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, 2011, Whamcloud, Inc.
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
65 #include <liblustre.h>
68 #include <lustre_dlm.h>
69 #include <obd_class.h>
72 #include "ldlm_internal.h"
74 int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
75 CFS_MODULE_PARM(ldlm_enqueue_min, "i", int, 0644,
76 "lock enqueue timeout minimum");
78 /* in client side, whether the cached locks will be canceled before replay */
79 unsigned int ldlm_cancel_unused_locks_before_replay = 1;
81 static void interrupted_completion_wait(void *data)
85 struct lock_wait_data {
86 struct ldlm_lock *lwd_lock;
90 struct ldlm_async_args {
91 struct lustre_handle lock_handle;
94 int ldlm_expired_completion_wait(void *data)
96 struct lock_wait_data *lwd = data;
97 struct ldlm_lock *lock = lwd->lwd_lock;
98 struct obd_import *imp;
99 struct obd_device *obd;
102 if (lock->l_conn_export == NULL) {
103 static cfs_time_t next_dump = 0, last_dump = 0;
105 if (ptlrpc_check_suspend())
108 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
109 CFS_DURATION_T"s ago); not entering recovery in "
110 "server code, just going back to sleep",
111 lock->l_last_activity,
112 cfs_time_sub(cfs_time_current_sec(),
113 lock->l_last_activity));
114 if (cfs_time_after(cfs_time_current(), next_dump)) {
115 last_dump = next_dump;
116 next_dump = cfs_time_shift(300);
117 ldlm_namespace_dump(D_DLMTRACE,
118 ldlm_lock_to_ns(lock));
120 libcfs_debug_dumplog();
125 obd = lock->l_conn_export->exp_obd;
126 imp = obd->u.cli.cl_import;
127 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
128 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
129 CFS_DURATION_T"s ago), entering recovery for %s@%s",
130 lock->l_last_activity,
131 cfs_time_sub(cfs_time_current_sec(), lock->l_last_activity),
132 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
136 EXPORT_SYMBOL(ldlm_expired_completion_wait);
138 /* We use the same basis for both server side and client side functions
139 from a single node. */
140 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
142 int timeout = at_get(ldlm_lock_to_ns_at(lock));
144 return obd_timeout / 2;
145 /* Since these are non-updating timeouts, we should be conservative.
146 It would be nice to have some kind of "early reply" mechanism for
147 lock callbacks too... */
148 timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */
149 return max(timeout, ldlm_enqueue_min);
151 EXPORT_SYMBOL(ldlm_get_enq_timeout);
154 * Helper function for ldlm_completion_ast(), updating timings when lock is
157 static int ldlm_completion_tail(struct ldlm_lock *lock)
162 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
163 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
166 delay = cfs_time_sub(cfs_time_current_sec(),
167 lock->l_last_activity);
168 LDLM_DEBUG(lock, "client-side enqueue: granted after "
169 CFS_DURATION_T"s", delay);
171 /* Update our time estimate */
172 at_measured(ldlm_lock_to_ns_at(lock),
180 * Implementation of ->l_completion_ast() for a client, that doesn't wait
181 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
182 * other threads that cannot block for long.
184 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
188 if (flags == LDLM_FL_WAIT_NOREPROC) {
189 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
193 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
194 LDLM_FL_BLOCK_CONV))) {
195 cfs_waitq_signal(&lock->l_waitq);
196 RETURN(ldlm_completion_tail(lock));
199 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
201 ldlm_reprocess_all(lock->l_resource);
204 EXPORT_SYMBOL(ldlm_completion_ast_async);
207 * Generic LDLM "completion" AST. This is called in several cases:
209 * - when a reply to an ENQUEUE RPC is received from the server
210 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
211 * this point (determined by flags);
213 * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
216 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
219 * - to force all locks when resource is destroyed (cleanup_resource());
221 * - during lock conversion (not used currently).
223 * If lock is not granted in the first case, this function waits until second
224 * or penultimate cases happen in some other thread.
227 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
229 /* XXX ALLOCATE - 160 bytes */
230 struct lock_wait_data lwd;
231 struct obd_device *obd;
232 struct obd_import *imp = NULL;
233 struct l_wait_info lwi;
238 if (flags == LDLM_FL_WAIT_NOREPROC) {
239 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
243 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
244 LDLM_FL_BLOCK_CONV))) {
245 cfs_waitq_signal(&lock->l_waitq);
249 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
254 obd = class_exp2obd(lock->l_conn_export);
256 /* if this is a local lock, then there is no import */
258 imp = obd->u.cli.cl_import;
261 /* Wait a long time for enqueue - server may have to callback a
262 lock from another client. Server will evict the other client if it
263 doesn't respond reasonably, and then give us the lock. */
264 timeout = ldlm_get_enq_timeout(lock) * 2;
268 if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
269 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
270 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
272 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
273 ldlm_expired_completion_wait,
274 interrupted_completion_wait, &lwd);
278 spin_lock(&imp->imp_lock);
279 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
280 spin_unlock(&imp->imp_lock);
283 if (ns_is_client(ldlm_lock_to_ns(lock)) &&
284 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
285 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
286 lock->l_flags |= LDLM_FL_FAIL_LOC;
289 /* Go to sleep until the lock is granted or cancelled. */
290 rc = l_wait_event(lock->l_waitq,
291 is_granted_or_cancelled(lock), &lwi);
295 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
300 RETURN(ldlm_completion_tail(lock));
302 EXPORT_SYMBOL(ldlm_completion_ast);
305 * A helper to build a blocking AST function
307 * Perform a common operation for blocking ASTs:
308 * defferred lock cancellation.
310 * \param lock the lock blocking or canceling AST was called on
312 * \see mdt_blocking_ast
313 * \see ldlm_blocking_ast
315 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
320 lock->l_flags |= LDLM_FL_CBPENDING;
321 do_ast = (!lock->l_readers && !lock->l_writers);
322 unlock_res_and_lock(lock);
325 struct lustre_handle lockh;
328 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
329 ldlm_lock2handle(lock, &lockh);
330 rc = ldlm_cli_cancel(&lockh);
332 CERROR("ldlm_cli_cancel: %d\n", rc);
334 LDLM_DEBUG(lock, "Lock still has references, will be "
339 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
342 * Server blocking AST
344 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
347 * \param lock the lock which blocks a request or cancelling lock
350 * \param flag indicates whether this cancelling or blocking callback
352 * \see ldlm_blocking_ast_nocheck
354 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
355 void *data, int flag)
359 if (flag == LDLM_CB_CANCELING) {
360 /* Don't need to do anything here. */
364 lock_res_and_lock(lock);
365 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
366 * that ldlm_blocking_ast is called just before intent_policy method
367 * takes the lr_lock, then by the time we get the lock, we might not
368 * be the correct blocking function anymore. So check, and return
370 if (lock->l_blocking_ast != ldlm_blocking_ast) {
371 unlock_res_and_lock(lock);
374 RETURN(ldlm_blocking_ast_nocheck(lock));
376 EXPORT_SYMBOL(ldlm_blocking_ast);
379 * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
380 * comment in filter_intent_policy() on why you may need this.
