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/
32 * This file contains Asynchronous System Trap (AST) handlers and related
33 * LDLM request-processing routines.
35 * An AST is a callback issued on a lock when its state is changed. There are
36 * several different types of ASTs (callbacks) registered for each lock:
38 * - completion AST: when a lock is enqueued by some process, but cannot be
39 * granted immediately due to other conflicting locks on the same resource,
40 * the completion AST is sent to notify the caller when the lock is
43 * - blocking AST: when a lock is granted to some process, if another process
44 * enqueues a conflicting (blocking) lock on a resource, a blocking AST is
45 * sent to notify the holder(s) of the lock(s) of the conflicting lock
46 * request. The lock holder(s) must release their lock(s) on that resource in
47 * a timely manner or be evicted by the server.
49 * - glimpse AST: this is used when a process wants information about a lock
50 * (i.e. the lock value block (LVB)) but does not necessarily require holding
51 * the lock. If the resource is locked, the lock holder(s) are sent glimpse
52 * ASTs and the LVB is returned to the caller, and lock holder(s) may CANCEL
53 * their lock(s) if they are idle. If the resource is not locked, the server
57 #define DEBUG_SUBSYSTEM S_LDLM
59 #include <linux/fs_struct.h>
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");
70 EXPORT_SYMBOL(ldlm_enqueue_min);
72 /* in client side, whether the cached locks will be canceled before replay */
73 unsigned int ldlm_cancel_unused_locks_before_replay = 1;
75 struct lock_wait_data {
76 struct ldlm_lock *lwd_lock;
80 struct ldlm_async_args {
81 struct lustre_handle lock_handle;
85 * ldlm_request_bufsize
87 * If opcode=LDLM_ENQUEUE, 1 slot is already occupied,
88 * LDLM_LOCKREQ_HANDLE -1 slots are available.
89 * Otherwise, LDLM_LOCKREQ_HANDLE slots are available.
94 * \retval size of the request buffer
96 static int ldlm_request_bufsize(int count, int type)
98 int avail = LDLM_LOCKREQ_HANDLES;
100 if (type == LDLM_ENQUEUE)
101 avail -= LDLM_ENQUEUE_CANCEL_OFF;
104 avail = (count - avail) * sizeof(struct lustre_handle);
108 return sizeof(struct ldlm_request) + avail;
111 static void ldlm_expired_completion_wait(struct lock_wait_data *lwd)
113 struct ldlm_lock *lock = lwd->lwd_lock;
114 struct obd_import *imp;
115 struct obd_device *obd;
118 if (lock->l_conn_export == NULL) {
119 static time64_t next_dump, last_dump;
122 "lock timed out (enqueued at %lld, %llds ago); not entering recovery in server code, just going back to sleep",
124 ktime_get_real_seconds() - lock->l_activity);
125 if (ktime_get_seconds() > next_dump) {
126 last_dump = next_dump;
127 next_dump = ktime_get_seconds() + 300;
128 ldlm_namespace_dump(D_DLMTRACE,
129 ldlm_lock_to_ns(lock));
131 libcfs_debug_dumplog();
136 obd = lock->l_conn_export->exp_obd;
137 imp = obd->u.cli.cl_import;
138 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
140 "lock timed out (enqueued at %lld, %llds ago), entering recovery for %s@%s",
142 ktime_get_real_seconds() - lock->l_activity,
143 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
148 int is_granted_or_cancelled_nolock(struct ldlm_lock *lock)
152 check_res_locked(lock->l_resource);
153 if (ldlm_is_granted(lock) && !ldlm_is_cp_reqd(lock))
155 else if (ldlm_is_failed(lock) || ldlm_is_cancel(lock))
159 EXPORT_SYMBOL(is_granted_or_cancelled_nolock);
162 * Calculate the Completion timeout (covering enqueue, BL AST, data flush,
163 * lock cancel, and their replies). Used for lock completion timeout on the
166 * \param[in] lock lock which is waiting the completion callback
168 * \retval timeout in seconds to wait for the server reply
171 * We use the same basis for both server side and client side functions
172 * from a single node.
174 static timeout_t ldlm_cp_timeout(struct ldlm_lock *lock)
177 struct obd_device *obd;
179 obd = class_exp2obd(lock->l_conn_export);
185 * Wait a long time for enqueue - server may have to callback a
186 * lock from another client. Server will evict the other client if it
187 * doesn't respond reasonably, and then give us the lock.
189 timeout = obd_at_get(obd, ldlm_lock_to_ns_at(lock));
190 return max(3 * timeout, (timeout_t)obd_get_ldlm_enqueue_min(obd));
194 * Helper function for ldlm_completion_ast(), updating timings when lock is
197 static int ldlm_completion_tail(struct ldlm_lock *lock, void *data)
201 if (ldlm_is_destroyed(lock) || ldlm_is_failed(lock)) {
202 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
204 } else if (data == NULL) {
205 LDLM_DEBUG(lock, "client-side enqueue: granted");
207 struct obd_device *obd = class_exp2obd(lock->l_conn_export);
208 /* Take into AT only CP RPC, not immediately granted locks */
211 /* Discard negative timeouts. We should also limit the
212 * maximum value of the timeout
214 if (ktime_get_real_seconds() > lock->l_activity)
215 delay = ktime_get_real_seconds() - lock->l_activity;
217 LDLM_DEBUG(lock, "client-side enqueue: granted after %ds",
219 /* Update our time estimate */
220 obd_at_measure(obd, ldlm_lock_to_ns_at(lock), delay);
226 * Implementation of ->l_completion_ast() for a client, that doesn't wait
227 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
228 * other threads that cannot block for long.
230 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
234 if (flags == LDLM_FL_WAIT_NOREPROC) {
235 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
239 if (!(flags & LDLM_FL_BLOCKED_MASK)) {
240 wake_up(&lock->l_waitq);
241 RETURN(ldlm_completion_tail(lock, data));
244 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, going forward");
245 ldlm_reprocess_all(lock->l_resource, 0);
248 EXPORT_SYMBOL(ldlm_completion_ast_async);
251 * Generic LDLM "completion" AST. This is called in several cases:
253 * - when a reply to an ENQUEUE RPC is received from the server
254 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
255 * this point (determined by flags);
257 * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
260 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
263 * - to force all locks when resource is destroyed (cleanup_resource());
265 * If lock is not granted in the first case, this function waits until second
266 * or penultimate cases happen in some other thread.
269 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
271 /* XXX ALLOCATE - 160 bytes */
272 struct lock_wait_data lwd = {
276 struct obd_device *obd;
277 struct obd_import *imp = NULL;
283 if (flags == LDLM_FL_WAIT_NOREPROC) {
284 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
288 if (!(flags & LDLM_FL_BLOCKED_MASK)) {
289 wake_up(&lock->l_waitq);
293 LDLM_DEBUG(lock, "client-side enqueue returned a blocked locksleeping");
297 obd = class_exp2obd(lock->l_conn_export);
299 /* if this is a local lock, then there is no import */
301 imp = obd->u.cli.cl_import;
303 timeout = ldlm_cp_timeout(lock);
306 lock->l_activity = ktime_get_real_seconds();
309 spin_lock(&imp->imp_lock);
310 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
311 spin_unlock(&imp->imp_lock);
314 if (ns_is_client(ldlm_lock_to_ns(lock)) &&
315 CFS_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
316 OBD_FAIL_LDLM_CP_BL_RACE | CFS_FAIL_ONCE)) {
317 ldlm_set_fail_loc(lock);
320 /* Go to sleep until the lock is granted or cancelled. */
321 if (ldlm_is_no_timeout(lock)) {
322 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
323 rc = l_wait_event_abortable(
325 is_granted_or_cancelled(lock));
327 if (wait_event_idle_timeout(
329 is_granted_or_cancelled(lock),
330 cfs_time_seconds(timeout)) == 0) {
331 ldlm_expired_completion_wait(&lwd);
332 rc = l_wait_event_abortable(
334 is_granted_or_cancelled(lock));
340 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
345 RETURN(ldlm_completion_tail(lock, data));
347 EXPORT_SYMBOL(ldlm_completion_ast);
350 * A helper to build a blocking AST function
352 * Perform a common operation for blocking ASTs:
353 * defferred lock cancellation.
355 * \param lock the lock blocking or canceling AST was called on
357 * \see mdt_blocking_ast
358 * \see ldlm_blocking_ast
360 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
366 ldlm_set_cbpending(lock);
367 do_ast = (!lock->l_readers && !lock->l_writers);
368 unlock_res_and_lock(lock);
371 struct lustre_handle lockh;
374 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
375 ldlm_lock2handle(lock, &lockh);
376 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
378 CERROR("ldlm_cli_cancel: %d\n", rc);
381 "Lock still has references, will be cancelled later");
385 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
388 * Server blocking AST
390 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
393 * \param lock the lock which blocks a request or cancelling lock
396 * \param flag indicates whether this cancelling or blocking callback
398 * \see ldlm_blocking_ast_nocheck
400 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
401 void *data, int flag)
405 if (flag == LDLM_CB_CANCELING) {
406 /* Don't need to do anything here. */
410 lock_res_and_lock(lock);
412 * Get this: if ldlm_blocking_ast is racing with intent_policy, such
413 * that ldlm_blocking_ast is called just before intent_policy method
414 * takes the lr_lock, then by the time we get the lock, we might not
415 * be the correct blocking function anymore. So check, and return
418 if (lock->l_blocking_ast != ldlm_blocking_ast) {
419 unlock_res_and_lock(lock);
422 RETURN(ldlm_blocking_ast_nocheck(lock));
424 EXPORT_SYMBOL(ldlm_blocking_ast);
427 * Implements ldlm_lock::l_glimpse_ast for extent locks acquired on the server.
