1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2002, 2003 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
25 #include <liblustre.h>
28 #include <linux/lustre_dlm.h>
29 #include <linux/obd_class.h>
30 #include <linux/obd.h>
32 #include "ldlm_internal.h"
34 static void interrupted_completion_wait(void *data)
38 struct lock_wait_data {
39 struct ldlm_lock *lwd_lock;
43 int ldlm_expired_completion_wait(void *data)
45 struct lock_wait_data *lwd = data;
46 struct ldlm_lock *lock = lwd->lwd_lock;
47 struct obd_import *imp;
48 struct obd_device *obd;
50 if (lock->l_conn_export == NULL) {
51 static unsigned long next_dump = 0, last_dump = 0;
53 if (ptlrpc_check_suspend())
56 LDLM_ERROR(lock, "lock timed out; not entering recovery in "
57 "server code, just going back to sleep");
58 if (time_after(jiffies, next_dump)) {
59 last_dump = next_dump;
60 next_dump = jiffies + 300 * HZ;
61 ldlm_namespace_dump(D_DLMTRACE,
62 lock->l_resource->lr_namespace);
64 portals_debug_dumplog();
69 obd = lock->l_conn_export->exp_obd;
70 imp = obd->u.cli.cl_import;
71 ptlrpc_fail_import(imp, lwd->lwd_generation);
72 LDLM_ERROR(lock, "lock timed out, entering recovery for %s@%s",
73 imp->imp_target_uuid.uuid,
74 imp->imp_connection->c_remote_uuid.uuid);
79 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
81 /* XXX ALLOCATE - 160 bytes */
82 struct lock_wait_data lwd;
83 unsigned long irqflags;
84 struct obd_device *obd;
85 struct obd_import *imp = NULL;
86 struct l_wait_info lwi;
90 if (flags == LDLM_FL_WAIT_NOREPROC)
93 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
94 LDLM_FL_BLOCK_CONV))) {
95 wake_up(&lock->l_waitq);
99 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
101 ldlm_lock_dump(D_OTHER, lock, 0);
102 ldlm_reprocess_all(lock->l_resource);
106 obd = class_exp2obd(lock->l_conn_export);
108 /* if this is a local lock, then there is no import */
110 imp = obd->u.cli.cl_import;
114 if (flags & LDLM_FL_NO_TIMEOUT) {
115 LDLM_DEBUG(lock, "waiting indefinitely for group lock\n");
116 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
118 lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ,
119 ldlm_expired_completion_wait,
120 interrupted_completion_wait, &lwd);
124 spin_lock_irqsave(&imp->imp_lock, irqflags);
125 lwd.lwd_generation = imp->imp_generation;
126 spin_unlock_irqrestore(&imp->imp_lock, irqflags);
129 /* Go to sleep until the lock is granted or cancelled. */
130 rc = l_wait_event(lock->l_waitq,
131 ((lock->l_req_mode == lock->l_granted_mode) ||
132 (lock->l_flags & LDLM_FL_FAILED)), &lwi);
134 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
135 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
140 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
145 LDLM_DEBUG(lock, "client-side enqueue waking up: granted");
149 static int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
150 struct ldlm_res_id res_id,
152 ldlm_policy_data_t *policy,
155 ldlm_blocking_callback blocking,
156 ldlm_completion_callback completion,
157 ldlm_glimpse_callback glimpse,
158 void *data, __u32 lvb_len,
160 struct lustre_handle *lockh)
162 struct ldlm_lock *lock;
167 CERROR("Trying to enqueue local lock in a shadow namespace\n");
171 lock = ldlm_lock_create(ns, NULL, res_id, type, mode, blocking,
172 completion, glimpse, data, lvb_len);
174 GOTO(out_nolock, err = -ENOMEM);
175 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
177 ldlm_lock_addref_internal(lock, mode);
178 ldlm_lock2handle(lock, lockh);
179 lock->l_flags |= LDLM_FL_LOCAL;