382 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
385 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
386 * that is rather subtle: with OST-side locking, it may so happen that
387 * _all_ extent locks are held by the OST. If client wants to obtain
388 * current file size it calls ll{,u}_glimpse_size(), and (as locks are
389 * on the server), dummy glimpse callback fires and does
390 * nothing. Client still receives correct file size due to the
391 * following fragment in filter_intent_policy():
393 * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
394 * if (rc != 0 && res->lr_namespace->ns_lvbo &&
395 * res->lr_namespace->ns_lvbo->lvbo_update) {
396 * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
399 * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
400 * returns correct file size to the client.
402 return -ELDLM_NO_LOCK_DATA;
404 EXPORT_SYMBOL(ldlm_glimpse_ast);
407 * Enqueue a local lock (typically on a server).
409 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
410 const struct ldlm_res_id *res_id,
411 ldlm_type_t type, ldlm_policy_data_t *policy,
412 ldlm_mode_t mode, __u64 *flags,
413 ldlm_blocking_callback blocking,
414 ldlm_completion_callback completion,
415 ldlm_glimpse_callback glimpse,
416 void *data, __u32 lvb_len, enum lvb_type lvb_type,
417 const __u64 *client_cookie,
418 struct lustre_handle *lockh)
420 struct ldlm_lock *lock;
422 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
423 .lcs_blocking = blocking,
424 .lcs_glimpse = glimpse,
428 LASSERT(!(*flags & LDLM_FL_REPLAY));
429 if (unlikely(ns_is_client(ns))) {
430 CERROR("Trying to enqueue local lock in a shadow namespace\n");
434 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
437 GOTO(out_nolock, err = -ENOMEM);
439 ldlm_lock2handle(lock, lockh);
441 /* NB: we don't have any lock now (lock_res_and_lock)
442 * because it's a new lock */
443 ldlm_lock_addref_internal_nolock(lock, mode);
444 lock->l_flags |= LDLM_FL_LOCAL;
445 if (*flags & LDLM_FL_ATOMIC_CB)
446 lock->l_flags |= LDLM_FL_ATOMIC_CB;
449 lock->l_policy_data = *policy;
450 if (client_cookie != NULL)
451 lock->l_client_cookie = *client_cookie;
452 if (type == LDLM_EXTENT)
453 lock->l_req_extent = policy->l_extent;
455 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
456 if (unlikely(err != ELDLM_OK))
460 *policy = lock->l_policy_data;
462 if (lock->l_completion_ast)
463 lock->l_completion_ast(lock, *flags, NULL);
465 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
468 LDLM_LOCK_RELEASE(lock);
472 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
474 static void failed_lock_cleanup(struct ldlm_namespace *ns,
475 struct ldlm_lock *lock, int mode)
479 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
480 lock_res_and_lock(lock);
481 /* Check that lock is not granted or failed, we might race. */
482 if ((lock->l_req_mode != lock->l_granted_mode) &&
483 !(lock->l_flags & LDLM_FL_FAILED)) {
484 /* Make sure that this lock will not be found by raced
485 * bl_ast and -EINVAL reply is sent to server anyways.
487 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
488 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
491 unlock_res_and_lock(lock);
495 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
496 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
498 LDLM_DEBUG(lock, "lock was granted or failed in race");
500 ldlm_lock_decref_internal(lock, mode);
502 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
503 * from llite/file.c/ll_file_flock(). */
504 /* This code makes for the fact that we do not have blocking handler on
505 * a client for flock locks. As such this is the place where we must
506 * completely kill failed locks. (interrupted and those that
507 * were waiting to be granted when server evicted us. */
508 if (lock->l_resource->lr_type == LDLM_FLOCK) {
509 lock_res_and_lock(lock);
510 ldlm_resource_unlink_lock(lock);
511 ldlm_lock_destroy_nolock(lock);
512 unlock_res_and_lock(lock);
517 * Finishing portion of client lock enqueue code.
519 * Called after receiving reply from server.
521 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
522 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
523 __u64 *flags, void *lvb, __u32 lvb_len,
524 struct lustre_handle *lockh,int rc)
526 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
527 int is_replay = *flags & LDLM_FL_REPLAY;
528 struct ldlm_lock *lock;
529 struct ldlm_reply *reply;
530 int cleanup_phase = 1;
534 lock = ldlm_handle2lock(lockh);
535 /* ldlm_cli_enqueue is holding a reference on this lock. */
537 LASSERT(type == LDLM_FLOCK);
541 LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
542 "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
544 if (rc != ELDLM_OK) {
546 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
547 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
549 if (rc != ELDLM_LOCK_ABORTED)
553 /* Before we return, swab the reply */
554 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
556 GOTO(cleanup, rc = -EPROTO);
559 LASSERT(lvb != NULL);
561 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
564 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
565 GOTO(cleanup, rc = size);
566 } else if (unlikely(size > lvb_len)) {
567 LDLM_ERROR(lock, "Replied LVB is larger than "
568 "expectation, expected = %d, replied = %d",
570 GOTO(cleanup, rc = -EINVAL);
574 if (rc == ELDLM_LOCK_ABORTED) {
576 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
578 GOTO(cleanup, rc = (rc != 0 ? rc : ELDLM_LOCK_ABORTED));
581 /* lock enqueued on the server */
584 lock_res_and_lock(lock);
585 /* Key change rehash lock in per-export hash with new key */
586 if (exp->exp_lock_hash) {
587 /* In the function below, .hs_keycmp resolves to
588 * ldlm_export_lock_keycmp() */
589 /* coverity[overrun-buffer-val] */
590 cfs_hash_rehash_key(exp->exp_lock_hash,
591 &lock->l_remote_handle,
595 lock->l_remote_handle = reply->lock_handle;
598 *flags = ldlm_flags_from_wire(reply->lock_flags);
599 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
601 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
602 * to wait with no timeout as well */
603 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
605 unlock_res_and_lock(lock);
607 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%llx\n",
608 lock, reply->lock_handle.cookie, *flags);
610 /* If enqueue returned a blocked lock but the completion handler has
611 * already run, then it fixed up the resource and we don't need to do it
613 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
614 int newmode = reply->lock_desc.l_req_mode;
616 if (newmode && newmode != lock->l_req_mode) {
617 LDLM_DEBUG(lock, "server returned different mode %s",
618 ldlm_lockname[newmode]);
619 lock->l_req_mode = newmode;
622 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
623 lock->l_resource->lr_name.name,
624 sizeof(struct ldlm_res_id))) {
625 CDEBUG(D_INFO, "remote intent success, locking "
626 "(%ld,%ld,%ld) instead of "
628 (long)reply->lock_desc.l_resource.lr_name.name[0],
629 (long)reply->lock_desc.l_resource.lr_name.name[1],
630 (long)reply->lock_desc.l_resource.lr_name.name[2],
631 (long)lock->l_resource->lr_name.name[0],
632 (long)lock->l_resource->lr_name.name[1],
633 (long)lock->l_resource->lr_name.name[2]);
635 rc = ldlm_lock_change_resource(ns, lock,
636 &reply->lock_desc.l_resource.lr_name);
637 if (rc || lock->l_resource == NULL)
638 GOTO(cleanup, rc = -ENOMEM);
639 LDLM_DEBUG(lock, "client-side enqueue, new resource");
642 if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
644 /* We assume lock type cannot change on server*/
645 ldlm_convert_policy_to_local(exp,
646 lock->l_resource->lr_type,
647 &reply->lock_desc.l_policy_data,
648 &lock->l_policy_data);
649 if (type != LDLM_PLAIN)
650 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
653 if ((*flags) & LDLM_FL_AST_SENT ||
654 /* Cancel extent locks as soon as possible on a liblustre client,
655 * because it cannot handle asynchronous ASTs robustly (see
657 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
658 lock_res_and_lock(lock);
659 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
660 unlock_res_and_lock(lock);
661 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
664 /* If the lock has already been granted by a completion AST, don't
665 * clobber the LVB with an older one. */
667 /* We must lock or a racing completion might update lvb without
668 * letting us know and we'll clobber the correct value.
