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 ", " 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));
139 return obd_timeout / 2;
140 /* Since these are non-updating timeouts, we should be conservative.
141 It would be nice to have some kind of "early reply" mechanism for
142 lock callbacks too... */
143 timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */
144 return max(timeout, ldlm_enqueue_min);
146 EXPORT_SYMBOL(ldlm_get_enq_timeout);
149 * Helper function for ldlm_completion_ast(), updating timings when lock is
152 static int ldlm_completion_tail(struct ldlm_lock *lock)
157 if (lock->l_flags & (LDLM_FL_DESTROYED | LDLM_FL_FAILED)) {
158 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
161 delay = cfs_time_sub(get_seconds(),
162 lock->l_last_activity);
163 LDLM_DEBUG(lock, "client-side enqueue: granted after "
164 CFS_DURATION_T"s", delay);
166 /* Update our time estimate */
167 at_measured(ldlm_lock_to_ns_at(lock),
175 * Implementation of ->l_completion_ast() for a client, that doesn't wait
176 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
177 * other threads that cannot block for long.
179 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
181 if (flags == LDLM_FL_WAIT_NOREPROC) {
182 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
186 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
187 LDLM_FL_BLOCK_CONV))) {
188 wake_up(&lock->l_waitq);
189 return ldlm_completion_tail(lock);
192 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, going forward");
193 ldlm_reprocess_all(lock->l_resource);
196 EXPORT_SYMBOL(ldlm_completion_ast_async);
199 * Generic LDLM "completion" AST. This is called in several cases:
201 * - when a reply to an ENQUEUE RPC is received from the server
202 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
203 * this point (determined by flags);
205 * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
208 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
211 * - to force all locks when resource is destroyed (cleanup_resource());
213 * - during lock conversion (not used currently).
215 * If lock is not granted in the first case, this function waits until second
216 * or penultimate cases happen in some other thread.
219 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
221 /* XXX ALLOCATE - 160 bytes */
222 struct lock_wait_data lwd;
223 struct obd_device *obd;
224 struct obd_import *imp = NULL;
225 struct l_wait_info lwi;
229 if (flags == LDLM_FL_WAIT_NOREPROC) {
230 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
234 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
235 LDLM_FL_BLOCK_CONV))) {
236 wake_up(&lock->l_waitq);
240 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, sleeping");
244 obd = class_exp2obd(lock->l_conn_export);
246 /* if this is a local lock, then there is no import */
248 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;
605 if (newmode && newmode != lock->l_req_mode) {
606 LDLM_DEBUG(lock, "server returned different mode %s",
607 ldlm_lockname[newmode]);
608 lock->l_req_mode = newmode;
611 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
612 &lock->l_resource->lr_name)) {
613 CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES
614 " instead of "DLDLMRES"\n",
615 PLDLMRES(&reply->lock_desc.l_resource),
616 PLDLMRES(lock->l_resource));
618 rc = ldlm_lock_change_resource(ns, lock,
619 &reply->lock_desc.l_resource.lr_name);
620 if (rc || lock->l_resource == NULL) {
624 LDLM_DEBUG(lock, "client-side enqueue, new resource");
627 if (!(type == LDLM_IBITS &&
628 !(exp_connect_flags(exp) & OBD_CONNECT_IBITS)))
629 /* We assume lock type cannot change on server*/
630 ldlm_convert_policy_to_local(exp,
631 lock->l_resource->lr_type,
632 &reply->lock_desc.l_policy_data,
633 &lock->l_policy_data);
634 if (type != LDLM_PLAIN)
636 "client-side enqueue, new policy data");
639 if ((*flags) & LDLM_FL_AST_SENT ||
640 /* Cancel extent locks as soon as possible on a liblustre client,
641 * because it cannot handle asynchronous ASTs robustly (see
643 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
644 lock_res_and_lock(lock);
645 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
646 unlock_res_and_lock(lock);
647 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
650 /* If the lock has already been granted by a completion AST, don't
651 * clobber the LVB with an older one. */
653 /* We must lock or a racing completion might update lvb without
654 * letting us know and we'll clobber the correct value.
655 * Cannot unlock after the check either, a that still leaves
656 * a tiny window for completion to get in */
657 lock_res_and_lock(lock);
658 if (lock->l_req_mode != lock->l_granted_mode)
659 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
660 lock->l_lvb_data, size);
661 unlock_res_and_lock(lock);
669 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
670 if (lock->l_completion_ast != NULL) {
671 int err = lock->l_completion_ast(lock, *flags, NULL);
679 if (lvb_len && lvb != NULL) {
680 /* Copy the LVB here, and not earlier, because the completion
681 * AST (if any) can override what we got in the reply */
682 memcpy(lvb, lock->l_lvb_data, lvb_len);
685 LDLM_DEBUG(lock, "client-side enqueue END");
687 if (cleanup_phase == 1 && rc)
688 failed_lock_cleanup(ns, lock, mode);
689 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
691 LDLM_LOCK_RELEASE(lock);
694 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
697 * Estimate number of lock handles that would fit into request of given
698 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
699 * a single page on the send/receive side. XXX: 512 should be changed to
700 * more adequate value.
