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, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 * This file contains Asynchronous System Trap (AST) handlers and related
38 * LDLM request-processing routines.
40 * An AST is a callback issued on a lock when its state is changed. There are
41 * several different types of ASTs (callbacks) registered for each lock:
43 * - completion AST: when a lock is enqueued by some process, but cannot be
44 * granted immediately due to other conflicting locks on the same resource,
45 * the completion AST is sent to notify the caller when the lock is
48 * - blocking AST: when a lock is granted to some process, if another process
49 * enqueues a conflicting (blocking) lock on a resource, a blocking AST is
50 * sent to notify the holder(s) of the lock(s) of the conflicting lock
51 * request. The lock holder(s) must release their lock(s) on that resource in
52 * a timely manner or be evicted by the server.
54 * - glimpse AST: this is used when a process wants information about a lock
55 * (i.e. the lock value block (LVB)) but does not necessarily require holding
56 * the lock. If the resource is locked, the lock holder(s) are sent glimpse
57 * ASTs and the LVB is returned to the caller, and lock holder(s) may CANCEL
58 * their lock(s) if they are idle. If the resource is not locked, the server
62 #define DEBUG_SUBSYSTEM S_LDLM
64 #include <lustre_dlm.h>
65 #include <obd_class.h>
68 #include "ldlm_internal.h"
70 unsigned int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
71 module_param(ldlm_enqueue_min, uint, 0644);
72 MODULE_PARM_DESC(ldlm_enqueue_min, "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;
91 * ldlm_request_bufsize
93 * If opcode=LDLM_ENQUEUE, 1 slot is already occupied,
94 * LDLM_LOCKREQ_HANDLE -1 slots are available.
95 * Otherwise, LDLM_LOCKREQ_HANDLE slots are available.
100 * \retval size of the request buffer
103 int ldlm_request_bufsize(int count, int type)
105 int avail = LDLM_LOCKREQ_HANDLES;
106 if (type == LDLM_ENQUEUE)
107 avail -= LDLM_ENQUEUE_CANCEL_OFF;
110 avail = (count - avail) * sizeof(struct lustre_handle);
114 return sizeof(struct ldlm_request) + avail;
117 int ldlm_expired_completion_wait(void *data)
119 struct lock_wait_data *lwd = data;
120 struct ldlm_lock *lock = lwd->lwd_lock;
121 struct obd_import *imp;
122 struct obd_device *obd;
125 if (lock->l_conn_export == NULL) {
126 static cfs_time_t next_dump = 0, last_dump = 0;
128 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
129 CFS_DURATION_T"s ago); not entering recovery in "
130 "server code, just going back to sleep",
131 lock->l_last_activity,
132 cfs_time_sub(cfs_time_current_sec(),
133 lock->l_last_activity));
134 if (cfs_time_after(cfs_time_current(), next_dump)) {
135 last_dump = next_dump;
136 next_dump = cfs_time_shift(300);
137 ldlm_namespace_dump(D_DLMTRACE,
138 ldlm_lock_to_ns(lock));
140 libcfs_debug_dumplog();
145 obd = lock->l_conn_export->exp_obd;
146 imp = obd->u.cli.cl_import;
147 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
148 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
149 CFS_DURATION_T"s ago), entering recovery for %s@%s",
150 lock->l_last_activity,
151 cfs_time_sub(cfs_time_current_sec(), lock->l_last_activity),
152 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
158 * Calculate the Completion timeout (covering enqueue, BL AST, data flush,
159 * lock cancel, and their replies). Used for lock completion timeout on the
162 * \param[in] lock lock which is waiting the completion callback
164 * \retval timeout in seconds to wait for the server reply
167 /* We use the same basis for both server side and client side functions
168 from a single node. */
169 static unsigned int ldlm_cp_timeout(struct ldlm_lock *lock)
171 unsigned int timeout;
176 /* Wait a long time for enqueue - server may have to callback a
177 * lock from another client. Server will evict the other client if it
178 * doesn't respond reasonably, and then give us the lock. */
179 timeout = at_get(ldlm_lock_to_ns_at(lock));
180 return max(3 * timeout, ldlm_enqueue_min);
184 * Helper function for ldlm_completion_ast(), updating timings when lock is
187 static int ldlm_completion_tail(struct ldlm_lock *lock, void *data)
192 if (ldlm_is_destroyed(lock) || ldlm_is_failed(lock)) {
193 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
195 } else if (data == NULL) {
196 LDLM_DEBUG(lock, "client-side enqueue: granted");
198 /* Take into AT only CP RPC, not immediately granted locks */
199 delay = cfs_time_sub(cfs_time_current_sec(),
200 lock->l_last_activity);
201 LDLM_DEBUG(lock, "client-side enqueue: granted after "
202 CFS_DURATION_T"s", delay);
204 /* Update our time estimate */
205 at_measured(ldlm_lock_to_ns_at(lock), delay);
211 * Implementation of ->l_completion_ast() for a client, that doesn't wait
212 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
213 * other threads that cannot block for long.
215 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
219 if (flags == LDLM_FL_WAIT_NOREPROC) {
220 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
224 if (!(flags & LDLM_FL_BLOCKED_MASK)) {
225 wake_up(&lock->l_waitq);
226 RETURN(ldlm_completion_tail(lock, data));
229 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
231 ldlm_reprocess_all(lock->l_resource);
234 EXPORT_SYMBOL(ldlm_completion_ast_async);
237 * Generic LDLM "completion" AST. This is called in several cases:
239 * - when a reply to an ENQUEUE RPC is received from the server
240 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
241 * this point (determined by flags);
243 * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
246 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
249 * - to force all locks when resource is destroyed (cleanup_resource());
251 * - during lock conversion (not used currently).
253 * If lock is not granted in the first case, this function waits until second
254 * or penultimate cases happen in some other thread.
257 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
259 /* XXX ALLOCATE - 160 bytes */
260 struct lock_wait_data lwd;
261 struct obd_device *obd;
262 struct obd_import *imp = NULL;
263 struct l_wait_info lwi;
268 if (flags == LDLM_FL_WAIT_NOREPROC) {
269 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
273 if (!(flags & LDLM_FL_BLOCKED_MASK)) {
274 wake_up(&lock->l_waitq);
278 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
283 obd = class_exp2obd(lock->l_conn_export);
285 /* if this is a local lock, then there is no import */
287 imp = obd->u.cli.cl_import;
290 timeout = ldlm_cp_timeout(lock);
293 lock->l_last_activity = cfs_time_current_sec();
295 if (ldlm_is_no_timeout(lock)) {
296 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
297 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
299 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
300 ldlm_expired_completion_wait,
301 interrupted_completion_wait, &lwd);
305 spin_lock(&imp->imp_lock);
306 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
307 spin_unlock(&imp->imp_lock);
310 if (ns_is_client(ldlm_lock_to_ns(lock)) &&
311 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
312 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
313 ldlm_set_fail_loc(lock);
316 /* Go to sleep until the lock is granted or cancelled. */
317 rc = l_wait_event(lock->l_waitq,
318 is_granted_or_cancelled(lock), &lwi);
322 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
327 RETURN(ldlm_completion_tail(lock, data));
329 EXPORT_SYMBOL(ldlm_completion_ast);
332 * A helper to build a blocking AST function
334 * Perform a common operation for blocking ASTs:
335 * defferred lock cancellation.
