4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 * This file contains Asynchronous System Trap (AST) handlers and related
34 * LDLM request-processing routines.
36 * An AST is a callback issued on a lock when its state is changed. There are
37 * several different types of ASTs (callbacks) registered for each lock:
39 * - completion AST: when a lock is enqueued by some process, but cannot be
40 * granted immediately due to other conflicting locks on the same resource,
41 * the completion AST is sent to notify the caller when the lock is
44 * - blocking AST: when a lock is granted to some process, if another process
45 * enqueues a conflicting (blocking) lock on a resource, a blocking AST is
46 * sent to notify the holder(s) of the lock(s) of the conflicting lock
47 * request. The lock holder(s) must release their lock(s) on that resource in
48 * a timely manner or be evicted by the server.
50 * - glimpse AST: this is used when a process wants information about a lock
51 * (i.e. the lock value block (LVB)) but does not necessarily require holding
52 * the lock. If the resource is locked, the lock holder(s) are sent glimpse
53 * ASTs and the LVB is returned to the caller, and lock holder(s) may CANCEL
54 * their lock(s) if they are idle. If the resource is not locked, the server
58 #define DEBUG_SUBSYSTEM S_LDLM
60 #include <lustre/lustre_errno.h>
62 #include <lustre_dlm.h>
63 #include <obd_class.h>
66 #include "ldlm_internal.h"
68 unsigned int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
69 module_param(ldlm_enqueue_min, uint, 0644);
70 MODULE_PARM_DESC(ldlm_enqueue_min, "lock enqueue timeout minimum");
72 /* in client side, whether the cached locks will be canceled before replay */
73 unsigned int ldlm_cancel_unused_locks_before_replay = 1;
75 static void interrupted_completion_wait(void *data)
79 struct lock_wait_data {
80 struct ldlm_lock *lwd_lock;
84 struct ldlm_async_args {
85 struct lustre_handle lock_handle;
89 * ldlm_request_bufsize
91 * If opcode=LDLM_ENQUEUE, 1 slot is already occupied,
92 * LDLM_LOCKREQ_HANDLE -1 slots are available.
93 * Otherwise, LDLM_LOCKREQ_HANDLE slots are available.
98 * \retval size of the request buffer
101 int ldlm_request_bufsize(int count, int type)
103 int avail = LDLM_LOCKREQ_HANDLES;
104 if (type == LDLM_ENQUEUE)
105 avail -= LDLM_ENQUEUE_CANCEL_OFF;
108 avail = (count - avail) * sizeof(struct lustre_handle);
112 return sizeof(struct ldlm_request) + avail;
115 int ldlm_expired_completion_wait(void *data)
117 struct lock_wait_data *lwd = data;
118 struct ldlm_lock *lock = lwd->lwd_lock;
119 struct obd_import *imp;
120 struct obd_device *obd;
123 if (lock->l_conn_export == NULL) {
124 static cfs_time_t next_dump = 0, last_dump = 0;
126 LDLM_ERROR(lock, "lock timed out (enqueued at %lld, %llds ago); "
127 "not entering recovery in server code, just going back to sleep",
128 (s64)lock->l_last_activity,
129 (s64)(ktime_get_real_seconds() -
130 lock->l_last_activity));
131 if (cfs_time_after(cfs_time_current(), next_dump)) {
132 last_dump = next_dump;
133 next_dump = cfs_time_shift(300);
134 ldlm_namespace_dump(D_DLMTRACE,
135 ldlm_lock_to_ns(lock));
137 libcfs_debug_dumplog();
142 obd = lock->l_conn_export->exp_obd;
143 imp = obd->u.cli.cl_import;
144 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
145 LDLM_ERROR(lock, "lock timed out (enqueued at %lld, %llds ago), entering recovery for %s@%s",
146 (s64)lock->l_last_activity,
147 (s64)(ktime_get_real_seconds() - lock->l_last_activity),
148 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
154 * Calculate the Completion timeout (covering enqueue, BL AST, data flush,
155 * lock cancel, and their replies). Used for lock completion timeout on the
158 * \param[in] lock lock which is waiting the completion callback
160 * \retval timeout in seconds to wait for the server reply
163 /* We use the same basis for both server side and client side functions
164 from a single node. */
165 static unsigned int ldlm_cp_timeout(struct ldlm_lock *lock)
167 unsigned int timeout;
172 /* Wait a long time for enqueue - server may have to callback a
173 * lock from another client. Server will evict the other client if it
174 * doesn't respond reasonably, and then give us the lock. */
175 timeout = at_get(ldlm_lock_to_ns_at(lock));
176 return max(3 * timeout, ldlm_enqueue_min);
180 * Helper function for ldlm_completion_ast(), updating timings when lock is
183 static int ldlm_completion_tail(struct ldlm_lock *lock, void *data)
188 if (ldlm_is_destroyed(lock) || ldlm_is_failed(lock)) {
189 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
191 } else if (data == NULL) {
192 LDLM_DEBUG(lock, "client-side enqueue: granted");
194 /* Take into AT only CP RPC, not immediately granted locks */
195 delay = ktime_get_real_seconds() - lock->l_last_activity;
196 LDLM_DEBUG(lock, "client-side enqueue: granted after %llds",
199 /* Update our time estimate */
200 at_measured(ldlm_lock_to_ns_at(lock), delay);
206 * Implementation of ->l_completion_ast() for a client, that doesn't wait
207 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
208 * other threads that cannot block for long.
210 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
214 if (flags == LDLM_FL_WAIT_NOREPROC) {
215 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
219 if (!(flags & LDLM_FL_BLOCKED_MASK)) {
220 wake_up(&lock->l_waitq);
221 RETURN(ldlm_completion_tail(lock, data));
224 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
226 ldlm_reprocess_all(lock->l_resource);
229 EXPORT_SYMBOL(ldlm_completion_ast_async);
232 * Generic LDLM "completion" AST. This is called in several cases:
234 * - when a reply to an ENQUEUE RPC is received from the server
235 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
236 * this point (determined by flags);
238 * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
241 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
244 * - to force all locks when resource is destroyed (cleanup_resource());
246 * - during lock conversion (not used currently).
248 * If lock is not granted in the first case, this function waits until second
249 * or penultimate cases happen in some other thread.
252 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
254 /* XXX ALLOCATE - 160 bytes */
255 struct lock_wait_data lwd;
256 struct obd_device *obd;
257 struct obd_import *imp = NULL;
258 struct l_wait_info lwi;
263 if (flags == LDLM_FL_WAIT_NOREPROC) {
264 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
268 if (!(flags & LDLM_FL_BLOCKED_MASK)) {
269 wake_up(&lock->l_waitq);
273 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
278 obd = class_exp2obd(lock->l_conn_export);
280 /* if this is a local lock, then there is no import */
282 imp = obd->u.cli.cl_import;
285 timeout = ldlm_cp_timeout(lock);
288 lock->l_last_activity = cfs_time_current_sec();
290 if (ldlm_is_no_timeout(lock)) {
291 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
292 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
294 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
295 ldlm_expired_completion_wait,
296 interrupted_completion_wait, &lwd);
300 spin_lock(&imp->imp_lock);
301 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
302 spin_unlock(&imp->imp_lock);
305 if (ns_is_client(ldlm_lock_to_ns(lock)) &&
306 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
307 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
308 ldlm_set_fail_loc(lock);
311 /* Go to sleep until the lock is granted or cancelled. */
312 rc = l_wait_event(lock->l_waitq,
313 is_granted_or_cancelled(lock), &lwi);
317 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
322 RETURN(ldlm_completion_tail(lock, data));
324 EXPORT_SYMBOL(ldlm_completion_ast);
327 * A helper to build a blocking AST function
329 * Perform a common operation for blocking ASTs:
330 * defferred lock cancellation.
