4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 * This file contains Asynchronous System Trap (AST) handlers and related
38 * LDLM request-processing routines.
40 * An AST is a callback issued on a lock when its state is changed. There are
41 * several different types of ASTs (callbacks) registered for each lock:
43 * - completion AST: when a lock is enqueued by some process, but cannot be
44 * granted immediately due to other conflicting locks on the same resource,
45 * the completion AST is sent to notify the caller when the lock is
48 * - blocking AST: when a lock is granted to some process, if another process
49 * enqueues a conflicting (blocking) lock on a resource, a blocking AST is
50 * sent to notify the holder(s) of the lock(s) of the conflicting lock
51 * request. The lock holder(s) must release their lock(s) on that resource in
52 * a timely manner or be evicted by the server.
54 * - glimpse AST: this is used when a process wants information about a lock
55 * (i.e. the lock value block (LVB)) but does not necessarily require holding
56 * the lock. If the resource is locked, the lock holder(s) are sent glimpse
57 * ASTs and the LVB is returned to the caller, and lock holder(s) may CANCEL
58 * their lock(s) if they are idle. If the resource is not locked, the server
62 #define DEBUG_SUBSYSTEM S_LDLM
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;
97 if (lock->l_conn_export == NULL) {
98 static cfs_time_t next_dump = 0, last_dump = 0;
100 if (ptlrpc_check_suspend())
103 LCONSOLE_WARN("lock timed out (enqueued at "CFS_TIME_T", "
104 CFS_DURATION_T"s ago)\n",
105 lock->l_last_activity,
106 cfs_time_sub(cfs_time_current_sec(),
107 lock->l_last_activity));
108 LDLM_DEBUG(lock, "lock timed out (enqueued at "CFS_TIME_T", "
109 CFS_DURATION_T"s ago); not entering recovery in "
110 "server code, just going back to sleep",
111 lock->l_last_activity,
112 cfs_time_sub(cfs_time_current_sec(),
113 lock->l_last_activity));
114 if (cfs_time_after(cfs_time_current(), next_dump)) {
115 last_dump = next_dump;
116 next_dump = cfs_time_shift(300);
117 ldlm_namespace_dump(D_DLMTRACE,
118 ldlm_lock_to_ns(lock));
120 libcfs_debug_dumplog();
125 obd = lock->l_conn_export->exp_obd;
126 imp = obd->u.cli.cl_import;
127 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
128 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
129 CFS_DURATION_T"s ago), entering recovery for %s@%s",
130 lock->l_last_activity,
131 cfs_time_sub(cfs_time_current_sec(), lock->l_last_activity),
132 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
136 EXPORT_SYMBOL(ldlm_expired_completion_wait);
138 /* We use the same basis for both server side and client side functions
139 from a single node. */
140 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
142 int timeout = at_get(ldlm_lock_to_ns_at(lock));
144 return obd_timeout / 2;
145 /* Since these are non-updating timeouts, we should be conservative.
146 It would be nice to have some kind of "early reply" mechanism for
147 lock callbacks too... */
148 timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */
149 return max(timeout, ldlm_enqueue_min);
151 EXPORT_SYMBOL(ldlm_get_enq_timeout);
154 * Helper function for ldlm_completion_ast(), updating timings when lock is
157 static int ldlm_completion_tail(struct ldlm_lock *lock)
162 if (lock->l_flags & (LDLM_FL_DESTROYED | LDLM_FL_FAILED)) {
163 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
166 delay = cfs_time_sub(cfs_time_current_sec(),
167 lock->l_last_activity);
168 LDLM_DEBUG(lock, "client-side enqueue: granted after "
169 CFS_DURATION_T"s", delay);
171 /* Update our time estimate */
172 at_measured(ldlm_lock_to_ns_at(lock),
180 * Implementation of ->l_completion_ast() for a client, that doesn't wait
181 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
182 * other threads that cannot block for long.
184 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
186 if (flags == LDLM_FL_WAIT_NOREPROC) {
187 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
191 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
192 LDLM_FL_BLOCK_CONV))) {
193 wake_up(&lock->l_waitq);
194 RETURN(ldlm_completion_tail(lock));
197 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
199 ldlm_reprocess_all(lock->l_resource);
202 EXPORT_SYMBOL(ldlm_completion_ast_async);
205 * Generic LDLM "completion" AST. This is called in several cases:
207 * - when a reply to an ENQUEUE RPC is received from the server
208 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
209 * this point (determined by flags);
211 * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
214 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
217 * - to force all locks when resource is destroyed (cleanup_resource());
219 * - during lock conversion (not used currently).
221 * If lock is not granted in the first case, this function waits until second
222 * or penultimate cases happen in some other thread.
225 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
227 /* XXX ALLOCATE - 160 bytes */
228 struct lock_wait_data lwd;
229 struct obd_device *obd;
230 struct obd_import *imp = NULL;
231 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_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
241 LDLM_FL_BLOCK_CONV))) {
242 wake_up(&lock->l_waitq);
246 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
251 obd = class_exp2obd(lock->l_conn_export);
253 /* if this is a local lock, then there is no import */
255 imp = obd->u.cli.cl_import;
258 /* Wait a long time for enqueue - server may have to callback a
259 lock from another client. Server will evict the other client if it
260 doesn't respond reasonably, and then give us the lock. */
261 timeout = ldlm_get_enq_timeout(lock) * 2;
265 if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
266 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
267 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
269 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
270 ldlm_expired_completion_wait,
271 interrupted_completion_wait, &lwd);
275 spin_lock(&imp->imp_lock);
276 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
277 spin_unlock(&imp->imp_lock);
280 if (ns_is_client(ldlm_lock_to_ns(lock)) &&
281 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
282 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
283 lock->l_flags |= LDLM_FL_FAIL_LOC;
286 /* Go to sleep until the lock is granted or cancelled. */
287 rc = l_wait_event(lock->l_waitq,
288 is_granted_or_cancelled(lock), &lwi);
292 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
297 RETURN(ldlm_completion_tail(lock));
299 EXPORT_SYMBOL(ldlm_completion_ast);
302 * A helper to build a blocking AST function
304 * Perform a common operation for blocking ASTs:
305 * defferred lock cancellation.
307 * \param lock the lock blocking or canceling AST was called on
309 * \see mdt_blocking_ast
310 * \see ldlm_blocking_ast
312 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
316 lock->l_flags |= LDLM_FL_CBPENDING;
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)
353 if (flag == LDLM_CB_CANCELING) {
354 /* Don't need to do anything here. */
358 lock_res_and_lock(lock);
359 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
360 * that ldlm_blocking_ast is called just before intent_policy method
361 * takes the lr_lock, then by the time we get the lock, we might not
362 * be the correct blocking function anymore. So check, and return
364 if (lock->l_blocking_ast != ldlm_blocking_ast) {
365 unlock_res_and_lock(lock);
368 RETURN(ldlm_blocking_ast_nocheck(lock));
370 EXPORT_SYMBOL(ldlm_blocking_ast);
373 * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
374 * comment in filter_intent_policy() on why you may need this.
