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
64 #include <lustre_dlm.h>
65 #include <obd_class.h>
68 #include "ldlm_internal.h"
70 unsigned int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
71 CFS_MODULE_PARM(ldlm_enqueue_min, "i", uint, 0644,
72 "lock enqueue timeout minimum");
74 /* in client side, whether the cached locks will be canceled before replay */
75 unsigned int ldlm_cancel_unused_locks_before_replay = 1;
77 static void interrupted_completion_wait(void *data)
81 struct lock_wait_data {
82 struct ldlm_lock *lwd_lock;
86 struct ldlm_async_args {
87 struct lustre_handle lock_handle;
90 int ldlm_expired_completion_wait(void *data)
92 struct lock_wait_data *lwd = data;
93 struct ldlm_lock *lock = lwd->lwd_lock;
94 struct obd_import *imp;
95 struct obd_device *obd;
98 if (lock->l_conn_export == NULL) {
99 static cfs_time_t next_dump = 0, last_dump = 0;
101 LCONSOLE_WARN("lock timed out (enqueued at "CFS_TIME_T", "
102 CFS_DURATION_T"s ago)\n",
103 lock->l_last_activity,
104 cfs_time_sub(cfs_time_current_sec(),
105 lock->l_last_activity));
106 LDLM_DEBUG(lock, "lock timed out (enqueued at "CFS_TIME_T", "
107 CFS_DURATION_T"s ago); not entering recovery in "
108 "server code, just going back to sleep",
109 lock->l_last_activity,
110 cfs_time_sub(cfs_time_current_sec(),
111 lock->l_last_activity));
112 if (cfs_time_after(cfs_time_current(), next_dump)) {
113 last_dump = next_dump;
114 next_dump = cfs_time_shift(300);
115 ldlm_namespace_dump(D_DLMTRACE,
116 ldlm_lock_to_ns(lock));
118 libcfs_debug_dumplog();
123 obd = lock->l_conn_export->exp_obd;
124 imp = obd->u.cli.cl_import;
125 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
126 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
127 CFS_DURATION_T"s ago), entering recovery for %s@%s",
128 lock->l_last_activity,
129 cfs_time_sub(cfs_time_current_sec(), lock->l_last_activity),
130 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
134 EXPORT_SYMBOL(ldlm_expired_completion_wait);
137 * Calculate the Completion timeout (covering enqueue, BL AST, data flush,
138 * lock cancel, and their replies). Used for lock completion timeout on the
141 * \param[in] lock lock which is waiting the completion callback
143 * \retval timeout in seconds to wait for the server reply
146 /* We use the same basis for both server side and client side functions
147 from a single node. */
148 static unsigned int ldlm_cp_timeout(struct ldlm_lock *lock)
150 unsigned int timeout;
155 /* Wait a long time for enqueue - server may have to callback a
156 * lock from another client. Server will evict the other client if it
157 * doesn't respond reasonably, and then give us the lock. */
158 timeout = at_get(ldlm_lock_to_ns_at(lock));
159 return max(3 * timeout, ldlm_enqueue_min);
163 * Helper function for ldlm_completion_ast(), updating timings when lock is
166 static int ldlm_completion_tail(struct ldlm_lock *lock, void *data)
171 if (ldlm_is_destroyed(lock) || ldlm_is_failed(lock)) {
172 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
174 } else if (data == NULL) {
175 LDLM_DEBUG(lock, "client-side enqueue: granted");
177 /* Take into AT only CP RPC, not immediately granted locks */
178 delay = cfs_time_sub(cfs_time_current_sec(),
179 lock->l_last_activity);
180 LDLM_DEBUG(lock, "client-side enqueue: granted after "
181 CFS_DURATION_T"s", delay);
183 /* Update our time estimate */
184 at_measured(ldlm_lock_to_ns_at(lock), delay);
190 * Implementation of ->l_completion_ast() for a client, that doesn't wait
191 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
192 * other threads that cannot block for long.
194 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
198 if (flags == LDLM_FL_WAIT_NOREPROC) {
199 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
203 if (!(flags & LDLM_FL_BLOCKED_MASK)) {
204 wake_up(&lock->l_waitq);
205 RETURN(ldlm_completion_tail(lock, data));
208 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
210 ldlm_reprocess_all(lock->l_resource);
213 EXPORT_SYMBOL(ldlm_completion_ast_async);
216 * Generic LDLM "completion" AST. This is called in several cases:
218 * - when a reply to an ENQUEUE RPC is received from the server
219 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
220 * this point (determined by flags);
222 * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
225 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
228 * - to force all locks when resource is destroyed (cleanup_resource());
230 * - during lock conversion (not used currently).
232 * If lock is not granted in the first case, this function waits until second
233 * or penultimate cases happen in some other thread.
236 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
238 /* XXX ALLOCATE - 160 bytes */
239 struct lock_wait_data lwd;
240 struct obd_device *obd;
241 struct obd_import *imp = NULL;
242 struct l_wait_info lwi;
247 if (flags == LDLM_FL_WAIT_NOREPROC) {
248 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
252 if (!(flags & LDLM_FL_BLOCKED_MASK)) {
253 wake_up(&lock->l_waitq);
257 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
262 obd = class_exp2obd(lock->l_conn_export);
264 /* if this is a local lock, then there is no import */
266 imp = obd->u.cli.cl_import;
269 timeout = ldlm_cp_timeout(lock);
272 lock->l_last_activity = cfs_time_current_sec();
274 if (ldlm_is_no_timeout(lock)) {
275 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
276 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
278 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
279 ldlm_expired_completion_wait,
280 interrupted_completion_wait, &lwd);
284 spin_lock(&imp->imp_lock);
285 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
286 spin_unlock(&imp->imp_lock);
289 if (ns_is_client(ldlm_lock_to_ns(lock)) &&
290 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
291 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
292 ldlm_set_fail_loc(lock);
295 /* Go to sleep until the lock is granted or cancelled. */
296 rc = l_wait_event(lock->l_waitq,
297 is_granted_or_cancelled(lock), &lwi);
301 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
306 RETURN(ldlm_completion_tail(lock, data));
308 EXPORT_SYMBOL(ldlm_completion_ast);
311 * A helper to build a blocking AST function
313 * Perform a common operation for blocking ASTs:
314 * defferred lock cancellation.
