1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ldlm/ldlm_lockd.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Phil Schwan <phil@clusterfs.com>
43 # define EXPORT_SYMTAB
45 #define DEBUG_SUBSYSTEM S_LDLM
48 # include <libcfs/libcfs.h>
50 # include <liblustre.h>
53 #include <lustre_dlm.h>
54 #include <obd_class.h>
55 #include <libcfs/list.h>
56 #include "ldlm_internal.h"
59 static int ldlm_num_threads;
60 CFS_MODULE_PARM(ldlm_num_threads, "i", int, 0444,
61 "number of DLM service threads to start");
64 extern cfs_mem_cache_t *ldlm_resource_slab;
65 extern cfs_mem_cache_t *ldlm_lock_slab;
66 static cfs_semaphore_t ldlm_ref_sem;
67 static int ldlm_refcount;
71 static struct ldlm_state *ldlm_state;
73 inline cfs_time_t round_timeout(cfs_time_t timeout)
75 return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1);
78 /* timeout for initial callback (AST) reply (bz10399) */
79 static inline unsigned int ldlm_get_rq_timeout(void)
82 unsigned int timeout = min(ldlm_timeout, obd_timeout / 3);
84 return timeout < 1 ? 1 : timeout;
88 /* w_l_spinlock protects both waiting_locks_list and expired_lock_thread */
89 static cfs_spinlock_t waiting_locks_spinlock; /* BH lock (timer) */
90 static cfs_list_t waiting_locks_list;
91 static cfs_timer_t waiting_locks_timer;
93 static struct expired_lock_thread {
94 cfs_waitq_t elt_waitq;
97 cfs_list_t elt_expired_locks;
98 } expired_lock_thread;
101 #define ELT_STOPPED 0
103 #define ELT_TERMINATE 2
105 struct ldlm_bl_pool {
106 cfs_spinlock_t blp_lock;
109 * blp_prio_list is used for callbacks that should be handled
110 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
113 cfs_list_t blp_prio_list;
116 * blp_list is used for all other callbacks which are likely
117 * to take longer to process.
121 cfs_waitq_t blp_waitq;
122 cfs_completion_t blp_comp;
123 cfs_atomic_t blp_num_threads;
124 cfs_atomic_t blp_busy_threads;
129 struct ldlm_bl_work_item {
130 cfs_list_t blwi_entry;
131 struct ldlm_namespace *blwi_ns;
132 struct ldlm_lock_desc blwi_ld;
133 struct ldlm_lock *blwi_lock;
134 cfs_list_t blwi_head;
136 cfs_completion_t blwi_comp;
138 int blwi_mem_pressure;
143 static inline int have_expired_locks(void)
148 cfs_spin_lock_bh(&waiting_locks_spinlock);
149 need_to_run = !cfs_list_empty(&expired_lock_thread.elt_expired_locks);
150 cfs_spin_unlock_bh(&waiting_locks_spinlock);
155 static int expired_lock_main(void *arg)
157 cfs_list_t *expired = &expired_lock_thread.elt_expired_locks;
158 struct l_wait_info lwi = { 0 };
162 cfs_daemonize("ldlm_elt");
164 expired_lock_thread.elt_state = ELT_READY;
165 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
168 l_wait_event(expired_lock_thread.elt_waitq,
169 have_expired_locks() ||
170 expired_lock_thread.elt_state == ELT_TERMINATE,
173 cfs_spin_lock_bh(&waiting_locks_spinlock);
174 if (expired_lock_thread.elt_dump) {
175 cfs_spin_unlock_bh(&waiting_locks_spinlock);
177 /* from waiting_locks_callback, but not in timer */
178 libcfs_debug_dumplog();
179 libcfs_run_lbug_upcall(__FILE__,
180 "waiting_locks_callback",
181 expired_lock_thread.elt_dump);
183 cfs_spin_lock_bh(&waiting_locks_spinlock);
184 expired_lock_thread.elt_dump = 0;
189 while (!cfs_list_empty(expired)) {
190 struct obd_export *export;
191 struct ldlm_lock *lock;
193 lock = cfs_list_entry(expired->next, struct ldlm_lock,
195 if ((void *)lock < LP_POISON + CFS_PAGE_SIZE &&
196 (void *)lock >= LP_POISON) {
197 cfs_spin_unlock_bh(&waiting_locks_spinlock);
198 CERROR("free lock on elt list %p\n", lock);
201 cfs_list_del_init(&lock->l_pending_chain);
202 if ((void *)lock->l_export < LP_POISON + CFS_PAGE_SIZE &&
203 (void *)lock->l_export >= LP_POISON) {
204 CERROR("lock with free export on elt list %p\n",
206 lock->l_export = NULL;
207 LDLM_ERROR(lock, "free export");
208 /* release extra ref grabbed by
209 * ldlm_add_waiting_lock() or
210 * ldlm_failed_ast() */
211 LDLM_LOCK_RELEASE(lock);
214 export = class_export_lock_get(lock->l_export, lock);
215 cfs_spin_unlock_bh(&waiting_locks_spinlock);
218 class_fail_export(export);
219 class_export_lock_put(export, lock);
221 /* release extra ref grabbed by ldlm_add_waiting_lock()
222 * or ldlm_failed_ast() */
223 LDLM_LOCK_RELEASE(lock);
225 cfs_spin_lock_bh(&waiting_locks_spinlock);
227 cfs_spin_unlock_bh(&waiting_locks_spinlock);
229 if (do_dump && obd_dump_on_eviction) {
230 CERROR("dump the log upon eviction\n");
231 libcfs_debug_dumplog();
234 if (expired_lock_thread.elt_state == ELT_TERMINATE)
238 expired_lock_thread.elt_state = ELT_STOPPED;
239 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
243 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
246 * Check if there is a request in the export request list
247 * which prevents the lock canceling.
249 static int ldlm_lock_busy(struct ldlm_lock *lock)
251 struct ptlrpc_request *req;
255 if (lock->l_export == NULL)
258 cfs_spin_lock_bh(&lock->l_export->exp_rpc_lock);
259 cfs_list_for_each_entry(req, &lock->l_export->exp_queued_rpc,
261 if (req->rq_ops->hpreq_lock_match) {
262 match = req->rq_ops->hpreq_lock_match(req, lock);
267 cfs_spin_unlock_bh(&lock->l_export->exp_rpc_lock);
271 /* This is called from within a timer interrupt and cannot schedule */
272 static void waiting_locks_callback(unsigned long unused)
274 struct ldlm_lock *lock, *last = NULL;
277 cfs_spin_lock_bh(&waiting_locks_spinlock);
278 while (!cfs_list_empty(&waiting_locks_list)) {
279 lock = cfs_list_entry(waiting_locks_list.next, struct ldlm_lock,
281 if (cfs_time_after(lock->l_callback_timeout,
282 cfs_time_current()) ||
283 (lock->l_req_mode == LCK_GROUP))
286 if (ptlrpc_check_suspend()) {
287 /* there is a case when we talk to one mds, holding
288 * lock from another mds. this way we easily can get
289 * here, if second mds is being recovered. so, we
290 * suspend timeouts. bug 6019 */
292 LDLM_ERROR(lock, "recharge timeout: %s@%s nid %s ",
293 lock->l_export->exp_client_uuid.uuid,
294 lock->l_export->exp_connection->c_remote_uuid.uuid,
295 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
297 cfs_list_del_init(&lock->l_pending_chain);
298 cfs_spin_unlock_bh(&waiting_locks_spinlock);
299 ldlm_add_waiting_lock(lock);
303 /* if timeout overlaps the activation time of suspended timeouts
304 * then extend it to give a chance for client to reconnect */
305 if (cfs_time_before(cfs_time_sub(lock->l_callback_timeout,
306 cfs_time_seconds(obd_timeout)/2),
307 ptlrpc_suspend_wakeup_time())) {
308 LDLM_ERROR(lock, "extend timeout due to recovery: %s@%s nid %s ",
309 lock->l_export->exp_client_uuid.uuid,
310 lock->l_export->exp_connection->c_remote_uuid.uuid,
311 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
313 cfs_list_del_init(&lock->l_pending_chain);
314 cfs_spin_unlock_bh(&waiting_locks_spinlock);
315 ldlm_add_waiting_lock(lock);
319 /* Check if we need to prolong timeout */
320 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
321 ldlm_lock_busy(lock)) {
324 if (lock->l_pending_chain.next == &waiting_locks_list)
329 cfs_spin_unlock_bh(&waiting_locks_spinlock);
330 LDLM_DEBUG(lock, "prolong the busy lock");
331 ldlm_refresh_waiting_lock(lock,
332 ldlm_get_enq_timeout(lock));
333 cfs_spin_lock_bh(&waiting_locks_spinlock);
336 LDLM_LOCK_RELEASE(lock);
340 LDLM_LOCK_RELEASE(lock);
343 lock->l_resource->lr_namespace->ns_timeouts++;
344 LDLM_ERROR(lock, "lock callback timer expired after %lds: "
345 "evicting client at %s ",
346 cfs_time_current_sec()- lock->l_last_activity,
348 lock->l_export->exp_connection->c_peer.nid));
352 /* no needs to take an extra ref on the lock since it was in
353 * the waiting_locks_list and ldlm_add_waiting_lock()
354 * already grabbed a ref */
355 cfs_list_del(&lock->l_pending_chain);
356 cfs_list_add(&lock->l_pending_chain,
357 &expired_lock_thread.elt_expired_locks);
360 if (!cfs_list_empty(&expired_lock_thread.elt_expired_locks)) {
361 if (obd_dump_on_timeout)
362 expired_lock_thread.elt_dump = __LINE__;
364 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
368 * Make sure the timer will fire again if we have any locks
371 if (!cfs_list_empty(&waiting_locks_list)) {
372 cfs_time_t timeout_rounded;
373 lock = cfs_list_entry(waiting_locks_list.next, struct ldlm_lock,
375 timeout_rounded = (cfs_time_t)round_timeout(lock->l_callback_timeout);
376 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
378 cfs_spin_unlock_bh(&waiting_locks_spinlock);
382 * Indicate that we're waiting for a client to call us back cancelling a given
383 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
384 * timer to fire appropriately. (We round up to the next second, to avoid
385 * floods of timer firings during periods of high lock contention and traffic).
