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, 2013, 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
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 LCONSOLE_WARN("lock timed out (enqueued at "CFS_TIME_T", "
106 CFS_DURATION_T"s ago)\n",
107 lock->l_last_activity,
108 cfs_time_sub(cfs_time_current_sec(),
109 lock->l_last_activity));
110 LDLM_DEBUG(lock, "lock timed out (enqueued at "CFS_TIME_T", "
111 CFS_DURATION_T"s ago); not entering recovery in "
112 "server code, just going back to sleep",
113 lock->l_last_activity,
114 cfs_time_sub(cfs_time_current_sec(),
115 lock->l_last_activity));
116 if (cfs_time_after(cfs_time_current(), next_dump)) {
117 last_dump = next_dump;
118 next_dump = cfs_time_shift(300);
119 ldlm_namespace_dump(D_DLMTRACE,
120 ldlm_lock_to_ns(lock));
122 libcfs_debug_dumplog();
127 obd = lock->l_conn_export->exp_obd;
128 imp = obd->u.cli.cl_import;
129 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
130 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
131 CFS_DURATION_T"s ago), entering recovery for %s@%s",
132 lock->l_last_activity,
133 cfs_time_sub(cfs_time_current_sec(), lock->l_last_activity),
134 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
138 EXPORT_SYMBOL(ldlm_expired_completion_wait);
140 /* We use the same basis for both server side and client side functions
141 from a single node. */
142 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
144 int timeout = at_get(ldlm_lock_to_ns_at(lock));
146 return obd_timeout / 2;
147 /* Since these are non-updating timeouts, we should be conservative.
148 It would be nice to have some kind of "early reply" mechanism for
149 lock callbacks too... */
150 timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */
151 return max(timeout, ldlm_enqueue_min);
153 EXPORT_SYMBOL(ldlm_get_enq_timeout);
156 * Helper function for ldlm_completion_ast(), updating timings when lock is
159 static int ldlm_completion_tail(struct ldlm_lock *lock)
164 if (ldlm_is_destroyed(lock) || ldlm_is_failed(lock)) {
165 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
168 delay = cfs_time_sub(cfs_time_current_sec(),
169 lock->l_last_activity);
170 LDLM_DEBUG(lock, "client-side enqueue: granted after "
171 CFS_DURATION_T"s", delay);
173 /* Update our time estimate */
174 at_measured(ldlm_lock_to_ns_at(lock),
182 * Implementation of ->l_completion_ast() for a client, that doesn't wait
183 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
184 * other threads that cannot block for long.
186 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
190 if (flags == LDLM_FL_WAIT_NOREPROC) {
191 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
195 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
196 LDLM_FL_BLOCK_CONV))) {
197 wake_up(&lock->l_waitq);
198 RETURN(ldlm_completion_tail(lock));
201 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
203 ldlm_reprocess_all(lock->l_resource);
206 EXPORT_SYMBOL(ldlm_completion_ast_async);
209 * Generic LDLM "completion" AST. This is called in several cases:
211 * - when a reply to an ENQUEUE RPC is received from the server
212 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
213 * this point (determined by flags);
215 * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
218 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
221 * - to force all locks when resource is destroyed (cleanup_resource());
223 * - during lock conversion (not used currently).
225 * If lock is not granted in the first case, this function waits until second
226 * or penultimate cases happen in some other thread.
229 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
231 /* XXX ALLOCATE - 160 bytes */
232 struct lock_wait_data lwd;
233 struct obd_device *obd;
234 struct obd_import *imp = NULL;
235 struct l_wait_info lwi;
240 if (flags == LDLM_FL_WAIT_NOREPROC) {
241 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
245 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
246 LDLM_FL_BLOCK_CONV))) {
247 wake_up(&lock->l_waitq);
251 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
256 obd = class_exp2obd(lock->l_conn_export);
258 /* if this is a local lock, then there is no import */
260 imp = obd->u.cli.cl_import;
263 /* Wait a long time for enqueue - server may have to callback a
264 lock from another client. Server will evict the other client if it
265 doesn't respond reasonably, and then give us the lock. */
266 timeout = ldlm_get_enq_timeout(lock) * 2;
270 if (ldlm_is_no_timeout(lock)) {
271 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
272 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
274 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
275 ldlm_expired_completion_wait,
276 interrupted_completion_wait, &lwd);
280 spin_lock(&imp->imp_lock);
281 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
282 spin_unlock(&imp->imp_lock);
285 if (ns_is_client(ldlm_lock_to_ns(lock)) &&
286 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
287 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
288 ldlm_set_fail_loc(lock);
291 /* Go to sleep until the lock is granted or cancelled. */
292 rc = l_wait_event(lock->l_waitq,
293 is_granted_or_cancelled(lock), &lwi);
297 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
302 RETURN(ldlm_completion_tail(lock));
304 EXPORT_SYMBOL(ldlm_completion_ast);
307 * A helper to build a blocking AST function
309 * Perform a common operation for blocking ASTs:
310 * defferred lock cancellation.
312 * \param lock the lock blocking or canceling AST was called on
314 * \see mdt_blocking_ast
315 * \see ldlm_blocking_ast
317 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
322 ldlm_set_cbpending(lock);
323 do_ast = (!lock->l_readers && !lock->l_writers);
324 unlock_res_and_lock(lock);
327 struct lustre_handle lockh;
330 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
331 ldlm_lock2handle(lock, &lockh);
332 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
334 CERROR("ldlm_cli_cancel: %d\n", rc);
336 LDLM_DEBUG(lock, "Lock still has references, will be "
341 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
344 * Server blocking AST
346 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
349 * \param lock the lock which blocks a request or cancelling lock
352 * \param flag indicates whether this cancelling or blocking callback
354 * \see ldlm_blocking_ast_nocheck
356 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
357 void *data, int flag)
361 if (flag == LDLM_CB_CANCELING) {
362 /* Don't need to do anything here. */
366 lock_res_and_lock(lock);
367 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
368 * that ldlm_blocking_ast is called just before intent_policy method
369 * takes the lr_lock, then by the time we get the lock, we might not
370 * be the correct blocking function anymore. So check, and return
372 if (lock->l_blocking_ast != ldlm_blocking_ast) {
373 unlock_res_and_lock(lock);
376 RETURN(ldlm_blocking_ast_nocheck(lock));
378 EXPORT_SYMBOL(ldlm_blocking_ast);
381 * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
382 * comment in filter_intent_policy() on why you may need this.
