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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 * This file contains Asynchronous System Trap (AST) handlers and related
34 * LDLM request-processing routines.
36 * An AST is a callback issued on a lock when its state is changed. There are
37 * several different types of ASTs (callbacks) registered for each lock:
39 * - completion AST: when a lock is enqueued by some process, but cannot be
40 * granted immediately due to other conflicting locks on the same resource,
41 * the completion AST is sent to notify the caller when the lock is
44 * - blocking AST: when a lock is granted to some process, if another process
45 * enqueues a conflicting (blocking) lock on a resource, a blocking AST is
46 * sent to notify the holder(s) of the lock(s) of the conflicting lock
47 * request. The lock holder(s) must release their lock(s) on that resource in
48 * a timely manner or be evicted by the server.
50 * - glimpse AST: this is used when a process wants information about a lock
51 * (i.e. the lock value block (LVB)) but does not necessarily require holding
52 * the lock. If the resource is locked, the lock holder(s) are sent glimpse
53 * ASTs and the LVB is returned to the caller, and lock holder(s) may CANCEL
54 * their lock(s) if they are idle. If the resource is not locked, the server
58 #define DEBUG_SUBSYSTEM S_LDLM
60 #include <lustre_errno.h>
61 #include <lustre_dlm.h>
62 #include <obd_class.h>
65 #include "ldlm_internal.h"
67 unsigned int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
68 module_param(ldlm_enqueue_min, uint, 0644);
69 MODULE_PARM_DESC(ldlm_enqueue_min, "lock enqueue timeout minimum");
71 /* in client side, whether the cached locks will be canceled before replay */
72 unsigned int ldlm_cancel_unused_locks_before_replay = 1;
74 static void interrupted_completion_wait(void *data)
78 struct lock_wait_data {
79 struct ldlm_lock *lwd_lock;
83 struct ldlm_async_args {
84 struct lustre_handle lock_handle;
88 * ldlm_request_bufsize
90 * If opcode=LDLM_ENQUEUE, 1 slot is already occupied,
91 * LDLM_LOCKREQ_HANDLE -1 slots are available.
92 * Otherwise, LDLM_LOCKREQ_HANDLE slots are available.
97 * \retval size of the request buffer
100 int ldlm_request_bufsize(int count, int type)
102 int avail = LDLM_LOCKREQ_HANDLES;
103 if (type == LDLM_ENQUEUE)
104 avail -= LDLM_ENQUEUE_CANCEL_OFF;
107 avail = (count - avail) * sizeof(struct lustre_handle);
111 return sizeof(struct ldlm_request) + avail;
114 int ldlm_expired_completion_wait(void *data)
116 struct lock_wait_data *lwd = data;
117 struct ldlm_lock *lock = lwd->lwd_lock;
118 struct obd_import *imp;
119 struct obd_device *obd;
122 if (lock->l_conn_export == NULL) {
123 static time64_t next_dump, last_dump;
125 LDLM_ERROR(lock, "lock timed out (enqueued at %lld, %llds ago); "
126 "not entering recovery in server code, just going back to sleep",
127 (s64)lock->l_last_activity,
128 (s64)(ktime_get_real_seconds() -
129 lock->l_last_activity));
130 if (ktime_get_seconds() > next_dump) {
131 last_dump = next_dump;
132 next_dump = ktime_get_seconds() + 300;
133 ldlm_namespace_dump(D_DLMTRACE,
134 ldlm_lock_to_ns(lock));
136 libcfs_debug_dumplog();
141 obd = lock->l_conn_export->exp_obd;
142 imp = obd->u.cli.cl_import;
143 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
144 LDLM_ERROR(lock, "lock timed out (enqueued at %lld, %llds ago), entering recovery for %s@%s",
145 (s64)lock->l_last_activity,
146 (s64)(ktime_get_real_seconds() - lock->l_last_activity),
147 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
153 * Calculate the Completion timeout (covering enqueue, BL AST, data flush,
154 * lock cancel, and their replies). Used for lock completion timeout on the
157 * \param[in] lock lock which is waiting the completion callback
159 * \retval timeout in seconds to wait for the server reply
162 /* We use the same basis for both server side and client side functions
163 from a single node. */
164 static time64_t ldlm_cp_timeout(struct ldlm_lock *lock)
171 /* Wait a long time for enqueue - server may have to callback a
172 * lock from another client. Server will evict the other client if it
173 * doesn't respond reasonably, and then give us the lock. */
174 timeout = at_get(ldlm_lock_to_ns_at(lock));
175 return max(3 * timeout, (time64_t) ldlm_enqueue_min);
179 * Helper function for ldlm_completion_ast(), updating timings when lock is
182 static int ldlm_completion_tail(struct ldlm_lock *lock, void *data)
187 if (ldlm_is_destroyed(lock) || ldlm_is_failed(lock)) {
188 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
190 } else if (data == NULL) {
191 LDLM_DEBUG(lock, "client-side enqueue: granted");
193 /* Take into AT only CP RPC, not immediately granted locks */
194 delay = ktime_get_real_seconds() - lock->l_last_activity;
195 LDLM_DEBUG(lock, "client-side enqueue: granted after %llds",
198 /* Update our time estimate */
199 at_measured(ldlm_lock_to_ns_at(lock), delay);
205 * Implementation of ->l_completion_ast() for a client, that doesn't wait
206 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
207 * other threads that cannot block for long.
209 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
213 if (flags == LDLM_FL_WAIT_NOREPROC) {
214 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
218 if (!(flags & LDLM_FL_BLOCKED_MASK)) {
219 wake_up(&lock->l_waitq);
220 RETURN(ldlm_completion_tail(lock, data));
223 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
225 ldlm_reprocess_all(lock->l_resource);
228 EXPORT_SYMBOL(ldlm_completion_ast_async);
231 * Generic LDLM "completion" AST. This is called in several cases:
233 * - when a reply to an ENQUEUE RPC is received from the server
234 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
235 * this point (determined by flags);
237 * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
240 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
243 * - to force all locks when resource is destroyed (cleanup_resource());
245 * - during lock conversion (not used currently).
247 * If lock is not granted in the first case, this function waits until second
248 * or penultimate cases happen in some other thread.
251 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
253 /* XXX ALLOCATE - 160 bytes */
254 struct lock_wait_data lwd;
255 struct obd_device *obd;
256 struct obd_import *imp = NULL;
257 struct l_wait_info lwi;
262 if (flags == LDLM_FL_WAIT_NOREPROC) {
263 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
267 if (!(flags & LDLM_FL_BLOCKED_MASK)) {
268 wake_up(&lock->l_waitq);
272 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
277 obd = class_exp2obd(lock->l_conn_export);
279 /* if this is a local lock, then there is no import */
281 imp = obd->u.cli.cl_import;
284 timeout = ldlm_cp_timeout(lock);
287 lock->l_last_activity = ktime_get_real_seconds();
289 if (ldlm_is_no_timeout(lock)) {
290 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
291 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
293 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
294 ldlm_expired_completion_wait,
295 interrupted_completion_wait, &lwd);
299 spin_lock(&imp->imp_lock);
300 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
301 spin_unlock(&imp->imp_lock);
304 if (ns_is_client(ldlm_lock_to_ns(lock)) &&
305 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
306 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
307 ldlm_set_fail_loc(lock);
310 /* Go to sleep until the lock is granted or cancelled. */
311 rc = l_wait_event(lock->l_waitq,
312 is_granted_or_cancelled(lock), &lwi);
316 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
321 RETURN(ldlm_completion_tail(lock, data));
323 EXPORT_SYMBOL(ldlm_completion_ast);
326 * A helper to build a blocking AST function
328 * Perform a common operation for blocking ASTs:
329 * defferred lock cancellation.
