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, 2017, 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_activity,
128 (s64)(ktime_get_real_seconds() -
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_activity,
146 (s64)(ktime_get_real_seconds() - lock->l_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_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 * If lock is not granted in the first case, this function waits until second
246 * or penultimate cases happen in some other thread.
249 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
251 /* XXX ALLOCATE - 160 bytes */
252 struct lock_wait_data lwd;
253 struct obd_device *obd;
254 struct obd_import *imp = NULL;
255 struct l_wait_info lwi;
260 if (flags == LDLM_FL_WAIT_NOREPROC) {
261 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
265 if (!(flags & LDLM_FL_BLOCKED_MASK)) {
266 wake_up(&lock->l_waitq);
270 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
275 obd = class_exp2obd(lock->l_conn_export);
277 /* if this is a local lock, then there is no import */
279 imp = obd->u.cli.cl_import;
282 timeout = ldlm_cp_timeout(lock);
285 lock->l_activity = ktime_get_real_seconds();
287 if (ldlm_is_no_timeout(lock)) {
288 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
289 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
291 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
292 ldlm_expired_completion_wait,
293 interrupted_completion_wait, &lwd);
297 spin_lock(&imp->imp_lock);
298 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
299 spin_unlock(&imp->imp_lock);
302 if (ns_is_client(ldlm_lock_to_ns(lock)) &&
303 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
304 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
305 ldlm_set_fail_loc(lock);
308 /* Go to sleep until the lock is granted or cancelled. */
309 rc = l_wait_event(lock->l_waitq,
310 is_granted_or_cancelled(lock), &lwi);
314 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
319 RETURN(ldlm_completion_tail(lock, data));
321 EXPORT_SYMBOL(ldlm_completion_ast);
324 * A helper to build a blocking AST function
326 * Perform a common operation for blocking ASTs:
327 * defferred lock cancellation.
329 * \param lock the lock blocking or canceling AST was called on
331 * \see mdt_blocking_ast
332 * \see ldlm_blocking_ast
334 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
339 ldlm_set_cbpending(lock);
340 do_ast = (!lock->l_readers && !lock->l_writers);
341 unlock_res_and_lock(lock);
344 struct lustre_handle lockh;
347 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
348 ldlm_lock2handle(lock, &lockh);
349 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
351 CERROR("ldlm_cli_cancel: %d\n", rc);
353 LDLM_DEBUG(lock, "Lock still has references, will be "
358 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
361 * Server blocking AST
363 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
366 * \param lock the lock which blocks a request or cancelling lock
369 * \param flag indicates whether this cancelling or blocking callback
371 * \see ldlm_blocking_ast_nocheck
373 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
374 void *data, int flag)
378 if (flag == LDLM_CB_CANCELING) {
379 /* Don't need to do anything here. */
383 lock_res_and_lock(lock);
384 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
385 * that ldlm_blocking_ast is called just before intent_policy method
386 * takes the lr_lock, then by the time we get the lock, we might not
387 * be the correct blocking function anymore. So check, and return
389 if (lock->l_blocking_ast != ldlm_blocking_ast) {
390 unlock_res_and_lock(lock);
393 RETURN(ldlm_blocking_ast_nocheck(lock));
395 EXPORT_SYMBOL(ldlm_blocking_ast);
398 * Implements ldlm_lock::l_glimpse_ast for extent locks acquired on the server.
400 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for that is
401 * rather subtle: with OST-side locking, it may so happen that _all_ extent
402 * locks are held by the OST. If client wants to obtain the current file size
403 * it calls ll_glimpse_size(), and (as all locks are held only on the server),
404 * this dummy glimpse callback fires and does nothing. The client still
405 * receives the correct file size due to the following fragment of code in
406 * ldlm_cb_interpret():
408 * if (rc == -ELDLM_NO_LOCK_DATA) {
409 * LDLM_DEBUG(lock, "lost race - client has a lock but no"
411 * ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
414 * That is, after the glimpse returns this error, ofd_lvbo_update() is called
415 * and returns the updated file attributes from the inode to the client.
417 * See also comment in ofd_intent_policy() on why servers must set a non-NULL
418 * l_glimpse_ast when grabbing DLM locks. Otherwise, the server will assume
419 * that the object is in the process of being destroyed.
421 * \param[in] lock DLM lock being glimpsed, unused
422 * \param[in] reqp pointer to ptlrpc_request, unused
424 * \retval -ELDLM_NO_LOCK_DATA to get attributes from disk object
426 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
428 return -ELDLM_NO_LOCK_DATA;
432 * Enqueue a local lock (typically on a server).
434 int ldlm_cli_enqueue_local(const struct lu_env *env,
435 struct ldlm_namespace *ns,
436 const struct ldlm_res_id *res_id,
437 enum ldlm_type type, union ldlm_policy_data *policy,
438 enum ldlm_mode mode, __u64 *flags,
439 ldlm_blocking_callback blocking,
440 ldlm_completion_callback completion,
441 ldlm_glimpse_callback glimpse,
442 void *data, __u32 lvb_len, enum lvb_type lvb_type,
443 const __u64 *client_cookie,
444 struct lustre_handle *lockh)
446 struct ldlm_lock *lock;
448 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
449 .lcs_blocking = blocking,
450 .lcs_glimpse = glimpse,
454 LASSERT(!(*flags & LDLM_FL_REPLAY));
455 if (unlikely(ns_is_client(ns))) {
456 CERROR("Trying to enqueue local lock in a shadow namespace\n");
460 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
463 GOTO(out_nolock, err = PTR_ERR(lock));
465 err = ldlm_lvbo_init(env, lock->l_resource);
467 LDLM_ERROR(lock, "delayed lvb init failed (rc %d)", err);
468 ldlm_lock_destroy_nolock(lock);
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(env, 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 const struct lu_env *env = NULL;
570 int is_replay = *flags & LDLM_FL_REPLAY;
571 struct ldlm_lock *lock;
572 struct ldlm_reply *reply;
573 int cleanup_phase = 1;
576 if (req && req->rq_svc_thread)
577 env = req->rq_svc_thread->t_env;
579 lock = ldlm_handle2lock(lockh);
580 /* ldlm_cli_enqueue is holding a reference on this lock. */
582 LASSERT(type == LDLM_FLOCK);
586 LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
587 "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
589 if (rc != ELDLM_OK) {
591 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
592 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
594 if (rc != ELDLM_LOCK_ABORTED)
598 /* Before we return, swab the reply */
599 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
601 GOTO(cleanup, rc = -EPROTO);
606 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
609 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
610 GOTO(cleanup, rc = size);
611 } else if (unlikely(size > lvb_len)) {
612 LDLM_ERROR(lock, "Replied LVB is larger than "
613 "expectation, expected = %d, replied = %d",
615 GOTO(cleanup, rc = -EINVAL);
620 if (rc == ELDLM_LOCK_ABORTED) {
621 if (lvb_len > 0 && lvb != NULL)
622 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
624 GOTO(cleanup, rc = rc ? : ELDLM_LOCK_ABORTED);
627 /* lock enqueued on the server */
630 lock_res_and_lock(lock);
631 /* Key change rehash lock in per-export hash with new key */
632 if (exp->exp_lock_hash) {
633 /* In the function below, .hs_keycmp resolves to
634 * ldlm_export_lock_keycmp() */
635 /* coverity[overrun-buffer-val] */
636 cfs_hash_rehash_key(exp->exp_lock_hash,
637 &lock->l_remote_handle,
641 lock->l_remote_handle = reply->lock_handle;
644 *flags = ldlm_flags_from_wire(reply->lock_flags);
645 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
646 LDLM_FL_INHERIT_MASK);
647 unlock_res_and_lock(lock);
649 CDEBUG(D_INFO, "local: %p, remote cookie: %#llx, flags: %#llx\n",
650 lock, reply->lock_handle.cookie, *flags);
652 /* If enqueue returned a blocked lock but the completion handler has
653 * already run, then it fixed up the resource and we don't need to do it
655 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
656 int newmode = reply->lock_desc.l_req_mode;
658 if (newmode && newmode != lock->l_req_mode) {
659 LDLM_DEBUG(lock, "server returned different mode %s",
660 ldlm_lockname[newmode]);
661 lock->l_req_mode = newmode;
664 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
665 &lock->l_resource->lr_name)) {
666 CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES
667 " instead of "DLDLMRES"\n",
668 PLDLMRES(&reply->lock_desc.l_resource),
669 PLDLMRES(lock->l_resource));
671 rc = ldlm_lock_change_resource(ns, lock,
672 &reply->lock_desc.l_resource.lr_name);
673 if (rc || lock->l_resource == NULL)
674 GOTO(cleanup, rc = -ENOMEM);
675 LDLM_DEBUG(lock, "client-side enqueue, new resource");
679 /* We assume lock type cannot change on server*/
680 ldlm_convert_policy_to_local(exp,
681 lock->l_resource->lr_type,
682 &reply->lock_desc.l_policy_data,
683 &lock->l_policy_data);
686 if (type != LDLM_PLAIN)
687 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
690 if ((*flags) & LDLM_FL_AST_SENT) {
691 lock_res_and_lock(lock);
692 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
693 unlock_res_and_lock(lock);
694 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
697 /* If the lock has already been granted by a completion AST, don't
698 * clobber the LVB with an older one. */
700 /* We must lock or a racing completion might update lvb without
701 * letting us know and we'll clobber the correct value.
