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);
481 if (*flags & LDLM_FL_CANCEL_ON_BLOCK)
482 ldlm_set_cancel_on_block(lock);
485 lock->l_policy_data = *policy;
486 if (client_cookie != NULL)
487 lock->l_client_cookie = *client_cookie;
488 if (type == LDLM_EXTENT) {
489 /* extent lock without policy is a bug */
493 lock->l_req_extent = policy->l_extent;
496 err = ldlm_lock_enqueue(env, ns, &lock, policy, flags);
497 if (unlikely(err != ELDLM_OK))
501 *policy = lock->l_policy_data;
503 if (lock->l_completion_ast)
504 lock->l_completion_ast(lock, *flags, NULL);
506 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
509 LDLM_LOCK_RELEASE(lock);
513 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
515 static void failed_lock_cleanup(struct ldlm_namespace *ns,
516 struct ldlm_lock *lock, int mode)
520 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
521 lock_res_and_lock(lock);
522 /* Check that lock is not granted or failed, we might race. */
523 if (!ldlm_is_granted(lock) && !ldlm_is_failed(lock)) {
524 /* Make sure that this lock will not be found by raced
525 * bl_ast and -EINVAL reply is sent to server anyways.
527 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
528 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
531 unlock_res_and_lock(lock);
535 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
536 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
538 LDLM_DEBUG(lock, "lock was granted or failed in race");
540 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
541 * from llite/file.c/ll_file_flock(). */
542 /* This code makes for the fact that we do not have blocking handler on
543 * a client for flock locks. As such this is the place where we must
544 * completely kill failed locks. (interrupted and those that
545 * were waiting to be granted when server evicted us. */
546 if (lock->l_resource->lr_type == LDLM_FLOCK) {
547 lock_res_and_lock(lock);
548 if (!ldlm_is_destroyed(lock)) {
549 ldlm_resource_unlink_lock(lock);
550 ldlm_lock_decref_internal_nolock(lock, mode);
551 ldlm_lock_destroy_nolock(lock);
553 unlock_res_and_lock(lock);
555 ldlm_lock_decref_internal(lock, mode);
560 * Finishing portion of client lock enqueue code.
562 * Called after receiving reply from server.
564 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
565 enum ldlm_type type, __u8 with_policy,
566 enum ldlm_mode mode, __u64 *flags, void *lvb,
567 __u32 lvb_len, const struct lustre_handle *lockh,
570 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
571 const struct lu_env *env = NULL;
572 int is_replay = *flags & LDLM_FL_REPLAY;
573 struct ldlm_lock *lock;
574 struct ldlm_reply *reply;
575 int cleanup_phase = 1;
578 if (req && req->rq_svc_thread)
579 env = req->rq_svc_thread->t_env;
581 lock = ldlm_handle2lock(lockh);
582 /* ldlm_cli_enqueue is holding a reference on this lock. */
584 LASSERT(type == LDLM_FLOCK);
588 LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
589 "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
591 if (rc != ELDLM_OK) {
593 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
594 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
596 if (rc != ELDLM_LOCK_ABORTED)
600 /* Before we return, swab the reply */
601 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
603 GOTO(cleanup, rc = -EPROTO);
608 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
611 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
612 GOTO(cleanup, rc = size);
613 } else if (unlikely(size > lvb_len)) {
614 LDLM_ERROR(lock, "Replied LVB is larger than "
615 "expectation, expected = %d, replied = %d",
617 GOTO(cleanup, rc = -EINVAL);
622 if (rc == ELDLM_LOCK_ABORTED) {
623 if (lvb_len > 0 && lvb != NULL)
624 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
626 GOTO(cleanup, rc = rc ? : ELDLM_LOCK_ABORTED);
629 /* lock enqueued on the server */
632 lock_res_and_lock(lock);
633 /* Key change rehash lock in per-export hash with new key */
634 if (exp->exp_lock_hash) {
635 /* In the function below, .hs_keycmp resolves to
636 * ldlm_export_lock_keycmp() */
637 /* coverity[overrun-buffer-val] */
638 cfs_hash_rehash_key(exp->exp_lock_hash,
639 &lock->l_remote_handle,
643 lock->l_remote_handle = reply->lock_handle;
646 *flags = ldlm_flags_from_wire(reply->lock_flags);
647 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
648 LDLM_FL_INHERIT_MASK);
649 unlock_res_and_lock(lock);
651 CDEBUG(D_INFO, "local: %p, remote cookie: %#llx, flags: %#llx\n",
652 lock, reply->lock_handle.cookie, *flags);
654 /* If enqueue returned a blocked lock but the completion handler has
655 * already run, then it fixed up the resource and we don't need to do it
657 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
658 int newmode = reply->lock_desc.l_req_mode;
660 if (newmode && newmode != lock->l_req_mode) {
661 LDLM_DEBUG(lock, "server returned different mode %s",
662 ldlm_lockname[newmode]);
663 lock->l_req_mode = newmode;
666 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
667 &lock->l_resource->lr_name)) {
668 CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES
669 " instead of "DLDLMRES"\n",
670 PLDLMRES(&reply->lock_desc.l_resource),
671 PLDLMRES(lock->l_resource));
673 rc = ldlm_lock_change_resource(ns, lock,
674 &reply->lock_desc.l_resource.lr_name);
675 if (rc || lock->l_resource == NULL)
676 GOTO(cleanup, rc = -ENOMEM);
677 LDLM_DEBUG(lock, "client-side enqueue, new resource");
681 /* We assume lock type cannot change on server*/
682 ldlm_convert_policy_to_local(exp,
683 lock->l_resource->lr_type,
684 &reply->lock_desc.l_policy_data,
685 &lock->l_policy_data);
688 if (type != LDLM_PLAIN)
689 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
692 if ((*flags) & LDLM_FL_AST_SENT) {
693 lock_res_and_lock(lock);
694 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
695 unlock_res_and_lock(lock);
696 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
699 /* If the lock has already been granted by a completion AST, don't
700 * clobber the LVB with an older one. */
702 /* We must lock or a racing completion might update lvb without
703 * letting us know and we'll clobber the correct value.
704 * Cannot unlock after the check either, a that still leaves
705 * a tiny window for completion to get in */
706 lock_res_and_lock(lock);
707 if (!ldlm_is_granted(lock))
708 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
709 lock->l_lvb_data, lvb_len);
710 unlock_res_and_lock(lock);
718 rc = ldlm_lock_enqueue(env, ns, &lock, NULL, flags);
719 if (lock->l_completion_ast != NULL) {
720 int err = lock->l_completion_ast(lock, *flags, NULL);
728 if (lvb_len > 0 && lvb != NULL) {
729 /* Copy the LVB here, and not earlier, because the completion
730 * AST (if any) can override what we got in the reply */
731 memcpy(lvb, lock->l_lvb_data, lvb_len);
734 LDLM_DEBUG(lock, "client-side enqueue END");
737 if (cleanup_phase == 1 && rc)
738 failed_lock_cleanup(ns, lock, mode);
739 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
741 LDLM_LOCK_RELEASE(lock);
744 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
747 * Estimate number of lock handles that would fit into request of given
748 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
749 * a single page on the send/receive side. XXX: 512 should be changed to
750 * more adequate value.
