4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ldlm/ldlm_lockd.c
34 * Author: Peter Braam <braam@clusterfs.com>
35 * Author: Phil Schwan <phil@clusterfs.com>
38 #define DEBUG_SUBSYSTEM S_LDLM
40 #include <linux/kthread.h>
41 #include <linux/list.h>
42 #include <libcfs/libcfs.h>
43 #include <lustre/lustre_errno.h>
44 #include <lustre_dlm.h>
45 #include <obd_class.h>
46 #include "ldlm_internal.h"
48 static int ldlm_num_threads;
49 module_param(ldlm_num_threads, int, 0444);
50 MODULE_PARM_DESC(ldlm_num_threads, "number of DLM service threads to start");
52 static char *ldlm_cpts;
53 module_param(ldlm_cpts, charp, 0444);
54 MODULE_PARM_DESC(ldlm_cpts, "CPU partitions ldlm threads should run on");
56 static struct mutex ldlm_ref_mutex;
57 static int ldlm_refcount;
59 struct kobject *ldlm_kobj;
60 struct kset *ldlm_ns_kset;
61 struct kset *ldlm_svc_kset;
65 static struct ldlm_state *ldlm_state;
67 static inline cfs_time_t round_timeout(cfs_time_t timeout)
69 return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1);
72 /* timeout for initial callback (AST) reply (bz10399) */
73 static inline unsigned int ldlm_get_rq_timeout(void)
76 unsigned int timeout = min(ldlm_timeout, obd_timeout / 3);
78 return timeout < 1 ? 1 : timeout;
83 #define ELT_TERMINATE 2
89 * blp_prio_list is used for callbacks that should be handled
90 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
93 struct list_head blp_prio_list;
96 * blp_list is used for all other callbacks which are likely
97 * to take longer to process.
99 struct list_head blp_list;
101 wait_queue_head_t blp_waitq;
102 struct completion blp_comp;
103 atomic_t blp_num_threads;
104 atomic_t blp_busy_threads;
109 struct ldlm_bl_work_item {
110 struct list_head blwi_entry;
111 struct ldlm_namespace *blwi_ns;
112 struct ldlm_lock_desc blwi_ld;
113 struct ldlm_lock *blwi_lock;
114 struct list_head blwi_head;
116 struct completion blwi_comp;
117 enum ldlm_cancel_flags blwi_flags;
118 int blwi_mem_pressure;
121 #ifdef HAVE_SERVER_SUPPORT
124 * Protects both waiting_locks_list and expired_lock_thread.
126 static spinlock_t waiting_locks_spinlock; /* BH lock (timer) */
129 * List for contended locks.
131 * As soon as a lock is contended, it gets placed on this list and
132 * expected time to get a response is filled in the lock. A special
133 * thread walks the list looking for locks that should be released and
134 * schedules client evictions for those that have not been released in
137 * All access to it should be under waiting_locks_spinlock.
139 static struct list_head waiting_locks_list;
140 static struct timer_list waiting_locks_timer;
142 static struct expired_lock_thread {
143 wait_queue_head_t elt_waitq;
146 struct list_head elt_expired_locks;
147 } expired_lock_thread;
149 static inline int have_expired_locks(void)
154 spin_lock_bh(&waiting_locks_spinlock);
155 need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
156 spin_unlock_bh(&waiting_locks_spinlock);
162 * Check expired lock list for expired locks and time them out.
164 static int expired_lock_main(void *arg)
166 struct list_head *expired = &expired_lock_thread.elt_expired_locks;
167 struct l_wait_info lwi = { 0 };
172 expired_lock_thread.elt_state = ELT_READY;
173 wake_up(&expired_lock_thread.elt_waitq);
176 l_wait_event(expired_lock_thread.elt_waitq,
177 have_expired_locks() ||
178 expired_lock_thread.elt_state == ELT_TERMINATE,
181 spin_lock_bh(&waiting_locks_spinlock);
182 if (expired_lock_thread.elt_dump) {
183 spin_unlock_bh(&waiting_locks_spinlock);
185 /* from waiting_locks_callback, but not in timer */
186 libcfs_debug_dumplog();
188 spin_lock_bh(&waiting_locks_spinlock);
189 expired_lock_thread.elt_dump = 0;
194 while (!list_empty(expired)) {
195 struct obd_export *export;
196 struct ldlm_lock *lock;
198 lock = list_entry(expired->next, struct ldlm_lock,
200 if ((void *)lock < LP_POISON + PAGE_SIZE &&
201 (void *)lock >= LP_POISON) {
202 spin_unlock_bh(&waiting_locks_spinlock);
203 CERROR("free lock on elt list %p\n", lock);
206 list_del_init(&lock->l_pending_chain);
207 if ((void *)lock->l_export <
208 LP_POISON + PAGE_SIZE &&
209 (void *)lock->l_export >= LP_POISON) {
210 CERROR("lock with free export on elt list %p\n",
212 lock->l_export = NULL;
213 LDLM_ERROR(lock, "free export");
214 /* release extra ref grabbed by
215 * ldlm_add_waiting_lock() or
216 * ldlm_failed_ast() */
217 LDLM_LOCK_RELEASE(lock);
221 if (ldlm_is_destroyed(lock)) {
222 /* release the lock refcount where
223 * waiting_locks_callback() founds */
224 LDLM_LOCK_RELEASE(lock);
227 export = class_export_lock_get(lock->l_export, lock);
228 spin_unlock_bh(&waiting_locks_spinlock);
230 spin_lock_bh(&export->exp_bl_list_lock);
231 list_del_init(&lock->l_exp_list);
232 spin_unlock_bh(&export->exp_bl_list_lock);
235 class_fail_export(export);
236 class_export_lock_put(export, lock);
238 /* release extra ref grabbed by ldlm_add_waiting_lock()
239 * or ldlm_failed_ast() */
240 LDLM_LOCK_RELEASE(lock);
242 spin_lock_bh(&waiting_locks_spinlock);
244 spin_unlock_bh(&waiting_locks_spinlock);
246 if (do_dump && obd_dump_on_eviction) {
247 CERROR("dump the log upon eviction\n");
248 libcfs_debug_dumplog();
251 if (expired_lock_thread.elt_state == ELT_TERMINATE)
255 expired_lock_thread.elt_state = ELT_STOPPED;
256 wake_up(&expired_lock_thread.elt_waitq);
260 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
261 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds);
264 * Check if there is a request in the export request list
265 * which prevents the lock canceling.
