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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2011, Whamcloud, Inc.
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_reprocess_all(lock->l_resource);
180 * Client side LDLM "completion" AST. This is called in several cases:
182 * - when a reply to an ENQUEUE rpc is received from the server
183 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
184 * this point (determined by flags);
186 * - when LDLM_CP_CALLBACK rpc comes to client to notify it that lock has
189 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
192 * - to force all locks when resource is destroyed (cleanup_resource());
194 * - during lock conversion (not used currently).
196 * If lock is not granted in the first case, this function waits until second
197 * or penultimate cases happen in some other thread.
200 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
202 /* XXX ALLOCATE - 160 bytes */
203 struct lock_wait_data lwd;
204 struct obd_device *obd;
205 struct obd_import *imp = NULL;
206 struct l_wait_info lwi;
211 if (flags == LDLM_FL_WAIT_NOREPROC) {
212 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
216 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
217 LDLM_FL_BLOCK_CONV))) {
218 cfs_waitq_signal(&lock->l_waitq);
222 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
227 obd = class_exp2obd(lock->l_conn_export);
229 /* if this is a local lock, then there is no import */
231 imp = obd->u.cli.cl_import;
234 /* Wait a long time for enqueue - server may have to callback a
235 lock from another client. Server will evict the other client if it
236 doesn't respond reasonably, and then give us the lock. */
237 timeout = ldlm_get_enq_timeout(lock) * 2;
241 if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
242 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
243 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
245 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
246 ldlm_expired_completion_wait,
247 interrupted_completion_wait, &lwd);
251 cfs_spin_lock(&imp->imp_lock);
252 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
253 cfs_spin_unlock(&imp->imp_lock);
256 if (ns_is_client(ldlm_lock_to_ns(lock)) &&
257 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
258 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
259 lock->l_flags |= LDLM_FL_FAIL_LOC;
262 /* Go to sleep until the lock is granted or cancelled. */
263 rc = l_wait_event(lock->l_waitq,
264 is_granted_or_cancelled(lock), &lwi);
268 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
273 RETURN(ldlm_completion_tail(lock));
277 * A helper to build a blocking ast function
279 * Perform a common operation for blocking asts:
280 * defferred lock cancellation.
282 * \param lock the lock blocking or canceling ast was called on
284 * \see mdt_blocking_ast
285 * \see ldlm_blocking_ast
287 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
292 lock->l_flags |= LDLM_FL_CBPENDING;
293 do_ast = (!lock->l_readers && !lock->l_writers);
294 unlock_res_and_lock(lock);
297 struct lustre_handle lockh;
300 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
301 ldlm_lock2handle(lock, &lockh);
302 rc = ldlm_cli_cancel(&lockh);
304 CERROR("ldlm_cli_cancel: %d\n", rc);
306 LDLM_DEBUG(lock, "Lock still has references, will be "
313 * Server blocking AST
315 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
318 * \param lock the lock which blocks a request or cancelling lock
321 * \param flag indicates whether this cancelling or blocking callback
323 * \see ldlm_blocking_ast_nocheck
325 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
326 void *data, int flag)
330 if (flag == LDLM_CB_CANCELING) {
331 /* Don't need to do anything here. */
335 lock_res_and_lock(lock);
336 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
337 * that ldlm_blocking_ast is called just before intent_policy method
338 * takes the lr_lock, then by the time we get the lock, we might not
339 * be the correct blocking function anymore. So check, and return
341 if (lock->l_blocking_ast != ldlm_blocking_ast) {
342 unlock_res_and_lock(lock);
345 RETURN(ldlm_blocking_ast_nocheck(lock));
349 * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
350 * comment in filter_intent_policy() on why you may need this.
352 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
355 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
356 * that is rather subtle: with OST-side locking, it may so happen that
357 * _all_ extent locks are held by the OST. If client wants to obtain
358 * current file size it calls ll{,u}_glimpse_size(), and (as locks are
359 * on the server), dummy glimpse callback fires and does
360 * nothing. Client still receives correct file size due to the
361 * following fragment in filter_intent_policy():
363 * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
364 * if (rc != 0 && res->lr_namespace->ns_lvbo &&
365 * res->lr_namespace->ns_lvbo->lvbo_update) {
366 * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
369 * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
370 * returns correct file size to the client.
372 return -ELDLM_NO_LOCK_DATA;
375 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
376 const struct ldlm_res_id *res_id,
377 ldlm_type_t type, ldlm_policy_data_t *policy,
378 ldlm_mode_t mode, int *flags,
379 ldlm_blocking_callback blocking,
380 ldlm_completion_callback completion,
381 ldlm_glimpse_callback glimpse,
382 void *data, __u32 lvb_len,
383 const __u64 *client_cookie,
384 struct lustre_handle *lockh)
386 struct ldlm_lock *lock;
388 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
389 .lcs_blocking = blocking,
390 .lcs_glimpse = glimpse,
394 LASSERT(!(*flags & LDLM_FL_REPLAY));
395 if (unlikely(ns_is_client(ns))) {
396 CERROR("Trying to enqueue local lock in a shadow namespace\n");
400 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len);
402 GOTO(out_nolock, err = -ENOMEM);
404 ldlm_lock2handle(lock, lockh);
406 /* NB: we don't have any lock now (lock_res_and_lock)
407 * because it's a new lock */
408 ldlm_lock_addref_internal_nolock(lock, mode);
409 lock->l_flags |= LDLM_FL_LOCAL;
410 if (*flags & LDLM_FL_ATOMIC_CB)
411 lock->l_flags |= LDLM_FL_ATOMIC_CB;
414 lock->l_policy_data = *policy;
415 if (client_cookie != NULL)
416 lock->l_client_cookie = *client_cookie;
417 if (type == LDLM_EXTENT)
418 lock->l_req_extent = policy->l_extent;
420 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
421 if (unlikely(err != ELDLM_OK))
425 *policy = lock->l_policy_data;
427 if (lock->l_completion_ast)
428 lock->l_completion_ast(lock, *flags, NULL);
430 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
433 LDLM_LOCK_RELEASE(lock);
438 static void failed_lock_cleanup(struct ldlm_namespace *ns,
439 struct ldlm_lock *lock, int mode)
443 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
444 lock_res_and_lock(lock);
445 /* Check that lock is not granted or failed, we might race. */
446 if ((lock->l_req_mode != lock->l_granted_mode) &&
447 !(lock->l_flags & LDLM_FL_FAILED)) {
448 /* Make sure that this lock will not be found by raced
449 * bl_ast and -EINVAL reply is sent to server anyways.