752 static inline int ldlm_req_handles_avail(int req_size, int off)
756 avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512) - req_size;
757 if (likely(avail >= 0))
758 avail /= (int)sizeof(struct lustre_handle);
761 avail += LDLM_LOCKREQ_HANDLES - off;
766 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
767 enum req_location loc,
770 __u32 size = req_capsule_msg_size(pill, loc);
771 return ldlm_req_handles_avail(size, off);
774 static inline int ldlm_format_handles_avail(struct obd_import *imp,
775 const struct req_format *fmt,
776 enum req_location loc, int off)
778 __u32 size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
779 return ldlm_req_handles_avail(size, off);
783 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
784 * \a count locks in \a cancels.
786 * This is to be called by functions preparing their own requests that
787 * might contain lists of locks to cancel in addition to actual operation
788 * that needs to be performed.
790 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
791 int version, int opc, int canceloff,
792 struct list_head *cancels, int count)
794 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
795 struct req_capsule *pill = &req->rq_pill;
796 struct ldlm_request *dlm = NULL;
797 struct list_head head = LIST_HEAD_INIT(head);
798 enum ldlm_lru_flags lru_flags;
799 int avail, to_free, pack = 0;
805 if (ns_connect_cancelset(ns)) {
806 /* Estimate the amount of available space in the request. */
807 req_capsule_filled_sizes(pill, RCL_CLIENT);
808 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
810 lru_flags = LDLM_LRU_FLAG_NO_WAIT | (ns_connect_lru_resize(ns) ?
811 LDLM_LRU_FLAG_LRUR : LDLM_LRU_FLAG_AGED);
812 to_free = !ns_connect_lru_resize(ns) &&
813 opc == LDLM_ENQUEUE ? 1 : 0;
815 /* Cancel LRU locks here _only_ if the server supports
816 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
817 * RPC, which will make us slower. */
819 count += ldlm_cancel_lru_local(ns, cancels, to_free,
826 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
827 ldlm_request_bufsize(pack, opc));
830 rc = ptlrpc_request_pack(req, version, opc);
832 ldlm_lock_list_put(cancels, l_bl_ast, count);
836 if (ns_connect_cancelset(ns)) {
838 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
840 /* Skip first lock handler in ldlm_request_pack(),
841 * this method will increment @lock_count according
842 * to the lock handle amount actually written to
844 dlm->lock_count = canceloff;
846 /* Pack into the request @pack lock handles. */
847 ldlm_cli_cancel_list(cancels, pack, req, 0);
848 /* Prepare and send separate cancel RPC for others. */
849 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
851 ldlm_lock_list_put(cancels, l_bl_ast, count);
855 EXPORT_SYMBOL(ldlm_prep_elc_req);
857 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
858 struct list_head *cancels, int count)
860 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
861 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
863 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
865 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
867 struct ptlrpc_request *req;
871 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
873 RETURN(ERR_PTR(-ENOMEM));
875 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
877 ptlrpc_request_free(req);
881 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
882 ptlrpc_request_set_replen(req);
885 EXPORT_SYMBOL(ldlm_enqueue_pack);
888 * Client-side lock enqueue.
890 * If a request has some specific initialisation it is passed in \a reqp,
891 * otherwise it is created in ldlm_cli_enqueue.
893 * Supports sync and async requests, pass \a async flag accordingly. If a
894 * request was created in ldlm_cli_enqueue and it is the async request,
895 * pass it to the caller in \a reqp.
897 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
898 struct ldlm_enqueue_info *einfo,
899 const struct ldlm_res_id *res_id,
900 union ldlm_policy_data const *policy, __u64 *flags,
901 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
902 struct lustre_handle *lockh, int async)
904 struct ldlm_namespace *ns;
905 struct ldlm_lock *lock;
906 struct ldlm_request *body;
907 int is_replay = *flags & LDLM_FL_REPLAY;
908 int req_passed_in = 1;
910 struct ptlrpc_request *req;
913 LASSERT(exp != NULL);
915 ns = exp->exp_obd->obd_namespace;
917 /* If we're replaying this lock, just check some invariants.
918 * If we're creating a new lock, get everything all setup nice. */
920 lock = ldlm_handle2lock_long(lockh, 0);
921 LASSERT(lock != NULL);
922 LDLM_DEBUG(lock, "client-side enqueue START");
923 LASSERT(exp == lock->l_conn_export);
925 const struct ldlm_callback_suite cbs = {
926 .lcs_completion = einfo->ei_cb_cp,
927 .lcs_blocking = einfo->ei_cb_bl,
928 .lcs_glimpse = einfo->ei_cb_gl
930 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
931 einfo->ei_mode, &cbs, einfo->ei_cbdata,
934 RETURN(PTR_ERR(lock));
936 if (einfo->ei_cb_created)
937 einfo->ei_cb_created(lock);
939 /* for the local lock, add the reference */
940 ldlm_lock_addref_internal(lock, einfo->ei_mode);
941 ldlm_lock2handle(lock, lockh);
943 lock->l_policy_data = *policy;
945 if (einfo->ei_type == LDLM_EXTENT) {
946 /* extent lock without policy is a bug */
950 lock->l_req_extent = policy->l_extent;
952 LDLM_DEBUG(lock, "client-side enqueue START, flags %#llx",
956 lock->l_conn_export = exp;
957 lock->l_export = NULL;
958 lock->l_blocking_ast = einfo->ei_cb_bl;
959 lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
960 lock->l_activity = ktime_get_real_seconds();
962 /* lock not sent to server yet */
963 if (reqp == NULL || *reqp == NULL) {
964 req = ldlm_enqueue_pack(exp, lvb_len);
966 failed_lock_cleanup(ns, lock, einfo->ei_mode);
967 LDLM_LOCK_RELEASE(lock);
968 RETURN(PTR_ERR(req));
978 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
980 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
981 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
984 if (*flags & LDLM_FL_NDELAY) {
985 DEBUG_REQ(D_DLMTRACE, req, "enque lock with no delay\n");
986 req->rq_no_resend = req->rq_no_delay = 1;
987 /* probably set a shorter timeout value and handle ETIMEDOUT
988 * in osc_lock_upcall() correctly */
989 /* lustre_msg_set_timeout(req, req->rq_timeout / 2); */
992 /* Dump lock data into the request buffer */
993 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
994 ldlm_lock2desc(lock, &body->lock_desc);
995 body->lock_flags = ldlm_flags_to_wire(*flags);
996 body->lock_handle[0] = *lockh;
998 /* extended LDLM opcodes in client stats */
999 if (exp->exp_obd->obd_svc_stats != NULL) {
1000 bool glimpse = *flags & LDLM_FL_HAS_INTENT;
1002 /* OST glimpse has no intent buffer */
1003 if (req_capsule_has_field(&req->rq_pill, &RMF_LDLM_INTENT,
1005 struct ldlm_intent *it;
1007 it = req_capsule_client_get(&req->rq_pill,
1009 glimpse = (it && (it->opc == IT_GLIMPSE));
1013 ldlm_svc_get_eopc(body, exp->exp_obd->obd_svc_stats);
1015 lprocfs_counter_incr(exp->exp_obd->obd_svc_stats,
1017 LDLM_GLIMPSE_ENQUEUE);
1021 LASSERT(reqp != NULL);
1025 LDLM_DEBUG(lock, "sending request");
1027 rc = ptlrpc_queue_wait(req);
1029 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
1030 einfo->ei_mode, flags, lvb, lvb_len,
1033 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
1034 * one reference that we took */
1036 LDLM_LOCK_RELEASE(lock);
1040 if (!req_passed_in && req != NULL) {
1041 ptlrpc_req_finished(req);
1048 EXPORT_SYMBOL(ldlm_cli_enqueue);
1051 * Client-side lock convert reply handling.