386 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
387 * if it has been added to the waiting list (1 is returned).
389 * Called with the namespace lock held.
391 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds)
394 cfs_time_t timeout_rounded;
396 if (!cfs_list_empty(&lock->l_pending_chain))
399 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
400 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
403 timeout = cfs_time_shift(seconds);
404 if (likely(cfs_time_after(timeout, lock->l_callback_timeout)))
405 lock->l_callback_timeout = timeout;
407 timeout_rounded = round_timeout(lock->l_callback_timeout);
409 if (cfs_time_before(timeout_rounded,
410 cfs_timer_deadline(&waiting_locks_timer)) ||
411 !cfs_timer_is_armed(&waiting_locks_timer)) {
412 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
414 /* if the new lock has a shorter timeout than something earlier on
415 the list, we'll wait the longer amount of time; no big deal. */
417 cfs_list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
421 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
424 int timeout = ldlm_get_enq_timeout(lock);
426 LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
428 cfs_spin_lock_bh(&waiting_locks_spinlock);
429 if (lock->l_destroyed) {
430 static cfs_time_t next;
431 cfs_spin_unlock_bh(&waiting_locks_spinlock);
432 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
433 if (cfs_time_after(cfs_time_current(), next)) {
434 next = cfs_time_shift(14400);
435 libcfs_debug_dumpstack(NULL);
440 ret = __ldlm_add_waiting_lock(lock, timeout);
442 /* grab ref on the lock if it has been added to the
445 cfs_spin_unlock_bh(&waiting_locks_spinlock);
447 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)",
448 ret == 0 ? "not re-" : "", timeout,
449 AT_OFF ? "off" : "on");
454 * Remove a lock from the pending list, likely because it had its cancellation
455 * callback arrive without incident. This adjusts the lock-timeout timer if
456 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
457 * As done by ldlm_del_waiting_lock(), the caller must release the lock
458 * reference when the lock is removed from any list (1 is returned).
460 * Called with namespace lock held.
462 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
464 cfs_list_t *list_next;
466 if (cfs_list_empty(&lock->l_pending_chain))
469 list_next = lock->l_pending_chain.next;
470 if (lock->l_pending_chain.prev == &waiting_locks_list) {
471 /* Removing the head of the list, adjust timer. */
472 if (list_next == &waiting_locks_list) {
473 /* No more, just cancel. */
474 cfs_timer_disarm(&waiting_locks_timer);
476 struct ldlm_lock *next;
477 next = cfs_list_entry(list_next, struct ldlm_lock,
479 cfs_timer_arm(&waiting_locks_timer,
480 round_timeout(next->l_callback_timeout));
483 cfs_list_del_init(&lock->l_pending_chain);
488 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
492 if (lock->l_export == NULL) {
493 /* We don't have a "waiting locks list" on clients. */
494 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
498 cfs_spin_lock_bh(&waiting_locks_spinlock);
499 ret = __ldlm_del_waiting_lock(lock);
500 cfs_spin_unlock_bh(&waiting_locks_spinlock);
502 /* release lock ref if it has indeed been removed
504 LDLM_LOCK_RELEASE(lock);
506 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
513 * Called with namespace lock held.
515 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
517 if (lock->l_export == NULL) {
518 /* We don't have a "waiting locks list" on clients. */
519 LDLM_DEBUG(lock, "client lock: no-op");
523 cfs_spin_lock_bh(&waiting_locks_spinlock);
525 if (cfs_list_empty(&lock->l_pending_chain)) {
526 cfs_spin_unlock_bh(&waiting_locks_spinlock);
527 LDLM_DEBUG(lock, "wasn't waiting");
531 /* we remove/add the lock to the waiting list, so no needs to
532 * release/take a lock reference */
533 __ldlm_del_waiting_lock(lock);
534 __ldlm_add_waiting_lock(lock, timeout);
535 cfs_spin_unlock_bh(&waiting_locks_spinlock);
537 LDLM_DEBUG(lock, "refreshed");
540 #else /* !__KERNEL__ */
542 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
544 LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
548 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
553 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
557 #endif /* __KERNEL__ */
559 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
560 const char *ast_type)
562 LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
563 "to a lock %s callback time out: rc %d\n",
564 lock->l_export->exp_obd->obd_name,
565 obd_export_nid2str(lock->l_export), ast_type, rc);
567 if (obd_dump_on_timeout)
568 libcfs_debug_dumplog();
570 cfs_spin_lock_bh(&waiting_locks_spinlock);
571 if (__ldlm_del_waiting_lock(lock) == 0)
572 /* the lock was not in any list, grab an extra ref before adding
573 * the lock to the expired list */
575 cfs_list_add(&lock->l_pending_chain,
576 &expired_lock_thread.elt_expired_locks);
577 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
578 cfs_spin_unlock_bh(&waiting_locks_spinlock);
580 class_fail_export(lock->l_export);
584 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
585 struct ptlrpc_request *req, int rc,
586 const char *ast_type)
588 lnet_process_id_t peer = req->rq_import->imp_connection->c_peer;
590 if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
591 LASSERT(lock->l_export);
592 if (lock->l_export->exp_libclient) {
593 LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
594 " timeout, just cancelling lock", ast_type,
595 libcfs_nid2str(peer.nid));
596 ldlm_lock_cancel(lock);
598 } else if (lock->l_flags & LDLM_FL_CANCEL) {
599 LDLM_DEBUG(lock, "%s AST timeout from nid %s, but "
600 "cancel was received (AST reply lost?)",
601 ast_type, libcfs_nid2str(peer.nid));
602 ldlm_lock_cancel(lock);
605 ldlm_del_waiting_lock(lock);
606 ldlm_failed_ast(lock, rc, ast_type);
610 LDLM_DEBUG(lock, "client (nid %s) returned %d"
611 " from %s AST - normal race",
612 libcfs_nid2str(peer.nid),
614 lustre_msg_get_status(req->rq_repmsg) : -1,
617 LDLM_ERROR(lock, "client (nid %s) returned %d "
618 "from %s AST", libcfs_nid2str(peer.nid),
619 (req->rq_repmsg != NULL) ?
620 lustre_msg_get_status(req->rq_repmsg) : 0,
622 ldlm_lock_cancel(lock);
623 /* Server-side AST functions are called from ldlm_reprocess_all,
624 * which needs to be told to please restart its reprocessing. */
631 static int ldlm_cb_interpret(const struct lu_env *env,
632 struct ptlrpc_request *req, void *data, int rc)
634 struct ldlm_cb_set_arg *arg;
635 struct ldlm_lock *lock;
638 LASSERT(data != NULL);
640 arg = req->rq_async_args.pointer_arg[0];
641 lock = req->rq_async_args.pointer_arg[1];
642 LASSERT(lock != NULL);
644 /* If client canceled the lock but the cancel has not
645 * been received yet, we need to update lvbo to have the
646 * proper attributes cached. */
647 if (rc == -EINVAL && arg->type == LDLM_BL_CALLBACK)
648 ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
649 rc = ldlm_handle_ast_error(lock, req, rc,
650 arg->type == LDLM_BL_CALLBACK
651 ? "blocking" : "completion");
654 LDLM_LOCK_RELEASE(lock);
657 cfs_atomic_set(&arg->restart, 1);
662 static inline int ldlm_bl_and_cp_ast_fini(struct ptlrpc_request *req,
663 struct ldlm_cb_set_arg *arg,
664 struct ldlm_lock *lock,
670 if (unlikely(instant_cancel)) {
671 rc = ptl_send_rpc(req, 1);
672 ptlrpc_req_finished(req);
674 /* If we cancelled the lock, we need to restart
675 * ldlm_reprocess_queue */
676 cfs_atomic_set(&arg->restart, 1);
679 ptlrpc_set_add_req(arg->set, req);
686 * Check if there are requests in the export request list which prevent
687 * the lock canceling and make these requests high priority ones.
689 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
691 struct ptlrpc_request *req;
694 if (lock->l_export == NULL) {
695 LDLM_DEBUG(lock, "client lock: no-op");
699 cfs_spin_lock_bh(&lock->l_export->exp_rpc_lock);
700 cfs_list_for_each_entry(req, &lock->l_export->exp_queued_rpc,
702 if (!req->rq_hp && req->rq_ops->hpreq_lock_match &&
703 req->rq_ops->hpreq_lock_match(req, lock))
704 ptlrpc_hpreq_reorder(req);
706 cfs_spin_unlock_bh(&lock->l_export->exp_rpc_lock);
711 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
712 * enqueued server lock conflicts with given one.
