4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 * This file contains Asynchronous System Trap (AST) handlers and related
38 * LDLM request-processing routines.
40 * An AST is a callback issued on a lock when its state is changed. There are
41 * several different types of ASTs (callbacks) registered for each lock:
43 * - completion AST: when a lock is enqueued by some process, but cannot be
44 * granted immediately due to other conflicting locks on the same resource,
45 * the completion AST is sent to notify the caller when the lock is
48 * - blocking AST: when a lock is granted to some process, if another process
49 * enqueues a conflicting (blocking) lock on a resource, a blocking AST is
50 * sent to notify the holder(s) of the lock(s) of the conflicting lock
51 * request. The lock holder(s) must release their lock(s) on that resource in
52 * a timely manner or be evicted by the server.
54 * - glimpse AST: this is used when a process wants information about a lock
55 * (i.e. the lock value block (LVB)) but does not necessarily require holding
56 * the lock. If the resource is locked, the lock holder(s) are sent glimpse
57 * ASTs and the LVB is returned to the caller, and lock holder(s) may CANCEL
58 * their lock(s) if they are idle. If the resource is not locked, the server
62 #define DEBUG_SUBSYSTEM S_LDLM
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 LCONSOLE_WARN("lock timed out (enqueued at "CFS_TIME_T", "
109 CFS_DURATION_T"s ago)\n",
110 lock->l_last_activity,
111 cfs_time_sub(cfs_time_current_sec(),
112 lock->l_last_activity));
113 LDLM_DEBUG(lock, "lock timed out (enqueued at "CFS_TIME_T", "
114 CFS_DURATION_T"s ago); not entering recovery in "
115 "server code, just going back to sleep",
116 lock->l_last_activity,
117 cfs_time_sub(cfs_time_current_sec(),
118 lock->l_last_activity));
119 if (cfs_time_after(cfs_time_current(), next_dump)) {
120 last_dump = next_dump;
121 next_dump = cfs_time_shift(300);
122 ldlm_namespace_dump(D_DLMTRACE,
123 ldlm_lock_to_ns(lock));
125 libcfs_debug_dumplog();
130 obd = lock->l_conn_export->exp_obd;
131 imp = obd->u.cli.cl_import;
132 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
133 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
134 CFS_DURATION_T"s ago), entering recovery for %s@%s",
135 lock->l_last_activity,
136 cfs_time_sub(cfs_time_current_sec(), lock->l_last_activity),
137 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
141 EXPORT_SYMBOL(ldlm_expired_completion_wait);
143 /* We use the same basis for both server side and client side functions
144 from a single node. */
145 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
147 int timeout = at_get(ldlm_lock_to_ns_at(lock));
149 return obd_timeout / 2;
150 /* Since these are non-updating timeouts, we should be conservative.
151 It would be nice to have some kind of "early reply" mechanism for
152 lock callbacks too... */
153 timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */
154 return max(timeout, ldlm_enqueue_min);
156 EXPORT_SYMBOL(ldlm_get_enq_timeout);
159 * Helper function for ldlm_completion_ast(), updating timings when lock is
162 static int ldlm_completion_tail(struct ldlm_lock *lock)
167 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
168 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
171 delay = cfs_time_sub(cfs_time_current_sec(),
172 lock->l_last_activity);
173 LDLM_DEBUG(lock, "client-side enqueue: granted after "
174 CFS_DURATION_T"s", delay);
176 /* Update our time estimate */
177 at_measured(ldlm_lock_to_ns_at(lock),
185 * Implementation of ->l_completion_ast() for a client, that doesn't wait
186 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
187 * other threads that cannot block for long.
189 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
193 if (flags == LDLM_FL_WAIT_NOREPROC) {
194 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
198 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
199 LDLM_FL_BLOCK_CONV))) {
200 cfs_waitq_signal(&lock->l_waitq);
201 RETURN(ldlm_completion_tail(lock));
204 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
206 ldlm_reprocess_all(lock->l_resource);
209 EXPORT_SYMBOL(ldlm_completion_ast_async);
212 * Generic LDLM "completion" AST. This is called in several cases:
214 * - when a reply to an ENQUEUE RPC is received from the server
215 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
216 * this point (determined by flags);
218 * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
221 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
224 * - to force all locks when resource is destroyed (cleanup_resource());
226 * - during lock conversion (not used currently).
228 * If lock is not granted in the first case, this function waits until second
229 * or penultimate cases happen in some other thread.
232 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
234 /* XXX ALLOCATE - 160 bytes */
235 struct lock_wait_data lwd;
236 struct obd_device *obd;
237 struct obd_import *imp = NULL;
238 struct l_wait_info lwi;
243 if (flags == LDLM_FL_WAIT_NOREPROC) {
244 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
248 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
249 LDLM_FL_BLOCK_CONV))) {
250 cfs_waitq_signal(&lock->l_waitq);
254 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
259 obd = class_exp2obd(lock->l_conn_export);
261 /* if this is a local lock, then there is no import */
263 imp = obd->u.cli.cl_import;
266 /* Wait a long time for enqueue - server may have to callback a
267 lock from another client. Server will evict the other client if it
268 doesn't respond reasonably, and then give us the lock. */
269 timeout = ldlm_get_enq_timeout(lock) * 2;
273 if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
274 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
275 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
277 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
278 ldlm_expired_completion_wait,
279 interrupted_completion_wait, &lwd);
283 spin_lock(&imp->imp_lock);
284 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
285 spin_unlock(&imp->imp_lock);
288 if (ns_is_client(ldlm_lock_to_ns(lock)) &&
289 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
290 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
291 lock->l_flags |= LDLM_FL_FAIL_LOC;
294 /* Go to sleep until the lock is granted or cancelled. */
295 rc = l_wait_event(lock->l_waitq,
296 is_granted_or_cancelled(lock), &lwi);
300 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
305 RETURN(ldlm_completion_tail(lock));
307 EXPORT_SYMBOL(ldlm_completion_ast);
310 * A helper to build a blocking AST function
312 * Perform a common operation for blocking ASTs:
313 * defferred lock cancellation.
315 * \param lock the lock blocking or canceling AST was called on
317 * \see mdt_blocking_ast
318 * \see ldlm_blocking_ast
320 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
325 lock->l_flags |= LDLM_FL_CBPENDING;
326 do_ast = (!lock->l_readers && !lock->l_writers);
327 unlock_res_and_lock(lock);
330 struct lustre_handle lockh;
333 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
334 ldlm_lock2handle(lock, &lockh);
335 rc = ldlm_cli_cancel(&lockh);
337 CERROR("ldlm_cli_cancel: %d\n", rc);
339 LDLM_DEBUG(lock, "Lock still has references, will be "
344 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
347 * Server blocking AST
349 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
352 * \param lock the lock which blocks a request or cancelling lock
355 * \param flag indicates whether this cancelling or blocking callback
357 * \see ldlm_blocking_ast_nocheck
359 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
360 void *data, int flag)
364 if (flag == LDLM_CB_CANCELING) {
365 /* Don't need to do anything here. */
369 lock_res_and_lock(lock);
370 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
371 * that ldlm_blocking_ast is called just before intent_policy method
372 * takes the lr_lock, then by the time we get the lock, we might not
373 * be the correct blocking function anymore. So check, and return
375 if (lock->l_blocking_ast != ldlm_blocking_ast) {
376 unlock_res_and_lock(lock);
379 RETURN(ldlm_blocking_ast_nocheck(lock));
381 EXPORT_SYMBOL(ldlm_blocking_ast);
384 * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
385 * comment in filter_intent_policy() on why you may need this.
