converting
) {
1982 lock
= list_entry(tmp
, struct ldlm_lock
, l_res_link
);
1984 if (iter(lock
, closure
) == LDLM_ITER_STOP
)
1985 GOTO(out
, rc
= LDLM_ITER_STOP
);
1988 list_for_each_safe(tmp
, next
, &res
->lr_waiting
) {
1989 lock
= list_entry(tmp
, struct ldlm_lock
, l_res_link
);
1991 if (iter(lock
, closure
) == LDLM_ITER_STOP
)
1992 GOTO(out
, rc
= LDLM_ITER_STOP
);
1998 EXPORT_SYMBOL(ldlm_resource_foreach
);
2000 struct iter_helper_data
{
2001 ldlm_iterator_t iter
;
2005 static int ldlm_iter_helper(struct ldlm_lock
*lock
, void *closure
)
2007 struct iter_helper_data
*helper
= closure
;
2008 return helper
->iter(lock
, helper
->closure
);
2011 static int ldlm_res_iter_helper(struct cfs_hash
*hs
, struct cfs_hash_bd
*bd
,
2012 struct hlist_node
*hnode
, void *arg
)
2015 struct ldlm_resource
*res
= cfs_hash_object(hs
, hnode
);
2017 return ldlm_resource_foreach(res
, ldlm_iter_helper
, arg
) ==
2021 void ldlm_namespace_foreach(struct ldlm_namespace
*ns
,
2022 ldlm_iterator_t iter
, void *closure
)
2025 struct iter_helper_data helper
= {
2030 cfs_hash_for_each_nolock(ns
->ns_rs_hash
,
2031 ldlm_res_iter_helper
, &helper
);
2034 EXPORT_SYMBOL(ldlm_namespace_foreach
);
2036 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2037 * return 0: find no resource
2038 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2041 int ldlm_resource_iterate(struct ldlm_namespace
*ns
,
2042 const struct ldlm_res_id
*res_id
,
2043 ldlm_iterator_t iter
, void *data
)
2045 struct ldlm_resource
*res
;
2049 CERROR("must pass in namespace\n");
2053 res
= ldlm_resource_get(ns
, NULL
, res_id
, 0, 0);
2057 LDLM_RESOURCE_ADDREF(res
);
2058 rc
= ldlm_resource_foreach(res
, iter
, data
);
2059 LDLM_RESOURCE_DELREF(res
);
2060 ldlm_resource_putref(res
);
2063 EXPORT_SYMBOL(ldlm_resource_iterate
);
2067 static int ldlm_chain_lock_for_replay(struct ldlm_lock
*lock
, void *closure
)
2069 struct list_head
*list
= closure
;
2071 /* we use l_pending_chain here, because it's unused on clients. */
2072 LASSERTF(list_empty(&lock
->l_pending_chain
),
2073 "lock %p next %p prev %p\n",
2074 lock
, &lock
->l_pending_chain
.next
,&lock
->l_pending_chain
.prev
);
2075 /* bug 9573: don't replay locks left after eviction, or
2076 * bug 17614: locks being actively cancelled. Get a reference
2077 * on a lock so that it does not disappear under us (e.g. due to cancel)
2079 if (!(lock
->l_flags
& (LDLM_FL_FAILED
|LDLM_FL_CANCELING
))) {
2080 list_add(&lock
->l_pending_chain
, list
);
2081 LDLM_LOCK_GET(lock
);
2084 return LDLM_ITER_CONTINUE
;
2087 static int replay_lock_interpret(const struct lu_env
*env
,
2088 struct ptlrpc_request
*req
,
2089 struct ldlm_async_args
*aa
, int rc
)
2091 struct ldlm_lock
*lock
;
2092 struct ldlm_reply
*reply
;
2093 struct obd_export
*exp
;
2095 atomic_dec(&req
->rq_import
->imp_replay_inflight
);
2100 reply
= req_capsule_server_get(&req
->rq_pill
, &RMF_DLM_REP
);
2102 GOTO(out
, rc
= -EPROTO
);
2104 lock
= ldlm_handle2lock(&aa
->lock_handle
);
2106 CERROR("received replay ack for unknown local cookie %#llx"
2107 " remote cookie %#llx from server %s id %s\n",
2108 aa
->lock_handle
.cookie
, reply
->lock_handle
.cookie
,
2109 req
->rq_export
->exp_client_uuid
.uuid
,
2110 libcfs_id2str(req
->rq_peer
));
2111 GOTO(out
, rc
= -ESTALE
);
2114 /* Key change rehash lock in per-export hash with new key */
2115 exp
= req
->rq_export
;
2116 if (exp
&& exp
->exp_lock_hash
) {
2117 /* In the function below, .hs_keycmp resolves to
2118 * ldlm_export_lock_keycmp() */
2119 /* coverity[overrun-buffer-val] */
2120 cfs_hash_rehash_key(exp
->exp_lock_hash
,
2121 &lock
->l_remote_handle
,
2122 &reply
->lock_handle
,
2125 lock
->l_remote_handle
= reply
->lock_handle
;
2128 LDLM_DEBUG(lock
, "replayed lock:");
2129 ptlrpc_import_recovery_state_machine(req
->rq_import
);
2130 LDLM_LOCK_PUT(lock
);
2133 ptlrpc_connect_import(req
->rq_import
);
2138 static int replay_one_lock(struct obd_import
*imp
, struct ldlm_lock
*lock
)
2140 struct ptlrpc_request
*req
;
2141 struct ldlm_async_args
*aa
;
2142 struct ldlm_request
*body
;
2145 /* Bug 11974: Do not replay a lock which is actively being canceled */
2146 if (lock
->l_flags
& LDLM_FL_CANCELING
) {
2147 LDLM_DEBUG(lock
, "Not replaying canceled lock:");
2151 /* If this is reply-less callback lock, we cannot replay it, since
2152 * server might have long dropped it, but notification of that event was
2153 * lost by network. (and server granted conflicting lock already) */
2154 if (lock
->l_flags
& LDLM_FL_CANCEL_ON_BLOCK
) {
2155 LDLM_DEBUG(lock
, "Not replaying reply-less lock:");
2156 ldlm_lock_cancel(lock
);
2161 * If granted mode matches the requested mode, this lock is granted.
