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_last_activity,
128 (s64)(ktime_get_real_seconds() -
129 lock->l_last_activity));
130 if (ktime_get_seconds() > next_dump) {
131 last_dump = next_dump;
132 next_dump = ktime_get_seconds() + 300;
133 ldlm_namespace_dump(D_DLMTRACE,
134 ldlm_lock_to_ns(lock));
136 libcfs_debug_dumplog();
141 obd = lock->l_conn_export->exp_obd;
142 imp = obd->u.cli.cl_import;
143 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
144 LDLM_ERROR(lock, "lock timed out (enqueued at %lld, %llds ago), entering recovery for %s@%s",
145 (s64)lock->l_last_activity,
146 (s64)(ktime_get_real_seconds() - lock->l_last_activity),
147 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
153 * Calculate the Completion timeout (covering enqueue, BL AST, data flush,
154 * lock cancel, and their replies). Used for lock completion timeout on the
157 * \param[in] lock lock which is waiting the completion callback
159 * \retval timeout in seconds to wait for the server reply
162 /* We use the same basis for both server side and client side functions
163 from a single node. */
164 static time64_t ldlm_cp_timeout(struct ldlm_lock *lock)
171 /* Wait a long time for enqueue - server may have to callback a
172 * lock from another client. Server will evict the other client if it
173 * doesn't respond reasonably, and then give us the lock. */
174 timeout = at_get(ldlm_lock_to_ns_at(lock));
175 return max(3 * timeout, (time64_t) ldlm_enqueue_min);
179 * Helper function for ldlm_completion_ast(), updating timings when lock is
182 static int ldlm_completion_tail(struct ldlm_lock *lock, void *data)
187 if (ldlm_is_destroyed(lock) || ldlm_is_failed(lock)) {
188 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
190 } else if (data == NULL) {
191 LDLM_DEBUG(lock, "client-side enqueue: granted");
193 /* Take into AT only CP RPC, not immediately granted locks */
194 delay = ktime_get_real_seconds() - lock->l_last_activity;
195 LDLM_DEBUG(lock, "client-side enqueue: granted after %llds",
198 /* Update our time estimate */
199 at_measured(ldlm_lock_to_ns_at(lock), delay);
205 * Implementation of ->l_completion_ast() for a client, that doesn't wait
206 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
207 * other threads that cannot block for long.
209 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
213 if (flags == LDLM_FL_WAIT_NOREPROC) {
214 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
218 if (!(flags & LDLM_FL_BLOCKED_MASK)) {
219 wake_up(&lock->l_waitq);
220 RETURN(ldlm_completion_tail(lock, data));
223 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
225 ldlm_reprocess_all(lock->l_resource);
228 EXPORT_SYMBOL(ldlm_completion_ast_async);
231 * Generic LDLM "completion" AST. This is called in several cases:
233 * - when a reply to an ENQUEUE RPC is received from the server
234 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
235 * this point (determined by flags);
237 * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
240 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
243 * - to force all locks when resource is destroyed (cleanup_resource());
245 * 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_last_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(struct ldlm_namespace *ns,
435 const struct ldlm_res_id *res_id,
436 enum ldlm_type type, union ldlm_policy_data *policy,
437 enum ldlm_mode mode, __u64 *flags,
438 ldlm_blocking_callback blocking,
439 ldlm_completion_callback completion,
440 ldlm_glimpse_callback glimpse,
441 void *data, __u32 lvb_len, enum lvb_type lvb_type,
442 const __u64 *client_cookie,
443 struct lustre_handle *lockh)
445 struct ldlm_lock *lock;
447 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
448 .lcs_blocking = blocking,
449 .lcs_glimpse = glimpse,
453 LASSERT(!(*flags & LDLM_FL_REPLAY));
454 if (unlikely(ns_is_client(ns))) {
455 CERROR("Trying to enqueue local lock in a shadow namespace\n");
459 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
462 GOTO(out_nolock, err = PTR_ERR(lock));
464 err = ldlm_lvbo_init(lock->l_resource);
466 LDLM_ERROR(lock, "delayed lvb init failed (rc %d)", err);
467 ldlm_lock_destroy_nolock(lock);
471 ldlm_lock2handle(lock, lockh);
473 /* NB: we don't have any lock now (lock_res_and_lock)
474 * because it's a new lock */
475 ldlm_lock_addref_internal_nolock(lock, mode);
476 ldlm_set_local(lock);
477 if (*flags & LDLM_FL_ATOMIC_CB)
478 ldlm_set_atomic_cb(lock);
481 lock->l_policy_data = *policy;
482 if (client_cookie != NULL)
483 lock->l_client_cookie = *client_cookie;
484 if (type == LDLM_EXTENT) {
485 /* extent lock without policy is a bug */
489 lock->l_req_extent = policy->l_extent;
492 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
493 if (unlikely(err != ELDLM_OK))
497 *policy = lock->l_policy_data;
499 if (lock->l_completion_ast)
500 lock->l_completion_ast(lock, *flags, NULL);
502 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
505 LDLM_LOCK_RELEASE(lock);
509 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
511 static void failed_lock_cleanup(struct ldlm_namespace *ns,
512 struct ldlm_lock *lock, int mode)
516 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
517 lock_res_and_lock(lock);
518 /* Check that lock is not granted or failed, we might race. */
519 if ((lock->l_req_mode != lock->l_granted_mode) &&
520 !ldlm_is_failed(lock)) {
521 /* Make sure that this lock will not be found by raced
522 * bl_ast and -EINVAL reply is sent to server anyways.
524 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
525 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
528 unlock_res_and_lock(lock);
532 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
533 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
535 LDLM_DEBUG(lock, "lock was granted or failed in race");
537 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
538 * from llite/file.c/ll_file_flock(). */
539 /* This code makes for the fact that we do not have blocking handler on
540 * a client for flock locks. As such this is the place where we must
541 * completely kill failed locks. (interrupted and those that
542 * were waiting to be granted when server evicted us. */
543 if (lock->l_resource->lr_type == LDLM_FLOCK) {
544 lock_res_and_lock(lock);
545 if (!ldlm_is_destroyed(lock)) {
546 ldlm_resource_unlink_lock(lock);
547 ldlm_lock_decref_internal_nolock(lock, mode);
548 ldlm_lock_destroy_nolock(lock);
550 unlock_res_and_lock(lock);
552 ldlm_lock_decref_internal(lock, mode);
557 * Finishing portion of client lock enqueue code.
559 * Called after receiving reply from server.