1053 * Finish client lock converting, checks for concurrent converts
1054 * and clear 'converting' flag so lock can be placed back into LRU.
1056 static int lock_convert_interpret(const struct lu_env *env,
1057 struct ptlrpc_request *req,
1060 struct ldlm_async_args *aa = args;
1061 struct ldlm_lock *lock;
1062 struct ldlm_reply *reply;
1066 lock = ldlm_handle2lock(&aa->lock_handle);
1068 LDLM_DEBUG_NOLOCK("convert ACK for unknown local cookie %#llx",
1069 aa->lock_handle.cookie);
1073 LDLM_DEBUG(lock, "CONVERTED lock:");
1078 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1080 GOTO(out, rc = -EPROTO);
1082 if (reply->lock_handle.cookie != aa->lock_handle.cookie) {
1083 LDLM_ERROR(lock, "convert ACK with wrong lock cookie %#llx"
1084 " but cookie %#llx from server %s id %s\n",
1085 aa->lock_handle.cookie, reply->lock_handle.cookie,
1086 req->rq_export->exp_client_uuid.uuid,
1087 libcfs_id2str(req->rq_peer));
1088 GOTO(out, rc = ELDLM_NO_LOCK_DATA);
1091 lock_res_and_lock(lock);
1092 /* Lock convert is sent for any new bits to drop, the converting flag
1093 * is dropped when ibits on server are the same as on client. Meanwhile
1094 * that can be so that more later convert will be replied first with
1095 * and clear converting flag, so in case of such race just exit here.
1096 * if lock has no converting bits then */
1097 if (!ldlm_is_converting(lock)) {
1098 LDLM_DEBUG(lock, "convert ACK for lock without converting flag,"
1099 " reply ibits %#llx",
1100 reply->lock_desc.l_policy_data.l_inodebits.bits);
1101 } else if (reply->lock_desc.l_policy_data.l_inodebits.bits !=
1102 lock->l_policy_data.l_inodebits.bits) {
1103 /* Compare server returned lock ibits and local lock ibits
1104 * if they are the same we consider convertion is done,
1105 * otherwise we have more converts inflight and keep
1108 LDLM_DEBUG(lock, "convert ACK with ibits %#llx\n",
1109 reply->lock_desc.l_policy_data.l_inodebits.bits);
1111 ldlm_clear_converting(lock);
1113 /* Concurrent BL AST may arrive and cause another convert
1114 * or cancel so just do nothing here if bl_ast is set,
1115 * finish with convert otherwise.
1117 if (!ldlm_is_bl_ast(lock)) {
1118 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
1120 /* Drop cancel_bits since there are no more converts
1121 * and put lock into LRU if it is still not used and
1124 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1125 if (!lock->l_readers && !lock->l_writers &&
1126 !ldlm_is_canceling(lock)) {
1127 spin_lock(&ns->ns_lock);
1128 /* there is check for list_empty() inside */
1129 ldlm_lock_remove_from_lru_nolock(lock);
1130 ldlm_lock_add_to_lru_nolock(lock);
1131 spin_unlock(&ns->ns_lock);
1135 unlock_res_and_lock(lock);
1140 lock_res_and_lock(lock);
1141 if (ldlm_is_converting(lock)) {
1142 ldlm_clear_converting(lock);
1143 ldlm_set_cbpending(lock);
1144 ldlm_set_bl_ast(lock);
1145 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1147 unlock_res_and_lock(lock);
1149 /* fallback to normal lock cancel. If rc means there is no
1150 * valid lock on server, do only local cancel */
1151 if (rc == ELDLM_NO_LOCK_DATA)
1156 rc = ldlm_cli_cancel(&aa->lock_handle, flag);
1158 LDLM_DEBUG(lock, "failed to cancel lock: rc = %d\n",
1161 LDLM_LOCK_PUT(lock);
1166 * Client-side IBITS lock convert.
1168 * Inform server that lock has been converted instead of canceling.
1169 * Server finishes convert on own side and does reprocess to grant
1170 * all related waiting locks.
1172 * Since convert means only ibits downgrading, client doesn't need to
1173 * wait for server reply to finish local converting process so this request
1174 * is made asynchronous.
1177 int ldlm_cli_convert(struct ldlm_lock *lock, __u32 *flags)
1179 struct ldlm_request *body;
1180 struct ptlrpc_request *req;
1181 struct ldlm_async_args *aa;
1182 struct obd_export *exp = lock->l_conn_export;
1187 LDLM_ERROR(lock, "convert must not be called on local locks.");
1191 /* this is better to check earlier and it is done so already,
1192 * but this check is kept too as final one to issue an error
1193 * if any new code will miss such check.
1195 if (!exp_connect_lock_convert(exp)) {
1196 LDLM_ERROR(lock, "server doesn't support lock convert\n");
1200 if (lock->l_resource->lr_type != LDLM_IBITS) {
1201 LDLM_ERROR(lock, "convert works with IBITS locks only.");
1205 LDLM_DEBUG(lock, "client-side convert");
1207 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1208 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1213 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1214 body->lock_handle[0] = lock->l_remote_handle;
1216 body->lock_desc.l_req_mode = lock->l_req_mode;
1217 body->lock_desc.l_granted_mode = lock->l_granted_mode;
1219 body->lock_desc.l_policy_data.l_inodebits.bits =
1220 lock->l_policy_data.l_inodebits.bits;
1221 body->lock_desc.l_policy_data.l_inodebits.cancel_bits = 0;
1223 body->lock_flags = ldlm_flags_to_wire(*flags);
1224 body->lock_count = 1;
1226 ptlrpc_request_set_replen(req);
1229 * Use cancel portals for convert as well as high-priority handling.
1231 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1232 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1234 ptlrpc_at_set_req_timeout(req);
1236 if (exp->exp_obd->obd_svc_stats != NULL)
1237 lprocfs_counter_incr(exp->exp_obd->obd_svc_stats,
1238 LDLM_CONVERT - LDLM_FIRST_OPC);
1240 aa = ptlrpc_req_async_args(req);
1241 ldlm_lock2handle(lock, &aa->lock_handle);
1242 req->rq_interpret_reply = lock_convert_interpret;
1244 ptlrpcd_add_req(req);
1249 * Cancel locks locally.