384 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
387 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
388 * that is rather subtle: with OST-side locking, it may so happen that
389 * _all_ extent locks are held by the OST. If client wants to obtain
390 * current file size it calls ll{,u}_glimpse_size(), and (as locks are
391 * on the server), dummy glimpse callback fires and does
392 * nothing. Client still receives correct file size due to the
393 * following fragment in filter_intent_policy():
395 * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
396 * if (rc != 0 && res->lr_namespace->ns_lvbo &&
397 * res->lr_namespace->ns_lvbo->lvbo_update) {
398 * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
401 * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
402 * returns correct file size to the client.
404 return -ELDLM_NO_LOCK_DATA;
406 EXPORT_SYMBOL(ldlm_glimpse_ast);
409 * Enqueue a local lock (typically on a server).
411 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
412 const struct ldlm_res_id *res_id,
413 ldlm_type_t type, ldlm_policy_data_t *policy,
414 ldlm_mode_t mode, __u64 *flags,
415 ldlm_blocking_callback blocking,
416 ldlm_completion_callback completion,
417 ldlm_glimpse_callback glimpse,
418 void *data, __u32 lvb_len, enum lvb_type lvb_type,
419 const __u64 *client_cookie,
420 struct lustre_handle *lockh)
422 struct ldlm_lock *lock;
424 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
425 .lcs_blocking = blocking,
426 .lcs_glimpse = glimpse,
430 LASSERT(!(*flags & LDLM_FL_REPLAY));
431 if (unlikely(ns_is_client(ns))) {
432 CERROR("Trying to enqueue local lock in a shadow namespace\n");
436 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
439 GOTO(out_nolock, err = PTR_ERR(lock));
441 ldlm_lock2handle(lock, lockh);
443 /* NB: we don't have any lock now (lock_res_and_lock)
444 * because it's a new lock */
445 ldlm_lock_addref_internal_nolock(lock, mode);
446 ldlm_set_local(lock);
447 if (*flags & LDLM_FL_ATOMIC_CB)
448 ldlm_set_atomic_cb(lock);
451 lock->l_policy_data = *policy;
452 if (client_cookie != NULL)
453 lock->l_client_cookie = *client_cookie;
454 if (type == LDLM_EXTENT) {
455 /* extent lock without policy is a bug */
459 lock->l_req_extent = policy->l_extent;
462 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
463 if (unlikely(err != ELDLM_OK))
467 *policy = lock->l_policy_data;
469 if (lock->l_completion_ast)
470 lock->l_completion_ast(lock, *flags, NULL);
472 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
475 LDLM_LOCK_RELEASE(lock);
479 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
481 static void failed_lock_cleanup(struct ldlm_namespace *ns,
482 struct ldlm_lock *lock, int mode)
486 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
487 lock_res_and_lock(lock);
488 /* Check that lock is not granted or failed, we might race. */
489 if ((lock->l_req_mode != lock->l_granted_mode) &&
490 !ldlm_is_failed(lock)) {
491 /* Make sure that this lock will not be found by raced
492 * bl_ast and -EINVAL reply is sent to server anyways.
494 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
495 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
498 unlock_res_and_lock(lock);
502 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
503 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
505 LDLM_DEBUG(lock, "lock was granted or failed in race");
507 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
508 * from llite/file.c/ll_file_flock(). */
509 /* This code makes for the fact that we do not have blocking handler on
510 * a client for flock locks. As such this is the place where we must
511 * completely kill failed locks. (interrupted and those that
512 * were waiting to be granted when server evicted us. */
513 if (lock->l_resource->lr_type == LDLM_FLOCK) {
514 lock_res_and_lock(lock);
515 if (!ldlm_is_destroyed(lock)) {
516 ldlm_resource_unlink_lock(lock);
517 ldlm_lock_decref_internal_nolock(lock, mode);
518 ldlm_lock_destroy_nolock(lock);
520 unlock_res_and_lock(lock);
522 ldlm_lock_decref_internal(lock, mode);
527 * Finishing portion of client lock enqueue code.
529 * Called after receiving reply from server.
531 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
532 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
533 __u64 *flags, void *lvb, __u32 lvb_len,
534 struct lustre_handle *lockh,int rc)
536 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
537 int is_replay = *flags & LDLM_FL_REPLAY;
538 struct ldlm_lock *lock;
539 struct ldlm_reply *reply;
540 int cleanup_phase = 1;
544 lock = ldlm_handle2lock(lockh);
545 /* ldlm_cli_enqueue is holding a reference on this lock. */
547 LASSERT(type == LDLM_FLOCK);
551 LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
552 "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
554 if (rc != ELDLM_OK) {
556 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
557 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
559 if (rc != ELDLM_LOCK_ABORTED)
563 /* Before we return, swab the reply */
564 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
566 GOTO(cleanup, rc = -EPROTO);
569 LASSERT(lvb != NULL);
571 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
574 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
575 GOTO(cleanup, rc = size);
576 } else if (unlikely(size > lvb_len)) {
577 LDLM_ERROR(lock, "Replied LVB is larger than "
578 "expectation, expected = %d, replied = %d",
580 GOTO(cleanup, rc = -EINVAL);
584 if (rc == ELDLM_LOCK_ABORTED) {
586 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
588 GOTO(cleanup, rc = (rc != 0 ? rc : ELDLM_LOCK_ABORTED));
591 /* lock enqueued on the server */
594 lock_res_and_lock(lock);
595 /* Key change rehash lock in per-export hash with new key */
596 if (exp->exp_lock_hash) {
597 /* In the function below, .hs_keycmp resolves to
598 * ldlm_export_lock_keycmp() */
599 /* coverity[overrun-buffer-val] */
600 cfs_hash_rehash_key(exp->exp_lock_hash,
601 &lock->l_remote_handle,
605 lock->l_remote_handle = reply->lock_handle;
608 *flags = ldlm_flags_from_wire(reply->lock_flags);
609 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
610 LDLM_FL_INHERIT_MASK);
611 unlock_res_and_lock(lock);
613 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: "LPX64"\n",
614 lock, reply->lock_handle.cookie, *flags);
616 /* If enqueue returned a blocked lock but the completion handler has
617 * already run, then it fixed up the resource and we don't need to do it
619 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
620 int newmode = reply->lock_desc.l_req_mode;
622 if (newmode && newmode != lock->l_req_mode) {
623 LDLM_DEBUG(lock, "server returned different mode %s",
624 ldlm_lockname[newmode]);
625 lock->l_req_mode = newmode;
628 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
629 &lock->l_resource->lr_name)) {
630 CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES
631 " instead of "DLDLMRES"\n",
632 PLDLMRES(&reply->lock_desc.l_resource),
633 PLDLMRES(lock->l_resource));
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 &&
643 !(exp_connect_flags(exp) & OBD_CONNECT_IBITS)))
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);
683 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
684 if (lock->l_completion_ast != NULL) {
685 int err = lock->l_completion_ast(lock, *flags, NULL);
693 if (lvb_len && lvb != NULL) {
694 /* Copy the LVB here, and not earlier, because the completion
695 * AST (if any) can override what we got in the reply */
696 memcpy(lvb, lock->l_lvb_data, lvb_len);
699 LDLM_DEBUG(lock, "client-side enqueue END");
702 if (cleanup_phase == 1 && rc)
703 failed_lock_cleanup(ns, lock, mode);
704 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
706 LDLM_LOCK_RELEASE(lock);
709 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
712 * Estimate number of lock handles that would fit into request of given
713 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
714 * a single page on the send/receive side. XXX: 512 should be changed to
715 * more adequate value.
