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 int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
71 CFS_MODULE_PARM(ldlm_enqueue_min, "i", int, 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);
136 /* We use the same basis for both server side and client side functions
137 from a single node. */
138 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
140 int timeout = at_get(ldlm_lock_to_ns_at(lock));
142 return obd_timeout / 2;
143 /* Since these are non-updating timeouts, we should be conservative.
144 It would be nice to have some kind of "early reply" mechanism for
145 lock callbacks too... */
146 timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */
147 return max(timeout, ldlm_enqueue_min);
149 EXPORT_SYMBOL(ldlm_get_enq_timeout);
152 * Helper function for ldlm_completion_ast(), updating timings when lock is
155 static int ldlm_completion_tail(struct ldlm_lock *lock)
160 if (ldlm_is_destroyed(lock) || ldlm_is_failed(lock)) {
161 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
164 delay = cfs_time_sub(cfs_time_current_sec(),
165 lock->l_last_activity);
166 LDLM_DEBUG(lock, "client-side enqueue: granted after "
167 CFS_DURATION_T"s", delay);
169 /* Update our time estimate */
170 at_measured(ldlm_lock_to_ns_at(lock),
178 * Implementation of ->l_completion_ast() for a client, that doesn't wait
179 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
180 * other threads that cannot block for long.
182 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
186 if (flags == LDLM_FL_WAIT_NOREPROC) {
187 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
191 if (!(flags & LDLM_FL_BLOCKED_MASK)) {
192 wake_up(&lock->l_waitq);
193 RETURN(ldlm_completion_tail(lock));
196 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
198 ldlm_reprocess_all(lock->l_resource);
201 EXPORT_SYMBOL(ldlm_completion_ast_async);
204 * Generic LDLM "completion" AST. This is called in several cases:
206 * - when a reply to an ENQUEUE RPC is received from the server
207 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
208 * this point (determined by flags);
210 * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
213 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
216 * - to force all locks when resource is destroyed (cleanup_resource());
218 * - during lock conversion (not used currently).
220 * If lock is not granted in the first case, this function waits until second
221 * or penultimate cases happen in some other thread.
224 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
226 /* XXX ALLOCATE - 160 bytes */
227 struct lock_wait_data lwd;
228 struct obd_device *obd;
229 struct obd_import *imp = NULL;
230 struct l_wait_info lwi;
235 if (flags == LDLM_FL_WAIT_NOREPROC) {
236 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
240 if (!(flags & LDLM_FL_BLOCKED_MASK)) {
241 wake_up(&lock->l_waitq);
245 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
250 obd = class_exp2obd(lock->l_conn_export);
252 /* if this is a local lock, then there is no import */
254 imp = obd->u.cli.cl_import;
257 /* Wait a long time for enqueue - server may have to callback a
258 lock from another client. Server will evict the other client if it
259 doesn't respond reasonably, and then give us the lock. */
260 timeout = ldlm_get_enq_timeout(lock) * 2;
264 if (ldlm_is_no_timeout(lock)) {
265 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
266 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
268 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
269 ldlm_expired_completion_wait,
270 interrupted_completion_wait, &lwd);
274 spin_lock(&imp->imp_lock);
275 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
276 spin_unlock(&imp->imp_lock);
279 if (ns_is_client(ldlm_lock_to_ns(lock)) &&
280 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
281 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
282 ldlm_set_fail_loc(lock);
285 /* Go to sleep until the lock is granted or cancelled. */
286 rc = l_wait_event(lock->l_waitq,
287 is_granted_or_cancelled(lock), &lwi);
291 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
296 RETURN(ldlm_completion_tail(lock));
298 EXPORT_SYMBOL(ldlm_completion_ast);
301 * A helper to build a blocking AST function
303 * Perform a common operation for blocking ASTs:
304 * defferred lock cancellation.
306 * \param lock the lock blocking or canceling AST was called on
308 * \see mdt_blocking_ast
309 * \see ldlm_blocking_ast
311 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
316 ldlm_set_cbpending(lock);
317 do_ast = (!lock->l_readers && !lock->l_writers);
318 unlock_res_and_lock(lock);
321 struct lustre_handle lockh;
324 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
325 ldlm_lock2handle(lock, &lockh);
326 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
328 CERROR("ldlm_cli_cancel: %d\n", rc);
330 LDLM_DEBUG(lock, "Lock still has references, will be "
335 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
338 * Server blocking AST
340 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
343 * \param lock the lock which blocks a request or cancelling lock
346 * \param flag indicates whether this cancelling or blocking callback
348 * \see ldlm_blocking_ast_nocheck
350 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
351 void *data, int flag)
355 if (flag == LDLM_CB_CANCELING) {
356 /* Don't need to do anything here. */
360 lock_res_and_lock(lock);
361 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
362 * that ldlm_blocking_ast is called just before intent_policy method
363 * takes the lr_lock, then by the time we get the lock, we might not
364 * be the correct blocking function anymore. So check, and return
366 if (lock->l_blocking_ast != ldlm_blocking_ast) {
367 unlock_res_and_lock(lock);
370 RETURN(ldlm_blocking_ast_nocheck(lock));
372 EXPORT_SYMBOL(ldlm_blocking_ast);
375 * Implements ldlm_lock::l_glimpse_ast for extent locks acquired on the server.
377 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for that is
378 * rather subtle: with OST-side locking, it may so happen that _all_ extent
379 * locks are held by the OST. If client wants to obtain the current file size
380 * it calls ll_glimpse_size(), and (as all locks are held only on the server),
381 * this dummy glimpse callback fires and does nothing. The client still
382 * receives the correct file size due to the following fragment of code in
383 * ldlm_cb_interpret():
385 * if (rc == -ELDLM_NO_LOCK_DATA) {
386 * LDLM_DEBUG(lock, "lost race - client has a lock but no"
388 * ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
391 * That is, after the glimpse returns this error, ofd_lvbo_update() is called
392 * and returns the updated file attributes from the inode to the client.
394 * See also comment in ofd_intent_policy() on why servers must set a non-NULL
395 * l_glimpse_ast when grabbing DLM locks. Otherwise, the server will assume
396 * that the object is in the process of being destroyed.