714 * Sends blocking ast rpc to the client owning that lock; arms timeout timer
715 * to wait for client response.
717 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
718 struct ldlm_lock_desc *desc,
719 void *data, int flag)
721 struct ldlm_cb_set_arg *arg = data;
722 struct ldlm_request *body;
723 struct ptlrpc_request *req;
724 int instant_cancel = 0;
728 if (flag == LDLM_CB_CANCELING)
729 /* Don't need to do anything here. */
733 LASSERT(data != NULL);
734 if (lock->l_export->exp_obd->obd_recovering != 0) {
735 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
736 ldlm_lock_dump(D_ERROR, lock, 0);
739 ldlm_lock_reorder_req(lock);
741 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
742 &RQF_LDLM_BL_CALLBACK,
743 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
747 req->rq_async_args.pointer_arg[0] = arg;
748 req->rq_async_args.pointer_arg[1] = lock;
749 req->rq_interpret_reply = ldlm_cb_interpret;
750 req->rq_no_resend = 1;
752 lock_res(lock->l_resource);
753 if (lock->l_granted_mode != lock->l_req_mode) {
754 /* this blocking AST will be communicated as part of the
755 * completion AST instead */
756 unlock_res(lock->l_resource);
757 ptlrpc_req_finished(req);
758 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
762 if (lock->l_destroyed) {
763 /* What's the point? */
764 unlock_res(lock->l_resource);
765 ptlrpc_req_finished(req);
769 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
772 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
773 body->lock_handle[0] = lock->l_remote_handle;
774 body->lock_desc = *desc;
775 body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
777 LDLM_DEBUG(lock, "server preparing blocking AST");
779 ptlrpc_request_set_replen(req);
780 if (instant_cancel) {
781 unlock_res(lock->l_resource);
782 ldlm_lock_cancel(lock);
784 LASSERT(lock->l_granted_mode == lock->l_req_mode);
785 ldlm_add_waiting_lock(lock);
786 unlock_res(lock->l_resource);
789 req->rq_send_state = LUSTRE_IMP_FULL;
790 /* ptlrpc_request_alloc_pack already set timeout */
792 req->rq_timeout = ldlm_get_rq_timeout();
794 if (lock->l_export && lock->l_export->exp_nid_stats &&
795 lock->l_export->exp_nid_stats->nid_ldlm_stats)
796 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
797 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
799 rc = ldlm_bl_and_cp_ast_fini(req, arg, lock, instant_cancel);
804 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
806 struct ldlm_cb_set_arg *arg = data;
807 struct ldlm_request *body;
808 struct ptlrpc_request *req;
809 long total_enqueue_wait;
810 int instant_cancel = 0;
814 LASSERT(lock != NULL);
815 LASSERT(data != NULL);
817 total_enqueue_wait = cfs_time_sub(cfs_time_current_sec(),
818 lock->l_last_activity);
820 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
821 &RQF_LDLM_CP_CALLBACK);
825 lock_res_and_lock(lock);
826 if (lock->l_resource->lr_lvb_len)
827 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT,
828 lock->l_resource->lr_lvb_len);
829 unlock_res_and_lock(lock);
831 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
833 ptlrpc_request_free(req);
837 req->rq_async_args.pointer_arg[0] = arg;
838 req->rq_async_args.pointer_arg[1] = lock;
839 req->rq_interpret_reply = ldlm_cb_interpret;
840 req->rq_no_resend = 1;
841 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
843 body->lock_handle[0] = lock->l_remote_handle;
844 body->lock_flags = flags;
845 ldlm_lock2desc(lock, &body->lock_desc);
846 if (lock->l_resource->lr_lvb_len) {
847 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
849 lock_res_and_lock(lock);
850 memcpy(lvb, lock->l_resource->lr_lvb_data,
851 lock->l_resource->lr_lvb_len);
852 unlock_res_and_lock(lock);
855 LDLM_DEBUG(lock, "server preparing completion AST (after %lds wait)",
858 /* Server-side enqueue wait time estimate, used in
859 __ldlm_add_waiting_lock to set future enqueue timers */
860 if (total_enqueue_wait < ldlm_get_enq_timeout(lock))
861 at_measured(&lock->l_resource->lr_namespace->ns_at_estimate,
864 /* bz18618. Don't add lock enqueue time we spend waiting for a
865 previous callback to fail. Locks waiting legitimately will
866 get extended by ldlm_refresh_waiting_lock regardless of the
867 estimate, so it's okay to underestimate here. */
868 LDLM_DEBUG(lock, "lock completed after %lus; estimate was %ds. "
869 "It is likely that a previous callback timed out.",
871 at_get(&lock->l_resource->lr_namespace->ns_at_estimate));
873 ptlrpc_request_set_replen(req);
875 req->rq_send_state = LUSTRE_IMP_FULL;
876 /* ptlrpc_request_pack already set timeout */
878 req->rq_timeout = ldlm_get_rq_timeout();
880 /* We only send real blocking ASTs after the lock is granted */
881 lock_res_and_lock(lock);
882 if (lock->l_flags & LDLM_FL_AST_SENT) {
883 body->lock_flags |= LDLM_FL_AST_SENT;
884 /* copy ast flags like LDLM_FL_DISCARD_DATA */
885 body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
887 /* We might get here prior to ldlm_handle_enqueue setting
888 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
889 * into waiting list, but this is safe and similar code in
890 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
891 * that would not only cancel the lock, but will also remove
892 * it from waiting list */
893 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
894 unlock_res_and_lock(lock);
895 ldlm_lock_cancel(lock);
897 lock_res_and_lock(lock);
899 /* start the lock-timeout clock */
900 ldlm_add_waiting_lock(lock);
903 unlock_res_and_lock(lock);
905 if (lock->l_export && lock->l_export->exp_nid_stats &&
906 lock->l_export->exp_nid_stats->nid_ldlm_stats)
907 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
908 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
910 rc = ldlm_bl_and_cp_ast_fini(req, arg, lock, instant_cancel);
915 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
917 struct ldlm_resource *res = lock->l_resource;
918 struct ldlm_request *body;
919 struct ptlrpc_request *req;
923 LASSERT(lock != NULL);
925 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
926 &RQF_LDLM_GL_CALLBACK,
927 LUSTRE_DLM_VERSION, LDLM_GL_CALLBACK);
932 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
933 body->lock_handle[0] = lock->l_remote_handle;
934 ldlm_lock2desc(lock, &body->lock_desc);
936 lock_res_and_lock(lock);
937 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
938 lock->l_resource->lr_lvb_len);
939 unlock_res_and_lock(lock);
940 res = lock->l_resource;
941 ptlrpc_request_set_replen(req);
944 req->rq_send_state = LUSTRE_IMP_FULL;
945 /* ptlrpc_request_alloc_pack already set timeout */
947 req->rq_timeout = ldlm_get_rq_timeout();
949 if (lock->l_export && lock->l_export->exp_nid_stats &&
950 lock->l_export->exp_nid_stats->nid_ldlm_stats)
951 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
952 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
954 rc = ptlrpc_queue_wait(req);
955 if (rc == -ELDLM_NO_LOCK_DATA)
956 LDLM_DEBUG(lock, "lost race - client has a lock but no inode");
958 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
960 rc = ldlm_res_lvbo_update(res, req, 1);
962 ptlrpc_req_finished(req);
964 ldlm_reprocess_all(res);
970 extern unsigned long long lu_time_stamp_get(void);
972 #define lu_time_stamp_get() time(NULL)
975 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
976 struct lprocfs_stats *srv_stats)
978 int lock_type = 0, op = 0;
980 lock_type = dlm_req->lock_desc.l_resource.lr_type;
984 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
987 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
988 op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
990 op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
993 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
996 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
1004 lprocfs_counter_incr(srv_stats, op);
1010 * Main server-side entry point into LDLM. This is called by ptlrpc service
1011 * threads to carry out client lock enqueueing requests.
