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)
182 * Wait a long time for enqueue - server may have to callback a
183 * lock from another client. Server will evict the other client if it
184 * doesn't respond reasonably, and then give us the lock.
186 timeout = at_get(ldlm_lock_to_ns_at(lock));
187 return max(3 * timeout, (timeout_t)ldlm_enqueue_min);
191 * Helper function for ldlm_completion_ast(), updating timings when lock is
194 static int ldlm_completion_tail(struct ldlm_lock *lock, void *data)
198 if (ldlm_is_destroyed(lock) || ldlm_is_failed(lock)) {
199 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
201 } else if (data == NULL) {
202 LDLM_DEBUG(lock, "client-side enqueue: granted");
204 /* Take into AT only CP RPC, not immediately granted locks */
207 /* Discard negative timeouts. We should also limit the
208 * maximum value of the timeout
210 if (ktime_get_real_seconds() > lock->l_activity)
211 delay = ktime_get_real_seconds() - lock->l_activity;
213 LDLM_DEBUG(lock, "client-side enqueue: granted after %ds",
215 /* Update our time estimate */
216 at_measured(ldlm_lock_to_ns_at(lock), delay);
222 * Implementation of ->l_completion_ast() for a client, that doesn't wait
223 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
224 * other threads that cannot block for long.
226 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
230 if (flags == LDLM_FL_WAIT_NOREPROC) {
231 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
235 if (!(flags & LDLM_FL_BLOCKED_MASK)) {
236 wake_up(&lock->l_waitq);
237 RETURN(ldlm_completion_tail(lock, data));
240 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, going forward");
241 ldlm_reprocess_all(lock->l_resource, 0);
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 = {
272 struct obd_device *obd;
273 struct obd_import *imp = NULL;
279 if (flags == LDLM_FL_WAIT_NOREPROC) {
280 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
284 if (!(flags & LDLM_FL_BLOCKED_MASK)) {
285 wake_up(&lock->l_waitq);
289 LDLM_DEBUG(lock, "client-side enqueue returned a blocked locksleeping");
293 obd = class_exp2obd(lock->l_conn_export);
295 /* if this is a local lock, then there is no import */
297 imp = obd->u.cli.cl_import;
299 timeout = ldlm_cp_timeout(lock);
302 lock->l_activity = ktime_get_real_seconds();
305 spin_lock(&imp->imp_lock);
306 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
307 spin_unlock(&imp->imp_lock);
310 if (ns_is_client(ldlm_lock_to_ns(lock)) &&
311 CFS_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
312 OBD_FAIL_LDLM_CP_BL_RACE | CFS_FAIL_ONCE)) {
313 ldlm_set_fail_loc(lock);
316 /* Go to sleep until the lock is granted or cancelled. */
317 if (ldlm_is_no_timeout(lock)) {
318 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
319 rc = l_wait_event_abortable(
321 is_granted_or_cancelled(lock));
323 if (wait_event_idle_timeout(
325 is_granted_or_cancelled(lock),
326 cfs_time_seconds(timeout)) == 0) {
327 ldlm_expired_completion_wait(&lwd);
328 rc = l_wait_event_abortable(
330 is_granted_or_cancelled(lock));
336 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
341 RETURN(ldlm_completion_tail(lock, data));
343 EXPORT_SYMBOL(ldlm_completion_ast);
346 * A helper to build a blocking AST function
348 * Perform a common operation for blocking ASTs:
349 * defferred lock cancellation.
351 * \param lock the lock blocking or canceling AST was called on
353 * \see mdt_blocking_ast
354 * \see ldlm_blocking_ast
356 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
362 ldlm_set_cbpending(lock);
363 do_ast = (!lock->l_readers && !lock->l_writers);
364 unlock_res_and_lock(lock);
367 struct lustre_handle lockh;
370 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
371 ldlm_lock2handle(lock, &lockh);
372 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
374 CERROR("ldlm_cli_cancel: %d\n", rc);
377 "Lock still has references, will be cancelled later");
381 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
384 * Server blocking AST
386 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
389 * \param lock the lock which blocks a request or cancelling lock
392 * \param flag indicates whether this cancelling or blocking callback
394 * \see ldlm_blocking_ast_nocheck
396 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
397 void *data, int flag)
401 if (flag == LDLM_CB_CANCELING) {
402 /* Don't need to do anything here. */
406 lock_res_and_lock(lock);
408 * Get this: if ldlm_blocking_ast is racing with intent_policy, such
409 * that ldlm_blocking_ast is called just before intent_policy method
410 * takes the lr_lock, then by the time we get the lock, we might not
411 * be the correct blocking function anymore. So check, and return
414 if (lock->l_blocking_ast != ldlm_blocking_ast) {
415 unlock_res_and_lock(lock);
418 RETURN(ldlm_blocking_ast_nocheck(lock));
420 EXPORT_SYMBOL(ldlm_blocking_ast);
423 * Implements ldlm_lock::l_glimpse_ast for extent locks acquired on the server.
425 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for that is
426 * rather subtle: with OST-side locking, it may so happen that _all_ extent
427 * locks are held by the OST. If client wants to obtain the current file size
428 * it calls ll_glimpse_size(), and (as all locks are held only on the server),
429 * this dummy glimpse callback fires and does nothing. The client still
430 * receives the correct file size due to the following fragment of code in
431 * ldlm_cb_interpret():
433 * if (rc == -ELDLM_NO_LOCK_DATA) {
434 * LDLM_DEBUG(lock, "lost race - client has a lock but no"
436 * ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
439 * That is, after the glimpse returns this error, ofd_lvbo_update() is called
440 * and returns the updated file attributes from the inode to the client.
442 * See also comment in ofd_intent_policy() on why servers must set a non-NULL
443 * l_glimpse_ast when grabbing DLM locks. Otherwise, the server will assume
444 * that the object is in the process of being destroyed.
446 * \param[in] lock DLM lock being glimpsed, unused
447 * \param[in] reqp pointer to ptlrpc_request, unused
449 * \retval -ELDLM_NO_LOCK_DATA to get attributes from disk object
451 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
453 return -ELDLM_NO_LOCK_DATA;
457 * Enqueue a local lock (typically on a server).
459 int ldlm_cli_enqueue_local(const struct lu_env *env,
460 struct ldlm_namespace *ns,
461 const struct ldlm_res_id *res_id,
462 enum ldlm_type type, union ldlm_policy_data *policy,
463 enum ldlm_mode mode, __u64 *flags,
464 ldlm_blocking_callback blocking,
465 ldlm_completion_callback completion,
466 ldlm_glimpse_callback glimpse,
467 void *data, __u32 lvb_len, enum lvb_type lvb_type,
468 const __u64 *client_cookie,
469 struct lustre_handle *lockh)
471 struct ldlm_lock *lock;
473 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
474 .lcs_blocking = blocking,
475 .lcs_glimpse = glimpse,
480 LASSERT(!(*flags & LDLM_FL_REPLAY));
481 if (unlikely(ns_is_client(ns))) {
482 CERROR("Trying to enqueue local lock in a shadow namespace\n");
486 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
489 GOTO(out_nolock, err = PTR_ERR(lock));
491 err = ldlm_lvbo_init(lock->l_resource);
493 LDLM_ERROR(lock, "delayed lvb init failed (rc %d)", err);
494 ldlm_lock_destroy_nolock(lock);
498 ldlm_lock2handle(lock, lockh);
501 * NB: we don't have any lock now (lock_res_and_lock)
502 * because it's a new lock
504 ldlm_lock_addref_internal_nolock(lock, mode);
505 ldlm_set_local(lock);
506 if (*flags & LDLM_FL_ATOMIC_CB)
507 ldlm_set_atomic_cb(lock);
509 if (*flags & LDLM_FL_CANCEL_ON_BLOCK)
510 ldlm_set_cancel_on_block(lock);
513 lock->l_policy_data = *policy;
514 if (client_cookie != NULL)
515 lock->l_client_cookie = *client_cookie;
516 if (type == LDLM_EXTENT) {
517 /* extent lock without policy is a bug */
521 lock->l_req_extent = policy->l_extent;
524 err = ldlm_lock_enqueue(env, ns, &lock, policy, flags);
525 if (unlikely(err != ELDLM_OK))
529 *policy = lock->l_policy_data;
531 if (lock->l_completion_ast)
532 lock->l_completion_ast(lock, *flags, NULL);
534 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
537 LDLM_LOCK_RELEASE(lock);
541 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
543 static void failed_lock_cleanup(struct ldlm_namespace *ns,
544 struct ldlm_lock *lock, int mode)
548 /* Set a flag to prevent us from sending a CANCEL (b=407) */
549 lock_res_and_lock(lock);
550 /* Check that lock is not granted or failed, we might race. */
551 if (!ldlm_is_granted(lock) && !ldlm_is_failed(lock)) {
553 * Make sure that this lock will not be found by raced
554 * bl_ast and -EINVAL reply is sent to server anyways.