332 * \param lock the lock blocking or canceling AST was called on
334 * \see mdt_blocking_ast
335 * \see ldlm_blocking_ast
337 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
342 ldlm_set_cbpending(lock);
343 do_ast = (!lock->l_readers && !lock->l_writers);
344 unlock_res_and_lock(lock);
347 struct lustre_handle lockh;
350 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
351 ldlm_lock2handle(lock, &lockh);
352 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
354 CERROR("ldlm_cli_cancel: %d\n", rc);
356 LDLM_DEBUG(lock, "Lock still has references, will be "
361 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
364 * Server blocking AST
366 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
369 * \param lock the lock which blocks a request or cancelling lock
372 * \param flag indicates whether this cancelling or blocking callback
374 * \see ldlm_blocking_ast_nocheck
376 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
377 void *data, int flag)
381 if (flag == LDLM_CB_CANCELING) {
382 /* Don't need to do anything here. */
386 lock_res_and_lock(lock);
387 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
388 * that ldlm_blocking_ast is called just before intent_policy method
389 * takes the lr_lock, then by the time we get the lock, we might not
390 * be the correct blocking function anymore. So check, and return
392 if (lock->l_blocking_ast != ldlm_blocking_ast) {
393 unlock_res_and_lock(lock);
396 RETURN(ldlm_blocking_ast_nocheck(lock));
398 EXPORT_SYMBOL(ldlm_blocking_ast);
401 * Implements ldlm_lock::l_glimpse_ast for extent locks acquired on the server.
403 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for that is
404 * rather subtle: with OST-side locking, it may so happen that _all_ extent
405 * locks are held by the OST. If client wants to obtain the current file size
406 * it calls ll_glimpse_size(), and (as all locks are held only on the server),
407 * this dummy glimpse callback fires and does nothing. The client still
408 * receives the correct file size due to the following fragment of code in
409 * ldlm_cb_interpret():
411 * if (rc == -ELDLM_NO_LOCK_DATA) {
412 * LDLM_DEBUG(lock, "lost race - client has a lock but no"
414 * ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
417 * That is, after the glimpse returns this error, ofd_lvbo_update() is called
418 * and returns the updated file attributes from the inode to the client.
420 * See also comment in ofd_intent_policy() on why servers must set a non-NULL
421 * l_glimpse_ast when grabbing DLM locks. Otherwise, the server will assume
422 * that the object is in the process of being destroyed.
424 * \param[in] lock DLM lock being glimpsed, unused
425 * \param[in] reqp pointer to ptlrpc_request, unused
427 * \retval -ELDLM_NO_LOCK_DATA to get attributes from disk object
429 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
431 return -ELDLM_NO_LOCK_DATA;
435 * Enqueue a local lock (typically on a server).
437 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
438 const struct ldlm_res_id *res_id,
439 enum ldlm_type type, union ldlm_policy_data *policy,
440 enum ldlm_mode mode, __u64 *flags,
441 ldlm_blocking_callback blocking,
442 ldlm_completion_callback completion,
443 ldlm_glimpse_callback glimpse,
444 void *data, __u32 lvb_len, enum lvb_type lvb_type,
445 const __u64 *client_cookie,
446 struct lustre_handle *lockh)
448 struct ldlm_lock *lock;
450 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
451 .lcs_blocking = blocking,
452 .lcs_glimpse = glimpse,
456 LASSERT(!(*flags & LDLM_FL_REPLAY));
457 if (unlikely(ns_is_client(ns))) {
458 CERROR("Trying to enqueue local lock in a shadow namespace\n");
462 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
465 GOTO(out_nolock, err = PTR_ERR(lock));
467 err = ldlm_lvbo_init(lock->l_resource);
469 LDLM_ERROR(lock, "delayed lvb init failed (rc %d)", err);
473 ldlm_lock2handle(lock, lockh);
475 /* NB: we don't have any lock now (lock_res_and_lock)
476 * because it's a new lock */
477 ldlm_lock_addref_internal_nolock(lock, mode);
478 ldlm_set_local(lock);
479 if (*flags & LDLM_FL_ATOMIC_CB)
480 ldlm_set_atomic_cb(lock);
483 lock->l_policy_data = *policy;
484 if (client_cookie != NULL)
485 lock->l_client_cookie = *client_cookie;
486 if (type == LDLM_EXTENT) {
487 /* extent lock without policy is a bug */
491 lock->l_req_extent = policy->l_extent;
494 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
495 if (unlikely(err != ELDLM_OK))
499 *policy = lock->l_policy_data;
501 if (lock->l_completion_ast)
502 lock->l_completion_ast(lock, *flags, NULL);
504 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
507 LDLM_LOCK_RELEASE(lock);
511 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
513 static void failed_lock_cleanup(struct ldlm_namespace *ns,
514 struct ldlm_lock *lock, int mode)
518 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
519 lock_res_and_lock(lock);
520 /* Check that lock is not granted or failed, we might race. */
521 if ((lock->l_req_mode != lock->l_granted_mode) &&
522 !ldlm_is_failed(lock)) {
523 /* Make sure that this lock will not be found by raced
524 * bl_ast and -EINVAL reply is sent to server anyways.
526 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
527 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
530 unlock_res_and_lock(lock);
534 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
535 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
537 LDLM_DEBUG(lock, "lock was granted or failed in race");
539 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
540 * from llite/file.c/ll_file_flock(). */
541 /* This code makes for the fact that we do not have blocking handler on
542 * a client for flock locks. As such this is the place where we must
543 * completely kill failed locks. (interrupted and those that
544 * were waiting to be granted when server evicted us. */
545 if (lock->l_resource->lr_type == LDLM_FLOCK) {
546 lock_res_and_lock(lock);
547 if (!ldlm_is_destroyed(lock)) {
548 ldlm_resource_unlink_lock(lock);
549 ldlm_lock_decref_internal_nolock(lock, mode);
550 ldlm_lock_destroy_nolock(lock);
552 unlock_res_and_lock(lock);
554 ldlm_lock_decref_internal(lock, mode);
559 * Finishing portion of client lock enqueue code.
561 * Called after receiving reply from server.