387 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
390 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
391 * that is rather subtle: with OST-side locking, it may so happen that
392 * _all_ extent locks are held by the OST. If client wants to obtain
393 * current file size it calls ll{,u}_glimpse_size(), and (as locks are
394 * on the server), dummy glimpse callback fires and does
395 * nothing. Client still receives correct file size due to the
396 * following fragment in filter_intent_policy():
398 * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
399 * if (rc != 0 && res->lr_namespace->ns_lvbo &&
400 * res->lr_namespace->ns_lvbo->lvbo_update) {
401 * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
404 * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
405 * returns correct file size to the client.
407 return -ELDLM_NO_LOCK_DATA;
409 EXPORT_SYMBOL(ldlm_glimpse_ast);
412 * Enqueue a local lock (typically on a server).
414 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
415 const struct ldlm_res_id *res_id,
416 ldlm_type_t type, ldlm_policy_data_t *policy,
417 ldlm_mode_t mode, __u64 *flags,
418 ldlm_blocking_callback blocking,
419 ldlm_completion_callback completion,
420 ldlm_glimpse_callback glimpse,
421 void *data, __u32 lvb_len, enum lvb_type lvb_type,
422 const __u64 *client_cookie,
423 struct lustre_handle *lockh)
425 struct ldlm_lock *lock;
427 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
428 .lcs_blocking = blocking,
429 .lcs_glimpse = glimpse,
433 LASSERT(!(*flags & LDLM_FL_REPLAY));
434 if (unlikely(ns_is_client(ns))) {
435 CERROR("Trying to enqueue local lock in a shadow namespace\n");
439 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
442 GOTO(out_nolock, err = -ENOMEM);
444 ldlm_lock2handle(lock, lockh);
446 /* NB: we don't have any lock now (lock_res_and_lock)
447 * because it's a new lock */
448 ldlm_lock_addref_internal_nolock(lock, mode);
449 lock->l_flags |= LDLM_FL_LOCAL;
450 if (*flags & LDLM_FL_ATOMIC_CB)
451 lock->l_flags |= LDLM_FL_ATOMIC_CB;
454 lock->l_policy_data = *policy;
455 if (client_cookie != NULL)
456 lock->l_client_cookie = *client_cookie;
457 if (type == LDLM_EXTENT)
458 lock->l_req_extent = policy->l_extent;
460 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
461 if (unlikely(err != ELDLM_OK))
465 *policy = lock->l_policy_data;
467 if (lock->l_completion_ast)
468 lock->l_completion_ast(lock, *flags, NULL);
470 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
473 LDLM_LOCK_RELEASE(lock);
477 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
479 static void failed_lock_cleanup(struct ldlm_namespace *ns,
480 struct ldlm_lock *lock, int mode)
484 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
485 lock_res_and_lock(lock);
486 /* Check that lock is not granted or failed, we might race. */
487 if ((lock->l_req_mode != lock->l_granted_mode) &&
488 !(lock->l_flags & LDLM_FL_FAILED)) {
489 /* Make sure that this lock will not be found by raced
490 * bl_ast and -EINVAL reply is sent to server anyways.
492 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
493 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
496 unlock_res_and_lock(lock);
500 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
501 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
503 LDLM_DEBUG(lock, "lock was granted or failed in race");
505 ldlm_lock_decref_internal(lock, mode);
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 ldlm_resource_unlink_lock(lock);
516 ldlm_lock_destroy_nolock(lock);
517 unlock_res_and_lock(lock);
522 * Finishing portion of client lock enqueue code.
524 * Called after receiving reply from server.
526 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
527 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
528 __u64 *flags, void *lvb, __u32 lvb_len,
529 struct lustre_handle *lockh,int rc)
531 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
532 int is_replay = *flags & LDLM_FL_REPLAY;
533 struct ldlm_lock *lock;
534 struct ldlm_reply *reply;
535 int cleanup_phase = 1;
539 lock = ldlm_handle2lock(lockh);
540 /* ldlm_cli_enqueue is holding a reference on this lock. */
542 LASSERT(type == LDLM_FLOCK);
546 LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
547 "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
549 if (rc != ELDLM_OK) {
551 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
552 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
554 if (rc != ELDLM_LOCK_ABORTED)
558 /* Before we return, swab the reply */
559 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
561 GOTO(cleanup, rc = -EPROTO);
564 LASSERT(lvb != NULL);
566 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
569 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
570 GOTO(cleanup, rc = size);
571 } else if (unlikely(size > lvb_len)) {
572 LDLM_ERROR(lock, "Replied LVB is larger than "
573 "expectation, expected = %d, replied = %d",
575 GOTO(cleanup, rc = -EINVAL);
579 if (rc == ELDLM_LOCK_ABORTED) {
581 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
583 GOTO(cleanup, rc = (rc != 0 ? rc : ELDLM_LOCK_ABORTED));
586 /* lock enqueued on the server */
589 lock_res_and_lock(lock);
590 /* Key change rehash lock in per-export hash with new key */
591 if (exp->exp_lock_hash) {
592 /* In the function below, .hs_keycmp resolves to
593 * ldlm_export_lock_keycmp() */
594 /* coverity[overrun-buffer-val] */
595 cfs_hash_rehash_key(exp->exp_lock_hash,
596 &lock->l_remote_handle,
600 lock->l_remote_handle = reply->lock_handle;
603 *flags = ldlm_flags_from_wire(reply->lock_flags);
604 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
606 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
607 * to wait with no timeout as well */
608 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
610 unlock_res_and_lock(lock);
612 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%llx\n",
613 lock, reply->lock_handle.cookie, *flags);
615 /* If enqueue returned a blocked lock but the completion handler has
616 * already run, then it fixed up the resource and we don't need to do it
618 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
619 int newmode = reply->lock_desc.l_req_mode;
621 if (newmode && newmode != lock->l_req_mode) {
622 LDLM_DEBUG(lock, "server returned different mode %s",
623 ldlm_lockname[newmode]);
624 lock->l_req_mode = newmode;
627 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
628 lock->l_resource->lr_name.name,
629 sizeof(struct ldlm_res_id))) {
630 CDEBUG(D_INFO, "remote intent success, locking "
631 "(%ld,%ld,%ld) instead of "
633 (long)reply->lock_desc.l_resource.lr_name.name[0],
634 (long)reply->lock_desc.l_resource.lr_name.name[1],
635 (long)reply->lock_desc.l_resource.lr_name.name[2],
636 (long)lock->l_resource->lr_name.name[0],
637 (long)lock->l_resource->lr_name.name[1],
638 (long)lock->l_resource->lr_name.name[2]);
640 rc = ldlm_lock_change_resource(ns, lock,
641 &reply->lock_desc.l_resource.lr_name);
642 if (rc || lock->l_resource == NULL)
643 GOTO(cleanup, rc = -ENOMEM);
644 LDLM_DEBUG(lock, "client-side enqueue, new resource");
647 if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
649 /* We assume lock type cannot change on server*/
650 ldlm_convert_policy_to_local(exp,
651 lock->l_resource->lr_type,
652 &reply->lock_desc.l_policy_data,
653 &lock->l_policy_data);
654 if (type != LDLM_PLAIN)
655 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
658 if ((*flags) & LDLM_FL_AST_SENT ||
659 /* Cancel extent locks as soon as possible on a liblustre client,
660 * because it cannot handle asynchronous ASTs robustly (see
662 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
663 lock_res_and_lock(lock);
664 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
665 unlock_res_and_lock(lock);
666 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
669 /* If the lock has already been granted by a completion AST, don't
670 * clobber the LVB with an older one. */
672 /* We must lock or a racing completion might update lvb without
673 * letting us know and we'll clobber the correct value.
