1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_LDLM
40 #include <liblustre.h>
43 #include <lustre_dlm.h>
44 #include <obd_class.h>
47 #include "ldlm_internal.h"
49 int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
50 CFS_MODULE_PARM(ldlm_enqueue_min, "i", int, 0644,
51 "lock enqueue timeout minimum");
53 /* in client side, whether the cached locks will be canceled before replay */
54 unsigned int ldlm_cancel_unused_locks_before_replay = 1;
56 static void interrupted_completion_wait(void *data)
60 struct lock_wait_data {
61 struct ldlm_lock *lwd_lock;
65 struct ldlm_async_args {
66 struct lustre_handle lock_handle;
69 int ldlm_expired_completion_wait(void *data)
71 struct lock_wait_data *lwd = data;
72 struct ldlm_lock *lock = lwd->lwd_lock;
73 struct obd_import *imp;
74 struct obd_device *obd;
77 if (lock->l_conn_export == NULL) {
78 static cfs_time_t next_dump = 0, last_dump = 0;
80 if (ptlrpc_check_suspend())
83 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
84 CFS_DURATION_T"s ago); not entering recovery in "
85 "server code, just going back to sleep",
86 lock->l_last_activity,
87 cfs_time_sub(cfs_time_current_sec(),
88 lock->l_last_activity));
89 if (cfs_time_after(cfs_time_current(), next_dump)) {
90 last_dump = next_dump;
91 next_dump = cfs_time_shift(300);
92 ldlm_namespace_dump(D_DLMTRACE,
93 ldlm_lock_to_ns(lock));
95 libcfs_debug_dumplog();
100 obd = lock->l_conn_export->exp_obd;
101 imp = obd->u.cli.cl_import;
102 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
103 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
104 CFS_DURATION_T"s ago), entering recovery for %s@%s",
105 lock->l_last_activity,
106 cfs_time_sub(cfs_time_current_sec(), lock->l_last_activity),
107 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
112 /* We use the same basis for both server side and client side functions
113 from a single node. */
114 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
116 int timeout = at_get(ldlm_lock_to_ns_at(lock));
118 return obd_timeout / 2;
119 /* Since these are non-updating timeouts, we should be conservative.
120 It would be nice to have some kind of "early reply" mechanism for
121 lock callbacks too... */
122 timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */
123 return max(timeout, ldlm_enqueue_min);
125 EXPORT_SYMBOL(ldlm_get_enq_timeout);
128 * Helper function for ldlm_completion_ast(), updating timings when lock is
131 static int ldlm_completion_tail(struct ldlm_lock *lock)
136 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
137 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
140 delay = cfs_time_sub(cfs_time_current_sec(),
141 lock->l_last_activity);
142 LDLM_DEBUG(lock, "client-side enqueue: granted after "
143 CFS_DURATION_T"s", delay);
145 /* Update our time estimate */
146 at_measured(ldlm_lock_to_ns_at(lock),
154 * Implementation of ->l_completion_ast() for a client, that doesn't wait
155 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
156 * other threads that cannot block for long.
158 int ldlm_completion_ast_async(struct ldlm_lock *lock, int flags, void *data)
162 if (flags == LDLM_FL_WAIT_NOREPROC) {
163 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
167 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
168 LDLM_FL_BLOCK_CONV))) {
169 cfs_waitq_signal(&lock->l_waitq);
170 RETURN(ldlm_completion_tail(lock));
173 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
175 ldlm_lock_dump(D_OTHER, lock, 0);
176 ldlm_reprocess_all(lock->l_resource);
181 * Client side LDLM "completion" AST. This is called in several cases:
183 * - when a reply to an ENQUEUE rpc is received from the server
184 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
185 * this point (determined by flags);
187 * - when LDLM_CP_CALLBACK rpc comes to client to notify it that lock has
190 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
193 * - to force all locks when resource is destroyed (cleanup_resource());
195 * - during lock conversion (not used currently).
197 * If lock is not granted in the first case, this function waits until second
198 * or penultimate cases happen in some other thread.
201 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
203 /* XXX ALLOCATE - 160 bytes */
204 struct lock_wait_data lwd;
205 struct obd_device *obd;
206 struct obd_import *imp = NULL;
207 struct l_wait_info lwi;
212 if (flags == LDLM_FL_WAIT_NOREPROC) {
213 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
217 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
218 LDLM_FL_BLOCK_CONV))) {
219 cfs_waitq_signal(&lock->l_waitq);
223 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
225 ldlm_lock_dump(D_OTHER, lock, 0);
229 obd = class_exp2obd(lock->l_conn_export);
231 /* if this is a local lock, then there is no import */
233 imp = obd->u.cli.cl_import;
236 /* Wait a long time for enqueue - server may have to callback a
237 lock from another client. Server will evict the other client if it
238 doesn't respond reasonably, and then give us the lock. */
239 timeout = ldlm_get_enq_timeout(lock) * 2;
243 if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
244 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
245 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
247 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
248 ldlm_expired_completion_wait,
249 interrupted_completion_wait, &lwd);
253 cfs_spin_lock(&imp->imp_lock);
254 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
255 cfs_spin_unlock(&imp->imp_lock);
258 if (ns_is_client(ldlm_lock_to_ns(lock)) &&
259 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
260 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
261 lock->l_flags |= LDLM_FL_FAIL_LOC;
264 /* Go to sleep until the lock is granted or cancelled. */
265 rc = l_wait_event(lock->l_waitq,
266 is_granted_or_cancelled(lock), &lwi);
270 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
275 RETURN(ldlm_completion_tail(lock));
279 * A helper to build a blocking ast function
281 * Perform a common operation for blocking asts:
282 * defferred lock cancellation.
284 * \param lock the lock blocking or canceling ast was called on
286 * \see mdt_blocking_ast
287 * \see ldlm_blocking_ast
289 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
294 lock->l_flags |= LDLM_FL_CBPENDING;
295 do_ast = (!lock->l_readers && !lock->l_writers);
296 unlock_res_and_lock(lock);
299 struct lustre_handle lockh;
302 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
303 ldlm_lock2handle(lock, &lockh);
304 rc = ldlm_cli_cancel(&lockh);
306 CERROR("ldlm_cli_cancel: %d\n", rc);
308 LDLM_DEBUG(lock, "Lock still has references, will be "
315 * Server blocking AST
317 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
320 * \param lock the lock which blocks a request or cancelling lock
323 * \param flag indicates whether this cancelling or blocking callback
325 * \see ldlm_blocking_ast_nocheck
327 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
328 void *data, int flag)
332 if (flag == LDLM_CB_CANCELING) {
333 /* Don't need to do anything here. */
337 lock_res_and_lock(lock);
338 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
339 * that ldlm_blocking_ast is called just before intent_policy method
340 * takes the lr_lock, then by the time we get the lock, we might not
341 * be the correct blocking function anymore. So check, and return
343 if (lock->l_blocking_ast != ldlm_blocking_ast) {
344 unlock_res_and_lock(lock);
347 RETURN(ldlm_blocking_ast_nocheck(lock));
351 * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
352 * comment in filter_intent_policy() on why you may need this.