563 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
564 enum ldlm_type type, __u8 with_policy,
565 enum ldlm_mode mode, __u64 *flags, void *lvb,
566 __u32 lvb_len, const struct lustre_handle *lockh,
569 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
570 int is_replay = *flags & LDLM_FL_REPLAY;
571 struct ldlm_lock *lock;
572 struct ldlm_reply *reply;
573 int cleanup_phase = 1;
576 lock = ldlm_handle2lock(lockh);
577 /* ldlm_cli_enqueue is holding a reference on this lock. */
579 LASSERT(type == LDLM_FLOCK);
583 LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
584 "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
586 if (rc != ELDLM_OK) {
588 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
589 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
591 if (rc != ELDLM_LOCK_ABORTED)
595 /* Before we return, swab the reply */
596 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
598 GOTO(cleanup, rc = -EPROTO);
603 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
606 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
607 GOTO(cleanup, rc = size);
608 } else if (unlikely(size > lvb_len)) {
609 LDLM_ERROR(lock, "Replied LVB is larger than "
610 "expectation, expected = %d, replied = %d",
612 GOTO(cleanup, rc = -EINVAL);
617 if (rc == ELDLM_LOCK_ABORTED) {
618 if (lvb_len > 0 && lvb != NULL)
619 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
621 GOTO(cleanup, rc = rc ? : ELDLM_LOCK_ABORTED);
624 /* lock enqueued on the server */
627 lock_res_and_lock(lock);
628 /* Key change rehash lock in per-export hash with new key */
629 if (exp->exp_lock_hash) {
630 /* In the function below, .hs_keycmp resolves to
631 * ldlm_export_lock_keycmp() */
632 /* coverity[overrun-buffer-val] */
633 cfs_hash_rehash_key(exp->exp_lock_hash,
634 &lock->l_remote_handle,
638 lock->l_remote_handle = reply->lock_handle;
641 *flags = ldlm_flags_from_wire(reply->lock_flags);
642 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
643 LDLM_FL_INHERIT_MASK);
644 unlock_res_and_lock(lock);
646 CDEBUG(D_INFO, "local: %p, remote cookie: %#llx, flags: %#llx\n",
647 lock, reply->lock_handle.cookie, *flags);
649 /* If enqueue returned a blocked lock but the completion handler has
650 * already run, then it fixed up the resource and we don't need to do it
652 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
653 int newmode = reply->lock_desc.l_req_mode;
655 if (newmode && newmode != lock->l_req_mode) {
656 LDLM_DEBUG(lock, "server returned different mode %s",
657 ldlm_lockname[newmode]);
658 lock->l_req_mode = newmode;
661 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
662 &lock->l_resource->lr_name)) {
663 CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES
664 " instead of "DLDLMRES"\n",
665 PLDLMRES(&reply->lock_desc.l_resource),
666 PLDLMRES(lock->l_resource));
668 rc = ldlm_lock_change_resource(ns, lock,
669 &reply->lock_desc.l_resource.lr_name);
670 if (rc || lock->l_resource == NULL)
671 GOTO(cleanup, rc = -ENOMEM);
672 LDLM_DEBUG(lock, "client-side enqueue, new resource");
675 if (!(type == LDLM_IBITS &&
676 !(exp_connect_flags(exp) & OBD_CONNECT_IBITS)))
677 /* We assume lock type cannot change on server*/
678 ldlm_convert_policy_to_local(exp,
679 lock->l_resource->lr_type,
680 &reply->lock_desc.l_policy_data,
681 &lock->l_policy_data);
682 if (type != LDLM_PLAIN)
683 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
686 if ((*flags) & LDLM_FL_AST_SENT) {
687 lock_res_and_lock(lock);
688 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
689 unlock_res_and_lock(lock);
690 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
693 /* If the lock has already been granted by a completion AST, don't
694 * clobber the LVB with an older one. */
696 /* We must lock or a racing completion might update lvb without
697 * letting us know and we'll clobber the correct value.
698 * Cannot unlock after the check either, a that still leaves
699 * a tiny window for completion to get in */
700 lock_res_and_lock(lock);
701 if (lock->l_req_mode != lock->l_granted_mode)
702 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
703 lock->l_lvb_data, lvb_len);
704 unlock_res_and_lock(lock);
712 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
713 if (lock->l_completion_ast != NULL) {
714 int err = lock->l_completion_ast(lock, *flags, NULL);
722 if (lvb_len > 0 && lvb != NULL) {
723 /* Copy the LVB here, and not earlier, because the completion
724 * AST (if any) can override what we got in the reply */
725 memcpy(lvb, lock->l_lvb_data, lvb_len);
728 LDLM_DEBUG(lock, "client-side enqueue END");
731 if (cleanup_phase == 1 && rc)
732 failed_lock_cleanup(ns, lock, mode);
733 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
735 LDLM_LOCK_RELEASE(lock);
738 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
741 * Estimate number of lock handles that would fit into request of given
742 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
743 * a single page on the send/receive side. XXX: 512 should be changed to
744 * more adequate value.
746 static inline int ldlm_req_handles_avail(int req_size, int off)
750 avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512) - req_size;
751 if (likely(avail >= 0))
752 avail /= (int)sizeof(struct lustre_handle);
755 avail += LDLM_LOCKREQ_HANDLES - off;
760 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
761 enum req_location loc,
764 __u32 size = req_capsule_msg_size(pill, loc);
765 return ldlm_req_handles_avail(size, off);
768 static inline int ldlm_format_handles_avail(struct obd_import *imp,
769 const struct req_format *fmt,
770 enum req_location loc, int off)
772 __u32 size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
773 return ldlm_req_handles_avail(size, off);
777 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
778 * \a count locks in \a cancels.
780 * This is to be called by functions preparing their own requests that
781 * might contain lists of locks to cancel in addition to actual operation
782 * that needs to be performed.
784 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
785 int version, int opc, int canceloff,
786 struct list_head *cancels, int count)
788 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
789 struct req_capsule *pill = &req->rq_pill;
790 struct ldlm_request *dlm = NULL;
791 struct list_head head = LIST_HEAD_INIT(head);
792 enum ldlm_lru_flags lru_flags;
793 int avail, to_free, pack = 0;
799 if (ns_connect_cancelset(ns)) {
800 /* Estimate the amount of available space in the request. */
801 req_capsule_filled_sizes(pill, RCL_CLIENT);
802 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
804 lru_flags = LDLM_LRU_FLAG_NO_WAIT | (ns_connect_lru_resize(ns) ?
805 LDLM_LRU_FLAG_LRUR : LDLM_LRU_FLAG_AGED);
806 to_free = !ns_connect_lru_resize(ns) &&
807 opc == LDLM_ENQUEUE ? 1 : 0;
809 /* Cancel LRU locks here _only_ if the server supports
810 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
811 * RPC, which will make us slower. */
813 count += ldlm_cancel_lru_local(ns, cancels, to_free,
820 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
821 ldlm_request_bufsize(pack, opc));
824 rc = ptlrpc_request_pack(req, version, opc);
826 ldlm_lock_list_put(cancels, l_bl_ast, count);
830 if (ns_connect_cancelset(ns)) {
832 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
834 /* Skip first lock handler in ldlm_request_pack(),
835 * this method will increment @lock_count according
836 * to the lock handle amount actually written to
838 dlm->lock_count = canceloff;
840 /* Pack into the request @pack lock handles. */
841 ldlm_cli_cancel_list(cancels, pack, req, 0);
842 /* Prepare and send separate cancel RPC for others. */
843 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
845 ldlm_lock_list_put(cancels, l_bl_ast, count);
849 EXPORT_SYMBOL(ldlm_prep_elc_req);
851 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
852 struct list_head *cancels, int count)
854 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
855 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
857 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
859 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
861 struct ptlrpc_request *req;
865 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
867 RETURN(ERR_PTR(-ENOMEM));
869 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
871 ptlrpc_request_free(req);
875 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
876 ptlrpc_request_set_replen(req);
879 EXPORT_SYMBOL(ldlm_enqueue_pack);
882 * Client-side lock enqueue.
884 * If a request has some specific initialisation it is passed in \a reqp,
885 * otherwise it is created in ldlm_cli_enqueue.
887 * Supports sync and async requests, pass \a async flag accordingly. If a
888 * request was created in ldlm_cli_enqueue and it is the async request,
889 * pass it to the caller in \a reqp.
891 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
892 struct ldlm_enqueue_info *einfo,
893 const struct ldlm_res_id *res_id,
894 union ldlm_policy_data const *policy, __u64 *flags,
895 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
896 struct lustre_handle *lockh, int async)
898 struct ldlm_namespace *ns;
899 struct ldlm_lock *lock;
900 struct ldlm_request *body;
901 int is_replay = *flags & LDLM_FL_REPLAY;
902 int req_passed_in = 1;
904 struct ptlrpc_request *req;
907 LASSERT(exp != NULL);
909 ns = exp->exp_obd->obd_namespace;
911 /* If we're replaying this lock, just check some invariants.
