1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2002-2004 Cluster File Systems, Inc.
5 * Author: Peter Braam <braam@clusterfs.com>
6 * Author: Phil Schwan <phil@clusterfs.com>
8 * This file is part of the Lustre file system, http://www.lustre.org
9 * Lustre is a trademark of Cluster File Systems, Inc.
11 * You may have signed or agreed to another license before downloading
12 * this software. If so, you are bound by the terms and conditions
13 * of that agreement, and the following does not apply to you. See the
14 * LICENSE file included with this distribution for more information.
16 * If you did not agree to a different license, then this copy of Lustre
17 * is open source software; you can redistribute it and/or modify it
18 * under the terms of version 2 of the GNU General Public License as
19 * published by the Free Software Foundation.
21 * In either case, Lustre is distributed in the hope that it will be
22 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
23 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
24 * license text for more details.
28 # define EXPORT_SYMTAB
30 #define DEBUG_SUBSYSTEM S_LDLM
33 # include <libcfs/libcfs.h>
35 # include <liblustre.h>
38 #include <lustre_dlm.h>
39 #include <obd_class.h>
40 #include <libcfs/list.h>
41 #include "ldlm_internal.h"
43 extern cfs_mem_cache_t *ldlm_resource_slab;
44 extern cfs_mem_cache_t *ldlm_lock_slab;
45 extern struct lustre_lock ldlm_handle_lock;
46 extern struct list_head ldlm_namespace_list;
48 extern struct semaphore ldlm_namespace_lock;
49 static struct semaphore ldlm_ref_sem;
50 static int ldlm_refcount;
54 static struct ldlm_state *ldlm_state;
56 inline cfs_time_t round_timeout(cfs_time_t timeout)
58 return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1);
61 /* timeout for initial callback (AST) reply */
62 static inline unsigned int ldlm_get_rq_timeout(unsigned int ldlm_timeout,
63 unsigned int obd_timeout)
65 unsigned int timeout = min(ldlm_timeout, obd_timeout / 3);
67 return timeout < 1 ? 1 : timeout;
71 /* w_l_spinlock protects both waiting_locks_list and expired_lock_thread */
72 static spinlock_t waiting_locks_spinlock; /* BH lock (timer) */
73 static struct list_head waiting_locks_list;
74 static cfs_timer_t waiting_locks_timer;
76 static struct expired_lock_thread {
77 cfs_waitq_t elt_waitq;
80 struct list_head elt_expired_locks;
81 } expired_lock_thread;
86 #define ELT_TERMINATE 2
90 struct list_head blp_list;
91 cfs_waitq_t blp_waitq;
92 atomic_t blp_num_threads;
93 struct completion blp_comp;
96 struct ldlm_bl_work_item {
97 struct list_head blwi_entry;
98 struct ldlm_namespace *blwi_ns;
99 struct ldlm_lock_desc blwi_ld;
100 struct ldlm_lock *blwi_lock;
105 static inline int have_expired_locks(void)
110 spin_lock_bh(&waiting_locks_spinlock);
111 need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
112 spin_unlock_bh(&waiting_locks_spinlock);
117 static int expired_lock_main(void *arg)
119 struct list_head *expired = &expired_lock_thread.elt_expired_locks;
120 struct l_wait_info lwi = { 0 };
124 cfs_daemonize("ldlm_elt");
126 expired_lock_thread.elt_state = ELT_READY;
127 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
130 l_wait_event(expired_lock_thread.elt_waitq,
131 have_expired_locks() ||
132 expired_lock_thread.elt_state == ELT_TERMINATE,
135 spin_lock_bh(&waiting_locks_spinlock);
136 if (expired_lock_thread.elt_dump) {
137 spin_unlock_bh(&waiting_locks_spinlock);
139 /* from waiting_locks_callback, but not in timer */
140 libcfs_debug_dumplog();
141 libcfs_run_lbug_upcall(__FILE__,
142 "waiting_locks_callback",
143 expired_lock_thread.elt_dump);
145 spin_lock_bh(&waiting_locks_spinlock);
146 expired_lock_thread.elt_dump = 0;
151 while (!list_empty(expired)) {
152 struct obd_export *export;
153 struct ldlm_lock *lock;
155 lock = list_entry(expired->next, struct ldlm_lock,
157 if ((void *)lock < LP_POISON + PAGE_SIZE &&
158 (void *)lock >= LP_POISON) {
159 spin_unlock_bh(&waiting_locks_spinlock);
160 CERROR("free lock on elt list %p\n", lock);
163 list_del_init(&lock->l_pending_chain);
164 if ((void *)lock->l_export < LP_POISON + PAGE_SIZE &&
165 (void *)lock->l_export >= LP_POISON) {
166 CERROR("lock with free export on elt list %p\n",
168 lock->l_export = NULL;
169 LDLM_ERROR(lock, "free export");
172 export = class_export_get(lock->l_export);
173 spin_unlock_bh(&waiting_locks_spinlock);
176 class_fail_export(export);
177 class_export_put(export);
178 spin_lock_bh(&waiting_locks_spinlock);
180 spin_unlock_bh(&waiting_locks_spinlock);
182 if (do_dump && obd_dump_on_eviction) {
183 CERROR("dump the log upon eviction\n");
184 libcfs_debug_dumplog();
187 if (expired_lock_thread.elt_state == ELT_TERMINATE)
191 expired_lock_thread.elt_state = ELT_STOPPED;
192 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
196 /* This is called from within a timer interrupt and cannot schedule */
197 static void waiting_locks_callback(unsigned long unused)
199 struct ldlm_lock *lock, *last = NULL;
201 spin_lock_bh(&waiting_locks_spinlock);
202 while (!list_empty(&waiting_locks_list)) {
203 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
206 if (cfs_time_after(lock->l_callback_timeout, cfs_time_current()) ||
207 (lock->l_req_mode == LCK_GROUP))
210 LDLM_ERROR(lock, "lock callback timer expired: evicting client "
212 lock->l_export->exp_client_uuid.uuid,
213 lock->l_export->exp_connection->c_remote_uuid.uuid,
214 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
217 LDLM_ERROR(lock, "waiting on lock multiple times");
218 CERROR("wll %p n/p %p/%p, l_pending %p n/p %p/%p\n",
220 waiting_locks_list.next, waiting_locks_list.prev,
221 &lock->l_pending_chain,
222 lock->l_pending_chain.next,
223 lock->l_pending_chain.prev);
225 CFS_INIT_LIST_HEAD(&waiting_locks_list); /* HACK */
226 expired_lock_thread.elt_dump = __LINE__;
229 CEMERG("would be an LBUG, but isn't (bug 5653)\n");
230 libcfs_debug_dumpstack(NULL);
231 /*blocks* libcfs_debug_dumplog(); */
232 /*blocks* libcfs_run_lbug_upcall(file, func, line); */
237 list_del(&lock->l_pending_chain);
238 list_add(&lock->l_pending_chain,
239 &expired_lock_thread.elt_expired_locks);
242 if (!list_empty(&expired_lock_thread.elt_expired_locks)) {
243 if (obd_dump_on_timeout)
244 expired_lock_thread.elt_dump = __LINE__;
246 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
250 * Make sure the timer will fire again if we have any locks
253 if (!list_empty(&waiting_locks_list)) {
254 cfs_time_t timeout_rounded;
255 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
257 timeout_rounded = (cfs_time_t)round_timeout(lock->l_callback_timeout);
258 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
260 spin_unlock_bh(&waiting_locks_spinlock);
264 * Indicate that we're waiting for a client to call us back cancelling a given
265 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
266 * timer to fire appropriately. (We round up to the next second, to avoid
267 * floods of timer firings during periods of high lock contention and traffic).