451 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
452 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
455 unlock_res_and_lock(lock);
459 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
460 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
462 LDLM_DEBUG(lock, "lock was granted or failed in race");
464 ldlm_lock_decref_internal(lock, mode);
466 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
467 * from llite/file.c/ll_file_flock(). */
468 /* This code makes for the fact that we do not have blocking handler on
469 * a client for flock locks. As such this is the place where we must
470 * completely kill failed locks. (interrupted and those that
471 * were waiting to be granted when server evicted us. */
472 if (lock->l_resource->lr_type == LDLM_FLOCK) {
473 lock_res_and_lock(lock);
474 ldlm_resource_unlink_lock(lock);
475 ldlm_lock_destroy_nolock(lock);
476 unlock_res_and_lock(lock);
480 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
481 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
482 int *flags, void *lvb, __u32 lvb_len,
483 struct lustre_handle *lockh,int rc)
485 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
486 int is_replay = *flags & LDLM_FL_REPLAY;
487 struct ldlm_lock *lock;
488 struct ldlm_reply *reply;
489 struct ost_lvb *tmplvb;
490 int cleanup_phase = 1;
493 lock = ldlm_handle2lock(lockh);
494 /* ldlm_cli_enqueue is holding a reference on this lock. */
496 LASSERT(type == LDLM_FLOCK);
500 if (rc != ELDLM_OK) {
502 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
503 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
504 if (rc == ELDLM_LOCK_ABORTED) {
505 /* Before we return, swab the reply */
506 reply = req_capsule_server_get(&req->rq_pill,
512 req_capsule_set_size(&req->rq_pill,
513 &RMF_DLM_LVB, RCL_SERVER,
515 tmplvb = req_capsule_server_get(&req->rq_pill,
518 GOTO(cleanup, rc = -EPROTO);
520 memcpy(lvb, tmplvb, lvb_len);
526 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
528 GOTO(cleanup, rc = -EPROTO);
530 /* lock enqueued on the server */
533 lock_res_and_lock(lock);
534 /* Key change rehash lock in per-export hash with new key */
535 if (exp->exp_lock_hash) {
536 cfs_hash_rehash_key(exp->exp_lock_hash,
537 &lock->l_remote_handle,
541 lock->l_remote_handle = reply->lock_handle;
544 *flags = reply->lock_flags;
545 lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
546 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
547 * to wait with no timeout as well */
548 lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
549 unlock_res_and_lock(lock);
551 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
552 lock, reply->lock_handle.cookie, *flags);
554 /* If enqueue returned a blocked lock but the completion handler has
555 * already run, then it fixed up the resource and we don't need to do it
557 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
558 int newmode = reply->lock_desc.l_req_mode;
560 if (newmode && newmode != lock->l_req_mode) {
561 LDLM_DEBUG(lock, "server returned different mode %s",
562 ldlm_lockname[newmode]);
563 lock->l_req_mode = newmode;
566 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
567 lock->l_resource->lr_name.name,
568 sizeof(struct ldlm_res_id))) {
569 CDEBUG(D_INFO, "remote intent success, locking "
570 "(%ld,%ld,%ld) instead of "
572 (long)reply->lock_desc.l_resource.lr_name.name[0],
573 (long)reply->lock_desc.l_resource.lr_name.name[1],
574 (long)reply->lock_desc.l_resource.lr_name.name[2],
575 (long)lock->l_resource->lr_name.name[0],
576 (long)lock->l_resource->lr_name.name[1],
577 (long)lock->l_resource->lr_name.name[2]);
579 rc = ldlm_lock_change_resource(ns, lock,
580 &reply->lock_desc.l_resource.lr_name);
581 if (rc || lock->l_resource == NULL)
582 GOTO(cleanup, rc = -ENOMEM);
583 LDLM_DEBUG(lock, "client-side enqueue, new resource");
586 if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
588 /* We assume lock type cannot change on server*/
589 ldlm_convert_policy_to_local(exp,
590 lock->l_resource->lr_type,
591 &reply->lock_desc.l_policy_data,
592 &lock->l_policy_data);
593 if (type != LDLM_PLAIN)
594 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
597 if ((*flags) & LDLM_FL_AST_SENT ||
598 /* Cancel extent locks as soon as possible on a liblustre client,
599 * because it cannot handle asynchronous ASTs robustly (see
601 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
602 lock_res_and_lock(lock);
603 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
604 unlock_res_and_lock(lock);
605 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
608 /* If the lock has already been granted by a completion AST, don't
609 * clobber the LVB with an older one. */
611 /* We must lock or a racing completion might update lvb
612 without letting us know and we'll clobber the correct value.