1013 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1014 struct ptlrpc_request *req,
1015 const struct ldlm_request *dlm_req,
1016 const struct ldlm_callback_suite *cbs)
1018 struct ldlm_reply *dlm_rep;
1020 ldlm_error_t err = ELDLM_OK;
1021 struct ldlm_lock *lock = NULL;
1022 void *cookie = NULL;
1026 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1028 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF);
1029 flags = dlm_req->lock_flags;
1031 LASSERT(req->rq_export);
1033 if (req->rq_rqbd->rqbd_service->srv_stats)
1034 ldlm_svc_get_eopc(dlm_req,
1035 req->rq_rqbd->rqbd_service->srv_stats);
1037 if (req->rq_export && req->rq_export->exp_nid_stats &&
1038 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1039 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1040 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1042 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1043 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1044 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1045 dlm_req->lock_desc.l_resource.lr_type);
1046 GOTO(out, rc = -EFAULT);
1049 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1050 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1051 dlm_req->lock_desc.l_req_mode &
1052 (dlm_req->lock_desc.l_req_mode-1))) {
1053 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1054 dlm_req->lock_desc.l_req_mode);
1055 GOTO(out, rc = -EFAULT);
1058 if (req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) {
1059 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1061 DEBUG_REQ(D_ERROR, req,
1062 "PLAIN lock request from IBITS client?");
1063 GOTO(out, rc = -EPROTO);
1065 } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1067 DEBUG_REQ(D_ERROR, req,
1068 "IBITS lock request from unaware client?");
1069 GOTO(out, rc = -EPROTO);
1073 /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
1074 against server's _CONNECT_SUPPORTED flags? (I don't want to use
1075 ibits for mgc/mgs) */
1077 /* INODEBITS_INTEROP: Perform conversion from plain lock to
1078 * inodebits lock if client does not support them. */
1079 if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) &&
1080 (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
1081 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
1082 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
1083 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1084 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
1085 dlm_req->lock_desc.l_req_mode = LCK_CR;
1089 if (unlikely(flags & LDLM_FL_REPLAY)) {
1090 /* Find an existing lock in the per-export lock hash */
1091 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1092 (void *)&dlm_req->lock_handle[0]);
1094 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie "
1095 LPX64, lock->l_handle.h_cookie);
1096 GOTO(existing_lock, rc = 0);
1100 /* The lock's callback data might be set in the policy function */
1101 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1102 dlm_req->lock_desc.l_resource.lr_type,
1103 dlm_req->lock_desc.l_req_mode,
1107 GOTO(out, rc = -ENOMEM);
1109 lock->l_last_activity = cfs_time_current_sec();
1110 lock->l_remote_handle = dlm_req->lock_handle[0];
1111 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1113 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1114 /* Don't enqueue a lock onto the export if it has already
1115 * been evicted. Cancel it now instead. (bug 3822) */
1116 if (req->rq_export->exp_failed) {
1117 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1118 GOTO(out, rc = -ENOTCONN);
1121 lock->l_export = class_export_lock_get(req->rq_export, lock);
1122 if (lock->l_export->exp_lock_hash)
1123 cfs_hash_add(lock->l_export->exp_lock_hash,
1124 &lock->l_remote_handle,
1129 if (flags & LDLM_FL_HAS_INTENT) {
1130 /* In this case, the reply buffer is allocated deep in
1131 * local_lock_enqueue by the policy function. */
1134 lock_res_and_lock(lock);
1135 if (lock->l_resource->lr_lvb_len) {
1136 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1138 lock->l_resource->lr_lvb_len);
1140 unlock_res_and_lock(lock);
1142 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1143 GOTO(out, rc = -ENOMEM);
1145 rc = req_capsule_server_pack(&req->rq_pill);
1150 if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
1151 lock->l_policy_data = dlm_req->lock_desc.l_policy_data;
1152 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1153 lock->l_req_extent = lock->l_policy_data.l_extent;
1155 err = ldlm_lock_enqueue(ns, &lock, cookie, (int *)&flags);
1159 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1160 dlm_rep->lock_flags = flags;
1162 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1163 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1165 /* We never send a blocking AST until the lock is granted, but
1166 * we can tell it right now */
1167 lock_res_and_lock(lock);
1169 /* Now take into account flags to be inherited from original lock
1170 request both in reply to client and in our own lock flags. */
1171 dlm_rep->lock_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1172 lock->l_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1174 /* Don't move a pending lock onto the export if it has already
1175 * been evicted. Cancel it now instead. (bug 5683) */
1176 if (unlikely(req->rq_export->exp_failed ||
1177 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1178 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1180 } else if (lock->l_flags & LDLM_FL_AST_SENT) {
1181 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
1182 if (lock->l_granted_mode == lock->l_req_mode) {
1184 * Only cancel lock if it was granted, because it would
1185 * be destroyed immediately and would never be granted
1186 * in the future, causing timeouts on client. Not
1187 * granted lock will be cancelled immediately after
1188 * sending completion AST.
1190 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1191 unlock_res_and_lock(lock);
1192 ldlm_lock_cancel(lock);
1193 lock_res_and_lock(lock);
1195 ldlm_add_waiting_lock(lock);
1198 /* Make sure we never ever grant usual metadata locks to liblustre
1200 if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1201 dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1202 req->rq_export->exp_libclient) {
1203 if (unlikely(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) ||
1204 !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1205 CERROR("Granting sync lock to libclient. "
1206 "req fl %d, rep fl %d, lock fl "LPX64"\n",
1207 dlm_req->lock_flags, dlm_rep->lock_flags,
1209 LDLM_ERROR(lock, "sync lock");
1210 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1211 struct ldlm_intent *it;
1213 it = req_capsule_client_get(&req->rq_pill,
1216 CERROR("This is intent %s ("LPU64")\n",
1217 ldlm_it2str(it->opc), it->opc);
1223 unlock_res_and_lock(lock);
1227 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1228 if (!req->rq_packed_final) {
1229 err = lustre_pack_reply(req, 1, NULL, NULL);
1234 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1235 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1237 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1238 "(err=%d, rc=%d)", err, rc);
1240 lock_res_and_lock(lock);
1242 if (lock->l_resource->lr_lvb_len > 0) {
1245 lvb = req_capsule_server_get(&req->rq_pill,
1247 LASSERTF(lvb != NULL, "req %p, lock %p\n",
1250 memcpy(lvb, lock->l_resource->lr_lvb_data,
1251 lock->l_resource->lr_lvb_len);
1254 ldlm_resource_unlink_lock(lock);
1255 ldlm_lock_destroy_nolock(lock);
1257 unlock_res_and_lock(lock);
1259 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1260 ldlm_reprocess_all(lock->l_resource);
1262 LDLM_LOCK_RELEASE(lock);
1265 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1271 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1272 ldlm_completion_callback completion_callback,
1273 ldlm_blocking_callback blocking_callback,
1274 ldlm_glimpse_callback glimpse_callback)
1276 struct ldlm_request *dlm_req;
1277 struct ldlm_callback_suite cbs = {
1278 .lcs_completion = completion_callback,
1279 .lcs_blocking = blocking_callback,
1280 .lcs_glimpse = glimpse_callback
1284 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1285 if (dlm_req != NULL) {
1286 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1287 req, dlm_req, &cbs);
1294 int ldlm_handle_convert0(struct ptlrpc_request *req,
1295 const struct ldlm_request *dlm_req)
1297 struct ldlm_reply *dlm_rep;
1298 struct ldlm_lock *lock;
1302 if (req->rq_export && req->rq_export->exp_nid_stats &&
1303 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1304 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1305 LDLM_CONVERT - LDLM_FIRST_OPC);
1307 rc = req_capsule_server_pack(&req->rq_pill);
1311 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1312 dlm_rep->lock_flags = dlm_req->lock_flags;
1314 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1316 req->rq_status = EINVAL;
1320 LDLM_DEBUG(lock, "server-side convert handler START");