669 * Cannot unlock after the check either, a that still leaves
670 * a tiny window for completion to get in */
671 lock_res_and_lock(lock);
672 if (lock->l_req_mode != lock->l_granted_mode)
673 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
674 lock->l_lvb_data, size);
675 unlock_res_and_lock(lock);
681 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
682 if (lock->l_completion_ast != NULL) {
683 int err = lock->l_completion_ast(lock, *flags, NULL);
691 if (lvb_len && lvb != NULL) {
692 /* Copy the LVB here, and not earlier, because the completion
693 * AST (if any) can override what we got in the reply */
694 memcpy(lvb, lock->l_lvb_data, lvb_len);
697 LDLM_DEBUG(lock, "client-side enqueue END");
700 if (cleanup_phase == 1 && rc)
701 failed_lock_cleanup(ns, lock, mode);
702 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
704 LDLM_LOCK_RELEASE(lock);
707 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
710 * Estimate number of lock handles that would fit into request of given
711 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
712 * a single page on the send/receive side. XXX: 512 should be changed to
713 * more adequate value.
715 static inline int ldlm_req_handles_avail(int req_size, int off)
719 avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size;
720 if (likely(avail >= 0))
721 avail /= (int)sizeof(struct lustre_handle);
724 avail += LDLM_LOCKREQ_HANDLES - off;
729 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
730 enum req_location loc,
733 int size = req_capsule_msg_size(pill, loc);
734 return ldlm_req_handles_avail(size, off);
737 static inline int ldlm_format_handles_avail(struct obd_import *imp,
738 const struct req_format *fmt,
739 enum req_location loc, int off)
741 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
742 return ldlm_req_handles_avail(size, off);
746 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
747 * \a count locks in \a cancels.
749 * This is to be called by functions preparing their own requests that
750 * might contain lists of locks to cancel in addition to actual operation
751 * that needs to be performed.
753 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
754 int version, int opc, int canceloff,
755 cfs_list_t *cancels, int count)
757 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
758 struct req_capsule *pill = &req->rq_pill;
759 struct ldlm_request *dlm = NULL;
760 int flags, avail, to_free, pack = 0;
767 if (ns_connect_cancelset(ns)) {
768 /* Estimate the amount of available space in the request. */
769 req_capsule_filled_sizes(pill, RCL_CLIENT);
770 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
772 flags = ns_connect_lru_resize(ns) ?
773 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
774 to_free = !ns_connect_lru_resize(ns) &&
775 opc == LDLM_ENQUEUE ? 1 : 0;
777 /* Cancel LRU locks here _only_ if the server supports
778 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
779 * RPC, which will make us slower. */
781 count += ldlm_cancel_lru_local(ns, cancels, to_free,
782 avail - count, 0, flags);
787 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
788 ldlm_request_bufsize(pack, opc));
791 rc = ptlrpc_request_pack(req, version, opc);
793 ldlm_lock_list_put(cancels, l_bl_ast, count);
797 if (ns_connect_cancelset(ns)) {
799 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
801 /* Skip first lock handler in ldlm_request_pack(),
802 * this method will incrment @lock_count according
803 * to the lock handle amount actually written to
805 dlm->lock_count = canceloff;
807 /* Pack into the request @pack lock handles. */
808 ldlm_cli_cancel_list(cancels, pack, req, 0);
809 /* Prepare and send separate cancel RPC for others. */
810 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
812 ldlm_lock_list_put(cancels, l_bl_ast, count);
816 EXPORT_SYMBOL(ldlm_prep_elc_req);
818 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
819 cfs_list_t *cancels, int count)
821 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
822 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
824 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
827 * Client-side lock enqueue.
829 * If a request has some specific initialisation it is passed in \a reqp,
830 * otherwise it is created in ldlm_cli_enqueue.
832 * Supports sync and async requests, pass \a async flag accordingly. If a
833 * request was created in ldlm_cli_enqueue and it is the async request,
834 * pass it to the caller in \a reqp.
836 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
837 struct ldlm_enqueue_info *einfo,
838 const struct ldlm_res_id *res_id,
839 ldlm_policy_data_t const *policy, __u64 *flags,
840 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
841 struct lustre_handle *lockh, int async)
843 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
844 struct ldlm_lock *lock;
845 struct ldlm_request *body;
846 int is_replay = *flags & LDLM_FL_REPLAY;
847 int req_passed_in = 1;
849 struct ptlrpc_request *req;
852 LASSERT(exp != NULL);
854 /* If we're replaying this lock, just check some invariants.
855 * If we're creating a new lock, get everything all setup nice. */
857 lock = ldlm_handle2lock_long(lockh, 0);
858 LASSERT(lock != NULL);
859 LDLM_DEBUG(lock, "client-side enqueue START");
860 LASSERT(exp == lock->l_conn_export);
862 const struct ldlm_callback_suite cbs = {
863 .lcs_completion = einfo->ei_cb_cp,
864 .lcs_blocking = einfo->ei_cb_bl,
865 .lcs_glimpse = einfo->ei_cb_gl,
866 .lcs_weigh = einfo->ei_cb_wg
868 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
869 einfo->ei_mode, &cbs, einfo->ei_cbdata,
873 /* for the local lock, add the reference */
874 ldlm_lock_addref_internal(lock, einfo->ei_mode);
875 ldlm_lock2handle(lock, lockh);
876 if (policy != NULL) {
877 /* INODEBITS_INTEROP: If the server does not support
878 * inodebits, we will request a plain lock in the
879 * descriptor (ldlm_lock2desc() below) but use an
880 * inodebits lock internally with both bits set.