561 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
562 enum ldlm_type type, __u8 with_policy,
563 enum ldlm_mode mode, __u64 *flags, void *lvb,
564 __u32 lvb_len, const struct lustre_handle *lockh,
567 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
568 int is_replay = *flags & LDLM_FL_REPLAY;
569 struct ldlm_lock *lock;
570 struct ldlm_reply *reply;
571 int cleanup_phase = 1;
574 lock = ldlm_handle2lock(lockh);
575 /* ldlm_cli_enqueue is holding a reference on this lock. */
577 LASSERT(type == LDLM_FLOCK);
581 LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
582 "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
584 if (rc != ELDLM_OK) {
586 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
587 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
589 if (rc != ELDLM_LOCK_ABORTED)
593 /* Before we return, swab the reply */
594 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
596 GOTO(cleanup, rc = -EPROTO);
601 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
604 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
605 GOTO(cleanup, rc = size);
606 } else if (unlikely(size > lvb_len)) {
607 LDLM_ERROR(lock, "Replied LVB is larger than "
608 "expectation, expected = %d, replied = %d",
610 GOTO(cleanup, rc = -EINVAL);
615 if (rc == ELDLM_LOCK_ABORTED) {
616 if (lvb_len > 0 && lvb != NULL)
617 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
619 GOTO(cleanup, rc = rc ? : ELDLM_LOCK_ABORTED);
622 /* lock enqueued on the server */
625 lock_res_and_lock(lock);
626 /* Key change rehash lock in per-export hash with new key */
627 if (exp->exp_lock_hash) {
628 /* In the function below, .hs_keycmp resolves to
629 * ldlm_export_lock_keycmp() */
630 /* coverity[overrun-buffer-val] */
631 cfs_hash_rehash_key(exp->exp_lock_hash,
632 &lock->l_remote_handle,
636 lock->l_remote_handle = reply->lock_handle;
639 *flags = ldlm_flags_from_wire(reply->lock_flags);
640 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
641 LDLM_FL_INHERIT_MASK);
642 unlock_res_and_lock(lock);
644 CDEBUG(D_INFO, "local: %p, remote cookie: %#llx, flags: %#llx\n",
645 lock, reply->lock_handle.cookie, *flags);
647 /* If enqueue returned a blocked lock but the completion handler has
648 * already run, then it fixed up the resource and we don't need to do it
650 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
651 int newmode = reply->lock_desc.l_req_mode;
653 if (newmode && newmode != lock->l_req_mode) {
654 LDLM_DEBUG(lock, "server returned different mode %s",
655 ldlm_lockname[newmode]);
656 lock->l_req_mode = newmode;
659 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
660 &lock->l_resource->lr_name)) {
661 CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES
662 " instead of "DLDLMRES"\n",
663 PLDLMRES(&reply->lock_desc.l_resource),
664 PLDLMRES(lock->l_resource));
666 rc = ldlm_lock_change_resource(ns, lock,
667 &reply->lock_desc.l_resource.lr_name);
668 if (rc || lock->l_resource == NULL)
669 GOTO(cleanup, rc = -ENOMEM);
670 LDLM_DEBUG(lock, "client-side enqueue, new resource");
674 /* We assume lock type cannot change on server*/
675 ldlm_convert_policy_to_local(exp,
676 lock->l_resource->lr_type,
677 &reply->lock_desc.l_policy_data,
678 &lock->l_policy_data);
681 if (type != LDLM_PLAIN)
682 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
685 if ((*flags) & LDLM_FL_AST_SENT) {
686 lock_res_and_lock(lock);
687 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
688 unlock_res_and_lock(lock);
689 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
692 /* If the lock has already been granted by a completion AST, don't
693 * clobber the LVB with an older one. */
695 /* We must lock or a racing completion might update lvb without
696 * letting us know and we'll clobber the correct value.
697 * Cannot unlock after the check either, a that still leaves
698 * a tiny window for completion to get in */
699 lock_res_and_lock(lock);
700 if (lock->l_req_mode != lock->l_granted_mode)
701 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
702 lock->l_lvb_data, lvb_len);
703 unlock_res_and_lock(lock);
711 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
712 if (lock->l_completion_ast != NULL) {
713 int err = lock->l_completion_ast(lock, *flags, NULL);
721 if (lvb_len > 0 && lvb != NULL) {
722 /* Copy the LVB here, and not earlier, because the completion
723 * AST (if any) can override what we got in the reply */
724 memcpy(lvb, lock->l_lvb_data, lvb_len);
727 LDLM_DEBUG(lock, "client-side enqueue END");
730 if (cleanup_phase == 1 && rc)
731 failed_lock_cleanup(ns, lock, mode);
732 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
734 LDLM_LOCK_RELEASE(lock);
737 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
740 * Estimate number of lock handles that would fit into request of given
741 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
742 * a single page on the send/receive side. XXX: 512 should be changed to
743 * more adequate value.
745 static inline int ldlm_req_handles_avail(int req_size, int off)
749 avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512) - req_size;
750 if (likely(avail >= 0))
751 avail /= (int)sizeof(struct lustre_handle);
754 avail += LDLM_LOCKREQ_HANDLES - off;
759 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
760 enum req_location loc,
763 __u32 size = req_capsule_msg_size(pill, loc);
764 return ldlm_req_handles_avail(size, off);
767 static inline int ldlm_format_handles_avail(struct obd_import *imp,
768 const struct req_format *fmt,
769 enum req_location loc, int off)
771 __u32 size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
772 return ldlm_req_handles_avail(size, off);
776 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
777 * \a count locks in \a cancels.
779 * This is to be called by functions preparing their own requests that
780 * might contain lists of locks to cancel in addition to actual operation
781 * that needs to be performed.
783 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
784 int version, int opc, int canceloff,
785 struct list_head *cancels, int count)
787 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
788 struct req_capsule *pill = &req->rq_pill;
789 struct ldlm_request *dlm = NULL;
790 struct list_head head = LIST_HEAD_INIT(head);
791 enum ldlm_lru_flags lru_flags;
792 int avail, to_free, pack = 0;
798 if (ns_connect_cancelset(ns)) {
799 /* Estimate the amount of available space in the request. */
800 req_capsule_filled_sizes(pill, RCL_CLIENT);
801 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
803 lru_flags = LDLM_LRU_FLAG_NO_WAIT | (ns_connect_lru_resize(ns) ?
804 LDLM_LRU_FLAG_LRUR : LDLM_LRU_FLAG_AGED);
805 to_free = !ns_connect_lru_resize(ns) &&
806 opc == LDLM_ENQUEUE ? 1 : 0;
808 /* Cancel LRU locks here _only_ if the server supports
809 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
810 * RPC, which will make us slower. */
812 count += ldlm_cancel_lru_local(ns, cancels, to_free,
819 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
820 ldlm_request_bufsize(pack, opc));
823 rc = ptlrpc_request_pack(req, version, opc);
825 ldlm_lock_list_put(cancels, l_bl_ast, count);
829 if (ns_connect_cancelset(ns)) {
831 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
833 /* Skip first lock handler in ldlm_request_pack(),
834 * this method will increment @lock_count according
835 * to the lock handle amount actually written to
837 dlm->lock_count = canceloff;
839 /* Pack into the request @pack lock handles. */
840 ldlm_cli_cancel_list(cancels, pack, req, 0);
841 /* Prepare and send separate cancel RPC for others. */
842 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
844 ldlm_lock_list_put(cancels, l_bl_ast, count);
848 EXPORT_SYMBOL(ldlm_prep_elc_req);
850 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
851 struct list_head *cancels, int count)
853 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
854 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
856 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
858 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
860 struct ptlrpc_request *req;
864 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
866 RETURN(ERR_PTR(-ENOMEM));
868 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
870 ptlrpc_request_free(req);
874 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
875 ptlrpc_request_set_replen(req);
878 EXPORT_SYMBOL(ldlm_enqueue_pack);
881 * Client-side lock enqueue.
883 * If a request has some specific initialisation it is passed in \a reqp,
884 * otherwise it is created in ldlm_cli_enqueue.
886 * Supports sync and async requests, pass \a async flag accordingly. If a
887 * request was created in ldlm_cli_enqueue and it is the async request,
888 * pass it to the caller in \a reqp.