398 * \param[in] lock DLM lock being glimpsed, unused
399 * \param[in] reqp pointer to ptlrpc_request, unused
401 * \retval -ELDLM_NO_LOCK_DATA to get attributes from disk object
403 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
405 return -ELDLM_NO_LOCK_DATA;
407 EXPORT_SYMBOL(ldlm_glimpse_ast);
410 * Enqueue a local lock (typically on a server).
412 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
413 const struct ldlm_res_id *res_id,
414 ldlm_type_t type, ldlm_policy_data_t *policy,
415 ldlm_mode_t mode, __u64 *flags,
416 ldlm_blocking_callback blocking,
417 ldlm_completion_callback completion,
418 ldlm_glimpse_callback glimpse,
419 void *data, __u32 lvb_len, enum lvb_type lvb_type,
420 const __u64 *client_cookie,
421 struct lustre_handle *lockh)
423 struct ldlm_lock *lock;
425 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
426 .lcs_blocking = blocking,
427 .lcs_glimpse = glimpse,
431 LASSERT(!(*flags & LDLM_FL_REPLAY));
432 if (unlikely(ns_is_client(ns))) {
433 CERROR("Trying to enqueue local lock in a shadow namespace\n");
437 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
440 GOTO(out_nolock, err = PTR_ERR(lock));
442 err = ldlm_lvbo_init(lock->l_resource);
444 LDLM_ERROR(lock, "delayed lvb init failed (rc %d)", err);
448 ldlm_lock2handle(lock, lockh);
450 /* NB: we don't have any lock now (lock_res_and_lock)
451 * because it's a new lock */
452 ldlm_lock_addref_internal_nolock(lock, mode);
453 ldlm_set_local(lock);
454 if (*flags & LDLM_FL_ATOMIC_CB)
455 ldlm_set_atomic_cb(lock);
458 lock->l_policy_data = *policy;
459 if (client_cookie != NULL)
460 lock->l_client_cookie = *client_cookie;
461 if (type == LDLM_EXTENT) {
462 /* extent lock without policy is a bug */
466 lock->l_req_extent = policy->l_extent;
469 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
470 if (unlikely(err != ELDLM_OK))
474 *policy = lock->l_policy_data;
476 if (lock->l_completion_ast)
477 lock->l_completion_ast(lock, *flags, NULL);
479 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
482 LDLM_LOCK_RELEASE(lock);
486 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
488 static void failed_lock_cleanup(struct ldlm_namespace *ns,
489 struct ldlm_lock *lock, int mode)
493 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
494 lock_res_and_lock(lock);
495 /* Check that lock is not granted or failed, we might race. */
496 if ((lock->l_req_mode != lock->l_granted_mode) &&
497 !ldlm_is_failed(lock)) {
498 /* Make sure that this lock will not be found by raced
499 * bl_ast and -EINVAL reply is sent to server anyways.
501 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
502 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
505 unlock_res_and_lock(lock);
509 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
510 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
512 LDLM_DEBUG(lock, "lock was granted or failed in race");
514 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
515 * from llite/file.c/ll_file_flock(). */
516 /* This code makes for the fact that we do not have blocking handler on
517 * a client for flock locks. As such this is the place where we must
518 * completely kill failed locks. (interrupted and those that
519 * were waiting to be granted when server evicted us. */
520 if (lock->l_resource->lr_type == LDLM_FLOCK) {
521 lock_res_and_lock(lock);
522 if (!ldlm_is_destroyed(lock)) {
523 ldlm_resource_unlink_lock(lock);
524 ldlm_lock_decref_internal_nolock(lock, mode);
525 ldlm_lock_destroy_nolock(lock);
527 unlock_res_and_lock(lock);
529 ldlm_lock_decref_internal(lock, mode);
534 * Finishing portion of client lock enqueue code.
536 * Called after receiving reply from server.
538 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
539 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
540 __u64 *flags, void *lvb, __u32 lvb_len,
541 struct lustre_handle *lockh,int rc)
543 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
544 int is_replay = *flags & LDLM_FL_REPLAY;
545 struct ldlm_lock *lock;
546 struct ldlm_reply *reply;
547 int cleanup_phase = 1;
551 lock = ldlm_handle2lock(lockh);
552 /* ldlm_cli_enqueue is holding a reference on this lock. */
554 LASSERT(type == LDLM_FLOCK);
558 LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
559 "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
561 if (rc != ELDLM_OK) {
563 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
564 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
566 if (rc != ELDLM_LOCK_ABORTED)
570 /* Before we return, swab the reply */
571 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
573 GOTO(cleanup, rc = -EPROTO);
576 LASSERT(lvb != NULL);
578 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
581 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
582 GOTO(cleanup, rc = size);
583 } else if (unlikely(size > lvb_len)) {
584 LDLM_ERROR(lock, "Replied LVB is larger than "
585 "expectation, expected = %d, replied = %d",
587 GOTO(cleanup, rc = -EINVAL);
591 if (rc == ELDLM_LOCK_ABORTED) {
593 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
595 GOTO(cleanup, rc = (rc != 0 ? rc : ELDLM_LOCK_ABORTED));
598 /* lock enqueued on the server */
601 lock_res_and_lock(lock);
602 /* Key change rehash lock in per-export hash with new key */
603 if (exp->exp_lock_hash) {
604 /* In the function below, .hs_keycmp resolves to
605 * ldlm_export_lock_keycmp() */
606 /* coverity[overrun-buffer-val] */
607 cfs_hash_rehash_key(exp->exp_lock_hash,
608 &lock->l_remote_handle,
612 lock->l_remote_handle = reply->lock_handle;
615 *flags = ldlm_flags_from_wire(reply->lock_flags);
616 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
617 LDLM_FL_INHERIT_MASK);
618 unlock_res_and_lock(lock);
620 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: "LPX64"\n",
621 lock, reply->lock_handle.cookie, *flags);
623 /* If enqueue returned a blocked lock but the completion handler has
624 * already run, then it fixed up the resource and we don't need to do it
626 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
627 int newmode = reply->lock_desc.l_req_mode;
629 if (newmode && newmode != lock->l_req_mode) {
630 LDLM_DEBUG(lock, "server returned different mode %s",
631 ldlm_lockname[newmode]);
632 lock->l_req_mode = newmode;
635 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
636 &lock->l_resource->lr_name)) {
637 CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES
638 " instead of "DLDLMRES"\n",
639 PLDLMRES(&reply->lock_desc.l_resource),
640 PLDLMRES(lock->l_resource));
642 rc = ldlm_lock_change_resource(ns, lock,
643 &reply->lock_desc.l_resource.lr_name);
644 if (rc || lock->l_resource == NULL)
645 GOTO(cleanup, rc = -ENOMEM);
646 LDLM_DEBUG(lock, "client-side enqueue, new resource");
649 if (!(type == LDLM_IBITS &&
650 !(exp_connect_flags(exp) & OBD_CONNECT_IBITS)))
651 /* We assume lock type cannot change on server*/
652 ldlm_convert_policy_to_local(exp,
653 lock->l_resource->lr_type,
654 &reply->lock_desc.l_policy_data,
655 &lock->l_policy_data);
656 if (type != LDLM_PLAIN)
657 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
660 if ((*flags) & LDLM_FL_AST_SENT) {
661 lock_res_and_lock(lock);
662 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
663 unlock_res_and_lock(lock);
664 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
667 /* If the lock has already been granted by a completion AST, don't
668 * clobber the LVB with an older one. */
670 /* We must lock or a racing completion might update lvb without
671 * letting us know and we'll clobber the correct value.