180 lock->l_lvb_swabber = lvb_swabber;
182 memcpy(&lock->l_policy_data, policy, sizeof(*policy));
183 if (type == LDLM_EXTENT)
184 memcpy(&lock->l_req_extent, &policy->l_extent,
185 sizeof(policy->l_extent));
187 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
192 memcpy(policy, &lock->l_policy_data, sizeof(*policy));
193 if ((*flags) & LDLM_FL_LOCK_CHANGED)
194 memcpy(&res_id, &lock->l_resource->lr_name, sizeof(res_id));
196 LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
199 if (lock->l_completion_ast)
200 lock->l_completion_ast(lock, *flags, NULL);
202 LDLM_DEBUG(lock, "client-side local enqueue END");
210 static void failed_lock_cleanup(struct ldlm_namespace *ns,
211 struct ldlm_lock *lock,
212 struct lustre_handle *lockh, int mode)
214 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
215 l_lock(&ns->ns_lock);
216 lock->l_flags |= LDLM_FL_LOCAL_ONLY;
217 LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
218 l_unlock(&ns->ns_lock);
220 ldlm_lock_decref_and_cancel(lockh, mode);
222 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
223 * from llite/file.c/ll_file_flock(). */
224 if (lock->l_resource->lr_type == LDLM_FLOCK) {
225 ldlm_lock_destroy(lock);
229 int ldlm_cli_enqueue(struct obd_export *exp,
230 struct ptlrpc_request *req,
231 struct ldlm_namespace *ns,
232 struct ldlm_res_id res_id,
234 ldlm_policy_data_t *policy,
237 ldlm_blocking_callback blocking,
238 ldlm_completion_callback completion,
239 ldlm_glimpse_callback glimpse,
244 struct lustre_handle *lockh)
246 struct ldlm_lock *lock;
247 struct ldlm_request *body;
248 struct ldlm_reply *reply;
249 int rc, size[3] = {0, sizeof(*body), lvb_len}, req_passed_in = 1;
250 int is_replay = *flags & LDLM_FL_REPLAY;
251 int cleanup_phase = 0;
256 rc = ldlm_cli_enqueue_local(ns, res_id, type, policy, mode,
257 flags, blocking, completion,
258 glimpse, data, lvb_len, lvb_swabber,
263 /* If we're replaying this lock, just check some invariants.
264 * If we're creating a new lock, get everything all setup nice. */
266 lock = ldlm_handle2lock(lockh);
267 LDLM_DEBUG(lock, "client-side enqueue START");
268 LASSERT(exp == lock->l_conn_export);
270 lock = ldlm_lock_create(ns, NULL, res_id, type, mode, blocking,
271 completion, glimpse, data, lvb_len);
274 /* for the local lock, add the reference */
275 ldlm_lock_addref_internal(lock, mode);
276 ldlm_lock2handle(lock, lockh);
277 lock->l_lvb_swabber = lvb_swabber;
279 memcpy(&lock->l_policy_data, policy, sizeof(*policy));
280 if (type == LDLM_EXTENT)
281 memcpy(&lock->l_req_extent, &policy->l_extent,
282 sizeof(policy->l_extent));
283 LDLM_DEBUG(lock, "client-side enqueue START");
286 /* lock not sent to server yet */
290 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
291 LDLM_ENQUEUE, 2, size, NULL);
293 GOTO(cleanup, rc = -ENOMEM);
297 LASSERTF(req->rq_reqmsg->buflens[MDS_REQ_INTENT_LOCKREQ_OFF] ==
298 sizeof(*body), "buflen[%d] = %d, not %d\n",
299 MDS_REQ_INTENT_LOCKREQ_OFF,
300 req->rq_reqmsg->buflens[MDS_REQ_INTENT_LOCKREQ_OFF],
303 /* Dump lock data into the request buffer */
304 body = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_LOCKREQ_OFF,
306 ldlm_lock2desc(lock, &body->lock_desc);
307 body->lock_flags = *flags;
309 memcpy(&body->lock_handle1, lockh, sizeof(*lockh));
311 /* Continue as normal. */
312 if (!req_passed_in) {
313 size[0] = sizeof(*reply);
314 req->rq_replen = lustre_msg_size(1 + (lvb_len > 0), size);
316 lock->l_conn_export = exp;
317 lock->l_export = NULL;
318 lock->l_blocking_ast = blocking;