674 * Cannot unlock after the check either, a that still leaves
675 * a tiny window for completion to get in */
676 lock_res_and_lock(lock);
677 if (lock->l_req_mode != lock->l_granted_mode)
678 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
679 lock->l_lvb_data, size);
680 unlock_res_and_lock(lock);
686 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
687 if (lock->l_completion_ast != NULL) {
688 int err = lock->l_completion_ast(lock, *flags, NULL);
696 if (lvb_len && lvb != NULL) {
697 /* Copy the LVB here, and not earlier, because the completion
698 * AST (if any) can override what we got in the reply */
699 memcpy(lvb, lock->l_lvb_data, lvb_len);
702 LDLM_DEBUG(lock, "client-side enqueue END");
705 if (cleanup_phase == 1 && rc)
706 failed_lock_cleanup(ns, lock, mode);
707 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
709 LDLM_LOCK_RELEASE(lock);
712 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
715 * Estimate number of lock handles that would fit into request of given
716 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
717 * a single page on the send/receive side. XXX: 512 should be changed to
718 * more adequate value.
720 static inline int ldlm_req_handles_avail(int req_size, int off)
724 avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size;
725 if (likely(avail >= 0))
726 avail /= (int)sizeof(struct lustre_handle);
729 avail += LDLM_LOCKREQ_HANDLES - off;
734 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
735 enum req_location loc,
738 int size = req_capsule_msg_size(pill, loc);
739 return ldlm_req_handles_avail(size, off);
742 static inline int ldlm_format_handles_avail(struct obd_import *imp,
743 const struct req_format *fmt,
744 enum req_location loc, int off)
746 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
747 return ldlm_req_handles_avail(size, off);
751 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
752 * \a count locks in \a cancels.
754 * This is to be called by functions preparing their own requests that
755 * might contain lists of locks to cancel in addition to actual operation
756 * that needs to be performed.
758 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
759 int version, int opc, int canceloff,
760 cfs_list_t *cancels, int count)
762 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
763 struct req_capsule *pill = &req->rq_pill;
764 struct ldlm_request *dlm = NULL;
765 int flags, avail, to_free, pack = 0;
772 if (ns_connect_cancelset(ns)) {
773 /* Estimate the amount of available space in the request. */
774 req_capsule_filled_sizes(pill, RCL_CLIENT);
775 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
777 flags = ns_connect_lru_resize(ns) ?
778 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
779 to_free = !ns_connect_lru_resize(ns) &&
780 opc == LDLM_ENQUEUE ? 1 : 0;
782 /* Cancel LRU locks here _only_ if the server supports
783 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
784 * RPC, which will make us slower. */
786 count += ldlm_cancel_lru_local(ns, cancels, to_free,
787 avail - count, 0, flags);
792 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
793 ldlm_request_bufsize(pack, opc));
796 rc = ptlrpc_request_pack(req, version, opc);
798 ldlm_lock_list_put(cancels, l_bl_ast, count);
802 if (ns_connect_cancelset(ns)) {
804 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
806 /* Skip first lock handler in ldlm_request_pack(),
807 * this method will incrment @lock_count according
808 * to the lock handle amount actually written to
810 dlm->lock_count = canceloff;
812 /* Pack into the request @pack lock handles. */
813 ldlm_cli_cancel_list(cancels, pack, req, 0);
814 /* Prepare and send separate cancel RPC for others. */
815 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
817 ldlm_lock_list_put(cancels, l_bl_ast, count);
821 EXPORT_SYMBOL(ldlm_prep_elc_req);
823 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
824 cfs_list_t *cancels, int count)
826 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
827 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
829 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
832 * Client-side lock enqueue.
834 * If a request has some specific initialisation it is passed in \a reqp,
835 * otherwise it is created in ldlm_cli_enqueue.
837 * Supports sync and async requests, pass \a async flag accordingly. If a
838 * request was created in ldlm_cli_enqueue and it is the async request,
839 * pass it to the caller in \a reqp.
841 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
842 struct ldlm_enqueue_info *einfo,
843 const struct ldlm_res_id *res_id,
844 ldlm_policy_data_t const *policy, __u64 *flags,
845 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
846 struct lustre_handle *lockh, int async)
848 struct ldlm_namespace *ns;
849 struct ldlm_lock *lock;
850 struct ldlm_request *body;
851 int is_replay = *flags & LDLM_FL_REPLAY;
852 int req_passed_in = 1;
854 struct ptlrpc_request *req;
857 LASSERT(exp != NULL);
859 ns = exp->exp_obd->obd_namespace;
861 /* If we're replaying this lock, just check some invariants.
862 * If we're creating a new lock, get everything all setup nice. */
864 lock = ldlm_handle2lock_long(lockh, 0);
865 LASSERT(lock != NULL);
866 LDLM_DEBUG(lock, "client-side enqueue START");
867 LASSERT(exp == lock->l_conn_export);
869 const struct ldlm_callback_suite cbs = {
870 .lcs_completion = einfo->ei_cb_cp,
871 .lcs_blocking = einfo->ei_cb_bl,
872 .lcs_glimpse = einfo->ei_cb_gl,
873 .lcs_weigh = einfo->ei_cb_wg
875 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
876 einfo->ei_mode, &cbs, einfo->ei_cbdata,
880 /* for the local lock, add the reference */
881 ldlm_lock_addref_internal(lock, einfo->ei_mode);
882 ldlm_lock2handle(lock, lockh);
883 if (policy != NULL) {
884 /* INODEBITS_INTEROP: If the server does not support
885 * inodebits, we will request a plain lock in the
886 * descriptor (ldlm_lock2desc() below) but use an
887 * inodebits lock internally with both bits set.
