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
99 int ldlm_request_bufsize(int count, int type)
101 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;
126 "lock timed out (enqueued at %lld, %llds ago); 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);
145 "lock timed out (enqueued at %lld, %llds ago), entering recovery for %s@%s",
146 (s64)lock->l_activity,
147 (s64)(ktime_get_real_seconds() - lock->l_activity),
148 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
153 int is_granted_or_cancelled_nolock(struct ldlm_lock *lock)
157 check_res_locked(lock->l_resource);
158 if (ldlm_is_granted(lock) && !ldlm_is_cp_reqd(lock))
160 else if (ldlm_is_failed(lock) || ldlm_is_cancel(lock))
164 EXPORT_SYMBOL(is_granted_or_cancelled_nolock);
167 * Calculate the Completion timeout (covering enqueue, BL AST, data flush,
168 * lock cancel, and their replies). Used for lock completion timeout on the
171 * \param[in] lock lock which is waiting the completion callback
173 * \retval timeout in seconds to wait for the server reply
176 * We use the same basis for both server side and client side functions
177 * from a single node.
179 static time64_t ldlm_cp_timeout(struct ldlm_lock *lock)
187 * Wait a long time for enqueue - server may have to callback a
188 * lock from another client. Server will evict the other client if it
189 * doesn't respond reasonably, and then give us the lock.
191 timeout = at_get(ldlm_lock_to_ns_at(lock));
192 return max(3 * timeout, (time64_t) ldlm_enqueue_min);
196 * Helper function for ldlm_completion_ast(), updating timings when lock is
199 static int ldlm_completion_tail(struct ldlm_lock *lock, void *data)
204 if (ldlm_is_destroyed(lock) || ldlm_is_failed(lock)) {
205 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
207 } else if (data == NULL) {
208 LDLM_DEBUG(lock, "client-side enqueue: granted");
210 /* Take into AT only CP RPC, not immediately granted locks */
211 delay = ktime_get_real_seconds() - lock->l_activity;
212 LDLM_DEBUG(lock, "client-side enqueue: granted after %llds",
215 /* Update our time estimate */
216 at_measured(ldlm_lock_to_ns_at(lock), delay);
222 * Implementation of ->l_completion_ast() for a client, that doesn't wait
223 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
224 * other threads that cannot block for long.
226 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
230 if (flags == LDLM_FL_WAIT_NOREPROC) {
231 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
235 if (!(flags & LDLM_FL_BLOCKED_MASK)) {
236 wake_up(&lock->l_waitq);
237 RETURN(ldlm_completion_tail(lock, data));
241 "client-side enqueue returned a blocked lock, going forward");
242 ldlm_reprocess_all(lock->l_resource, NULL);
245 EXPORT_SYMBOL(ldlm_completion_ast_async);
248 * Generic LDLM "completion" AST. This is called in several cases:
250 * - when a reply to an ENQUEUE RPC is received from the server
251 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
252 * this point (determined by flags);
254 * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
257 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
260 * - to force all locks when resource is destroyed (cleanup_resource());
262 * If lock is not granted in the first case, this function waits until second
263 * or penultimate cases happen in some other thread.
266 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
268 /* XXX ALLOCATE - 160 bytes */
269 struct lock_wait_data lwd;
270 struct obd_device *obd;
271 struct obd_import *imp = NULL;
272 struct l_wait_info lwi;
278 if (flags == LDLM_FL_WAIT_NOREPROC) {
279 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
283 if (!(flags & LDLM_FL_BLOCKED_MASK)) {
284 wake_up(&lock->l_waitq);
288 LDLM_DEBUG(lock, "client-side enqueue returned a blocked locksleeping");
292 obd = class_exp2obd(lock->l_conn_export);
294 /* if this is a local lock, then there is no import */
296 imp = obd->u.cli.cl_import;
298 timeout = ldlm_cp_timeout(lock);
301 lock->l_activity = ktime_get_real_seconds();
303 if (ldlm_is_no_timeout(lock)) {
304 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
305 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
307 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
308 ldlm_expired_completion_wait,
309 interrupted_completion_wait, &lwd);
313 spin_lock(&imp->imp_lock);
314 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
315 spin_unlock(&imp->imp_lock);
318 if (ns_is_client(ldlm_lock_to_ns(lock)) &&
319 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
320 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
321 ldlm_set_fail_loc(lock);
324 /* Go to sleep until the lock is granted or cancelled. */
325 rc = l_wait_event(lock->l_waitq,
326 is_granted_or_cancelled(lock), &lwi);
330 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
335 RETURN(ldlm_completion_tail(lock, data));
337 EXPORT_SYMBOL(ldlm_completion_ast);
340 * A helper to build a blocking AST function
342 * Perform a common operation for blocking ASTs:
343 * defferred lock cancellation.
345 * \param lock the lock blocking or canceling AST was called on
347 * \see mdt_blocking_ast
348 * \see ldlm_blocking_ast
350 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
356 ldlm_set_cbpending(lock);
357 do_ast = (!lock->l_readers && !lock->l_writers);
358 unlock_res_and_lock(lock);
361 struct lustre_handle lockh;
364 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
365 ldlm_lock2handle(lock, &lockh);
366 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
368 CERROR("ldlm_cli_cancel: %d\n", rc);
371 "Lock still has references, will be cancelled later");
375 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
378 * Server blocking AST
380 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
383 * \param lock the lock which blocks a request or cancelling lock
386 * \param flag indicates whether this cancelling or blocking callback
388 * \see ldlm_blocking_ast_nocheck
390 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
391 void *data, int flag)
395 if (flag == LDLM_CB_CANCELING) {
396 /* Don't need to do anything here. */
400 lock_res_and_lock(lock);
402 * Get this: if ldlm_blocking_ast is racing with intent_policy, such
403 * that ldlm_blocking_ast is called just before intent_policy method
404 * takes the lr_lock, then by the time we get the lock, we might not
405 * be the correct blocking function anymore. So check, and return
408 if (lock->l_blocking_ast != ldlm_blocking_ast) {
409 unlock_res_and_lock(lock);
412 RETURN(ldlm_blocking_ast_nocheck(lock));
414 EXPORT_SYMBOL(ldlm_blocking_ast);
417 * Implements ldlm_lock::l_glimpse_ast for extent locks acquired on the server.
419 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for that is
420 * rather subtle: with OST-side locking, it may so happen that _all_ extent
421 * locks are held by the OST. If client wants to obtain the current file size
422 * it calls ll_glimpse_size(), and (as all locks are held only on the server),
423 * this dummy glimpse callback fires and does nothing. The client still
424 * receives the correct file size due to the following fragment of code in
425 * ldlm_cb_interpret():
427 * if (rc == -ELDLM_NO_LOCK_DATA) {
428 * LDLM_DEBUG(lock, "lost race - client has a lock but no"
430 * ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
433 * That is, after the glimpse returns this error, ofd_lvbo_update() is called
434 * and returns the updated file attributes from the inode to the client.
436 * See also comment in ofd_intent_policy() on why servers must set a non-NULL
437 * l_glimpse_ast when grabbing DLM locks. Otherwise, the server will assume
438 * that the object is in the process of being destroyed.
440 * \param[in] lock DLM lock being glimpsed, unused
441 * \param[in] reqp pointer to ptlrpc_request, unused
443 * \retval -ELDLM_NO_LOCK_DATA to get attributes from disk object
445 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
447 return -ELDLM_NO_LOCK_DATA;
451 * Enqueue a local lock (typically on a server).