613 Cannot unlock after the check either, a that still leaves
614 a tiny window for completion to get in */
615 lock_res_and_lock(lock);
616 if (lock->l_req_mode != lock->l_granted_mode) {
618 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
619 RCL_SERVER, lvb_len);
620 tmplvb = req_capsule_server_get(&req->rq_pill,
622 if (tmplvb == NULL) {
623 unlock_res_and_lock(lock);
624 GOTO(cleanup, rc = -EPROTO);
626 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
628 unlock_res_and_lock(lock);
632 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
633 if (lock->l_completion_ast != NULL) {
634 int err = lock->l_completion_ast(lock, *flags, NULL);
642 if (lvb_len && lvb != NULL) {
643 /* Copy the LVB here, and not earlier, because the completion
644 * AST (if any) can override what we got in the reply */
645 memcpy(lvb, lock->l_lvb_data, lvb_len);
648 LDLM_DEBUG(lock, "client-side enqueue END");
651 if (cleanup_phase == 1 && rc)
652 failed_lock_cleanup(ns, lock, mode);
653 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
655 LDLM_LOCK_RELEASE(lock);
659 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
660 * a single page on the send/receive side. XXX: 512 should be changed
661 * to more adequate value. */
662 static inline int ldlm_req_handles_avail(int req_size, int off)
666 avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size;
667 if (likely(avail >= 0))
668 avail /= (int)sizeof(struct lustre_handle);
671 avail += LDLM_LOCKREQ_HANDLES - off;
676 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
677 enum req_location loc,
680 int size = req_capsule_msg_size(pill, loc);
681 return ldlm_req_handles_avail(size, off);
684 static inline int ldlm_format_handles_avail(struct obd_import *imp,
685 const struct req_format *fmt,
686 enum req_location loc, int off)
688 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
689 return ldlm_req_handles_avail(size, off);
692 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
693 * @count locks in @cancels. */
694 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
695 int version, int opc, int canceloff,
696 cfs_list_t *cancels, int count)
698 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
699 struct req_capsule *pill = &req->rq_pill;
700 struct ldlm_request *dlm = NULL;
701 int flags, avail, to_free, pack = 0;
708 if (exp_connect_cancelset(exp)) {
709 /* Estimate the amount of available space in the request. */
710 req_capsule_filled_sizes(pill, RCL_CLIENT);
711 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
713 flags = ns_connect_lru_resize(ns) ?
714 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
715 to_free = !ns_connect_lru_resize(ns) &&
716 opc == LDLM_ENQUEUE ? 1 : 0;
718 /* Cancel lru locks here _only_ if the server supports
719 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
720 * rpc, what will make us slower. */
722 count += ldlm_cancel_lru_local(ns, cancels, to_free,
723 avail - count, 0, flags);
728 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
729 ldlm_request_bufsize(pack, opc));
732 rc = ptlrpc_request_pack(req, version, opc);
734 ldlm_lock_list_put(cancels, l_bl_ast, count);
738 if (exp_connect_cancelset(exp)) {
740 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
742 /* Skip first lock handler in ldlm_request_pack(),
743 * this method will incrment @lock_count according
744 * to the lock handle amount actually written to
746 dlm->lock_count = canceloff;
748 /* Pack into the request @pack lock handles. */
749 ldlm_cli_cancel_list(cancels, pack, req, 0);
750 /* Prepare and send separate cancel rpc for others. */
751 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
753 ldlm_lock_list_put(cancels, l_bl_ast, count);
758 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
759 cfs_list_t *cancels, int count)
761 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
762 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
765 /* If a request has some specific initialisation it is passed in @reqp,
766 * otherwise it is created in ldlm_cli_enqueue.
768 * Supports sync and async requests, pass @async flag accordingly. If a
769 * request was created in ldlm_cli_enqueue and it is the async request,
770 * pass it to the caller in @reqp. */
771 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
772 struct ldlm_enqueue_info *einfo,
773 const struct ldlm_res_id *res_id,
774 ldlm_policy_data_t const *policy, int *flags,
775 void *lvb, __u32 lvb_len, struct lustre_handle *lockh,
778 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
779 struct ldlm_lock *lock;
780 struct ldlm_request *body;
781 int is_replay = *flags & LDLM_FL_REPLAY;
782 int req_passed_in = 1;
784 struct ptlrpc_request *req;
787 LASSERT(exp != NULL);
789 /* If we're replaying this lock, just check some invariants.
790 * If we're creating a new lock, get everything all setup nice. */
792 lock = ldlm_handle2lock_long(lockh, 0);
793 LASSERT(lock != NULL);
794 LDLM_DEBUG(lock, "client-side enqueue START");
795 LASSERT(exp == lock->l_conn_export);
797 const struct ldlm_callback_suite cbs = {
798 .lcs_completion = einfo->ei_cb_cp,
799 .lcs_blocking = einfo->ei_cb_bl,
800 .lcs_glimpse = einfo->ei_cb_gl,
801 .lcs_weigh = einfo->ei_cb_wg
803 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
804 einfo->ei_mode, &cbs, einfo->ei_cbdata,
808 /* for the local lock, add the reference */
809 ldlm_lock_addref_internal(lock, einfo->ei_mode);
810 ldlm_lock2handle(lock, lockh);
811 if (policy != NULL) {
812 /* INODEBITS_INTEROP: If the server does not support
813 * inodebits, we will request a plain lock in the
814 * descriptor (ldlm_lock2desc() below) but use an
815 * inodebits lock internally with both bits set.
817 if (einfo->ei_type == LDLM_IBITS &&
818 !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
819 lock->l_policy_data.l_inodebits.bits =
820 MDS_INODELOCK_LOOKUP |
821 MDS_INODELOCK_UPDATE;
823 lock->l_policy_data = *policy;
826 if (einfo->ei_type == LDLM_EXTENT)
827 lock->l_req_extent = policy->l_extent;
828 LDLM_DEBUG(lock, "client-side enqueue START");
831 lock->l_conn_export = exp;
832 lock->l_export = NULL;
833 lock->l_blocking_ast = einfo->ei_cb_bl;
834 lock->l_flags |= (*flags & LDLM_FL_NO_LRU);
836 /* lock not sent to server yet */
838 if (reqp == NULL || *reqp == NULL) {
839 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
844 failed_lock_cleanup(ns, lock, einfo->ei_mode);
845 LDLM_LOCK_RELEASE(lock);
855 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
857 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
858 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
861 /* Dump lock data into the request buffer */
862 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
863 ldlm_lock2desc(lock, &body->lock_desc);
864 body->lock_flags = *flags;
865 body->lock_handle[0] = *lockh;
867 /* Continue as normal. */
868 if (!req_passed_in) {
870 req_capsule_extend(&req->rq_pill,
871 &RQF_LDLM_ENQUEUE_LVB);
872 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
873 RCL_SERVER, lvb_len);
875 ptlrpc_request_set_replen(req);
879 * Liblustre client doesn't get extent locks, except for O_APPEND case
880 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
881 * [i_size, OBD_OBJECT_EOF] lock is taken.