316 * \param lock the lock blocking or canceling AST was called on
318 * \see mdt_blocking_ast
319 * \see ldlm_blocking_ast
321 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
326 ldlm_set_cbpending(lock);
327 do_ast = (!lock->l_readers && !lock->l_writers);
328 unlock_res_and_lock(lock);
331 struct lustre_handle lockh;
334 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
335 ldlm_lock2handle(lock, &lockh);
336 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
338 CERROR("ldlm_cli_cancel: %d\n", rc);
340 LDLM_DEBUG(lock, "Lock still has references, will be "
345 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
348 * Server blocking AST
350 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
353 * \param lock the lock which blocks a request or cancelling lock
356 * \param flag indicates whether this cancelling or blocking callback
358 * \see ldlm_blocking_ast_nocheck
360 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
361 void *data, int flag)
365 if (flag == LDLM_CB_CANCELING) {
366 /* Don't need to do anything here. */
370 lock_res_and_lock(lock);
371 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
372 * that ldlm_blocking_ast is called just before intent_policy method
373 * takes the lr_lock, then by the time we get the lock, we might not
374 * be the correct blocking function anymore. So check, and return
376 if (lock->l_blocking_ast != ldlm_blocking_ast) {
377 unlock_res_and_lock(lock);
380 RETURN(ldlm_blocking_ast_nocheck(lock));
382 EXPORT_SYMBOL(ldlm_blocking_ast);
385 * Implements ldlm_lock::l_glimpse_ast for extent locks acquired on the server.
387 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for that is
388 * rather subtle: with OST-side locking, it may so happen that _all_ extent
389 * locks are held by the OST. If client wants to obtain the current file size
390 * it calls ll_glimpse_size(), and (as all locks are held only on the server),
391 * this dummy glimpse callback fires and does nothing. The client still
392 * receives the correct file size due to the following fragment of code in
393 * ldlm_cb_interpret():
395 * if (rc == -ELDLM_NO_LOCK_DATA) {
396 * LDLM_DEBUG(lock, "lost race - client has a lock but no"
398 * ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
401 * That is, after the glimpse returns this error, ofd_lvbo_update() is called
402 * and returns the updated file attributes from the inode to the client.
404 * See also comment in ofd_intent_policy() on why servers must set a non-NULL
405 * l_glimpse_ast when grabbing DLM locks. Otherwise, the server will assume
406 * that the object is in the process of being destroyed.
408 * \param[in] lock DLM lock being glimpsed, unused
409 * \param[in] reqp pointer to ptlrpc_request, unused
411 * \retval -ELDLM_NO_LOCK_DATA to get attributes from disk object
413 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
415 return -ELDLM_NO_LOCK_DATA;
417 EXPORT_SYMBOL(ldlm_glimpse_ast);
420 * Enqueue a local lock (typically on a server).
422 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
423 const struct ldlm_res_id *res_id,
424 ldlm_type_t type, ldlm_policy_data_t *policy,
425 ldlm_mode_t mode, __u64 *flags,
426 ldlm_blocking_callback blocking,
427 ldlm_completion_callback completion,
428 ldlm_glimpse_callback glimpse,
429 void *data, __u32 lvb_len, enum lvb_type lvb_type,
430 const __u64 *client_cookie,
431 struct lustre_handle *lockh)
433 struct ldlm_lock *lock;
435 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
436 .lcs_blocking = blocking,
437 .lcs_glimpse = glimpse,
441 LASSERT(!(*flags & LDLM_FL_REPLAY));
442 if (unlikely(ns_is_client(ns))) {
443 CERROR("Trying to enqueue local lock in a shadow namespace\n");
447 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
450 GOTO(out_nolock, err = PTR_ERR(lock));
452 err = ldlm_lvbo_init(lock->l_resource);
454 LDLM_ERROR(lock, "delayed lvb init failed (rc %d)", err);
458 ldlm_lock2handle(lock, lockh);
460 /* NB: we don't have any lock now (lock_res_and_lock)
461 * because it's a new lock */
462 ldlm_lock_addref_internal_nolock(lock, mode);
463 ldlm_set_local(lock);
464 if (*flags & LDLM_FL_ATOMIC_CB)
465 ldlm_set_atomic_cb(lock);
468 lock->l_policy_data = *policy;
469 if (client_cookie != NULL)
470 lock->l_client_cookie = *client_cookie;
471 if (type == LDLM_EXTENT) {
472 /* extent lock without policy is a bug */
476 lock->l_req_extent = policy->l_extent;
479 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
480 if (unlikely(err != ELDLM_OK))
484 *policy = lock->l_policy_data;
486 if (lock->l_completion_ast)
487 lock->l_completion_ast(lock, *flags, NULL);
489 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
492 LDLM_LOCK_RELEASE(lock);
496 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
498 static void failed_lock_cleanup(struct ldlm_namespace *ns,
499 struct ldlm_lock *lock, int mode)
503 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
504 lock_res_and_lock(lock);
505 /* Check that lock is not granted or failed, we might race. */
506 if ((lock->l_req_mode != lock->l_granted_mode) &&
507 !ldlm_is_failed(lock)) {
508 /* Make sure that this lock will not be found by raced
509 * bl_ast and -EINVAL reply is sent to server anyways.
511 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
512 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
515 unlock_res_and_lock(lock);
519 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
520 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
522 LDLM_DEBUG(lock, "lock was granted or failed in race");
524 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
525 * from llite/file.c/ll_file_flock(). */
526 /* This code makes for the fact that we do not have blocking handler on
527 * a client for flock locks. As such this is the place where we must
528 * completely kill failed locks. (interrupted and those that
529 * were waiting to be granted when server evicted us. */
530 if (lock->l_resource->lr_type == LDLM_FLOCK) {
531 lock_res_and_lock(lock);
532 if (!ldlm_is_destroyed(lock)) {
533 ldlm_resource_unlink_lock(lock);
534 ldlm_lock_decref_internal_nolock(lock, mode);
535 ldlm_lock_destroy_nolock(lock);
537 unlock_res_and_lock(lock);
539 ldlm_lock_decref_internal(lock, mode);
544 * Finishing portion of client lock enqueue code.
546 * Called after receiving reply from server.