453 int ldlm_cli_enqueue_local(const struct lu_env *env,
454 struct ldlm_namespace *ns,
455 const struct ldlm_res_id *res_id,
456 enum ldlm_type type, union ldlm_policy_data *policy,
457 enum ldlm_mode mode, __u64 *flags,
458 ldlm_blocking_callback blocking,
459 ldlm_completion_callback completion,
460 ldlm_glimpse_callback glimpse,
461 void *data, __u32 lvb_len, enum lvb_type lvb_type,
462 const __u64 *client_cookie,
463 struct lustre_handle *lockh)
465 struct ldlm_lock *lock;
467 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
468 .lcs_blocking = blocking,
469 .lcs_glimpse = glimpse,
474 LASSERT(!(*flags & LDLM_FL_REPLAY));
475 if (unlikely(ns_is_client(ns))) {
476 CERROR("Trying to enqueue local lock in a shadow namespace\n");
480 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
483 GOTO(out_nolock, err = PTR_ERR(lock));
485 err = ldlm_lvbo_init(lock->l_resource);
487 LDLM_ERROR(lock, "delayed lvb init failed (rc %d)", err);
488 ldlm_lock_destroy_nolock(lock);
492 ldlm_lock2handle(lock, lockh);
495 * NB: we don't have any lock now (lock_res_and_lock)
496 * because it's a new lock
498 ldlm_lock_addref_internal_nolock(lock, mode);
499 ldlm_set_local(lock);
500 if (*flags & LDLM_FL_ATOMIC_CB)
501 ldlm_set_atomic_cb(lock);
503 if (*flags & LDLM_FL_CANCEL_ON_BLOCK)
504 ldlm_set_cancel_on_block(lock);
507 lock->l_policy_data = *policy;
508 if (client_cookie != NULL)
509 lock->l_client_cookie = *client_cookie;
510 if (type == LDLM_EXTENT) {
511 /* extent lock without policy is a bug */
515 lock->l_req_extent = policy->l_extent;
518 err = ldlm_lock_enqueue(env, ns, &lock, policy, flags);
519 if (unlikely(err != ELDLM_OK))
523 *policy = lock->l_policy_data;
525 if (lock->l_completion_ast)
526 lock->l_completion_ast(lock, *flags, NULL);
528 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
531 LDLM_LOCK_RELEASE(lock);
535 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
537 static void failed_lock_cleanup(struct ldlm_namespace *ns,
538 struct ldlm_lock *lock, int mode)
542 /* Set a flag to prevent us from sending a CANCEL (b=407) */
543 lock_res_and_lock(lock);
544 /* Check that lock is not granted or failed, we might race. */
545 if (!ldlm_is_granted(lock) && !ldlm_is_failed(lock)) {
547 * Make sure that this lock will not be found by raced
548 * bl_ast and -EINVAL reply is sent to server anyways.
551 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
552 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
555 unlock_res_and_lock(lock);
559 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
561 LDLM_DEBUG(lock, "lock was granted or failed in race");
564 * XXX - HACK because we shouldn't call ldlm_lock_destroy()
565 * from llite/file.c/ll_file_flock().
568 * This code makes for the fact that we do not have blocking handler on
569 * a client for flock locks. As such this is the place where we must
570 * completely kill failed locks. (interrupted and those that
571 * were waiting to be granted when server evicted us.
573 if (lock->l_resource->lr_type == LDLM_FLOCK) {
574 lock_res_and_lock(lock);
575 if (!ldlm_is_destroyed(lock)) {
576 ldlm_resource_unlink_lock(lock);
577 ldlm_lock_decref_internal_nolock(lock, mode);
578 ldlm_lock_destroy_nolock(lock);
580 unlock_res_and_lock(lock);
582 ldlm_lock_decref_internal(lock, mode);
587 * Finishing portion of client lock enqueue code.
589 * Called after receiving reply from server.
591 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
592 enum ldlm_type type, __u8 with_policy,
593 enum ldlm_mode mode, __u64 *flags, void *lvb,
594 __u32 lvb_len, const struct lustre_handle *lockh,
597 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
598 const struct lu_env *env = NULL;
599 int is_replay = *flags & LDLM_FL_REPLAY;
600 struct ldlm_lock *lock;
601 struct ldlm_reply *reply;
602 int cleanup_phase = 1;
606 if (req && req->rq_svc_thread)
607 env = req->rq_svc_thread->t_env;
609 lock = ldlm_handle2lock(lockh);
610 /* ldlm_cli_enqueue is holding a reference on this lock. */
612 LASSERT(type == LDLM_FLOCK);
616 LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
617 "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
619 if (rc != ELDLM_OK) {
621 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
622 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
624 if (rc != ELDLM_LOCK_ABORTED)
628 /* Before we return, swab the reply */
629 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
631 GOTO(cleanup, rc = -EPROTO);
636 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
639 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
640 GOTO(cleanup, rc = size);
641 } else if (unlikely(size > lvb_len)) {
643 "Replied LVB is larger than expectation, expected = %d, replied = %d",
645 GOTO(cleanup, rc = -EINVAL);
650 if (rc == ELDLM_LOCK_ABORTED) {
651 if (lvb_len > 0 && lvb != NULL)
652 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
654 GOTO(cleanup, rc = rc ? : ELDLM_LOCK_ABORTED);
657 /* lock enqueued on the server */
660 lock_res_and_lock(lock);
661 /* Key change rehash lock in per-export hash with new key */
662 if (exp->exp_lock_hash) {
664 * In the function below, .hs_keycmp resolves to
665 * ldlm_export_lock_keycmp()
667 /* coverity[overrun-buffer-val] */
668 cfs_hash_rehash_key(exp->exp_lock_hash,
669 &lock->l_remote_handle,
673 lock->l_remote_handle = reply->lock_handle;
676 *flags = ldlm_flags_from_wire(reply->lock_flags);
677 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
678 LDLM_FL_INHERIT_MASK);
679 unlock_res_and_lock(lock);
681 CDEBUG(D_INFO, "local: %p, remote cookie: %#llx, flags: %#llx\n",
682 lock, reply->lock_handle.cookie, *flags);
685 * If enqueue returned a blocked lock but the completion handler has
686 * already run, then it fixed up the resource and we don't need to do it
689 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
690 int newmode = reply->lock_desc.l_req_mode;
693 if (newmode && newmode != lock->l_req_mode) {
694 LDLM_DEBUG(lock, "server returned different mode %s",
695 ldlm_lockname[newmode]);
696 lock->l_req_mode = newmode;
699 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
700 &lock->l_resource->lr_name)) {
702 "remote intent success, locking "DLDLMRES", instead of "DLDLMRES"\n",
703 PLDLMRES(&reply->lock_desc.l_resource),
704 PLDLMRES(lock->l_resource));
706 rc = ldlm_lock_change_resource(ns, lock,
707 &reply->lock_desc.l_resource.lr_name);
708 if (rc || lock->l_resource == NULL)
709 GOTO(cleanup, rc = -ENOMEM);
710 LDLM_DEBUG(lock, "client-side enqueue, new resource");
714 /* We assume lock type cannot change on server*/
715 ldlm_convert_policy_to_local(exp,
716 lock->l_resource->lr_type,
717 &reply->lock_desc.l_policy_data,
718 &lock->l_policy_data);
721 if (type != LDLM_PLAIN)
723 "client-side enqueue, new policy data");
726 if ((*flags) & LDLM_FL_AST_SENT) {
727 lock_res_and_lock(lock);
728 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
729 unlock_res_and_lock(lock);
730 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
734 * If the lock has already been granted by a completion AST, don't
735 * clobber the LVB with an older one.
739 * We must lock or a racing completion might update lvb without
740 * letting us know and we'll clobber the correct value.
741 * Cannot unlock after the check either, a that still leaves
742 * a tiny window for completion to get in
744 lock_res_and_lock(lock);
745 if (!ldlm_is_granted(lock))
746 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
747 lock->l_lvb_data, lvb_len);
748 unlock_res_and_lock(lock);
756 rc = ldlm_lock_enqueue(env, ns, &lock, NULL, flags);
757 if (lock->l_completion_ast != NULL) {
758 int err = lock->l_completion_ast(lock, *flags, NULL);
767 if (lvb_len > 0 && lvb != NULL) {
769 * Copy the LVB here, and not earlier, because the completion
770 * AST (if any) can override what we got in the reply
772 memcpy(lvb, lock->l_lvb_data, lvb_len);
775 LDLM_DEBUG(lock, "client-side enqueue END");
778 if (cleanup_phase == 1 && rc)
779 failed_lock_cleanup(ns, lock, mode);
780 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
782 LDLM_LOCK_RELEASE(lock);
785 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
788 * Estimate number of lock handles that would fit into request of given
789 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
790 * a single page on the send/receive side. XXX: 512 should be changed to
791 * more adequate value.