717 static inline int ldlm_req_handles_avail(int req_size, int off)
721 avail = min_t(int, LDLM_MAXREQSIZE, PAGE_CACHE_SIZE - 512) - req_size;
722 if (likely(avail >= 0))
723 avail /= (int)sizeof(struct lustre_handle);
726 avail += LDLM_LOCKREQ_HANDLES - off;
731 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
732 enum req_location loc,
735 int size = req_capsule_msg_size(pill, loc);
736 return ldlm_req_handles_avail(size, off);
739 static inline int ldlm_format_handles_avail(struct obd_import *imp,
740 const struct req_format *fmt,
741 enum req_location loc, int off)
743 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
744 return ldlm_req_handles_avail(size, off);
748 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
749 * \a count locks in \a cancels.
751 * This is to be called by functions preparing their own requests that
752 * might contain lists of locks to cancel in addition to actual operation
753 * that needs to be performed.
755 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
756 int version, int opc, int canceloff,
757 struct list_head *cancels, int count)
759 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
760 struct req_capsule *pill = &req->rq_pill;
761 struct ldlm_request *dlm = NULL;
762 struct list_head head = LIST_HEAD_INIT(head);
763 int flags, avail, to_free, pack = 0;
769 if (ns_connect_cancelset(ns)) {
770 /* Estimate the amount of available space in the request. */
771 req_capsule_filled_sizes(pill, RCL_CLIENT);
772 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
774 flags = ns_connect_lru_resize(ns) ?
775 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
776 to_free = !ns_connect_lru_resize(ns) &&
777 opc == LDLM_ENQUEUE ? 1 : 0;
779 /* Cancel LRU locks here _only_ if the server supports
780 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
781 * RPC, which will make us slower. */
783 count += ldlm_cancel_lru_local(ns, cancels, to_free,
784 avail - count, 0, flags);
789 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
790 ldlm_request_bufsize(pack, opc));
793 rc = ptlrpc_request_pack(req, version, opc);
795 ldlm_lock_list_put(cancels, l_bl_ast, count);
799 if (ns_connect_cancelset(ns)) {
801 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
803 /* Skip first lock handler in ldlm_request_pack(),
804 * this method will incrment @lock_count according
805 * to the lock handle amount actually written to
807 dlm->lock_count = canceloff;
809 /* Pack into the request @pack lock handles. */
810 ldlm_cli_cancel_list(cancels, pack, req, 0);
811 /* Prepare and send separate cancel RPC for others. */
812 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
814 ldlm_lock_list_put(cancels, l_bl_ast, count);
818 EXPORT_SYMBOL(ldlm_prep_elc_req);
820 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
821 cfs_list_t *cancels, int count)
823 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
824 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
826 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
828 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
830 struct ptlrpc_request *req;
834 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
836 RETURN(ERR_PTR(-ENOMEM));
838 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
840 ptlrpc_request_free(req);
844 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
845 ptlrpc_request_set_replen(req);
848 EXPORT_SYMBOL(ldlm_enqueue_pack);
851 * Client-side lock enqueue.
853 * If a request has some specific initialisation it is passed in \a reqp,
854 * otherwise it is created in ldlm_cli_enqueue.
856 * Supports sync and async requests, pass \a async flag accordingly. If a
857 * request was created in ldlm_cli_enqueue and it is the async request,
858 * pass it to the caller in \a reqp.
860 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
861 struct ldlm_enqueue_info *einfo,
862 const struct ldlm_res_id *res_id,
863 ldlm_policy_data_t const *policy, __u64 *flags,
864 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
865 struct lustre_handle *lockh, int async)
867 struct ldlm_namespace *ns;
868 struct ldlm_lock *lock;
869 struct ldlm_request *body;
870 int is_replay = *flags & LDLM_FL_REPLAY;
871 int req_passed_in = 1;
873 struct ptlrpc_request *req;
876 LASSERT(exp != NULL);
878 ns = exp->exp_obd->obd_namespace;
880 /* If we're replaying this lock, just check some invariants.
881 * If we're creating a new lock, get everything all setup nice. */
883 lock = ldlm_handle2lock_long(lockh, 0);
884 LASSERT(lock != NULL);
885 LDLM_DEBUG(lock, "client-side enqueue START");
886 LASSERT(exp == lock->l_conn_export);
888 const struct ldlm_callback_suite cbs = {
889 .lcs_completion = einfo->ei_cb_cp,
890 .lcs_blocking = einfo->ei_cb_bl,
891 .lcs_glimpse = einfo->ei_cb_gl
893 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
894 einfo->ei_mode, &cbs, einfo->ei_cbdata,
897 RETURN(PTR_ERR(lock));
898 /* for the local lock, add the reference */
899 ldlm_lock_addref_internal(lock, einfo->ei_mode);
900 ldlm_lock2handle(lock, lockh);
902 lock->l_policy_data = *policy;
904 if (einfo->ei_type == LDLM_EXTENT) {
905 /* extent lock without policy is a bug */
909 lock->l_req_extent = policy->l_extent;
911 LDLM_DEBUG(lock, "client-side enqueue START, flags "LPX64"\n",
915 lock->l_conn_export = exp;
916 lock->l_export = NULL;
917 lock->l_blocking_ast = einfo->ei_cb_bl;
918 lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
920 /* lock not sent to server yet */
922 if (reqp == NULL || *reqp == NULL) {
923 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
928 failed_lock_cleanup(ns, lock, einfo->ei_mode);
929 LDLM_LOCK_RELEASE(lock);
939 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
941 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
942 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
945 /* Dump lock data into the request buffer */
946 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
947 ldlm_lock2desc(lock, &body->lock_desc);
948 body->lock_flags = ldlm_flags_to_wire(*flags);
949 body->lock_handle[0] = *lockh;
951 /* Continue as normal. */
952 if (!req_passed_in) {
954 req_capsule_extend(&req->rq_pill,
955 &RQF_LDLM_ENQUEUE_LVB);
956 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
958 ptlrpc_request_set_replen(req);
962 * Liblustre client doesn't get extent locks, except for O_APPEND case
963 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
964 * [i_size, OBD_OBJECT_EOF] lock is taken.