672 * Cannot unlock after the check either, a that still leaves
673 * a tiny window for completion to get in */
674 lock_res_and_lock(lock);
675 if (lock->l_req_mode != lock->l_granted_mode)
676 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
677 lock->l_lvb_data, size);
678 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, PAGE_CACHE_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 struct list_head *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 struct list_head head = LIST_HEAD_INIT(head);
766 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 struct list_head *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);
831 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
833 struct ptlrpc_request *req;
837 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
839 RETURN(ERR_PTR(-ENOMEM));
841 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
843 ptlrpc_request_free(req);
847 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
848 ptlrpc_request_set_replen(req);
851 EXPORT_SYMBOL(ldlm_enqueue_pack);
854 * Client-side lock enqueue.
856 * If a request has some specific initialisation it is passed in \a reqp,
857 * otherwise it is created in ldlm_cli_enqueue.
859 * Supports sync and async requests, pass \a async flag accordingly. If a
860 * request was created in ldlm_cli_enqueue and it is the async request,
861 * pass it to the caller in \a reqp.
863 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
864 struct ldlm_enqueue_info *einfo,
865 const struct ldlm_res_id *res_id,
866 ldlm_policy_data_t const *policy, __u64 *flags,
867 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
868 struct lustre_handle *lockh, int async)
870 struct ldlm_namespace *ns;
871 struct ldlm_lock *lock;
872 struct ldlm_request *body;
873 int is_replay = *flags & LDLM_FL_REPLAY;
874 int req_passed_in = 1;
876 struct ptlrpc_request *req;
879 LASSERT(exp != NULL);
881 ns = exp->exp_obd->obd_namespace;
883 /* If we're replaying this lock, just check some invariants.
884 * If we're creating a new lock, get everything all setup nice. */
886 lock = ldlm_handle2lock_long(lockh, 0);
887 LASSERT(lock != NULL);
888 LDLM_DEBUG(lock, "client-side enqueue START");
889 LASSERT(exp == lock->l_conn_export);
891 const struct ldlm_callback_suite cbs = {
892 .lcs_completion = einfo->ei_cb_cp,
893 .lcs_blocking = einfo->ei_cb_bl,
894 .lcs_glimpse = einfo->ei_cb_gl
896 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
897 einfo->ei_mode, &cbs, einfo->ei_cbdata,
900 RETURN(PTR_ERR(lock));
901 /* for the local lock, add the reference */
902 ldlm_lock_addref_internal(lock, einfo->ei_mode);
903 ldlm_lock2handle(lock, lockh);
905 lock->l_policy_data = *policy;
907 if (einfo->ei_type == LDLM_EXTENT) {
908 /* extent lock without policy is a bug */
912 lock->l_req_extent = policy->l_extent;
914 LDLM_DEBUG(lock, "client-side enqueue START, flags "LPX64"\n",
918 lock->l_conn_export = exp;
919 lock->l_export = NULL;
920 lock->l_blocking_ast = einfo->ei_cb_bl;
921 lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
923 /* lock not sent to server yet */
925 if (reqp == NULL || *reqp == NULL) {
926 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
931 failed_lock_cleanup(ns, lock, einfo->ei_mode);
932 LDLM_LOCK_RELEASE(lock);
942 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
944 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
945 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
948 /* Dump lock data into the request buffer */
949 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
950 ldlm_lock2desc(lock, &body->lock_desc);
951 body->lock_flags = ldlm_flags_to_wire(*flags);
952 body->lock_handle[0] = *lockh;
954 /* Continue as normal. */
955 if (!req_passed_in) {
957 req_capsule_extend(&req->rq_pill,
958 &RQF_LDLM_ENQUEUE_LVB);
959 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
961 ptlrpc_request_set_replen(req);
965 LASSERT(reqp != NULL);
969 LDLM_DEBUG(lock, "sending request");
971 rc = ptlrpc_queue_wait(req);
973 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
974 einfo->ei_mode, flags, lvb, lvb_len,
977 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
978 * one reference that we took */
980 LDLM_LOCK_RELEASE(lock);
984 if (!req_passed_in && req != NULL) {
985 ptlrpc_req_finished(req);
992 EXPORT_SYMBOL(ldlm_cli_enqueue);
994 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
997 struct ldlm_resource *res;
1000 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1001 CERROR("Trying to cancel local lock\n");
1004 LDLM_DEBUG(lock, "client-side local convert");
1006 res = ldlm_lock_convert(lock, new_mode, flags);
1008 ldlm_reprocess_all(res);
1011 rc = LUSTRE_EDEADLK;
1013 LDLM_DEBUG(lock, "client-side local convert handler END");
1014 LDLM_LOCK_PUT(lock);
1018 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
1019 * conversion of locks which are on the waiting or converting queue */
1020 /* Caller of this code is supposed to take care of lock readers/writers
1022 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
1024 struct ldlm_request *body;
1025 struct ldlm_reply *reply;
1026 struct ldlm_lock *lock;
1027 struct ldlm_resource *res;
1028 struct ptlrpc_request *req;
1032 lock = ldlm_handle2lock(lockh);
1039 if (lock->l_conn_export == NULL)
1040 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
1042 LDLM_DEBUG(lock, "client-side convert");
1044 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
1045 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1048 LDLM_LOCK_PUT(lock);
1052 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1053 body->lock_handle[0] = lock->l_remote_handle;
1055 body->lock_desc.l_req_mode = new_mode;
1056 body->lock_flags = ldlm_flags_to_wire(*flags);
1059 ptlrpc_request_set_replen(req);
1060 rc = ptlrpc_queue_wait(req);
1064 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1066 GOTO(out, rc = -EPROTO);
1069 GOTO(out, rc = req->rq_status);
1071 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
1073 ldlm_reprocess_all(res);
1074 /* Go to sleep until the lock is granted. */
1075 /* FIXME: or cancelled. */
1076 if (lock->l_completion_ast) {
1077 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1083 rc = LUSTRE_EDEADLK;
1087 LDLM_LOCK_PUT(lock);
1088 ptlrpc_req_finished(req);
1091 EXPORT_SYMBOL(ldlm_cli_convert);
1094 * Cancel locks locally.