429 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for that is
430 * rather subtle: with OST-side locking, it may so happen that _all_ extent
431 * locks are held by the OST. If client wants to obtain the current file size
432 * it calls ll_glimpse_size(), and (as all locks are held only on the server),
433 * this dummy glimpse callback fires and does nothing. The client still
434 * receives the correct file size due to the following fragment of code in
435 * ldlm_cb_interpret():
437 * if (rc == -ELDLM_NO_LOCK_DATA) {
438 * LDLM_DEBUG(lock, "lost race - client has a lock but no"
440 * ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
443 * That is, after the glimpse returns this error, ofd_lvbo_update() is called
444 * and returns the updated file attributes from the inode to the client.
446 * See also comment in ofd_intent_policy() on why servers must set a non-NULL
447 * l_glimpse_ast when grabbing DLM locks. Otherwise, the server will assume
448 * that the object is in the process of being destroyed.
450 * \param[in] lock DLM lock being glimpsed, unused
451 * \param[in] reqp pointer to ptlrpc_request, unused
453 * \retval -ELDLM_NO_LOCK_DATA to get attributes from disk object
455 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
457 return -ELDLM_NO_LOCK_DATA;
461 * Enqueue a local lock (typically on a server).
463 int ldlm_cli_enqueue_local(const struct lu_env *env,
464 struct ldlm_namespace *ns,
465 const struct ldlm_res_id *res_id,
466 enum ldlm_type type, union ldlm_policy_data *policy,
467 enum ldlm_mode mode, __u64 *flags,
468 ldlm_blocking_callback blocking,
469 ldlm_completion_callback completion,
470 ldlm_glimpse_callback glimpse,
471 void *data, __u32 lvb_len, enum lvb_type lvb_type,
472 const __u64 *client_cookie,
473 struct lustre_handle *lockh)
475 struct ldlm_lock *lock;
477 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
478 .lcs_blocking = blocking,
479 .lcs_glimpse = glimpse,
484 LASSERT(!(*flags & LDLM_FL_REPLAY));
485 if (unlikely(ns_is_client(ns))) {
486 CERROR("Trying to enqueue local lock in a shadow namespace\n");
490 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
493 GOTO(out_nolock, err = PTR_ERR(lock));
495 err = ldlm_lvbo_init(lock->l_resource);
497 LDLM_ERROR(lock, "delayed lvb init failed (rc %d)", err);
498 ldlm_lock_destroy_nolock(lock);
502 ldlm_lock2handle(lock, lockh);
505 * NB: we don't have any lock now (lock_res_and_lock)
506 * because it's a new lock
508 ldlm_lock_addref_internal_nolock(lock, mode);
509 ldlm_set_local(lock);
510 if (*flags & LDLM_FL_ATOMIC_CB)
511 ldlm_set_atomic_cb(lock);
513 if (*flags & LDLM_FL_CANCEL_ON_BLOCK)
514 ldlm_set_cancel_on_block(lock);
517 lock->l_policy_data = *policy;
518 if (client_cookie != NULL)
519 lock->l_client_cookie = *client_cookie;
520 if (type == LDLM_EXTENT) {
521 /* extent lock without policy is a bug */
525 lock->l_req_extent = policy->l_extent;
528 err = ldlm_lock_enqueue(env, ns, &lock, policy, flags);
529 if (unlikely(err != ELDLM_OK))
533 *policy = lock->l_policy_data;
535 if (lock->l_completion_ast)
536 lock->l_completion_ast(lock, *flags, NULL);
538 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
541 LDLM_LOCK_RELEASE(lock);
545 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
547 static void failed_lock_cleanup(struct ldlm_namespace *ns,
548 struct ldlm_lock *lock, int mode)
552 /* Set a flag to prevent us from sending a CANCEL (b=407) */
553 lock_res_and_lock(lock);
554 /* Check that lock is not granted or failed, we might race. */
555 if (!ldlm_is_granted(lock) && !ldlm_is_failed(lock)) {
557 * Make sure that this lock will not be found by raced
558 * bl_ast and -EINVAL reply is sent to server anyways.
561 lock->l_flags |= LDLM_FL_FAILED |
562 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
563 if (!(ldlm_is_bl_ast(lock) &&
564 lock->l_remote_handle.cookie != 0))
565 lock->l_flags |= LDLM_FL_LOCAL_ONLY;
568 unlock_res_and_lock(lock);
572 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
574 LDLM_DEBUG(lock, "lock was granted or failed in race");
577 * XXX - HACK because we shouldn't call ldlm_lock_destroy()
578 * from llite/file.c/ll_file_flock().
581 * This code makes for the fact that we do not have blocking handler on
582 * a client for flock locks. As such this is the place where we must
583 * completely kill failed locks. (interrupted and those that
584 * were waiting to be granted when server evicted us.
586 if (lock->l_resource->lr_type == LDLM_FLOCK) {
587 lock_res_and_lock(lock);
588 if (!ldlm_is_destroyed(lock)) {
589 ldlm_resource_unlink_lock(lock);
590 ldlm_lock_decref_internal_nolock(lock, mode);
591 ldlm_lock_destroy_nolock(lock);
593 unlock_res_and_lock(lock);
595 ldlm_lock_decref_internal(lock, mode);
599 static bool ldlm_request_slot_needed(struct ldlm_enqueue_info *einfo)
601 /* exclude EXTENT locks and DOM-only IBITS locks because they
602 * are asynchronous and don't wait on server being blocked.
604 return einfo->ei_req_slot &&
605 (einfo->ei_type == LDLM_FLOCK ||
606 (einfo->ei_type == LDLM_IBITS &&
607 einfo->ei_inodebits != MDS_INODELOCK_DOM));
611 * Finishing portion of client lock enqueue code.
613 * Called after receiving reply from server.
615 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct req_capsule *pill,
616 struct ldlm_enqueue_info *einfo,
617 __u8 with_policy, __u64 *ldlm_flags, void *lvb,
618 __u32 lvb_len, const struct lustre_handle *lockh,
619 int rc, bool request_slot)
621 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
622 const struct lu_env *env = NULL;
623 int is_replay = *ldlm_flags & LDLM_FL_REPLAY;
624 struct ldlm_lock *lock;
625 struct ldlm_reply *reply;
626 int cleanup_phase = 1;
630 if (req_capsule_ptlreq(pill)) {
631 struct ptlrpc_request *req = pill->rc_req;
634 obd_put_request_slot(&req->rq_import->imp_obd->u.cli);
636 ptlrpc_put_mod_rpc_slot(req);
638 if (req && req->rq_svc_thread)
639 env = req->rq_svc_thread->t_env;
642 lock = ldlm_handle2lock(lockh);
643 /* ldlm_cli_enqueue is holding a reference on this lock. */
645 LASSERT(einfo->ei_type == LDLM_FLOCK);
649 LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
650 "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
652 if (rc != ELDLM_OK) {
654 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
655 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
657 if (rc != ELDLM_LOCK_ABORTED)
661 /* Before we return, swab the reply */
662 reply = req_capsule_server_get(pill, &RMF_DLM_REP);
664 GOTO(cleanup, rc = -EPROTO);
669 size = req_capsule_get_size(pill, &RMF_DLM_LVB, RCL_SERVER);
671 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
672 GOTO(cleanup, rc = size);
673 } else if (unlikely(size > lvb_len)) {
675 "Replied LVB is larger than expectation, expected = %d, replied = %d",
677 GOTO(cleanup, rc = -EINVAL);
682 if (rc == ELDLM_LOCK_ABORTED) {
683 if (lvb_len > 0 && lvb != NULL)
684 rc = ldlm_fill_lvb(lock, pill, RCL_SERVER,
686 GOTO(cleanup, rc = rc ? : ELDLM_LOCK_ABORTED);
689 /* lock enqueued on the server */
692 lock_res_and_lock(lock);
693 /* Key change rehash lock in per-export hash with new key */
694 if (exp->exp_lock_hash) {
696 * In the function below, .hs_keycmp resolves to
697 * ldlm_export_lock_keycmp()
699 cfs_hash_rehash_key(exp->exp_lock_hash,
700 &lock->l_remote_handle,
704 lock->l_remote_handle = reply->lock_handle;
707 *ldlm_flags = ldlm_flags_from_wire(reply->lock_flags);
708 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
709 LDLM_FL_INHERIT_MASK);
710 unlock_res_and_lock(lock);
712 CDEBUG(D_INFO, "local: %p, remote cookie: %#llx, flags: %#llx\n",
713 lock, reply->lock_handle.cookie, *ldlm_flags);
716 * If enqueue returned a blocked lock but the completion handler has
717 * already run, then it fixed up the resource and we don't need to do it
720 if ((*ldlm_flags) & LDLM_FL_LOCK_CHANGED) {
721 int newmode = reply->lock_desc.l_req_mode;
724 if (newmode && newmode != lock->l_req_mode) {
725 LDLM_DEBUG(lock, "server returned different mode %s",
726 ldlm_lockname[newmode]);
727 lock->l_req_mode = newmode;
730 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
731 &lock->l_resource->lr_name)) {
733 "remote intent success, locking "DLDLMRES", instead of "DLDLMRES"\n",
734 PLDLMRES(&reply->lock_desc.l_resource),
735 PLDLMRES(lock->l_resource));
737 rc = ldlm_lock_change_resource(ns, lock,
738 &reply->lock_desc.l_resource.lr_name);
739 if (rc || lock->l_resource == NULL)
740 GOTO(cleanup, rc = -ENOMEM);
741 LDLM_DEBUG(lock, "client-side enqueue, new resource");
745 /* We assume lock type cannot change on server*/
746 ldlm_convert_policy_to_local(exp,
747 lock->l_resource->lr_type,
748 &reply->lock_desc.l_policy_data,
749 &lock->l_policy_data);
752 if (einfo->ei_type != LDLM_PLAIN)
754 "client-side enqueue, new policy data");
757 if ((*ldlm_flags) & LDLM_FL_AST_SENT) {
758 lock_res_and_lock(lock);
759 ldlm_bl_desc2lock(&reply->lock_desc, lock);
760 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
761 unlock_res_and_lock(lock);
762 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
766 * If the lock has already been granted by a completion AST, don't
767 * clobber the LVB with an older one.