1251 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1252 * \retval LDLM_FL_CANCELING otherwise;
1253 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1255 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1257 __u64 rc = LDLM_FL_LOCAL_ONLY;
1260 if (lock->l_conn_export) {
1263 LDLM_DEBUG(lock, "client-side cancel");
1264 /* Set this flag to prevent others from getting new references*/
1265 lock_res_and_lock(lock);
1266 ldlm_set_cbpending(lock);
1267 local_only = !!(lock->l_flags &
1268 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1269 ldlm_cancel_callback(lock);
1270 rc = (ldlm_is_bl_ast(lock)) ?
1271 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1272 unlock_res_and_lock(lock);
1275 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1277 rc = LDLM_FL_LOCAL_ONLY;
1279 ldlm_lock_cancel(lock);
1281 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1282 LDLM_ERROR(lock, "Trying to cancel local lock");
1285 LDLM_DEBUG(lock, "server-side local cancel");
1286 ldlm_lock_cancel(lock);
1287 ldlm_reprocess_all(lock->l_resource);
1294 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1296 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1297 struct list_head *head, int count)
1299 struct ldlm_request *dlm;
1300 struct ldlm_lock *lock;
1301 int max, packed = 0;
1304 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1305 LASSERT(dlm != NULL);
1307 /* Check the room in the request buffer. */
1308 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1309 sizeof(struct ldlm_request);
1310 max /= sizeof(struct lustre_handle);
1311 max += LDLM_LOCKREQ_HANDLES;
1312 LASSERT(max >= dlm->lock_count + count);
1314 /* XXX: it would be better to pack lock handles grouped by resource.
1315 * so that the server cancel would call filter_lvbo_update() less
1317 list_for_each_entry(lock, head, l_bl_ast) {
1320 LASSERT(lock->l_conn_export);
1321 /* Pack the lock handle to the given request buffer. */
1322 LDLM_DEBUG(lock, "packing");
1323 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1326 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1331 * Prepare and send a batched cancel RPC. It will include \a count lock
1332 * handles of locks given in \a cancels list. */
1333 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1334 int count, enum ldlm_cancel_flags flags)
1336 struct ptlrpc_request *req = NULL;
1337 struct obd_import *imp;
1342 LASSERT(exp != NULL);
1345 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1347 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1350 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1351 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1356 imp = class_exp2cliimp(exp);
1357 if (imp == NULL || imp->imp_invalid) {
1359 "skipping cancel on invalid import %p\n", imp);
1363 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1365 GOTO(out, rc = -ENOMEM);
1367 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1368 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1369 ldlm_request_bufsize(count, LDLM_CANCEL));
1371 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1373 ptlrpc_request_free(req);
1377 /* If OSP want cancel cross-MDT lock, let's not block it in
1378 * in recovery, otherwise the lock will not released, if
1379 * the remote target is also in recovery, and it also need
1380 * this lock, it might cause deadlock. */
1381 if (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS &&
1382 exp->exp_obd->obd_lu_dev != NULL &&
1383 exp->exp_obd->obd_lu_dev->ld_site != NULL) {
1384 struct lu_device *top_dev;
1386 top_dev = exp->exp_obd->obd_lu_dev->ld_site->ls_top_dev;
1387 if (top_dev != NULL &&
1388 top_dev->ld_obd->obd_recovering)
1389 req->rq_allow_replay = 1;
1392 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1393 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1394 ptlrpc_at_set_req_timeout(req);
1396 ldlm_cancel_pack(req, cancels, count);
1398 ptlrpc_request_set_replen(req);
1399 if (flags & LCF_ASYNC) {
1400 ptlrpcd_add_req(req);
1405 rc = ptlrpc_queue_wait(req);
1406 if (rc == LUSTRE_ESTALE) {
1407 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1408 "out of sync -- not fatal\n",
1409 libcfs_nid2str(req->rq_import->
1410 imp_connection->c_peer.nid));
1412 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1413 req->rq_import_generation == imp->imp_generation) {
1414 ptlrpc_req_finished(req);
1416 } else if (rc != ELDLM_OK) {
1417 /* -ESHUTDOWN is common on umount */
1418 CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1419 "Got rc %d from cancel RPC: "
1420 "canceling anyway\n", rc);
1427 ptlrpc_req_finished(req);
1430 return sent ? sent : rc;
1433 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1435 LASSERT(imp != NULL);
1436 return &imp->imp_obd->obd_namespace->ns_pool;
1440 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1442 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1444 struct obd_device *obd;
1448 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1449 !imp_connect_lru_resize(req->rq_import)))
1452 * Do nothing for corner cases.
1457 /* In some cases RPC may contain SLV and limit zeroed out. This
1458 * is the case when server does not support LRU resize feature.
1459 * This is also possible in some recovery cases when server-side
1460 * reqs have no reference to the OBD export and thus access to
1461 * server-side namespace is not possible. */
1462 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1463 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1464 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1465 "(SLV: %llu, Limit: %u)",
1466 lustre_msg_get_slv(req->rq_repmsg),
1467 lustre_msg_get_limit(req->rq_repmsg));
1471 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1472 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1473 obd = req->rq_import->imp_obd;
1475 /* Set new SLV and limit in OBD fields to make them accessible
1476 * to the pool thread. We do not access obd_namespace and pool
1477 * directly here as there is no reliable way to make sure that
1478 * they are still alive at cleanup time. Evil races are possible
1479 * which may cause Oops at that time. */
1480 write_lock(&obd->obd_pool_lock);
1481 obd->obd_pool_slv = new_slv;
1482 obd->obd_pool_limit = new_limit;
1483 write_unlock(&obd->obd_pool_lock);
1489 * Client side lock cancel.
1491 * Lock must not have any readers or writers by this time.
1493 int ldlm_cli_cancel(const struct lustre_handle *lockh,
1494 enum ldlm_cancel_flags cancel_flags)
1496 struct obd_export *exp;
1497 enum ldlm_lru_flags lru_flags;
1498 int avail, count = 1;
1500 struct ldlm_namespace *ns;
1501 struct ldlm_lock *lock;
1502 struct list_head cancels = LIST_HEAD_INIT(cancels);
1506 lock = ldlm_handle2lock_long(lockh, 0);
1508 LDLM_DEBUG_NOLOCK("lock is already being destroyed");
1512 /* Convert lock bits instead of cancel for IBITS locks */
1513 if (cancel_flags & LCF_CONVERT) {
1514 LASSERT(lock->l_resource->lr_type == LDLM_IBITS);
1515 LASSERT(lock->l_policy_data.l_inodebits.cancel_bits != 0);
1517 rc = ldlm_cli_dropbits(lock,
1518 lock->l_policy_data.l_inodebits.cancel_bits);
1520 LDLM_LOCK_RELEASE(lock);
1525 lock_res_and_lock(lock);
1526 /* Lock is being canceled and the caller doesn't want to wait */
1527 if (ldlm_is_canceling(lock)) {
1528 if (cancel_flags & LCF_ASYNC) {
1529 unlock_res_and_lock(lock);
1531 struct l_wait_info lwi = { 0 };
1533 unlock_res_and_lock(lock);
1534 l_wait_event(lock->l_waitq, is_bl_done(lock), &lwi);
1536 LDLM_LOCK_RELEASE(lock);
1540 /* Lock is being converted, cancel it immediately.