548 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
549 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
550 __u64 *flags, void *lvb, __u32 lvb_len,
551 struct lustre_handle *lockh,int rc)
553 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
554 int is_replay = *flags & LDLM_FL_REPLAY;
555 struct ldlm_lock *lock;
556 struct ldlm_reply *reply;
557 int cleanup_phase = 1;
560 lock = ldlm_handle2lock(lockh);
561 /* ldlm_cli_enqueue is holding a reference on this lock. */
563 LASSERT(type == LDLM_FLOCK);
567 LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
568 "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
570 if (rc != ELDLM_OK) {
572 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
573 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
575 if (rc != ELDLM_LOCK_ABORTED)
579 /* Before we return, swab the reply */
580 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
582 GOTO(cleanup, rc = -EPROTO);
587 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
590 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
591 GOTO(cleanup, rc = size);
592 } else if (unlikely(size > lvb_len)) {
593 LDLM_ERROR(lock, "Replied LVB is larger than "
594 "expectation, expected = %d, replied = %d",
596 GOTO(cleanup, rc = -EINVAL);
601 if (rc == ELDLM_LOCK_ABORTED) {
602 if (lvb_len > 0 && lvb != NULL)
603 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
605 GOTO(cleanup, rc = rc ? : ELDLM_LOCK_ABORTED);
608 /* lock enqueued on the server */
611 lock_res_and_lock(lock);
612 /* Key change rehash lock in per-export hash with new key */
613 if (exp->exp_lock_hash) {
614 /* In the function below, .hs_keycmp resolves to
615 * ldlm_export_lock_keycmp() */
616 /* coverity[overrun-buffer-val] */
617 cfs_hash_rehash_key(exp->exp_lock_hash,
618 &lock->l_remote_handle,
622 lock->l_remote_handle = reply->lock_handle;
625 *flags = ldlm_flags_from_wire(reply->lock_flags);
626 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
627 LDLM_FL_INHERIT_MASK);
628 unlock_res_and_lock(lock);
630 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: "LPX64"\n",
631 lock, reply->lock_handle.cookie, *flags);
633 /* If enqueue returned a blocked lock but the completion handler has
634 * already run, then it fixed up the resource and we don't need to do it
636 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
637 int newmode = reply->lock_desc.l_req_mode;
639 if (newmode && newmode != lock->l_req_mode) {
640 LDLM_DEBUG(lock, "server returned different mode %s",
641 ldlm_lockname[newmode]);
642 lock->l_req_mode = newmode;
645 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
646 &lock->l_resource->lr_name)) {
647 CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES
648 " instead of "DLDLMRES"\n",
649 PLDLMRES(&reply->lock_desc.l_resource),
650 PLDLMRES(lock->l_resource));
652 rc = ldlm_lock_change_resource(ns, lock,
653 &reply->lock_desc.l_resource.lr_name);
654 if (rc || lock->l_resource == NULL)
655 GOTO(cleanup, rc = -ENOMEM);
656 LDLM_DEBUG(lock, "client-side enqueue, new resource");
659 if (!(type == LDLM_IBITS &&
660 !(exp_connect_flags(exp) & OBD_CONNECT_IBITS)))
661 /* We assume lock type cannot change on server*/
662 ldlm_convert_policy_to_local(exp,
663 lock->l_resource->lr_type,
664 &reply->lock_desc.l_policy_data,
665 &lock->l_policy_data);
666 if (type != LDLM_PLAIN)
667 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
670 if ((*flags) & LDLM_FL_AST_SENT) {
671 lock_res_and_lock(lock);
672 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
673 unlock_res_and_lock(lock);
674 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
677 /* If the lock has already been granted by a completion AST, don't
678 * clobber the LVB with an older one. */
680 /* We must lock or a racing completion might update lvb without
681 * letting us know and we'll clobber the correct value.
682 * Cannot unlock after the check either, a that still leaves
683 * a tiny window for completion to get in */
684 lock_res_and_lock(lock);
685 if (lock->l_req_mode != lock->l_granted_mode)
686 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
687 lock->l_lvb_data, lvb_len);
688 unlock_res_and_lock(lock);
696 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
697 if (lock->l_completion_ast != NULL) {
698 int err = lock->l_completion_ast(lock, *flags, NULL);
706 if (lvb_len > 0 && lvb != NULL) {
707 /* Copy the LVB here, and not earlier, because the completion
708 * AST (if any) can override what we got in the reply */
709 memcpy(lvb, lock->l_lvb_data, lvb_len);
712 LDLM_DEBUG(lock, "client-side enqueue END");
715 if (cleanup_phase == 1 && rc)
716 failed_lock_cleanup(ns, lock, mode);
717 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
719 LDLM_LOCK_RELEASE(lock);
722 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
725 * Estimate number of lock handles that would fit into request of given
726 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
727 * a single page on the send/receive side. XXX: 512 should be changed to
728 * more adequate value.
730 static inline int ldlm_req_handles_avail(int req_size, int off)
734 avail = min_t(int, LDLM_MAXREQSIZE, PAGE_CACHE_SIZE - 512) - req_size;
735 if (likely(avail >= 0))
736 avail /= (int)sizeof(struct lustre_handle);
739 avail += LDLM_LOCKREQ_HANDLES - off;
744 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
745 enum req_location loc,
748 __u32 size = req_capsule_msg_size(pill, loc);
749 return ldlm_req_handles_avail(size, off);
752 static inline int ldlm_format_handles_avail(struct obd_import *imp,
753 const struct req_format *fmt,
754 enum req_location loc, int off)
756 __u32 size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
757 return ldlm_req_handles_avail(size, off);
761 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
762 * \a count locks in \a cancels.
764 * This is to be called by functions preparing their own requests that
765 * might contain lists of locks to cancel in addition to actual operation
766 * that needs to be performed.
768 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
769 int version, int opc, int canceloff,
770 struct list_head *cancels, int count)
772 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
773 struct req_capsule *pill = &req->rq_pill;
774 struct ldlm_request *dlm = NULL;
775 struct list_head head = LIST_HEAD_INIT(head);
776 int flags, avail, to_free, pack = 0;
782 if (ns_connect_cancelset(ns)) {
783 /* Estimate the amount of available space in the request. */
784 req_capsule_filled_sizes(pill, RCL_CLIENT);
785 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
787 flags = ns_connect_lru_resize(ns) ?
788 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
789 to_free = !ns_connect_lru_resize(ns) &&
790 opc == LDLM_ENQUEUE ? 1 : 0;
792 /* Cancel LRU locks here _only_ if the server supports
793 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
794 * RPC, which will make us slower. */
796 count += ldlm_cancel_lru_local(ns, cancels, to_free,
797 avail - count, 0, flags);
802 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
803 ldlm_request_bufsize(pack, opc));
806 rc = ptlrpc_request_pack(req, version, opc);
808 ldlm_lock_list_put(cancels, l_bl_ast, count);
812 if (ns_connect_cancelset(ns)) {
814 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
816 /* Skip first lock handler in ldlm_request_pack(),
817 * this method will incrment @lock_count according
818 * to the lock handle amount actually written to
820 dlm->lock_count = canceloff;
822 /* Pack into the request @pack lock handles. */
823 ldlm_cli_cancel_list(cancels, pack, req, 0);
824 /* Prepare and send separate cancel RPC for others. */
825 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
827 ldlm_lock_list_put(cancels, l_bl_ast, count);
831 EXPORT_SYMBOL(ldlm_prep_elc_req);
833 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
834 struct list_head *cancels, int count)
836 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
837 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
839 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
841 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
843 struct ptlrpc_request *req;
847 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
849 RETURN(ERR_PTR(-ENOMEM));
851 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
853 ptlrpc_request_free(req);
857 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
858 ptlrpc_request_set_replen(req);
861 EXPORT_SYMBOL(ldlm_enqueue_pack);
864 * Client-side lock enqueue.
866 * If a request has some specific initialisation it is passed in \a reqp,
867 * otherwise it is created in ldlm_cli_enqueue.
869 * Supports sync and async requests, pass \a async flag accordingly. If a
870 * request was created in ldlm_cli_enqueue and it is the async request,
871 * pass it to the caller in \a reqp.
873 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
874 struct ldlm_enqueue_info *einfo,
875 const struct ldlm_res_id *res_id,
876 ldlm_policy_data_t const *policy, __u64 *flags,
877 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
878 struct lustre_handle *lockh, int async)
880 struct ldlm_namespace *ns;
881 struct ldlm_lock *lock;
882 struct ldlm_request *body;
883 int is_replay = *flags & LDLM_FL_REPLAY;
884 int req_passed_in = 1;
886 struct ptlrpc_request *req;
889 LASSERT(exp != NULL);
891 ns = exp->exp_obd->obd_namespace;
893 /* If we're replaying this lock, just check some invariants.