557 lock->l_flags |= LDLM_FL_FAILED |
558 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
559 if (!(ldlm_is_bl_ast(lock) &&
560 lock->l_remote_handle.cookie != 0))
561 lock->l_flags |= LDLM_FL_LOCAL_ONLY;
564 unlock_res_and_lock(lock);
568 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
570 LDLM_DEBUG(lock, "lock was granted or failed in race");
573 * XXX - HACK because we shouldn't call ldlm_lock_destroy()
574 * from llite/file.c/ll_file_flock().
577 * This code makes for the fact that we do not have blocking handler on
578 * a client for flock locks. As such this is the place where we must
579 * completely kill failed locks. (interrupted and those that
580 * were waiting to be granted when server evicted us.
582 if (lock->l_resource->lr_type == LDLM_FLOCK) {
583 lock_res_and_lock(lock);
584 if (!ldlm_is_destroyed(lock)) {
585 ldlm_resource_unlink_lock(lock);
586 ldlm_lock_decref_internal_nolock(lock, mode);
587 ldlm_lock_destroy_nolock(lock);
589 unlock_res_and_lock(lock);
591 ldlm_lock_decref_internal(lock, mode);
595 static bool ldlm_request_slot_needed(struct ldlm_enqueue_info *einfo)
597 /* exclude EXTENT locks and DOM-only IBITS locks because they
598 * are asynchronous and don't wait on server being blocked.
600 return einfo->ei_req_slot &&
601 (einfo->ei_type == LDLM_FLOCK ||
602 (einfo->ei_type == LDLM_IBITS &&
603 einfo->ei_inodebits != MDS_INODELOCK_DOM));
607 * Finishing portion of client lock enqueue code.
609 * Called after receiving reply from server.
611 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct req_capsule *pill,
612 struct ldlm_enqueue_info *einfo,
613 __u8 with_policy, __u64 *ldlm_flags, void *lvb,
614 __u32 lvb_len, const struct lustre_handle *lockh,
615 int rc, bool request_slot)
617 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
618 const struct lu_env *env = NULL;
619 int is_replay = *ldlm_flags & LDLM_FL_REPLAY;
620 struct ldlm_lock *lock;
621 struct ldlm_reply *reply;
622 int cleanup_phase = 1;
626 if (req_capsule_ptlreq(pill)) {
627 struct ptlrpc_request *req = pill->rc_req;
630 obd_put_request_slot(&req->rq_import->imp_obd->u.cli);
632 ptlrpc_put_mod_rpc_slot(req);
634 if (req && req->rq_svc_thread)
635 env = req->rq_svc_thread->t_env;
638 lock = ldlm_handle2lock(lockh);
639 /* ldlm_cli_enqueue is holding a reference on this lock. */
641 LASSERT(einfo->ei_type == LDLM_FLOCK);
645 LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
646 "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
648 if (rc != ELDLM_OK) {
650 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
651 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
653 if (rc != ELDLM_LOCK_ABORTED)
657 /* Before we return, swab the reply */
658 reply = req_capsule_server_get(pill, &RMF_DLM_REP);
660 GOTO(cleanup, rc = -EPROTO);
665 size = req_capsule_get_size(pill, &RMF_DLM_LVB, RCL_SERVER);
667 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
668 GOTO(cleanup, rc = size);
669 } else if (unlikely(size > lvb_len)) {
671 "Replied LVB is larger than expectation, expected = %d, replied = %d",
673 GOTO(cleanup, rc = -EINVAL);
678 if (rc == ELDLM_LOCK_ABORTED) {
679 if (lvb_len > 0 && lvb != NULL)
680 rc = ldlm_fill_lvb(lock, pill, RCL_SERVER,
682 GOTO(cleanup, rc = rc ? : ELDLM_LOCK_ABORTED);
685 /* lock enqueued on the server */
688 lock_res_and_lock(lock);
689 /* Key change rehash lock in per-export hash with new key */
690 if (exp->exp_lock_hash) {
692 * In the function below, .hs_keycmp resolves to
693 * ldlm_export_lock_keycmp()
695 cfs_hash_rehash_key(exp->exp_lock_hash,
696 &lock->l_remote_handle,
700 lock->l_remote_handle = reply->lock_handle;
703 *ldlm_flags = ldlm_flags_from_wire(reply->lock_flags);
704 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
705 LDLM_FL_INHERIT_MASK);
706 unlock_res_and_lock(lock);
708 CDEBUG(D_INFO, "local: %p, remote cookie: %#llx, flags: %#llx\n",
709 lock, reply->lock_handle.cookie, *ldlm_flags);
712 * If enqueue returned a blocked lock but the completion handler has
713 * already run, then it fixed up the resource and we don't need to do it
716 if ((*ldlm_flags) & LDLM_FL_LOCK_CHANGED) {
717 int newmode = reply->lock_desc.l_req_mode;
720 if (newmode && newmode != lock->l_req_mode) {
721 LDLM_DEBUG(lock, "server returned different mode %s",
722 ldlm_lockname[newmode]);
723 lock->l_req_mode = newmode;
726 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
727 &lock->l_resource->lr_name)) {
729 "remote intent success, locking "DLDLMRES", instead of "DLDLMRES"\n",
730 PLDLMRES(&reply->lock_desc.l_resource),
731 PLDLMRES(lock->l_resource));
733 rc = ldlm_lock_change_resource(ns, lock,
734 &reply->lock_desc.l_resource.lr_name);
735 if (rc || lock->l_resource == NULL)
736 GOTO(cleanup, rc = -ENOMEM);
737 LDLM_DEBUG(lock, "client-side enqueue, new resource");
741 /* We assume lock type cannot change on server*/
742 ldlm_convert_policy_to_local(exp,
743 lock->l_resource->lr_type,
744 &reply->lock_desc.l_policy_data,
745 &lock->l_policy_data);
748 if (einfo->ei_type != LDLM_PLAIN)
750 "client-side enqueue, new policy data");
753 if ((*ldlm_flags) & LDLM_FL_AST_SENT) {
754 lock_res_and_lock(lock);
755 ldlm_bl_desc2lock(&reply->lock_desc, lock);
756 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
757 unlock_res_and_lock(lock);
758 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
762 * If the lock has already been granted by a completion AST, don't
763 * clobber the LVB with an older one.
767 * We must lock or a racing completion might update lvb without
768 * letting us know and we'll clobber the correct value.
769 * Cannot unlock after the check either, a that still leaves
770 * a tiny window for completion to get in
772 lock_res_and_lock(lock);
773 if (!ldlm_is_granted(lock))
774 rc = ldlm_fill_lvb(lock, pill, RCL_SERVER,
775 lock->l_lvb_data, lvb_len);
776 unlock_res_and_lock(lock);
784 rc = ldlm_lock_enqueue(env, ns, &lock, NULL, ldlm_flags);
785 if (lock->l_completion_ast != NULL) {
786 int err = lock->l_completion_ast(lock, *ldlm_flags,
796 if (lvb_len > 0 && lvb != NULL) {
798 * Copy the LVB here, and not earlier, because the completion
799 * AST (if any) can override what we got in the reply
801 memcpy(lvb, lock->l_lvb_data, lvb_len);
804 LDLM_DEBUG(lock, "client-side enqueue END");
807 if (cleanup_phase == 1 && rc)
808 failed_lock_cleanup(ns, lock, einfo->ei_mode);
809 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
811 LDLM_LOCK_RELEASE(lock);
814 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
817 * Estimate number of lock handles that would fit into request of given
818 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
819 * a single page on the send/receive side. XXX: 512 should be changed to
820 * more adequate value.
822 static inline int ldlm_req_handles_avail(int req_size, int off)
826 avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512) - req_size;
827 if (likely(avail >= 0))
828 avail /= (int)sizeof(struct lustre_handle);
831 avail += LDLM_LOCKREQ_HANDLES - off;
836 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
837 enum req_location loc,
840 __u32 size = req_capsule_msg_size(pill, loc);
842 return ldlm_req_handles_avail(size, off);
845 static inline int ldlm_format_handles_avail(struct obd_import *imp,
846 const struct req_format *fmt,
847 enum req_location loc, int off)
849 __u32 size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
851 return ldlm_req_handles_avail(size, off);
855 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
856 * \a count locks in \a cancels.