1322 lock->l_last_activity = cfs_time_current_sec();
1323 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1324 &dlm_rep->lock_flags);
1326 if (ldlm_del_waiting_lock(lock))
1327 LDLM_DEBUG(lock, "converted waiting lock");
1330 req->rq_status = EDEADLOCK;
1335 if (!req->rq_status)
1336 ldlm_reprocess_all(lock->l_resource);
1337 LDLM_DEBUG(lock, "server-side convert handler END");
1338 LDLM_LOCK_PUT(lock);
1340 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1345 int ldlm_handle_convert(struct ptlrpc_request *req)
1348 struct ldlm_request *dlm_req;
1350 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1351 if (dlm_req != NULL) {
1352 rc = ldlm_handle_convert0(req, dlm_req);
1354 CERROR ("Can't unpack dlm_req\n");
1360 /* Cancel all the locks whos handles are packed into ldlm_request */
1361 int ldlm_request_cancel(struct ptlrpc_request *req,
1362 const struct ldlm_request *dlm_req, int first)
1364 struct ldlm_resource *res, *pres = NULL;
1365 struct ldlm_lock *lock;
1366 int i, count, done = 0;
1369 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1373 /* There is no lock on the server at the replay time,
1374 * skip lock cancelling to make replay tests to pass. */
1375 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1378 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1379 "starting at %d", count, first);
1381 for (i = first; i < count; i++) {
1382 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1384 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1385 "lock (cookie "LPU64")",
1386 dlm_req->lock_handle[i].cookie);
1390 res = lock->l_resource;
1395 ldlm_reprocess_all(pres);
1396 LDLM_RESOURCE_DELREF(pres);
1397 ldlm_resource_putref(pres);
1400 ldlm_resource_getref(res);
1401 LDLM_RESOURCE_ADDREF(res);
1402 ldlm_res_lvbo_update(res, NULL, 1);
1406 ldlm_lock_cancel(lock);
1407 LDLM_LOCK_PUT(lock);
1410 ldlm_reprocess_all(pres);
1411 LDLM_RESOURCE_DELREF(pres);
1412 ldlm_resource_putref(pres);
1414 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1418 int ldlm_handle_cancel(struct ptlrpc_request *req)
1420 struct ldlm_request *dlm_req;
1424 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1425 if (dlm_req == NULL) {
1426 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1430 if (req->rq_export && req->rq_export->exp_nid_stats &&
1431 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1432 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1433 LDLM_CANCEL - LDLM_FIRST_OPC);
1435 rc = req_capsule_server_pack(&req->rq_pill);
1439 if (!ldlm_request_cancel(req, dlm_req, 0))
1440 req->rq_status = ESTALE;
1442 RETURN(ptlrpc_reply(req));
1445 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1446 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1451 LDLM_DEBUG(lock, "client blocking AST callback handler");
1453 lock_res_and_lock(lock);
1454 lock->l_flags |= LDLM_FL_CBPENDING;
1456 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
1457 lock->l_flags |= LDLM_FL_CANCEL;
1459 do_ast = (!lock->l_readers && !lock->l_writers);
1460 unlock_res_and_lock(lock);
1463 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1464 lock, lock->l_blocking_ast);
1465 if (lock->l_blocking_ast != NULL)
1466 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1469 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1473 LDLM_DEBUG(lock, "client blocking callback handler END");
1474 LDLM_LOCK_RELEASE(lock);
1478 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1479 struct ldlm_namespace *ns,
1480 struct ldlm_request *dlm_req,
1481 struct ldlm_lock *lock)
1483 CFS_LIST_HEAD(ast_list);
1486 LDLM_DEBUG(lock, "client completion callback handler START");
1488 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1489 int to = cfs_time_seconds(1);
1491 cfs_schedule_timeout_and_set_state(
1492 CFS_TASK_INTERRUPTIBLE, to);
1493 if (lock->l_granted_mode == lock->l_req_mode ||
1499 lock_res_and_lock(lock);
1500 if (lock->l_destroyed ||
1501 lock->l_granted_mode == lock->l_req_mode) {
1502 /* bug 11300: the lock has already been granted */
1503 unlock_res_and_lock(lock);
1504 LDLM_DEBUG(lock, "Double grant race happened");
1505 LDLM_LOCK_RELEASE(lock);
1510 /* If we receive the completion AST before the actual enqueue returned,
1511 * then we might need to switch lock modes, resources, or extents. */
1512 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1513 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1514 LDLM_DEBUG(lock, "completion AST, new lock mode");
1517 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1518 lock->l_policy_data = dlm_req->lock_desc.l_policy_data;
1519 LDLM_DEBUG(lock, "completion AST, new policy data");
1522 ldlm_resource_unlink_lock(lock);
1523 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1524 &lock->l_resource->lr_name,
1525 sizeof(lock->l_resource->lr_name)) != 0) {
1526 unlock_res_and_lock(lock);
1527 if (ldlm_lock_change_resource(ns, lock,
1528 &dlm_req->lock_desc.l_resource.lr_name) != 0) {
1529 LDLM_ERROR(lock, "Failed to allocate resource");
1530 LDLM_LOCK_RELEASE(lock);
1534 LDLM_DEBUG(lock, "completion AST, new resource");
1535 CERROR("change resource!\n");
1536 lock_res_and_lock(lock);
1539 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1540 /* BL_AST locks are not needed in lru.
1541 * let ldlm_cancel_lru() be fast. */
1542 ldlm_lock_remove_from_lru(lock);
1543 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1544 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1547 if (lock->l_lvb_len) {
1548 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
1549 RCL_CLIENT) < lock->l_lvb_len) {
1550 LDLM_ERROR(lock, "completion AST did not contain "
1553 void *lvb = req_capsule_client_get(&req->rq_pill,
1555 memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
1559 ldlm_grant_lock(lock, &ast_list);
1560 unlock_res_and_lock(lock);
1562 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1564 /* Let Enqueue to call osc_lock_upcall() and initialize
1566 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1568 ldlm_run_ast_work(&ast_list, LDLM_WORK_CP_AST);
1570 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1572 LDLM_LOCK_RELEASE(lock);
1576 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1577 struct ldlm_namespace *ns,
1578 struct ldlm_request *dlm_req,
1579 struct ldlm_lock *lock)
1584 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1586 if (lock->l_glimpse_ast != NULL)
1587 rc = lock->l_glimpse_ast(lock, req);
1589 if (req->rq_repmsg != NULL) {
1592 req->rq_status = rc;
1596 lock_res_and_lock(lock);
1597 if (lock->l_granted_mode == LCK_PW &&
1598 !lock->l_readers && !lock->l_writers &&
1599 cfs_time_after(cfs_time_current(),
1600 cfs_time_add(lock->l_last_used,
1601 cfs_time_seconds(10)))) {
1602 unlock_res_and_lock(lock);
1603 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1604 ldlm_handle_bl_callback(ns, NULL, lock);
1609 unlock_res_and_lock(lock);
1610 LDLM_LOCK_RELEASE(lock);
1614 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1616 if (req->rq_no_reply)
1619 req->rq_status = rc;
1620 if (!req->rq_packed_final) {
1621 rc = lustre_pack_reply(req, 1, NULL, NULL);
1625 return ptlrpc_reply(req);
1629 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi, int mode)
1631 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1634 cfs_spin_lock(&blp->blp_lock);
1635 if (blwi->blwi_lock && blwi->blwi_lock->l_flags & LDLM_FL_DISCARD_DATA) {
1636 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1637 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1639 /* other blocking callbacks are added to the regular list */
1640 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1642 cfs_spin_unlock(&blp->blp_lock);
1644 cfs_waitq_signal(&blp->blp_waitq);
1646 /* can not use blwi->blwi_mode as blwi could be already freed in
1648 if (mode == LDLM_SYNC)
1649 cfs_wait_for_completion(&blwi->blwi_comp);
1654 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1655 struct ldlm_namespace *ns,
1656 struct ldlm_lock_desc *ld,
1657 cfs_list_t *cancels, int count,
1658 struct ldlm_lock *lock,
1661 cfs_init_completion(&blwi->blwi_comp);
1662 CFS_INIT_LIST_HEAD(&blwi->blwi_head);
1664 if (cfs_memory_pressure_get())
1665 blwi->blwi_mem_pressure = 1;
1668 blwi->blwi_mode = mode;
1670 blwi->blwi_ld = *ld;
1672 cfs_list_add(&blwi->blwi_head, cancels);
1673 cfs_list_del_init(cancels);
1674 blwi->blwi_count = count;
1676 blwi->blwi_lock = lock;
1680 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
1681 struct ldlm_lock_desc *ld, struct ldlm_lock *lock,
1682 cfs_list_t *cancels, int count, int mode)
1686 if (cancels && count == 0)
1689 if (mode == LDLM_SYNC) {
1690 /* if it is synchronous call do minimum mem alloc, as it could
1691 * be triggered from kernel shrinker
1693 struct ldlm_bl_work_item blwi;
1694 memset(&blwi, 0, sizeof(blwi));
1695 init_blwi(&blwi, ns, ld, cancels, count, lock, LDLM_SYNC);
1696 RETURN(__ldlm_bl_to_thread(&blwi, LDLM_SYNC));
1698 struct ldlm_bl_work_item *blwi;
1699 OBD_ALLOC(blwi, sizeof(*blwi));