702 static inline int ldlm_req_handles_avail(int req_size, int off)
706 avail = min_t(int, LDLM_MAXREQSIZE, PAGE_CACHE_SIZE - 512) - req_size;
707 if (likely(avail >= 0))
708 avail /= (int)sizeof(struct lustre_handle);
711 avail += LDLM_LOCKREQ_HANDLES - off;
716 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
717 enum req_location loc,
720 int size = req_capsule_msg_size(pill, loc);
721 return ldlm_req_handles_avail(size, off);
724 static inline int ldlm_format_handles_avail(struct obd_import *imp,
725 const struct req_format *fmt,
726 enum req_location loc, int off)
728 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
729 return ldlm_req_handles_avail(size, off);
733 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
734 * \a count locks in \a cancels.
736 * This is to be called by functions preparing their own requests that
737 * might contain lists of locks to cancel in addition to actual operation
738 * that needs to be performed.
740 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
741 int version, int opc, int canceloff,
742 struct list_head *cancels, int count)
744 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
745 struct req_capsule *pill = &req->rq_pill;
746 struct ldlm_request *dlm = NULL;
747 int flags, avail, to_free, pack = 0;
753 if (ns_connect_cancelset(ns)) {
754 /* Estimate the amount of available space in the request. */
755 req_capsule_filled_sizes(pill, RCL_CLIENT);
756 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
758 flags = ns_connect_lru_resize(ns) ?
759 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
760 to_free = !ns_connect_lru_resize(ns) &&
761 opc == LDLM_ENQUEUE ? 1 : 0;
763 /* Cancel LRU locks here _only_ if the server supports
764 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
765 * RPC, which will make us slower. */
767 count += ldlm_cancel_lru_local(ns, cancels, to_free,
768 avail - count, 0, flags);
773 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
774 ldlm_request_bufsize(pack, opc));
777 rc = ptlrpc_request_pack(req, version, opc);
779 ldlm_lock_list_put(cancels, l_bl_ast, count);
783 if (ns_connect_cancelset(ns)) {
785 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
787 /* Skip first lock handler in ldlm_request_pack(),
788 * this method will increment @lock_count according
789 * to the lock handle amount actually written to
791 dlm->lock_count = canceloff;
793 /* Pack into the request @pack lock handles. */
794 ldlm_cli_cancel_list(cancels, pack, req, 0);
795 /* Prepare and send separate cancel RPC for others. */
796 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
798 ldlm_lock_list_put(cancels, l_bl_ast, count);
802 EXPORT_SYMBOL(ldlm_prep_elc_req);
804 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
805 struct list_head *cancels, int count)
807 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
808 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
810 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
812 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
814 struct ptlrpc_request *req;
817 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
819 return ERR_PTR(-ENOMEM);
821 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
823 ptlrpc_request_free(req);
827 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
828 ptlrpc_request_set_replen(req);
831 EXPORT_SYMBOL(ldlm_enqueue_pack);
834 * Client-side lock enqueue.
836 * If a request has some specific initialisation it is passed in \a reqp,
837 * otherwise it is created in ldlm_cli_enqueue.
839 * Supports sync and async requests, pass \a async flag accordingly. If a
840 * request was created in ldlm_cli_enqueue and it is the async request,
841 * pass it to the caller in \a reqp.
843 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
844 struct ldlm_enqueue_info *einfo,
845 const struct ldlm_res_id *res_id,
846 ldlm_policy_data_t const *policy, __u64 *flags,
847 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
848 struct lustre_handle *lockh, int async)
850 struct ldlm_namespace *ns;
851 struct ldlm_lock *lock;
852 struct ldlm_request *body;
853 int is_replay = *flags & LDLM_FL_REPLAY;
854 int req_passed_in = 1;
856 struct ptlrpc_request *req;
858 LASSERT(exp != NULL);
860 ns = exp->exp_obd->obd_namespace;
862 /* If we're replaying this lock, just check some invariants.
863 * If we're creating a new lock, get everything all setup nice. */
865 lock = ldlm_handle2lock_long(lockh, 0);
866 LASSERT(lock != NULL);
867 LDLM_DEBUG(lock, "client-side enqueue START");
868 LASSERT(exp == lock->l_conn_export);
870 const struct ldlm_callback_suite cbs = {
871 .lcs_completion = einfo->ei_cb_cp,
872 .lcs_blocking = einfo->ei_cb_bl,
873 .lcs_glimpse = einfo->ei_cb_gl
875 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
876 einfo->ei_mode, &cbs, einfo->ei_cbdata,
880 /* for the local lock, add the reference */
881 ldlm_lock_addref_internal(lock, einfo->ei_mode);
882 ldlm_lock2handle(lock, lockh);
884 lock->l_policy_data = *policy;
886 if (einfo->ei_type == LDLM_EXTENT)
887 lock->l_req_extent = policy->l_extent;
888 LDLM_DEBUG(lock, "client-side enqueue START, flags %llx\n",
892 lock->l_conn_export = exp;
893 lock->l_export = NULL;
894 lock->l_blocking_ast = einfo->ei_cb_bl;
895 lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
897 /* lock not sent to server yet */
899 if (reqp == NULL || *reqp == NULL) {
900 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
905 failed_lock_cleanup(ns, lock, einfo->ei_mode);
906 LDLM_LOCK_RELEASE(lock);
916 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
918 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
919 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
922 /* Dump lock data into the request buffer */
923 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
924 ldlm_lock2desc(lock, &body->lock_desc);
925 body->lock_flags = ldlm_flags_to_wire(*flags);
926 body->lock_handle[0] = *lockh;
928 /* Continue as normal. */
929 if (!req_passed_in) {
931 req_capsule_extend(&req->rq_pill,
932 &RQF_LDLM_ENQUEUE_LVB);
933 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
935 ptlrpc_request_set_replen(req);
939 * Liblustre client doesn't get extent locks, except for O_APPEND case
940 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
941 * [i_size, OBD_OBJECT_EOF] lock is taken.