858 * This is to be called by functions preparing their own requests that
859 * might contain lists of locks to cancel in addition to actual operation
860 * that needs to be performed.
862 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
863 int version, int opc, int canceloff,
864 struct list_head *cancels, int count)
866 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
867 struct req_capsule *pill = &req->rq_pill;
868 struct ldlm_request *dlm = NULL;
870 int avail, to_free = 0, pack = 0;
877 if (ns_connect_cancelset(ns)) {
878 /* Estimate the amount of available space in the request. */
879 req_capsule_filled_sizes(pill, RCL_CLIENT);
880 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
882 /* If we have reached the limit, free +1 slot for the new one */
883 if (!ns_connect_lru_resize(ns) && opc == LDLM_ENQUEUE &&
884 ns->ns_nr_unused >= ns->ns_max_unused)
888 * Cancel LRU locks here _only_ if the server supports
889 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
890 * RPC, which will make us slower.
893 count += ldlm_cancel_lru_local(ns, cancels, to_free,
895 LDLM_LRU_FLAG_NO_WAIT);
900 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
901 ldlm_request_bufsize(pack, opc));
904 rc = ptlrpc_request_pack(req, version, opc);
906 ldlm_lock_list_put(cancels, l_bl_ast, count);
910 if (ns_connect_cancelset(ns)) {
912 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
915 * Skip first lock handler in ldlm_request_pack(),
916 * this method will increment @lock_count according
917 * to the lock handle amount actually written to
920 dlm->lock_count = canceloff;
922 /* Pack into the request @pack lock handles. */
923 ldlm_cli_cancel_list(cancels, pack, req, 0);
924 /* Prepare and send separate cancel RPC for others. */
925 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
927 ldlm_lock_list_put(cancels, l_bl_ast, count);
931 EXPORT_SYMBOL(ldlm_prep_elc_req);
933 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
934 struct list_head *cancels, int count)
936 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
937 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
939 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
941 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
943 struct ptlrpc_request *req;
948 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
950 RETURN(ERR_PTR(-ENOMEM));
952 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
954 ptlrpc_request_free(req);
958 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
959 ptlrpc_request_set_replen(req);
962 EXPORT_SYMBOL(ldlm_enqueue_pack);
965 * Client-side lock enqueue.
967 * If a request has some specific initialisation it is passed in \a reqp,
968 * otherwise it is created in ldlm_cli_enqueue.
970 * Supports sync and async requests, pass \a async flag accordingly. If a
971 * request was created in ldlm_cli_enqueue and it is the async request,
972 * pass it to the caller in \a reqp.
974 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
975 struct ldlm_enqueue_info *einfo,
976 const struct ldlm_res_id *res_id,
977 union ldlm_policy_data const *policy, __u64 *flags,
978 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
979 struct lustre_handle *lockh, int async)
981 struct ldlm_namespace *ns;
982 struct ldlm_lock *lock;
983 struct ldlm_request *body;
984 int is_replay = *flags & LDLM_FL_REPLAY;
985 int req_passed_in = 1;
988 struct ptlrpc_request *req;
992 LASSERT(exp != NULL);
994 ns = exp->exp_obd->obd_namespace;
997 * If we're replaying this lock, just check some invariants.
998 * If we're creating a new lock, get everything all setup nice.
1001 lock = ldlm_handle2lock_long(lockh, 0);
1002 LASSERT(lock != NULL);
1003 LDLM_DEBUG(lock, "client-side enqueue START");
1004 LASSERT(exp == lock->l_conn_export);
1006 const struct ldlm_callback_suite cbs = {
1007 .lcs_completion = einfo->ei_cb_cp,
1008 .lcs_blocking = einfo->ei_cb_bl,
1009 .lcs_glimpse = einfo->ei_cb_gl
1011 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
1012 einfo->ei_mode, &cbs, einfo->ei_cbdata,
1015 RETURN(PTR_ERR(lock));
1017 if (einfo->ei_cb_created)
1018 einfo->ei_cb_created(lock);
1020 /* for the local lock, add the reference */
1021 ldlm_lock_addref_internal(lock, einfo->ei_mode);
1022 ldlm_lock2handle(lock, lockh);
1024 lock->l_policy_data = *policy;
1026 if (einfo->ei_type == LDLM_EXTENT) {
1027 /* extent lock without policy is a bug */
1031 lock->l_req_extent = policy->l_extent;
1033 LDLM_DEBUG(lock, "client-side enqueue START, flags %#llx",
1037 lock->l_conn_export = exp;
1038 lock->l_export = NULL;
1039 lock->l_blocking_ast = einfo->ei_cb_bl;
1040 lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
1041 lock->l_activity = ktime_get_real_seconds();
1043 /* lock not sent to server yet */
1044 if (reqp == NULL || *reqp == NULL) {
1045 req = ldlm_enqueue_pack(exp, lvb_len);
1047 failed_lock_cleanup(ns, lock, einfo->ei_mode);
1048 LDLM_LOCK_RELEASE(lock);
1049 RETURN(PTR_ERR(req));
1059 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
1061 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
1062 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
1065 if (*flags & LDLM_FL_NDELAY) {
1066 DEBUG_REQ(D_DLMTRACE, req, "enqueue lock with no delay");
1067 req->rq_no_resend = req->rq_no_delay = 1;
1069 * probably set a shorter timeout value and handle ETIMEDOUT
1070 * in osc_lock_upcall() correctly
1072 /* lustre_msg_set_timeout(req, req->rq_timeout / 2); */
1075 /* Dump lock data into the request buffer */
1076 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1077 ldlm_lock2desc(lock, &body->lock_desc);
1078 body->lock_flags = ldlm_flags_to_wire(*flags);
1079 body->lock_handle[0] = *lockh;
1081 /* extended LDLM opcodes in client stats */
1082 if (exp->exp_obd->obd_svc_stats != NULL) {
1083 /* glimpse is intent with no intent buffer */
1084 if (*flags & LDLM_FL_HAS_INTENT &&
1085 !req_capsule_has_field(&req->rq_pill, &RMF_LDLM_INTENT,
1087 lprocfs_counter_incr(exp->exp_obd->obd_svc_stats,
1089 LDLM_GLIMPSE_ENQUEUE);
1091 ldlm_svc_get_eopc(body, exp->exp_obd->obd_svc_stats);
1094 /* It is important to obtain modify RPC slot first (if applicable), so
1095 * that threads that are waiting for a modify RPC slot are not polluting
1096 * our rpcs in flight counter. */
1098 if (einfo->ei_mod_slot)
1099 ptlrpc_get_mod_rpc_slot(req);
1101 need_req_slot = ldlm_request_slot_needed(einfo);
1103 if (need_req_slot) {
1104 rc = obd_get_request_slot(&req->rq_import->imp_obd->u.cli);
1106 if (einfo->ei_mod_slot)
1107 ptlrpc_put_mod_rpc_slot(req);
1108 failed_lock_cleanup(ns, lock, einfo->ei_mode);
1109 LDLM_LOCK_RELEASE(lock);
1111 ptlrpc_req_finished(req);
1117 LASSERT(reqp != NULL);
1121 LDLM_DEBUG(lock, "sending request");
1123 rc = ptlrpc_queue_wait(req);
1125 err = ldlm_cli_enqueue_fini(exp, &req->rq_pill, einfo, policy ? 1 : 0,
1126 flags, lvb, lvb_len, lockh, rc,
1130 * If ldlm_cli_enqueue_fini did not find the lock, we need to free
1131 * one reference that we took
1134 LDLM_LOCK_RELEASE(lock);
1139 if (!req_passed_in && req != NULL) {
1140 ptlrpc_req_finished(req);
1147 EXPORT_SYMBOL(ldlm_cli_enqueue);
1150 * Client-side IBITS lock create and pack for WBC EX lock request.
