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 2008 Sun Microsystems, Inc. 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 static void interrupted_completion_wait(void *data)
57 struct lock_wait_data {
58 struct ldlm_lock *lwd_lock;
62 struct ldlm_async_args {
63 struct lustre_handle lock_handle;
66 int ldlm_expired_completion_wait(void *data)
68 struct lock_wait_data *lwd = data;
69 struct ldlm_lock *lock = lwd->lwd_lock;
70 struct obd_import *imp;
71 struct obd_device *obd;
74 if (lock->l_conn_export == NULL) {
75 static cfs_time_t next_dump = 0, last_dump = 0;
77 if (ptlrpc_check_suspend())
80 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
81 CFS_DURATION_T"s ago); not entering recovery in "
82 "server code, just going back to sleep",
83 lock->l_last_activity,
84 cfs_time_sub(cfs_time_current_sec(),
85 lock->l_last_activity));
86 if (cfs_time_after(cfs_time_current(), next_dump)) {
87 last_dump = next_dump;
88 next_dump = cfs_time_shift(300);
89 ldlm_namespace_dump(D_DLMTRACE,
90 lock->l_resource->lr_namespace);
92 libcfs_debug_dumplog();
97 obd = lock->l_conn_export->exp_obd;
98 imp = obd->u.cli.cl_import;
99 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
100 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
101 CFS_DURATION_T"s ago), entering recovery for %s@%s",
102 lock->l_last_activity,
103 cfs_time_sub(cfs_time_current_sec(), lock->l_last_activity),
104 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
109 /* We use the same basis for both server side and client side functions
110 from a single node. */
111 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
113 int timeout = at_get(&lock->l_resource->lr_namespace->ns_at_estimate);
115 return obd_timeout / 2;
116 /* Since these are non-updating timeouts, we should be conservative.
117 It would be nice to have some kind of "early reply" mechanism for
118 lock callbacks too... */
119 timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */
120 return max(timeout, ldlm_enqueue_min);
122 EXPORT_SYMBOL(ldlm_get_enq_timeout);
125 * Helper function for ldlm_completion_ast(), updating timings when lock is
128 static int ldlm_completion_tail(struct ldlm_lock *lock)
133 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
134 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
137 delay = cfs_time_sub(cfs_time_current_sec(),
138 lock->l_last_activity);
139 LDLM_DEBUG(lock, "client-side enqueue: granted after "
140 CFS_DURATION_T"s", delay);
142 /* Update our time estimate */
143 at_add(&lock->l_resource->lr_namespace->ns_at_estimate, delay);
150 * Implementation of ->l_completion_ast() for a client, that doesn't wait
151 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
152 * other threads that cannot block for long.
154 int ldlm_completion_ast_async(struct ldlm_lock *lock, int flags, void *data)
158 if (flags == LDLM_FL_WAIT_NOREPROC) {
159 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
163 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
164 LDLM_FL_BLOCK_CONV))) {
165 cfs_waitq_signal(&lock->l_waitq);
166 RETURN(ldlm_completion_tail(lock));
169 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
171 ldlm_lock_dump(D_OTHER, lock, 0);
172 ldlm_reprocess_all(lock->l_resource);
177 * Client side LDLM "completion" AST. This is called in several cases:
179 * - when a reply to an ENQUEUE rpc is received from the server
180 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
181 * this point (determined by flags);
183 * - when LDLM_CP_CALLBACK rpc comes to client to notify it that lock has
186 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
189 * - to force all locks when resource is destroyed (cleanup_resource());
191 * - during lock conversion (not used currently).
193 * If lock is not granted in the first case, this function waits until second
194 * or penultimate cases happen in some other thread.
197 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
199 /* XXX ALLOCATE - 160 bytes */
200 struct lock_wait_data lwd;
201 struct obd_device *obd;
202 struct obd_import *imp = NULL;
203 struct l_wait_info lwi;
208 if (flags == LDLM_FL_WAIT_NOREPROC) {
209 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
213 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
214 LDLM_FL_BLOCK_CONV))) {
215 cfs_waitq_signal(&lock->l_waitq);
219 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
221 ldlm_lock_dump(D_OTHER, lock, 0);
225 obd = class_exp2obd(lock->l_conn_export);
227 /* if this is a local lock, then there is no import */
229 imp = obd->u.cli.cl_import;
232 /* Wait a long time for enqueue - server may have to callback a
233 lock from another client. Server will evict the other client if it
234 doesn't respond reasonably, and then give us the lock. */
235 timeout = ldlm_get_enq_timeout(lock) * 2;
239 if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
240 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
241 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
243 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
244 ldlm_expired_completion_wait,
245 interrupted_completion_wait, &lwd);
249 spin_lock(&imp->imp_lock);
250 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
251 spin_unlock(&imp->imp_lock);
254 if (ns_is_client(lock->l_resource->lr_namespace) &&
255 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
256 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
257 lock->l_flags |= LDLM_FL_FAIL_LOC;
260 /* Go to sleep until the lock is granted or cancelled. */
261 rc = l_wait_event(lock->l_waitq,
262 is_granted_or_cancelled(lock), &lwi);
266 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
271 RETURN(ldlm_completion_tail(lock));
275 * A helper to build a blocking ast function
277 * Perform a common operation for blocking asts:
278 * defferred lock cancellation.
280 * \param lock the lock blocking or canceling ast was called on
282 * \see mdt_blocking_ast
283 * \see ldlm_blocking_ast
285 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
290 lock->l_flags |= LDLM_FL_CBPENDING;
291 do_ast = (!lock->l_readers && !lock->l_writers);
292 unlock_res_and_lock(lock);
295 struct lustre_handle lockh;
298 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
299 ldlm_lock2handle(lock, &lockh);
300 rc = ldlm_cli_cancel(&lockh);
302 CERROR("ldlm_cli_cancel: %d\n", rc);
304 LDLM_DEBUG(lock, "Lock still has references, will be "
311 * Server blocking AST
313 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
316 * \param lock the lock which blocks a request or cancelling lock
319 * \param flag indicates whether this cancelling or blocking callback
321 * \see ldlm_blocking_ast_nocheck
323 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
324 void *data, int flag)
328 if (flag == LDLM_CB_CANCELING) {
329 /* Don't need to do anything here. */
333 lock_res_and_lock(lock);
334 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
335 * that ldlm_blocking_ast is called just before intent_policy method
336 * takes the ns_lock, then by the time we get the lock, we might not
337 * be the correct blocking function anymore. So check, and return
339 if (lock->l_blocking_ast != ldlm_blocking_ast) {
340 unlock_res_and_lock(lock);
343 RETURN(ldlm_blocking_ast_nocheck(lock));
347 * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
348 * comment in filter_intent_policy() on why you may need this.
