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 struct lock_wait_data {
75 struct ldlm_lock *lwd_lock;
79 struct ldlm_async_args {
80 struct lustre_handle lock_handle;
84 * ldlm_request_bufsize
86 * If opcode=LDLM_ENQUEUE, 1 slot is already occupied,
87 * LDLM_LOCKREQ_HANDLE -1 slots are available.
88 * Otherwise, LDLM_LOCKREQ_HANDLE slots are available.
93 * \retval size of the request buffer
95 int ldlm_request_bufsize(int count, int type)
97 int avail = LDLM_LOCKREQ_HANDLES;
99 if (type == LDLM_ENQUEUE)
100 avail -= LDLM_ENQUEUE_CANCEL_OFF;
103 avail = (count - avail) * sizeof(struct lustre_handle);
107 return sizeof(struct ldlm_request) + avail;
110 void ldlm_expired_completion_wait(struct lock_wait_data *lwd)
112 struct ldlm_lock *lock = lwd->lwd_lock;
113 struct obd_import *imp;
114 struct obd_device *obd;
117 if (lock->l_conn_export == NULL) {
118 static time64_t next_dump, last_dump;
121 "lock timed out (enqueued at %lld, %llds ago); not entering recovery in server code, just going back to sleep",
123 ktime_get_real_seconds() - lock->l_activity);
124 if (ktime_get_seconds() > next_dump) {
125 last_dump = next_dump;
126 next_dump = ktime_get_seconds() + 300;
127 ldlm_namespace_dump(D_DLMTRACE,
128 ldlm_lock_to_ns(lock));
130 libcfs_debug_dumplog();
135 obd = lock->l_conn_export->exp_obd;
136 imp = obd->u.cli.cl_import;
137 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
139 "lock timed out (enqueued at %lld, %llds ago), entering recovery for %s@%s",
141 ktime_get_real_seconds() - lock->l_activity,
142 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
147 int is_granted_or_cancelled_nolock(struct ldlm_lock *lock)
151 check_res_locked(lock->l_resource);
152 if (ldlm_is_granted(lock) && !ldlm_is_cp_reqd(lock))
154 else if (ldlm_is_failed(lock) || ldlm_is_cancel(lock))
158 EXPORT_SYMBOL(is_granted_or_cancelled_nolock);
161 * Calculate the Completion timeout (covering enqueue, BL AST, data flush,
162 * lock cancel, and their replies). Used for lock completion timeout on the
165 * \param[in] lock lock which is waiting the completion callback
167 * \retval timeout in seconds to wait for the server reply
170 * We use the same basis for both server side and client side functions
171 * from a single node.
173 static timeout_t ldlm_cp_timeout(struct ldlm_lock *lock)
181 * Wait a long time for enqueue - server may have to callback a
182 * lock from another client. Server will evict the other client if it
183 * doesn't respond reasonably, and then give us the lock.
185 timeout = at_get(ldlm_lock_to_ns_at(lock));
186 return max(3 * timeout, (timeout_t)ldlm_enqueue_min);
190 * Helper function for ldlm_completion_ast(), updating timings when lock is
193 static int ldlm_completion_tail(struct ldlm_lock *lock, void *data)
197 if (ldlm_is_destroyed(lock) || ldlm_is_failed(lock)) {
198 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
200 } else if (data == NULL) {
201 LDLM_DEBUG(lock, "client-side enqueue: granted");
203 /* Take into AT only CP RPC, not immediately granted locks */
206 /* Discard negative timeouts. We should also limit the
207 * maximum value of the timeout
209 if (ktime_get_real_seconds() > lock->l_activity)
210 delay = ktime_get_real_seconds() - lock->l_activity;
212 LDLM_DEBUG(lock, "client-side enqueue: granted after %ds",
214 /* Update our time estimate */
215 at_measured(ldlm_lock_to_ns_at(lock), delay);
221 * Implementation of ->l_completion_ast() for a client, that doesn't wait
222 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
223 * other threads that cannot block for long.
225 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
229 if (flags == LDLM_FL_WAIT_NOREPROC) {
230 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
234 if (!(flags & LDLM_FL_BLOCKED_MASK)) {
235 wake_up(&lock->l_waitq);
236 RETURN(ldlm_completion_tail(lock, data));
240 "client-side enqueue returned a blocked lock, going forward");
241 ldlm_reprocess_all(lock->l_resource, NULL);
244 EXPORT_SYMBOL(ldlm_completion_ast_async);
247 * Generic LDLM "completion" AST. This is called in several cases:
249 * - when a reply to an ENQUEUE RPC is received from the server
250 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
251 * this point (determined by flags);
253 * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
256 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
259 * - to force all locks when resource is destroyed (cleanup_resource());
261 * If lock is not granted in the first case, this function waits until second
262 * or penultimate cases happen in some other thread.
265 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
267 /* XXX ALLOCATE - 160 bytes */
268 struct lock_wait_data lwd;
269 struct obd_device *obd;
270 struct obd_import *imp = NULL;
276 if (flags == LDLM_FL_WAIT_NOREPROC) {
277 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
281 if (!(flags & LDLM_FL_BLOCKED_MASK)) {
282 wake_up(&lock->l_waitq);
286 LDLM_DEBUG(lock, "client-side enqueue returned a blocked locksleeping");
290 obd = class_exp2obd(lock->l_conn_export);
292 /* if this is a local lock, then there is no import */
294 imp = obd->u.cli.cl_import;
296 timeout = ldlm_cp_timeout(lock);
299 lock->l_activity = ktime_get_real_seconds();
302 spin_lock(&imp->imp_lock);
303 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
304 spin_unlock(&imp->imp_lock);
307 if (ns_is_client(ldlm_lock_to_ns(lock)) &&
308 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
309 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
310 ldlm_set_fail_loc(lock);
313 /* Go to sleep until the lock is granted or cancelled. */
314 if (ldlm_is_no_timeout(lock)) {
315 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
316 rc = l_wait_event_abortable(
318 is_granted_or_cancelled(lock));
320 if (wait_event_idle_timeout(
322 is_granted_or_cancelled(lock),
323 cfs_time_seconds(timeout)) == 0) {
324 ldlm_expired_completion_wait(&lwd);
325 rc = l_wait_event_abortable(
327 is_granted_or_cancelled(lock));
333 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
338 RETURN(ldlm_completion_tail(lock, data));
340 EXPORT_SYMBOL(ldlm_completion_ast);
343 * A helper to build a blocking AST function
345 * Perform a common operation for blocking ASTs:
346 * defferred lock cancellation.
348 * \param lock the lock blocking or canceling AST was called on
350 * \see mdt_blocking_ast
351 * \see ldlm_blocking_ast
353 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
359 ldlm_set_cbpending(lock);
360 do_ast = (!lock->l_readers && !lock->l_writers);
361 unlock_res_and_lock(lock);
364 struct lustre_handle lockh;
367 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
368 ldlm_lock2handle(lock, &lockh);
369 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
371 CERROR("ldlm_cli_cancel: %d\n", rc);
374 "Lock still has references, will be cancelled later");
378 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
381 * Server blocking AST
383 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
386 * \param lock the lock which blocks a request or cancelling lock
389 * \param flag indicates whether this cancelling or blocking callback
391 * \see ldlm_blocking_ast_nocheck
393 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
394 void *data, int flag)
398 if (flag == LDLM_CB_CANCELING) {
399 /* Don't need to do anything here. */
403 lock_res_and_lock(lock);
405 * Get this: if ldlm_blocking_ast is racing with intent_policy, such
406 * that ldlm_blocking_ast is called just before intent_policy method
407 * takes the lr_lock, then by the time we get the lock, we might not
408 * be the correct blocking function anymore. So check, and return
411 if (lock->l_blocking_ast != ldlm_blocking_ast) {
412 unlock_res_and_lock(lock);
415 RETURN(ldlm_blocking_ast_nocheck(lock));
417 EXPORT_SYMBOL(ldlm_blocking_ast);
420 * Implements ldlm_lock::l_glimpse_ast for extent locks acquired on the server.