1096 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1097 * \retval LDLM_FL_CANCELING otherwise;
1098 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1100 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1102 __u64 rc = LDLM_FL_LOCAL_ONLY;
1105 if (lock->l_conn_export) {
1108 LDLM_DEBUG(lock, "client-side cancel");
1109 /* Set this flag to prevent others from getting new references*/
1110 lock_res_and_lock(lock);
1111 ldlm_set_cbpending(lock);
1112 local_only = !!(lock->l_flags &
1113 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1114 ldlm_cancel_callback(lock);
1115 rc = (ldlm_is_bl_ast(lock)) ?
1116 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1117 unlock_res_and_lock(lock);
1120 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1122 rc = LDLM_FL_LOCAL_ONLY;
1124 ldlm_lock_cancel(lock);
1126 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1127 LDLM_ERROR(lock, "Trying to cancel local lock");
1130 LDLM_DEBUG(lock, "server-side local cancel");
1131 ldlm_lock_cancel(lock);
1132 ldlm_reprocess_all(lock->l_resource);
1139 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1141 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1142 struct list_head *head, int count)
1144 struct ldlm_request *dlm;
1145 struct ldlm_lock *lock;
1146 int max, packed = 0;
1149 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1150 LASSERT(dlm != NULL);
1152 /* Check the room in the request buffer. */
1153 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1154 sizeof(struct ldlm_request);
1155 max /= sizeof(struct lustre_handle);
1156 max += LDLM_LOCKREQ_HANDLES;
1157 LASSERT(max >= dlm->lock_count + count);
1159 /* XXX: it would be better to pack lock handles grouped by resource.
1160 * so that the server cancel would call filter_lvbo_update() less
1162 list_for_each_entry(lock, head, l_bl_ast) {
1165 LASSERT(lock->l_conn_export);
1166 /* Pack the lock handle to the given request buffer. */
1167 LDLM_DEBUG(lock, "packing");
1168 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1171 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1176 * Prepare and send a batched cancel RPC. It will include \a count lock
1177 * handles of locks given in \a cancels list. */
1178 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1179 int count, ldlm_cancel_flags_t flags)
1181 struct ptlrpc_request *req = NULL;
1182 struct obd_import *imp;
1187 LASSERT(exp != NULL);
1190 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1192 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1195 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1196 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1201 imp = class_exp2cliimp(exp);
1202 if (imp == NULL || imp->imp_invalid) {
1204 "skipping cancel on invalid import %p\n", imp);
1208 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1210 GOTO(out, rc = -ENOMEM);
1212 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1213 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1214 ldlm_request_bufsize(count, LDLM_CANCEL));
1216 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1218 ptlrpc_request_free(req);
1222 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1223 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1224 ptlrpc_at_set_req_timeout(req);
1226 ldlm_cancel_pack(req, cancels, count);
1228 ptlrpc_request_set_replen(req);
1229 if (flags & LCF_ASYNC) {
1230 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1234 rc = ptlrpc_queue_wait(req);
1236 if (rc == LUSTRE_ESTALE) {
1237 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1238 "out of sync -- not fatal\n",
1239 libcfs_nid2str(req->rq_import->
1240 imp_connection->c_peer.nid));
1242 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1243 req->rq_import_generation == imp->imp_generation) {
1244 ptlrpc_req_finished(req);
1246 } else if (rc != ELDLM_OK) {
1247 /* -ESHUTDOWN is common on umount */
1248 CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1249 "Got rc %d from cancel RPC: "
1250 "canceling anyway\n", rc);
1257 ptlrpc_req_finished(req);
1260 return sent ? sent : rc;
1262 EXPORT_SYMBOL(ldlm_cli_cancel_req);
1264 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1266 LASSERT(imp != NULL);
1267 return &imp->imp_obd->obd_namespace->ns_pool;
1271 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1273 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1275 struct obd_device *obd;
1279 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1280 !imp_connect_lru_resize(req->rq_import)))
1283 * Do nothing for corner cases.
1288 /* In some cases RPC may contain SLV and limit zeroed out. This
1289 * is the case when server does not support LRU resize feature.
1290 * This is also possible in some recovery cases when server-side
1291 * reqs have no reference to the OBD export and thus access to
1292 * server-side namespace is not possible. */
1293 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1294 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1295 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1296 "(SLV: "LPU64", Limit: %u)",
1297 lustre_msg_get_slv(req->rq_repmsg),
1298 lustre_msg_get_limit(req->rq_repmsg));
1302 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1303 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1304 obd = req->rq_import->imp_obd;
1306 /* Set new SLV and limit in OBD fields to make them accessible
1307 * to the pool thread. We do not access obd_namespace and pool
1308 * directly here as there is no reliable way to make sure that
1309 * they are still alive at cleanup time. Evil races are possible
1310 * which may cause Oops at that time. */
1311 write_lock(&obd->obd_pool_lock);
1312 obd->obd_pool_slv = new_slv;
1313 obd->obd_pool_limit = new_limit;
1314 write_unlock(&obd->obd_pool_lock);
1318 EXPORT_SYMBOL(ldlm_cli_update_pool);
1321 * Client side lock cancel.
1323 * Lock must not have any readers or writers by this time.