702 * Cannot unlock after the check either, a that still leaves
703 * a tiny window for completion to get in */
704 lock_res_and_lock(lock);
705 if (lock->l_req_mode != lock->l_granted_mode)
706 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
707 lock->l_lvb_data, lvb_len);
708 unlock_res_and_lock(lock);
716 rc = ldlm_lock_enqueue(env, ns, &lock, NULL, flags);
717 if (lock->l_completion_ast != NULL) {
718 int err = lock->l_completion_ast(lock, *flags, NULL);
726 if (lvb_len > 0 && lvb != NULL) {
727 /* Copy the LVB here, and not earlier, because the completion
728 * AST (if any) can override what we got in the reply */
729 memcpy(lvb, lock->l_lvb_data, lvb_len);
732 LDLM_DEBUG(lock, "client-side enqueue END");
735 if (cleanup_phase == 1 && rc)
736 failed_lock_cleanup(ns, lock, mode);
737 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
739 LDLM_LOCK_RELEASE(lock);
742 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
745 * Estimate number of lock handles that would fit into request of given
746 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
747 * a single page on the send/receive side. XXX: 512 should be changed to
748 * more adequate value.
750 static inline int ldlm_req_handles_avail(int req_size, int off)
754 avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512) - req_size;
755 if (likely(avail >= 0))
756 avail /= (int)sizeof(struct lustre_handle);
759 avail += LDLM_LOCKREQ_HANDLES - off;
764 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
765 enum req_location loc,
768 __u32 size = req_capsule_msg_size(pill, loc);
769 return ldlm_req_handles_avail(size, off);
772 static inline int ldlm_format_handles_avail(struct obd_import *imp,
773 const struct req_format *fmt,
774 enum req_location loc, int off)
776 __u32 size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
777 return ldlm_req_handles_avail(size, off);
781 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
782 * \a count locks in \a cancels.
784 * This is to be called by functions preparing their own requests that
785 * might contain lists of locks to cancel in addition to actual operation
786 * that needs to be performed.
788 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
789 int version, int opc, int canceloff,
790 struct list_head *cancels, int count)
792 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
793 struct req_capsule *pill = &req->rq_pill;
794 struct ldlm_request *dlm = NULL;
795 struct list_head head = LIST_HEAD_INIT(head);
796 enum ldlm_lru_flags lru_flags;
797 int avail, to_free, pack = 0;
803 if (ns_connect_cancelset(ns)) {
804 /* Estimate the amount of available space in the request. */
805 req_capsule_filled_sizes(pill, RCL_CLIENT);
806 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
808 lru_flags = LDLM_LRU_FLAG_NO_WAIT | (ns_connect_lru_resize(ns) ?
809 LDLM_LRU_FLAG_LRUR : LDLM_LRU_FLAG_AGED);
810 to_free = !ns_connect_lru_resize(ns) &&
811 opc == LDLM_ENQUEUE ? 1 : 0;
813 /* Cancel LRU locks here _only_ if the server supports
814 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
815 * RPC, which will make us slower. */
817 count += ldlm_cancel_lru_local(ns, cancels, to_free,
824 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
825 ldlm_request_bufsize(pack, opc));
828 rc = ptlrpc_request_pack(req, version, opc);
830 ldlm_lock_list_put(cancels, l_bl_ast, count);
834 if (ns_connect_cancelset(ns)) {
836 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
838 /* Skip first lock handler in ldlm_request_pack(),
839 * this method will increment @lock_count according
840 * to the lock handle amount actually written to
842 dlm->lock_count = canceloff;
844 /* Pack into the request @pack lock handles. */
845 ldlm_cli_cancel_list(cancels, pack, req, 0);
846 /* Prepare and send separate cancel RPC for others. */
847 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
849 ldlm_lock_list_put(cancels, l_bl_ast, count);
853 EXPORT_SYMBOL(ldlm_prep_elc_req);
855 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
856 struct list_head *cancels, int count)
858 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
859 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
861 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
863 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
865 struct ptlrpc_request *req;
869 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
871 RETURN(ERR_PTR(-ENOMEM));
873 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
875 ptlrpc_request_free(req);
879 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
880 ptlrpc_request_set_replen(req);
883 EXPORT_SYMBOL(ldlm_enqueue_pack);
886 * Client-side lock enqueue.
888 * If a request has some specific initialisation it is passed in \a reqp,
889 * otherwise it is created in ldlm_cli_enqueue.
891 * Supports sync and async requests, pass \a async flag accordingly. If a
892 * request was created in ldlm_cli_enqueue and it is the async request,
893 * pass it to the caller in \a reqp.
895 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
896 struct ldlm_enqueue_info *einfo,
897 const struct ldlm_res_id *res_id,
898 union ldlm_policy_data const *policy, __u64 *flags,
899 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
900 struct lustre_handle *lockh, int async)
902 struct ldlm_namespace *ns;
903 struct ldlm_lock *lock;
904 struct ldlm_request *body;
905 int is_replay = *flags & LDLM_FL_REPLAY;
906 int req_passed_in = 1;
908 struct ptlrpc_request *req;
911 LASSERT(exp != NULL);
913 ns = exp->exp_obd->obd_namespace;
915 /* If we're replaying this lock, just check some invariants.