943 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
944 policy->l_extent.end == OBD_OBJECT_EOF));
947 LASSERT(reqp != NULL);
951 LDLM_DEBUG(lock, "sending request");
953 rc = ptlrpc_queue_wait(req);
955 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
956 einfo->ei_mode, flags, lvb, lvb_len,
959 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
960 * one reference that we took */
962 LDLM_LOCK_RELEASE(lock);
966 if (!req_passed_in && req != NULL) {
967 ptlrpc_req_finished(req);
974 EXPORT_SYMBOL(ldlm_cli_enqueue);
976 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
979 struct ldlm_resource *res;
982 if (ns_is_client(ldlm_lock_to_ns(lock))) {
983 CERROR("Trying to cancel local lock\n");
986 LDLM_DEBUG(lock, "client-side local convert");
988 res = ldlm_lock_convert(lock, new_mode, flags);
990 ldlm_reprocess_all(res);
995 LDLM_DEBUG(lock, "client-side local convert handler END");
1000 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
1001 * conversion of locks which are on the waiting or converting queue */
1002 /* Caller of this code is supposed to take care of lock readers/writers
1004 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
1006 struct ldlm_request *body;
1007 struct ldlm_reply *reply;
1008 struct ldlm_lock *lock;
1009 struct ldlm_resource *res;
1010 struct ptlrpc_request *req;
1013 lock = ldlm_handle2lock(lockh);
1020 if (lock->l_conn_export == NULL)
1021 return ldlm_cli_convert_local(lock, new_mode, flags);
1023 LDLM_DEBUG(lock, "client-side convert");
1025 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
1026 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1029 LDLM_LOCK_PUT(lock);
1033 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1034 body->lock_handle[0] = lock->l_remote_handle;
1036 body->lock_desc.l_req_mode = new_mode;
1037 body->lock_flags = ldlm_flags_to_wire(*flags);
1040 ptlrpc_request_set_replen(req);
1041 rc = ptlrpc_queue_wait(req);
1045 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1046 if (reply == NULL) {
1051 if (req->rq_status) {
1052 rc = req->rq_status;
1056 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
1058 ldlm_reprocess_all(res);
1059 /* Go to sleep until the lock is granted. */
1060 /* FIXME: or cancelled. */
1061 if (lock->l_completion_ast) {
1062 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1068 rc = LUSTRE_EDEADLK;
1071 LDLM_LOCK_PUT(lock);
1072 ptlrpc_req_finished(req);
1075 EXPORT_SYMBOL(ldlm_cli_convert);
1078 * Cancel locks locally.
1080 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1081 * \retval LDLM_FL_CANCELING otherwise;
1082 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1084 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1086 __u64 rc = LDLM_FL_LOCAL_ONLY;
1088 if (lock->l_conn_export) {
1091 LDLM_DEBUG(lock, "client-side cancel");
1092 /* Set this flag to prevent others from getting new references*/
1093 lock_res_and_lock(lock);
1094 lock->l_flags |= LDLM_FL_CBPENDING;
1095 local_only = !!(lock->l_flags &
1096 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1097 ldlm_cancel_callback(lock);
1098 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
1099 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1100 unlock_res_and_lock(lock);
1103 CDEBUG(D_DLMTRACE, "not sending request (at caller's instruction)\n");
1104 rc = LDLM_FL_LOCAL_ONLY;
1106 ldlm_lock_cancel(lock);
1108 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1109 LDLM_ERROR(lock, "Trying to cancel local lock");
1112 LDLM_DEBUG(lock, "server-side local cancel");
1113 ldlm_lock_cancel(lock);
1114 ldlm_reprocess_all(lock->l_resource);
1121 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1123 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1124 struct list_head *head, int count)
1126 struct ldlm_request *dlm;
1127 struct ldlm_lock *lock;
1128 int max, packed = 0;
1130 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1131 LASSERT(dlm != NULL);
1133 /* Check the room in the request buffer. */
1134 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1135 sizeof(struct ldlm_request);
1136 max /= sizeof(struct lustre_handle);
1137 max += LDLM_LOCKREQ_HANDLES;
1138 LASSERT(max >= dlm->lock_count + count);
1140 /* XXX: it would be better to pack lock handles grouped by resource.
