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)
633 "client-side enqueue, new policy data");
636 if ((*flags) & LDLM_FL_AST_SENT ||
637 /* Cancel extent locks as soon as possible on a liblustre client,
638 * because it cannot handle asynchronous ASTs robustly (see
640 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
641 lock_res_and_lock(lock);
642 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
643 unlock_res_and_lock(lock);
644 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
647 /* If the lock has already been granted by a completion AST, don't
648 * clobber the LVB with an older one. */
650 /* We must lock or a racing completion might update lvb without
651 * letting us know and we'll clobber the correct value.
652 * Cannot unlock after the check either, a that still leaves
653 * a tiny window for completion to get in */
654 lock_res_and_lock(lock);
655 if (lock->l_req_mode != lock->l_granted_mode)
656 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
657 lock->l_lvb_data, size);
658 unlock_res_and_lock(lock);
666 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
667 if (lock->l_completion_ast != NULL) {
668 int err = lock->l_completion_ast(lock, *flags, NULL);
676 if (lvb_len && lvb != NULL) {
677 /* Copy the LVB here, and not earlier, because the completion
678 * AST (if any) can override what we got in the reply */
679 memcpy(lvb, lock->l_lvb_data, lvb_len);
682 LDLM_DEBUG(lock, "client-side enqueue END");
684 if (cleanup_phase == 1 && rc)
685 failed_lock_cleanup(ns, lock, mode);
686 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
688 LDLM_LOCK_RELEASE(lock);
691 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
694 * Estimate number of lock handles that would fit into request of given
695 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
696 * a single page on the send/receive side. XXX: 512 should be changed to
697 * more adequate value.
699 static inline int ldlm_req_handles_avail(int req_size, int off)
703 avail = min_t(int, LDLM_MAXREQSIZE, PAGE_CACHE_SIZE - 512) - req_size;
704 if (likely(avail >= 0))
705 avail /= (int)sizeof(struct lustre_handle);
708 avail += LDLM_LOCKREQ_HANDLES - off;
713 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
714 enum req_location loc,
717 int size = req_capsule_msg_size(pill, loc);
718 return ldlm_req_handles_avail(size, off);
721 static inline int ldlm_format_handles_avail(struct obd_import *imp,
722 const struct req_format *fmt,
723 enum req_location loc, int off)
725 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
726 return ldlm_req_handles_avail(size, off);
730 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
731 * \a count locks in \a cancels.
733 * This is to be called by functions preparing their own requests that
734 * might contain lists of locks to cancel in addition to actual operation
735 * that needs to be performed.
737 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
738 int version, int opc, int canceloff,
739 struct list_head *cancels, int count)
741 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
742 struct req_capsule *pill = &req->rq_pill;
743 struct ldlm_request *dlm = NULL;
744 int flags, avail, to_free, pack = 0;
750 if (ns_connect_cancelset(ns)) {
751 /* Estimate the amount of available space in the request. */
752 req_capsule_filled_sizes(pill, RCL_CLIENT);
753 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
755 flags = ns_connect_lru_resize(ns) ?
756 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
757 to_free = !ns_connect_lru_resize(ns) &&
758 opc == LDLM_ENQUEUE ? 1 : 0;
760 /* Cancel LRU locks here _only_ if the server supports
761 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
762 * RPC, which will make us slower. */
764 count += ldlm_cancel_lru_local(ns, cancels, to_free,
765 avail - count, 0, flags);
770 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
771 ldlm_request_bufsize(pack, opc));
774 rc = ptlrpc_request_pack(req, version, opc);
776 ldlm_lock_list_put(cancels, l_bl_ast, count);
780 if (ns_connect_cancelset(ns)) {
782 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
784 /* Skip first lock handler in ldlm_request_pack(),
785 * this method will increment @lock_count according
786 * to the lock handle amount actually written to
788 dlm->lock_count = canceloff;
790 /* Pack into the request @pack lock handles. */
791 ldlm_cli_cancel_list(cancels, pack, req, 0);
792 /* Prepare and send separate cancel RPC for others. */
793 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
795 ldlm_lock_list_put(cancels, l_bl_ast, count);
799 EXPORT_SYMBOL(ldlm_prep_elc_req);
801 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
802 struct list_head *cancels, int count)
804 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
805 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
807 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
809 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
811 struct ptlrpc_request *req;
814 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
816 return ERR_PTR(-ENOMEM);
818 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
820 ptlrpc_request_free(req);
824 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
825 ptlrpc_request_set_replen(req);
828 EXPORT_SYMBOL(ldlm_enqueue_pack);
831 * Client-side lock enqueue.
833 * If a request has some specific initialisation it is passed in \a reqp,
834 * otherwise it is created in ldlm_cli_enqueue.
836 * Supports sync and async requests, pass \a async flag accordingly. If a
837 * request was created in ldlm_cli_enqueue and it is the async request,
838 * pass it to the caller in \a reqp.