883 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
884 policy->l_extent.end == OBD_OBJECT_EOF));
887 LASSERT(reqp != NULL);
891 LDLM_DEBUG(lock, "sending request");
893 rc = ptlrpc_queue_wait(req);
895 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
896 einfo->ei_mode, flags, lvb, lvb_len,
899 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
900 * one reference that we took */
902 LDLM_LOCK_RELEASE(lock);
906 if (!req_passed_in && req != NULL) {
907 ptlrpc_req_finished(req);
915 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
918 struct ldlm_resource *res;
921 if (ns_is_client(ldlm_lock_to_ns(lock))) {
922 CERROR("Trying to cancel local lock\n");
925 LDLM_DEBUG(lock, "client-side local convert");
927 res = ldlm_lock_convert(lock, new_mode, flags);
929 ldlm_reprocess_all(res);
934 LDLM_DEBUG(lock, "client-side local convert handler END");
939 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
940 * conversion of locks which are on the waiting or converting queue */
941 /* Caller of this code is supposed to take care of lock readers/writers
943 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
945 struct ldlm_request *body;
946 struct ldlm_reply *reply;
947 struct ldlm_lock *lock;
948 struct ldlm_resource *res;
949 struct ptlrpc_request *req;
953 lock = ldlm_handle2lock(lockh);
960 if (lock->l_conn_export == NULL)
961 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
963 LDLM_DEBUG(lock, "client-side convert");
965 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
966 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
973 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
974 body->lock_handle[0] = lock->l_remote_handle;
976 body->lock_desc.l_req_mode = new_mode;
977 body->lock_flags = *flags;
980 ptlrpc_request_set_replen(req);
981 rc = ptlrpc_queue_wait(req);
985 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
987 GOTO(out, rc = -EPROTO);
990 GOTO(out, rc = req->rq_status);
992 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
994 ldlm_reprocess_all(res);
995 /* Go to sleep until the lock is granted. */
996 /* FIXME: or cancelled. */
997 if (lock->l_completion_ast) {
998 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1008 LDLM_LOCK_PUT(lock);
1009 ptlrpc_req_finished(req);
1013 /* Cancel locks locally.
1015 * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
1016 * LDLM_FL_CANCELING otherwise;
1017 * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
1018 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
1020 int rc = LDLM_FL_LOCAL_ONLY;
1023 if (lock->l_conn_export) {
1026 LDLM_DEBUG(lock, "client-side cancel");
1027 /* Set this flag to prevent others from getting new references*/
1028 lock_res_and_lock(lock);
1029 lock->l_flags |= LDLM_FL_CBPENDING;
1030 local_only = (lock->l_flags &
1031 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1032 ldlm_cancel_callback(lock);
1033 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
1034 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1035 unlock_res_and_lock(lock);
1038 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1040 rc = LDLM_FL_LOCAL_ONLY;
1042 ldlm_lock_cancel(lock);
1044 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1045 LDLM_ERROR(lock, "Trying to cancel local lock");
1048 LDLM_DEBUG(lock, "server-side local cancel");
1049 ldlm_lock_cancel(lock);
1050 ldlm_reprocess_all(lock->l_resource);
1056 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
1057 of the request @req. */
1058 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1059 cfs_list_t *head, int count)
1061 struct ldlm_request *dlm;
1062 struct ldlm_lock *lock;
1063 int max, packed = 0;
1066 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1067 LASSERT(dlm != NULL);
1069 /* Check the room in the request buffer. */
1070 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1071 sizeof(struct ldlm_request);
1072 max /= sizeof(struct lustre_handle);
1073 max += LDLM_LOCKREQ_HANDLES;
1074 LASSERT(max >= dlm->lock_count + count);
1076 /* XXX: it would be better to pack lock handles grouped by resource.
1077 * so that the server cancel would call filter_lvbo_update() less
1079 cfs_list_for_each_entry(lock, head, l_bl_ast) {
1082 LASSERT(lock->l_conn_export);
1083 /* Pack the lock handle to the given request buffer. */
1084 LDLM_DEBUG(lock, "packing");
1085 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1088 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1092 /* Prepare and send a batched cancel rpc, it will include count lock handles
1093 * of locks given in @head. */
1094 int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels,
1095 int count, ldlm_cancel_flags_t flags)
1097 struct ptlrpc_request *req = NULL;
1098 struct obd_import *imp;
1103 LASSERT(exp != NULL);
1106 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1108 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1111 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1112 &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 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, PDL_POLICY_LOCAL, -1);
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;
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 obd->obd_pool_slv = new_slv;
1232 obd->obd_pool_limit = new_limit;
1233 cfs_write_unlock(&obd->obd_pool_lock);
1237 EXPORT_SYMBOL(ldlm_cli_update_pool);
1239 int ldlm_cli_cancel(struct lustre_handle *lockh)
1241 struct obd_export *exp;
1242 int avail, flags, count = 1, rc = 0;
1243 struct ldlm_namespace *ns;
1244 struct ldlm_lock *lock;
1245 CFS_LIST_HEAD(cancels);
1248 /* concurrent cancels on the same handle can happen */
1249 lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1251 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1255 rc = ldlm_cli_cancel_local(lock);
1256 if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) {
1257 LDLM_LOCK_RELEASE(lock);
1258 RETURN(rc < 0 ? rc : 0);
1260 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1261 * rpc which goes to canceld portal, so we can cancel other lru locks
1262 * here and send them all as one LDLM_CANCEL rpc. */
1263 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1264 cfs_list_add(&lock->l_bl_ast, &cancels);
1266 exp = lock->l_conn_export;
1267 if (exp_connect_cancelset(exp)) {
1268 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1273 ns = ldlm_lock_to_ns(lock);
1274 flags = ns_connect_lru_resize(ns) ?