331 * \param lock the lock blocking or canceling AST was called on
333 * \see mdt_blocking_ast
334 * \see ldlm_blocking_ast
336 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
341 ldlm_set_cbpending(lock);
342 do_ast = (!lock->l_readers && !lock->l_writers);
343 unlock_res_and_lock(lock);
346 struct lustre_handle lockh;
349 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
350 ldlm_lock2handle(lock, &lockh);
351 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
353 CERROR("ldlm_cli_cancel: %d\n", rc);
355 LDLM_DEBUG(lock, "Lock still has references, will be "
360 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
363 * Server blocking AST
365 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
368 * \param lock the lock which blocks a request or cancelling lock
371 * \param flag indicates whether this cancelling or blocking callback
373 * \see ldlm_blocking_ast_nocheck
375 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
376 void *data, int flag)
380 if (flag == LDLM_CB_CANCELING) {
381 /* Don't need to do anything here. */
385 lock_res_and_lock(lock);
386 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
387 * that ldlm_blocking_ast is called just before intent_policy method
388 * takes the lr_lock, then by the time we get the lock, we might not
389 * be the correct blocking function anymore. So check, and return
391 if (lock->l_blocking_ast != ldlm_blocking_ast) {
392 unlock_res_and_lock(lock);
395 RETURN(ldlm_blocking_ast_nocheck(lock));
397 EXPORT_SYMBOL(ldlm_blocking_ast);
400 * Implements ldlm_lock::l_glimpse_ast for extent locks acquired on the server.
402 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for that is
403 * rather subtle: with OST-side locking, it may so happen that _all_ extent
404 * locks are held by the OST. If client wants to obtain the current file size
405 * it calls ll_glimpse_size(), and (as all locks are held only on the server),
406 * this dummy glimpse callback fires and does nothing. The client still
407 * receives the correct file size due to the following fragment of code in
408 * ldlm_cb_interpret():
410 * if (rc == -ELDLM_NO_LOCK_DATA) {
411 * LDLM_DEBUG(lock, "lost race - client has a lock but no"
413 * ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
416 * That is, after the glimpse returns this error, ofd_lvbo_update() is called
417 * and returns the updated file attributes from the inode to the client.
419 * See also comment in ofd_intent_policy() on why servers must set a non-NULL
420 * l_glimpse_ast when grabbing DLM locks. Otherwise, the server will assume
421 * that the object is in the process of being destroyed.
423 * \param[in] lock DLM lock being glimpsed, unused
424 * \param[in] reqp pointer to ptlrpc_request, unused
426 * \retval -ELDLM_NO_LOCK_DATA to get attributes from disk object
428 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
430 return -ELDLM_NO_LOCK_DATA;
434 * Enqueue a local lock (typically on a server).
436 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
437 const struct ldlm_res_id *res_id,
438 enum ldlm_type type, union ldlm_policy_data *policy,
439 enum ldlm_mode mode, __u64 *flags,
440 ldlm_blocking_callback blocking,
441 ldlm_completion_callback completion,
442 ldlm_glimpse_callback glimpse,
443 void *data, __u32 lvb_len, enum lvb_type lvb_type,
444 const __u64 *client_cookie,
445 struct lustre_handle *lockh)
447 struct ldlm_lock *lock;
449 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
450 .lcs_blocking = blocking,
451 .lcs_glimpse = glimpse,
455 LASSERT(!(*flags & LDLM_FL_REPLAY));
456 if (unlikely(ns_is_client(ns))) {
457 CERROR("Trying to enqueue local lock in a shadow namespace\n");
461 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
464 GOTO(out_nolock, err = PTR_ERR(lock));
466 err = ldlm_lvbo_init(lock->l_resource);
468 LDLM_ERROR(lock, "delayed lvb init failed (rc %d)", err);
472 ldlm_lock2handle(lock, lockh);
474 /* NB: we don't have any lock now (lock_res_and_lock)
475 * because it's a new lock */
476 ldlm_lock_addref_internal_nolock(lock, mode);
477 ldlm_set_local(lock);
478 if (*flags & LDLM_FL_ATOMIC_CB)
479 ldlm_set_atomic_cb(lock);
482 lock->l_policy_data = *policy;
483 if (client_cookie != NULL)
484 lock->l_client_cookie = *client_cookie;
485 if (type == LDLM_EXTENT) {
486 /* extent lock without policy is a bug */
490 lock->l_req_extent = policy->l_extent;
493 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
494 if (unlikely(err != ELDLM_OK))
498 *policy = lock->l_policy_data;
500 if (lock->l_completion_ast)
501 lock->l_completion_ast(lock, *flags, NULL);
503 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
506 LDLM_LOCK_RELEASE(lock);
510 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
512 static void failed_lock_cleanup(struct ldlm_namespace *ns,
513 struct ldlm_lock *lock, int mode)
517 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
518 lock_res_and_lock(lock);
519 /* Check that lock is not granted or failed, we might race. */
520 if ((lock->l_req_mode != lock->l_granted_mode) &&
521 !ldlm_is_failed(lock)) {
522 /* Make sure that this lock will not be found by raced
523 * bl_ast and -EINVAL reply is sent to server anyways.
525 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
526 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
529 unlock_res_and_lock(lock);
533 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
534 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
536 LDLM_DEBUG(lock, "lock was granted or failed in race");
538 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
539 * from llite/file.c/ll_file_flock(). */
540 /* This code makes for the fact that we do not have blocking handler on
541 * a client for flock locks. As such this is the place where we must
542 * completely kill failed locks. (interrupted and those that
543 * were waiting to be granted when server evicted us. */
544 if (lock->l_resource->lr_type == LDLM_FLOCK) {
545 lock_res_and_lock(lock);
546 if (!ldlm_is_destroyed(lock)) {
547 ldlm_resource_unlink_lock(lock);
548 ldlm_lock_decref_internal_nolock(lock, mode);
549 ldlm_lock_destroy_nolock(lock);
551 unlock_res_and_lock(lock);
553 ldlm_lock_decref_internal(lock, mode);
558 * Finishing portion of client lock enqueue code.
560 * Called after receiving reply from server.