269 * Called with the namespace lock held.
271 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock)
273 cfs_time_t timeout_rounded;
275 if (!list_empty(&lock->l_pending_chain))
278 lock->l_callback_timeout =cfs_time_add(cfs_time_current(),
279 cfs_time_seconds(obd_timeout)/2);
281 timeout_rounded = round_timeout(lock->l_callback_timeout);
283 if (cfs_time_before(timeout_rounded, cfs_timer_deadline(&waiting_locks_timer)) ||
284 !cfs_timer_is_armed(&waiting_locks_timer)) {
285 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
288 list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
292 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
296 LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
298 spin_lock_bh(&waiting_locks_spinlock);
299 if (lock->l_destroyed) {
300 static cfs_time_t next;
301 spin_unlock_bh(&waiting_locks_spinlock);
302 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
303 if (cfs_time_after(cfs_time_current(), next)) {
304 next = cfs_time_shift(14400);
305 libcfs_debug_dumpstack(NULL);
310 ret = __ldlm_add_waiting_lock(lock);
311 spin_unlock_bh(&waiting_locks_spinlock);
313 LDLM_DEBUG(lock, "%sadding to wait list",
314 ret == 0 ? "not re-" : "");
319 * Remove a lock from the pending list, likely because it had its cancellation
320 * callback arrive without incident. This adjusts the lock-timeout timer if
321 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
323 * Called with namespace lock held.
325 int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
327 struct list_head *list_next;
329 if (list_empty(&lock->l_pending_chain))
332 list_next = lock->l_pending_chain.next;
333 if (lock->l_pending_chain.prev == &waiting_locks_list) {
334 /* Removing the head of the list, adjust timer. */
335 if (list_next == &waiting_locks_list) {
336 /* No more, just cancel. */
337 cfs_timer_disarm(&waiting_locks_timer);
339 struct ldlm_lock *next;
340 next = list_entry(list_next, struct ldlm_lock,
342 cfs_timer_arm(&waiting_locks_timer,
343 round_timeout(next->l_callback_timeout));
346 list_del_init(&lock->l_pending_chain);
351 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
355 if (lock->l_export == NULL) {
356 /* We don't have a "waiting locks list" on clients. */
357 LDLM_DEBUG(lock, "client lock: no-op");
361 spin_lock_bh(&waiting_locks_spinlock);
362 ret = __ldlm_del_waiting_lock(lock);
363 spin_unlock_bh(&waiting_locks_spinlock);
365 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
372 * Called with namespace lock held.
374 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock)
376 if (lock->l_export == NULL) {
377 /* We don't have a "waiting locks list" on clients. */
378 LDLM_DEBUG(lock, "client lock: no-op");
382 spin_lock_bh(&waiting_locks_spinlock);
384 if (list_empty(&lock->l_pending_chain)) {
385 spin_unlock_bh(&waiting_locks_spinlock);
386 LDLM_DEBUG(lock, "wasn't waiting");
390 __ldlm_del_waiting_lock(lock);
391 __ldlm_add_waiting_lock(lock);
392 spin_unlock_bh(&waiting_locks_spinlock);
394 LDLM_DEBUG(lock, "refreshed");
398 #else /* !__KERNEL__ */
400 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
402 LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
406 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
411 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock)
415 #endif /* __KERNEL__ */
417 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
418 const char *ast_type)
420 struct ptlrpc_connection *conn = lock->l_export->exp_connection;
421 char *str = libcfs_nid2str(conn->c_peer.nid);
423 LCONSOLE_ERROR("A client on nid %s was evicted from service %s.\n",
424 str, lock->l_export->exp_obd->obd_name);
426 LDLM_ERROR(lock, "%s AST failed (%d): evicting client %s@%s NID %s"
427 " (%s)", ast_type, rc, lock->l_export->exp_client_uuid.uuid,
428 conn->c_remote_uuid.uuid, libcfs_nid2str(conn->c_peer.nid),
431 if (obd_dump_on_timeout)
432 libcfs_debug_dumplog();
433 class_fail_export(lock->l_export);
436 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
437 struct ptlrpc_request *req, int rc,
438 const char *ast_type)
440 lnet_process_id_t peer = req->rq_import->imp_connection->c_peer;
442 if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
443 LASSERT(lock->l_export);
444 if (lock->l_export->exp_libclient) {
445 LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
446 " timeout, just cancelling lock", ast_type,
447 libcfs_nid2str(peer.nid));
448 ldlm_lock_cancel(lock);
450 } else if (lock->l_flags & LDLM_FL_CANCEL) {
451 LDLM_DEBUG(lock, "%s AST timeout from nid %s, but "
452 "cancel was received (AST reply lost?)",
453 ast_type, libcfs_nid2str(peer.nid));
454 ldlm_lock_cancel(lock);
457 ldlm_del_waiting_lock(lock);
458 ldlm_failed_ast(lock, rc, ast_type);
462 LDLM_DEBUG(lock, "client (nid %s) returned %d"
463 " from %s AST - normal race",
464 libcfs_nid2str(peer.nid),
465 lustre_msg_get_status(req->rq_repmsg),
468 LDLM_ERROR(lock, "client (nid %s) returned %d "
469 "from %s AST", libcfs_nid2str(peer.nid),
470 (req->rq_repmsg != NULL) ?