376 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
379 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
380 * that is rather subtle: with OST-side locking, it may so happen that
381 * _all_ extent locks are held by the OST. If client wants to obtain
382 * current file size it calls ll{,u}_glimpse_size(), and (as locks are
383 * on the server), dummy glimpse callback fires and does
384 * nothing. Client still receives correct file size due to the
385 * following fragment in filter_intent_policy():
387 * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
388 * if (rc != 0 && res->lr_namespace->ns_lvbo &&
389 * res->lr_namespace->ns_lvbo->lvbo_update) {
390 * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
393 * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
394 * returns correct file size to the client.
396 return -ELDLM_NO_LOCK_DATA;
398 EXPORT_SYMBOL(ldlm_glimpse_ast);
401 * Enqueue a local lock (typically on a server).
403 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
404 const struct ldlm_res_id *res_id,
405 ldlm_type_t type, ldlm_policy_data_t *policy,
406 ldlm_mode_t mode, __u64 *flags,
407 ldlm_blocking_callback blocking,
408 ldlm_completion_callback completion,
409 ldlm_glimpse_callback glimpse,
410 void *data, __u32 lvb_len, enum lvb_type lvb_type,
411 const __u64 *client_cookie,
412 struct lustre_handle *lockh)
414 struct ldlm_lock *lock;
416 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
417 .lcs_blocking = blocking,
418 .lcs_glimpse = glimpse,
421 LASSERT(!(*flags & LDLM_FL_REPLAY));
422 if (unlikely(ns_is_client(ns))) {
423 CERROR("Trying to enqueue local lock in a shadow namespace\n");
427 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
430 GOTO(out_nolock, err = -ENOMEM);
432 ldlm_lock2handle(lock, lockh);
434 /* NB: we don't have any lock now (lock_res_and_lock)
435 * because it's a new lock */
436 ldlm_lock_addref_internal_nolock(lock, mode);
437 lock->l_flags |= LDLM_FL_LOCAL;
438 if (*flags & LDLM_FL_ATOMIC_CB)
439 lock->l_flags |= LDLM_FL_ATOMIC_CB;
442 lock->l_policy_data = *policy;
443 if (client_cookie != NULL)
444 lock->l_client_cookie = *client_cookie;
445 if (type == LDLM_EXTENT)
446 lock->l_req_extent = policy->l_extent;
448 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
449 if (unlikely(err != ELDLM_OK))
453 *policy = lock->l_policy_data;
455 if (lock->l_completion_ast)
456 lock->l_completion_ast(lock, *flags, NULL);
458 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
461 LDLM_LOCK_RELEASE(lock);
465 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
467 static void failed_lock_cleanup(struct ldlm_namespace *ns,
468 struct ldlm_lock *lock, int mode)
472 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
473 lock_res_and_lock(lock);
474 /* Check that lock is not granted or failed, we might race. */
475 if ((lock->l_req_mode != lock->l_granted_mode) &&
476 !(lock->l_flags & LDLM_FL_FAILED)) {
477 /* Make sure that this lock will not be found by raced
478 * bl_ast and -EINVAL reply is sent to server anyways.
480 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
481 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
484 unlock_res_and_lock(lock);
488 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
489 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
491 LDLM_DEBUG(lock, "lock was granted or failed in race");
493 ldlm_lock_decref_internal(lock, mode);
495 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
496 * from llite/file.c/ll_file_flock(). */
497 /* This code makes for the fact that we do not have blocking handler on
498 * a client for flock locks. As such this is the place where we must
499 * completely kill failed locks. (interrupted and those that
500 * were waiting to be granted when server evicted us. */
501 if (lock->l_resource->lr_type == LDLM_FLOCK) {
502 lock_res_and_lock(lock);
503 ldlm_resource_unlink_lock(lock);
504 ldlm_lock_destroy_nolock(lock);
505 unlock_res_and_lock(lock);
510 * Finishing portion of client lock enqueue code.
512 * Called after receiving reply from server.
514 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
515 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
516 __u64 *flags, void *lvb, __u32 lvb_len,
517 struct lustre_handle *lockh,int rc)
519 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
520 int is_replay = *flags & LDLM_FL_REPLAY;
521 struct ldlm_lock *lock;
522 struct ldlm_reply *reply;
523 int cleanup_phase = 1;
526 lock = ldlm_handle2lock(lockh);
527 /* ldlm_cli_enqueue is holding a reference on this lock. */
529 LASSERT(type == LDLM_FLOCK);
533 LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
534 "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
536 if (rc != ELDLM_OK) {
538 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
539 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
541 if (rc != ELDLM_LOCK_ABORTED)
545 /* Before we return, swab the reply */
546 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
548 GOTO(cleanup, rc = -EPROTO);
551 LASSERT(lvb != NULL);
553 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
556 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
557 GOTO(cleanup, rc = size);
558 } else if (unlikely(size > lvb_len)) {
559 LDLM_ERROR(lock, "Replied LVB is larger than "
560 "expectation, expected = %d, replied = %d",
562 GOTO(cleanup, rc = -EINVAL);
566 if (rc == ELDLM_LOCK_ABORTED) {
568 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
570 GOTO(cleanup, rc = (rc != 0 ? rc : ELDLM_LOCK_ABORTED));
573 /* lock enqueued on the server */
576 lock_res_and_lock(lock);
577 /* Key change rehash lock in per-export hash with new key */
578 if (exp->exp_lock_hash) {
579 /* In the function below, .hs_keycmp resolves to
580 * ldlm_export_lock_keycmp() */
581 /* coverity[overrun-buffer-val] */
582 cfs_hash_rehash_key(exp->exp_lock_hash,
583 &lock->l_remote_handle,
587 lock->l_remote_handle = reply->lock_handle;
590 *flags = ldlm_flags_from_wire(reply->lock_flags);
591 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
593 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
594 * to wait with no timeout as well */
595 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
597 unlock_res_and_lock(lock);
599 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%llx\n",
600 lock, reply->lock_handle.cookie, *flags);
602 /* If enqueue returned a blocked lock but the completion handler has
603 * already run, then it fixed up the resource and we don't need to do it
605 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
606 int newmode = reply->lock_desc.l_req_mode;
608 if (newmode && newmode != lock->l_req_mode) {
609 LDLM_DEBUG(lock, "server returned different mode %s",
610 ldlm_lockname[newmode]);
611 lock->l_req_mode = newmode;
614 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
615 lock->l_resource->lr_name.name,
616 sizeof(struct ldlm_res_id))) {
617 CDEBUG(D_INFO, "remote intent success, locking "
618 "(%ld,%ld,%ld) instead of "
620 (long)reply->lock_desc.l_resource.lr_name.name[0],
621 (long)reply->lock_desc.l_resource.lr_name.name[1],
622 (long)reply->lock_desc.l_resource.lr_name.name[2],
623 (long)lock->l_resource->lr_name.name[0],
624 (long)lock->l_resource->lr_name.name[1],
625 (long)lock->l_resource->lr_name.name[2]);
627 rc = ldlm_lock_change_resource(ns, lock,
628 &reply->lock_desc.l_resource.lr_name);
629 if (rc || lock->l_resource == NULL)
630 GOTO(cleanup, rc = -ENOMEM);
631 LDLM_DEBUG(lock, "client-side enqueue, new resource");
634 if (!(type == LDLM_IBITS &&
635 !(exp_connect_flags(exp) & OBD_CONNECT_IBITS)))
636 /* We assume lock type cannot change on server*/
637 ldlm_convert_policy_to_local(exp,
638 lock->l_resource->lr_type,
639 &reply->lock_desc.l_policy_data,
640 &lock->l_policy_data);
641 if (type != LDLM_PLAIN)
642 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
645 if ((*flags) & LDLM_FL_AST_SENT ||
646 /* Cancel extent locks as soon as possible on a liblustre client,
647 * because it cannot handle asynchronous ASTs robustly (see
649 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
650 lock_res_and_lock(lock);
651 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
652 unlock_res_and_lock(lock);
653 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
656 /* If the lock has already been granted by a completion AST, don't
657 * clobber the LVB with an older one. */
659 /* We must lock or a racing completion might update lvb without
660 * letting us know and we'll clobber the correct value.