840 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
841 struct ldlm_enqueue_info *einfo,
842 const struct ldlm_res_id *res_id,
843 ldlm_policy_data_t const *policy, __u64 *flags,
844 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
845 struct lustre_handle *lockh, int async)
847 struct ldlm_namespace *ns;
848 struct ldlm_lock *lock;
849 struct ldlm_request *body;
850 int is_replay = *flags & LDLM_FL_REPLAY;
851 int req_passed_in = 1;
853 struct ptlrpc_request *req;
855 LASSERT(exp != NULL);
857 ns = exp->exp_obd->obd_namespace;
859 /* If we're replaying this lock, just check some invariants.
860 * If we're creating a new lock, get everything all setup nice. */
862 lock = ldlm_handle2lock_long(lockh, 0);
863 LASSERT(lock != NULL);
864 LDLM_DEBUG(lock, "client-side enqueue START");
865 LASSERT(exp == lock->l_conn_export);
867 const struct ldlm_callback_suite cbs = {
868 .lcs_completion = einfo->ei_cb_cp,
869 .lcs_blocking = einfo->ei_cb_bl,
870 .lcs_glimpse = einfo->ei_cb_gl
872 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
873 einfo->ei_mode, &cbs, einfo->ei_cbdata,
877 /* for the local lock, add the reference */
878 ldlm_lock_addref_internal(lock, einfo->ei_mode);
879 ldlm_lock2handle(lock, lockh);
881 lock->l_policy_data = *policy;
883 if (einfo->ei_type == LDLM_EXTENT)
884 lock->l_req_extent = policy->l_extent;
885 LDLM_DEBUG(lock, "client-side enqueue START, flags %llx\n",
889 lock->l_conn_export = exp;
890 lock->l_export = NULL;
891 lock->l_blocking_ast = einfo->ei_cb_bl;
892 lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
894 /* lock not sent to server yet */
896 if (reqp == NULL || *reqp == NULL) {
897 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
902 failed_lock_cleanup(ns, lock, einfo->ei_mode);
903 LDLM_LOCK_RELEASE(lock);
913 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
915 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
916 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
919 /* Dump lock data into the request buffer */
920 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
921 ldlm_lock2desc(lock, &body->lock_desc);
922 body->lock_flags = ldlm_flags_to_wire(*flags);
923 body->lock_handle[0] = *lockh;
925 /* Continue as normal. */
926 if (!req_passed_in) {
928 req_capsule_extend(&req->rq_pill,
929 &RQF_LDLM_ENQUEUE_LVB);
930 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
932 ptlrpc_request_set_replen(req);
936 * Liblustre client doesn't get extent locks, except for O_APPEND case
937 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
938 * [i_size, OBD_OBJECT_EOF] lock is taken.
940 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
941 policy->l_extent.end == OBD_OBJECT_EOF));
944 LASSERT(reqp != NULL);
948 LDLM_DEBUG(lock, "sending request");
950 rc = ptlrpc_queue_wait(req);
952 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
953 einfo->ei_mode, flags, lvb, lvb_len,
956 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
957 * one reference that we took */
959 LDLM_LOCK_RELEASE(lock);
963 if (!req_passed_in && req != NULL) {
964 ptlrpc_req_finished(req);
971 EXPORT_SYMBOL(ldlm_cli_enqueue);
973 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
976 struct ldlm_resource *res;
979 if (ns_is_client(ldlm_lock_to_ns(lock))) {
980 CERROR("Trying to cancel local lock\n");
983 LDLM_DEBUG(lock, "client-side local convert");
985 res = ldlm_lock_convert(lock, new_mode, flags);
987 ldlm_reprocess_all(res);
992 LDLM_DEBUG(lock, "client-side local convert handler END");
997 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
998 * conversion of locks which are on the waiting or converting queue */
999 /* Caller of this code is supposed to take care of lock readers/writers
1001 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
1003 struct ldlm_request *body;
1004 struct ldlm_reply *reply;
1005 struct ldlm_lock *lock;
1006 struct ldlm_resource *res;
1007 struct ptlrpc_request *req;
1010 lock = ldlm_handle2lock(lockh);
1017 if (lock->l_conn_export == NULL)
1018 return ldlm_cli_convert_local(lock, new_mode, flags);
1020 LDLM_DEBUG(lock, "client-side convert");
1022 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
1023 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1026 LDLM_LOCK_PUT(lock);
1030 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1031 body->lock_handle[0] = lock->l_remote_handle;
1033 body->lock_desc.l_req_mode = new_mode;
1034 body->lock_flags = ldlm_flags_to_wire(*flags);
1037 ptlrpc_request_set_replen(req);
1038 rc = ptlrpc_queue_wait(req);
1042 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1044 GOTO(out, rc = -EPROTO);
1047 GOTO(out, rc = req->rq_status);
1049 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
1051 ldlm_reprocess_all(res);
1052 /* Go to sleep until the lock is granted. */
1053 /* FIXME: or cancelled. */
1054 if (lock->l_completion_ast) {
1055 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1061 rc = LUSTRE_EDEADLK;
1064 LDLM_LOCK_PUT(lock);
1065 ptlrpc_req_finished(req);
1068 EXPORT_SYMBOL(ldlm_cli_convert);
1071 * Cancel locks locally.