793 static inline int ldlm_req_handles_avail(int req_size, int off)
797 avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512) - req_size;
798 if (likely(avail >= 0))
799 avail /= (int)sizeof(struct lustre_handle);
802 avail += LDLM_LOCKREQ_HANDLES - off;
807 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
808 enum req_location loc,
811 __u32 size = req_capsule_msg_size(pill, loc);
813 return ldlm_req_handles_avail(size, off);
816 static inline int ldlm_format_handles_avail(struct obd_import *imp,
817 const struct req_format *fmt,
818 enum req_location loc, int off)
820 __u32 size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
822 return ldlm_req_handles_avail(size, off);
826 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
827 * \a count locks in \a cancels.
829 * This is to be called by functions preparing their own requests that
830 * might contain lists of locks to cancel in addition to actual operation
831 * that needs to be performed.
833 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
834 int version, int opc, int canceloff,
835 struct list_head *cancels, int count)
837 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
838 struct req_capsule *pill = &req->rq_pill;
839 struct ldlm_request *dlm = NULL;
840 struct list_head head = LIST_HEAD_INIT(head);
841 enum ldlm_lru_flags lru_flags;
842 int avail, to_free, pack = 0;
849 if (ns_connect_cancelset(ns)) {
850 /* Estimate the amount of available space in the request. */
851 req_capsule_filled_sizes(pill, RCL_CLIENT);
852 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
854 lru_flags = LDLM_LRU_FLAG_NO_WAIT | (ns_connect_lru_resize(ns) ?
855 LDLM_LRU_FLAG_LRUR : LDLM_LRU_FLAG_AGED);
856 to_free = !ns_connect_lru_resize(ns) &&
857 opc == LDLM_ENQUEUE ? 1 : 0;
860 * Cancel LRU locks here _only_ if the server supports
861 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
862 * RPC, which will make us slower.
865 count += ldlm_cancel_lru_local(ns, cancels, to_free,
872 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
873 ldlm_request_bufsize(pack, opc));
876 rc = ptlrpc_request_pack(req, version, opc);
878 ldlm_lock_list_put(cancels, l_bl_ast, count);
882 if (ns_connect_cancelset(ns)) {
884 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
887 * Skip first lock handler in ldlm_request_pack(),
888 * this method will increment @lock_count according
889 * to the lock handle amount actually written to
892 dlm->lock_count = canceloff;
894 /* Pack into the request @pack lock handles. */
895 ldlm_cli_cancel_list(cancels, pack, req, 0);
896 /* Prepare and send separate cancel RPC for others. */
897 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
899 ldlm_lock_list_put(cancels, l_bl_ast, count);
903 EXPORT_SYMBOL(ldlm_prep_elc_req);
905 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
906 struct list_head *cancels, int count)
908 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
909 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
911 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
913 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
915 struct ptlrpc_request *req;
920 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
922 RETURN(ERR_PTR(-ENOMEM));
924 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
926 ptlrpc_request_free(req);
930 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
931 ptlrpc_request_set_replen(req);
934 EXPORT_SYMBOL(ldlm_enqueue_pack);
937 * Client-side lock enqueue.
939 * If a request has some specific initialisation it is passed in \a reqp,
940 * otherwise it is created in ldlm_cli_enqueue.
942 * Supports sync and async requests, pass \a async flag accordingly. If a
943 * request was created in ldlm_cli_enqueue and it is the async request,
944 * pass it to the caller in \a reqp.
946 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
947 struct ldlm_enqueue_info *einfo,
948 const struct ldlm_res_id *res_id,
949 union ldlm_policy_data const *policy, __u64 *flags,
950 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
951 struct lustre_handle *lockh, int async)
953 struct ldlm_namespace *ns;
954 struct ldlm_lock *lock;
955 struct ldlm_request *body;
956 int is_replay = *flags & LDLM_FL_REPLAY;
957 int req_passed_in = 1;
959 struct ptlrpc_request *req;
963 LASSERT(exp != NULL);
965 ns = exp->exp_obd->obd_namespace;
968 * If we're replaying this lock, just check some invariants.
969 * If we're creating a new lock, get everything all setup nice.
972 lock = ldlm_handle2lock_long(lockh, 0);
973 LASSERT(lock != NULL);
974 LDLM_DEBUG(lock, "client-side enqueue START");
975 LASSERT(exp == lock->l_conn_export);
977 const struct ldlm_callback_suite cbs = {
978 .lcs_completion = einfo->ei_cb_cp,
979 .lcs_blocking = einfo->ei_cb_bl,
980 .lcs_glimpse = einfo->ei_cb_gl
982 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
983 einfo->ei_mode, &cbs, einfo->ei_cbdata,
986 RETURN(PTR_ERR(lock));
988 if (einfo->ei_cb_created)
989 einfo->ei_cb_created(lock);
991 /* for the local lock, add the reference */
992 ldlm_lock_addref_internal(lock, einfo->ei_mode);
993 ldlm_lock2handle(lock, lockh);
995 lock->l_policy_data = *policy;
997 if (einfo->ei_type == LDLM_EXTENT) {
998 /* extent lock without policy is a bug */
1002 lock->l_req_extent = policy->l_extent;
1004 LDLM_DEBUG(lock, "client-side enqueue START, flags %#llx",
1008 lock->l_conn_export = exp;
1009 lock->l_export = NULL;
1010 lock->l_blocking_ast = einfo->ei_cb_bl;
1011 lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL |
1012 LDLM_FL_ATOMIC_CB));
1013 lock->l_activity = ktime_get_real_seconds();
1015 /* lock not sent to server yet */
1016 if (reqp == NULL || *reqp == NULL) {
1017 req = ldlm_enqueue_pack(exp, lvb_len);
1019 failed_lock_cleanup(ns, lock, einfo->ei_mode);
1020 LDLM_LOCK_RELEASE(lock);
1021 RETURN(PTR_ERR(req));
1031 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
1033 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
1034 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
1037 if (*flags & LDLM_FL_NDELAY) {
1038 DEBUG_REQ(D_DLMTRACE, req, "enqueue lock with no delay");
1039 req->rq_no_resend = req->rq_no_delay = 1;
1041 * probably set a shorter timeout value and handle ETIMEDOUT
1042 * in osc_lock_upcall() correctly
1044 /* lustre_msg_set_timeout(req, req->rq_timeout / 2); */
1047 /* Dump lock data into the request buffer */
1048 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1049 ldlm_lock2desc(lock, &body->lock_desc);
1050 body->lock_flags = ldlm_flags_to_wire(*flags);
1051 body->lock_handle[0] = *lockh;
1053 /* extended LDLM opcodes in client stats */
1054 if (exp->exp_obd->obd_svc_stats != NULL) {
1055 bool glimpse = *flags & LDLM_FL_HAS_INTENT;
1057 /* OST glimpse has no intent buffer */
1058 if (req_capsule_has_field(&req->rq_pill, &RMF_LDLM_INTENT,
1060 struct ldlm_intent *it;
1062 it = req_capsule_client_get(&req->rq_pill,
1064 glimpse = (it && (it->opc == IT_GLIMPSE));
1068 ldlm_svc_get_eopc(body, exp->exp_obd->obd_svc_stats);
1070 lprocfs_counter_incr(exp->exp_obd->obd_svc_stats,
1072 LDLM_GLIMPSE_ENQUEUE);
1076 LASSERT(reqp != NULL);
1080 LDLM_DEBUG(lock, "sending request");
1082 rc = ptlrpc_queue_wait(req);
1084 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
1085 einfo->ei_mode, flags, lvb, lvb_len,
1089 * If ldlm_cli_enqueue_fini did not find the lock, we need to free
1090 * one reference that we took
1093 LDLM_LOCK_RELEASE(lock);
1097 if (!req_passed_in && req != NULL) {
1098 ptlrpc_req_finished(req);
1105 EXPORT_SYMBOL(ldlm_cli_enqueue);
1108 * Client-side lock convert reply handling.
1110 * Finish client lock converting, checks for concurrent converts
1111 * and clear 'converting' flag so lock can be placed back into LRU.