337 * \param lock the lock blocking or canceling AST was called on
339 * \see mdt_blocking_ast
340 * \see ldlm_blocking_ast
342 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
347 ldlm_set_cbpending(lock);
348 do_ast = (!lock->l_readers && !lock->l_writers);
349 unlock_res_and_lock(lock);
352 struct lustre_handle lockh;
355 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
356 ldlm_lock2handle(lock, &lockh);
357 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
359 CERROR("ldlm_cli_cancel: %d\n", rc);
361 LDLM_DEBUG(lock, "Lock still has references, will be "
366 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
369 * Server blocking AST
371 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
374 * \param lock the lock which blocks a request or cancelling lock
377 * \param flag indicates whether this cancelling or blocking callback
379 * \see ldlm_blocking_ast_nocheck
381 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
382 void *data, int flag)
386 if (flag == LDLM_CB_CANCELING) {
387 /* Don't need to do anything here. */
391 lock_res_and_lock(lock);
392 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
393 * that ldlm_blocking_ast is called just before intent_policy method
394 * takes the lr_lock, then by the time we get the lock, we might not
395 * be the correct blocking function anymore. So check, and return
397 if (lock->l_blocking_ast != ldlm_blocking_ast) {
398 unlock_res_and_lock(lock);
401 RETURN(ldlm_blocking_ast_nocheck(lock));
403 EXPORT_SYMBOL(ldlm_blocking_ast);
406 * Implements ldlm_lock::l_glimpse_ast for extent locks acquired on the server.
408 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for that is
409 * rather subtle: with OST-side locking, it may so happen that _all_ extent
410 * locks are held by the OST. If client wants to obtain the current file size
411 * it calls ll_glimpse_size(), and (as all locks are held only on the server),
412 * this dummy glimpse callback fires and does nothing. The client still
413 * receives the correct file size due to the following fragment of code in
414 * ldlm_cb_interpret():
416 * if (rc == -ELDLM_NO_LOCK_DATA) {
417 * LDLM_DEBUG(lock, "lost race - client has a lock but no"
419 * ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
422 * That is, after the glimpse returns this error, ofd_lvbo_update() is called
423 * and returns the updated file attributes from the inode to the client.
425 * See also comment in ofd_intent_policy() on why servers must set a non-NULL
426 * l_glimpse_ast when grabbing DLM locks. Otherwise, the server will assume
427 * that the object is in the process of being destroyed.
429 * \param[in] lock DLM lock being glimpsed, unused
430 * \param[in] reqp pointer to ptlrpc_request, unused
432 * \retval -ELDLM_NO_LOCK_DATA to get attributes from disk object
434 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
436 return -ELDLM_NO_LOCK_DATA;
440 * Enqueue a local lock (typically on a server).
442 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
443 const struct ldlm_res_id *res_id,
444 enum ldlm_type type, union ldlm_policy_data *policy,
445 enum ldlm_mode mode, __u64 *flags,
446 ldlm_blocking_callback blocking,
447 ldlm_completion_callback completion,
448 ldlm_glimpse_callback glimpse,
449 void *data, __u32 lvb_len, enum lvb_type lvb_type,
450 const __u64 *client_cookie,
451 struct lustre_handle *lockh)
453 struct ldlm_lock *lock;
455 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
456 .lcs_blocking = blocking,
457 .lcs_glimpse = glimpse,
461 LASSERT(!(*flags & LDLM_FL_REPLAY));
462 if (unlikely(ns_is_client(ns))) {
463 CERROR("Trying to enqueue local lock in a shadow namespace\n");
467 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
470 GOTO(out_nolock, err = PTR_ERR(lock));
472 err = ldlm_lvbo_init(lock->l_resource);
474 LDLM_ERROR(lock, "delayed lvb init failed (rc %d)", err);
478 ldlm_lock2handle(lock, lockh);
480 /* NB: we don't have any lock now (lock_res_and_lock)
481 * because it's a new lock */
482 ldlm_lock_addref_internal_nolock(lock, mode);
483 ldlm_set_local(lock);
484 if (*flags & LDLM_FL_ATOMIC_CB)
485 ldlm_set_atomic_cb(lock);
488 lock->l_policy_data = *policy;
489 if (client_cookie != NULL)
490 lock->l_client_cookie = *client_cookie;
491 if (type == LDLM_EXTENT) {
492 /* extent lock without policy is a bug */
496 lock->l_req_extent = policy->l_extent;
499 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
500 if (unlikely(err != ELDLM_OK))
504 *policy = lock->l_policy_data;
506 if (lock->l_completion_ast)
507 lock->l_completion_ast(lock, *flags, NULL);
509 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
512 LDLM_LOCK_RELEASE(lock);
516 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
518 static void failed_lock_cleanup(struct ldlm_namespace *ns,
519 struct ldlm_lock *lock, int mode)
523 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
524 lock_res_and_lock(lock);
525 /* Check that lock is not granted or failed, we might race. */
526 if ((lock->l_req_mode != lock->l_granted_mode) &&
527 !ldlm_is_failed(lock)) {
528 /* Make sure that this lock will not be found by raced
529 * bl_ast and -EINVAL reply is sent to server anyways.
531 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
532 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
535 unlock_res_and_lock(lock);
539 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
540 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
542 LDLM_DEBUG(lock, "lock was granted or failed in race");
544 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
545 * from llite/file.c/ll_file_flock(). */
546 /* This code makes for the fact that we do not have blocking handler on
547 * a client for flock locks. As such this is the place where we must
548 * completely kill failed locks. (interrupted and those that
549 * were waiting to be granted when server evicted us. */
550 if (lock->l_resource->lr_type == LDLM_FLOCK) {
551 lock_res_and_lock(lock);
552 if (!ldlm_is_destroyed(lock)) {
553 ldlm_resource_unlink_lock(lock);
554 ldlm_lock_decref_internal_nolock(lock, mode);
555 ldlm_lock_destroy_nolock(lock);
557 unlock_res_and_lock(lock);
559 ldlm_lock_decref_internal(lock, mode);
564 * Finishing portion of client lock enqueue code.
566 * Called after receiving reply from server.