422 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for that is
423 * rather subtle: with OST-side locking, it may so happen that _all_ extent
424 * locks are held by the OST. If client wants to obtain the current file size
425 * it calls ll_glimpse_size(), and (as all locks are held only on the server),
426 * this dummy glimpse callback fires and does nothing. The client still
427 * receives the correct file size due to the following fragment of code in
428 * ldlm_cb_interpret():
430 * if (rc == -ELDLM_NO_LOCK_DATA) {
431 * LDLM_DEBUG(lock, "lost race - client has a lock but no"
433 * ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
436 * That is, after the glimpse returns this error, ofd_lvbo_update() is called
437 * and returns the updated file attributes from the inode to the client.
439 * See also comment in ofd_intent_policy() on why servers must set a non-NULL
440 * l_glimpse_ast when grabbing DLM locks. Otherwise, the server will assume
441 * that the object is in the process of being destroyed.
443 * \param[in] lock DLM lock being glimpsed, unused
444 * \param[in] reqp pointer to ptlrpc_request, unused
446 * \retval -ELDLM_NO_LOCK_DATA to get attributes from disk object
448 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
450 return -ELDLM_NO_LOCK_DATA;
454 * Enqueue a local lock (typically on a server).
456 int ldlm_cli_enqueue_local(const struct lu_env *env,
457 struct ldlm_namespace *ns,
458 const struct ldlm_res_id *res_id,
459 enum ldlm_type type, union ldlm_policy_data *policy,
460 enum ldlm_mode mode, __u64 *flags,
461 ldlm_blocking_callback blocking,
462 ldlm_completion_callback completion,
463 ldlm_glimpse_callback glimpse,
464 void *data, __u32 lvb_len, enum lvb_type lvb_type,
465 const __u64 *client_cookie,
466 struct lustre_handle *lockh)
468 struct ldlm_lock *lock;
470 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
471 .lcs_blocking = blocking,
472 .lcs_glimpse = glimpse,
477 LASSERT(!(*flags & LDLM_FL_REPLAY));
478 if (unlikely(ns_is_client(ns))) {
479 CERROR("Trying to enqueue local lock in a shadow namespace\n");
483 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
486 GOTO(out_nolock, err = PTR_ERR(lock));
488 err = ldlm_lvbo_init(lock->l_resource);
490 LDLM_ERROR(lock, "delayed lvb init failed (rc %d)", err);
491 ldlm_lock_destroy_nolock(lock);
495 ldlm_lock2handle(lock, lockh);
498 * NB: we don't have any lock now (lock_res_and_lock)
499 * because it's a new lock
501 ldlm_lock_addref_internal_nolock(lock, mode);
502 ldlm_set_local(lock);
503 if (*flags & LDLM_FL_ATOMIC_CB)
504 ldlm_set_atomic_cb(lock);
506 if (*flags & LDLM_FL_CANCEL_ON_BLOCK)
507 ldlm_set_cancel_on_block(lock);
510 lock->l_policy_data = *policy;
511 if (client_cookie != NULL)
512 lock->l_client_cookie = *client_cookie;
513 if (type == LDLM_EXTENT) {
514 /* extent lock without policy is a bug */
518 lock->l_req_extent = policy->l_extent;
521 err = ldlm_lock_enqueue(env, ns, &lock, policy, flags);
522 if (unlikely(err != ELDLM_OK))
526 *policy = lock->l_policy_data;
528 if (lock->l_completion_ast)
529 lock->l_completion_ast(lock, *flags, NULL);
531 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
534 LDLM_LOCK_RELEASE(lock);
538 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
540 static void failed_lock_cleanup(struct ldlm_namespace *ns,
541 struct ldlm_lock *lock, int mode)
545 /* Set a flag to prevent us from sending a CANCEL (b=407) */
546 lock_res_and_lock(lock);
547 /* Check that lock is not granted or failed, we might race. */
548 if (!ldlm_is_granted(lock) && !ldlm_is_failed(lock)) {
550 * Make sure that this lock will not be found by raced
551 * bl_ast and -EINVAL reply is sent to server anyways.
554 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
555 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
558 unlock_res_and_lock(lock);
562 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
564 LDLM_DEBUG(lock, "lock was granted or failed in race");
567 * XXX - HACK because we shouldn't call ldlm_lock_destroy()
568 * from llite/file.c/ll_file_flock().
571 * This code makes for the fact that we do not have blocking handler on
572 * a client for flock locks. As such this is the place where we must
573 * completely kill failed locks. (interrupted and those that
574 * were waiting to be granted when server evicted us.
576 if (lock->l_resource->lr_type == LDLM_FLOCK) {
577 lock_res_and_lock(lock);
578 if (!ldlm_is_destroyed(lock)) {
579 ldlm_resource_unlink_lock(lock);
580 ldlm_lock_decref_internal_nolock(lock, mode);
581 ldlm_lock_destroy_nolock(lock);
583 unlock_res_and_lock(lock);
585 ldlm_lock_decref_internal(lock, mode);
589 static bool ldlm_request_slot_needed(enum ldlm_type type)
591 return type == LDLM_FLOCK || type == LDLM_IBITS;
595 * Finishing portion of client lock enqueue code.
597 * Called after receiving reply from server.
599 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
600 enum ldlm_type type, __u8 with_policy,
601 enum ldlm_mode mode, __u64 *flags, void *lvb,
602 __u32 lvb_len, const struct lustre_handle *lockh,
605 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
606 const struct lu_env *env = NULL;
607 int is_replay = *flags & LDLM_FL_REPLAY;
608 struct ldlm_lock *lock;
609 struct ldlm_reply *reply;
610 int cleanup_phase = 1;
614 if (ldlm_request_slot_needed(type))
615 obd_put_request_slot(&req->rq_import->imp_obd->u.cli);
617 ptlrpc_put_mod_rpc_slot(req);
619 if (req && req->rq_svc_thread)
620 env = req->rq_svc_thread->t_env;
622 lock = ldlm_handle2lock(lockh);
623 /* ldlm_cli_enqueue is holding a reference on this lock. */
625 LASSERT(type == LDLM_FLOCK);
629 LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
630 "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
632 if (rc != ELDLM_OK) {
634 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
635 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
637 if (rc != ELDLM_LOCK_ABORTED)
641 /* Before we return, swab the reply */
642 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
644 GOTO(cleanup, rc = -EPROTO);
649 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
652 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
653 GOTO(cleanup, rc = size);
654 } else if (unlikely(size > lvb_len)) {
656 "Replied LVB is larger than expectation, expected = %d, replied = %d",
658 GOTO(cleanup, rc = -EINVAL);
663 if (rc == ELDLM_LOCK_ABORTED) {
664 if (lvb_len > 0 && lvb != NULL)
665 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
667 GOTO(cleanup, rc = rc ? : ELDLM_LOCK_ABORTED);
670 /* lock enqueued on the server */
673 lock_res_and_lock(lock);
674 /* Key change rehash lock in per-export hash with new key */
675 if (exp->exp_lock_hash) {
677 * In the function below, .hs_keycmp resolves to
678 * ldlm_export_lock_keycmp()
680 /* coverity[overrun-buffer-val] */
681 cfs_hash_rehash_key(exp->exp_lock_hash,
682 &lock->l_remote_handle,
686 lock->l_remote_handle = reply->lock_handle;
689 *flags = ldlm_flags_from_wire(reply->lock_flags);
690 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
691 LDLM_FL_INHERIT_MASK);
692 unlock_res_and_lock(lock);
694 CDEBUG(D_INFO, "local: %p, remote cookie: %#llx, flags: %#llx\n",
695 lock, reply->lock_handle.cookie, *flags);
698 * If enqueue returned a blocked lock but the completion handler has
699 * already run, then it fixed up the resource and we don't need to do it
702 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
703 int newmode = reply->lock_desc.l_req_mode;
706 if (newmode && newmode != lock->l_req_mode) {
707 LDLM_DEBUG(lock, "server returned different mode %s",
708 ldlm_lockname[newmode]);
709 lock->l_req_mode = newmode;
712 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
713 &lock->l_resource->lr_name)) {
715 "remote intent success, locking "DLDLMRES", instead of "DLDLMRES"\n",
716 PLDLMRES(&reply->lock_desc.l_resource),
717 PLDLMRES(lock->l_resource));
719 rc = ldlm_lock_change_resource(ns, lock,
720 &reply->lock_desc.l_resource.lr_name);
721 if (rc || lock->l_resource == NULL)
722 GOTO(cleanup, rc = -ENOMEM);
723 LDLM_DEBUG(lock, "client-side enqueue, new resource");
727 /* We assume lock type cannot change on server*/
728 ldlm_convert_policy_to_local(exp,
729 lock->l_resource->lr_type,
730 &reply->lock_desc.l_policy_data,
731 &lock->l_policy_data);
734 if (type != LDLM_PLAIN)
736 "client-side enqueue, new policy data");
739 if ((*flags) & LDLM_FL_AST_SENT) {
740 lock_res_and_lock(lock);
741 ldlm_bl_desc2lock(&reply->lock_desc, lock);
742 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
743 unlock_res_and_lock(lock);
744 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
748 * If the lock has already been granted by a completion AST, don't
749 * clobber the LVB with an older one.