1702 init_blwi(blwi, ns, ld, cancels, count, lock, LDLM_ASYNC);
1704 RETURN(__ldlm_bl_to_thread(blwi, LDLM_ASYNC));
1710 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1711 struct ldlm_lock *lock)
1714 RETURN(ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LDLM_ASYNC));
1720 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1721 cfs_list_t *cancels, int count, int mode)
1724 RETURN(ldlm_bl_to_thread(ns, ld, NULL, cancels, count, mode));
1730 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
1731 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
1733 struct obd_device *obd = req->rq_export->exp_obd;
1740 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
1742 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
1744 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
1746 DEBUG_REQ(D_IOCTL, req, "no set_info key");
1749 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
1751 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
1753 DEBUG_REQ(D_IOCTL, req, "no set_info val");
1756 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
1759 /* We are responsible for swabbing contents of val */
1761 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
1762 /* Pass it on to mdc (the "export" in this case) */
1763 rc = obd_set_info_async(req->rq_export,
1764 sizeof(KEY_HSM_COPYTOOL_SEND),
1765 KEY_HSM_COPYTOOL_SEND,
1768 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
1773 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
1774 const char *msg, int rc,
1775 struct lustre_handle *handle)
1777 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
1778 "%s: [nid %s] [rc %d] [lock "LPX64"]",
1779 msg, libcfs_id2str(req->rq_peer), rc,
1780 handle ? handle->cookie : 0);
1781 if (req->rq_no_reply)
1782 CWARN("No reply was sent, maybe cause bug 21636.\n");
1784 CWARN("Send reply failed, maybe cause bug 21636.\n");
1787 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
1788 static int ldlm_callback_handler(struct ptlrpc_request *req)
1790 struct ldlm_namespace *ns;
1791 struct ldlm_request *dlm_req;
1792 struct ldlm_lock *lock;
1796 /* Requests arrive in sender's byte order. The ptlrpc service
1797 * handler has already checked and, if necessary, byte-swapped the
1798 * incoming request message body, but I am responsible for the
1799 * message buffers. */
1801 /* do nothing for sec context finalize */
1802 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
1805 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
1807 if (req->rq_export == NULL) {
1808 rc = ldlm_callback_reply(req, -ENOTCONN);
1809 ldlm_callback_errmsg(req, "Operate on unconnected server",
1814 LASSERT(req->rq_export != NULL);
1815 LASSERT(req->rq_export->exp_obd != NULL);
1817 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1818 case LDLM_BL_CALLBACK:
1819 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK))
1822 case LDLM_CP_CALLBACK:
1823 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK))
1826 case LDLM_GL_CALLBACK:
1827 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK))
1831 rc = ldlm_handle_setinfo(req);
1832 ldlm_callback_reply(req, rc);
1834 case OBD_LOG_CANCEL: /* remove this eventually - for 1.4.0 compat */
1835 CERROR("shouldn't be handling OBD_LOG_CANCEL on DLM thread\n");
1836 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
1837 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
1839 rc = llog_origin_handle_cancel(req);
1840 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
1842 ldlm_callback_reply(req, rc);
1844 case OBD_QC_CALLBACK:
1845 req_capsule_set(&req->rq_pill, &RQF_QC_CALLBACK);
1846 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_QC_CALLBACK_NET))
1848 rc = target_handle_qc_callback(req);
1849 ldlm_callback_reply(req, rc);
1853 /* reply in handler */
1854 req_capsule_set(&req->rq_pill, &RQF_MDS_QUOTA_DQACQ);
1855 rc = target_handle_dqacq_callback(req);
1857 case LLOG_ORIGIN_HANDLE_CREATE:
1858 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
1859 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1861 rc = llog_origin_handle_create(req);
1862 ldlm_callback_reply(req, rc);
1864 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1865 req_capsule_set(&req->rq_pill,
1866 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
1867 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1869 rc = llog_origin_handle_next_block(req);
1870 ldlm_callback_reply(req, rc);
1872 case LLOG_ORIGIN_HANDLE_READ_HEADER:
1873 req_capsule_set(&req->rq_pill,
1874 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
1875 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1877 rc = llog_origin_handle_read_header(req);
1878 ldlm_callback_reply(req, rc);
1880 case LLOG_ORIGIN_HANDLE_CLOSE:
1881 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1883 rc = llog_origin_handle_close(req);
1884 ldlm_callback_reply(req, rc);
1887 CERROR("unknown opcode %u\n",
1888 lustre_msg_get_opc(req->rq_reqmsg));
1889 ldlm_callback_reply(req, -EPROTO);
1893 ns = req->rq_export->exp_obd->obd_namespace;
1894 LASSERT(ns != NULL);
1896 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
1898 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1899 if (dlm_req == NULL) {
1900 rc = ldlm_callback_reply(req, -EPROTO);
1901 ldlm_callback_errmsg(req, "Operate without parameter", rc,
1906 /* Force a known safe race, send a cancel to the server for a lock
1907 * which the server has already started a blocking callback on. */
1908 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
1909 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
1910 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0]);
1912 CERROR("ldlm_cli_cancel: %d\n", rc);
1915 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
1917 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
1918 "disappeared\n", dlm_req->lock_handle[0].cookie);
1919 rc = ldlm_callback_reply(req, -EINVAL);
1920 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
1921 &dlm_req->lock_handle[0]);
1925 if ((lock->l_flags & LDLM_FL_FAIL_LOC) &&
1926 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
1927 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
1929 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1930 lock_res_and_lock(lock);
1931 lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1932 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
1933 /* If somebody cancels lock and cache is already dropped,
1934 * or lock is failed before cp_ast received on client,
1935 * we can tell the server we have no lock. Otherwise, we
1936 * should send cancel after dropping the cache. */
1937 if (((lock->l_flags & LDLM_FL_CANCELING) &&
1938 (lock->l_flags & LDLM_FL_BL_DONE)) ||
1939 (lock->l_flags & LDLM_FL_FAILED)) {
1940 LDLM_DEBUG(lock, "callback on lock "
1941 LPX64" - lock disappeared\n",
1942 dlm_req->lock_handle[0].cookie);
1943 unlock_res_and_lock(lock);
1944 LDLM_LOCK_RELEASE(lock);
1945 rc = ldlm_callback_reply(req, -EINVAL);
1946 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
1947 &dlm_req->lock_handle[0]);
1950 /* BL_AST locks are not needed in lru.
1951 * let ldlm_cancel_lru() be fast. */
1952 ldlm_lock_remove_from_lru(lock);
1953 lock->l_flags |= LDLM_FL_BL_AST;
1955 unlock_res_and_lock(lock);
1957 /* We want the ost thread to get this reply so that it can respond
1958 * to ost requests (write cache writeback) that might be triggered
1961 * But we'd also like to be able to indicate in the reply that we're
1962 * cancelling right now, because it's unused, or have an intent result
1963 * in the reply, so we might have to push the responsibility for sending
1964 * the reply down into the AST handlers, alas. */
1966 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1967 case LDLM_BL_CALLBACK:
1968 CDEBUG(D_INODE, "blocking ast\n");
1969 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
1970 if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)) {
1971 rc = ldlm_callback_reply(req, 0);
1972 if (req->rq_no_reply || rc)
1973 ldlm_callback_errmsg(req, "Normal process", rc,
1974 &dlm_req->lock_handle[0]);
1976 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
1977 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1979 case LDLM_CP_CALLBACK:
1980 CDEBUG(D_INODE, "completion ast\n");
1981 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
1982 ldlm_callback_reply(req, 0);
1983 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1985 case LDLM_GL_CALLBACK:
1986 CDEBUG(D_INODE, "glimpse ast\n");
1987 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
1988 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1991 LBUG(); /* checked above */
1997 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2002 /* Requests arrive in sender's byte order. The ptlrpc service
2003 * handler has already checked and, if necessary, byte-swapped the
2004 * incoming request message body, but I am responsible for the
2005 * message buffers. */
2007 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2009 if (req->rq_export == NULL) {