966 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
967 policy->l_extent.end == OBD_OBJECT_EOF));
970 LASSERT(reqp != NULL);
974 LDLM_DEBUG(lock, "sending request");
976 rc = ptlrpc_queue_wait(req);
978 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
979 einfo->ei_mode, flags, lvb, lvb_len,
982 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
983 * one reference that we took */
985 LDLM_LOCK_RELEASE(lock);
989 if (!req_passed_in && req != NULL) {
990 ptlrpc_req_finished(req);
997 EXPORT_SYMBOL(ldlm_cli_enqueue);
999 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
1002 struct ldlm_resource *res;
1005 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1006 CERROR("Trying to cancel local lock\n");
1009 LDLM_DEBUG(lock, "client-side local convert");
1011 res = ldlm_lock_convert(lock, new_mode, flags);
1013 ldlm_reprocess_all(res);
1016 rc = LUSTRE_EDEADLK;
1018 LDLM_DEBUG(lock, "client-side local convert handler END");
1019 LDLM_LOCK_PUT(lock);
1023 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
1024 * conversion of locks which are on the waiting or converting queue */
1025 /* Caller of this code is supposed to take care of lock readers/writers
1027 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
1029 struct ldlm_request *body;
1030 struct ldlm_reply *reply;
1031 struct ldlm_lock *lock;
1032 struct ldlm_resource *res;
1033 struct ptlrpc_request *req;
1037 lock = ldlm_handle2lock(lockh);
1044 if (lock->l_conn_export == NULL)
1045 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
1047 LDLM_DEBUG(lock, "client-side convert");
1049 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
1050 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1053 LDLM_LOCK_PUT(lock);
1057 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1058 body->lock_handle[0] = lock->l_remote_handle;
1060 body->lock_desc.l_req_mode = new_mode;
1061 body->lock_flags = ldlm_flags_to_wire(*flags);
1064 ptlrpc_request_set_replen(req);
1065 rc = ptlrpc_queue_wait(req);
1069 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1071 GOTO(out, rc = -EPROTO);
1074 GOTO(out, rc = req->rq_status);
1076 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
1078 ldlm_reprocess_all(res);
1079 /* Go to sleep until the lock is granted. */
1080 /* FIXME: or cancelled. */
1081 if (lock->l_completion_ast) {
1082 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1088 rc = LUSTRE_EDEADLK;
1092 LDLM_LOCK_PUT(lock);
1093 ptlrpc_req_finished(req);
1096 EXPORT_SYMBOL(ldlm_cli_convert);
1099 * Cancel locks locally.
1101 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1102 * \retval LDLM_FL_CANCELING otherwise;
1103 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1105 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1107 __u64 rc = LDLM_FL_LOCAL_ONLY;
1110 if (lock->l_conn_export) {
1113 LDLM_DEBUG(lock, "client-side cancel");
1114 /* Set this flag to prevent others from getting new references*/
1115 lock_res_and_lock(lock);
1116 ldlm_set_cbpending(lock);
1117 local_only = !!(lock->l_flags &
1118 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1119 ldlm_cancel_callback(lock);
1120 rc = (ldlm_is_bl_ast(lock)) ?
1121 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1122 unlock_res_and_lock(lock);
1125 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1127 rc = LDLM_FL_LOCAL_ONLY;
1129 ldlm_lock_cancel(lock);
1131 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1132 LDLM_ERROR(lock, "Trying to cancel local lock");
1135 LDLM_DEBUG(lock, "server-side local cancel");
1136 ldlm_lock_cancel(lock);
1137 ldlm_reprocess_all(lock->l_resource);
1144 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1146 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1147 cfs_list_t *head, int count)
1149 struct ldlm_request *dlm;
1150 struct ldlm_lock *lock;
1151 int max, packed = 0;
1154 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1155 LASSERT(dlm != NULL);
1157 /* Check the room in the request buffer. */
1158 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1159 sizeof(struct ldlm_request);
1160 max /= sizeof(struct lustre_handle);
1161 max += LDLM_LOCKREQ_HANDLES;
1162 LASSERT(max >= dlm->lock_count + count);
1164 /* XXX: it would be better to pack lock handles grouped by resource.
1165 * so that the server cancel would call filter_lvbo_update() less
1167 cfs_list_for_each_entry(lock, head, l_bl_ast) {
1170 LASSERT(lock->l_conn_export);
1171 /* Pack the lock handle to the given request buffer. */
1172 LDLM_DEBUG(lock, "packing");
1173 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1176 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1181 * Prepare and send a batched cancel RPC. It will include \a count lock
1182 * handles of locks given in \a cancels list. */
1183 int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels,
1184 int count, ldlm_cancel_flags_t flags)
1186 struct ptlrpc_request *req = NULL;
1187 struct obd_import *imp;
1192 LASSERT(exp != NULL);
1195 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1197 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1200 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1201 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1206 imp = class_exp2cliimp(exp);
1207 if (imp == NULL || imp->imp_invalid) {
1209 "skipping cancel on invalid import %p\n", imp);
1213 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1215 GOTO(out, rc = -ENOMEM);
1217 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1218 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1219 ldlm_request_bufsize(count, LDLM_CANCEL));
1221 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1223 ptlrpc_request_free(req);
1227 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1228 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1229 ptlrpc_at_set_req_timeout(req);
1231 ldlm_cancel_pack(req, cancels, count);
1233 ptlrpc_request_set_replen(req);
1234 if (flags & LCF_ASYNC) {
1235 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1239 rc = ptlrpc_queue_wait(req);
1241 if (rc == LUSTRE_ESTALE) {
1242 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1243 "out of sync -- not fatal\n",
1244 libcfs_nid2str(req->rq_import->
1245 imp_connection->c_peer.nid));
1247 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1248 req->rq_import_generation == imp->imp_generation) {
1249 ptlrpc_req_finished(req);
1251 } else if (rc != ELDLM_OK) {
1252 /* -ESHUTDOWN is common on umount */
1253 CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1254 "Got rc %d from cancel RPC: "
1255 "canceling anyway\n", rc);
1262 ptlrpc_req_finished(req);
1265 return sent ? sent : rc;
1267 EXPORT_SYMBOL(ldlm_cli_cancel_req);
1269 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1271 LASSERT(imp != NULL);
1272 return &imp->imp_obd->obd_namespace->ns_pool;
1276 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1278 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1280 struct obd_device *obd;
1284 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1285 !imp_connect_lru_resize(req->rq_import)))
1288 * Do nothing for corner cases.
1293 /* In some cases RPC may contain SLV and limit zeroed out. This
1294 * is the case when server does not support LRU resize feature.