1073 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1074 * \retval LDLM_FL_CANCELING otherwise;
1075 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1077 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1079 __u64 rc = LDLM_FL_LOCAL_ONLY;
1081 if (lock->l_conn_export) {
1084 LDLM_DEBUG(lock, "client-side cancel");
1085 /* Set this flag to prevent others from getting new references*/
1086 lock_res_and_lock(lock);
1087 lock->l_flags |= LDLM_FL_CBPENDING;
1088 local_only = !!(lock->l_flags &
1089 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1090 ldlm_cancel_callback(lock);
1091 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
1092 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1093 unlock_res_and_lock(lock);
1096 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1098 rc = LDLM_FL_LOCAL_ONLY;
1100 ldlm_lock_cancel(lock);
1102 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1103 LDLM_ERROR(lock, "Trying to cancel local lock");
1106 LDLM_DEBUG(lock, "server-side local cancel");
1107 ldlm_lock_cancel(lock);
1108 ldlm_reprocess_all(lock->l_resource);
1115 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1117 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1118 struct list_head *head, int count)
1120 struct ldlm_request *dlm;
1121 struct ldlm_lock *lock;
1122 int max, packed = 0;
1124 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1125 LASSERT(dlm != NULL);
1127 /* Check the room in the request buffer. */
1128 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1129 sizeof(struct ldlm_request);
1130 max /= sizeof(struct lustre_handle);
1131 max += LDLM_LOCKREQ_HANDLES;
1132 LASSERT(max >= dlm->lock_count + count);
1134 /* XXX: it would be better to pack lock handles grouped by resource.
1135 * so that the server cancel would call filter_lvbo_update() less
1137 list_for_each_entry(lock, head, l_bl_ast) {
1140 LASSERT(lock->l_conn_export);
1141 /* Pack the lock handle to the given request buffer. */
1142 LDLM_DEBUG(lock, "packing");
1143 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1146 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1150 * Prepare and send a batched cancel RPC. It will include \a count lock
1151 * handles of locks given in \a cancels list. */
1152 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1153 int count, ldlm_cancel_flags_t flags)
1155 struct ptlrpc_request *req = NULL;
1156 struct obd_import *imp;
1160 LASSERT(exp != NULL);
1163 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1165 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1168 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1169 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1174 imp = class_exp2cliimp(exp);
1175 if (imp == NULL || imp->imp_invalid) {
1177 "skipping cancel on invalid import %p\n", imp);
1181 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1183 GOTO(out, rc = -ENOMEM);
1185 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1186 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1187 ldlm_request_bufsize(count, LDLM_CANCEL));
1189 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1191 ptlrpc_request_free(req);
1195 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1196 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1197 ptlrpc_at_set_req_timeout(req);
1199 ldlm_cancel_pack(req, cancels, count);
1201 ptlrpc_request_set_replen(req);
1202 if (flags & LCF_ASYNC) {
1203 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1207 rc = ptlrpc_queue_wait(req);
1209 if (rc == LUSTRE_ESTALE) {
1210 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1211 "out of sync -- not fatal\n",
1212 libcfs_nid2str(req->rq_import->
1213 imp_connection->c_peer.nid));
1215 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1216 req->rq_import_generation == imp->imp_generation) {
1217 ptlrpc_req_finished(req);
1219 } else if (rc != ELDLM_OK) {
1220 /* -ESHUTDOWN is common on umount */
1221 CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1222 "Got rc %d from cancel RPC: "
1223 "canceling anyway\n", rc);
1230 ptlrpc_req_finished(req);
1232 return sent ? sent : rc;
1234 EXPORT_SYMBOL(ldlm_cli_cancel_req);
1236 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1238 LASSERT(imp != NULL);
1239 return &imp->imp_obd->obd_namespace->ns_pool;
1243 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1245 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1247 struct obd_device *obd;
1251 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1252 !imp_connect_lru_resize(req->rq_import))) {
1254 * Do nothing for corner cases.
1259 /* In some cases RPC may contain SLV and limit zeroed out. This
1260 * is the case when server does not support LRU resize feature.
1261 * This is also possible in some recovery cases when server-side
1262 * reqs have no reference to the OBD export and thus access to
1263 * server-side namespace is not possible. */
1264 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1265 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1266 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found (SLV: %llu, Limit: %u)",
1267 lustre_msg_get_slv(req->rq_repmsg),
1268 lustre_msg_get_limit(req->rq_repmsg));
1272 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1273 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1274 obd = req->rq_import->imp_obd;
1276 /* Set new SLV and limit in OBD fields to make them accessible
1277 * to the pool thread. We do not access obd_namespace and pool
1278 * directly here as there is no reliable way to make sure that
1279 * they are still alive at cleanup time. Evil races are possible
1280 * which may cause Oops at that time. */
1281 write_lock(&obd->obd_pool_lock);
1282 obd->obd_pool_slv = new_slv;
1283 obd->obd_pool_limit = new_limit;
1284 write_unlock(&obd->obd_pool_lock);
1288 EXPORT_SYMBOL(ldlm_cli_update_pool);
1291 * Client side lock cancel.