890 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
891 struct ldlm_enqueue_info *einfo,
892 const struct ldlm_res_id *res_id,
893 union ldlm_policy_data const *policy, __u64 *flags,
894 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
895 struct lustre_handle *lockh, int async)
897 struct ldlm_namespace *ns;
898 struct ldlm_lock *lock;
899 struct ldlm_request *body;
900 int is_replay = *flags & LDLM_FL_REPLAY;
901 int req_passed_in = 1;
903 struct ptlrpc_request *req;
906 LASSERT(exp != NULL);
908 ns = exp->exp_obd->obd_namespace;
910 /* If we're replaying this lock, just check some invariants.
911 * If we're creating a new lock, get everything all setup nice. */
913 lock = ldlm_handle2lock_long(lockh, 0);
914 LASSERT(lock != NULL);
915 LDLM_DEBUG(lock, "client-side enqueue START");
916 LASSERT(exp == lock->l_conn_export);
918 const struct ldlm_callback_suite cbs = {
919 .lcs_completion = einfo->ei_cb_cp,
920 .lcs_blocking = einfo->ei_cb_bl,
921 .lcs_glimpse = einfo->ei_cb_gl
923 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
924 einfo->ei_mode, &cbs, einfo->ei_cbdata,
927 RETURN(PTR_ERR(lock));
928 /* for the local lock, add the reference */
929 ldlm_lock_addref_internal(lock, einfo->ei_mode);
930 ldlm_lock2handle(lock, lockh);
932 lock->l_policy_data = *policy;
934 if (einfo->ei_type == LDLM_EXTENT) {
935 /* extent lock without policy is a bug */
939 lock->l_req_extent = policy->l_extent;
941 LDLM_DEBUG(lock, "client-side enqueue START, flags %#llx",
945 lock->l_conn_export = exp;
946 lock->l_export = NULL;
947 lock->l_blocking_ast = einfo->ei_cb_bl;
948 lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
949 lock->l_last_activity = ktime_get_real_seconds();
951 /* lock not sent to server yet */
952 if (reqp == NULL || *reqp == NULL) {
953 req = ldlm_enqueue_pack(exp, lvb_len);
955 failed_lock_cleanup(ns, lock, einfo->ei_mode);
956 LDLM_LOCK_RELEASE(lock);
957 RETURN(PTR_ERR(req));
967 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
969 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
970 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
973 if (*flags & LDLM_FL_NDELAY) {
974 DEBUG_REQ(D_DLMTRACE, req, "enque lock with no delay\n");
975 req->rq_no_resend = req->rq_no_delay = 1;
976 /* probably set a shorter timeout value and handle ETIMEDOUT
977 * in osc_lock_upcall() correctly */
978 /* lustre_msg_set_timeout(req, req->rq_timeout / 2); */
981 /* Dump lock data into the request buffer */
982 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
983 ldlm_lock2desc(lock, &body->lock_desc);
984 body->lock_flags = ldlm_flags_to_wire(*flags);
985 body->lock_handle[0] = *lockh;
987 /* extended LDLM opcodes in client stats */
988 if (exp->exp_obd->obd_svc_stats != NULL) {
989 bool glimpse = *flags & LDLM_FL_HAS_INTENT;
991 /* OST glimpse has no intent buffer */
992 if (req_capsule_has_field(&req->rq_pill, &RMF_LDLM_INTENT,
994 struct ldlm_intent *it;
996 it = req_capsule_client_get(&req->rq_pill,
998 glimpse = (it && (it->opc == IT_GLIMPSE));
1002 ldlm_svc_get_eopc(body, exp->exp_obd->obd_svc_stats);
1004 lprocfs_counter_incr(exp->exp_obd->obd_svc_stats,
1006 LDLM_GLIMPSE_ENQUEUE);
1010 LASSERT(reqp != NULL);
1014 LDLM_DEBUG(lock, "sending request");
1016 rc = ptlrpc_queue_wait(req);
1018 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
1019 einfo->ei_mode, flags, lvb, lvb_len,
1022 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
1023 * one reference that we took */
1025 LDLM_LOCK_RELEASE(lock);
1029 if (!req_passed_in && req != NULL) {
1030 ptlrpc_req_finished(req);
1037 EXPORT_SYMBOL(ldlm_cli_enqueue);
1040 * Client-side lock convert reply handling.
1042 * Finish client lock converting, checks for concurrent converts
1043 * and clear 'converting' flag so lock can be placed back into LRU.
1045 static int lock_convert_interpret(const struct lu_env *env,
1046 struct ptlrpc_request *req,
1047 struct ldlm_async_args *aa, int rc)
1049 struct ldlm_lock *lock;
1050 struct ldlm_reply *reply;
1054 lock = ldlm_handle2lock(&aa->lock_handle);
1056 LDLM_DEBUG_NOLOCK("convert ACK for unknown local cookie %#llx",
1057 aa->lock_handle.cookie);
1061 LDLM_DEBUG(lock, "CONVERTED lock:");
1066 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1068 GOTO(out, rc = -EPROTO);
1070 if (reply->lock_handle.cookie != aa->lock_handle.cookie) {
1071 LDLM_ERROR(lock, "convert ACK with wrong lock cookie %#llx"
1072 " but cookie %#llx from server %s id %s\n",
1073 aa->lock_handle.cookie, reply->lock_handle.cookie,
1074 req->rq_export->exp_client_uuid.uuid,
1075 libcfs_id2str(req->rq_peer));
1076 GOTO(out, rc = -ESTALE);
1079 lock_res_and_lock(lock);
1080 /* Lock convert is sent for any new bits to drop, the converting flag
1081 * is dropped when ibits on server are the same as on client. Meanwhile
1082 * that can be so that more later convert will be replied first with
1083 * and clear converting flag, so in case of such race just exit here.
1084 * if lock has no converting bits then */
1085 if (!ldlm_is_converting(lock)) {
1086 LDLM_DEBUG(lock, "convert ACK for lock without converting flag,"
1087 " reply ibits %#llx",
1088 reply->lock_desc.l_policy_data.l_inodebits.bits);
1089 } else if (reply->lock_desc.l_policy_data.l_inodebits.bits !=
1090 lock->l_policy_data.l_inodebits.bits) {
1091 /* Compare server returned lock ibits and local lock ibits
1092 * if they are the same we consider convertion is done,
1093 * otherwise we have more converts inflight and keep
1096 LDLM_DEBUG(lock, "convert ACK with ibits %#llx\n",
1097 reply->lock_desc.l_policy_data.l_inodebits.bits);
1099 ldlm_clear_converting(lock);
1101 /* Concurrent BL AST has arrived, it may cause another convert
1102 * or cancel so just exit here.
1104 if (!ldlm_is_bl_ast(lock)) {
1105 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
1107 /* Drop cancel_bits since there are no more converts
1108 * and put lock into LRU if it is not there yet.
1110 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1111 spin_lock(&ns->ns_lock);
1112 if (!list_empty(&lock->l_lru))
1113 ldlm_lock_remove_from_lru_nolock(lock);
1114 ldlm_lock_add_to_lru_nolock(lock);
1115 spin_unlock(&ns->ns_lock);
1118 unlock_res_and_lock(lock);
1121 lock_res_and_lock(lock);
1122 if (ldlm_is_converting(lock)) {
1123 LASSERT(list_empty(&lock->l_lru));
1124 ldlm_clear_converting(lock);
1125 ldlm_set_cbpending(lock);
1126 ldlm_set_bl_ast(lock);
1128 unlock_res_and_lock(lock);
1131 LDLM_LOCK_PUT(lock);
1136 * Client-side IBITS lock convert.
1138 * Inform server that lock has been converted instead of canceling.
1139 * Server finishes convert on own side and does reprocess to grant
1140 * all related waiting locks.