562 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
563 enum ldlm_type type, __u8 with_policy,
564 enum ldlm_mode mode, __u64 *flags, void *lvb,
565 __u32 lvb_len, const struct lustre_handle *lockh,
568 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
569 int is_replay = *flags & LDLM_FL_REPLAY;
570 struct ldlm_lock *lock;
571 struct ldlm_reply *reply;
572 int cleanup_phase = 1;
575 lock = ldlm_handle2lock(lockh);
576 /* ldlm_cli_enqueue is holding a reference on this lock. */
578 LASSERT(type == LDLM_FLOCK);
582 LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
583 "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
585 if (rc != ELDLM_OK) {
587 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
588 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
590 if (rc != ELDLM_LOCK_ABORTED)
594 /* Before we return, swab the reply */
595 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
597 GOTO(cleanup, rc = -EPROTO);
602 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
605 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
606 GOTO(cleanup, rc = size);
607 } else if (unlikely(size > lvb_len)) {
608 LDLM_ERROR(lock, "Replied LVB is larger than "
609 "expectation, expected = %d, replied = %d",
611 GOTO(cleanup, rc = -EINVAL);
616 if (rc == ELDLM_LOCK_ABORTED) {
617 if (lvb_len > 0 && lvb != NULL)
618 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
620 GOTO(cleanup, rc = rc ? : ELDLM_LOCK_ABORTED);
623 /* lock enqueued on the server */
626 lock_res_and_lock(lock);
627 /* Key change rehash lock in per-export hash with new key */
628 if (exp->exp_lock_hash) {
629 /* In the function below, .hs_keycmp resolves to
630 * ldlm_export_lock_keycmp() */
631 /* coverity[overrun-buffer-val] */
632 cfs_hash_rehash_key(exp->exp_lock_hash,
633 &lock->l_remote_handle,
637 lock->l_remote_handle = reply->lock_handle;
640 *flags = ldlm_flags_from_wire(reply->lock_flags);
641 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
642 LDLM_FL_INHERIT_MASK);
643 unlock_res_and_lock(lock);
645 CDEBUG(D_INFO, "local: %p, remote cookie: %#llx, flags: %#llx\n",
646 lock, reply->lock_handle.cookie, *flags);
648 /* If enqueue returned a blocked lock but the completion handler has
649 * already run, then it fixed up the resource and we don't need to do it
651 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
652 int newmode = reply->lock_desc.l_req_mode;
654 if (newmode && newmode != lock->l_req_mode) {
655 LDLM_DEBUG(lock, "server returned different mode %s",
656 ldlm_lockname[newmode]);
657 lock->l_req_mode = newmode;
660 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
661 &lock->l_resource->lr_name)) {
662 CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES
663 " instead of "DLDLMRES"\n",
664 PLDLMRES(&reply->lock_desc.l_resource),
665 PLDLMRES(lock->l_resource));
667 rc = ldlm_lock_change_resource(ns, lock,
668 &reply->lock_desc.l_resource.lr_name);
669 if (rc || lock->l_resource == NULL)
670 GOTO(cleanup, rc = -ENOMEM);
671 LDLM_DEBUG(lock, "client-side enqueue, new resource");
675 /* We assume lock type cannot change on server*/
676 ldlm_convert_policy_to_local(exp,
677 lock->l_resource->lr_type,
678 &reply->lock_desc.l_policy_data,
679 &lock->l_policy_data);
682 if (type != LDLM_PLAIN)
683 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
686 if ((*flags) & LDLM_FL_AST_SENT) {
687 lock_res_and_lock(lock);
688 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
689 unlock_res_and_lock(lock);
690 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
693 /* If the lock has already been granted by a completion AST, don't
694 * clobber the LVB with an older one. */
696 /* We must lock or a racing completion might update lvb without
697 * letting us know and we'll clobber the correct value.
698 * Cannot unlock after the check either, a that still leaves
699 * a tiny window for completion to get in */
700 lock_res_and_lock(lock);
701 if (lock->l_req_mode != lock->l_granted_mode)
702 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
703 lock->l_lvb_data, lvb_len);
704 unlock_res_and_lock(lock);
712 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
713 if (lock->l_completion_ast != NULL) {
714 int err = lock->l_completion_ast(lock, *flags, NULL);
722 if (lvb_len > 0 && lvb != NULL) {
723 /* Copy the LVB here, and not earlier, because the completion
724 * AST (if any) can override what we got in the reply */
725 memcpy(lvb, lock->l_lvb_data, lvb_len);
728 LDLM_DEBUG(lock, "client-side enqueue END");
731 if (cleanup_phase == 1 && rc)
732 failed_lock_cleanup(ns, lock, mode);
733 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
735 LDLM_LOCK_RELEASE(lock);
738 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
741 * Estimate number of lock handles that would fit into request of given
742 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
743 * a single page on the send/receive side. XXX: 512 should be changed to
744 * more adequate value.
746 static inline int ldlm_req_handles_avail(int req_size, int off)
750 avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512) - req_size;
751 if (likely(avail >= 0))
752 avail /= (int)sizeof(struct lustre_handle);
755 avail += LDLM_LOCKREQ_HANDLES - off;
760 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
761 enum req_location loc,
764 __u32 size = req_capsule_msg_size(pill, loc);
765 return ldlm_req_handles_avail(size, off);
768 static inline int ldlm_format_handles_avail(struct obd_import *imp,
769 const struct req_format *fmt,
770 enum req_location loc, int off)
772 __u32 size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
773 return ldlm_req_handles_avail(size, off);
777 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
778 * \a count locks in \a cancels.
780 * This is to be called by functions preparing their own requests that
781 * might contain lists of locks to cancel in addition to actual operation
782 * that needs to be performed.
784 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
785 int version, int opc, int canceloff,
786 struct list_head *cancels, int count)
788 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
789 struct req_capsule *pill = &req->rq_pill;
790 struct ldlm_request *dlm = NULL;
791 struct list_head head = LIST_HEAD_INIT(head);
792 enum ldlm_lru_flags lru_flags;
793 int avail, to_free, pack = 0;
799 if (ns_connect_cancelset(ns)) {
800 /* Estimate the amount of available space in the request. */
801 req_capsule_filled_sizes(pill, RCL_CLIENT);
802 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
804 lru_flags = LDLM_LRU_FLAG_NO_WAIT | (ns_connect_lru_resize(ns) ?
805 LDLM_LRU_FLAG_LRUR : LDLM_LRU_FLAG_AGED);
806 to_free = !ns_connect_lru_resize(ns) &&
807 opc == LDLM_ENQUEUE ? 1 : 0;
809 /* Cancel LRU locks here _only_ if the server supports
810 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
811 * RPC, which will make us slower. */
813 count += ldlm_cancel_lru_local(ns, cancels, to_free,
820 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
821 ldlm_request_bufsize(pack, opc));
824 rc = ptlrpc_request_pack(req, version, opc);
826 ldlm_lock_list_put(cancels, l_bl_ast, count);
830 if (ns_connect_cancelset(ns)) {
832 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
834 /* Skip first lock handler in ldlm_request_pack(),
835 * this method will increment @lock_count according
836 * to the lock handle amount actually written to
838 dlm->lock_count = canceloff;
840 /* Pack into the request @pack lock handles. */
841 ldlm_cli_cancel_list(cancels, pack, req, 0);
842 /* Prepare and send separate cancel RPC for others. */
843 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
845 ldlm_lock_list_put(cancels, l_bl_ast, count);
849 EXPORT_SYMBOL(ldlm_prep_elc_req);
851 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
852 struct list_head *cancels, int count)
854 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
855 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
857 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
859 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
861 struct ptlrpc_request *req;
865 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
867 RETURN(ERR_PTR(-ENOMEM));
869 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
871 ptlrpc_request_free(req);
875 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
876 ptlrpc_request_set_replen(req);
879 EXPORT_SYMBOL(ldlm_enqueue_pack);
882 * Client-side lock enqueue.
884 * If a request has some specific initialisation it is passed in \a reqp,
885 * otherwise it is created in ldlm_cli_enqueue.
887 * Supports sync and async requests, pass \a async flag accordingly. If a
888 * request was created in ldlm_cli_enqueue and it is the async request,
889 * pass it to the caller in \a reqp.
891 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
892 struct ldlm_enqueue_info *einfo,
893 const struct ldlm_res_id *res_id,
894 union ldlm_policy_data const *policy, __u64 *flags,
895 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
896 struct lustre_handle *lockh, int async)
898 struct ldlm_namespace *ns;
899 struct ldlm_lock *lock;
900 struct ldlm_request *body;
901 int is_replay = *flags & LDLM_FL_REPLAY;
902 int req_passed_in = 1;
904 struct ptlrpc_request *req;
907 LASSERT(exp != NULL);
909 ns = exp->exp_obd->obd_namespace;
911 /* If we're replaying this lock, just check some invariants.