350 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
353 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
354 * that is rather subtle: with OST-side locking, it may so happen that
355 * _all_ extent locks are held by the OST. If client wants to obtain
356 * current file size it calls ll{,u}_glimpse_size(), and (as locks are
357 * on the server), dummy glimpse callback fires and does
358 * nothing. Client still receives correct file size due to the
359 * following fragment in filter_intent_policy():
361 * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
362 * if (rc != 0 && res->lr_namespace->ns_lvbo &&
363 * res->lr_namespace->ns_lvbo->lvbo_update) {
364 * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
367 * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
368 * returns correct file size to the client.
370 return -ELDLM_NO_LOCK_DATA;
373 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
374 const struct ldlm_res_id *res_id,
375 ldlm_type_t type, ldlm_policy_data_t *policy,
376 ldlm_mode_t mode, int *flags,
377 ldlm_blocking_callback blocking,
378 ldlm_completion_callback completion,
379 ldlm_glimpse_callback glimpse,
380 void *data, __u32 lvb_len, void *lvb_swabber,
381 const __u64 *client_cookie,
382 struct lustre_handle *lockh)
384 struct ldlm_lock *lock;
386 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
387 .lcs_blocking = blocking,
388 .lcs_glimpse = glimpse,
392 LASSERT(!(*flags & LDLM_FL_REPLAY));
393 if (unlikely(ns_is_client(ns))) {
394 CERROR("Trying to enqueue local lock in a shadow namespace\n");
398 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len);
400 GOTO(out_nolock, err = -ENOMEM);
401 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
403 ldlm_lock_addref_internal(lock, mode);
404 ldlm_lock2handle(lock, lockh);
405 lock_res_and_lock(lock);
406 lock->l_flags |= LDLM_FL_LOCAL;
407 if (*flags & LDLM_FL_ATOMIC_CB)
408 lock->l_flags |= LDLM_FL_ATOMIC_CB;
409 lock->l_lvb_swabber = lvb_swabber;
410 unlock_res_and_lock(lock);
412 lock->l_policy_data = *policy;
413 if (client_cookie != NULL)
414 lock->l_client_cookie = *client_cookie;
415 if (type == LDLM_EXTENT)
416 lock->l_req_extent = policy->l_extent;
418 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
419 if (unlikely(err != ELDLM_OK))
423 *policy = lock->l_policy_data;
425 LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
428 if (lock->l_completion_ast)
429 lock->l_completion_ast(lock, *flags, NULL);
431 LDLM_DEBUG(lock, "client-side local enqueue END");
434 LDLM_LOCK_RELEASE(lock);
439 static void failed_lock_cleanup(struct ldlm_namespace *ns,
440 struct ldlm_lock *lock,
441 struct lustre_handle *lockh, 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 |
457 unlock_res_and_lock(lock);
461 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
462 "LDLM_FL_ATOMIC_CB");
463 ldlm_lock_decref_and_cancel(lockh, mode);
465 LDLM_DEBUG(lock, "lock was granted or failed in race");
466 ldlm_lock_decref(lockh, mode);
469 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
470 * from llite/file.c/ll_file_flock(). */
471 if (lock->l_resource->lr_type == LDLM_FLOCK) {
472 ldlm_lock_destroy(lock);
476 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
477 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
478 int *flags, void *lvb, __u32 lvb_len,
479 void *lvb_swabber, struct lustre_handle *lockh,int rc)
481 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
482 int is_replay = *flags & LDLM_FL_REPLAY;
483 struct lustre_handle old_hash_key;
484 struct ldlm_lock *lock;
485 struct ldlm_reply *reply;
486 int cleanup_phase = 1;
489 lock = ldlm_handle2lock(lockh);
490 /* ldlm_cli_enqueue is holding a reference on this lock. */
492 LASSERT(type == LDLM_FLOCK);
496 if (rc != ELDLM_OK) {
498 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
499 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
500 if (rc == ELDLM_LOCK_ABORTED) {
501 /* Before we return, swab the reply */
502 reply = req_capsule_server_get(&req->rq_pill,
507 struct ost_lvb *tmplvb;
509 req_capsule_set_size(&req->rq_pill,
510 &RMF_DLM_LVB, RCL_SERVER,
512 tmplvb = req_capsule_server_swab_get(&req->rq_pill,
516 GOTO(cleanup, rc = -EPROTO);
518 memcpy(lvb, tmplvb, lvb_len);
524 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
526 GOTO(cleanup, rc = -EPROTO);
528 /* lock enqueued on the server */
531 lock_res_and_lock(lock);
532 old_hash_key = lock->l_remote_handle;
533 lock->l_remote_handle = reply->lock_handle;
535 /* Key change rehash lock in per-export hash with new key */
536 if (exp->exp_lock_hash)
537 lustre_hash_rehash_key(exp->exp_lock_hash, &old_hash_key,
538 &lock->l_remote_handle,
541 *flags = reply->lock_flags;
542 lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
543 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
544 * to wait with no timeout as well */
545 lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
546 unlock_res_and_lock(lock);
548 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
549 lock, reply->lock_handle.cookie, *flags);
551 /* If enqueue returned a blocked lock but the completion handler has
552 * already run, then it fixed up the resource and we don't need to do it
554 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
555 int newmode = reply->lock_desc.l_req_mode;
557 if (newmode && newmode != lock->l_req_mode) {
558 LDLM_DEBUG(lock, "server returned different mode %s",
559 ldlm_lockname[newmode]);
560 lock->l_req_mode = newmode;
563 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
564 lock->l_resource->lr_name.name,
565 sizeof(struct ldlm_res_id))) {
566 CDEBUG(D_INFO, "remote intent success, locking "
567 "(%ld,%ld,%ld) instead of "
569 (long)reply->lock_desc.l_resource.lr_name.name[0],
570 (long)reply->lock_desc.l_resource.lr_name.name[1],
571 (long)reply->lock_desc.l_resource.lr_name.name[2],
572 (long)lock->l_resource->lr_name.name[0],
573 (long)lock->l_resource->lr_name.name[1],
574 (long)lock->l_resource->lr_name.name[2]);
576 rc = ldlm_lock_change_resource(ns, lock,
577 &reply->lock_desc.l_resource.lr_name);
578 if (rc || lock->l_resource == NULL)
579 GOTO(cleanup, rc = -ENOMEM);
580 LDLM_DEBUG(lock, "client-side enqueue, new resource");
583 if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