1142 * Since convert means only ibits downgrading, client doesn't need to
1143 * wait for server reply to finish local converting process so this request
1144 * is made asynchronous.
1147 int ldlm_cli_convert(struct ldlm_lock *lock, __u32 *flags)
1149 struct ldlm_request *body;
1150 struct ptlrpc_request *req;
1151 struct ldlm_async_args *aa;
1152 struct obd_export *exp = lock->l_conn_export;
1157 LDLM_ERROR(lock, "convert must not be called on local locks.");
1161 if (lock->l_resource->lr_type != LDLM_IBITS) {
1162 LDLM_ERROR(lock, "convert works with IBITS locks only.");
1166 LDLM_DEBUG(lock, "client-side convert");
1168 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1169 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1174 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1175 body->lock_handle[0] = lock->l_remote_handle;
1177 body->lock_desc.l_req_mode = lock->l_req_mode;
1178 body->lock_desc.l_granted_mode = lock->l_granted_mode;
1180 body->lock_desc.l_policy_data.l_inodebits.bits =
1181 lock->l_policy_data.l_inodebits.bits;
1182 body->lock_desc.l_policy_data.l_inodebits.cancel_bits = 0;
1184 body->lock_flags = ldlm_flags_to_wire(*flags);
1185 body->lock_count = 1;
1187 ptlrpc_request_set_replen(req);
1189 /* That could be useful to use cancel portals for convert as well
1190 * as high-priority handling. This will require changes in
1191 * ldlm_cancel_handler to understand convert RPC as well.
1193 * req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1194 * req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1196 ptlrpc_at_set_req_timeout(req);
1198 if (exp->exp_obd->obd_svc_stats != NULL)
1199 lprocfs_counter_incr(exp->exp_obd->obd_svc_stats,
1200 LDLM_CONVERT - LDLM_FIRST_OPC);
1202 aa = ptlrpc_req_async_args(req);
1203 ldlm_lock2handle(lock, &aa->lock_handle);
1204 req->rq_interpret_reply = (ptlrpc_interpterer_t)lock_convert_interpret;
1206 ptlrpcd_add_req(req);
1211 * Cancel locks locally.
1213 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1214 * \retval LDLM_FL_CANCELING otherwise;
1215 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1217 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1219 __u64 rc = LDLM_FL_LOCAL_ONLY;
1222 if (lock->l_conn_export) {
1225 LDLM_DEBUG(lock, "client-side cancel");
1226 /* Set this flag to prevent others from getting new references*/
1227 lock_res_and_lock(lock);
1228 ldlm_set_cbpending(lock);
1229 local_only = !!(lock->l_flags &
1230 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1231 ldlm_cancel_callback(lock);
1232 rc = (ldlm_is_bl_ast(lock)) ?
1233 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1234 unlock_res_and_lock(lock);
1237 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1239 rc = LDLM_FL_LOCAL_ONLY;
1241 ldlm_lock_cancel(lock);
1243 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1244 LDLM_ERROR(lock, "Trying to cancel local lock");
1247 LDLM_DEBUG(lock, "server-side local cancel");
1248 ldlm_lock_cancel(lock);
1249 ldlm_reprocess_all(lock->l_resource);
1256 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1258 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1259 struct list_head *head, int count)
1261 struct ldlm_request *dlm;
1262 struct ldlm_lock *lock;
1263 int max, packed = 0;
1266 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1267 LASSERT(dlm != NULL);
1269 /* Check the room in the request buffer. */
1270 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1271 sizeof(struct ldlm_request);
1272 max /= sizeof(struct lustre_handle);
1273 max += LDLM_LOCKREQ_HANDLES;
1274 LASSERT(max >= dlm->lock_count + count);
1276 /* XXX: it would be better to pack lock handles grouped by resource.
1277 * so that the server cancel would call filter_lvbo_update() less
1279 list_for_each_entry(lock, head, l_bl_ast) {
1282 LASSERT(lock->l_conn_export);
1283 /* Pack the lock handle to the given request buffer. */
1284 LDLM_DEBUG(lock, "packing");
1285 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1288 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1293 * Prepare and send a batched cancel RPC. It will include \a count lock
1294 * handles of locks given in \a cancels list. */
1295 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1296 int count, enum ldlm_cancel_flags flags)
1298 struct ptlrpc_request *req = NULL;
1299 struct obd_import *imp;
1304 LASSERT(exp != NULL);
1307 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1309 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1312 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1313 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1318 imp = class_exp2cliimp(exp);
1319 if (imp == NULL || imp->imp_invalid) {
1321 "skipping cancel on invalid import %p\n", imp);
1325 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1327 GOTO(out, rc = -ENOMEM);
1329 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1330 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1331 ldlm_request_bufsize(count, LDLM_CANCEL));
1333 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1335 ptlrpc_request_free(req);
1339 /* If OSP want cancel cross-MDT lock, let's not block it in
1340 * in recovery, otherwise the lock will not released, if
1341 * the remote target is also in recovery, and it also need
1342 * this lock, it might cause deadlock. */
1343 if (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS &&
1344 exp->exp_obd->obd_lu_dev != NULL &&
1345 exp->exp_obd->obd_lu_dev->ld_site != NULL) {
1346 struct lu_device *top_dev;
1348 top_dev = exp->exp_obd->obd_lu_dev->ld_site->ls_top_dev;
1349 if (top_dev != NULL &&
1350 top_dev->ld_obd->obd_recovering)
1351 req->rq_allow_replay = 1;
1354 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1355 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1356 ptlrpc_at_set_req_timeout(req);
1358 ldlm_cancel_pack(req, cancels, count);
1360 ptlrpc_request_set_replen(req);
1361 if (flags & LCF_ASYNC) {
1362 ptlrpcd_add_req(req);
1367 rc = ptlrpc_queue_wait(req);
1368 if (rc == LUSTRE_ESTALE) {
1369 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1370 "out of sync -- not fatal\n",
1371 libcfs_nid2str(req->rq_import->
1372 imp_connection->c_peer.nid));
1374 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1375 req->rq_import_generation == imp->imp_generation) {
1376 ptlrpc_req_finished(req);
1378 } else if (rc != ELDLM_OK) {
1379 /* -ESHUTDOWN is common on umount */
1380 CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1381 "Got rc %d from cancel RPC: "
1382 "canceling anyway\n", rc);
1389 ptlrpc_req_finished(req);
1392 return sent ? sent : rc;
1395 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1397 LASSERT(imp != NULL);
1398 return &imp->imp_obd->obd_namespace->ns_pool;
1402 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1404 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1406 struct obd_device *obd;
1410 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1411 !imp_connect_lru_resize(req->rq_import)))
1414 * Do nothing for corner cases.
1419 /* In some cases RPC may contain SLV and limit zeroed out. This
1420 * is the case when server does not support LRU resize feature.
1421 * This is also possible in some recovery cases when server-side
1422 * reqs have no reference to the OBD export and thus access to
1423 * server-side namespace is not possible. */
1424 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1425 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1426 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1427 "(SLV: %llu, Limit: %u)",
1428 lustre_msg_get_slv(req->rq_repmsg),
1429 lustre_msg_get_limit(req->rq_repmsg));
1433 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1434 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1435 obd = req->rq_import->imp_obd;
1437 /* Set new SLV and limit in OBD fields to make them accessible
1438 * to the pool thread. We do not access obd_namespace and pool
1439 * directly here as there is no reliable way to make sure that
1440 * they are still alive at cleanup time. Evil races are possible
1441 * which may cause Oops at that time. */
1442 write_lock(&obd->obd_pool_lock);
1443 obd->obd_pool_slv = new_slv;
1444 obd->obd_pool_limit = new_limit;
1445 write_unlock(&obd->obd_pool_lock);
1451 * Client side lock cancel.