912 * If we're creating a new lock, get everything all setup nice. */
914 lock = ldlm_handle2lock_long(lockh, 0);
915 LASSERT(lock != NULL);
916 LDLM_DEBUG(lock, "client-side enqueue START");
917 LASSERT(exp == lock->l_conn_export);
919 const struct ldlm_callback_suite cbs = {
920 .lcs_completion = einfo->ei_cb_cp,
921 .lcs_blocking = einfo->ei_cb_bl,
922 .lcs_glimpse = einfo->ei_cb_gl
924 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
925 einfo->ei_mode, &cbs, einfo->ei_cbdata,
928 RETURN(PTR_ERR(lock));
929 /* for the local lock, add the reference */
930 ldlm_lock_addref_internal(lock, einfo->ei_mode);
931 ldlm_lock2handle(lock, lockh);
933 lock->l_policy_data = *policy;
935 if (einfo->ei_type == LDLM_EXTENT) {
936 /* extent lock without policy is a bug */
940 lock->l_req_extent = policy->l_extent;
942 LDLM_DEBUG(lock, "client-side enqueue START, flags %#llx",
946 lock->l_conn_export = exp;
947 lock->l_export = NULL;
948 lock->l_blocking_ast = einfo->ei_cb_bl;
949 lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
950 lock->l_last_activity = cfs_time_current_sec();
952 /* lock not sent to server yet */
953 if (reqp == NULL || *reqp == NULL) {
954 req = ldlm_enqueue_pack(exp, lvb_len);
956 failed_lock_cleanup(ns, lock, einfo->ei_mode);
957 LDLM_LOCK_RELEASE(lock);
958 RETURN(PTR_ERR(req));
968 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
970 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
971 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
974 /* Dump lock data into the request buffer */
975 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
976 ldlm_lock2desc(lock, &body->lock_desc);
977 body->lock_flags = ldlm_flags_to_wire(*flags);
978 body->lock_handle[0] = *lockh;
981 LASSERT(reqp != NULL);
985 LDLM_DEBUG(lock, "sending request");
987 rc = ptlrpc_queue_wait(req);
989 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
990 einfo->ei_mode, flags, lvb, lvb_len,
993 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
994 * one reference that we took */
996 LDLM_LOCK_RELEASE(lock);
1000 if (!req_passed_in && req != NULL) {
1001 ptlrpc_req_finished(req);
1008 EXPORT_SYMBOL(ldlm_cli_enqueue);
1010 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
1013 struct ldlm_resource *res;
1016 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1017 CERROR("Trying to cancel local lock\n");
1020 LDLM_DEBUG(lock, "client-side local convert");
1022 res = ldlm_lock_convert(lock, new_mode, flags);
1024 ldlm_reprocess_all(res);
1027 rc = LUSTRE_EDEADLK;
1029 LDLM_DEBUG(lock, "client-side local convert handler END");
1030 LDLM_LOCK_PUT(lock);
1034 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
1035 * conversion of locks which are on the waiting or converting queue */
1036 /* Caller of this code is supposed to take care of lock readers/writers
1038 int ldlm_cli_convert(const struct lustre_handle *lockh, int new_mode,
1041 struct ldlm_request *body;
1042 struct ldlm_reply *reply;
1043 struct ldlm_lock *lock;
1044 struct ldlm_resource *res;
1045 struct ptlrpc_request *req;
1049 lock = ldlm_handle2lock(lockh);
1056 if (lock->l_conn_export == NULL)
1057 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
1059 LDLM_DEBUG(lock, "client-side convert");
1061 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
1062 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1065 LDLM_LOCK_PUT(lock);
1069 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1070 body->lock_handle[0] = lock->l_remote_handle;
1072 body->lock_desc.l_req_mode = new_mode;
1073 body->lock_flags = ldlm_flags_to_wire(*flags);
1076 ptlrpc_request_set_replen(req);
1077 rc = ptlrpc_queue_wait(req);
1081 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1083 GOTO(out, rc = -EPROTO);
1086 GOTO(out, rc = req->rq_status);
1088 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
1090 ldlm_reprocess_all(res);
1091 /* Go to sleep until the lock is granted. */
1092 /* FIXME: or cancelled. */
1093 if (lock->l_completion_ast) {
1094 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1100 rc = LUSTRE_EDEADLK;
1104 LDLM_LOCK_PUT(lock);
1105 ptlrpc_req_finished(req);
1110 * Cancel locks locally.
1112 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1113 * \retval LDLM_FL_CANCELING otherwise;
1114 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1116 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1118 __u64 rc = LDLM_FL_LOCAL_ONLY;
1121 if (lock->l_conn_export) {
1124 LDLM_DEBUG(lock, "client-side cancel");
1125 /* Set this flag to prevent others from getting new references*/
1126 lock_res_and_lock(lock);
1127 ldlm_set_cbpending(lock);
1128 local_only = !!(lock->l_flags &
1129 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1130 ldlm_cancel_callback(lock);
1131 rc = (ldlm_is_bl_ast(lock)) ?
1132 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1133 unlock_res_and_lock(lock);
1136 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1138 rc = LDLM_FL_LOCAL_ONLY;
1140 ldlm_lock_cancel(lock);
1142 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1143 LDLM_ERROR(lock, "Trying to cancel local lock");
1146 LDLM_DEBUG(lock, "server-side local cancel");
1147 ldlm_lock_cancel(lock);
1148 ldlm_reprocess_all(lock->l_resource);
1155 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1157 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1158 struct list_head *head, int count)
1160 struct ldlm_request *dlm;
1161 struct ldlm_lock *lock;
1162 int max, packed = 0;
1165 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1166 LASSERT(dlm != NULL);
1168 /* Check the room in the request buffer. */
1169 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1170 sizeof(struct ldlm_request);
1171 max /= sizeof(struct lustre_handle);
1172 max += LDLM_LOCKREQ_HANDLES;
1173 LASSERT(max >= dlm->lock_count + count);
1175 /* XXX: it would be better to pack lock handles grouped by resource.
1176 * so that the server cancel would call filter_lvbo_update() less
1178 list_for_each_entry(lock, head, l_bl_ast) {
1181 LASSERT(lock->l_conn_export);
1182 /* Pack the lock handle to the given request buffer. */
1183 LDLM_DEBUG(lock, "packing");
1184 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1187 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1192 * Prepare and send a batched cancel RPC. It will include \a count lock
1193 * handles of locks given in \a cancels list. */
1194 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1195 int count, enum ldlm_cancel_flags flags)
1197 struct ptlrpc_request *req = NULL;
1198 struct obd_import *imp;
1203 LASSERT(exp != NULL);
1206 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1208 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1211 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1212 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1217 imp = class_exp2cliimp(exp);
1218 if (imp == NULL || imp->imp_invalid) {
1220 "skipping cancel on invalid import %p\n", imp);
1224 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1226 GOTO(out, rc = -ENOMEM);
1228 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1229 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1230 ldlm_request_bufsize(count, LDLM_CANCEL));
1232 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1234 ptlrpc_request_free(req);
1238 /* If OSP want cancel cross-MDT lock, let's not block it in
1239 * in recovery, otherwise the lock will not released, if
1240 * the remote target is also in recovery, and it also need
1241 * this lock, it might cause deadlock. */
1242 if (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS &&
1243 exp->exp_obd->obd_lu_dev != NULL &&
1244 exp->exp_obd->obd_lu_dev->ld_site != NULL) {
1245 struct lu_device *top_dev;
1247 top_dev = exp->exp_obd->obd_lu_dev->ld_site->ls_top_dev;
1248 if (top_dev != NULL &&
1249 top_dev->ld_obd->obd_recovering)
1250 req->rq_allow_replay = 1;
1253 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1254 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1255 ptlrpc_at_set_req_timeout(req);
1257 ldlm_cancel_pack(req, cancels, count);
1259 ptlrpc_request_set_replen(req);
1260 if (flags & LCF_ASYNC) {
1261 ptlrpcd_add_req(req);
1266 rc = ptlrpc_queue_wait(req);
1267 if (rc == LUSTRE_ESTALE) {
1268 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1269 "out of sync -- not fatal\n",
1270 libcfs_nid2str(req->rq_import->
1271 imp_connection->c_peer.nid));
1273 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1274 req->rq_import_generation == imp->imp_generation) {
1275 ptlrpc_req_finished(req);
1277 } else if (rc != ELDLM_OK) {
1278 /* -ESHUTDOWN is common on umount */
1279 CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1280 "Got rc %d from cancel RPC: "
1281 "canceling anyway\n", rc);
1288 ptlrpc_req_finished(req);
1291 return sent ? sent : rc;
1294 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1296 LASSERT(imp != NULL);
1297 return &imp->imp_obd->obd_namespace->ns_pool;
1301 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1303 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1305 struct obd_device *obd;
1309 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1310 !imp_connect_lru_resize(req->rq_import)))
1313 * Do nothing for corner cases.