1152 int ldlm_cli_lock_create_pack(struct obd_export *exp,
1153 struct ldlm_request *dlmreq,
1154 struct ldlm_enqueue_info *einfo,
1155 const struct ldlm_res_id *res_id,
1156 union ldlm_policy_data const *policy,
1157 __u64 *flags, void *lvb, __u32 lvb_len,
1158 enum lvb_type lvb_type,
1159 struct lustre_handle *lockh)
1161 const struct ldlm_callback_suite cbs = {
1162 .lcs_completion = einfo->ei_cb_cp,
1163 .lcs_blocking = einfo->ei_cb_bl,
1164 .lcs_glimpse = einfo->ei_cb_gl
1166 struct ldlm_namespace *ns;
1167 struct ldlm_lock *lock;
1171 LASSERT(exp != NULL);
1172 LASSERT(!(*flags & LDLM_FL_REPLAY));
1174 ns = exp->exp_obd->obd_namespace;
1175 lock = ldlm_lock_create(ns, res_id, einfo->ei_type, einfo->ei_mode,
1176 &cbs, einfo->ei_cbdata, lvb_len, lvb_type);
1178 RETURN(PTR_ERR(lock));
1180 if (einfo->ei_cb_created)
1181 einfo->ei_cb_created(lock);
1183 /* For the local lock, add the reference */
1184 ldlm_lock_addref_internal(lock, einfo->ei_mode);
1185 ldlm_lock2handle(lock, lockh);
1187 lock->l_policy_data = *policy;
1189 LDLM_DEBUG(lock, "client-side enqueue START, flags %#llx", *flags);
1190 lock->l_conn_export = exp;
1191 lock->l_export = NULL;
1192 lock->l_blocking_ast = einfo->ei_cb_bl;
1193 lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL |
1194 LDLM_FL_ATOMIC_CB));
1195 lock->l_activity = ktime_get_real_seconds();
1197 ldlm_lock2desc(lock, &dlmreq->lock_desc);
1198 dlmreq->lock_flags = ldlm_flags_to_wire(*flags);
1199 dlmreq->lock_handle[0] = *lockh;
1203 EXPORT_SYMBOL(ldlm_cli_lock_create_pack);
1206 * Client-side IBITS lock convert.
1208 * Inform server that lock has been converted instead of canceling.
1209 * Server finishes convert on own side and does reprocess to grant
1210 * all related waiting locks.
1212 * Since convert means only ibits downgrading, client doesn't need to
1213 * wait for server reply to finish local converting process so this request
1214 * is made asynchronous.
1217 int ldlm_cli_convert_req(struct ldlm_lock *lock, __u32 *flags, __u64 new_bits)
1219 struct ldlm_request *body;
1220 struct ptlrpc_request *req;
1221 struct obd_export *exp = lock->l_conn_export;
1225 LASSERT(exp != NULL);
1228 * this is better to check earlier and it is done so already,
1229 * but this check is kept too as final one to issue an error
1230 * if any new code will miss such check.
1232 if (!exp_connect_lock_convert(exp)) {
1233 LDLM_ERROR(lock, "server doesn't support lock convert\n");
1237 if (lock->l_resource->lr_type != LDLM_IBITS) {
1238 LDLM_ERROR(lock, "convert works with IBITS locks only.");
1242 LDLM_DEBUG(lock, "client-side convert");
1244 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1245 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1250 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1251 body->lock_handle[0] = lock->l_remote_handle;
1253 body->lock_desc.l_req_mode = lock->l_req_mode;
1254 body->lock_desc.l_granted_mode = lock->l_granted_mode;
1256 body->lock_desc.l_policy_data.l_inodebits.bits = new_bits;
1257 body->lock_desc.l_policy_data.l_inodebits.cancel_bits = 0;
1259 body->lock_flags = ldlm_flags_to_wire(*flags);
1260 body->lock_count = 1;
1262 ptlrpc_request_set_replen(req);
1265 * Use cancel portals for convert as well as high-priority handling.
1267 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1268 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1270 ptlrpc_at_set_req_timeout(req);
1272 if (exp->exp_obd->obd_svc_stats != NULL)
1273 lprocfs_counter_incr(exp->exp_obd->obd_svc_stats,
1274 LDLM_CONVERT - LDLM_FIRST_OPC);
1276 ptlrpcd_add_req(req);
1281 * Cancel locks locally.
1283 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1284 * \retval LDLM_FL_CANCELING otherwise;
1285 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1287 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1289 __u64 rc = LDLM_FL_LOCAL_ONLY;
1293 if (lock->l_conn_export) {
1296 LDLM_DEBUG(lock, "client-side cancel");
1297 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL_LOCAL,
1300 /* Set this flag to prevent others from getting new references*/
1301 lock_res_and_lock(lock);
1302 ldlm_set_cbpending(lock);
1303 local_only = !!(lock->l_flags &
1304 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1305 ldlm_cancel_callback(lock);
1306 rc = (ldlm_is_bl_ast(lock)) ?
1307 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1308 unlock_res_and_lock(lock);
1312 "not sending request (at caller's instruction)\n");
1313 rc = LDLM_FL_LOCAL_ONLY;
1315 ldlm_lock_cancel(lock);
1317 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1318 LDLM_ERROR(lock, "Trying to cancel local lock");
1321 LDLM_DEBUG(lock, "server-side local cancel");
1322 ldlm_lock_cancel(lock);
1323 ldlm_reprocess_all(lock->l_resource,
1324 lock->l_policy_data.l_inodebits.bits);
1330 static inline int __ldlm_pack_lock(struct ldlm_lock *lock,
1331 struct ldlm_request *dlm)
1333 LASSERT(lock->l_conn_export);
1334 lock_res_and_lock(lock);
1335 if (ldlm_is_ast_sent(lock)) {
1336 unlock_res_and_lock(lock);
1339 ldlm_set_ast_sent(lock);
1340 unlock_res_and_lock(lock);
1342 /* Pack the lock handle to the given request buffer. */
1343 LDLM_DEBUG(lock, "packing");
1344 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1348 #define ldlm_cancel_pack(req, head, count) \
1349 _ldlm_cancel_pack(req, NULL, head, count)
1351 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1353 static int _ldlm_cancel_pack(struct ptlrpc_request *req, struct ldlm_lock *lock,
1354 struct list_head *head, int count)
1356 struct ldlm_request *dlm;
1357 int max, packed = 0;
1361 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1362 LASSERT(dlm != NULL);
1364 /* Check the room in the request buffer. */
1365 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1366 sizeof(struct ldlm_request);
1367 max /= sizeof(struct lustre_handle);
1368 max += LDLM_LOCKREQ_HANDLES;
1369 LASSERT(max >= dlm->lock_count + count);
1372 * XXX: it would be better to pack lock handles grouped by resource.
1373 * so that the server cancel would call filter_lvbo_update() less
1376 if (lock) { /* only pack one lock */
1377 packed = __ldlm_pack_lock(lock, dlm);
1379 list_for_each_entry(lock, head, l_bl_ast) {
1382 packed += __ldlm_pack_lock(lock, dlm);
1385 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1390 * Prepare and send a batched cancel RPC. It will include \a count lock
1391 * handles of locks given in \a cancels list.
1393 int ldlm_cli_cancel_req(struct obd_export *exp, struct ldlm_lock *lock,
1394 struct list_head *head, int count,
1395 enum ldlm_cancel_flags flags)
1397 struct ptlrpc_request *req = NULL;
1398 struct obd_import *imp;
1404 LASSERT(exp != NULL);
1406 LASSERT(head == NULL || lock == NULL);
1408 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1410 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1413 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1414 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1419 imp = class_exp2cliimp(exp);
1420 if (imp == NULL || imp->imp_invalid) {
1422 "skipping cancel on invalid import %p\n", imp);
1426 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1428 GOTO(out, rc = -ENOMEM);
1430 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1431 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1432 ldlm_request_bufsize(count, LDLM_CANCEL));
1434 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1436 ptlrpc_request_free(req);
1441 * If OSP want cancel cross-MDT lock, let's not block it in
1442 * in recovery, otherwise the lock will not released, if
1443 * the remote target is also in recovery, and it also need
1444 * this lock, it might cause deadlock.
1446 if (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS &&
1447 exp->exp_obd->obd_lu_dev != NULL &&
1448 exp->exp_obd->obd_lu_dev->ld_site != NULL) {
1449 struct lu_device *top_dev;
1451 top_dev = exp->exp_obd->obd_lu_dev->ld_site->ls_top_dev;
1452 if (top_dev != NULL &&
1453 top_dev->ld_obd->obd_recovering)
1454 req->rq_allow_replay = 1;
1457 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1458 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1459 ptlrpc_at_set_req_timeout(req);
1461 rc = _ldlm_cancel_pack(req, lock, head, count);
1463 ptlrpc_req_finished(req);
1468 ptlrpc_request_set_replen(req);
1469 if (flags & LCF_ASYNC) {
1470 ptlrpcd_add_req(req);
1475 rc = ptlrpc_queue_wait(req);
1476 if (rc == LUSTRE_ESTALE) {
1478 "client/server (nid %s) out of sync -- not fatal\n",
1479 libcfs_nidstr(&req->rq_import->imp_connection->c_peer.nid));
1481 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1482 req->rq_import_generation == imp->imp_generation) {
1483 ptlrpc_req_finished(req);
1485 } else if (rc != ELDLM_OK) {
1486 /* -ESHUTDOWN is common on umount */
1487 CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1488 "Got rc %d from cancel RPC: canceling anyway\n",
1496 ptlrpc_req_finished(req);
1499 return sent ? sent : rc;
1502 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1504 LASSERT(imp != NULL);
1505 return &imp->imp_obd->obd_namespace->ns_pool;
1509 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1511 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1513 struct ldlm_namespace *ns;
1514 struct obd_device *obd;
1515 __u64 new_slv, ratio;
1519 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1520 !imp_connect_lru_resize(req->rq_import)))
1521 /* Do nothing for corner cases. */
1525 * In some cases RPC may contain SLV and limit zeroed out. This
1526 * is the case when server does not support LRU resize feature.