585 lock->l_policy_data =
586 reply->lock_desc.l_policy_data;
587 if (type != LDLM_PLAIN)
588 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
591 if ((*flags) & LDLM_FL_AST_SENT ||
592 /* Cancel extent locks as soon as possible on a liblustre client,
593 * because it cannot handle asynchronous ASTs robustly (see
595 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
596 lock_res_and_lock(lock);
597 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
598 unlock_res_and_lock(lock);
599 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
602 /* If the lock has already been granted by a completion AST, don't
603 * clobber the LVB with an older one. */
604 if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
607 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
609 tmplvb = req_capsule_server_swab_get(&req->rq_pill,
613 GOTO(cleanup, rc = -EPROTO);
614 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
618 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
619 if (lock->l_completion_ast != NULL) {
620 int err = lock->l_completion_ast(lock, *flags, NULL);
623 if (rc && type != LDLM_FLOCK) /* bug 9425, bug 10250 */
628 if (lvb_len && lvb != NULL) {
629 /* Copy the LVB here, and not earlier, because the completion
630 * AST (if any) can override what we got in the reply */
631 memcpy(lvb, lock->l_lvb_data, lvb_len);
634 LDLM_DEBUG(lock, "client-side enqueue END");
637 if (cleanup_phase == 1 && rc)
638 failed_lock_cleanup(ns, lock, lockh, mode);
639 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
641 LDLM_LOCK_RELEASE(lock);
645 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
646 * a single page on the send/receive side. XXX: 512 should be changed
647 * to more adequate value. */
648 static inline int ldlm_req_handles_avail(int req_size, int off)
652 avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size;
653 if (likely(avail >= 0))
654 avail /= (int)sizeof(struct lustre_handle);
657 avail += LDLM_LOCKREQ_HANDLES - off;
662 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
663 enum req_location loc,
666 int size = req_capsule_msg_size(pill, loc);
667 return ldlm_req_handles_avail(size, off);
670 static inline int ldlm_format_handles_avail(struct obd_import *imp,
671 const struct req_format *fmt,
672 enum req_location loc, int off)
674 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
675 return ldlm_req_handles_avail(size, off);
678 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
679 * @count locks in @cancels. */
680 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
681 int version, int opc, int canceloff,
682 struct list_head *cancels, int count)
684 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
685 struct req_capsule *pill = &req->rq_pill;
686 struct ldlm_request *dlm = NULL;
687 int flags, avail, to_free, bufcount, pack = 0;
694 if (exp_connect_cancelset(exp)) {
695 /* Estimate the amount of available space in the request. */
696 bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT);
697 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
699 flags = ns_connect_lru_resize(ns) ?
700 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
701 to_free = !ns_connect_lru_resize(ns) &&
702 opc == LDLM_ENQUEUE ? 1 : 0;
704 /* Cancel lru locks here _only_ if the server supports
705 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
706 * rpc, what will make us slower. */
708 count += ldlm_cancel_lru_local(ns, cancels, to_free,
709 avail - count, 0, flags);
714 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
715 ldlm_request_bufsize(pack, opc));
718 rc = ptlrpc_request_pack(req, version, opc);
720 ldlm_lock_list_put(cancels, l_bl_ast, count);
724 if (exp_connect_cancelset(exp)) {
726 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
728 /* Skip first lock handler in ldlm_request_pack(),
729 * this method will incrment @lock_count according
730 * to the lock handle amount actually written to
732 dlm->lock_count = canceloff;
734 /* Pack into the request @pack lock handles. */
735 ldlm_cli_cancel_list(cancels, pack, req, 0);
736 /* Prepare and send separate cancel rpc for others. */
737 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
739 ldlm_lock_list_put(cancels, l_bl_ast, count);
744 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
745 struct list_head *cancels, int count)
747 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
748 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
751 /* If a request has some specific initialisation it is passed in @reqp,
752 * otherwise it is created in ldlm_cli_enqueue.
754 * Supports sync and async requests, pass @async flag accordingly. If a
755 * request was created in ldlm_cli_enqueue and it is the async request,
756 * pass it to the caller in @reqp. */
757 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
758 struct ldlm_enqueue_info *einfo,
759 const struct ldlm_res_id *res_id,
760 ldlm_policy_data_t *policy, int *flags,
761 void *lvb, __u32 lvb_len, void *lvb_swabber,
762 struct lustre_handle *lockh, int async)
764 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
765 struct ldlm_lock *lock;
766 struct ldlm_request *body;
767 int is_replay = *flags & LDLM_FL_REPLAY;
768 int req_passed_in = 1;
770 struct ptlrpc_request *req;
773 LASSERT(exp != NULL);
775 /* If we're replaying this lock, just check some invariants.
776 * If we're creating a new lock, get everything all setup nice. */
778 lock = ldlm_handle2lock_long(lockh, 0);
779 LASSERT(lock != NULL);
780 LDLM_DEBUG(lock, "client-side enqueue START");
781 LASSERT(exp == lock->l_conn_export);
783 const struct ldlm_callback_suite cbs = {
784 .lcs_completion = einfo->ei_cb_cp,
785 .lcs_blocking = einfo->ei_cb_bl,
786 .lcs_glimpse = einfo->ei_cb_gl,
787 .lcs_weigh = einfo->ei_cb_wg
789 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
790 einfo->ei_mode, &cbs, einfo->ei_cbdata,
794 /* for the local lock, add the reference */
795 ldlm_lock_addref_internal(lock, einfo->ei_mode);
796 ldlm_lock2handle(lock, lockh);
797 lock->l_lvb_swabber = lvb_swabber;
798 if (policy != NULL) {
799 /* INODEBITS_INTEROP: If the server does not support
800 * inodebits, we will request a plain lock in the
801 * descriptor (ldlm_lock2desc() below) but use an
802 * inodebits lock internally with both bits set.