354 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
357 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
358 * that is rather subtle: with OST-side locking, it may so happen that
359 * _all_ extent locks are held by the OST. If client wants to obtain
360 * current file size it calls ll{,u}_glimpse_size(), and (as locks are
361 * on the server), dummy glimpse callback fires and does
362 * nothing. Client still receives correct file size due to the
363 * following fragment in filter_intent_policy():
365 * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
366 * if (rc != 0 && res->lr_namespace->ns_lvbo &&
367 * res->lr_namespace->ns_lvbo->lvbo_update) {
368 * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
371 * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
372 * returns correct file size to the client.
374 return -ELDLM_NO_LOCK_DATA;
377 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
378 const struct ldlm_res_id *res_id,
379 ldlm_type_t type, ldlm_policy_data_t *policy,
380 ldlm_mode_t mode, int *flags,
381 ldlm_blocking_callback blocking,
382 ldlm_completion_callback completion,
383 ldlm_glimpse_callback glimpse,
384 void *data, __u32 lvb_len,
385 const __u64 *client_cookie,
386 struct lustre_handle *lockh)
388 struct ldlm_lock *lock;
390 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
391 .lcs_blocking = blocking,
392 .lcs_glimpse = glimpse,
396 LASSERT(!(*flags & LDLM_FL_REPLAY));
397 if (unlikely(ns_is_client(ns))) {
398 CERROR("Trying to enqueue local lock in a shadow namespace\n");
402 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len);
404 GOTO(out_nolock, err = -ENOMEM);
406 ldlm_lock2handle(lock, lockh);
408 /* NB: we don't have any lock now (lock_res_and_lock)
409 * because it's a new lock */
410 ldlm_lock_addref_internal_nolock(lock, mode);
411 lock->l_flags |= LDLM_FL_LOCAL;
412 if (*flags & LDLM_FL_ATOMIC_CB)
413 lock->l_flags |= LDLM_FL_ATOMIC_CB;
416 lock->l_policy_data = *policy;
417 if (client_cookie != NULL)
418 lock->l_client_cookie = *client_cookie;
419 if (type == LDLM_EXTENT)
420 lock->l_req_extent = policy->l_extent;
422 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
423 if (unlikely(err != ELDLM_OK))
427 *policy = lock->l_policy_data;
429 if (lock->l_completion_ast)
430 lock->l_completion_ast(lock, *flags, NULL);
432 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
435 LDLM_LOCK_RELEASE(lock);
440 static void failed_lock_cleanup(struct ldlm_namespace *ns,
441 struct ldlm_lock *lock, int mode)
445 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
446 lock_res_and_lock(lock);
447 /* Check that lock is not granted or failed, we might race. */
448 if ((lock->l_req_mode != lock->l_granted_mode) &&
449 !(lock->l_flags & LDLM_FL_FAILED)) {
450 /* Make sure that this lock will not be found by raced
451 * bl_ast and -EINVAL reply is sent to server anyways.
453 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
454 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
457 unlock_res_and_lock(lock);
461 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
462 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
464 LDLM_DEBUG(lock, "lock was granted or failed in race");
466 ldlm_lock_decref_internal(lock, mode);
468 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
469 * from llite/file.c/ll_file_flock(). */
470 /* This code makes for the fact that we do not have blocking handler on
471 * a client for flock locks. As such this is the place where we must
472 * completely kill failed locks. (interrupted and those that
473 * were waiting to be granted when server evicted us. */
474 if (lock->l_resource->lr_type == LDLM_FLOCK) {
475 lock_res_and_lock(lock);
476 ldlm_resource_unlink_lock(lock);
477 ldlm_lock_destroy_nolock(lock);
478 unlock_res_and_lock(lock);
482 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
483 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
484 int *flags, void *lvb, __u32 lvb_len,
485 struct lustre_handle *lockh,int rc)
487 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
488 int is_replay = *flags & LDLM_FL_REPLAY;
489 struct ldlm_lock *lock;
490 struct ldlm_reply *reply;
491 struct ost_lvb *tmplvb;
492 int cleanup_phase = 1;
495 lock = ldlm_handle2lock(lockh);
496 /* ldlm_cli_enqueue is holding a reference on this lock. */
498 LASSERT(type == LDLM_FLOCK);
502 if (rc != ELDLM_OK) {
504 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
505 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
506 if (rc == ELDLM_LOCK_ABORTED) {
507 /* Before we return, swab the reply */
508 reply = req_capsule_server_get(&req->rq_pill,
514 req_capsule_set_size(&req->rq_pill,
515 &RMF_DLM_LVB, RCL_SERVER,
517 tmplvb = req_capsule_server_get(&req->rq_pill,
520 GOTO(cleanup, rc = -EPROTO);
522 memcpy(lvb, tmplvb, lvb_len);
528 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
530 GOTO(cleanup, rc = -EPROTO);
532 /* lock enqueued on the server */
535 lock_res_and_lock(lock);
536 /* Key change rehash lock in per-export hash with new key */
537 if (exp->exp_lock_hash) {
538 cfs_hash_rehash_key(exp->exp_lock_hash,
539 &lock->l_remote_handle,
543 lock->l_remote_handle = reply->lock_handle;
546 *flags = reply->lock_flags;
547 lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
548 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
549 * to wait with no timeout as well */
550 lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
551 unlock_res_and_lock(lock);
553 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
554 lock, reply->lock_handle.cookie, *flags);
556 /* If enqueue returned a blocked lock but the completion handler has
557 * already run, then it fixed up the resource and we don't need to do it
559 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
560 int newmode = reply->lock_desc.l_req_mode;
562 if (newmode && newmode != lock->l_req_mode) {
563 LDLM_DEBUG(lock, "server returned different mode %s",
564 ldlm_lockname[newmode]);
565 lock->l_req_mode = newmode;
568 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
569 lock->l_resource->lr_name.name,
570 sizeof(struct ldlm_res_id))) {
571 CDEBUG(D_INFO, "remote intent success, locking "
572 "(%ld,%ld,%ld) instead of "
574 (long)reply->lock_desc.l_resource.lr_name.name[0],
575 (long)reply->lock_desc.l_resource.lr_name.name[1],
576 (long)reply->lock_desc.l_resource.lr_name.name[2],
577 (long)lock->l_resource->lr_name.name[0],
578 (long)lock->l_resource->lr_name.name[1],
579 (long)lock->l_resource->lr_name.name[2]);
581 rc = ldlm_lock_change_resource(ns, lock,
582 &reply->lock_desc.l_resource.lr_name);
583 if (rc || lock->l_resource == NULL)
584 GOTO(cleanup, rc = -ENOMEM);
585 LDLM_DEBUG(lock, "client-side enqueue, new resource");
588 if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
590 lock->l_policy_data =
591 reply->lock_desc.l_policy_data;
592 if (type != LDLM_PLAIN)
593 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
596 if ((*flags) & LDLM_FL_AST_SENT ||
597 /* Cancel extent locks as soon as possible on a liblustre client,
598 * because it cannot handle asynchronous ASTs robustly (see
600 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
601 lock_res_and_lock(lock);
602 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
603 unlock_res_and_lock(lock);
604 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
607 /* If the lock has already been granted by a completion AST, don't
608 * clobber the LVB with an older one. */
610 /* We must lock or a racing completion might update lvb
611 without letting us know and we'll clobber the correct value.