912 * If we're creating a new lock, get everything all setup nice. */
914 lock = ldlm_handle2lock_long(lockh, 0);
915 LASSERT(lock != NULL);
916 LDLM_DEBUG(lock, "client-side enqueue START");
917 LASSERT(exp == lock->l_conn_export);
919 const struct ldlm_callback_suite cbs = {
920 .lcs_completion = einfo->ei_cb_cp,
921 .lcs_blocking = einfo->ei_cb_bl,
922 .lcs_glimpse = einfo->ei_cb_gl
924 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
925 einfo->ei_mode, &cbs, einfo->ei_cbdata,
928 RETURN(PTR_ERR(lock));
929 /* for the local lock, add the reference */
930 ldlm_lock_addref_internal(lock, einfo->ei_mode);
931 ldlm_lock2handle(lock, lockh);
933 lock->l_policy_data = *policy;
935 if (einfo->ei_type == LDLM_EXTENT) {
936 /* extent lock without policy is a bug */
940 lock->l_req_extent = policy->l_extent;
942 LDLM_DEBUG(lock, "client-side enqueue START, flags %#llx",
946 lock->l_conn_export = exp;
947 lock->l_export = NULL;
948 lock->l_blocking_ast = einfo->ei_cb_bl;
949 lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
950 lock->l_last_activity = ktime_get_real_seconds();
952 /* lock not sent to server yet */
953 if (reqp == NULL || *reqp == NULL) {
954 req = ldlm_enqueue_pack(exp, lvb_len);
956 failed_lock_cleanup(ns, lock, einfo->ei_mode);
957 LDLM_LOCK_RELEASE(lock);
958 RETURN(PTR_ERR(req));
968 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
970 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
971 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
974 /* Dump lock data into the request buffer */
975 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
976 ldlm_lock2desc(lock, &body->lock_desc);
977 body->lock_flags = ldlm_flags_to_wire(*flags);
978 body->lock_handle[0] = *lockh;
980 /* extended LDLM opcodes in client stats */
981 if (exp->exp_obd->obd_svc_stats != NULL) {
982 bool glimpse = *flags & LDLM_FL_HAS_INTENT;
984 /* OST glimpse has no intent buffer */
985 if (req_capsule_has_field(&req->rq_pill, &RMF_LDLM_INTENT,
987 struct ldlm_intent *it;
989 it = req_capsule_client_get(&req->rq_pill,
991 glimpse = (it && (it->opc == IT_GLIMPSE));
995 ldlm_svc_get_eopc(body, exp->exp_obd->obd_svc_stats);
997 lprocfs_counter_incr(exp->exp_obd->obd_svc_stats,
999 LDLM_GLIMPSE_ENQUEUE);
1003 LASSERT(reqp != NULL);
1007 LDLM_DEBUG(lock, "sending request");
1009 rc = ptlrpc_queue_wait(req);
1011 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
1012 einfo->ei_mode, flags, lvb, lvb_len,
1015 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
1016 * one reference that we took */
1018 LDLM_LOCK_RELEASE(lock);
1022 if (!req_passed_in && req != NULL) {
1023 ptlrpc_req_finished(req);
1030 EXPORT_SYMBOL(ldlm_cli_enqueue);
1032 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
1035 struct ldlm_resource *res;
1038 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1039 CERROR("Trying to cancel local lock\n");
1042 LDLM_DEBUG(lock, "client-side local convert");
1044 res = ldlm_lock_convert(lock, new_mode, flags);
1046 ldlm_reprocess_all(res);
1049 rc = LUSTRE_EDEADLK;
1051 LDLM_DEBUG(lock, "client-side local convert handler END");
1052 LDLM_LOCK_PUT(lock);
1056 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
1057 * conversion of locks which are on the waiting or converting queue */
1058 /* Caller of this code is supposed to take care of lock readers/writers
1060 int ldlm_cli_convert(const struct lustre_handle *lockh, int new_mode,
1063 struct ldlm_request *body;
1064 struct ldlm_reply *reply;
1065 struct ldlm_lock *lock;
1066 struct ldlm_resource *res;
1067 struct ptlrpc_request *req;
1071 lock = ldlm_handle2lock(lockh);
1078 if (lock->l_conn_export == NULL)
1079 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
1081 LDLM_DEBUG(lock, "client-side convert");
1083 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
1084 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1087 LDLM_LOCK_PUT(lock);
1091 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1092 body->lock_handle[0] = lock->l_remote_handle;
1094 body->lock_desc.l_req_mode = new_mode;
1095 body->lock_flags = ldlm_flags_to_wire(*flags);
1098 ptlrpc_request_set_replen(req);
1099 rc = ptlrpc_queue_wait(req);
1103 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1105 GOTO(out, rc = -EPROTO);
1108 GOTO(out, rc = req->rq_status);
1110 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
1112 ldlm_reprocess_all(res);
1113 /* Go to sleep until the lock is granted. */
1114 /* FIXME: or cancelled. */
1115 if (lock->l_completion_ast) {
1116 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1122 rc = LUSTRE_EDEADLK;
1126 LDLM_LOCK_PUT(lock);
1127 ptlrpc_req_finished(req);
1132 * Cancel locks locally.
1134 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1135 * \retval LDLM_FL_CANCELING otherwise;
1136 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1138 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1140 __u64 rc = LDLM_FL_LOCAL_ONLY;
1143 if (lock->l_conn_export) {
1146 LDLM_DEBUG(lock, "client-side cancel");
1147 /* Set this flag to prevent others from getting new references*/
1148 lock_res_and_lock(lock);
1149 ldlm_set_cbpending(lock);
1150 local_only = !!(lock->l_flags &
1151 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1152 ldlm_cancel_callback(lock);
1153 rc = (ldlm_is_bl_ast(lock)) ?
1154 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1155 unlock_res_and_lock(lock);
1158 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1160 rc = LDLM_FL_LOCAL_ONLY;
1162 ldlm_lock_cancel(lock);
1164 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1165 LDLM_ERROR(lock, "Trying to cancel local lock");
1168 LDLM_DEBUG(lock, "server-side local cancel");
1169 ldlm_lock_cancel(lock);
1170 ldlm_reprocess_all(lock->l_resource);
1177 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1179 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1180 struct list_head *head, int count)
1182 struct ldlm_request *dlm;
1183 struct ldlm_lock *lock;
1184 int max, packed = 0;
1187 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1188 LASSERT(dlm != NULL);
1190 /* Check the room in the request buffer. */
1191 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1192 sizeof(struct ldlm_request);
1193 max /= sizeof(struct lustre_handle);
1194 max += LDLM_LOCKREQ_HANDLES;
1195 LASSERT(max >= dlm->lock_count + count);
1197 /* XXX: it would be better to pack lock handles grouped by resource.
1198 * so that the server cancel would call filter_lvbo_update() less
1200 list_for_each_entry(lock, head, l_bl_ast) {
1203 LASSERT(lock->l_conn_export);
1204 /* Pack the lock handle to the given request buffer. */
1205 LDLM_DEBUG(lock, "packing");
1206 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1209 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1214 * Prepare and send a batched cancel RPC. It will include \a count lock
1215 * handles of locks given in \a cancels list. */
1216 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1217 int count, enum ldlm_cancel_flags flags)
1219 struct ptlrpc_request *req = NULL;
1220 struct obd_import *imp;
1225 LASSERT(exp != NULL);
1228 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1230 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1233 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1234 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1239 imp = class_exp2cliimp(exp);
1240 if (imp == NULL || imp->imp_invalid) {
1242 "skipping cancel on invalid import %p\n", imp);
1246 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1248 GOTO(out, rc = -ENOMEM);
1250 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1251 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1252 ldlm_request_bufsize(count, LDLM_CANCEL));
1254 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1256 ptlrpc_request_free(req);
1260 /* If OSP want cancel cross-MDT lock, let's not block it in
1261 * in recovery, otherwise the lock will not released, if
1262 * the remote target is also in recovery, and it also need
1263 * this lock, it might cause deadlock. */
1264 if (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS &&
1265 exp->exp_obd->obd_lu_dev != NULL &&
1266 exp->exp_obd->obd_lu_dev->ld_site != NULL) {
1267 struct lu_device *top_dev;
1269 top_dev = exp->exp_obd->obd_lu_dev->ld_site->ls_top_dev;
1270 if (top_dev != NULL &&
1271 top_dev->ld_obd->obd_recovering)
1272 req->rq_allow_replay = 1;
1275 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1276 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1277 ptlrpc_at_set_req_timeout(req);
1279 ldlm_cancel_pack(req, cancels, count);
1281 ptlrpc_request_set_replen(req);
1282 if (flags & LCF_ASYNC) {
1283 ptlrpcd_add_req(req);
1288 rc = ptlrpc_queue_wait(req);
1289 if (rc == LUSTRE_ESTALE) {
1290 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1291 "out of sync -- not fatal\n",
1292 libcfs_nid2str(req->rq_import->
1293 imp_connection->c_peer.nid));
1295 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1296 req->rq_import_generation == imp->imp_generation) {
1297 ptlrpc_req_finished(req);
1299 } else if (rc != ELDLM_OK) {
1300 /* -ESHUTDOWN is common on umount */
1301 CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1302 "Got rc %d from cancel RPC: "
1303 "canceling anyway\n", rc);
1310 ptlrpc_req_finished(req);
1313 return sent ? sent : rc;
1316 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1318 LASSERT(imp != NULL);
1319 return &imp->imp_obd->obd_namespace->ns_pool;
1323 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1325 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1327 struct obd_device *obd;
1331 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1332 !imp_connect_lru_resize(req->rq_import)))
1335 * Do nothing for corner cases.