267 static int ldlm_lock_busy(struct ldlm_lock *lock)
269 struct ptlrpc_request *req;
273 if (lock->l_export == NULL)
276 spin_lock_bh(&lock->l_export->exp_rpc_lock);
277 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
279 if (req->rq_ops->hpreq_lock_match) {
280 match = req->rq_ops->hpreq_lock_match(req, lock);
285 spin_unlock_bh(&lock->l_export->exp_rpc_lock);
289 /* This is called from within a timer interrupt and cannot schedule */
290 static void waiting_locks_callback(unsigned long unused)
292 struct ldlm_lock *lock;
295 spin_lock_bh(&waiting_locks_spinlock);
296 while (!list_empty(&waiting_locks_list)) {
297 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
299 if (cfs_time_after(lock->l_callback_timeout,
300 cfs_time_current()) ||
301 (lock->l_req_mode == LCK_GROUP))
304 /* Check if we need to prolong timeout */
305 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
306 ldlm_lock_busy(lock)) {
309 if (lock->l_pending_chain.next == &waiting_locks_list)
314 spin_unlock_bh(&waiting_locks_spinlock);
315 LDLM_DEBUG(lock, "prolong the busy lock");
316 ldlm_refresh_waiting_lock(lock,
317 ldlm_bl_timeout(lock) >> 1);
318 spin_lock_bh(&waiting_locks_spinlock);
321 LDLM_LOCK_RELEASE(lock);
325 LDLM_LOCK_RELEASE(lock);
328 ldlm_lock_to_ns(lock)->ns_timeouts++;
329 LDLM_ERROR(lock, "lock callback timer expired after %llds: "
330 "evicting client at %s ",
331 ktime_get_real_seconds() - lock->l_last_activity,
333 lock->l_export->exp_connection->c_peer.nid));
335 /* no needs to take an extra ref on the lock since it was in
336 * the waiting_locks_list and ldlm_add_waiting_lock()
337 * already grabbed a ref */
338 list_del(&lock->l_pending_chain);
339 list_add(&lock->l_pending_chain,
340 &expired_lock_thread.elt_expired_locks);
344 if (!list_empty(&expired_lock_thread.elt_expired_locks)) {
345 if (obd_dump_on_timeout && need_dump)
346 expired_lock_thread.elt_dump = __LINE__;
348 wake_up(&expired_lock_thread.elt_waitq);
352 * Make sure the timer will fire again if we have any locks
355 if (!list_empty(&waiting_locks_list)) {
356 cfs_time_t timeout_rounded;
357 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
359 timeout_rounded = (cfs_time_t)round_timeout(lock->l_callback_timeout);
360 mod_timer(&waiting_locks_timer, timeout_rounded);
362 spin_unlock_bh(&waiting_locks_spinlock);
366 * Add lock to the list of contended locks.
368 * Indicate that we're waiting for a client to call us back cancelling a given
369 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
370 * timer to fire appropriately. (We round up to the next second, to avoid
371 * floods of timer firings during periods of high lock contention and traffic).
372 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
373 * if it has been added to the waiting list (1 is returned).
375 * Called with the namespace lock held.
377 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds)
380 cfs_time_t timeout_rounded;
382 if (!list_empty(&lock->l_pending_chain))
385 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
386 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
389 timeout = cfs_time_shift(seconds);
390 if (likely(cfs_time_after(timeout, lock->l_callback_timeout)))
391 lock->l_callback_timeout = timeout;
393 timeout_rounded = round_timeout(lock->l_callback_timeout);
395 if (cfs_time_before(timeout_rounded, waiting_locks_timer.expires) ||
396 !timer_pending(&waiting_locks_timer)) {
397 mod_timer(&waiting_locks_timer, timeout_rounded);
399 /* if the new lock has a shorter timeout than something earlier on
400 the list, we'll wait the longer amount of time; no big deal. */
402 list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
406 static void ldlm_add_blocked_lock(struct ldlm_lock *lock)
408 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
409 if (list_empty(&lock->l_exp_list)) {
410 if (lock->l_granted_mode != lock->l_req_mode)
411 list_add_tail(&lock->l_exp_list,
412 &lock->l_export->exp_bl_list);
414 list_add(&lock->l_exp_list,
415 &lock->l_export->exp_bl_list);
417 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
419 /* A blocked lock is added. Adjust the position in
420 * the stale list if the export is in the list.
421 * If export is stale and not in the list - it is being
422 * processed and will be placed on the right position
423 * on obd_stale_export_put(). */
424 if (!list_empty(&lock->l_export->exp_stale_list))
425 obd_stale_export_adjust(lock->l_export);
428 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
431 int timeout = ldlm_bl_timeout(lock);
433 /* NB: must be called with hold of lock_res_and_lock() */
434 LASSERT(ldlm_is_res_locked(lock));
435 LASSERT(!ldlm_is_cancel_on_block(lock));
437 /* Do not put cross-MDT lock in the waiting list, since we
438 * will not evict it due to timeout for now */
439 if (lock->l_export != NULL &&
440 (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS))
443 spin_lock_bh(&waiting_locks_spinlock);
444 if (ldlm_is_cancel(lock)) {
445 spin_unlock_bh(&waiting_locks_spinlock);
449 if (ldlm_is_destroyed(lock)) {
450 static cfs_time_t next;
452 spin_unlock_bh(&waiting_locks_spinlock);
453 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
454 if (cfs_time_after(cfs_time_current(), next)) {
455 next = cfs_time_shift(14400);
456 libcfs_debug_dumpstack(NULL);
461 ldlm_set_waited(lock);
462 lock->l_last_activity = ktime_get_real_seconds();
463 ret = __ldlm_add_waiting_lock(lock, timeout);
465 /* grab ref on the lock if it has been added to the
469 spin_unlock_bh(&waiting_locks_spinlock);
472 ldlm_add_blocked_lock(lock);
474 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)",
475 ret == 0 ? "not re-" : "", timeout,
476 AT_OFF ? "off" : "on");
481 * Remove a lock from the pending list, likely because it had its cancellation
482 * callback arrive without incident. This adjusts the lock-timeout timer if
483 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
484 * As done by ldlm_del_waiting_lock(), the caller must release the lock
485 * reference when the lock is removed from any list (1 is returned).
487 * Called with namespace lock held.
489 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
491 struct list_head *list_next;
493 if (list_empty(&lock->l_pending_chain))
496 list_next = lock->l_pending_chain.next;
497 if (lock->l_pending_chain.prev == &waiting_locks_list) {
498 /* Removing the head of the list, adjust timer. */
499 if (list_next == &waiting_locks_list) {
500 /* No more, just cancel. */
501 del_timer(&waiting_locks_timer);
503 struct ldlm_lock *next;
504 next = list_entry(list_next, struct ldlm_lock,
506 mod_timer(&waiting_locks_timer,
507 round_timeout(next->l_callback_timeout));
510 list_del_init(&lock->l_pending_chain);
515 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
519 if (lock->l_export == NULL) {
520 /* We don't have a "waiting locks list" on clients. */
521 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
525 spin_lock_bh(&waiting_locks_spinlock);
526 ret = __ldlm_del_waiting_lock(lock);
527 ldlm_clear_waited(lock);
528 spin_unlock_bh(&waiting_locks_spinlock);
530 /* remove the lock out of export blocking list */
531 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
532 list_del_init(&lock->l_exp_list);
533 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
536 /* release lock ref if it has indeed been removed
538 LDLM_LOCK_RELEASE(lock);
541 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
546 * Prolong the contended lock waiting time.
548 * Called with namespace lock held.