2010 struct ldlm_request *dlm_req;
2012 CERROR("%s from %s arrived at %lu with bad export cookie "
2014 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2015 libcfs_nid2str(req->rq_peer.nid),
2016 req->rq_arrival_time.tv_sec,
2017 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2019 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2020 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2021 dlm_req = req_capsule_client_get(&req->rq_pill,
2023 if (dlm_req != NULL)
2024 ldlm_lock_dump_handle(D_ERROR,
2025 &dlm_req->lock_handle[0]);
2027 ldlm_callback_reply(req, -ENOTCONN);
2031 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2033 /* XXX FIXME move this back to mds/handler.c, bug 249 */
2035 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2036 CDEBUG(D_INODE, "cancel\n");
2037 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL))
2039 rc = ldlm_handle_cancel(req);
2043 case OBD_LOG_CANCEL:
2044 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
2045 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
2047 rc = llog_origin_handle_cancel(req);
2048 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
2050 ldlm_callback_reply(req, rc);
2053 CERROR("invalid opcode %d\n",
2054 lustre_msg_get_opc(req->rq_reqmsg));
2055 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2056 ldlm_callback_reply(req, -EINVAL);
2062 void ldlm_revoke_lock_cb(void *obj, void *data)
2064 cfs_list_t *rpc_list = data;
2065 struct ldlm_lock *lock = obj;
2067 lock_res_and_lock(lock);
2069 if (lock->l_req_mode != lock->l_granted_mode) {
2070 unlock_res_and_lock(lock);
2074 LASSERT(lock->l_resource);
2075 if (lock->l_resource->lr_type != LDLM_IBITS &&
2076 lock->l_resource->lr_type != LDLM_PLAIN) {
2077 unlock_res_and_lock(lock);
2081 if (lock->l_flags & LDLM_FL_AST_SENT) {
2082 unlock_res_and_lock(lock);
2086 LASSERT(lock->l_blocking_ast);
2087 LASSERT(!lock->l_blocking_lock);
2089 lock->l_flags |= LDLM_FL_AST_SENT;
2090 if (lock->l_export && lock->l_export->exp_lock_hash &&
2091 !cfs_hlist_unhashed(&lock->l_exp_hash))
2092 cfs_hash_del(lock->l_export->exp_lock_hash,
2093 &lock->l_remote_handle, &lock->l_exp_hash);
2094 cfs_list_add_tail(&lock->l_rk_ast, rpc_list);
2095 LDLM_LOCK_GET(lock);
2097 unlock_res_and_lock(lock);
2100 void ldlm_revoke_export_locks(struct obd_export *exp)
2102 cfs_list_t rpc_list;
2105 CFS_INIT_LIST_HEAD(&rpc_list);
2106 cfs_hash_for_each_empty(exp->exp_lock_hash,
2107 ldlm_revoke_lock_cb, &rpc_list);
2108 ldlm_run_ast_work(&rpc_list, LDLM_WORK_REVOKE_AST);
2114 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
2116 struct ldlm_bl_work_item *blwi = NULL;
2117 static unsigned int num_bl = 0;
2119 cfs_spin_lock(&blp->blp_lock);
2120 /* process a request from the blp_list at least every blp_num_threads */
2121 if (!cfs_list_empty(&blp->blp_list) &&
2122 (cfs_list_empty(&blp->blp_prio_list) || num_bl == 0))
2123 blwi = cfs_list_entry(blp->blp_list.next,
2124 struct ldlm_bl_work_item, blwi_entry);
2126 if (!cfs_list_empty(&blp->blp_prio_list))
2127 blwi = cfs_list_entry(blp->blp_prio_list.next,
2128 struct ldlm_bl_work_item,
2132 if (++num_bl >= cfs_atomic_read(&blp->blp_num_threads))
2134 cfs_list_del(&blwi->blwi_entry);
2136 cfs_spin_unlock(&blp->blp_lock);
2141 /* This only contains temporary data until the thread starts */
2142 struct ldlm_bl_thread_data {
2143 char bltd_name[CFS_CURPROC_COMM_MAX];
2144 struct ldlm_bl_pool *bltd_blp;
2145 cfs_completion_t bltd_comp;
2149 static int ldlm_bl_thread_main(void *arg);
2151 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
2153 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2156 cfs_init_completion(&bltd.bltd_comp);
2157 rc = cfs_kernel_thread(ldlm_bl_thread_main, &bltd, 0);
2159 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %d\n",
2160 cfs_atomic_read(&blp->blp_num_threads), rc);
2163 cfs_wait_for_completion(&bltd.bltd_comp);
2168 static int ldlm_bl_thread_main(void *arg)
2170 struct ldlm_bl_pool *blp;
2174 struct ldlm_bl_thread_data *bltd = arg;
2176 blp = bltd->bltd_blp;
2179 cfs_atomic_inc_return(&blp->blp_num_threads) - 1;
2180 cfs_atomic_inc(&blp->blp_busy_threads);
2182 snprintf(bltd->bltd_name, sizeof(bltd->bltd_name) - 1,
2183 "ldlm_bl_%02d", bltd->bltd_num);
2184 cfs_daemonize(bltd->bltd_name);
2186 cfs_complete(&bltd->bltd_comp);
2187 /* cannot use bltd after this, it is only on caller's stack */
2191 struct l_wait_info lwi = { 0 };
2192 struct ldlm_bl_work_item *blwi = NULL;
2194 blwi = ldlm_bl_get_work(blp);
2199 cfs_atomic_dec(&blp->blp_busy_threads);
2200 l_wait_event_exclusive(blp->blp_waitq,
2201 (blwi = ldlm_bl_get_work(blp)) != NULL,
2203 busy = cfs_atomic_inc_return(&blp->blp_busy_threads);
2205 if (blwi->blwi_ns == NULL)
2206 /* added by ldlm_cleanup() */
2209 /* Not fatal if racy and have a few too many threads */
2210 if (unlikely(busy < blp->blp_max_threads &&
2211 busy >= cfs_atomic_read(&blp->blp_num_threads)))
2212 /* discard the return value, we tried */
2213 ldlm_bl_thread_start(blp);
2215 if (blwi->blwi_ns == NULL)
2216 /* added by ldlm_cleanup() */
2219 if (blwi->blwi_mem_pressure)
2220 cfs_memory_pressure_set();
2222 if (blwi->blwi_count) {
2224 /* The special case when we cancel locks in lru
2225 * asynchronously, we pass the list of locks here.
2226 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2227 * canceled locally yet. */
2228 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2231 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL, 0);
2233 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2236 if (blwi->blwi_mem_pressure)
2237 cfs_memory_pressure_clr();
2239 if (blwi->blwi_mode == LDLM_ASYNC)
2240 OBD_FREE(blwi, sizeof(*blwi));
2242 cfs_complete(&blwi->blwi_comp);
2245 cfs_atomic_dec(&blp->blp_busy_threads);
2246 cfs_atomic_dec(&blp->blp_num_threads);
2247 cfs_complete(&blp->blp_comp);
2253 static int ldlm_setup(void);
2254 static int ldlm_cleanup(void);
2256 int ldlm_get_ref(void)
2260 cfs_mutex_down(&ldlm_ref_sem);
2261 if (++ldlm_refcount == 1) {
2266 cfs_mutex_up(&ldlm_ref_sem);
2271 void ldlm_put_ref(void)
2274 cfs_mutex_down(&ldlm_ref_sem);
2275 if (ldlm_refcount == 1) {
2276 int rc = ldlm_cleanup();
2278 CERROR("ldlm_cleanup failed: %d\n", rc);
2284 cfs_mutex_up(&ldlm_ref_sem);
2290 * Export handle<->lock hash operations.
2293 ldlm_export_lock_hash(cfs_hash_t *hs, void *key, unsigned mask)
2295 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2299 ldlm_export_lock_key(cfs_hlist_node_t *hnode)
2301 struct ldlm_lock *lock;
2304 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2305 RETURN(&lock->l_remote_handle);
2309 ldlm_export_lock_compare(void *key, cfs_hlist_node_t *hnode)
2312 RETURN(lustre_handle_equal(ldlm_export_lock_key(hnode), key));
2316 ldlm_export_lock_get(cfs_hlist_node_t *hnode)
2318 struct ldlm_lock *lock;
2321 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2322 LDLM_LOCK_GET(lock);
2328 ldlm_export_lock_put(cfs_hlist_node_t *hnode)
2330 struct ldlm_lock *lock;
2333 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2334 LDLM_LOCK_RELEASE(lock);
2339 static cfs_hash_ops_t ldlm_export_lock_ops = {
2340 .hs_hash = ldlm_export_lock_hash,
2341 .hs_key = ldlm_export_lock_key,
2342 .hs_compare = ldlm_export_lock_compare,
2343 .hs_get = ldlm_export_lock_get,
2344 .hs_put = ldlm_export_lock_put
2347 int ldlm_init_export(struct obd_export *exp)
2351 exp->exp_lock_hash =
2352 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2353 HASH_EXP_LOCK_CUR_BITS, HASH_EXP_LOCK_MAX_BITS,
2354 &ldlm_export_lock_ops, CFS_HASH_REHASH);
2356 if (!exp->exp_lock_hash)
2361 EXPORT_SYMBOL(ldlm_init_export);
2363 void ldlm_destroy_export(struct obd_export *exp)
2366 cfs_hash_putref(exp->exp_lock_hash);
2367 exp->exp_lock_hash = NULL;
2370 EXPORT_SYMBOL(ldlm_destroy_export);
2372 static int ldlm_setup(void)
2374 struct ldlm_bl_pool *blp;
2376 int ldlm_min_threads = LDLM_THREADS_AUTO_MIN;
2377 int ldlm_max_threads = LDLM_THREADS_AUTO_MAX;
2383 if (ldlm_state != NULL)
2386 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2387 if (ldlm_state == NULL)
2391 rc = ldlm_proc_setup();
2397 if (ldlm_num_threads) {
2398 /* If ldlm_num_threads is set, it is the min and the max. */
2399 if (ldlm_num_threads > LDLM_THREADS_AUTO_MAX)
2400 ldlm_num_threads = LDLM_THREADS_AUTO_MAX;
2401 if (ldlm_num_threads < LDLM_THREADS_AUTO_MIN)
2402 ldlm_num_threads = LDLM_THREADS_AUTO_MIN;
2403 ldlm_min_threads = ldlm_max_threads = ldlm_num_threads;
2407 ldlm_state->ldlm_cb_service =
2408 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
2409 LDLM_MAXREPSIZE, LDLM_CB_REQUEST_PORTAL,
2410 LDLM_CB_REPLY_PORTAL, 2,
2411 ldlm_callback_handler, "ldlm_cbd",
2412 ldlm_svc_proc_dir, NULL,
2413 ldlm_min_threads, ldlm_max_threads,
2415 LCT_MD_THREAD|LCT_DT_THREAD, NULL);
2417 if (!ldlm_state->ldlm_cb_service) {
2418 CERROR("failed to start service\n");
2419 GOTO(out_proc, rc = -ENOMEM);
2422 ldlm_state->ldlm_cancel_service =
2423 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