894 * If we're creating a new lock, get everything all setup nice. */
896 lock = ldlm_handle2lock_long(lockh, 0);
897 LASSERT(lock != NULL);
898 LDLM_DEBUG(lock, "client-side enqueue START");
899 LASSERT(exp == lock->l_conn_export);
901 const struct ldlm_callback_suite cbs = {
902 .lcs_completion = einfo->ei_cb_cp,
903 .lcs_blocking = einfo->ei_cb_bl,
904 .lcs_glimpse = einfo->ei_cb_gl
906 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
907 einfo->ei_mode, &cbs, einfo->ei_cbdata,
910 RETURN(PTR_ERR(lock));
911 /* for the local lock, add the reference */
912 ldlm_lock_addref_internal(lock, einfo->ei_mode);
913 ldlm_lock2handle(lock, lockh);
915 lock->l_policy_data = *policy;
917 if (einfo->ei_type == LDLM_EXTENT) {
918 /* extent lock without policy is a bug */
922 lock->l_req_extent = policy->l_extent;
924 LDLM_DEBUG(lock, "client-side enqueue START, flags "LPX64"\n",
928 lock->l_conn_export = exp;
929 lock->l_export = NULL;
930 lock->l_blocking_ast = einfo->ei_cb_bl;
931 lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
932 lock->l_last_activity = cfs_time_current_sec();
934 /* lock not sent to server yet */
936 if (reqp == NULL || *reqp == NULL) {
937 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
942 failed_lock_cleanup(ns, lock, einfo->ei_mode);
943 LDLM_LOCK_RELEASE(lock);
953 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
955 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
956 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
959 /* Dump lock data into the request buffer */
960 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
961 ldlm_lock2desc(lock, &body->lock_desc);
962 body->lock_flags = ldlm_flags_to_wire(*flags);
963 body->lock_handle[0] = *lockh;
965 /* Continue as normal. */
966 if (!req_passed_in) {
968 req_capsule_extend(&req->rq_pill,
969 &RQF_LDLM_ENQUEUE_LVB);
970 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
972 ptlrpc_request_set_replen(req);
976 LASSERT(reqp != NULL);
980 LDLM_DEBUG(lock, "sending request");
982 rc = ptlrpc_queue_wait(req);
984 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
985 einfo->ei_mode, flags, lvb, lvb_len,
988 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
989 * one reference that we took */
991 LDLM_LOCK_RELEASE(lock);
995 if (!req_passed_in && req != NULL) {
996 ptlrpc_req_finished(req);
1003 EXPORT_SYMBOL(ldlm_cli_enqueue);
1005 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
1008 struct ldlm_resource *res;
1011 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1012 CERROR("Trying to cancel local lock\n");
1015 LDLM_DEBUG(lock, "client-side local convert");
1017 res = ldlm_lock_convert(lock, new_mode, flags);
1019 ldlm_reprocess_all(res);
1022 rc = LUSTRE_EDEADLK;
1024 LDLM_DEBUG(lock, "client-side local convert handler END");
1025 LDLM_LOCK_PUT(lock);
1029 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
1030 * conversion of locks which are on the waiting or converting queue */
1031 /* Caller of this code is supposed to take care of lock readers/writers
1033 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
1035 struct ldlm_request *body;
1036 struct ldlm_reply *reply;
1037 struct ldlm_lock *lock;
1038 struct ldlm_resource *res;
1039 struct ptlrpc_request *req;
1043 lock = ldlm_handle2lock(lockh);
1050 if (lock->l_conn_export == NULL)
1051 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
1053 LDLM_DEBUG(lock, "client-side convert");
1055 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
1056 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1059 LDLM_LOCK_PUT(lock);
1063 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1064 body->lock_handle[0] = lock->l_remote_handle;
1066 body->lock_desc.l_req_mode = new_mode;
1067 body->lock_flags = ldlm_flags_to_wire(*flags);
1070 ptlrpc_request_set_replen(req);
1071 rc = ptlrpc_queue_wait(req);
1075 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1077 GOTO(out, rc = -EPROTO);
1080 GOTO(out, rc = req->rq_status);
1082 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
1084 ldlm_reprocess_all(res);
1085 /* Go to sleep until the lock is granted. */
1086 /* FIXME: or cancelled. */
1087 if (lock->l_completion_ast) {
1088 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1094 rc = LUSTRE_EDEADLK;
1098 LDLM_LOCK_PUT(lock);
1099 ptlrpc_req_finished(req);
1102 EXPORT_SYMBOL(ldlm_cli_convert);
1105 * Cancel locks locally.
1107 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1108 * \retval LDLM_FL_CANCELING otherwise;
1109 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1111 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1113 __u64 rc = LDLM_FL_LOCAL_ONLY;
1116 if (lock->l_conn_export) {
1119 LDLM_DEBUG(lock, "client-side cancel");
1120 /* Set this flag to prevent others from getting new references*/
1121 lock_res_and_lock(lock);
1122 ldlm_set_cbpending(lock);
1123 local_only = !!(lock->l_flags &
1124 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1125 ldlm_cancel_callback(lock);
1126 rc = (ldlm_is_bl_ast(lock)) ?
1127 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1128 unlock_res_and_lock(lock);
1131 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1133 rc = LDLM_FL_LOCAL_ONLY;
1135 ldlm_lock_cancel(lock);
1137 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1138 LDLM_ERROR(lock, "Trying to cancel local lock");
1141 LDLM_DEBUG(lock, "server-side local cancel");
1142 ldlm_lock_cancel(lock);
1143 ldlm_reprocess_all(lock->l_resource);
1150 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1152 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1153 struct list_head *head, int count)
1155 struct ldlm_request *dlm;
1156 struct ldlm_lock *lock;
1157 int max, packed = 0;
1160 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1161 LASSERT(dlm != NULL);
1163 /* Check the room in the request buffer. */
1164 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1165 sizeof(struct ldlm_request);
1166 max /= sizeof(struct lustre_handle);
1167 max += LDLM_LOCKREQ_HANDLES;
1168 LASSERT(max >= dlm->lock_count + count);
1170 /* XXX: it would be better to pack lock handles grouped by resource.