661 * Cannot unlock after the check either, a that still leaves
662 * a tiny window for completion to get in */
663 lock_res_and_lock(lock);
664 if (lock->l_req_mode != lock->l_granted_mode)
665 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
666 lock->l_lvb_data, size);
667 unlock_res_and_lock(lock);
675 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
676 if (lock->l_completion_ast != NULL) {
677 int err = lock->l_completion_ast(lock, *flags, NULL);
685 if (lvb_len && lvb != NULL) {
686 /* Copy the LVB here, and not earlier, because the completion
687 * AST (if any) can override what we got in the reply */
688 memcpy(lvb, lock->l_lvb_data, lvb_len);
691 LDLM_DEBUG(lock, "client-side enqueue END");
694 if (cleanup_phase == 1 && rc)
695 failed_lock_cleanup(ns, lock, mode);
696 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
698 LDLM_LOCK_RELEASE(lock);
701 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
704 * Estimate number of lock handles that would fit into request of given
705 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
706 * a single page on the send/receive side. XXX: 512 should be changed to
707 * more adequate value.
709 static inline int ldlm_req_handles_avail(int req_size, int off)
713 avail = min_t(int, LDLM_MAXREQSIZE, PAGE_CACHE_SIZE - 512) - req_size;
714 if (likely(avail >= 0))
715 avail /= (int)sizeof(struct lustre_handle);
718 avail += LDLM_LOCKREQ_HANDLES - off;
723 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
724 enum req_location loc,
727 int size = req_capsule_msg_size(pill, loc);
728 return ldlm_req_handles_avail(size, off);
731 static inline int ldlm_format_handles_avail(struct obd_import *imp,
732 const struct req_format *fmt,
733 enum req_location loc, int off)
735 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
736 return ldlm_req_handles_avail(size, off);
740 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
741 * \a count locks in \a cancels.
743 * This is to be called by functions preparing their own requests that
744 * might contain lists of locks to cancel in addition to actual operation
745 * that needs to be performed.
747 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
748 int version, int opc, int canceloff,
749 struct list_head *cancels, int count)
751 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
752 struct req_capsule *pill = &req->rq_pill;
753 struct ldlm_request *dlm = NULL;
754 int flags, avail, to_free, pack = 0;
760 if (ns_connect_cancelset(ns)) {
761 /* Estimate the amount of available space in the request. */
762 req_capsule_filled_sizes(pill, RCL_CLIENT);
763 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
765 flags = ns_connect_lru_resize(ns) ?
766 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
767 to_free = !ns_connect_lru_resize(ns) &&
768 opc == LDLM_ENQUEUE ? 1 : 0;
770 /* Cancel LRU locks here _only_ if the server supports
771 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
772 * RPC, which will make us slower. */
774 count += ldlm_cancel_lru_local(ns, cancels, to_free,
775 avail - count, 0, flags);
780 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
781 ldlm_request_bufsize(pack, opc));
784 rc = ptlrpc_request_pack(req, version, opc);
786 ldlm_lock_list_put(cancels, l_bl_ast, count);
790 if (ns_connect_cancelset(ns)) {
792 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
794 /* Skip first lock handler in ldlm_request_pack(),
795 * this method will incrment @lock_count according
796 * to the lock handle amount actually written to
798 dlm->lock_count = canceloff;
800 /* Pack into the request @pack lock handles. */
801 ldlm_cli_cancel_list(cancels, pack, req, 0);
802 /* Prepare and send separate cancel RPC for others. */
803 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
805 ldlm_lock_list_put(cancels, l_bl_ast, count);
809 EXPORT_SYMBOL(ldlm_prep_elc_req);
811 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
812 struct list_head *cancels, int count)
814 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
815 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
817 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
819 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
821 struct ptlrpc_request *req;
824 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
826 RETURN(ERR_PTR(-ENOMEM));
828 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
830 ptlrpc_request_free(req);
834 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
835 ptlrpc_request_set_replen(req);
838 EXPORT_SYMBOL(ldlm_enqueue_pack);
841 * Client-side lock enqueue.
843 * If a request has some specific initialisation it is passed in \a reqp,
844 * otherwise it is created in ldlm_cli_enqueue.
846 * Supports sync and async requests, pass \a async flag accordingly. If a
847 * request was created in ldlm_cli_enqueue and it is the async request,
848 * pass it to the caller in \a reqp.
850 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
851 struct ldlm_enqueue_info *einfo,
852 const struct ldlm_res_id *res_id,
853 ldlm_policy_data_t const *policy, __u64 *flags,
854 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
855 struct lustre_handle *lockh, int async)
857 struct ldlm_namespace *ns;
858 struct ldlm_lock *lock;
859 struct ldlm_request *body;
860 int is_replay = *flags & LDLM_FL_REPLAY;
861 int req_passed_in = 1;
863 struct ptlrpc_request *req;
865 LASSERT(exp != NULL);
867 ns = exp->exp_obd->obd_namespace;
869 /* If we're replaying this lock, just check some invariants.