550 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
552 if (lock->l_export == NULL) {
553 /* We don't have a "waiting locks list" on clients. */
554 LDLM_DEBUG(lock, "client lock: no-op");
558 if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) {
559 /* We don't have a "waiting locks list" on OSP. */
560 LDLM_DEBUG(lock, "MDS-MDS lock: no-op");
564 spin_lock_bh(&waiting_locks_spinlock);
566 if (list_empty(&lock->l_pending_chain)) {
567 spin_unlock_bh(&waiting_locks_spinlock);
568 LDLM_DEBUG(lock, "wasn't waiting");
572 /* we remove/add the lock to the waiting list, so no needs to
573 * release/take a lock reference */
574 __ldlm_del_waiting_lock(lock);
575 __ldlm_add_waiting_lock(lock, timeout);
576 spin_unlock_bh(&waiting_locks_spinlock);
578 LDLM_DEBUG(lock, "refreshed");
581 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
583 #else /* HAVE_SERVER_SUPPORT */
585 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
590 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
595 #endif /* !HAVE_SERVER_SUPPORT */
597 #ifdef HAVE_SERVER_SUPPORT
600 * Calculate the per-export Blocking timeout (covering BL AST, data flush,
601 * lock cancel, and their replies). Used for lock callback timeout and AST
604 * \param[in] lock lock which is getting the blocking callback
606 * \retval timeout in seconds to wait for the client reply
608 unsigned int ldlm_bl_timeout(struct ldlm_lock *lock)
610 unsigned int timeout;
613 return obd_timeout / 2;
615 /* Since these are non-updating timeouts, we should be conservative.
616 * Take more than usually, 150%
617 * It would be nice to have some kind of "early reply" mechanism for
618 * lock callbacks too... */
619 timeout = at_get(&lock->l_export->exp_bl_lock_at);
620 return max(timeout + (timeout >> 1), ldlm_enqueue_min);
622 EXPORT_SYMBOL(ldlm_bl_timeout);
625 * Perform lock cleanup if AST sending failed.
627 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
628 const char *ast_type)
630 LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
631 "to a lock %s callback time out: rc %d\n",
632 lock->l_export->exp_obd->obd_name,
633 obd_export_nid2str(lock->l_export), ast_type, rc);
635 if (obd_dump_on_timeout)
636 libcfs_debug_dumplog();
637 spin_lock_bh(&waiting_locks_spinlock);
638 if (__ldlm_del_waiting_lock(lock) == 0)
639 /* the lock was not in any list, grab an extra ref before adding
640 * the lock to the expired list */
642 list_add(&lock->l_pending_chain,
643 &expired_lock_thread.elt_expired_locks);
644 wake_up(&expired_lock_thread.elt_waitq);
645 spin_unlock_bh(&waiting_locks_spinlock);
649 * Perform lock cleanup if AST reply came with error.
651 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
652 struct ptlrpc_request *req, int rc,
653 const char *ast_type)
655 struct lnet_process_id peer = req->rq_import->imp_connection->c_peer;
657 if (!req->rq_replied || (rc && rc != -EINVAL)) {
658 if (lock->l_export && lock->l_export->exp_libclient) {
660 "%s AST (req@%p x%llu) to liblustre client (nid %s) timeout, just cancelling lock",
661 ast_type, req, req->rq_xid,
662 libcfs_nid2str(peer.nid));
663 ldlm_lock_cancel(lock);
665 } else if (ldlm_is_cancel(lock)) {
667 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
668 ast_type, req, req->rq_xid,
669 libcfs_nid2str(peer.nid));
670 ldlm_lock_cancel(lock);
674 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
675 libcfs_nid2str(peer.nid),
676 req->rq_replied ? "returned error from" :
677 "failed to reply to",
678 ast_type, req, req->rq_xid,
679 (req->rq_repmsg != NULL) ?
680 lustre_msg_get_status(req->rq_repmsg) : 0,
682 ldlm_failed_ast(lock, rc, ast_type);
688 struct ldlm_resource *res = lock->l_resource;
691 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
692 libcfs_nid2str(peer.nid),
694 lustre_msg_get_status(req->rq_repmsg) : -1,
695 ast_type, req, req->rq_xid);
697 /* update lvbo to return proper attributes.
699 ldlm_resource_getref(res);
700 ldlm_res_lvbo_update(res, NULL, 1);
701 ldlm_resource_putref(res);
703 ldlm_lock_cancel(lock);
710 static int ldlm_cb_interpret(const struct lu_env *env,
711 struct ptlrpc_request *req, void *data, int rc)
713 struct ldlm_cb_async_args *ca = data;
714 struct ldlm_lock *lock = ca->ca_lock;
715 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
718 LASSERT(lock != NULL);
721 case LDLM_GL_CALLBACK:
722 /* Update the LVB from disk if the AST failed
723 * (this is a legal race)
725 * - Glimpse callback of local lock just returns
726 * -ELDLM_NO_LOCK_DATA.
727 * - Glimpse callback of remote lock might return
728 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
730 if (unlikely(arg->gl_interpret_reply)) {
731 rc = arg->gl_interpret_reply(env, req, data, rc);
732 } else if (rc == -ELDLM_NO_LOCK_DATA) {
733 LDLM_DEBUG(lock, "lost race - client has a lock but no "
735 ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
736 } else if (rc != 0) {
737 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
739 rc = ldlm_res_lvbo_update(lock->l_resource, req, 1);
742 case LDLM_BL_CALLBACK:
744 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
746 case LDLM_CP_CALLBACK:
748 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
751 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
756 /* release extra reference taken in ldlm_ast_fini() */
757 LDLM_LOCK_RELEASE(lock);
760 atomic_inc(&arg->restart);
765 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
767 struct ldlm_cb_async_args *ca = data;
768 struct ldlm_lock *lock = ca->ca_lock;
770 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
773 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
774 struct ldlm_cb_set_arg *arg,
775 struct ldlm_lock *lock,
781 if (unlikely(instant_cancel)) {
782 rc = ptl_send_rpc(req, 1);
783 ptlrpc_req_finished(req);
785 atomic_inc(&arg->restart);
788 ptlrpc_set_add_req(arg->set, req);
795 * Check if there are requests in the export request list which prevent
796 * the lock canceling and make these requests high priority ones.
798 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
800 struct ptlrpc_request *req;
803 if (lock->l_export == NULL) {
804 LDLM_DEBUG(lock, "client lock: no-op");
808 spin_lock_bh(&lock->l_export->exp_rpc_lock);
809 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
811 /* Do not process requests that were not yet added to there
812 * incoming queue or were already removed from there for
813 * processing. We evaluate ptlrpc_nrs_req_can_move() without
814 * holding svcpt->scp_req_lock, and then redo the check with
815 * the lock held once we need to obtain a reliable result.
817 if (ptlrpc_nrs_req_can_move(req) &&
818 req->rq_ops->hpreq_lock_match &&
819 req->rq_ops->hpreq_lock_match(req, lock))
820 ptlrpc_nrs_req_hp_move(req);
822 spin_unlock_bh(&lock->l_export->exp_rpc_lock);
827 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
828 * enqueued server lock conflicts with given one.