1527 * This is also possible in some recovery cases when server-side
1528 * reqs have no reference to the OBD export and thus access to
1529 * server-side namespace is not possible.
1531 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1532 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1533 DEBUG_REQ(D_HA, req,
1534 "Zero SLV or limit found (SLV=%llu, limit=%u)",
1535 lustre_msg_get_slv(req->rq_repmsg),
1536 lustre_msg_get_limit(req->rq_repmsg));
1540 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1541 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1542 obd = req->rq_import->imp_obd;
1544 read_lock(&obd->obd_pool_lock);
1545 if (obd->obd_pool_slv == new_slv &&
1546 obd->obd_pool_limit == new_limit) {
1547 read_unlock(&obd->obd_pool_lock);
1550 read_unlock(&obd->obd_pool_lock);
1553 * OBD device keeps the new pool attributes before they are handled by
1556 write_lock(&obd->obd_pool_lock);
1557 obd->obd_pool_slv = new_slv;
1558 obd->obd_pool_limit = new_limit;
1559 write_unlock(&obd->obd_pool_lock);
1562 * Check if an urgent pool recalc is needed, let it to be a change of
1563 * SLV on 10%. It is applicable to LRU resize enabled case only.
1565 ns = obd->obd_namespace;
1566 if (!ns_connect_lru_resize(ns) ||
1567 ldlm_pool_get_slv(&ns->ns_pool) < new_slv)
1570 ratio = 100 * new_slv / ldlm_pool_get_slv(&ns->ns_pool);
1571 if (100 - ratio >= ns->ns_recalc_pct &&
1572 !ns->ns_stopping && !ns->ns_rpc_recalc) {
1573 bool recalc = false;
1575 spin_lock(&ns->ns_lock);
1576 if (!ns->ns_stopping && !ns->ns_rpc_recalc) {
1577 ldlm_namespace_get(ns);
1579 ns->ns_rpc_recalc = 1;
1581 spin_unlock(&ns->ns_lock);
1583 ldlm_bl_to_thread_ns(ns);
1589 int ldlm_cli_convert(struct ldlm_lock *lock,
1590 enum ldlm_cancel_flags cancel_flags)
1594 LASSERT(!lock->l_readers && !lock->l_writers);
1595 LDLM_DEBUG(lock, "client lock convert START");
1597 if (lock->l_resource->lr_type == LDLM_IBITS) {
1598 lock_res_and_lock(lock);
1600 rc = ldlm_cli_inodebits_convert(lock, cancel_flags);
1601 } while (rc == -EAGAIN);
1602 unlock_res_and_lock(lock);
1605 LDLM_DEBUG(lock, "client lock convert END");
1608 EXPORT_SYMBOL(ldlm_cli_convert);
1611 * Client side lock cancel.
1613 * Lock must not have any readers or writers by this time.
1615 int ldlm_cli_cancel(const struct lustre_handle *lockh,
1616 enum ldlm_cancel_flags flags)
1618 struct obd_export *exp;
1619 int avail, count = 1, separate = 0;
1620 enum ldlm_lru_flags lru_flags = 0;
1622 struct ldlm_namespace *ns;
1623 struct ldlm_lock *lock;
1628 lock = ldlm_handle2lock_long(lockh, 0);
1630 LDLM_DEBUG_NOLOCK("lock is already being destroyed");
1634 lock_res_and_lock(lock);
1635 LASSERT(!ldlm_is_converting(lock));
1637 if (ldlm_is_bl_ast(lock)) {
1638 if (ldlm_is_ast_sent(lock)) {
1639 unlock_res_and_lock(lock);
1640 LDLM_LOCK_RELEASE(lock);
1643 if (ldlm_is_canceling(lock))
1645 } else if (ldlm_is_canceling(lock)) {
1646 /* Lock is being canceled and the caller doesn't want to wait */
1647 if (flags & LCF_ASYNC) {
1648 unlock_res_and_lock(lock);
1650 unlock_res_and_lock(lock);
1651 wait_event_idle(lock->l_waitq, is_bl_done(lock));
1653 LDLM_LOCK_RELEASE(lock);
1657 ldlm_set_canceling(lock);
1658 unlock_res_and_lock(lock);
1660 if (flags & LCF_LOCAL)
1661 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_LOCAL_CANCEL_PAUSE,
1664 rc = ldlm_cli_cancel_local(lock);
1665 if (rc == LDLM_FL_LOCAL_ONLY || flags & LCF_LOCAL) {
1666 LDLM_LOCK_RELEASE(lock);
1668 } else if (rc == LDLM_FL_BL_AST) {
1669 /* BL_AST lock must not wait. */
1670 lru_flags |= LDLM_LRU_FLAG_NO_WAIT;
1673 exp = lock->l_conn_export;
1674 /* If a lock has been taken from lru for a batched cancel and a later
1675 * BL_AST came, send a CANCEL RPC individually for it right away, not
1676 * waiting for the batch to be handled. */
1678 ldlm_cli_cancel_req(exp, lock, NULL, 1, flags);
1679 LDLM_LOCK_RELEASE(lock);
1683 LASSERT(list_empty(&lock->l_bl_ast));
1684 list_add(&lock->l_bl_ast, &cancels);
1686 * This is a LDLM_CANCEL RPC which goes to canceld portal,
1687 * so we can cancel other LRU locks here and send them all
1688 * as one LDLM_CANCEL RPC.
1690 if (exp_connect_cancelset(exp)) {
1691 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1696 ns = ldlm_lock_to_ns(lock);
1697 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1698 LCF_BL_AST, lru_flags);
1700 ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1704 EXPORT_SYMBOL(ldlm_cli_cancel);
1707 * Locally cancel up to \a count locks in list \a cancels.
1708 * Return the number of cancelled locks.
1710 int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
1711 enum ldlm_cancel_flags cancel_flags)
1714 struct ldlm_lock *lock, *next;
1715 int left = 0, bl_ast = 0;
1719 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1723 if (cancel_flags & LCF_LOCAL) {
1724 rc = LDLM_FL_LOCAL_ONLY;
1725 ldlm_lock_cancel(lock);
1727 rc = ldlm_cli_cancel_local(lock);
1730 * Until we have compound requests and can send LDLM_CANCEL
1731 * requests batched with generic RPCs, we need to send cancels
1732 * with the LDLM_FL_BL_AST flag in a separate RPC from
1733 * the one being generated now.
1735 if (!(cancel_flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1736 LDLM_DEBUG(lock, "Cancel lock separately");
1737 list_move(&lock->l_bl_ast, &head);
1741 if (rc == LDLM_FL_LOCAL_ONLY) {
1742 /* CANCEL RPC should not be sent to server. */
1743 list_del_init(&lock->l_bl_ast);
1744 LDLM_LOCK_RELEASE(lock);
1750 ldlm_cli_cancel_list(&head, bl_ast, NULL, cancel_flags);
1757 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1758 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1759 * readahead requests, ...)
1761 static enum ldlm_policy_res
1762 ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1765 enum ldlm_policy_res result = LDLM_POLICY_CANCEL_LOCK;
1768 * don't check @added & @min since we want to process all locks
1770 * It's fine to not take lock to access lock->l_resource since
1771 * the lock has already been granted so it won't change.
1773 switch (lock->l_resource->lr_type) {
1776 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) != 0)
1780 result = LDLM_POLICY_SKIP_LOCK;
1788 * Callback function for LRU-resize policy. Decides whether to keep
1789 * \a lock in LRU for \a added in current scan and \a min number of locks
1790 * to be preferably canceled.
1792 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1794 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1796 static enum ldlm_policy_res ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1797 struct ldlm_lock *lock,
1800 ktime_t cur = ktime_get();
1801 struct ldlm_pool *pl = &ns->ns_pool;
1806 return LDLM_POLICY_CANCEL_LOCK;
1809 * Despite of the LV, It doesn't make sense to keep the lock which
1810 * is unused for ns_max_age time.