804 if (einfo->ei_type == LDLM_IBITS &&
805 !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
806 lock->l_policy_data.l_inodebits.bits =
807 MDS_INODELOCK_LOOKUP |
808 MDS_INODELOCK_UPDATE;
810 lock->l_policy_data = *policy;
813 if (einfo->ei_type == LDLM_EXTENT)
814 lock->l_req_extent = policy->l_extent;
815 LDLM_DEBUG(lock, "client-side enqueue START");
818 /* lock not sent to server yet */
820 if (reqp == NULL || *reqp == NULL) {
821 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
826 failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode);
827 LDLM_LOCK_RELEASE(lock);
837 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
839 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
840 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
843 lock->l_conn_export = exp;
844 lock->l_export = NULL;
845 lock->l_blocking_ast = einfo->ei_cb_bl;
847 /* Dump lock data into the request buffer */
848 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
849 ldlm_lock2desc(lock, &body->lock_desc);
850 body->lock_flags = *flags;
851 body->lock_handle[0] = *lockh;
853 /* Continue as normal. */
854 if (!req_passed_in) {
856 req_capsule_extend(&req->rq_pill,
857 &RQF_LDLM_ENQUEUE_LVB);
858 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
859 RCL_SERVER, lvb_len);
861 ptlrpc_request_set_replen(req);
865 * Liblustre client doesn't get extent locks, except for O_APPEND case
866 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
867 * [i_size, OBD_OBJECT_EOF] lock is taken.
869 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
870 policy->l_extent.end == OBD_OBJECT_EOF));
873 LASSERT(reqp != NULL);
877 LDLM_DEBUG(lock, "sending request");
879 rc = ptlrpc_queue_wait(req);
881 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
882 einfo->ei_mode, flags, lvb, lvb_len,
883 lvb_swabber, lockh, rc);
885 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
886 * one reference that we took */
888 LDLM_LOCK_RELEASE(lock);
892 if (!req_passed_in && req != NULL) {
893 ptlrpc_req_finished(req);
901 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
904 struct ldlm_resource *res;
907 if (ns_is_client(lock->l_resource->lr_namespace)) {
908 CERROR("Trying to cancel local lock\n");
911 LDLM_DEBUG(lock, "client-side local convert");
913 res = ldlm_lock_convert(lock, new_mode, flags);
915 ldlm_reprocess_all(res);
920 LDLM_DEBUG(lock, "client-side local convert handler END");
925 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
926 * conversion of locks which are on the waiting or converting queue */
927 /* Caller of this code is supposed to take care of lock readers/writers
929 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
931 struct ldlm_request *body;
932 struct ldlm_reply *reply;
933 struct ldlm_lock *lock;
934 struct ldlm_resource *res;
935 struct ptlrpc_request *req;
939 lock = ldlm_handle2lock(lockh);
946 if (lock->l_conn_export == NULL)
947 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
949 LDLM_DEBUG(lock, "client-side convert");
951 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
952 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
959 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
960 body->lock_handle[0] = lock->l_remote_handle;
962 body->lock_desc.l_req_mode = new_mode;
963 body->lock_flags = *flags;
966 ptlrpc_request_set_replen(req);
967 rc = ptlrpc_queue_wait(req);
971 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
973 GOTO(out, rc = -EPROTO);
976 GOTO(out, rc = req->rq_status);
978 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
980 ldlm_reprocess_all(res);
981 /* Go to sleep until the lock is granted. */
982 /* FIXME: or cancelled. */
983 if (lock->l_completion_ast) {
984 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
995 ptlrpc_req_finished(req);
999 /* Cancel locks locally.
1001 * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
1002 * LDLM_FL_CANCELING otherwise;
1003 * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
1004 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
1006 int rc = LDLM_FL_LOCAL_ONLY;
1009 if (lock->l_conn_export) {
1012 LDLM_DEBUG(lock, "client-side cancel");
1013 /* Set this flag to prevent others from getting new references*/
1014 lock_res_and_lock(lock);
1015 lock->l_flags |= LDLM_FL_CBPENDING;
1016 local_only = (lock->l_flags &
1017 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1018 ldlm_cancel_callback(lock);
1019 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
1020 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1021 unlock_res_and_lock(lock);
1024 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1026 rc = LDLM_FL_LOCAL_ONLY;
1028 ldlm_lock_cancel(lock);
1030 if (ns_is_client(lock->l_resource->lr_namespace)) {
1031 LDLM_ERROR(lock, "Trying to cancel local lock");
1034 LDLM_DEBUG(lock, "server-side local cancel");
1035 ldlm_lock_cancel(lock);
1036 ldlm_reprocess_all(lock->l_resource);
1037 LDLM_DEBUG(lock, "server-side local cancel handler END");
1043 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
1044 of the request @req. */
1045 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1046 struct list_head *head, int count)
1048 struct ldlm_request *dlm;
1049 struct ldlm_lock *lock;
1050 int max, packed = 0;
1053 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1054 LASSERT(dlm != NULL);
1056 /* Check the room in the request buffer. */
1057 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1058 sizeof(struct ldlm_request);
1059 max /= sizeof(struct lustre_handle);
1060 max += LDLM_LOCKREQ_HANDLES;
1061 LASSERT(max >= dlm->lock_count + count);
1063 /* XXX: it would be better to pack lock handles grouped by resource.
1064 * so that the server cancel would call filter_lvbo_update() less
1066 list_for_each_entry(lock, head, l_bl_ast) {
1069 LASSERT(lock->l_conn_export);
1070 /* Pack the lock handle to the given request buffer. */
1071 LDLM_DEBUG(lock, "packing");
1072 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1075 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1079 /* Prepare and send a batched cancel rpc, it will include count lock handles
1080 * of locks given in @head. */
1081 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1082 int count, int flags)
1084 struct ptlrpc_request *req = NULL;
1085 struct obd_import *imp;
1090 LASSERT(exp != NULL);
1093 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, obd_fail_val);
1095 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1098 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1099 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1106 imp = class_exp2cliimp(exp);
1107 if (imp == NULL || imp->imp_invalid) {
1109 "skipping cancel on invalid import %p\n", imp);
1113 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1115 GOTO(out, rc = -ENOMEM);
1117 bufcount = req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1118 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1119 ldlm_request_bufsize(count, LDLM_CANCEL));
1121 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1123 ptlrpc_request_free(req);
1126 req->rq_no_resend = 1;
1127 req->rq_no_delay = 1;
1129 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1130 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1131 ptlrpc_at_set_req_timeout(req);
1133 ldlm_cancel_pack(req, cancels, count);
1135 ptlrpc_request_set_replen(req);
1136 if (flags & LDLM_FL_ASYNC) {
1137 ptlrpcd_add_req(req, PSCOPE_OTHER);
1141 rc = ptlrpc_queue_wait(req);
1144 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1145 "out of sync -- not fatal\n",
1146 libcfs_nid2str(req->rq_import->
1147 imp_connection->c_peer.nid));
1149 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1150 req->rq_import_generation == imp->imp_generation) {
1151 ptlrpc_req_finished(req);
1153 } else if (rc != ELDLM_OK) {
1154 CERROR("Got rc %d from cancel RPC: canceling "
1162 ptlrpc_req_finished(req);
1165 return sent ? sent : rc;
1168 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1170 LASSERT(imp != NULL);
1171 return &imp->imp_obd->obd_namespace->ns_pool;
1175 * Update client's obd pool related fields with new SLV and Limit from \a req.