1141 * so that the server cancel would call filter_lvbo_update() less
1143 list_for_each_entry(lock, head, l_bl_ast) {
1146 LASSERT(lock->l_conn_export);
1147 /* Pack the lock handle to the given request buffer. */
1148 LDLM_DEBUG(lock, "packing");
1149 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1152 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1156 * Prepare and send a batched cancel RPC. It will include \a count lock
1157 * handles of locks given in \a cancels list. */
1158 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1159 int count, ldlm_cancel_flags_t flags)
1161 struct ptlrpc_request *req = NULL;
1162 struct obd_import *imp;
1166 LASSERT(exp != NULL);
1169 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1171 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1174 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1175 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1180 imp = class_exp2cliimp(exp);
1181 if (imp == NULL || imp->imp_invalid) {
1183 "skipping cancel on invalid import %p\n", imp);
1187 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1193 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1194 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1195 ldlm_request_bufsize(count, LDLM_CANCEL));
1197 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1199 ptlrpc_request_free(req);
1203 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1204 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1205 ptlrpc_at_set_req_timeout(req);
1207 ldlm_cancel_pack(req, cancels, count);
1209 ptlrpc_request_set_replen(req);
1210 if (flags & LCF_ASYNC) {
1211 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1215 rc = ptlrpc_queue_wait(req);
1217 if (rc == LUSTRE_ESTALE) {
1218 CDEBUG(D_DLMTRACE, "client/server (nid %s) out of sync -- not fatal\n",
1219 libcfs_nid2str(req->rq_import->
1220 imp_connection->c_peer.nid));
1222 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1223 req->rq_import_generation == imp->imp_generation) {
1224 ptlrpc_req_finished(req);
1226 } else if (rc != ELDLM_OK) {
1227 /* -ESHUTDOWN is common on umount */
1228 CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1229 "Got rc %d from cancel RPC: canceling anyway\n",
1237 ptlrpc_req_finished(req);
1239 return sent ? sent : rc;
1241 EXPORT_SYMBOL(ldlm_cli_cancel_req);
1243 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1245 LASSERT(imp != NULL);
1246 return &imp->imp_obd->obd_namespace->ns_pool;
1250 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1252 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1254 struct obd_device *obd;
1258 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1259 !imp_connect_lru_resize(req->rq_import))) {
1261 * Do nothing for corner cases.
1266 /* In some cases RPC may contain SLV and limit zeroed out. This
1267 * is the case when server does not support LRU resize feature.
1268 * This is also possible in some recovery cases when server-side
1269 * reqs have no reference to the OBD export and thus access to
1270 * server-side namespace is not possible. */
1271 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1272 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1273 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found (SLV: %llu, Limit: %u)",
1274 lustre_msg_get_slv(req->rq_repmsg),
1275 lustre_msg_get_limit(req->rq_repmsg));
1279 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1280 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1281 obd = req->rq_import->imp_obd;
1283 /* Set new SLV and limit in OBD fields to make them accessible
1284 * to the pool thread. We do not access obd_namespace and pool
1285 * directly here as there is no reliable way to make sure that
1286 * they are still alive at cleanup time. Evil races are possible
1287 * which may cause Oops at that time. */
1288 write_lock(&obd->obd_pool_lock);
1289 obd->obd_pool_slv = new_slv;
1290 obd->obd_pool_limit = new_limit;
1291 write_unlock(&obd->obd_pool_lock);
1295 EXPORT_SYMBOL(ldlm_cli_update_pool);
1298 * Client side lock cancel.
1300 * Lock must not have any readers or writers by this time.
1302 int ldlm_cli_cancel(struct lustre_handle *lockh,
1303 ldlm_cancel_flags_t cancel_flags)
1305 struct obd_export *exp;
1306 int avail, flags, count = 1;
1308 struct ldlm_namespace *ns;
1309 struct ldlm_lock *lock;
1312 /* concurrent cancels on the same handle can happen */
1313 lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1315 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1319 rc = ldlm_cli_cancel_local(lock);
1320 if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
1321 LDLM_LOCK_RELEASE(lock);
1324 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1325 * RPC which goes to canceld portal, so we can cancel other LRU locks
1326 * here and send them all as one LDLM_CANCEL RPC. */
1327 LASSERT(list_empty(&lock->l_bl_ast));
1328 list_add(&lock->l_bl_ast, &cancels);
1330 exp = lock->l_conn_export;
1331 if (exp_connect_cancelset(exp)) {
1332 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1337 ns = ldlm_lock_to_ns(lock);
1338 flags = ns_connect_lru_resize(ns) ?
1339 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1340 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1343 ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1346 EXPORT_SYMBOL(ldlm_cli_cancel);
1349 * Locally cancel up to \a count locks in list \a cancels.
1350 * Return the number of cancelled locks.
