1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2002 Cluster File Systems, Inc.
6 * This file is part of Lustre, http://www.lustre.org.
8 * Lustre is free software; you can redistribute it and/or
9 * modify it under the terms of version 2 of the GNU General Public
10 * License as published by the Free Software Foundation.
12 * Lustre is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License
18 * along with Lustre; if not, write to the Free Software
19 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
22 #define DEBUG_SUBSYSTEM S_LDLM
24 #include <linux/lustre_dlm.h>
25 #include <linux/obd_class.h>
26 #include <linux/obd.h>
28 static int interrupted_completion_wait(void *data)
33 static int expired_completion_wait(void *data)
35 struct ldlm_lock *lock = data;
36 struct ptlrpc_connection *conn;
37 struct obd_device *obd;
40 CERROR("NULL lock\n");
41 else if (!lock->l_connh)
42 CERROR("lock %p has NULL connh\n", lock);
43 else if (!(obd = class_conn2obd(lock->l_connh)))
44 CERROR("lock %p has NULL obd\n", lock);
45 else if (!(conn = obd->u.cli.cl_import.imp_connection))
46 CERROR("lock %p has NULL connection\n", lock);
48 class_signal_connection_failure(conn);
52 int ldlm_completion_ast(struct ldlm_lock *lock, int flags)
54 struct l_wait_info lwi =
55 LWI_TIMEOUT_INTR(obd_timeout * HZ, expired_completion_wait,
56 interrupted_completion_wait, lock);
60 if (flags == LDLM_FL_WAIT_NOREPROC)
64 wake_up(&lock->l_waitq);
68 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
72 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
75 ldlm_reprocess_all(lock->l_resource);
78 /* Go to sleep until the lock is granted or cancelled. */
79 rc = l_wait_event(lock->l_waitq,
80 ((lock->l_req_mode == lock->l_granted_mode) ||
81 (lock->l_flags & LDLM_FL_DESTROYED)), &lwi);
83 if (lock->l_flags & LDLM_FL_DESTROYED) {
84 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
89 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
94 LDLM_DEBUG(lock, "client-side enqueue waking up: granted");
98 static int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
99 struct lustre_handle *parent_lockh,
102 void *cookie, int cookielen,
105 ldlm_completion_callback completion,
106 ldlm_blocking_callback blocking,
109 struct lustre_handle *lockh)
111 struct ldlm_lock *lock;
116 CERROR("Trying to enqueue local lock in a shadow namespace\n");
120 lock = ldlm_lock_create(ns, parent_lockh, res_id, type, mode, data,
123 GOTO(out_nolock, err = -ENOMEM);
124 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
126 ldlm_lock_addref_internal(lock, mode);
127 ldlm_lock2handle(lock, lockh);
128 lock->l_connh = NULL;
130 err = ldlm_lock_enqueue(lock, cookie, cookielen, flags, completion,
135 if (type == LDLM_EXTENT)
136 memcpy(cookie, &lock->l_extent, sizeof(lock->l_extent));
137 if ((*flags) & LDLM_FL_LOCK_CHANGED)
138 memcpy(res_id, lock->l_resource->lr_name, sizeof(*res_id));
140 LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
143 if (lock->l_completion_ast)
144 lock->l_completion_ast(lock, *flags);
146 LDLM_DEBUG(lock, "client-side local enqueue END");
154 int ldlm_cli_enqueue(struct lustre_handle *connh,
155 struct ptlrpc_request *req,
156 struct ldlm_namespace *ns,
157 struct lustre_handle *parent_lock_handle,
160 void *cookie, int cookielen,
163 ldlm_completion_callback completion,
164 ldlm_blocking_callback blocking,
167 struct lustre_handle *lockh)
169 struct ldlm_lock *lock;
170 struct ldlm_request *body;
171 struct ldlm_reply *reply;
172 int rc, size = sizeof(*body), req_passed_in = 1, is_replay;
175 is_replay = *flags & LDLM_FL_REPLAY;
176 LASSERT(connh != NULL || !is_replay);
179 return ldlm_cli_enqueue_local(ns, parent_lock_handle, res_id,
180 type, cookie, cookielen, mode,
181 flags, completion, blocking, data,
184 /* If we're replaying this lock, just check some invariants.