1293 * Lock must not have any readers or writers by this time.
1295 int ldlm_cli_cancel(struct lustre_handle *lockh,
1296 ldlm_cancel_flags_t cancel_flags)
1298 struct obd_export *exp;
1299 int avail, flags, count = 1;
1301 struct ldlm_namespace *ns;
1302 struct ldlm_lock *lock;
1305 /* concurrent cancels on the same handle can happen */
1306 lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1308 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1312 rc = ldlm_cli_cancel_local(lock);
1313 if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
1314 LDLM_LOCK_RELEASE(lock);
1317 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1318 * RPC which goes to canceld portal, so we can cancel other LRU locks
1319 * here and send them all as one LDLM_CANCEL RPC. */
1320 LASSERT(list_empty(&lock->l_bl_ast));
1321 list_add(&lock->l_bl_ast, &cancels);
1323 exp = lock->l_conn_export;
1324 if (exp_connect_cancelset(exp)) {
1325 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1330 ns = ldlm_lock_to_ns(lock);
1331 flags = ns_connect_lru_resize(ns) ?
1332 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1333 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1336 ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1339 EXPORT_SYMBOL(ldlm_cli_cancel);
1342 * Locally cancel up to \a count locks in list \a cancels.
1343 * Return the number of cancelled locks.
1345 int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
1346 ldlm_cancel_flags_t flags)
1349 struct ldlm_lock *lock, *next;
1350 int left = 0, bl_ast = 0;
1354 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1358 if (flags & LCF_LOCAL) {
1359 rc = LDLM_FL_LOCAL_ONLY;
1360 ldlm_lock_cancel(lock);
1362 rc = ldlm_cli_cancel_local(lock);
1364 /* Until we have compound requests and can send LDLM_CANCEL
1365 * requests batched with generic RPCs, we need to send cancels
1366 * with the LDLM_FL_BL_AST flag in a separate RPC from
1367 * the one being generated now. */
1368 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1369 LDLM_DEBUG(lock, "Cancel lock separately");
1370 list_del_init(&lock->l_bl_ast);
1371 list_add(&lock->l_bl_ast, &head);
1375 if (rc == LDLM_FL_LOCAL_ONLY) {
1376 /* CANCEL RPC should not be sent to server. */
1377 list_del_init(&lock->l_bl_ast);
1378 LDLM_LOCK_RELEASE(lock);
1384 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1389 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
1392 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1393 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1394 * readahead requests, ...)
1396 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1397 struct ldlm_lock *lock,
1398 int unused, int added,
1401 ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1402 ldlm_cancel_for_recovery cb = ns->ns_cancel_for_recovery;
1403 lock_res_and_lock(lock);
1405 /* don't check added & count since we want to process all locks
1406 * from unused list */
1407 switch (lock->l_resource->lr_type) {
1413 result = LDLM_POLICY_SKIP_LOCK;
1414 lock->l_flags |= LDLM_FL_SKIPPED;
1418 unlock_res_and_lock(lock);
1423 * Callback function for LRU-resize policy. Decides whether to keep
1424 * \a lock in LRU for current \a LRU size \a unused, added in current
1425 * scan \a added and number of locks to be preferably canceled \a count.
1427 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1429 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1431 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1432 struct ldlm_lock *lock,
1433 int unused, int added,
1436 unsigned long cur = cfs_time_current();
1437 struct ldlm_pool *pl = &ns->ns_pool;
1441 /* Stop LRU processing when we reach past @count or have checked all
1443 if (count && added >= count)
1444 return LDLM_POLICY_KEEP_LOCK;
1446 slv = ldlm_pool_get_slv(pl);
1447 lvf = ldlm_pool_get_lvf(pl);
1448 la = cfs_duration_sec(cfs_time_sub(cur,
1449 lock->l_last_used));
1450 lv = lvf * la * unused;
1452 /* Inform pool about current CLV to see it via proc. */
1453 ldlm_pool_set_clv(pl, lv);
1455 /* Stop when SLV is not yet come from server or lv is smaller than
1457 return (slv == 0 || lv < slv) ?
1458 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1462 * Callback function for proc used policy. Makes decision whether to keep
1463 * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1464 * added and number of locks to be preferably canceled \a count.
1466 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1468 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1470 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1471 struct ldlm_lock *lock,
1472 int unused, int added,
1475 /* Stop LRU processing when we reach past @count or have checked all
1477 return (added >= count) ?
1478 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1482 * Callback function for aged policy. Makes decision whether to keep \a lock in
1483 * LRU for current LRU size \a unused, added in current scan \a added and
1484 * number of locks to be preferably canceled \a count.