471 lustre_msg_get_status(req->rq_repmsg) : 0,
473 ldlm_lock_cancel(lock);
474 /* Server-side AST functions are called from ldlm_reprocess_all,
475 * which needs to be told to please restart its reprocessing. */
482 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
483 struct ldlm_lock_desc *desc,
484 void *data, int flag)
486 struct ldlm_request *body;
487 struct ptlrpc_request *req;
488 int size[] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
489 [DLM_LOCKREQ_OFF] = sizeof(*body) };
490 int instant_cancel = 0, rc = 0;
493 if (flag == LDLM_CB_CANCELING) {
494 /* Don't need to do anything here. */
500 req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
501 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK, 2, size,
506 lock_res(lock->l_resource);
507 if (lock->l_granted_mode != lock->l_req_mode) {
508 /* this blocking AST will be communicated as part of the
509 * completion AST instead */
510 unlock_res(lock->l_resource);
511 ptlrpc_req_finished(req);
512 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
516 if (lock->l_destroyed) {
517 /* What's the point? */
518 unlock_res(lock->l_resource);
519 ptlrpc_req_finished(req);
524 if (CURRENT_SECONDS - lock->l_export->exp_last_request_time > 30){
525 unlock_res(lock->l_resource);
526 ptlrpc_req_finished(req);
527 ldlm_failed_ast(lock, -ETIMEDOUT, "Not-attempted blocking");
532 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
535 body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
536 body->lock_handle1 = lock->l_remote_handle;
537 body->lock_desc = *desc;
538 body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
540 LDLM_DEBUG(lock, "server preparing blocking AST");
541 ptlrpc_req_set_repsize(req, 1, NULL);
542 if (instant_cancel) {
543 unlock_res(lock->l_resource);
544 ldlm_lock_cancel(lock);
545 } else if (lock->l_granted_mode == lock->l_req_mode) {
546 ldlm_add_waiting_lock(lock);
547 unlock_res(lock->l_resource);
550 req->rq_send_state = LUSTRE_IMP_FULL;
551 req->rq_timeout = ldlm_get_rq_timeout(ldlm_timeout, obd_timeout);
552 if (unlikely(instant_cancel)) {
553 rc = ptl_send_rpc(req, 1);
555 rc = ptlrpc_queue_wait(req);
558 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
560 ptlrpc_req_finished(req);
562 /* If we cancelled the lock, we need to restart ldlm_reprocess_queue */
563 if (!rc && instant_cancel)
569 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
571 struct ldlm_request *body;
572 struct ptlrpc_request *req;
573 struct timeval granted_time;
574 long total_enqueue_wait;
575 int size[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
576 [DLM_LOCKREQ_OFF] = sizeof(*body) };
577 int rc = 0, buffers = 2, instant_cancel = 0;
580 LASSERT(lock != NULL);
582 do_gettimeofday(&granted_time);
583 total_enqueue_wait = cfs_timeval_sub(&granted_time,
584 &lock->l_enqueued_time, NULL);
586 if (total_enqueue_wait / 1000000 > obd_timeout)
587 LDLM_ERROR(lock, "enqueue wait took %luus from %lu",
588 total_enqueue_wait, lock->l_enqueued_time.tv_sec);
590 lock_res_and_lock(lock);
591 if (lock->l_resource->lr_lvb_len) {
592 size[DLM_REQ_REC_OFF] = lock->l_resource->lr_lvb_len;
595 unlock_res_and_lock(lock);
597 req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
598 LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK, buffers,
603 body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
604 body->lock_handle1 = lock->l_remote_handle;
605 body->lock_flags = flags;
606 ldlm_lock2desc(lock, &body->lock_desc);
611 lvb = lustre_msg_buf(req->rq_reqmsg, DLM_REQ_REC_OFF,
612 lock->l_resource->lr_lvb_len);
613 lock_res_and_lock(lock);
614 memcpy(lvb, lock->l_resource->lr_lvb_data,
615 lock->l_resource->lr_lvb_len);
616 unlock_res_and_lock(lock);
619 LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
621 ptlrpc_req_set_repsize(req, 1, NULL);
623 req->rq_send_state = LUSTRE_IMP_FULL;
624 req->rq_timeout = ldlm_get_rq_timeout(ldlm_timeout, obd_timeout);
626 /* We only send real blocking ASTs after the lock is granted */
627 lock_res_and_lock(lock);
628 if (lock->l_flags & LDLM_FL_AST_SENT) {
629 body->lock_flags |= LDLM_FL_AST_SENT;
631 /* We might get here prior to ldlm_handle_enqueue setting
632 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
633 * into waiting list, but this is safe and similar code in
634 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
635 * that would not only cancel the lock, but will also remove
636 * it from waiting list */
637 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
638 unlock_res_and_lock(lock);
639 ldlm_lock_cancel(lock);
641 lock_res_and_lock(lock);
643 ldlm_add_waiting_lock(lock); /* start the lock-timeout
647 unlock_res_and_lock(lock);
649 rc = ptlrpc_queue_wait(req);
651 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
653 ptlrpc_req_finished(req);
655 /* If we cancelled the lock, we need to restart ldlm_reprocess_queue */
656 if (!rc && instant_cancel)
662 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
664 struct ldlm_resource *res = lock->l_resource;
665 struct ldlm_request *body;
666 struct ptlrpc_request *req;
667 int size[] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
668 [DLM_LOCKREQ_OFF] = sizeof(*body) };
672 LASSERT(lock != NULL);
674 req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
675 LUSTRE_DLM_VERSION, LDLM_GL_CALLBACK, 2, size,
680 body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
681 body->lock_handle1 = lock->l_remote_handle;
682 ldlm_lock2desc(lock, &body->lock_desc);
684 lock_res_and_lock(lock);
685 size[REPLY_REC_OFF] = lock->l_resource->lr_lvb_len;
686 unlock_res_and_lock(lock);
687 res = lock->l_resource;
688 ptlrpc_req_set_repsize(req, 2, size);
690 req->rq_send_state = LUSTRE_IMP_FULL;
691 req->rq_timeout = ldlm_get_rq_timeout(ldlm_timeout, obd_timeout);
693 rc = ptlrpc_queue_wait(req);
694 if (rc == -ELDLM_NO_LOCK_DATA)
695 LDLM_DEBUG(lock, "lost race - client has a lock but no inode");
697 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
699 rc = res->lr_namespace->ns_lvbo->lvbo_update
700 (res, req->rq_repmsg, REPLY_REC_OFF, 1);
701 ptlrpc_req_finished(req);
705 static struct ldlm_lock *
706 find_existing_lock(struct obd_export *exp,
707 const struct lustre_handle *remote_hdl)
709 struct list_head *iter;
711 spin_lock(&exp->exp_ldlm_data.led_lock);
712 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
713 struct ldlm_lock *lock;
714 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
715 if (lock->l_remote_handle.cookie == remote_hdl->cookie) {
717 spin_unlock(&exp->exp_ldlm_data.led_lock);
721 spin_unlock(&exp->exp_ldlm_data.led_lock);
727 * Main server-side entry point into LDLM. This is called by ptlrpc service
728 * threads to carry out client lock enqueueing requests.