1812 if (ktime_after(cur, ktime_add(lock->l_last_used, ns->ns_max_age)))
1813 return LDLM_POLICY_CANCEL_LOCK;
1815 slv = ldlm_pool_get_slv(pl);
1816 lvf = ldlm_pool_get_lvf(pl);
1817 la = div_u64(ktime_to_ns(ktime_sub(cur, lock->l_last_used)),
1819 lv = lvf * la * ns->ns_nr_unused >> 8;
1821 /* Inform pool about current CLV to see it via debugfs. */
1822 ldlm_pool_set_clv(pl, lv);
1825 * Stop when SLV is not yet come from server or lv is smaller than
1828 if (slv == 0 || lv < slv)
1829 return LDLM_POLICY_KEEP_LOCK;
1831 return LDLM_POLICY_CANCEL_LOCK;
1834 static enum ldlm_policy_res
1835 ldlm_cancel_lrur_no_wait_policy(struct ldlm_namespace *ns,
1836 struct ldlm_lock *lock,
1839 enum ldlm_policy_res result;
1841 result = ldlm_cancel_lrur_policy(ns, lock, added, min);
1842 if (result == LDLM_POLICY_KEEP_LOCK)
1845 return ldlm_cancel_no_wait_policy(ns, lock, added, min);
1849 * Callback function for aged policy. Decides whether to keep
1850 * \a lock in LRU for \a added in current scan and \a min number of locks
1851 * to be preferably canceled.
1853 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1855 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1857 static enum ldlm_policy_res ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1858 struct ldlm_lock *lock,
1861 if ((added >= min) &&
1862 ktime_before(ktime_get(),
1863 ktime_add(lock->l_last_used, ns->ns_max_age)))
1864 return LDLM_POLICY_KEEP_LOCK;
1866 return LDLM_POLICY_CANCEL_LOCK;
1869 static enum ldlm_policy_res
1870 ldlm_cancel_aged_no_wait_policy(struct ldlm_namespace *ns,
1871 struct ldlm_lock *lock,
1874 enum ldlm_policy_res result;
1876 result = ldlm_cancel_aged_policy(ns, lock, added, min);
1877 if (result == LDLM_POLICY_KEEP_LOCK)
1880 return ldlm_cancel_no_wait_policy(ns, lock, added, min);
1883 typedef enum ldlm_policy_res
1884 (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1885 int added, int min);
1887 static ldlm_cancel_lru_policy_t
1888 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, enum ldlm_lru_flags lru_flags)
1890 if (ns_connect_lru_resize(ns)) {
1891 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1892 return ldlm_cancel_lrur_no_wait_policy;
1894 return ldlm_cancel_lrur_policy;
1896 if (lru_flags & LDLM_LRU_FLAG_NO_WAIT)
1897 return ldlm_cancel_aged_no_wait_policy;
1899 return ldlm_cancel_aged_policy;
1904 * - Free space in LRU for \a min new locks,
1905 * redundant unused locks are canceled locally;
1906 * - also cancel locally unused aged locks;
1907 * - do not cancel more than \a max locks;
1908 * - if some locks are cancelled, try to cancel at least \a batch locks
1909 * - GET the found locks and add them into the \a cancels list.
1911 * A client lock can be added to the l_bl_ast list only when it is
1912 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1913 * CANCEL. There are the following use cases:
1914 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1915 * ldlm_cli_cancel(), which check and set this flag properly. As any
1916 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1917 * later without any special locking.
1919 * Locks are cancelled according to the LRU resize policy (SLV from server)
1920 * if LRU resize is enabled; otherwise, the "aged policy" is used;
1923 * ----------------------------------------
1925 * flags & LDLM_LRU_FLAG_NO_WAIT - cancel locks w/o sending any RPCs or waiting
1926 * for any outstanding RPC to complete.
1928 * flags & LDLM_CANCEL_CLEANUP - when cancelling read locks, do not check for
1929 * other read locks covering the same pages, just
1930 * discard those pages.
1932 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns,
1933 struct list_head *cancels,
1934 int min, int max, int batch,
1935 enum ldlm_lru_flags lru_flags)
1937 ldlm_cancel_lru_policy_t pf;
1939 int no_wait = lru_flags & LDLM_LRU_FLAG_NO_WAIT;
1943 * Let only 1 thread to proceed. However, not for those which have the
1944 * @max limit given (ELC), as LRU may be left not cleaned up in full.
1947 if (test_and_set_bit(LDLM_LRU_CANCEL, &ns->ns_flags))
1949 } else if (test_bit(LDLM_LRU_CANCEL, &ns->ns_flags))
1952 LASSERT(ergo(max, min <= max));
1953 /* No sense to give @batch for ELC */
1954 LASSERT(ergo(max, batch == 0));
1956 if (!ns_connect_lru_resize(ns))
1957 min = max_t(int, min, ns->ns_nr_unused - ns->ns_max_unused);
1959 /* If at least 1 lock is to be cancelled, cancel at least @batch locks */
1960 if (min && min < batch)
1963 pf = ldlm_cancel_lru_policy(ns, lru_flags);
1964 LASSERT(pf != NULL);
1966 /* For any flags, stop scanning if @max is reached. */
1967 while (!list_empty(&ns->ns_unused_list) && (max == 0 || added < max)) {
1968 struct ldlm_lock *lock;
1969 struct list_head *item, *next;
1970 enum ldlm_policy_res result;
1971 ktime_t last_use = ktime_set(0, 0);
1973 spin_lock(&ns->ns_lock);
1974 item = no_wait ? ns->ns_last_pos : &ns->ns_unused_list;
1975 for (item = item->next, next = item->next;
1976 item != &ns->ns_unused_list;
1977 item = next, next = item->next) {
1978 lock = list_entry(item, struct ldlm_lock, l_lru);
1980 /* No locks which got blocking requests. */
1981 LASSERT(!ldlm_is_bl_ast(lock));
1983 if (!ldlm_is_canceling(lock))
1987 * Somebody is already doing CANCEL. No need for this
1988 * lock in LRU, do not traverse it again.
1990 ldlm_lock_remove_from_lru_nolock(lock);
1992 if (item == &ns->ns_unused_list) {
1993 spin_unlock(&ns->ns_lock);
1997 last_use = lock->l_last_used;
1999 LDLM_LOCK_GET(lock);
2000 spin_unlock(&ns->ns_lock);
2001 lu_ref_add(&lock->l_reference, __FUNCTION__, current);
2004 * Pass the lock through the policy filter and see if it
2005 * should stay in LRU.
2007 * Even for shrinker policy we stop scanning if
2008 * we find a lock that should stay in the cache.
2009 * We should take into account lock age anyway
2010 * as a new lock is a valuable resource even if
2011 * it has a low weight.
2013 * That is, for shrinker policy we drop only
2014 * old locks, but additionally choose them by
2015 * their weight. Big extent locks will stay in
2018 result = pf(ns, lock, added, min);
2019 if (result == LDLM_POLICY_KEEP_LOCK) {
2020 lu_ref_del(&lock->l_reference, __func__, current);
2021 LDLM_LOCK_RELEASE(lock);
2025 if (result == LDLM_POLICY_SKIP_LOCK) {
2026 lu_ref_del(&lock->l_reference, __func__, current);
2028 spin_lock(&ns->ns_lock);
2029 if (!list_empty(&lock->l_lru) &&
2030 lock->l_lru.prev == ns->ns_last_pos)
2031 ns->ns_last_pos = &lock->l_lru;
2032 spin_unlock(&ns->ns_lock);
2035 LDLM_LOCK_RELEASE(lock);
2039 lock_res_and_lock(lock);
2040 /* Check flags again under the lock. */
2041 if (ldlm_is_canceling(lock) ||
2042 ldlm_lock_remove_from_lru_check(lock, last_use) == 0) {
2044 * Another thread is removing lock from LRU, or
2045 * somebody is already doing CANCEL, or there
2046 * is a blocking request which will send cancel
2047 * by itself, or the lock is no longer unused or
2048 * the lock has been used since the pf() call and
2049 * pages could be put under it.
2051 unlock_res_and_lock(lock);
2052 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
2053 LDLM_LOCK_RELEASE(lock);
2056 LASSERT(!lock->l_readers && !lock->l_writers);
2059 * If we have chosen to cancel this lock voluntarily, we
2060 * better send cancel notification to server, so that it
2061 * frees appropriate state. This might lead to a race
2062 * where while we are doing cancel here, server is also
2063 * silently cancelling this lock.