771 * We must lock or a racing completion might update lvb without
772 * letting us know and we'll clobber the correct value.
773 * Cannot unlock after the check either, a that still leaves
774 * a tiny window for completion to get in
776 lock_res_and_lock(lock);
777 if (!ldlm_is_granted(lock))
778 rc = ldlm_fill_lvb(lock, pill, RCL_SERVER,
779 lock->l_lvb_data, lvb_len);
780 unlock_res_and_lock(lock);
788 rc = ldlm_lock_enqueue(env, ns, &lock, NULL, ldlm_flags);
789 if (lock->l_completion_ast != NULL) {
790 int err = lock->l_completion_ast(lock, *ldlm_flags,
800 if (lvb_len > 0 && lvb != NULL) {
802 * Copy the LVB here, and not earlier, because the completion
803 * AST (if any) can override what we got in the reply
805 memcpy(lvb, lock->l_lvb_data, lvb_len);
808 LDLM_DEBUG(lock, "client-side enqueue END");
811 if (cleanup_phase == 1 && rc)
812 failed_lock_cleanup(ns, lock, einfo->ei_mode);
813 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
815 LDLM_LOCK_RELEASE(lock);
818 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
821 * Estimate number of lock handles that would fit into request of given
822 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
823 * a single page on the send/receive side. XXX: 512 should be changed to
824 * more adequate value.
826 static inline int ldlm_req_handles_avail(int req_size, int off)
830 avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512) - req_size;
831 if (likely(avail >= 0))
832 avail /= (int)sizeof(struct lustre_handle);
835 avail += LDLM_LOCKREQ_HANDLES - off;
840 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
841 enum req_location loc,
844 __u32 size = req_capsule_msg_size(pill, loc);
846 return ldlm_req_handles_avail(size, off);
849 static inline int ldlm_format_handles_avail(struct obd_import *imp,
850 const struct req_format *fmt,
851 enum req_location loc, int off)
853 __u32 size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
855 return ldlm_req_handles_avail(size, off);
859 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
860 * \a count locks in \a cancels.
862 * This is to be called by functions preparing their own requests that
863 * might contain lists of locks to cancel in addition to actual operation
864 * that needs to be performed.
866 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
867 int version, int opc, int canceloff,
868 struct list_head *cancels, int count)
870 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
871 struct req_capsule *pill = &req->rq_pill;
872 struct ldlm_request *dlm = NULL;
874 int avail, to_free = 0, pack = 0;
881 if (ns_connect_cancelset(ns)) {
882 /* Estimate the amount of available space in the request. */
883 req_capsule_filled_sizes(pill, RCL_CLIENT);
884 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
886 /* If we have reached the limit, free +1 slot for the new one */
887 if (!ns_connect_lru_resize(ns) && opc == LDLM_ENQUEUE &&
888 ns->ns_nr_unused >= ns->ns_max_unused)
892 * Cancel LRU locks here _only_ if the server supports
893 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
894 * RPC, which will make us slower.
897 count += ldlm_cancel_lru_local(ns, cancels, to_free,
899 LDLM_LRU_FLAG_NO_WAIT);
904 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
905 ldlm_request_bufsize(pack, opc));
908 rc = ptlrpc_request_pack(req, version, opc);
910 ldlm_lock_list_put(cancels, l_bl_ast, count);
914 if (ns_connect_cancelset(ns)) {
916 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
919 * Skip first lock handler in ldlm_request_pack(),
920 * this method will increment @lock_count according
921 * to the lock handle amount actually written to
924 dlm->lock_count = canceloff;
926 /* Pack into the request @pack lock handles. */
927 ldlm_cli_cancel_list(cancels, pack, req, 0);
928 /* Prepare and send separate cancel RPC for others. */
929 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
931 ldlm_lock_list_put(cancels, l_bl_ast, count);
935 EXPORT_SYMBOL(ldlm_prep_elc_req);
937 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
938 struct list_head *cancels, int count)
940 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
941 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
943 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
945 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
947 struct ptlrpc_request *req;
952 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
954 RETURN(ERR_PTR(-ENOMEM));
956 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
958 ptlrpc_request_free(req);
962 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
963 ptlrpc_request_set_replen(req);
966 EXPORT_SYMBOL(ldlm_enqueue_pack);
969 * Client-side lock enqueue.
971 * If a request has some specific initialisation it is passed in \a reqp,
972 * otherwise it is created in ldlm_cli_enqueue.
974 * Supports sync and async requests, pass \a async flag accordingly. If a
975 * request was created in ldlm_cli_enqueue and it is the async request,
976 * pass it to the caller in \a reqp.
978 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
979 struct ldlm_enqueue_info *einfo,
980 const struct ldlm_res_id *res_id,
981 union ldlm_policy_data const *policy, __u64 *flags,
982 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
983 struct lustre_handle *lockh, int async)
985 struct ldlm_namespace *ns;
986 struct ldlm_lock *lock;
987 struct ldlm_request *body;
988 int is_replay = *flags & LDLM_FL_REPLAY;
989 int req_passed_in = 1;
992 struct ptlrpc_request *req;
996 LASSERT(exp != NULL);
998 ns = exp->exp_obd->obd_namespace;
1001 * If we're replaying this lock, just check some invariants.
1002 * If we're creating a new lock, get everything all setup nice.
1005 lock = ldlm_handle2lock_long(lockh, 0);
1006 LASSERT(lock != NULL);
1007 LDLM_DEBUG(lock, "client-side enqueue START");
1008 LASSERT(exp == lock->l_conn_export);
1010 const struct ldlm_callback_suite cbs = {
1011 .lcs_completion = einfo->ei_cb_cp,
1012 .lcs_blocking = einfo->ei_cb_bl,
1013 .lcs_glimpse = einfo->ei_cb_gl
1015 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
1016 einfo->ei_mode, &cbs, einfo->ei_cbdata,
1019 RETURN(PTR_ERR(lock));
1021 if (einfo->ei_cb_created)
1022 einfo->ei_cb_created(lock);
1024 /* for the local lock, add the reference */
1025 ldlm_lock_addref_internal(lock, einfo->ei_mode);
1026 ldlm_lock2handle(lock, lockh);
1028 lock->l_policy_data = *policy;
1030 if (einfo->ei_type == LDLM_EXTENT) {
1031 /* extent lock without policy is a bug */
1035 lock->l_req_extent = policy->l_extent;
1037 LDLM_DEBUG(lock, "client-side enqueue START, flags %#llx",
1041 lock->l_conn_export = exp;
1042 lock->l_export = NULL;
1043 lock->l_blocking_ast = einfo->ei_cb_bl;
1044 lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
1045 lock->l_activity = ktime_get_real_seconds();
1047 /* lock not sent to server yet */
1048 if (reqp == NULL || *reqp == NULL) {
1049 req = ldlm_enqueue_pack(exp, lvb_len);
1051 failed_lock_cleanup(ns, lock, einfo->ei_mode);
1052 LDLM_LOCK_RELEASE(lock);
1053 RETURN(PTR_ERR(req));
1063 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
1065 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
1066 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
1069 if (*flags & LDLM_FL_NDELAY) {
1070 DEBUG_REQ(D_DLMTRACE, req, "enqueue lock with no delay");
1071 req->rq_no_resend = req->rq_no_delay = 1;
1073 * probably set a shorter timeout value and handle ETIMEDOUT
1074 * in osc_lock_upcall() correctly
1076 /* lustre_msg_set_timeout(req, req->rq_timeout / 2); */
1079 /* Dump lock data into the request buffer */
1080 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1081 ldlm_lock2desc(lock, &body->lock_desc);
1082 body->lock_flags = ldlm_flags_to_wire(*flags);
1083 body->lock_handle[0] = *lockh;
1085 /* extended LDLM opcodes in client stats */
1086 if (exp->exp_obd->obd_svc_stats != NULL) {
1087 /* glimpse is intent with no intent buffer */
1088 if (*flags & LDLM_FL_HAS_INTENT &&
1089 !req_capsule_has_field(&req->rq_pill, &RMF_LDLM_INTENT,
1091 lprocfs_counter_incr(exp->exp_obd->obd_svc_stats,
1093 LDLM_GLIMPSE_ENQUEUE);
1095 ldlm_svc_get_eopc(body, exp->exp_obd->obd_svc_stats);
1098 /* It is important to obtain modify RPC slot first (if applicable), so
1099 * that threads that are waiting for a modify RPC slot are not polluting
1100 * our rpcs in flight counter. */
1102 if (einfo->ei_mod_slot)
1103 ptlrpc_get_mod_rpc_slot(req);
1105 need_req_slot = ldlm_request_slot_needed(einfo);
1107 if (need_req_slot) {
1108 rc = obd_get_request_slot(&req->rq_import->imp_obd->u.cli);
1110 if (einfo->ei_mod_slot)
1111 ptlrpc_put_mod_rpc_slot(req);
1112 failed_lock_cleanup(ns, lock, einfo->ei_mode);
1113 LDLM_LOCK_RELEASE(lock);
1115 ptlrpc_req_finished(req);
1121 LASSERT(reqp != NULL);
1125 LDLM_DEBUG(lock, "sending request");
1127 rc = ptlrpc_queue_wait(req);
1129 err = ldlm_cli_enqueue_fini(exp, &req->rq_pill, einfo, policy ? 1 : 0,
1130 flags, lvb, lvb_len, lockh, rc,
1134 * If ldlm_cli_enqueue_fini did not find the lock, we need to free
1135 * one reference that we took
1138 LDLM_LOCK_RELEASE(lock);
1143 if (!req_passed_in && req != NULL) {
1144 ptlrpc_req_finished(req);
1151 EXPORT_SYMBOL(ldlm_cli_enqueue);
1154 * Client-side IBITS lock create and pack for WBC EX lock request.