916 * If we're creating a new lock, get everything all setup nice. */
918 lock = ldlm_handle2lock_long(lockh, 0);
919 LASSERT(lock != NULL);
920 LDLM_DEBUG(lock, "client-side enqueue START");
921 LASSERT(exp == lock->l_conn_export);
923 const struct ldlm_callback_suite cbs = {
924 .lcs_completion = einfo->ei_cb_cp,
925 .lcs_blocking = einfo->ei_cb_bl,
926 .lcs_glimpse = einfo->ei_cb_gl
928 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
929 einfo->ei_mode, &cbs, einfo->ei_cbdata,
932 RETURN(PTR_ERR(lock));
933 /* for the local lock, add the reference */
934 ldlm_lock_addref_internal(lock, einfo->ei_mode);
935 ldlm_lock2handle(lock, lockh);
937 lock->l_policy_data = *policy;
939 if (einfo->ei_type == LDLM_EXTENT) {
940 /* extent lock without policy is a bug */
944 lock->l_req_extent = policy->l_extent;
946 LDLM_DEBUG(lock, "client-side enqueue START, flags %#llx",
950 lock->l_conn_export = exp;
951 lock->l_export = NULL;
952 lock->l_blocking_ast = einfo->ei_cb_bl;
953 lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
954 lock->l_activity = ktime_get_real_seconds();
956 /* lock not sent to server yet */
957 if (reqp == NULL || *reqp == NULL) {
958 req = ldlm_enqueue_pack(exp, lvb_len);
960 failed_lock_cleanup(ns, lock, einfo->ei_mode);
961 LDLM_LOCK_RELEASE(lock);
962 RETURN(PTR_ERR(req));
972 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
974 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
975 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
978 if (*flags & LDLM_FL_NDELAY) {
979 DEBUG_REQ(D_DLMTRACE, req, "enque lock with no delay\n");
980 req->rq_no_resend = req->rq_no_delay = 1;
981 /* probably set a shorter timeout value and handle ETIMEDOUT
982 * in osc_lock_upcall() correctly */
983 /* lustre_msg_set_timeout(req, req->rq_timeout / 2); */
986 /* Dump lock data into the request buffer */
987 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
988 ldlm_lock2desc(lock, &body->lock_desc);
989 body->lock_flags = ldlm_flags_to_wire(*flags);
990 body->lock_handle[0] = *lockh;
992 /* extended LDLM opcodes in client stats */
993 if (exp->exp_obd->obd_svc_stats != NULL) {
994 bool glimpse = *flags & LDLM_FL_HAS_INTENT;
996 /* OST glimpse has no intent buffer */
997 if (req_capsule_has_field(&req->rq_pill, &RMF_LDLM_INTENT,
999 struct ldlm_intent *it;
1001 it = req_capsule_client_get(&req->rq_pill,
1003 glimpse = (it && (it->opc == IT_GLIMPSE));
1007 ldlm_svc_get_eopc(body, exp->exp_obd->obd_svc_stats);
1009 lprocfs_counter_incr(exp->exp_obd->obd_svc_stats,
1011 LDLM_GLIMPSE_ENQUEUE);
1015 LASSERT(reqp != NULL);
1019 LDLM_DEBUG(lock, "sending request");
1021 rc = ptlrpc_queue_wait(req);
1023 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
1024 einfo->ei_mode, flags, lvb, lvb_len,
1027 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
1028 * one reference that we took */
1030 LDLM_LOCK_RELEASE(lock);
1034 if (!req_passed_in && req != NULL) {
1035 ptlrpc_req_finished(req);
1042 EXPORT_SYMBOL(ldlm_cli_enqueue);
1045 * Client-side lock convert reply handling.
1047 * Finish client lock converting, checks for concurrent converts
1048 * and clear 'converting' flag so lock can be placed back into LRU.
1050 static int lock_convert_interpret(const struct lu_env *env,
1051 struct ptlrpc_request *req,
1054 struct ldlm_async_args *aa = args;
1055 struct ldlm_lock *lock;
1056 struct ldlm_reply *reply;
1060 lock = ldlm_handle2lock(&aa->lock_handle);
1062 LDLM_DEBUG_NOLOCK("convert ACK for unknown local cookie %#llx",
1063 aa->lock_handle.cookie);
1067 LDLM_DEBUG(lock, "CONVERTED lock:");
1072 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1074 GOTO(out, rc = -EPROTO);
1076 if (reply->lock_handle.cookie != aa->lock_handle.cookie) {
1077 LDLM_ERROR(lock, "convert ACK with wrong lock cookie %#llx"
1078 " but cookie %#llx from server %s id %s\n",
1079 aa->lock_handle.cookie, reply->lock_handle.cookie,
1080 req->rq_export->exp_client_uuid.uuid,
1081 libcfs_id2str(req->rq_peer));
1082 GOTO(out, rc = ELDLM_NO_LOCK_DATA);
1085 lock_res_and_lock(lock);
1086 /* Lock convert is sent for any new bits to drop, the converting flag
1087 * is dropped when ibits on server are the same as on client. Meanwhile
1088 * that can be so that more later convert will be replied first with
1089 * and clear converting flag, so in case of such race just exit here.
1090 * if lock has no converting bits then */
1091 if (!ldlm_is_converting(lock)) {
1092 LDLM_DEBUG(lock, "convert ACK for lock without converting flag,"
1093 " reply ibits %#llx",
1094 reply->lock_desc.l_policy_data.l_inodebits.bits);
1095 } else if (reply->lock_desc.l_policy_data.l_inodebits.bits !=
1096 lock->l_policy_data.l_inodebits.bits) {
1097 /* Compare server returned lock ibits and local lock ibits
1098 * if they are the same we consider convertion is done,
1099 * otherwise we have more converts inflight and keep
1102 LDLM_DEBUG(lock, "convert ACK with ibits %#llx\n",
1103 reply->lock_desc.l_policy_data.l_inodebits.bits);
1105 ldlm_clear_converting(lock);
1107 /* Concurrent BL AST may arrive and cause another convert
1108 * or cancel so just do nothing here if bl_ast is set,
1109 * finish with convert otherwise.
1111 if (!ldlm_is_bl_ast(lock)) {
1112 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
1114 /* Drop cancel_bits since there are no more converts
1115 * and put lock into LRU if it is still not used and
1118 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1119 if (!lock->l_readers && !lock->l_writers &&
1120 !ldlm_is_canceling(lock)) {
1121 spin_lock(&ns->ns_lock);
1122 /* there is check for list_empty() inside */
1123 ldlm_lock_remove_from_lru_nolock(lock);
1124 ldlm_lock_add_to_lru_nolock(lock);
1125 spin_unlock(&ns->ns_lock);
1129 unlock_res_and_lock(lock);
1134 lock_res_and_lock(lock);
1135 if (ldlm_is_converting(lock)) {
1136 ldlm_clear_converting(lock);
1137 ldlm_set_cbpending(lock);
1138 ldlm_set_bl_ast(lock);
1139 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1141 unlock_res_and_lock(lock);
1143 /* fallback to normal lock cancel. If rc means there is no
1144 * valid lock on server, do only local cancel */
1145 if (rc == ELDLM_NO_LOCK_DATA)
1150 rc = ldlm_cli_cancel(&aa->lock_handle, flag);
1152 LDLM_DEBUG(lock, "failed to cancel lock: rc = %d\n",
1155 LDLM_LOCK_PUT(lock);
1160 * Client-side IBITS lock convert.
1162 * Inform server that lock has been converted instead of canceling.
1163 * Server finishes convert on own side and does reprocess to grant
1164 * all related waiting locks.
1166 * Since convert means only ibits downgrading, client doesn't need to
1167 * wait for server reply to finish local converting process so this request
1168 * is made asynchronous.