730 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
731 struct ptlrpc_request *req,
732 const struct ldlm_request *dlm_req,
733 const struct ldlm_callback_suite *cbs)
735 struct ldlm_reply *dlm_rep;
736 int size[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
737 [DLM_LOCKREPLY_OFF] = sizeof(*dlm_rep) };
740 ldlm_error_t err = ELDLM_OK;
741 struct ldlm_lock *lock = NULL;
745 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
747 flags = dlm_req->lock_flags;
749 LASSERT(req->rq_export);
751 if (dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
752 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE) {
753 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d\n",
754 dlm_req->lock_desc.l_resource.lr_type);
755 GOTO(out, rc = -EFAULT);
758 if (dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
759 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
760 dlm_req->lock_desc.l_req_mode & (dlm_req->lock_desc.l_req_mode-1)) {
761 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d\n",
762 dlm_req->lock_desc.l_req_mode);
763 GOTO(out, rc = -EFAULT);
766 if (req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) {
767 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN) {
768 DEBUG_REQ(D_ERROR, req,
769 "PLAIN lock request from IBITS client?\n");
770 GOTO(out, rc = -EPROTO);
772 } else if (dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) {
773 DEBUG_REQ(D_ERROR, req,
774 "IBITS lock request from unaware client?\n");
775 GOTO(out, rc = -EPROTO);
779 /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
780 against server's _CONNECT_SUPPORTED flags? (I don't want to use
781 ibits for mgc/mgs) */
783 /* INODEBITS_INTEROP: Perform conversion from plain lock to
784 * inodebits lock if client does not support them. */
785 if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) &&
786 (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
787 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
788 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
789 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
790 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
791 dlm_req->lock_desc.l_req_mode = LCK_CR;
795 if (flags & LDLM_FL_REPLAY) {
796 lock = find_existing_lock(req->rq_export,
797 &dlm_req->lock_handle1);
799 DEBUG_REQ(D_HA, req, "found existing lock cookie "LPX64,
800 lock->l_handle.h_cookie);
801 GOTO(existing_lock, rc = 0);
805 /* The lock's callback data might be set in the policy function */
806 lock = ldlm_lock_create(ns, &dlm_req->lock_handle2,
807 dlm_req->lock_desc.l_resource.lr_name,
808 dlm_req->lock_desc.l_resource.lr_type,
809 dlm_req->lock_desc.l_req_mode,
810 cbs->lcs_blocking, cbs->lcs_completion,
811 cbs->lcs_glimpse, NULL, 0);
813 GOTO(out, rc = -ENOMEM);
815 do_gettimeofday(&lock->l_enqueued_time);
816 lock->l_remote_handle = dlm_req->lock_handle1;
817 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
819 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
820 /* Don't enqueue a lock onto the export if it has already
821 * been evicted. Cancel it now instead. (bug 3822) */
822 if (req->rq_export->exp_failed) {
823 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
824 GOTO(out, rc = -ENOTCONN);
826 lock->l_export = class_export_get(req->rq_export);
827 spin_lock(&lock->l_export->exp_ldlm_data.led_lock);
828 list_add(&lock->l_export_chain,
829 &lock->l_export->exp_ldlm_data.led_held_locks);
830 spin_unlock(&lock->l_export->exp_ldlm_data.led_lock);
834 if (flags & LDLM_FL_HAS_INTENT) {
835 /* In this case, the reply buffer is allocated deep in
836 * local_lock_enqueue by the policy function. */
841 lock_res_and_lock(lock);
842 if (lock->l_resource->lr_lvb_len) {
843 size[DLM_REPLY_REC_OFF] = lock->l_resource->lr_lvb_len;
846 unlock_res_and_lock(lock);
848 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
849 GOTO(out, rc = -ENOMEM);
851 rc = lustre_pack_reply(req, buffers, size, NULL);
856 if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
857 lock->l_policy_data = dlm_req->lock_desc.l_policy_data;
858 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
859 lock->l_req_extent = lock->l_policy_data.l_extent;
861 err = ldlm_lock_enqueue(ns, &lock, cookie, &flags);
865 dlm_rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
867 dlm_rep->lock_flags = flags;
869 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
870 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
872 /* We never send a blocking AST until the lock is granted, but
873 * we can tell it right now */
874 lock_res_and_lock(lock);
876 /* Now take into account flags to be inherited from original lock
877 request both in reply to client and in our own lock flags. */
878 dlm_rep->lock_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
879 lock->l_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
881 /* Don't move a pending lock onto the export if it has already
882 * been evicted. Cancel it now instead. (bug 5683) */
883 if (req->rq_export->exp_failed ||
884 OBD_FAIL_CHECK_ONCE(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT)) {
885 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
887 } else if (lock->l_flags & LDLM_FL_AST_SENT) {
888 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
889 if (lock->l_granted_mode == lock->l_req_mode) {
891 * Only cancel lock if it was granted, because it would
892 * be destroyed immediatelly and would never be granted
893 * in the future, causing timeouts on client. Not
894 * granted lock will be cancelled immediatelly after
895 * sending completion AST.
897 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
898 unlock_res_and_lock(lock);
899 ldlm_lock_cancel(lock);
900 lock_res_and_lock(lock);
901 } else if (lock->l_granted_mode == lock->l_req_mode)
902 ldlm_add_waiting_lock(lock);
905 /* Make sure we never ever grant usual metadata locks to liblustre
907 if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
908 dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
909 req->rq_export->exp_libclient) {
910 if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) ||
911 !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK)) {
912 CERROR("Granting sync lock to libclient. "
913 "req fl %d, rep fl %d, lock fl %d\n",
914 dlm_req->lock_flags, dlm_rep->lock_flags,
916 LDLM_ERROR(lock, "sync lock");
917 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