1156 int ldlm_cli_lock_create_pack(struct obd_export *exp,
1157 struct ldlm_request *dlmreq,
1158 struct ldlm_enqueue_info *einfo,
1159 const struct ldlm_res_id *res_id,
1160 union ldlm_policy_data const *policy,
1161 __u64 *flags, void *lvb, __u32 lvb_len,
1162 enum lvb_type lvb_type,
1163 struct lustre_handle *lockh)
1165 const struct ldlm_callback_suite cbs = {
1166 .lcs_completion = einfo->ei_cb_cp,
1167 .lcs_blocking = einfo->ei_cb_bl,
1168 .lcs_glimpse = einfo->ei_cb_gl
1170 struct ldlm_namespace *ns;
1171 struct ldlm_lock *lock;
1175 LASSERT(exp != NULL);
1176 LASSERT(!(*flags & LDLM_FL_REPLAY));
1178 ns = exp->exp_obd->obd_namespace;
1179 lock = ldlm_lock_create(ns, res_id, einfo->ei_type, einfo->ei_mode,
1180 &cbs, einfo->ei_cbdata, lvb_len, lvb_type);
1182 RETURN(PTR_ERR(lock));
1184 if (einfo->ei_cb_created)
1185 einfo->ei_cb_created(lock);
1187 /* For the local lock, add the reference */
1188 ldlm_lock_addref_internal(lock, einfo->ei_mode);
1189 ldlm_lock2handle(lock, lockh);
1191 lock->l_policy_data = *policy;
1193 LDLM_DEBUG(lock, "client-side enqueue START, flags %#llx", *flags);
1194 lock->l_conn_export = exp;
1195 lock->l_export = NULL;
1196 lock->l_blocking_ast = einfo->ei_cb_bl;
1197 lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL |
1198 LDLM_FL_ATOMIC_CB));
1199 lock->l_activity = ktime_get_real_seconds();
1201 ldlm_lock2desc(lock, &dlmreq->lock_desc);
1202 dlmreq->lock_flags = ldlm_flags_to_wire(*flags);
1203 dlmreq->lock_handle[0] = *lockh;
1207 EXPORT_SYMBOL(ldlm_cli_lock_create_pack);
1210 * Client-side IBITS lock convert.
1212 * Inform server that lock has been converted instead of canceling.
1213 * Server finishes convert on own side and does reprocess to grant
1214 * all related waiting locks.
1216 * Since convert means only ibits downgrading, client doesn't need to
1217 * wait for server reply to finish local converting process so this request
1218 * is made asynchronous.
1221 int ldlm_cli_convert_req(struct ldlm_lock *lock, __u32 *flags, __u64 new_bits)
1223 struct ldlm_request *body;
1224 struct ptlrpc_request *req;
1225 struct obd_export *exp = lock->l_conn_export;
1229 LASSERT(exp != NULL);
1232 * this is better to check earlier and it is done so already,
1233 * but this check is kept too as final one to issue an error
1234 * if any new code will miss such check.
1236 if (!exp_connect_lock_convert(exp)) {
1237 LDLM_ERROR(lock, "server doesn't support lock convert\n");
1241 if (lock->l_resource->lr_type != LDLM_IBITS) {
1242 LDLM_ERROR(lock, "convert works with IBITS locks only.");
1246 LDLM_DEBUG(lock, "client-side convert");
1248 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1249 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1254 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1255 body->lock_handle[0] = lock->l_remote_handle;
1257 body->lock_desc.l_req_mode = lock->l_req_mode;
1258 body->lock_desc.l_granted_mode = lock->l_granted_mode;
1260 body->lock_desc.l_policy_data.l_inodebits.bits = new_bits;
1261 body->lock_desc.l_policy_data.l_inodebits.cancel_bits = 0;
1263 body->lock_flags = ldlm_flags_to_wire(*flags);
1264 body->lock_count = 1;
1266 ptlrpc_request_set_replen(req);
1269 * Use cancel portals for convert as well as high-priority handling.
1271 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1272 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1274 ptlrpc_at_set_req_timeout(req);
1276 if (exp->exp_obd->obd_svc_stats != NULL)
1277 lprocfs_counter_incr(exp->exp_obd->obd_svc_stats,
1278 LDLM_CONVERT - LDLM_FIRST_OPC);
1280 ptlrpcd_add_req(req);
1285 * Cancel locks locally.
1287 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1288 * \retval LDLM_FL_CANCELING otherwise;
1289 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1291 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1293 __u64 rc = LDLM_FL_LOCAL_ONLY;
1297 if (lock->l_conn_export) {
1300 LDLM_DEBUG(lock, "client-side cancel");
1301 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL_LOCAL,
1304 /* Set this flag to prevent others from getting new references*/
1305 lock_res_and_lock(lock);
1306 ldlm_set_cbpending(lock);
1307 local_only = !!(lock->l_flags &
1308 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1309 ldlm_cancel_callback(lock);
1310 rc = (ldlm_is_bl_ast(lock)) ?
1311 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1312 unlock_res_and_lock(lock);
1316 "not sending request (at caller's instruction)\n");
1317 rc = LDLM_FL_LOCAL_ONLY;
1319 ldlm_lock_cancel(lock);
1321 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1322 LDLM_ERROR(lock, "Trying to cancel local lock");
1325 LDLM_DEBUG(lock, "server-side local cancel");
1326 ldlm_lock_cancel(lock);
1327 ldlm_reprocess_all(lock->l_resource,
1328 lock->l_policy_data.l_inodebits.bits);
1334 static inline int __ldlm_pack_lock(struct ldlm_lock *lock,
1335 struct ldlm_request *dlm)
1337 LASSERT(lock->l_conn_export);
1338 lock_res_and_lock(lock);
1339 if (ldlm_is_ast_sent(lock)) {
1340 unlock_res_and_lock(lock);
1343 ldlm_set_ast_sent(lock);
1344 unlock_res_and_lock(lock);
1346 /* Pack the lock handle to the given request buffer. */
1347 LDLM_DEBUG(lock, "packing");
1348 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1352 #define ldlm_cancel_pack(req, head, count) \
1353 _ldlm_cancel_pack(req, NULL, head, count)
1355 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1357 static int _ldlm_cancel_pack(struct ptlrpc_request *req, struct ldlm_lock *lock,
1358 struct list_head *head, int count)
1360 struct ldlm_request *dlm;
1361 int max, packed = 0;
1365 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1366 LASSERT(dlm != NULL);
1368 /* Check the room in the request buffer. */
1369 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1370 sizeof(struct ldlm_request);
1371 max /= sizeof(struct lustre_handle);
1372 max += LDLM_LOCKREQ_HANDLES;
1373 LASSERT(max >= dlm->lock_count + count);
1376 * XXX: it would be better to pack lock handles grouped by resource.
1377 * so that the server cancel would call filter_lvbo_update() less
1380 if (lock) { /* only pack one lock */
1381 packed = __ldlm_pack_lock(lock, dlm);
1383 list_for_each_entry(lock, head, l_bl_ast) {
1386 packed += __ldlm_pack_lock(lock, dlm);
1389 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1394 * Prepare and send a batched cancel RPC. It will include \a count lock
1395 * handles of locks given in \a cancels list.