1171 int ldlm_cli_convert(struct ldlm_lock *lock, __u32 *flags)
1173 struct ldlm_request *body;
1174 struct ptlrpc_request *req;
1175 struct ldlm_async_args *aa;
1176 struct obd_export *exp = lock->l_conn_export;
1181 LDLM_ERROR(lock, "convert must not be called on local locks.");
1185 /* this is better to check earlier and it is done so already,
1186 * but this check is kept too as final one to issue an error
1187 * if any new code will miss such check.
1189 if (!exp_connect_lock_convert(exp)) {
1190 LDLM_ERROR(lock, "server doesn't support lock convert\n");
1194 if (lock->l_resource->lr_type != LDLM_IBITS) {
1195 LDLM_ERROR(lock, "convert works with IBITS locks only.");
1199 LDLM_DEBUG(lock, "client-side convert");
1201 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1202 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1207 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1208 body->lock_handle[0] = lock->l_remote_handle;
1210 body->lock_desc.l_req_mode = lock->l_req_mode;
1211 body->lock_desc.l_granted_mode = lock->l_granted_mode;
1213 body->lock_desc.l_policy_data.l_inodebits.bits =
1214 lock->l_policy_data.l_inodebits.bits;
1215 body->lock_desc.l_policy_data.l_inodebits.cancel_bits = 0;
1217 body->lock_flags = ldlm_flags_to_wire(*flags);
1218 body->lock_count = 1;
1220 ptlrpc_request_set_replen(req);
1223 * Use cancel portals for convert as well as high-priority handling.
1225 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1226 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1228 ptlrpc_at_set_req_timeout(req);
1230 if (exp->exp_obd->obd_svc_stats != NULL)
1231 lprocfs_counter_incr(exp->exp_obd->obd_svc_stats,
1232 LDLM_CONVERT - LDLM_FIRST_OPC);
1234 aa = ptlrpc_req_async_args(req);
1235 ldlm_lock2handle(lock, &aa->lock_handle);
1236 req->rq_interpret_reply = lock_convert_interpret;
1238 ptlrpcd_add_req(req);
1243 * Cancel locks locally.
1245 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1246 * \retval LDLM_FL_CANCELING otherwise;
1247 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1249 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1251 __u64 rc = LDLM_FL_LOCAL_ONLY;
1254 if (lock->l_conn_export) {
1257 LDLM_DEBUG(lock, "client-side cancel");
1258 /* Set this flag to prevent others from getting new references*/
1259 lock_res_and_lock(lock);
1260 ldlm_set_cbpending(lock);
1261 local_only = !!(lock->l_flags &
1262 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1263 ldlm_cancel_callback(lock);
1264 rc = (ldlm_is_bl_ast(lock)) ?
1265 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1266 unlock_res_and_lock(lock);
1269 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1271 rc = LDLM_FL_LOCAL_ONLY;
1273 ldlm_lock_cancel(lock);
1275 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1276 LDLM_ERROR(lock, "Trying to cancel local lock");
1279 LDLM_DEBUG(lock, "server-side local cancel");
1280 ldlm_lock_cancel(lock);
1281 ldlm_reprocess_all(lock->l_resource);
1288 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1290 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1291 struct list_head *head, int count)
1293 struct ldlm_request *dlm;
1294 struct ldlm_lock *lock;
1295 int max, packed = 0;
1298 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1299 LASSERT(dlm != NULL);
1301 /* Check the room in the request buffer. */
1302 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1303 sizeof(struct ldlm_request);
1304 max /= sizeof(struct lustre_handle);
1305 max += LDLM_LOCKREQ_HANDLES;
1306 LASSERT(max >= dlm->lock_count + count);
1308 /* XXX: it would be better to pack lock handles grouped by resource.
1309 * so that the server cancel would call filter_lvbo_update() less
1311 list_for_each_entry(lock, head, l_bl_ast) {
1314 LASSERT(lock->l_conn_export);
1315 /* Pack the lock handle to the given request buffer. */
1316 LDLM_DEBUG(lock, "packing");
1317 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1320 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1325 * Prepare and send a batched cancel RPC. It will include \a count lock
1326 * handles of locks given in \a cancels list. */
1327 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1328 int count, enum ldlm_cancel_flags flags)
1330 struct ptlrpc_request *req = NULL;
1331 struct obd_import *imp;
1336 LASSERT(exp != NULL);
1339 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1341 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1344 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1345 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1350 imp = class_exp2cliimp(exp);
1351 if (imp == NULL || imp->imp_invalid) {
1353 "skipping cancel on invalid import %p\n", imp);
1357 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1359 GOTO(out, rc = -ENOMEM);
1361 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1362 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1363 ldlm_request_bufsize(count, LDLM_CANCEL));
1365 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1367 ptlrpc_request_free(req);
1371 /* If OSP want cancel cross-MDT lock, let's not block it in
1372 * in recovery, otherwise the lock will not released, if
1373 * the remote target is also in recovery, and it also need
1374 * this lock, it might cause deadlock. */
1375 if (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS &&
1376 exp->exp_obd->obd_lu_dev != NULL &&
1377 exp->exp_obd->obd_lu_dev->ld_site != NULL) {
1378 struct lu_device *top_dev;
1380 top_dev = exp->exp_obd->obd_lu_dev->ld_site->ls_top_dev;
1381 if (top_dev != NULL &&
1382 top_dev->ld_obd->obd_recovering)
1383 req->rq_allow_replay = 1;
1386 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1387 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1388 ptlrpc_at_set_req_timeout(req);
1390 ldlm_cancel_pack(req, cancels, count);
1392 ptlrpc_request_set_replen(req);
1393 if (flags & LCF_ASYNC) {
1394 ptlrpcd_add_req(req);
1399 rc = ptlrpc_queue_wait(req);
1400 if (rc == LUSTRE_ESTALE) {
1401 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1402 "out of sync -- not fatal\n",
1403 libcfs_nid2str(req->rq_import->
1404 imp_connection->c_peer.nid));
1406 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1407 req->rq_import_generation == imp->imp_generation) {
1408 ptlrpc_req_finished(req);
1410 } else if (rc != ELDLM_OK) {
1411 /* -ESHUTDOWN is common on umount */
1412 CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1413 "Got rc %d from cancel RPC: "
1414 "canceling anyway\n", rc);
1421 ptlrpc_req_finished(req);
1424 return sent ? sent : rc;
1427 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1429 LASSERT(imp != NULL);
1430 return &imp->imp_obd->obd_namespace->ns_pool;
1434 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1436 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1438 struct obd_device *obd;
1442 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1443 !imp_connect_lru_resize(req->rq_import)))
1446 * Do nothing for corner cases.
1451 /* In some cases RPC may contain SLV and limit zeroed out. This
1452 * is the case when server does not support LRU resize feature.
1453 * This is also possible in some recovery cases when server-side
1454 * reqs have no reference to the OBD export and thus access to
1455 * server-side namespace is not possible. */
1456 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1457 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1458 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1459 "(SLV: %llu, Limit: %u)",
1460 lustre_msg_get_slv(req->rq_repmsg),
1461 lustre_msg_get_limit(req->rq_repmsg));
1465 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1466 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1467 obd = req->rq_import->imp_obd;
1469 /* Set new SLV and limit in OBD fields to make them accessible
1470 * to the pool thread. We do not access obd_namespace and pool
1471 * directly here as there is no reliable way to make sure that
1472 * they are still alive at cleanup time. Evil races are possible
1473 * which may cause Oops at that time. */
1474 write_lock(&obd->obd_pool_lock);
1475 obd->obd_pool_slv = new_slv;
1476 obd->obd_pool_limit = new_limit;
1477 write_unlock(&obd->obd_pool_lock);
1483 * Client side lock cancel.