2163 * If they differ, but we have a granted mode, then we were granted
2164 * one mode and now want another: ergo, converting.
2166 * If we haven't been granted anything and are on a resource list,
2167 * then we're blocked/waiting.
2169 * If we haven't been granted anything and we're NOT on a resource list,
2170 * then we haven't got a reply yet and don't have a known disposition.
2171 * This happens whenever a lock enqueue is the request that triggers
2174 if (lock
->l_granted_mode
== lock
->l_req_mode
)
2175 flags
= LDLM_FL_REPLAY
| LDLM_FL_BLOCK_GRANTED
;
2176 else if (lock
->l_granted_mode
)
2177 flags
= LDLM_FL_REPLAY
| LDLM_FL_BLOCK_CONV
;
2178 else if (!list_empty(&lock
->l_res_link
))
2179 flags
= LDLM_FL_REPLAY
| LDLM_FL_BLOCK_WAIT
;
2181 flags
= LDLM_FL_REPLAY
;
2183 req
= ptlrpc_request_alloc_pack(imp
, &RQF_LDLM_ENQUEUE
,
2184 LUSTRE_DLM_VERSION
, LDLM_ENQUEUE
);
2188 /* We're part of recovery, so don't wait for it. */
2189 req
->rq_send_state
= LUSTRE_IMP_REPLAY_LOCKS
;
2191 body
= req_capsule_client_get(&req
->rq_pill
, &RMF_DLM_REQ
);
2192 ldlm_lock2desc(lock
, &body
->lock_desc
);
2193 body
->lock_flags
= ldlm_flags_to_wire(flags
);
2195 ldlm_lock2handle(lock
, &body
->lock_handle
[0]);
2196 if (lock
->l_lvb_len
> 0)
2197 req_capsule_extend(&req
->rq_pill
, &RQF_LDLM_ENQUEUE_LVB
);
2198 req_capsule_set_size(&req
->rq_pill
, &RMF_DLM_LVB
, RCL_SERVER
,
2200 ptlrpc_request_set_replen(req
);
2201 /* notify the server we've replayed all requests.
2202 * also, we mark the request to be put on a dedicated
2203 * queue to be processed after all request replayes.
2205 lustre_msg_set_flags(req
->rq_reqmsg
, MSG_REQ_REPLAY_DONE
);
2207 LDLM_DEBUG(lock
, "replaying lock:");
2209 atomic_inc(&req
->rq_import
->imp_replay_inflight
);
2210 CLASSERT(sizeof(*aa
) <= sizeof(req
->rq_async_args
));
2211 aa
= ptlrpc_req_async_args(req
);
2212 aa
->lock_handle
= body
->lock_handle
[0];
2213 req
->rq_interpret_reply
= (ptlrpc_interpterer_t
)replay_lock_interpret
;
2214 ptlrpcd_add_req(req
, PDL_POLICY_LOCAL
, -1);
2220 * Cancel as many unused locks as possible before replay. since we are
2221 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2224 * Called only in recovery before replaying locks. there is no need to
2225 * replay locks that are unused. since the clients may hold thousands of
2226 * cached unused locks, dropping the unused locks can greatly reduce the
2227 * load on the servers at recovery time.
2229 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace
*ns
)
2234 CDEBUG(D_DLMTRACE
, "Dropping as many unused locks as possible before"
2235 "replay for namespace %s (%d)\n",
2236 ldlm_ns_name(ns
), ns
->ns_nr_unused
);
2238 /* We don't need to care whether or not LRU resize is enabled
2239 * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2240 * count parameter */
2241 canceled
= ldlm_cancel_lru_local(ns
, &cancels
, ns
->ns_nr_unused
, 0,
2242 LCF_LOCAL
, LDLM_CANCEL_NO_WAIT
);
2244 CDEBUG(D_DLMTRACE
, "Canceled %d unused locks from namespace %s\n",
2245 canceled
, ldlm_ns_name(ns
));
2248 int ldlm_replay_locks(struct obd_import
*imp
)
2250 struct ldlm_namespace
*ns
= imp
->imp_obd
->obd_namespace
;
2252 struct ldlm_lock
*lock
, *next
;
2255 LASSERT(atomic_read(&imp
->imp_replay_inflight
) == 0);
2257 /* don't replay locks if import failed recovery */
2258 if (imp
->imp_vbr_failed
)
2261 /* ensure this doesn't fall to 0 before all have been queued */
2262 atomic_inc(&imp
->imp_replay_inflight
);
2264 if (ldlm_cancel_unused_locks_before_replay
)
2265 ldlm_cancel_unused_locks_for_replay(ns
);
2267 ldlm_namespace_foreach(ns
, ldlm_chain_lock_for_replay
, &list
);
2269 list_for_each_entry_safe(lock
, next
, &list
, l_pending_chain
) {
2270 list_del_init(&lock
->l_pending_chain
);
2272 LDLM_LOCK_RELEASE(lock
);
2273 continue; /* or try to do the rest? */
2275 rc
= replay_one_lock(imp
, lock
);
2276 LDLM_LOCK_RELEASE(lock
);
2279 atomic_dec(&imp
->imp_replay_inflight
);
2283 EXPORT_SYMBOL(ldlm_replay_locks
);