830 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
831 * to wait for client response.
833 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
834 struct ldlm_lock_desc *desc,
835 void *data, int flag)
837 struct ldlm_cb_async_args *ca;
838 struct ldlm_cb_set_arg *arg = data;
839 struct ldlm_request *body;
840 struct ptlrpc_request *req;
841 int instant_cancel = 0;
845 if (flag == LDLM_CB_CANCELING)
846 /* Don't need to do anything here. */
849 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
850 LDLM_DEBUG(lock, "dropping BL AST");
855 LASSERT(data != NULL);
856 if (lock->l_export->exp_obd->obd_recovering != 0)
857 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
859 ldlm_lock_reorder_req(lock);
861 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
862 &RQF_LDLM_BL_CALLBACK,
863 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
867 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
868 ca = ptlrpc_req_async_args(req);
869 ca->ca_set_arg = arg;
872 req->rq_interpret_reply = ldlm_cb_interpret;
874 lock_res_and_lock(lock);
875 if (ldlm_is_destroyed(lock)) {
876 /* What's the point? */
877 unlock_res_and_lock(lock);
878 ptlrpc_req_finished(req);
882 if (lock->l_granted_mode != lock->l_req_mode) {
883 /* this blocking AST will be communicated as part of the
884 * completion AST instead */
885 ldlm_add_blocked_lock(lock);
886 ldlm_set_waited(lock);
887 unlock_res_and_lock(lock);
889 ptlrpc_req_finished(req);
890 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
894 if (ldlm_is_cancel_on_block(lock))
897 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
898 body->lock_handle[0] = lock->l_remote_handle;
899 body->lock_desc = *desc;
900 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
902 LDLM_DEBUG(lock, "server preparing blocking AST");
904 ptlrpc_request_set_replen(req);
905 ldlm_set_cbpending(lock);
906 if (instant_cancel) {
907 unlock_res_and_lock(lock);
908 ldlm_lock_cancel(lock);
910 req->rq_no_resend = 1;
912 LASSERT(lock->l_granted_mode == lock->l_req_mode);
913 ldlm_add_waiting_lock(lock);
914 unlock_res_and_lock(lock);
916 /* Do not resend after lock callback timeout */
917 req->rq_delay_limit = ldlm_bl_timeout(lock);
918 req->rq_resend_cb = ldlm_update_resend;
921 req->rq_send_state = LUSTRE_IMP_FULL;
922 /* ptlrpc_request_alloc_pack already set timeout */
924 req->rq_timeout = ldlm_get_rq_timeout();
926 lock->l_last_activity = ktime_get_real_seconds();
928 if (lock->l_export && lock->l_export->exp_nid_stats &&
929 lock->l_export->exp_nid_stats->nid_ldlm_stats)
930 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
931 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
933 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
939 * ->l_completion_ast callback for a remote lock in server namespace.
941 * Sends AST to the client notifying it of lock granting. If initial
942 * lock response was not sent yet, instead of sending another RPC, just
943 * mark the lock as granted and client will understand
945 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
947 struct ldlm_cb_set_arg *arg = data;
948 struct ldlm_request *body;
949 struct ptlrpc_request *req;
950 struct ldlm_cb_async_args *ca;
951 int instant_cancel = 0;
956 LASSERT(lock != NULL);
957 LASSERT(data != NULL);
959 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
960 LDLM_DEBUG(lock, "dropping CP AST");
964 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
965 &RQF_LDLM_CP_CALLBACK);
969 /* server namespace, doesn't need lock */
970 lvb_len = ldlm_lvbo_size(lock);
971 /* LU-3124 & LU-2187: to not return layout in completion AST because
972 * it may deadlock for LU-2187, or client may not have enough space
973 * for large layout. The layout will be returned to client with an
974 * extra RPC to fetch xattr.lov */
975 if (ldlm_has_layout(lock))
978 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
979 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
981 ptlrpc_request_free(req);
985 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
986 ca = ptlrpc_req_async_args(req);
987 ca->ca_set_arg = arg;
990 req->rq_interpret_reply = ldlm_cb_interpret;
991 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
993 body->lock_handle[0] = lock->l_remote_handle;
994 body->lock_flags = ldlm_flags_to_wire(flags);
995 ldlm_lock2desc(lock, &body->lock_desc);
997 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
999 lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len);
1001 /* We still need to send the RPC to wake up the blocked
1002 * enqueue thread on the client.
1004 * Consider old client, there is no better way to notify
1005 * the failure, just zero-sized the LVB, then the client
1006 * will fail out as "-EPROTO". */
1007 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1011 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1016 lock->l_last_activity = ktime_get_real_seconds();
1018 LDLM_DEBUG(lock, "server preparing completion AST");
1020 ptlrpc_request_set_replen(req);
1022 req->rq_send_state = LUSTRE_IMP_FULL;
1023 /* ptlrpc_request_pack already set timeout */
1025 req->rq_timeout = ldlm_get_rq_timeout();
1027 /* We only send real blocking ASTs after the lock is granted */
1028 lock_res_and_lock(lock);
1029 if (ldlm_is_ast_sent(lock)) {
1030 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1031 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1032 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1035 /* We might get here prior to ldlm_handle_enqueue setting
1036 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1037 * into waiting list, but this is safe and similar code in
1038 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1039 * that would not only cancel the lock, but will also remove
1040 * it from waiting list */
1041 if (ldlm_is_cancel_on_block(lock)) {
1042 unlock_res_and_lock(lock);
1043 ldlm_lock_cancel(lock);
1046 req->rq_no_resend = 1;
1048 lock_res_and_lock(lock);
1050 /* start the lock-timeout clock */
1051 ldlm_add_waiting_lock(lock);
1052 /* Do not resend after lock callback timeout */
1053 req->rq_delay_limit = ldlm_bl_timeout(lock);
1054 req->rq_resend_cb = ldlm_update_resend;
1057 unlock_res_and_lock(lock);
1059 if (lock->l_export && lock->l_export->exp_nid_stats &&
1060 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1061 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1062 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1064 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1066 RETURN(lvb_len < 0 ? lvb_len : rc);
1070 * Server side ->l_glimpse_ast handler for client locks.
1072 * Sends glimpse AST to the client and waits for reply. Then updates
1073 * lvbo with the result.