1541 * When convert will end, it releases lock and it will be gone.
1543 if (ldlm_is_converting(lock)) {
1544 /* set back flags removed by convert */
1545 ldlm_set_cbpending(lock);
1546 ldlm_set_bl_ast(lock);
1549 ldlm_set_canceling(lock);
1550 unlock_res_and_lock(lock);
1552 if (cancel_flags & LCF_LOCAL)
1553 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_LOCAL_CANCEL_PAUSE,
1556 rc = ldlm_cli_cancel_local(lock);
1557 if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
1558 LDLM_LOCK_RELEASE(lock);
1561 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1562 * RPC which goes to canceld portal, so we can cancel other LRU locks
1563 * here and send them all as one LDLM_CANCEL RPC. */
1564 LASSERT(list_empty(&lock->l_bl_ast));
1565 list_add(&lock->l_bl_ast, &cancels);
1567 exp = lock->l_conn_export;
1568 if (exp_connect_cancelset(exp)) {
1569 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1574 ns = ldlm_lock_to_ns(lock);
1575 lru_flags = ns_connect_lru_resize(ns) ?
1576 LDLM_LRU_FLAG_LRUR : LDLM_LRU_FLAG_AGED;
1577 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1578 LCF_BL_AST, lru_flags);
1580 ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1583 EXPORT_SYMBOL(ldlm_cli_cancel);
1586 * Locally cancel up to \a count locks in list \a cancels.
1587 * Return the number of cancelled locks.
1589 int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
1590 enum ldlm_cancel_flags cancel_flags)
1592 struct list_head head = LIST_HEAD_INIT(head);
1593 struct ldlm_lock *lock, *next;
1594 int left = 0, bl_ast = 0;
1598 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1602 if (cancel_flags & LCF_LOCAL) {
1603 rc = LDLM_FL_LOCAL_ONLY;
1604 ldlm_lock_cancel(lock);
1606 rc = ldlm_cli_cancel_local(lock);
1608 /* Until we have compound requests and can send LDLM_CANCEL
1609 * requests batched with generic RPCs, we need to send cancels
1610 * with the LDLM_FL_BL_AST flag in a separate RPC from
1611 * the one being generated now. */
1612 if (!(cancel_flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1613 LDLM_DEBUG(lock, "Cancel lock separately");
1614 list_del_init(&lock->l_bl_ast);
1615 list_add(&lock->l_bl_ast, &head);
1619 if (rc == LDLM_FL_LOCAL_ONLY) {
1620 /* CANCEL RPC should not be sent to server. */
1621 list_del_init(&lock->l_bl_ast);
1622 LDLM_LOCK_RELEASE(lock);
1628 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1635 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1636 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1637 * readahead requests, ...)
1639 static enum ldlm_policy_res
1640 ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1641 int unused, int added, int count)
1643 enum ldlm_policy_res result = LDLM_POLICY_CANCEL_LOCK;
1645 /* don't check added & count since we want to process all locks
1647 * It's fine to not take lock to access lock->l_resource since
1648 * the lock has already been granted so it won't change. */
1649 switch (lock->l_resource->lr_type) {
1652 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) != 0)
1655 result = LDLM_POLICY_SKIP_LOCK;
1663 * Callback function for LRU-resize policy. Decides whether to keep
1664 * \a lock in LRU for current \a LRU size \a unused, added in current
1665 * scan \a added and number of locks to be preferably canceled \a count.
1667 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1669 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1671 static enum ldlm_policy_res ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1672 struct ldlm_lock *lock,
1673 int unused, int added,
1676 ktime_t cur = ktime_get();
1677 struct ldlm_pool *pl = &ns->ns_pool;
1681 /* Stop LRU processing when we reach past @count or have checked all
1683 if (count && added >= count)
1684 return LDLM_POLICY_KEEP_LOCK;
1686 /* Despite of the LV, It doesn't make sense to keep the lock which
1687 * is unused for ns_max_age time.
1689 if (ktime_after(ktime_get(),
1690 ktime_add(lock->l_last_used, ns->ns_max_age)))
1691 return LDLM_POLICY_CANCEL_LOCK;
1693 slv = ldlm_pool_get_slv(pl);
1694 lvf = ldlm_pool_get_lvf(pl);
1695 la = div_u64(ktime_to_ns(ktime_sub(cur, lock->l_last_used)),
1697 lv = lvf * la * unused;
1699 /* Inform pool about current CLV to see it via debugfs. */
1700 ldlm_pool_set_clv(pl, lv);
1702 /* Stop when SLV is not yet come from server or lv is smaller than
1704 if (slv == 0 || lv < slv)
1705 return LDLM_POLICY_KEEP_LOCK;
1707 return LDLM_POLICY_CANCEL_LOCK;
1710 static enum ldlm_policy_res
1711 ldlm_cancel_lrur_no_wait_policy(struct ldlm_namespace *ns,
1712 struct ldlm_lock *lock,
1713 int unused, int added,
1716 enum ldlm_policy_res result;
1718 result = ldlm_cancel_lrur_policy(ns, lock, unused, added, count);
1719 if (result == LDLM_POLICY_KEEP_LOCK)
1722 return ldlm_cancel_no_wait_policy(ns, lock, unused, added, count);
1726 * Callback function for debugfs used policy. Makes decision whether to keep
1727 * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1728 * added and number of locks to be preferably canceled \a count.
1730 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1732 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1734 static enum ldlm_policy_res ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1735 struct ldlm_lock *lock,
1736 int unused, int added,
1739 /* Stop LRU processing when we reach past @count or have checked all
1741 return (added >= count) ?
1742 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1746 * Callback function for aged policy. Makes decision whether to keep \a lock in
1747 * LRU for current LRU size \a unused, added in current scan \a added and
1748 * number of locks to be preferably canceled \a count.
1750 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1752 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1754 static enum ldlm_policy_res ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1755 struct ldlm_lock *lock,
1756 int unused, int added,
1759 if ((added >= count) &&
1760 ktime_before(ktime_get(),
1761 ktime_add(lock->l_last_used, ns->ns_max_age)))
1762 return LDLM_POLICY_KEEP_LOCK;
1764 return LDLM_POLICY_CANCEL_LOCK;
1767 static enum ldlm_policy_res
1768 ldlm_cancel_aged_no_wait_policy(struct ldlm_namespace *ns,
1769 struct ldlm_lock *lock,
1770 int unused, int added, int count)
1772 enum ldlm_policy_res result;
1774 result = ldlm_cancel_aged_policy(ns, lock, unused, added, count);
1775 if (result == LDLM_POLICY_KEEP_LOCK)
1778 return ldlm_cancel_no_wait_policy(ns, lock, unused, added, count);
1782 * Callback function for default policy. Makes decision whether to keep \a lock
1783 * in LRU for current LRU size \a unused, added in current scan \a added and
1784 * number of locks to be preferably canceled \a count.
1786 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1788 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1791 enum ldlm_policy_res ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1792 struct ldlm_lock *lock,
1793 int unused, int added,
1796 /* Stop LRU processing when we reach past count or have checked all
1798 return (added >= count) ?