1325 int ldlm_cli_cancel(struct lustre_handle *lockh,
1326 ldlm_cancel_flags_t cancel_flags)
1328 struct obd_export *exp;
1329 int avail, flags, count = 1;
1331 struct ldlm_namespace *ns;
1332 struct ldlm_lock *lock;
1333 struct list_head cancels = LIST_HEAD_INIT(cancels);
1336 /* concurrent cancels on the same handle can happen */
1337 lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1339 LDLM_DEBUG_NOLOCK("lock is already being destroyed");
1343 rc = ldlm_cli_cancel_local(lock);
1344 if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
1345 LDLM_LOCK_RELEASE(lock);
1348 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1349 * RPC which goes to canceld portal, so we can cancel other LRU locks
1350 * here and send them all as one LDLM_CANCEL RPC. */
1351 LASSERT(list_empty(&lock->l_bl_ast));
1352 list_add(&lock->l_bl_ast, &cancels);
1354 exp = lock->l_conn_export;
1355 if (exp_connect_cancelset(exp)) {
1356 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1361 ns = ldlm_lock_to_ns(lock);
1362 flags = ns_connect_lru_resize(ns) ?
1363 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1364 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1367 ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1370 EXPORT_SYMBOL(ldlm_cli_cancel);
1373 * Locally cancel up to \a count locks in list \a cancels.
1374 * Return the number of cancelled locks.
1376 int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
1377 ldlm_cancel_flags_t flags)
1379 struct list_head head = LIST_HEAD_INIT(head);
1380 struct ldlm_lock *lock, *next;
1381 int left = 0, bl_ast = 0;
1385 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1389 if (flags & LCF_LOCAL) {
1390 rc = LDLM_FL_LOCAL_ONLY;
1391 ldlm_lock_cancel(lock);
1393 rc = ldlm_cli_cancel_local(lock);
1395 /* Until we have compound requests and can send LDLM_CANCEL
1396 * requests batched with generic RPCs, we need to send cancels
1397 * with the LDLM_FL_BL_AST flag in a separate RPC from
1398 * the one being generated now. */
1399 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1400 LDLM_DEBUG(lock, "Cancel lock separately");
1401 list_del_init(&lock->l_bl_ast);
1402 list_add(&lock->l_bl_ast, &head);
1406 if (rc == LDLM_FL_LOCAL_ONLY) {
1407 /* CANCEL RPC should not be sent to server. */
1408 list_del_init(&lock->l_bl_ast);
1409 LDLM_LOCK_RELEASE(lock);
1415 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1420 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
1423 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1424 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1425 * readahead requests, ...)
1427 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1428 struct ldlm_lock *lock,
1429 int unused, int added,
1432 ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1434 /* don't check added & count since we want to process all locks
1436 * It's fine to not take lock to access lock->l_resource since
1437 * the lock has already been granted so it won't change. */
1438 switch (lock->l_resource->lr_type) {
1441 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) != 0)
1444 result = LDLM_POLICY_SKIP_LOCK;
1445 lock_res_and_lock(lock);
1446 ldlm_set_skipped(lock);
1447 unlock_res_and_lock(lock);
1455 * Callback function for LRU-resize policy. Decides whether to keep
1456 * \a lock in LRU for current \a LRU size \a unused, added in current
1457 * scan \a added and number of locks to be preferably canceled \a count.
1459 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1461 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1463 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1464 struct ldlm_lock *lock,
1465 int unused, int added,
1468 cfs_time_t cur = cfs_time_current();
1469 struct ldlm_pool *pl = &ns->ns_pool;
1473 /* Stop LRU processing when we reach past @count or have checked all
1475 if (count && added >= count)
1476 return LDLM_POLICY_KEEP_LOCK;
1478 slv = ldlm_pool_get_slv(pl);
1479 lvf = ldlm_pool_get_lvf(pl);
1480 la = cfs_duration_sec(cfs_time_sub(cur,
1481 lock->l_last_used));
1482 lv = lvf * la * unused;
1484 /* Inform pool about current CLV to see it via proc. */
1485 ldlm_pool_set_clv(pl, lv);
1487 /* Stop when SLV is not yet come from server or lv is smaller than
1489 if (slv == 0 || lv < slv)
1490 return LDLM_POLICY_KEEP_LOCK;
1492 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) == 0)
1493 return LDLM_POLICY_KEEP_LOCK;
1495 return LDLM_POLICY_CANCEL_LOCK;
1499 * Callback function for proc used policy. Makes decision whether to keep
1500 * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1501 * added and number of locks to be preferably canceled \a count.
1503 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1505 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1507 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1508 struct ldlm_lock *lock,
1509 int unused, int added,
1512 /* Stop LRU processing when we reach past @count or have checked all
1514 return (added >= count) ?
1515 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1519 * Callback function for aged policy. Makes decision whether to keep \a lock in
1520 * LRU for current LRU size \a unused, added in current scan \a added and
1521 * number of locks to be preferably canceled \a count.
1523 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1525 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1527 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1528 struct ldlm_lock *lock,
1529 int unused, int added,
1532 if ((added >= count) &&
1533 cfs_time_before(cfs_time_current(),
1534 cfs_time_add(lock->l_last_used, ns->ns_max_age)))
1535 return LDLM_POLICY_KEEP_LOCK;
1537 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) == 0)
1538 return LDLM_POLICY_KEEP_LOCK;
1540 return LDLM_POLICY_CANCEL_LOCK;
1544 * Callback function for default policy. Makes decision whether to keep \a lock
1545 * in LRU for current LRU size \a unused, added in current scan \a added and
1546 * number of locks to be preferably canceled \a count.
1548 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1550 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1552 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1553 struct ldlm_lock *lock,
1554 int unused, int added,
1557 /* Stop LRU processing when we reach past count or have checked all
1559 return (added >= count) ?
1560 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1563 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1564 struct ldlm_lock *, int,
1567 static ldlm_cancel_lru_policy_t
1568 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1570 if (flags & LDLM_CANCEL_NO_WAIT)
1571 return ldlm_cancel_no_wait_policy;
1573 if (ns_connect_lru_resize(ns)) {
1574 if (flags & LDLM_CANCEL_SHRINK)
1575 /* We kill passed number of old locks. */
1576 return ldlm_cancel_passed_policy;
1577 else if (flags & LDLM_CANCEL_LRUR)
1578 return ldlm_cancel_lrur_policy;
1579 else if (flags & LDLM_CANCEL_PASSED)
1580 return ldlm_cancel_passed_policy;
1582 if (flags & LDLM_CANCEL_AGED)
1583 return ldlm_cancel_aged_policy;
1586 return ldlm_cancel_default_policy;
1590 * - Free space in LRU for \a count new locks,
1591 * redundant unused locks are canceled locally;
1592 * - also cancel locally unused aged locks;
1593 * - do not cancel more than \a max locks;
1594 * - GET the found locks and add them into the \a cancels list.