1397 int ldlm_cli_cancel_req(struct obd_export *exp, struct ldlm_lock *lock,
1398 struct list_head *head, int count,
1399 enum ldlm_cancel_flags flags)
1401 struct ptlrpc_request *req = NULL;
1402 struct obd_import *imp;
1408 LASSERT(exp != NULL);
1410 LASSERT(head == NULL || lock == NULL);
1412 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1414 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1417 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1418 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1423 imp = class_exp2cliimp(exp);
1424 if (imp == NULL || imp->imp_invalid) {
1426 "skipping cancel on invalid import %p\n", imp);
1430 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1432 GOTO(out, rc = -ENOMEM);
1434 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1435 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1436 ldlm_request_bufsize(count, LDLM_CANCEL));
1438 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1440 ptlrpc_request_free(req);
1445 * If OSP want cancel cross-MDT lock, let's not block it in
1446 * in recovery, otherwise the lock will not released, if
1447 * the remote target is also in recovery, and it also need
1448 * this lock, it might cause deadlock.
1450 if (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS &&
1451 exp->exp_obd->obd_lu_dev != NULL &&
1452 exp->exp_obd->obd_lu_dev->ld_site != NULL) {
1453 struct lu_device *top_dev;
1455 top_dev = exp->exp_obd->obd_lu_dev->ld_site->ls_top_dev;
1456 if (top_dev != NULL &&
1457 top_dev->ld_obd->obd_recovering)
1458 req->rq_allow_replay = 1;
1461 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1462 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1463 ptlrpc_at_set_req_timeout(req);
1465 rc = _ldlm_cancel_pack(req, lock, head, count);
1467 ptlrpc_req_finished(req);
1472 ptlrpc_request_set_replen(req);
1473 if (flags & LCF_ASYNC) {
1474 ptlrpcd_add_req(req);
1479 rc = ptlrpc_queue_wait(req);
1480 if (rc == LUSTRE_ESTALE) {
1482 "client/server (nid %s) out of sync -- not fatal\n",
1483 libcfs_nidstr(&req->rq_import->imp_connection->c_peer.nid));
1485 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1486 req->rq_import_generation == imp->imp_generation) {
1487 ptlrpc_req_finished(req);
1489 } else if (rc != ELDLM_OK) {
1490 /* -ESHUTDOWN is common on umount */
1491 CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1492 "Got rc %d from cancel RPC: canceling anyway\n",
1500 ptlrpc_req_finished(req);
1503 return sent ? sent : rc;
1506 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1508 LASSERT(imp != NULL);
1509 return &imp->imp_obd->obd_namespace->ns_pool;
1513 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1515 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1517 struct ldlm_namespace *ns;
1518 struct obd_device *obd;
1519 __u64 new_slv, ratio;
1523 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1524 !imp_connect_lru_resize(req->rq_import)))
1525 /* Do nothing for corner cases. */
1529 * In some cases RPC may contain SLV and limit zeroed out. This
1530 * is the case when server does not support LRU resize feature.
1531 * This is also possible in some recovery cases when server-side
1532 * reqs have no reference to the OBD export and thus access to
1533 * server-side namespace is not possible.
1535 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1536 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1537 DEBUG_REQ(D_HA, req,
1538 "Zero SLV or limit found (SLV=%llu, limit=%u)",
1539 lustre_msg_get_slv(req->rq_repmsg),
1540 lustre_msg_get_limit(req->rq_repmsg));
1544 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1545 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1546 obd = req->rq_import->imp_obd;
1548 read_lock(&obd->obd_pool_lock);
1549 if (obd->obd_pool_slv == new_slv &&
1550 obd->obd_pool_limit == new_limit) {
1551 read_unlock(&obd->obd_pool_lock);
1554 read_unlock(&obd->obd_pool_lock);
1557 * OBD device keeps the new pool attributes before they are handled by
1560 write_lock(&obd->obd_pool_lock);
1561 obd->obd_pool_slv = new_slv;
1562 obd->obd_pool_limit = new_limit;
1563 write_unlock(&obd->obd_pool_lock);
1566 * Check if an urgent pool recalc is needed, let it to be a change of
1567 * SLV on 10%. It is applicable to LRU resize enabled case only.
1569 ns = obd->obd_namespace;
1570 if (!ns_connect_lru_resize(ns) ||
1571 ldlm_pool_get_slv(&ns->ns_pool) < new_slv)
1574 ratio = 100 * new_slv / ldlm_pool_get_slv(&ns->ns_pool);
1575 if (100 - ratio >= ns->ns_recalc_pct &&
1576 !ns->ns_stopping && !ns->ns_rpc_recalc) {
1577 bool recalc = false;
1579 spin_lock(&ns->ns_lock);
1580 if (!ns->ns_stopping && !ns->ns_rpc_recalc) {
1581 ldlm_namespace_get(ns);
1583 ns->ns_rpc_recalc = 1;
1585 spin_unlock(&ns->ns_lock);
1587 ldlm_bl_to_thread_ns(ns);
1593 int ldlm_cli_convert(struct ldlm_lock *lock,
1594 enum ldlm_cancel_flags cancel_flags)
1598 LASSERT(!lock->l_readers && !lock->l_writers);
1599 LDLM_DEBUG(lock, "client lock convert START");
1601 if (lock->l_resource->lr_type == LDLM_IBITS) {
1602 lock_res_and_lock(lock);
1604 rc = ldlm_cli_inodebits_convert(lock, cancel_flags);
1605 } while (rc == -EAGAIN);
1606 unlock_res_and_lock(lock);
1609 LDLM_DEBUG(lock, "client lock convert END");
1612 EXPORT_SYMBOL(ldlm_cli_convert);
1615 * Client side lock cancel.
1617 * Lock must not have any readers or writers by this time.
1619 int ldlm_cli_cancel(const struct lustre_handle *lockh,
1620 enum ldlm_cancel_flags flags)
1622 struct obd_export *exp;
1623 int avail, count = 1, separate = 0;
1624 enum ldlm_lru_flags lru_flags = 0;
1626 struct ldlm_namespace *ns;
1627 struct ldlm_lock *lock;
1632 lock = ldlm_handle2lock_long(lockh, 0);
1634 LDLM_DEBUG_NOLOCK("lock is already being destroyed");
1638 lock_res_and_lock(lock);
1639 LASSERT(!ldlm_is_converting(lock));
1641 if (ldlm_is_bl_ast(lock)) {
1642 if (ldlm_is_ast_sent(lock)) {
1643 unlock_res_and_lock(lock);
1644 LDLM_LOCK_RELEASE(lock);
1647 if (ldlm_is_canceling(lock))
1649 } else if (ldlm_is_canceling(lock)) {
1650 /* Lock is being canceled and the caller doesn't want to wait */
1651 if (flags & LCF_ASYNC) {
1652 unlock_res_and_lock(lock);
1654 unlock_res_and_lock(lock);
1655 wait_event_idle(lock->l_waitq, is_bl_done(lock));
1657 LDLM_LOCK_RELEASE(lock);
1661 ldlm_set_canceling(lock);
1662 unlock_res_and_lock(lock);
1664 if (flags & LCF_LOCAL)
1665 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_LOCAL_CANCEL_PAUSE,
1668 rc = ldlm_cli_cancel_local(lock);
1669 if (rc == LDLM_FL_LOCAL_ONLY || flags & LCF_LOCAL) {
1670 LDLM_LOCK_RELEASE(lock);
1672 } else if (rc == LDLM_FL_BL_AST) {
1673 /* BL_AST lock must not wait. */
1674 lru_flags |= LDLM_LRU_FLAG_NO_WAIT;
1677 exp = lock->l_conn_export;
1678 /* If a lock has been taken from lru for a batched cancel and a later
1679 * BL_AST came, send a CANCEL RPC individually for it right away, not
1680 * waiting for the batch to be handled. */
1682 ldlm_cli_cancel_req(exp, lock, NULL, 1, flags);
1683 LDLM_LOCK_RELEASE(lock);
1687 LASSERT(list_empty(&lock->l_bl_ast));
1688 list_add(&lock->l_bl_ast, &cancels);
1690 * This is a LDLM_CANCEL RPC which goes to canceld portal,
1691 * so we can cancel other LRU locks here and send them all
1692 * as one LDLM_CANCEL RPC.
1694 if (exp_connect_cancelset(exp)) {
1695 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1700 ns = ldlm_lock_to_ns(lock);
1701 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1702 LCF_BL_AST, lru_flags);
1704 ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1708 EXPORT_SYMBOL(ldlm_cli_cancel);
1711 * Locally cancel up to \a count locks in list \a cancels.
1712 * Return the number of cancelled locks.
1714 int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
1715 enum ldlm_cancel_flags cancel_flags)
1718 struct ldlm_lock *lock, *next;
1719 int left = 0, bl_ast = 0;
1723 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1727 if (cancel_flags & LCF_LOCAL) {
1728 rc = LDLM_FL_LOCAL_ONLY;
1729 ldlm_lock_cancel(lock);
1731 rc = ldlm_cli_cancel_local(lock);
1734 * Until we have compound requests and can send LDLM_CANCEL
1735 * requests batched with generic RPCs, we need to send cancels
1736 * with the LDLM_FL_BL_AST flag in a separate RPC from
1737 * the one being generated now.