1295 * This is also possible in some recovery cases when server-side
1296 * reqs have no reference to the OBD export and thus access to
1297 * server-side namespace is not possible. */
1298 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1299 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1300 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1301 "(SLV: "LPU64", Limit: %u)",
1302 lustre_msg_get_slv(req->rq_repmsg),
1303 lustre_msg_get_limit(req->rq_repmsg));
1307 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1308 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1309 obd = req->rq_import->imp_obd;
1311 /* Set new SLV and limit in OBD fields to make them accessible
1312 * to the pool thread. We do not access obd_namespace and pool
1313 * directly here as there is no reliable way to make sure that
1314 * they are still alive at cleanup time. Evil races are possible
1315 * which may cause Oops at that time. */
1316 write_lock(&obd->obd_pool_lock);
1317 obd->obd_pool_slv = new_slv;
1318 obd->obd_pool_limit = new_limit;
1319 write_unlock(&obd->obd_pool_lock);
1323 EXPORT_SYMBOL(ldlm_cli_update_pool);
1326 * Client side lock cancel.
1328 * Lock must not have any readers or writers by this time.
1330 int ldlm_cli_cancel(struct lustre_handle *lockh,
1331 ldlm_cancel_flags_t cancel_flags)
1333 struct obd_export *exp;
1334 int avail, flags, count = 1;
1336 struct ldlm_namespace *ns;
1337 struct ldlm_lock *lock;
1338 struct list_head cancels = LIST_HEAD_INIT(cancels);
1341 /* concurrent cancels on the same handle can happen */
1342 lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1344 LDLM_DEBUG_NOLOCK("lock is already being destroyed");
1348 rc = ldlm_cli_cancel_local(lock);
1349 if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
1350 LDLM_LOCK_RELEASE(lock);
1353 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1354 * RPC which goes to canceld portal, so we can cancel other LRU locks
1355 * here and send them all as one LDLM_CANCEL RPC. */
1356 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1357 cfs_list_add(&lock->l_bl_ast, &cancels);
1359 exp = lock->l_conn_export;
1360 if (exp_connect_cancelset(exp)) {
1361 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1366 ns = ldlm_lock_to_ns(lock);
1367 flags = ns_connect_lru_resize(ns) ?
1368 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1369 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1372 ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1375 EXPORT_SYMBOL(ldlm_cli_cancel);
1378 * Locally cancel up to \a count locks in list \a cancels.
1379 * Return the number of cancelled locks.
1381 int ldlm_cli_cancel_list_local(cfs_list_t *cancels, int count,
1382 ldlm_cancel_flags_t flags)
1384 struct list_head head = LIST_HEAD_INIT(head);
1385 struct ldlm_lock *lock, *next;
1386 int left = 0, bl_ast = 0;
1390 cfs_list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1394 if (flags & LCF_LOCAL) {
1395 rc = LDLM_FL_LOCAL_ONLY;
1396 ldlm_lock_cancel(lock);
1398 rc = ldlm_cli_cancel_local(lock);
1400 /* Until we have compound requests and can send LDLM_CANCEL
1401 * requests batched with generic RPCs, we need to send cancels
1402 * with the LDLM_FL_BL_AST flag in a separate RPC from
1403 * the one being generated now. */
1404 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1405 LDLM_DEBUG(lock, "Cancel lock separately");
1406 cfs_list_del_init(&lock->l_bl_ast);
1407 cfs_list_add(&lock->l_bl_ast, &head);
1411 if (rc == LDLM_FL_LOCAL_ONLY) {
1412 /* CANCEL RPC should not be sent to server. */
1413 cfs_list_del_init(&lock->l_bl_ast);
1414 LDLM_LOCK_RELEASE(lock);
1420 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1425 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
1428 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1429 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1430 * readahead requests, ...)
1432 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1433 struct ldlm_lock *lock,
1434 int unused, int added,
1437 ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1439 /* don't check added & count since we want to process all locks
1441 * It's fine to not take lock to access lock->l_resource since
1442 * the lock has already been granted so it won't change. */
1443 switch (lock->l_resource->lr_type) {
1446 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) != 0)
1449 result = LDLM_POLICY_SKIP_LOCK;
1450 lock_res_and_lock(lock);
1451 ldlm_set_skipped(lock);
1452 unlock_res_and_lock(lock);
1460 * Callback function for LRU-resize policy. Decides whether to keep
1461 * \a lock in LRU for current \a LRU size \a unused, added in current
1462 * scan \a added and number of locks to be preferably canceled \a count.
1464 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1466 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1468 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1469 struct ldlm_lock *lock,
1470 int unused, int added,
1473 cfs_time_t cur = cfs_time_current();
1474 struct ldlm_pool *pl = &ns->ns_pool;
1478 /* Stop LRU processing when we reach past @count or have checked all
1480 if (count && added >= count)
1481 return LDLM_POLICY_KEEP_LOCK;
1483 slv = ldlm_pool_get_slv(pl);
1484 lvf = ldlm_pool_get_lvf(pl);
1485 la = cfs_duration_sec(cfs_time_sub(cur,
1486 lock->l_last_used));
1487 lv = lvf * la * unused;
1489 /* Inform pool about current CLV to see it via proc. */
1490 ldlm_pool_set_clv(pl, lv);
1492 /* Stop when SLV is not yet come from server or lv is smaller than
1494 if (slv == 0 || lv < slv)
1495 return LDLM_POLICY_KEEP_LOCK;
1497 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) == 0)
1498 return LDLM_POLICY_KEEP_LOCK;
1500 return LDLM_POLICY_CANCEL_LOCK;
1504 * Callback function for proc used policy. Makes decision whether to keep
1505 * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1506 * added and number of locks to be preferably canceled \a count.
1508 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1510 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1512 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1513 struct ldlm_lock *lock,
1514 int unused, int added,
1517 /* Stop LRU processing when we reach past @count or have checked all
1519 return (added >= count) ?
1520 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1524 * Callback function for aged policy. Makes decision whether to keep \a lock in
1525 * LRU for current LRU size \a unused, added in current scan \a added and
1526 * number of locks to be preferably canceled \a count.
1528 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1530 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1532 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1533 struct ldlm_lock *lock,
1534 int unused, int added,
1538 return LDLM_POLICY_KEEP_LOCK;
1540 if (cfs_time_before(cfs_time_current(),
1541 cfs_time_add(lock->l_last_used, ns->ns_max_age)))
1542 return LDLM_POLICY_KEEP_LOCK;
1544 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) == 0)
1545 return LDLM_POLICY_KEEP_LOCK;
1547 return LDLM_POLICY_CANCEL_LOCK;
1551 * Callback function for default policy. Makes decision whether to keep \a lock
1552 * in LRU for current LRU size \a unused, added in current scan \a added and
1553 * number of locks to be preferably canceled \a count.