320 LDLM_DEBUG(lock, "sending request");
321 rc = ptlrpc_queue_wait(req);
323 if (rc != ELDLM_OK) {
325 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
326 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
327 if (rc == ELDLM_LOCK_ABORTED) {
328 /* Before we return, swab the reply */
329 reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
330 lustre_swab_ldlm_reply);
332 CERROR("Can't unpack ldlm_reply\n");
337 tmplvb = lustre_swab_repbuf(req, 1, lvb_len,
340 GOTO(cleanup, rc = -EPROTO);
342 memcpy(lvb, tmplvb, lvb_len);
348 reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
349 lustre_swab_ldlm_reply);
351 CERROR("Can't unpack ldlm_reply\n");
352 GOTO(cleanup, rc = -EPROTO);
355 /* lock enqueued on the server */
358 memcpy(&lock->l_remote_handle, &reply->lock_handle,
359 sizeof(lock->l_remote_handle));
360 *flags = reply->lock_flags;
362 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
363 lock, reply->lock_handle.cookie, *flags);
365 /* If enqueue returned a blocked lock but the completion handler has
366 * already run, then it fixed up the resource and we don't need to do it
368 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
369 int newmode = reply->lock_desc.l_req_mode;
371 if (newmode && newmode != lock->l_req_mode) {
372 LDLM_DEBUG(lock, "server returned different mode %s",
373 ldlm_lockname[newmode]);
374 lock->l_req_mode = newmode;
377 if (reply->lock_desc.l_resource.lr_name.name[0] !=
378 lock->l_resource->lr_name.name[0] ||
379 reply->lock_desc.l_resource.lr_name.name[1] !=
380 lock->l_resource->lr_name.name[1]) {
381 CDEBUG(D_INFO, "remote intent success, locking %ld "
383 (long)reply->lock_desc.l_resource.lr_name.name[0],
384 (long)lock->l_resource->lr_name.name[0]);
386 ldlm_lock_change_resource(ns, lock,
387 reply->lock_desc.l_resource.lr_name);
388 if (lock->l_resource == NULL) {
390 GOTO(cleanup, rc = -ENOMEM);
392 LDLM_DEBUG(lock, "client-side enqueue, new resource");
395 memcpy(&lock->l_policy_data,
396 &reply->lock_desc.l_policy_data,
397 sizeof(reply->lock_desc.l_policy_data));
398 if (type != LDLM_PLAIN)
399 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
402 if ((*flags) & LDLM_FL_AST_SENT) {
403 l_lock(&ns->ns_lock);
404 lock->l_flags |= LDLM_FL_CBPENDING;
405 l_unlock(&ns->ns_lock);
406 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
409 /* If the lock has already been granted by a completion AST, don't
410 * clobber the LVB with an older one. */
411 if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
413 tmplvb = lustre_swab_repbuf(req, 1, lvb_len, lvb_swabber);
414 if (tmplvb == NULL) {
416 GOTO(cleanup, rc = -EPROTO);
418 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
422 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
423 if (lock->l_completion_ast != NULL) {
424 int err = lock->l_completion_ast(lock, *flags, NULL);
432 if (lvb_len && lvb != NULL) {
433 /* Copy the LVB here, and not earlier, because the completion
434 * AST (if any) can override what we got in the reply */
435 memcpy(lvb, lock->l_lvb_data, lvb_len);
438 LDLM_DEBUG(lock, "client-side enqueue END");
441 switch (cleanup_phase) {
444 failed_lock_cleanup(ns, lock, lockh, mode);
446 if (!req_passed_in && req != NULL)
447 ptlrpc_req_finished(req);
454 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
457 struct ldlm_resource *res;
460 if (lock->l_resource->lr_namespace->ns_client) {
461 CERROR("Trying to cancel local lock\n");
464 LDLM_DEBUG(lock, "client-side local convert");
466 res = ldlm_lock_convert(lock, new_mode, flags);