1799 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1802 typedef enum ldlm_policy_res
1803 (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1804 int unused, int added, int count);
1806 static ldlm_cancel_lru_policy_t
1807 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, enum ldlm_lru_flags lru_flags)
1809 if (ns_connect_lru_resize(ns)) {
1810 if (lru_flags & LDLM_LRU_FLAG_SHRINK)
1811 /* We kill passed number of old locks. */
1812 return ldlm_cancel_passed_policy;
1813 if (lru_flags & LDLM_LRU_FLAG_LRUR) {
1814 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1815 return ldlm_cancel_lrur_no_wait_policy;
1817 return ldlm_cancel_lrur_policy;
1819 if (lru_flags & LDLM_LRU_FLAG_PASSED)
1820 return ldlm_cancel_passed_policy;
1822 if (lru_flags & LDLM_LRU_FLAG_AGED) {
1823 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1824 return ldlm_cancel_aged_no_wait_policy;
1826 return ldlm_cancel_aged_policy;
1829 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1830 return ldlm_cancel_no_wait_policy;
1832 return ldlm_cancel_default_policy;
1836 * - Free space in LRU for \a count new locks,
1837 * redundant unused locks are canceled locally;
1838 * - also cancel locally unused aged locks;
1839 * - do not cancel more than \a max locks;
1840 * - GET the found locks and add them into the \a cancels list.
1842 * A client lock can be added to the l_bl_ast list only when it is
1843 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1844 * CANCEL. There are the following use cases:
1845 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1846 * ldlm_cli_cancel(), which check and set this flag properly. As any
1847 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1848 * later without any special locking.
1850 * Calling policies for enabled LRU resize:
1851 * ----------------------------------------
1852 * flags & LDLM_LRU_FLAG_LRUR - use LRU resize policy (SLV from server) to
1853 * cancel not more than \a count locks;
1855 * flags & LDLM_LRU_FLAG_PASSED - cancel \a count number of old locks (located
1856 * at the beginning of LRU list);
1858 * flags & LDLM_LRU_FLAG_SHRINK - cancel not more than \a count locks according
1859 * to memory pressre policy function;
1861 * flags & LDLM_LRU_FLAG_AGED - cancel \a count locks according to "aged policy"
1863 * flags & LDLM_LRU_FLAG_NO_WAIT - cancel as many unused locks as possible
1864 * (typically before replaying locks) w/o
1865 * sending any RPCs or waiting for any
1866 * outstanding RPC to complete.
1868 * flags & LDLM_CANCEL_CLEANUP - when cancelling read locks, do not check for
1869 * other read locks covering the same pages, just
1870 * discard those pages.
1872 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns,
1873 struct list_head *cancels, int count, int max,
1874 enum ldlm_lru_flags lru_flags)
1876 ldlm_cancel_lru_policy_t pf;
1878 int no_wait = lru_flags & LDLM_LRU_FLAG_NO_WAIT;
1882 if (!ns_connect_lru_resize(ns))
1883 count += ns->ns_nr_unused - ns->ns_max_unused;
1885 pf = ldlm_cancel_lru_policy(ns, lru_flags);
1886 LASSERT(pf != NULL);
1888 /* For any flags, stop scanning if @max is reached. */
1889 while (!list_empty(&ns->ns_unused_list) && (max == 0 || added < max)) {
1890 struct ldlm_lock *lock;
1891 struct list_head *item, *next;
1892 enum ldlm_policy_res result;
1893 ktime_t last_use = ktime_set(0, 0);
1895 spin_lock(&ns->ns_lock);
1896 item = no_wait ? ns->ns_last_pos : &ns->ns_unused_list;
1897 for (item = item->next, next = item->next;
1898 item != &ns->ns_unused_list;
1899 item = next, next = item->next) {
1900 lock = list_entry(item, struct ldlm_lock, l_lru);
1902 /* No locks which got blocking requests. */
1903 LASSERT(!ldlm_is_bl_ast(lock));
1905 if (!ldlm_is_canceling(lock) &&
1906 !ldlm_is_converting(lock))
1909 /* Somebody is already doing CANCEL. No need for this
1910 * lock in LRU, do not traverse it again. */
1911 ldlm_lock_remove_from_lru_nolock(lock);
1913 if (item == &ns->ns_unused_list) {
1914 spin_unlock(&ns->ns_lock);
1918 last_use = lock->l_last_used;
1920 LDLM_LOCK_GET(lock);
1921 spin_unlock(&ns->ns_lock);
1922 lu_ref_add(&lock->l_reference, __FUNCTION__, current);
1924 /* Pass the lock through the policy filter and see if it
1925 * should stay in LRU.
1927 * Even for shrinker policy we stop scanning if
1928 * we find a lock that should stay in the cache.
1929 * We should take into account lock age anyway
1930 * as a new lock is a valuable resource even if
1931 * it has a low weight.
1933 * That is, for shrinker policy we drop only
1934 * old locks, but additionally choose them by
1935 * their weight. Big extent locks will stay in
1937 result = pf(ns, lock, ns->ns_nr_unused, added, count);
1938 if (result == LDLM_POLICY_KEEP_LOCK) {
1939 lu_ref_del(&lock->l_reference, __func__, current);
1940 LDLM_LOCK_RELEASE(lock);
1944 if (result == LDLM_POLICY_SKIP_LOCK) {
1945 lu_ref_del(&lock->l_reference, __func__, current);
1947 spin_lock(&ns->ns_lock);
1948 if (!list_empty(&lock->l_lru) &&
1949 lock->l_lru.prev == ns->ns_last_pos)
1950 ns->ns_last_pos = &lock->l_lru;
1951 spin_unlock(&ns->ns_lock);
1954 LDLM_LOCK_RELEASE(lock);
1958 lock_res_and_lock(lock);
1959 /* Check flags again under the lock. */
1960 if (ldlm_is_canceling(lock) || ldlm_is_converting(lock) ||
1961 ldlm_lock_remove_from_lru_check(lock, last_use) == 0) {
1962 /* Another thread is removing lock from LRU, or
1963 * somebody is already doing CANCEL, or there
1964 * is a blocking request which will send cancel
1965 * by itself, or the lock is no longer unused or
1966 * the lock has been used since the pf() call and
1967 * pages could be put under it. */
1968 unlock_res_and_lock(lock);
1969 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1970 LDLM_LOCK_RELEASE(lock);
1973 LASSERT(!lock->l_readers && !lock->l_writers);
1975 /* If we have chosen to cancel this lock voluntarily, we
1976 * better send cancel notification to server, so that it
1977 * frees appropriate state. This might lead to a race
1978 * where while we are doing cancel here, server is also
1979 * silently cancelling this lock. */
1980 ldlm_clear_cancel_on_block(lock);
1982 /* Setting the CBPENDING flag is a little misleading,
1983 * but prevents an important race; namely, once
1984 * CBPENDING is set, the lock can accumulate no more
1985 * readers/writers. Since readers and writers are
1986 * already zero here, ldlm_lock_decref() won't see
1987 * this flag and call l_blocking_ast */
1988 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1990 if ((lru_flags & LDLM_LRU_FLAG_CLEANUP) &&
1991 (lock->l_resource->lr_type == LDLM_EXTENT ||
1992 ldlm_has_dom(lock)) && lock->l_granted_mode == LCK_PR)
1993 ldlm_set_discard_data(lock);
1995 /* We can't re-add to l_lru as it confuses the
1996 * refcounting in ldlm_lock_remove_from_lru() if an AST
1997 * arrives after we drop lr_lock below. We use l_bl_ast
1998 * and can't use l_pending_chain as it is used both on
1999 * server and client nevertheless bug 5666 says it is
2000 * used only on server */
2001 LASSERT(list_empty(&lock->l_bl_ast));
2002 list_add(&lock->l_bl_ast, cancels);
2003 unlock_res_and_lock(lock);
2004 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
2010 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
2012 enum ldlm_cancel_flags cancel_flags,
2013 enum ldlm_lru_flags lru_flags)
2017 added = ldlm_prepare_lru_list(ns, cancels, count, max, lru_flags);
2021 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
2025 * Cancel at least \a nr locks from given namespace LRU.