1318 /* In some cases RPC may contain SLV and limit zeroed out. This
1319 * is the case when server does not support LRU resize feature.
1320 * This is also possible in some recovery cases when server-side
1321 * reqs have no reference to the OBD export and thus access to
1322 * server-side namespace is not possible. */
1323 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1324 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1325 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1326 "(SLV: %llu, Limit: %u)",
1327 lustre_msg_get_slv(req->rq_repmsg),
1328 lustre_msg_get_limit(req->rq_repmsg));
1332 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1333 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1334 obd = req->rq_import->imp_obd;
1336 /* Set new SLV and limit in OBD fields to make them accessible
1337 * to the pool thread. We do not access obd_namespace and pool
1338 * directly here as there is no reliable way to make sure that
1339 * they are still alive at cleanup time. Evil races are possible
1340 * which may cause Oops at that time. */
1341 write_lock(&obd->obd_pool_lock);
1342 obd->obd_pool_slv = new_slv;
1343 obd->obd_pool_limit = new_limit;
1344 write_unlock(&obd->obd_pool_lock);
1350 * Client side lock cancel.
1352 * Lock must not have any readers or writers by this time.
1354 int ldlm_cli_cancel(const struct lustre_handle *lockh,
1355 enum ldlm_cancel_flags cancel_flags)
1357 struct obd_export *exp;
1358 enum ldlm_lru_flags lru_flags;
1359 int avail, count = 1;
1361 struct ldlm_namespace *ns;
1362 struct ldlm_lock *lock;
1363 struct list_head cancels = LIST_HEAD_INIT(cancels);
1366 lock = ldlm_handle2lock_long(lockh, 0);
1368 LDLM_DEBUG_NOLOCK("lock is already being destroyed");
1372 lock_res_and_lock(lock);
1373 /* Lock is being canceled and the caller doesn't want to wait */
1374 if (ldlm_is_canceling(lock)) {
1375 if (cancel_flags & LCF_ASYNC) {
1376 unlock_res_and_lock(lock);
1378 struct l_wait_info lwi = { 0 };
1380 unlock_res_and_lock(lock);
1381 l_wait_event(lock->l_waitq, is_bl_done(lock), &lwi);
1383 LDLM_LOCK_RELEASE(lock);
1387 ldlm_set_canceling(lock);
1388 unlock_res_and_lock(lock);
1390 rc = ldlm_cli_cancel_local(lock);
1391 if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
1392 LDLM_LOCK_RELEASE(lock);
1395 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1396 * RPC which goes to canceld portal, so we can cancel other LRU locks
1397 * here and send them all as one LDLM_CANCEL RPC. */
1398 LASSERT(list_empty(&lock->l_bl_ast));
1399 list_add(&lock->l_bl_ast, &cancels);
1401 exp = lock->l_conn_export;
1402 if (exp_connect_cancelset(exp)) {
1403 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1408 ns = ldlm_lock_to_ns(lock);
1409 lru_flags = ns_connect_lru_resize(ns) ?
1410 LDLM_LRU_FLAG_LRUR : LDLM_LRU_FLAG_AGED;
1411 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1412 LCF_BL_AST, lru_flags);
1414 ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1417 EXPORT_SYMBOL(ldlm_cli_cancel);
1420 * Locally cancel up to \a count locks in list \a cancels.
1421 * Return the number of cancelled locks.
1423 int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
1424 enum ldlm_cancel_flags cancel_flags)
1426 struct list_head head = LIST_HEAD_INIT(head);
1427 struct ldlm_lock *lock, *next;
1428 int left = 0, bl_ast = 0;
1432 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1436 if (cancel_flags & LCF_LOCAL) {
1437 rc = LDLM_FL_LOCAL_ONLY;
1438 ldlm_lock_cancel(lock);
1440 rc = ldlm_cli_cancel_local(lock);
1442 /* Until we have compound requests and can send LDLM_CANCEL
1443 * requests batched with generic RPCs, we need to send cancels
1444 * with the LDLM_FL_BL_AST flag in a separate RPC from
1445 * the one being generated now. */
1446 if (!(cancel_flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1447 LDLM_DEBUG(lock, "Cancel lock separately");
1448 list_del_init(&lock->l_bl_ast);
1449 list_add(&lock->l_bl_ast, &head);
1453 if (rc == LDLM_FL_LOCAL_ONLY) {
1454 /* CANCEL RPC should not be sent to server. */
1455 list_del_init(&lock->l_bl_ast);
1456 LDLM_LOCK_RELEASE(lock);
1462 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1469 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1470 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1471 * readahead requests, ...)
1473 static enum ldlm_policy_res
1474 ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1475 int unused, int added, int count)
1477 enum ldlm_policy_res result = LDLM_POLICY_CANCEL_LOCK;
1479 /* don't check added & count since we want to process all locks
1481 * It's fine to not take lock to access lock->l_resource since
1482 * the lock has already been granted so it won't change. */
1483 switch (lock->l_resource->lr_type) {
1486 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) != 0)
1489 result = LDLM_POLICY_SKIP_LOCK;
1497 * Callback function for LRU-resize policy. Decides whether to keep
1498 * \a lock in LRU for current \a LRU size \a unused, added in current
1499 * scan \a added and number of locks to be preferably canceled \a count.
1501 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1503 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1505 static enum ldlm_policy_res ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1506 struct ldlm_lock *lock,
1507 int unused, int added,
1510 ktime_t cur = ktime_get();
1511 struct ldlm_pool *pl = &ns->ns_pool;
1515 /* Stop LRU processing when we reach past @count or have checked all
1517 if (count && added >= count)
1518 return LDLM_POLICY_KEEP_LOCK;
1520 /* Despite of the LV, It doesn't make sense to keep the lock which
1521 * is unused for ns_max_age time.
1523 if (ktime_after(ktime_get(),
1524 ktime_add(lock->l_last_used, ns->ns_max_age)))
1525 return LDLM_POLICY_CANCEL_LOCK;
1527 slv = ldlm_pool_get_slv(pl);
1528 lvf = ldlm_pool_get_lvf(pl);
1529 la = ktime_to_ns(ktime_sub(cur, lock->l_last_used)) / NSEC_PER_SEC;
1530 lv = lvf * la * unused;
1532 /* Inform pool about current CLV to see it via proc. */
1533 ldlm_pool_set_clv(pl, lv);
1535 /* Stop when SLV is not yet come from server or lv is smaller than
1537 if (slv == 0 || lv < slv)
1538 return LDLM_POLICY_KEEP_LOCK;
1540 return LDLM_POLICY_CANCEL_LOCK;
1543 static enum ldlm_policy_res
1544 ldlm_cancel_lrur_no_wait_policy(struct ldlm_namespace *ns,
1545 struct ldlm_lock *lock,
1546 int unused, int added,
1549 enum ldlm_policy_res result;
1551 result = ldlm_cancel_lrur_policy(ns, lock, unused, added, count);
1552 if (result == LDLM_POLICY_KEEP_LOCK)
1555 return ldlm_cancel_no_wait_policy(ns, lock, unused, added, count);
1559 * Callback function for proc used policy. Makes decision whether to keep
1560 * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1561 * added and number of locks to be preferably canceled \a count.