882 if (einfo->ei_type == LDLM_IBITS &&
883 !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
884 lock->l_policy_data.l_inodebits.bits =
885 MDS_INODELOCK_LOOKUP |
886 MDS_INODELOCK_UPDATE;
888 lock->l_policy_data = *policy;
891 if (einfo->ei_type == LDLM_EXTENT)
892 lock->l_req_extent = policy->l_extent;
893 LDLM_DEBUG(lock, "client-side enqueue START, flags %llx\n",
897 lock->l_conn_export = exp;
898 lock->l_export = NULL;
899 lock->l_blocking_ast = einfo->ei_cb_bl;
900 lock->l_flags |= (*flags & LDLM_FL_NO_LRU);
902 /* lock not sent to server yet */
904 if (reqp == NULL || *reqp == NULL) {
905 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
910 failed_lock_cleanup(ns, lock, einfo->ei_mode);
911 LDLM_LOCK_RELEASE(lock);
921 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
923 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
924 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
927 /* Dump lock data into the request buffer */
928 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
929 ldlm_lock2desc(lock, &body->lock_desc);
930 body->lock_flags = ldlm_flags_to_wire(*flags);
931 body->lock_handle[0] = *lockh;
933 /* Continue as normal. */
934 if (!req_passed_in) {
936 req_capsule_extend(&req->rq_pill,
937 &RQF_LDLM_ENQUEUE_LVB);
938 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
940 ptlrpc_request_set_replen(req);
944 * Liblustre client doesn't get extent locks, except for O_APPEND case
945 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
946 * [i_size, OBD_OBJECT_EOF] lock is taken.
948 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
949 policy->l_extent.end == OBD_OBJECT_EOF));
952 LASSERT(reqp != NULL);
956 LDLM_DEBUG(lock, "sending request");
958 rc = ptlrpc_queue_wait(req);
960 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
961 einfo->ei_mode, flags, lvb, lvb_len,
964 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
965 * one reference that we took */
967 LDLM_LOCK_RELEASE(lock);
971 if (!req_passed_in && req != NULL) {
972 ptlrpc_req_finished(req);
979 EXPORT_SYMBOL(ldlm_cli_enqueue);
981 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
984 struct ldlm_resource *res;
987 if (ns_is_client(ldlm_lock_to_ns(lock))) {
988 CERROR("Trying to cancel local lock\n");
991 LDLM_DEBUG(lock, "client-side local convert");
993 res = ldlm_lock_convert(lock, new_mode, flags);
995 ldlm_reprocess_all(res);
1000 LDLM_DEBUG(lock, "client-side local convert handler END");
1001 LDLM_LOCK_PUT(lock);
1005 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
1006 * conversion of locks which are on the waiting or converting queue */
1007 /* Caller of this code is supposed to take care of lock readers/writers
1009 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
1011 struct ldlm_request *body;
1012 struct ldlm_reply *reply;
1013 struct ldlm_lock *lock;
1014 struct ldlm_resource *res;
1015 struct ptlrpc_request *req;
1019 lock = ldlm_handle2lock(lockh);
1026 if (lock->l_conn_export == NULL)
1027 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
1029 LDLM_DEBUG(lock, "client-side convert");
1031 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
1032 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1035 LDLM_LOCK_PUT(lock);
1039 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1040 body->lock_handle[0] = lock->l_remote_handle;
1042 body->lock_desc.l_req_mode = new_mode;
1043 body->lock_flags = ldlm_flags_to_wire(*flags);
1046 ptlrpc_request_set_replen(req);
1047 rc = ptlrpc_queue_wait(req);
1051 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1053 GOTO(out, rc = -EPROTO);
1056 GOTO(out, rc = req->rq_status);
1058 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
1060 ldlm_reprocess_all(res);
1061 /* Go to sleep until the lock is granted. */
1062 /* FIXME: or cancelled. */
1063 if (lock->l_completion_ast) {
1064 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1074 LDLM_LOCK_PUT(lock);
1075 ptlrpc_req_finished(req);
1078 EXPORT_SYMBOL(ldlm_cli_convert);
1081 * Cancel locks locally.
1083 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1084 * \retval LDLM_FL_CANCELING otherwise;
1085 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1087 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1089 __u64 rc = LDLM_FL_LOCAL_ONLY;
1092 if (lock->l_conn_export) {
1095 LDLM_DEBUG(lock, "client-side cancel");
1096 /* Set this flag to prevent others from getting new references*/
1097 lock_res_and_lock(lock);
1098 lock->l_flags |= LDLM_FL_CBPENDING;
1099 local_only = !!(lock->l_flags &
1100 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1101 ldlm_cancel_callback(lock);
1102 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
1103 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1104 unlock_res_and_lock(lock);
1107 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1109 rc = LDLM_FL_LOCAL_ONLY;
1111 ldlm_lock_cancel(lock);
1113 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1114 LDLM_ERROR(lock, "Trying to cancel local lock");
1117 LDLM_DEBUG(lock, "server-side local cancel");
1118 ldlm_lock_cancel(lock);
1119 ldlm_reprocess_all(lock->l_resource);
1126 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1128 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1129 cfs_list_t *head, int count)
1131 struct ldlm_request *dlm;
1132 struct ldlm_lock *lock;
1133 int max, packed = 0;
1136 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1137 LASSERT(dlm != NULL);
1139 /* Check the room in the request buffer. */
1140 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1141 sizeof(struct ldlm_request);
1142 max /= sizeof(struct lustre_handle);
1143 max += LDLM_LOCKREQ_HANDLES;
1144 LASSERT(max >= dlm->lock_count + count);
1146 /* XXX: it would be better to pack lock handles grouped by resource.