1275 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1276 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1279 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1283 /* XXX until we will have compound requests and can cut cancels from generic rpc
1284 * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */
1285 int ldlm_cli_cancel_list_local(cfs_list_t *cancels, int count,
1286 ldlm_cancel_flags_t flags)
1288 CFS_LIST_HEAD(head);
1289 struct ldlm_lock *lock, *next;
1290 int left = 0, bl_ast = 0, rc;
1293 cfs_list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1297 if (flags & LCF_LOCAL) {
1298 rc = LDLM_FL_LOCAL_ONLY;
1299 ldlm_lock_cancel(lock);
1301 rc = ldlm_cli_cancel_local(lock);
1303 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1304 LDLM_DEBUG(lock, "Cancel lock separately");
1305 cfs_list_del_init(&lock->l_bl_ast);
1306 cfs_list_add(&lock->l_bl_ast, &head);
1310 if (rc == LDLM_FL_LOCAL_ONLY) {
1311 /* CANCEL RPC should not be sent to server. */
1312 cfs_list_del_init(&lock->l_bl_ast);
1313 LDLM_LOCK_RELEASE(lock);
1320 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1327 * Cancel as many locks as possible w/o sending any rpcs (e.g. to write back
1328 * dirty data, to close a file, ...) or waiting for any rpcs in-flight (e.g.
1329 * readahead requests, ...)
1331 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1332 struct ldlm_lock *lock,
1333 int unused, int added,
1336 ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1337 ldlm_cancel_for_recovery cb = ns->ns_cancel_for_recovery;
1338 lock_res_and_lock(lock);
1340 /* don't check added & count since we want to process all locks
1341 * from unused list */
1342 switch (lock->l_resource->lr_type) {
1348 result = LDLM_POLICY_SKIP_LOCK;
1349 lock->l_flags |= LDLM_FL_SKIPPED;
1353 unlock_res_and_lock(lock);
1358 * Callback function for lru-resize policy. Makes decision whether to keep
1359 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1360 * \a added and number of locks to be preferably canceled \a count.
1362 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1364 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1366 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1367 struct ldlm_lock *lock,
1368 int unused, int added,
1371 cfs_time_t cur = cfs_time_current();
1372 struct ldlm_pool *pl = &ns->ns_pool;
1377 * Stop lru processing when we reached passed @count or checked all
1380 if (count && added >= count)
1381 return LDLM_POLICY_KEEP_LOCK;
1383 slv = ldlm_pool_get_slv(pl);
1384 lvf = ldlm_pool_get_lvf(pl);
1385 la = cfs_duration_sec(cfs_time_sub(cur,
1386 lock->l_last_used));
1389 * Stop when slv is not yet come from server or lv is smaller than
1392 lv = lvf * la * unused;
1395 * Inform pool about current CLV to see it via proc.
1397 ldlm_pool_set_clv(pl, lv);
1398 return (slv == 0 || lv < slv) ?
1399 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1403 * Callback function for proc used policy. Makes decision whether to keep
1404 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1405 * \a added and number of locks to be preferably canceled \a count.
1407 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1409 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1411 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1412 struct ldlm_lock *lock,
1413 int unused, int added,
1417 * Stop lru processing when we reached passed @count or checked all
1420 return (added >= count) ?
1421 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1425 * Callback function for aged policy. Makes decision whether to keep
1426 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1427 * \a added and number of locks to be preferably canceled \a count.
1429 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1431 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1433 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1434 struct ldlm_lock *lock,
1435 int unused, int added,
1439 * Stop lru processing if young lock is found and we reached passed
1442 return ((added >= count) &&
1443 cfs_time_before(cfs_time_current(),
1444 cfs_time_add(lock->l_last_used,
1446 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1450 * Callback function for default policy. Makes decision whether to keep
1451 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1452 * \a added and number of locks to be preferably canceled \a count.
1454 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1456 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1458 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1459 struct ldlm_lock *lock,
1460 int unused, int added,
1464 * Stop lru processing when we reached passed @count or checked all
1467 return (added >= count) ?
1468 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1471 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1472 struct ldlm_lock *, int,
1475 static ldlm_cancel_lru_policy_t
1476 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1478 if (flags & LDLM_CANCEL_NO_WAIT)
1479 return ldlm_cancel_no_wait_policy;
1481 if (ns_connect_lru_resize(ns)) {
1482 if (flags & LDLM_CANCEL_SHRINK)
1483 /* We kill passed number of old locks. */
1484 return ldlm_cancel_passed_policy;
1485 else if (flags & LDLM_CANCEL_LRUR)
1486 return ldlm_cancel_lrur_policy;
1487 else if (flags & LDLM_CANCEL_PASSED)
1488 return ldlm_cancel_passed_policy;
1490 if (flags & LDLM_CANCEL_AGED)
1491 return ldlm_cancel_aged_policy;
1494 return ldlm_cancel_default_policy;
1497 /* - Free space in lru for @count new locks,
1498 * redundant unused locks are canceled locally;
1499 * - also cancel locally unused aged locks;
1500 * - do not cancel more than @max locks;
1501 * - GET the found locks and add them into the @cancels list.
1503 * A client lock can be added to the l_bl_ast list only when it is
1504 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
1505 * There are the following use cases: ldlm_cancel_resource_local(),
1506 * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
1507 * flag properly. As any attempt to cancel a lock rely on this flag,
1508 * l_bl_ast list is accessed later without any special locking.