870 * If we're creating a new lock, get everything all setup nice. */
872 lock = ldlm_handle2lock_long(lockh, 0);
873 LASSERT(lock != NULL);
874 LDLM_DEBUG(lock, "client-side enqueue START");
875 LASSERT(exp == lock->l_conn_export);
877 const struct ldlm_callback_suite cbs = {
878 .lcs_completion = einfo->ei_cb_cp,
879 .lcs_blocking = einfo->ei_cb_bl,
880 .lcs_glimpse = einfo->ei_cb_gl
882 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
883 einfo->ei_mode, &cbs, einfo->ei_cbdata,
887 /* for the local lock, add the reference */
888 ldlm_lock_addref_internal(lock, einfo->ei_mode);
889 ldlm_lock2handle(lock, lockh);
890 if (policy != NULL) {
891 /* INODEBITS_INTEROP: If the server does not support
892 * inodebits, we will request a plain lock in the
893 * descriptor (ldlm_lock2desc() below) but use an
894 * inodebits lock internally with both bits set.
896 if (einfo->ei_type == LDLM_IBITS &&
897 !(exp_connect_flags(exp) &
899 lock->l_policy_data.l_inodebits.bits =
900 MDS_INODELOCK_LOOKUP |
901 MDS_INODELOCK_UPDATE;
903 lock->l_policy_data = *policy;
906 if (einfo->ei_type == LDLM_EXTENT)
907 lock->l_req_extent = policy->l_extent;
908 LDLM_DEBUG(lock, "client-side enqueue START, flags %llx\n",
912 lock->l_conn_export = exp;
913 lock->l_export = NULL;
914 lock->l_blocking_ast = einfo->ei_cb_bl;
915 lock->l_flags |= (*flags & LDLM_FL_NO_LRU);
917 /* lock not sent to server yet */
919 if (reqp == NULL || *reqp == NULL) {
920 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
925 failed_lock_cleanup(ns, lock, einfo->ei_mode);
926 LDLM_LOCK_RELEASE(lock);
936 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
938 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
939 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
942 /* Dump lock data into the request buffer */
943 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
944 ldlm_lock2desc(lock, &body->lock_desc);
945 body->lock_flags = ldlm_flags_to_wire(*flags);
946 body->lock_handle[0] = *lockh;
948 /* Continue as normal. */
949 if (!req_passed_in) {
951 req_capsule_extend(&req->rq_pill,
952 &RQF_LDLM_ENQUEUE_LVB);
953 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
955 ptlrpc_request_set_replen(req);
959 * Liblustre client doesn't get extent locks, except for O_APPEND case
960 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
961 * [i_size, OBD_OBJECT_EOF] lock is taken.
963 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
964 policy->l_extent.end == OBD_OBJECT_EOF));
967 LASSERT(reqp != NULL);
971 LDLM_DEBUG(lock, "sending request");
973 rc = ptlrpc_queue_wait(req);
975 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
976 einfo->ei_mode, flags, lvb, lvb_len,
979 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
980 * one reference that we took */
982 LDLM_LOCK_RELEASE(lock);
986 if (!req_passed_in && req != NULL) {
987 ptlrpc_req_finished(req);
994 EXPORT_SYMBOL(ldlm_cli_enqueue);
996 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
999 struct ldlm_resource *res;
1002 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1003 CERROR("Trying to cancel local lock\n");
1006 LDLM_DEBUG(lock, "client-side local convert");
1008 res = ldlm_lock_convert(lock, new_mode, flags);
1010 ldlm_reprocess_all(res);
1013 rc = LUSTRE_EDEADLK;
1015 LDLM_DEBUG(lock, "client-side local convert handler END");
1016 LDLM_LOCK_PUT(lock);
1020 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
1021 * conversion of locks which are on the waiting or converting queue */
1022 /* Caller of this code is supposed to take care of lock readers/writers
1024 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
1026 struct ldlm_request *body;
1027 struct ldlm_reply *reply;
1028 struct ldlm_lock *lock;
1029 struct ldlm_resource *res;
1030 struct ptlrpc_request *req;
1033 lock = ldlm_handle2lock(lockh);
1040 if (lock->l_conn_export == NULL)
1041 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
1043 LDLM_DEBUG(lock, "client-side convert");
1045 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
1046 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1049 LDLM_LOCK_PUT(lock);
1053 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1054 body->lock_handle[0] = lock->l_remote_handle;
1056 body->lock_desc.l_req_mode = new_mode;
1057 body->lock_flags = ldlm_flags_to_wire(*flags);
1060 ptlrpc_request_set_replen(req);
1061 rc = ptlrpc_queue_wait(req);
1065 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1067 GOTO(out, rc = -EPROTO);
1070 GOTO(out, rc = req->rq_status);
1072 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
1074 ldlm_reprocess_all(res);
1075 /* Go to sleep until the lock is granted. */
1076 /* FIXME: or cancelled. */
1077 if (lock->l_completion_ast) {
1078 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1084 rc = LUSTRE_EDEADLK;
1088 LDLM_LOCK_PUT(lock);
1089 ptlrpc_req_finished(req);
1092 EXPORT_SYMBOL(ldlm_cli_convert);
1095 * Cancel locks locally.
1097 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1098 * \retval LDLM_FL_CANCELING otherwise;
1099 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1101 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1103 __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 lock->l_flags |= LDLM_FL_CBPENDING;
1112 local_only = !!(lock->l_flags &
1113 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1114 ldlm_cancel_callback(lock);
1115 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
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;
1148 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1149 LASSERT(dlm != NULL);
1151 /* Check the room in the request buffer. */
1152 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1153 sizeof(struct ldlm_request);
1154 max /= sizeof(struct lustre_handle);
1155 max += LDLM_LOCKREQ_HANDLES;
1156 LASSERT(max >= dlm->lock_count + count);
1158 /* XXX: it would be better to pack lock handles grouped by resource.