468 ldlm_reprocess_all(res);
473 LDLM_DEBUG(lock, "client-side local convert handler END");
478 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
479 * conversion of locks which are on the waiting or converting queue */
480 /* Caller of this code is supposed to take care of lock readers/writers
482 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
484 struct ldlm_request *body;
485 struct ldlm_reply *reply;
486 struct ldlm_lock *lock;
487 struct ldlm_resource *res;
488 struct ptlrpc_request *req = NULL;
489 int rc, size = sizeof(*body);
492 lock = ldlm_handle2lock(lockh);
499 if (lock->l_conn_export == NULL)
500 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
502 LDLM_DEBUG(lock, "client-side convert");
504 req = ptlrpc_prep_req(class_exp2cliimp(lock->l_conn_export),
505 LUSTRE_DLM_VERSION, LDLM_CONVERT, 1, &size, NULL);
507 GOTO(out, rc = -ENOMEM);
509 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
510 memcpy(&body->lock_handle1, &lock->l_remote_handle,
511 sizeof(body->lock_handle1));
513 body->lock_desc.l_req_mode = new_mode;
514 body->lock_flags = *flags;
516 size = sizeof(*reply);
517 req->rq_replen = lustre_msg_size(1, &size);
519 rc = ptlrpc_queue_wait(req);
523 reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
524 lustre_swab_ldlm_reply);
526 CERROR ("Can't unpack ldlm_reply\n");
527 GOTO (out, rc = -EPROTO);
531 GOTO(out, rc = req->rq_status);
533 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
535 ldlm_reprocess_all(res);
536 /* Go to sleep until the lock is granted. */
537 /* FIXME: or cancelled. */
538 if (lock->l_completion_ast) {
539 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
550 ptlrpc_req_finished(req);
554 int ldlm_cli_cancel(struct lustre_handle *lockh)
556 struct ptlrpc_request *req;
557 struct ldlm_lock *lock;
558 struct ldlm_request *body;
559 int rc = 0, size = sizeof(*body);
562 /* concurrent cancels on the same handle can happen */
563 lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
567 if (lock->l_conn_export) {
569 struct obd_import *imp;
571 LDLM_DEBUG(lock, "client-side cancel");
572 /* Set this flag to prevent others from getting new references*/
573 l_lock(&lock->l_resource->lr_namespace->ns_lock);
574 lock->l_flags |= LDLM_FL_CBPENDING;
575 local_only = lock->l_flags & LDLM_FL_LOCAL_ONLY;
576 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
577 ldlm_cancel_callback(lock);
580 CDEBUG(D_INFO, "not sending request (at caller's "
586 imp = class_exp2cliimp(lock->l_conn_export);
587 if (imp == NULL || imp->imp_invalid) {
588 CDEBUG(D_HA, "skipping cancel on invalid import %p\n",
593 req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_CANCEL,
596 GOTO(out, rc = -ENOMEM);
597 req->rq_no_resend = 1;
599 /* XXX FIXME bug 249 */
600 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
601 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
603 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
604 memcpy(&body->lock_handle1, &lock->l_remote_handle,
605 sizeof(body->lock_handle1));
607 req->rq_replen = lustre_msg_size(0, NULL);
609 rc = ptlrpc_queue_wait(req);
612 char str[PTL_NALFMT_SIZE];
613 CERROR("client/server (nid %s) out of sync"
615 ptlrpc_peernid2str(&req->rq_import->
616 imp_connection->c_peer, str));
617 } else if (rc == -ETIMEDOUT) {
618 ptlrpc_req_finished(req);
620 } else if (rc != ELDLM_OK) {
621 CERROR("Got rc %d from cancel RPC: canceling "
625 ptlrpc_req_finished(req);
627 ldlm_lock_cancel(lock);
629 if (lock->l_resource->lr_namespace->ns_client) {