1113 static int lock_convert_interpret(const struct lu_env *env,
1114 struct ptlrpc_request *req,
1117 struct ldlm_async_args *aa = args;
1118 struct ldlm_lock *lock;
1119 struct ldlm_reply *reply;
1123 lock = ldlm_handle2lock(&aa->lock_handle);
1125 LDLM_DEBUG_NOLOCK("convert ACK for unknown local cookie %#llx",
1126 aa->lock_handle.cookie);
1130 LDLM_DEBUG(lock, "CONVERTED lock:");
1135 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1137 GOTO(out, rc = -EPROTO);
1139 if (reply->lock_handle.cookie != aa->lock_handle.cookie) {
1141 "convert ACK with wrong lock cookie %#llx but cookie %#llx from server %s id %s\n",
1142 aa->lock_handle.cookie, reply->lock_handle.cookie,
1143 req->rq_export->exp_client_uuid.uuid,
1144 libcfs_id2str(req->rq_peer));
1145 GOTO(out, rc = ELDLM_NO_LOCK_DATA);
1148 lock_res_and_lock(lock);
1150 * Lock convert is sent for any new bits to drop, the converting flag
1151 * is dropped when ibits on server are the same as on client. Meanwhile
1152 * that can be so that more later convert will be replied first with
1153 * and clear converting flag, so in case of such race just exit here.
1154 * if lock has no converting bits then
1156 if (!ldlm_is_converting(lock)) {
1158 "convert ACK for lock without converting flag, reply ibits %#llx",
1159 reply->lock_desc.l_policy_data.l_inodebits.bits);
1160 } else if (reply->lock_desc.l_policy_data.l_inodebits.bits !=
1161 lock->l_policy_data.l_inodebits.bits) {
1163 * Compare server returned lock ibits and local lock ibits
1164 * if they are the same we consider convertion is done,
1165 * otherwise we have more converts inflight and keep
1168 LDLM_DEBUG(lock, "convert ACK with ibits %#llx\n",
1169 reply->lock_desc.l_policy_data.l_inodebits.bits);
1171 ldlm_clear_converting(lock);
1174 * Concurrent BL AST may arrive and cause another convert
1175 * or cancel so just do nothing here if bl_ast is set,
1176 * finish with convert otherwise.
1178 if (!ldlm_is_bl_ast(lock)) {
1179 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
1182 * Drop cancel_bits since there are no more converts
1183 * and put lock into LRU if it is still not used and
1186 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1187 if (!lock->l_readers && !lock->l_writers &&
1188 !ldlm_is_canceling(lock)) {
1189 spin_lock(&ns->ns_lock);
1190 /* there is check for list_empty() inside */
1191 ldlm_lock_remove_from_lru_nolock(lock);
1192 ldlm_lock_add_to_lru_nolock(lock);
1193 spin_unlock(&ns->ns_lock);
1197 unlock_res_and_lock(lock);
1202 lock_res_and_lock(lock);
1203 if (ldlm_is_converting(lock)) {
1204 ldlm_clear_converting(lock);
1205 ldlm_set_cbpending(lock);
1206 ldlm_set_bl_ast(lock);
1207 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1209 unlock_res_and_lock(lock);
1212 * fallback to normal lock cancel. If rc means there is no
1213 * valid lock on server, do only local cancel
1215 if (rc == ELDLM_NO_LOCK_DATA)
1220 rc = ldlm_cli_cancel(&aa->lock_handle, flag);
1222 LDLM_DEBUG(lock, "failed to cancel lock: rc = %d\n",
1225 LDLM_LOCK_PUT(lock);
1230 * Client-side IBITS lock convert.
1232 * Inform server that lock has been converted instead of canceling.
1233 * Server finishes convert on own side and does reprocess to grant
1234 * all related waiting locks.
1236 * Since convert means only ibits downgrading, client doesn't need to
1237 * wait for server reply to finish local converting process so this request
1238 * is made asynchronous.
1241 int ldlm_cli_convert(struct ldlm_lock *lock, __u32 *flags)
1243 struct ldlm_request *body;
1244 struct ptlrpc_request *req;
1245 struct ldlm_async_args *aa;
1246 struct obd_export *exp = lock->l_conn_export;
1251 LDLM_ERROR(lock, "convert must not be called on local locks.");
1256 * this is better to check earlier and it is done so already,
1257 * but this check is kept too as final one to issue an error
1258 * if any new code will miss such check.
1260 if (!exp_connect_lock_convert(exp)) {
1261 LDLM_ERROR(lock, "server doesn't support lock convert\n");
1265 if (lock->l_resource->lr_type != LDLM_IBITS) {
1266 LDLM_ERROR(lock, "convert works with IBITS locks only.");
1270 LDLM_DEBUG(lock, "client-side convert");
1272 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1273 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1278 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1279 body->lock_handle[0] = lock->l_remote_handle;
1281 body->lock_desc.l_req_mode = lock->l_req_mode;
1282 body->lock_desc.l_granted_mode = lock->l_granted_mode;
1284 body->lock_desc.l_policy_data.l_inodebits.bits =
1285 lock->l_policy_data.l_inodebits.bits;
1286 body->lock_desc.l_policy_data.l_inodebits.cancel_bits = 0;
1288 body->lock_flags = ldlm_flags_to_wire(*flags);
1289 body->lock_count = 1;
1291 ptlrpc_request_set_replen(req);
1294 * Use cancel portals for convert as well as high-priority handling.
1296 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1297 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1299 ptlrpc_at_set_req_timeout(req);
1301 if (exp->exp_obd->obd_svc_stats != NULL)
1302 lprocfs_counter_incr(exp->exp_obd->obd_svc_stats,
1303 LDLM_CONVERT - LDLM_FIRST_OPC);
1305 aa = ptlrpc_req_async_args(aa, req);
1306 ldlm_lock2handle(lock, &aa->lock_handle);
1307 req->rq_interpret_reply = lock_convert_interpret;
1309 ptlrpcd_add_req(req);
1314 * Cancel locks locally.
1316 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1317 * \retval LDLM_FL_CANCELING otherwise;
1318 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1320 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1322 __u64 rc = LDLM_FL_LOCAL_ONLY;
1326 if (lock->l_conn_export) {
1329 LDLM_DEBUG(lock, "client-side cancel");
1330 /* Set this flag to prevent others from getting new references*/
1331 lock_res_and_lock(lock);
1332 ldlm_set_cbpending(lock);
1333 local_only = !!(lock->l_flags &
1334 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1335 ldlm_cancel_callback(lock);
1336 rc = (ldlm_is_bl_ast(lock)) ?
1337 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1338 unlock_res_and_lock(lock);
1342 "not sending request (at caller's instruction)\n");
1343 rc = LDLM_FL_LOCAL_ONLY;
1345 ldlm_lock_cancel(lock);
1347 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1348 LDLM_ERROR(lock, "Trying to cancel local lock");
1351 LDLM_DEBUG(lock, "server-side local cancel");
1352 ldlm_lock_cancel(lock);
1353 ldlm_reprocess_all(lock->l_resource, lock);
1360 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1362 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1363 struct list_head *head, int count)
1365 struct ldlm_request *dlm;
1366 struct ldlm_lock *lock;
1367 int max, packed = 0;
1371 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1372 LASSERT(dlm != NULL);
1374 /* Check the room in the request buffer. */
1375 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1376 sizeof(struct ldlm_request);
1377 max /= sizeof(struct lustre_handle);
1378 max += LDLM_LOCKREQ_HANDLES;
1379 LASSERT(max >= dlm->lock_count + count);
1382 * XXX: it would be better to pack lock handles grouped by resource.
1383 * so that the server cancel would call filter_lvbo_update() less
1386 list_for_each_entry(lock, head, l_bl_ast) {
1389 LASSERT(lock->l_conn_export);
1390 /* Pack the lock handle to the given request buffer. */
1391 LDLM_DEBUG(lock, "packing");
1392 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1395 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1400 * Prepare and send a batched cancel RPC. It will include \a count lock
1401 * handles of locks given in \a cancels list.
1403 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1404 int count, enum ldlm_cancel_flags flags)
1406 struct ptlrpc_request *req = NULL;
1407 struct obd_import *imp;
1413 LASSERT(exp != NULL);
1416 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1418 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1421 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1422 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1427 imp = class_exp2cliimp(exp);
1428 if (imp == NULL || imp->imp_invalid) {
1430 "skipping cancel on invalid import %p\n", imp);
1434 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1436 GOTO(out, rc = -ENOMEM);
1438 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1439 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1440 ldlm_request_bufsize(count, LDLM_CANCEL));
1442 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1444 ptlrpc_request_free(req);
1449 * If OSP want cancel cross-MDT lock, let's not block it in
1450 * in recovery, otherwise the lock will not released, if
1451 * the remote target is also in recovery, and it also need
1452 * this lock, it might cause deadlock.