1075 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1077 struct ldlm_cb_set_arg *arg = data;
1078 struct ldlm_request *body;
1079 struct ptlrpc_request *req;
1080 struct ldlm_cb_async_args *ca;
1082 struct req_format *req_fmt;
1085 LASSERT(lock != NULL);
1087 if (arg->gl_desc != NULL)
1088 /* There is a glimpse descriptor to pack */
1089 req_fmt = &RQF_LDLM_GL_DESC_CALLBACK;
1091 req_fmt = &RQF_LDLM_GL_CALLBACK;
1093 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1094 req_fmt, LUSTRE_DLM_VERSION,
1100 if (arg->gl_desc != NULL) {
1101 /* copy the GL descriptor */
1102 union ldlm_gl_desc *desc;
1103 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1104 *desc = *arg->gl_desc;
1107 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1108 body->lock_handle[0] = lock->l_remote_handle;
1109 ldlm_lock2desc(lock, &body->lock_desc);
1111 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1112 ca = ptlrpc_req_async_args(req);
1113 ca->ca_set_arg = arg;
1116 /* server namespace, doesn't need lock */
1117 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1118 ldlm_lvbo_size(lock));
1119 ptlrpc_request_set_replen(req);
1121 req->rq_send_state = LUSTRE_IMP_FULL;
1122 /* ptlrpc_request_alloc_pack already set timeout */
1124 req->rq_timeout = ldlm_get_rq_timeout();
1126 lock->l_last_activity = ktime_get_real_seconds();
1128 req->rq_interpret_reply = ldlm_cb_interpret;
1130 if (lock->l_export && lock->l_export->exp_nid_stats &&
1131 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1132 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1133 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1135 rc = ldlm_ast_fini(req, arg, lock, 0);
1140 int ldlm_glimpse_locks(struct ldlm_resource *res,
1141 struct list_head *gl_work_list)
1146 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1148 if (rc == -ERESTART)
1149 ldlm_reprocess_all(res);
1153 EXPORT_SYMBOL(ldlm_glimpse_locks);
1155 /* return LDLM lock associated with a lock callback request */
1156 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1158 struct ldlm_cb_async_args *ca;
1159 struct ldlm_lock *lock;
1162 ca = ptlrpc_req_async_args(req);
1165 RETURN(ERR_PTR(-EFAULT));
1169 EXPORT_SYMBOL(ldlm_request_lock);
1171 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
1172 struct lprocfs_stats *srv_stats)
1174 int lock_type = 0, op = 0;
1176 lock_type = dlm_req->lock_desc.l_resource.lr_type;
1178 switch (lock_type) {
1180 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
1183 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
1184 op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
1186 op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
1189 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
1192 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
1200 lprocfs_counter_incr(srv_stats, op);
1206 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1207 * service threads to carry out client lock enqueueing requests.
1209 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1210 struct ptlrpc_request *req,
1211 const struct ldlm_request *dlm_req,
1212 const struct ldlm_callback_suite *cbs)
1214 struct ldlm_reply *dlm_rep;
1216 enum ldlm_error err = ELDLM_OK;
1217 struct ldlm_lock *lock = NULL;
1218 void *cookie = NULL;
1220 struct ldlm_resource *res = NULL;
1223 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1225 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1226 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1228 LASSERT(req->rq_export);
1230 if (ptlrpc_req2svc(req)->srv_stats != NULL)
1231 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1233 if (req->rq_export && req->rq_export->exp_nid_stats &&
1234 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1235 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1236 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1238 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1239 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1240 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1241 dlm_req->lock_desc.l_resource.lr_type);
1242 GOTO(out, rc = -EFAULT);
1245 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1246 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1247 dlm_req->lock_desc.l_req_mode &
1248 (dlm_req->lock_desc.l_req_mode-1))) {
1249 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1250 dlm_req->lock_desc.l_req_mode);
1251 GOTO(out, rc = -EFAULT);
1254 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) {
1255 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1257 DEBUG_REQ(D_ERROR, req,
1258 "PLAIN lock request from IBITS client?");
1259 GOTO(out, rc = -EPROTO);
1261 } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1263 DEBUG_REQ(D_ERROR, req,
1264 "IBITS lock request from unaware client?");
1265 GOTO(out, rc = -EPROTO);
1268 if (unlikely((flags & LDLM_FL_REPLAY) ||
1269 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1270 /* Find an existing lock in the per-export lock hash */
1271 /* In the function below, .hs_keycmp resolves to
1272 * ldlm_export_lock_keycmp() */
1273 /* coverity[overrun-buffer-val] */
1274 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1275 (void *)&dlm_req->lock_handle[0]);
1277 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie %#llx",
1278 lock->l_handle.h_cookie);
1279 flags |= LDLM_FL_RESENT;
1280 GOTO(existing_lock, rc = 0);
1283 if (ldlm_reclaim_full()) {
1284 DEBUG_REQ(D_DLMTRACE, req, "Too many granted locks, "
1285 "reject current enqueue request and let the "
1286 "client retry later.\n");
1287 GOTO(out, rc = -EINPROGRESS);
1291 /* The lock's callback data might be set in the policy function */
1292 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1293 dlm_req->lock_desc.l_resource.lr_type,
1294 dlm_req->lock_desc.l_req_mode,
1295 cbs, NULL, 0, LVB_T_NONE);
1302 lock->l_remote_handle = dlm_req->lock_handle[0];
1303 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1305 /* Initialize resource lvb but not for a lock being replayed since
1306 * Client already got lvb sent in this case.
1307 * This must occur early since some policy methods assume resource
1308 * lvb is available (lr_lvb_data != NULL).
1310 res = lock->l_resource;
1311 if (!(flags & LDLM_FL_REPLAY)) {
1312 /* non-replayed lock, delayed lvb init may need to be done */
1313 rc = ldlm_lvbo_init(res);
1315 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1320 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1321 /* Don't enqueue a lock onto the export if it is been disonnected
1322 * due to eviction (bug 3822) or server umount (bug 24324).
1323 * Cancel it now instead. */
1324 if (req->rq_export->exp_disconnected) {
1325 LDLM_ERROR(lock, "lock on disconnected export %p",
1327 GOTO(out, rc = -ENOTCONN);
1330 lock->l_export = class_export_lock_get(req->rq_export, lock);
1331 if (lock->l_export->exp_lock_hash)
1332 cfs_hash_add(lock->l_export->exp_lock_hash,
1333 &lock->l_remote_handle,
1336 /* Inherit the enqueue flags before the operation, because we do not
1337 * keep the res lock on return and next operations (BL AST) may proceed
1339 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1340 LDLM_FL_INHERIT_MASK);
1342 ldlm_convert_policy_to_local(req->rq_export,
1343 dlm_req->lock_desc.l_resource.lr_type,
1344 &dlm_req->lock_desc.l_policy_data,
1345 &lock->l_policy_data);
1346 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1347 lock->l_req_extent = lock->l_policy_data.l_extent;
1351 if (flags & LDLM_FL_HAS_INTENT) {
1352 /* In this case, the reply buffer is allocated deep in
1353 * local_lock_enqueue by the policy function. */
1356 /* based on the assumption that lvb size never changes during
1357 * resource life time otherwise it need resource->lr_lock's
1359 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1360 RCL_SERVER, ldlm_lvbo_size(lock));
1362 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1363 GOTO(out, rc = -ENOMEM);
1365 rc = req_capsule_server_pack(&req->rq_pill);
1370 err = ldlm_lock_enqueue(ns, &lock, cookie, &flags);
1377 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1379 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1380 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1382 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1383 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1385 /* We never send a blocking AST until the lock is granted, but
1386 * we can tell it right now */
1387 lock_res_and_lock(lock);
1389 /* Now take into account flags to be inherited from original lock
1390 request both in reply to client and in our own lock flags. */
1391 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1392 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1394 /* Don't move a pending lock onto the export if it has already been
1395 * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1396 * Cancel it now instead. */
1397 if (unlikely(req->rq_export->exp_disconnected ||
1398 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1399 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1401 } else if (ldlm_is_ast_sent(lock)) {
1402 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1403 if (lock->l_granted_mode == lock->l_req_mode) {
1405 * Only cancel lock if it was granted, because it would
1406 * be destroyed immediately and would never be granted
1407 * in the future, causing timeouts on client. Not
1408 * granted lock will be cancelled immediately after
1409 * sending completion AST.