612 Cannot unlock after the check either, a that still leaves
613 a tiny window for completion to get in */
614 lock_res_and_lock(lock);
615 if (lock->l_req_mode != lock->l_granted_mode) {
617 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
618 RCL_SERVER, lvb_len);
619 tmplvb = req_capsule_server_get(&req->rq_pill,
621 if (tmplvb == NULL) {
622 unlock_res_and_lock(lock);
623 GOTO(cleanup, rc = -EPROTO);
625 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
627 unlock_res_and_lock(lock);
631 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
632 if (lock->l_completion_ast != NULL) {
633 int err = lock->l_completion_ast(lock, *flags, NULL);
641 if (lvb_len && lvb != NULL) {
642 /* Copy the LVB here, and not earlier, because the completion
643 * AST (if any) can override what we got in the reply */
644 memcpy(lvb, lock->l_lvb_data, lvb_len);
647 LDLM_DEBUG(lock, "client-side enqueue END");
650 if (cleanup_phase == 1 && rc)
651 failed_lock_cleanup(ns, lock, mode);
652 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
654 LDLM_LOCK_RELEASE(lock);
658 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
659 * a single page on the send/receive side. XXX: 512 should be changed
660 * to more adequate value. */
661 static inline int ldlm_req_handles_avail(int req_size, int off)
665 avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size;
666 if (likely(avail >= 0))
667 avail /= (int)sizeof(struct lustre_handle);
670 avail += LDLM_LOCKREQ_HANDLES - off;
675 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
676 enum req_location loc,
679 int size = req_capsule_msg_size(pill, loc);
680 return ldlm_req_handles_avail(size, off);
683 static inline int ldlm_format_handles_avail(struct obd_import *imp,
684 const struct req_format *fmt,
685 enum req_location loc, int off)
687 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
688 return ldlm_req_handles_avail(size, off);
691 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
692 * @count locks in @cancels. */
693 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
694 int version, int opc, int canceloff,
695 cfs_list_t *cancels, int count)
697 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
698 struct req_capsule *pill = &req->rq_pill;
699 struct ldlm_request *dlm = NULL;
700 int flags, avail, to_free, bufcount, pack = 0;
707 if (exp_connect_cancelset(exp)) {
708 /* Estimate the amount of available space in the request. */
709 bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT);
710 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
712 flags = ns_connect_lru_resize(ns) ?
713 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
714 to_free = !ns_connect_lru_resize(ns) &&
715 opc == LDLM_ENQUEUE ? 1 : 0;
717 /* Cancel lru locks here _only_ if the server supports
718 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
719 * rpc, what will make us slower. */
721 count += ldlm_cancel_lru_local(ns, cancels, to_free,
722 avail - count, 0, flags);
727 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
728 ldlm_request_bufsize(pack, opc));
731 rc = ptlrpc_request_pack(req, version, opc);
733 ldlm_lock_list_put(cancels, l_bl_ast, count);
737 if (exp_connect_cancelset(exp)) {
739 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
741 /* Skip first lock handler in ldlm_request_pack(),
742 * this method will incrment @lock_count according
743 * to the lock handle amount actually written to
745 dlm->lock_count = canceloff;
747 /* Pack into the request @pack lock handles. */
748 ldlm_cli_cancel_list(cancels, pack, req, 0);
749 /* Prepare and send separate cancel rpc for others. */
750 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
752 ldlm_lock_list_put(cancels, l_bl_ast, count);
757 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
758 cfs_list_t *cancels, int count)
760 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
761 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
764 /* If a request has some specific initialisation it is passed in @reqp,
765 * otherwise it is created in ldlm_cli_enqueue.
767 * Supports sync and async requests, pass @async flag accordingly. If a
768 * request was created in ldlm_cli_enqueue and it is the async request,
769 * pass it to the caller in @reqp. */
770 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
771 struct ldlm_enqueue_info *einfo,
772 const struct ldlm_res_id *res_id,
773 ldlm_policy_data_t const *policy, int *flags,
774 void *lvb, __u32 lvb_len, struct lustre_handle *lockh,
777 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
778 struct ldlm_lock *lock;
779 struct ldlm_request *body;
780 int is_replay = *flags & LDLM_FL_REPLAY;
781 int req_passed_in = 1;
783 struct ptlrpc_request *req;
786 LASSERT(exp != NULL);
788 /* If we're replaying this lock, just check some invariants.
789 * If we're creating a new lock, get everything all setup nice. */
791 lock = ldlm_handle2lock_long(lockh, 0);
792 LASSERT(lock != NULL);
793 LDLM_DEBUG(lock, "client-side enqueue START");
794 LASSERT(exp == lock->l_conn_export);
796 const struct ldlm_callback_suite cbs = {
797 .lcs_completion = einfo->ei_cb_cp,
798 .lcs_blocking = einfo->ei_cb_bl,
799 .lcs_glimpse = einfo->ei_cb_gl,
800 .lcs_weigh = einfo->ei_cb_wg
802 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
803 einfo->ei_mode, &cbs, einfo->ei_cbdata,
807 /* for the local lock, add the reference */
808 ldlm_lock_addref_internal(lock, einfo->ei_mode);
809 ldlm_lock2handle(lock, lockh);
810 if (policy != NULL) {
811 /* INODEBITS_INTEROP: If the server does not support
812 * inodebits, we will request a plain lock in the
813 * descriptor (ldlm_lock2desc() below) but use an
814 * inodebits lock internally with both bits set.