1454 if (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS &&
1455 exp->exp_obd->obd_lu_dev != NULL &&
1456 exp->exp_obd->obd_lu_dev->ld_site != NULL) {
1457 struct lu_device *top_dev;
1459 top_dev = exp->exp_obd->obd_lu_dev->ld_site->ls_top_dev;
1460 if (top_dev != NULL &&
1461 top_dev->ld_obd->obd_recovering)
1462 req->rq_allow_replay = 1;
1465 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1466 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1467 ptlrpc_at_set_req_timeout(req);
1469 ldlm_cancel_pack(req, cancels, count);
1471 ptlrpc_request_set_replen(req);
1472 if (flags & LCF_ASYNC) {
1473 ptlrpcd_add_req(req);
1478 rc = ptlrpc_queue_wait(req);
1479 if (rc == LUSTRE_ESTALE) {
1481 "client/server (nid %s) out of sync -- not fatal\n",
1482 libcfs_nid2str(req->rq_import->imp_connection->c_peer.nid));
1484 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1485 req->rq_import_generation == imp->imp_generation) {
1486 ptlrpc_req_finished(req);
1488 } else if (rc != ELDLM_OK) {
1489 /* -ESHUTDOWN is common on umount */
1490 CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1491 "Got rc %d from cancel RPC: canceling anyway\n",
1499 ptlrpc_req_finished(req);
1502 return sent ? sent : rc;
1505 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1507 LASSERT(imp != NULL);
1508 return &imp->imp_obd->obd_namespace->ns_pool;
1512 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1514 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1516 struct obd_device *obd;
1521 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1522 !imp_connect_lru_resize(req->rq_import)))
1523 /* Do nothing for corner cases. */
1527 * In some cases RPC may contain SLV and limit zeroed out. This
1528 * is the case when server does not support LRU resize feature.
1529 * This is also possible in some recovery cases when server-side
1530 * reqs have no reference to the OBD export and thus access to
1531 * server-side namespace is not possible.
1533 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1534 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1535 DEBUG_REQ(D_HA, req,
1536 "Zero SLV or limit found (SLV=%llu, limit=%u)",
1537 lustre_msg_get_slv(req->rq_repmsg),
1538 lustre_msg_get_limit(req->rq_repmsg));
1542 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1543 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1544 obd = req->rq_import->imp_obd;
1547 * Set new SLV and limit in OBD fields to make them accessible
1548 * to the pool thread. We do not access obd_namespace and pool
1549 * directly here as there is no reliable way to make sure that
1550 * they are still alive at cleanup time. Evil races are possible
1551 * which may cause Oops at that time.
1553 write_lock(&obd->obd_pool_lock);
1554 obd->obd_pool_slv = new_slv;
1555 obd->obd_pool_limit = new_limit;
1556 write_unlock(&obd->obd_pool_lock);
1562 * Client side lock cancel.
1564 * Lock must not have any readers or writers by this time.
1566 int ldlm_cli_cancel(const struct lustre_handle *lockh,
1567 enum ldlm_cancel_flags cancel_flags)
1569 struct obd_export *exp;
1570 enum ldlm_lru_flags lru_flags;
1571 int avail, count = 1;
1573 struct ldlm_namespace *ns;
1574 struct ldlm_lock *lock;
1575 struct list_head cancels = LIST_HEAD_INIT(cancels);
1579 lock = ldlm_handle2lock_long(lockh, 0);
1581 LDLM_DEBUG_NOLOCK("lock is already being destroyed");
1585 /* Convert lock bits instead of cancel for IBITS locks */
1586 if (cancel_flags & LCF_CONVERT) {
1587 LASSERT(lock->l_resource->lr_type == LDLM_IBITS);
1588 LASSERT(lock->l_policy_data.l_inodebits.cancel_bits != 0);
1590 rc = ldlm_cli_dropbits(lock,
1591 lock->l_policy_data.l_inodebits.cancel_bits);
1593 LDLM_LOCK_RELEASE(lock);
1598 lock_res_and_lock(lock);
1599 /* Lock is being canceled and the caller doesn't want to wait */
1600 if (ldlm_is_canceling(lock)) {
1601 if (cancel_flags & LCF_ASYNC) {
1602 unlock_res_and_lock(lock);
1604 struct l_wait_info lwi = { 0 };
1606 unlock_res_and_lock(lock);
1607 l_wait_event(lock->l_waitq, is_bl_done(lock), &lwi);
1609 LDLM_LOCK_RELEASE(lock);
1614 * Lock is being converted, cancel it immediately.
1615 * When convert will end, it releases lock and it will be gone.
1617 if (ldlm_is_converting(lock)) {
1618 /* set back flags removed by convert */
1619 ldlm_set_cbpending(lock);
1620 ldlm_set_bl_ast(lock);
1623 ldlm_set_canceling(lock);
1624 unlock_res_and_lock(lock);
1626 if (cancel_flags & LCF_LOCAL)
1627 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_LOCAL_CANCEL_PAUSE,
1630 rc = ldlm_cli_cancel_local(lock);
1631 if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
1632 LDLM_LOCK_RELEASE(lock);
1636 * Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1637 * RPC which goes to canceld portal, so we can cancel other LRU locks
1638 * here and send them all as one LDLM_CANCEL RPC.
1640 LASSERT(list_empty(&lock->l_bl_ast));
1641 list_add(&lock->l_bl_ast, &cancels);
1643 exp = lock->l_conn_export;
1644 if (exp_connect_cancelset(exp)) {
1645 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1650 ns = ldlm_lock_to_ns(lock);
1651 lru_flags = ns_connect_lru_resize(ns) ?
1652 LDLM_LRU_FLAG_LRUR : LDLM_LRU_FLAG_AGED;
1653 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1654 LCF_BL_AST, lru_flags);
1656 ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1659 EXPORT_SYMBOL(ldlm_cli_cancel);
1662 * Locally cancel up to \a count locks in list \a cancels.
1663 * Return the number of cancelled locks.
1665 int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
1666 enum ldlm_cancel_flags cancel_flags)
1668 struct list_head head = LIST_HEAD_INIT(head);
1669 struct ldlm_lock *lock, *next;
1670 int left = 0, bl_ast = 0;
1674 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1678 if (cancel_flags & LCF_LOCAL) {
1679 rc = LDLM_FL_LOCAL_ONLY;
1680 ldlm_lock_cancel(lock);
1682 rc = ldlm_cli_cancel_local(lock);
1685 * Until we have compound requests and can send LDLM_CANCEL
1686 * requests batched with generic RPCs, we need to send cancels
1687 * with the LDLM_FL_BL_AST flag in a separate RPC from
1688 * the one being generated now.
1690 if (!(cancel_flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1691 LDLM_DEBUG(lock, "Cancel lock separately");
1692 list_move(&lock->l_bl_ast, &head);
1696 if (rc == LDLM_FL_LOCAL_ONLY) {
1697 /* CANCEL RPC should not be sent to server. */
1698 list_del_init(&lock->l_bl_ast);
1699 LDLM_LOCK_RELEASE(lock);
1705 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1712 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1713 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1714 * readahead requests, ...)
1716 static enum ldlm_policy_res
1717 ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1718 int unused, int added, int count)
1720 enum ldlm_policy_res result = LDLM_POLICY_CANCEL_LOCK;
1723 * don't check added & count since we want to process all locks
1725 * It's fine to not take lock to access lock->l_resource since
1726 * the lock has already been granted so it won't change.
1728 switch (lock->l_resource->lr_type) {
1731 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) != 0)
1735 result = LDLM_POLICY_SKIP_LOCK;
1743 * Callback function for LRU-resize policy. Decides whether to keep
1744 * \a lock in LRU for current \a LRU size \a unused, added in current
1745 * scan \a added and number of locks to be preferably canceled \a count.
1747 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1749 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1751 static enum ldlm_policy_res ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1752 struct ldlm_lock *lock,
1753 int unused, int added,
1756 ktime_t cur = ktime_get();
1757 struct ldlm_pool *pl = &ns->ns_pool;
1762 * Stop LRU processing when we reach past @count or have checked all
1765 if (count && added >= count)
1766 return LDLM_POLICY_KEEP_LOCK;
1769 * Despite of the LV, It doesn't make sense to keep the lock which
1770 * is unused for ns_max_age time.