185 * If we're creating a new lock, get everything all setup nice. */
187 lock = ldlm_handle2lock(lockh);
188 LDLM_DEBUG(lock, "client-side enqueue START");
189 LASSERT(connh == lock->l_connh);
191 lock = ldlm_lock_create(ns, parent_lock_handle, res_id, type,
192 mode, data, data_len);
194 GOTO(out_nolock, rc = -ENOMEM);
195 LDLM_DEBUG(lock, "client-side enqueue START");
196 /* for the local lock, add the reference */
197 ldlm_lock_addref_internal(lock, mode);
198 ldlm_lock2handle(lock, lockh);
199 if (type == LDLM_EXTENT)
200 memcpy(&lock->l_extent, cookie,
201 sizeof(body->lock_desc.l_extent));
205 req = ptlrpc_prep_req(class_conn2cliimp(connh), LDLM_ENQUEUE, 1,
208 GOTO(out, rc = -ENOMEM);
210 } else if (req->rq_reqmsg->buflens[0] != sizeof(*body))
213 /* Dump lock data into the request buffer */
214 body = lustre_msg_buf(req->rq_reqmsg, 0);
215 ldlm_lock2desc(lock, &body->lock_desc);
216 body->lock_flags = *flags;
218 memcpy(&body->lock_handle1, lockh, sizeof(*lockh));
219 if (parent_lock_handle)
220 memcpy(&body->lock_handle2, parent_lock_handle,
221 sizeof(body->lock_handle2));
223 /* Continue as normal. */
224 if (!req_passed_in) {
225 size = sizeof(*reply);
226 req->rq_replen = lustre_msg_size(1, &size);
228 lock->l_connh = connh;
229 lock->l_export = NULL;
231 LDLM_DEBUG(lock, "sending request");
232 rc = ptlrpc_queue_wait(req);
234 if (rc != ELDLM_OK) {
236 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
237 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
238 ldlm_lock_decref(lockh, mode);
239 /* FIXME: if we've already received a completion AST, this will
241 ldlm_lock_destroy(lock);
245 reply = lustre_msg_buf(req->rq_repmsg, 0);
246 memcpy(&lock->l_remote_handle, &reply->lock_handle,
247 sizeof(lock->l_remote_handle));
248 *flags = reply->lock_flags;
250 CDEBUG(D_INFO, "local: %p, remote: %p, flags: %d\n", lock,
251 (void *)(unsigned long)reply->lock_handle.addr, *flags);
252 if (type == LDLM_EXTENT) {
253 CDEBUG(D_INFO, "requested extent: "LPU64" -> "LPU64", got "
254 "extent "LPU64" -> "LPU64"\n",
255 body->lock_desc.l_extent.start,
256 body->lock_desc.l_extent.end,
257 reply->lock_extent.start, reply->lock_extent.end);
258 cookie = &reply->lock_extent; /* FIXME bug 267 */
259 cookielen = sizeof(reply->lock_extent);
262 /* If enqueue returned a blocked lock but the completion handler has
263 * already run, then it fixed up the resource and we don't need to do it
265 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
266 int newmode = reply->lock_mode;
268 if (newmode && newmode != lock->l_req_mode) {
269 LDLM_DEBUG(lock, "server returned different mode %s",
270 ldlm_lockname[newmode]);
271 lock->l_req_mode = newmode;
274 if (reply->lock_resource_name[0] !=
275 lock->l_resource->lr_name[0]) {
276 CDEBUG(D_INFO, "remote intent success, locking %ld "
278 (long)reply->lock_resource_name[0],
279 (long)lock->l_resource->lr_name[0]);
281 ldlm_lock_change_resource(lock,
282 reply->lock_resource_name);
283 if (lock->l_resource == NULL) {
287 LDLM_DEBUG(lock, "client-side enqueue, new resource");
292 rc = ldlm_lock_enqueue(lock, cookie, cookielen, flags,
293 completion, blocking);
294 if (lock->l_completion_ast)
295 lock->l_completion_ast(lock, *flags);
299 ptlrpc_req_finished(req);
301 LDLM_DEBUG(lock, "client-side enqueue END");
309 int ldlm_match_or_enqueue(struct lustre_handle *connh,
310 struct ptlrpc_request *req,
311 struct ldlm_namespace *ns,
312 struct lustre_handle *parent_lock_handle,
315 void *cookie, int cookielen,
318 ldlm_completion_callback completion,