1453 * Lock must not have any readers or writers by this time.
1455 int ldlm_cli_cancel(const struct lustre_handle *lockh,
1456 enum ldlm_cancel_flags cancel_flags)
1458 struct obd_export *exp;
1459 enum ldlm_lru_flags lru_flags;
1460 int avail, count = 1;
1462 struct ldlm_namespace *ns;
1463 struct ldlm_lock *lock;
1464 struct list_head cancels = LIST_HEAD_INIT(cancels);
1468 lock = ldlm_handle2lock_long(lockh, 0);
1470 LDLM_DEBUG_NOLOCK("lock is already being destroyed");
1474 /* Convert lock bits instead of cancel for IBITS locks */
1475 if (cancel_flags & LCF_CONVERT) {
1476 LASSERT(lock->l_resource->lr_type == LDLM_IBITS);
1477 LASSERT(lock->l_policy_data.l_inodebits.cancel_bits != 0);
1479 rc = ldlm_cli_dropbits(lock,
1480 lock->l_policy_data.l_inodebits.cancel_bits);
1482 LDLM_LOCK_RELEASE(lock);
1487 lock_res_and_lock(lock);
1488 /* Lock is being canceled and the caller doesn't want to wait */
1489 if (ldlm_is_canceling(lock)) {
1490 if (cancel_flags & LCF_ASYNC) {
1491 unlock_res_and_lock(lock);
1493 struct l_wait_info lwi = { 0 };
1495 unlock_res_and_lock(lock);
1496 l_wait_event(lock->l_waitq, is_bl_done(lock), &lwi);
1498 LDLM_LOCK_RELEASE(lock);
1502 /* Lock is being converted, cancel it immediately.
1503 * When convert will end, it releases lock and it will be gone.
1505 if (ldlm_is_converting(lock)) {
1506 /* set back flags removed by convert */
1507 ldlm_set_cbpending(lock);
1508 ldlm_set_bl_ast(lock);
1511 ldlm_set_canceling(lock);
1512 unlock_res_and_lock(lock);
1514 rc = ldlm_cli_cancel_local(lock);
1515 if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
1516 LDLM_LOCK_RELEASE(lock);
1519 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1520 * RPC which goes to canceld portal, so we can cancel other LRU locks
1521 * here and send them all as one LDLM_CANCEL RPC. */
1522 LASSERT(list_empty(&lock->l_bl_ast));
1523 list_add(&lock->l_bl_ast, &cancels);
1525 exp = lock->l_conn_export;
1526 if (exp_connect_cancelset(exp)) {
1527 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1532 ns = ldlm_lock_to_ns(lock);
1533 lru_flags = ns_connect_lru_resize(ns) ?
1534 LDLM_LRU_FLAG_LRUR : LDLM_LRU_FLAG_AGED;
1535 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1536 LCF_BL_AST, lru_flags);
1538 ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1541 EXPORT_SYMBOL(ldlm_cli_cancel);
1544 * Locally cancel up to \a count locks in list \a cancels.
1545 * Return the number of cancelled locks.
1547 int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
1548 enum ldlm_cancel_flags cancel_flags)
1550 struct list_head head = LIST_HEAD_INIT(head);
1551 struct ldlm_lock *lock, *next;
1552 int left = 0, bl_ast = 0;
1556 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1560 if (cancel_flags & LCF_LOCAL) {
1561 rc = LDLM_FL_LOCAL_ONLY;
1562 ldlm_lock_cancel(lock);
1564 rc = ldlm_cli_cancel_local(lock);
1566 /* Until we have compound requests and can send LDLM_CANCEL
1567 * requests batched with generic RPCs, we need to send cancels
1568 * with the LDLM_FL_BL_AST flag in a separate RPC from
1569 * the one being generated now. */
1570 if (!(cancel_flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1571 LDLM_DEBUG(lock, "Cancel lock separately");
1572 list_del_init(&lock->l_bl_ast);
1573 list_add(&lock->l_bl_ast, &head);
1577 if (rc == LDLM_FL_LOCAL_ONLY) {
1578 /* CANCEL RPC should not be sent to server. */
1579 list_del_init(&lock->l_bl_ast);
1580 LDLM_LOCK_RELEASE(lock);
1586 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1593 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1594 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1595 * readahead requests, ...)
1597 static enum ldlm_policy_res
1598 ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1599 int unused, int added, int count)
1601 enum ldlm_policy_res result = LDLM_POLICY_CANCEL_LOCK;
1603 /* don't check added & count since we want to process all locks
1605 * It's fine to not take lock to access lock->l_resource since
1606 * the lock has already been granted so it won't change. */
1607 switch (lock->l_resource->lr_type) {
1610 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) != 0)
1613 result = LDLM_POLICY_SKIP_LOCK;
1621 * Callback function for LRU-resize policy. Decides whether to keep
1622 * \a lock in LRU for current \a LRU size \a unused, added in current
1623 * scan \a added and number of locks to be preferably canceled \a count.
1625 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1627 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1629 static enum ldlm_policy_res ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1630 struct ldlm_lock *lock,
1631 int unused, int added,
1634 ktime_t cur = ktime_get();
1635 struct ldlm_pool *pl = &ns->ns_pool;
1639 /* Stop LRU processing when we reach past @count or have checked all
1641 if (count && added >= count)
1642 return LDLM_POLICY_KEEP_LOCK;
1644 /* Despite of the LV, It doesn't make sense to keep the lock which
1645 * is unused for ns_max_age time.
1647 if (ktime_after(ktime_get(),
1648 ktime_add(lock->l_last_used, ns->ns_max_age)))
1649 return LDLM_POLICY_CANCEL_LOCK;
1651 slv = ldlm_pool_get_slv(pl);
1652 lvf = ldlm_pool_get_lvf(pl);
1653 la = div_u64(ktime_to_ns(ktime_sub(cur, lock->l_last_used)),
1655 lv = lvf * la * unused;
1657 /* Inform pool about current CLV to see it via debugfs. */
1658 ldlm_pool_set_clv(pl, lv);
1660 /* Stop when SLV is not yet come from server or lv is smaller than
1662 if (slv == 0 || lv < slv)
1663 return LDLM_POLICY_KEEP_LOCK;
1665 return LDLM_POLICY_CANCEL_LOCK;
1668 static enum ldlm_policy_res
1669 ldlm_cancel_lrur_no_wait_policy(struct ldlm_namespace *ns,
1670 struct ldlm_lock *lock,
1671 int unused, int added,
1674 enum ldlm_policy_res result;
1676 result = ldlm_cancel_lrur_policy(ns, lock, unused, added, count);
1677 if (result == LDLM_POLICY_KEEP_LOCK)
1680 return ldlm_cancel_no_wait_policy(ns, lock, unused, added, count);
1684 * Callback function for debugfs used policy. Makes decision whether to keep
1685 * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1686 * added and number of locks to be preferably canceled \a count.
1688 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1690 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1692 static enum ldlm_policy_res ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1693 struct ldlm_lock *lock,
1694 int unused, int added,
1697 /* Stop LRU processing when we reach past @count or have checked all
1699 return (added >= count) ?
1700 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1704 * Callback function for aged policy. Makes decision whether to keep \a lock in
1705 * LRU for current LRU size \a unused, added in current scan \a added and
1706 * number of locks to be preferably canceled \a count.