1486 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1488 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1490 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1491 struct ldlm_lock *lock,
1492 int unused, int added,
1495 /* Stop LRU processing if young lock is found and we reach past count */
1496 return ((added >= count) &&
1497 time_before(cfs_time_current(),
1498 cfs_time_add(lock->l_last_used, ns->ns_max_age))) ?
1499 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1503 * Callback function for default policy. Makes decision whether to keep \a lock
1504 * in LRU for current LRU size \a unused, added in current scan \a added and
1505 * number of locks to be preferably canceled \a count.
1507 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1509 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1511 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1512 struct ldlm_lock *lock,
1513 int unused, int added,
1516 /* Stop LRU processing when we reach past count or have checked all
1518 return (added >= count) ?
1519 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1522 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1523 struct ldlm_lock *, int,
1526 static ldlm_cancel_lru_policy_t
1527 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1529 if (flags & LDLM_CANCEL_NO_WAIT)
1530 return ldlm_cancel_no_wait_policy;
1532 if (ns_connect_lru_resize(ns)) {
1533 if (flags & LDLM_CANCEL_SHRINK)
1534 /* We kill passed number of old locks. */
1535 return ldlm_cancel_passed_policy;
1536 else if (flags & LDLM_CANCEL_LRUR)
1537 return ldlm_cancel_lrur_policy;
1538 else if (flags & LDLM_CANCEL_PASSED)
1539 return ldlm_cancel_passed_policy;
1541 if (flags & LDLM_CANCEL_AGED)
1542 return ldlm_cancel_aged_policy;
1545 return ldlm_cancel_default_policy;
1549 * - Free space in LRU for \a count new locks,
1550 * redundant unused locks are canceled locally;
1551 * - also cancel locally unused aged locks;
1552 * - do not cancel more than \a max locks;
1553 * - GET the found locks and add them into the \a cancels list.
1555 * A client lock can be added to the l_bl_ast list only when it is
1556 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1557 * CANCEL. There are the following use cases:
1558 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1559 * ldlm_cli_cancel(), which check and set this flag properly. As any
1560 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1561 * later without any special locking.
1563 * Calling policies for enabled LRU resize:
1564 * ----------------------------------------
1565 * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to
1566 * cancel not more than \a count locks;
1568 * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at
1569 * the beginning of LRU list);
1571 * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to
1572 * memory pressure policy function;
1574 * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy".
1576 * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1577 * (typically before replaying locks) w/o
1578 * sending any RPCs or waiting for any
1579 * outstanding RPC to complete.
1581 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, struct list_head *cancels,
1582 int count, int max, int flags)
1584 ldlm_cancel_lru_policy_t pf;
1585 struct ldlm_lock *lock, *next;
1586 int added = 0, unused, remained;
1588 spin_lock(&ns->ns_lock);
1589 unused = ns->ns_nr_unused;
1592 if (!ns_connect_lru_resize(ns))
1593 count += unused - ns->ns_max_unused;
1595 pf = ldlm_cancel_lru_policy(ns, flags);
1596 LASSERT(pf != NULL);
1598 while (!list_empty(&ns->ns_unused_list)) {
1599 ldlm_policy_res_t result;
1601 /* all unused locks */
1602 if (remained-- <= 0)
1605 /* For any flags, stop scanning if @max is reached. */
1606 if (max && added >= max)
1609 list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1611 /* No locks which got blocking requests. */
1612 LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1614 if (flags & LDLM_CANCEL_NO_WAIT &&
1615 lock->l_flags & LDLM_FL_SKIPPED)
1616 /* already processed */
1619 /* Somebody is already doing CANCEL. No need for this
1620 * lock in LRU, do not traverse it again. */
1621 if (!(lock->l_flags & LDLM_FL_CANCELING))
1624 ldlm_lock_remove_from_lru_nolock(lock);
1626 if (&lock->l_lru == &ns->ns_unused_list)
1629 LDLM_LOCK_GET(lock);
1630 spin_unlock(&ns->ns_lock);
1631 lu_ref_add(&lock->l_reference, __func__, current);
1633 /* Pass the lock through the policy filter and see if it
1634 * should stay in LRU.
1636 * Even for shrinker policy we stop scanning if
1637 * we find a lock that should stay in the cache.
1638 * We should take into account lock age anyway
1639 * as a new lock is a valuable resource even if
1640 * it has a low weight.