918 struct ldlm_intent *it;
919 it = lustre_msg_buf(req->rq_reqmsg,
923 CERROR("This is intent %s ("LPU64")\n",
924 ldlm_it2str(it->opc), it->opc);
930 unlock_res_and_lock(lock);
934 req->rq_status = err;
935 if (req->rq_reply_state == NULL) {
936 err = lustre_pack_reply(req, 1, NULL, NULL);
942 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
943 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
945 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
946 "(err=%d, rc=%d)", err, rc);
949 lock_res_and_lock(lock);
950 size[DLM_REPLY_REC_OFF] = lock->l_resource->lr_lvb_len;
951 if (size[DLM_REPLY_REC_OFF] > 0) {
952 void *lvb = lustre_msg_buf(req->rq_repmsg,
954 size[DLM_REPLY_REC_OFF]);
955 LASSERTF(lvb != NULL, "req %p, lock %p\n",
958 memcpy(lvb, lock->l_resource->lr_lvb_data,
959 size[DLM_REPLY_REC_OFF]);
961 unlock_res_and_lock(lock);
963 lock_res_and_lock(lock);
964 ldlm_resource_unlink_lock(lock);
965 ldlm_lock_destroy_nolock(lock);
966 unlock_res_and_lock(lock);
969 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
970 ldlm_reprocess_all(lock->l_resource);
975 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
981 int ldlm_handle_enqueue(struct ptlrpc_request *req,
982 ldlm_completion_callback completion_callback,
983 ldlm_blocking_callback blocking_callback,
984 ldlm_glimpse_callback glimpse_callback)
987 struct ldlm_request *dlm_req;
988 struct ldlm_callback_suite cbs = {
989 .lcs_completion = completion_callback,
990 .lcs_blocking = blocking_callback,
991 .lcs_glimpse = glimpse_callback
995 dlm_req = lustre_swab_reqbuf(req, DLM_LOCKREQ_OFF,
996 sizeof *dlm_req, lustre_swab_ldlm_request);
997 if (dlm_req != NULL) {
998 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1001 CERROR ("Can't unpack dlm_req\n");
1007 int ldlm_handle_convert0(struct ptlrpc_request *req,
1008 const struct ldlm_request *dlm_req)
1010 struct ldlm_reply *dlm_rep;
1011 struct ldlm_lock *lock;
1013 int size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1014 [DLM_LOCKREPLY_OFF] = sizeof(*dlm_rep) };
1017 dlm_req = lustre_swab_reqbuf(req, DLM_LOCKREQ_OFF, sizeof(*dlm_req),
1018 lustre_swab_ldlm_request);
1019 if (dlm_req == NULL) {
1020 CERROR ("Can't unpack dlm_req\n");
1024 rc = lustre_pack_reply(req, 2, size, NULL);
1026 CERROR("out of memory\n");
1029 dlm_rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
1031 dlm_rep->lock_flags = dlm_req->lock_flags;
1033 lock = ldlm_handle2lock(&dlm_req->lock_handle1);
1035 req->rq_status = EINVAL;
1039 LDLM_DEBUG(lock, "server-side convert handler START");
1041 do_gettimeofday(&lock->l_enqueued_time);
1042 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1043 &dlm_rep->lock_flags);
1045 if (ldlm_del_waiting_lock(lock))
1046 LDLM_DEBUG(lock, "converted waiting lock");
1049 req->rq_status = EDEADLOCK;
1054 if (!req->rq_status)
1055 ldlm_reprocess_all(lock->l_resource);
1056 LDLM_DEBUG(lock, "server-side convert handler END");
1057 LDLM_LOCK_PUT(lock);
1059 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1064 int ldlm_handle_convert(struct ptlrpc_request *req)
1067 struct ldlm_request *dlm_req;
1069 dlm_req = lustre_swab_reqbuf(req, 0, sizeof *dlm_req,
1070 lustre_swab_ldlm_request);
1071 if (dlm_req != NULL) {
1072 rc = ldlm_handle_convert0(req, dlm_req);
1074 CERROR ("Can't unpack dlm_req\n");
1080 int ldlm_handle_cancel(struct ptlrpc_request *req)
1082 struct ldlm_request *dlm_req;
1083 struct ldlm_lock *lock;
1084 struct ldlm_resource *res;
1088 dlm_req = lustre_swab_reqbuf(req, DLM_LOCKREQ_OFF, sizeof(*dlm_req),
1089 lustre_swab_ldlm_request);
1090 if (dlm_req == NULL) {
1091 CERROR("bad request buffer for cancel\n");
1095 rc = lustre_pack_reply(req, 1, NULL, NULL);
1097 CERROR("out of memory\n");
1101 lock = ldlm_handle2lock(&dlm_req->lock_handle1);
1103 CERROR("received cancel for unknown lock cookie "LPX64
1104 " from client %s id %s\n",
1105 dlm_req->lock_handle1.cookie,
1106 req->rq_export->exp_client_uuid.uuid,
1107 libcfs_id2str(req->rq_peer));
1108 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock "
1110 dlm_req->lock_handle1.cookie);
1111 req->rq_status = ESTALE;
1113 LDLM_DEBUG(lock, "server-side cancel handler START");
1114 res = lock->l_resource;
1115 if (res && res->lr_namespace->ns_lvbo &&
1116 res->lr_namespace->ns_lvbo->lvbo_update) {
1117 (void)res->lr_namespace->ns_lvbo->lvbo_update
1119 //(res, req->rq_reqmsg, 1);
1122 ldlm_lock_cancel(lock);
1123 if (ldlm_del_waiting_lock(lock))
1124 CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
1125 req->rq_status = rc;
1128 if (ptlrpc_reply(req) != 0)
1132 ldlm_reprocess_all(lock->l_resource);
1133 LDLM_DEBUG(lock, "server-side cancel handler END");
1134 LDLM_LOCK_PUT(lock);
1140 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1141 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1146 LDLM_DEBUG(lock, "client blocking AST callback handler START");
1148 lock_res_and_lock(lock);
1149 lock->l_flags |= LDLM_FL_CBPENDING;
1151 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
1152 lock->l_flags |= LDLM_FL_CANCEL;
1154 do_ast = (!lock->l_readers && !lock->l_writers);
1155 unlock_res_and_lock(lock);
1158 LDLM_DEBUG(lock, "already unused, calling "
1159 "callback (%p)", lock->l_blocking_ast);
1160 if (lock->l_blocking_ast != NULL)
1161 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1164 LDLM_DEBUG(lock, "Lock still has references, will be"
1165 " cancelled later");
1168 LDLM_DEBUG(lock, "client blocking callback handler END");
1169 LDLM_LOCK_PUT(lock);
1173 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1174 struct ldlm_namespace *ns,
1175 struct ldlm_request *dlm_req,
1176 struct ldlm_lock *lock)
1178 CFS_LIST_HEAD(ast_list);
1181 LDLM_DEBUG(lock, "client completion callback handler START");
1183 lock_res_and_lock(lock);
1185 /* If we receive the completion AST before the actual enqueue returned,
1186 * then we might need to switch lock modes, resources, or extents. */
1187 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1188 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1189 LDLM_DEBUG(lock, "completion AST, new lock mode");
1192 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1193 lock->l_policy_data = dlm_req->lock_desc.l_policy_data;
1194 LDLM_DEBUG(lock, "completion AST, new policy data");
1197 ldlm_resource_unlink_lock(lock);