1596 * A client lock can be added to the l_bl_ast list only when it is
1597 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1598 * CANCEL. There are the following use cases:
1599 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1600 * ldlm_cli_cancel(), which check and set this flag properly. As any
1601 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1602 * later without any special locking.
1604 * Calling policies for enabled LRU resize:
1605 * ----------------------------------------
1606 * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to
1607 * cancel not more than \a count locks;
1609 * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at
1610 * the beginning of LRU list);
1612 * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to
1613 * memory pressre policy function;
1615 * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy".
1617 * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1618 * (typically before replaying locks) w/o
1619 * sending any RPCs or waiting for any
1620 * outstanding RPC to complete.
1622 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns,
1623 struct list_head *cancels, int count, int max,
1626 ldlm_cancel_lru_policy_t pf;
1627 struct ldlm_lock *lock, *next;
1628 int added = 0, unused, remained;
1631 spin_lock(&ns->ns_lock);
1632 unused = ns->ns_nr_unused;
1635 if (!ns_connect_lru_resize(ns))
1636 count += unused - ns->ns_max_unused;
1638 pf = ldlm_cancel_lru_policy(ns, flags);
1639 LASSERT(pf != NULL);
1641 while (!list_empty(&ns->ns_unused_list)) {
1642 ldlm_policy_res_t result;
1644 /* all unused locks */
1645 if (remained-- <= 0)
1648 /* For any flags, stop scanning if @max is reached. */
1649 if (max && added >= max)
1652 list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1654 /* No locks which got blocking requests. */
1655 LASSERT(!ldlm_is_bl_ast(lock));
1657 if (flags & LDLM_CANCEL_NO_WAIT &&
1658 ldlm_is_skipped(lock))
1659 /* already processed */
1662 /* Somebody is already doing CANCEL. No need for this
1663 * lock in LRU, do not traverse it again. */
1664 if (!ldlm_is_canceling(lock))
1667 ldlm_lock_remove_from_lru_nolock(lock);
1669 if (&lock->l_lru == &ns->ns_unused_list)
1672 LDLM_LOCK_GET(lock);
1673 spin_unlock(&ns->ns_lock);
1674 lu_ref_add(&lock->l_reference, __FUNCTION__, current);
1676 /* Pass the lock through the policy filter and see if it
1677 * should stay in LRU.
1679 * Even for shrinker policy we stop scanning if
1680 * we find a lock that should stay in the cache.
1681 * We should take into account lock age anyway
1682 * as a new lock is a valuable resource even if
1683 * it has a low weight.
1685 * That is, for shrinker policy we drop only
1686 * old locks, but additionally choose them by
1687 * their weight. Big extent locks will stay in
1689 result = pf(ns, lock, unused, added, count);
1690 if (result == LDLM_POLICY_KEEP_LOCK) {
1691 lu_ref_del(&lock->l_reference,
1692 __FUNCTION__, current);
1693 LDLM_LOCK_RELEASE(lock);
1694 spin_lock(&ns->ns_lock);
1697 if (result == LDLM_POLICY_SKIP_LOCK) {
1698 lu_ref_del(&lock->l_reference,
1700 LDLM_LOCK_RELEASE(lock);
1701 spin_lock(&ns->ns_lock);
1705 lock_res_and_lock(lock);
1706 /* Check flags again under the lock. */
1707 if (ldlm_is_canceling(lock) ||
1708 (ldlm_lock_remove_from_lru(lock) == 0)) {
1709 /* Another thread is removing lock from LRU, or
1710 * somebody is already doing CANCEL, or there
1711 * is a blocking request which will send cancel
1712 * by itself, or the lock is no longer unused. */
1713 unlock_res_and_lock(lock);
1714 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1715 LDLM_LOCK_RELEASE(lock);
1716 spin_lock(&ns->ns_lock);
1719 LASSERT(!lock->l_readers && !lock->l_writers);
1721 /* If we have chosen to cancel this lock voluntarily, we
1722 * better send cancel notification to server, so that it
1723 * frees appropriate state. This might lead to a race
1724 * where while we are doing cancel here, server is also
1725 * silently cancelling this lock. */
1726 ldlm_clear_cancel_on_block(lock);
1728 /* Setting the CBPENDING flag is a little misleading,
1729 * but prevents an important race; namely, once
1730 * CBPENDING is set, the lock can accumulate no more
1731 * readers/writers. Since readers and writers are
1732 * already zero here, ldlm_lock_decref() won't see
1733 * this flag and call l_blocking_ast */
1734 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1736 /* We can't re-add to l_lru as it confuses the
1737 * refcounting in ldlm_lock_remove_from_lru() if an AST
1738 * arrives after we drop lr_lock below. We use l_bl_ast
1739 * and can't use l_pending_chain as it is used both on
1740 * server and client nevertheless bug 5666 says it is
1741 * used only on server */
1742 LASSERT(list_empty(&lock->l_bl_ast));
1743 list_add(&lock->l_bl_ast, cancels);
1744 unlock_res_and_lock(lock);
1745 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1746 spin_lock(&ns->ns_lock);
1750 spin_unlock(&ns->ns_lock);
1754 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1755 int count, int max, ldlm_cancel_flags_t cancel_flags,
1759 added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1762 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1766 * Cancel at least \a nr locks from given namespace LRU.
1768 * When called with LCF_ASYNC the blocking callback will be handled
1769 * in a thread and this function will return after the thread has been
1770 * asked to call the callback. When called with LCF_ASYNC the blocking
1771 * callback will be performed in this function.
1773 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
1774 ldlm_cancel_flags_t cancel_flags,
1777 struct list_head cancels = LIST_HEAD_INIT(cancels);
1781 /* Just prepare the list of locks, do not actually cancel them yet.