1708 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1710 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1712 static enum ldlm_policy_res ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1713 struct ldlm_lock *lock,
1714 int unused, int added,
1717 if ((added >= count) &&
1718 ktime_before(ktime_get(),
1719 ktime_add(lock->l_last_used, ns->ns_max_age)))
1720 return LDLM_POLICY_KEEP_LOCK;
1722 return LDLM_POLICY_CANCEL_LOCK;
1725 static enum ldlm_policy_res
1726 ldlm_cancel_aged_no_wait_policy(struct ldlm_namespace *ns,
1727 struct ldlm_lock *lock,
1728 int unused, int added, int count)
1730 enum ldlm_policy_res result;
1732 result = ldlm_cancel_aged_policy(ns, lock, unused, added, count);
1733 if (result == LDLM_POLICY_KEEP_LOCK)
1736 return ldlm_cancel_no_wait_policy(ns, lock, unused, added, count);
1740 * Callback function for default policy. Makes decision whether to keep \a lock
1741 * in LRU for current LRU size \a unused, added in current scan \a added and
1742 * number of locks to be preferably canceled \a count.
1744 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1746 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1749 enum ldlm_policy_res ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1750 struct ldlm_lock *lock,
1751 int unused, int added,
1754 /* Stop LRU processing when we reach past count or have checked all
1756 return (added >= count) ?
1757 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1760 typedef enum ldlm_policy_res
1761 (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1762 int unused, int added, int count);
1764 static ldlm_cancel_lru_policy_t
1765 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, enum ldlm_lru_flags lru_flags)
1767 if (ns_connect_lru_resize(ns)) {
1768 if (lru_flags & LDLM_LRU_FLAG_SHRINK)
1769 /* We kill passed number of old locks. */
1770 return ldlm_cancel_passed_policy;
1771 if (lru_flags & LDLM_LRU_FLAG_LRUR) {
1772 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1773 return ldlm_cancel_lrur_no_wait_policy;
1775 return ldlm_cancel_lrur_policy;
1777 if (lru_flags & LDLM_LRU_FLAG_PASSED)
1778 return ldlm_cancel_passed_policy;
1780 if (lru_flags & LDLM_LRU_FLAG_AGED) {
1781 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1782 return ldlm_cancel_aged_no_wait_policy;
1784 return ldlm_cancel_aged_policy;
1787 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1788 return ldlm_cancel_no_wait_policy;
1790 return ldlm_cancel_default_policy;
1794 * - Free space in LRU for \a count new locks,
1795 * redundant unused locks are canceled locally;
1796 * - also cancel locally unused aged locks;
1797 * - do not cancel more than \a max locks;
1798 * - GET the found locks and add them into the \a cancels list.
1800 * A client lock can be added to the l_bl_ast list only when it is
1801 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1802 * CANCEL. There are the following use cases:
1803 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1804 * ldlm_cli_cancel(), which check and set this flag properly. As any
1805 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1806 * later without any special locking.
1808 * Calling policies for enabled LRU resize:
1809 * ----------------------------------------
1810 * flags & LDLM_LRU_FLAG_LRUR - use LRU resize policy (SLV from server) to
1811 * cancel not more than \a count locks;
1813 * flags & LDLM_LRU_FLAG_PASSED - cancel \a count number of old locks (located
1814 * at the beginning of LRU list);
1816 * flags & LDLM_LRU_FLAG_SHRINK - cancel not more than \a count locks according
1817 * to memory pressre policy function;
1819 * flags & LDLM_LRU_FLAG_AGED - cancel \a count locks according to "aged policy"
1821 * flags & LDLM_LRU_FLAG_NO_WAIT - cancel as many unused locks as possible
1822 * (typically before replaying locks) w/o
1823 * sending any RPCs or waiting for any
1824 * outstanding RPC to complete.
1826 * flags & LDLM_CANCEL_CLEANUP - when cancelling read locks, do not check for
1827 * other read locks covering the same pages, just
1828 * discard those pages.
1830 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns,
1831 struct list_head *cancels, int count, int max,
1832 enum ldlm_lru_flags lru_flags)
1834 ldlm_cancel_lru_policy_t pf;
1836 int no_wait = lru_flags & LDLM_LRU_FLAG_NO_WAIT;
1840 if (!ns_connect_lru_resize(ns))
1841 count += ns->ns_nr_unused - ns->ns_max_unused;
1843 pf = ldlm_cancel_lru_policy(ns, lru_flags);
1844 LASSERT(pf != NULL);
1846 /* For any flags, stop scanning if @max is reached. */
1847 while (!list_empty(&ns->ns_unused_list) && (max == 0 || added < max)) {
1848 struct ldlm_lock *lock;
1849 struct list_head *item, *next;
1850 enum ldlm_policy_res result;
1851 ktime_t last_use = ktime_set(0, 0);
1853 spin_lock(&ns->ns_lock);
1854 item = no_wait ? ns->ns_last_pos : &ns->ns_unused_list;
1855 for (item = item->next, next = item->next;
1856 item != &ns->ns_unused_list;
1857 item = next, next = item->next) {
1858 lock = list_entry(item, struct ldlm_lock, l_lru);
1860 /* No locks which got blocking requests. */
1861 LASSERT(!ldlm_is_bl_ast(lock));
1863 if (!ldlm_is_canceling(lock) ||
1864 !ldlm_is_converting(lock))
1867 /* Somebody is already doing CANCEL. No need for this
1868 * lock in LRU, do not traverse it again. */
1869 ldlm_lock_remove_from_lru_nolock(lock);
1871 if (item == &ns->ns_unused_list) {
1872 spin_unlock(&ns->ns_lock);
1876 last_use = lock->l_last_used;
1878 LDLM_LOCK_GET(lock);
1879 spin_unlock(&ns->ns_lock);
1880 lu_ref_add(&lock->l_reference, __FUNCTION__, current);
1882 /* Pass the lock through the policy filter and see if it
1883 * should stay in LRU.
1885 * Even for shrinker policy we stop scanning if
1886 * we find a lock that should stay in the cache.
1887 * We should take into account lock age anyway
1888 * as a new lock is a valuable resource even if
1889 * it has a low weight.