2027 * When called with LCF_ASYNC the blocking callback will be handled
2028 * in a thread and this function will return after the thread has been
2029 * asked to call the callback. When called with LCF_ASYNC the blocking
2030 * callback will be performed in this function.
2032 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
2033 enum ldlm_cancel_flags cancel_flags,
2034 enum ldlm_lru_flags lru_flags)
2036 struct list_head cancels = LIST_HEAD_INIT(cancels);
2040 /* Just prepare the list of locks, do not actually cancel them yet.
2041 * Locks are cancelled later in a separate thread. */
2042 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, lru_flags);
2043 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
2051 * Find and cancel locally unused locks found on resource, matched to the
2052 * given policy, mode. GET the found locks and add them into the \a cancels
2055 int ldlm_cancel_resource_local(struct ldlm_resource *res,
2056 struct list_head *cancels,
2057 union ldlm_policy_data *policy,
2058 enum ldlm_mode mode, __u64 lock_flags,
2059 enum ldlm_cancel_flags cancel_flags,
2062 struct ldlm_lock *lock;
2068 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
2069 if (opaque != NULL && lock->l_ast_data != opaque) {
2070 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
2071 lock->l_ast_data, opaque);
2075 if (lock->l_readers || lock->l_writers)
2078 /* If somebody is already doing CANCEL, or blocking AST came,
2079 * or lock is being converted then skip this lock. */
2080 if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock) ||
2081 ldlm_is_converting(lock))
2084 if (lockmode_compat(lock->l_granted_mode, mode))
2087 /* If policy is given and this is IBITS lock, add to list only
2088 * those locks that match by policy.
2089 * Skip locks with DoM bit always to don't flush data.
2091 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
2092 (!(lock->l_policy_data.l_inodebits.bits &
2093 policy->l_inodebits.bits) || ldlm_has_dom(lock)))
2096 /* See CBPENDING comment in ldlm_cancel_lru */
2097 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
2100 LASSERT(list_empty(&lock->l_bl_ast));
2101 list_add(&lock->l_bl_ast, cancels);
2102 LDLM_LOCK_GET(lock);
2107 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
2109 EXPORT_SYMBOL(ldlm_cancel_resource_local);
2112 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
2114 * If \a req is NULL, send CANCEL request to server with handles of locks
2115 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
2116 * separately per lock.
2117 * If \a req is not NULL, put handles of locks in \a cancels into the request
2118 * buffer at the offset \a off.
2119 * Destroy \a cancels at the end.
2121 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
2122 struct ptlrpc_request *req,
2123 enum ldlm_cancel_flags flags)
2125 struct ldlm_lock *lock;
2129 if (list_empty(cancels) || count == 0)
2132 /* XXX: requests (both batched and not) could be sent in parallel.
2133 * Usually it is enough to have just 1 RPC, but it is possible that
2134 * there are too many locks to be cancelled in LRU or on a resource.
2135 * It would also speed up the case when the server does not support
2138 LASSERT(!list_empty(cancels));
2139 lock = list_entry(cancels->next, struct ldlm_lock,
2141 LASSERT(lock->l_conn_export);
2143 if (exp_connect_cancelset(lock->l_conn_export)) {
2146 ldlm_cancel_pack(req, cancels, count);
2148 res = ldlm_cli_cancel_req(lock->l_conn_export,
2152 res = ldlm_cli_cancel_req(lock->l_conn_export,
2157 CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
2158 "ldlm_cli_cancel_list: %d\n", res);
2163 ldlm_lock_list_put(cancels, l_bl_ast, res);
2165 LASSERT(count == 0);
2168 EXPORT_SYMBOL(ldlm_cli_cancel_list);
2171 * Cancel all locks on a resource that have 0 readers/writers.
2173 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
2174 * to notify the server. */
2175 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
2176 const struct ldlm_res_id *res_id,
2177 union ldlm_policy_data *policy,
2178 enum ldlm_mode mode,
2179 enum ldlm_cancel_flags flags, void *opaque)
2181 struct ldlm_resource *res;
2182 struct list_head cancels = LIST_HEAD_INIT(cancels);
2187 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2189 /* This is not a problem. */
2190 CDEBUG(D_INFO, "No resource %llu\n", res_id->name[0]);
2194 LDLM_RESOURCE_ADDREF(res);
2195 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
2196 0, flags | LCF_BL_AST, opaque);
2197 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
2199 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
2202 LDLM_RESOURCE_DELREF(res);
2203 ldlm_resource_putref(res);
2206 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2208 struct ldlm_cli_cancel_arg {
2214 ldlm_cli_hash_cancel_unused(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2215 struct hlist_node *hnode, void *arg)
2217 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2218 struct ldlm_cli_cancel_arg *lc = arg;
2220 ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
2221 NULL, LCK_MINMODE, lc->lc_flags,
2223 /* must return 0 for hash iteration */
2228 * Cancel all locks on a namespace (or a specific resource, if given)
2229 * that have 0 readers/writers.