1782 * Locks are cancelled later in a separate thread. */
1783 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1784 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
1792 * Find and cancel locally unused locks found on resource, matched to the
1793 * given policy, mode. GET the found locks and add them into the \a cancels
1796 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1797 struct list_head *cancels,
1798 ldlm_policy_data_t *policy,
1799 ldlm_mode_t mode, __u64 lock_flags,
1800 ldlm_cancel_flags_t cancel_flags, void *opaque)
1802 struct ldlm_lock *lock;
1807 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1808 if (opaque != NULL && lock->l_ast_data != opaque) {
1809 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1810 lock->l_ast_data, opaque);
1815 if (lock->l_readers || lock->l_writers)
1818 /* If somebody is already doing CANCEL, or blocking AST came,
1819 * skip this lock. */
1820 if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock))
1823 if (lockmode_compat(lock->l_granted_mode, mode))
1826 /* If policy is given and this is IBITS lock, add to list only
1827 * those locks that match by policy. */
1828 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1829 !(lock->l_policy_data.l_inodebits.bits &
1830 policy->l_inodebits.bits))
1833 /* See CBPENDING comment in ldlm_cancel_lru */
1834 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1837 LASSERT(list_empty(&lock->l_bl_ast));
1838 list_add(&lock->l_bl_ast, cancels);
1839 LDLM_LOCK_GET(lock);
1844 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
1846 EXPORT_SYMBOL(ldlm_cancel_resource_local);
1849 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1851 * If \a req is NULL, send CANCEL request to server with handles of locks
1852 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1853 * separately per lock.
1854 * If \a req is not NULL, put handles of locks in \a cancels into the request
1855 * buffer at the offset \a off.
1856 * Destroy \a cancels at the end.
1858 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1859 struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1861 struct ldlm_lock *lock;
1865 if (list_empty(cancels) || count == 0)
1868 /* XXX: requests (both batched and not) could be sent in parallel.
1869 * Usually it is enough to have just 1 RPC, but it is possible that
1870 * there are too many locks to be cancelled in LRU or on a resource.
1871 * It would also speed up the case when the server does not support
1874 LASSERT(!list_empty(cancels));
1875 lock = list_entry(cancels->next, struct ldlm_lock,
1877 LASSERT(lock->l_conn_export);
1879 if (exp_connect_cancelset(lock->l_conn_export)) {
1882 ldlm_cancel_pack(req, cancels, count);
1884 res = ldlm_cli_cancel_req(lock->l_conn_export,
1888 res = ldlm_cli_cancel_req(lock->l_conn_export,
1893 CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1894 "ldlm_cli_cancel_list: %d\n", res);
1899 ldlm_lock_list_put(cancels, l_bl_ast, res);
1901 LASSERT(count == 0);
1904 EXPORT_SYMBOL(ldlm_cli_cancel_list);
1907 * Cancel all locks on a resource that have 0 readers/writers.
1909 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1910 * to notify the server. */
1911 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1912 const struct ldlm_res_id *res_id,
1913 ldlm_policy_data_t *policy,
1915 ldlm_cancel_flags_t flags,
1918 struct ldlm_resource *res;
1919 struct list_head cancels = LIST_HEAD_INIT(cancels);
1924 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1926 /* This is not a problem. */
1927 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1931 LDLM_RESOURCE_ADDREF(res);
1932 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1933 0, flags | LCF_BL_AST, opaque);
1934 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1936 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
1939 LDLM_RESOURCE_DELREF(res);
1940 ldlm_resource_putref(res);
1943 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
1945 struct ldlm_cli_cancel_arg {
1950 static int ldlm_cli_hash_cancel_unused(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1951 struct hlist_node *hnode, void *arg)
1953 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1954 struct ldlm_cli_cancel_arg *lc = arg;
1956 ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1957 NULL, LCK_MINMODE, lc->lc_flags,
1959 /* must return 0 for hash iteration */
1964 * Cancel all locks on a namespace (or a specific resource, if given)
1965 * that have 0 readers/writers.
1967 * If flags & LCF_LOCAL, throw the locks away without trying
1968 * to notify the server. */
1969 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1970 const struct ldlm_res_id *res_id,
1971 ldlm_cancel_flags_t flags, void *opaque)
1973 struct ldlm_cli_cancel_arg arg = {
1975 .lc_opaque = opaque,
1983 if (res_id != NULL) {
1984 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1988 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1989 ldlm_cli_hash_cancel_unused, &arg);
1993 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1995 /* Lock iterators. */
1997 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
2000 struct list_head *tmp, *next;
2001 struct ldlm_lock *lock;
2002 int rc = LDLM_ITER_CONTINUE;
2007 RETURN(LDLM_ITER_CONTINUE);
2010 list_for_each_safe(tmp, next, &res->lr_granted) {
2011 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2013 if (iter(lock, closure) == LDLM_ITER_STOP)
2014 GOTO(out, rc = LDLM_ITER_STOP);
2017 list_for_each_safe(tmp, next, &res->lr_converting) {
2018 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2020 if (iter(lock, closure) == LDLM_ITER_STOP)
2021 GOTO(out, rc = LDLM_ITER_STOP);
2024 list_for_each_safe(tmp, next, &res->lr_waiting) {
2025 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2027 if (iter(lock, closure) == LDLM_ITER_STOP)
2028 GOTO(out, rc = LDLM_ITER_STOP);
2034 EXPORT_SYMBOL(ldlm_resource_foreach);
2036 struct iter_helper_data {
2037 ldlm_iterator_t iter;
2041 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2043 struct iter_helper_data *helper = closure;
2044 return helper->iter(lock, helper->closure);
2047 static int ldlm_res_iter_helper(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2048 struct hlist_node *hnode, void *arg)
2051 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2053 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2057 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2058 ldlm_iterator_t iter, void *closure)
2061 struct iter_helper_data helper = { .iter = iter, .closure = closure };
2063 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2064 ldlm_res_iter_helper, &helper);
2067 EXPORT_SYMBOL(ldlm_namespace_foreach);
2069 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2070 * return 0: find no resource
2071 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2074 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2075 const struct ldlm_res_id *res_id,
2076 ldlm_iterator_t iter, void *data)
2078 struct ldlm_resource *res;