1891 * That is, for shrinker policy we drop only
1892 * old locks, but additionally choose them by
1893 * their weight. Big extent locks will stay in
1895 result = pf(ns, lock, ns->ns_nr_unused, added, count);
1896 if (result == LDLM_POLICY_KEEP_LOCK) {
1897 lu_ref_del(&lock->l_reference, __func__, current);
1898 LDLM_LOCK_RELEASE(lock);
1902 if (result == LDLM_POLICY_SKIP_LOCK) {
1903 lu_ref_del(&lock->l_reference, __func__, current);
1904 LDLM_LOCK_RELEASE(lock);
1906 spin_lock(&ns->ns_lock);
1907 if (!list_empty(&lock->l_lru) &&
1908 lock->l_lru.prev == ns->ns_last_pos)
1909 ns->ns_last_pos = &lock->l_lru;
1910 spin_unlock(&ns->ns_lock);
1915 lock_res_and_lock(lock);
1916 /* Check flags again under the lock. */
1917 if (ldlm_is_canceling(lock) || ldlm_is_converting(lock) ||
1918 ldlm_lock_remove_from_lru_check(lock, last_use) == 0) {
1919 /* Another thread is removing lock from LRU, or
1920 * somebody is already doing CANCEL, or there
1921 * is a blocking request which will send cancel
1922 * by itself, or the lock is no longer unused or
1923 * the lock has been used since the pf() call and
1924 * pages could be put under it. */
1925 unlock_res_and_lock(lock);
1926 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1927 LDLM_LOCK_RELEASE(lock);
1930 LASSERT(!lock->l_readers && !lock->l_writers);
1932 /* If we have chosen to cancel this lock voluntarily, we
1933 * better send cancel notification to server, so that it
1934 * frees appropriate state. This might lead to a race
1935 * where while we are doing cancel here, server is also
1936 * silently cancelling this lock. */
1937 ldlm_clear_cancel_on_block(lock);
1939 /* Setting the CBPENDING flag is a little misleading,
1940 * but prevents an important race; namely, once
1941 * CBPENDING is set, the lock can accumulate no more
1942 * readers/writers. Since readers and writers are
1943 * already zero here, ldlm_lock_decref() won't see
1944 * this flag and call l_blocking_ast */
1945 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1947 if ((lru_flags & LDLM_LRU_FLAG_CLEANUP) &&
1948 (lock->l_resource->lr_type == LDLM_EXTENT ||
1949 ldlm_has_dom(lock)) && lock->l_granted_mode == LCK_PR)
1950 ldlm_set_discard_data(lock);
1952 /* We can't re-add to l_lru as it confuses the
1953 * refcounting in ldlm_lock_remove_from_lru() if an AST
1954 * arrives after we drop lr_lock below. We use l_bl_ast
1955 * and can't use l_pending_chain as it is used both on
1956 * server and client nevertheless bug 5666 says it is
1957 * used only on server */
1958 LASSERT(list_empty(&lock->l_bl_ast));
1959 list_add(&lock->l_bl_ast, cancels);
1960 unlock_res_and_lock(lock);
1961 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1967 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1969 enum ldlm_cancel_flags cancel_flags,
1970 enum ldlm_lru_flags lru_flags)
1974 added = ldlm_prepare_lru_list(ns, cancels, count, max, lru_flags);
1978 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1982 * Cancel at least \a nr locks from given namespace LRU.
1984 * When called with LCF_ASYNC the blocking callback will be handled
1985 * in a thread and this function will return after the thread has been
1986 * asked to call the callback. When called with LCF_ASYNC the blocking
1987 * callback will be performed in this function.
1989 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
1990 enum ldlm_cancel_flags cancel_flags,
1991 enum ldlm_lru_flags lru_flags)
1993 struct list_head cancels = LIST_HEAD_INIT(cancels);
1997 /* Just prepare the list of locks, do not actually cancel them yet.
1998 * Locks are cancelled later in a separate thread. */
1999 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, lru_flags);
2000 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
2008 * Find and cancel locally unused locks found on resource, matched to the
2009 * given policy, mode. GET the found locks and add them into the \a cancels
2012 int ldlm_cancel_resource_local(struct ldlm_resource *res,
2013 struct list_head *cancels,
2014 union ldlm_policy_data *policy,
2015 enum ldlm_mode mode, __u64 lock_flags,
2016 enum ldlm_cancel_flags cancel_flags,
2019 struct ldlm_lock *lock;
2024 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
2025 if (opaque != NULL && lock->l_ast_data != opaque) {
2026 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
2027 lock->l_ast_data, opaque);
2032 if (lock->l_readers || lock->l_writers)
2035 /* If somebody is already doing CANCEL, or blocking AST came,
2036 * skip this lock. */
2037 if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock))
2040 if (lockmode_compat(lock->l_granted_mode, mode))
2043 /* If policy is given and this is IBITS lock, add to list only
2044 * those locks that match by policy. */
2045 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
2046 !(lock->l_policy_data.l_inodebits.bits &
2047 policy->l_inodebits.bits))
2050 /* See CBPENDING comment in ldlm_cancel_lru */
2051 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
2054 LASSERT(list_empty(&lock->l_bl_ast));
2055 list_add(&lock->l_bl_ast, cancels);
2056 LDLM_LOCK_GET(lock);
2061 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
2063 EXPORT_SYMBOL(ldlm_cancel_resource_local);
2066 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
2068 * If \a req is NULL, send CANCEL request to server with handles of locks
2069 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
2070 * separately per lock.
2071 * If \a req is not NULL, put handles of locks in \a cancels into the request
2072 * buffer at the offset \a off.
2073 * Destroy \a cancels at the end.
2075 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
2076 struct ptlrpc_request *req,
2077 enum ldlm_cancel_flags flags)
2079 struct ldlm_lock *lock;
2083 if (list_empty(cancels) || count == 0)
2086 /* XXX: requests (both batched and not) could be sent in parallel.
2087 * Usually it is enough to have just 1 RPC, but it is possible that
2088 * there are too many locks to be cancelled in LRU or on a resource.
2089 * It would also speed up the case when the server does not support
2092 LASSERT(!list_empty(cancels));
2093 lock = list_entry(cancels->next, struct ldlm_lock,
2095 LASSERT(lock->l_conn_export);
2097 if (exp_connect_cancelset(lock->l_conn_export)) {
2100 ldlm_cancel_pack(req, cancels, count);
2102 res = ldlm_cli_cancel_req(lock->l_conn_export,
2106 res = ldlm_cli_cancel_req(lock->l_conn_export,
2111 CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
2112 "ldlm_cli_cancel_list: %d\n", res);
2117 ldlm_lock_list_put(cancels, l_bl_ast, res);
2119 LASSERT(count == 0);
2122 EXPORT_SYMBOL(ldlm_cli_cancel_list);
2125 * Cancel all locks on a resource that have 0 readers/writers.
2127 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
2128 * to notify the server. */
2129 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
2130 const struct ldlm_res_id *res_id,
2131 union ldlm_policy_data *policy,
2132 enum ldlm_mode mode,
2133 enum ldlm_cancel_flags flags, void *opaque)
2135 struct ldlm_resource *res;
2136 struct list_head cancels = LIST_HEAD_INIT(cancels);
2141 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2143 /* This is not a problem. */
2144 CDEBUG(D_INFO, "No resource %llu\n", res_id->name[0]);
2148 LDLM_RESOURCE_ADDREF(res);
2149 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
2150 0, flags | LCF_BL_AST, opaque);
2151 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
2153 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
2156 LDLM_RESOURCE_DELREF(res);
2157 ldlm_resource_putref(res);
2160 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2162 struct ldlm_cli_cancel_arg {
2168 ldlm_cli_hash_cancel_unused(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2169 struct hlist_node *hnode, void *arg)
2171 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2172 struct ldlm_cli_cancel_arg *lc = arg;
2174 ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
2175 NULL, LCK_MINMODE, lc->lc_flags,
2177 /* must return 0 for hash iteration */
2182 * Cancel all locks on a namespace (or a specific resource, if given)
2183 * that have 0 readers/writers.