1147 * so that the server cancel would call filter_lvbo_update() less
1149 cfs_list_for_each_entry(lock, head, l_bl_ast) {
1152 LASSERT(lock->l_conn_export);
1153 /* Pack the lock handle to the given request buffer. */
1154 LDLM_DEBUG(lock, "packing");
1155 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1158 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1163 * Prepare and send a batched cancel RPC. It will include \a count lock
1164 * handles of locks given in \a cancels list. */
1165 int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels,
1166 int count, ldlm_cancel_flags_t flags)
1168 struct ptlrpc_request *req = NULL;
1169 struct obd_import *imp;
1174 LASSERT(exp != NULL);
1177 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1179 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1182 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1183 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1188 imp = class_exp2cliimp(exp);
1189 if (imp == NULL || imp->imp_invalid) {
1191 "skipping cancel on invalid import %p\n", imp);
1195 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1197 GOTO(out, rc = -ENOMEM);
1199 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1200 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1201 ldlm_request_bufsize(count, LDLM_CANCEL));
1203 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1205 ptlrpc_request_free(req);
1209 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1210 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1211 ptlrpc_at_set_req_timeout(req);
1213 ldlm_cancel_pack(req, cancels, count);
1215 ptlrpc_request_set_replen(req);
1216 if (flags & LCF_ASYNC) {
1217 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1221 rc = ptlrpc_queue_wait(req);
1224 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1225 "out of sync -- not fatal\n",
1226 libcfs_nid2str(req->rq_import->
1227 imp_connection->c_peer.nid));
1229 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1230 req->rq_import_generation == imp->imp_generation) {
1231 ptlrpc_req_finished(req);
1233 } else if (rc != ELDLM_OK) {
1234 if (rc != -ESHUTDOWN)
1235 CERROR("Got rc %d from cancel RPC: canceling "
1243 ptlrpc_req_finished(req);
1246 return sent ? sent : rc;
1248 EXPORT_SYMBOL(ldlm_cli_cancel_req);
1250 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1252 LASSERT(imp != NULL);
1253 return &imp->imp_obd->obd_namespace->ns_pool;
1257 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1259 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1261 struct obd_device *obd;
1265 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1266 !imp_connect_lru_resize(req->rq_import)))
1269 * Do nothing for corner cases.
1274 /* In some cases RPC may contain SLV and limit zeroed out. This
1275 * is the case when server does not support LRU resize feature.
1276 * This is also possible in some recovery cases when server-side
1277 * reqs have no reference to the OBD export and thus access to
1278 * server-side namespace is not possible. */
1279 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1280 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1281 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1282 "(SLV: "LPU64", Limit: %u)",
1283 lustre_msg_get_slv(req->rq_repmsg),
1284 lustre_msg_get_limit(req->rq_repmsg));
1288 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1289 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1290 obd = req->rq_import->imp_obd;
1292 /* Set new SLV and limit in OBD fields to make them accessible
1293 * to the pool thread. We do not access obd_namespace and pool
1294 * directly here as there is no reliable way to make sure that
1295 * they are still alive at cleanup time. Evil races are possible
1296 * which may cause Oops at that time. */
1297 write_lock(&obd->obd_pool_lock);
1298 obd->obd_pool_slv = new_slv;
1299 obd->obd_pool_limit = new_limit;
1300 write_unlock(&obd->obd_pool_lock);
1304 EXPORT_SYMBOL(ldlm_cli_update_pool);
1307 * Client side lock cancel.
1309 * Lock must not have any readers or writers by this time.
1311 int ldlm_cli_cancel(struct lustre_handle *lockh)
1313 struct obd_export *exp;
1314 int avail, flags, count = 1;
1316 struct ldlm_namespace *ns;
1317 struct ldlm_lock *lock;
1318 CFS_LIST_HEAD(cancels);
1321 /* concurrent cancels on the same handle can happen */
1322 lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1324 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1328 rc = ldlm_cli_cancel_local(lock);
1329 if (rc == LDLM_FL_LOCAL_ONLY) {
1330 LDLM_LOCK_RELEASE(lock);
1333 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1334 * RPC which goes to canceld portal, so we can cancel other LRU locks
1335 * here and send them all as one LDLM_CANCEL RPC. */
1336 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1337 cfs_list_add(&lock->l_bl_ast, &cancels);
1339 exp = lock->l_conn_export;
1340 if (exp_connect_cancelset(exp)) {
1341 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1346 ns = ldlm_lock_to_ns(lock);
1347 flags = ns_connect_lru_resize(ns) ?
1348 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1349 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1352 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1355 EXPORT_SYMBOL(ldlm_cli_cancel);
1358 * Locally cancel up to \a count locks in list \a cancels.
1359 * Return the number of cancelled locks.
1361 int ldlm_cli_cancel_list_local(cfs_list_t *cancels, int count,
1362 ldlm_cancel_flags_t flags)
1364 CFS_LIST_HEAD(head);
1365 struct ldlm_lock *lock, *next;
1366 int left = 0, bl_ast = 0;
1370 cfs_list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1374 if (flags & LCF_LOCAL) {
1375 rc = LDLM_FL_LOCAL_ONLY;
1376 ldlm_lock_cancel(lock);
1378 rc = ldlm_cli_cancel_local(lock);
1380 /* Until we have compound requests and can send LDLM_CANCEL
1381 * requests batched with generic RPCs, we need to send cancels
1382 * with the LDLM_FL_BL_AST flag in a separate RPC from
1383 * the one being generated now. */
1384 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1385 LDLM_DEBUG(lock, "Cancel lock separately");
1386 cfs_list_del_init(&lock->l_bl_ast);
1387 cfs_list_add(&lock->l_bl_ast, &head);
1391 if (rc == LDLM_FL_LOCAL_ONLY) {
1392 /* CANCEL RPC should not be sent to server. */
1393 cfs_list_del_init(&lock->l_bl_ast);
1394 LDLM_LOCK_RELEASE(lock);
1400 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1405 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
1408 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1409 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1410 * readahead requests, ...)
1412 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1413 struct ldlm_lock *lock,
1414 int unused, int added,
1417 ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1418 ldlm_cancel_for_recovery cb = ns->ns_cancel_for_recovery;
1419 lock_res_and_lock(lock);
1421 /* don't check added & count since we want to process all locks
1422 * from unused list */
1423 switch (lock->l_resource->lr_type) {
1429 result = LDLM_POLICY_SKIP_LOCK;
1430 lock->l_flags |= LDLM_FL_SKIPPED;
1434 unlock_res_and_lock(lock);
1439 * Callback function for LRU-resize policy. Decides whether to keep
1440 * \a lock in LRU for current \a LRU size \a unused, added in current
1441 * scan \a added and number of locks to be preferably canceled \a count.
1443 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1445 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1447 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1448 struct ldlm_lock *lock,
1449 int unused, int added,
1452 cfs_time_t cur = cfs_time_current();
1453 struct ldlm_pool *pl = &ns->ns_pool;
1457 /* Stop LRU processing when we reach past @count or have checked all
1459 if (count && added >= count)
1460 return LDLM_POLICY_KEEP_LOCK;
1462 slv = ldlm_pool_get_slv(pl);
1463 lvf = ldlm_pool_get_lvf(pl);
1464 la = cfs_duration_sec(cfs_time_sub(cur,
1465 lock->l_last_used));
1466 lv = lvf * la * unused;
1468 /* Inform pool about current CLV to see it via proc. */
1469 ldlm_pool_set_clv(pl, lv);
1471 /* Stop when SLV is not yet come from server or lv is smaller than
1473 return (slv == 0 || lv < slv) ?
1474 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1478 * Callback function for proc used policy. Makes decision whether to keep
1479 * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1480 * added and number of locks to be preferably canceled \a count.
1482 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1484 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1486 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1487 struct ldlm_lock *lock,
1488 int unused, int added,
1491 /* Stop LRU processing when we reach past @count or have checked all
1493 return (added >= count) ?