1411 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1412 unlock_res_and_lock(lock);
1413 ldlm_lock_cancel(lock);
1414 lock_res_and_lock(lock);
1416 ldlm_add_waiting_lock(lock);
1419 /* Make sure we never ever grant usual metadata locks to liblustre
1421 if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1422 dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1423 req->rq_export->exp_libclient) {
1424 if (unlikely(!ldlm_is_cancel_on_block(lock) ||
1425 !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1426 CERROR("Granting sync lock to libclient. "
1427 "req fl %d, rep fl %d, lock fl %#llx\n",
1428 dlm_req->lock_flags, dlm_rep->lock_flags,
1430 LDLM_ERROR(lock, "sync lock");
1431 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1432 struct ldlm_intent *it;
1434 it = req_capsule_client_get(&req->rq_pill,
1437 CERROR("This is intent %s (%llu)\n",
1438 ldlm_it2str(it->opc), it->opc);
1444 unlock_res_and_lock(lock);
1448 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1449 if (!req->rq_packed_final) {
1450 err = lustre_pack_reply(req, 1, NULL, NULL);
1455 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1456 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1458 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1459 "(err=%d, rc=%d)", err, rc);
1462 if (req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1464 ldlm_lvbo_size(lock) > 0) {
1468 buf = req_capsule_server_get(&req->rq_pill,
1470 LASSERTF(buf != NULL, "req %p, lock %p\n",
1472 buflen = req_capsule_get_size(&req->rq_pill,
1473 &RMF_DLM_LVB, RCL_SERVER);
1474 /* non-replayed lock, delayed lvb init may
1475 * need to be occur now */
1476 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1477 buflen = ldlm_lvbo_fill(lock, buf,
1483 buflen, RCL_SERVER);
1486 } else if (flags & LDLM_FL_REPLAY) {
1487 /* no LVB resend upon replay */
1501 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1502 if (lock->l_export) {
1503 ldlm_lock_cancel(lock);
1505 lock_res_and_lock(lock);
1506 ldlm_resource_unlink_lock(lock);
1507 ldlm_lock_destroy_nolock(lock);
1508 unlock_res_and_lock(lock);
1513 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1514 ldlm_reprocess_all(lock->l_resource);
1516 LDLM_LOCK_RELEASE(lock);
1519 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1526 * Old-style LDLM main entry point for server code enqueue.
1528 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1529 ldlm_completion_callback completion_callback,
1530 ldlm_blocking_callback blocking_callback,
1531 ldlm_glimpse_callback glimpse_callback)
1533 struct ldlm_request *dlm_req;
1534 struct ldlm_callback_suite cbs = {
1535 .lcs_completion = completion_callback,
1536 .lcs_blocking = blocking_callback,
1537 .lcs_glimpse = glimpse_callback
1541 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1542 if (dlm_req != NULL) {
1543 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1544 req, dlm_req, &cbs);
1552 * Main LDLM entry point for server code to process lock conversion requests.
1554 int ldlm_handle_convert0(struct ptlrpc_request *req,
1555 const struct ldlm_request *dlm_req)
1557 struct ldlm_reply *dlm_rep;
1558 struct ldlm_lock *lock;
1562 if (req->rq_export && req->rq_export->exp_nid_stats &&
1563 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1564 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1565 LDLM_CONVERT - LDLM_FIRST_OPC);
1567 rc = req_capsule_server_pack(&req->rq_pill);
1571 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1572 dlm_rep->lock_flags = dlm_req->lock_flags;
1574 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1576 req->rq_status = LUSTRE_EINVAL;
1580 LDLM_DEBUG(lock, "server-side convert handler START");
1582 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1583 &dlm_rep->lock_flags);
1585 if (ldlm_del_waiting_lock(lock))
1586 LDLM_DEBUG(lock, "converted waiting lock");
1589 req->rq_status = LUSTRE_EDEADLK;
1594 if (!req->rq_status)
1595 ldlm_reprocess_all(lock->l_resource);
1596 LDLM_DEBUG(lock, "server-side convert handler END");
1597 LDLM_LOCK_PUT(lock);
1599 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1605 * Old-style main LDLM entry point for server code to process lock conversion
1608 int ldlm_handle_convert(struct ptlrpc_request *req)
1611 struct ldlm_request *dlm_req;
1613 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1614 if (dlm_req != NULL) {
1615 rc = ldlm_handle_convert0(req, dlm_req);
1617 CERROR ("Can't unpack dlm_req\n");
1624 * Cancel all the locks whose handles are packed into ldlm_request
1626 * Called by server code expecting such combined cancel activity
1629 int ldlm_request_cancel(struct ptlrpc_request *req,
1630 const struct ldlm_request *dlm_req,
1631 int first, enum lustre_at_flags flags)
1633 struct ldlm_resource *res, *pres = NULL;
1634 struct ldlm_lock *lock;
1635 int i, count, done = 0;
1638 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1642 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1645 /* There is no lock on the server at the replay time,
1646 * skip lock cancelling to make replay tests to pass. */
1647 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1650 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1651 "starting at %d", count, first);
1653 for (i = first; i < count; i++) {
1654 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1656 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1657 "lock (cookie %llu)",
1658 dlm_req->lock_handle[i].cookie);
1662 res = lock->l_resource;
1665 /* This code is an optimization to only attempt lock
1666 * granting on the resource (that could be CPU-expensive)
1667 * after we are done cancelling lock in that resource. */
1670 ldlm_reprocess_all(pres);
1671 LDLM_RESOURCE_DELREF(pres);
1672 ldlm_resource_putref(pres);
1675 ldlm_resource_getref(res);
1676 LDLM_RESOURCE_ADDREF(res);
1677 ldlm_res_lvbo_update(res, NULL, 1);
1682 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock)) {
1683 time64_t delay = ktime_get_real_seconds() -
1684 lock->l_last_activity;
1685 LDLM_DEBUG(lock, "server cancels blocked lock after %llds",
1687 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1689 ldlm_lock_cancel(lock);
1690 LDLM_LOCK_PUT(lock);
1693 ldlm_reprocess_all(pres);
1694 LDLM_RESOURCE_DELREF(pres);
1695 ldlm_resource_putref(pres);
1697 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1700 EXPORT_SYMBOL(ldlm_request_cancel);
1703 * Main LDLM entry point for server code to cancel locks.