2082 LASSERTF(ns != NULL, "must pass in namespace\n");
2084 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2088 LDLM_RESOURCE_ADDREF(res);
2089 rc = ldlm_resource_foreach(res, iter, data);
2090 LDLM_RESOURCE_DELREF(res);
2091 ldlm_resource_putref(res);
2094 EXPORT_SYMBOL(ldlm_resource_iterate);
2098 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2100 struct list_head *list = closure;
2102 /* we use l_pending_chain here, because it's unused on clients. */
2103 LASSERTF(list_empty(&lock->l_pending_chain),
2104 "lock %p next %p prev %p\n",
2105 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2106 /* bug 9573: don't replay locks left after eviction, or
2107 * bug 17614: locks being actively cancelled. Get a reference
2108 * on a lock so that it does not disapear under us (e.g. due to cancel)
2110 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
2111 list_add(&lock->l_pending_chain, list);
2112 LDLM_LOCK_GET(lock);
2115 return LDLM_ITER_CONTINUE;
2118 static int replay_lock_interpret(const struct lu_env *env,
2119 struct ptlrpc_request *req,
2120 struct ldlm_async_args *aa, int rc)
2122 struct ldlm_lock *lock;
2123 struct ldlm_reply *reply;
2124 struct obd_export *exp;
2127 atomic_dec(&req->rq_import->imp_replay_inflight);
2131 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2133 GOTO(out, rc = -EPROTO);
2135 lock = ldlm_handle2lock(&aa->lock_handle);
2137 CERROR("received replay ack for unknown local cookie "LPX64
2138 " remote cookie "LPX64 " from server %s id %s\n",
2139 aa->lock_handle.cookie, reply->lock_handle.cookie,
2140 req->rq_export->exp_client_uuid.uuid,
2141 libcfs_id2str(req->rq_peer));
2142 GOTO(out, rc = -ESTALE);
2145 /* Key change rehash lock in per-export hash with new key */
2146 exp = req->rq_export;
2147 if (exp && exp->exp_lock_hash) {
2148 /* In the function below, .hs_keycmp resolves to
2149 * ldlm_export_lock_keycmp() */
2150 /* coverity[overrun-buffer-val] */
2151 cfs_hash_rehash_key(exp->exp_lock_hash,
2152 &lock->l_remote_handle,
2153 &reply->lock_handle,
2156 lock->l_remote_handle = reply->lock_handle;
2159 LDLM_DEBUG(lock, "replayed lock:");
2160 ptlrpc_import_recovery_state_machine(req->rq_import);
2161 LDLM_LOCK_PUT(lock);
2164 ptlrpc_connect_import(req->rq_import);
2169 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2171 struct ptlrpc_request *req;
2172 struct ldlm_async_args *aa;
2173 struct ldlm_request *body;
2178 /* Bug 11974: Do not replay a lock which is actively being canceled */
2179 if (ldlm_is_canceling(lock)) {
2180 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2184 /* If this is reply-less callback lock, we cannot replay it, since
2185 * server might have long dropped it, but notification of that event was
2186 * lost by network. (and server granted conflicting lock already) */
2187 if (ldlm_is_cancel_on_block(lock)) {
2188 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2189 ldlm_lock_cancel(lock);
2194 * If granted mode matches the requested mode, this lock is granted.
2196 * If they differ, but we have a granted mode, then we were granted
2197 * one mode and now want another: ergo, converting.
2199 * If we haven't been granted anything and are on a resource list,
2200 * then we're blocked/waiting.
2202 * If we haven't been granted anything and we're NOT on a resource list,
2203 * then we haven't got a reply yet and don't have a known disposition.
2204 * This happens whenever a lock enqueue is the request that triggers
2207 if (lock->l_granted_mode == lock->l_req_mode)
2208 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2209 else if (lock->l_granted_mode)
2210 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2211 else if (!list_empty(&lock->l_res_link))
2212 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2214 flags = LDLM_FL_REPLAY;
2216 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2217 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2221 /* We're part of recovery, so don't wait for it. */
2222 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2224 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2225 ldlm_lock2desc(lock, &body->lock_desc);
2226 body->lock_flags = ldlm_flags_to_wire(flags);
2228 ldlm_lock2handle(lock, &body->lock_handle[0]);
2229 if (lock->l_lvb_len > 0)
2230 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2231 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2233 ptlrpc_request_set_replen(req);
2234 /* notify the server we've replayed all requests.
2235 * also, we mark the request to be put on a dedicated
2236 * queue to be processed after all request replayes.
2238 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2240 LDLM_DEBUG(lock, "replaying lock:");
2242 atomic_inc(&req->rq_import->imp_replay_inflight);
2243 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2244 aa = ptlrpc_req_async_args(req);
2245 aa->lock_handle = body->lock_handle[0];
2246 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2247 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2253 * Cancel as many unused locks as possible before replay. since we are
2254 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2257 * Called only in recovery before replaying locks. there is no need to
2258 * replay locks that are unused. since the clients may hold thousands of
2259 * cached unused locks, dropping the unused locks can greatly reduce the
2260 * load on the servers at recovery time.
2262 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2265 struct list_head cancels = LIST_HEAD_INIT(cancels);
2267 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2268 "replay for namespace %s (%d)\n",
2269 ldlm_ns_name(ns), ns->ns_nr_unused);
2271 /* We don't need to care whether or not LRU resize is enabled
2272 * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2273 * count parameter */
2274 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2275 LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2277 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2278 canceled, ldlm_ns_name(ns));
2281 int ldlm_replay_locks(struct obd_import *imp)
2283 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2284 struct list_head list = LIST_HEAD_INIT(list);
2285 struct ldlm_lock *lock, *next;
2290 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2292 /* don't replay locks if import failed recovery */
2293 if (imp->imp_vbr_failed)
2296 /* ensure this doesn't fall to 0 before all have been queued */
2297 atomic_inc(&imp->imp_replay_inflight);
2299 if (ldlm_cancel_unused_locks_before_replay)
2300 ldlm_cancel_unused_locks_for_replay(ns);
2302 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2304 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2305 list_del_init(&lock->l_pending_chain);
2307 LDLM_LOCK_RELEASE(lock);
2308 continue; /* or try to do the rest? */
2310 rc = replay_one_lock(imp, lock);
2311 LDLM_LOCK_RELEASE(lock);
2314 atomic_dec(&imp->imp_replay_inflight);
2318 EXPORT_SYMBOL(ldlm_replay_locks);