1563 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1565 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1567 static enum ldlm_policy_res ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1568 struct ldlm_lock *lock,
1569 int unused, int added,
1572 /* Stop LRU processing when we reach past @count or have checked all
1574 return (added >= count) ?
1575 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1579 * Callback function for aged policy. Makes decision whether to keep \a lock in
1580 * LRU for current LRU size \a unused, added in current scan \a added and
1581 * number of locks to be preferably canceled \a count.
1583 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1585 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1587 static enum ldlm_policy_res ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1588 struct ldlm_lock *lock,
1589 int unused, int added,
1592 if ((added >= count) &&
1593 ktime_before(ktime_get(),
1594 ktime_add(lock->l_last_used, ns->ns_max_age)))
1595 return LDLM_POLICY_KEEP_LOCK;
1597 return LDLM_POLICY_CANCEL_LOCK;
1600 static enum ldlm_policy_res
1601 ldlm_cancel_aged_no_wait_policy(struct ldlm_namespace *ns,
1602 struct ldlm_lock *lock,
1603 int unused, int added, int count)
1605 enum ldlm_policy_res result;
1607 result = ldlm_cancel_aged_policy(ns, lock, unused, added, count);
1608 if (result == LDLM_POLICY_KEEP_LOCK)
1611 return ldlm_cancel_no_wait_policy(ns, lock, unused, added, count);
1615 * Callback function for default policy. Makes decision whether to keep \a lock
1616 * in LRU for current LRU size \a unused, added in current scan \a added and
1617 * number of locks to be preferably canceled \a count.
1619 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1621 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1624 enum ldlm_policy_res ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1625 struct ldlm_lock *lock,
1626 int unused, int added,
1629 /* Stop LRU processing when we reach past count or have checked all
1631 return (added >= count) ?
1632 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1635 typedef enum ldlm_policy_res
1636 (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1637 int unused, int added, int count);
1639 static ldlm_cancel_lru_policy_t
1640 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, enum ldlm_lru_flags lru_flags)
1642 if (ns_connect_lru_resize(ns)) {
1643 if (lru_flags & LDLM_LRU_FLAG_SHRINK)
1644 /* We kill passed number of old locks. */
1645 return ldlm_cancel_passed_policy;
1646 if (lru_flags & LDLM_LRU_FLAG_LRUR) {
1647 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1648 return ldlm_cancel_lrur_no_wait_policy;
1650 return ldlm_cancel_lrur_policy;
1652 if (lru_flags & LDLM_LRU_FLAG_PASSED)
1653 return ldlm_cancel_passed_policy;
1655 if (lru_flags & LDLM_LRU_FLAG_AGED) {
1656 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1657 return ldlm_cancel_aged_no_wait_policy;
1659 return ldlm_cancel_aged_policy;
1662 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1663 return ldlm_cancel_no_wait_policy;
1665 return ldlm_cancel_default_policy;
1669 * - Free space in LRU for \a count new locks,
1670 * redundant unused locks are canceled locally;
1671 * - also cancel locally unused aged locks;
1672 * - do not cancel more than \a max locks;
1673 * - GET the found locks and add them into the \a cancels list.
1675 * A client lock can be added to the l_bl_ast list only when it is
1676 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1677 * CANCEL. There are the following use cases:
1678 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1679 * ldlm_cli_cancel(), which check and set this flag properly. As any
1680 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1681 * later without any special locking.
1683 * Calling policies for enabled LRU resize:
1684 * ----------------------------------------
1685 * flags & LDLM_LRU_FLAG_LRUR - use LRU resize policy (SLV from server) to
1686 * cancel not more than \a count locks;
1688 * flags & LDLM_LRU_FLAG_PASSED - cancel \a count number of old locks (located
1689 * at the beginning of LRU list);
1691 * flags & LDLM_LRU_FLAG_SHRINK - cancel not more than \a count locks according
1692 * to memory pressre policy function;
1694 * flags & LDLM_LRU_FLAG_AGED - cancel \a count locks according to "aged policy"
1696 * flags & LDLM_LRU_FLAG_NO_WAIT - cancel as many unused locks as possible
1697 * (typically before replaying locks) w/o
1698 * sending any RPCs or waiting for any
1699 * outstanding RPC to complete.
1701 * flags & LDLM_CANCEL_CLEANUP - when cancelling read locks, do not check for
1702 * other read locks covering the same pages, just
1703 * discard those pages.
1705 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns,
1706 struct list_head *cancels, int count, int max,
1707 enum ldlm_lru_flags lru_flags)
1709 ldlm_cancel_lru_policy_t pf;
1711 int no_wait = lru_flags & LDLM_LRU_FLAG_NO_WAIT;
1715 if (!ns_connect_lru_resize(ns))
1716 count += ns->ns_nr_unused - ns->ns_max_unused;
1718 pf = ldlm_cancel_lru_policy(ns, lru_flags);
1719 LASSERT(pf != NULL);
1721 /* For any flags, stop scanning if @max is reached. */
1722 while (!list_empty(&ns->ns_unused_list) && (max == 0 || added < max)) {
1723 struct ldlm_lock *lock;
1724 struct list_head *item, *next;
1725 enum ldlm_policy_res result;
1726 ktime_t last_use = ktime_set(0, 0);
1728 spin_lock(&ns->ns_lock);
1729 item = no_wait ? ns->ns_last_pos : &ns->ns_unused_list;
1730 for (item = item->next, next = item->next;
1731 item != &ns->ns_unused_list;
1732 item = next, next = item->next) {
1733 lock = list_entry(item, struct ldlm_lock, l_lru);
1735 /* No locks which got blocking requests. */
1736 LASSERT(!ldlm_is_bl_ast(lock));
1738 if (!ldlm_is_canceling(lock))
1741 /* Somebody is already doing CANCEL. No need for this
1742 * lock in LRU, do not traverse it again. */
1743 ldlm_lock_remove_from_lru_nolock(lock);
1745 if (item == &ns->ns_unused_list) {
1746 spin_unlock(&ns->ns_lock);
1750 last_use = lock->l_last_used;
1752 LDLM_LOCK_GET(lock);
1753 spin_unlock(&ns->ns_lock);
1754 lu_ref_add(&lock->l_reference, __FUNCTION__, current);
1756 /* Pass the lock through the policy filter and see if it
1757 * should stay in LRU.
1759 * Even for shrinker policy we stop scanning if
1760 * we find a lock that should stay in the cache.
1761 * We should take into account lock age anyway
1762 * as a new lock is a valuable resource even if
1763 * it has a low weight.