1485 * Lock must not have any readers or writers by this time.
1487 int ldlm_cli_cancel(const struct lustre_handle *lockh,
1488 enum ldlm_cancel_flags cancel_flags)
1490 struct obd_export *exp;
1491 enum ldlm_lru_flags lru_flags;
1492 int avail, count = 1;
1494 struct ldlm_namespace *ns;
1495 struct ldlm_lock *lock;
1496 struct list_head cancels = LIST_HEAD_INIT(cancels);
1500 lock = ldlm_handle2lock_long(lockh, 0);
1502 LDLM_DEBUG_NOLOCK("lock is already being destroyed");
1506 /* Convert lock bits instead of cancel for IBITS locks */
1507 if (cancel_flags & LCF_CONVERT) {
1508 LASSERT(lock->l_resource->lr_type == LDLM_IBITS);
1509 LASSERT(lock->l_policy_data.l_inodebits.cancel_bits != 0);
1511 rc = ldlm_cli_dropbits(lock,
1512 lock->l_policy_data.l_inodebits.cancel_bits);
1514 LDLM_LOCK_RELEASE(lock);
1519 lock_res_and_lock(lock);
1520 /* Lock is being canceled and the caller doesn't want to wait */
1521 if (ldlm_is_canceling(lock)) {
1522 if (cancel_flags & LCF_ASYNC) {
1523 unlock_res_and_lock(lock);
1525 struct l_wait_info lwi = { 0 };
1527 unlock_res_and_lock(lock);
1528 l_wait_event(lock->l_waitq, is_bl_done(lock), &lwi);
1530 LDLM_LOCK_RELEASE(lock);
1534 /* Lock is being converted, cancel it immediately.
1535 * When convert will end, it releases lock and it will be gone.
1537 if (ldlm_is_converting(lock)) {
1538 /* set back flags removed by convert */
1539 ldlm_set_cbpending(lock);
1540 ldlm_set_bl_ast(lock);
1543 ldlm_set_canceling(lock);
1544 unlock_res_and_lock(lock);
1546 rc = ldlm_cli_cancel_local(lock);
1547 if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
1548 LDLM_LOCK_RELEASE(lock);
1551 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1552 * RPC which goes to canceld portal, so we can cancel other LRU locks
1553 * here and send them all as one LDLM_CANCEL RPC. */
1554 LASSERT(list_empty(&lock->l_bl_ast));
1555 list_add(&lock->l_bl_ast, &cancels);
1557 exp = lock->l_conn_export;
1558 if (exp_connect_cancelset(exp)) {
1559 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1564 ns = ldlm_lock_to_ns(lock);
1565 lru_flags = ns_connect_lru_resize(ns) ?
1566 LDLM_LRU_FLAG_LRUR : LDLM_LRU_FLAG_AGED;
1567 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1568 LCF_BL_AST, lru_flags);
1570 ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1573 EXPORT_SYMBOL(ldlm_cli_cancel);
1576 * Locally cancel up to \a count locks in list \a cancels.
1577 * Return the number of cancelled locks.
1579 int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
1580 enum ldlm_cancel_flags cancel_flags)
1582 struct list_head head = LIST_HEAD_INIT(head);
1583 struct ldlm_lock *lock, *next;
1584 int left = 0, bl_ast = 0;
1588 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1592 if (cancel_flags & LCF_LOCAL) {
1593 rc = LDLM_FL_LOCAL_ONLY;
1594 ldlm_lock_cancel(lock);
1596 rc = ldlm_cli_cancel_local(lock);
1598 /* Until we have compound requests and can send LDLM_CANCEL
1599 * requests batched with generic RPCs, we need to send cancels
1600 * with the LDLM_FL_BL_AST flag in a separate RPC from
1601 * the one being generated now. */
1602 if (!(cancel_flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1603 LDLM_DEBUG(lock, "Cancel lock separately");
1604 list_del_init(&lock->l_bl_ast);
1605 list_add(&lock->l_bl_ast, &head);
1609 if (rc == LDLM_FL_LOCAL_ONLY) {
1610 /* CANCEL RPC should not be sent to server. */
1611 list_del_init(&lock->l_bl_ast);
1612 LDLM_LOCK_RELEASE(lock);
1618 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1625 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1626 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1627 * readahead requests, ...)
1629 static enum ldlm_policy_res
1630 ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1631 int unused, int added, int count)
1633 enum ldlm_policy_res result = LDLM_POLICY_CANCEL_LOCK;
1635 /* don't check added & count since we want to process all locks
1637 * It's fine to not take lock to access lock->l_resource since
1638 * the lock has already been granted so it won't change. */
1639 switch (lock->l_resource->lr_type) {
1642 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) != 0)
1645 result = LDLM_POLICY_SKIP_LOCK;
1653 * Callback function for LRU-resize policy. Decides whether to keep
1654 * \a lock in LRU for current \a LRU size \a unused, added in current
1655 * scan \a added and number of locks to be preferably canceled \a count.
1657 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1659 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1661 static enum ldlm_policy_res ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1662 struct ldlm_lock *lock,
1663 int unused, int added,
1666 ktime_t cur = ktime_get();
1667 struct ldlm_pool *pl = &ns->ns_pool;
1671 /* Stop LRU processing when we reach past @count or have checked all
1673 if (count && added >= count)
1674 return LDLM_POLICY_KEEP_LOCK;
1676 /* Despite of the LV, It doesn't make sense to keep the lock which
1677 * is unused for ns_max_age time.
1679 if (ktime_after(ktime_get(),
1680 ktime_add(lock->l_last_used, ns->ns_max_age)))
1681 return LDLM_POLICY_CANCEL_LOCK;
1683 slv = ldlm_pool_get_slv(pl);
1684 lvf = ldlm_pool_get_lvf(pl);
1685 la = div_u64(ktime_to_ns(ktime_sub(cur, lock->l_last_used)),
1687 lv = lvf * la * unused;
1689 /* Inform pool about current CLV to see it via debugfs. */
1690 ldlm_pool_set_clv(pl, lv);
1692 /* Stop when SLV is not yet come from server or lv is smaller than
1694 if (slv == 0 || lv < slv)
1695 return LDLM_POLICY_KEEP_LOCK;
1697 return LDLM_POLICY_CANCEL_LOCK;
1700 static enum ldlm_policy_res
1701 ldlm_cancel_lrur_no_wait_policy(struct ldlm_namespace *ns,
1702 struct ldlm_lock *lock,
1703 int unused, int added,
1706 enum ldlm_policy_res result;
1708 result = ldlm_cancel_lrur_policy(ns, lock, unused, added, count);
1709 if (result == LDLM_POLICY_KEEP_LOCK)
1712 return ldlm_cancel_no_wait_policy(ns, lock, unused, added, count);
1716 * Callback function for debugfs used policy. Makes decision whether to keep
1717 * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1718 * added and number of locks to be preferably canceled \a count.
1720 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1722 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1724 static enum ldlm_policy_res ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1725 struct ldlm_lock *lock,
1726 int unused, int added,
1729 /* Stop LRU processing when we reach past @count or have checked all
1731 return (added >= count) ?
1732 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1736 * Callback function for aged policy. Makes decision whether to keep \a lock in
1737 * LRU for current LRU size \a unused, added in current scan \a added and
1738 * number of locks to be preferably canceled \a count.