1739 if (!(cancel_flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1740 LDLM_DEBUG(lock, "Cancel lock separately");
1741 list_move(&lock->l_bl_ast, &head);
1745 if (rc == LDLM_FL_LOCAL_ONLY) {
1746 /* CANCEL RPC should not be sent to server. */
1747 list_del_init(&lock->l_bl_ast);
1748 LDLM_LOCK_RELEASE(lock);
1754 ldlm_cli_cancel_list(&head, bl_ast, NULL, cancel_flags);
1761 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1762 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1763 * readahead requests, ...)
1765 static enum ldlm_policy_res
1766 ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1769 enum ldlm_policy_res result = LDLM_POLICY_CANCEL_LOCK;
1772 * don't check @added & @min since we want to process all locks
1774 * It's fine to not take lock to access lock->l_resource since
1775 * the lock has already been granted so it won't change.
1777 switch (lock->l_resource->lr_type) {
1780 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) != 0)
1784 result = LDLM_POLICY_SKIP_LOCK;
1792 * Callback function for LRU-resize policy. Decides whether to keep
1793 * \a lock in LRU for \a added in current scan and \a min number of locks
1794 * to be preferably canceled.
1796 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1798 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1800 static enum ldlm_policy_res ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1801 struct ldlm_lock *lock,
1804 ktime_t cur = ktime_get();
1805 struct ldlm_pool *pl = &ns->ns_pool;
1810 return LDLM_POLICY_CANCEL_LOCK;
1813 * Despite of the LV, It doesn't make sense to keep the lock which
1814 * is unused for ns_max_age time.
1816 if (ktime_after(cur, ktime_add(lock->l_last_used, ns->ns_max_age)))
1817 return LDLM_POLICY_CANCEL_LOCK;
1819 slv = ldlm_pool_get_slv(pl);
1820 lvf = ldlm_pool_get_lvf(pl);
1821 la = div_u64(ktime_to_ns(ktime_sub(cur, lock->l_last_used)),
1823 lv = lvf * la * ns->ns_nr_unused >> 8;
1825 /* Inform pool about current CLV to see it via debugfs. */
1826 ldlm_pool_set_clv(pl, lv);
1829 * Stop when SLV is not yet come from server or lv is smaller than
1832 if (slv == 0 || lv < slv)
1833 return LDLM_POLICY_KEEP_LOCK;
1835 return LDLM_POLICY_CANCEL_LOCK;
1838 static enum ldlm_policy_res
1839 ldlm_cancel_lrur_no_wait_policy(struct ldlm_namespace *ns,
1840 struct ldlm_lock *lock,
1843 enum ldlm_policy_res result;
1845 result = ldlm_cancel_lrur_policy(ns, lock, added, min);
1846 if (result == LDLM_POLICY_KEEP_LOCK)
1849 return ldlm_cancel_no_wait_policy(ns, lock, added, min);
1853 * Callback function for aged policy. Decides whether to keep
1854 * \a lock in LRU for \a added in current scan and \a min number of locks
1855 * to be preferably canceled.
1857 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1859 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1861 static enum ldlm_policy_res ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1862 struct ldlm_lock *lock,
1865 if ((added >= min) &&
1866 ktime_before(ktime_get(),
1867 ktime_add(lock->l_last_used, ns->ns_max_age)))
1868 return LDLM_POLICY_KEEP_LOCK;
1870 return LDLM_POLICY_CANCEL_LOCK;
1873 static enum ldlm_policy_res
1874 ldlm_cancel_aged_no_wait_policy(struct ldlm_namespace *ns,
1875 struct ldlm_lock *lock,
1878 enum ldlm_policy_res result;
1880 result = ldlm_cancel_aged_policy(ns, lock, added, min);
1881 if (result == LDLM_POLICY_KEEP_LOCK)
1884 return ldlm_cancel_no_wait_policy(ns, lock, added, min);
1887 typedef enum ldlm_policy_res
1888 (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1889 int added, int min);
1891 static ldlm_cancel_lru_policy_t
1892 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, enum ldlm_lru_flags lru_flags)
1894 if (ns_connect_lru_resize(ns)) {
1895 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1896 return ldlm_cancel_lrur_no_wait_policy;
1898 return ldlm_cancel_lrur_policy;
1900 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1901 return ldlm_cancel_aged_no_wait_policy;
1903 return ldlm_cancel_aged_policy;
1908 * - Free space in LRU for \a min new locks,
1909 * redundant unused locks are canceled locally;
1910 * - also cancel locally unused aged locks;
1911 * - do not cancel more than \a max locks;
1912 * - if some locks are cancelled, try to cancel at least \a batch locks
1913 * - GET the found locks and add them into the \a cancels list.
1915 * A client lock can be added to the l_bl_ast list only when it is
1916 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1917 * CANCEL. There are the following use cases:
1918 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1919 * ldlm_cli_cancel(), which check and set this flag properly. As any
1920 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1921 * later without any special locking.
1923 * Locks are cancelled according to the LRU resize policy (SLV from server)
1924 * if LRU resize is enabled; otherwise, the "aged policy" is used;
1927 * ----------------------------------------
1929 * flags & LDLM_LRU_FLAG_NO_WAIT - cancel locks w/o sending any RPCs or waiting
1930 * for any outstanding RPC to complete.
1932 * flags & LDLM_CANCEL_CLEANUP - when cancelling read locks, do not check for
1933 * other read locks covering the same pages, just
1934 * discard those pages.
1936 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns,
1937 struct list_head *cancels,
1938 int min, int max, int batch,
1939 enum ldlm_lru_flags lru_flags)
1941 ldlm_cancel_lru_policy_t pf;
1943 int no_wait = lru_flags & LDLM_LRU_FLAG_NO_WAIT;
1947 * Let only 1 thread to proceed. However, not for those which have the
1948 * @max limit given (ELC), as LRU may be left not cleaned up in full.
1951 if (test_and_set_bit(LDLM_LRU_CANCEL, &ns->ns_flags))
1953 } else if (test_bit(LDLM_LRU_CANCEL, &ns->ns_flags))
1956 LASSERT(ergo(max, min <= max));
1957 /* No sense to give @batch for ELC */
1958 LASSERT(ergo(max, batch == 0));
1960 if (!ns_connect_lru_resize(ns))
1961 min = max_t(int, min, ns->ns_nr_unused - ns->ns_max_unused);
1963 /* If at least 1 lock is to be cancelled, cancel at least @batch locks */
1964 if (min && min < batch)
1967 pf = ldlm_cancel_lru_policy(ns, lru_flags);
1968 LASSERT(pf != NULL);
1970 /* For any flags, stop scanning if @max is reached. */
1971 while (!list_empty(&ns->ns_unused_list) && (max == 0 || added < max)) {
1972 struct ldlm_lock *lock;
1973 struct list_head *item, *next;
1974 enum ldlm_policy_res result;
1975 ktime_t last_use = ktime_set(0, 0);
1977 spin_lock(&ns->ns_lock);
1978 item = no_wait ? ns->ns_last_pos : &ns->ns_unused_list;
1979 for (item = item->next, next = item->next;
1980 item != &ns->ns_unused_list;
1981 item = next, next = item->next) {
1982 lock = list_entry(item, struct ldlm_lock, l_lru);
1984 /* No locks which got blocking requests. */
1985 LASSERT(!ldlm_is_bl_ast(lock));
1987 if (!ldlm_is_canceling(lock))
1991 * Somebody is already doing CANCEL. No need for this
1992 * lock in LRU, do not traverse it again.
1994 ldlm_lock_remove_from_lru_nolock(lock);
1996 if (item == &ns->ns_unused_list) {
1997 spin_unlock(&ns->ns_lock);
2001 last_use = lock->l_last_used;
2003 LDLM_LOCK_GET(lock);
2004 spin_unlock(&ns->ns_lock);
2005 lu_ref_add(&lock->l_reference, __FUNCTION__, current);
2008 * Pass the lock through the policy filter and see if it
2009 * should stay in LRU.
2011 * Even for shrinker policy we stop scanning if
2012 * we find a lock that should stay in the cache.
2013 * We should take into account lock age anyway
2014 * as a new lock is a valuable resource even if
2015 * it has a low weight.
2017 * That is, for shrinker policy we drop only
2018 * old locks, but additionally choose them by
2019 * their weight. Big extent locks will stay in
2022 result = pf(ns, lock, added, min);
2023 if (result == LDLM_POLICY_KEEP_LOCK) {
2024 lu_ref_del(&lock->l_reference, __func__, current);
2025 LDLM_LOCK_RELEASE(lock);
2029 if (result == LDLM_POLICY_SKIP_LOCK) {
2030 lu_ref_del(&lock->l_reference, __func__, current);
2032 spin_lock(&ns->ns_lock);
2033 if (!list_empty(&lock->l_lru) &&
2034 lock->l_lru.prev == ns->ns_last_pos)
2035 ns->ns_last_pos = &lock->l_lru;
2036 spin_unlock(&ns->ns_lock);
2039 LDLM_LOCK_RELEASE(lock);
2043 lock_res_and_lock(lock);
2044 /* Check flags again under the lock. */
2045 if (ldlm_is_canceling(lock) ||
2046 ldlm_lock_remove_from_lru_check(lock, last_use) == 0) {
2048 * Another thread is removing lock from LRU, or
2049 * somebody is already doing CANCEL, or there
2050 * is a blocking request which will send cancel
2051 * by itself, or the lock is no longer unused or
2052 * the lock has been used since the pf() call and
2053 * pages could be put under it.