1705 * Typically gets called from service handler on LDLM_CANCEL opc.
1707 int ldlm_handle_cancel(struct ptlrpc_request *req)
1709 struct ldlm_request *dlm_req;
1713 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1714 if (dlm_req == NULL) {
1715 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1719 if (req->rq_export && req->rq_export->exp_nid_stats &&
1720 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1721 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1722 LDLM_CANCEL - LDLM_FIRST_OPC);
1724 rc = req_capsule_server_pack(&req->rq_pill);
1728 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1729 req->rq_status = LUSTRE_ESTALE;
1731 RETURN(ptlrpc_reply(req));
1733 #endif /* HAVE_SERVER_SUPPORT */
1736 * Callback handler for receiving incoming blocking ASTs.
1738 * This can only happen on client side.
1740 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1741 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1746 LDLM_DEBUG(lock, "client blocking AST callback handler");
1748 lock_res_and_lock(lock);
1749 ldlm_set_cbpending(lock);
1751 if (ldlm_is_cancel_on_block(lock))
1752 ldlm_set_cancel(lock);
1754 do_ast = (!lock->l_readers && !lock->l_writers);
1755 unlock_res_and_lock(lock);
1758 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1759 lock, lock->l_blocking_ast);
1760 if (lock->l_blocking_ast != NULL)
1761 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1764 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1768 LDLM_DEBUG(lock, "client blocking callback handler END");
1769 LDLM_LOCK_RELEASE(lock);
1774 * Callback handler for receiving incoming completion ASTs.
1776 * This only can happen on client side.
1778 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1779 struct ldlm_namespace *ns,
1780 struct ldlm_request *dlm_req,
1781 struct ldlm_lock *lock)
1783 struct list_head ast_list;
1788 LDLM_DEBUG(lock, "client completion callback handler START");
1790 INIT_LIST_HEAD(&ast_list);
1791 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1792 int to = cfs_time_seconds(1);
1794 set_current_state(TASK_INTERRUPTIBLE);
1795 schedule_timeout(to);
1796 if (lock->l_granted_mode == lock->l_req_mode ||
1797 ldlm_is_destroyed(lock))
1802 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1804 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1805 GOTO(out, rc = lvb_len);
1806 } else if (lvb_len > 0) {
1807 if (lock->l_lvb_len > 0) {
1808 /* for extent lock, lvb contains ost_lvb{}. */
1809 LASSERT(lock->l_lvb_data != NULL);
1811 if (unlikely(lock->l_lvb_len < lvb_len)) {
1812 LDLM_ERROR(lock, "Replied LVB is larger than "
1813 "expectation, expected = %d, "
1815 lock->l_lvb_len, lvb_len);
1816 GOTO(out, rc = -EINVAL);
1821 lock_res_and_lock(lock);
1822 if (ldlm_is_destroyed(lock) ||
1823 lock->l_granted_mode == lock->l_req_mode) {
1824 /* bug 11300: the lock has already been granted */
1825 unlock_res_and_lock(lock);
1826 LDLM_DEBUG(lock, "Double grant race happened");
1830 /* If we receive the completion AST before the actual enqueue returned,
1831 * then we might need to switch lock modes, resources, or extents. */
1832 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1833 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1834 LDLM_DEBUG(lock, "completion AST, new lock mode");
1837 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1838 ldlm_convert_policy_to_local(req->rq_export,
1839 dlm_req->lock_desc.l_resource.lr_type,
1840 &dlm_req->lock_desc.l_policy_data,
1841 &lock->l_policy_data);
1842 LDLM_DEBUG(lock, "completion AST, new policy data");
1845 ldlm_resource_unlink_lock(lock);
1846 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1847 &lock->l_resource->lr_name,
1848 sizeof(lock->l_resource->lr_name)) != 0) {
1849 unlock_res_and_lock(lock);
1850 rc = ldlm_lock_change_resource(ns, lock,
1851 &dlm_req->lock_desc.l_resource.lr_name);
1853 LDLM_ERROR(lock, "Failed to allocate resource");
1856 LDLM_DEBUG(lock, "completion AST, new resource");
1857 CERROR("change resource!\n");
1858 lock_res_and_lock(lock);
1861 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1862 /* BL_AST locks are not needed in LRU.
1863 * Let ldlm_cancel_lru() be fast. */
1864 ldlm_lock_remove_from_lru(lock);
1865 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1866 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1869 if (lock->l_lvb_len > 0) {
1870 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1871 lock->l_lvb_data, lvb_len);
1873 unlock_res_and_lock(lock);
1878 ldlm_grant_lock(lock, &ast_list);
1879 unlock_res_and_lock(lock);
1881 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1883 /* Let Enqueue to call osc_lock_upcall() and initialize
1885 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1887 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1889 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1895 lock_res_and_lock(lock);
1896 ldlm_set_failed(lock);
1897 unlock_res_and_lock(lock);
1898 wake_up(&lock->l_waitq);
1900 LDLM_LOCK_RELEASE(lock);
1904 * Callback handler for receiving incoming glimpse ASTs.
1906 * This only can happen on client side. After handling the glimpse AST
1907 * we also consider dropping the lock here if it is unused locally for a
1910 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1911 struct ldlm_namespace *ns,
1912 struct ldlm_request *dlm_req,
1913 struct ldlm_lock *lock)
1918 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1920 if (lock->l_glimpse_ast != NULL)
1921 rc = lock->l_glimpse_ast(lock, req);
1923 if (req->rq_repmsg != NULL) {
1926 req->rq_status = rc;
1930 lock_res_and_lock(lock);
1931 if (lock->l_granted_mode == LCK_PW &&
1932 !lock->l_readers && !lock->l_writers &&
1933 cfs_time_after(cfs_time_current(),
1934 cfs_time_add(lock->l_last_used,
1935 cfs_time_seconds(10)))) {
1936 unlock_res_and_lock(lock);
1937 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1938 ldlm_handle_bl_callback(ns, NULL, lock);
1943 unlock_res_and_lock(lock);
1944 LDLM_LOCK_RELEASE(lock);
1948 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1950 if (req->rq_no_reply)
1953 req->rq_status = rc;
1954 if (!req->rq_packed_final) {
1955 rc = lustre_pack_reply(req, 1, NULL, NULL);
1959 return ptlrpc_reply(req);
1962 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
1963 enum ldlm_cancel_flags cancel_flags)
1965 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1968 spin_lock(&blp->blp_lock);
1969 if (blwi->blwi_lock &&
1970 ldlm_is_discard_data(blwi->blwi_lock)) {
1971 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1972 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1974 /* other blocking callbacks are added to the regular list */
1975 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1977 spin_unlock(&blp->blp_lock);
1979 wake_up(&blp->blp_waitq);
1981 /* can not check blwi->blwi_flags as blwi could be already freed in
1983 if (!(cancel_flags & LCF_ASYNC))
1984 wait_for_completion(&blwi->blwi_comp);
1989 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1990 struct ldlm_namespace *ns,
1991 struct ldlm_lock_desc *ld,
1992 struct list_head *cancels, int count,
1993 struct ldlm_lock *lock,
1994 enum ldlm_cancel_flags cancel_flags)
1996 init_completion(&blwi->blwi_comp);
1997 INIT_LIST_HEAD(&blwi->blwi_head);
1999 if (memory_pressure_get())
2000 blwi->blwi_mem_pressure = 1;
2003 blwi->blwi_flags = cancel_flags;
2005 blwi->blwi_ld = *ld;
2007 list_add(&blwi->blwi_head, cancels);
2008 list_del_init(cancels);
2009 blwi->blwi_count = count;
2011 blwi->blwi_lock = lock;
2016 * Queues a list of locks \a cancels containing \a count locks
2017 * for later processing by a blocking thread. If \a count is zero,
2018 * then the lock referenced as \a lock is queued instead.