1510 * Calling policies for enabled lru resize:
1511 * ----------------------------------------
1512 * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to
1513 * cancel not more than @count locks;
1515 * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at
1516 * the beginning of lru list);
1518 * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
1519 * memory pressre policy function;
1521 * flags & LDLM_CANCEL_AGED - cancel alocks according to "aged policy".
1523 * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1524 * (typically before replaying locks) w/o
1525 * sending any rpcs or waiting for any
1526 * outstanding rpc to complete.
1528 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels,
1529 int count, int max, int flags)
1531 ldlm_cancel_lru_policy_t pf;
1532 struct ldlm_lock *lock, *next;
1533 int added = 0, unused, remained;
1536 cfs_spin_lock(&ns->ns_lock);
1537 unused = ns->ns_nr_unused;
1540 if (!ns_connect_lru_resize(ns))
1541 count += unused - ns->ns_max_unused;
1543 pf = ldlm_cancel_lru_policy(ns, flags);
1544 LASSERT(pf != NULL);
1546 while (!cfs_list_empty(&ns->ns_unused_list)) {
1547 ldlm_policy_res_t result;
1549 /* all unused locks */
1550 if (remained-- <= 0)
1553 /* For any flags, stop scanning if @max is reached. */
1554 if (max && added >= max)
1557 cfs_list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1559 /* No locks which got blocking requests. */
1560 LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1562 if (flags & LDLM_CANCEL_NO_WAIT &&
1563 lock->l_flags & LDLM_FL_SKIPPED)
1564 /* already processed */
1567 /* Somebody is already doing CANCEL. No need in this
1568 * lock in lru, do not traverse it again. */
1569 if (!(lock->l_flags & LDLM_FL_CANCELING))
1572 ldlm_lock_remove_from_lru_nolock(lock);
1574 if (&lock->l_lru == &ns->ns_unused_list)
1577 LDLM_LOCK_GET(lock);
1578 cfs_spin_unlock(&ns->ns_lock);
1579 lu_ref_add(&lock->l_reference, __FUNCTION__, cfs_current());
1581 /* Pass the lock through the policy filter and see if it
1582 * should stay in lru.
1584 * Even for shrinker policy we stop scanning if
1585 * we find a lock that should stay in the cache.
1586 * We should take into account lock age anyway
1587 * as new lock even if it is small of weight is
1588 * valuable resource.
1590 * That is, for shrinker policy we drop only
1591 * old locks, but additionally chose them by
1592 * their weight. Big extent locks will stay in
1594 result = pf(ns, lock, unused, added, count);
1595 if (result == LDLM_POLICY_KEEP_LOCK) {
1596 lu_ref_del(&lock->l_reference,
1597 __FUNCTION__, cfs_current());
1598 LDLM_LOCK_RELEASE(lock);
1599 cfs_spin_lock(&ns->ns_lock);
1602 if (result == LDLM_POLICY_SKIP_LOCK) {
1603 lu_ref_del(&lock->l_reference,
1604 __FUNCTION__, cfs_current());
1605 LDLM_LOCK_RELEASE(lock);
1606 cfs_spin_lock(&ns->ns_lock);
1610 lock_res_and_lock(lock);
1611 /* Check flags again under the lock. */
1612 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1613 (ldlm_lock_remove_from_lru(lock) == 0)) {
1614 /* other thread is removing lock from lru or
1615 * somebody is already doing CANCEL or
1616 * there is a blocking request which will send
1617 * cancel by itseft or the lock is matched
1618 * is already not unused. */
1619 unlock_res_and_lock(lock);
1620 lu_ref_del(&lock->l_reference,
1621 __FUNCTION__, cfs_current());
1622 LDLM_LOCK_RELEASE(lock);
1623 cfs_spin_lock(&ns->ns_lock);
1626 LASSERT(!lock->l_readers && !lock->l_writers);
1628 /* If we have chosen to cancel this lock voluntarily, we
1629 * better send cancel notification to server, so that it
1630 * frees appropriate state. This might lead to a race
1631 * where while we are doing cancel here, server is also
1632 * silently cancelling this lock. */
1633 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1635 /* Setting the CBPENDING flag is a little misleading,
1636 * but prevents an important race; namely, once
1637 * CBPENDING is set, the lock can accumulate no more
1638 * readers/writers. Since readers and writers are
1639 * already zero here, ldlm_lock_decref() won't see
1640 * this flag and call l_blocking_ast */
1641 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1643 /* We can't re-add to l_lru as it confuses the
1644 * refcounting in ldlm_lock_remove_from_lru() if an AST
1645 * arrives after we drop lr_lock below. We use l_bl_ast
1646 * and can't use l_pending_chain as it is used both on
1647 * server and client nevertheless bug 5666 says it is
1648 * used only on server */
1649 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1650 cfs_list_add(&lock->l_bl_ast, cancels);
1651 unlock_res_and_lock(lock);
1652 lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current());
1653 cfs_spin_lock(&ns->ns_lock);
1657 cfs_spin_unlock(&ns->ns_lock);
1661 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, cfs_list_t *cancels,
1662 int count, int max, ldlm_cancel_flags_t cancel_flags,
1666 added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1669 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1672 /* when called with LDLM_ASYNC the blocking callback will be handled
1673 * in a thread and this function will return after the thread has been
1674 * asked to call the callback. when called with LDLM_SYNC the blocking
1675 * callback will be performed in this function. */
1676 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t mode,
1679 CFS_LIST_HEAD(cancels);
1684 mode = LDLM_SYNC; /* force to be sync in user space */
1686 /* Just prepare the list of locks, do not actually cancel them yet.