630 LDLM_ERROR(lock, "Trying to cancel local lock\n");
633 LDLM_DEBUG(lock, "client-side local cancel");
634 ldlm_lock_cancel(lock);
635 ldlm_reprocess_all(lock->l_resource);
636 LDLM_DEBUG(lock, "client-side local cancel handler END");
645 /* when called with LDLM_ASYNC the blocking callback will be handled
646 * in a thread and this function will return after the thread has been
647 * asked to call the callback. when called with LDLM_SYNC the blocking
648 * callback will be performed in this function. */
649 int ldlm_cancel_lru(struct ldlm_namespace *ns, ldlm_sync_t sync)
651 struct list_head *tmp, *next;
652 struct ldlm_lock *lock;
658 sync = LDLM_SYNC; /* force to be sync in user space */
661 l_lock(&ns->ns_lock);
662 count = ns->ns_nr_unused - ns->ns_max_unused;
665 l_unlock(&ns->ns_lock);
669 list_for_each_safe(tmp, next, &ns->ns_unused_list) {
671 lock = list_entry(tmp, struct ldlm_lock, l_lru);
673 LASSERT(!lock->l_readers && !lock->l_writers);
675 /* Setting the CBPENDING flag is a little misleading, but
676 * prevents an important race; namely, once CBPENDING is set,
677 * the lock can accumulate no more readers/writers. Since
678 * readers and writers are already zero here, ldlm_lock_decref
679 * won't see this flag and call l_blocking_ast */
680 lock->l_flags |= LDLM_FL_CBPENDING;
682 LDLM_LOCK_GET(lock); /* dropped by bl thread */
683 ldlm_lock_remove_from_lru(lock);
684 if (sync != LDLM_ASYNC || ldlm_bl_to_thread(ns, NULL, lock))
685 list_add(&lock->l_lru, &cblist);
690 l_unlock(&ns->ns_lock);
692 list_for_each_safe(tmp, next, &cblist) {
693 lock = list_entry(tmp, struct ldlm_lock, l_lru);
694 list_del_init(&lock->l_lru);
695 ldlm_handle_bl_callback(ns, NULL, lock);
700 static int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
701 struct ldlm_res_id res_id, int flags,
704 struct ldlm_resource *res;
705 struct list_head *tmp, *next, list = LIST_HEAD_INIT(list);
706 struct ldlm_ast_work *w;
709 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
711 /* This is not a problem. */
712 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id.name[0]);
716 l_lock(&ns->ns_lock);
717 list_for_each(tmp, &res->lr_granted) {
718 struct ldlm_lock *lock;
719 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
721 if (opaque != NULL && lock->l_ast_data != opaque) {
722 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
723 lock->l_ast_data, opaque);
727 if (lock->l_readers || lock->l_writers) {
728 if (flags & LDLM_FL_CONFIG_CHANGE)
729 lock->l_flags |= LDLM_FL_CBPENDING;
730 else if (flags & LDLM_FL_WARN)
731 LDLM_ERROR(lock, "lock in use");
735 /* See CBPENDING comment in ldlm_cancel_lru */
736 lock->l_flags |= LDLM_FL_CBPENDING;
738 OBD_ALLOC(w, sizeof(*w));
741 w->w_lock = LDLM_LOCK_GET(lock);
743 list_add(&w->w_list, &list);
745 l_unlock(&ns->ns_lock);
747 list_for_each_safe(tmp, next, &list) {
748 struct lustre_handle lockh;
750 w = list_entry(tmp, struct ldlm_ast_work, w_list);
752 if (flags & LDLM_FL_LOCAL_ONLY) {
753 ldlm_lock_cancel(w->w_lock);
755 ldlm_lock2handle(w->w_lock, &lockh);
756 rc = ldlm_cli_cancel(&lockh);
758 CERROR("ldlm_cli_cancel: %d\n", rc);
760 list_del(&w->w_list);
761 LDLM_LOCK_PUT(w->w_lock);
762 OBD_FREE(w, sizeof(*w));
765 ldlm_resource_putref(res);
770 static inline int have_no_nsresource(struct ldlm_namespace *ns)
774 spin_lock(&ns->ns_counter_lock);
775 if (ns->ns_resources == 0)
777 spin_unlock(&ns->ns_counter_lock);
782 /* Cancel all locks on a namespace (or a specific resource, if given)
783 * that have 0 readers/writers.