889 if (einfo->ei_type == LDLM_IBITS &&
890 !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
891 lock->l_policy_data.l_inodebits.bits =
892 MDS_INODELOCK_LOOKUP |
893 MDS_INODELOCK_UPDATE;
895 lock->l_policy_data = *policy;
898 if (einfo->ei_type == LDLM_EXTENT)
899 lock->l_req_extent = policy->l_extent;
900 LDLM_DEBUG(lock, "client-side enqueue START, flags %llx\n",
904 lock->l_conn_export = exp;
905 lock->l_export = NULL;
906 lock->l_blocking_ast = einfo->ei_cb_bl;
907 lock->l_flags |= (*flags & LDLM_FL_NO_LRU);
909 /* lock not sent to server yet */
911 if (reqp == NULL || *reqp == NULL) {
912 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
917 failed_lock_cleanup(ns, lock, einfo->ei_mode);
918 LDLM_LOCK_RELEASE(lock);
928 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
930 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
931 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
934 /* Dump lock data into the request buffer */
935 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
936 ldlm_lock2desc(lock, &body->lock_desc);
937 body->lock_flags = ldlm_flags_to_wire(*flags);
938 body->lock_handle[0] = *lockh;
940 /* Continue as normal. */
941 if (!req_passed_in) {
943 req_capsule_extend(&req->rq_pill,
944 &RQF_LDLM_ENQUEUE_LVB);
945 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
947 ptlrpc_request_set_replen(req);
951 * Liblustre client doesn't get extent locks, except for O_APPEND case
952 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
953 * [i_size, OBD_OBJECT_EOF] lock is taken.
955 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
956 policy->l_extent.end == OBD_OBJECT_EOF));
959 LASSERT(reqp != NULL);
963 LDLM_DEBUG(lock, "sending request");
965 rc = ptlrpc_queue_wait(req);
967 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
968 einfo->ei_mode, flags, lvb, lvb_len,
971 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
972 * one reference that we took */
974 LDLM_LOCK_RELEASE(lock);
978 if (!req_passed_in && req != NULL) {
979 ptlrpc_req_finished(req);
986 EXPORT_SYMBOL(ldlm_cli_enqueue);
988 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
991 struct ldlm_resource *res;
994 if (ns_is_client(ldlm_lock_to_ns(lock))) {
995 CERROR("Trying to cancel local lock\n");
998 LDLM_DEBUG(lock, "client-side local convert");
1000 res = ldlm_lock_convert(lock, new_mode, flags);
1002 ldlm_reprocess_all(res);
1007 LDLM_DEBUG(lock, "client-side local convert handler END");
1008 LDLM_LOCK_PUT(lock);
1012 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
1013 * conversion of locks which are on the waiting or converting queue */
1014 /* Caller of this code is supposed to take care of lock readers/writers
1016 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
1018 struct ldlm_request *body;
1019 struct ldlm_reply *reply;
1020 struct ldlm_lock *lock;
1021 struct ldlm_resource *res;
1022 struct ptlrpc_request *req;
1026 lock = ldlm_handle2lock(lockh);
1033 if (lock->l_conn_export == NULL)
1034 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
1036 LDLM_DEBUG(lock, "client-side convert");
1038 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
1039 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1042 LDLM_LOCK_PUT(lock);
1046 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1047 body->lock_handle[0] = lock->l_remote_handle;
1049 body->lock_desc.l_req_mode = new_mode;
1050 body->lock_flags = ldlm_flags_to_wire(*flags);
1053 ptlrpc_request_set_replen(req);
1054 rc = ptlrpc_queue_wait(req);
1058 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1060 GOTO(out, rc = -EPROTO);
1063 GOTO(out, rc = req->rq_status);
1065 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
1067 ldlm_reprocess_all(res);
1068 /* Go to sleep until the lock is granted. */
1069 /* FIXME: or cancelled. */
1070 if (lock->l_completion_ast) {
1071 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1081 LDLM_LOCK_PUT(lock);
1082 ptlrpc_req_finished(req);
1085 EXPORT_SYMBOL(ldlm_cli_convert);
1088 * Cancel locks locally.
1090 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1091 * \retval LDLM_FL_CANCELING otherwise;
1092 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1094 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1096 __u64 rc = LDLM_FL_LOCAL_ONLY;
1099 if (lock->l_conn_export) {
1102 LDLM_DEBUG(lock, "client-side cancel");
1103 /* Set this flag to prevent others from getting new references*/
1104 lock_res_and_lock(lock);
1105 lock->l_flags |= LDLM_FL_CBPENDING;
1106 local_only = !!(lock->l_flags &
1107 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1108 ldlm_cancel_callback(lock);
1109 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
1110 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1111 unlock_res_and_lock(lock);
1114 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1116 rc = LDLM_FL_LOCAL_ONLY;
1118 ldlm_lock_cancel(lock);
1120 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1121 LDLM_ERROR(lock, "Trying to cancel local lock");
1124 LDLM_DEBUG(lock, "server-side local cancel");
1125 ldlm_lock_cancel(lock);
1126 ldlm_reprocess_all(lock->l_resource);
1133 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1135 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1136 cfs_list_t *head, int count)
1138 struct ldlm_request *dlm;
1139 struct ldlm_lock *lock;
1140 int max, packed = 0;
1143 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1144 LASSERT(dlm != NULL);
1146 /* Check the room in the request buffer. */
1147 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1148 sizeof(struct ldlm_request);
1149 max /= sizeof(struct lustre_handle);
1150 max += LDLM_LOCKREQ_HANDLES;
1151 LASSERT(max >= dlm->lock_count + count);
1153 /* XXX: it would be better to pack lock handles grouped by resource.