753 * We must lock or a racing completion might update lvb without
754 * letting us know and we'll clobber the correct value.
755 * Cannot unlock after the check either, a that still leaves
756 * a tiny window for completion to get in
758 lock_res_and_lock(lock);
759 if (!ldlm_is_granted(lock))
760 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
761 lock->l_lvb_data, lvb_len);
762 unlock_res_and_lock(lock);
770 rc = ldlm_lock_enqueue(env, ns, &lock, NULL, flags);
771 if (lock->l_completion_ast != NULL) {
772 int err = lock->l_completion_ast(lock, *flags, NULL);
781 if (lvb_len > 0 && lvb != NULL) {
783 * Copy the LVB here, and not earlier, because the completion
784 * AST (if any) can override what we got in the reply
786 memcpy(lvb, lock->l_lvb_data, lvb_len);
789 LDLM_DEBUG(lock, "client-side enqueue END");
792 if (cleanup_phase == 1 && rc)
793 failed_lock_cleanup(ns, lock, mode);
794 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
796 LDLM_LOCK_RELEASE(lock);
799 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
802 * Estimate number of lock handles that would fit into request of given
803 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
804 * a single page on the send/receive side. XXX: 512 should be changed to
805 * more adequate value.
807 static inline int ldlm_req_handles_avail(int req_size, int off)
811 avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512) - req_size;
812 if (likely(avail >= 0))
813 avail /= (int)sizeof(struct lustre_handle);
816 avail += LDLM_LOCKREQ_HANDLES - off;
821 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
822 enum req_location loc,
825 __u32 size = req_capsule_msg_size(pill, loc);
827 return ldlm_req_handles_avail(size, off);
830 static inline int ldlm_format_handles_avail(struct obd_import *imp,
831 const struct req_format *fmt,
832 enum req_location loc, int off)
834 __u32 size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
836 return ldlm_req_handles_avail(size, off);
840 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
841 * \a count locks in \a cancels.
843 * This is to be called by functions preparing their own requests that
844 * might contain lists of locks to cancel in addition to actual operation
845 * that needs to be performed.
847 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
848 int version, int opc, int canceloff,
849 struct list_head *cancels, int count)
851 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
852 struct req_capsule *pill = &req->rq_pill;
853 struct ldlm_request *dlm = NULL;
855 enum ldlm_lru_flags lru_flags;
856 int avail, to_free, pack = 0;
863 if (ns_connect_cancelset(ns)) {
864 /* Estimate the amount of available space in the request. */
865 req_capsule_filled_sizes(pill, RCL_CLIENT);
866 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
868 lru_flags = LDLM_LRU_FLAG_NO_WAIT | (ns_connect_lru_resize(ns) ?
869 LDLM_LRU_FLAG_LRUR : LDLM_LRU_FLAG_AGED);
870 to_free = !ns_connect_lru_resize(ns) &&
871 opc == LDLM_ENQUEUE ? 1 : 0;
874 * Cancel LRU locks here _only_ if the server supports
875 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
876 * RPC, which will make us slower.
879 count += ldlm_cancel_lru_local(ns, cancels, to_free,
886 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
887 ldlm_request_bufsize(pack, opc));
890 rc = ptlrpc_request_pack(req, version, opc);
892 ldlm_lock_list_put(cancels, l_bl_ast, count);
896 if (ns_connect_cancelset(ns)) {
898 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
901 * Skip first lock handler in ldlm_request_pack(),
902 * this method will increment @lock_count according
903 * to the lock handle amount actually written to
906 dlm->lock_count = canceloff;
908 /* Pack into the request @pack lock handles. */
909 ldlm_cli_cancel_list(cancels, pack, req, 0);
910 /* Prepare and send separate cancel RPC for others. */
911 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
913 ldlm_lock_list_put(cancels, l_bl_ast, count);
917 EXPORT_SYMBOL(ldlm_prep_elc_req);
919 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
920 struct list_head *cancels, int count)
922 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
923 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
925 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
927 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
929 struct ptlrpc_request *req;
934 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
936 RETURN(ERR_PTR(-ENOMEM));
938 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
940 ptlrpc_request_free(req);
944 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
945 ptlrpc_request_set_replen(req);
948 EXPORT_SYMBOL(ldlm_enqueue_pack);
951 * Client-side lock enqueue.
953 * If a request has some specific initialisation it is passed in \a reqp,
954 * otherwise it is created in ldlm_cli_enqueue.
956 * Supports sync and async requests, pass \a async flag accordingly. If a
957 * request was created in ldlm_cli_enqueue and it is the async request,
958 * pass it to the caller in \a reqp.
960 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
961 struct ldlm_enqueue_info *einfo,
962 const struct ldlm_res_id *res_id,
963 union ldlm_policy_data const *policy, __u64 *flags,
964 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
965 struct lustre_handle *lockh, int async)
967 struct ldlm_namespace *ns;
968 struct ldlm_lock *lock;
969 struct ldlm_request *body;
970 int is_replay = *flags & LDLM_FL_REPLAY;
971 int req_passed_in = 1;
973 struct ptlrpc_request *req;
977 LASSERT(exp != NULL);
979 ns = exp->exp_obd->obd_namespace;
982 * If we're replaying this lock, just check some invariants.
983 * If we're creating a new lock, get everything all setup nice.