2065 ldlm_clear_cancel_on_block(lock);
2068 * Setting the CBPENDING flag is a little misleading,
2069 * but prevents an important race; namely, once
2070 * CBPENDING is set, the lock can accumulate no more
2071 * readers/writers. Since readers and writers are
2072 * already zero here, ldlm_lock_decref() won't see
2073 * this flag and call l_blocking_ast
2075 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
2077 if ((lru_flags & LDLM_LRU_FLAG_CLEANUP) &&
2078 (lock->l_resource->lr_type == LDLM_EXTENT ||
2079 ldlm_has_dom(lock)) && lock->l_granted_mode == LCK_PR)
2080 ldlm_set_discard_data(lock);
2083 * We can't re-add to l_lru as it confuses the
2084 * refcounting in ldlm_lock_remove_from_lru() if an AST
2085 * arrives after we drop lr_lock below. We use l_bl_ast
2086 * and can't use l_pending_chain as it is used both on
2087 * server and client nevertheless b=5666 says it is
2088 * used only on server
2090 LASSERT(list_empty(&lock->l_bl_ast));
2091 list_add(&lock->l_bl_ast, cancels);
2092 unlock_res_and_lock(lock);
2093 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
2095 /* Once a lock added, batch the requested amount */
2101 clear_bit(LDLM_LRU_CANCEL, &ns->ns_flags);
2106 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
2108 enum ldlm_cancel_flags cancel_flags,
2109 enum ldlm_lru_flags lru_flags)
2113 added = ldlm_prepare_lru_list(ns, cancels, min, max, 0, lru_flags);
2117 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
2121 * Cancel at least \a min locks from given namespace LRU.
2123 * When called with LCF_ASYNC the blocking callback will be handled
2124 * in a thread and this function will return after the thread has been
2125 * asked to call the callback. When called with LCF_ASYNC the blocking
2126 * callback will be performed in this function.
2128 int ldlm_cancel_lru(struct ldlm_namespace *ns, int min,
2129 enum ldlm_cancel_flags cancel_flags,
2130 enum ldlm_lru_flags lru_flags)
2138 * Just prepare the list of locks, do not actually cancel them yet.
2139 * Locks are cancelled later in a separate thread.
2141 count = ldlm_prepare_lru_list(ns, &cancels, min, 0,
2142 ns->ns_cancel_batch, lru_flags);
2143 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
2151 * Find and cancel locally unused locks found on resource, matched to the
2152 * given policy, mode. GET the found locks and add them into the \a cancels
2155 int ldlm_cancel_resource_local(struct ldlm_resource *res,
2156 struct list_head *cancels,
2157 union ldlm_policy_data *policy,
2158 enum ldlm_mode mode, __u64 lock_flags,
2159 enum ldlm_cancel_flags cancel_flags,
2162 struct ldlm_lock *lock;
2168 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
2169 if (opaque != NULL && lock->l_ast_data != opaque) {
2170 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
2171 lock->l_ast_data, opaque);
2175 if (lock->l_readers || lock->l_writers)
2179 * If somebody is already doing CANCEL, or blocking AST came
2180 * then skip this lock.
2182 if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock))
2185 if (lockmode_compat(lock->l_granted_mode, mode))
2189 * If policy is given and this is IBITS lock, add to list only
2190 * those locks that match by policy.
2192 if (policy && (lock->l_resource->lr_type == LDLM_IBITS)) {
2193 if (!(lock->l_policy_data.l_inodebits.bits &
2194 policy->l_inodebits.bits))
2196 /* Skip locks with DoM bit if it is not set in policy
2197 * to don't flush data by side-bits. Lock convert will
2198 * drop those bits separately.
2200 if (ldlm_has_dom(lock) &&
2201 !(policy->l_inodebits.bits & MDS_INODELOCK_DOM))
2205 /* See CBPENDING comment in ldlm_cancel_lru */
2206 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
2208 LASSERT(list_empty(&lock->l_bl_ast));
2209 list_add(&lock->l_bl_ast, cancels);
2210 LDLM_LOCK_GET(lock);
2215 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
2217 EXPORT_SYMBOL(ldlm_cancel_resource_local);
2220 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
2222 * If \a req is NULL, send CANCEL request to server with handles of locks
2223 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
2224 * separately per lock.
2225 * If \a req is not NULL, put handles of locks in \a cancels into the request
2226 * buffer at the offset \a off.
2227 * Destroy \a cancels at the end.
2229 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
2230 struct ptlrpc_request *req,
2231 enum ldlm_cancel_flags flags)
2233 struct ldlm_lock *lock;
2238 if (list_empty(cancels) || count == 0)
2242 * XXX: requests (both batched and not) could be sent in parallel.
2243 * Usually it is enough to have just 1 RPC, but it is possible that
2244 * there are too many locks to be cancelled in LRU or on a resource.
2245 * It would also speed up the case when the server does not support
2249 LASSERT(!list_empty(cancels));
2250 lock = list_first_entry(cancels, struct ldlm_lock, l_bl_ast);
2251 LASSERT(lock->l_conn_export);
2253 if (exp_connect_cancelset(lock->l_conn_export)) {
2256 ldlm_cancel_pack(req, cancels, count);
2258 res = ldlm_cli_cancel_req(lock->l_conn_export,
2259 NULL, cancels, count,
2262 res = ldlm_cli_cancel_req(lock->l_conn_export,
2263 NULL, cancels, 1, flags);
2267 CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
2268 "ldlm_cli_cancel_list: %d\n", res);
2273 ldlm_lock_list_put(cancels, l_bl_ast, res);
2275 LASSERT(count == 0);
2278 EXPORT_SYMBOL(ldlm_cli_cancel_list);
2281 * Cancel all locks on a resource that have 0 readers/writers.
2283 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
2284 * to notify the server.
2286 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
2287 const struct ldlm_res_id *res_id,
2288 union ldlm_policy_data *policy,
2289 enum ldlm_mode mode,
2290 enum ldlm_cancel_flags flags, void *opaque)
2292 struct ldlm_resource *res;
2299 res = ldlm_resource_get(ns, res_id, 0, 0);
2301 /* This is not a problem. */
2302 CDEBUG(D_INFO, "No resource %llu\n", res_id->name[0]);
2306 LDLM_RESOURCE_ADDREF(res);
2307 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
2308 0, flags | LCF_BL_AST, opaque);
2309 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
2311 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
2314 LDLM_RESOURCE_DELREF(res);
2315 ldlm_resource_putref(res);
2318 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2320 struct ldlm_cli_cancel_arg {
2326 ldlm_cli_hash_cancel_unused(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2327 struct hlist_node *hnode, void *arg)
2329 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2330 struct ldlm_cli_cancel_arg *lc = arg;
2332 ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
2333 NULL, LCK_MINMODE, lc->lc_flags,
2335 /* must return 0 for hash iteration */
2340 * Cancel all locks on a namespace (or a specific resource, if given)
2341 * that have 0 readers/writers.
2343 * If flags & LCF_LOCAL, throw the locks away without trying
2344 * to notify the server.
2346 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
2347 const struct ldlm_res_id *res_id,
2348 enum ldlm_cancel_flags flags, void *opaque)
2350 struct ldlm_cli_cancel_arg arg = {
2352 .lc_opaque = opaque,
2360 if (res_id != NULL) {
2361 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
2365 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2366 ldlm_cli_hash_cancel_unused, &arg, 0);
2371 /* Lock iterators. */
2373 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
2376 struct ldlm_lock *tmp;
2377 struct ldlm_lock *lock;
2378 int rc = LDLM_ITER_CONTINUE;
2382 RETURN(LDLM_ITER_CONTINUE);
2385 list_for_each_entry_safe(lock, tmp, &res->lr_granted, l_res_link) {
2386 if (iter(lock, closure) == LDLM_ITER_STOP)
2387 GOTO(out, rc = LDLM_ITER_STOP);
2390 list_for_each_entry_safe(lock, tmp, &res->lr_waiting, l_res_link) {
2391 if (iter(lock, closure) == LDLM_ITER_STOP)
2392 GOTO(out, rc = LDLM_ITER_STOP);
2399 struct iter_helper_data {
2400 ldlm_iterator_t iter;
2404 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2406 struct iter_helper_data *helper = closure;
2408 return helper->iter(lock, helper->closure);
2411 static int ldlm_res_iter_helper(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2412 struct hlist_node *hnode, void *arg)
2415 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2417 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2421 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2422 ldlm_iterator_t iter, void *closure)
2425 struct iter_helper_data helper = { .iter = iter, .closure = closure };
2427 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2428 ldlm_res_iter_helper, &helper, 0);
2433 * non-blocking function to manipulate a lock whose cb_data is being put away.