319 ldlm_blocking_callback blocking,
322 struct lustre_handle *lockh)
326 rc = ldlm_lock_match(ns, res_id, type, cookie, cookielen, mode, lockh);
328 rc = ldlm_cli_enqueue(connh, req, ns,
329 parent_lock_handle, res_id, type, cookie,
330 cookielen, mode, flags, completion,
331 blocking, data, data_len, lockh);
333 CERROR("ldlm_cli_enqueue: err: %d\n", rc);
339 int ldlm_cli_replay_enqueue(struct ldlm_lock *lock)
341 struct lustre_handle lockh;
342 int flags = LDLM_FL_REPLAY;
343 ldlm_lock2handle(lock, &lockh);
344 return ldlm_cli_enqueue(lock->l_connh, NULL, NULL, NULL, NULL,
345 lock->l_resource->lr_type, NULL, 0, -1, &flags,
346 NULL, NULL, NULL, 0, &lockh);
349 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
353 if (lock->l_resource->lr_namespace->ns_client) {
354 CERROR("Trying to cancel local lock\n");
357 LDLM_DEBUG(lock, "client-side local convert");
359 ldlm_lock_convert(lock, new_mode, flags);
360 ldlm_reprocess_all(lock->l_resource);
362 LDLM_DEBUG(lock, "client-side local convert handler END");
367 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
368 * conversion of locks which are on the waiting or converting queue */
369 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
371 struct ldlm_request *body;
372 struct lustre_handle *connh;
373 struct ldlm_reply *reply;
374 struct ldlm_lock *lock;
375 struct ldlm_resource *res;
376 struct ptlrpc_request *req;
377 int rc, size = sizeof(*body);
380 lock = ldlm_handle2lock(lockh);
386 connh = lock->l_connh;
389 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
391 LDLM_DEBUG(lock, "client-side convert");
393 req = ptlrpc_prep_req(class_conn2cliimp(connh), LDLM_CONVERT, 1, &size,
396 GOTO(out, rc = -ENOMEM);
398 body = lustre_msg_buf(req->rq_reqmsg, 0);
399 memcpy(&body->lock_handle1, &lock->l_remote_handle,
400 sizeof(body->lock_handle1));
402 body->lock_desc.l_req_mode = new_mode;
403 body->lock_flags = *flags;
405 size = sizeof(*reply);
406 req->rq_replen = lustre_msg_size(1, &size);
408 rc = ptlrpc_queue_wait(req);
412 reply = lustre_msg_buf(req->rq_repmsg, 0);
413 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
415 ldlm_reprocess_all(res);
416 /* Go to sleep until the lock is granted. */
417 /* FIXME: or cancelled. */
418 if (lock->l_completion_ast)
419 lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC);
423 ptlrpc_req_finished(req);
427 int ldlm_cli_cancel(struct lustre_handle *lockh)
429 struct ptlrpc_request *req;
430 struct ldlm_lock *lock;
431 struct ldlm_request *body;
432 int rc = 0, size = sizeof(*body);
435 /* concurrent cancels on the same handle can happen */
436 lock = __ldlm_handle2lock(lockh, 0, LDLM_FL_CANCELING);
441 LDLM_DEBUG(lock, "client-side cancel");
442 /* Set this flag to prevent others from getting new references*/
443 l_lock(&lock->l_resource->lr_namespace->ns_lock);
444 lock->l_flags |= LDLM_FL_CBPENDING;
445 ldlm_cancel_callback(lock);
446 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
448 req = ptlrpc_prep_req(class_conn2cliimp(lock->l_connh),
449 LDLM_CANCEL, 1, &size, NULL);
451 GOTO(out, rc = -ENOMEM);
453 /* XXX FIXME bug 249 */
454 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
455 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
457 body = lustre_msg_buf(req->rq_reqmsg, 0);
458 memcpy(&body->lock_handle1, &lock->l_remote_handle,
459 sizeof(body->lock_handle1));
461 req->rq_replen = lustre_msg_size(0, NULL);
463 rc = ptlrpc_queue_wait(req);
464 ptlrpc_req_finished(req);
468 ldlm_lock_cancel(lock);
470 LDLM_DEBUG(lock, "client-side local cancel");