816 if (einfo->ei_type == LDLM_IBITS &&
817 !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
818 lock->l_policy_data.l_inodebits.bits =
819 MDS_INODELOCK_LOOKUP |
820 MDS_INODELOCK_UPDATE;
822 lock->l_policy_data = *policy;
825 if (einfo->ei_type == LDLM_EXTENT)
826 lock->l_req_extent = policy->l_extent;
827 LDLM_DEBUG(lock, "client-side enqueue START");
830 /* lock not sent to server yet */
832 if (reqp == NULL || *reqp == NULL) {
833 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
838 failed_lock_cleanup(ns, lock, einfo->ei_mode);
839 LDLM_LOCK_RELEASE(lock);
849 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
851 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
852 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
855 lock->l_conn_export = exp;
856 lock->l_export = NULL;
857 lock->l_blocking_ast = einfo->ei_cb_bl;
859 /* Dump lock data into the request buffer */
860 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
861 ldlm_lock2desc(lock, &body->lock_desc);
862 body->lock_flags = *flags;
863 body->lock_handle[0] = *lockh;
865 /* Continue as normal. */
866 if (!req_passed_in) {
868 req_capsule_extend(&req->rq_pill,
869 &RQF_LDLM_ENQUEUE_LVB);
870 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
871 RCL_SERVER, lvb_len);
873 ptlrpc_request_set_replen(req);
877 * Liblustre client doesn't get extent locks, except for O_APPEND case
878 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
879 * [i_size, OBD_OBJECT_EOF] lock is taken.
881 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
882 policy->l_extent.end == OBD_OBJECT_EOF));
885 LASSERT(reqp != NULL);
889 LDLM_DEBUG(lock, "sending request");
891 rc = ptlrpc_queue_wait(req);
893 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
894 einfo->ei_mode, flags, lvb, lvb_len,
897 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
898 * one reference that we took */
900 LDLM_LOCK_RELEASE(lock);
904 if (!req_passed_in && req != NULL) {
905 ptlrpc_req_finished(req);
913 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
916 struct ldlm_resource *res;
919 if (ns_is_client(ldlm_lock_to_ns(lock))) {
920 CERROR("Trying to cancel local lock\n");
923 LDLM_DEBUG(lock, "client-side local convert");
925 res = ldlm_lock_convert(lock, new_mode, flags);
927 ldlm_reprocess_all(res);
932 LDLM_DEBUG(lock, "client-side local convert handler END");
937 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
938 * conversion of locks which are on the waiting or converting queue */
939 /* Caller of this code is supposed to take care of lock readers/writers
941 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
943 struct ldlm_request *body;
944 struct ldlm_reply *reply;
945 struct ldlm_lock *lock;
946 struct ldlm_resource *res;
947 struct ptlrpc_request *req;
951 lock = ldlm_handle2lock(lockh);
958 if (lock->l_conn_export == NULL)
959 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
961 LDLM_DEBUG(lock, "client-side convert");
963 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
964 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
971 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
972 body->lock_handle[0] = lock->l_remote_handle;
974 body->lock_desc.l_req_mode = new_mode;
975 body->lock_flags = *flags;
978 ptlrpc_request_set_replen(req);
979 rc = ptlrpc_queue_wait(req);
983 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
985 GOTO(out, rc = -EPROTO);
988 GOTO(out, rc = req->rq_status);
990 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
992 ldlm_reprocess_all(res);
993 /* Go to sleep until the lock is granted. */
994 /* FIXME: or cancelled. */
995 if (lock->l_completion_ast) {
996 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1006 LDLM_LOCK_PUT(lock);
1007 ptlrpc_req_finished(req);
1011 /* Cancel locks locally.
1013 * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
1014 * LDLM_FL_CANCELING otherwise;
1015 * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
1016 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
1018 int rc = LDLM_FL_LOCAL_ONLY;
1021 if (lock->l_conn_export) {
1024 LDLM_DEBUG(lock, "client-side cancel");
1025 /* Set this flag to prevent others from getting new references*/
1026 lock_res_and_lock(lock);
1027 lock->l_flags |= LDLM_FL_CBPENDING;
1028 local_only = (lock->l_flags &
1029 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1030 ldlm_cancel_callback(lock);
1031 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
1032 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1033 unlock_res_and_lock(lock);
1036 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1038 rc = LDLM_FL_LOCAL_ONLY;
1040 ldlm_lock_cancel(lock);
1042 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1043 LDLM_ERROR(lock, "Trying to cancel local lock");
1046 LDLM_DEBUG(lock, "server-side local cancel");
1047 ldlm_lock_cancel(lock);
1048 ldlm_reprocess_all(lock->l_resource);
1054 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
1055 of the request @req. */
1056 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1057 cfs_list_t *head, int count)
1059 struct ldlm_request *dlm;
1060 struct ldlm_lock *lock;
1061 int max, packed = 0;
1064 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1065 LASSERT(dlm != NULL);
1067 /* Check the room in the request buffer. */
1068 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1069 sizeof(struct ldlm_request);
1070 max /= sizeof(struct lustre_handle);
1071 max += LDLM_LOCKREQ_HANDLES;
1072 LASSERT(max >= dlm->lock_count + count);
1074 /* XXX: it would be better to pack lock handles grouped by resource.
1075 * so that the server cancel would call filter_lvbo_update() less
1077 cfs_list_for_each_entry(lock, head, l_bl_ast) {
1080 LASSERT(lock->l_conn_export);
1081 /* Pack the lock handle to the given request buffer. */
1082 LDLM_DEBUG(lock, "packing");
1083 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1086 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1090 /* Prepare and send a batched cancel rpc, it will include count lock handles
1091 * of locks given in @head. */
1092 int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels,
1093 int count, ldlm_cancel_flags_t flags)
1095 struct ptlrpc_request *req = NULL;
1096 struct obd_import *imp;
1101 LASSERT(exp != NULL);
1104 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, obd_fail_val);
1106 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1109 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1110 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1117 imp = class_exp2cliimp(exp);
1118 if (imp == NULL || imp->imp_invalid) {
1120 "skipping cancel on invalid import %p\n", imp);
1124 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1126 GOTO(out, rc = -ENOMEM);
1128 bufcount = req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1129 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1130 ldlm_request_bufsize(count, LDLM_CANCEL));
1132 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1134 ptlrpc_request_free(req);
1137 req->rq_no_resend = 1;
1138 req->rq_no_delay = 1;
1140 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1141 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1142 ptlrpc_at_set_req_timeout(req);
1144 ldlm_cancel_pack(req, cancels, count);
1146 ptlrpc_request_set_replen(req);
1147 if (flags & LCF_ASYNC) {
1148 ptlrpcd_add_req(req, PSCOPE_OTHER);
1152 rc = ptlrpc_queue_wait(req);
1155 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1156 "out of sync -- not fatal\n",
1157 libcfs_nid2str(req->rq_import->
1158 imp_connection->c_peer.nid));
1160 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1161 req->rq_import_generation == imp->imp_generation) {
1162 ptlrpc_req_finished(req);
1164 } else if (rc != ELDLM_OK) {
1165 CERROR("Got rc %d from cancel RPC: canceling "
1173 ptlrpc_req_finished(req);
1176 return sent ? sent : rc;
1179 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1181 LASSERT(imp != NULL);
1182 return &imp->imp_obd->obd_namespace->ns_pool;
1186 * Update client's obd pool related fields with new SLV and Limit from \a req.