2055 unlock_res_and_lock(lock);
2056 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
2057 LDLM_LOCK_RELEASE(lock);
2060 LASSERT(!lock->l_readers && !lock->l_writers);
2063 * If we have chosen to cancel this lock voluntarily, we
2064 * better send cancel notification to server, so that it
2065 * frees appropriate state. This might lead to a race
2066 * where while we are doing cancel here, server is also
2067 * silently cancelling this lock.
2069 ldlm_clear_cancel_on_block(lock);
2072 * Setting the CBPENDING flag is a little misleading,
2073 * but prevents an important race; namely, once
2074 * CBPENDING is set, the lock can accumulate no more
2075 * readers/writers. Since readers and writers are
2076 * already zero here, ldlm_lock_decref() won't see
2077 * this flag and call l_blocking_ast
2079 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
2081 if ((lru_flags & LDLM_LRU_FLAG_CLEANUP) &&
2082 (lock->l_resource->lr_type == LDLM_EXTENT ||
2083 ldlm_has_dom(lock)) && lock->l_granted_mode == LCK_PR)
2084 ldlm_set_discard_data(lock);
2087 * We can't re-add to l_lru as it confuses the
2088 * refcounting in ldlm_lock_remove_from_lru() if an AST
2089 * arrives after we drop lr_lock below. We use l_bl_ast
2090 * and can't use l_pending_chain as it is used both on
2091 * server and client nevertheless b=5666 says it is
2092 * used only on server
2094 LASSERT(list_empty(&lock->l_bl_ast));
2095 list_add(&lock->l_bl_ast, cancels);
2096 unlock_res_and_lock(lock);
2097 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
2099 /* Once a lock added, batch the requested amount */
2105 clear_bit(LDLM_LRU_CANCEL, &ns->ns_flags);
2110 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
2112 enum ldlm_cancel_flags cancel_flags,
2113 enum ldlm_lru_flags lru_flags)
2117 added = ldlm_prepare_lru_list(ns, cancels, min, max, 0, lru_flags);
2121 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
2125 * Cancel at least \a min locks from given namespace LRU.
2127 * When called with LCF_ASYNC the blocking callback will be handled
2128 * in a thread and this function will return after the thread has been
2129 * asked to call the callback. When called with LCF_ASYNC the blocking
2130 * callback will be performed in this function.
2132 int ldlm_cancel_lru(struct ldlm_namespace *ns, int min,
2133 enum ldlm_cancel_flags cancel_flags,
2134 enum ldlm_lru_flags lru_flags)
2142 * Just prepare the list of locks, do not actually cancel them yet.
2143 * Locks are cancelled later in a separate thread.
2145 count = ldlm_prepare_lru_list(ns, &cancels, min, 0,
2146 ns->ns_cancel_batch, lru_flags);
2147 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
2155 * Find and cancel locally unused locks found on resource, matched to the
2156 * given policy, mode. GET the found locks and add them into the \a cancels
2159 int ldlm_cancel_resource_local(struct ldlm_resource *res,
2160 struct list_head *cancels,
2161 union ldlm_policy_data *policy,
2162 enum ldlm_mode mode, __u64 lock_flags,
2163 enum ldlm_cancel_flags cancel_flags,
2166 struct ldlm_lock *lock;
2172 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
2173 if (opaque != NULL && lock->l_ast_data != opaque) {
2174 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
2175 lock->l_ast_data, opaque);
2179 if (lock->l_readers || lock->l_writers)
2183 * If somebody is already doing CANCEL, or blocking AST came
2184 * then skip this lock.
2186 if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock))
2189 if (lockmode_compat(lock->l_granted_mode, mode))
2193 * If policy is given and this is IBITS lock, add to list only
2194 * those locks that match by policy.
2196 if (policy && (lock->l_resource->lr_type == LDLM_IBITS)) {
2197 if (!(lock->l_policy_data.l_inodebits.bits &
2198 policy->l_inodebits.bits))
2200 /* Skip locks with DoM bit if it is not set in policy
2201 * to don't flush data by side-bits. Lock convert will
2202 * drop those bits separately.
2204 if (ldlm_has_dom(lock) &&
2205 !(policy->l_inodebits.bits & MDS_INODELOCK_DOM))
2209 /* See CBPENDING comment in ldlm_cancel_lru */
2210 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
2212 LASSERT(list_empty(&lock->l_bl_ast));
2213 list_add(&lock->l_bl_ast, cancels);
2214 LDLM_LOCK_GET(lock);
2219 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
2221 EXPORT_SYMBOL(ldlm_cancel_resource_local);
2224 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
2226 * If \a req is NULL, send CANCEL request to server with handles of locks
2227 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
2228 * separately per lock.
2229 * If \a req is not NULL, put handles of locks in \a cancels into the request
2230 * buffer at the offset \a off.
2231 * Destroy \a cancels at the end.
2233 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
2234 struct ptlrpc_request *req,
2235 enum ldlm_cancel_flags flags)
2237 struct ldlm_lock *lock;
2242 if (list_empty(cancels) || count == 0)
2246 * XXX: requests (both batched and not) could be sent in parallel.
2247 * Usually it is enough to have just 1 RPC, but it is possible that
2248 * there are too many locks to be cancelled in LRU or on a resource.
2249 * It would also speed up the case when the server does not support
2253 LASSERT(!list_empty(cancels));
2254 lock = list_first_entry(cancels, struct ldlm_lock, l_bl_ast);
2255 LASSERT(lock->l_conn_export);
2257 if (exp_connect_cancelset(lock->l_conn_export)) {
2260 ldlm_cancel_pack(req, cancels, count);
2262 res = ldlm_cli_cancel_req(lock->l_conn_export,
2263 NULL, cancels, count,
2266 res = ldlm_cli_cancel_req(lock->l_conn_export,
2267 NULL, cancels, 1, flags);
2271 CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
2272 "ldlm_cli_cancel_list: %d\n", res);
2277 ldlm_lock_list_put(cancels, l_bl_ast, res);
2279 LASSERT(count == 0);
2282 EXPORT_SYMBOL(ldlm_cli_cancel_list);
2285 * Cancel all locks on a resource that have 0 readers/writers.
2287 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
2288 * to notify the server.
2290 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
2291 const struct ldlm_res_id *res_id,
2292 union ldlm_policy_data *policy,
2293 enum ldlm_mode mode,
2294 enum ldlm_cancel_flags flags, void *opaque)
2296 struct ldlm_resource *res;
2303 res = ldlm_resource_get(ns, res_id, 0, 0);
2305 /* This is not a problem. */
2306 CDEBUG(D_INFO, "No resource %llu\n", res_id->name[0]);
2310 LDLM_RESOURCE_ADDREF(res);
2311 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
2312 0, flags | LCF_BL_AST, opaque);
2313 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
2315 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
2318 LDLM_RESOURCE_DELREF(res);
2319 ldlm_resource_putref(res);
2322 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2324 struct ldlm_cli_cancel_arg {
2330 ldlm_cli_hash_cancel_unused(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2331 struct hlist_node *hnode, void *arg)
2333 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2334 struct ldlm_cli_cancel_arg *lc = arg;
2336 ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
2337 NULL, LCK_MINMODE, lc->lc_flags,
2339 /* must return 0 for hash iteration */
2344 * Cancel all locks on a namespace (or a specific resource, if given)
2345 * that have 0 readers/writers.
2347 * If flags & LCF_LOCAL, throw the locks away without trying
2348 * to notify the server.
2350 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
2351 const struct ldlm_res_id *res_id,
2352 enum ldlm_cancel_flags flags, void *opaque)
2354 struct ldlm_cli_cancel_arg arg = {
2356 .lc_opaque = opaque,
2364 if (res_id != NULL) {
2365 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
2369 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2370 ldlm_cli_hash_cancel_unused, &arg, 0);
2375 /* Lock iterators. */
2377 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
2380 struct ldlm_lock *tmp;
2381 struct ldlm_lock *lock;
2382 int rc = LDLM_ITER_CONTINUE;
2386 RETURN(LDLM_ITER_CONTINUE);
2389 list_for_each_entry_safe(lock, tmp, &res->lr_granted, l_res_link) {
2390 if (iter(lock, closure) == LDLM_ITER_STOP)
2391 GOTO(out, rc = LDLM_ITER_STOP);
2394 list_for_each_entry_safe(lock, tmp, &res->lr_waiting, l_res_link) {
2395 if (iter(lock, closure) == LDLM_ITER_STOP)
2396 GOTO(out, rc = LDLM_ITER_STOP);
2403 struct iter_helper_data {
2404 ldlm_iterator_t iter;
2408 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2410 struct iter_helper_data *helper = closure;
2412 return helper->iter(lock, helper->closure);
2415 static int ldlm_res_iter_helper(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2416 struct hlist_node *hnode, void *arg)
2419 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2421 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2425 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2426 ldlm_iterator_t iter, void *closure)
2429 struct iter_helper_data helper = { .iter = iter, .closure = closure };
2431 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2432 ldlm_res_iter_helper, &helper, 0);
2437 * non-blocking function to manipulate a lock whose cb_data is being put away.