1494 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1498 * Callback function for aged policy. Makes decision whether to keep \a lock in
1499 * LRU for current LRU size \a unused, added in current scan \a added and
1500 * number of locks to be preferably canceled \a count.
1502 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1504 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1506 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1507 struct ldlm_lock *lock,
1508 int unused, int added,
1511 /* Stop LRU processing if young lock is found and we reach past count */
1512 return ((added >= count) &&
1513 cfs_time_before(cfs_time_current(),
1514 cfs_time_add(lock->l_last_used,
1516 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1520 * Callback function for default policy. Makes decision whether to keep \a lock
1521 * in LRU for current LRU size \a unused, added in current scan \a added and
1522 * number of locks to be preferably canceled \a count.
1524 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1526 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1528 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1529 struct ldlm_lock *lock,
1530 int unused, int added,
1533 /* Stop LRU processing when we reach past count or have checked all
1535 return (added >= count) ?
1536 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1539 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1540 struct ldlm_lock *, int,
1543 static ldlm_cancel_lru_policy_t
1544 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1546 if (flags & LDLM_CANCEL_NO_WAIT)
1547 return ldlm_cancel_no_wait_policy;
1549 if (ns_connect_lru_resize(ns)) {
1550 if (flags & LDLM_CANCEL_SHRINK)
1551 /* We kill passed number of old locks. */
1552 return ldlm_cancel_passed_policy;
1553 else if (flags & LDLM_CANCEL_LRUR)
1554 return ldlm_cancel_lrur_policy;
1555 else if (flags & LDLM_CANCEL_PASSED)
1556 return ldlm_cancel_passed_policy;
1558 if (flags & LDLM_CANCEL_AGED)
1559 return ldlm_cancel_aged_policy;
1562 return ldlm_cancel_default_policy;
1566 * - Free space in LRU for \a count new locks,
1567 * redundant unused locks are canceled locally;
1568 * - also cancel locally unused aged locks;
1569 * - do not cancel more than \a max locks;
1570 * - GET the found locks and add them into the \a cancels list.
1572 * A client lock can be added to the l_bl_ast list only when it is
1573 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1574 * CANCEL. There are the following use cases:
1575 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1576 * ldlm_cli_cancel(), which check and set this flag properly. As any
1577 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1578 * later without any special locking.
1580 * Calling policies for enabled LRU resize:
1581 * ----------------------------------------
1582 * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to
1583 * cancel not more than \a count locks;
1585 * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at
1586 * the beginning of LRU list);
1588 * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to
1589 * memory pressre policy function;
1591 * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy".
1593 * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1594 * (typically before replaying locks) w/o
1595 * sending any RPCs or waiting for any
1596 * outstanding RPC to complete.
1598 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels,
1599 int count, int max, int flags)
1601 ldlm_cancel_lru_policy_t pf;
1602 struct ldlm_lock *lock, *next;
1603 int added = 0, unused, remained;
1606 spin_lock(&ns->ns_lock);
1607 unused = ns->ns_nr_unused;
1610 if (!ns_connect_lru_resize(ns))
1611 count += unused - ns->ns_max_unused;
1613 pf = ldlm_cancel_lru_policy(ns, flags);
1614 LASSERT(pf != NULL);
1616 while (!cfs_list_empty(&ns->ns_unused_list)) {
1617 ldlm_policy_res_t result;
1619 /* all unused locks */
1620 if (remained-- <= 0)
1623 /* For any flags, stop scanning if @max is reached. */
1624 if (max && added >= max)
1627 cfs_list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1629 /* No locks which got blocking requests. */
1630 LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1632 if (flags & LDLM_CANCEL_NO_WAIT &&
1633 lock->l_flags & LDLM_FL_SKIPPED)
1634 /* already processed */
1637 /* Somebody is already doing CANCEL. No need for this
1638 * lock in LRU, do not traverse it again. */
1639 if (!(lock->l_flags & LDLM_FL_CANCELING))
1642 ldlm_lock_remove_from_lru_nolock(lock);
1644 if (&lock->l_lru == &ns->ns_unused_list)
1647 LDLM_LOCK_GET(lock);
1648 spin_unlock(&ns->ns_lock);
1649 lu_ref_add(&lock->l_reference, __FUNCTION__, cfs_current());
1651 /* Pass the lock through the policy filter and see if it
1652 * should stay in LRU.
1654 * Even for shrinker policy we stop scanning if
1655 * we find a lock that should stay in the cache.
1656 * We should take into account lock age anyway
1657 * as a new lock is a valuable resource even if
1658 * it has a low weight.
1660 * That is, for shrinker policy we drop only
1661 * old locks, but additionally choose them by
1662 * their weight. Big extent locks will stay in
1664 result = pf(ns, lock, unused, added, count);
1665 if (result == LDLM_POLICY_KEEP_LOCK) {
1666 lu_ref_del(&lock->l_reference,
1667 __FUNCTION__, cfs_current());
1668 LDLM_LOCK_RELEASE(lock);
1669 spin_lock(&ns->ns_lock);
1672 if (result == LDLM_POLICY_SKIP_LOCK) {
1673 lu_ref_del(&lock->l_reference,
1674 __func__, cfs_current());
1675 LDLM_LOCK_RELEASE(lock);
1676 spin_lock(&ns->ns_lock);
1680 lock_res_and_lock(lock);
1681 /* Check flags again under the lock. */
1682 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1683 (ldlm_lock_remove_from_lru(lock) == 0)) {
1684 /* Another thread is removing lock from LRU, or
1685 * somebody is already doing CANCEL, or there
1686 * is a blocking request which will send cancel
1687 * by itself, or the lock is no longer unused. */
1688 unlock_res_and_lock(lock);
1689 lu_ref_del(&lock->l_reference,
1690 __FUNCTION__, cfs_current());
1691 LDLM_LOCK_RELEASE(lock);
1692 spin_lock(&ns->ns_lock);
1695 LASSERT(!lock->l_readers && !lock->l_writers);
1697 /* If we have chosen to cancel this lock voluntarily, we
1698 * better send cancel notification to server, so that it
1699 * frees appropriate state. This might lead to a race
1700 * where while we are doing cancel here, server is also
1701 * silently cancelling this lock. */
1702 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1704 /* Setting the CBPENDING flag is a little misleading,
1705 * but prevents an important race; namely, once
1706 * CBPENDING is set, the lock can accumulate no more
1707 * readers/writers. Since readers and writers are
1708 * already zero here, ldlm_lock_decref() won't see
1709 * this flag and call l_blocking_ast */
1710 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1712 /* We can't re-add to l_lru as it confuses the
1713 * refcounting in ldlm_lock_remove_from_lru() if an AST
1714 * arrives after we drop lr_lock below. We use l_bl_ast
1715 * and can't use l_pending_chain as it is used both on
1716 * server and client nevertheless bug 5666 says it is
1717 * used only on server */
1718 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1719 cfs_list_add(&lock->l_bl_ast, cancels);
1720 unlock_res_and_lock(lock);
1721 lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current());
1722 spin_lock(&ns->ns_lock);
1726 spin_unlock(&ns->ns_lock);
1730 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, cfs_list_t *cancels,
1731 int count, int max, ldlm_cancel_flags_t cancel_flags,
1735 added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1738 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1742 * Cancel at least \a nr locks from given namespace LRU.