785 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
786 * to notify the server.
787 * If flags & LDLM_FL_NO_CALLBACK, don't run the cancel callback.
788 * If flags & LDLM_FL_WARN, print a warning if some locks are still in use.
789 * If flags & LDLM_FL_CONFIG_CHANGE, mark all locks as having a pending callback
791 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
792 struct ldlm_res_id *res_id, int flags, void *opaque)
795 struct l_wait_info lwi = { 0 };
802 RETURN(ldlm_cli_cancel_unused_resource(ns, *res_id, flags,
805 l_lock(&ns->ns_lock);
806 for (i = 0; i < RES_HASH_SIZE; i++) {
807 struct list_head *tmp, *next;
808 list_for_each_safe(tmp, next, &(ns->ns_hash[i])) {
810 struct ldlm_resource *res;
811 res = list_entry(tmp, struct ldlm_resource, lr_hash);
812 ldlm_resource_getref(res);
813 l_unlock(&ns->ns_lock);
815 rc = ldlm_cli_cancel_unused_resource(ns, res->lr_name,
818 CERROR("cancel_unused_res ("LPU64"): %d\n",
819 res->lr_name.name[0], rc);
821 l_lock(&ns->ns_lock);
823 ldlm_resource_putref(res);
826 l_unlock(&ns->ns_lock);
827 if (flags & LDLM_FL_CONFIG_CHANGE)
828 l_wait_event(ns->ns_waitq, have_no_nsresource(ns), &lwi);
833 /* Lock iterators. */
835 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
838 struct list_head *tmp, *next;
839 struct ldlm_lock *lock;
840 int rc = LDLM_ITER_CONTINUE;
841 struct ldlm_namespace *ns = res->lr_namespace;
846 RETURN(LDLM_ITER_CONTINUE);
848 l_lock(&ns->ns_lock);
849 list_for_each_safe(tmp, next, &res->lr_granted) {
850 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
852 if (iter(lock, closure) == LDLM_ITER_STOP)
853 GOTO(out, rc = LDLM_ITER_STOP);
856 list_for_each_safe(tmp, next, &res->lr_converting) {
857 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
859 if (iter(lock, closure) == LDLM_ITER_STOP)
860 GOTO(out, rc = LDLM_ITER_STOP);
863 list_for_each_safe(tmp, next, &res->lr_waiting) {
864 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
866 if (iter(lock, closure) == LDLM_ITER_STOP)
867 GOTO(out, rc = LDLM_ITER_STOP);
870 l_unlock(&ns->ns_lock);
874 struct iter_helper_data {
875 ldlm_iterator_t iter;
879 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
881 struct iter_helper_data *helper = closure;
882 return helper->iter(lock, helper->closure);
885 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
887 return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
890 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
893 struct iter_helper_data helper = { iter: iter, closure: closure };
894 return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
897 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
898 ldlm_res_iterator_t iter, void *closure)
900 int i, rc = LDLM_ITER_CONTINUE;
902 l_lock(&ns->ns_lock);
903 for (i = 0; i < RES_HASH_SIZE; i++) {
904 struct list_head *tmp, *next;
905 list_for_each_safe(tmp, next, &(ns->ns_hash[i])) {
906 struct ldlm_resource *res =
907 list_entry(tmp, struct ldlm_resource, lr_hash);
909 ldlm_resource_getref(res);
910 rc = iter(res, closure);
911 ldlm_resource_putref(res);
912 if (rc == LDLM_ITER_STOP)
917 l_unlock(&ns->ns_lock);
921 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
922 void ldlm_change_cbdata(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
923 ldlm_iterator_t iter, void *data)
925 struct ldlm_resource *res;
929 CERROR("must pass in namespace\n");
933 res = ldlm_resource_get(ns, NULL, *res_id, 0, 0);
939 l_lock(&ns->ns_lock);
940 ldlm_resource_foreach(res, iter, data);
941 l_unlock(&ns->ns_lock);