2438 * return 0: find no resource
2439 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2442 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2443 const struct ldlm_res_id *res_id,
2444 ldlm_iterator_t iter, void *data)
2446 struct ldlm_resource *res;
2451 LASSERTF(ns != NULL, "must pass in namespace\n");
2453 res = ldlm_resource_get(ns, res_id, 0, 0);
2457 LDLM_RESOURCE_ADDREF(res);
2458 rc = ldlm_resource_foreach(res, iter, data);
2459 LDLM_RESOURCE_DELREF(res);
2460 ldlm_resource_putref(res);
2463 EXPORT_SYMBOL(ldlm_resource_iterate);
2466 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2468 struct list_head *list = closure;
2470 /* we use l_pending_chain here, because it's unused on clients. */
2471 LASSERTF(list_empty(&lock->l_pending_chain),
2472 "lock %p next %p prev %p\n",
2473 lock, &lock->l_pending_chain.next,
2474 &lock->l_pending_chain.prev);
2476 * b=9573: don't replay locks left after eviction, or
2477 * b=17614: locks being actively cancelled. Get a reference
2478 * on a lock so that it does not disapear under us (e.g. due to cancel)
2480 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_BL_DONE))) {
2481 list_add(&lock->l_pending_chain, list);
2482 LDLM_LOCK_GET(lock);
2485 return LDLM_ITER_CONTINUE;
2488 static int replay_lock_interpret(const struct lu_env *env,
2489 struct ptlrpc_request *req, void *args, int rc)
2491 struct ldlm_async_args *aa = args;
2492 struct ldlm_lock *lock;
2493 struct ldlm_reply *reply;
2494 struct obd_export *exp;
2497 atomic_dec(&req->rq_import->imp_replay_inflight);
2498 wake_up(&req->rq_import->imp_replay_waitq);
2503 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2505 GOTO(out, rc = -EPROTO);
2507 lock = ldlm_handle2lock(&aa->lock_handle);
2509 CERROR("received replay ack for unknown local cookie %#llx remote cookie %#llx from server %s id %s\n",
2510 aa->lock_handle.cookie, reply->lock_handle.cookie,
2511 req->rq_export->exp_client_uuid.uuid,
2512 libcfs_idstr(&req->rq_peer));
2513 GOTO(out, rc = -ESTALE);
2516 /* Key change rehash lock in per-export hash with new key */
2517 exp = req->rq_export;
2518 if (exp && exp->exp_lock_hash) {
2520 * In the function below, .hs_keycmp resolves to
2521 * ldlm_export_lock_keycmp()
2523 cfs_hash_rehash_key(exp->exp_lock_hash,
2524 &lock->l_remote_handle,
2525 &reply->lock_handle,
2528 lock->l_remote_handle = reply->lock_handle;
2531 LDLM_DEBUG(lock, "replayed lock:");
2532 ptlrpc_import_recovery_state_machine(req->rq_import);
2533 LDLM_LOCK_PUT(lock);
2536 ptlrpc_connect_import(req->rq_import);
2541 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2543 struct ptlrpc_request *req;
2544 struct ldlm_async_args *aa;
2545 struct ldlm_request *body;
2551 /* b=11974: Do not replay a lock which is actively being canceled */
2552 if (ldlm_is_bl_done(lock)) {
2553 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2558 * If this is reply-less callback lock, we cannot replay it, since
2559 * server might have long dropped it, but notification of that event was
2560 * lost by network. (and server granted conflicting lock already)
2562 if (ldlm_is_cancel_on_block(lock)) {
2563 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2564 ldlm_lock_cancel(lock);
2569 * If granted mode matches the requested mode, this lock is granted.
2571 * If we haven't been granted anything and are on a resource list,
2572 * then we're blocked/waiting.
2574 * If we haven't been granted anything and we're NOT on a resource list,
2575 * then we haven't got a reply yet and don't have a known disposition.
2576 * This happens whenever a lock enqueue is the request that triggers
2579 if (ldlm_is_granted(lock))
2580 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2581 else if (!list_empty(&lock->l_res_link))
2582 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2584 flags = LDLM_FL_REPLAY;
2586 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2587 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2591 /* We're part of recovery, so don't wait for it. */
2592 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2593 /* If the state changed while we were prepared, don't wait */
2594 req->rq_no_delay = 1;
2596 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2597 ldlm_lock2desc(lock, &body->lock_desc);
2598 body->lock_flags = ldlm_flags_to_wire(flags);
2600 ldlm_lock2handle(lock, &body->lock_handle[0]);
2601 if (lock->l_lvb_len > 0)
2602 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2603 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2605 ptlrpc_request_set_replen(req);
2607 * notify the server we've replayed all requests.
2608 * also, we mark the request to be put on a dedicated
2609 * queue to be processed after all request replayes.
2612 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2614 LDLM_DEBUG(lock, "replaying lock:");
2616 atomic_inc(&imp->imp_replay_inflight);
2617 aa = ptlrpc_req_async_args(aa, req);
2618 aa->lock_handle = body->lock_handle[0];
2619 req->rq_interpret_reply = replay_lock_interpret;
2620 ptlrpcd_add_req(req);
2626 * Cancel as many unused locks as possible before replay. since we are
2627 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2630 * Called only in recovery before replaying locks. there is no need to
2631 * replay locks that are unused. since the clients may hold thousands of
2632 * cached unused locks, dropping the unused locks can greatly reduce the
2633 * load on the servers at recovery time.
2635 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2641 "Dropping as many unused locks as possible before replay for namespace %s (%d)\n",
2642 ldlm_ns_name(ns), ns->ns_nr_unused);
2644 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_REPLAY_PAUSE, cfs_fail_val);
2647 * We don't need to care whether or not LRU resize is enabled
2648 * because the LDLM_LRU_FLAG_NO_WAIT policy doesn't use the
2651 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2652 LCF_LOCAL, LDLM_LRU_FLAG_NO_WAIT);
2654 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2655 canceled, ldlm_ns_name(ns));
2658 static int lock_can_replay(struct obd_import *imp)
2660 struct client_obd *cli = &imp->imp_obd->u.cli;
2662 CDEBUG(D_HA, "check lock replay limit, inflights = %u(%u)\n",
2663 atomic_read(&imp->imp_replay_inflight) - 1,
2664 cli->cl_max_rpcs_in_flight);
2666 /* +1 due to ldlm_lock_replay() increment */
2667 return atomic_read(&imp->imp_replay_inflight) <
2668 1 + min_t(u32, cli->cl_max_rpcs_in_flight, 8);
2671 static int __ldlm_replay_locks(struct obd_import *imp, bool rate_limit)
2673 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2675 struct ldlm_lock *lock, *next;
2680 while (atomic_read(&imp->imp_replay_inflight) != 1)
2683 /* don't replay locks if import failed recovery */
2684 if (imp->imp_vbr_failed)
2687 if (ldlm_cancel_unused_locks_before_replay)
2688 ldlm_cancel_unused_locks_for_replay(ns);
2690 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2692 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2693 list_del_init(&lock->l_pending_chain);
2694 /* If we disconnected in the middle - cleanup and let
2695 * reconnection to happen again. LU-14027 */
2696 if (rc || (imp->imp_state != LUSTRE_IMP_REPLAY_LOCKS)) {
2697 LDLM_LOCK_RELEASE(lock);
2700 rc = replay_one_lock(imp, lock);
2701 LDLM_LOCK_RELEASE(lock);
2704 wait_event_idle_exclusive(imp->imp_replay_waitq,
2705 lock_can_replay(imp));
2712 * Lock replay uses rate control and can sleep waiting so
2713 * must be in separate thread from ptlrpcd itself
2715 static int ldlm_lock_replay_thread(void *data)
2717 struct obd_import *imp = data;
2719 unshare_fs_struct();
2720 CDEBUG(D_HA, "lock replay thread %s to %s@%s\n",
2721 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
2722 imp->imp_connection->c_remote_uuid.uuid);
2724 __ldlm_replay_locks(imp, true);
2725 atomic_dec(&imp->imp_replay_inflight);
2726 ptlrpc_import_recovery_state_machine(imp);
2727 class_import_put(imp);
2732 int ldlm_replay_locks(struct obd_import *imp)
2734 struct task_struct *task;
2737 /* ensure this doesn't fall to 0 before all have been queued */
2738 if (atomic_inc_return(&imp->imp_replay_inflight) > 1) {
2739 atomic_dec(&imp->imp_replay_inflight);
2742 class_import_get(imp);
2744 task = kthread_run(ldlm_lock_replay_thread, imp, "ldlm_lock_replay");
2747 CDEBUG(D_HA, "can't start lock replay thread: rc = %d\n", rc);
2749 /* run lock replay without rate control */
2750 rc = __ldlm_replay_locks(imp, false);
2751 atomic_dec(&imp->imp_replay_inflight);
2752 class_import_put(imp);