1177 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1179 struct obd_device *obd;
1180 __u64 old_slv, new_slv;
1183 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1184 !imp_connect_lru_resize(req->rq_import)))
1187 * Do nothing for corner cases.
1193 * In some cases RPC may contain slv and limit zeroed out. This is
1194 * the case when server does not support lru resize feature. This is
1195 * also possible in some recovery cases when server side reqs have no
1196 * ref to obd export and thus access to server side namespace is no
1199 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1200 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1201 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1202 "(SLV: "LPU64", Limit: %u)",
1203 lustre_msg_get_slv(req->rq_repmsg),
1204 lustre_msg_get_limit(req->rq_repmsg));
1208 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1209 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1210 obd = req->rq_import->imp_obd;
1213 * Set new SLV and Limit to obd fields to make accessible for pool
1214 * thread. We do not access obd_namespace and pool directly here
1215 * as there is no reliable way to make sure that they are still
1216 * alive in cleanup time. Evil races are possible which may cause
1217 * oops in that time.
1219 write_lock(&obd->obd_pool_lock);
1220 old_slv = obd->obd_pool_slv;
1221 obd->obd_pool_slv = new_slv;
1222 obd->obd_pool_limit = new_limit;
1223 write_unlock(&obd->obd_pool_lock);
1227 EXPORT_SYMBOL(ldlm_cli_update_pool);
1229 int ldlm_cli_cancel(struct lustre_handle *lockh)
1231 struct obd_export *exp;
1232 int avail, flags, count = 1, rc = 0;
1233 struct ldlm_namespace *ns;
1234 struct ldlm_lock *lock;
1235 CFS_LIST_HEAD(cancels);
1238 /* concurrent cancels on the same handle can happen */
1239 lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1241 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1245 rc = ldlm_cli_cancel_local(lock);
1246 if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) {
1247 LDLM_LOCK_RELEASE(lock);
1248 RETURN(rc < 0 ? rc : 0);
1250 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1251 * rpc which goes to canceld portal, so we can cancel other lru locks
1252 * here and send them all as one LDLM_CANCEL rpc. */
1253 LASSERT(list_empty(&lock->l_bl_ast));
1254 list_add(&lock->l_bl_ast, &cancels);
1256 exp = lock->l_conn_export;
1257 if (exp_connect_cancelset(exp)) {
1258 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1263 ns = lock->l_resource->lr_namespace;
1264 flags = ns_connect_lru_resize(ns) ?
1265 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1266 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1267 LDLM_FL_BL_AST, flags);
1269 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1273 /* XXX until we will have compound requests and can cut cancels from generic rpc
1274 * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */
1275 static int ldlm_cancel_list(struct list_head *cancels, int count, int flags)
1277 CFS_LIST_HEAD(head);
1278 struct ldlm_lock *lock, *next;
1279 int left = 0, bl_ast = 0, rc;
1282 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1286 if (flags & LDLM_FL_LOCAL_ONLY) {
1287 rc = LDLM_FL_LOCAL_ONLY;
1288 ldlm_lock_cancel(lock);
1290 rc = ldlm_cli_cancel_local(lock);
1292 if (!(flags & LDLM_FL_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1293 LDLM_DEBUG(lock, "Cancel lock separately");
1294 list_del_init(&lock->l_bl_ast);
1295 list_add(&lock->l_bl_ast, &head);
1299 if (rc == LDLM_FL_LOCAL_ONLY) {
1300 /* CANCEL RPC should not be sent to server. */
1301 list_del_init(&lock->l_bl_ast);
1302 LDLM_LOCK_RELEASE(lock);
1309 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1316 * Callback function for lru-resize policy. Makes decision whether to keep
1317 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1318 * \a added and number of locks to be preferably canceled \a count.
1320 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1322 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1324 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1325 struct ldlm_lock *lock,
1326 int unused, int added,
1329 cfs_time_t cur = cfs_time_current();
1330 struct ldlm_pool *pl = &ns->ns_pool;
1335 * Stop lru processing when we reached passed @count or checked all
1338 if (count && added >= count)
1339 return LDLM_POLICY_KEEP_LOCK;
1341 slv = ldlm_pool_get_slv(pl);
1342 lvf = ldlm_pool_get_lvf(pl);
1343 la = cfs_duration_sec(cfs_time_sub(cur,
1344 lock->l_last_used));
1347 * Stop when slv is not yet come from server or lv is smaller than
1350 lv = lvf * la * unused;
1353 * Inform pool about current CLV to see it via proc.
1355 ldlm_pool_set_clv(pl, lv);
1356 return (slv == 1 || lv < slv) ?
1357 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1361 * Callback function for proc used policy. Makes decision whether to keep
1362 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1363 * \a added and number of locks to be preferably canceled \a count.
1365 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1367 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1369 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1370 struct ldlm_lock *lock,
1371 int unused, int added,
1375 * Stop lru processing when we reached passed @count or checked all
1378 return (added >= count) ?
1379 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1383 * Callback function for aged policy. Makes decision whether to keep
1384 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1385 * \a added and number of locks to be preferably canceled \a count.
1387 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1389 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1391 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1392 struct ldlm_lock *lock,
1393 int unused, int added,
1397 * Stop lru processing if young lock is found and we reached passed
1400 return ((added >= count) &&
1401 cfs_time_before(cfs_time_current(),
1402 cfs_time_add(lock->l_last_used,
1404 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1408 * Callback function for default policy. Makes decision whether to keep
1409 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1410 * \a added and number of locks to be preferably canceled \a count.