1744 * When called with LDLM_ASYNC the blocking callback will be handled
1745 * in a thread and this function will return after the thread has been
1746 * asked to call the callback. When called with LDLM_SYNC the blocking
1747 * callback will be performed in this function.
1749 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t mode,
1752 CFS_LIST_HEAD(cancels);
1757 mode = LDLM_SYNC; /* force to be sync in user space */
1759 /* Just prepare the list of locks, do not actually cancel them yet.
1760 * Locks are cancelled later in a separate thread. */
1761 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1762 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, mode);
1770 * Find and cancel locally unused locks found on resource, matched to the
1771 * given policy, mode. GET the found locks and add them into the \a cancels
1774 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1775 cfs_list_t *cancels,
1776 ldlm_policy_data_t *policy,
1777 ldlm_mode_t mode, int lock_flags,
1778 ldlm_cancel_flags_t cancel_flags, void *opaque)
1780 struct ldlm_lock *lock;
1785 cfs_list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1786 if (opaque != NULL && lock->l_ast_data != opaque) {
1787 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1788 lock->l_ast_data, opaque);
1793 if (lock->l_readers || lock->l_writers)
1796 /* If somebody is already doing CANCEL, or blocking AST came,
1797 * skip this lock. */
1798 if (lock->l_flags & LDLM_FL_BL_AST ||
1799 lock->l_flags & LDLM_FL_CANCELING)
1802 if (lockmode_compat(lock->l_granted_mode, mode))
1805 /* If policy is given and this is IBITS lock, add to list only
1806 * those locks that match by policy. */
1807 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1808 !(lock->l_policy_data.l_inodebits.bits &
1809 policy->l_inodebits.bits))
1812 /* See CBPENDING comment in ldlm_cancel_lru */
1813 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1816 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1817 cfs_list_add(&lock->l_bl_ast, cancels);
1818 LDLM_LOCK_GET(lock);
1823 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
1825 EXPORT_SYMBOL(ldlm_cancel_resource_local);
1828 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1830 * If \a req is NULL, send CANCEL request to server with handles of locks
1831 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1832 * separately per lock.
1833 * If \a req is not NULL, put handles of locks in \a cancels into the request
1834 * buffer at the offset \a off.
1835 * Destroy \a cancels at the end.
1837 int ldlm_cli_cancel_list(cfs_list_t *cancels, int count,
1838 struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1840 struct ldlm_lock *lock;
1844 if (cfs_list_empty(cancels) || count == 0)
1847 /* XXX: requests (both batched and not) could be sent in parallel.
1848 * Usually it is enough to have just 1 RPC, but it is possible that
1849 * there are too many locks to be cancelled in LRU or on a resource.
1850 * It would also speed up the case when the server does not support
1853 LASSERT(!cfs_list_empty(cancels));
1854 lock = cfs_list_entry(cancels->next, struct ldlm_lock,
1856 LASSERT(lock->l_conn_export);
1858 if (exp_connect_cancelset(lock->l_conn_export)) {
1861 ldlm_cancel_pack(req, cancels, count);
1863 res = ldlm_cli_cancel_req(lock->l_conn_export,
1867 res = ldlm_cli_cancel_req(lock->l_conn_export,
1872 if (res != -ESHUTDOWN)
1873 CERROR("ldlm_cli_cancel_list: %d\n", res);
1878 ldlm_lock_list_put(cancels, l_bl_ast, res);
1880 LASSERT(count == 0);
1883 EXPORT_SYMBOL(ldlm_cli_cancel_list);
1886 * Cancel all locks on a resource that have 0 readers/writers.
1888 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1889 * to notify the server. */
1890 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1891 const struct ldlm_res_id *res_id,
1892 ldlm_policy_data_t *policy,
1894 ldlm_cancel_flags_t flags,
1897 struct ldlm_resource *res;
1898 CFS_LIST_HEAD(cancels);
1903 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1905 /* This is not a problem. */
1906 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1910 LDLM_RESOURCE_ADDREF(res);
1911 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1912 0, flags | LCF_BL_AST, opaque);
1913 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1915 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1917 LDLM_RESOURCE_DELREF(res);
1918 ldlm_resource_putref(res);
1921 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
1923 struct ldlm_cli_cancel_arg {
1928 static int ldlm_cli_hash_cancel_unused(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1929 cfs_hlist_node_t *hnode, void *arg)
1931 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1932 struct ldlm_cli_cancel_arg *lc = arg;
1935 rc = ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1937 lc->lc_flags, lc->lc_opaque);
1939 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1940 res->lr_name.name[0], rc);
1942 /* must return 0 for hash iteration */
1947 * Cancel all locks on a namespace (or a specific resource, if given)
1948 * that have 0 readers/writers.