568 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
569 enum ldlm_type type, __u8 with_policy,
570 enum ldlm_mode mode, __u64 *flags, void *lvb,
571 __u32 lvb_len, const struct lustre_handle *lockh,
574 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
575 int is_replay = *flags & LDLM_FL_REPLAY;
576 struct ldlm_lock *lock;
577 struct ldlm_reply *reply;
578 int cleanup_phase = 1;
581 lock = ldlm_handle2lock(lockh);
582 /* ldlm_cli_enqueue is holding a reference on this lock. */
584 LASSERT(type == LDLM_FLOCK);
588 LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
589 "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
591 if (rc != ELDLM_OK) {
593 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
594 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
596 if (rc != ELDLM_LOCK_ABORTED)
600 /* Before we return, swab the reply */
601 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
603 GOTO(cleanup, rc = -EPROTO);
608 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
611 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
612 GOTO(cleanup, rc = size);
613 } else if (unlikely(size > lvb_len)) {
614 LDLM_ERROR(lock, "Replied LVB is larger than "
615 "expectation, expected = %d, replied = %d",
617 GOTO(cleanup, rc = -EINVAL);
622 if (rc == ELDLM_LOCK_ABORTED) {
623 if (lvb_len > 0 && lvb != NULL)
624 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
626 GOTO(cleanup, rc = rc ? : ELDLM_LOCK_ABORTED);
629 /* lock enqueued on the server */
632 lock_res_and_lock(lock);
633 /* Key change rehash lock in per-export hash with new key */
634 if (exp->exp_lock_hash) {
635 /* In the function below, .hs_keycmp resolves to
636 * ldlm_export_lock_keycmp() */
637 /* coverity[overrun-buffer-val] */
638 cfs_hash_rehash_key(exp->exp_lock_hash,
639 &lock->l_remote_handle,
643 lock->l_remote_handle = reply->lock_handle;
646 *flags = ldlm_flags_from_wire(reply->lock_flags);
647 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
648 LDLM_FL_INHERIT_MASK);
649 unlock_res_and_lock(lock);
651 CDEBUG(D_INFO, "local: %p, remote cookie: %#llx, flags: %#llx\n",
652 lock, reply->lock_handle.cookie, *flags);
654 /* If enqueue returned a blocked lock but the completion handler has
655 * already run, then it fixed up the resource and we don't need to do it
657 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
658 int newmode = reply->lock_desc.l_req_mode;
660 if (newmode && newmode != lock->l_req_mode) {
661 LDLM_DEBUG(lock, "server returned different mode %s",
662 ldlm_lockname[newmode]);
663 lock->l_req_mode = newmode;
666 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
667 &lock->l_resource->lr_name)) {
668 CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES
669 " instead of "DLDLMRES"\n",
670 PLDLMRES(&reply->lock_desc.l_resource),
671 PLDLMRES(lock->l_resource));
673 rc = ldlm_lock_change_resource(ns, lock,
674 &reply->lock_desc.l_resource.lr_name);
675 if (rc || lock->l_resource == NULL)
676 GOTO(cleanup, rc = -ENOMEM);
677 LDLM_DEBUG(lock, "client-side enqueue, new resource");
680 if (!(type == LDLM_IBITS &&
681 !(exp_connect_flags(exp) & OBD_CONNECT_IBITS)))
682 /* We assume lock type cannot change on server*/
683 ldlm_convert_policy_to_local(exp,
684 lock->l_resource->lr_type,
685 &reply->lock_desc.l_policy_data,
686 &lock->l_policy_data);
687 if (type != LDLM_PLAIN)
688 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
691 if ((*flags) & LDLM_FL_AST_SENT) {
692 lock_res_and_lock(lock);
693 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
694 unlock_res_and_lock(lock);
695 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
698 /* If the lock has already been granted by a completion AST, don't
699 * clobber the LVB with an older one. */
701 /* We must lock or a racing completion might update lvb without
702 * letting us know and we'll clobber the correct value.
703 * Cannot unlock after the check either, a that still leaves
704 * a tiny window for completion to get in */
705 lock_res_and_lock(lock);
706 if (lock->l_req_mode != lock->l_granted_mode)
707 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
708 lock->l_lvb_data, lvb_len);
709 unlock_res_and_lock(lock);
717 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
718 if (lock->l_completion_ast != NULL) {
719 int err = lock->l_completion_ast(lock, *flags, NULL);
727 if (lvb_len > 0 && lvb != NULL) {
728 /* Copy the LVB here, and not earlier, because the completion
729 * AST (if any) can override what we got in the reply */
730 memcpy(lvb, lock->l_lvb_data, lvb_len);
733 LDLM_DEBUG(lock, "client-side enqueue END");
736 if (cleanup_phase == 1 && rc)
737 failed_lock_cleanup(ns, lock, mode);
738 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
740 LDLM_LOCK_RELEASE(lock);
743 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
746 * Estimate number of lock handles that would fit into request of given
747 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
748 * a single page on the send/receive side. XXX: 512 should be changed to
749 * more adequate value.
751 static inline int ldlm_req_handles_avail(int req_size, int off)
755 avail = min_t(int, LDLM_MAXREQSIZE, PAGE_CACHE_SIZE - 512) - req_size;
756 if (likely(avail >= 0))
757 avail /= (int)sizeof(struct lustre_handle);
760 avail += LDLM_LOCKREQ_HANDLES - off;
765 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
766 enum req_location loc,
769 __u32 size = req_capsule_msg_size(pill, loc);
770 return ldlm_req_handles_avail(size, off);
773 static inline int ldlm_format_handles_avail(struct obd_import *imp,
774 const struct req_format *fmt,
775 enum req_location loc, int off)
777 __u32 size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
778 return ldlm_req_handles_avail(size, off);
782 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
783 * \a count locks in \a cancels.
785 * This is to be called by functions preparing their own requests that
786 * might contain lists of locks to cancel in addition to actual operation
787 * that needs to be performed.
789 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
790 int version, int opc, int canceloff,
791 struct list_head *cancels, int count)
793 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
794 struct req_capsule *pill = &req->rq_pill;
795 struct ldlm_request *dlm = NULL;
796 struct list_head head = LIST_HEAD_INIT(head);
797 enum ldlm_lru_flags lru_flags;
798 int avail, to_free, pack = 0;
804 if (ns_connect_cancelset(ns)) {
805 /* Estimate the amount of available space in the request. */
806 req_capsule_filled_sizes(pill, RCL_CLIENT);
807 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
809 lru_flags = ns_connect_lru_resize(ns) ?
810 LDLM_LRU_FLAG_LRUR_NO_WAIT : LDLM_LRU_FLAG_AGED;
811 to_free = !ns_connect_lru_resize(ns) &&
812 opc == LDLM_ENQUEUE ? 1 : 0;
814 /* Cancel LRU locks here _only_ if the server supports
815 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
816 * RPC, which will make us slower. */
818 count += ldlm_cancel_lru_local(ns, cancels, to_free,
825 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
826 ldlm_request_bufsize(pack, opc));
829 rc = ptlrpc_request_pack(req, version, opc);
831 ldlm_lock_list_put(cancels, l_bl_ast, count);
835 if (ns_connect_cancelset(ns)) {
837 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
839 /* Skip first lock handler in ldlm_request_pack(),
840 * this method will increment @lock_count according
841 * to the lock handle amount actually written to
843 dlm->lock_count = canceloff;
845 /* Pack into the request @pack lock handles. */
846 ldlm_cli_cancel_list(cancels, pack, req, 0);
847 /* Prepare and send separate cancel RPC for others. */
848 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
850 ldlm_lock_list_put(cancels, l_bl_ast, count);
854 EXPORT_SYMBOL(ldlm_prep_elc_req);
856 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
857 struct list_head *cancels, int count)
859 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
860 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
862 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
864 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
866 struct ptlrpc_request *req;
870 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
872 RETURN(ERR_PTR(-ENOMEM));
874 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
876 ptlrpc_request_free(req);
880 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
881 ptlrpc_request_set_replen(req);
884 EXPORT_SYMBOL(ldlm_enqueue_pack);
887 * Client-side lock enqueue.
889 * If a request has some specific initialisation it is passed in \a reqp,
890 * otherwise it is created in ldlm_cli_enqueue.
892 * Supports sync and async requests, pass \a async flag accordingly. If a
893 * request was created in ldlm_cli_enqueue and it is the async request,
894 * pass it to the caller in \a reqp.
896 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
897 struct ldlm_enqueue_info *einfo,
898 const struct ldlm_res_id *res_id,
899 union ldlm_policy_data const *policy, __u64 *flags,
900 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
901 struct lustre_handle *lockh, int async)
903 struct ldlm_namespace *ns;
904 struct ldlm_lock *lock;
905 struct ldlm_request *body;
906 int is_replay = *flags & LDLM_FL_REPLAY;
907 int req_passed_in = 1;
909 struct ptlrpc_request *req;
912 LASSERT(exp != NULL);
914 ns = exp->exp_obd->obd_namespace;
916 /* If we're replaying this lock, just check some invariants.