1772 if (ktime_after(ktime_get(),
1773 ktime_add(lock->l_last_used, ns->ns_max_age)))
1774 return LDLM_POLICY_CANCEL_LOCK;
1776 slv = ldlm_pool_get_slv(pl);
1777 lvf = ldlm_pool_get_lvf(pl);
1778 la = div_u64(ktime_to_ns(ktime_sub(cur, lock->l_last_used)),
1780 lv = lvf * la * unused;
1782 /* Inform pool about current CLV to see it via debugfs. */
1783 ldlm_pool_set_clv(pl, lv);
1786 * Stop when SLV is not yet come from server or lv is smaller than
1789 if (slv == 0 || lv < slv)
1790 return LDLM_POLICY_KEEP_LOCK;
1792 return LDLM_POLICY_CANCEL_LOCK;
1795 static enum ldlm_policy_res
1796 ldlm_cancel_lrur_no_wait_policy(struct ldlm_namespace *ns,
1797 struct ldlm_lock *lock,
1798 int unused, int added,
1801 enum ldlm_policy_res result;
1803 result = ldlm_cancel_lrur_policy(ns, lock, unused, added, count);
1804 if (result == LDLM_POLICY_KEEP_LOCK)
1807 return ldlm_cancel_no_wait_policy(ns, lock, unused, added, count);
1811 * Callback function for debugfs used policy. Makes decision whether to keep
1812 * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1813 * added and number of locks to be preferably canceled \a count.
1815 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1817 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1819 static enum ldlm_policy_res ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1820 struct ldlm_lock *lock,
1821 int unused, int added,
1825 * Stop LRU processing when we reach past @count or have checked all
1828 return (added >= count) ?
1829 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1833 * Callback function for aged policy. Makes decision whether to keep \a lock in
1834 * LRU for current LRU size \a unused, added in current scan \a added and
1835 * number of locks to be preferably canceled \a count.
1837 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1839 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1841 static enum ldlm_policy_res ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1842 struct ldlm_lock *lock,
1843 int unused, int added,
1846 if ((added >= count) &&
1847 ktime_before(ktime_get(),
1848 ktime_add(lock->l_last_used, ns->ns_max_age)))
1849 return LDLM_POLICY_KEEP_LOCK;
1851 return LDLM_POLICY_CANCEL_LOCK;
1854 static enum ldlm_policy_res
1855 ldlm_cancel_aged_no_wait_policy(struct ldlm_namespace *ns,
1856 struct ldlm_lock *lock,
1857 int unused, int added, int count)
1859 enum ldlm_policy_res result;
1861 result = ldlm_cancel_aged_policy(ns, lock, unused, added, count);
1862 if (result == LDLM_POLICY_KEEP_LOCK)
1865 return ldlm_cancel_no_wait_policy(ns, lock, unused, added, count);
1869 * Callback function for default policy. Makes decision whether to keep \a lock
1870 * in LRU for current LRU size \a unused, added in current scan \a added and
1871 * number of locks to be preferably canceled \a count.
1873 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1875 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1878 enum ldlm_policy_res ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1879 struct ldlm_lock *lock,
1880 int unused, int added,
1884 * Stop LRU processing when we reach past count or have checked all
1887 return (added >= count) ?
1888 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1891 typedef enum ldlm_policy_res
1892 (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1893 int unused, int added, int count);
1895 static ldlm_cancel_lru_policy_t
1896 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, enum ldlm_lru_flags lru_flags)
1898 if (ns_connect_lru_resize(ns)) {
1899 if (lru_flags & LDLM_LRU_FLAG_SHRINK)
1900 /* We kill passed number of old locks. */
1901 return ldlm_cancel_passed_policy;
1902 if (lru_flags & LDLM_LRU_FLAG_LRUR) {
1903 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1904 return ldlm_cancel_lrur_no_wait_policy;
1906 return ldlm_cancel_lrur_policy;
1908 if (lru_flags & LDLM_LRU_FLAG_PASSED)
1909 return ldlm_cancel_passed_policy;
1911 if (lru_flags & LDLM_LRU_FLAG_AGED) {
1912 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1913 return ldlm_cancel_aged_no_wait_policy;
1915 return ldlm_cancel_aged_policy;
1918 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1919 return ldlm_cancel_no_wait_policy;
1921 return ldlm_cancel_default_policy;
1925 * - Free space in LRU for \a count new locks,
1926 * redundant unused locks are canceled locally;
1927 * - also cancel locally unused aged locks;
1928 * - do not cancel more than \a max locks;
1929 * - GET the found locks and add them into the \a cancels list.
1931 * A client lock can be added to the l_bl_ast list only when it is
1932 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1933 * CANCEL. There are the following use cases:
1934 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1935 * ldlm_cli_cancel(), which check and set this flag properly. As any
1936 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1937 * later without any special locking.
1939 * Calling policies for enabled LRU resize:
1940 * ----------------------------------------
1941 * flags & LDLM_LRU_FLAG_LRUR - use LRU resize policy (SLV from server) to
1942 * cancel not more than \a count locks;
1944 * flags & LDLM_LRU_FLAG_PASSED - cancel \a count number of old locks (located
1945 * at the beginning of LRU list);
1947 * flags & LDLM_LRU_FLAG_SHRINK - cancel not more than \a count locks according
1948 * to memory pressre policy function;
1950 * flags & LDLM_LRU_FLAG_AGED - cancel \a count locks according to "aged policy"
1952 * flags & LDLM_LRU_FLAG_NO_WAIT - cancel as many unused locks as possible
1953 * (typically before replaying locks) w/o
1954 * sending any RPCs or waiting for any
1955 * outstanding RPC to complete.
1957 * flags & LDLM_CANCEL_CLEANUP - when cancelling read locks, do not check for
1958 * other read locks covering the same pages, just
1959 * discard those pages.
1961 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns,
1962 struct list_head *cancels, int count, int max,
1963 enum ldlm_lru_flags lru_flags)
1965 ldlm_cancel_lru_policy_t pf;
1967 int no_wait = lru_flags & LDLM_LRU_FLAG_NO_WAIT;
1971 if (!ns_connect_lru_resize(ns))
1972 count += ns->ns_nr_unused - ns->ns_max_unused;
1974 pf = ldlm_cancel_lru_policy(ns, lru_flags);
1975 LASSERT(pf != NULL);
1977 /* For any flags, stop scanning if @max is reached. */
1978 while (!list_empty(&ns->ns_unused_list) && (max == 0 || added < max)) {
1979 struct ldlm_lock *lock;
1980 struct list_head *item, *next;
1981 enum ldlm_policy_res result;
1982 ktime_t last_use = ktime_set(0, 0);
1984 spin_lock(&ns->ns_lock);
1985 item = no_wait ? ns->ns_last_pos : &ns->ns_unused_list;
1986 for (item = item->next, next = item->next;
1987 item != &ns->ns_unused_list;
1988 item = next, next = item->next) {
1989 lock = list_entry(item, struct ldlm_lock, l_lru);
1991 /* No locks which got blocking requests. */
1992 LASSERT(!ldlm_is_bl_ast(lock));
1994 if (!ldlm_is_canceling(lock) &&
1995 !ldlm_is_converting(lock))
1999 * Somebody is already doing CANCEL. No need for this
2000 * lock in LRU, do not traverse it again.
2002 ldlm_lock_remove_from_lru_nolock(lock);
2004 if (item == &ns->ns_unused_list) {
2005 spin_unlock(&ns->ns_lock);
2009 last_use = lock->l_last_used;
2011 LDLM_LOCK_GET(lock);
2012 spin_unlock(&ns->ns_lock);
2013 lu_ref_add(&lock->l_reference, __FUNCTION__, current);
2016 * Pass the lock through the policy filter and see if it
2017 * should stay in LRU.
2019 * Even for shrinker policy we stop scanning if
2020 * we find a lock that should stay in the cache.
2021 * We should take into account lock age anyway
2022 * as a new lock is a valuable resource even if
2023 * it has a low weight.