1642 * That is, for shrinker policy we drop only
1643 * old locks, but additionally choose them by
1644 * their weight. Big extent locks will stay in
1646 result = pf(ns, lock, unused, added, count);
1647 if (result == LDLM_POLICY_KEEP_LOCK) {
1648 lu_ref_del(&lock->l_reference,
1650 LDLM_LOCK_RELEASE(lock);
1651 spin_lock(&ns->ns_lock);
1654 if (result == LDLM_POLICY_SKIP_LOCK) {
1655 lu_ref_del(&lock->l_reference,
1657 LDLM_LOCK_RELEASE(lock);
1658 spin_lock(&ns->ns_lock);
1662 lock_res_and_lock(lock);
1663 /* Check flags again under the lock. */
1664 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1665 (ldlm_lock_remove_from_lru(lock) == 0)) {
1666 /* Another thread is removing lock from LRU, or
1667 * somebody is already doing CANCEL, or there
1668 * is a blocking request which will send cancel
1669 * by itself, or the lock is no longer unused. */
1670 unlock_res_and_lock(lock);
1671 lu_ref_del(&lock->l_reference,
1673 LDLM_LOCK_RELEASE(lock);
1674 spin_lock(&ns->ns_lock);
1677 LASSERT(!lock->l_readers && !lock->l_writers);
1679 /* If we have chosen to cancel this lock voluntarily, we
1680 * better send cancel notification to server, so that it
1681 * frees appropriate state. This might lead to a race
1682 * where while we are doing cancel here, server is also
1683 * silently cancelling this lock. */
1684 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1686 /* Setting the CBPENDING flag is a little misleading,
1687 * but prevents an important race; namely, once
1688 * CBPENDING is set, the lock can accumulate no more
1689 * readers/writers. Since readers and writers are
1690 * already zero here, ldlm_lock_decref() won't see
1691 * this flag and call l_blocking_ast */
1692 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1694 /* We can't re-add to l_lru as it confuses the
1695 * refcounting in ldlm_lock_remove_from_lru() if an AST
1696 * arrives after we drop lr_lock below. We use l_bl_ast
1697 * and can't use l_pending_chain as it is used both on
1698 * server and client nevertheless bug 5666 says it is
1699 * used only on server */
1700 LASSERT(list_empty(&lock->l_bl_ast));
1701 list_add(&lock->l_bl_ast, cancels);
1702 unlock_res_and_lock(lock);
1703 lu_ref_del(&lock->l_reference, __func__, current);
1704 spin_lock(&ns->ns_lock);
1708 spin_unlock(&ns->ns_lock);
1712 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1713 int count, int max, ldlm_cancel_flags_t cancel_flags,
1717 added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1720 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1724 * Cancel at least \a nr locks from given namespace LRU.
1726 * When called with LCF_ASYNC the blocking callback will be handled
1727 * in a thread and this function will return after the thread has been
1728 * asked to call the callback. When called with LCF_ASYNC the blocking
1729 * callback will be performed in this function.
1731 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
1732 ldlm_cancel_flags_t cancel_flags,
1738 /* Just prepare the list of locks, do not actually cancel them yet.
1739 * Locks are cancelled later in a separate thread. */
1740 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1741 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
1749 * Find and cancel locally unused locks found on resource, matched to the
1750 * given policy, mode. GET the found locks and add them into the \a cancels
1753 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1754 struct list_head *cancels,
1755 ldlm_policy_data_t *policy,
1756 ldlm_mode_t mode, __u64 lock_flags,
1757 ldlm_cancel_flags_t cancel_flags, void *opaque)
1759 struct ldlm_lock *lock;
1763 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1764 if (opaque != NULL && lock->l_ast_data != opaque) {
1765 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1766 lock->l_ast_data, opaque);
1771 if (lock->l_readers || lock->l_writers)
1774 /* If somebody is already doing CANCEL, or blocking AST came,
1775 * skip this lock. */
1776 if (lock->l_flags & LDLM_FL_BL_AST ||
1777 lock->l_flags & LDLM_FL_CANCELING)
1780 if (lockmode_compat(lock->l_granted_mode, mode))
1783 /* If policy is given and this is IBITS lock, add to list only
1784 * those locks that match by policy. */
1785 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1786 !(lock->l_policy_data.l_inodebits.bits &
1787 policy->l_inodebits.bits))
1790 /* See CBPENDING comment in ldlm_cancel_lru */
1791 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1794 LASSERT(list_empty(&lock->l_bl_ast));
1795 list_add(&lock->l_bl_ast, cancels);
1796 LDLM_LOCK_GET(lock);
1801 return ldlm_cli_cancel_list_local(cancels, count, cancel_flags);
1803 EXPORT_SYMBOL(ldlm_cancel_resource_local);
1806 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1808 * If \a req is NULL, send CANCEL request to server with handles of locks
1809 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1810 * separately per lock.
1811 * If \a req is not NULL, put handles of locks in \a cancels into the request
1812 * buffer at the offset \a off.
1813 * Destroy \a cancels at the end.
1815 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1816 struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1818 struct ldlm_lock *lock;
1821 if (list_empty(cancels) || count == 0)
1824 /* XXX: requests (both batched and not) could be sent in parallel.
1825 * Usually it is enough to have just 1 RPC, but it is possible that
1826 * there are too many locks to be cancelled in LRU or on a resource.
1827 * It would also speed up the case when the server does not support
1830 LASSERT(!list_empty(cancels));
1831 lock = list_entry(cancels->next, struct ldlm_lock,
1833 LASSERT(lock->l_conn_export);
1835 if (exp_connect_cancelset(lock->l_conn_export)) {
1838 ldlm_cancel_pack(req, cancels, count);
1840 res = ldlm_cli_cancel_req(lock->l_conn_export,
1844 res = ldlm_cli_cancel_req(lock->l_conn_export,
1849 CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1850 "ldlm_cli_cancel_list: %d\n", res);
1855 ldlm_lock_list_put(cancels, l_bl_ast, res);
1857 LASSERT(count == 0);
1860 EXPORT_SYMBOL(ldlm_cli_cancel_list);
1863 * Cancel all locks on a resource that have 0 readers/writers.