1154 * so that the server cancel would call filter_lvbo_update() less
1156 cfs_list_for_each_entry(lock, head, l_bl_ast) {
1159 LASSERT(lock->l_conn_export);
1160 /* Pack the lock handle to the given request buffer. */
1161 LDLM_DEBUG(lock, "packing");
1162 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1165 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1170 * Prepare and send a batched cancel RPC. It will include \a count lock
1171 * handles of locks given in \a cancels list. */
1172 int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels,
1173 int count, ldlm_cancel_flags_t flags)
1175 struct ptlrpc_request *req = NULL;
1176 struct obd_import *imp;
1181 LASSERT(exp != NULL);
1184 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1186 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1189 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1190 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1195 imp = class_exp2cliimp(exp);
1196 if (imp == NULL || imp->imp_invalid) {
1198 "skipping cancel on invalid import %p\n", imp);
1202 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1204 GOTO(out, rc = -ENOMEM);
1206 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1207 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1208 ldlm_request_bufsize(count, LDLM_CANCEL));
1210 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1212 ptlrpc_request_free(req);
1216 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1217 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1218 ptlrpc_at_set_req_timeout(req);
1220 ldlm_cancel_pack(req, cancels, count);
1222 ptlrpc_request_set_replen(req);
1223 if (flags & LCF_ASYNC) {
1224 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1228 rc = ptlrpc_queue_wait(req);
1231 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1232 "out of sync -- not fatal\n",
1233 libcfs_nid2str(req->rq_import->
1234 imp_connection->c_peer.nid));
1236 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1237 req->rq_import_generation == imp->imp_generation) {
1238 ptlrpc_req_finished(req);
1240 } else if (rc != ELDLM_OK) {
1241 /* -ESHUTDOWN is common on umount */
1242 CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1243 "Got rc %d from cancel RPC: "
1244 "canceling anyway\n", rc);
1251 ptlrpc_req_finished(req);
1254 return sent ? sent : rc;
1256 EXPORT_SYMBOL(ldlm_cli_cancel_req);
1258 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1260 LASSERT(imp != NULL);
1261 return &imp->imp_obd->obd_namespace->ns_pool;
1265 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1267 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1269 struct obd_device *obd;
1273 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1274 !imp_connect_lru_resize(req->rq_import)))
1277 * Do nothing for corner cases.
1282 /* In some cases RPC may contain SLV and limit zeroed out. This
1283 * is the case when server does not support LRU resize feature.
1284 * This is also possible in some recovery cases when server-side
1285 * reqs have no reference to the OBD export and thus access to
1286 * server-side namespace is not possible. */
1287 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1288 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1289 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1290 "(SLV: "LPU64", Limit: %u)",
1291 lustre_msg_get_slv(req->rq_repmsg),
1292 lustre_msg_get_limit(req->rq_repmsg));
1296 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1297 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1298 obd = req->rq_import->imp_obd;
1300 /* Set new SLV and limit in OBD fields to make them accessible
1301 * to the pool thread. We do not access obd_namespace and pool
1302 * directly here as there is no reliable way to make sure that
1303 * they are still alive at cleanup time. Evil races are possible
1304 * which may cause Oops at that time. */
1305 write_lock(&obd->obd_pool_lock);
1306 obd->obd_pool_slv = new_slv;
1307 obd->obd_pool_limit = new_limit;
1308 write_unlock(&obd->obd_pool_lock);
1312 EXPORT_SYMBOL(ldlm_cli_update_pool);
1315 * Client side lock cancel.
1317 * Lock must not have any readers or writers by this time.
1319 int ldlm_cli_cancel(struct lustre_handle *lockh)
1321 struct obd_export *exp;
1322 int avail, flags, count = 1;
1324 struct ldlm_namespace *ns;
1325 struct ldlm_lock *lock;
1326 CFS_LIST_HEAD(cancels);
1329 /* concurrent cancels on the same handle can happen */
1330 lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1332 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1336 rc = ldlm_cli_cancel_local(lock);
1337 if (rc == LDLM_FL_LOCAL_ONLY) {
1338 LDLM_LOCK_RELEASE(lock);
1341 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1342 * RPC which goes to canceld portal, so we can cancel other LRU locks
1343 * here and send them all as one LDLM_CANCEL RPC. */
1344 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1345 cfs_list_add(&lock->l_bl_ast, &cancels);
1347 exp = lock->l_conn_export;
1348 if (exp_connect_cancelset(exp)) {
1349 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1354 ns = ldlm_lock_to_ns(lock);
1355 flags = ns_connect_lru_resize(ns) ?
1356 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1357 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1360 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1363 EXPORT_SYMBOL(ldlm_cli_cancel);
1366 * Locally cancel up to \a count locks in list \a cancels.
1367 * Return the number of cancelled locks.
1369 int ldlm_cli_cancel_list_local(cfs_list_t *cancels, int count,
1370 ldlm_cancel_flags_t flags)
1372 CFS_LIST_HEAD(head);
1373 struct ldlm_lock *lock, *next;
1374 int left = 0, bl_ast = 0;
1378 cfs_list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1382 if (flags & LCF_LOCAL) {
1383 rc = LDLM_FL_LOCAL_ONLY;
1384 ldlm_lock_cancel(lock);
1386 rc = ldlm_cli_cancel_local(lock);
1388 /* Until we have compound requests and can send LDLM_CANCEL
1389 * requests batched with generic RPCs, we need to send cancels
1390 * with the LDLM_FL_BL_AST flag in a separate RPC from
1391 * the one being generated now. */
1392 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1393 LDLM_DEBUG(lock, "Cancel lock separately");
1394 cfs_list_del_init(&lock->l_bl_ast);
1395 cfs_list_add(&lock->l_bl_ast, &head);
1399 if (rc == LDLM_FL_LOCAL_ONLY) {
1400 /* CANCEL RPC should not be sent to server. */
1401 cfs_list_del_init(&lock->l_bl_ast);
1402 LDLM_LOCK_RELEASE(lock);
1408 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1413 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
1416 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1417 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1418 * readahead requests, ...)
1420 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1421 struct ldlm_lock *lock,
1422 int unused, int added,
1425 ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1426 ldlm_cancel_for_recovery cb = ns->ns_cancel_for_recovery;
1427 lock_res_and_lock(lock);
1429 /* don't check added & count since we want to process all locks
1430 * from unused list */
1431 switch (lock->l_resource->lr_type) {
1437 result = LDLM_POLICY_SKIP_LOCK;
1438 lock->l_flags |= LDLM_FL_SKIPPED;
1442 unlock_res_and_lock(lock);
1447 * Callback function for LRU-resize policy. Decides whether to keep
1448 * \a lock in LRU for current \a LRU size \a unused, added in current
1449 * scan \a added and number of locks to be preferably canceled \a count.
1451 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1453 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1455 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1456 struct ldlm_lock *lock,
1457 int unused, int added,
1460 cfs_time_t cur = cfs_time_current();
1461 struct ldlm_pool *pl = &ns->ns_pool;
1465 /* Stop LRU processing when we reach past @count or have checked all
1467 if (count && added >= count)
1468 return LDLM_POLICY_KEEP_LOCK;
1470 slv = ldlm_pool_get_slv(pl);
1471 lvf = ldlm_pool_get_lvf(pl);
1472 la = cfs_duration_sec(cfs_time_sub(cur,
1473 lock->l_last_used));
1474 lv = lvf * la * unused;
1476 /* Inform pool about current CLV to see it via proc. */
1477 ldlm_pool_set_clv(pl, lv);
1479 /* Stop when SLV is not yet come from server or lv is smaller than
1481 return (slv == 0 || lv < slv) ?
1482 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1486 * Callback function for proc used policy. Makes decision whether to keep
1487 * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1488 * added and number of locks to be preferably canceled \a count.
1490 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1492 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1494 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1495 struct ldlm_lock *lock,
1496 int unused, int added,
1499 /* Stop LRU processing when we reach past @count or have checked all
1501 return (added >= count) ?