986 lock = ldlm_handle2lock_long(lockh, 0);
987 LASSERT(lock != NULL);
988 LDLM_DEBUG(lock, "client-side enqueue START");
989 LASSERT(exp == lock->l_conn_export);
991 const struct ldlm_callback_suite cbs = {
992 .lcs_completion = einfo->ei_cb_cp,
993 .lcs_blocking = einfo->ei_cb_bl,
994 .lcs_glimpse = einfo->ei_cb_gl
996 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
997 einfo->ei_mode, &cbs, einfo->ei_cbdata,
1000 RETURN(PTR_ERR(lock));
1002 if (einfo->ei_cb_created)
1003 einfo->ei_cb_created(lock);
1005 /* for the local lock, add the reference */
1006 ldlm_lock_addref_internal(lock, einfo->ei_mode);
1007 ldlm_lock2handle(lock, lockh);
1009 lock->l_policy_data = *policy;
1011 if (einfo->ei_type == LDLM_EXTENT) {
1012 /* extent lock without policy is a bug */
1016 lock->l_req_extent = policy->l_extent;
1018 LDLM_DEBUG(lock, "client-side enqueue START, flags %#llx",
1022 lock->l_conn_export = exp;
1023 lock->l_export = NULL;
1024 lock->l_blocking_ast = einfo->ei_cb_bl;
1025 lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL |
1026 LDLM_FL_ATOMIC_CB));
1027 lock->l_activity = ktime_get_real_seconds();
1029 /* lock not sent to server yet */
1030 if (reqp == NULL || *reqp == NULL) {
1031 req = ldlm_enqueue_pack(exp, lvb_len);
1033 failed_lock_cleanup(ns, lock, einfo->ei_mode);
1034 LDLM_LOCK_RELEASE(lock);
1035 RETURN(PTR_ERR(req));
1045 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
1047 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
1048 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
1051 if (*flags & LDLM_FL_NDELAY) {
1052 DEBUG_REQ(D_DLMTRACE, req, "enqueue lock with no delay");
1053 req->rq_no_resend = req->rq_no_delay = 1;
1055 * probably set a shorter timeout value and handle ETIMEDOUT
1056 * in osc_lock_upcall() correctly
1058 /* lustre_msg_set_timeout(req, req->rq_timeout / 2); */
1061 /* Dump lock data into the request buffer */
1062 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1063 ldlm_lock2desc(lock, &body->lock_desc);
1064 body->lock_flags = ldlm_flags_to_wire(*flags);
1065 body->lock_handle[0] = *lockh;
1067 /* extended LDLM opcodes in client stats */
1068 if (exp->exp_obd->obd_svc_stats != NULL) {
1069 bool glimpse = *flags & LDLM_FL_HAS_INTENT;
1071 /* OST glimpse has no intent buffer */
1072 if (req_capsule_has_field(&req->rq_pill, &RMF_LDLM_INTENT,
1074 struct ldlm_intent *it;
1076 it = req_capsule_client_get(&req->rq_pill,
1078 glimpse = (it && (it->opc == IT_GLIMPSE));
1082 ldlm_svc_get_eopc(body, exp->exp_obd->obd_svc_stats);
1084 lprocfs_counter_incr(exp->exp_obd->obd_svc_stats,
1086 LDLM_GLIMPSE_ENQUEUE);
1089 /* It is important to obtain modify RPC slot first (if applicable), so
1090 * that threads that are waiting for a modify RPC slot are not polluting
1091 * our rpcs in flight counter. */
1093 if (einfo->ei_enq_slot)
1094 ptlrpc_get_mod_rpc_slot(req);
1096 if (ldlm_request_slot_needed(einfo->ei_type)) {
1097 rc = obd_get_request_slot(&req->rq_import->imp_obd->u.cli);
1099 if (einfo->ei_enq_slot)
1100 ptlrpc_put_mod_rpc_slot(req);
1101 failed_lock_cleanup(ns, lock, einfo->ei_mode);
1102 LDLM_LOCK_RELEASE(lock);
1108 LASSERT(reqp != NULL);
1112 LDLM_DEBUG(lock, "sending request");
1114 rc = ptlrpc_queue_wait(req);
1116 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
1117 einfo->ei_mode, flags, lvb, lvb_len,
1121 * If ldlm_cli_enqueue_fini did not find the lock, we need to free
1122 * one reference that we took
1125 LDLM_LOCK_RELEASE(lock);
1130 if (!req_passed_in && req != NULL) {
1131 ptlrpc_req_finished(req);
1138 EXPORT_SYMBOL(ldlm_cli_enqueue);
1141 * Client-side IBITS lock convert.
1143 * Inform server that lock has been converted instead of canceling.
1144 * Server finishes convert on own side and does reprocess to grant
1145 * all related waiting locks.
1147 * Since convert means only ibits downgrading, client doesn't need to
1148 * wait for server reply to finish local converting process so this request
1149 * is made asynchronous.
1152 int ldlm_cli_convert_req(struct ldlm_lock *lock, __u32 *flags, __u64 new_bits)
1154 struct ldlm_request *body;
1155 struct ptlrpc_request *req;
1156 struct obd_export *exp = lock->l_conn_export;
1160 LASSERT(exp != NULL);
1163 * this is better to check earlier and it is done so already,
1164 * but this check is kept too as final one to issue an error
1165 * if any new code will miss such check.
1167 if (!exp_connect_lock_convert(exp)) {
1168 LDLM_ERROR(lock, "server doesn't support lock convert\n");
1172 if (lock->l_resource->lr_type != LDLM_IBITS) {
1173 LDLM_ERROR(lock, "convert works with IBITS locks only.");
1177 LDLM_DEBUG(lock, "client-side convert");
1179 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1180 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1185 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1186 body->lock_handle[0] = lock->l_remote_handle;
1188 body->lock_desc.l_req_mode = lock->l_req_mode;
1189 body->lock_desc.l_granted_mode = lock->l_granted_mode;
1191 body->lock_desc.l_policy_data.l_inodebits.bits = new_bits;
1192 body->lock_desc.l_policy_data.l_inodebits.cancel_bits = 0;
1194 body->lock_flags = ldlm_flags_to_wire(*flags);
1195 body->lock_count = 1;
1197 ptlrpc_request_set_replen(req);
1200 * Use cancel portals for convert as well as high-priority handling.
1202 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1203 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1205 ptlrpc_at_set_req_timeout(req);
1207 if (exp->exp_obd->obd_svc_stats != NULL)
1208 lprocfs_counter_incr(exp->exp_obd->obd_svc_stats,
1209 LDLM_CONVERT - LDLM_FIRST_OPC);
1211 ptlrpcd_add_req(req);
1216 * Cancel locks locally.
1218 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1219 * \retval LDLM_FL_CANCELING otherwise;
1220 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1222 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1224 __u64 rc = LDLM_FL_LOCAL_ONLY;
1228 if (lock->l_conn_export) {
1231 LDLM_DEBUG(lock, "client-side cancel");
1232 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL_LOCAL,
1235 /* Set this flag to prevent others from getting new references*/
1236 lock_res_and_lock(lock);
1237 ldlm_set_cbpending(lock);
1238 local_only = !!(lock->l_flags &
1239 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1240 ldlm_cancel_callback(lock);
1241 rc = (ldlm_is_bl_ast(lock)) ?
1242 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1243 unlock_res_and_lock(lock);
1247 "not sending request (at caller's instruction)\n");
1248 rc = LDLM_FL_LOCAL_ONLY;
1250 ldlm_lock_cancel(lock);
1252 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1253 LDLM_ERROR(lock, "Trying to cancel local lock");
1256 LDLM_DEBUG(lock, "server-side local cancel");
1257 ldlm_lock_cancel(lock);
1258 ldlm_reprocess_all(lock->l_resource, lock);
1265 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1267 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1268 struct list_head *head, int count)
1270 struct ldlm_request *dlm;
1271 struct ldlm_lock *lock;
1272 int max, packed = 0;
1276 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1277 LASSERT(dlm != NULL);
1279 /* Check the room in the request buffer. */
1280 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1281 sizeof(struct ldlm_request);
1282 max /= sizeof(struct lustre_handle);
1283 max += LDLM_LOCKREQ_HANDLES;
1284 LASSERT(max >= dlm->lock_count + count);
1287 * XXX: it would be better to pack lock handles grouped by resource.
1288 * so that the server cancel would call filter_lvbo_update() less
1291 list_for_each_entry(lock, head, l_bl_ast) {
1294 LASSERT(lock->l_conn_export);
1295 /* Pack the lock handle to the given request buffer. */
1296 LDLM_DEBUG(lock, "packing");
1297 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1300 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1305 * Prepare and send a batched cancel RPC. It will include \a count lock
1306 * handles of locks given in \a cancels list.
1308 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1309 int count, enum ldlm_cancel_flags flags)
1311 struct ptlrpc_request *req = NULL;
1312 struct obd_import *imp;
1318 LASSERT(exp != NULL);
1321 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1323 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1326 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1327 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1332 imp = class_exp2cliimp(exp);
1333 if (imp == NULL || imp->imp_invalid) {
1335 "skipping cancel on invalid import %p\n", imp);
1339 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1341 GOTO(out, rc = -ENOMEM);
1343 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1344 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1345 ldlm_request_bufsize(count, LDLM_CANCEL));
1347 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1349 ptlrpc_request_free(req);
1354 * If OSP want cancel cross-MDT lock, let's not block it in
1355 * in recovery, otherwise the lock will not released, if
1356 * the remote target is also in recovery, and it also need
1357 * this lock, it might cause deadlock.