1352 int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
1353 ldlm_cancel_flags_t flags)
1356 struct ldlm_lock *lock, *next;
1357 int left = 0, bl_ast = 0;
1361 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1365 if (flags & LCF_LOCAL) {
1366 rc = LDLM_FL_LOCAL_ONLY;
1367 ldlm_lock_cancel(lock);
1369 rc = ldlm_cli_cancel_local(lock);
1371 /* Until we have compound requests and can send LDLM_CANCEL
1372 * requests batched with generic RPCs, we need to send cancels
1373 * with the LDLM_FL_BL_AST flag in a separate RPC from
1374 * the one being generated now. */
1375 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1376 LDLM_DEBUG(lock, "Cancel lock separately");
1377 list_del_init(&lock->l_bl_ast);
1378 list_add(&lock->l_bl_ast, &head);
1382 if (rc == LDLM_FL_LOCAL_ONLY) {
1383 /* CANCEL RPC should not be sent to server. */
1384 list_del_init(&lock->l_bl_ast);
1385 LDLM_LOCK_RELEASE(lock);
1391 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1396 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
1399 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1400 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1401 * readahead requests, ...)
1403 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1404 struct ldlm_lock *lock,
1405 int unused, int added,
1408 ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1409 ldlm_cancel_for_recovery cb = ns->ns_cancel_for_recovery;
1410 lock_res_and_lock(lock);
1412 /* don't check added & count since we want to process all locks
1413 * from unused list */
1414 switch (lock->l_resource->lr_type) {
1420 result = LDLM_POLICY_SKIP_LOCK;
1421 lock->l_flags |= LDLM_FL_SKIPPED;
1425 unlock_res_and_lock(lock);
1430 * Callback function for LRU-resize policy. Decides whether to keep
1431 * \a lock in LRU for current \a LRU size \a unused, added in current
1432 * scan \a added and number of locks to be preferably canceled \a count.
1434 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1436 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1438 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1439 struct ldlm_lock *lock,
1440 int unused, int added,
1443 unsigned long cur = cfs_time_current();
1444 struct ldlm_pool *pl = &ns->ns_pool;
1448 /* Stop LRU processing when we reach past @count or have checked all
1450 if (count && added >= count)
1451 return LDLM_POLICY_KEEP_LOCK;
1453 slv = ldlm_pool_get_slv(pl);
1454 lvf = ldlm_pool_get_lvf(pl);
1455 la = cfs_duration_sec(cfs_time_sub(cur,
1456 lock->l_last_used));
1457 lv = lvf * la * unused;
1459 /* Inform pool about current CLV to see it via proc. */
1460 ldlm_pool_set_clv(pl, lv);
1462 /* Stop when SLV is not yet come from server or lv is smaller than
1464 return (slv == 0 || lv < slv) ?
1465 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1469 * Callback function for proc used policy. Makes decision whether to keep
1470 * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1471 * added and number of locks to be preferably canceled \a count.
1473 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1475 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1477 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1478 struct ldlm_lock *lock,
1479 int unused, int added,
1482 /* Stop LRU processing when we reach past @count or have checked all
1484 return (added >= count) ?
1485 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1489 * Callback function for aged policy. Makes decision whether to keep \a lock in
1490 * LRU for current LRU size \a unused, added in current scan \a added and
1491 * number of locks to be preferably canceled \a count.
1493 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1495 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1497 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1498 struct ldlm_lock *lock,
1499 int unused, int added,
1502 /* Stop LRU processing if young lock is found and we reach past count */
1503 return ((added >= count) &&
1504 time_before(cfs_time_current(),
1505 cfs_time_add(lock->l_last_used, ns->ns_max_age))) ?
1506 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1510 * Callback function for default policy. Makes decision whether to keep \a lock
1511 * in LRU for current LRU size \a unused, added in current scan \a added and
1512 * number of locks to be preferably canceled \a count.
1514 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1516 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1518 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1519 struct ldlm_lock *lock,
1520 int unused, int added,
1523 /* Stop LRU processing when we reach past count or have checked all
1525 return (added >= count) ?
1526 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1529 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1530 struct ldlm_lock *, int,
1533 static ldlm_cancel_lru_policy_t
1534 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1536 if (flags & LDLM_CANCEL_NO_WAIT)
1537 return ldlm_cancel_no_wait_policy;
1539 if (ns_connect_lru_resize(ns)) {
1540 if (flags & LDLM_CANCEL_SHRINK)
1541 /* We kill passed number of old locks. */
1542 return ldlm_cancel_passed_policy;
1543 else if (flags & LDLM_CANCEL_LRUR)
1544 return ldlm_cancel_lrur_policy;
1545 else if (flags & LDLM_CANCEL_PASSED)
1546 return ldlm_cancel_passed_policy;
1548 if (flags & LDLM_CANCEL_AGED)
1549 return ldlm_cancel_aged_policy;
1552 return ldlm_cancel_default_policy;
1556 * - Free space in LRU for \a count new locks,
1557 * redundant unused locks are canceled locally;
1558 * - also cancel locally unused aged locks;
1559 * - do not cancel more than \a max locks;
1560 * - GET the found locks and add them into the \a cancels list.
1562 * A client lock can be added to the l_bl_ast list only when it is
1563 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1564 * CANCEL. There are the following use cases:
1565 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1566 * ldlm_cli_cancel(), which check and set this flag properly. As any
1567 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1568 * later without any special locking.
1570 * Calling policies for enabled LRU resize:
1571 * ----------------------------------------
1572 * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to
1573 * cancel not more than \a count locks;
1575 * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at
1576 * the beginning of LRU list);
1578 * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to
1579 * memory pressure policy function;
1581 * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy".