2025 * That is, for shrinker policy we drop only
2026 * old locks, but additionally choose them by
2027 * their weight. Big extent locks will stay in
2030 result = pf(ns, lock, ns->ns_nr_unused, added, count);
2031 if (result == LDLM_POLICY_KEEP_LOCK) {
2032 lu_ref_del(&lock->l_reference, __func__, current);
2033 LDLM_LOCK_RELEASE(lock);
2037 if (result == LDLM_POLICY_SKIP_LOCK) {
2038 lu_ref_del(&lock->l_reference, __func__, current);
2040 spin_lock(&ns->ns_lock);
2041 if (!list_empty(&lock->l_lru) &&
2042 lock->l_lru.prev == ns->ns_last_pos)
2043 ns->ns_last_pos = &lock->l_lru;
2044 spin_unlock(&ns->ns_lock);
2047 LDLM_LOCK_RELEASE(lock);
2051 lock_res_and_lock(lock);
2052 /* Check flags again under the lock. */
2053 if (ldlm_is_canceling(lock) || ldlm_is_converting(lock) ||
2054 ldlm_lock_remove_from_lru_check(lock, last_use) == 0) {
2056 * Another thread is removing lock from LRU, or
2057 * somebody is already doing CANCEL, or there
2058 * is a blocking request which will send cancel
2059 * by itself, or the lock is no longer unused or
2060 * the lock has been used since the pf() call and
2061 * pages could be put under it.
2063 unlock_res_and_lock(lock);
2064 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
2065 LDLM_LOCK_RELEASE(lock);
2068 LASSERT(!lock->l_readers && !lock->l_writers);
2071 * If we have chosen to cancel this lock voluntarily, we
2072 * better send cancel notification to server, so that it
2073 * frees appropriate state. This might lead to a race
2074 * where while we are doing cancel here, server is also
2075 * silently cancelling this lock.
2077 ldlm_clear_cancel_on_block(lock);
2080 * Setting the CBPENDING flag is a little misleading,
2081 * but prevents an important race; namely, once
2082 * CBPENDING is set, the lock can accumulate no more
2083 * readers/writers. Since readers and writers are
2084 * already zero here, ldlm_lock_decref() won't see
2085 * this flag and call l_blocking_ast
2087 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
2089 if ((lru_flags & LDLM_LRU_FLAG_CLEANUP) &&
2090 (lock->l_resource->lr_type == LDLM_EXTENT ||
2091 ldlm_has_dom(lock)) && lock->l_granted_mode == LCK_PR)
2092 ldlm_set_discard_data(lock);
2095 * We can't re-add to l_lru as it confuses the
2096 * refcounting in ldlm_lock_remove_from_lru() if an AST
2097 * arrives after we drop lr_lock below. We use l_bl_ast
2098 * and can't use l_pending_chain as it is used both on
2099 * server and client nevertheless b=5666 says it is
2100 * used only on server
2102 LASSERT(list_empty(&lock->l_bl_ast));
2103 list_add(&lock->l_bl_ast, cancels);
2104 unlock_res_and_lock(lock);
2105 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
2111 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
2113 enum ldlm_cancel_flags cancel_flags,
2114 enum ldlm_lru_flags lru_flags)
2118 added = ldlm_prepare_lru_list(ns, cancels, count, max, lru_flags);
2122 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
2126 * Cancel at least \a nr locks from given namespace LRU.
2128 * When called with LCF_ASYNC the blocking callback will be handled
2129 * in a thread and this function will return after the thread has been
2130 * asked to call the callback. When called with LCF_ASYNC the blocking
2131 * callback will be performed in this function.
2133 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
2134 enum ldlm_cancel_flags cancel_flags,
2135 enum ldlm_lru_flags lru_flags)
2137 struct list_head cancels = LIST_HEAD_INIT(cancels);
2143 * Just prepare the list of locks, do not actually cancel them yet.
2144 * Locks are cancelled later in a separate thread.
2146 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, lru_flags);
2147 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
2155 * Find and cancel locally unused locks found on resource, matched to the
2156 * given policy, mode. GET the found locks and add them into the \a cancels
2159 int ldlm_cancel_resource_local(struct ldlm_resource *res,
2160 struct list_head *cancels,
2161 union ldlm_policy_data *policy,
2162 enum ldlm_mode mode, __u64 lock_flags,
2163 enum ldlm_cancel_flags cancel_flags,
2166 struct ldlm_lock *lock;
2172 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
2173 if (opaque != NULL && lock->l_ast_data != opaque) {
2174 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
2175 lock->l_ast_data, opaque);
2179 if (lock->l_readers || lock->l_writers)
2183 * If somebody is already doing CANCEL, or blocking AST came,
2184 * or lock is being converted then skip this lock.
2186 if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock) ||
2187 ldlm_is_converting(lock))
2190 if (lockmode_compat(lock->l_granted_mode, mode))
2194 * If policy is given and this is IBITS lock, add to list only
2195 * those locks that match by policy.
2197 if (policy && (lock->l_resource->lr_type == LDLM_IBITS)) {
2198 if (!(lock->l_policy_data.l_inodebits.bits &
2199 policy->l_inodebits.bits))
2201 /* Skip locks with DoM bit if it is not set in policy
2202 * to don't flush data by side-bits. Lock convert will
2203 * drop those bits separately.
2205 if (ldlm_has_dom(lock) &&
2206 !(policy->l_inodebits.bits & MDS_INODELOCK_DOM))
2210 /* See CBPENDING comment in ldlm_cancel_lru */
2211 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
2214 LASSERT(list_empty(&lock->l_bl_ast));
2215 list_add(&lock->l_bl_ast, cancels);
2216 LDLM_LOCK_GET(lock);
2221 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
2223 EXPORT_SYMBOL(ldlm_cancel_resource_local);
2226 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
2228 * If \a req is NULL, send CANCEL request to server with handles of locks
2229 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
2230 * separately per lock.
2231 * If \a req is not NULL, put handles of locks in \a cancels into the request
2232 * buffer at the offset \a off.
2233 * Destroy \a cancels at the end.
2235 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
2236 struct ptlrpc_request *req,
2237 enum ldlm_cancel_flags flags)
2239 struct ldlm_lock *lock;
2244 if (list_empty(cancels) || count == 0)
2248 * XXX: requests (both batched and not) could be sent in parallel.
2249 * Usually it is enough to have just 1 RPC, but it is possible that
2250 * there are too many locks to be cancelled in LRU or on a resource.
2251 * It would also speed up the case when the server does not support
2255 LASSERT(!list_empty(cancels));
2256 lock = list_entry(cancels->next, struct ldlm_lock,
2258 LASSERT(lock->l_conn_export);
2260 if (exp_connect_cancelset(lock->l_conn_export)) {
2263 ldlm_cancel_pack(req, cancels, count);
2265 res = ldlm_cli_cancel_req(lock->l_conn_export,
2269 res = ldlm_cli_cancel_req(lock->l_conn_export,
2274 CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
2275 "ldlm_cli_cancel_list: %d\n", res);
2280 ldlm_lock_list_put(cancels, l_bl_ast, res);
2282 LASSERT(count == 0);
2285 EXPORT_SYMBOL(ldlm_cli_cancel_list);
2288 * Cancel all locks on a resource that have 0 readers/writers.
2290 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
2291 * to notify the server.
2293 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
2294 const struct ldlm_res_id *res_id,
2295 union ldlm_policy_data *policy,
2296 enum ldlm_mode mode,
2297 enum ldlm_cancel_flags flags, void *opaque)
2299 struct ldlm_resource *res;
2300 struct list_head cancels = LIST_HEAD_INIT(cancels);
2306 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2308 /* This is not a problem. */
2309 CDEBUG(D_INFO, "No resource %llu\n", res_id->name[0]);
2313 LDLM_RESOURCE_ADDREF(res);
2314 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
2315 0, flags | LCF_BL_AST, opaque);
2316 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
2318 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
2321 LDLM_RESOURCE_DELREF(res);
2322 ldlm_resource_putref(res);
2325 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2327 struct ldlm_cli_cancel_arg {
2333 ldlm_cli_hash_cancel_unused(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2334 struct hlist_node *hnode, void *arg)
2336 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2337 struct ldlm_cli_cancel_arg *lc = arg;
2339 ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
2340 NULL, LCK_MINMODE, lc->lc_flags,
2342 /* must return 0 for hash iteration */
2347 * Cancel all locks on a namespace (or a specific resource, if given)
2348 * that have 0 readers/writers.
2350 * If flags & LCF_LOCAL, throw the locks away without trying
2351 * to notify the server.