1740 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1742 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1744 static enum ldlm_policy_res ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1745 struct ldlm_lock *lock,
1746 int unused, int added,
1749 if ((added >= count) &&
1750 ktime_before(ktime_get(),
1751 ktime_add(lock->l_last_used, ns->ns_max_age)))
1752 return LDLM_POLICY_KEEP_LOCK;
1754 return LDLM_POLICY_CANCEL_LOCK;
1757 static enum ldlm_policy_res
1758 ldlm_cancel_aged_no_wait_policy(struct ldlm_namespace *ns,
1759 struct ldlm_lock *lock,
1760 int unused, int added, int count)
1762 enum ldlm_policy_res result;
1764 result = ldlm_cancel_aged_policy(ns, lock, unused, added, count);
1765 if (result == LDLM_POLICY_KEEP_LOCK)
1768 return ldlm_cancel_no_wait_policy(ns, lock, unused, added, count);
1772 * Callback function for default policy. Makes decision whether to keep \a lock
1773 * in LRU for current LRU size \a unused, added in current scan \a added and
1774 * number of locks to be preferably canceled \a count.
1776 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1778 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1781 enum ldlm_policy_res ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1782 struct ldlm_lock *lock,
1783 int unused, int added,
1786 /* Stop LRU processing when we reach past count or have checked all
1788 return (added >= count) ?
1789 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1792 typedef enum ldlm_policy_res
1793 (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1794 int unused, int added, int count);
1796 static ldlm_cancel_lru_policy_t
1797 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, enum ldlm_lru_flags lru_flags)
1799 if (ns_connect_lru_resize(ns)) {
1800 if (lru_flags & LDLM_LRU_FLAG_SHRINK)
1801 /* We kill passed number of old locks. */
1802 return ldlm_cancel_passed_policy;
1803 if (lru_flags & LDLM_LRU_FLAG_LRUR) {
1804 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1805 return ldlm_cancel_lrur_no_wait_policy;
1807 return ldlm_cancel_lrur_policy;
1809 if (lru_flags & LDLM_LRU_FLAG_PASSED)
1810 return ldlm_cancel_passed_policy;
1812 if (lru_flags & LDLM_LRU_FLAG_AGED) {
1813 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1814 return ldlm_cancel_aged_no_wait_policy;
1816 return ldlm_cancel_aged_policy;
1819 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1820 return ldlm_cancel_no_wait_policy;
1822 return ldlm_cancel_default_policy;
1826 * - Free space in LRU for \a count new locks,
1827 * redundant unused locks are canceled locally;
1828 * - also cancel locally unused aged locks;
1829 * - do not cancel more than \a max locks;
1830 * - GET the found locks and add them into the \a cancels list.
1832 * A client lock can be added to the l_bl_ast list only when it is
1833 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1834 * CANCEL. There are the following use cases:
1835 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1836 * ldlm_cli_cancel(), which check and set this flag properly. As any
1837 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1838 * later without any special locking.
1840 * Calling policies for enabled LRU resize:
1841 * ----------------------------------------
1842 * flags & LDLM_LRU_FLAG_LRUR - use LRU resize policy (SLV from server) to
1843 * cancel not more than \a count locks;
1845 * flags & LDLM_LRU_FLAG_PASSED - cancel \a count number of old locks (located
1846 * at the beginning of LRU list);
1848 * flags & LDLM_LRU_FLAG_SHRINK - cancel not more than \a count locks according
1849 * to memory pressre policy function;
1851 * flags & LDLM_LRU_FLAG_AGED - cancel \a count locks according to "aged policy"
1853 * flags & LDLM_LRU_FLAG_NO_WAIT - cancel as many unused locks as possible
1854 * (typically before replaying locks) w/o
1855 * sending any RPCs or waiting for any
1856 * outstanding RPC to complete.
1858 * flags & LDLM_CANCEL_CLEANUP - when cancelling read locks, do not check for
1859 * other read locks covering the same pages, just
1860 * discard those pages.
1862 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns,
1863 struct list_head *cancels, int count, int max,
1864 enum ldlm_lru_flags lru_flags)
1866 ldlm_cancel_lru_policy_t pf;
1868 int no_wait = lru_flags & LDLM_LRU_FLAG_NO_WAIT;
1872 if (!ns_connect_lru_resize(ns))
1873 count += ns->ns_nr_unused - ns->ns_max_unused;
1875 pf = ldlm_cancel_lru_policy(ns, lru_flags);
1876 LASSERT(pf != NULL);
1878 /* For any flags, stop scanning if @max is reached. */
1879 while (!list_empty(&ns->ns_unused_list) && (max == 0 || added < max)) {
1880 struct ldlm_lock *lock;
1881 struct list_head *item, *next;
1882 enum ldlm_policy_res result;
1883 ktime_t last_use = ktime_set(0, 0);
1885 spin_lock(&ns->ns_lock);
1886 item = no_wait ? ns->ns_last_pos : &ns->ns_unused_list;
1887 for (item = item->next, next = item->next;
1888 item != &ns->ns_unused_list;
1889 item = next, next = item->next) {
1890 lock = list_entry(item, struct ldlm_lock, l_lru);
1892 /* No locks which got blocking requests. */
1893 LASSERT(!ldlm_is_bl_ast(lock));
1895 if (!ldlm_is_canceling(lock) &&
1896 !ldlm_is_converting(lock))
1899 /* Somebody is already doing CANCEL. No need for this
1900 * lock in LRU, do not traverse it again. */
1901 ldlm_lock_remove_from_lru_nolock(lock);
1903 if (item == &ns->ns_unused_list) {
1904 spin_unlock(&ns->ns_lock);
1908 last_use = lock->l_last_used;
1910 LDLM_LOCK_GET(lock);
1911 spin_unlock(&ns->ns_lock);
1912 lu_ref_add(&lock->l_reference, __FUNCTION__, current);
1914 /* Pass the lock through the policy filter and see if it
1915 * should stay in LRU.
1917 * Even for shrinker policy we stop scanning if
1918 * we find a lock that should stay in the cache.
1919 * We should take into account lock age anyway
1920 * as a new lock is a valuable resource even if
1921 * it has a low weight.