1583 * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1584 * (typically before replaying locks) w/o
1585 * sending any RPCs or waiting for any
1586 * outstanding RPC to complete.
1588 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, struct list_head *cancels,
1589 int count, int max, int flags)
1591 ldlm_cancel_lru_policy_t pf;
1592 struct ldlm_lock *lock, *next;
1593 int added = 0, unused, remained;
1595 spin_lock(&ns->ns_lock);
1596 unused = ns->ns_nr_unused;
1599 if (!ns_connect_lru_resize(ns))
1600 count += unused - ns->ns_max_unused;
1602 pf = ldlm_cancel_lru_policy(ns, flags);
1603 LASSERT(pf != NULL);
1605 while (!list_empty(&ns->ns_unused_list)) {
1606 ldlm_policy_res_t result;
1608 /* all unused locks */
1609 if (remained-- <= 0)
1612 /* For any flags, stop scanning if @max is reached. */
1613 if (max && added >= max)
1616 list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1618 /* No locks which got blocking requests. */
1619 LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1621 if (flags & LDLM_CANCEL_NO_WAIT &&
1622 lock->l_flags & LDLM_FL_SKIPPED)
1623 /* already processed */
1626 /* Somebody is already doing CANCEL. No need for this
1627 * lock in LRU, do not traverse it again. */
1628 if (!(lock->l_flags & LDLM_FL_CANCELING))
1631 ldlm_lock_remove_from_lru_nolock(lock);
1633 if (&lock->l_lru == &ns->ns_unused_list)
1636 LDLM_LOCK_GET(lock);
1637 spin_unlock(&ns->ns_lock);
1638 lu_ref_add(&lock->l_reference, __func__, current);
1640 /* Pass the lock through the policy filter and see if it
1641 * should stay in LRU.
1643 * Even for shrinker policy we stop scanning if
1644 * we find a lock that should stay in the cache.
1645 * We should take into account lock age anyway
1646 * as a new lock is a valuable resource even if
1647 * it has a low weight.
1649 * That is, for shrinker policy we drop only
1650 * old locks, but additionally choose them by
1651 * their weight. Big extent locks will stay in
1653 result = pf(ns, lock, unused, added, count);
1654 if (result == LDLM_POLICY_KEEP_LOCK) {
1655 lu_ref_del(&lock->l_reference,
1657 LDLM_LOCK_RELEASE(lock);
1658 spin_lock(&ns->ns_lock);
1661 if (result == LDLM_POLICY_SKIP_LOCK) {
1662 lu_ref_del(&lock->l_reference,
1664 LDLM_LOCK_RELEASE(lock);
1665 spin_lock(&ns->ns_lock);
1669 lock_res_and_lock(lock);
1670 /* Check flags again under the lock. */
1671 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1672 (ldlm_lock_remove_from_lru(lock) == 0)) {
1673 /* Another thread is removing lock from LRU, or
1674 * somebody is already doing CANCEL, or there
1675 * is a blocking request which will send cancel
1676 * by itself, or the lock is no longer unused. */
1677 unlock_res_and_lock(lock);
1678 lu_ref_del(&lock->l_reference,
1680 LDLM_LOCK_RELEASE(lock);
1681 spin_lock(&ns->ns_lock);
1684 LASSERT(!lock->l_readers && !lock->l_writers);
1686 /* If we have chosen to cancel this lock voluntarily, we
1687 * better send cancel notification to server, so that it
1688 * frees appropriate state. This might lead to a race
1689 * where while we are doing cancel here, server is also
1690 * silently cancelling this lock. */
1691 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1693 /* Setting the CBPENDING flag is a little misleading,
1694 * but prevents an important race; namely, once
1695 * CBPENDING is set, the lock can accumulate no more
1696 * readers/writers. Since readers and writers are
1697 * already zero here, ldlm_lock_decref() won't see
1698 * this flag and call l_blocking_ast */
1699 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1701 /* We can't re-add to l_lru as it confuses the
1702 * refcounting in ldlm_lock_remove_from_lru() if an AST
1703 * arrives after we drop lr_lock below. We use l_bl_ast
1704 * and can't use l_pending_chain as it is used both on
1705 * server and client nevertheless bug 5666 says it is
1706 * used only on server */
1707 LASSERT(list_empty(&lock->l_bl_ast));
1708 list_add(&lock->l_bl_ast, cancels);
1709 unlock_res_and_lock(lock);
1710 lu_ref_del(&lock->l_reference, __func__, current);
1711 spin_lock(&ns->ns_lock);
1715 spin_unlock(&ns->ns_lock);
1719 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1720 int count, int max, ldlm_cancel_flags_t cancel_flags,
1724 added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1727 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1731 * Cancel at least \a nr locks from given namespace LRU.
1733 * When called with LCF_ASYNC the blocking callback will be handled
1734 * in a thread and this function will return after the thread has been
1735 * asked to call the callback. When called with LCF_ASYNC the blocking
1736 * callback will be performed in this function.
1738 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
1739 ldlm_cancel_flags_t cancel_flags,
1745 /* Just prepare the list of locks, do not actually cancel them yet.