1950 * If flags & LCF_LOCAL, throw the locks away without trying
1951 * to notify the server. */
1952 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1953 const struct ldlm_res_id *res_id,
1954 ldlm_cancel_flags_t flags, void *opaque)
1956 struct ldlm_cli_cancel_arg arg = {
1958 .lc_opaque = opaque,
1966 if (res_id != NULL) {
1967 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1971 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1972 ldlm_cli_hash_cancel_unused, &arg);
1976 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1978 /* Lock iterators. */
1980 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1983 cfs_list_t *tmp, *next;
1984 struct ldlm_lock *lock;
1985 int rc = LDLM_ITER_CONTINUE;
1990 RETURN(LDLM_ITER_CONTINUE);
1993 cfs_list_for_each_safe(tmp, next, &res->lr_granted) {
1994 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1996 if (iter(lock, closure) == LDLM_ITER_STOP)
1997 GOTO(out, rc = LDLM_ITER_STOP);
2000 cfs_list_for_each_safe(tmp, next, &res->lr_converting) {
2001 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
2003 if (iter(lock, closure) == LDLM_ITER_STOP)
2004 GOTO(out, rc = LDLM_ITER_STOP);
2007 cfs_list_for_each_safe(tmp, next, &res->lr_waiting) {
2008 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
2010 if (iter(lock, closure) == LDLM_ITER_STOP)
2011 GOTO(out, rc = LDLM_ITER_STOP);
2017 EXPORT_SYMBOL(ldlm_resource_foreach);
2019 struct iter_helper_data {
2020 ldlm_iterator_t iter;
2024 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2026 struct iter_helper_data *helper = closure;
2027 return helper->iter(lock, helper->closure);
2030 static int ldlm_res_iter_helper(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2031 cfs_hlist_node_t *hnode, void *arg)
2034 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2036 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2040 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2041 ldlm_iterator_t iter, void *closure)
2044 struct iter_helper_data helper = { iter: iter, closure: closure };
2046 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2047 ldlm_res_iter_helper, &helper);
2050 EXPORT_SYMBOL(ldlm_namespace_foreach);
2052 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2053 * return 0: find no resource
2054 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2057 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2058 const struct ldlm_res_id *res_id,
2059 ldlm_iterator_t iter, void *data)
2061 struct ldlm_resource *res;
2066 CERROR("must pass in namespace\n");
2070 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2074 LDLM_RESOURCE_ADDREF(res);
2075 rc = ldlm_resource_foreach(res, iter, data);
2076 LDLM_RESOURCE_DELREF(res);
2077 ldlm_resource_putref(res);
2080 EXPORT_SYMBOL(ldlm_resource_iterate);
2084 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2086 cfs_list_t *list = closure;
2088 /* we use l_pending_chain here, because it's unused on clients. */
2089 LASSERTF(cfs_list_empty(&lock->l_pending_chain),
2090 "lock %p next %p prev %p\n",
2091 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2092 /* bug 9573: don't replay locks left after eviction, or
2093 * bug 17614: locks being actively cancelled. Get a reference
2094 * on a lock so that it does not disapear under us (e.g. due to cancel)
2096 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
2097 cfs_list_add(&lock->l_pending_chain, list);
2098 LDLM_LOCK_GET(lock);
2101 return LDLM_ITER_CONTINUE;
2104 static int replay_lock_interpret(const struct lu_env *env,
2105 struct ptlrpc_request *req,
2106 struct ldlm_async_args *aa, int rc)
2108 struct ldlm_lock *lock;
2109 struct ldlm_reply *reply;
2110 struct obd_export *exp;
2113 cfs_atomic_dec(&req->rq_import->imp_replay_inflight);
2118 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2120 GOTO(out, rc = -EPROTO);
2122 lock = ldlm_handle2lock(&aa->lock_handle);
2124 CERROR("received replay ack for unknown local cookie "LPX64
2125 " remote cookie "LPX64 " from server %s id %s\n",
2126 aa->lock_handle.cookie, reply->lock_handle.cookie,
2127 req->rq_export->exp_client_uuid.uuid,
2128 libcfs_id2str(req->rq_peer));
2129 GOTO(out, rc = -ESTALE);
2132 /* Key change rehash lock in per-export hash with new key */
2133 exp = req->rq_export;
2134 if (exp && exp->exp_lock_hash) {
2135 /* In the function below, .hs_keycmp resolves to
2136 * ldlm_export_lock_keycmp() */
2137 /* coverity[overrun-buffer-val] */
2138 cfs_hash_rehash_key(exp->exp_lock_hash,
2139 &lock->l_remote_handle,
2140 &reply->lock_handle,
2143 lock->l_remote_handle = reply->lock_handle;
2146 LDLM_DEBUG(lock, "replayed lock:");
2147 ptlrpc_import_recovery_state_machine(req->rq_import);
2148 LDLM_LOCK_PUT(lock);
2151 ptlrpc_connect_import(req->rq_import);
2156 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2158 struct ptlrpc_request *req;
2159 struct ldlm_async_args *aa;
2160 struct ldlm_request *body;
2165 /* Bug 11974: Do not replay a lock which is actively being canceled */
2166 if (lock->l_flags & LDLM_FL_CANCELING) {
2167 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2171 /* If this is reply-less callback lock, we cannot replay it, since
2172 * server might have long dropped it, but notification of that event was
2173 * lost by network. (and server granted conflicting lock already) */
2174 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2175 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2176 ldlm_lock_cancel(lock);
2181 * If granted mode matches the requested mode, this lock is granted.
2183 * If they differ, but we have a granted mode, then we were granted
2184 * one mode and now want another: ergo, converting.
2186 * If we haven't been granted anything and are on a resource list,
2187 * then we're blocked/waiting.
2189 * If we haven't been granted anything and we're NOT on a resource list,
2190 * then we haven't got a reply yet and don't have a known disposition.
2191 * This happens whenever a lock enqueue is the request that triggers
2194 if (lock->l_granted_mode == lock->l_req_mode)
2195 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2196 else if (lock->l_granted_mode)
2197 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2198 else if (!cfs_list_empty(&lock->l_res_link))
2199 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2201 flags = LDLM_FL_REPLAY;
2203 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2204 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2208 /* We're part of recovery, so don't wait for it. */
2209 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2211 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2212 ldlm_lock2desc(lock, &body->lock_desc);
2213 body->lock_flags = ldlm_flags_to_wire(flags);
2215 ldlm_lock2handle(lock, &body->lock_handle[0]);
2216 if (lock->l_lvb_len > 0)
2217 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2218 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2220 ptlrpc_request_set_replen(req);
2221 /* notify the server we've replayed all requests.
2222 * also, we mark the request to be put on a dedicated
2223 * queue to be processed after all request replayes.
2225 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2227 LDLM_DEBUG(lock, "replaying lock:");
2229 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2230 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2231 aa = ptlrpc_req_async_args(req);
2232 aa->lock_handle = body->lock_handle[0];
2233 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2234 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2240 * Cancel as many unused locks as possible before replay. since we are
2241 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2244 * Called only in recovery before replaying locks. there is no need to
2245 * replay locks that are unused. since the clients may hold thousands of
2246 * cached unused locks, dropping the unused locks can greatly reduce the
2247 * load on the servers at recovery time.
2249 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2252 CFS_LIST_HEAD(cancels);
2254 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2255 "replay for namespace %s (%d)\n",
2256 ldlm_ns_name(ns), ns->ns_nr_unused);
2258 /* We don't need to care whether or not LRU resize is enabled
2259 * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2260 * count parameter */
2261 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2262 LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2264 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2265 canceled, ldlm_ns_name(ns));
2268 int ldlm_replay_locks(struct obd_import *imp)
2270 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2271 CFS_LIST_HEAD(list);
2272 struct ldlm_lock *lock, *next;
2277 LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0);
2279 /* don't replay locks if import failed recovery */
2280 if (imp->imp_vbr_failed)
2283 /* ensure this doesn't fall to 0 before all have been queued */
2284 cfs_atomic_inc(&imp->imp_replay_inflight);
2286 if (ldlm_cancel_unused_locks_before_replay)
2287 ldlm_cancel_unused_locks_for_replay(ns);
2289 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2291 cfs_list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2292 cfs_list_del_init(&lock->l_pending_chain);
2294 LDLM_LOCK_RELEASE(lock);
2295 continue; /* or try to do the rest? */
2297 rc = replay_one_lock(imp, lock);
2298 LDLM_LOCK_RELEASE(lock);
2301 cfs_atomic_dec(&imp->imp_replay_inflight);
2305 EXPORT_SYMBOL(ldlm_replay_locks);