1159 * so that the server cancel would call filter_lvbo_update() less
1161 list_for_each_entry(lock, head, l_bl_ast) {
1164 LASSERT(lock->l_conn_export);
1165 /* Pack the lock handle to the given request buffer. */
1166 LDLM_DEBUG(lock, "packing");
1167 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1170 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1175 * Prepare and send a batched cancel RPC. It will include \a count lock
1176 * handles of locks given in \a cancels list. */
1177 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1178 int count, ldlm_cancel_flags_t flags)
1180 struct ptlrpc_request *req = NULL;
1181 struct obd_import *imp;
1185 LASSERT(exp != NULL);
1188 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1190 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1193 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1194 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1199 imp = class_exp2cliimp(exp);
1200 if (imp == NULL || imp->imp_invalid) {
1202 "skipping cancel on invalid import %p\n", imp);
1206 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1208 GOTO(out, rc = -ENOMEM);
1210 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1211 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1212 ldlm_request_bufsize(count, LDLM_CANCEL));
1214 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1216 ptlrpc_request_free(req);
1220 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1221 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1222 ptlrpc_at_set_req_timeout(req);
1224 ldlm_cancel_pack(req, cancels, count);
1226 ptlrpc_request_set_replen(req);
1227 if (flags & LCF_ASYNC) {
1228 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1232 rc = ptlrpc_queue_wait(req);
1234 if (rc == LUSTRE_ESTALE) {
1235 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1236 "out of sync -- not fatal\n",
1237 libcfs_nid2str(req->rq_import->
1238 imp_connection->c_peer.nid));
1240 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1241 req->rq_import_generation == imp->imp_generation) {
1242 ptlrpc_req_finished(req);
1244 } else if (rc != ELDLM_OK) {
1245 /* -ESHUTDOWN is common on umount */
1246 CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1247 "Got rc %d from cancel RPC: "
1248 "canceling anyway\n", rc);
1255 ptlrpc_req_finished(req);
1258 return sent ? sent : rc;
1260 EXPORT_SYMBOL(ldlm_cli_cancel_req);
1262 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1264 LASSERT(imp != NULL);
1265 return &imp->imp_obd->obd_namespace->ns_pool;
1269 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1271 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1273 struct obd_device *obd;
1277 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1278 !imp_connect_lru_resize(req->rq_import)))
1281 * Do nothing for corner cases.
1286 /* In some cases RPC may contain SLV and limit zeroed out. This
1287 * is the case when server does not support LRU resize feature.
1288 * This is also possible in some recovery cases when server-side
1289 * reqs have no reference to the OBD export and thus access to
1290 * server-side namespace is not possible. */
1291 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1292 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1293 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1294 "(SLV: "LPU64", Limit: %u)",
1295 lustre_msg_get_slv(req->rq_repmsg),
1296 lustre_msg_get_limit(req->rq_repmsg));
1300 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1301 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1302 obd = req->rq_import->imp_obd;
1304 /* Set new SLV and limit in OBD fields to make them accessible
1305 * to the pool thread. We do not access obd_namespace and pool
1306 * directly here as there is no reliable way to make sure that
1307 * they are still alive at cleanup time. Evil races are possible
1308 * which may cause Oops at that time. */
1309 write_lock(&obd->obd_pool_lock);
1310 obd->obd_pool_slv = new_slv;
1311 obd->obd_pool_limit = new_limit;
1312 write_unlock(&obd->obd_pool_lock);
1316 EXPORT_SYMBOL(ldlm_cli_update_pool);
1319 * Client side lock cancel.
1321 * Lock must not have any readers or writers by this time.
1323 int ldlm_cli_cancel(struct lustre_handle *lockh,
1324 ldlm_cancel_flags_t cancel_flags)
1326 struct obd_export *exp;
1327 int avail, flags, count = 1;
1329 struct ldlm_namespace *ns;
1330 struct ldlm_lock *lock;
1333 /* concurrent cancels on the same handle can happen */
1334 lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1336 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1340 rc = ldlm_cli_cancel_local(lock);
1341 if (rc == LDLM_FL_LOCAL_ONLY) {
1342 LDLM_LOCK_RELEASE(lock);
1345 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1346 * RPC which goes to canceld portal, so we can cancel other LRU locks
1347 * here and send them all as one LDLM_CANCEL RPC. */
1348 LASSERT(list_empty(&lock->l_bl_ast));
1349 list_add(&lock->l_bl_ast, &cancels);
1351 exp = lock->l_conn_export;
1352 if (exp_connect_cancelset(exp)) {
1353 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1358 ns = ldlm_lock_to_ns(lock);
1359 flags = ns_connect_lru_resize(ns) ?
1360 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1361 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1364 ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1367 EXPORT_SYMBOL(ldlm_cli_cancel);
1370 * Locally cancel up to \a count locks in list \a cancels.
1371 * Return the number of cancelled locks.
1373 int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
1374 ldlm_cancel_flags_t flags)
1377 struct ldlm_lock *lock, *next;
1378 int left = 0, bl_ast = 0;
1382 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1386 if (flags & LCF_LOCAL) {
1387 rc = LDLM_FL_LOCAL_ONLY;
1388 ldlm_lock_cancel(lock);
1390 rc = ldlm_cli_cancel_local(lock);
1392 /* Until we have compound requests and can send LDLM_CANCEL
1393 * requests batched with generic RPCs, we need to send cancels
1394 * with the LDLM_FL_BL_AST flag in a separate RPC from
1395 * the one being generated now. */
1396 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1397 LDLM_DEBUG(lock, "Cancel lock separately");
1398 list_del_init(&lock->l_bl_ast);
1399 list_add(&lock->l_bl_ast, &head);
1403 if (rc == LDLM_FL_LOCAL_ONLY) {
1404 /* CANCEL RPC should not be sent to server. */
1405 list_del_init(&lock->l_bl_ast);
1406 LDLM_LOCK_RELEASE(lock);
1412 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1417 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
1420 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1421 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1422 * readahead requests, ...)
1424 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1425 struct ldlm_lock *lock,
1426 int unused, int added,
1429 ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1430 ldlm_cancel_for_recovery cb = ns->ns_cancel_for_recovery;
1431 lock_res_and_lock(lock);
1433 /* don't check added & count since we want to process all locks
1434 * from unused list */
1435 switch (lock->l_resource->lr_type) {
1441 result = LDLM_POLICY_SKIP_LOCK;
1442 lock->l_flags |= LDLM_FL_SKIPPED;
1446 unlock_res_and_lock(lock);
1451 * Callback function for LRU-resize policy. Decides whether to keep
1452 * \a lock in LRU for current \a LRU size \a unused, added in current
1453 * scan \a added and number of locks to be preferably canceled \a count.
1455 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1457 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1459 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1460 struct ldlm_lock *lock,
1461 int unused, int added,
1464 cfs_time_t cur = cfs_time_current();
1465 struct ldlm_pool *pl = &ns->ns_pool;
1469 /* Stop LRU processing when we reach past @count or have checked all
1471 if (count && added >= count)
1472 return LDLM_POLICY_KEEP_LOCK;
1474 slv = ldlm_pool_get_slv(pl);
1475 lvf = ldlm_pool_get_lvf(pl);
1476 la = cfs_duration_sec(cfs_time_sub(cur,
1477 lock->l_last_used));
1478 lv = lvf * la * unused;
1480 /* Inform pool about current CLV to see it via proc. */
1481 ldlm_pool_set_clv(pl, lv);
1483 /* Stop when SLV is not yet come from server or lv is smaller than
1485 return (slv == 0 || lv < slv) ?
1486 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1490 * Callback function for proc used policy. Makes decision whether to keep
1491 * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1492 * added and number of locks to be preferably canceled \a count.