1555 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1557 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1559 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1560 struct ldlm_lock *lock,
1561 int unused, int added,
1564 /* Stop LRU processing when we reach past count or have checked all
1566 return (added >= count) ?
1567 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1570 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1571 struct ldlm_lock *, int,
1574 static ldlm_cancel_lru_policy_t
1575 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1577 if (flags & LDLM_CANCEL_NO_WAIT)
1578 return ldlm_cancel_no_wait_policy;
1580 if (ns_connect_lru_resize(ns)) {
1581 if (flags & LDLM_CANCEL_SHRINK)
1582 /* We kill passed number of old locks. */
1583 return ldlm_cancel_passed_policy;
1584 else if (flags & LDLM_CANCEL_LRUR)
1585 return ldlm_cancel_lrur_policy;
1586 else if (flags & LDLM_CANCEL_PASSED)
1587 return ldlm_cancel_passed_policy;
1589 if (flags & LDLM_CANCEL_AGED)
1590 return ldlm_cancel_aged_policy;
1593 return ldlm_cancel_default_policy;
1597 * - Free space in LRU for \a count new locks,
1598 * redundant unused locks are canceled locally;
1599 * - also cancel locally unused aged locks;
1600 * - do not cancel more than \a max locks;
1601 * - GET the found locks and add them into the \a cancels list.
1603 * A client lock can be added to the l_bl_ast list only when it is
1604 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1605 * CANCEL. There are the following use cases:
1606 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1607 * ldlm_cli_cancel(), which check and set this flag properly. As any
1608 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1609 * later without any special locking.
1611 * Calling policies for enabled LRU resize:
1612 * ----------------------------------------
1613 * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to
1614 * cancel not more than \a count locks;
1616 * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at
1617 * the beginning of LRU list);
1619 * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to
1620 * memory pressre policy function;
1622 * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy".
1624 * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1625 * (typically before replaying locks) w/o
1626 * sending any RPCs or waiting for any
1627 * outstanding RPC to complete.
1629 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels,
1630 int count, int max, int flags)
1632 ldlm_cancel_lru_policy_t pf;
1633 struct ldlm_lock *lock, *next;
1634 int added = 0, unused, remained;
1637 spin_lock(&ns->ns_lock);
1638 unused = ns->ns_nr_unused;
1641 if (!ns_connect_lru_resize(ns))
1642 count += unused - ns->ns_max_unused;
1644 pf = ldlm_cancel_lru_policy(ns, flags);
1645 LASSERT(pf != NULL);
1647 while (!cfs_list_empty(&ns->ns_unused_list)) {
1648 ldlm_policy_res_t result;
1650 /* all unused locks */
1651 if (remained-- <= 0)
1654 /* For any flags, stop scanning if @max is reached. */
1655 if (max && added >= max)
1658 cfs_list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1660 /* No locks which got blocking requests. */
1661 LASSERT(!ldlm_is_bl_ast(lock));
1663 if (flags & LDLM_CANCEL_NO_WAIT &&
1664 ldlm_is_skipped(lock))
1665 /* already processed */
1668 /* Somebody is already doing CANCEL. No need for this
1669 * lock in LRU, do not traverse it again. */
1670 if (!ldlm_is_canceling(lock))
1673 ldlm_lock_remove_from_lru_nolock(lock);
1675 if (&lock->l_lru == &ns->ns_unused_list)
1678 LDLM_LOCK_GET(lock);
1679 spin_unlock(&ns->ns_lock);
1680 lu_ref_add(&lock->l_reference, __FUNCTION__, current);
1682 /* Pass the lock through the policy filter and see if it
1683 * should stay in LRU.
1685 * Even for shrinker policy we stop scanning if
1686 * we find a lock that should stay in the cache.
1687 * We should take into account lock age anyway
1688 * as a new lock is a valuable resource even if
1689 * it has a low weight.
1691 * That is, for shrinker policy we drop only
1692 * old locks, but additionally choose them by
1693 * their weight. Big extent locks will stay in
1695 result = pf(ns, lock, unused, added, count);
1696 if (result == LDLM_POLICY_KEEP_LOCK) {
1697 lu_ref_del(&lock->l_reference,
1698 __FUNCTION__, current);
1699 LDLM_LOCK_RELEASE(lock);
1700 spin_lock(&ns->ns_lock);
1703 if (result == LDLM_POLICY_SKIP_LOCK) {
1704 lu_ref_del(&lock->l_reference,
1706 LDLM_LOCK_RELEASE(lock);
1707 spin_lock(&ns->ns_lock);
1711 lock_res_and_lock(lock);
1712 /* Check flags again under the lock. */
1713 if (ldlm_is_canceling(lock) ||
1714 (ldlm_lock_remove_from_lru(lock) == 0)) {
1715 /* Another thread is removing lock from LRU, or
1716 * somebody is already doing CANCEL, or there
1717 * is a blocking request which will send cancel
1718 * by itself, or the lock is no longer unused. */
1719 unlock_res_and_lock(lock);
1720 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1721 LDLM_LOCK_RELEASE(lock);
1722 spin_lock(&ns->ns_lock);
1725 LASSERT(!lock->l_readers && !lock->l_writers);
1727 /* If we have chosen to cancel this lock voluntarily, we
1728 * better send cancel notification to server, so that it
1729 * frees appropriate state. This might lead to a race
1730 * where while we are doing cancel here, server is also
1731 * silently cancelling this lock. */
1732 ldlm_clear_cancel_on_block(lock);
1734 /* Setting the CBPENDING flag is a little misleading,
1735 * but prevents an important race; namely, once
1736 * CBPENDING is set, the lock can accumulate no more
1737 * readers/writers. Since readers and writers are
1738 * already zero here, ldlm_lock_decref() won't see
1739 * this flag and call l_blocking_ast */
1740 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1742 /* We can't re-add to l_lru as it confuses the
1743 * refcounting in ldlm_lock_remove_from_lru() if an AST
1744 * arrives after we drop lr_lock below. We use l_bl_ast
1745 * and can't use l_pending_chain as it is used both on
1746 * server and client nevertheless bug 5666 says it is
1747 * used only on server */
1748 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1749 cfs_list_add(&lock->l_bl_ast, cancels);
1750 unlock_res_and_lock(lock);
1751 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1752 spin_lock(&ns->ns_lock);
1756 spin_unlock(&ns->ns_lock);
1760 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, cfs_list_t *cancels,
1761 int count, int max, ldlm_cancel_flags_t cancel_flags,
1765 added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1768 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1772 * Cancel at least \a nr locks from given namespace LRU.