1765 * That is, for shrinker policy we drop only
1766 * old locks, but additionally choose them by
1767 * their weight. Big extent locks will stay in
1769 result = pf(ns, lock, ns->ns_nr_unused, added, count);
1770 if (result == LDLM_POLICY_KEEP_LOCK) {
1771 lu_ref_del(&lock->l_reference, __func__, current);
1772 LDLM_LOCK_RELEASE(lock);
1776 if (result == LDLM_POLICY_SKIP_LOCK) {
1777 lu_ref_del(&lock->l_reference, __func__, current);
1778 LDLM_LOCK_RELEASE(lock);
1780 spin_lock(&ns->ns_lock);
1781 if (!list_empty(&lock->l_lru) &&
1782 lock->l_lru.prev == ns->ns_last_pos)
1783 ns->ns_last_pos = &lock->l_lru;
1784 spin_unlock(&ns->ns_lock);
1789 lock_res_and_lock(lock);
1790 /* Check flags again under the lock. */
1791 if (ldlm_is_canceling(lock) ||
1792 ldlm_lock_remove_from_lru_check(lock, last_use) == 0) {
1793 /* Another thread is removing lock from LRU, or
1794 * somebody is already doing CANCEL, or there
1795 * is a blocking request which will send cancel
1796 * by itself, or the lock is no longer unused or
1797 * the lock has been used since the pf() call and
1798 * pages could be put under it. */
1799 unlock_res_and_lock(lock);
1800 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1801 LDLM_LOCK_RELEASE(lock);
1804 LASSERT(!lock->l_readers && !lock->l_writers);
1806 /* If we have chosen to cancel this lock voluntarily, we
1807 * better send cancel notification to server, so that it
1808 * frees appropriate state. This might lead to a race
1809 * where while we are doing cancel here, server is also
1810 * silently cancelling this lock. */
1811 ldlm_clear_cancel_on_block(lock);
1813 /* Setting the CBPENDING flag is a little misleading,
1814 * but prevents an important race; namely, once
1815 * CBPENDING is set, the lock can accumulate no more
1816 * readers/writers. Since readers and writers are
1817 * already zero here, ldlm_lock_decref() won't see
1818 * this flag and call l_blocking_ast */
1819 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1821 if ((lru_flags & LDLM_LRU_FLAG_CLEANUP) &&
1822 lock->l_resource->lr_type == LDLM_EXTENT &&
1823 lock->l_granted_mode == LCK_PR)
1824 ldlm_set_discard_data(lock);
1826 /* We can't re-add to l_lru as it confuses the
1827 * refcounting in ldlm_lock_remove_from_lru() if an AST
1828 * arrives after we drop lr_lock below. We use l_bl_ast
1829 * and can't use l_pending_chain as it is used both on
1830 * server and client nevertheless bug 5666 says it is
1831 * used only on server */
1832 LASSERT(list_empty(&lock->l_bl_ast));
1833 list_add(&lock->l_bl_ast, cancels);
1834 unlock_res_and_lock(lock);
1835 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1841 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1843 enum ldlm_cancel_flags cancel_flags,
1844 enum ldlm_lru_flags lru_flags)
1848 added = ldlm_prepare_lru_list(ns, cancels, count, max, lru_flags);
1852 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1856 * Cancel at least \a nr locks from given namespace LRU.
1858 * When called with LCF_ASYNC the blocking callback will be handled
1859 * in a thread and this function will return after the thread has been
1860 * asked to call the callback. When called with LCF_ASYNC the blocking
1861 * callback will be performed in this function.
1863 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
1864 enum ldlm_cancel_flags cancel_flags,
1865 enum ldlm_lru_flags lru_flags)
1867 struct list_head cancels = LIST_HEAD_INIT(cancels);
1871 /* Just prepare the list of locks, do not actually cancel them yet.
1872 * Locks are cancelled later in a separate thread. */
1873 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, lru_flags);
1874 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
1882 * Find and cancel locally unused locks found on resource, matched to the
1883 * given policy, mode. GET the found locks and add them into the \a cancels
1886 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1887 struct list_head *cancels,
1888 union ldlm_policy_data *policy,
1889 enum ldlm_mode mode, __u64 lock_flags,
1890 enum ldlm_cancel_flags cancel_flags,
1893 struct ldlm_lock *lock;
1898 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1899 if (opaque != NULL && lock->l_ast_data != opaque) {
1900 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1901 lock->l_ast_data, opaque);
1906 if (lock->l_readers || lock->l_writers)
1909 /* If somebody is already doing CANCEL, or blocking AST came,
1910 * skip this lock. */
1911 if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock))
1914 if (lockmode_compat(lock->l_granted_mode, mode))
1917 /* If policy is given and this is IBITS lock, add to list only
1918 * those locks that match by policy. */
1919 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1920 !(lock->l_policy_data.l_inodebits.bits &
1921 policy->l_inodebits.bits))
1924 /* See CBPENDING comment in ldlm_cancel_lru */
1925 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1928 LASSERT(list_empty(&lock->l_bl_ast));
1929 list_add(&lock->l_bl_ast, cancels);
1930 LDLM_LOCK_GET(lock);
1935 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
1937 EXPORT_SYMBOL(ldlm_cancel_resource_local);
1940 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1942 * If \a req is NULL, send CANCEL request to server with handles of locks
1943 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1944 * separately per lock.
1945 * If \a req is not NULL, put handles of locks in \a cancels into the request
1946 * buffer at the offset \a off.
1947 * Destroy \a cancels at the end.
1949 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1950 struct ptlrpc_request *req,
1951 enum ldlm_cancel_flags flags)
1953 struct ldlm_lock *lock;
1957 if (list_empty(cancels) || count == 0)
1960 /* XXX: requests (both batched and not) could be sent in parallel.
1961 * Usually it is enough to have just 1 RPC, but it is possible that
1962 * there are too many locks to be cancelled in LRU or on a resource.
1963 * It would also speed up the case when the server does not support
1966 LASSERT(!list_empty(cancels));
1967 lock = list_entry(cancels->next, struct ldlm_lock,
1969 LASSERT(lock->l_conn_export);
1971 if (exp_connect_cancelset(lock->l_conn_export)) {
1974 ldlm_cancel_pack(req, cancels, count);
1976 res = ldlm_cli_cancel_req(lock->l_conn_export,
1980 res = ldlm_cli_cancel_req(lock->l_conn_export,
1985 CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1986 "ldlm_cli_cancel_list: %d\n", res);
1991 ldlm_lock_list_put(cancels, l_bl_ast, res);
1993 LASSERT(count == 0);
1996 EXPORT_SYMBOL(ldlm_cli_cancel_list);
1999 * Cancel all locks on a resource that have 0 readers/writers.
2001 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
2002 * to notify the server. */
2003 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
2004 const struct ldlm_res_id *res_id,
2005 union ldlm_policy_data *policy,
2006 enum ldlm_mode mode,
2007 enum ldlm_cancel_flags flags, void *opaque)
2009 struct ldlm_resource *res;
2010 struct list_head cancels = LIST_HEAD_INIT(cancels);
2015 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2017 /* This is not a problem. */
2018 CDEBUG(D_INFO, "No resource %llu\n", res_id->name[0]);
2022 LDLM_RESOURCE_ADDREF(res);
2023 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
2024 0, flags | LCF_BL_AST, opaque);
2025 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
2027 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
2030 LDLM_RESOURCE_DELREF(res);
2031 ldlm_resource_putref(res);
2034 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2036 struct ldlm_cli_cancel_arg {
2042 ldlm_cli_hash_cancel_unused(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2043 struct hlist_node *hnode, void *arg)
2045 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2046 struct ldlm_cli_cancel_arg *lc = arg;
2048 ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
2049 NULL, LCK_MINMODE, lc->lc_flags,
2051 /* must return 0 for hash iteration */
2056 * Cancel all locks on a namespace (or a specific resource, if given)
2057 * that have 0 readers/writers.