1359 if (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS &&
1360 exp->exp_obd->obd_lu_dev != NULL &&
1361 exp->exp_obd->obd_lu_dev->ld_site != NULL) {
1362 struct lu_device *top_dev;
1364 top_dev = exp->exp_obd->obd_lu_dev->ld_site->ls_top_dev;
1365 if (top_dev != NULL &&
1366 top_dev->ld_obd->obd_recovering)
1367 req->rq_allow_replay = 1;
1370 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1371 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1372 ptlrpc_at_set_req_timeout(req);
1374 ldlm_cancel_pack(req, cancels, count);
1376 ptlrpc_request_set_replen(req);
1377 if (flags & LCF_ASYNC) {
1378 ptlrpcd_add_req(req);
1383 rc = ptlrpc_queue_wait(req);
1384 if (rc == LUSTRE_ESTALE) {
1386 "client/server (nid %s) out of sync -- not fatal\n",
1387 libcfs_nid2str(req->rq_import->imp_connection->c_peer.nid));
1389 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1390 req->rq_import_generation == imp->imp_generation) {
1391 ptlrpc_req_finished(req);
1393 } else if (rc != ELDLM_OK) {
1394 /* -ESHUTDOWN is common on umount */
1395 CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1396 "Got rc %d from cancel RPC: canceling anyway\n",
1404 ptlrpc_req_finished(req);
1407 return sent ? sent : rc;
1410 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1412 LASSERT(imp != NULL);
1413 return &imp->imp_obd->obd_namespace->ns_pool;
1417 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1419 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1421 struct obd_device *obd;
1426 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1427 !imp_connect_lru_resize(req->rq_import)))
1428 /* Do nothing for corner cases. */
1432 * In some cases RPC may contain SLV and limit zeroed out. This
1433 * is the case when server does not support LRU resize feature.
1434 * This is also possible in some recovery cases when server-side
1435 * reqs have no reference to the OBD export and thus access to
1436 * server-side namespace is not possible.
1438 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1439 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1440 DEBUG_REQ(D_HA, req,
1441 "Zero SLV or limit found (SLV=%llu, limit=%u)",
1442 lustre_msg_get_slv(req->rq_repmsg),
1443 lustre_msg_get_limit(req->rq_repmsg));
1447 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1448 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1449 obd = req->rq_import->imp_obd;
1452 * Set new SLV and limit in OBD fields to make them accessible
1453 * to the pool thread. We do not access obd_namespace and pool
1454 * directly here as there is no reliable way to make sure that
1455 * they are still alive at cleanup time. Evil races are possible
1456 * which may cause Oops at that time.
1458 write_lock(&obd->obd_pool_lock);
1459 obd->obd_pool_slv = new_slv;
1460 obd->obd_pool_limit = new_limit;
1461 write_unlock(&obd->obd_pool_lock);
1466 int ldlm_cli_convert(struct ldlm_lock *lock,
1467 enum ldlm_cancel_flags cancel_flags)
1471 LASSERT(!lock->l_readers && !lock->l_writers);
1472 LDLM_DEBUG(lock, "client lock convert START");
1474 if (lock->l_resource->lr_type == LDLM_IBITS) {
1475 lock_res_and_lock(lock);
1477 rc = ldlm_cli_inodebits_convert(lock, cancel_flags);
1478 } while (rc == -EAGAIN);
1479 unlock_res_and_lock(lock);
1482 LDLM_DEBUG(lock, "client lock convert END");
1485 EXPORT_SYMBOL(ldlm_cli_convert);
1488 * Client side lock cancel.
1490 * Lock must not have any readers or writers by this time.
1492 int ldlm_cli_cancel(const struct lustre_handle *lockh,
1493 enum ldlm_cancel_flags cancel_flags)
1495 struct obd_export *exp;
1496 enum ldlm_lru_flags lru_flags;
1497 int avail, count = 1;
1499 struct ldlm_namespace *ns;
1500 struct ldlm_lock *lock;
1505 lock = ldlm_handle2lock_long(lockh, 0);
1507 LDLM_DEBUG_NOLOCK("lock is already being destroyed");
1511 lock_res_and_lock(lock);
1512 LASSERT(!ldlm_is_converting(lock));
1514 /* Lock is being canceled and the caller doesn't want to wait */
1515 if (ldlm_is_canceling(lock)) {
1516 if (cancel_flags & LCF_ASYNC) {
1517 unlock_res_and_lock(lock);
1519 unlock_res_and_lock(lock);
1520 wait_event_idle(lock->l_waitq, is_bl_done(lock));
1522 LDLM_LOCK_RELEASE(lock);
1526 ldlm_set_canceling(lock);
1527 unlock_res_and_lock(lock);
1529 if (cancel_flags & LCF_LOCAL)
1530 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_LOCAL_CANCEL_PAUSE,
1533 rc = ldlm_cli_cancel_local(lock);
1534 if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
1535 LDLM_LOCK_RELEASE(lock);
1539 * Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1540 * RPC which goes to canceld portal, so we can cancel other LRU locks
1541 * here and send them all as one LDLM_CANCEL RPC.
1543 LASSERT(list_empty(&lock->l_bl_ast));
1544 list_add(&lock->l_bl_ast, &cancels);
1546 exp = lock->l_conn_export;
1547 if (exp_connect_cancelset(exp)) {
1548 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1553 ns = ldlm_lock_to_ns(lock);
1554 lru_flags = ns_connect_lru_resize(ns) ?
1555 LDLM_LRU_FLAG_LRUR : LDLM_LRU_FLAG_AGED;
1556 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1557 LCF_BL_AST, lru_flags);
1559 ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1562 EXPORT_SYMBOL(ldlm_cli_cancel);
1565 * Locally cancel up to \a count locks in list \a cancels.
1566 * Return the number of cancelled locks.
1568 int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
1569 enum ldlm_cancel_flags cancel_flags)
1572 struct ldlm_lock *lock, *next;
1573 int left = 0, bl_ast = 0;
1577 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1581 if (cancel_flags & LCF_LOCAL) {
1582 rc = LDLM_FL_LOCAL_ONLY;
1583 ldlm_lock_cancel(lock);
1585 rc = ldlm_cli_cancel_local(lock);
1588 * Until we have compound requests and can send LDLM_CANCEL
1589 * requests batched with generic RPCs, we need to send cancels
1590 * with the LDLM_FL_BL_AST flag in a separate RPC from
1591 * the one being generated now.
1593 if (!(cancel_flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1594 LDLM_DEBUG(lock, "Cancel lock separately");
1595 list_move(&lock->l_bl_ast, &head);
1599 if (rc == LDLM_FL_LOCAL_ONLY) {
1600 /* CANCEL RPC should not be sent to server. */
1601 list_del_init(&lock->l_bl_ast);
1602 LDLM_LOCK_RELEASE(lock);
1608 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1615 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1616 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1617 * readahead requests, ...)
1619 static enum ldlm_policy_res
1620 ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1621 int unused, int added, int count)
1623 enum ldlm_policy_res result = LDLM_POLICY_CANCEL_LOCK;
1626 * don't check added & count since we want to process all locks
1628 * It's fine to not take lock to access lock->l_resource since
1629 * the lock has already been granted so it won't change.
1631 switch (lock->l_resource->lr_type) {
1634 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) != 0)
1638 result = LDLM_POLICY_SKIP_LOCK;
1646 * Callback function for LRU-resize policy. Decides whether to keep
1647 * \a lock in LRU for current \a LRU size \a unused, added in current
1648 * scan \a added and number of locks to be preferably canceled \a count.
1650 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1652 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1654 static enum ldlm_policy_res ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1655 struct ldlm_lock *lock,
1656 int unused, int added,
1659 ktime_t cur = ktime_get();
1660 struct ldlm_pool *pl = &ns->ns_pool;
1665 * Stop LRU processing when we reach past @count or have checked all
1668 if (count && added >= count)
1669 return LDLM_POLICY_KEEP_LOCK;
1672 * Despite of the LV, It doesn't make sense to keep the lock which
1673 * is unused for ns_max_age time.