1687 * Locks are cancelled later in a separate thread. */
1688 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1689 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, mode);
1696 /* Find and cancel locally unused locks found on resource, matched to the
1697 * given policy, mode. GET the found locks and add them into the @cancels
1699 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1700 cfs_list_t *cancels,
1701 ldlm_policy_data_t *policy,
1702 ldlm_mode_t mode, int lock_flags,
1703 ldlm_cancel_flags_t cancel_flags, void *opaque)
1705 struct ldlm_lock *lock;
1710 cfs_list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1711 if (opaque != NULL && lock->l_ast_data != opaque) {
1712 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1713 lock->l_ast_data, opaque);
1718 if (lock->l_readers || lock->l_writers)
1721 /* If somebody is already doing CANCEL, or blocking ast came,
1722 * skip this lock. */
1723 if (lock->l_flags & LDLM_FL_BL_AST ||
1724 lock->l_flags & LDLM_FL_CANCELING)
1727 if (lockmode_compat(lock->l_granted_mode, mode))
1730 /* If policy is given and this is IBITS lock, add to list only
1731 * those locks that match by policy. */
1732 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1733 !(lock->l_policy_data.l_inodebits.bits &
1734 policy->l_inodebits.bits))
1737 /* See CBPENDING comment in ldlm_cancel_lru */
1738 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1741 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1742 cfs_list_add(&lock->l_bl_ast, cancels);
1743 LDLM_LOCK_GET(lock);
1748 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
1751 /* If @req is NULL, send CANCEL request to server with handles of locks
1752 * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1753 * separately per lock.
1754 * If @req is not NULL, put handles of locks in @cancels into the request
1755 * buffer at the offset @off.
1756 * Destroy @cancels at the end. */
1757 int ldlm_cli_cancel_list(cfs_list_t *cancels, int count,
1758 struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1760 struct ldlm_lock *lock;
1764 if (cfs_list_empty(cancels) || count == 0)
1767 /* XXX: requests (both batched and not) could be sent in parallel.
1768 * Usually it is enough to have just 1 RPC, but it is possible that
1769 * there are to many locks to be cancelled in LRU or on a resource.
1770 * It would also speed up the case when the server does not support
1773 LASSERT(!cfs_list_empty(cancels));
1774 lock = cfs_list_entry(cancels->next, struct ldlm_lock,
1776 LASSERT(lock->l_conn_export);
1778 if (exp_connect_cancelset(lock->l_conn_export)) {
1781 ldlm_cancel_pack(req, cancels, count);
1783 res = ldlm_cli_cancel_req(lock->l_conn_export,
1787 res = ldlm_cli_cancel_req(lock->l_conn_export,
1792 CERROR("ldlm_cli_cancel_list: %d\n", res);
1797 ldlm_lock_list_put(cancels, l_bl_ast, res);
1799 LASSERT(count == 0);
1803 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1804 const struct ldlm_res_id *res_id,
1805 ldlm_policy_data_t *policy,
1807 ldlm_cancel_flags_t flags,
1810 struct ldlm_resource *res;
1811 CFS_LIST_HEAD(cancels);
1816 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1818 /* This is not a problem. */
1819 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1823 LDLM_RESOURCE_ADDREF(res);
1824 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1825 0, flags | LCF_BL_AST, opaque);
1826 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1828 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1830 LDLM_RESOURCE_DELREF(res);
1831 ldlm_resource_putref(res);
1835 struct ldlm_cli_cancel_arg {
1840 static int ldlm_cli_hash_cancel_unused(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1841 cfs_hlist_node_t *hnode, void *arg)
1843 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1844 struct ldlm_cli_cancel_arg *lc = arg;
1847 rc = ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1849 lc->lc_flags, lc->lc_opaque);
1851 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1852 res->lr_name.name[0], rc);
1854 /* must return 0 for hash iteration */
1858 /* Cancel all locks on a namespace (or a specific resource, if given)
1859 * that have 0 readers/writers.
1861 * If flags & LCF_LOCAL, throw the locks away without trying
1862 * to notify the server. */
1863 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1864 const struct ldlm_res_id *res_id,
1865 ldlm_cancel_flags_t flags, void *opaque)
1867 struct ldlm_cli_cancel_arg arg = {
1869 .lc_opaque = opaque,
1877 if (res_id != NULL) {
1878 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1882 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1883 ldlm_cli_hash_cancel_unused, &arg);
1888 /* Lock iterators. */
1890 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1893 cfs_list_t *tmp, *next;
1894 struct ldlm_lock *lock;
1895 int rc = LDLM_ITER_CONTINUE;
1900 RETURN(LDLM_ITER_CONTINUE);
1903 cfs_list_for_each_safe(tmp, next, &res->lr_granted) {
1904 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1906 if (iter(lock, closure) == LDLM_ITER_STOP)
1907 GOTO(out, rc = LDLM_ITER_STOP);
1910 cfs_list_for_each_safe(tmp, next, &res->lr_converting) {
1911 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1913 if (iter(lock, closure) == LDLM_ITER_STOP)
1914 GOTO(out, rc = LDLM_ITER_STOP);
1917 cfs_list_for_each_safe(tmp, next, &res->lr_waiting) {
1918 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1920 if (iter(lock, closure) == LDLM_ITER_STOP)
1921 GOTO(out, rc = LDLM_ITER_STOP);
1928 struct iter_helper_data {
1929 ldlm_iterator_t iter;
1933 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1935 struct iter_helper_data *helper = closure;
1936 return helper->iter(lock, helper->closure);
1939 static int ldlm_res_iter_helper(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1940 cfs_hlist_node_t *hnode, void *arg)
1943 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1945 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
1949 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
1950 ldlm_iterator_t iter, void *closure)
1953 struct iter_helper_data helper = { iter: iter, closure: closure };
1955 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1956 ldlm_res_iter_helper, &helper);
1960 /* non-blocking function to manipulate a lock whose cb_data is being put away.