1865 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1866 * to notify the server. */
1867 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1868 const struct ldlm_res_id *res_id,
1869 ldlm_policy_data_t *policy,
1871 ldlm_cancel_flags_t flags,
1874 struct ldlm_resource *res;
1879 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1881 /* This is not a problem. */
1882 CDEBUG(D_INFO, "No resource %llu\n", res_id->name[0]);
1886 LDLM_RESOURCE_ADDREF(res);
1887 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1888 0, flags | LCF_BL_AST, opaque);
1889 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1891 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
1894 LDLM_RESOURCE_DELREF(res);
1895 ldlm_resource_putref(res);
1898 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
1900 struct ldlm_cli_cancel_arg {
1905 static int ldlm_cli_hash_cancel_unused(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1906 struct hlist_node *hnode, void *arg)
1908 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1909 struct ldlm_cli_cancel_arg *lc = arg;
1911 ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1913 lc->lc_flags, lc->lc_opaque);
1914 /* must return 0 for hash iteration */
1919 * Cancel all locks on a namespace (or a specific resource, if given)
1920 * that have 0 readers/writers.
1922 * If flags & LCF_LOCAL, throw the locks away without trying
1923 * to notify the server. */
1924 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1925 const struct ldlm_res_id *res_id,
1926 ldlm_cancel_flags_t flags, void *opaque)
1928 struct ldlm_cli_cancel_arg arg = {
1930 .lc_opaque = opaque,
1936 if (res_id != NULL) {
1937 return ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1941 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1942 ldlm_cli_hash_cancel_unused, &arg);
1946 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1948 /* Lock iterators. */
1950 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1953 struct list_head *tmp, *next;
1954 struct ldlm_lock *lock;
1955 int rc = LDLM_ITER_CONTINUE;
1958 return LDLM_ITER_CONTINUE;
1961 list_for_each_safe(tmp, next, &res->lr_granted) {
1962 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1964 if (iter(lock, closure) == LDLM_ITER_STOP)
1965 GOTO(out, rc = LDLM_ITER_STOP);
1968 list_for_each_safe(tmp, next, &res->lr_converting) {
1969 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1971 if (iter(lock, closure) == LDLM_ITER_STOP)
1972 GOTO(out, rc = LDLM_ITER_STOP);
1975 list_for_each_safe(tmp, next, &res->lr_waiting) {
1976 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1978 if (iter(lock, closure) == LDLM_ITER_STOP)
1979 GOTO(out, rc = LDLM_ITER_STOP);
1985 EXPORT_SYMBOL(ldlm_resource_foreach);
1987 struct iter_helper_data {
1988 ldlm_iterator_t iter;
1992 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1994 struct iter_helper_data *helper = closure;
1995 return helper->iter(lock, helper->closure);
1998 static int ldlm_res_iter_helper(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1999 struct hlist_node *hnode, void *arg)
2002 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2004 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2008 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2009 ldlm_iterator_t iter, void *closure)
2012 struct iter_helper_data helper = {
2017 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2018 ldlm_res_iter_helper, &helper);
2021 EXPORT_SYMBOL(ldlm_namespace_foreach);
2023 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2024 * return 0: find no resource
2025 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2028 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2029 const struct ldlm_res_id *res_id,
2030 ldlm_iterator_t iter, void *data)
2032 struct ldlm_resource *res;
2036 CERROR("must pass in namespace\n");
2040 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2044 LDLM_RESOURCE_ADDREF(res);
2045 rc = ldlm_resource_foreach(res, iter, data);
2046 LDLM_RESOURCE_DELREF(res);
2047 ldlm_resource_putref(res);
2050 EXPORT_SYMBOL(ldlm_resource_iterate);
2054 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2056 struct list_head *list = closure;
2058 /* we use l_pending_chain here, because it's unused on clients. */
2059 LASSERTF(list_empty(&lock->l_pending_chain),
2060 "lock %p next %p prev %p\n",
2061 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2062 /* bug 9573: don't replay locks left after eviction, or
2063 * bug 17614: locks being actively cancelled. Get a reference
2064 * on a lock so that it does not disappear under us (e.g. due to cancel)
2066 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
2067 list_add(&lock->l_pending_chain, list);
2068 LDLM_LOCK_GET(lock);
2071 return LDLM_ITER_CONTINUE;
2074 static int replay_lock_interpret(const struct lu_env *env,
2075 struct ptlrpc_request *req,
2076 struct ldlm_async_args *aa, int rc)
2078 struct ldlm_lock *lock;
2079 struct ldlm_reply *reply;
2080 struct obd_export *exp;
2082 atomic_dec(&req->rq_import->imp_replay_inflight);
2087 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2089 GOTO(out, rc = -EPROTO);
2091 lock = ldlm_handle2lock(&aa->lock_handle);
2093 CERROR("received replay ack for unknown local cookie %#llx"
2094 " remote cookie %#llx from server %s id %s\n",
2095 aa->lock_handle.cookie, reply->lock_handle.cookie,