1171 * so that the server cancel would call filter_lvbo_update() less
1173 list_for_each_entry(lock, head, l_bl_ast) {
1176 LASSERT(lock->l_conn_export);
1177 /* Pack the lock handle to the given request buffer. */
1178 LDLM_DEBUG(lock, "packing");
1179 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1182 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1187 * Prepare and send a batched cancel RPC. It will include \a count lock
1188 * handles of locks given in \a cancels list. */
1189 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1190 int count, ldlm_cancel_flags_t flags)
1192 struct ptlrpc_request *req = NULL;
1193 struct obd_import *imp;
1198 LASSERT(exp != NULL);
1201 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1203 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1206 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1207 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1212 imp = class_exp2cliimp(exp);
1213 if (imp == NULL || imp->imp_invalid) {
1215 "skipping cancel on invalid import %p\n", imp);
1219 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1221 GOTO(out, rc = -ENOMEM);
1223 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1224 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1225 ldlm_request_bufsize(count, LDLM_CANCEL));
1227 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1229 ptlrpc_request_free(req);
1233 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1234 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1235 ptlrpc_at_set_req_timeout(req);
1237 ldlm_cancel_pack(req, cancels, count);
1239 ptlrpc_request_set_replen(req);
1240 if (flags & LCF_ASYNC) {
1241 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1245 rc = ptlrpc_queue_wait(req);
1247 if (rc == LUSTRE_ESTALE) {
1248 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1249 "out of sync -- not fatal\n",
1250 libcfs_nid2str(req->rq_import->
1251 imp_connection->c_peer.nid));
1253 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1254 req->rq_import_generation == imp->imp_generation) {
1255 ptlrpc_req_finished(req);
1257 } else if (rc != ELDLM_OK) {
1258 /* -ESHUTDOWN is common on umount */
1259 CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1260 "Got rc %d from cancel RPC: "
1261 "canceling anyway\n", rc);
1268 ptlrpc_req_finished(req);
1271 return sent ? sent : rc;
1273 EXPORT_SYMBOL(ldlm_cli_cancel_req);
1275 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1277 LASSERT(imp != NULL);
1278 return &imp->imp_obd->obd_namespace->ns_pool;
1282 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1284 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1286 struct obd_device *obd;
1290 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1291 !imp_connect_lru_resize(req->rq_import)))
1294 * Do nothing for corner cases.
1299 /* In some cases RPC may contain SLV and limit zeroed out. This
1300 * is the case when server does not support LRU resize feature.
1301 * This is also possible in some recovery cases when server-side
1302 * reqs have no reference to the OBD export and thus access to
1303 * server-side namespace is not possible. */
1304 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1305 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1306 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1307 "(SLV: "LPU64", Limit: %u)",
1308 lustre_msg_get_slv(req->rq_repmsg),
1309 lustre_msg_get_limit(req->rq_repmsg));
1313 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1314 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1315 obd = req->rq_import->imp_obd;
1317 /* Set new SLV and limit in OBD fields to make them accessible
1318 * to the pool thread. We do not access obd_namespace and pool
1319 * directly here as there is no reliable way to make sure that
1320 * they are still alive at cleanup time. Evil races are possible
1321 * which may cause Oops at that time. */
1322 write_lock(&obd->obd_pool_lock);
1323 obd->obd_pool_slv = new_slv;
1324 obd->obd_pool_limit = new_limit;
1325 write_unlock(&obd->obd_pool_lock);
1329 EXPORT_SYMBOL(ldlm_cli_update_pool);
1332 * Client side lock cancel.
1334 * Lock must not have any readers or writers by this time.
1336 int ldlm_cli_cancel(struct lustre_handle *lockh,
1337 ldlm_cancel_flags_t cancel_flags)
1339 struct obd_export *exp;
1340 int avail, flags, count = 1;
1342 struct ldlm_namespace *ns;
1343 struct ldlm_lock *lock;
1344 struct list_head cancels = LIST_HEAD_INIT(cancels);
1347 /* concurrent cancels on the same handle can happen */
1348 lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1350 LDLM_DEBUG_NOLOCK("lock is already being destroyed");
1354 rc = ldlm_cli_cancel_local(lock);
1355 if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
1356 LDLM_LOCK_RELEASE(lock);
1359 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1360 * RPC which goes to canceld portal, so we can cancel other LRU locks
1361 * here and send them all as one LDLM_CANCEL RPC. */
1362 LASSERT(list_empty(&lock->l_bl_ast));
1363 list_add(&lock->l_bl_ast, &cancels);
1365 exp = lock->l_conn_export;
1366 if (exp_connect_cancelset(exp)) {
1367 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1372 ns = ldlm_lock_to_ns(lock);
1373 flags = ns_connect_lru_resize(ns) ?
1374 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1375 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1378 ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1381 EXPORT_SYMBOL(ldlm_cli_cancel);
1384 * Locally cancel up to \a count locks in list \a cancels.
1385 * Return the number of cancelled locks.
1387 int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
1388 ldlm_cancel_flags_t flags)
1390 struct list_head head = LIST_HEAD_INIT(head);
1391 struct ldlm_lock *lock, *next;
1392 int left = 0, bl_ast = 0;
1396 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1400 if (flags & LCF_LOCAL) {
1401 rc = LDLM_FL_LOCAL_ONLY;
1402 ldlm_lock_cancel(lock);
1404 rc = ldlm_cli_cancel_local(lock);
1406 /* Until we have compound requests and can send LDLM_CANCEL
1407 * requests batched with generic RPCs, we need to send cancels
1408 * with the LDLM_FL_BL_AST flag in a separate RPC from
1409 * the one being generated now. */
1410 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1411 LDLM_DEBUG(lock, "Cancel lock separately");
1412 list_del_init(&lock->l_bl_ast);
1413 list_add(&lock->l_bl_ast, &head);
1417 if (rc == LDLM_FL_LOCAL_ONLY) {
1418 /* CANCEL RPC should not be sent to server. */
1419 list_del_init(&lock->l_bl_ast);
1420 LDLM_LOCK_RELEASE(lock);
1426 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1431 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
1434 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1435 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1436 * readahead requests, ...)
1438 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1439 struct ldlm_lock *lock,
1440 int unused, int added,
1443 ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1445 /* don't check added & count since we want to process all locks
1447 * It's fine to not take lock to access lock->l_resource since
1448 * the lock has already been granted so it won't change. */
1449 switch (lock->l_resource->lr_type) {
1452 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) != 0)
1455 result = LDLM_POLICY_SKIP_LOCK;
1456 lock_res_and_lock(lock);
1457 ldlm_set_skipped(lock);
1458 unlock_res_and_lock(lock);
1466 * Callback function for LRU-resize policy. Decides whether to keep
1467 * \a lock in LRU for current \a LRU size \a unused, added in current
1468 * scan \a added and number of locks to be preferably canceled \a count.
1470 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1472 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1474 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1475 struct ldlm_lock *lock,
1476 int unused, int added,
1479 cfs_time_t cur = cfs_time_current();
1480 struct ldlm_pool *pl = &ns->ns_pool;
1484 /* Stop LRU processing when we reach past @count or have checked all
1486 if (count && added >= count)
1487 return LDLM_POLICY_KEEP_LOCK;
1489 slv = ldlm_pool_get_slv(pl);
1490 lvf = ldlm_pool_get_lvf(pl);
1491 la = cfs_duration_sec(cfs_time_sub(cur,
1492 lock->l_last_used));
1493 lv = lvf * la * unused;
1495 /* Inform pool about current CLV to see it via proc. */
1496 ldlm_pool_set_clv(pl, lv);
1498 /* Stop when SLV is not yet come from server or lv is smaller than
1500 if (slv == 0 || lv < slv)
1501 return LDLM_POLICY_KEEP_LOCK;
1503 return LDLM_POLICY_CANCEL_LOCK;
1507 * Callback function for proc used policy. Makes decision whether to keep
1508 * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1509 * added and number of locks to be preferably canceled \a count.