2424 LDLM_MAXREPSIZE, LDLM_CANCEL_REQUEST_PORTAL,
2425 LDLM_CANCEL_REPLY_PORTAL, 6,
2426 ldlm_cancel_handler, "ldlm_canceld",
2427 ldlm_svc_proc_dir, NULL,
2428 ldlm_min_threads, ldlm_max_threads,
2430 LCT_MD_THREAD|LCT_DT_THREAD|LCT_CL_THREAD,
2433 if (!ldlm_state->ldlm_cancel_service) {
2434 CERROR("failed to start service\n");
2435 GOTO(out_proc, rc = -ENOMEM);
2438 OBD_ALLOC(blp, sizeof(*blp));
2440 GOTO(out_proc, rc = -ENOMEM);
2441 ldlm_state->ldlm_bl_pool = blp;
2443 cfs_spin_lock_init(&blp->blp_lock);
2444 CFS_INIT_LIST_HEAD(&blp->blp_list);
2445 CFS_INIT_LIST_HEAD(&blp->blp_prio_list);
2446 cfs_waitq_init(&blp->blp_waitq);
2447 cfs_atomic_set(&blp->blp_num_threads, 0);
2448 cfs_atomic_set(&blp->blp_busy_threads, 0);
2449 blp->blp_min_threads = ldlm_min_threads;
2450 blp->blp_max_threads = ldlm_max_threads;
2453 for (i = 0; i < blp->blp_min_threads; i++) {
2454 rc = ldlm_bl_thread_start(blp);
2456 GOTO(out_thread, rc);
2459 rc = ptlrpc_start_threads(NULL, ldlm_state->ldlm_cancel_service);
2461 GOTO(out_thread, rc);
2463 rc = ptlrpc_start_threads(NULL, ldlm_state->ldlm_cb_service);
2465 GOTO(out_thread, rc);
2467 CFS_INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
2468 expired_lock_thread.elt_state = ELT_STOPPED;
2469 cfs_waitq_init(&expired_lock_thread.elt_waitq);
2471 CFS_INIT_LIST_HEAD(&waiting_locks_list);
2472 cfs_spin_lock_init(&waiting_locks_spinlock);
2473 cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0);
2475 rc = cfs_kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FILES);
2477 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
2478 GOTO(out_thread, rc);
2481 cfs_wait_event(expired_lock_thread.elt_waitq,
2482 expired_lock_thread.elt_state == ELT_READY);
2486 rc = ldlm_pools_init();
2488 GOTO(out_thread, rc);
2494 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2495 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2500 ldlm_proc_cleanup();
2503 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
2508 static int ldlm_cleanup(void)
2511 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2515 if (!cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
2516 !cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
2517 CERROR("ldlm still has namespaces; clean these up first.\n");
2518 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
2519 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
2528 while (cfs_atomic_read(&blp->blp_num_threads) > 0) {
2529 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
2531 cfs_init_completion(&blp->blp_comp);
2533 cfs_spin_lock(&blp->blp_lock);
2534 cfs_list_add_tail(&blwi.blwi_entry, &blp->blp_list);
2535 cfs_waitq_signal(&blp->blp_waitq);
2536 cfs_spin_unlock(&blp->blp_lock);
2538 cfs_wait_for_completion(&blp->blp_comp);
2540 OBD_FREE(blp, sizeof(*blp));
2542 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2543 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2544 ldlm_proc_cleanup();
2546 expired_lock_thread.elt_state = ELT_TERMINATE;
2547 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
2548 cfs_wait_event(expired_lock_thread.elt_waitq,
2549 expired_lock_thread.elt_state == ELT_STOPPED);
2551 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2552 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2555 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
2561 int __init ldlm_init(void)
2563 cfs_init_mutex(&ldlm_ref_sem);
2564 cfs_init_mutex(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
2565 cfs_init_mutex(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
2566 ldlm_resource_slab = cfs_mem_cache_create("ldlm_resources",
2567 sizeof(struct ldlm_resource), 0,
2568 CFS_SLAB_HWCACHE_ALIGN);
2569 if (ldlm_resource_slab == NULL)
2572 ldlm_lock_slab = cfs_mem_cache_create("ldlm_locks",
2573 sizeof(struct ldlm_lock), 0,
2574 CFS_SLAB_HWCACHE_ALIGN | CFS_SLAB_DESTROY_BY_RCU);
2575 if (ldlm_lock_slab == NULL) {
2576 cfs_mem_cache_destroy(ldlm_resource_slab);
2580 ldlm_interval_slab = cfs_mem_cache_create("interval_node",
2581 sizeof(struct ldlm_interval),
2582 0, CFS_SLAB_HWCACHE_ALIGN);
2583 if (ldlm_interval_slab == NULL) {
2584 cfs_mem_cache_destroy(ldlm_resource_slab);
2585 cfs_mem_cache_destroy(ldlm_lock_slab);
2588 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2589 class_export_dump_hook = ldlm_dump_export_locks;
2594 void __exit ldlm_exit(void)
2598 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
2599 rc = cfs_mem_cache_destroy(ldlm_resource_slab);
2600 LASSERTF(rc == 0, "couldn't free ldlm resource slab\n");
2602 /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
2603 * synchronize_rcu() to wait a grace period elapsed, so that
2604 * ldlm_lock_free() get a chance to be called. */
2607 rc = cfs_mem_cache_destroy(ldlm_lock_slab);
2608 LASSERTF(rc == 0, "couldn't free ldlm lock slab\n");
2609 rc = cfs_mem_cache_destroy(ldlm_interval_slab);
2610 LASSERTF(rc == 0, "couldn't free interval node slab\n");
2614 EXPORT_SYMBOL(ldlm_extent_shift_kms);
2617 EXPORT_SYMBOL(ldlm_get_processing_policy);
2618 EXPORT_SYMBOL(ldlm_lock2desc);
2619 EXPORT_SYMBOL(ldlm_register_intent);
2620 EXPORT_SYMBOL(ldlm_lockname);
2621 EXPORT_SYMBOL(ldlm_typename);
2622 EXPORT_SYMBOL(ldlm_lock2handle);
2623 EXPORT_SYMBOL(__ldlm_handle2lock);
2624 EXPORT_SYMBOL(ldlm_lock_get);
2625 EXPORT_SYMBOL(ldlm_lock_put);
2626 EXPORT_SYMBOL(ldlm_lock_match);
2627 EXPORT_SYMBOL(ldlm_lock_cancel);
2628 EXPORT_SYMBOL(ldlm_lock_addref);
2629 EXPORT_SYMBOL(ldlm_lock_addref_try);
2630 EXPORT_SYMBOL(ldlm_lock_decref);
2631 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
2632 EXPORT_SYMBOL(ldlm_lock_change_resource);
2633 EXPORT_SYMBOL(ldlm_it2str);
2634 EXPORT_SYMBOL(ldlm_lock_dump);
2635 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2636 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
2637 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
2638 EXPORT_SYMBOL(ldlm_lock_allow_match);
2639 EXPORT_SYMBOL(ldlm_lock_downgrade);
2640 EXPORT_SYMBOL(ldlm_lock_convert);
2642 /* ldlm_request.c */
2643 EXPORT_SYMBOL(ldlm_completion_ast_async);
2644 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
2645 EXPORT_SYMBOL(ldlm_completion_ast);
2646 EXPORT_SYMBOL(ldlm_blocking_ast);
2647 EXPORT_SYMBOL(ldlm_glimpse_ast);
2648 EXPORT_SYMBOL(ldlm_expired_completion_wait);
2649 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
2650 EXPORT_SYMBOL(ldlm_prep_elc_req);
2651 EXPORT_SYMBOL(ldlm_cli_convert);
2652 EXPORT_SYMBOL(ldlm_cli_enqueue);
2653 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
2654 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
2655 EXPORT_SYMBOL(ldlm_cli_cancel);
2656 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
2657 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2658 EXPORT_SYMBOL(ldlm_cli_cancel_req);
2659 EXPORT_SYMBOL(ldlm_replay_locks);
2660 EXPORT_SYMBOL(ldlm_resource_foreach);
2661 EXPORT_SYMBOL(ldlm_namespace_foreach);
2662 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
2663 EXPORT_SYMBOL(ldlm_resource_iterate);
2664 EXPORT_SYMBOL(ldlm_cancel_resource_local);
2665 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
2666 EXPORT_SYMBOL(ldlm_cli_cancel_list);
2669 EXPORT_SYMBOL(ldlm_server_blocking_ast);
2670 EXPORT_SYMBOL(ldlm_server_completion_ast);
2671 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
2672 EXPORT_SYMBOL(ldlm_handle_enqueue);
2673 EXPORT_SYMBOL(ldlm_handle_enqueue0);
2674 EXPORT_SYMBOL(ldlm_handle_cancel);
2675 EXPORT_SYMBOL(ldlm_request_cancel);
2676 EXPORT_SYMBOL(ldlm_handle_convert);
2677 EXPORT_SYMBOL(ldlm_handle_convert0);
2678 EXPORT_SYMBOL(ldlm_del_waiting_lock);
2679 EXPORT_SYMBOL(ldlm_get_ref);
2680 EXPORT_SYMBOL(ldlm_put_ref);
2681 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
2682 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2684 /* ldlm_resource.c */
2685 EXPORT_SYMBOL(ldlm_namespace_new);
2686 EXPORT_SYMBOL(ldlm_namespace_cleanup);
2687 EXPORT_SYMBOL(ldlm_namespace_free);
2688 EXPORT_SYMBOL(ldlm_namespace_dump);
2689 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
2690 EXPORT_SYMBOL(ldlm_resource_get);
2691 EXPORT_SYMBOL(ldlm_resource_putref);
2692 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
2695 EXPORT_SYMBOL(client_import_add_conn);
2696 EXPORT_SYMBOL(client_import_del_conn);
2697 EXPORT_SYMBOL(client_obd_setup);
2698 EXPORT_SYMBOL(client_obd_cleanup);
2699 EXPORT_SYMBOL(client_connect_import);
2700 EXPORT_SYMBOL(client_disconnect_export);
2701 EXPORT_SYMBOL(server_disconnect_export);
2702 EXPORT_SYMBOL(target_stop_recovery_thread);
2703 EXPORT_SYMBOL(target_handle_connect);
2704 EXPORT_SYMBOL(target_cleanup_recovery);
2705 EXPORT_SYMBOL(target_destroy_export);
2706 EXPORT_SYMBOL(target_cancel_recovery_timer);
2707 EXPORT_SYMBOL(target_send_reply);
2708 EXPORT_SYMBOL(target_queue_recovery_request);
2709 EXPORT_SYMBOL(target_handle_ping);
2710 EXPORT_SYMBOL(target_pack_pool_reply);
2711 EXPORT_SYMBOL(target_handle_disconnect);
2714 EXPORT_SYMBOL(lock_res_and_lock);
2715 EXPORT_SYMBOL(unlock_res_and_lock);