1188 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1190 struct obd_device *obd;
1191 __u64 old_slv, new_slv;
1194 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1195 !imp_connect_lru_resize(req->rq_import)))
1198 * Do nothing for corner cases.
1204 * In some cases RPC may contain slv and limit zeroed out. This is
1205 * the case when server does not support lru resize feature. This is
1206 * also possible in some recovery cases when server side reqs have no
1207 * ref to obd export and thus access to server side namespace is no
1210 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1211 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1212 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1213 "(SLV: "LPU64", Limit: %u)",
1214 lustre_msg_get_slv(req->rq_repmsg),
1215 lustre_msg_get_limit(req->rq_repmsg));
1219 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1220 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1221 obd = req->rq_import->imp_obd;
1224 * Set new SLV and Limit to obd fields to make accessible for pool
1225 * thread. We do not access obd_namespace and pool directly here
1226 * as there is no reliable way to make sure that they are still
1227 * alive in cleanup time. Evil races are possible which may cause
1228 * oops in that time.
1230 cfs_write_lock(&obd->obd_pool_lock);
1231 old_slv = obd->obd_pool_slv;
1232 obd->obd_pool_slv = new_slv;
1233 obd->obd_pool_limit = new_limit;
1234 cfs_write_unlock(&obd->obd_pool_lock);
1238 EXPORT_SYMBOL(ldlm_cli_update_pool);
1240 int ldlm_cli_cancel(struct lustre_handle *lockh)
1242 struct obd_export *exp;
1243 int avail, flags, count = 1, rc = 0;
1244 struct ldlm_namespace *ns;
1245 struct ldlm_lock *lock;
1246 CFS_LIST_HEAD(cancels);
1249 /* concurrent cancels on the same handle can happen */
1250 lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1252 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1256 rc = ldlm_cli_cancel_local(lock);
1257 if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) {
1258 LDLM_LOCK_RELEASE(lock);
1259 RETURN(rc < 0 ? rc : 0);
1261 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1262 * rpc which goes to canceld portal, so we can cancel other lru locks
1263 * here and send them all as one LDLM_CANCEL rpc. */
1264 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1265 cfs_list_add(&lock->l_bl_ast, &cancels);
1267 exp = lock->l_conn_export;
1268 if (exp_connect_cancelset(exp)) {
1269 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1274 ns = ldlm_lock_to_ns(lock);
1275 flags = ns_connect_lru_resize(ns) ?
1276 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1277 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1280 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1284 /* XXX until we will have compound requests and can cut cancels from generic rpc
1285 * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */
1286 int ldlm_cli_cancel_list_local(cfs_list_t *cancels, int count,
1287 ldlm_cancel_flags_t flags)
1289 CFS_LIST_HEAD(head);
1290 struct ldlm_lock *lock, *next;
1291 int left = 0, bl_ast = 0, rc;
1294 cfs_list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1298 if (flags & LCF_LOCAL) {
1299 rc = LDLM_FL_LOCAL_ONLY;
1300 ldlm_lock_cancel(lock);
1302 rc = ldlm_cli_cancel_local(lock);
1304 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1305 LDLM_DEBUG(lock, "Cancel lock separately");
1306 cfs_list_del_init(&lock->l_bl_ast);
1307 cfs_list_add(&lock->l_bl_ast, &head);
1311 if (rc == LDLM_FL_LOCAL_ONLY) {
1312 /* CANCEL RPC should not be sent to server. */
1313 cfs_list_del_init(&lock->l_bl_ast);
1314 LDLM_LOCK_RELEASE(lock);
1321 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1328 * Cancel as many locks as possible w/o sending any rpcs (e.g. to write back
1329 * dirty data, to close a file, ...) or waiting for any rpcs in-flight (e.g.
1330 * readahead requests, ...)
1332 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1333 struct ldlm_lock *lock,
1334 int unused, int added,
1337 ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1338 ldlm_cancel_for_recovery cb = ns->ns_cancel_for_recovery;
1339 lock_res_and_lock(lock);
1341 /* don't check added & count since we want to process all locks
1342 * from unused list */
1343 switch (lock->l_resource->lr_type) {
1349 result = LDLM_POLICY_SKIP_LOCK;
1350 lock->l_flags |= LDLM_FL_SKIPPED;
1354 unlock_res_and_lock(lock);
1359 * Callback function for lru-resize policy. Makes decision whether to keep
1360 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1361 * \a added and number of locks to be preferably canceled \a count.
1363 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1365 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1367 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1368 struct ldlm_lock *lock,
1369 int unused, int added,
1372 cfs_time_t cur = cfs_time_current();
1373 struct ldlm_pool *pl = &ns->ns_pool;
1378 * Stop lru processing when we reached passed @count or checked all
1381 if (count && added >= count)
1382 return LDLM_POLICY_KEEP_LOCK;
1384 slv = ldlm_pool_get_slv(pl);
1385 lvf = ldlm_pool_get_lvf(pl);
1386 la = cfs_duration_sec(cfs_time_sub(cur,
1387 lock->l_last_used));
1390 * Stop when slv is not yet come from server or lv is smaller than
1393 lv = lvf * la * unused;
1396 * Inform pool about current CLV to see it via proc.
1398 ldlm_pool_set_clv(pl, lv);
1399 return (slv == 0 || lv < slv) ?
1400 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1404 * Callback function for proc used policy. Makes decision whether to keep
1405 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1406 * \a added and number of locks to be preferably canceled \a count.
1408 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1410 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1412 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1413 struct ldlm_lock *lock,
1414 int unused, int added,
1418 * Stop lru processing when we reached passed @count or checked all
1421 return (added >= count) ?
1422 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1426 * Callback function for aged policy. Makes decision whether to keep
1427 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1428 * \a added and number of locks to be preferably canceled \a count.