1746 * Locks are cancelled later in a separate thread. */
1747 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1748 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
1756 * Find and cancel locally unused locks found on resource, matched to the
1757 * given policy, mode. GET the found locks and add them into the \a cancels
1760 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1761 struct list_head *cancels,
1762 ldlm_policy_data_t *policy,
1763 ldlm_mode_t mode, __u64 lock_flags,
1764 ldlm_cancel_flags_t cancel_flags, void *opaque)
1766 struct ldlm_lock *lock;
1770 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1771 if (opaque != NULL && lock->l_ast_data != opaque) {
1772 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1773 lock->l_ast_data, opaque);
1778 if (lock->l_readers || lock->l_writers)
1781 /* If somebody is already doing CANCEL, or blocking AST came,
1782 * skip this lock. */
1783 if (lock->l_flags & LDLM_FL_BL_AST ||
1784 lock->l_flags & LDLM_FL_CANCELING)
1787 if (lockmode_compat(lock->l_granted_mode, mode))
1790 /* If policy is given and this is IBITS lock, add to list only
1791 * those locks that match by policy. */
1792 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1793 !(lock->l_policy_data.l_inodebits.bits &
1794 policy->l_inodebits.bits))
1797 /* See CBPENDING comment in ldlm_cancel_lru */
1798 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1801 LASSERT(list_empty(&lock->l_bl_ast));
1802 list_add(&lock->l_bl_ast, cancels);
1803 LDLM_LOCK_GET(lock);
1808 return ldlm_cli_cancel_list_local(cancels, count, cancel_flags);
1810 EXPORT_SYMBOL(ldlm_cancel_resource_local);
1813 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1815 * If \a req is NULL, send CANCEL request to server with handles of locks
1816 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1817 * separately per lock.
1818 * If \a req is not NULL, put handles of locks in \a cancels into the request
1819 * buffer at the offset \a off.
1820 * Destroy \a cancels at the end.
1822 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1823 struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1825 struct ldlm_lock *lock;
1828 if (list_empty(cancels) || count == 0)
1831 /* XXX: requests (both batched and not) could be sent in parallel.
1832 * Usually it is enough to have just 1 RPC, but it is possible that
1833 * there are too many locks to be cancelled in LRU or on a resource.
1834 * It would also speed up the case when the server does not support
1837 LASSERT(!list_empty(cancels));
1838 lock = list_entry(cancels->next, struct ldlm_lock,
1840 LASSERT(lock->l_conn_export);
1842 if (exp_connect_cancelset(lock->l_conn_export)) {
1845 ldlm_cancel_pack(req, cancels, count);
1847 res = ldlm_cli_cancel_req(lock->l_conn_export,
1851 res = ldlm_cli_cancel_req(lock->l_conn_export,
1856 CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1857 "ldlm_cli_cancel_list: %d\n", res);
1862 ldlm_lock_list_put(cancels, l_bl_ast, res);
1864 LASSERT(count == 0);
1867 EXPORT_SYMBOL(ldlm_cli_cancel_list);
1870 * Cancel all locks on a resource that have 0 readers/writers.
1872 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1873 * to notify the server. */
1874 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1875 const struct ldlm_res_id *res_id,
1876 ldlm_policy_data_t *policy,
1878 ldlm_cancel_flags_t flags,
1881 struct ldlm_resource *res;
1886 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1888 /* This is not a problem. */
1889 CDEBUG(D_INFO, "No resource %llu\n", res_id->name[0]);
1893 LDLM_RESOURCE_ADDREF(res);
1894 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1895 0, flags | LCF_BL_AST, opaque);
1896 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1898 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
1901 LDLM_RESOURCE_DELREF(res);
1902 ldlm_resource_putref(res);
1905 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
1907 struct ldlm_cli_cancel_arg {
1912 static int ldlm_cli_hash_cancel_unused(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1913 struct hlist_node *hnode, void *arg)
1915 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1916 struct ldlm_cli_cancel_arg *lc = arg;
1918 ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1920 lc->lc_flags, lc->lc_opaque);
1921 /* must return 0 for hash iteration */
1926 * Cancel all locks on a namespace (or a specific resource, if given)
1927 * that have 0 readers/writers.