2059 * If flags & LCF_LOCAL, throw the locks away without trying
2060 * to notify the server. */
2061 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
2062 const struct ldlm_res_id *res_id,
2063 enum ldlm_cancel_flags flags, void *opaque)
2065 struct ldlm_cli_cancel_arg arg = {
2067 .lc_opaque = opaque,
2075 if (res_id != NULL) {
2076 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
2080 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2081 ldlm_cli_hash_cancel_unused, &arg, 0);
2086 /* Lock iterators. */
2088 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
2091 struct list_head *tmp, *next;
2092 struct ldlm_lock *lock;
2093 int rc = LDLM_ITER_CONTINUE;
2098 RETURN(LDLM_ITER_CONTINUE);
2101 list_for_each_safe(tmp, next, &res->lr_granted) {
2102 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2104 if (iter(lock, closure) == LDLM_ITER_STOP)
2105 GOTO(out, rc = LDLM_ITER_STOP);
2108 list_for_each_safe(tmp, next, &res->lr_converting) {
2109 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2111 if (iter(lock, closure) == LDLM_ITER_STOP)
2112 GOTO(out, rc = LDLM_ITER_STOP);
2115 list_for_each_safe(tmp, next, &res->lr_waiting) {
2116 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2118 if (iter(lock, closure) == LDLM_ITER_STOP)
2119 GOTO(out, rc = LDLM_ITER_STOP);
2126 struct iter_helper_data {
2127 ldlm_iterator_t iter;
2131 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2133 struct iter_helper_data *helper = closure;
2134 return helper->iter(lock, helper->closure);
2137 static int ldlm_res_iter_helper(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2138 struct hlist_node *hnode, void *arg)
2141 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2143 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2147 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2148 ldlm_iterator_t iter, void *closure)
2151 struct iter_helper_data helper = { .iter = iter, .closure = closure };
2153 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2154 ldlm_res_iter_helper, &helper, 0);
2158 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2159 * return 0: find no resource
2160 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2163 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2164 const struct ldlm_res_id *res_id,
2165 ldlm_iterator_t iter, void *data)
2167 struct ldlm_resource *res;
2171 LASSERTF(ns != NULL, "must pass in namespace\n");
2173 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2177 LDLM_RESOURCE_ADDREF(res);
2178 rc = ldlm_resource_foreach(res, iter, data);
2179 LDLM_RESOURCE_DELREF(res);
2180 ldlm_resource_putref(res);
2183 EXPORT_SYMBOL(ldlm_resource_iterate);
2187 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2189 struct list_head *list = closure;
2191 /* we use l_pending_chain here, because it's unused on clients. */
2192 LASSERTF(list_empty(&lock->l_pending_chain),
2193 "lock %p next %p prev %p\n",
2194 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2195 /* bug 9573: don't replay locks left after eviction, or
2196 * bug 17614: locks being actively cancelled. Get a reference
2197 * on a lock so that it does not disapear under us (e.g. due to cancel)
2199 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_BL_DONE))) {
2200 list_add(&lock->l_pending_chain, list);
2201 LDLM_LOCK_GET(lock);
2204 return LDLM_ITER_CONTINUE;
2207 static int replay_lock_interpret(const struct lu_env *env,
2208 struct ptlrpc_request *req,
2209 struct ldlm_async_args *aa, int rc)
2211 struct ldlm_lock *lock;
2212 struct ldlm_reply *reply;
2213 struct obd_export *exp;
2216 atomic_dec(&req->rq_import->imp_replay_inflight);
2220 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2222 GOTO(out, rc = -EPROTO);
2224 lock = ldlm_handle2lock(&aa->lock_handle);
2226 CERROR("received replay ack for unknown local cookie %#llx"
2227 " remote cookie %#llx from server %s id %s\n",
2228 aa->lock_handle.cookie, reply->lock_handle.cookie,
2229 req->rq_export->exp_client_uuid.uuid,
2230 libcfs_id2str(req->rq_peer));
2231 GOTO(out, rc = -ESTALE);
2234 /* Key change rehash lock in per-export hash with new key */
2235 exp = req->rq_export;
2236 if (exp && exp->exp_lock_hash) {
2237 /* In the function below, .hs_keycmp resolves to
2238 * ldlm_export_lock_keycmp() */
2239 /* coverity[overrun-buffer-val] */
2240 cfs_hash_rehash_key(exp->exp_lock_hash,
2241 &lock->l_remote_handle,
2242 &reply->lock_handle,
2245 lock->l_remote_handle = reply->lock_handle;
2248 LDLM_DEBUG(lock, "replayed lock:");
2249 ptlrpc_import_recovery_state_machine(req->rq_import);
2250 LDLM_LOCK_PUT(lock);
2253 ptlrpc_connect_import(req->rq_import);
2258 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2260 struct ptlrpc_request *req;
2261 struct ldlm_async_args *aa;
2262 struct ldlm_request *body;
2267 /* Bug 11974: Do not replay a lock which is actively being canceled */
2268 if (ldlm_is_bl_done(lock)) {
2269 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2273 /* If this is reply-less callback lock, we cannot replay it, since
2274 * server might have long dropped it, but notification of that event was
2275 * lost by network. (and server granted conflicting lock already) */
2276 if (ldlm_is_cancel_on_block(lock)) {
2277 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2278 ldlm_lock_cancel(lock);
2283 * If granted mode matches the requested mode, this lock is granted.
2285 * If they differ, but we have a granted mode, then we were granted
2286 * one mode and now want another: ergo, converting.
2288 * If we haven't been granted anything and are on a resource list,
2289 * then we're blocked/waiting.
2291 * If we haven't been granted anything and we're NOT on a resource list,
2292 * then we haven't got a reply yet and don't have a known disposition.
2293 * This happens whenever a lock enqueue is the request that triggers
2296 if (lock->l_granted_mode == lock->l_req_mode)
2297 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2298 else if (lock->l_granted_mode)
2299 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2300 else if (!list_empty(&lock->l_res_link))
2301 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2303 flags = LDLM_FL_REPLAY;
2305 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2306 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2310 /* We're part of recovery, so don't wait for it. */
2311 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2313 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2314 ldlm_lock2desc(lock, &body->lock_desc);
2315 body->lock_flags = ldlm_flags_to_wire(flags);
2317 ldlm_lock2handle(lock, &body->lock_handle[0]);
2318 if (lock->l_lvb_len > 0)
2319 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2320 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2322 ptlrpc_request_set_replen(req);
2323 /* notify the server we've replayed all requests.
2324 * also, we mark the request to be put on a dedicated
2325 * queue to be processed after all request replayes.
2327 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2329 LDLM_DEBUG(lock, "replaying lock:");
2331 atomic_inc(&req->rq_import->imp_replay_inflight);
2332 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2333 aa = ptlrpc_req_async_args(req);
2334 aa->lock_handle = body->lock_handle[0];
2335 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2336 ptlrpcd_add_req(req);
2342 * Cancel as many unused locks as possible before replay. since we are
2343 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2346 * Called only in recovery before replaying locks. there is no need to
2347 * replay locks that are unused. since the clients may hold thousands of
2348 * cached unused locks, dropping the unused locks can greatly reduce the
2349 * load on the servers at recovery time.
2351 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2354 struct list_head cancels = LIST_HEAD_INIT(cancels);
2356 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2357 "replay for namespace %s (%d)\n",
2358 ldlm_ns_name(ns), ns->ns_nr_unused);
2360 /* We don't need to care whether or not LRU resize is enabled
2361 * because the LDLM_LRU_FLAG_NO_WAIT policy doesn't use the
2362 * count parameter */
2363 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2364 LCF_LOCAL, LDLM_LRU_FLAG_NO_WAIT);
2366 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2367 canceled, ldlm_ns_name(ns));
2370 int ldlm_replay_locks(struct obd_import *imp)
2372 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2373 struct list_head list = LIST_HEAD_INIT(list);
2374 struct ldlm_lock *lock, *next;
2379 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2381 /* don't replay locks if import failed recovery */
2382 if (imp->imp_vbr_failed)
2385 /* ensure this doesn't fall to 0 before all have been queued */
2386 atomic_inc(&imp->imp_replay_inflight);
2388 if (ldlm_cancel_unused_locks_before_replay)
2389 ldlm_cancel_unused_locks_for_replay(ns);
2391 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2393 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2394 list_del_init(&lock->l_pending_chain);
2396 LDLM_LOCK_RELEASE(lock);
2397 continue; /* or try to do the rest? */
2399 rc = replay_one_lock(imp, lock);
2400 LDLM_LOCK_RELEASE(lock);
2403 atomic_dec(&imp->imp_replay_inflight);