2434 * return 0: find no resource
2435 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2438 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2439 const struct ldlm_res_id *res_id,
2440 ldlm_iterator_t iter, void *data)
2442 struct ldlm_resource *res;
2447 LASSERTF(ns != NULL, "must pass in namespace\n");
2449 res = ldlm_resource_get(ns, res_id, 0, 0);
2453 LDLM_RESOURCE_ADDREF(res);
2454 rc = ldlm_resource_foreach(res, iter, data);
2455 LDLM_RESOURCE_DELREF(res);
2456 ldlm_resource_putref(res);
2459 EXPORT_SYMBOL(ldlm_resource_iterate);
2462 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2464 struct list_head *list = closure;
2466 /* we use l_pending_chain here, because it's unused on clients. */
2467 LASSERTF(list_empty(&lock->l_pending_chain),
2468 "lock %p next %p prev %p\n",
2469 lock, &lock->l_pending_chain.next,
2470 &lock->l_pending_chain.prev);
2472 * b=9573: don't replay locks left after eviction, or
2473 * b=17614: locks being actively cancelled. Get a reference
2474 * on a lock so that it does not disapear under us (e.g. due to cancel)
2476 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_BL_DONE))) {
2477 list_add(&lock->l_pending_chain, list);
2478 LDLM_LOCK_GET(lock);
2481 return LDLM_ITER_CONTINUE;
2484 static int replay_lock_interpret(const struct lu_env *env,
2485 struct ptlrpc_request *req, void *args, int rc)
2487 struct ldlm_async_args *aa = args;
2488 struct ldlm_lock *lock;
2489 struct ldlm_reply *reply;
2490 struct obd_export *exp;
2493 atomic_dec(&req->rq_import->imp_replay_inflight);
2494 wake_up(&req->rq_import->imp_replay_waitq);
2499 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2501 GOTO(out, rc = -EPROTO);
2503 lock = ldlm_handle2lock(&aa->lock_handle);
2505 CERROR("received replay ack for unknown local cookie %#llx remote cookie %#llx from server %s id %s\n",
2506 aa->lock_handle.cookie, reply->lock_handle.cookie,
2507 req->rq_export->exp_client_uuid.uuid,
2508 libcfs_idstr(&req->rq_peer));
2509 GOTO(out, rc = -ESTALE);
2512 /* Key change rehash lock in per-export hash with new key */
2513 exp = req->rq_export;
2514 if (exp && exp->exp_lock_hash) {
2516 * In the function below, .hs_keycmp resolves to
2517 * ldlm_export_lock_keycmp()
2519 cfs_hash_rehash_key(exp->exp_lock_hash,
2520 &lock->l_remote_handle,
2521 &reply->lock_handle,
2524 lock->l_remote_handle = reply->lock_handle;
2527 LDLM_DEBUG(lock, "replayed lock:");
2528 ptlrpc_import_recovery_state_machine(req->rq_import);
2529 LDLM_LOCK_PUT(lock);
2532 ptlrpc_connect_import(req->rq_import);
2537 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2539 struct ptlrpc_request *req;
2540 struct ldlm_async_args *aa;
2541 struct ldlm_request *body;
2547 /* b=11974: Do not replay a lock which is actively being canceled */
2548 if (ldlm_is_bl_done(lock)) {
2549 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2554 * If this is reply-less callback lock, we cannot replay it, since
2555 * server might have long dropped it, but notification of that event was
2556 * lost by network. (and server granted conflicting lock already)
2558 if (ldlm_is_cancel_on_block(lock)) {
2559 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2560 ldlm_lock_cancel(lock);
2565 * If granted mode matches the requested mode, this lock is granted.
2567 * If we haven't been granted anything and are on a resource list,
2568 * then we're blocked/waiting.
2570 * If we haven't been granted anything and we're NOT on a resource list,
2571 * then we haven't got a reply yet and don't have a known disposition.
2572 * This happens whenever a lock enqueue is the request that triggers
2575 if (ldlm_is_granted(lock))
2576 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2577 else if (!list_empty(&lock->l_res_link))
2578 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2580 flags = LDLM_FL_REPLAY;
2582 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2583 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2587 /* We're part of recovery, so don't wait for it. */
2588 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2589 /* If the state changed while we were prepared, don't wait */
2590 req->rq_no_delay = 1;
2592 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2593 ldlm_lock2desc(lock, &body->lock_desc);
2594 body->lock_flags = ldlm_flags_to_wire(flags);
2596 ldlm_lock2handle(lock, &body->lock_handle[0]);
2597 if (lock->l_lvb_len > 0)
2598 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2599 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2601 ptlrpc_request_set_replen(req);
2603 * notify the server we've replayed all requests.
2604 * also, we mark the request to be put on a dedicated
2605 * queue to be processed after all request replayes.
2608 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2610 LDLM_DEBUG(lock, "replaying lock:");
2612 atomic_inc(&imp->imp_replay_inflight);
2613 aa = ptlrpc_req_async_args(aa, req);
2614 aa->lock_handle = body->lock_handle[0];
2615 req->rq_interpret_reply = replay_lock_interpret;
2616 ptlrpcd_add_req(req);
2622 * Cancel as many unused locks as possible before replay. since we are
2623 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2626 * Called only in recovery before replaying locks. there is no need to
2627 * replay locks that are unused. since the clients may hold thousands of
2628 * cached unused locks, dropping the unused locks can greatly reduce the
2629 * load on the servers at recovery time.
2631 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2637 "Dropping as many unused locks as possible before replay for namespace %s (%d)\n",
2638 ldlm_ns_name(ns), ns->ns_nr_unused);
2640 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_REPLAY_PAUSE, cfs_fail_val);
2643 * We don't need to care whether or not LRU resize is enabled
2644 * because the LDLM_LRU_FLAG_NO_WAIT policy doesn't use the
2647 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2648 LCF_LOCAL, LDLM_LRU_FLAG_NO_WAIT);
2650 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2651 canceled, ldlm_ns_name(ns));
2654 static int lock_can_replay(struct obd_import *imp)
2656 struct client_obd *cli = &imp->imp_obd->u.cli;
2658 CDEBUG(D_HA, "check lock replay limit, inflights = %u(%u)\n",
2659 atomic_read(&imp->imp_replay_inflight) - 1,
2660 cli->cl_max_rpcs_in_flight);
2662 /* +1 due to ldlm_lock_replay() increment */
2663 return atomic_read(&imp->imp_replay_inflight) <
2664 1 + min_t(u32, cli->cl_max_rpcs_in_flight, 8);
2667 static int __ldlm_replay_locks(struct obd_import *imp, bool rate_limit)
2669 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2671 struct ldlm_lock *lock, *next;
2676 while (atomic_read(&imp->imp_replay_inflight) != 1)
2679 /* don't replay locks if import failed recovery */
2680 if (imp->imp_vbr_failed)
2683 if (ldlm_cancel_unused_locks_before_replay)
2684 ldlm_cancel_unused_locks_for_replay(ns);
2686 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2688 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2689 list_del_init(&lock->l_pending_chain);
2690 /* If we disconnected in the middle - cleanup and let
2691 * reconnection to happen again. LU-14027 */
2692 if (rc || (imp->imp_state != LUSTRE_IMP_REPLAY_LOCKS)) {
2693 LDLM_LOCK_RELEASE(lock);
2696 rc = replay_one_lock(imp, lock);
2697 LDLM_LOCK_RELEASE(lock);
2700 wait_event_idle_exclusive(imp->imp_replay_waitq,
2701 lock_can_replay(imp));
2708 * Lock replay uses rate control and can sleep waiting so
2709 * must be in separate thread from ptlrpcd itself
2711 static int ldlm_lock_replay_thread(void *data)
2713 struct obd_import *imp = data;
2715 unshare_fs_struct();
2716 CDEBUG(D_HA, "lock replay thread %s to %s@%s\n",
2717 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
2718 imp->imp_connection->c_remote_uuid.uuid);
2720 __ldlm_replay_locks(imp, true);
2721 atomic_dec(&imp->imp_replay_inflight);
2722 ptlrpc_import_recovery_state_machine(imp);
2723 class_import_put(imp);
2728 int ldlm_replay_locks(struct obd_import *imp)
2730 struct task_struct *task;
2733 /* ensure this doesn't fall to 0 before all have been queued */
2734 if (atomic_inc_return(&imp->imp_replay_inflight) > 1) {
2735 atomic_dec(&imp->imp_replay_inflight);
2738 class_import_get(imp);
2740 task = kthread_run(ldlm_lock_replay_thread, imp, "ldlm_lock_replay");
2743 CDEBUG(D_HA, "can't start lock replay thread: rc = %d\n", rc);
2745 /* run lock replay without rate control */
2746 rc = __ldlm_replay_locks(imp, false);
2747 atomic_dec(&imp->imp_replay_inflight);
2748 class_import_put(imp);