1494 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1496 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1498 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1499 struct ldlm_lock *lock,
1500 int unused, int added,
1503 /* Stop LRU processing when we reach past @count or have checked all
1505 return (added >= count) ?
1506 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1510 * Callback function for aged policy. Makes decision whether to keep \a lock in
1511 * LRU for current LRU size \a unused, added in current scan \a added and
1512 * number of locks to be preferably canceled \a count.
1514 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1516 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1518 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1519 struct ldlm_lock *lock,
1520 int unused, int added,
1523 /* Stop LRU processing if young lock is found and we reach past count */
1524 return ((added >= count) &&
1525 cfs_time_before(cfs_time_current(),
1526 cfs_time_add(lock->l_last_used,
1528 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1532 * Callback function for default policy. Makes decision whether to keep \a lock
1533 * in LRU for current LRU size \a unused, added in current scan \a added and
1534 * number of locks to be preferably canceled \a count.
1536 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1538 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1540 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1541 struct ldlm_lock *lock,
1542 int unused, int added,
1545 /* Stop LRU processing when we reach past count or have checked all
1547 return (added >= count) ?
1548 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1551 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1552 struct ldlm_lock *, int,
1555 static ldlm_cancel_lru_policy_t
1556 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1558 if (flags & LDLM_CANCEL_NO_WAIT)
1559 return ldlm_cancel_no_wait_policy;
1561 if (ns_connect_lru_resize(ns)) {
1562 if (flags & LDLM_CANCEL_SHRINK)
1563 /* We kill passed number of old locks. */
1564 return ldlm_cancel_passed_policy;
1565 else if (flags & LDLM_CANCEL_LRUR)
1566 return ldlm_cancel_lrur_policy;
1567 else if (flags & LDLM_CANCEL_PASSED)
1568 return ldlm_cancel_passed_policy;
1570 if (flags & LDLM_CANCEL_AGED)
1571 return ldlm_cancel_aged_policy;
1574 return ldlm_cancel_default_policy;
1578 * - Free space in LRU for \a count new locks,
1579 * redundant unused locks are canceled locally;
1580 * - also cancel locally unused aged locks;
1581 * - do not cancel more than \a max locks;
1582 * - GET the found locks and add them into the \a cancels list.
1584 * A client lock can be added to the l_bl_ast list only when it is
1585 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1586 * CANCEL. There are the following use cases:
1587 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1588 * ldlm_cli_cancel(), which check and set this flag properly. As any
1589 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1590 * later without any special locking.
1592 * Calling policies for enabled LRU resize:
1593 * ----------------------------------------
1594 * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to
1595 * cancel not more than \a count locks;
1597 * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at
1598 * the beginning of LRU list);
1600 * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to
1601 * memory pressre policy function;
1603 * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy".
1605 * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1606 * (typically before replaying locks) w/o
1607 * sending any RPCs or waiting for any
1608 * outstanding RPC to complete.
1610 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, struct list_head *cancels,
1611 int count, int max, int flags)
1613 ldlm_cancel_lru_policy_t pf;
1614 struct ldlm_lock *lock, *next;
1615 int added = 0, unused, remained;
1617 spin_lock(&ns->ns_lock);
1618 unused = ns->ns_nr_unused;
1621 if (!ns_connect_lru_resize(ns))
1622 count += unused - ns->ns_max_unused;
1624 pf = ldlm_cancel_lru_policy(ns, flags);
1625 LASSERT(pf != NULL);
1627 while (!list_empty(&ns->ns_unused_list)) {
1628 ldlm_policy_res_t result;
1630 /* all unused locks */
1631 if (remained-- <= 0)
1634 /* For any flags, stop scanning if @max is reached. */
1635 if (max && added >= max)
1638 list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1640 /* No locks which got blocking requests. */
1641 LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1643 if (flags & LDLM_CANCEL_NO_WAIT &&
1644 lock->l_flags & LDLM_FL_SKIPPED)
1645 /* already processed */
1648 /* Somebody is already doing CANCEL. No need for this
1649 * lock in LRU, do not traverse it again. */
1650 if (!(lock->l_flags & LDLM_FL_CANCELING))
1653 ldlm_lock_remove_from_lru_nolock(lock);
1655 if (&lock->l_lru == &ns->ns_unused_list)
1658 LDLM_LOCK_GET(lock);
1659 spin_unlock(&ns->ns_lock);
1660 lu_ref_add(&lock->l_reference, __FUNCTION__, current);
1662 /* Pass the lock through the policy filter and see if it
1663 * should stay in LRU.
1665 * Even for shrinker policy we stop scanning if
1666 * we find a lock that should stay in the cache.
1667 * We should take into account lock age anyway
1668 * as a new lock is a valuable resource even if
1669 * it has a low weight.
1671 * That is, for shrinker policy we drop only
1672 * old locks, but additionally choose them by
1673 * their weight. Big extent locks will stay in
1675 result = pf(ns, lock, unused, added, count);
1676 if (result == LDLM_POLICY_KEEP_LOCK) {
1677 lu_ref_del(&lock->l_reference,
1678 __FUNCTION__, current);
1679 LDLM_LOCK_RELEASE(lock);
1680 spin_lock(&ns->ns_lock);
1683 if (result == LDLM_POLICY_SKIP_LOCK) {
1684 lu_ref_del(&lock->l_reference,
1686 LDLM_LOCK_RELEASE(lock);
1687 spin_lock(&ns->ns_lock);
1691 lock_res_and_lock(lock);
1692 /* Check flags again under the lock. */
1693 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1694 (ldlm_lock_remove_from_lru(lock) == 0)) {
1695 /* Another thread is removing lock from LRU, or
1696 * somebody is already doing CANCEL, or there
1697 * is a blocking request which will send cancel
1698 * by itself, or the lock is no longer unused. */
1699 unlock_res_and_lock(lock);
1700 lu_ref_del(&lock->l_reference,
1701 __FUNCTION__, current);
1702 LDLM_LOCK_RELEASE(lock);
1703 spin_lock(&ns->ns_lock);
1706 LASSERT(!lock->l_readers && !lock->l_writers);
1708 /* If we have chosen to cancel this lock voluntarily, we
1709 * better send cancel notification to server, so that it
1710 * frees appropriate state. This might lead to a race
1711 * where while we are doing cancel here, server is also
1712 * silently cancelling this lock. */
1713 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1715 /* Setting the CBPENDING flag is a little misleading,
1716 * but prevents an important race; namely, once
1717 * CBPENDING is set, the lock can accumulate no more
1718 * readers/writers. Since readers and writers are
1719 * already zero here, ldlm_lock_decref() won't see
1720 * this flag and call l_blocking_ast */
1721 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1723 /* We can't re-add to l_lru as it confuses the
1724 * refcounting in ldlm_lock_remove_from_lru() if an AST
1725 * arrives after we drop lr_lock below. We use l_bl_ast
1726 * and can't use l_pending_chain as it is used both on
1727 * server and client nevertheless bug 5666 says it is
1728 * used only on server */
1729 LASSERT(list_empty(&lock->l_bl_ast));
1730 list_add(&lock->l_bl_ast, cancels);
1731 unlock_res_and_lock(lock);
1732 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1733 spin_lock(&ns->ns_lock);
1737 spin_unlock(&ns->ns_lock);
1741 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1742 int count, int max, ldlm_cancel_flags_t cancel_flags,
1746 added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1749 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1753 * Cancel at least \a nr locks from given namespace LRU.