1675 if (ktime_after(ktime_get(),
1676 ktime_add(lock->l_last_used, ns->ns_max_age)))
1677 return LDLM_POLICY_CANCEL_LOCK;
1679 slv = ldlm_pool_get_slv(pl);
1680 lvf = ldlm_pool_get_lvf(pl);
1681 la = div_u64(ktime_to_ns(ktime_sub(cur, lock->l_last_used)),
1683 lv = lvf * la * unused;
1685 /* Inform pool about current CLV to see it via debugfs. */
1686 ldlm_pool_set_clv(pl, lv);
1689 * Stop when SLV is not yet come from server or lv is smaller than
1692 if (slv == 0 || lv < slv)
1693 return LDLM_POLICY_KEEP_LOCK;
1695 return LDLM_POLICY_CANCEL_LOCK;
1698 static enum ldlm_policy_res
1699 ldlm_cancel_lrur_no_wait_policy(struct ldlm_namespace *ns,
1700 struct ldlm_lock *lock,
1701 int unused, int added,
1704 enum ldlm_policy_res result;
1706 result = ldlm_cancel_lrur_policy(ns, lock, unused, added, count);
1707 if (result == LDLM_POLICY_KEEP_LOCK)
1710 return ldlm_cancel_no_wait_policy(ns, lock, unused, added, count);
1714 * Callback function for debugfs used policy. Makes decision whether to keep
1715 * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1716 * added and number of locks to be preferably canceled \a count.
1718 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1720 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1722 static enum ldlm_policy_res ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1723 struct ldlm_lock *lock,
1724 int unused, int added,
1728 * Stop LRU processing when we reach past @count or have checked all
1731 return (added >= count) ?
1732 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1736 * Callback function for aged policy. Makes decision whether to keep \a lock in
1737 * LRU for current LRU size \a unused, added in current scan \a added and
1738 * number of locks to be preferably canceled \a count.
1740 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1742 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1744 static enum ldlm_policy_res ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1745 struct ldlm_lock *lock,
1746 int unused, int added,
1749 if ((added >= count) &&
1750 ktime_before(ktime_get(),
1751 ktime_add(lock->l_last_used, ns->ns_max_age)))
1752 return LDLM_POLICY_KEEP_LOCK;
1754 return LDLM_POLICY_CANCEL_LOCK;
1757 static enum ldlm_policy_res
1758 ldlm_cancel_aged_no_wait_policy(struct ldlm_namespace *ns,
1759 struct ldlm_lock *lock,
1760 int unused, int added, int count)
1762 enum ldlm_policy_res result;
1764 result = ldlm_cancel_aged_policy(ns, lock, unused, added, count);
1765 if (result == LDLM_POLICY_KEEP_LOCK)
1768 return ldlm_cancel_no_wait_policy(ns, lock, unused, added, count);
1772 * Callback function for default policy. Makes decision whether to keep \a lock
1773 * in LRU for current LRU size \a unused, added in current scan \a added and
1774 * number of locks to be preferably canceled \a count.
1776 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1778 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1781 enum ldlm_policy_res ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1782 struct ldlm_lock *lock,
1783 int unused, int added,
1787 * Stop LRU processing when we reach past count or have checked all
1790 return (added >= count) ?
1791 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1794 typedef enum ldlm_policy_res
1795 (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1796 int unused, int added, int count);
1798 static ldlm_cancel_lru_policy_t
1799 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, enum ldlm_lru_flags lru_flags)
1801 if (ns_connect_lru_resize(ns)) {
1802 if (lru_flags & LDLM_LRU_FLAG_SHRINK)
1803 /* We kill passed number of old locks. */
1804 return ldlm_cancel_passed_policy;
1805 if (lru_flags & LDLM_LRU_FLAG_LRUR) {
1806 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1807 return ldlm_cancel_lrur_no_wait_policy;
1809 return ldlm_cancel_lrur_policy;
1811 if (lru_flags & LDLM_LRU_FLAG_PASSED)
1812 return ldlm_cancel_passed_policy;
1814 if (lru_flags & LDLM_LRU_FLAG_AGED) {
1815 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1816 return ldlm_cancel_aged_no_wait_policy;
1818 return ldlm_cancel_aged_policy;
1821 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1822 return ldlm_cancel_no_wait_policy;
1824 return ldlm_cancel_default_policy;
1828 * - Free space in LRU for \a count new locks,
1829 * redundant unused locks are canceled locally;
1830 * - also cancel locally unused aged locks;
1831 * - do not cancel more than \a max locks;
1832 * - GET the found locks and add them into the \a cancels list.
1834 * A client lock can be added to the l_bl_ast list only when it is
1835 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1836 * CANCEL. There are the following use cases:
1837 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1838 * ldlm_cli_cancel(), which check and set this flag properly. As any
1839 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1840 * later without any special locking.
1842 * Calling policies for enabled LRU resize:
1843 * ----------------------------------------
1844 * flags & LDLM_LRU_FLAG_LRUR - use LRU resize policy (SLV from server) to
1845 * cancel not more than \a count locks;
1847 * flags & LDLM_LRU_FLAG_PASSED - cancel \a count number of old locks (located
1848 * at the beginning of LRU list);
1850 * flags & LDLM_LRU_FLAG_SHRINK - cancel not more than \a count locks according
1851 * to memory pressre policy function;
1853 * flags & LDLM_LRU_FLAG_AGED - cancel \a count locks according to "aged policy"
1855 * flags & LDLM_LRU_FLAG_NO_WAIT - cancel as many unused locks as possible
1856 * (typically before replaying locks) w/o
1857 * sending any RPCs or waiting for any
1858 * outstanding RPC to complete.
1860 * flags & LDLM_CANCEL_CLEANUP - when cancelling read locks, do not check for
1861 * other read locks covering the same pages, just
1862 * discard those pages.
1864 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns,
1865 struct list_head *cancels, int count, int max,
1866 enum ldlm_lru_flags lru_flags)
1868 ldlm_cancel_lru_policy_t pf;
1870 int no_wait = lru_flags & LDLM_LRU_FLAG_NO_WAIT;
1874 if (!ns_connect_lru_resize(ns))
1875 count += ns->ns_nr_unused - ns->ns_max_unused;
1877 pf = ldlm_cancel_lru_policy(ns, lru_flags);
1878 LASSERT(pf != NULL);
1880 /* For any flags, stop scanning if @max is reached. */
1881 while (!list_empty(&ns->ns_unused_list) && (max == 0 || added < max)) {
1882 struct ldlm_lock *lock;
1883 struct list_head *item, *next;
1884 enum ldlm_policy_res result;
1885 ktime_t last_use = ktime_set(0, 0);
1887 spin_lock(&ns->ns_lock);
1888 item = no_wait ? ns->ns_last_pos : &ns->ns_unused_list;
1889 for (item = item->next, next = item->next;
1890 item != &ns->ns_unused_list;
1891 item = next, next = item->next) {
1892 lock = list_entry(item, struct ldlm_lock, l_lru);
1894 /* No locks which got blocking requests. */
1895 LASSERT(!ldlm_is_bl_ast(lock));
1897 if (!ldlm_is_canceling(lock))
1901 * Somebody is already doing CANCEL. No need for this
1902 * lock in LRU, do not traverse it again.
1904 ldlm_lock_remove_from_lru_nolock(lock);
1906 if (item == &ns->ns_unused_list) {
1907 spin_unlock(&ns->ns_lock);
1911 last_use = lock->l_last_used;
1913 LDLM_LOCK_GET(lock);
1914 spin_unlock(&ns->ns_lock);
1915 lu_ref_add(&lock->l_reference, __FUNCTION__, current);
1918 * Pass the lock through the policy filter and see if it
1919 * should stay in LRU.
1921 * Even for shrinker policy we stop scanning if
1922 * we find a lock that should stay in the cache.
1923 * We should take into account lock age anyway
1924 * as a new lock is a valuable resource even if
1925 * it has a low weight.