1923 * That is, for shrinker policy we drop only
1924 * old locks, but additionally choose them by
1925 * their weight. Big extent locks will stay in
1927 result = pf(ns, lock, ns->ns_nr_unused, added, count);
1928 if (result == LDLM_POLICY_KEEP_LOCK) {
1929 lu_ref_del(&lock->l_reference, __func__, current);
1930 LDLM_LOCK_RELEASE(lock);
1934 if (result == LDLM_POLICY_SKIP_LOCK) {
1935 lu_ref_del(&lock->l_reference, __func__, current);
1937 spin_lock(&ns->ns_lock);
1938 if (!list_empty(&lock->l_lru) &&
1939 lock->l_lru.prev == ns->ns_last_pos)
1940 ns->ns_last_pos = &lock->l_lru;
1941 spin_unlock(&ns->ns_lock);
1944 LDLM_LOCK_RELEASE(lock);
1948 lock_res_and_lock(lock);
1949 /* Check flags again under the lock. */
1950 if (ldlm_is_canceling(lock) || ldlm_is_converting(lock) ||
1951 ldlm_lock_remove_from_lru_check(lock, last_use) == 0) {
1952 /* Another thread is removing lock from LRU, or
1953 * somebody is already doing CANCEL, or there
1954 * is a blocking request which will send cancel
1955 * by itself, or the lock is no longer unused or
1956 * the lock has been used since the pf() call and
1957 * pages could be put under it. */
1958 unlock_res_and_lock(lock);
1959 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1960 LDLM_LOCK_RELEASE(lock);
1963 LASSERT(!lock->l_readers && !lock->l_writers);
1965 /* If we have chosen to cancel this lock voluntarily, we
1966 * better send cancel notification to server, so that it
1967 * frees appropriate state. This might lead to a race
1968 * where while we are doing cancel here, server is also
1969 * silently cancelling this lock. */
1970 ldlm_clear_cancel_on_block(lock);
1972 /* Setting the CBPENDING flag is a little misleading,
1973 * but prevents an important race; namely, once
1974 * CBPENDING is set, the lock can accumulate no more
1975 * readers/writers. Since readers and writers are
1976 * already zero here, ldlm_lock_decref() won't see
1977 * this flag and call l_blocking_ast */
1978 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1980 if ((lru_flags & LDLM_LRU_FLAG_CLEANUP) &&
1981 (lock->l_resource->lr_type == LDLM_EXTENT ||
1982 ldlm_has_dom(lock)) && lock->l_granted_mode == LCK_PR)
1983 ldlm_set_discard_data(lock);
1985 /* We can't re-add to l_lru as it confuses the
1986 * refcounting in ldlm_lock_remove_from_lru() if an AST
1987 * arrives after we drop lr_lock below. We use l_bl_ast
1988 * and can't use l_pending_chain as it is used both on
1989 * server and client nevertheless bug 5666 says it is
1990 * used only on server */
1991 LASSERT(list_empty(&lock->l_bl_ast));
1992 list_add(&lock->l_bl_ast, cancels);
1993 unlock_res_and_lock(lock);
1994 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
2000 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
2002 enum ldlm_cancel_flags cancel_flags,
2003 enum ldlm_lru_flags lru_flags)
2007 added = ldlm_prepare_lru_list(ns, cancels, count, max, lru_flags);
2011 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
2015 * Cancel at least \a nr locks from given namespace LRU.
2017 * When called with LCF_ASYNC the blocking callback will be handled
2018 * in a thread and this function will return after the thread has been
2019 * asked to call the callback. When called with LCF_ASYNC the blocking
2020 * callback will be performed in this function.
2022 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
2023 enum ldlm_cancel_flags cancel_flags,
2024 enum ldlm_lru_flags lru_flags)
2026 struct list_head cancels = LIST_HEAD_INIT(cancels);
2030 /* Just prepare the list of locks, do not actually cancel them yet.
2031 * Locks are cancelled later in a separate thread. */
2032 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, lru_flags);
2033 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
2041 * Find and cancel locally unused locks found on resource, matched to the
2042 * given policy, mode. GET the found locks and add them into the \a cancels
2045 int ldlm_cancel_resource_local(struct ldlm_resource *res,
2046 struct list_head *cancels,
2047 union ldlm_policy_data *policy,
2048 enum ldlm_mode mode, __u64 lock_flags,
2049 enum ldlm_cancel_flags cancel_flags,
2052 struct ldlm_lock *lock;
2058 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
2059 if (opaque != NULL && lock->l_ast_data != opaque) {
2060 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
2061 lock->l_ast_data, opaque);
2065 if (lock->l_readers || lock->l_writers)
2068 /* If somebody is already doing CANCEL, or blocking AST came,
2069 * or lock is being converted then skip this lock. */
2070 if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock) ||
2071 ldlm_is_converting(lock))
2074 if (lockmode_compat(lock->l_granted_mode, mode))
2077 /* If policy is given and this is IBITS lock, add to list only
2078 * those locks that match by policy.
2079 * Skip locks with DoM bit always to don't flush data.
2081 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
2082 (!(lock->l_policy_data.l_inodebits.bits &
2083 policy->l_inodebits.bits) || ldlm_has_dom(lock)))
2086 /* See CBPENDING comment in ldlm_cancel_lru */
2087 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
2090 LASSERT(list_empty(&lock->l_bl_ast));
2091 list_add(&lock->l_bl_ast, cancels);
2092 LDLM_LOCK_GET(lock);
2097 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
2099 EXPORT_SYMBOL(ldlm_cancel_resource_local);
2102 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
2104 * If \a req is NULL, send CANCEL request to server with handles of locks
2105 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
2106 * separately per lock.
2107 * If \a req is not NULL, put handles of locks in \a cancels into the request
2108 * buffer at the offset \a off.
2109 * Destroy \a cancels at the end.
2111 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
2112 struct ptlrpc_request *req,
2113 enum ldlm_cancel_flags flags)
2115 struct ldlm_lock *lock;
2119 if (list_empty(cancels) || count == 0)
2122 /* XXX: requests (both batched and not) could be sent in parallel.
2123 * Usually it is enough to have just 1 RPC, but it is possible that
2124 * there are too many locks to be cancelled in LRU or on a resource.
2125 * It would also speed up the case when the server does not support
2128 LASSERT(!list_empty(cancels));
2129 lock = list_entry(cancels->next, struct ldlm_lock,
2131 LASSERT(lock->l_conn_export);
2133 if (exp_connect_cancelset(lock->l_conn_export)) {
2136 ldlm_cancel_pack(req, cancels, count);
2138 res = ldlm_cli_cancel_req(lock->l_conn_export,
2142 res = ldlm_cli_cancel_req(lock->l_conn_export,
2147 CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
2148 "ldlm_cli_cancel_list: %d\n", res);
2153 ldlm_lock_list_put(cancels, l_bl_ast, res);
2155 LASSERT(count == 0);
2158 EXPORT_SYMBOL(ldlm_cli_cancel_list);
2161 * Cancel all locks on a resource that have 0 readers/writers.
2163 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
2164 * to notify the server. */
2165 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
2166 const struct ldlm_res_id *res_id,
2167 union ldlm_policy_data *policy,
2168 enum ldlm_mode mode,
2169 enum ldlm_cancel_flags flags, void *opaque)
2171 struct ldlm_resource *res;
2172 struct list_head cancels = LIST_HEAD_INIT(cancels);
2177 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2179 /* This is not a problem. */
2180 CDEBUG(D_INFO, "No resource %llu\n", res_id->name[0]);
2184 LDLM_RESOURCE_ADDREF(res);
2185 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
2186 0, flags | LCF_BL_AST, opaque);
2187 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
2189 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
2192 LDLM_RESOURCE_DELREF(res);
2193 ldlm_resource_putref(res);
2196 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2198 struct ldlm_cli_cancel_arg {
2204 ldlm_cli_hash_cancel_unused(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2205 struct hlist_node *hnode, void *arg)
2207 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2208 struct ldlm_cli_cancel_arg *lc = arg;
2210 ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
2211 NULL, LCK_MINMODE, lc->lc_flags,
2213 /* must return 0 for hash iteration */
2218 * Cancel all locks on a namespace (or a specific resource, if given)
2219 * that have 0 readers/writers.