1502 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1506 * Callback function for aged policy. Makes decision whether to keep \a lock in
1507 * LRU for current LRU size \a unused, added in current scan \a added and
1508 * number of locks to be preferably canceled \a count.
1510 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1512 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1514 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1515 struct ldlm_lock *lock,
1516 int unused, int added,
1519 /* Stop LRU processing if young lock is found and we reach past count */
1520 return ((added >= count) &&
1521 cfs_time_before(cfs_time_current(),
1522 cfs_time_add(lock->l_last_used,
1524 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1528 * Callback function for default policy. Makes decision whether to keep \a lock
1529 * in LRU for current LRU size \a unused, added in current scan \a added and
1530 * number of locks to be preferably canceled \a count.
1532 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1534 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1536 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1537 struct ldlm_lock *lock,
1538 int unused, int added,
1541 /* Stop LRU processing when we reach past count or have checked all
1543 return (added >= count) ?
1544 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1547 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1548 struct ldlm_lock *, int,
1551 static ldlm_cancel_lru_policy_t
1552 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1554 if (flags & LDLM_CANCEL_NO_WAIT)
1555 return ldlm_cancel_no_wait_policy;
1557 if (ns_connect_lru_resize(ns)) {
1558 if (flags & LDLM_CANCEL_SHRINK)
1559 /* We kill passed number of old locks. */
1560 return ldlm_cancel_passed_policy;
1561 else if (flags & LDLM_CANCEL_LRUR)
1562 return ldlm_cancel_lrur_policy;
1563 else if (flags & LDLM_CANCEL_PASSED)
1564 return ldlm_cancel_passed_policy;
1566 if (flags & LDLM_CANCEL_AGED)
1567 return ldlm_cancel_aged_policy;
1570 return ldlm_cancel_default_policy;
1574 * - Free space in LRU for \a count new locks,
1575 * redundant unused locks are canceled locally;
1576 * - also cancel locally unused aged locks;
1577 * - do not cancel more than \a max locks;
1578 * - GET the found locks and add them into the \a cancels list.
1580 * A client lock can be added to the l_bl_ast list only when it is
1581 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1582 * CANCEL. There are the following use cases:
1583 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1584 * ldlm_cli_cancel(), which check and set this flag properly. As any
1585 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1586 * later without any special locking.
1588 * Calling policies for enabled LRU resize:
1589 * ----------------------------------------
1590 * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to
1591 * cancel not more than \a count locks;
1593 * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at
1594 * the beginning of LRU list);
1596 * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to
1597 * memory pressre policy function;
1599 * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy".
1601 * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1602 * (typically before replaying locks) w/o
1603 * sending any RPCs or waiting for any
1604 * outstanding RPC to complete.
1606 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels,
1607 int count, int max, int flags)
1609 ldlm_cancel_lru_policy_t pf;
1610 struct ldlm_lock *lock, *next;
1611 int added = 0, unused, remained;
1614 spin_lock(&ns->ns_lock);
1615 unused = ns->ns_nr_unused;
1618 if (!ns_connect_lru_resize(ns))
1619 count += unused - ns->ns_max_unused;
1621 pf = ldlm_cancel_lru_policy(ns, flags);
1622 LASSERT(pf != NULL);
1624 while (!cfs_list_empty(&ns->ns_unused_list)) {
1625 ldlm_policy_res_t result;
1627 /* all unused locks */
1628 if (remained-- <= 0)
1631 /* For any flags, stop scanning if @max is reached. */
1632 if (max && added >= max)
1635 cfs_list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1637 /* No locks which got blocking requests. */
1638 LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1640 if (flags & LDLM_CANCEL_NO_WAIT &&
1641 lock->l_flags & LDLM_FL_SKIPPED)
1642 /* already processed */
1645 /* Somebody is already doing CANCEL. No need for this
1646 * lock in LRU, do not traverse it again. */
1647 if (!(lock->l_flags & LDLM_FL_CANCELING))
1650 ldlm_lock_remove_from_lru_nolock(lock);
1652 if (&lock->l_lru == &ns->ns_unused_list)
1655 LDLM_LOCK_GET(lock);
1656 spin_unlock(&ns->ns_lock);
1657 lu_ref_add(&lock->l_reference, __FUNCTION__, cfs_current());
1659 /* Pass the lock through the policy filter and see if it
1660 * should stay in LRU.
1662 * Even for shrinker policy we stop scanning if
1663 * we find a lock that should stay in the cache.
1664 * We should take into account lock age anyway
1665 * as a new lock is a valuable resource even if
1666 * it has a low weight.
1668 * That is, for shrinker policy we drop only
1669 * old locks, but additionally choose them by
1670 * their weight. Big extent locks will stay in
1672 result = pf(ns, lock, unused, added, count);
1673 if (result == LDLM_POLICY_KEEP_LOCK) {
1674 lu_ref_del(&lock->l_reference,
1675 __FUNCTION__, cfs_current());
1676 LDLM_LOCK_RELEASE(lock);
1677 spin_lock(&ns->ns_lock);
1680 if (result == LDLM_POLICY_SKIP_LOCK) {
1681 lu_ref_del(&lock->l_reference,
1682 __func__, cfs_current());
1683 LDLM_LOCK_RELEASE(lock);
1684 spin_lock(&ns->ns_lock);
1688 lock_res_and_lock(lock);
1689 /* Check flags again under the lock. */
1690 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1691 (ldlm_lock_remove_from_lru(lock) == 0)) {
1692 /* Another thread is removing lock from LRU, or
1693 * somebody is already doing CANCEL, or there
1694 * is a blocking request which will send cancel
1695 * by itself, or the lock is no longer unused. */
1696 unlock_res_and_lock(lock);
1697 lu_ref_del(&lock->l_reference,
1698 __FUNCTION__, cfs_current());
1699 LDLM_LOCK_RELEASE(lock);
1700 spin_lock(&ns->ns_lock);
1703 LASSERT(!lock->l_readers && !lock->l_writers);
1705 /* If we have chosen to cancel this lock voluntarily, we
1706 * better send cancel notification to server, so that it
1707 * frees appropriate state. This might lead to a race
1708 * where while we are doing cancel here, server is also
1709 * silently cancelling this lock. */
1710 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1712 /* Setting the CBPENDING flag is a little misleading,
1713 * but prevents an important race; namely, once
1714 * CBPENDING is set, the lock can accumulate no more
1715 * readers/writers. Since readers and writers are
1716 * already zero here, ldlm_lock_decref() won't see
1717 * this flag and call l_blocking_ast */
1718 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1720 /* We can't re-add to l_lru as it confuses the
1721 * refcounting in ldlm_lock_remove_from_lru() if an AST
1722 * arrives after we drop lr_lock below. We use l_bl_ast
1723 * and can't use l_pending_chain as it is used both on
1724 * server and client nevertheless bug 5666 says it is
1725 * used only on server */
1726 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1727 cfs_list_add(&lock->l_bl_ast, cancels);
1728 unlock_res_and_lock(lock);
1729 lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current());
1730 spin_lock(&ns->ns_lock);
1734 spin_unlock(&ns->ns_lock);
1738 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, cfs_list_t *cancels,
1739 int count, int max, ldlm_cancel_flags_t cancel_flags,
1743 added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1746 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1750 * Cancel at least \a nr locks from given namespace LRU.