1412 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1414 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1416 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1417 struct ldlm_lock *lock,
1418 int unused, int added,
1422 * Stop lru processing when we reached passed @count or checked all
1425 return (added >= count) ?
1426 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1429 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1430 struct ldlm_lock *, int,
1433 static ldlm_cancel_lru_policy_t
1434 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1436 if (ns_connect_lru_resize(ns)) {
1437 if (flags & LDLM_CANCEL_SHRINK)
1438 /* We kill passed number of old locks. */
1439 return ldlm_cancel_passed_policy;
1440 else if (flags & LDLM_CANCEL_LRUR)
1441 return ldlm_cancel_lrur_policy;
1442 else if (flags & LDLM_CANCEL_PASSED)
1443 return ldlm_cancel_passed_policy;
1445 if (flags & LDLM_CANCEL_AGED)
1446 return ldlm_cancel_aged_policy;
1449 return ldlm_cancel_default_policy;
1452 /* - Free space in lru for @count new locks,
1453 * redundant unused locks are canceled locally;
1454 * - also cancel locally unused aged locks;
1455 * - do not cancel more than @max locks;
1456 * - GET the found locks and add them into the @cancels list.
1458 * A client lock can be added to the l_bl_ast list only when it is
1459 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
1460 * There are the following use cases: ldlm_cancel_resource_local(),
1461 * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
1462 * flag properly. As any attempt to cancel a lock rely on this flag,
1463 * l_bl_ast list is accessed later without any special locking.
1465 * Calling policies for enabled lru resize:
1466 * ----------------------------------------
1467 * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to
1468 * cancel not more than @count locks;
1470 * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at
1471 * the beginning of lru list);
1473 * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
1474 * memory pressre policy function;
1476 * flags & LDLM_CANCEL_AGED - cancel alocks according to "aged policy".
1478 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1479 int count, int max, int cancel_flags, int flags)
1481 ldlm_cancel_lru_policy_t pf;
1482 struct ldlm_lock *lock, *next;
1483 int added = 0, unused;
1486 spin_lock(&ns->ns_unused_lock);
1487 unused = ns->ns_nr_unused;
1489 if (!ns_connect_lru_resize(ns))
1490 count += unused - ns->ns_max_unused;
1492 pf = ldlm_cancel_lru_policy(ns, flags);
1493 LASSERT(pf != NULL);
1495 while (!list_empty(&ns->ns_unused_list)) {
1496 /* For any flags, stop scanning if @max is reached. */
1497 if (max && added >= max)
1500 list_for_each_entry_safe(lock, next, &ns->ns_unused_list, l_lru){
1501 /* No locks which got blocking requests. */
1502 LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1504 /* Somebody is already doing CANCEL. No need in this
1505 * lock in lru, do not traverse it again. */
1506 if (!(lock->l_flags & LDLM_FL_CANCELING))
1509 ldlm_lock_remove_from_lru_nolock(lock);
1511 if (&lock->l_lru == &ns->ns_unused_list)
1514 LDLM_LOCK_GET(lock);
1515 spin_unlock(&ns->ns_unused_lock);
1516 lu_ref_add(&lock->l_reference, __FUNCTION__, cfs_current());
1518 /* Pass the lock through the policy filter and see if it
1519 * should stay in lru.
1521 * Even for shrinker policy we stop scanning if
1522 * we find a lock that should stay in the cache.
1523 * We should take into account lock age anyway
1524 * as new lock even if it is small of weight is
1525 * valuable resource.
1527 * That is, for shrinker policy we drop only
1528 * old locks, but additionally chose them by
1529 * their weight. Big extent locks will stay in
1531 if (pf(ns, lock, unused, added, count) ==
1532 LDLM_POLICY_KEEP_LOCK) {
1533 lu_ref_del(&lock->l_reference,
1534 __FUNCTION__, cfs_current());
1535 LDLM_LOCK_RELEASE(lock);
1536 spin_lock(&ns->ns_unused_lock);
1540 lock_res_and_lock(lock);
1541 /* Check flags again under the lock. */
1542 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1543 (ldlm_lock_remove_from_lru(lock) == 0)) {
1544 /* other thread is removing lock from lru or
1545 * somebody is already doing CANCEL or
1546 * there is a blocking request which will send
1547 * cancel by itseft or the lock is matched
1548 * is already not unused. */
1549 unlock_res_and_lock(lock);
1550 lu_ref_del(&lock->l_reference,
1551 __FUNCTION__, cfs_current());
1552 LDLM_LOCK_RELEASE(lock);
1553 spin_lock(&ns->ns_unused_lock);
1556 LASSERT(!lock->l_readers && !lock->l_writers);
1558 /* If we have chosen to cancel this lock voluntarily, we
1559 * better send cancel notification to server, so that it
1560 * frees appropriate state. This might lead to a race
1561 * where while we are doing cancel here, server is also
1562 * silently cancelling this lock. */
1563 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1565 /* Setting the CBPENDING flag is a little misleading,
1566 * but prevents an important race; namely, once
1567 * CBPENDING is set, the lock can accumulate no more
1568 * readers/writers. Since readers and writers are
1569 * already zero here, ldlm_lock_decref() won't see
1570 * this flag and call l_blocking_ast */
1571 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1573 /* We can't re-add to l_lru as it confuses the
1574 * refcounting in ldlm_lock_remove_from_lru() if an AST
1575 * arrives after we drop ns_lock below. We use l_bl_ast
1576 * and can't use l_pending_chain as it is used both on
1577 * server and client nevertheless bug 5666 says it is
1578 * used only on server */
1579 LASSERT(list_empty(&lock->l_bl_ast));
1580 list_add(&lock->l_bl_ast, cancels);
1581 unlock_res_and_lock(lock);
1582 lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current());
1583 spin_lock(&ns->ns_unused_lock);
1587 spin_unlock(&ns->ns_unused_lock);
1588 RETURN(ldlm_cancel_list(cancels, added, cancel_flags));
1591 /* when called with LDLM_ASYNC the blocking callback will be handled
1592 * in a thread and this function will return after the thread has been
1593 * asked to call the callback. when called with LDLM_SYNC the blocking
1594 * callback will be performed in this function. */
1595 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync,
1598 CFS_LIST_HEAD(cancels);
1603 sync = LDLM_SYNC; /* force to be sync in user space */