917 * If we're creating a new lock, get everything all setup nice. */
919 lock = ldlm_handle2lock_long(lockh, 0);
920 LASSERT(lock != NULL);
921 LDLM_DEBUG(lock, "client-side enqueue START");
922 LASSERT(exp == lock->l_conn_export);
924 const struct ldlm_callback_suite cbs = {
925 .lcs_completion = einfo->ei_cb_cp,
926 .lcs_blocking = einfo->ei_cb_bl,
927 .lcs_glimpse = einfo->ei_cb_gl
929 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
930 einfo->ei_mode, &cbs, einfo->ei_cbdata,
933 RETURN(PTR_ERR(lock));
934 /* for the local lock, add the reference */
935 ldlm_lock_addref_internal(lock, einfo->ei_mode);
936 ldlm_lock2handle(lock, lockh);
938 lock->l_policy_data = *policy;
940 if (einfo->ei_type == LDLM_EXTENT) {
941 /* extent lock without policy is a bug */
945 lock->l_req_extent = policy->l_extent;
947 LDLM_DEBUG(lock, "client-side enqueue START, flags %#llx",
951 lock->l_conn_export = exp;
952 lock->l_export = NULL;
953 lock->l_blocking_ast = einfo->ei_cb_bl;
954 lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
955 lock->l_last_activity = cfs_time_current_sec();
957 /* lock not sent to server yet */
959 if (reqp == NULL || *reqp == NULL) {
960 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
965 failed_lock_cleanup(ns, lock, einfo->ei_mode);
966 LDLM_LOCK_RELEASE(lock);
976 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
978 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
979 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
982 /* Dump lock data into the request buffer */
983 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
984 ldlm_lock2desc(lock, &body->lock_desc);
985 body->lock_flags = ldlm_flags_to_wire(*flags);
986 body->lock_handle[0] = *lockh;
988 /* Continue as normal. */
989 if (!req_passed_in) {
991 req_capsule_extend(&req->rq_pill,
992 &RQF_LDLM_ENQUEUE_LVB);
993 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
995 ptlrpc_request_set_replen(req);
999 LASSERT(reqp != NULL);
1003 LDLM_DEBUG(lock, "sending request");
1005 rc = ptlrpc_queue_wait(req);
1007 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
1008 einfo->ei_mode, flags, lvb, lvb_len,
1011 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
1012 * one reference that we took */
1014 LDLM_LOCK_RELEASE(lock);
1018 if (!req_passed_in && req != NULL) {
1019 ptlrpc_req_finished(req);
1026 EXPORT_SYMBOL(ldlm_cli_enqueue);
1028 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
1031 struct ldlm_resource *res;
1034 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1035 CERROR("Trying to cancel local lock\n");
1038 LDLM_DEBUG(lock, "client-side local convert");
1040 res = ldlm_lock_convert(lock, new_mode, flags);
1042 ldlm_reprocess_all(res);
1045 rc = LUSTRE_EDEADLK;
1047 LDLM_DEBUG(lock, "client-side local convert handler END");
1048 LDLM_LOCK_PUT(lock);
1052 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
1053 * conversion of locks which are on the waiting or converting queue */
1054 /* Caller of this code is supposed to take care of lock readers/writers
1056 int ldlm_cli_convert(const struct lustre_handle *lockh, int new_mode,
1059 struct ldlm_request *body;
1060 struct ldlm_reply *reply;
1061 struct ldlm_lock *lock;
1062 struct ldlm_resource *res;
1063 struct ptlrpc_request *req;
1067 lock = ldlm_handle2lock(lockh);
1074 if (lock->l_conn_export == NULL)
1075 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
1077 LDLM_DEBUG(lock, "client-side convert");
1079 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
1080 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1083 LDLM_LOCK_PUT(lock);
1087 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1088 body->lock_handle[0] = lock->l_remote_handle;
1090 body->lock_desc.l_req_mode = new_mode;
1091 body->lock_flags = ldlm_flags_to_wire(*flags);
1094 ptlrpc_request_set_replen(req);
1095 rc = ptlrpc_queue_wait(req);
1099 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1101 GOTO(out, rc = -EPROTO);
1104 GOTO(out, rc = req->rq_status);
1106 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
1108 ldlm_reprocess_all(res);
1109 /* Go to sleep until the lock is granted. */
1110 /* FIXME: or cancelled. */
1111 if (lock->l_completion_ast) {
1112 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1118 rc = LUSTRE_EDEADLK;
1122 LDLM_LOCK_PUT(lock);
1123 ptlrpc_req_finished(req);
1128 * Cancel locks locally.
1130 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1131 * \retval LDLM_FL_CANCELING otherwise;
1132 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1134 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1136 __u64 rc = LDLM_FL_LOCAL_ONLY;
1139 if (lock->l_conn_export) {
1142 LDLM_DEBUG(lock, "client-side cancel");
1143 /* Set this flag to prevent others from getting new references*/
1144 lock_res_and_lock(lock);
1145 ldlm_set_cbpending(lock);
1146 local_only = !!(lock->l_flags &
1147 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1148 ldlm_cancel_callback(lock);
1149 rc = (ldlm_is_bl_ast(lock)) ?
1150 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1151 unlock_res_and_lock(lock);
1154 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1156 rc = LDLM_FL_LOCAL_ONLY;
1158 ldlm_lock_cancel(lock);
1160 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1161 LDLM_ERROR(lock, "Trying to cancel local lock");
1164 LDLM_DEBUG(lock, "server-side local cancel");
1165 ldlm_lock_cancel(lock);
1166 ldlm_reprocess_all(lock->l_resource);
1173 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1175 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1176 struct list_head *head, int count)
1178 struct ldlm_request *dlm;
1179 struct ldlm_lock *lock;
1180 int max, packed = 0;
1183 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1184 LASSERT(dlm != NULL);
1186 /* Check the room in the request buffer. */
1187 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1188 sizeof(struct ldlm_request);
1189 max /= sizeof(struct lustre_handle);
1190 max += LDLM_LOCKREQ_HANDLES;
1191 LASSERT(max >= dlm->lock_count + count);
1193 /* XXX: it would be better to pack lock handles grouped by resource.
1194 * so that the server cancel would call filter_lvbo_update() less
1196 list_for_each_entry(lock, head, l_bl_ast) {
1199 LASSERT(lock->l_conn_export);
1200 /* Pack the lock handle to the given request buffer. */
1201 LDLM_DEBUG(lock, "packing");
1202 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1205 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1210 * Prepare and send a batched cancel RPC. It will include \a count lock
1211 * handles of locks given in \a cancels list. */
1212 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1213 int count, enum ldlm_cancel_flags flags)
1215 struct ptlrpc_request *req = NULL;
1216 struct obd_import *imp;
1221 LASSERT(exp != NULL);
1224 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1226 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1229 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1230 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1235 imp = class_exp2cliimp(exp);
1236 if (imp == NULL || imp->imp_invalid) {
1238 "skipping cancel on invalid import %p\n", imp);
1242 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1244 GOTO(out, rc = -ENOMEM);
1246 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1247 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1248 ldlm_request_bufsize(count, LDLM_CANCEL));
1250 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1252 ptlrpc_request_free(req);
1256 /* If OSP want cancel cross-MDT lock, let's not block it in
1257 * in recovery, otherwise the lock will not released, if
1258 * the remote target is also in recovery, and it also need
1259 * this lock, it might cause deadlock. */
1260 if (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS &&
1261 exp->exp_obd->obd_lu_dev != NULL &&
1262 exp->exp_obd->obd_lu_dev->ld_site != NULL) {
1263 struct lu_device *top_dev;
1265 top_dev = exp->exp_obd->obd_lu_dev->ld_site->ls_top_dev;
1266 if (top_dev != NULL &&
1267 top_dev->ld_obd->obd_recovering)
1268 req->rq_allow_replay = 1;
1271 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1272 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1273 ptlrpc_at_set_req_timeout(req);
1275 ldlm_cancel_pack(req, cancels, count);
1277 ptlrpc_request_set_replen(req);
1278 if (flags & LCF_ASYNC) {
1279 ptlrpcd_add_req(req);
1284 rc = ptlrpc_queue_wait(req);
1285 if (rc == LUSTRE_ESTALE) {
1286 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1287 "out of sync -- not fatal\n",
1288 libcfs_nid2str(req->rq_import->
1289 imp_connection->c_peer.nid));
1291 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1292 req->rq_import_generation == imp->imp_generation) {
1293 ptlrpc_req_finished(req);
1295 } else if (rc != ELDLM_OK) {
1296 /* -ESHUTDOWN is common on umount */
1297 CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1298 "Got rc %d from cancel RPC: "
1299 "canceling anyway\n", rc);
1306 ptlrpc_req_finished(req);
1309 return sent ? sent : rc;
1312 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1314 LASSERT(imp != NULL);
1315 return &imp->imp_obd->obd_namespace->ns_pool;
1319 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1321 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1323 struct obd_device *obd;
1327 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1328 !imp_connect_lru_resize(req->rq_import)))
1331 * Do nothing for corner cases.