1752 * When called with LDLM_ASYNC the blocking callback will be handled
1753 * in a thread and this function will return after the thread has been
1754 * asked to call the callback. When called with LDLM_SYNC the blocking
1755 * callback will be performed in this function.
1757 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t mode,
1760 CFS_LIST_HEAD(cancels);
1765 mode = LDLM_SYNC; /* force to be sync in user space */
1767 /* Just prepare the list of locks, do not actually cancel them yet.
1768 * Locks are cancelled later in a separate thread. */
1769 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1770 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, mode);
1778 * Find and cancel locally unused locks found on resource, matched to the
1779 * given policy, mode. GET the found locks and add them into the \a cancels
1782 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1783 cfs_list_t *cancels,
1784 ldlm_policy_data_t *policy,
1785 ldlm_mode_t mode, int lock_flags,
1786 ldlm_cancel_flags_t cancel_flags, void *opaque)
1788 struct ldlm_lock *lock;
1793 cfs_list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1794 if (opaque != NULL && lock->l_ast_data != opaque) {
1795 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1796 lock->l_ast_data, opaque);
1801 if (lock->l_readers || lock->l_writers)
1804 /* If somebody is already doing CANCEL, or blocking AST came,
1805 * skip this lock. */
1806 if (lock->l_flags & LDLM_FL_BL_AST ||
1807 lock->l_flags & LDLM_FL_CANCELING)
1810 if (lockmode_compat(lock->l_granted_mode, mode))
1813 /* If policy is given and this is IBITS lock, add to list only
1814 * those locks that match by policy. */
1815 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1816 !(lock->l_policy_data.l_inodebits.bits &
1817 policy->l_inodebits.bits))
1820 /* See CBPENDING comment in ldlm_cancel_lru */
1821 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1824 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1825 cfs_list_add(&lock->l_bl_ast, cancels);
1826 LDLM_LOCK_GET(lock);
1831 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
1833 EXPORT_SYMBOL(ldlm_cancel_resource_local);
1836 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1838 * If \a req is NULL, send CANCEL request to server with handles of locks
1839 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1840 * separately per lock.
1841 * If \a req is not NULL, put handles of locks in \a cancels into the request
1842 * buffer at the offset \a off.
1843 * Destroy \a cancels at the end.
1845 int ldlm_cli_cancel_list(cfs_list_t *cancels, int count,
1846 struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1848 struct ldlm_lock *lock;
1852 if (cfs_list_empty(cancels) || count == 0)
1855 /* XXX: requests (both batched and not) could be sent in parallel.
1856 * Usually it is enough to have just 1 RPC, but it is possible that
1857 * there are too many locks to be cancelled in LRU or on a resource.
1858 * It would also speed up the case when the server does not support
1861 LASSERT(!cfs_list_empty(cancels));
1862 lock = cfs_list_entry(cancels->next, struct ldlm_lock,
1864 LASSERT(lock->l_conn_export);
1866 if (exp_connect_cancelset(lock->l_conn_export)) {
1869 ldlm_cancel_pack(req, cancels, count);
1871 res = ldlm_cli_cancel_req(lock->l_conn_export,
1875 res = ldlm_cli_cancel_req(lock->l_conn_export,
1880 CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1881 "ldlm_cli_cancel_list: %d\n", res);
1886 ldlm_lock_list_put(cancels, l_bl_ast, res);
1888 LASSERT(count == 0);
1891 EXPORT_SYMBOL(ldlm_cli_cancel_list);
1894 * Cancel all locks on a resource that have 0 readers/writers.
1896 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1897 * to notify the server. */
1898 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1899 const struct ldlm_res_id *res_id,
1900 ldlm_policy_data_t *policy,
1902 ldlm_cancel_flags_t flags,
1905 struct ldlm_resource *res;
1906 CFS_LIST_HEAD(cancels);
1911 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1913 /* This is not a problem. */
1914 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1918 LDLM_RESOURCE_ADDREF(res);
1919 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1920 0, flags | LCF_BL_AST, opaque);
1921 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1923 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1925 LDLM_RESOURCE_DELREF(res);
1926 ldlm_resource_putref(res);
1929 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
1931 struct ldlm_cli_cancel_arg {
1936 static int ldlm_cli_hash_cancel_unused(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1937 cfs_hlist_node_t *hnode, void *arg)
1939 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1940 struct ldlm_cli_cancel_arg *lc = arg;
1943 rc = ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1945 lc->lc_flags, lc->lc_opaque);
1947 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1948 res->lr_name.name[0], rc);
1950 /* must return 0 for hash iteration */
1955 * Cancel all locks on a namespace (or a specific resource, if given)
1956 * that have 0 readers/writers.