1340 /* In some cases RPC may contain SLV and limit zeroed out. This
1341 * is the case when server does not support LRU resize feature.
1342 * This is also possible in some recovery cases when server-side
1343 * reqs have no reference to the OBD export and thus access to
1344 * server-side namespace is not possible. */
1345 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1346 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1347 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1348 "(SLV: %llu, Limit: %u)",
1349 lustre_msg_get_slv(req->rq_repmsg),
1350 lustre_msg_get_limit(req->rq_repmsg));
1354 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1355 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1356 obd = req->rq_import->imp_obd;
1358 /* Set new SLV and limit in OBD fields to make them accessible
1359 * to the pool thread. We do not access obd_namespace and pool
1360 * directly here as there is no reliable way to make sure that
1361 * they are still alive at cleanup time. Evil races are possible
1362 * which may cause Oops at that time. */
1363 write_lock(&obd->obd_pool_lock);
1364 obd->obd_pool_slv = new_slv;
1365 obd->obd_pool_limit = new_limit;
1366 write_unlock(&obd->obd_pool_lock);
1372 * Client side lock cancel.
1374 * Lock must not have any readers or writers by this time.
1376 int ldlm_cli_cancel(const struct lustre_handle *lockh,
1377 enum ldlm_cancel_flags cancel_flags)
1379 struct obd_export *exp;
1380 enum ldlm_lru_flags lru_flags;
1381 int avail, count = 1;
1383 struct ldlm_namespace *ns;
1384 struct ldlm_lock *lock;
1385 struct list_head cancels = LIST_HEAD_INIT(cancels);
1388 lock = ldlm_handle2lock_long(lockh, 0);
1390 LDLM_DEBUG_NOLOCK("lock is already being destroyed");
1394 lock_res_and_lock(lock);
1395 /* Lock is being canceled and the caller doesn't want to wait */
1396 if (ldlm_is_canceling(lock) && (cancel_flags & LCF_ASYNC)) {
1397 unlock_res_and_lock(lock);
1398 LDLM_LOCK_RELEASE(lock);
1402 ldlm_set_canceling(lock);
1403 unlock_res_and_lock(lock);
1405 rc = ldlm_cli_cancel_local(lock);
1406 if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
1407 LDLM_LOCK_RELEASE(lock);
1410 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1411 * RPC which goes to canceld portal, so we can cancel other LRU locks
1412 * here and send them all as one LDLM_CANCEL RPC. */
1413 LASSERT(list_empty(&lock->l_bl_ast));
1414 list_add(&lock->l_bl_ast, &cancels);
1416 exp = lock->l_conn_export;
1417 if (exp_connect_cancelset(exp)) {
1418 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1423 ns = ldlm_lock_to_ns(lock);
1424 lru_flags = ns_connect_lru_resize(ns) ?
1425 LDLM_LRU_FLAG_LRUR : LDLM_LRU_FLAG_AGED;
1426 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1427 LCF_BL_AST, lru_flags);
1429 ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1432 EXPORT_SYMBOL(ldlm_cli_cancel);
1435 * Locally cancel up to \a count locks in list \a cancels.
1436 * Return the number of cancelled locks.
1438 int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
1439 enum ldlm_cancel_flags cancel_flags)
1441 struct list_head head = LIST_HEAD_INIT(head);
1442 struct ldlm_lock *lock, *next;
1443 int left = 0, bl_ast = 0;
1447 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1451 if (cancel_flags & LCF_LOCAL) {
1452 rc = LDLM_FL_LOCAL_ONLY;
1453 ldlm_lock_cancel(lock);
1455 rc = ldlm_cli_cancel_local(lock);
1457 /* Until we have compound requests and can send LDLM_CANCEL
1458 * requests batched with generic RPCs, we need to send cancels
1459 * with the LDLM_FL_BL_AST flag in a separate RPC from
1460 * the one being generated now. */
1461 if (!(cancel_flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1462 LDLM_DEBUG(lock, "Cancel lock separately");
1463 list_del_init(&lock->l_bl_ast);
1464 list_add(&lock->l_bl_ast, &head);
1468 if (rc == LDLM_FL_LOCAL_ONLY) {
1469 /* CANCEL RPC should not be sent to server. */
1470 list_del_init(&lock->l_bl_ast);
1471 LDLM_LOCK_RELEASE(lock);
1477 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1484 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1485 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1486 * readahead requests, ...)
1488 static enum ldlm_policy_res
1489 ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1490 int unused, int added, int count)
1492 enum ldlm_policy_res result = LDLM_POLICY_CANCEL_LOCK;
1494 /* don't check added & count since we want to process all locks
1496 * It's fine to not take lock to access lock->l_resource since
1497 * the lock has already been granted so it won't change. */
1498 switch (lock->l_resource->lr_type) {
1501 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) != 0)
1504 result = LDLM_POLICY_SKIP_LOCK;
1505 lock_res_and_lock(lock);
1506 ldlm_set_skipped(lock);
1507 unlock_res_and_lock(lock);
1515 * Callback function for LRU-resize policy. Decides whether to keep
1516 * \a lock in LRU for current \a LRU size \a unused, added in current
1517 * scan \a added and number of locks to be preferably canceled \a count.
1519 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1521 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1523 static enum ldlm_policy_res ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1524 struct ldlm_lock *lock,
1525 int unused, int added,
1528 ktime_t cur = ktime_get();
1529 struct ldlm_pool *pl = &ns->ns_pool;
1533 /* Stop LRU processing when we reach past @count or have checked all
1535 if (count && added >= count)
1536 return LDLM_POLICY_KEEP_LOCK;
1538 /* Despite of the LV, It doesn't make sense to keep the lock which
1539 * is unused for ns_max_age time.
1541 if (ktime_after(ktime_get(),
1542 ktime_add(lock->l_last_used, ns->ns_max_age)))
1543 return LDLM_POLICY_CANCEL_LOCK;
1545 slv = ldlm_pool_get_slv(pl);
1546 lvf = ldlm_pool_get_lvf(pl);
1547 la = ktime_to_ns(ktime_sub(cur, lock->l_last_used)) / NSEC_PER_SEC;
1548 lv = lvf * la * unused;
1550 /* Inform pool about current CLV to see it via proc. */
1551 ldlm_pool_set_clv(pl, lv);
1553 /* Stop when SLV is not yet come from server or lv is smaller than
1555 if (slv == 0 || lv < slv)
1556 return LDLM_POLICY_KEEP_LOCK;
1558 return LDLM_POLICY_CANCEL_LOCK;
1561 static enum ldlm_policy_res
1562 ldlm_cancel_lrur_no_wait_policy(struct ldlm_namespace *ns,
1563 struct ldlm_lock *lock,
1564 int unused, int added,
1567 enum ldlm_policy_res result;
1569 result = ldlm_cancel_lrur_policy(ns, lock, unused, added, count);
1570 if (result == LDLM_POLICY_KEEP_LOCK)
1573 return ldlm_cancel_no_wait_policy(ns, lock, unused, added, count);
1577 * Callback function for proc used policy. Makes decision whether to keep
1578 * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1579 * added and number of locks to be preferably canceled \a count.