1198 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1199 &lock->l_resource->lr_name,
1200 sizeof(lock->l_resource->lr_name)) != 0) {
1201 unlock_res_and_lock(lock);
1202 ldlm_lock_change_resource(ns, lock,
1203 dlm_req->lock_desc.l_resource.lr_name);
1204 LDLM_DEBUG(lock, "completion AST, new resource");
1205 CERROR("change resource!\n");
1206 lock_res_and_lock(lock);
1209 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1210 lock->l_flags |= LDLM_FL_CBPENDING;
1211 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1214 if (lock->l_lvb_len) {
1216 lvb = lustre_swab_reqbuf(req, DLM_REQ_REC_OFF, lock->l_lvb_len,
1217 lock->l_lvb_swabber);
1219 LDLM_ERROR(lock, "completion AST did not contain "
1222 memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
1226 ldlm_grant_lock(lock, &ast_list);
1227 unlock_res_and_lock(lock);
1229 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1231 ldlm_run_cp_ast_work(&ast_list);
1233 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1235 LDLM_LOCK_PUT(lock);
1239 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1240 struct ldlm_namespace *ns,
1241 struct ldlm_request *dlm_req,
1242 struct ldlm_lock *lock)
1247 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1249 if (lock->l_glimpse_ast != NULL)
1250 rc = lock->l_glimpse_ast(lock, req);
1252 if (req->rq_repmsg != NULL) {
1255 req->rq_status = rc;
1259 lock_res_and_lock(lock);
1260 if (lock->l_granted_mode == LCK_PW &&
1261 !lock->l_readers && !lock->l_writers &&
1262 cfs_time_after(cfs_time_current(),
1263 cfs_time_add(lock->l_last_used, cfs_time_seconds(10)))) {
1264 unlock_res_and_lock(lock);
1265 if (ldlm_bl_to_thread(ns, NULL, lock))
1266 ldlm_handle_bl_callback(ns, NULL, lock);
1271 unlock_res_and_lock(lock);
1272 LDLM_LOCK_PUT(lock);
1276 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1278 req->rq_status = rc;
1279 if (req->rq_reply_state == NULL) {
1280 rc = lustre_pack_reply(req, 1, NULL, NULL);
1284 return ptlrpc_reply(req);
1287 int ldlm_bl_to_thread(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1288 struct ldlm_lock *lock)
1291 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1292 struct ldlm_bl_work_item *blwi;
1295 OBD_ALLOC(blwi, sizeof(*blwi));
1301 blwi->blwi_ld = *ld;
1302 blwi->blwi_lock = lock;
1304 spin_lock(&blp->blp_lock);
1305 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1306 cfs_waitq_signal(&blp->blp_waitq);
1307 spin_unlock(&blp->blp_lock);
1315 static int ldlm_callback_handler(struct ptlrpc_request *req)
1317 struct ldlm_namespace *ns;
1318 struct ldlm_request *dlm_req;
1319 struct ldlm_lock *lock;
1323 /* Requests arrive in sender's byte order. The ptlrpc service
1324 * handler has already checked and, if necessary, byte-swapped the
1325 * incoming request message body, but I am responsible for the
1326 * message buffers. */
1328 if (req->rq_export == NULL) {
1329 struct ldlm_request *dlm_req;
1331 CDEBUG(D_RPCTRACE, "operation %d from %s with bad "
1332 "export cookie "LPX64"; this is "
1333 "normal if this node rebooted with a lock held\n",
1334 lustre_msg_get_opc(req->rq_reqmsg),
1335 libcfs_id2str(req->rq_peer),
1336 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
1338 dlm_req = lustre_swab_reqbuf(req, DLM_LOCKREQ_OFF,
1340 lustre_swab_ldlm_request);
1341 if (dlm_req != NULL)
1342 CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
1343 dlm_req->lock_handle1.cookie);
1345 ldlm_callback_reply(req, -ENOTCONN);
1349 LASSERT(req->rq_export != NULL);
1350 LASSERT(req->rq_export->exp_obd != NULL);
1352 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1353 case LDLM_BL_CALLBACK:
1354 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1356 case LDLM_CP_CALLBACK:
1357 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
1359 case LDLM_GL_CALLBACK:
1360 OBD_FAIL_RETURN(OBD_FAIL_LDLM_GL_CALLBACK, 0);
1362 case OBD_LOG_CANCEL: /* remove this eventually - for 1.4.0 compat */
1363 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1364 rc = llog_origin_handle_cancel(req);
1365 ldlm_callback_reply(req, rc);
1367 case OBD_QC_CALLBACK:
1368 OBD_FAIL_RETURN(OBD_FAIL_OBD_QC_CALLBACK_NET, 0);
1369 rc = target_handle_qc_callback(req);
1370 ldlm_callback_reply(req, rc);
1374 /* reply in handler */
1375 rc = target_handle_dqacq_callback(req);
1377 case LLOG_ORIGIN_HANDLE_CREATE:
1378 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1379 rc = llog_origin_handle_create(req);
1380 ldlm_callback_reply(req, rc);
1382 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1383 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1384 rc = llog_origin_handle_next_block(req);
1385 ldlm_callback_reply(req, rc);
1387 case LLOG_ORIGIN_HANDLE_READ_HEADER:
1388 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1389 rc = llog_origin_handle_read_header(req);
1390 ldlm_callback_reply(req, rc);
1392 case LLOG_ORIGIN_HANDLE_CLOSE:
1393 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1394 rc = llog_origin_handle_close(req);
1395 ldlm_callback_reply(req, rc);
1398 CERROR("unknown opcode %u\n",
1399 lustre_msg_get_opc(req->rq_reqmsg));
1400 ldlm_callback_reply(req, -EPROTO);
1404 ns = req->rq_export->exp_obd->obd_namespace;
1405 LASSERT(ns != NULL);
1407 dlm_req = lustre_swab_reqbuf(req, DLM_LOCKREQ_OFF, sizeof(*dlm_req),
1408 lustre_swab_ldlm_request);
1409 if (dlm_req == NULL) {
1410 CERROR ("can't unpack dlm_req\n");
1411 ldlm_callback_reply(req, -EPROTO);
1415 lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle1);
1417 CDEBUG(D_INODE, "callback on lock "LPX64" - lock disappeared\n",
1418 dlm_req->lock_handle1.cookie);
1419 ldlm_callback_reply(req, -EINVAL);
1423 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1424 lock_res_and_lock(lock);
1425 lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1426 unlock_res_and_lock(lock);
1428 /* We want the ost thread to get this reply so that it can respond
1429 * to ost requests (write cache writeback) that might be triggered
1432 * But we'd also like to be able to indicate in the reply that we're
1433 * cancelling right now, because it's unused, or have an intent result
1434 * in the reply, so we might have to push the responsibility for sending
1435 * the reply down into the AST handlers, alas. */
1437 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1438 case LDLM_BL_CALLBACK:
1439 CDEBUG(D_INODE, "blocking ast\n");