1755 * When called with LCF_ASYNC the blocking callback will be handled
1756 * in a thread and this function will return after the thread has been
1757 * asked to call the callback. When called with LCF_ASYNC the blocking
1758 * callback will be performed in this function.
1760 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
1761 ldlm_cancel_flags_t cancel_flags,
1767 /* Just prepare the list of locks, do not actually cancel them yet.
1768 * Locks are cancelled later in a separate thread. */
1769 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1770 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
1778 * Find and cancel locally unused locks found on resource, matched to the
1779 * given policy, mode. GET the found locks and add them into the \a cancels
1782 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1783 struct list_head *cancels,
1784 ldlm_policy_data_t *policy,
1785 ldlm_mode_t mode, int lock_flags,
1786 ldlm_cancel_flags_t cancel_flags, void *opaque)
1788 struct ldlm_lock *lock;
1792 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1793 if (opaque != NULL && lock->l_ast_data != opaque) {
1794 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1795 lock->l_ast_data, opaque);
1800 if (lock->l_readers || lock->l_writers)
1803 /* If somebody is already doing CANCEL, or blocking AST came,
1804 * skip this lock. */
1805 if (lock->l_flags & LDLM_FL_BL_AST ||
1806 lock->l_flags & LDLM_FL_CANCELING)
1809 if (lockmode_compat(lock->l_granted_mode, mode))
1812 /* If policy is given and this is IBITS lock, add to list only
1813 * those locks that match by policy. */
1814 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1815 !(lock->l_policy_data.l_inodebits.bits &
1816 policy->l_inodebits.bits))
1819 /* See CBPENDING comment in ldlm_cancel_lru */
1820 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1823 LASSERT(list_empty(&lock->l_bl_ast));
1824 list_add(&lock->l_bl_ast, cancels);
1825 LDLM_LOCK_GET(lock);
1830 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
1832 EXPORT_SYMBOL(ldlm_cancel_resource_local);
1835 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1837 * If \a req is NULL, send CANCEL request to server with handles of locks
1838 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1839 * separately per lock.
1840 * If \a req is not NULL, put handles of locks in \a cancels into the request
1841 * buffer at the offset \a off.
1842 * Destroy \a cancels at the end.
1844 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1845 struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1847 struct ldlm_lock *lock;
1850 if (list_empty(cancels) || count == 0)
1853 /* XXX: requests (both batched and not) could be sent in parallel.
1854 * Usually it is enough to have just 1 RPC, but it is possible that
1855 * there are too many locks to be cancelled in LRU or on a resource.
1856 * It would also speed up the case when the server does not support
1859 LASSERT(!list_empty(cancels));
1860 lock = list_entry(cancels->next, struct ldlm_lock,
1862 LASSERT(lock->l_conn_export);
1864 if (exp_connect_cancelset(lock->l_conn_export)) {
1867 ldlm_cancel_pack(req, cancels, count);
1869 res = ldlm_cli_cancel_req(lock->l_conn_export,
1873 res = ldlm_cli_cancel_req(lock->l_conn_export,
1878 CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1879 "ldlm_cli_cancel_list: %d\n", res);
1884 ldlm_lock_list_put(cancels, l_bl_ast, res);
1886 LASSERT(count == 0);
1889 EXPORT_SYMBOL(ldlm_cli_cancel_list);
1892 * Cancel all locks on a resource that have 0 readers/writers.
1894 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1895 * to notify the server. */
1896 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1897 const struct ldlm_res_id *res_id,
1898 ldlm_policy_data_t *policy,
1900 ldlm_cancel_flags_t flags,
1903 struct ldlm_resource *res;
1908 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1910 /* This is not a problem. */
1911 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1915 LDLM_RESOURCE_ADDREF(res);
1916 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1917 0, flags | LCF_BL_AST, opaque);
1918 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1920 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1922 LDLM_RESOURCE_DELREF(res);
1923 ldlm_resource_putref(res);
1926 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
1928 struct ldlm_cli_cancel_arg {
1933 static int ldlm_cli_hash_cancel_unused(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1934 struct hlist_node *hnode, void *arg)
1936 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1937 struct ldlm_cli_cancel_arg *lc = arg;
1940 rc = ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1942 lc->lc_flags, lc->lc_opaque);
1944 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1945 res->lr_name.name[0], rc);
1947 /* must return 0 for hash iteration */
1952 * Cancel all locks on a namespace (or a specific resource, if given)
1953 * that have 0 readers/writers.