1581 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1583 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1585 static enum ldlm_policy_res ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1586 struct ldlm_lock *lock,
1587 int unused, int added,
1590 /* Stop LRU processing when we reach past @count or have checked all
1592 return (added >= count) ?
1593 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1597 * Callback function for aged policy. Makes decision whether to keep \a lock in
1598 * LRU for current LRU size \a unused, added in current scan \a added and
1599 * number of locks to be preferably canceled \a count.
1601 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1603 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1605 static enum ldlm_policy_res ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1606 struct ldlm_lock *lock,
1607 int unused, int added,
1610 if ((added >= count) &&
1611 ktime_before(ktime_get(),
1612 ktime_add(lock->l_last_used, ns->ns_max_age)))
1613 return LDLM_POLICY_KEEP_LOCK;
1615 return LDLM_POLICY_CANCEL_LOCK;
1618 static enum ldlm_policy_res
1619 ldlm_cancel_aged_no_wait_policy(struct ldlm_namespace *ns,
1620 struct ldlm_lock *lock,
1621 int unused, int added, int count)
1623 enum ldlm_policy_res result;
1625 result = ldlm_cancel_aged_policy(ns, lock, unused, added, count);
1626 if (result == LDLM_POLICY_KEEP_LOCK)
1629 return ldlm_cancel_no_wait_policy(ns, lock, unused, added, count);
1633 * Callback function for default policy. Makes decision whether to keep \a lock
1634 * in LRU for current LRU size \a unused, added in current scan \a added and
1635 * number of locks to be preferably canceled \a count.
1637 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1639 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1642 enum ldlm_policy_res ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1643 struct ldlm_lock *lock,
1644 int unused, int added,
1647 /* Stop LRU processing when we reach past count or have checked all
1649 return (added >= count) ?
1650 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1653 typedef enum ldlm_policy_res
1654 (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1655 int unused, int added, int count);
1657 static ldlm_cancel_lru_policy_t
1658 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, enum ldlm_lru_flags lru_flags)
1660 if (ns_connect_lru_resize(ns)) {
1661 if (lru_flags & LDLM_LRU_FLAG_SHRINK)
1662 /* We kill passed number of old locks. */
1663 return ldlm_cancel_passed_policy;
1664 if (lru_flags & LDLM_LRU_FLAG_LRUR) {
1665 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1666 return ldlm_cancel_lrur_no_wait_policy;
1668 return ldlm_cancel_lrur_policy;
1670 if (lru_flags & LDLM_LRU_FLAG_PASSED)
1671 return ldlm_cancel_passed_policy;
1673 if (lru_flags & LDLM_LRU_FLAG_AGED) {
1674 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1675 return ldlm_cancel_aged_no_wait_policy;
1677 return ldlm_cancel_aged_policy;
1680 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1681 return ldlm_cancel_no_wait_policy;
1683 return ldlm_cancel_default_policy;
1687 * - Free space in LRU for \a count new locks,
1688 * redundant unused locks are canceled locally;
1689 * - also cancel locally unused aged locks;
1690 * - do not cancel more than \a max locks;
1691 * - GET the found locks and add them into the \a cancels list.
1693 * A client lock can be added to the l_bl_ast list only when it is
1694 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1695 * CANCEL. There are the following use cases:
1696 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1697 * ldlm_cli_cancel(), which check and set this flag properly. As any
1698 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1699 * later without any special locking.
1701 * Calling policies for enabled LRU resize:
1702 * ----------------------------------------
1703 * flags & LDLM_LRU_FLAG_LRUR - use LRU resize policy (SLV from server) to
1704 * cancel not more than \a count locks;
1706 * flags & LDLM_LRU_FLAG_PASSED - cancel \a count number of old locks (located
1707 * at the beginning of LRU list);
1709 * flags & LDLM_LRU_FLAG_SHRINK - cancel not more than \a count locks according
1710 * to memory pressre policy function;
1712 * flags & LDLM_LRU_FLAG_AGED - cancel \a count locks according to "aged policy"
1714 * flags & LDLM_LRU_FLAG_NO_WAIT - cancel as many unused locks as possible
1715 * (typically before replaying locks) w/o
1716 * sending any RPCs or waiting for any
1717 * outstanding RPC to complete.
1719 * flags & LDLM_CANCEL_CLEANUP - when cancelling read locks, do not check for
1720 * other read locks covering the same pages, just
1721 * discard those pages.
1723 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns,
1724 struct list_head *cancels, int count, int max,
1725 enum ldlm_lru_flags lru_flags)
1727 ldlm_cancel_lru_policy_t pf;
1728 struct ldlm_lock *lock, *next;
1729 int added = 0, unused, remained;
1730 int no_wait = lru_flags & LDLM_LRU_FLAG_NO_WAIT;
1733 spin_lock(&ns->ns_lock);
1734 unused = ns->ns_nr_unused;
1737 if (!ns_connect_lru_resize(ns))
1738 count += unused - ns->ns_max_unused;
1740 pf = ldlm_cancel_lru_policy(ns, lru_flags);
1741 LASSERT(pf != NULL);
1743 while (!list_empty(&ns->ns_unused_list)) {
1744 enum ldlm_policy_res result;
1745 ktime_t last_use = ktime_set(0, 0);
1747 /* all unused locks */
1748 if (remained-- <= 0)
1751 /* For any flags, stop scanning if @max is reached. */
1752 if (max && added >= max)
1755 list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1757 /* No locks which got blocking requests. */
1758 LASSERT(!ldlm_is_bl_ast(lock));
1760 if (no_wait && ldlm_is_skipped(lock))
1761 /* already processed */
1764 last_use = lock->l_last_used;
1766 /* Somebody is already doing CANCEL. No need for this
1767 * lock in LRU, do not traverse it again. */
1768 if (!ldlm_is_canceling(lock))
1771 ldlm_lock_remove_from_lru_nolock(lock);
1773 if (&lock->l_lru == &ns->ns_unused_list)
1776 LDLM_LOCK_GET(lock);
1777 spin_unlock(&ns->ns_lock);
1778 lu_ref_add(&lock->l_reference, __FUNCTION__, current);
1780 /* Pass the lock through the policy filter and see if it
1781 * should stay in LRU.
1783 * Even for shrinker policy we stop scanning if
1784 * we find a lock that should stay in the cache.
1785 * We should take into account lock age anyway
1786 * as a new lock is a valuable resource even if
1787 * it has a low weight.