1511 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1513 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1515 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1516 struct ldlm_lock *lock,
1517 int unused, int added,
1520 /* Stop LRU processing when we reach past @count or have checked all
1522 return (added >= count) ?
1523 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1527 * Callback function for aged policy. Makes decision whether to keep \a lock in
1528 * LRU for current LRU size \a unused, added in current scan \a added and
1529 * number of locks to be preferably canceled \a count.
1531 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1533 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1535 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1536 struct ldlm_lock *lock,
1537 int unused, int added,
1540 if ((added >= count) &&
1541 cfs_time_before(cfs_time_current(),
1542 cfs_time_add(lock->l_last_used, ns->ns_max_age)))
1543 return LDLM_POLICY_KEEP_LOCK;
1545 return LDLM_POLICY_CANCEL_LOCK;
1549 * Callback function for default policy. Makes decision whether to keep \a lock
1550 * in LRU for current LRU size \a unused, added in current scan \a added and
1551 * number of locks to be preferably canceled \a count.
1553 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1555 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1557 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1558 struct ldlm_lock *lock,
1559 int unused, int added,
1562 /* Stop LRU processing when we reach past count or have checked all
1564 return (added >= count) ?
1565 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1568 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1569 struct ldlm_lock *, int,
1572 static ldlm_cancel_lru_policy_t
1573 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1575 if (flags & LDLM_CANCEL_NO_WAIT)
1576 return ldlm_cancel_no_wait_policy;
1578 if (ns_connect_lru_resize(ns)) {
1579 if (flags & LDLM_CANCEL_SHRINK)
1580 /* We kill passed number of old locks. */
1581 return ldlm_cancel_passed_policy;
1582 else if (flags & LDLM_CANCEL_LRUR)
1583 return ldlm_cancel_lrur_policy;
1584 else if (flags & LDLM_CANCEL_PASSED)
1585 return ldlm_cancel_passed_policy;
1587 if (flags & LDLM_CANCEL_AGED)
1588 return ldlm_cancel_aged_policy;
1591 return ldlm_cancel_default_policy;
1595 * - Free space in LRU for \a count new locks,
1596 * redundant unused locks are canceled locally;
1597 * - also cancel locally unused aged locks;
1598 * - do not cancel more than \a max locks;
1599 * - GET the found locks and add them into the \a cancels list.
1601 * A client lock can be added to the l_bl_ast list only when it is
1602 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1603 * CANCEL. There are the following use cases:
1604 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1605 * ldlm_cli_cancel(), which check and set this flag properly. As any
1606 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1607 * later without any special locking.
1609 * Calling policies for enabled LRU resize:
1610 * ----------------------------------------
1611 * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to
1612 * cancel not more than \a count locks;
1614 * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at
1615 * the beginning of LRU list);
1617 * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to
1618 * memory pressre policy function;
1620 * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy".
1622 * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1623 * (typically before replaying locks) w/o
1624 * sending any RPCs or waiting for any
1625 * outstanding RPC to complete.
1627 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns,
1628 struct list_head *cancels, int count, int max,
1631 ldlm_cancel_lru_policy_t pf;
1632 struct ldlm_lock *lock, *next;
1633 int added = 0, unused, remained;
1636 spin_lock(&ns->ns_lock);
1637 unused = ns->ns_nr_unused;
1640 if (!ns_connect_lru_resize(ns))
1641 count += unused - ns->ns_max_unused;
1643 pf = ldlm_cancel_lru_policy(ns, flags);
1644 LASSERT(pf != NULL);
1646 while (!list_empty(&ns->ns_unused_list)) {
1647 ldlm_policy_res_t result;
1649 /* all unused locks */
1650 if (remained-- <= 0)
1653 /* For any flags, stop scanning if @max is reached. */
1654 if (max && added >= max)
1657 list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1659 /* No locks which got blocking requests. */
1660 LASSERT(!ldlm_is_bl_ast(lock));
1662 if (flags & LDLM_CANCEL_NO_WAIT &&
1663 ldlm_is_skipped(lock))
1664 /* already processed */
1667 /* Somebody is already doing CANCEL. No need for this
1668 * lock in LRU, do not traverse it again. */
1669 if (!ldlm_is_canceling(lock))
1672 ldlm_lock_remove_from_lru_nolock(lock);
1674 if (&lock->l_lru == &ns->ns_unused_list)
1677 LDLM_LOCK_GET(lock);
1678 spin_unlock(&ns->ns_lock);
1679 lu_ref_add(&lock->l_reference, __FUNCTION__, current);
1681 /* Pass the lock through the policy filter and see if it
1682 * should stay in LRU.
1684 * Even for shrinker policy we stop scanning if
1685 * we find a lock that should stay in the cache.
1686 * We should take into account lock age anyway
1687 * as a new lock is a valuable resource even if
1688 * it has a low weight.
1690 * That is, for shrinker policy we drop only
1691 * old locks, but additionally choose them by
1692 * their weight. Big extent locks will stay in
1694 result = pf(ns, lock, unused, added, count);
1695 if (result == LDLM_POLICY_KEEP_LOCK) {
1696 lu_ref_del(&lock->l_reference,
1697 __FUNCTION__, current);
1698 LDLM_LOCK_RELEASE(lock);
1699 spin_lock(&ns->ns_lock);
1702 if (result == LDLM_POLICY_SKIP_LOCK) {
1703 lu_ref_del(&lock->l_reference,
1705 LDLM_LOCK_RELEASE(lock);
1706 spin_lock(&ns->ns_lock);
1710 lock_res_and_lock(lock);
1711 /* Check flags again under the lock. */
1712 if (ldlm_is_canceling(lock) ||
1713 (ldlm_lock_remove_from_lru(lock) == 0)) {
1714 /* Another thread is removing lock from LRU, or
1715 * somebody is already doing CANCEL, or there
1716 * is a blocking request which will send cancel
1717 * by itself, or the lock is no longer unused. */
1718 unlock_res_and_lock(lock);
1719 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1720 LDLM_LOCK_RELEASE(lock);
1721 spin_lock(&ns->ns_lock);
1724 LASSERT(!lock->l_readers && !lock->l_writers);
1726 /* If we have chosen to cancel this lock voluntarily, we
1727 * better send cancel notification to server, so that it
1728 * frees appropriate state. This might lead to a race
1729 * where while we are doing cancel here, server is also
1730 * silently cancelling this lock. */
1731 ldlm_clear_cancel_on_block(lock);
1733 /* Setting the CBPENDING flag is a little misleading,
1734 * but prevents an important race; namely, once
1735 * CBPENDING is set, the lock can accumulate no more
1736 * readers/writers. Since readers and writers are
1737 * already zero here, ldlm_lock_decref() won't see
1738 * this flag and call l_blocking_ast */
1739 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1741 /* We can't re-add to l_lru as it confuses the
1742 * refcounting in ldlm_lock_remove_from_lru() if an AST
1743 * arrives after we drop lr_lock below. We use l_bl_ast
1744 * and can't use l_pending_chain as it is used both on
1745 * server and client nevertheless bug 5666 says it is
1746 * used only on server */
1747 LASSERT(list_empty(&lock->l_bl_ast));
1748 list_add(&lock->l_bl_ast, cancels);
1749 unlock_res_and_lock(lock);
1750 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1751 spin_lock(&ns->ns_lock);
1755 spin_unlock(&ns->ns_lock);
1759 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1760 int count, int max, ldlm_cancel_flags_t cancel_flags,
1764 added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1767 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1771 * Cancel at least \a nr locks from given namespace LRU.