2353 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
2354 const struct ldlm_res_id *res_id,
2355 enum ldlm_cancel_flags flags, void *opaque)
2357 struct ldlm_cli_cancel_arg arg = {
2359 .lc_opaque = opaque,
2367 if (res_id != NULL) {
2368 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
2372 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2373 ldlm_cli_hash_cancel_unused, &arg, 0);
2378 /* Lock iterators. */
2380 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
2383 struct list_head *tmp, *next;
2384 struct ldlm_lock *lock;
2385 int rc = LDLM_ITER_CONTINUE;
2390 RETURN(LDLM_ITER_CONTINUE);
2393 list_for_each_safe(tmp, next, &res->lr_granted) {
2394 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2396 if (iter(lock, closure) == LDLM_ITER_STOP)
2397 GOTO(out, rc = LDLM_ITER_STOP);
2400 list_for_each_safe(tmp, next, &res->lr_waiting) {
2401 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2403 if (iter(lock, closure) == LDLM_ITER_STOP)
2404 GOTO(out, rc = LDLM_ITER_STOP);
2411 struct iter_helper_data {
2412 ldlm_iterator_t iter;
2416 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2418 struct iter_helper_data *helper = closure;
2420 return helper->iter(lock, helper->closure);
2423 static int ldlm_res_iter_helper(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2424 struct hlist_node *hnode, void *arg)
2427 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2429 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2433 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2434 ldlm_iterator_t iter, void *closure)
2437 struct iter_helper_data helper = { .iter = iter, .closure = closure };
2439 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2440 ldlm_res_iter_helper, &helper, 0);
2445 * non-blocking function to manipulate a lock whose cb_data is being put away.
2446 * return 0: find no resource
2447 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2450 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2451 const struct ldlm_res_id *res_id,
2452 ldlm_iterator_t iter, void *data)
2454 struct ldlm_resource *res;
2459 LASSERTF(ns != NULL, "must pass in namespace\n");
2461 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2465 LDLM_RESOURCE_ADDREF(res);
2466 rc = ldlm_resource_foreach(res, iter, data);
2467 LDLM_RESOURCE_DELREF(res);
2468 ldlm_resource_putref(res);
2471 EXPORT_SYMBOL(ldlm_resource_iterate);
2474 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2476 struct list_head *list = closure;
2478 /* we use l_pending_chain here, because it's unused on clients. */
2479 LASSERTF(list_empty(&lock->l_pending_chain),
2480 "lock %p next %p prev %p\n",
2481 lock, &lock->l_pending_chain.next,
2482 &lock->l_pending_chain.prev);
2484 * b=9573: don't replay locks left after eviction, or
2485 * b=17614: locks being actively cancelled. Get a reference
2486 * on a lock so that it does not disapear under us (e.g. due to cancel)
2488 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_BL_DONE))) {
2489 list_add(&lock->l_pending_chain, list);
2490 LDLM_LOCK_GET(lock);
2493 return LDLM_ITER_CONTINUE;
2496 static int replay_lock_interpret(const struct lu_env *env,
2497 struct ptlrpc_request *req, void *args, int rc)
2499 struct ldlm_async_args *aa = args;
2500 struct ldlm_lock *lock;
2501 struct ldlm_reply *reply;
2502 struct obd_export *exp;
2505 atomic_dec(&req->rq_import->imp_replay_inflight);
2509 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2511 GOTO(out, rc = -EPROTO);
2513 lock = ldlm_handle2lock(&aa->lock_handle);
2515 CERROR("received replay ack for unknown local cookie %#llx remote cookie %#llx from server %s id %s\n",
2516 aa->lock_handle.cookie, reply->lock_handle.cookie,
2517 req->rq_export->exp_client_uuid.uuid,
2518 libcfs_id2str(req->rq_peer));
2519 GOTO(out, rc = -ESTALE);
2522 /* Key change rehash lock in per-export hash with new key */
2523 exp = req->rq_export;
2524 if (exp && exp->exp_lock_hash) {
2526 * In the function below, .hs_keycmp resolves to
2527 * ldlm_export_lock_keycmp()
2529 /* coverity[overrun-buffer-val] */
2530 cfs_hash_rehash_key(exp->exp_lock_hash,
2531 &lock->l_remote_handle,
2532 &reply->lock_handle,
2535 lock->l_remote_handle = reply->lock_handle;
2538 LDLM_DEBUG(lock, "replayed lock:");
2539 ptlrpc_import_recovery_state_machine(req->rq_import);
2540 LDLM_LOCK_PUT(lock);
2543 ptlrpc_connect_import(req->rq_import);
2548 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2550 struct ptlrpc_request *req;
2551 struct ldlm_async_args *aa;
2552 struct ldlm_request *body;
2558 /* b=11974: Do not replay a lock which is actively being canceled */
2559 if (ldlm_is_bl_done(lock)) {
2560 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2565 * If this is reply-less callback lock, we cannot replay it, since
2566 * server might have long dropped it, but notification of that event was
2567 * lost by network. (and server granted conflicting lock already)
2569 if (ldlm_is_cancel_on_block(lock)) {
2570 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2571 ldlm_lock_cancel(lock);
2576 * If granted mode matches the requested mode, this lock is granted.
2578 * If we haven't been granted anything and are on a resource list,
2579 * then we're blocked/waiting.
2581 * If we haven't been granted anything and we're NOT on a resource list,
2582 * then we haven't got a reply yet and don't have a known disposition.
2583 * This happens whenever a lock enqueue is the request that triggers
2586 if (ldlm_is_granted(lock))
2587 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2588 else if (!list_empty(&lock->l_res_link))
2589 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2591 flags = LDLM_FL_REPLAY;
2593 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2594 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2598 /* We're part of recovery, so don't wait for it. */
2599 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2601 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2602 ldlm_lock2desc(lock, &body->lock_desc);
2603 body->lock_flags = ldlm_flags_to_wire(flags);
2605 ldlm_lock2handle(lock, &body->lock_handle[0]);
2606 if (lock->l_lvb_len > 0)
2607 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2608 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2610 ptlrpc_request_set_replen(req);
2612 * notify the server we've replayed all requests.
2613 * also, we mark the request to be put on a dedicated
2614 * queue to be processed after all request replayes.
2617 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2619 LDLM_DEBUG(lock, "replaying lock:");
2621 atomic_inc(&req->rq_import->imp_replay_inflight);
2622 aa = ptlrpc_req_async_args(aa, req);
2623 aa->lock_handle = body->lock_handle[0];
2624 req->rq_interpret_reply = replay_lock_interpret;
2625 ptlrpcd_add_req(req);
2631 * Cancel as many unused locks as possible before replay. since we are
2632 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2635 * Called only in recovery before replaying locks. there is no need to
2636 * replay locks that are unused. since the clients may hold thousands of
2637 * cached unused locks, dropping the unused locks can greatly reduce the
2638 * load on the servers at recovery time.
2640 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2643 struct list_head cancels = LIST_HEAD_INIT(cancels);
2646 "Dropping as many unused locks as possible before replay for namespace %s (%d)\n",
2647 ldlm_ns_name(ns), ns->ns_nr_unused);
2650 * We don't need to care whether or not LRU resize is enabled
2651 * because the LDLM_LRU_FLAG_NO_WAIT policy doesn't use the
2654 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2655 LCF_LOCAL, LDLM_LRU_FLAG_NO_WAIT);
2657 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2658 canceled, ldlm_ns_name(ns));
2661 int ldlm_replay_locks(struct obd_import *imp)
2663 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2664 struct list_head list = LIST_HEAD_INIT(list);
2665 struct ldlm_lock *lock, *next;
2670 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2672 /* don't replay locks if import failed recovery */
2673 if (imp->imp_vbr_failed)
2676 /* ensure this doesn't fall to 0 before all have been queued */
2677 atomic_inc(&imp->imp_replay_inflight);
2679 if (ldlm_cancel_unused_locks_before_replay)
2680 ldlm_cancel_unused_locks_for_replay(ns);
2682 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2684 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2685 list_del_init(&lock->l_pending_chain);
2687 LDLM_LOCK_RELEASE(lock);
2688 continue; /* or try to do the rest? */
2690 rc = replay_one_lock(imp, lock);
2691 LDLM_LOCK_RELEASE(lock);
2694 atomic_dec(&imp->imp_replay_inflight);