1336 /* In some cases RPC may contain SLV and limit zeroed out. This
1337 * is the case when server does not support LRU resize feature.
1338 * This is also possible in some recovery cases when server-side
1339 * reqs have no reference to the OBD export and thus access to
1340 * server-side namespace is not possible. */
1341 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1342 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1343 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1344 "(SLV: %llu, Limit: %u)",
1345 lustre_msg_get_slv(req->rq_repmsg),
1346 lustre_msg_get_limit(req->rq_repmsg));
1350 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1351 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1352 obd = req->rq_import->imp_obd;
1354 /* Set new SLV and limit in OBD fields to make them accessible
1355 * to the pool thread. We do not access obd_namespace and pool
1356 * directly here as there is no reliable way to make sure that
1357 * they are still alive at cleanup time. Evil races are possible
1358 * which may cause Oops at that time. */
1359 write_lock(&obd->obd_pool_lock);
1360 obd->obd_pool_slv = new_slv;
1361 obd->obd_pool_limit = new_limit;
1362 write_unlock(&obd->obd_pool_lock);
1368 * Client side lock cancel.
1370 * Lock must not have any readers or writers by this time.
1372 int ldlm_cli_cancel(const struct lustre_handle *lockh,
1373 enum ldlm_cancel_flags cancel_flags)
1375 struct obd_export *exp;
1376 enum ldlm_lru_flags lru_flags;
1377 int avail, count = 1;
1379 struct ldlm_namespace *ns;
1380 struct ldlm_lock *lock;
1381 struct list_head cancels = LIST_HEAD_INIT(cancels);
1384 lock = ldlm_handle2lock_long(lockh, 0);
1386 LDLM_DEBUG_NOLOCK("lock is already being destroyed");
1390 lock_res_and_lock(lock);
1391 /* Lock is being canceled and the caller doesn't want to wait */
1392 if (ldlm_is_canceling(lock) && (cancel_flags & LCF_ASYNC)) {
1393 unlock_res_and_lock(lock);
1394 LDLM_LOCK_RELEASE(lock);
1398 ldlm_set_canceling(lock);
1399 unlock_res_and_lock(lock);
1401 rc = ldlm_cli_cancel_local(lock);
1402 if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
1403 LDLM_LOCK_RELEASE(lock);
1406 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1407 * RPC which goes to canceld portal, so we can cancel other LRU locks
1408 * here and send them all as one LDLM_CANCEL RPC. */
1409 LASSERT(list_empty(&lock->l_bl_ast));
1410 list_add(&lock->l_bl_ast, &cancels);
1412 exp = lock->l_conn_export;
1413 if (exp_connect_cancelset(exp)) {
1414 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1419 ns = ldlm_lock_to_ns(lock);
1420 lru_flags = ns_connect_lru_resize(ns) ?
1421 LDLM_LRU_FLAG_LRUR : LDLM_LRU_FLAG_AGED;
1422 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1423 LCF_BL_AST, lru_flags);
1425 ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1428 EXPORT_SYMBOL(ldlm_cli_cancel);
1431 * Locally cancel up to \a count locks in list \a cancels.
1432 * Return the number of cancelled locks.
1434 int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
1435 enum ldlm_cancel_flags cancel_flags)
1437 struct list_head head = LIST_HEAD_INIT(head);
1438 struct ldlm_lock *lock, *next;
1439 int left = 0, bl_ast = 0;
1443 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1447 if (cancel_flags & LCF_LOCAL) {
1448 rc = LDLM_FL_LOCAL_ONLY;
1449 ldlm_lock_cancel(lock);
1451 rc = ldlm_cli_cancel_local(lock);
1453 /* Until we have compound requests and can send LDLM_CANCEL
1454 * requests batched with generic RPCs, we need to send cancels
1455 * with the LDLM_FL_BL_AST flag in a separate RPC from
1456 * the one being generated now. */
1457 if (!(cancel_flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1458 LDLM_DEBUG(lock, "Cancel lock separately");
1459 list_del_init(&lock->l_bl_ast);
1460 list_add(&lock->l_bl_ast, &head);
1464 if (rc == LDLM_FL_LOCAL_ONLY) {
1465 /* CANCEL RPC should not be sent to server. */
1466 list_del_init(&lock->l_bl_ast);
1467 LDLM_LOCK_RELEASE(lock);
1473 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1480 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1481 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1482 * readahead requests, ...)
1484 static enum ldlm_policy_res
1485 ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1486 int unused, int added, int count)
1488 enum ldlm_policy_res result = LDLM_POLICY_CANCEL_LOCK;
1490 /* don't check added & count since we want to process all locks
1492 * It's fine to not take lock to access lock->l_resource since
1493 * the lock has already been granted so it won't change. */
1494 switch (lock->l_resource->lr_type) {
1497 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) != 0)
1500 result = LDLM_POLICY_SKIP_LOCK;
1501 lock_res_and_lock(lock);
1502 ldlm_set_skipped(lock);
1503 unlock_res_and_lock(lock);
1511 * Callback function for LRU-resize policy. Decides whether to keep
1512 * \a lock in LRU for current \a LRU size \a unused, added in current
1513 * scan \a added and number of locks to be preferably canceled \a count.
1515 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1517 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1519 static enum ldlm_policy_res ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1520 struct ldlm_lock *lock,
1521 int unused, int added,
1524 cfs_time_t cur = cfs_time_current();
1525 struct ldlm_pool *pl = &ns->ns_pool;
1529 /* Stop LRU processing when we reach past @count or have checked all
1531 if (count && added >= count)
1532 return LDLM_POLICY_KEEP_LOCK;
1534 /* Despite of the LV, It doesn't make sense to keep the lock which
1535 * is unused for ns_max_age time. */
1536 if (cfs_time_after(cfs_time_current(),
1537 cfs_time_add(lock->l_last_used, ns->ns_max_age)))
1538 return LDLM_POLICY_CANCEL_LOCK;
1540 slv = ldlm_pool_get_slv(pl);
1541 lvf = ldlm_pool_get_lvf(pl);
1542 la = cfs_duration_sec(cfs_time_sub(cur,
1543 lock->l_last_used));
1544 lv = lvf * la * unused;
1546 /* Inform pool about current CLV to see it via proc. */
1547 ldlm_pool_set_clv(pl, lv);
1549 /* Stop when SLV is not yet come from server or lv is smaller than
1551 if (slv == 0 || lv < slv)
1552 return LDLM_POLICY_KEEP_LOCK;
1554 return LDLM_POLICY_CANCEL_LOCK;
1557 static enum ldlm_policy_res
1558 ldlm_cancel_lrur_no_wait_policy(struct ldlm_namespace *ns,
1559 struct ldlm_lock *lock,
1560 int unused, int added,
1563 enum ldlm_policy_res result;
1565 result = ldlm_cancel_lrur_policy(ns, lock, unused, added, count);
1566 if (result == LDLM_POLICY_KEEP_LOCK)
1569 return ldlm_cancel_no_wait_policy(ns, lock, unused, added, count);
1573 * Callback function for proc used policy. Makes decision whether to keep
1574 * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1575 * added and number of locks to be preferably canceled \a count.