1430 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1432 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1434 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1435 struct ldlm_lock *lock,
1436 int unused, int added,
1440 * Stop lru processing if young lock is found and we reached passed
1443 return ((added >= count) &&
1444 cfs_time_before(cfs_time_current(),
1445 cfs_time_add(lock->l_last_used,
1447 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1451 * Callback function for default policy. Makes decision whether to keep
1452 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1453 * \a added and number of locks to be preferably canceled \a count.
1455 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1457 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1459 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1460 struct ldlm_lock *lock,
1461 int unused, int added,
1465 * Stop lru processing when we reached passed @count or checked all
1468 return (added >= count) ?
1469 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1472 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1473 struct ldlm_lock *, int,
1476 static ldlm_cancel_lru_policy_t
1477 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1479 if (flags & LDLM_CANCEL_NO_WAIT)
1480 return ldlm_cancel_no_wait_policy;
1482 if (ns_connect_lru_resize(ns)) {
1483 if (flags & LDLM_CANCEL_SHRINK)
1484 /* We kill passed number of old locks. */
1485 return ldlm_cancel_passed_policy;
1486 else if (flags & LDLM_CANCEL_LRUR)
1487 return ldlm_cancel_lrur_policy;
1488 else if (flags & LDLM_CANCEL_PASSED)
1489 return ldlm_cancel_passed_policy;
1491 if (flags & LDLM_CANCEL_AGED)
1492 return ldlm_cancel_aged_policy;
1495 return ldlm_cancel_default_policy;
1498 /* - Free space in lru for @count new locks,
1499 * redundant unused locks are canceled locally;
1500 * - also cancel locally unused aged locks;
1501 * - do not cancel more than @max locks;
1502 * - GET the found locks and add them into the @cancels list.
1504 * A client lock can be added to the l_bl_ast list only when it is
1505 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
1506 * There are the following use cases: ldlm_cancel_resource_local(),
1507 * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
1508 * flag properly. As any attempt to cancel a lock rely on this flag,
1509 * l_bl_ast list is accessed later without any special locking.
1511 * Calling policies for enabled lru resize:
1512 * ----------------------------------------
1513 * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to
1514 * cancel not more than @count locks;
1516 * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at
1517 * the beginning of lru list);
1519 * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
1520 * memory pressre policy function;
1522 * flags & LDLM_CANCEL_AGED - cancel alocks according to "aged policy".
1524 * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1525 * (typically before replaying locks) w/o
1526 * sending any rpcs or waiting for any
1527 * outstanding rpc to complete.
1529 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels,
1530 int count, int max, int flags)
1532 ldlm_cancel_lru_policy_t pf;
1533 struct ldlm_lock *lock, *next;
1534 int added = 0, unused, remained;
1537 cfs_spin_lock(&ns->ns_lock);
1538 unused = ns->ns_nr_unused;
1541 if (!ns_connect_lru_resize(ns))
1542 count += unused - ns->ns_max_unused;
1544 pf = ldlm_cancel_lru_policy(ns, flags);
1545 LASSERT(pf != NULL);
1547 while (!cfs_list_empty(&ns->ns_unused_list)) {
1548 ldlm_policy_res_t result;
1550 /* all unused locks */
1551 if (remained-- <= 0)
1554 /* For any flags, stop scanning if @max is reached. */
1555 if (max && added >= max)
1558 cfs_list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1560 /* No locks which got blocking requests. */
1561 LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1563 if (flags & LDLM_CANCEL_NO_WAIT &&
1564 lock->l_flags & LDLM_FL_SKIPPED)
1565 /* already processed */
1568 /* Somebody is already doing CANCEL. No need in this
1569 * lock in lru, do not traverse it again. */
1570 if (!(lock->l_flags & LDLM_FL_CANCELING))
1573 ldlm_lock_remove_from_lru_nolock(lock);
1575 if (&lock->l_lru == &ns->ns_unused_list)
1578 LDLM_LOCK_GET(lock);
1579 cfs_spin_unlock(&ns->ns_lock);
1580 lu_ref_add(&lock->l_reference, __FUNCTION__, cfs_current());
1582 /* Pass the lock through the policy filter and see if it
1583 * should stay in lru.
1585 * Even for shrinker policy we stop scanning if
1586 * we find a lock that should stay in the cache.
1587 * We should take into account lock age anyway
1588 * as new lock even if it is small of weight is
1589 * valuable resource.
1591 * That is, for shrinker policy we drop only
1592 * old locks, but additionally chose them by
1593 * their weight. Big extent locks will stay in
1595 result = pf(ns, lock, unused, added, count);
1596 if (result == LDLM_POLICY_KEEP_LOCK) {
1597 lu_ref_del(&lock->l_reference,
1598 __FUNCTION__, cfs_current());
1599 LDLM_LOCK_RELEASE(lock);
1600 cfs_spin_lock(&ns->ns_lock);
1603 if (result == LDLM_POLICY_SKIP_LOCK) {
1604 lu_ref_del(&lock->l_reference,
1605 __FUNCTION__, cfs_current());
1606 LDLM_LOCK_RELEASE(lock);
1607 cfs_spin_lock(&ns->ns_lock);
1611 lock_res_and_lock(lock);
1612 /* Check flags again under the lock. */
1613 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1614 (ldlm_lock_remove_from_lru(lock) == 0)) {
1615 /* other thread is removing lock from lru or
1616 * somebody is already doing CANCEL or
1617 * there is a blocking request which will send
1618 * cancel by itseft or the lock is matched
1619 * is already not unused. */
1620 unlock_res_and_lock(lock);
1621 lu_ref_del(&lock->l_reference,
1622 __FUNCTION__, cfs_current());
1623 LDLM_LOCK_RELEASE(lock);
1624 cfs_spin_lock(&ns->ns_lock);
1627 LASSERT(!lock->l_readers && !lock->l_writers);
1629 /* If we have chosen to cancel this lock voluntarily, we
1630 * better send cancel notification to server, so that it
1631 * frees appropriate state. This might lead to a race
1632 * where while we are doing cancel here, server is also
1633 * silently cancelling this lock. */
1634 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1636 /* Setting the CBPENDING flag is a little misleading,
1637 * but prevents an important race; namely, once
1638 * CBPENDING is set, the lock can accumulate no more
1639 * readers/writers. Since readers and writers are
1640 * already zero here, ldlm_lock_decref() won't see
1641 * this flag and call l_blocking_ast */
1642 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1644 /* We can't re-add to l_lru as it confuses the
1645 * refcounting in ldlm_lock_remove_from_lru() if an AST
1646 * arrives after we drop lr_lock below. We use l_bl_ast
1647 * and can't use l_pending_chain as it is used both on
1648 * server and client nevertheless bug 5666 says it is
1649 * used only on server */
1650 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1651 cfs_list_add(&lock->l_bl_ast, cancels);
1652 unlock_res_and_lock(lock);
1653 lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current());
1654 cfs_spin_lock(&ns->ns_lock);
1658 cfs_spin_unlock(&ns->ns_lock);
1662 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, cfs_list_t *cancels,
1663 int count, int max, ldlm_cancel_flags_t cancel_flags,
1667 added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1670 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1673 /* when called with LDLM_ASYNC the blocking callback will be handled
1674 * in a thread and this function will return after the thread has been
1675 * asked to call the callback. when called with LDLM_SYNC the blocking
1676 * callback will be performed in this function. */
1677 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t mode,
1680 CFS_LIST_HEAD(cancels);
1685 mode = LDLM_SYNC; /* force to be sync in user space */
1687 /* Just prepare the list of locks, do not actually cancel them yet.