1789 * That is, for shrinker policy we drop only
1790 * old locks, but additionally choose them by
1791 * their weight. Big extent locks will stay in
1793 result = pf(ns, lock, unused, added, count);
1794 if (result == LDLM_POLICY_KEEP_LOCK) {
1795 lu_ref_del(&lock->l_reference,
1796 __FUNCTION__, current);
1797 LDLM_LOCK_RELEASE(lock);
1798 spin_lock(&ns->ns_lock);
1801 if (result == LDLM_POLICY_SKIP_LOCK) {
1802 lu_ref_del(&lock->l_reference,
1804 LDLM_LOCK_RELEASE(lock);
1805 spin_lock(&ns->ns_lock);
1809 lock_res_and_lock(lock);
1810 /* Check flags again under the lock. */
1811 if (ldlm_is_canceling(lock) ||
1812 ldlm_lock_remove_from_lru_check(lock, last_use) == 0) {
1813 /* Another thread is removing lock from LRU, or
1814 * somebody is already doing CANCEL, or there
1815 * is a blocking request which will send cancel
1816 * by itself, or the lock is no longer unused or
1817 * the lock has been used since the pf() call and
1818 * pages could be put under it. */
1819 unlock_res_and_lock(lock);
1820 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1821 LDLM_LOCK_RELEASE(lock);
1822 spin_lock(&ns->ns_lock);
1825 LASSERT(!lock->l_readers && !lock->l_writers);
1827 /* If we have chosen to cancel this lock voluntarily, we
1828 * better send cancel notification to server, so that it
1829 * frees appropriate state. This might lead to a race
1830 * where while we are doing cancel here, server is also
1831 * silently cancelling this lock. */
1832 ldlm_clear_cancel_on_block(lock);
1834 /* Setting the CBPENDING flag is a little misleading,
1835 * but prevents an important race; namely, once
1836 * CBPENDING is set, the lock can accumulate no more
1837 * readers/writers. Since readers and writers are
1838 * already zero here, ldlm_lock_decref() won't see
1839 * this flag and call l_blocking_ast */
1840 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1842 if ((lru_flags & LDLM_LRU_FLAG_CLEANUP) &&
1843 (lock->l_resource->lr_type == LDLM_EXTENT ||
1844 ldlm_has_dom(lock)) && lock->l_granted_mode == LCK_PR)
1845 ldlm_set_discard_data(lock);
1847 /* We can't re-add to l_lru as it confuses the
1848 * refcounting in ldlm_lock_remove_from_lru() if an AST
1849 * arrives after we drop lr_lock below. We use l_bl_ast
1850 * and can't use l_pending_chain as it is used both on
1851 * server and client nevertheless bug 5666 says it is
1852 * used only on server */
1853 LASSERT(list_empty(&lock->l_bl_ast));
1854 list_add(&lock->l_bl_ast, cancels);
1855 unlock_res_and_lock(lock);
1856 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1857 spin_lock(&ns->ns_lock);
1861 spin_unlock(&ns->ns_lock);
1865 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1867 enum ldlm_cancel_flags cancel_flags,
1868 enum ldlm_lru_flags lru_flags)
1872 added = ldlm_prepare_lru_list(ns, cancels, count, max, lru_flags);
1876 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1880 * Cancel at least \a nr locks from given namespace LRU.
1882 * When called with LCF_ASYNC the blocking callback will be handled
1883 * in a thread and this function will return after the thread has been
1884 * asked to call the callback. When called with LCF_ASYNC the blocking
1885 * callback will be performed in this function.
1887 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
1888 enum ldlm_cancel_flags cancel_flags,
1889 enum ldlm_lru_flags lru_flags)
1891 struct list_head cancels = LIST_HEAD_INIT(cancels);
1895 /* Just prepare the list of locks, do not actually cancel them yet.
1896 * Locks are cancelled later in a separate thread. */
1897 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, lru_flags);
1898 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
1906 * Find and cancel locally unused locks found on resource, matched to the
1907 * given policy, mode. GET the found locks and add them into the \a cancels
1910 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1911 struct list_head *cancels,
1912 union ldlm_policy_data *policy,
1913 enum ldlm_mode mode, __u64 lock_flags,
1914 enum ldlm_cancel_flags cancel_flags,
1917 struct ldlm_lock *lock;
1922 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1923 if (opaque != NULL && lock->l_ast_data != opaque) {
1924 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1925 lock->l_ast_data, opaque);
1930 if (lock->l_readers || lock->l_writers)
1933 /* If somebody is already doing CANCEL, or blocking AST came,
1934 * skip this lock. */
1935 if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock))
1938 if (lockmode_compat(lock->l_granted_mode, mode))
1941 /* If policy is given and this is IBITS lock, add to list only
1942 * those locks that match by policy. */
1943 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1944 !(lock->l_policy_data.l_inodebits.bits &
1945 policy->l_inodebits.bits))
1948 /* See CBPENDING comment in ldlm_cancel_lru */
1949 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1952 LASSERT(list_empty(&lock->l_bl_ast));
1953 list_add(&lock->l_bl_ast, cancels);
1954 LDLM_LOCK_GET(lock);
1959 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
1961 EXPORT_SYMBOL(ldlm_cancel_resource_local);
1964 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1966 * If \a req is NULL, send CANCEL request to server with handles of locks
1967 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1968 * separately per lock.
1969 * If \a req is not NULL, put handles of locks in \a cancels into the request
1970 * buffer at the offset \a off.
1971 * Destroy \a cancels at the end.
1973 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1974 struct ptlrpc_request *req,
1975 enum ldlm_cancel_flags flags)
1977 struct ldlm_lock *lock;
1981 if (list_empty(cancels) || count == 0)
1984 /* XXX: requests (both batched and not) could be sent in parallel.
1985 * Usually it is enough to have just 1 RPC, but it is possible that
1986 * there are too many locks to be cancelled in LRU or on a resource.
1987 * It would also speed up the case when the server does not support
1990 LASSERT(!list_empty(cancels));
1991 lock = list_entry(cancels->next, struct ldlm_lock,
1993 LASSERT(lock->l_conn_export);
1995 if (exp_connect_cancelset(lock->l_conn_export)) {
1998 ldlm_cancel_pack(req, cancels, count);
2000 res = ldlm_cli_cancel_req(lock->l_conn_export,
2004 res = ldlm_cli_cancel_req(lock->l_conn_export,
2009 CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
2010 "ldlm_cli_cancel_list: %d\n", res);
2015 ldlm_lock_list_put(cancels, l_bl_ast, res);
2017 LASSERT(count == 0);
2020 EXPORT_SYMBOL(ldlm_cli_cancel_list);
2023 * Cancel all locks on a resource that have 0 readers/writers.
2025 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
2026 * to notify the server. */
2027 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
2028 const struct ldlm_res_id *res_id,
2029 union ldlm_policy_data *policy,
2030 enum ldlm_mode mode,
2031 enum ldlm_cancel_flags flags, void *opaque)
2033 struct ldlm_resource *res;
2034 struct list_head cancels = LIST_HEAD_INIT(cancels);
2039 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2041 /* This is not a problem. */
2042 CDEBUG(D_INFO, "No resource %llu\n", res_id->name[0]);
2046 LDLM_RESOURCE_ADDREF(res);
2047 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
2048 0, flags | LCF_BL_AST, opaque);
2049 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
2051 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
2054 LDLM_RESOURCE_DELREF(res);
2055 ldlm_resource_putref(res);
2058 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2060 struct ldlm_cli_cancel_arg {
2066 ldlm_cli_hash_cancel_unused(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2067 struct hlist_node *hnode, void *arg)
2069 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2070 struct ldlm_cli_cancel_arg *lc = arg;
2072 ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
2073 NULL, LCK_MINMODE, lc->lc_flags,
2075 /* must return 0 for hash iteration */
2080 * Cancel all locks on a namespace (or a specific resource, if given)
2081 * that have 0 readers/writers.