1577 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1579 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1581 static enum ldlm_policy_res ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1582 struct ldlm_lock *lock,
1583 int unused, int added,
1586 /* Stop LRU processing when we reach past @count or have checked all
1588 return (added >= count) ?
1589 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1593 * Callback function for aged policy. Makes decision whether to keep \a lock in
1594 * LRU for current LRU size \a unused, added in current scan \a added and
1595 * number of locks to be preferably canceled \a count.
1597 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1599 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1601 static enum ldlm_policy_res ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1602 struct ldlm_lock *lock,
1603 int unused, int added,
1606 if ((added >= count) &&
1607 cfs_time_before(cfs_time_current(),
1608 cfs_time_add(lock->l_last_used, ns->ns_max_age)))
1609 return LDLM_POLICY_KEEP_LOCK;
1611 return LDLM_POLICY_CANCEL_LOCK;
1615 * Callback function for default policy. Makes decision whether to keep \a lock
1616 * in LRU for current LRU size \a unused, added in current scan \a added and
1617 * number of locks to be preferably canceled \a count.
1619 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1621 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1624 enum ldlm_policy_res ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1625 struct ldlm_lock *lock,
1626 int unused, int added,
1629 /* Stop LRU processing when we reach past count or have checked all
1631 return (added >= count) ?
1632 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1635 typedef enum ldlm_policy_res
1636 (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1637 int unused, int added, int count);
1639 static ldlm_cancel_lru_policy_t
1640 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, enum ldlm_lru_flags lru_flags)
1642 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1643 return ldlm_cancel_no_wait_policy;
1645 if (ns_connect_lru_resize(ns)) {
1646 if (lru_flags & LDLM_LRU_FLAG_SHRINK)
1647 /* We kill passed number of old locks. */
1648 return ldlm_cancel_passed_policy;
1649 if (lru_flags & LDLM_LRU_FLAG_LRUR)
1650 return ldlm_cancel_lrur_policy;
1651 if (lru_flags & LDLM_LRU_FLAG_PASSED)
1652 return ldlm_cancel_passed_policy;
1653 else if (lru_flags & LDLM_LRU_FLAG_LRUR_NO_WAIT)
1654 return ldlm_cancel_lrur_no_wait_policy;
1656 if (lru_flags & LDLM_LRU_FLAG_AGED)
1657 return ldlm_cancel_aged_policy;
1660 return ldlm_cancel_default_policy;
1664 * - Free space in LRU for \a count new locks,
1665 * redundant unused locks are canceled locally;
1666 * - also cancel locally unused aged locks;
1667 * - do not cancel more than \a max locks;
1668 * - GET the found locks and add them into the \a cancels list.
1670 * A client lock can be added to the l_bl_ast list only when it is
1671 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1672 * CANCEL. There are the following use cases:
1673 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1674 * ldlm_cli_cancel(), which check and set this flag properly. As any
1675 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1676 * later without any special locking.
1678 * Calling policies for enabled LRU resize:
1679 * ----------------------------------------
1680 * flags & LDLM_LRU_FLAG_LRUR - use LRU resize policy (SLV from server) to
1681 * cancel not more than \a count locks;
1683 * flags & LDLM_LRU_FLAG_PASSED - cancel \a count number of old locks (located
1684 * at the beginning of LRU list);
1686 * flags & LDLM_LRU_FLAG_SHRINK - cancel not more than \a count locks according
1687 * to memory pressre policy function;
1689 * flags & LDLM_LRU_FLAG_AGED - cancel \a count locks according to "aged policy"
1691 * flags & LDLM_LRU_FLAG_NO_WAIT - cancel as many unused locks as possible
1692 * (typically before replaying locks) w/o
1693 * sending any RPCs or waiting for any
1694 * outstanding RPC to complete.
1696 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns,
1697 struct list_head *cancels, int count, int max,
1698 enum ldlm_lru_flags lru_flags)
1700 ldlm_cancel_lru_policy_t pf;
1701 struct ldlm_lock *lock, *next;
1702 int added = 0, unused, remained;
1703 int no_wait = lru_flags & (LDLM_LRU_FLAG_NO_WAIT |
1704 LDLM_LRU_FLAG_LRUR_NO_WAIT);
1707 spin_lock(&ns->ns_lock);
1708 unused = ns->ns_nr_unused;
1711 if (!ns_connect_lru_resize(ns))
1712 count += unused - ns->ns_max_unused;
1714 pf = ldlm_cancel_lru_policy(ns, lru_flags);
1715 LASSERT(pf != NULL);
1717 while (!list_empty(&ns->ns_unused_list)) {
1718 enum ldlm_policy_res result;
1719 cfs_time_t last_use = 0;
1721 /* all unused locks */
1722 if (remained-- <= 0)
1725 /* For any flags, stop scanning if @max is reached. */
1726 if (max && added >= max)
1729 list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1731 /* No locks which got blocking requests. */
1732 LASSERT(!ldlm_is_bl_ast(lock));
1734 if (no_wait && ldlm_is_skipped(lock))
1735 /* already processed */
1738 last_use = lock->l_last_used;
1739 if (last_use == cfs_time_current())
1742 /* Somebody is already doing CANCEL. No need for this
1743 * lock in LRU, do not traverse it again. */
1744 if (!ldlm_is_canceling(lock))
1747 ldlm_lock_remove_from_lru_nolock(lock);
1749 if (&lock->l_lru == &ns->ns_unused_list)
1752 LDLM_LOCK_GET(lock);
1753 spin_unlock(&ns->ns_lock);
1754 lu_ref_add(&lock->l_reference, __FUNCTION__, current);
1756 /* Pass the lock through the policy filter and see if it
1757 * should stay in LRU.
1759 * Even for shrinker policy we stop scanning if
1760 * we find a lock that should stay in the cache.
1761 * We should take into account lock age anyway
1762 * as a new lock is a valuable resource even if
1763 * it has a low weight.