2231 * If flags & LCF_LOCAL, throw the locks away without trying
2232 * to notify the server. */
2233 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
2234 const struct ldlm_res_id *res_id,
2235 enum ldlm_cancel_flags flags, void *opaque)
2237 struct ldlm_cli_cancel_arg arg = {
2239 .lc_opaque = opaque,
2247 if (res_id != NULL) {
2248 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
2252 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2253 ldlm_cli_hash_cancel_unused, &arg, 0);
2258 /* Lock iterators. */
2260 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
2263 struct list_head *tmp, *next;
2264 struct ldlm_lock *lock;
2265 int rc = LDLM_ITER_CONTINUE;
2270 RETURN(LDLM_ITER_CONTINUE);
2273 list_for_each_safe(tmp, next, &res->lr_granted) {
2274 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2276 if (iter(lock, closure) == LDLM_ITER_STOP)
2277 GOTO(out, rc = LDLM_ITER_STOP);
2280 list_for_each_safe(tmp, next, &res->lr_waiting) {
2281 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2283 if (iter(lock, closure) == LDLM_ITER_STOP)
2284 GOTO(out, rc = LDLM_ITER_STOP);
2291 struct iter_helper_data {
2292 ldlm_iterator_t iter;
2296 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2298 struct iter_helper_data *helper = closure;
2299 return helper->iter(lock, helper->closure);
2302 static int ldlm_res_iter_helper(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2303 struct hlist_node *hnode, void *arg)
2306 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2308 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2312 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2313 ldlm_iterator_t iter, void *closure)
2316 struct iter_helper_data helper = { .iter = iter, .closure = closure };
2318 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2319 ldlm_res_iter_helper, &helper, 0);
2323 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2324 * return 0: find no resource
2325 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2328 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2329 const struct ldlm_res_id *res_id,
2330 ldlm_iterator_t iter, void *data)
2332 struct ldlm_resource *res;
2336 LASSERTF(ns != NULL, "must pass in namespace\n");
2338 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2342 LDLM_RESOURCE_ADDREF(res);
2343 rc = ldlm_resource_foreach(res, iter, data);
2344 LDLM_RESOURCE_DELREF(res);
2345 ldlm_resource_putref(res);
2348 EXPORT_SYMBOL(ldlm_resource_iterate);
2352 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2354 struct list_head *list = closure;
2356 /* we use l_pending_chain here, because it's unused on clients. */
2357 LASSERTF(list_empty(&lock->l_pending_chain),
2358 "lock %p next %p prev %p\n",
2359 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2360 /* bug 9573: don't replay locks left after eviction, or
2361 * bug 17614: locks being actively cancelled. Get a reference
2362 * on a lock so that it does not disapear under us (e.g. due to cancel)
2364 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_BL_DONE))) {
2365 list_add(&lock->l_pending_chain, list);
2366 LDLM_LOCK_GET(lock);
2369 return LDLM_ITER_CONTINUE;
2372 static int replay_lock_interpret(const struct lu_env *env,
2373 struct ptlrpc_request *req, void *args, int rc)
2375 struct ldlm_async_args *aa = args;
2376 struct ldlm_lock *lock;
2377 struct ldlm_reply *reply;
2378 struct obd_export *exp;
2381 atomic_dec(&req->rq_import->imp_replay_inflight);
2385 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2387 GOTO(out, rc = -EPROTO);
2389 lock = ldlm_handle2lock(&aa->lock_handle);
2391 CERROR("received replay ack for unknown local cookie %#llx"
2392 " remote cookie %#llx from server %s id %s\n",
2393 aa->lock_handle.cookie, reply->lock_handle.cookie,
2394 req->rq_export->exp_client_uuid.uuid,
2395 libcfs_id2str(req->rq_peer));
2396 GOTO(out, rc = -ESTALE);
2399 /* Key change rehash lock in per-export hash with new key */
2400 exp = req->rq_export;
2401 if (exp && exp->exp_lock_hash) {
2402 /* In the function below, .hs_keycmp resolves to
2403 * ldlm_export_lock_keycmp() */
2404 /* coverity[overrun-buffer-val] */
2405 cfs_hash_rehash_key(exp->exp_lock_hash,
2406 &lock->l_remote_handle,
2407 &reply->lock_handle,
2410 lock->l_remote_handle = reply->lock_handle;
2413 LDLM_DEBUG(lock, "replayed lock:");
2414 ptlrpc_import_recovery_state_machine(req->rq_import);
2415 LDLM_LOCK_PUT(lock);
2418 ptlrpc_connect_import(req->rq_import);
2423 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2425 struct ptlrpc_request *req;
2426 struct ldlm_async_args *aa;
2427 struct ldlm_request *body;
2432 /* Bug 11974: Do not replay a lock which is actively being canceled */
2433 if (ldlm_is_bl_done(lock)) {
2434 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2438 /* If this is reply-less callback lock, we cannot replay it, since
2439 * server might have long dropped it, but notification of that event was
2440 * lost by network. (and server granted conflicting lock already) */
2441 if (ldlm_is_cancel_on_block(lock)) {
2442 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2443 ldlm_lock_cancel(lock);
2448 * If granted mode matches the requested mode, this lock is granted.
2450 * If we haven't been granted anything and are on a resource list,
2451 * then we're blocked/waiting.
2453 * If we haven't been granted anything and we're NOT on a resource list,
2454 * then we haven't got a reply yet and don't have a known disposition.
2455 * This happens whenever a lock enqueue is the request that triggers
2458 if (ldlm_is_granted(lock))
2459 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2460 else if (!list_empty(&lock->l_res_link))
2461 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2463 flags = LDLM_FL_REPLAY;
2465 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2466 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2470 /* We're part of recovery, so don't wait for it. */
2471 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2473 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2474 ldlm_lock2desc(lock, &body->lock_desc);
2475 body->lock_flags = ldlm_flags_to_wire(flags);
2477 ldlm_lock2handle(lock, &body->lock_handle[0]);
2478 if (lock->l_lvb_len > 0)
2479 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2480 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2482 ptlrpc_request_set_replen(req);
2483 /* notify the server we've replayed all requests.
2484 * also, we mark the request to be put on a dedicated
2485 * queue to be processed after all request replayes.
2487 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2489 LDLM_DEBUG(lock, "replaying lock:");
2491 atomic_inc(&req->rq_import->imp_replay_inflight);
2492 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2493 aa = ptlrpc_req_async_args(req);
2494 aa->lock_handle = body->lock_handle[0];
2495 req->rq_interpret_reply = replay_lock_interpret;
2496 ptlrpcd_add_req(req);
2502 * Cancel as many unused locks as possible before replay. since we are
2503 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2506 * Called only in recovery before replaying locks. there is no need to
2507 * replay locks that are unused. since the clients may hold thousands of
2508 * cached unused locks, dropping the unused locks can greatly reduce the
2509 * load on the servers at recovery time.
2511 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2514 struct list_head cancels = LIST_HEAD_INIT(cancels);
2516 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2517 "replay for namespace %s (%d)\n",
2518 ldlm_ns_name(ns), ns->ns_nr_unused);
2520 /* We don't need to care whether or not LRU resize is enabled
2521 * because the LDLM_LRU_FLAG_NO_WAIT policy doesn't use the
2522 * count parameter */
2523 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2524 LCF_LOCAL, LDLM_LRU_FLAG_NO_WAIT);
2526 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2527 canceled, ldlm_ns_name(ns));
2530 int ldlm_replay_locks(struct obd_import *imp)
2532 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2533 struct list_head list = LIST_HEAD_INIT(list);
2534 struct ldlm_lock *lock, *next;
2539 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2541 /* don't replay locks if import failed recovery */
2542 if (imp->imp_vbr_failed)
2545 /* ensure this doesn't fall to 0 before all have been queued */
2546 atomic_inc(&imp->imp_replay_inflight);
2548 if (ldlm_cancel_unused_locks_before_replay)
2549 ldlm_cancel_unused_locks_for_replay(ns);
2551 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2553 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2554 list_del_init(&lock->l_pending_chain);
2556 LDLM_LOCK_RELEASE(lock);
2557 continue; /* or try to do the rest? */
2559 rc = replay_one_lock(imp, lock);
2560 LDLM_LOCK_RELEASE(lock);
2563 atomic_dec(&imp->imp_replay_inflight);