1440 if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK))
1441 ldlm_callback_reply(req, 0);
1442 if (ldlm_bl_to_thread(ns, &dlm_req->lock_desc, lock))
1443 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1445 case LDLM_CP_CALLBACK:
1446 CDEBUG(D_INODE, "completion ast\n");
1447 ldlm_callback_reply(req, 0);
1448 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1450 case LDLM_GL_CALLBACK:
1451 CDEBUG(D_INODE, "glimpse ast\n");
1452 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1455 LBUG(); /* checked above */
1461 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1466 /* Requests arrive in sender's byte order. The ptlrpc service
1467 * handler has already checked and, if necessary, byte-swapped the
1468 * incoming request message body, but I am responsible for the
1469 * message buffers. */
1471 if (req->rq_export == NULL) {
1472 struct ldlm_request *dlm_req;
1474 CERROR("operation %d from %s with bad export cookie "LPU64"\n",
1475 lustre_msg_get_opc(req->rq_reqmsg),
1476 libcfs_id2str(req->rq_peer),
1477 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
1479 dlm_req = lustre_swab_reqbuf(req, DLM_LOCKREQ_OFF,
1481 lustre_swab_ldlm_request);
1482 if (dlm_req != NULL)
1483 ldlm_lock_dump_handle(D_ERROR, &dlm_req->lock_handle1);
1485 ldlm_callback_reply(req, -ENOTCONN);
1489 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1491 /* XXX FIXME move this back to mds/handler.c, bug 249 */
1493 CDEBUG(D_INODE, "cancel\n");
1494 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1495 rc = ldlm_handle_cancel(req);
1499 case OBD_LOG_CANCEL:
1500 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1501 rc = llog_origin_handle_cancel(req);
1502 ldlm_callback_reply(req, rc);
1505 CERROR("invalid opcode %d\n",
1506 lustre_msg_get_opc(req->rq_reqmsg));
1507 ldlm_callback_reply(req, -EINVAL);
1513 void ldlm_revoke_export_locks(struct obd_export *exp)
1515 struct list_head *locklist = &exp->exp_ldlm_data.led_held_locks;
1516 struct list_head rpc_list;
1517 struct ldlm_lock *lock, *next;
1518 struct ldlm_lock_desc desc;
1521 INIT_LIST_HEAD(&rpc_list);
1523 spin_lock(&exp->exp_ldlm_data.led_lock);
1524 list_for_each_entry_safe(lock, next, locklist, l_export_chain) {
1525 lock_res_and_lock(lock);
1526 if (lock->l_req_mode != lock->l_granted_mode) {
1527 unlock_res_and_lock(lock);
1531 LASSERT(lock->l_resource);
1532 if (lock->l_resource->lr_type != LDLM_IBITS &&
1533 lock->l_resource->lr_type != LDLM_PLAIN) {
1534 unlock_res_and_lock(lock);
1538 if (lock->l_flags & LDLM_FL_AST_SENT) {
1539 unlock_res_and_lock(lock);
1543 LASSERT(lock->l_blocking_ast);
1544 LASSERT(!lock->l_blocking_lock);
1546 lock->l_flags |= LDLM_FL_AST_SENT;
1547 unlock_res_and_lock(lock);
1549 list_move(&lock->l_export_chain, &rpc_list);
1551 spin_unlock(&exp->exp_ldlm_data.led_lock);
1553 while (!list_empty(&rpc_list)) {
1554 lock = list_entry(rpc_list.next, struct ldlm_lock,
1556 list_del_init(&lock->l_export_chain);
1558 /* the desc just pretend to exclusive */
1559 ldlm_lock2desc(lock, &desc);
1560 desc.l_req_mode = LCK_EX;
1561 desc.l_granted_mode = 0;
1563 lock->l_blocking_ast(lock, &desc, NULL, LDLM_CB_BLOCKING);
1569 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1571 struct ldlm_bl_work_item *blwi = NULL;
1573 spin_lock(&blp->blp_lock);
1574 if (!list_empty(&blp->blp_list)) {
1575 blwi = list_entry(blp->blp_list.next, struct ldlm_bl_work_item,
1577 list_del(&blwi->blwi_entry);
1579 spin_unlock(&blp->blp_lock);
1584 struct ldlm_bl_thread_data {
1586 struct ldlm_bl_pool *bltd_blp;
1589 static int ldlm_bl_thread_main(void *arg)
1591 struct ldlm_bl_thread_data *bltd = arg;
1592 struct ldlm_bl_pool *blp = bltd->bltd_blp;
1596 char name[CFS_CURPROC_COMM_MAX];
1597 snprintf(name, sizeof(name) - 1, "ldlm_bl_%02d",
1599 cfs_daemonize(name);
1602 atomic_inc(&blp->blp_num_threads);
1603 complete(&blp->blp_comp);
1606 struct l_wait_info lwi = { 0 };
1607 struct ldlm_bl_work_item *blwi = NULL;
1609 l_wait_event_exclusive(blp->blp_waitq,
1610 (blwi = ldlm_bl_get_work(blp)) != NULL,
1613 if (blwi->blwi_ns == NULL)
1616 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1618 OBD_FREE(blwi, sizeof(*blwi));
1621 atomic_dec(&blp->blp_num_threads);
1622 complete(&blp->blp_comp);
1628 static int ldlm_setup(void);
1629 static int ldlm_cleanup(int force);
1631 int ldlm_get_ref(void)
1635 mutex_down(&ldlm_ref_sem);
1636 if (++ldlm_refcount == 1) {
1641 mutex_up(&ldlm_ref_sem);
1646 void ldlm_put_ref(int force)
1649 mutex_down(&ldlm_ref_sem);
1650 if (ldlm_refcount == 1) {
1651 int rc = ldlm_cleanup(force);
1653 CERROR("ldlm_cleanup failed: %d\n", rc);
1659 mutex_up(&ldlm_ref_sem);
1664 static int ldlm_setup(void)
1666 struct ldlm_bl_pool *blp;
1673 if (ldlm_state != NULL)
1676 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
1677 if (ldlm_state == NULL)
1681 rc = ldlm_proc_setup();
1686 ldlm_state->ldlm_cb_service =
1687 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1688 LDLM_MAXREPSIZE, LDLM_CB_REQUEST_PORTAL,
1689 LDLM_CB_REPLY_PORTAL, ldlm_timeout * 900,
1690 ldlm_callback_handler, "ldlm_cbd",
1691 ldlm_svc_proc_dir, NULL, LDLM_NUM_THREADS,
1692 LCT_MD_THREAD|LCT_DT_THREAD);
1694 if (!ldlm_state->ldlm_cb_service) {
1695 CERROR("failed to start service\n");
1696 GOTO(out_proc, rc = -ENOMEM);
1699 ldlm_state->ldlm_cancel_service =
1700 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1701 LDLM_MAXREPSIZE, LDLM_CANCEL_REQUEST_PORTAL,
1702 LDLM_CANCEL_REPLY_PORTAL, 30000,
1703 ldlm_cancel_handler, "ldlm_canceld",
1704 ldlm_svc_proc_dir, NULL, LDLM_NUM_THREADS,
1705 LCT_MD_THREAD|LCT_DT_THREAD);
1707 if (!ldlm_state->ldlm_cancel_service) {
1708 CERROR("failed to start service\n");
1709 GOTO(out_proc, rc = -ENOMEM);
1712 OBD_ALLOC(blp, sizeof(*blp));
1714 GOTO(out_proc, rc = -ENOMEM);
1715 ldlm_state->ldlm_bl_pool = blp;
1717 atomic_set(&blp->blp_num_threads, 0);
1718 cfs_waitq_init(&blp->blp_waitq);
1719 spin_lock_init(&blp->blp_lock);
1721 CFS_INIT_LIST_HEAD(&blp->blp_list);
1724 for (i = 0; i < LDLM_NUM_THREADS; i++) {
1725 struct ldlm_bl_thread_data bltd = {
1729 init_completion(&blp->blp_comp);
1730 rc = cfs_kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1732 CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
1733 GOTO(out_thread, rc);
1735 wait_for_completion(&blp->blp_comp);
1738 rc = ptlrpc_start_threads(NULL, ldlm_state->ldlm_cancel_service,