1774 * When called with LCF_ASYNC the blocking callback will be handled
1775 * in a thread and this function will return after the thread has been
1776 * asked to call the callback. When called with LCF_ASYNC the blocking
1777 * callback will be performed in this function.
1779 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
1780 ldlm_cancel_flags_t cancel_flags,
1783 struct list_head cancels = LIST_HEAD_INIT(cancels);
1788 cancel_flags &= ~LCF_ASYNC; /* force to be sync in user space */
1790 /* Just prepare the list of locks, do not actually cancel them yet.
1791 * Locks are cancelled later in a separate thread. */
1792 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1793 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
1801 * Find and cancel locally unused locks found on resource, matched to the
1802 * given policy, mode. GET the found locks and add them into the \a cancels
1805 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1806 cfs_list_t *cancels,
1807 ldlm_policy_data_t *policy,
1808 ldlm_mode_t mode, __u64 lock_flags,
1809 ldlm_cancel_flags_t cancel_flags, void *opaque)
1811 struct ldlm_lock *lock;
1816 cfs_list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1817 if (opaque != NULL && lock->l_ast_data != opaque) {
1818 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1819 lock->l_ast_data, opaque);
1824 if (lock->l_readers || lock->l_writers)
1827 /* If somebody is already doing CANCEL, or blocking AST came,
1828 * skip this lock. */
1829 if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock))
1832 if (lockmode_compat(lock->l_granted_mode, mode))
1835 /* If policy is given and this is IBITS lock, add to list only
1836 * those locks that match by policy. */
1837 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1838 !(lock->l_policy_data.l_inodebits.bits &
1839 policy->l_inodebits.bits))
1842 /* See CBPENDING comment in ldlm_cancel_lru */
1843 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1846 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1847 cfs_list_add(&lock->l_bl_ast, cancels);
1848 LDLM_LOCK_GET(lock);
1853 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
1855 EXPORT_SYMBOL(ldlm_cancel_resource_local);
1858 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1860 * If \a req is NULL, send CANCEL request to server with handles of locks
1861 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1862 * separately per lock.
1863 * If \a req is not NULL, put handles of locks in \a cancels into the request
1864 * buffer at the offset \a off.
1865 * Destroy \a cancels at the end.
1867 int ldlm_cli_cancel_list(cfs_list_t *cancels, int count,
1868 struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1870 struct ldlm_lock *lock;
1874 if (cfs_list_empty(cancels) || count == 0)
1877 /* XXX: requests (both batched and not) could be sent in parallel.
1878 * Usually it is enough to have just 1 RPC, but it is possible that
1879 * there are too many locks to be cancelled in LRU or on a resource.
1880 * It would also speed up the case when the server does not support
1883 LASSERT(!cfs_list_empty(cancels));
1884 lock = cfs_list_entry(cancels->next, struct ldlm_lock,
1886 LASSERT(lock->l_conn_export);
1888 if (exp_connect_cancelset(lock->l_conn_export)) {
1891 ldlm_cancel_pack(req, cancels, count);
1893 res = ldlm_cli_cancel_req(lock->l_conn_export,
1897 res = ldlm_cli_cancel_req(lock->l_conn_export,
1902 CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1903 "ldlm_cli_cancel_list: %d\n", res);
1908 ldlm_lock_list_put(cancels, l_bl_ast, res);
1910 LASSERT(count == 0);
1913 EXPORT_SYMBOL(ldlm_cli_cancel_list);
1916 * Cancel all locks on a resource that have 0 readers/writers.
1918 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1919 * to notify the server. */
1920 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1921 const struct ldlm_res_id *res_id,
1922 ldlm_policy_data_t *policy,
1924 ldlm_cancel_flags_t flags,
1927 struct ldlm_resource *res;
1928 struct list_head cancels = LIST_HEAD_INIT(cancels);
1933 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1935 /* This is not a problem. */
1936 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1940 LDLM_RESOURCE_ADDREF(res);
1941 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1942 0, flags | LCF_BL_AST, opaque);
1943 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1945 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
1948 LDLM_RESOURCE_DELREF(res);
1949 ldlm_resource_putref(res);
1952 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
1954 struct ldlm_cli_cancel_arg {
1959 static int ldlm_cli_hash_cancel_unused(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1960 cfs_hlist_node_t *hnode, void *arg)
1962 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1963 struct ldlm_cli_cancel_arg *lc = arg;
1965 ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1966 NULL, LCK_MINMODE, lc->lc_flags,
1968 /* must return 0 for hash iteration */
1973 * Cancel all locks on a namespace (or a specific resource, if given)
1974 * that have 0 readers/writers.
1976 * If flags & LCF_LOCAL, throw the locks away without trying
1977 * to notify the server. */
1978 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1979 const struct ldlm_res_id *res_id,
1980 ldlm_cancel_flags_t flags, void *opaque)
1982 struct ldlm_cli_cancel_arg arg = {
1984 .lc_opaque = opaque,
1992 if (res_id != NULL) {
1993 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1997 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1998 ldlm_cli_hash_cancel_unused, &arg);
2002 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
2004 /* Lock iterators. */
2006 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
2009 cfs_list_t *tmp, *next;
2010 struct ldlm_lock *lock;
2011 int rc = LDLM_ITER_CONTINUE;
2016 RETURN(LDLM_ITER_CONTINUE);
2019 cfs_list_for_each_safe(tmp, next, &res->lr_granted) {
2020 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
2022 if (iter(lock, closure) == LDLM_ITER_STOP)
2023 GOTO(out, rc = LDLM_ITER_STOP);
2026 cfs_list_for_each_safe(tmp, next, &res->lr_converting) {
2027 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
2029 if (iter(lock, closure) == LDLM_ITER_STOP)
2030 GOTO(out, rc = LDLM_ITER_STOP);
2033 cfs_list_for_each_safe(tmp, next, &res->lr_waiting) {
2034 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
2036 if (iter(lock, closure) == LDLM_ITER_STOP)
2037 GOTO(out, rc = LDLM_ITER_STOP);
2043 EXPORT_SYMBOL(ldlm_resource_foreach);
2045 struct iter_helper_data {
2046 ldlm_iterator_t iter;
2050 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2052 struct iter_helper_data *helper = closure;
2053 return helper->iter(lock, helper->closure);
2056 static int ldlm_res_iter_helper(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2057 cfs_hlist_node_t *hnode, void *arg)
2060 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2062 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2066 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2067 ldlm_iterator_t iter, void *closure)
2070 struct iter_helper_data helper = { .iter = iter, .closure = closure };
2072 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2073 ldlm_res_iter_helper, &helper);
2076 EXPORT_SYMBOL(ldlm_namespace_foreach);