2083 * If flags & LCF_LOCAL, throw the locks away without trying
2084 * to notify the server. */
2085 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
2086 const struct ldlm_res_id *res_id,
2087 enum ldlm_cancel_flags flags, void *opaque)
2089 struct ldlm_cli_cancel_arg arg = {
2091 .lc_opaque = opaque,
2099 if (res_id != NULL) {
2100 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
2104 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2105 ldlm_cli_hash_cancel_unused, &arg, 0);
2110 /* Lock iterators. */
2112 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
2115 struct list_head *tmp, *next;
2116 struct ldlm_lock *lock;
2117 int rc = LDLM_ITER_CONTINUE;
2122 RETURN(LDLM_ITER_CONTINUE);
2125 list_for_each_safe(tmp, next, &res->lr_granted) {
2126 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2128 if (iter(lock, closure) == LDLM_ITER_STOP)
2129 GOTO(out, rc = LDLM_ITER_STOP);
2132 list_for_each_safe(tmp, next, &res->lr_converting) {
2133 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2135 if (iter(lock, closure) == LDLM_ITER_STOP)
2136 GOTO(out, rc = LDLM_ITER_STOP);
2139 list_for_each_safe(tmp, next, &res->lr_waiting) {
2140 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2142 if (iter(lock, closure) == LDLM_ITER_STOP)
2143 GOTO(out, rc = LDLM_ITER_STOP);
2150 struct iter_helper_data {
2151 ldlm_iterator_t iter;
2155 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2157 struct iter_helper_data *helper = closure;
2158 return helper->iter(lock, helper->closure);
2161 static int ldlm_res_iter_helper(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2162 struct hlist_node *hnode, void *arg)
2165 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2167 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2171 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2172 ldlm_iterator_t iter, void *closure)
2175 struct iter_helper_data helper = { .iter = iter, .closure = closure };
2177 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2178 ldlm_res_iter_helper, &helper, 0);
2182 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2183 * return 0: find no resource
2184 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2187 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2188 const struct ldlm_res_id *res_id,
2189 ldlm_iterator_t iter, void *data)
2191 struct ldlm_resource *res;
2195 LASSERTF(ns != NULL, "must pass in namespace\n");
2197 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2201 LDLM_RESOURCE_ADDREF(res);
2202 rc = ldlm_resource_foreach(res, iter, data);
2203 LDLM_RESOURCE_DELREF(res);
2204 ldlm_resource_putref(res);
2207 EXPORT_SYMBOL(ldlm_resource_iterate);
2211 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2213 struct list_head *list = closure;
2215 /* we use l_pending_chain here, because it's unused on clients. */
2216 LASSERTF(list_empty(&lock->l_pending_chain),
2217 "lock %p next %p prev %p\n",
2218 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2219 /* bug 9573: don't replay locks left after eviction, or
2220 * bug 17614: locks being actively cancelled. Get a reference
2221 * on a lock so that it does not disapear under us (e.g. due to cancel)
2223 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_BL_DONE))) {
2224 list_add(&lock->l_pending_chain, list);
2225 LDLM_LOCK_GET(lock);
2228 return LDLM_ITER_CONTINUE;
2231 static int replay_lock_interpret(const struct lu_env *env,
2232 struct ptlrpc_request *req,
2233 struct ldlm_async_args *aa, int rc)
2235 struct ldlm_lock *lock;
2236 struct ldlm_reply *reply;
2237 struct obd_export *exp;
2240 atomic_dec(&req->rq_import->imp_replay_inflight);
2244 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2246 GOTO(out, rc = -EPROTO);
2248 lock = ldlm_handle2lock(&aa->lock_handle);
2250 CERROR("received replay ack for unknown local cookie %#llx"
2251 " remote cookie %#llx from server %s id %s\n",
2252 aa->lock_handle.cookie, reply->lock_handle.cookie,
2253 req->rq_export->exp_client_uuid.uuid,
2254 libcfs_id2str(req->rq_peer));
2255 GOTO(out, rc = -ESTALE);
2258 /* Key change rehash lock in per-export hash with new key */
2259 exp = req->rq_export;
2260 if (exp && exp->exp_lock_hash) {
2261 /* In the function below, .hs_keycmp resolves to
2262 * ldlm_export_lock_keycmp() */
2263 /* coverity[overrun-buffer-val] */
2264 cfs_hash_rehash_key(exp->exp_lock_hash,
2265 &lock->l_remote_handle,
2266 &reply->lock_handle,
2269 lock->l_remote_handle = reply->lock_handle;
2272 LDLM_DEBUG(lock, "replayed lock:");
2273 ptlrpc_import_recovery_state_machine(req->rq_import);
2274 LDLM_LOCK_PUT(lock);
2277 ptlrpc_connect_import(req->rq_import);
2282 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2284 struct ptlrpc_request *req;
2285 struct ldlm_async_args *aa;
2286 struct ldlm_request *body;
2291 /* Bug 11974: Do not replay a lock which is actively being canceled */
2292 if (ldlm_is_bl_done(lock)) {
2293 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2297 /* If this is reply-less callback lock, we cannot replay it, since
2298 * server might have long dropped it, but notification of that event was
2299 * lost by network. (and server granted conflicting lock already) */
2300 if (ldlm_is_cancel_on_block(lock)) {
2301 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2302 ldlm_lock_cancel(lock);
2307 * If granted mode matches the requested mode, this lock is granted.
2309 * If they differ, but we have a granted mode, then we were granted
2310 * one mode and now want another: ergo, converting.
2312 * If we haven't been granted anything and are on a resource list,
2313 * then we're blocked/waiting.
2315 * If we haven't been granted anything and we're NOT on a resource list,
2316 * then we haven't got a reply yet and don't have a known disposition.
2317 * This happens whenever a lock enqueue is the request that triggers
2320 if (lock->l_granted_mode == lock->l_req_mode)
2321 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2322 else if (lock->l_granted_mode)
2323 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2324 else if (!list_empty(&lock->l_res_link))
2325 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2327 flags = LDLM_FL_REPLAY;
2329 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2330 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2334 /* We're part of recovery, so don't wait for it. */
2335 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2337 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2338 ldlm_lock2desc(lock, &body->lock_desc);
2339 body->lock_flags = ldlm_flags_to_wire(flags);
2341 ldlm_lock2handle(lock, &body->lock_handle[0]);
2342 if (lock->l_lvb_len > 0)
2343 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2344 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2346 ptlrpc_request_set_replen(req);
2347 /* notify the server we've replayed all requests.
2348 * also, we mark the request to be put on a dedicated
2349 * queue to be processed after all request replayes.
2351 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2353 LDLM_DEBUG(lock, "replaying lock:");
2355 atomic_inc(&req->rq_import->imp_replay_inflight);
2356 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2357 aa = ptlrpc_req_async_args(req);
2358 aa->lock_handle = body->lock_handle[0];
2359 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2360 ptlrpcd_add_req(req);
2366 * Cancel as many unused locks as possible before replay. since we are
2367 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2370 * Called only in recovery before replaying locks. there is no need to
2371 * replay locks that are unused. since the clients may hold thousands of
2372 * cached unused locks, dropping the unused locks can greatly reduce the
2373 * load on the servers at recovery time.
2375 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2378 struct list_head cancels = LIST_HEAD_INIT(cancels);
2380 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2381 "replay for namespace %s (%d)\n",
2382 ldlm_ns_name(ns), ns->ns_nr_unused);
2384 /* We don't need to care whether or not LRU resize is enabled
2385 * because the LDLM_LRU_FLAG_NO_WAIT policy doesn't use the
2386 * count parameter */
2387 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2388 LCF_LOCAL, LDLM_LRU_FLAG_NO_WAIT);
2390 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2391 canceled, ldlm_ns_name(ns));
2394 int ldlm_replay_locks(struct obd_import *imp)
2396 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2397 struct list_head list = LIST_HEAD_INIT(list);
2398 struct ldlm_lock *lock, *next;
2403 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2405 /* don't replay locks if import failed recovery */
2406 if (imp->imp_vbr_failed)
2409 /* ensure this doesn't fall to 0 before all have been queued */
2410 atomic_inc(&imp->imp_replay_inflight);
2412 if (ldlm_cancel_unused_locks_before_replay)
2413 ldlm_cancel_unused_locks_for_replay(ns);
2415 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2417 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2418 list_del_init(&lock->l_pending_chain);
2420 LDLM_LOCK_RELEASE(lock);
2421 continue; /* or try to do the rest? */
2423 rc = replay_one_lock(imp, lock);
2424 LDLM_LOCK_RELEASE(lock);
2427 atomic_dec(&imp->imp_replay_inflight);