1605 count = ldlm_cancel_lru_local(ns, &cancels, nr, 0, 0, flags);
1606 if (sync == LDLM_ASYNC) {
1607 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count);
1612 /* If an error occured in ASYNC mode, or this is SYNC mode,
1613 * cancel the list. */
1614 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1618 /* Find and cancel locally unused locks found on resource, matched to the
1619 * given policy, mode. GET the found locks and add them into the @cancels
1621 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1622 struct list_head *cancels,
1623 ldlm_policy_data_t *policy,
1624 ldlm_mode_t mode, int lock_flags,
1625 int cancel_flags, void *opaque)
1627 struct ldlm_lock *lock;
1632 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1633 if (opaque != NULL && lock->l_ast_data != opaque) {
1634 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1635 lock->l_ast_data, opaque);
1640 if (lock->l_readers || lock->l_writers) {
1641 if (cancel_flags & LDLM_FL_WARN) {
1642 LDLM_ERROR(lock, "lock in use");
1648 /* If somebody is already doing CANCEL, or blocking ast came,
1649 * skip this lock. */
1650 if (lock->l_flags & LDLM_FL_BL_AST ||
1651 lock->l_flags & LDLM_FL_CANCELING)
1654 if (lockmode_compat(lock->l_granted_mode, mode))
1657 /* If policy is given and this is IBITS lock, add to list only
1658 * those locks that match by policy. */
1659 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1660 !(lock->l_policy_data.l_inodebits.bits &
1661 policy->l_inodebits.bits))
1664 /* See CBPENDING comment in ldlm_cancel_lru */
1665 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1668 LASSERT(list_empty(&lock->l_bl_ast));
1669 list_add(&lock->l_bl_ast, cancels);
1670 LDLM_LOCK_GET(lock);
1675 RETURN(ldlm_cancel_list(cancels, count, cancel_flags));
1678 /* If @req is NULL, send CANCEL request to server with handles of locks
1679 * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1680 * separately per lock.
1681 * If @req is not NULL, put handles of locks in @cancels into the request
1682 * buffer at the offset @off.
1683 * Destroy @cancels at the end. */
1684 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1685 struct ptlrpc_request *req, int flags)
1687 struct ldlm_lock *lock;
1691 if (list_empty(cancels) || count == 0)
1694 /* XXX: requests (both batched and not) could be sent in parallel.
1695 * Usually it is enough to have just 1 RPC, but it is possible that
1696 * there are to many locks to be cancelled in LRU or on a resource.
1697 * It would also speed up the case when the server does not support
1700 LASSERT(!list_empty(cancels));
1701 lock = list_entry(cancels->next, struct ldlm_lock, l_bl_ast);
1702 LASSERT(lock->l_conn_export);
1704 if (exp_connect_cancelset(lock->l_conn_export)) {
1707 ldlm_cancel_pack(req, cancels, count);
1709 res = ldlm_cli_cancel_req(lock->l_conn_export,
1713 res = ldlm_cli_cancel_req(lock->l_conn_export,
1718 CERROR("ldlm_cli_cancel_list: %d\n", res);
1723 ldlm_lock_list_put(cancels, l_bl_ast, res);
1725 LASSERT(count == 0);
1729 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1730 const struct ldlm_res_id *res_id,
1731 ldlm_policy_data_t *policy,
1732 ldlm_mode_t mode, int flags, void *opaque)
1734 struct ldlm_resource *res;
1735 CFS_LIST_HEAD(cancels);
1740 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1742 /* This is not a problem. */
1743 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1747 LDLM_RESOURCE_ADDREF(res);
1748 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1750 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1752 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1754 LDLM_RESOURCE_DELREF(res);
1755 ldlm_resource_putref(res);
1759 static inline int have_no_nsresource(struct ldlm_namespace *ns)
1761 int no_resource = 0;
1763 spin_lock(&ns->ns_hash_lock);
1764 if (ns->ns_resources == 0)
1766 spin_unlock(&ns->ns_hash_lock);
1768 RETURN(no_resource);
1771 /* Cancel all locks on a namespace (or a specific resource, if given)
1772 * that have 0 readers/writers.
1774 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1775 * to notify the server. */
1776 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1777 const struct ldlm_res_id *res_id,
1778 int flags, void *opaque)
1787 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1791 spin_lock(&ns->ns_hash_lock);
1792 for (i = 0; i < RES_HASH_SIZE; i++) {
1793 struct list_head *tmp;
1794 tmp = ns->ns_hash[i].next;
1795 while (tmp != &(ns->ns_hash[i])) {
1796 struct ldlm_resource *res;
1799 res = list_entry(tmp, struct ldlm_resource, lr_hash);
1800 ldlm_resource_getref(res);
1801 spin_unlock(&ns->ns_hash_lock);
1803 LDLM_RESOURCE_ADDREF(res);
1804 rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name,
1809 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1810 res->lr_name.name[0], rc);
1812 LDLM_RESOURCE_DELREF(res);
1813 spin_lock(&ns->ns_hash_lock);
1815 ldlm_resource_putref_locked(res);
1818 spin_unlock(&ns->ns_hash_lock);
1823 /* Lock iterators. */
1825 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1828 struct list_head *tmp, *next;
1829 struct ldlm_lock *lock;
1830 int rc = LDLM_ITER_CONTINUE;
1835 RETURN(LDLM_ITER_CONTINUE);
1838 list_for_each_safe(tmp, next, &res->lr_granted) {
1839 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1841 if (iter(lock, closure) == LDLM_ITER_STOP)
1842 GOTO(out, rc = LDLM_ITER_STOP);
1845 list_for_each_safe(tmp, next, &res->lr_converting) {
1846 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1848 if (iter(lock, closure) == LDLM_ITER_STOP)
1849 GOTO(out, rc = LDLM_ITER_STOP);
1852 list_for_each_safe(tmp, next, &res->lr_waiting) {
1853 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1855 if (iter(lock, closure) == LDLM_ITER_STOP)
1856 GOTO(out, rc = LDLM_ITER_STOP);
1863 struct iter_helper_data {
1864 ldlm_iterator_t iter;
1868 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1870 struct iter_helper_data *helper = closure;
1871 return helper->iter(lock, helper->closure);
1874 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
1876 return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
1879 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
1882 struct iter_helper_data helper = { iter: iter, closure: closure };
1883 return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
1886 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
1887 ldlm_res_iterator_t iter, void *closure)