1741 GOTO(out_thread, rc);
1743 rc = ptlrpc_start_threads(NULL, ldlm_state->ldlm_cb_service,
1746 GOTO(out_thread, rc);
1748 CFS_INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
1749 expired_lock_thread.elt_state = ELT_STOPPED;
1750 cfs_waitq_init(&expired_lock_thread.elt_waitq);
1752 CFS_INIT_LIST_HEAD(&waiting_locks_list);
1753 spin_lock_init(&waiting_locks_spinlock);
1754 cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0);
1756 rc = cfs_kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FILES);
1758 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
1759 GOTO(out_thread, rc);
1762 wait_event(expired_lock_thread.elt_waitq,
1763 expired_lock_thread.elt_state == ELT_READY);
1770 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1771 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1776 ldlm_proc_cleanup();
1779 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1784 static int ldlm_cleanup(int force)
1787 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1791 if (!list_empty(&ldlm_namespace_list)) {
1792 CERROR("ldlm still has namespaces; clean these up first.\n");
1793 ldlm_dump_all_namespaces(D_DLMTRACE);
1798 while (atomic_read(&blp->blp_num_threads) > 0) {
1799 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
1801 init_completion(&blp->blp_comp);
1803 spin_lock(&blp->blp_lock);
1804 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
1805 cfs_waitq_signal(&blp->blp_waitq);
1806 spin_unlock(&blp->blp_lock);
1808 wait_for_completion(&blp->blp_comp);
1810 OBD_FREE(blp, sizeof(*blp));
1812 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1813 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1814 ldlm_proc_cleanup();
1816 expired_lock_thread.elt_state = ELT_TERMINATE;
1817 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
1818 wait_event(expired_lock_thread.elt_waitq,
1819 expired_lock_thread.elt_state == ELT_STOPPED);
1821 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1822 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1825 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1831 int __init ldlm_init(void)
1833 init_mutex(&ldlm_ref_sem);
1834 init_mutex(&ldlm_namespace_lock);
1835 ldlm_resource_slab = cfs_mem_cache_create("ldlm_resources",
1836 sizeof(struct ldlm_resource), 0,
1837 SLAB_HWCACHE_ALIGN);
1838 if (ldlm_resource_slab == NULL)
1841 ldlm_lock_slab = cfs_mem_cache_create("ldlm_locks",
1842 sizeof(struct ldlm_lock), 0,
1843 SLAB_HWCACHE_ALIGN);
1844 if (ldlm_lock_slab == NULL) {
1845 cfs_mem_cache_destroy(ldlm_resource_slab);
1852 void __exit ldlm_exit(void)
1856 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
1857 rc = cfs_mem_cache_destroy(ldlm_resource_slab);
1858 LASSERTF(rc == 0, "couldn't free ldlm resource slab\n");
1859 rc = cfs_mem_cache_destroy(ldlm_lock_slab);
1860 LASSERTF(rc == 0, "couldn't free ldlm lock slab\n");
1864 EXPORT_SYMBOL(ldlm_extent_shift_kms);
1867 EXPORT_SYMBOL(ldlm_get_processing_policy);
1868 EXPORT_SYMBOL(ldlm_lock2desc);
1869 EXPORT_SYMBOL(ldlm_register_intent);
1870 EXPORT_SYMBOL(ldlm_lockname);
1871 EXPORT_SYMBOL(ldlm_typename);
1872 EXPORT_SYMBOL(ldlm_lock2handle);
1873 EXPORT_SYMBOL(__ldlm_handle2lock);
1874 EXPORT_SYMBOL(ldlm_lock_get);
1875 EXPORT_SYMBOL(ldlm_lock_put);
1876 EXPORT_SYMBOL(ldlm_lock_match);
1877 EXPORT_SYMBOL(ldlm_lock_cancel);
1878 EXPORT_SYMBOL(ldlm_lock_addref);
1879 EXPORT_SYMBOL(ldlm_lock_decref);
1880 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
1881 EXPORT_SYMBOL(ldlm_lock_change_resource);
1882 EXPORT_SYMBOL(ldlm_lock_set_data);
1883 EXPORT_SYMBOL(ldlm_it2str);
1884 EXPORT_SYMBOL(ldlm_lock_dump);
1885 EXPORT_SYMBOL(ldlm_lock_dump_handle);
1886 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
1887 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
1888 EXPORT_SYMBOL(ldlm_lock_allow_match);
1890 /* ldlm_request.c */
1891 EXPORT_SYMBOL(ldlm_completion_ast);
1892 EXPORT_SYMBOL(ldlm_blocking_ast);
1893 EXPORT_SYMBOL(ldlm_glimpse_ast);
1894 EXPORT_SYMBOL(ldlm_expired_completion_wait);
1895 EXPORT_SYMBOL(ldlm_cli_convert);
1896 EXPORT_SYMBOL(ldlm_cli_enqueue);
1897 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
1898 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
1899 EXPORT_SYMBOL(ldlm_cli_cancel);
1900 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1901 EXPORT_SYMBOL(ldlm_cli_join_lru);
1902 EXPORT_SYMBOL(ldlm_replay_locks);
1903 EXPORT_SYMBOL(ldlm_resource_foreach);
1904 EXPORT_SYMBOL(ldlm_namespace_foreach);
1905 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
1906 EXPORT_SYMBOL(ldlm_resource_iterate);
1909 EXPORT_SYMBOL(ldlm_server_blocking_ast);
1910 EXPORT_SYMBOL(ldlm_server_completion_ast);
1911 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1912 EXPORT_SYMBOL(ldlm_handle_enqueue);
1913 EXPORT_SYMBOL(ldlm_handle_enqueue0);
1914 EXPORT_SYMBOL(ldlm_handle_cancel);
1915 EXPORT_SYMBOL(ldlm_handle_convert);
1916 EXPORT_SYMBOL(ldlm_handle_convert0);
1917 EXPORT_SYMBOL(ldlm_del_waiting_lock);
1918 EXPORT_SYMBOL(ldlm_get_ref);
1919 EXPORT_SYMBOL(ldlm_put_ref);
1920 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
1921 EXPORT_SYMBOL(ldlm_revoke_export_locks);
1923 /* ldlm_resource.c */
1924 EXPORT_SYMBOL(ldlm_namespace_new);
1925 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1926 EXPORT_SYMBOL(ldlm_namespace_free);
1927 EXPORT_SYMBOL(ldlm_namespace_dump);
1928 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1929 EXPORT_SYMBOL(ldlm_resource_get);
1930 EXPORT_SYMBOL(ldlm_resource_putref);
1931 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1934 EXPORT_SYMBOL(client_import_add_conn);
1935 EXPORT_SYMBOL(client_import_del_conn);
1936 EXPORT_SYMBOL(client_obd_setup);
1937 EXPORT_SYMBOL(client_obd_cleanup);
1938 EXPORT_SYMBOL(client_connect_import);
1939 EXPORT_SYMBOL(client_disconnect_export);
1940 EXPORT_SYMBOL(target_start_recovery_thread);
1941 EXPORT_SYMBOL(target_stop_recovery_thread);
1942 EXPORT_SYMBOL(target_handle_connect);
1943 EXPORT_SYMBOL(target_cleanup_recovery);
1944 EXPORT_SYMBOL(target_destroy_export);
1945 EXPORT_SYMBOL(target_cancel_recovery_timer);
1946 EXPORT_SYMBOL(target_send_reply);
1947 EXPORT_SYMBOL(target_queue_recovery_request);
1948 EXPORT_SYMBOL(target_handle_ping);
1949 EXPORT_SYMBOL(target_handle_disconnect);
1952 EXPORT_SYMBOL(lock_res_and_lock);
1953 EXPORT_SYMBOL(unlock_res_and_lock);