1688 * Locks are cancelled later in a separate thread. */
1689 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1690 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, mode);
1697 /* Find and cancel locally unused locks found on resource, matched to the
1698 * given policy, mode. GET the found locks and add them into the @cancels
1700 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1701 cfs_list_t *cancels,
1702 ldlm_policy_data_t *policy,
1703 ldlm_mode_t mode, int lock_flags,
1704 ldlm_cancel_flags_t cancel_flags, void *opaque)
1706 struct ldlm_lock *lock;
1711 cfs_list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1712 if (opaque != NULL && lock->l_ast_data != opaque) {
1713 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1714 lock->l_ast_data, opaque);
1719 if (lock->l_readers || lock->l_writers)
1722 /* If somebody is already doing CANCEL, or blocking ast came,
1723 * skip this lock. */
1724 if (lock->l_flags & LDLM_FL_BL_AST ||
1725 lock->l_flags & LDLM_FL_CANCELING)
1728 if (lockmode_compat(lock->l_granted_mode, mode))
1731 /* If policy is given and this is IBITS lock, add to list only
1732 * those locks that match by policy. */
1733 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1734 !(lock->l_policy_data.l_inodebits.bits &
1735 policy->l_inodebits.bits))
1738 /* See CBPENDING comment in ldlm_cancel_lru */
1739 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1742 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1743 cfs_list_add(&lock->l_bl_ast, cancels);
1744 LDLM_LOCK_GET(lock);
1749 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
1752 /* If @req is NULL, send CANCEL request to server with handles of locks
1753 * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1754 * separately per lock.
1755 * If @req is not NULL, put handles of locks in @cancels into the request
1756 * buffer at the offset @off.
1757 * Destroy @cancels at the end. */
1758 int ldlm_cli_cancel_list(cfs_list_t *cancels, int count,
1759 struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1761 struct ldlm_lock *lock;
1765 if (cfs_list_empty(cancels) || count == 0)
1768 /* XXX: requests (both batched and not) could be sent in parallel.
1769 * Usually it is enough to have just 1 RPC, but it is possible that
1770 * there are to many locks to be cancelled in LRU or on a resource.
1771 * It would also speed up the case when the server does not support
1774 LASSERT(!cfs_list_empty(cancels));
1775 lock = cfs_list_entry(cancels->next, struct ldlm_lock,
1777 LASSERT(lock->l_conn_export);
1779 if (exp_connect_cancelset(lock->l_conn_export)) {
1782 ldlm_cancel_pack(req, cancels, count);
1784 res = ldlm_cli_cancel_req(lock->l_conn_export,
1788 res = ldlm_cli_cancel_req(lock->l_conn_export,
1793 CERROR("ldlm_cli_cancel_list: %d\n", res);
1798 ldlm_lock_list_put(cancels, l_bl_ast, res);
1800 LASSERT(count == 0);
1804 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1805 const struct ldlm_res_id *res_id,
1806 ldlm_policy_data_t *policy,
1808 ldlm_cancel_flags_t flags,
1811 struct ldlm_resource *res;
1812 CFS_LIST_HEAD(cancels);
1817 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1819 /* This is not a problem. */
1820 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1824 LDLM_RESOURCE_ADDREF(res);
1825 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1826 0, flags | LCF_BL_AST, opaque);
1827 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1829 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1831 LDLM_RESOURCE_DELREF(res);
1832 ldlm_resource_putref(res);
1836 struct ldlm_cli_cancel_arg {
1841 static int ldlm_cli_hash_cancel_unused(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1842 cfs_hlist_node_t *hnode, void *arg)
1844 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1845 struct ldlm_cli_cancel_arg *lc = arg;
1848 rc = ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1850 lc->lc_flags, lc->lc_opaque);
1852 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1853 res->lr_name.name[0], rc);
1855 /* must return 0 for hash iteration */
1859 /* Cancel all locks on a namespace (or a specific resource, if given)
1860 * that have 0 readers/writers.
1862 * If flags & LCF_LOCAL, throw the locks away without trying
1863 * to notify the server. */
1864 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1865 const struct ldlm_res_id *res_id,
1866 ldlm_cancel_flags_t flags, void *opaque)
1868 struct ldlm_cli_cancel_arg arg = {
1870 .lc_opaque = opaque,
1878 if (res_id != NULL) {
1879 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1883 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1884 ldlm_cli_hash_cancel_unused, &arg);
1889 /* Lock iterators. */
1891 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1894 cfs_list_t *tmp, *next;
1895 struct ldlm_lock *lock;
1896 int rc = LDLM_ITER_CONTINUE;
1901 RETURN(LDLM_ITER_CONTINUE);
1904 cfs_list_for_each_safe(tmp, next, &res->lr_granted) {
1905 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1907 if (iter(lock, closure) == LDLM_ITER_STOP)
1908 GOTO(out, rc = LDLM_ITER_STOP);
1911 cfs_list_for_each_safe(tmp, next, &res->lr_converting) {
1912 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1914 if (iter(lock, closure) == LDLM_ITER_STOP)
1915 GOTO(out, rc = LDLM_ITER_STOP);
1918 cfs_list_for_each_safe(tmp, next, &res->lr_waiting) {
1919 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1921 if (iter(lock, closure) == LDLM_ITER_STOP)
1922 GOTO(out, rc = LDLM_ITER_STOP);
1929 struct iter_helper_data {
1930 ldlm_iterator_t iter;
1934 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1936 struct iter_helper_data *helper = closure;
1937 return helper->iter(lock, helper->closure);
1940 static int ldlm_res_iter_helper(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1941 cfs_hlist_node_t *hnode, void *arg)
1944 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1946 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
1950 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
1951 ldlm_iterator_t iter, void *closure)
1954 struct iter_helper_data helper = { iter: iter, closure: closure };
1956 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1957 ldlm_res_iter_helper, &helper);
1961 /* non-blocking function to manipulate a lock whose cb_data is being put away.