2185 * If flags & LCF_LOCAL, throw the locks away without trying
2186 * to notify the server. */
2187 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
2188 const struct ldlm_res_id *res_id,
2189 enum ldlm_cancel_flags flags, void *opaque)
2191 struct ldlm_cli_cancel_arg arg = {
2193 .lc_opaque = opaque,
2201 if (res_id != NULL) {
2202 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
2206 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2207 ldlm_cli_hash_cancel_unused, &arg, 0);
2212 /* Lock iterators. */
2214 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
2217 struct list_head *tmp, *next;
2218 struct ldlm_lock *lock;
2219 int rc = LDLM_ITER_CONTINUE;
2224 RETURN(LDLM_ITER_CONTINUE);
2227 list_for_each_safe(tmp, next, &res->lr_granted) {
2228 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2230 if (iter(lock, closure) == LDLM_ITER_STOP)
2231 GOTO(out, rc = LDLM_ITER_STOP);
2234 list_for_each_safe(tmp, next, &res->lr_waiting) {
2235 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2237 if (iter(lock, closure) == LDLM_ITER_STOP)
2238 GOTO(out, rc = LDLM_ITER_STOP);
2245 struct iter_helper_data {
2246 ldlm_iterator_t iter;
2250 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2252 struct iter_helper_data *helper = closure;
2253 return helper->iter(lock, helper->closure);
2256 static int ldlm_res_iter_helper(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2257 struct hlist_node *hnode, void *arg)
2260 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2262 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2266 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2267 ldlm_iterator_t iter, void *closure)
2270 struct iter_helper_data helper = { .iter = iter, .closure = closure };
2272 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2273 ldlm_res_iter_helper, &helper, 0);
2277 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2278 * return 0: find no resource
2279 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2282 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2283 const struct ldlm_res_id *res_id,
2284 ldlm_iterator_t iter, void *data)
2286 struct ldlm_resource *res;
2290 LASSERTF(ns != NULL, "must pass in namespace\n");
2292 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2296 LDLM_RESOURCE_ADDREF(res);
2297 rc = ldlm_resource_foreach(res, iter, data);
2298 LDLM_RESOURCE_DELREF(res);
2299 ldlm_resource_putref(res);
2302 EXPORT_SYMBOL(ldlm_resource_iterate);
2306 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2308 struct list_head *list = closure;
2310 /* we use l_pending_chain here, because it's unused on clients. */
2311 LASSERTF(list_empty(&lock->l_pending_chain),
2312 "lock %p next %p prev %p\n",
2313 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2314 /* bug 9573: don't replay locks left after eviction, or
2315 * bug 17614: locks being actively cancelled. Get a reference
2316 * on a lock so that it does not disapear under us (e.g. due to cancel)
2318 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_BL_DONE))) {
2319 list_add(&lock->l_pending_chain, list);
2320 LDLM_LOCK_GET(lock);
2323 return LDLM_ITER_CONTINUE;
2326 static int replay_lock_interpret(const struct lu_env *env,
2327 struct ptlrpc_request *req,
2328 struct ldlm_async_args *aa, int rc)
2330 struct ldlm_lock *lock;
2331 struct ldlm_reply *reply;
2332 struct obd_export *exp;
2335 atomic_dec(&req->rq_import->imp_replay_inflight);
2339 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2341 GOTO(out, rc = -EPROTO);
2343 lock = ldlm_handle2lock(&aa->lock_handle);
2345 CERROR("received replay ack for unknown local cookie %#llx"
2346 " remote cookie %#llx from server %s id %s\n",
2347 aa->lock_handle.cookie, reply->lock_handle.cookie,
2348 req->rq_export->exp_client_uuid.uuid,
2349 libcfs_id2str(req->rq_peer));
2350 GOTO(out, rc = -ESTALE);
2353 /* Key change rehash lock in per-export hash with new key */
2354 exp = req->rq_export;
2355 if (exp && exp->exp_lock_hash) {
2356 /* In the function below, .hs_keycmp resolves to
2357 * ldlm_export_lock_keycmp() */
2358 /* coverity[overrun-buffer-val] */
2359 cfs_hash_rehash_key(exp->exp_lock_hash,
2360 &lock->l_remote_handle,
2361 &reply->lock_handle,
2364 lock->l_remote_handle = reply->lock_handle;
2367 LDLM_DEBUG(lock, "replayed lock:");
2368 ptlrpc_import_recovery_state_machine(req->rq_import);
2369 LDLM_LOCK_PUT(lock);
2372 ptlrpc_connect_import(req->rq_import);
2377 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2379 struct ptlrpc_request *req;
2380 struct ldlm_async_args *aa;
2381 struct ldlm_request *body;
2386 /* Bug 11974: Do not replay a lock which is actively being canceled */
2387 if (ldlm_is_bl_done(lock)) {
2388 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2392 /* If this is reply-less callback lock, we cannot replay it, since
2393 * server might have long dropped it, but notification of that event was
2394 * lost by network. (and server granted conflicting lock already) */
2395 if (ldlm_is_cancel_on_block(lock)) {
2396 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2397 ldlm_lock_cancel(lock);
2402 * If granted mode matches the requested mode, this lock is granted.
2404 * If we haven't been granted anything and are on a resource list,
2405 * then we're blocked/waiting.
2407 * If we haven't been granted anything and we're NOT on a resource list,
2408 * then we haven't got a reply yet and don't have a known disposition.
2409 * This happens whenever a lock enqueue is the request that triggers
2412 if (lock->l_granted_mode == lock->l_req_mode)
2413 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2414 else if (!list_empty(&lock->l_res_link))
2415 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2417 flags = LDLM_FL_REPLAY;
2419 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2420 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2424 /* We're part of recovery, so don't wait for it. */
2425 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2427 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2428 ldlm_lock2desc(lock, &body->lock_desc);
2429 body->lock_flags = ldlm_flags_to_wire(flags);
2431 ldlm_lock2handle(lock, &body->lock_handle[0]);
2432 if (lock->l_lvb_len > 0)
2433 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2434 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2436 ptlrpc_request_set_replen(req);
2437 /* notify the server we've replayed all requests.
2438 * also, we mark the request to be put on a dedicated
2439 * queue to be processed after all request replayes.
2441 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2443 LDLM_DEBUG(lock, "replaying lock:");
2445 atomic_inc(&req->rq_import->imp_replay_inflight);
2446 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2447 aa = ptlrpc_req_async_args(req);
2448 aa->lock_handle = body->lock_handle[0];
2449 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2450 ptlrpcd_add_req(req);
2456 * Cancel as many unused locks as possible before replay. since we are
2457 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2460 * Called only in recovery before replaying locks. there is no need to
2461 * replay locks that are unused. since the clients may hold thousands of
2462 * cached unused locks, dropping the unused locks can greatly reduce the
2463 * load on the servers at recovery time.
2465 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2468 struct list_head cancels = LIST_HEAD_INIT(cancels);
2470 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2471 "replay for namespace %s (%d)\n",
2472 ldlm_ns_name(ns), ns->ns_nr_unused);
2474 /* We don't need to care whether or not LRU resize is enabled
2475 * because the LDLM_LRU_FLAG_NO_WAIT policy doesn't use the
2476 * count parameter */
2477 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2478 LCF_LOCAL, LDLM_LRU_FLAG_NO_WAIT);
2480 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2481 canceled, ldlm_ns_name(ns));
2484 int ldlm_replay_locks(struct obd_import *imp)
2486 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2487 struct list_head list = LIST_HEAD_INIT(list);
2488 struct ldlm_lock *lock, *next;
2493 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2495 /* don't replay locks if import failed recovery */
2496 if (imp->imp_vbr_failed)
2499 /* ensure this doesn't fall to 0 before all have been queued */
2500 atomic_inc(&imp->imp_replay_inflight);
2502 if (ldlm_cancel_unused_locks_before_replay)
2503 ldlm_cancel_unused_locks_for_replay(ns);
2505 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2507 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2508 list_del_init(&lock->l_pending_chain);
2510 LDLM_LOCK_RELEASE(lock);
2511 continue; /* or try to do the rest? */
2513 rc = replay_one_lock(imp, lock);
2514 LDLM_LOCK_RELEASE(lock);
2517 atomic_dec(&imp->imp_replay_inflight);