1889 int i, rc = LDLM_ITER_CONTINUE;
1890 struct ldlm_resource *res;
1891 struct list_head *tmp;
1894 spin_lock(&ns->ns_hash_lock);
1895 for (i = 0; i < RES_HASH_SIZE; i++) {
1896 tmp = ns->ns_hash[i].next;
1897 while (tmp != &(ns->ns_hash[i])) {
1898 res = list_entry(tmp, struct ldlm_resource, lr_hash);
1899 ldlm_resource_getref(res);
1900 spin_unlock(&ns->ns_hash_lock);
1901 LDLM_RESOURCE_ADDREF(res);
1903 rc = iter(res, closure);
1905 LDLM_RESOURCE_DELREF(res);
1906 spin_lock(&ns->ns_hash_lock);
1908 ldlm_resource_putref_locked(res);
1909 if (rc == LDLM_ITER_STOP)
1914 spin_unlock(&ns->ns_hash_lock);
1918 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
1919 void ldlm_resource_iterate(struct ldlm_namespace *ns,
1920 const struct ldlm_res_id *res_id,
1921 ldlm_iterator_t iter, void *data)
1923 struct ldlm_resource *res;
1927 CERROR("must pass in namespace\n");
1931 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1937 LDLM_RESOURCE_ADDREF(res);
1938 ldlm_resource_foreach(res, iter, data);
1939 LDLM_RESOURCE_DELREF(res);
1940 ldlm_resource_putref(res);
1946 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1948 struct list_head *list = closure;
1950 /* we use l_pending_chain here, because it's unused on clients. */
1951 LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n",
1952 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1953 /* bug 9573: don't replay locks left after eviction, or
1954 * bug 17614: locks being actively cancelled. Get a reference
1955 * on a lock so that it does not disapear under us (e.g. due to cancel)
1957 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
1958 list_add(&lock->l_pending_chain, list);
1959 LDLM_LOCK_GET(lock);
1962 return LDLM_ITER_CONTINUE;
1965 static int replay_lock_interpret(const struct lu_env *env,
1966 struct ptlrpc_request *req,
1967 struct ldlm_async_args *aa, int rc)
1969 struct lustre_handle old_hash_key;
1970 struct ldlm_lock *lock;
1971 struct ldlm_reply *reply;
1972 struct obd_export *exp;
1975 atomic_dec(&req->rq_import->imp_replay_inflight);
1980 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1982 GOTO(out, rc = -EPROTO);
1984 lock = ldlm_handle2lock(&aa->lock_handle);
1986 CERROR("received replay ack for unknown local cookie "LPX64
1987 " remote cookie "LPX64 " from server %s id %s\n",
1988 aa->lock_handle.cookie, reply->lock_handle.cookie,
1989 req->rq_export->exp_client_uuid.uuid,
1990 libcfs_id2str(req->rq_peer));
1991 GOTO(out, rc = -ESTALE);
1994 old_hash_key = lock->l_remote_handle;
1995 lock->l_remote_handle = reply->lock_handle;
1997 /* Key change rehash lock in per-export hash with new key */
1998 exp = req->rq_export;
1999 if (exp && exp->exp_lock_hash)
2000 lustre_hash_rehash_key(exp->exp_lock_hash, &old_hash_key,
2001 &lock->l_remote_handle,
2004 LDLM_DEBUG(lock, "replayed lock:");
2005 ptlrpc_import_recovery_state_machine(req->rq_import);
2006 LDLM_LOCK_PUT(lock);
2009 ptlrpc_connect_import(req->rq_import, NULL);
2014 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2016 struct ptlrpc_request *req;
2017 struct ldlm_async_args *aa;
2018 struct ldlm_request *body;
2023 /* Bug 11974: Do not replay a lock which is actively being canceled */
2024 if (lock->l_flags & LDLM_FL_CANCELING) {
2025 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2029 /* If this is reply-less callback lock, we cannot replay it, since
2030 * server might have long dropped it, but notification of that event was
2031 * lost by network. (and server granted conflicting lock already) */
2032 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2033 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2034 ldlm_lock_cancel(lock);
2038 * If granted mode matches the requested mode, this lock is granted.
2040 * If they differ, but we have a granted mode, then we were granted
2041 * one mode and now want another: ergo, converting.
2043 * If we haven't been granted anything and are on a resource list,
2044 * then we're blocked/waiting.
2046 * If we haven't been granted anything and we're NOT on a resource list,
2047 * then we haven't got a reply yet and don't have a known disposition.
2048 * This happens whenever a lock enqueue is the request that triggers
2051 if (lock->l_granted_mode == lock->l_req_mode)
2052 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2053 else if (lock->l_granted_mode)
2054 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2055 else if (!list_empty(&lock->l_res_link))
2056 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2058 flags = LDLM_FL_REPLAY;
2060 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2061 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2065 /* We're part of recovery, so don't wait for it. */
2066 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2068 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2069 ldlm_lock2desc(lock, &body->lock_desc);
2070 body->lock_flags = flags;
2072 ldlm_lock2handle(lock, &body->lock_handle[0]);
2073 if (lock->l_lvb_len != 0) {
2074 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2075 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2078 ptlrpc_request_set_replen(req);
2079 /* notify the server we've replayed all requests.
2080 * also, we mark the request to be put on a dedicated
2081 * queue to be processed after all request replayes.
2083 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2085 LDLM_DEBUG(lock, "replaying lock:");
2087 atomic_inc(&req->rq_import->imp_replay_inflight);
2088 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2089 aa = ptlrpc_req_async_args(req);
2090 aa->lock_handle = body->lock_handle[0];
2091 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2092 ptlrpcd_add_req(req, PSCOPE_OTHER);
2097 int ldlm_replay_locks(struct obd_import *imp)
2099 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2100 CFS_LIST_HEAD(list);
2101 struct ldlm_lock *lock, *next;
2106 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2108 /* ensure this doesn't fall to 0 before all have been queued */
2109 atomic_inc(&imp->imp_replay_inflight);
2111 (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2113 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2114 list_del_init(&lock->l_pending_chain);
2116 LDLM_LOCK_PUT(lock);
2117 continue; /* or try to do the rest? */
2119 rc = replay_one_lock(imp, lock);
2120 LDLM_LOCK_PUT(lock);
2123 atomic_dec(&imp->imp_replay_inflight);