1929 * If flags & LCF_LOCAL, throw the locks away without trying
1930 * to notify the server. */
1931 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1932 const struct ldlm_res_id *res_id,
1933 ldlm_cancel_flags_t flags, void *opaque)
1935 struct ldlm_cli_cancel_arg arg = {
1937 .lc_opaque = opaque,
1943 if (res_id != NULL) {
1944 return ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1948 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1949 ldlm_cli_hash_cancel_unused, &arg);
1953 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1955 /* Lock iterators. */
1957 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1960 struct list_head *tmp, *next;
1961 struct ldlm_lock *lock;
1962 int rc = LDLM_ITER_CONTINUE;
1965 return LDLM_ITER_CONTINUE;
1968 list_for_each_safe(tmp, next, &res->lr_granted) {
1969 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1971 if (iter(lock, closure) == LDLM_ITER_STOP) {
1972 rc = LDLM_ITER_STOP;
1977 list_for_each_safe(tmp, next, &res->lr_converting) {
1978 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1980 if (iter(lock, closure) == LDLM_ITER_STOP) {
1981 rc = LDLM_ITER_STOP;
1986 list_for_each_safe(tmp, next, &res->lr_waiting) {
1987 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1989 if (iter(lock, closure) == LDLM_ITER_STOP) {
1990 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);
2101 if (reply == NULL) {
2106 lock = ldlm_handle2lock(&aa->lock_handle);
2108 CERROR("received replay ack for unknown local cookie %#llx remote cookie %#llx from server %s id %s\n",
2109 aa->lock_handle.cookie, reply->lock_handle.cookie,
2110 req->rq_export->exp_client_uuid.uuid,
2111 libcfs_id2str(req->rq_peer));
2116 /* Key change rehash lock in per-export hash with new key */
2117 exp = req->rq_export;
2118 if (exp && exp->exp_lock_hash) {
2119 /* In the function below, .hs_keycmp resolves to
2120 * ldlm_export_lock_keycmp() */
2121 /* coverity[overrun-buffer-val] */
2122 cfs_hash_rehash_key(exp->exp_lock_hash,
2123 &lock->l_remote_handle,
2124 &reply->lock_handle,
2127 lock->l_remote_handle = reply->lock_handle;
2130 LDLM_DEBUG(lock, "replayed lock:");
2131 ptlrpc_import_recovery_state_machine(req->rq_import);
2132 LDLM_LOCK_PUT(lock);
2135 ptlrpc_connect_import(req->rq_import);
2140 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2142 struct ptlrpc_request *req;
2143 struct ldlm_async_args *aa;
2144 struct ldlm_request *body;
2147 /* Bug 11974: Do not replay a lock which is actively being canceled */
2148 if (lock->l_flags & LDLM_FL_CANCELING) {
2149 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2153 /* If this is reply-less callback lock, we cannot replay it, since
2154 * server might have long dropped it, but notification of that event was
2155 * lost by network. (and server granted conflicting lock already) */
2156 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2157 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2158 ldlm_lock_cancel(lock);
2163 * If granted mode matches the requested mode, this lock is granted.
2165 * If they differ, but we have a granted mode, then we were granted
2166 * one mode and now want another: ergo, converting.
2168 * If we haven't been granted anything and are on a resource list,
2169 * then we're blocked/waiting.
2171 * If we haven't been granted anything and we're NOT on a resource list,
2172 * then we haven't got a reply yet and don't have a known disposition.
2173 * This happens whenever a lock enqueue is the request that triggers
2176 if (lock->l_granted_mode == lock->l_req_mode)
2177 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2178 else if (lock->l_granted_mode)
2179 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2180 else if (!list_empty(&lock->l_res_link))
2181 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2183 flags = LDLM_FL_REPLAY;
2185 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2186 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2190 /* We're part of recovery, so don't wait for it. */
2191 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2193 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2194 ldlm_lock2desc(lock, &body->lock_desc);
2195 body->lock_flags = ldlm_flags_to_wire(flags);
2197 ldlm_lock2handle(lock, &body->lock_handle[0]);
2198 if (lock->l_lvb_len > 0)
2199 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2200 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2202 ptlrpc_request_set_replen(req);
2203 /* notify the server we've replayed all requests.
2204 * also, we mark the request to be put on a dedicated
2205 * queue to be processed after all request replayes.
2207 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2209 LDLM_DEBUG(lock, "replaying lock:");
2211 atomic_inc(&req->rq_import->imp_replay_inflight);
2212 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2213 aa = ptlrpc_req_async_args(req);
2214 aa->lock_handle = body->lock_handle[0];
2215 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2216 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2222 * Cancel as many unused locks as possible before replay. since we are
2223 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2226 * Called only in recovery before replaying locks. there is no need to
2227 * replay locks that are unused. since the clients may hold thousands of
2228 * cached unused locks, dropping the unused locks can greatly reduce the
2229 * load on the servers at recovery time.
2231 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2236 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before replay for namespace %s (%d)\n",
2237 ldlm_ns_name(ns), ns->ns_nr_unused);
2239 /* We don't need to care whether or not LRU resize is enabled
2240 * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2241 * count parameter */
2242 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2243 LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2245 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2246 canceled, ldlm_ns_name(ns));
2249 int ldlm_replay_locks(struct obd_import *imp)
2251 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2253 struct ldlm_lock *lock, *next;
2256 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2258 /* don't replay locks if import failed recovery */
2259 if (imp->imp_vbr_failed)
2262 /* ensure this doesn't fall to 0 before all have been queued */
2263 atomic_inc(&imp->imp_replay_inflight);
2265 if (ldlm_cancel_unused_locks_before_replay)
2266 ldlm_cancel_unused_locks_for_replay(ns);
2268 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2270 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2271 list_del_init(&lock->l_pending_chain);
2273 LDLM_LOCK_RELEASE(lock);
2274 continue; /* or try to do the rest? */
2276 rc = replay_one_lock(imp, lock);
2277 LDLM_LOCK_RELEASE(lock);
2280 atomic_dec(&imp->imp_replay_inflight);
2284 EXPORT_SYMBOL(ldlm_replay_locks);