2221 * If flags & LCF_LOCAL, throw the locks away without trying
2222 * to notify the server. */
2223 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
2224 const struct ldlm_res_id *res_id,
2225 enum ldlm_cancel_flags flags, void *opaque)
2227 struct ldlm_cli_cancel_arg arg = {
2229 .lc_opaque = opaque,
2237 if (res_id != NULL) {
2238 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
2242 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2243 ldlm_cli_hash_cancel_unused, &arg, 0);
2248 /* Lock iterators. */
2250 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
2253 struct list_head *tmp, *next;
2254 struct ldlm_lock *lock;
2255 int rc = LDLM_ITER_CONTINUE;
2260 RETURN(LDLM_ITER_CONTINUE);
2263 list_for_each_safe(tmp, next, &res->lr_granted) {
2264 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2266 if (iter(lock, closure) == LDLM_ITER_STOP)
2267 GOTO(out, rc = LDLM_ITER_STOP);
2270 list_for_each_safe(tmp, next, &res->lr_waiting) {
2271 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2273 if (iter(lock, closure) == LDLM_ITER_STOP)
2274 GOTO(out, rc = LDLM_ITER_STOP);
2281 struct iter_helper_data {
2282 ldlm_iterator_t iter;
2286 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2288 struct iter_helper_data *helper = closure;
2289 return helper->iter(lock, helper->closure);
2292 static int ldlm_res_iter_helper(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2293 struct hlist_node *hnode, void *arg)
2296 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2298 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2302 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2303 ldlm_iterator_t iter, void *closure)
2306 struct iter_helper_data helper = { .iter = iter, .closure = closure };
2308 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2309 ldlm_res_iter_helper, &helper, 0);
2313 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2314 * return 0: find no resource
2315 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2318 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2319 const struct ldlm_res_id *res_id,
2320 ldlm_iterator_t iter, void *data)
2322 struct ldlm_resource *res;
2326 LASSERTF(ns != NULL, "must pass in namespace\n");
2328 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2332 LDLM_RESOURCE_ADDREF(res);
2333 rc = ldlm_resource_foreach(res, iter, data);
2334 LDLM_RESOURCE_DELREF(res);
2335 ldlm_resource_putref(res);
2338 EXPORT_SYMBOL(ldlm_resource_iterate);
2342 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2344 struct list_head *list = closure;
2346 /* we use l_pending_chain here, because it's unused on clients. */
2347 LASSERTF(list_empty(&lock->l_pending_chain),
2348 "lock %p next %p prev %p\n",
2349 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2350 /* bug 9573: don't replay locks left after eviction, or
2351 * bug 17614: locks being actively cancelled. Get a reference
2352 * on a lock so that it does not disapear under us (e.g. due to cancel)
2354 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_BL_DONE))) {
2355 list_add(&lock->l_pending_chain, list);
2356 LDLM_LOCK_GET(lock);
2359 return LDLM_ITER_CONTINUE;
2362 static int replay_lock_interpret(const struct lu_env *env,
2363 struct ptlrpc_request *req, void *args, int rc)
2365 struct ldlm_async_args *aa = args;
2366 struct ldlm_lock *lock;
2367 struct ldlm_reply *reply;
2368 struct obd_export *exp;
2371 atomic_dec(&req->rq_import->imp_replay_inflight);
2375 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2377 GOTO(out, rc = -EPROTO);
2379 lock = ldlm_handle2lock(&aa->lock_handle);
2381 CERROR("received replay ack for unknown local cookie %#llx"
2382 " remote cookie %#llx from server %s id %s\n",
2383 aa->lock_handle.cookie, reply->lock_handle.cookie,
2384 req->rq_export->exp_client_uuid.uuid,
2385 libcfs_id2str(req->rq_peer));
2386 GOTO(out, rc = -ESTALE);
2389 /* Key change rehash lock in per-export hash with new key */
2390 exp = req->rq_export;
2391 if (exp && exp->exp_lock_hash) {
2392 /* In the function below, .hs_keycmp resolves to
2393 * ldlm_export_lock_keycmp() */
2394 /* coverity[overrun-buffer-val] */
2395 cfs_hash_rehash_key(exp->exp_lock_hash,
2396 &lock->l_remote_handle,
2397 &reply->lock_handle,
2400 lock->l_remote_handle = reply->lock_handle;
2403 LDLM_DEBUG(lock, "replayed lock:");
2404 ptlrpc_import_recovery_state_machine(req->rq_import);
2405 LDLM_LOCK_PUT(lock);
2408 ptlrpc_connect_import(req->rq_import);
2413 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2415 struct ptlrpc_request *req;
2416 struct ldlm_async_args *aa;
2417 struct ldlm_request *body;
2422 /* Bug 11974: Do not replay a lock which is actively being canceled */
2423 if (ldlm_is_bl_done(lock)) {
2424 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2428 /* If this is reply-less callback lock, we cannot replay it, since
2429 * server might have long dropped it, but notification of that event was
2430 * lost by network. (and server granted conflicting lock already) */
2431 if (ldlm_is_cancel_on_block(lock)) {
2432 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2433 ldlm_lock_cancel(lock);
2438 * If granted mode matches the requested mode, this lock is granted.
2440 * If we haven't been granted anything and are on a resource list,
2441 * then we're blocked/waiting.
2443 * If we haven't been granted anything and we're NOT on a resource list,
2444 * then we haven't got a reply yet and don't have a known disposition.
2445 * This happens whenever a lock enqueue is the request that triggers
2448 if (lock->l_granted_mode == lock->l_req_mode)
2449 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2450 else if (!list_empty(&lock->l_res_link))
2451 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2453 flags = LDLM_FL_REPLAY;
2455 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2456 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2460 /* We're part of recovery, so don't wait for it. */
2461 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2463 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2464 ldlm_lock2desc(lock, &body->lock_desc);
2465 body->lock_flags = ldlm_flags_to_wire(flags);
2467 ldlm_lock2handle(lock, &body->lock_handle[0]);
2468 if (lock->l_lvb_len > 0)
2469 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2470 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2472 ptlrpc_request_set_replen(req);
2473 /* notify the server we've replayed all requests.
2474 * also, we mark the request to be put on a dedicated
2475 * queue to be processed after all request replayes.
2477 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2479 LDLM_DEBUG(lock, "replaying lock:");
2481 atomic_inc(&req->rq_import->imp_replay_inflight);
2482 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2483 aa = ptlrpc_req_async_args(req);
2484 aa->lock_handle = body->lock_handle[0];
2485 req->rq_interpret_reply = replay_lock_interpret;
2486 ptlrpcd_add_req(req);
2492 * Cancel as many unused locks as possible before replay. since we are
2493 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2496 * Called only in recovery before replaying locks. there is no need to
2497 * replay locks that are unused. since the clients may hold thousands of
2498 * cached unused locks, dropping the unused locks can greatly reduce the
2499 * load on the servers at recovery time.
2501 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2504 struct list_head cancels = LIST_HEAD_INIT(cancels);
2506 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2507 "replay for namespace %s (%d)\n",
2508 ldlm_ns_name(ns), ns->ns_nr_unused);
2510 /* We don't need to care whether or not LRU resize is enabled
2511 * because the LDLM_LRU_FLAG_NO_WAIT policy doesn't use the
2512 * count parameter */
2513 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2514 LCF_LOCAL, LDLM_LRU_FLAG_NO_WAIT);
2516 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2517 canceled, ldlm_ns_name(ns));
2520 int ldlm_replay_locks(struct obd_import *imp)
2522 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2523 struct list_head list = LIST_HEAD_INIT(list);
2524 struct ldlm_lock *lock, *next;
2529 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2531 /* don't replay locks if import failed recovery */
2532 if (imp->imp_vbr_failed)
2535 /* ensure this doesn't fall to 0 before all have been queued */
2536 atomic_inc(&imp->imp_replay_inflight);
2538 if (ldlm_cancel_unused_locks_before_replay)
2539 ldlm_cancel_unused_locks_for_replay(ns);
2541 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2543 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2544 list_del_init(&lock->l_pending_chain);
2546 LDLM_LOCK_RELEASE(lock);
2547 continue; /* or try to do the rest? */
2549 rc = replay_one_lock(imp, lock);
2550 LDLM_LOCK_RELEASE(lock);
2553 atomic_dec(&imp->imp_replay_inflight);