1773 * When called with LCF_ASYNC the blocking callback will be handled
1774 * in a thread and this function will return after the thread has been
1775 * asked to call the callback. When called with LCF_ASYNC the blocking
1776 * callback will be performed in this function.
1778 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
1779 ldlm_cancel_flags_t cancel_flags,
1782 struct list_head cancels = LIST_HEAD_INIT(cancels);
1786 /* Just prepare the list of locks, do not actually cancel them yet.
1787 * Locks are cancelled later in a separate thread. */
1788 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1789 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
1797 * Find and cancel locally unused locks found on resource, matched to the
1798 * given policy, mode. GET the found locks and add them into the \a cancels
1801 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1802 struct list_head *cancels,
1803 ldlm_policy_data_t *policy,
1804 ldlm_mode_t mode, __u64 lock_flags,
1805 ldlm_cancel_flags_t cancel_flags, void *opaque)
1807 struct ldlm_lock *lock;
1812 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1813 if (opaque != NULL && lock->l_ast_data != opaque) {
1814 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1815 lock->l_ast_data, opaque);
1820 if (lock->l_readers || lock->l_writers)
1823 /* If somebody is already doing CANCEL, or blocking AST came,
1824 * skip this lock. */
1825 if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock))
1828 if (lockmode_compat(lock->l_granted_mode, mode))
1831 /* If policy is given and this is IBITS lock, add to list only
1832 * those locks that match by policy. */
1833 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1834 !(lock->l_policy_data.l_inodebits.bits &
1835 policy->l_inodebits.bits))
1838 /* See CBPENDING comment in ldlm_cancel_lru */
1839 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1842 LASSERT(list_empty(&lock->l_bl_ast));
1843 list_add(&lock->l_bl_ast, cancels);
1844 LDLM_LOCK_GET(lock);
1849 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
1851 EXPORT_SYMBOL(ldlm_cancel_resource_local);
1854 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1856 * If \a req is NULL, send CANCEL request to server with handles of locks
1857 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1858 * separately per lock.
1859 * If \a req is not NULL, put handles of locks in \a cancels into the request
1860 * buffer at the offset \a off.
1861 * Destroy \a cancels at the end.
1863 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1864 struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1866 struct ldlm_lock *lock;
1870 if (list_empty(cancels) || count == 0)
1873 /* XXX: requests (both batched and not) could be sent in parallel.
1874 * Usually it is enough to have just 1 RPC, but it is possible that
1875 * there are too many locks to be cancelled in LRU or on a resource.
1876 * It would also speed up the case when the server does not support
1879 LASSERT(!list_empty(cancels));
1880 lock = list_entry(cancels->next, struct ldlm_lock,
1882 LASSERT(lock->l_conn_export);
1884 if (exp_connect_cancelset(lock->l_conn_export)) {
1887 ldlm_cancel_pack(req, cancels, count);
1889 res = ldlm_cli_cancel_req(lock->l_conn_export,
1893 res = ldlm_cli_cancel_req(lock->l_conn_export,
1898 CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1899 "ldlm_cli_cancel_list: %d\n", res);
1904 ldlm_lock_list_put(cancels, l_bl_ast, res);
1906 LASSERT(count == 0);
1909 EXPORT_SYMBOL(ldlm_cli_cancel_list);
1912 * Cancel all locks on a resource that have 0 readers/writers.
1914 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1915 * to notify the server. */
1916 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1917 const struct ldlm_res_id *res_id,
1918 ldlm_policy_data_t *policy,
1920 ldlm_cancel_flags_t flags,
1923 struct ldlm_resource *res;
1924 struct list_head cancels = LIST_HEAD_INIT(cancels);
1929 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1931 /* This is not a problem. */
1932 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1936 LDLM_RESOURCE_ADDREF(res);
1937 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1938 0, flags | LCF_BL_AST, opaque);
1939 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1941 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
1944 LDLM_RESOURCE_DELREF(res);
1945 ldlm_resource_putref(res);
1948 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
1950 struct ldlm_cli_cancel_arg {
1955 static int ldlm_cli_hash_cancel_unused(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1956 struct hlist_node *hnode, void *arg)
1958 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1959 struct ldlm_cli_cancel_arg *lc = arg;
1961 ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1962 NULL, LCK_MINMODE, lc->lc_flags,
1964 /* must return 0 for hash iteration */
1969 * Cancel all locks on a namespace (or a specific resource, if given)
1970 * that have 0 readers/writers.