1958 * If flags & LCF_LOCAL, throw the locks away without trying
1959 * to notify the server. */
1960 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1961 const struct ldlm_res_id *res_id,
1962 ldlm_cancel_flags_t flags, void *opaque)
1964 struct ldlm_cli_cancel_arg arg = {
1966 .lc_opaque = opaque,
1974 if (res_id != NULL) {
1975 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1979 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1980 ldlm_cli_hash_cancel_unused, &arg);
1984 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1986 /* Lock iterators. */
1988 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1991 cfs_list_t *tmp, *next;
1992 struct ldlm_lock *lock;
1993 int rc = LDLM_ITER_CONTINUE;
1998 RETURN(LDLM_ITER_CONTINUE);
2001 cfs_list_for_each_safe(tmp, next, &res->lr_granted) {
2002 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
2004 if (iter(lock, closure) == LDLM_ITER_STOP)
2005 GOTO(out, rc = LDLM_ITER_STOP);
2008 cfs_list_for_each_safe(tmp, next, &res->lr_converting) {
2009 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
2011 if (iter(lock, closure) == LDLM_ITER_STOP)
2012 GOTO(out, rc = LDLM_ITER_STOP);
2015 cfs_list_for_each_safe(tmp, next, &res->lr_waiting) {
2016 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
2018 if (iter(lock, closure) == LDLM_ITER_STOP)
2019 GOTO(out, rc = LDLM_ITER_STOP);
2025 EXPORT_SYMBOL(ldlm_resource_foreach);
2027 struct iter_helper_data {
2028 ldlm_iterator_t iter;
2032 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2034 struct iter_helper_data *helper = closure;
2035 return helper->iter(lock, helper->closure);
2038 static int ldlm_res_iter_helper(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2039 cfs_hlist_node_t *hnode, void *arg)
2042 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2044 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2048 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2049 ldlm_iterator_t iter, void *closure)
2052 struct iter_helper_data helper = { iter: iter, closure: closure };
2054 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2055 ldlm_res_iter_helper, &helper);
2058 EXPORT_SYMBOL(ldlm_namespace_foreach);
2060 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2061 * return 0: find no resource
2062 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2065 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2066 const struct ldlm_res_id *res_id,
2067 ldlm_iterator_t iter, void *data)
2069 struct ldlm_resource *res;
2074 CERROR("must pass in namespace\n");
2078 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2082 LDLM_RESOURCE_ADDREF(res);
2083 rc = ldlm_resource_foreach(res, iter, data);
2084 LDLM_RESOURCE_DELREF(res);
2085 ldlm_resource_putref(res);
2088 EXPORT_SYMBOL(ldlm_resource_iterate);
2092 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2094 cfs_list_t *list = closure;
2096 /* we use l_pending_chain here, because it's unused on clients. */
2097 LASSERTF(cfs_list_empty(&lock->l_pending_chain),
2098 "lock %p next %p prev %p\n",
2099 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2100 /* bug 9573: don't replay locks left after eviction, or
2101 * bug 17614: locks being actively cancelled. Get a reference
2102 * on a lock so that it does not disapear under us (e.g. due to cancel)
2104 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
2105 cfs_list_add(&lock->l_pending_chain, list);
2106 LDLM_LOCK_GET(lock);
2109 return LDLM_ITER_CONTINUE;
2112 static int replay_lock_interpret(const struct lu_env *env,
2113 struct ptlrpc_request *req,
2114 struct ldlm_async_args *aa, int rc)
2116 struct ldlm_lock *lock;
2117 struct ldlm_reply *reply;
2118 struct obd_export *exp;
2121 cfs_atomic_dec(&req->rq_import->imp_replay_inflight);
2126 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2128 GOTO(out, rc = -EPROTO);
2130 lock = ldlm_handle2lock(&aa->lock_handle);
2132 CERROR("received replay ack for unknown local cookie "LPX64
2133 " remote cookie "LPX64 " from server %s id %s\n",
2134 aa->lock_handle.cookie, reply->lock_handle.cookie,
2135 req->rq_export->exp_client_uuid.uuid,
2136 libcfs_id2str(req->rq_peer));
2137 GOTO(out, rc = -ESTALE);
2140 /* Key change rehash lock in per-export hash with new key */
2141 exp = req->rq_export;
2142 if (exp && exp->exp_lock_hash) {
2143 /* In the function below, .hs_keycmp resolves to
2144 * ldlm_export_lock_keycmp() */
2145 /* coverity[overrun-buffer-val] */
2146 cfs_hash_rehash_key(exp->exp_lock_hash,
2147 &lock->l_remote_handle,
2148 &reply->lock_handle,
2151 lock->l_remote_handle = reply->lock_handle;
2154 LDLM_DEBUG(lock, "replayed lock:");
2155 ptlrpc_import_recovery_state_machine(req->rq_import);
2156 LDLM_LOCK_PUT(lock);
2159 ptlrpc_connect_import(req->rq_import);
2164 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2166 struct ptlrpc_request *req;
2167 struct ldlm_async_args *aa;
2168 struct ldlm_request *body;
2173 /* Bug 11974: Do not replay a lock which is actively being canceled */
2174 if (lock->l_flags & LDLM_FL_CANCELING) {
2175 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2179 /* If this is reply-less callback lock, we cannot replay it, since
2180 * server might have long dropped it, but notification of that event was
2181 * lost by network. (and server granted conflicting lock already) */
2182 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2183 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2184 ldlm_lock_cancel(lock);
2189 * If granted mode matches the requested mode, this lock is granted.
2191 * If they differ, but we have a granted mode, then we were granted
2192 * one mode and now want another: ergo, converting.
2194 * If we haven't been granted anything and are on a resource list,
2195 * then we're blocked/waiting.
2197 * If we haven't been granted anything and we're NOT on a resource list,
2198 * then we haven't got a reply yet and don't have a known disposition.
2199 * This happens whenever a lock enqueue is the request that triggers
2202 if (lock->l_granted_mode == lock->l_req_mode)
2203 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2204 else if (lock->l_granted_mode)
2205 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2206 else if (!cfs_list_empty(&lock->l_res_link))
2207 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2209 flags = LDLM_FL_REPLAY;
2211 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2212 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2216 /* We're part of recovery, so don't wait for it. */
2217 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2219 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2220 ldlm_lock2desc(lock, &body->lock_desc);
2221 body->lock_flags = ldlm_flags_to_wire(flags);
2223 ldlm_lock2handle(lock, &body->lock_handle[0]);
2224 if (lock->l_lvb_len > 0)
2225 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2226 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2228 ptlrpc_request_set_replen(req);
2229 /* notify the server we've replayed all requests.
2230 * also, we mark the request to be put on a dedicated
2231 * queue to be processed after all request replayes.
2233 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2235 LDLM_DEBUG(lock, "replaying lock:");
2237 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2238 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2239 aa = ptlrpc_req_async_args(req);
2240 aa->lock_handle = body->lock_handle[0];
2241 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2242 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2248 * Cancel as many unused locks as possible before replay. since we are
2249 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2252 * Called only in recovery before replaying locks. there is no need to
2253 * replay locks that are unused. since the clients may hold thousands of
2254 * cached unused locks, dropping the unused locks can greatly reduce the
2255 * load on the servers at recovery time.
2257 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2260 CFS_LIST_HEAD(cancels);
2262 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2263 "replay for namespace %s (%d)\n",
2264 ldlm_ns_name(ns), ns->ns_nr_unused);
2266 /* We don't need to care whether or not LRU resize is enabled
2267 * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2268 * count parameter */
2269 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2270 LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2272 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2273 canceled, ldlm_ns_name(ns));
2276 int ldlm_replay_locks(struct obd_import *imp)
2278 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2279 CFS_LIST_HEAD(list);
2280 struct ldlm_lock *lock, *next;
2285 LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0);
2287 /* don't replay locks if import failed recovery */
2288 if (imp->imp_vbr_failed)
2291 /* ensure this doesn't fall to 0 before all have been queued */
2292 cfs_atomic_inc(&imp->imp_replay_inflight);
2294 if (ldlm_cancel_unused_locks_before_replay)
2295 ldlm_cancel_unused_locks_for_replay(ns);
2297 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2299 cfs_list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2300 cfs_list_del_init(&lock->l_pending_chain);
2302 LDLM_LOCK_RELEASE(lock);
2303 continue; /* or try to do the rest? */
2305 rc = replay_one_lock(imp, lock);
2306 LDLM_LOCK_RELEASE(lock);
2309 cfs_atomic_dec(&imp->imp_replay_inflight);
2313 EXPORT_SYMBOL(ldlm_replay_locks);