1765 * That is, for shrinker policy we drop only
1766 * old locks, but additionally choose them by
1767 * their weight. Big extent locks will stay in
1769 result = pf(ns, lock, unused, added, count);
1770 if (result == LDLM_POLICY_KEEP_LOCK) {
1771 lu_ref_del(&lock->l_reference,
1772 __FUNCTION__, current);
1773 LDLM_LOCK_RELEASE(lock);
1774 spin_lock(&ns->ns_lock);
1777 if (result == LDLM_POLICY_SKIP_LOCK) {
1778 lu_ref_del(&lock->l_reference,
1780 LDLM_LOCK_RELEASE(lock);
1781 spin_lock(&ns->ns_lock);
1785 lock_res_and_lock(lock);
1786 /* Check flags again under the lock. */
1787 if (ldlm_is_canceling(lock) ||
1788 ldlm_lock_remove_from_lru_check(lock, last_use) == 0) {
1789 /* Another thread is removing lock from LRU, or
1790 * somebody is already doing CANCEL, or there
1791 * is a blocking request which will send cancel
1792 * by itself, or the lock is no longer unused or
1793 * the lock has been used since the pf() call and
1794 * pages could be put under it. */
1795 unlock_res_and_lock(lock);
1796 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1797 LDLM_LOCK_RELEASE(lock);
1798 spin_lock(&ns->ns_lock);
1801 LASSERT(!lock->l_readers && !lock->l_writers);
1803 /* If we have chosen to cancel this lock voluntarily, we
1804 * better send cancel notification to server, so that it
1805 * frees appropriate state. This might lead to a race
1806 * where while we are doing cancel here, server is also
1807 * silently cancelling this lock. */
1808 ldlm_clear_cancel_on_block(lock);
1810 /* Setting the CBPENDING flag is a little misleading,
1811 * but prevents an important race; namely, once
1812 * CBPENDING is set, the lock can accumulate no more
1813 * readers/writers. Since readers and writers are
1814 * already zero here, ldlm_lock_decref() won't see
1815 * this flag and call l_blocking_ast */
1816 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1818 /* We can't re-add to l_lru as it confuses the
1819 * refcounting in ldlm_lock_remove_from_lru() if an AST
1820 * arrives after we drop lr_lock below. We use l_bl_ast
1821 * and can't use l_pending_chain as it is used both on
1822 * server and client nevertheless bug 5666 says it is
1823 * used only on server */
1824 LASSERT(list_empty(&lock->l_bl_ast));
1825 list_add(&lock->l_bl_ast, cancels);
1826 unlock_res_and_lock(lock);
1827 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1828 spin_lock(&ns->ns_lock);
1832 spin_unlock(&ns->ns_lock);
1836 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1838 enum ldlm_cancel_flags cancel_flags,
1839 enum ldlm_lru_flags lru_flags)
1843 added = ldlm_prepare_lru_list(ns, cancels, count, max, lru_flags);
1847 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1851 * Cancel at least \a nr locks from given namespace LRU.
1853 * When called with LCF_ASYNC the blocking callback will be handled
1854 * in a thread and this function will return after the thread has been
1855 * asked to call the callback. When called with LCF_ASYNC the blocking
1856 * callback will be performed in this function.
1858 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
1859 enum ldlm_cancel_flags cancel_flags,
1860 enum ldlm_lru_flags lru_flags)
1862 struct list_head cancels = LIST_HEAD_INIT(cancels);
1866 /* Just prepare the list of locks, do not actually cancel them yet.
1867 * Locks are cancelled later in a separate thread. */
1868 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, lru_flags);
1869 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
1877 * Find and cancel locally unused locks found on resource, matched to the
1878 * given policy, mode. GET the found locks and add them into the \a cancels
1881 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1882 struct list_head *cancels,
1883 union ldlm_policy_data *policy,
1884 enum ldlm_mode mode, __u64 lock_flags,
1885 enum ldlm_cancel_flags cancel_flags,
1888 struct ldlm_lock *lock;
1893 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1894 if (opaque != NULL && lock->l_ast_data != opaque) {
1895 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1896 lock->l_ast_data, opaque);
1901 if (lock->l_readers || lock->l_writers)
1904 /* If somebody is already doing CANCEL, or blocking AST came,
1905 * skip this lock. */
1906 if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock))
1909 if (lockmode_compat(lock->l_granted_mode, mode))
1912 /* If policy is given and this is IBITS lock, add to list only
1913 * those locks that match by policy. */
1914 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1915 !(lock->l_policy_data.l_inodebits.bits &
1916 policy->l_inodebits.bits))
1919 /* See CBPENDING comment in ldlm_cancel_lru */
1920 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1923 LASSERT(list_empty(&lock->l_bl_ast));
1924 list_add(&lock->l_bl_ast, cancels);
1925 LDLM_LOCK_GET(lock);
1930 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
1932 EXPORT_SYMBOL(ldlm_cancel_resource_local);
1935 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1937 * If \a req is NULL, send CANCEL request to server with handles of locks
1938 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1939 * separately per lock.
1940 * If \a req is not NULL, put handles of locks in \a cancels into the request
1941 * buffer at the offset \a off.
1942 * Destroy \a cancels at the end.
1944 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1945 struct ptlrpc_request *req,
1946 enum ldlm_cancel_flags flags)
1948 struct ldlm_lock *lock;
1952 if (list_empty(cancels) || count == 0)
1955 /* XXX: requests (both batched and not) could be sent in parallel.
1956 * Usually it is enough to have just 1 RPC, but it is possible that
1957 * there are too many locks to be cancelled in LRU or on a resource.
1958 * It would also speed up the case when the server does not support
1961 LASSERT(!list_empty(cancels));
1962 lock = list_entry(cancels->next, struct ldlm_lock,
1964 LASSERT(lock->l_conn_export);
1966 if (exp_connect_cancelset(lock->l_conn_export)) {
1969 ldlm_cancel_pack(req, cancels, count);
1971 res = ldlm_cli_cancel_req(lock->l_conn_export,
1975 res = ldlm_cli_cancel_req(lock->l_conn_export,
1980 CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1981 "ldlm_cli_cancel_list: %d\n", res);
1986 ldlm_lock_list_put(cancels, l_bl_ast, res);
1988 LASSERT(count == 0);
1991 EXPORT_SYMBOL(ldlm_cli_cancel_list);
1994 * Cancel all locks on a resource that have 0 readers/writers.
1996 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1997 * to notify the server. */
1998 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1999 const struct ldlm_res_id *res_id,
2000 union ldlm_policy_data *policy,
2001 enum ldlm_mode mode,
2002 enum ldlm_cancel_flags flags, void *opaque)
2004 struct ldlm_resource *res;
2005 struct list_head cancels = LIST_HEAD_INIT(cancels);
2010 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2012 /* This is not a problem. */
2013 CDEBUG(D_INFO, "No resource %llu\n", res_id->name[0]);
2017 LDLM_RESOURCE_ADDREF(res);
2018 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
2019 0, flags | LCF_BL_AST, opaque);
2020 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
2022 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
2025 LDLM_RESOURCE_DELREF(res);
2026 ldlm_resource_putref(res);
2029 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2031 struct ldlm_cli_cancel_arg {
2037 ldlm_cli_hash_cancel_unused(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2038 struct hlist_node *hnode, void *arg)
2040 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2041 struct ldlm_cli_cancel_arg *lc = arg;
2043 ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
2044 NULL, LCK_MINMODE, lc->lc_flags,
2046 /* must return 0 for hash iteration */
2051 * Cancel all locks on a namespace (or a specific resource, if given)
2052 * that have 0 readers/writers.