1972 * If flags & LCF_LOCAL, throw the locks away without trying
1973 * to notify the server. */
1974 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1975 const struct ldlm_res_id *res_id,
1976 ldlm_cancel_flags_t flags, void *opaque)
1978 struct ldlm_cli_cancel_arg arg = {
1980 .lc_opaque = opaque,
1988 if (res_id != NULL) {
1989 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1993 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1994 ldlm_cli_hash_cancel_unused, &arg);
1998 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
2000 /* Lock iterators. */
2002 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
2005 struct list_head *tmp, *next;
2006 struct ldlm_lock *lock;
2007 int rc = LDLM_ITER_CONTINUE;
2012 RETURN(LDLM_ITER_CONTINUE);
2015 list_for_each_safe(tmp, next, &res->lr_granted) {
2016 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2018 if (iter(lock, closure) == LDLM_ITER_STOP)
2019 GOTO(out, rc = LDLM_ITER_STOP);
2022 list_for_each_safe(tmp, next, &res->lr_converting) {
2023 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2025 if (iter(lock, closure) == LDLM_ITER_STOP)
2026 GOTO(out, rc = LDLM_ITER_STOP);
2029 list_for_each_safe(tmp, next, &res->lr_waiting) {
2030 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2032 if (iter(lock, closure) == LDLM_ITER_STOP)
2033 GOTO(out, rc = LDLM_ITER_STOP);
2039 EXPORT_SYMBOL(ldlm_resource_foreach);
2041 struct iter_helper_data {
2042 ldlm_iterator_t iter;
2046 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2048 struct iter_helper_data *helper = closure;
2049 return helper->iter(lock, helper->closure);
2052 static int ldlm_res_iter_helper(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2053 struct hlist_node *hnode, void *arg)
2056 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2058 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2062 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2063 ldlm_iterator_t iter, void *closure)
2066 struct iter_helper_data helper = { .iter = iter, .closure = closure };
2068 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2069 ldlm_res_iter_helper, &helper);
2072 EXPORT_SYMBOL(ldlm_namespace_foreach);
2074 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2075 * return 0: find no resource
2076 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2079 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2080 const struct ldlm_res_id *res_id,
2081 ldlm_iterator_t iter, void *data)
2083 struct ldlm_resource *res;
2087 LASSERTF(ns != NULL, "must pass in namespace\n");
2089 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2093 LDLM_RESOURCE_ADDREF(res);
2094 rc = ldlm_resource_foreach(res, iter, data);
2095 LDLM_RESOURCE_DELREF(res);
2096 ldlm_resource_putref(res);
2099 EXPORT_SYMBOL(ldlm_resource_iterate);
2103 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2105 struct list_head *list = closure;
2107 /* we use l_pending_chain here, because it's unused on clients. */
2108 LASSERTF(list_empty(&lock->l_pending_chain),
2109 "lock %p next %p prev %p\n",
2110 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2111 /* bug 9573: don't replay locks left after eviction, or
2112 * bug 17614: locks being actively cancelled. Get a reference
2113 * on a lock so that it does not disapear under us (e.g. due to cancel)
2115 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
2116 list_add(&lock->l_pending_chain, list);
2117 LDLM_LOCK_GET(lock);
2120 return LDLM_ITER_CONTINUE;
2123 static int replay_lock_interpret(const struct lu_env *env,
2124 struct ptlrpc_request *req,
2125 struct ldlm_async_args *aa, int rc)
2127 struct ldlm_lock *lock;
2128 struct ldlm_reply *reply;
2129 struct obd_export *exp;
2132 atomic_dec(&req->rq_import->imp_replay_inflight);
2136 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2138 GOTO(out, rc = -EPROTO);
2140 lock = ldlm_handle2lock(&aa->lock_handle);
2142 CERROR("received replay ack for unknown local cookie "LPX64
2143 " remote cookie "LPX64 " from server %s id %s\n",
2144 aa->lock_handle.cookie, reply->lock_handle.cookie,
2145 req->rq_export->exp_client_uuid.uuid,
2146 libcfs_id2str(req->rq_peer));
2147 GOTO(out, rc = -ESTALE);
2150 /* Key change rehash lock in per-export hash with new key */
2151 exp = req->rq_export;
2152 if (exp && exp->exp_lock_hash) {
2153 /* In the function below, .hs_keycmp resolves to
2154 * ldlm_export_lock_keycmp() */
2155 /* coverity[overrun-buffer-val] */
2156 cfs_hash_rehash_key(exp->exp_lock_hash,
2157 &lock->l_remote_handle,
2158 &reply->lock_handle,
2161 lock->l_remote_handle = reply->lock_handle;
2164 LDLM_DEBUG(lock, "replayed lock:");
2165 ptlrpc_import_recovery_state_machine(req->rq_import);
2166 LDLM_LOCK_PUT(lock);
2169 ptlrpc_connect_import(req->rq_import);
2174 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2176 struct ptlrpc_request *req;
2177 struct ldlm_async_args *aa;
2178 struct ldlm_request *body;
2183 /* Bug 11974: Do not replay a lock which is actively being canceled */
2184 if (ldlm_is_canceling(lock)) {
2185 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2189 /* If this is reply-less callback lock, we cannot replay it, since
2190 * server might have long dropped it, but notification of that event was
2191 * lost by network. (and server granted conflicting lock already) */
2192 if (ldlm_is_cancel_on_block(lock)) {
2193 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2194 ldlm_lock_cancel(lock);
2199 * If granted mode matches the requested mode, this lock is granted.
2201 * If they differ, but we have a granted mode, then we were granted
2202 * one mode and now want another: ergo, converting.
2204 * If we haven't been granted anything and are on a resource list,
2205 * then we're blocked/waiting.
2207 * If we haven't been granted anything and we're NOT on a resource list,
2208 * then we haven't got a reply yet and don't have a known disposition.
2209 * This happens whenever a lock enqueue is the request that triggers
2212 if (lock->l_granted_mode == lock->l_req_mode)
2213 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2214 else if (lock->l_granted_mode)
2215 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2216 else if (!list_empty(&lock->l_res_link))
2217 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2219 flags = LDLM_FL_REPLAY;
2221 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2222 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2226 /* We're part of recovery, so don't wait for it. */
2227 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2229 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2230 ldlm_lock2desc(lock, &body->lock_desc);
2231 body->lock_flags = ldlm_flags_to_wire(flags);
2233 ldlm_lock2handle(lock, &body->lock_handle[0]);
2234 if (lock->l_lvb_len > 0)
2235 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2236 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2238 ptlrpc_request_set_replen(req);
2239 /* notify the server we've replayed all requests.
2240 * also, we mark the request to be put on a dedicated
2241 * queue to be processed after all request replayes.
2243 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2245 LDLM_DEBUG(lock, "replaying lock:");
2247 atomic_inc(&req->rq_import->imp_replay_inflight);
2248 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2249 aa = ptlrpc_req_async_args(req);
2250 aa->lock_handle = body->lock_handle[0];
2251 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2252 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2258 * Cancel as many unused locks as possible before replay. since we are
2259 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2262 * Called only in recovery before replaying locks. there is no need to
2263 * replay locks that are unused. since the clients may hold thousands of
2264 * cached unused locks, dropping the unused locks can greatly reduce the
2265 * load on the servers at recovery time.
2267 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2270 struct list_head cancels = LIST_HEAD_INIT(cancels);
2272 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2273 "replay for namespace %s (%d)\n",
2274 ldlm_ns_name(ns), ns->ns_nr_unused);
2276 /* We don't need to care whether or not LRU resize is enabled
2277 * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2278 * count parameter */
2279 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2280 LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2282 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2283 canceled, ldlm_ns_name(ns));
2286 int ldlm_replay_locks(struct obd_import *imp)
2288 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2289 struct list_head list = LIST_HEAD_INIT(list);
2290 struct ldlm_lock *lock, *next;
2295 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2297 /* don't replay locks if import failed recovery */
2298 if (imp->imp_vbr_failed)
2301 /* ensure this doesn't fall to 0 before all have been queued */
2302 atomic_inc(&imp->imp_replay_inflight);
2304 if (ldlm_cancel_unused_locks_before_replay)
2305 ldlm_cancel_unused_locks_for_replay(ns);
2307 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2309 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2310 list_del_init(&lock->l_pending_chain);
2312 LDLM_LOCK_RELEASE(lock);
2313 continue; /* or try to do the rest? */
2315 rc = replay_one_lock(imp, lock);
2316 LDLM_LOCK_RELEASE(lock);
2319 atomic_dec(&imp->imp_replay_inflight);
2323 EXPORT_SYMBOL(ldlm_replay_locks);