1927 * That is, for shrinker policy we drop only
1928 * old locks, but additionally choose them by
1929 * their weight. Big extent locks will stay in
1932 result = pf(ns, lock, ns->ns_nr_unused, added, count);
1933 if (result == LDLM_POLICY_KEEP_LOCK) {
1934 lu_ref_del(&lock->l_reference, __func__, current);
1935 LDLM_LOCK_RELEASE(lock);
1939 if (result == LDLM_POLICY_SKIP_LOCK) {
1940 lu_ref_del(&lock->l_reference, __func__, current);
1942 spin_lock(&ns->ns_lock);
1943 if (!list_empty(&lock->l_lru) &&
1944 lock->l_lru.prev == ns->ns_last_pos)
1945 ns->ns_last_pos = &lock->l_lru;
1946 spin_unlock(&ns->ns_lock);
1949 LDLM_LOCK_RELEASE(lock);
1953 lock_res_and_lock(lock);
1954 /* Check flags again under the lock. */
1955 if (ldlm_is_canceling(lock) ||
1956 ldlm_lock_remove_from_lru_check(lock, last_use) == 0) {
1958 * Another thread is removing lock from LRU, or
1959 * somebody is already doing CANCEL, or there
1960 * is a blocking request which will send cancel
1961 * by itself, or the lock is no longer unused or
1962 * the lock has been used since the pf() call and
1963 * pages could be put under it.
1965 unlock_res_and_lock(lock);
1966 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1967 LDLM_LOCK_RELEASE(lock);
1970 LASSERT(!lock->l_readers && !lock->l_writers);
1973 * If we have chosen to cancel this lock voluntarily, we
1974 * better send cancel notification to server, so that it
1975 * frees appropriate state. This might lead to a race
1976 * where while we are doing cancel here, server is also
1977 * silently cancelling this lock.
1979 ldlm_clear_cancel_on_block(lock);
1982 * Setting the CBPENDING flag is a little misleading,
1983 * but prevents an important race; namely, once
1984 * CBPENDING is set, the lock can accumulate no more
1985 * readers/writers. Since readers and writers are
1986 * already zero here, ldlm_lock_decref() won't see
1987 * this flag and call l_blocking_ast
1989 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1991 if ((lru_flags & LDLM_LRU_FLAG_CLEANUP) &&
1992 (lock->l_resource->lr_type == LDLM_EXTENT ||
1993 ldlm_has_dom(lock)) && lock->l_granted_mode == LCK_PR)
1994 ldlm_set_discard_data(lock);
1997 * We can't re-add to l_lru as it confuses the
1998 * refcounting in ldlm_lock_remove_from_lru() if an AST
1999 * arrives after we drop lr_lock below. We use l_bl_ast
2000 * and can't use l_pending_chain as it is used both on
2001 * server and client nevertheless b=5666 says it is
2002 * used only on server
2004 LASSERT(list_empty(&lock->l_bl_ast));
2005 list_add(&lock->l_bl_ast, cancels);
2006 unlock_res_and_lock(lock);
2007 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
2013 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
2015 enum ldlm_cancel_flags cancel_flags,
2016 enum ldlm_lru_flags lru_flags)
2020 added = ldlm_prepare_lru_list(ns, cancels, count, max, lru_flags);
2024 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
2028 * Cancel at least \a nr locks from given namespace LRU.
2030 * When called with LCF_ASYNC the blocking callback will be handled
2031 * in a thread and this function will return after the thread has been
2032 * asked to call the callback. When called with LCF_ASYNC the blocking
2033 * callback will be performed in this function.
2035 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
2036 enum ldlm_cancel_flags cancel_flags,
2037 enum ldlm_lru_flags lru_flags)
2045 * Just prepare the list of locks, do not actually cancel them yet.
2046 * Locks are cancelled later in a separate thread.
2048 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, lru_flags);
2049 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
2057 * Find and cancel locally unused locks found on resource, matched to the
2058 * given policy, mode. GET the found locks and add them into the \a cancels
2061 int ldlm_cancel_resource_local(struct ldlm_resource *res,
2062 struct list_head *cancels,
2063 union ldlm_policy_data *policy,
2064 enum ldlm_mode mode, __u64 lock_flags,
2065 enum ldlm_cancel_flags cancel_flags,
2068 struct ldlm_lock *lock;
2074 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
2075 if (opaque != NULL && lock->l_ast_data != opaque) {
2076 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
2077 lock->l_ast_data, opaque);
2081 if (lock->l_readers || lock->l_writers)
2085 * If somebody is already doing CANCEL, or blocking AST came
2086 * then skip this lock.
2088 if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock))
2091 if (lockmode_compat(lock->l_granted_mode, mode))
2095 * If policy is given and this is IBITS lock, add to list only
2096 * those locks that match by policy.
2098 if (policy && (lock->l_resource->lr_type == LDLM_IBITS)) {
2099 if (!(lock->l_policy_data.l_inodebits.bits &
2100 policy->l_inodebits.bits))
2102 /* Skip locks with DoM bit if it is not set in policy
2103 * to don't flush data by side-bits. Lock convert will
2104 * drop those bits separately.
2106 if (ldlm_has_dom(lock) &&
2107 !(policy->l_inodebits.bits & MDS_INODELOCK_DOM))
2111 /* See CBPENDING comment in ldlm_cancel_lru */
2112 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
2114 LASSERT(list_empty(&lock->l_bl_ast));
2115 list_add(&lock->l_bl_ast, cancels);
2116 LDLM_LOCK_GET(lock);
2121 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
2123 EXPORT_SYMBOL(ldlm_cancel_resource_local);
2126 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
2128 * If \a req is NULL, send CANCEL request to server with handles of locks
2129 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
2130 * separately per lock.
2131 * If \a req is not NULL, put handles of locks in \a cancels into the request
2132 * buffer at the offset \a off.
2133 * Destroy \a cancels at the end.
2135 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
2136 struct ptlrpc_request *req,
2137 enum ldlm_cancel_flags flags)
2139 struct ldlm_lock *lock;
2144 if (list_empty(cancels) || count == 0)
2148 * XXX: requests (both batched and not) could be sent in parallel.
2149 * Usually it is enough to have just 1 RPC, but it is possible that
2150 * there are too many locks to be cancelled in LRU or on a resource.
2151 * It would also speed up the case when the server does not support
2155 LASSERT(!list_empty(cancels));
2156 lock = list_entry(cancels->next, struct ldlm_lock,
2158 LASSERT(lock->l_conn_export);
2160 if (exp_connect_cancelset(lock->l_conn_export)) {
2163 ldlm_cancel_pack(req, cancels, count);
2165 res = ldlm_cli_cancel_req(lock->l_conn_export,
2169 res = ldlm_cli_cancel_req(lock->l_conn_export,
2174 CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
2175 "ldlm_cli_cancel_list: %d\n", res);
2180 ldlm_lock_list_put(cancels, l_bl_ast, res);
2182 LASSERT(count == 0);
2185 EXPORT_SYMBOL(ldlm_cli_cancel_list);
2188 * Cancel all locks on a resource that have 0 readers/writers.
2190 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
2191 * to notify the server.
2193 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
2194 const struct ldlm_res_id *res_id,
2195 union ldlm_policy_data *policy,
2196 enum ldlm_mode mode,
2197 enum ldlm_cancel_flags flags, void *opaque)
2199 struct ldlm_resource *res;
2206 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2208 /* This is not a problem. */
2209 CDEBUG(D_INFO, "No resource %llu\n", res_id->name[0]);
2213 LDLM_RESOURCE_ADDREF(res);
2214 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
2215 0, flags | LCF_BL_AST, opaque);
2216 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
2218 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
2221 LDLM_RESOURCE_DELREF(res);
2222 ldlm_resource_putref(res);
2225 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2227 struct ldlm_cli_cancel_arg {
2233 ldlm_cli_hash_cancel_unused(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2234 struct hlist_node *hnode, void *arg)
2236 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2237 struct ldlm_cli_cancel_arg *lc = arg;
2239 ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
2240 NULL, LCK_MINMODE, lc->lc_flags,
2242 /* must return 0 for hash iteration */
2247 * Cancel all locks on a namespace (or a specific resource, if given)
2248 * that have 0 readers/writers.
2250 * If flags & LCF_LOCAL, throw the locks away without trying
2251 * to notify the server.