2078 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2079 * return 0: find no resource
2080 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2083 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2084 const struct ldlm_res_id *res_id,
2085 ldlm_iterator_t iter, void *data)
2087 struct ldlm_resource *res;
2091 LASSERTF(ns != NULL, "must pass in namespace\n");
2093 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2097 LDLM_RESOURCE_ADDREF(res);
2098 rc = ldlm_resource_foreach(res, iter, data);
2099 LDLM_RESOURCE_DELREF(res);
2100 ldlm_resource_putref(res);
2103 EXPORT_SYMBOL(ldlm_resource_iterate);
2107 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2109 cfs_list_t *list = closure;
2111 /* we use l_pending_chain here, because it's unused on clients. */
2112 LASSERTF(cfs_list_empty(&lock->l_pending_chain),
2113 "lock %p next %p prev %p\n",
2114 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2115 /* bug 9573: don't replay locks left after eviction, or
2116 * bug 17614: locks being actively cancelled. Get a reference
2117 * on a lock so that it does not disapear under us (e.g. due to cancel)
2119 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
2120 cfs_list_add(&lock->l_pending_chain, list);
2121 LDLM_LOCK_GET(lock);
2124 return LDLM_ITER_CONTINUE;
2127 static int replay_lock_interpret(const struct lu_env *env,
2128 struct ptlrpc_request *req,
2129 struct ldlm_async_args *aa, int rc)
2131 struct ldlm_lock *lock;
2132 struct ldlm_reply *reply;
2133 struct obd_export *exp;
2136 atomic_dec(&req->rq_import->imp_replay_inflight);
2140 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2142 GOTO(out, rc = -EPROTO);
2144 lock = ldlm_handle2lock(&aa->lock_handle);
2146 CERROR("received replay ack for unknown local cookie "LPX64
2147 " remote cookie "LPX64 " from server %s id %s\n",
2148 aa->lock_handle.cookie, reply->lock_handle.cookie,
2149 req->rq_export->exp_client_uuid.uuid,
2150 libcfs_id2str(req->rq_peer));
2151 GOTO(out, rc = -ESTALE);
2154 /* Key change rehash lock in per-export hash with new key */
2155 exp = req->rq_export;
2156 if (exp && exp->exp_lock_hash) {
2157 /* In the function below, .hs_keycmp resolves to
2158 * ldlm_export_lock_keycmp() */
2159 /* coverity[overrun-buffer-val] */
2160 cfs_hash_rehash_key(exp->exp_lock_hash,
2161 &lock->l_remote_handle,
2162 &reply->lock_handle,
2165 lock->l_remote_handle = reply->lock_handle;
2168 LDLM_DEBUG(lock, "replayed lock:");
2169 ptlrpc_import_recovery_state_machine(req->rq_import);
2170 LDLM_LOCK_PUT(lock);
2173 ptlrpc_connect_import(req->rq_import);
2178 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2180 struct ptlrpc_request *req;
2181 struct ldlm_async_args *aa;
2182 struct ldlm_request *body;
2187 /* Bug 11974: Do not replay a lock which is actively being canceled */
2188 if (ldlm_is_canceling(lock)) {
2189 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2193 /* If this is reply-less callback lock, we cannot replay it, since
2194 * server might have long dropped it, but notification of that event was
2195 * lost by network. (and server granted conflicting lock already) */
2196 if (ldlm_is_cancel_on_block(lock)) {
2197 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2198 ldlm_lock_cancel(lock);
2203 * If granted mode matches the requested mode, this lock is granted.
2205 * If they differ, but we have a granted mode, then we were granted
2206 * one mode and now want another: ergo, converting.
2208 * If we haven't been granted anything and are on a resource list,
2209 * then we're blocked/waiting.
2211 * If we haven't been granted anything and we're NOT on a resource list,
2212 * then we haven't got a reply yet and don't have a known disposition.
2213 * This happens whenever a lock enqueue is the request that triggers
2216 if (lock->l_granted_mode == lock->l_req_mode)
2217 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2218 else if (lock->l_granted_mode)
2219 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2220 else if (!cfs_list_empty(&lock->l_res_link))
2221 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2223 flags = LDLM_FL_REPLAY;
2225 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2226 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2230 /* We're part of recovery, so don't wait for it. */
2231 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2233 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2234 ldlm_lock2desc(lock, &body->lock_desc);
2235 body->lock_flags = ldlm_flags_to_wire(flags);
2237 ldlm_lock2handle(lock, &body->lock_handle[0]);
2238 if (lock->l_lvb_len > 0)
2239 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2240 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2242 ptlrpc_request_set_replen(req);
2243 /* notify the server we've replayed all requests.
2244 * also, we mark the request to be put on a dedicated
2245 * queue to be processed after all request replayes.
2247 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2249 LDLM_DEBUG(lock, "replaying lock:");
2251 atomic_inc(&req->rq_import->imp_replay_inflight);
2252 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2253 aa = ptlrpc_req_async_args(req);
2254 aa->lock_handle = body->lock_handle[0];
2255 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2256 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2262 * Cancel as many unused locks as possible before replay. since we are
2263 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2266 * Called only in recovery before replaying locks. there is no need to
2267 * replay locks that are unused. since the clients may hold thousands of
2268 * cached unused locks, dropping the unused locks can greatly reduce the
2269 * load on the servers at recovery time.
2271 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2274 struct list_head cancels = LIST_HEAD_INIT(cancels);
2276 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2277 "replay for namespace %s (%d)\n",
2278 ldlm_ns_name(ns), ns->ns_nr_unused);
2280 /* We don't need to care whether or not LRU resize is enabled
2281 * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2282 * count parameter */
2283 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2284 LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2286 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2287 canceled, ldlm_ns_name(ns));
2290 int ldlm_replay_locks(struct obd_import *imp)
2292 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2293 struct list_head list = LIST_HEAD_INIT(list);
2294 struct ldlm_lock *lock, *next;
2299 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2301 /* don't replay locks if import failed recovery */
2302 if (imp->imp_vbr_failed)
2305 /* ensure this doesn't fall to 0 before all have been queued */
2306 atomic_inc(&imp->imp_replay_inflight);
2308 if (ldlm_cancel_unused_locks_before_replay)
2309 ldlm_cancel_unused_locks_for_replay(ns);
2311 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2313 cfs_list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2314 cfs_list_del_init(&lock->l_pending_chain);
2316 LDLM_LOCK_RELEASE(lock);
2317 continue; /* or try to do the rest? */
2319 rc = replay_one_lock(imp, lock);
2320 LDLM_LOCK_RELEASE(lock);
2323 atomic_dec(&imp->imp_replay_inflight);
2327 EXPORT_SYMBOL(ldlm_replay_locks);