2054 * If flags & LCF_LOCAL, throw the locks away without trying
2055 * to notify the server. */
2056 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
2057 const struct ldlm_res_id *res_id,
2058 enum ldlm_cancel_flags flags, void *opaque)
2060 struct ldlm_cli_cancel_arg arg = {
2062 .lc_opaque = opaque,
2070 if (res_id != NULL) {
2071 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
2075 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2076 ldlm_cli_hash_cancel_unused, &arg, 0);
2081 /* Lock iterators. */
2083 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
2086 struct list_head *tmp, *next;
2087 struct ldlm_lock *lock;
2088 int rc = LDLM_ITER_CONTINUE;
2093 RETURN(LDLM_ITER_CONTINUE);
2096 list_for_each_safe(tmp, next, &res->lr_granted) {
2097 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2099 if (iter(lock, closure) == LDLM_ITER_STOP)
2100 GOTO(out, rc = LDLM_ITER_STOP);
2103 list_for_each_safe(tmp, next, &res->lr_converting) {
2104 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2106 if (iter(lock, closure) == LDLM_ITER_STOP)
2107 GOTO(out, rc = LDLM_ITER_STOP);
2110 list_for_each_safe(tmp, next, &res->lr_waiting) {
2111 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2113 if (iter(lock, closure) == LDLM_ITER_STOP)
2114 GOTO(out, rc = LDLM_ITER_STOP);
2121 struct iter_helper_data {
2122 ldlm_iterator_t iter;
2126 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2128 struct iter_helper_data *helper = closure;
2129 return helper->iter(lock, helper->closure);
2132 static int ldlm_res_iter_helper(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2133 struct hlist_node *hnode, void *arg)
2136 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2138 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2142 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2143 ldlm_iterator_t iter, void *closure)
2146 struct iter_helper_data helper = { .iter = iter, .closure = closure };
2148 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2149 ldlm_res_iter_helper, &helper, 0);
2153 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2154 * return 0: find no resource
2155 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2158 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2159 const struct ldlm_res_id *res_id,
2160 ldlm_iterator_t iter, void *data)
2162 struct ldlm_resource *res;
2166 LASSERTF(ns != NULL, "must pass in namespace\n");
2168 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2172 LDLM_RESOURCE_ADDREF(res);
2173 rc = ldlm_resource_foreach(res, iter, data);
2174 LDLM_RESOURCE_DELREF(res);
2175 ldlm_resource_putref(res);
2178 EXPORT_SYMBOL(ldlm_resource_iterate);
2182 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2184 struct list_head *list = closure;
2186 /* we use l_pending_chain here, because it's unused on clients. */
2187 LASSERTF(list_empty(&lock->l_pending_chain),
2188 "lock %p next %p prev %p\n",
2189 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2190 /* bug 9573: don't replay locks left after eviction, or
2191 * bug 17614: locks being actively cancelled. Get a reference
2192 * on a lock so that it does not disapear under us (e.g. due to cancel)
2194 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_BL_DONE))) {
2195 list_add(&lock->l_pending_chain, list);
2196 LDLM_LOCK_GET(lock);
2199 return LDLM_ITER_CONTINUE;
2202 static int replay_lock_interpret(const struct lu_env *env,
2203 struct ptlrpc_request *req,
2204 struct ldlm_async_args *aa, int rc)
2206 struct ldlm_lock *lock;
2207 struct ldlm_reply *reply;
2208 struct obd_export *exp;
2211 atomic_dec(&req->rq_import->imp_replay_inflight);
2215 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2217 GOTO(out, rc = -EPROTO);
2219 lock = ldlm_handle2lock(&aa->lock_handle);
2221 CERROR("received replay ack for unknown local cookie %#llx"
2222 " remote cookie %#llx from server %s id %s\n",
2223 aa->lock_handle.cookie, reply->lock_handle.cookie,
2224 req->rq_export->exp_client_uuid.uuid,
2225 libcfs_id2str(req->rq_peer));
2226 GOTO(out, rc = -ESTALE);
2229 /* Key change rehash lock in per-export hash with new key */
2230 exp = req->rq_export;
2231 if (exp && exp->exp_lock_hash) {
2232 /* In the function below, .hs_keycmp resolves to
2233 * ldlm_export_lock_keycmp() */
2234 /* coverity[overrun-buffer-val] */
2235 cfs_hash_rehash_key(exp->exp_lock_hash,
2236 &lock->l_remote_handle,
2237 &reply->lock_handle,
2240 lock->l_remote_handle = reply->lock_handle;
2243 LDLM_DEBUG(lock, "replayed lock:");
2244 ptlrpc_import_recovery_state_machine(req->rq_import);
2245 LDLM_LOCK_PUT(lock);
2248 ptlrpc_connect_import(req->rq_import);
2253 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2255 struct ptlrpc_request *req;
2256 struct ldlm_async_args *aa;
2257 struct ldlm_request *body;
2262 /* Bug 11974: Do not replay a lock which is actively being canceled */
2263 if (ldlm_is_bl_done(lock)) {
2264 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2268 /* If this is reply-less callback lock, we cannot replay it, since
2269 * server might have long dropped it, but notification of that event was
2270 * lost by network. (and server granted conflicting lock already) */
2271 if (ldlm_is_cancel_on_block(lock)) {
2272 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2273 ldlm_lock_cancel(lock);
2278 * If granted mode matches the requested mode, this lock is granted.
2280 * If they differ, but we have a granted mode, then we were granted
2281 * one mode and now want another: ergo, converting.
2283 * If we haven't been granted anything and are on a resource list,
2284 * then we're blocked/waiting.
2286 * If we haven't been granted anything and we're NOT on a resource list,
2287 * then we haven't got a reply yet and don't have a known disposition.
2288 * This happens whenever a lock enqueue is the request that triggers
2291 if (lock->l_granted_mode == lock->l_req_mode)
2292 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2293 else if (lock->l_granted_mode)
2294 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2295 else if (!list_empty(&lock->l_res_link))
2296 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2298 flags = LDLM_FL_REPLAY;
2300 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2301 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2305 /* We're part of recovery, so don't wait for it. */
2306 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2308 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2309 ldlm_lock2desc(lock, &body->lock_desc);
2310 body->lock_flags = ldlm_flags_to_wire(flags);
2312 ldlm_lock2handle(lock, &body->lock_handle[0]);
2313 if (lock->l_lvb_len > 0)
2314 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2315 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2317 ptlrpc_request_set_replen(req);
2318 /* notify the server we've replayed all requests.
2319 * also, we mark the request to be put on a dedicated
2320 * queue to be processed after all request replayes.
2322 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2324 LDLM_DEBUG(lock, "replaying lock:");
2326 atomic_inc(&req->rq_import->imp_replay_inflight);
2327 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2328 aa = ptlrpc_req_async_args(req);
2329 aa->lock_handle = body->lock_handle[0];
2330 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2331 ptlrpcd_add_req(req);
2337 * Cancel as many unused locks as possible before replay. since we are
2338 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2341 * Called only in recovery before replaying locks. there is no need to
2342 * replay locks that are unused. since the clients may hold thousands of
2343 * cached unused locks, dropping the unused locks can greatly reduce the
2344 * load on the servers at recovery time.
2346 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2349 struct list_head cancels = LIST_HEAD_INIT(cancels);
2351 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2352 "replay for namespace %s (%d)\n",
2353 ldlm_ns_name(ns), ns->ns_nr_unused);
2355 /* We don't need to care whether or not LRU resize is enabled
2356 * because the LDLM_LRU_FLAG_NO_WAIT policy doesn't use the
2357 * count parameter */
2358 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2359 LCF_LOCAL, LDLM_LRU_FLAG_NO_WAIT);
2361 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2362 canceled, ldlm_ns_name(ns));
2365 int ldlm_replay_locks(struct obd_import *imp)
2367 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2368 struct list_head list = LIST_HEAD_INIT(list);
2369 struct ldlm_lock *lock, *next;
2374 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2376 /* don't replay locks if import failed recovery */
2377 if (imp->imp_vbr_failed)
2380 /* ensure this doesn't fall to 0 before all have been queued */
2381 atomic_inc(&imp->imp_replay_inflight);
2383 if (ldlm_cancel_unused_locks_before_replay)
2384 ldlm_cancel_unused_locks_for_replay(ns);
2386 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2388 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2389 list_del_init(&lock->l_pending_chain);
2391 LDLM_LOCK_RELEASE(lock);
2392 continue; /* or try to do the rest? */
2394 rc = replay_one_lock(imp, lock);
2395 LDLM_LOCK_RELEASE(lock);
2398 atomic_dec(&imp->imp_replay_inflight);