2253 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
2254 const struct ldlm_res_id *res_id,
2255 enum ldlm_cancel_flags flags, void *opaque)
2257 struct ldlm_cli_cancel_arg arg = {
2259 .lc_opaque = opaque,
2267 if (res_id != NULL) {
2268 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
2272 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2273 ldlm_cli_hash_cancel_unused, &arg, 0);
2278 /* Lock iterators. */
2280 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
2283 struct list_head *tmp, *next;
2284 struct ldlm_lock *lock;
2285 int rc = LDLM_ITER_CONTINUE;
2290 RETURN(LDLM_ITER_CONTINUE);
2293 list_for_each_safe(tmp, next, &res->lr_granted) {
2294 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2296 if (iter(lock, closure) == LDLM_ITER_STOP)
2297 GOTO(out, rc = LDLM_ITER_STOP);
2300 list_for_each_safe(tmp, next, &res->lr_waiting) {
2301 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2303 if (iter(lock, closure) == LDLM_ITER_STOP)
2304 GOTO(out, rc = LDLM_ITER_STOP);
2311 struct iter_helper_data {
2312 ldlm_iterator_t iter;
2316 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2318 struct iter_helper_data *helper = closure;
2320 return helper->iter(lock, helper->closure);
2323 static int ldlm_res_iter_helper(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2324 struct hlist_node *hnode, void *arg)
2327 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2329 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2333 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2334 ldlm_iterator_t iter, void *closure)
2337 struct iter_helper_data helper = { .iter = iter, .closure = closure };
2339 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2340 ldlm_res_iter_helper, &helper, 0);
2345 * non-blocking function to manipulate a lock whose cb_data is being put away.
2346 * return 0: find no resource
2347 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2350 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2351 const struct ldlm_res_id *res_id,
2352 ldlm_iterator_t iter, void *data)
2354 struct ldlm_resource *res;
2359 LASSERTF(ns != NULL, "must pass in namespace\n");
2361 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2365 LDLM_RESOURCE_ADDREF(res);
2366 rc = ldlm_resource_foreach(res, iter, data);
2367 LDLM_RESOURCE_DELREF(res);
2368 ldlm_resource_putref(res);
2371 EXPORT_SYMBOL(ldlm_resource_iterate);
2374 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2376 struct list_head *list = closure;
2378 /* we use l_pending_chain here, because it's unused on clients. */
2379 LASSERTF(list_empty(&lock->l_pending_chain),
2380 "lock %p next %p prev %p\n",
2381 lock, &lock->l_pending_chain.next,
2382 &lock->l_pending_chain.prev);
2384 * b=9573: don't replay locks left after eviction, or
2385 * b=17614: locks being actively cancelled. Get a reference
2386 * on a lock so that it does not disapear under us (e.g. due to cancel)
2388 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_BL_DONE))) {
2389 list_add(&lock->l_pending_chain, list);
2390 LDLM_LOCK_GET(lock);
2393 return LDLM_ITER_CONTINUE;
2396 static int replay_lock_interpret(const struct lu_env *env,
2397 struct ptlrpc_request *req, void *args, int rc)
2399 struct ldlm_async_args *aa = args;
2400 struct ldlm_lock *lock;
2401 struct ldlm_reply *reply;
2402 struct obd_export *exp;
2405 atomic_dec(&req->rq_import->imp_replay_inflight);
2409 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2411 GOTO(out, rc = -EPROTO);
2413 lock = ldlm_handle2lock(&aa->lock_handle);
2415 CERROR("received replay ack for unknown local cookie %#llx remote cookie %#llx from server %s id %s\n",
2416 aa->lock_handle.cookie, reply->lock_handle.cookie,
2417 req->rq_export->exp_client_uuid.uuid,
2418 libcfs_id2str(req->rq_peer));
2419 GOTO(out, rc = -ESTALE);
2422 /* Key change rehash lock in per-export hash with new key */
2423 exp = req->rq_export;
2424 if (exp && exp->exp_lock_hash) {
2426 * In the function below, .hs_keycmp resolves to
2427 * ldlm_export_lock_keycmp()
2429 /* coverity[overrun-buffer-val] */
2430 cfs_hash_rehash_key(exp->exp_lock_hash,
2431 &lock->l_remote_handle,
2432 &reply->lock_handle,
2435 lock->l_remote_handle = reply->lock_handle;
2438 LDLM_DEBUG(lock, "replayed lock:");
2439 ptlrpc_import_recovery_state_machine(req->rq_import);
2440 LDLM_LOCK_PUT(lock);
2443 ptlrpc_connect_import(req->rq_import);
2448 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2450 struct ptlrpc_request *req;
2451 struct ldlm_async_args *aa;
2452 struct ldlm_request *body;
2458 /* b=11974: Do not replay a lock which is actively being canceled */
2459 if (ldlm_is_bl_done(lock)) {
2460 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2465 * If this is reply-less callback lock, we cannot replay it, since
2466 * server might have long dropped it, but notification of that event was
2467 * lost by network. (and server granted conflicting lock already)
2469 if (ldlm_is_cancel_on_block(lock)) {
2470 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2471 ldlm_lock_cancel(lock);
2476 * If granted mode matches the requested mode, this lock is granted.
2478 * If we haven't been granted anything and are on a resource list,
2479 * then we're blocked/waiting.
2481 * If we haven't been granted anything and we're NOT on a resource list,
2482 * then we haven't got a reply yet and don't have a known disposition.
2483 * This happens whenever a lock enqueue is the request that triggers
2486 if (ldlm_is_granted(lock))
2487 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2488 else if (!list_empty(&lock->l_res_link))
2489 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2491 flags = LDLM_FL_REPLAY;
2493 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2494 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2498 /* We're part of recovery, so don't wait for it. */
2499 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2501 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2502 ldlm_lock2desc(lock, &body->lock_desc);
2503 body->lock_flags = ldlm_flags_to_wire(flags);
2505 ldlm_lock2handle(lock, &body->lock_handle[0]);
2506 if (lock->l_lvb_len > 0)
2507 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2508 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2510 ptlrpc_request_set_replen(req);
2512 * notify the server we've replayed all requests.
2513 * also, we mark the request to be put on a dedicated
2514 * queue to be processed after all request replayes.
2517 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2519 LDLM_DEBUG(lock, "replaying lock:");
2521 atomic_inc(&req->rq_import->imp_replay_inflight);
2522 aa = ptlrpc_req_async_args(aa, req);
2523 aa->lock_handle = body->lock_handle[0];
2524 req->rq_interpret_reply = replay_lock_interpret;
2525 ptlrpcd_add_req(req);
2531 * Cancel as many unused locks as possible before replay. since we are
2532 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2535 * Called only in recovery before replaying locks. there is no need to
2536 * replay locks that are unused. since the clients may hold thousands of
2537 * cached unused locks, dropping the unused locks can greatly reduce the
2538 * load on the servers at recovery time.
2540 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2546 "Dropping as many unused locks as possible before replay for namespace %s (%d)\n",
2547 ldlm_ns_name(ns), ns->ns_nr_unused);
2550 * We don't need to care whether or not LRU resize is enabled
2551 * because the LDLM_LRU_FLAG_NO_WAIT policy doesn't use the
2554 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2555 LCF_LOCAL, LDLM_LRU_FLAG_NO_WAIT);
2557 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2558 canceled, ldlm_ns_name(ns));
2561 int ldlm_replay_locks(struct obd_import *imp)
2563 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2565 struct ldlm_lock *lock, *next;
2570 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2572 /* don't replay locks if import failed recovery */
2573 if (imp->imp_vbr_failed)
2576 /* ensure this doesn't fall to 0 before all have been queued */
2577 atomic_inc(&imp->imp_replay_inflight);
2579 if (ldlm_cancel_unused_locks_before_replay)
2580 ldlm_cancel_unused_locks_for_replay(ns);
2582 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2584 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2585 list_del_init(&lock->l_pending_chain);
2587 LDLM_LOCK_RELEASE(lock);
2588 continue; /* or try to do the rest? */
2590 rc = replay_one_lock(imp, lock);
2591 LDLM_LOCK_RELEASE(lock);
2594 atomic_dec(&imp->imp_replay_inflight);