471 if (lock->l_resource->lr_namespace->ns_client) {
472 CERROR("Trying to cancel local lock\n");
475 ldlm_lock_cancel(lock);
476 ldlm_reprocess_all(lock->l_resource);
477 LDLM_DEBUG(lock, "client-side local cancel handler END");
480 lock->l_flags |= LDLM_FL_CANCELING;
488 int ldlm_cancel_lru(struct ldlm_namespace *ns)
490 struct list_head *tmp, *next, list = LIST_HEAD_INIT(list);
492 struct ldlm_ast_work *w;
495 l_lock(&ns->ns_lock);
496 count = ns->ns_nr_unused - ns->ns_max_unused;
499 l_unlock(&ns->ns_lock);
503 list_for_each_safe(tmp, next, &ns->ns_unused_list) {
504 struct ldlm_lock *lock;
505 lock = list_entry(tmp, struct ldlm_lock, l_lru);
507 LASSERT(!lock->l_readers && !lock->l_writers);
509 /* Setting the CBPENDING flag is a little misleading, but
510 * prevents an important race; namely, once CBPENDING is set,
511 * the lock can accumulate no more readers/writers. Since
512 * readers and writers are already zero here, ldlm_lock_decref
513 * won't see this flag and call l_blocking_ast */
514 lock->l_flags |= LDLM_FL_CBPENDING;
516 OBD_ALLOC(w, sizeof(*w));
519 w->w_lock = LDLM_LOCK_GET(lock);
520 list_add(&w->w_list, &list);
521 list_del_init(&lock->l_lru);
526 l_unlock(&ns->ns_lock);
528 list_for_each_safe(tmp, next, &list) {
529 struct lustre_handle lockh;
531 w = list_entry(tmp, struct ldlm_ast_work, w_list);
533 ldlm_lock2handle(w->w_lock, &lockh);
534 rc = ldlm_cli_cancel(&lockh);
536 CDEBUG(D_INFO, "ldlm_cli_cancel: %d\n", rc);
538 list_del(&w->w_list);
539 LDLM_LOCK_PUT(w->w_lock);
540 OBD_FREE(w, sizeof(*w));
546 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
547 __u64 *res_id, int flags)
549 struct ldlm_resource *res;
550 struct list_head *tmp, *next, list = LIST_HEAD_INIT(list);
551 struct ldlm_ast_work *w;
554 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
556 /* This is not a problem. */
557 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id[0]);
561 l_lock(&ns->ns_lock);
562 list_for_each(tmp, &res->lr_granted) {
563 struct ldlm_lock *lock;
564 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
566 if (lock->l_readers || lock->l_writers)
569 /* See CBPENDING comment in ldlm_cancel_lru */
570 lock->l_flags |= LDLM_FL_CBPENDING;
572 OBD_ALLOC(w, sizeof(*w));
575 w->w_lock = LDLM_LOCK_GET(lock);
576 list_add(&w->w_list, &list);
578 l_unlock(&ns->ns_lock);
580 list_for_each_safe(tmp, next, &list) {
581 struct lustre_handle lockh;
583 w = list_entry(tmp, struct ldlm_ast_work, w_list);
585 /* Prevent the cancel callback from being called by setting
586 * LDLM_FL_CANCEL in the lock. Very sneaky. -p */
587 if (flags & LDLM_FL_NO_CALLBACK)
588 w->w_lock->l_flags |= LDLM_FL_CANCEL;
590 if (flags & LDLM_FL_LOCAL_ONLY) {
591 ldlm_lock_cancel(w->w_lock);
593 ldlm_lock2handle(w->w_lock, &lockh);
594 rc = ldlm_cli_cancel(&lockh);
596 CERROR("ldlm_cli_cancel: %d\n", rc);
598 list_del(&w->w_list);
599 LDLM_LOCK_PUT(w->w_lock);
600 OBD_FREE(w, sizeof(*w));
603 ldlm_resource_put(res);
608 /* Cancel all locks on a namespace (or a specific resource, if given) that have
611 * If 'local_only' is true, throw the locks away without trying to notify the
613 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns, __u64 *res_id,
619 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, flags));
621 l_lock(&ns->ns_lock);
622 for (i = 0; i < RES_HASH_SIZE; i++) {
623 struct list_head *tmp, *pos;
624 list_for_each_safe(tmp, pos, &(ns->ns_hash[i])) {
626 struct ldlm_resource *res;
627 res = list_entry(tmp, struct ldlm_resource, lr_hash);
628 ldlm_resource_getref(res);