1961 * return 0: find no resource
1962 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
1965 int ldlm_resource_iterate(struct ldlm_namespace *ns,
1966 const struct ldlm_res_id *res_id,
1967 ldlm_iterator_t iter, void *data)
1969 struct ldlm_resource *res;
1974 CERROR("must pass in namespace\n");
1978 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1982 LDLM_RESOURCE_ADDREF(res);
1983 rc = ldlm_resource_foreach(res, iter, data);
1984 LDLM_RESOURCE_DELREF(res);
1985 ldlm_resource_putref(res);
1991 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1993 cfs_list_t *list = closure;
1995 /* we use l_pending_chain here, because it's unused on clients. */
1996 LASSERTF(cfs_list_empty(&lock->l_pending_chain),
1997 "lock %p next %p prev %p\n",
1998 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1999 /* bug 9573: don't replay locks left after eviction, or
2000 * bug 17614: locks being actively cancelled. Get a reference
2001 * on a lock so that it does not disapear under us (e.g. due to cancel)
2003 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
2004 cfs_list_add(&lock->l_pending_chain, list);
2005 LDLM_LOCK_GET(lock);
2008 return LDLM_ITER_CONTINUE;
2011 static int replay_lock_interpret(const struct lu_env *env,
2012 struct ptlrpc_request *req,
2013 struct ldlm_async_args *aa, int rc)
2015 struct ldlm_lock *lock;
2016 struct ldlm_reply *reply;
2017 struct obd_export *exp;
2020 cfs_atomic_dec(&req->rq_import->imp_replay_inflight);
2025 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2027 GOTO(out, rc = -EPROTO);
2029 lock = ldlm_handle2lock(&aa->lock_handle);
2031 CERROR("received replay ack for unknown local cookie "LPX64
2032 " remote cookie "LPX64 " from server %s id %s\n",
2033 aa->lock_handle.cookie, reply->lock_handle.cookie,
2034 req->rq_export->exp_client_uuid.uuid,
2035 libcfs_id2str(req->rq_peer));
2036 GOTO(out, rc = -ESTALE);
2039 /* Key change rehash lock in per-export hash with new key */
2040 exp = req->rq_export;
2041 if (exp && exp->exp_lock_hash) {
2042 cfs_hash_rehash_key(exp->exp_lock_hash,
2043 &lock->l_remote_handle,
2044 &reply->lock_handle,
2047 lock->l_remote_handle = reply->lock_handle;
2050 LDLM_DEBUG(lock, "replayed lock:");
2051 ptlrpc_import_recovery_state_machine(req->rq_import);
2052 LDLM_LOCK_PUT(lock);
2055 ptlrpc_connect_import(req->rq_import);
2060 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2062 struct ptlrpc_request *req;
2063 struct ldlm_async_args *aa;
2064 struct ldlm_request *body;
2069 /* Bug 11974: Do not replay a lock which is actively being canceled */
2070 if (lock->l_flags & LDLM_FL_CANCELING) {
2071 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2075 /* If this is reply-less callback lock, we cannot replay it, since
2076 * server might have long dropped it, but notification of that event was
2077 * lost by network. (and server granted conflicting lock already) */
2078 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2079 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2080 ldlm_lock_cancel(lock);
2084 * If granted mode matches the requested mode, this lock is granted.
2086 * If they differ, but we have a granted mode, then we were granted
2087 * one mode and now want another: ergo, converting.
2089 * If we haven't been granted anything and are on a resource list,
2090 * then we're blocked/waiting.
2092 * If we haven't been granted anything and we're NOT on a resource list,
2093 * then we haven't got a reply yet and don't have a known disposition.
2094 * This happens whenever a lock enqueue is the request that triggers
2097 if (lock->l_granted_mode == lock->l_req_mode)
2098 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2099 else if (lock->l_granted_mode)
2100 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2101 else if (!cfs_list_empty(&lock->l_res_link))
2102 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2104 flags = LDLM_FL_REPLAY;
2106 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2107 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2111 /* We're part of recovery, so don't wait for it. */
2112 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2114 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2115 ldlm_lock2desc(lock, &body->lock_desc);
2116 body->lock_flags = flags;
2118 ldlm_lock2handle(lock, &body->lock_handle[0]);
2119 if (lock->l_lvb_len != 0) {
2120 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2121 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2124 ptlrpc_request_set_replen(req);
2125 /* notify the server we've replayed all requests.
2126 * also, we mark the request to be put on a dedicated
2127 * queue to be processed after all request replayes.
2129 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2131 LDLM_DEBUG(lock, "replaying lock:");
2133 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2134 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2135 aa = ptlrpc_req_async_args(req);
2136 aa->lock_handle = body->lock_handle[0];
2137 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2138 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2144 * Cancel as many unused locks as possible before replay. since we are
2145 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2148 * Called only in recovery before replaying locks. there is no need to
2149 * replay locks that are unused. since the clients may hold thousands of
2150 * cached unused locks, dropping the unused locks can greatly reduce the
2151 * load on the servers at recovery time.
2153 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2156 CFS_LIST_HEAD(cancels);
2158 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2159 "replay for namespace %s (%d)\n",
2160 ldlm_ns_name(ns), ns->ns_nr_unused);
2162 /* We don't need to care whether or not LRU resize is enabled
2163 * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2164 * count parameter */
2165 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2166 LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2168 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2169 canceled, ldlm_ns_name(ns));
2172 int ldlm_replay_locks(struct obd_import *imp)
2174 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2175 CFS_LIST_HEAD(list);
2176 struct ldlm_lock *lock, *next;
2181 LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0);
2183 /* don't replay locks if import failed recovery */
2184 if (imp->imp_vbr_failed)
2187 /* ensure this doesn't fall to 0 before all have been queued */
2188 cfs_atomic_inc(&imp->imp_replay_inflight);
2190 if (ldlm_cancel_unused_locks_before_replay)
2191 ldlm_cancel_unused_locks_for_replay(ns);
2193 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2195 cfs_list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2196 cfs_list_del_init(&lock->l_pending_chain);
2198 LDLM_LOCK_RELEASE(lock);
2199 continue; /* or try to do the rest? */
2201 rc = replay_one_lock(imp, lock);
2202 LDLM_LOCK_RELEASE(lock);
2205 cfs_atomic_dec(&imp->imp_replay_inflight);