942 ldlm_resource_putref(res);
948 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
950 struct list_head *list = closure;
952 /* we use l_pending_chain here, because it's unused on clients. */
953 list_add(&lock->l_pending_chain, list);
954 return LDLM_ITER_CONTINUE;
957 static int replay_lock_interpret(struct ptlrpc_request *req,
960 struct ldlm_lock *lock;
961 struct ldlm_reply *reply;
963 atomic_dec(&req->rq_import->imp_replay_inflight);
967 lock = req->rq_async_args.pointer_arg[0];
968 LASSERT(lock != NULL);
970 reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
971 lustre_swab_ldlm_reply);
973 CERROR("Can't unpack ldlm_reply\n");
974 GOTO (out, rc = -EPROTO);
977 memcpy(&lock->l_remote_handle, &reply->lock_handle,
978 sizeof(lock->l_remote_handle));
979 LDLM_DEBUG(lock, "replayed lock:");
980 ptlrpc_import_recovery_state_machine(req->rq_import);
983 ptlrpc_connect_import(req->rq_import, NULL);
988 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
990 struct ptlrpc_request *req;
991 struct ldlm_request *body;
992 struct ldlm_reply *reply;
994 int size[2] = {0, sizeof(*body)};
998 * If granted mode matches the requested mode, this lock is granted.
1000 * If they differ, but we have a granted mode, then we were granted
1001 * one mode and now want another: ergo, converting.
1003 * If we haven't been granted anything and are on a resource list,
1004 * then we're blocked/waiting.
1006 * If we haven't been granted anything and we're NOT on a resource list,
1007 * then we haven't got a reply yet and don't have a known disposition.
1008 * This happens whenever a lock enqueue is the request that triggers
1011 if (lock->l_granted_mode == lock->l_req_mode)
1012 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
1013 else if (lock->l_granted_mode)
1014 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
1015 else if (!list_empty(&lock->l_res_link))
1016 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
1018 flags = LDLM_FL_REPLAY;
1020 req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
1025 /* We're part of recovery, so don't wait for it. */
1026 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
1028 body = lustre_msg_buf(req->rq_reqmsg, 1, sizeof (*body));
1029 ldlm_lock2desc(lock, &body->lock_desc);
1030 body->lock_flags = flags;
1032 ldlm_lock2handle(lock, &body->lock_handle1);
1033 size[0] = sizeof(*reply);
1034 if (lock->l_lvb_len != 0) {
1036 size[1] = lock->l_lvb_len;
1038 req->rq_replen = lustre_msg_size(buffers, size);
1040 LDLM_DEBUG(lock, "replaying lock:");
1042 imp->imp_locks_replayed++;
1043 atomic_inc(&req->rq_import->imp_replay_inflight);
1044 req->rq_async_args.pointer_arg[0] = lock;
1045 req->rq_interpret_reply = replay_lock_interpret;
1046 ptlrpcd_add_req(req);
1051 int ldlm_replay_locks(struct obd_import *imp)
1053 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
1054 struct list_head list, *pos, *next;
1055 struct ldlm_lock *lock;
1059 INIT_LIST_HEAD(&list);
1061 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1062 LASSERT(ns != NULL);
1064 /* ensure this doesn't fall to 0 before all have been queued */
1065 atomic_inc(&imp->imp_replay_inflight);
1067 l_lock(&ns->ns_lock);
1068 (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
1070 list_for_each_safe(pos, next, &list) {
1071 lock = list_entry(pos, struct ldlm_lock, l_pending_chain);
1072 rc = replay_one_lock(imp, lock);
1074 break; /* or try to do the rest? */
1076 l_unlock(&ns->ns_lock);
1078 atomic_dec(&imp->imp_replay_inflight);