2096 req->rq_export->exp_client_uuid.uuid,
2097 libcfs_id2str(req->rq_peer));
2098 GOTO(out, rc = -ESTALE);
2101 /* Key change rehash lock in per-export hash with new key */
2102 exp = req->rq_export;
2103 if (exp && exp->exp_lock_hash) {
2104 /* In the function below, .hs_keycmp resolves to
2105 * ldlm_export_lock_keycmp() */
2106 /* coverity[overrun-buffer-val] */
2107 cfs_hash_rehash_key(exp->exp_lock_hash,
2108 &lock->l_remote_handle,
2109 &reply->lock_handle,
2112 lock->l_remote_handle = reply->lock_handle;
2115 LDLM_DEBUG(lock, "replayed lock:");
2116 ptlrpc_import_recovery_state_machine(req->rq_import);
2117 LDLM_LOCK_PUT(lock);
2120 ptlrpc_connect_import(req->rq_import);
2125 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2127 struct ptlrpc_request *req;
2128 struct ldlm_async_args *aa;
2129 struct ldlm_request *body;
2132 /* Bug 11974: Do not replay a lock which is actively being canceled */
2133 if (lock->l_flags & LDLM_FL_CANCELING) {
2134 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2138 /* If this is reply-less callback lock, we cannot replay it, since
2139 * server might have long dropped it, but notification of that event was
2140 * lost by network. (and server granted conflicting lock already) */
2141 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2142 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2143 ldlm_lock_cancel(lock);
2148 * If granted mode matches the requested mode, this lock is granted.
2150 * If they differ, but we have a granted mode, then we were granted
2151 * one mode and now want another: ergo, converting.
2153 * If we haven't been granted anything and are on a resource list,
2154 * then we're blocked/waiting.
2156 * If we haven't been granted anything and we're NOT on a resource list,
2157 * then we haven't got a reply yet and don't have a known disposition.
2158 * This happens whenever a lock enqueue is the request that triggers
2161 if (lock->l_granted_mode == lock->l_req_mode)
2162 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2163 else if (lock->l_granted_mode)
2164 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2165 else if (!list_empty(&lock->l_res_link))
2166 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2168 flags = LDLM_FL_REPLAY;
2170 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2171 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2175 /* We're part of recovery, so don't wait for it. */
2176 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2178 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2179 ldlm_lock2desc(lock, &body->lock_desc);
2180 body->lock_flags = ldlm_flags_to_wire(flags);
2182 ldlm_lock2handle(lock, &body->lock_handle[0]);
2183 if (lock->l_lvb_len > 0)
2184 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2185 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2187 ptlrpc_request_set_replen(req);
2188 /* notify the server we've replayed all requests.
2189 * also, we mark the request to be put on a dedicated
2190 * queue to be processed after all request replayes.
2192 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2194 LDLM_DEBUG(lock, "replaying lock:");
2196 atomic_inc(&req->rq_import->imp_replay_inflight);
2197 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2198 aa = ptlrpc_req_async_args(req);
2199 aa->lock_handle = body->lock_handle[0];
2200 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2201 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2207 * Cancel as many unused locks as possible before replay. since we are
2208 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2211 * Called only in recovery before replaying locks. there is no need to
2212 * replay locks that are unused. since the clients may hold thousands of
2213 * cached unused locks, dropping the unused locks can greatly reduce the
2214 * load on the servers at recovery time.
2216 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2221 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2222 "replay for namespace %s (%d)\n",
2223 ldlm_ns_name(ns), ns->ns_nr_unused);
2225 /* We don't need to care whether or not LRU resize is enabled
2226 * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2227 * count parameter */
2228 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2229 LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2231 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2232 canceled, ldlm_ns_name(ns));
2235 int ldlm_replay_locks(struct obd_import *imp)
2237 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2239 struct ldlm_lock *lock, *next;
2242 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2244 /* don't replay locks if import failed recovery */
2245 if (imp->imp_vbr_failed)
2248 /* ensure this doesn't fall to 0 before all have been queued */
2249 atomic_inc(&imp->imp_replay_inflight);
2251 if (ldlm_cancel_unused_locks_before_replay)
2252 ldlm_cancel_unused_locks_for_replay(ns);
2254 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2256 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2257 list_del_init(&lock->l_pending_chain);
2259 LDLM_LOCK_RELEASE(lock);
2260 continue; /* or try to do the rest? */
2262 rc = replay_one_lock(imp, lock);
2263 LDLM_LOCK_RELEASE(lock);
2266 atomic_dec(&imp->imp_replay_inflight);
2270 EXPORT_SYMBOL(ldlm_replay_locks);