1962 * return 0: find no resource
1963 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
1966 int ldlm_resource_iterate(struct ldlm_namespace *ns,
1967 const struct ldlm_res_id *res_id,
1968 ldlm_iterator_t iter, void *data)
1970 struct ldlm_resource *res;
1975 CERROR("must pass in namespace\n");
1979 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1983 LDLM_RESOURCE_ADDREF(res);
1984 rc = ldlm_resource_foreach(res, iter, data);
1985 LDLM_RESOURCE_DELREF(res);
1986 ldlm_resource_putref(res);
1992 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1994 cfs_list_t *list = closure;
1996 /* we use l_pending_chain here, because it's unused on clients. */
1997 LASSERTF(cfs_list_empty(&lock->l_pending_chain),
1998 "lock %p next %p prev %p\n",
1999 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2000 /* bug 9573: don't replay locks left after eviction, or
2001 * bug 17614: locks being actively cancelled. Get a reference
2002 * on a lock so that it does not disapear under us (e.g. due to cancel)
2004 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
2005 cfs_list_add(&lock->l_pending_chain, list);
2006 LDLM_LOCK_GET(lock);
2009 return LDLM_ITER_CONTINUE;
2012 static int replay_lock_interpret(const struct lu_env *env,
2013 struct ptlrpc_request *req,
2014 struct ldlm_async_args *aa, int rc)
2016 struct ldlm_lock *lock;
2017 struct ldlm_reply *reply;
2018 struct obd_export *exp;
2021 cfs_atomic_dec(&req->rq_import->imp_replay_inflight);
2026 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2028 GOTO(out, rc = -EPROTO);
2030 lock = ldlm_handle2lock(&aa->lock_handle);
2032 CERROR("received replay ack for unknown local cookie "LPX64
2033 " remote cookie "LPX64 " from server %s id %s\n",
2034 aa->lock_handle.cookie, reply->lock_handle.cookie,
2035 req->rq_export->exp_client_uuid.uuid,
2036 libcfs_id2str(req->rq_peer));
2037 GOTO(out, rc = -ESTALE);
2040 /* Key change rehash lock in per-export hash with new key */
2041 exp = req->rq_export;
2042 if (exp && exp->exp_lock_hash) {
2043 cfs_hash_rehash_key(exp->exp_lock_hash,
2044 &lock->l_remote_handle,
2045 &reply->lock_handle,
2048 lock->l_remote_handle = reply->lock_handle;
2051 LDLM_DEBUG(lock, "replayed lock:");
2052 ptlrpc_import_recovery_state_machine(req->rq_import);
2053 LDLM_LOCK_PUT(lock);
2056 ptlrpc_connect_import(req->rq_import, NULL);
2061 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2063 struct ptlrpc_request *req;
2064 struct ldlm_async_args *aa;
2065 struct ldlm_request *body;
2070 /* Bug 11974: Do not replay a lock which is actively being canceled */
2071 if (lock->l_flags & LDLM_FL_CANCELING) {
2072 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2076 /* If this is reply-less callback lock, we cannot replay it, since
2077 * server might have long dropped it, but notification of that event was
2078 * lost by network. (and server granted conflicting lock already) */
2079 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2080 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2081 ldlm_lock_cancel(lock);
2085 * If granted mode matches the requested mode, this lock is granted.
2087 * If they differ, but we have a granted mode, then we were granted
2088 * one mode and now want another: ergo, converting.
2090 * If we haven't been granted anything and are on a resource list,
2091 * then we're blocked/waiting.
2093 * If we haven't been granted anything and we're NOT on a resource list,
2094 * then we haven't got a reply yet and don't have a known disposition.
2095 * This happens whenever a lock enqueue is the request that triggers
2098 if (lock->l_granted_mode == lock->l_req_mode)
2099 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2100 else if (lock->l_granted_mode)
2101 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2102 else if (!cfs_list_empty(&lock->l_res_link))
2103 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2105 flags = LDLM_FL_REPLAY;
2107 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2108 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2112 /* We're part of recovery, so don't wait for it. */
2113 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2115 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2116 ldlm_lock2desc(lock, &body->lock_desc);
2117 body->lock_flags = flags;
2119 ldlm_lock2handle(lock, &body->lock_handle[0]);
2120 if (lock->l_lvb_len != 0) {
2121 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2122 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2125 ptlrpc_request_set_replen(req);
2126 /* notify the server we've replayed all requests.
2127 * also, we mark the request to be put on a dedicated
2128 * queue to be processed after all request replayes.
2130 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2132 LDLM_DEBUG(lock, "replaying lock:");
2134 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2135 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2136 aa = ptlrpc_req_async_args(req);
2137 aa->lock_handle = body->lock_handle[0];
2138 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2139 ptlrpcd_add_req(req, PSCOPE_OTHER);
2145 * Cancel as many unused locks as possible before replay. since we are
2146 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2149 * Called only in recovery before replaying locks. there is no need to
2150 * replay locks that are unused. since the clients may hold thousands of
2151 * cached unused locks, dropping the unused locks can greatly reduce the
2152 * load on the servers at recovery time.
2154 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2157 CFS_LIST_HEAD(cancels);
2159 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2160 "replay for namespace %s (%d)\n",
2161 ldlm_ns_name(ns), ns->ns_nr_unused);
2163 /* We don't need to care whether or not LRU resize is enabled
2164 * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2165 * count parameter */
2166 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2167 LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2169 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2170 canceled, ldlm_ns_name(ns));
2173 int ldlm_replay_locks(struct obd_import *imp)
2175 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2176 CFS_LIST_HEAD(list);
2177 struct ldlm_lock *lock, *next;
2182 LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0);
2184 /* don't replay locks if import failed recovery */
2185 if (imp->imp_vbr_failed)
2188 /* ensure this doesn't fall to 0 before all have been queued */
2189 cfs_atomic_inc(&imp->imp_replay_inflight);
2191 if (ldlm_cancel_unused_locks_before_replay)
2192 ldlm_cancel_unused_locks_for_replay(ns);
2194 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2196 cfs_list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2197 cfs_list_del_init(&lock->l_pending_chain);
2199 LDLM_LOCK_RELEASE(lock);
2200 continue; /* or try to do the rest? */
2202 rc = replay_one_lock(imp, lock);
2203 LDLM_LOCK_RELEASE(lock);
2206 cfs_atomic_dec(&imp->imp_replay_inflight);