630 rc = ldlm_cli_cancel_unused_resource(ns, res->lr_name,
634 CERROR("cancel_unused_res ("LPU64"): %d\n",
635 res->lr_name[0], rc);
636 ldlm_resource_put(res);
639 l_unlock(&ns->ns_lock);
644 /* Lock iterators. */
646 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
649 struct list_head *tmp, *next;
650 struct ldlm_lock *lock;
651 int rc = LDLM_ITER_CONTINUE;
652 struct ldlm_namespace *ns = res->lr_namespace;
657 RETURN(LDLM_ITER_CONTINUE);
659 l_lock(&ns->ns_lock);
660 list_for_each_safe(tmp, next, &res->lr_granted) {
661 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
663 if (iter(lock, closure) == LDLM_ITER_STOP)
664 GOTO(out, rc = LDLM_ITER_STOP);
667 list_for_each_safe(tmp, next, &res->lr_converting) {
668 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
670 if (iter(lock, closure) == LDLM_ITER_STOP)
671 GOTO(out, rc = LDLM_ITER_STOP);
674 list_for_each_safe(tmp, next, &res->lr_waiting) {
675 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
677 if (iter(lock, closure) == LDLM_ITER_STOP)
678 GOTO(out, rc = LDLM_ITER_STOP);
681 l_unlock(&ns->ns_lock);
685 struct iter_helper_data {
686 ldlm_iterator_t iter;
690 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
692 struct iter_helper_data *helper = closure;
693 return helper->iter(lock, helper->closure);
696 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
699 int i, rc = LDLM_ITER_CONTINUE;
700 struct iter_helper_data helper = { iter: iter, closure: closure };
702 l_lock(&ns->ns_lock);
703 for (i = 0; i < RES_HASH_SIZE; i++) {
704 struct list_head *tmp, *next;
705 list_for_each_safe(tmp, next, &(ns->ns_hash[i])) {
706 struct ldlm_resource *res =
707 list_entry(tmp, struct ldlm_resource, lr_hash);
709 ldlm_resource_getref(res);
710 rc = ldlm_resource_foreach(res, ldlm_iter_helper,
712 ldlm_resource_put(res);
713 if (rc == LDLM_ITER_STOP)
718 l_unlock(&ns->ns_lock);
724 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
726 struct list_head *list = closure;
728 /* we use l_pending_chain here, because it's unused on clients. */
729 list_add(&lock->l_pending_chain, list);
730 return LDLM_ITER_CONTINUE;
733 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock,
736 struct ptlrpc_request *req;
737 struct ldlm_request *body;
738 struct ldlm_reply *reply;
740 int flags = LDLM_FL_REPLAY;
742 flags |= lock->l_flags &
743 (LDLM_FL_BLOCK_GRANTED|LDLM_FL_BLOCK_CONV|LDLM_FL_BLOCK_WAIT);
745 size = sizeof(*body);
746 req = ptlrpc_prep_req(imp, LDLM_ENQUEUE, 1, &size, NULL);
750 body = lustre_msg_buf(req->rq_reqmsg, 0);
751 ldlm_lock2desc(lock, &body->lock_desc);
752 body->lock_flags = flags;
754 ldlm_lock2handle(lock, &body->lock_handle1);
755 size = sizeof(*reply);
756 req->rq_replen = lustre_msg_size(1, &size);
759 req->rq_reqmsg->flags |= MSG_LAST_REPLAY;
761 LDLM_DEBUG(lock, "replaying lock:");
762 rc = ptlrpc_queue_wait(req);
766 reply = lustre_msg_buf(req->rq_repmsg, 0);
767 memcpy(&lock->l_remote_handle, &reply->lock_handle,
768 sizeof(lock->l_remote_handle));
769 LDLM_DEBUG(lock, "replayed lock:");
771 ptlrpc_req_finished(req);
775 int ldlm_replay_locks(struct obd_import *imp)
777 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
778 struct list_head list, *pos, *next;
779 struct ldlm_lock *lock;
783 INIT_LIST_HEAD(&list);
785 l_lock(&ns->ns_lock);
786 (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
788 list_for_each_safe(pos, next, &list) {
789 lock = list_entry(pos, struct ldlm_lock, l_pending_chain);
790 rc = replay_one_lock(imp, lock, (next == &list));
792 break; /* or try to do the rest? */
794 l_unlock(&ns->ns_lock);