1955 * If flags & LCF_LOCAL, throw the locks away without trying
1956 * to notify the server. */
1957 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1958 const struct ldlm_res_id *res_id,
1959 ldlm_cancel_flags_t flags, void *opaque)
1961 struct ldlm_cli_cancel_arg arg = {
1963 .lc_opaque = opaque,
1969 if (res_id != NULL) {
1970 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1974 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1975 ldlm_cli_hash_cancel_unused, &arg);
1979 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1981 /* Lock iterators. */
1983 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1986 struct list_head *tmp, *next;
1987 struct ldlm_lock *lock;
1988 int rc = LDLM_ITER_CONTINUE;
1991 RETURN(LDLM_ITER_CONTINUE);
1994 list_for_each_safe(tmp, next, &res->lr_granted) {
1995 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1997 if (iter(lock, closure) == LDLM_ITER_STOP)
1998 GOTO(out, rc = LDLM_ITER_STOP);
2001 list_for_each_safe(tmp, next, &res->lr_converting) {
2002 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2004 if (iter(lock, closure) == LDLM_ITER_STOP)
2005 GOTO(out, rc = LDLM_ITER_STOP);
2008 list_for_each_safe(tmp, next, &res->lr_waiting) {
2009 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2011 if (iter(lock, closure) == LDLM_ITER_STOP)
2012 GOTO(out, rc = LDLM_ITER_STOP);
2018 EXPORT_SYMBOL(ldlm_resource_foreach);
2020 struct iter_helper_data {
2021 ldlm_iterator_t iter;
2025 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2027 struct iter_helper_data *helper = closure;
2028 return helper->iter(lock, helper->closure);
2031 static int ldlm_res_iter_helper(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2032 struct hlist_node *hnode, void *arg)
2035 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2037 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2041 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2042 ldlm_iterator_t iter, void *closure)
2045 struct iter_helper_data helper = {
2050 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2051 ldlm_res_iter_helper, &helper);
2054 EXPORT_SYMBOL(ldlm_namespace_foreach);
2056 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2057 * return 0: find no resource
2058 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2061 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2062 const struct ldlm_res_id *res_id,
2063 ldlm_iterator_t iter, void *data)
2065 struct ldlm_resource *res;
2069 CERROR("must pass in namespace\n");
2073 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2077 LDLM_RESOURCE_ADDREF(res);
2078 rc = ldlm_resource_foreach(res, iter, data);
2079 LDLM_RESOURCE_DELREF(res);
2080 ldlm_resource_putref(res);
2083 EXPORT_SYMBOL(ldlm_resource_iterate);
2087 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2089 struct list_head *list = closure;
2091 /* we use l_pending_chain here, because it's unused on clients. */
2092 LASSERTF(list_empty(&lock->l_pending_chain),
2093 "lock %p next %p prev %p\n",
2094 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2095 /* bug 9573: don't replay locks left after eviction, or
2096 * bug 17614: locks being actively cancelled. Get a reference
2097 * on a lock so that it does not disapear under us (e.g. due to cancel)
2099 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
2100 list_add(&lock->l_pending_chain, list);
2101 LDLM_LOCK_GET(lock);
2104 return LDLM_ITER_CONTINUE;
2107 static int replay_lock_interpret(const struct lu_env *env,
2108 struct ptlrpc_request *req,
2109 struct ldlm_async_args *aa, int rc)
2111 struct ldlm_lock *lock;
2112 struct ldlm_reply *reply;
2113 struct obd_export *exp;
2115 atomic_dec(&req->rq_import->imp_replay_inflight);
2120 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2122 GOTO(out, rc = -EPROTO);
2124 lock = ldlm_handle2lock(&aa->lock_handle);
2126 CERROR("received replay ack for unknown local cookie "LPX64
2127 " remote cookie "LPX64 " from server %s id %s\n",
2128 aa->lock_handle.cookie, reply->lock_handle.cookie,
2129 req->rq_export->exp_client_uuid.uuid,
2130 libcfs_id2str(req->rq_peer));
2131 GOTO(out, rc = -ESTALE);
2134 /* Key change rehash lock in per-export hash with new key */
2135 exp = req->rq_export;
2136 if (exp && exp->exp_lock_hash) {
2137 /* In the function below, .hs_keycmp resolves to
2138 * ldlm_export_lock_keycmp() */
2139 /* coverity[overrun-buffer-val] */
2140 cfs_hash_rehash_key(exp->exp_lock_hash,
2141 &lock->l_remote_handle,
2142 &reply->lock_handle,
2145 lock->l_remote_handle = reply->lock_handle;
2148 LDLM_DEBUG(lock, "replayed lock:");
2149 ptlrpc_import_recovery_state_machine(req->rq_import);
2150 LDLM_LOCK_PUT(lock);
2153 ptlrpc_connect_import(req->rq_import);
2158 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2160 struct ptlrpc_request *req;
2161 struct ldlm_async_args *aa;
2162 struct ldlm_request *body;
2165 /* Bug 11974: Do not replay a lock which is actively being canceled */
2166 if (lock->l_flags & LDLM_FL_CANCELING) {
2167 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2171 /* If this is reply-less callback lock, we cannot replay it, since
2172 * server might have long dropped it, but notification of that event was
2173 * lost by network. (and server granted conflicting lock already) */
2174 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2175 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2176 ldlm_lock_cancel(lock);
2181 * If granted mode matches the requested mode, this lock is granted.
2183 * If they differ, but we have a granted mode, then we were granted
2184 * one mode and now want another: ergo, converting.
2186 * If we haven't been granted anything and are on a resource list,
2187 * then we're blocked/waiting.
2189 * If we haven't been granted anything and we're NOT on a resource list,
2190 * then we haven't got a reply yet and don't have a known disposition.
2191 * This happens whenever a lock enqueue is the request that triggers
2194 if (lock->l_granted_mode == lock->l_req_mode)
2195 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2196 else if (lock->l_granted_mode)
2197 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2198 else if (!list_empty(&lock->l_res_link))
2199 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2201 flags = LDLM_FL_REPLAY;
2203 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2204 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2208 /* We're part of recovery, so don't wait for it. */
2209 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2211 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2212 ldlm_lock2desc(lock, &body->lock_desc);
2213 body->lock_flags = ldlm_flags_to_wire(flags);
2215 ldlm_lock2handle(lock, &body->lock_handle[0]);
2216 if (lock->l_lvb_len > 0)
2217 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2218 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2220 ptlrpc_request_set_replen(req);
2221 /* notify the server we've replayed all requests.
2222 * also, we mark the request to be put on a dedicated
2223 * queue to be processed after all request replayes.
2225 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2227 LDLM_DEBUG(lock, "replaying lock:");
2229 atomic_inc(&req->rq_import->imp_replay_inflight);
2230 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2231 aa = ptlrpc_req_async_args(req);
2232 aa->lock_handle = body->lock_handle[0];
2233 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2234 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2240 * Cancel as many unused locks as possible before replay. since we are
2241 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2244 * Called only in recovery before replaying locks. there is no need to
2245 * replay locks that are unused. since the clients may hold thousands of
2246 * cached unused locks, dropping the unused locks can greatly reduce the
2247 * load on the servers at recovery time.
2249 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2254 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2255 "replay for namespace %s (%d)\n",
2256 ldlm_ns_name(ns), ns->ns_nr_unused);
2258 /* We don't need to care whether or not LRU resize is enabled
2259 * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2260 * count parameter */
2261 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2262 LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2264 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2265 canceled, ldlm_ns_name(ns));
2268 int ldlm_replay_locks(struct obd_import *imp)
2270 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2272 struct ldlm_lock *lock, *next;
2275 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2277 /* don't replay locks if import failed recovery */
2278 if (imp->imp_vbr_failed)
2281 /* ensure this doesn't fall to 0 before all have been queued */
2282 atomic_inc(&imp->imp_replay_inflight);
2284 if (ldlm_cancel_unused_locks_before_replay)
2285 ldlm_cancel_unused_locks_for_replay(ns);
2287 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2289 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2290 list_del_init(&lock->l_pending_chain);
2292 LDLM_LOCK_RELEASE(lock);
2293 continue; /* or try to do the rest? */
2295 rc = replay_one_lock(imp, lock);
2296 LDLM_LOCK_RELEASE(lock);
2299 atomic_dec(&imp->imp_replay_inflight);
2303 EXPORT_SYMBOL(ldlm_replay_locks);