2020 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2021 * If list addition fails an error is returned and caller is supposed to
2022 * call ->l_blocking_ast itself.
2024 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2025 struct ldlm_lock_desc *ld,
2026 struct ldlm_lock *lock,
2027 struct list_head *cancels, int count,
2028 enum ldlm_cancel_flags cancel_flags)
2032 if (cancels && count == 0)
2035 if (cancel_flags & LCF_ASYNC) {
2036 struct ldlm_bl_work_item *blwi;
2038 OBD_ALLOC(blwi, sizeof(*blwi));
2041 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2043 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2045 /* if it is synchronous call do minimum mem alloc, as it could
2046 * be triggered from kernel shrinker
2048 struct ldlm_bl_work_item blwi;
2050 memset(&blwi, 0, sizeof(blwi));
2051 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2052 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2057 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2058 struct ldlm_lock *lock)
2060 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2063 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2064 struct list_head *cancels, int count,
2065 enum ldlm_cancel_flags cancel_flags)
2067 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2070 int ldlm_bl_thread_wakeup(void)
2072 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2076 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2077 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2079 struct obd_device *obd = req->rq_export->exp_obd;
2086 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2088 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2090 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2092 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2095 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2097 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2099 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2102 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2105 /* We are responsible for swabbing contents of val */
2107 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2108 /* Pass it on to mdc (the "export" in this case) */
2109 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2111 sizeof(KEY_HSM_COPYTOOL_SEND),
2112 KEY_HSM_COPYTOOL_SEND,
2115 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2120 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2121 const char *msg, int rc,
2122 const struct lustre_handle *handle)
2124 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2125 "%s: [nid %s] [rc %d] [lock %#llx]",
2126 msg, libcfs_id2str(req->rq_peer), rc,
2127 handle ? handle->cookie : 0);
2128 if (req->rq_no_reply)
2129 CWARN("No reply was sent, maybe cause bug 21636.\n");
2131 CWARN("Send reply failed, maybe cause bug 21636.\n");
2134 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2135 static int ldlm_callback_handler(struct ptlrpc_request *req)
2137 struct ldlm_namespace *ns;
2138 struct ldlm_request *dlm_req;
2139 struct ldlm_lock *lock;
2143 /* Requests arrive in sender's byte order. The ptlrpc service
2144 * handler has already checked and, if necessary, byte-swapped the
2145 * incoming request message body, but I am responsible for the
2146 * message buffers. */
2148 /* do nothing for sec context finalize */
2149 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2152 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2154 if (req->rq_export == NULL) {
2155 rc = ldlm_callback_reply(req, -ENOTCONN);
2156 ldlm_callback_errmsg(req, "Operate on unconnected server",
2161 LASSERT(req->rq_export != NULL);
2162 LASSERT(req->rq_export->exp_obd != NULL);
2164 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2165 case LDLM_BL_CALLBACK:
2166 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2168 ldlm_callback_reply(req, -(int)cfs_fail_err);
2172 case LDLM_CP_CALLBACK:
2173 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2176 case LDLM_GL_CALLBACK:
2177 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2181 rc = ldlm_handle_setinfo(req);
2182 ldlm_callback_reply(req, rc);
2184 case LLOG_ORIGIN_HANDLE_CREATE:
2185 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
2186 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2188 rc = llog_origin_handle_open(req);
2189 ldlm_callback_reply(req, rc);
2191 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2192 req_capsule_set(&req->rq_pill,
2193 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
2194 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2196 rc = llog_origin_handle_next_block(req);
2197 ldlm_callback_reply(req, rc);
2199 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2200 req_capsule_set(&req->rq_pill,
2201 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
2202 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2204 rc = llog_origin_handle_read_header(req);
2205 ldlm_callback_reply(req, rc);
2207 case LLOG_ORIGIN_HANDLE_CLOSE:
2208 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2210 rc = llog_origin_handle_close(req);
2211 ldlm_callback_reply(req, rc);
2214 CERROR("unknown opcode %u\n",
2215 lustre_msg_get_opc(req->rq_reqmsg));
2216 ldlm_callback_reply(req, -EPROTO);
2220 ns = req->rq_export->exp_obd->obd_namespace;
2221 LASSERT(ns != NULL);
2223 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2225 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2226 if (dlm_req == NULL) {
2227 rc = ldlm_callback_reply(req, -EPROTO);
2228 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2233 /* Force a known safe race, send a cancel to the server for a lock
2234 * which the server has already started a blocking callback on. */
2235 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2236 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2237 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2239 CERROR("ldlm_cli_cancel: %d\n", rc);
2242 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2244 CDEBUG(D_DLMTRACE, "callback on lock %#llx - lock "
2245 "disappeared\n", dlm_req->lock_handle[0].cookie);
2246 rc = ldlm_callback_reply(req, -EINVAL);
2247 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2248 &dlm_req->lock_handle[0]);
2252 if (ldlm_is_fail_loc(lock) &&
2253 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2254 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2256 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2257 lock_res_and_lock(lock);
2258 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2260 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2261 /* If somebody cancels lock and cache is already dropped,
2262 * or lock is failed before cp_ast received on client,
2263 * we can tell the server we have no lock. Otherwise, we
2264 * should send cancel after dropping the cache. */
2265 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2266 ldlm_is_failed(lock)) {
2267 LDLM_DEBUG(lock, "callback on lock %llx - lock disappeared",
2268 dlm_req->lock_handle[0].cookie);
2269 unlock_res_and_lock(lock);
2270 LDLM_LOCK_RELEASE(lock);
2271 rc = ldlm_callback_reply(req, -EINVAL);
2272 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2273 &dlm_req->lock_handle[0]);
2276 /* BL_AST locks are not needed in LRU.
2277 * Let ldlm_cancel_lru() be fast. */
2278 ldlm_lock_remove_from_lru(lock);
2279 ldlm_set_bl_ast(lock);
2281 unlock_res_and_lock(lock);
2283 /* We want the ost thread to get this reply so that it can respond
2284 * to ost requests (write cache writeback) that might be triggered
2287 * But we'd also like to be able to indicate in the reply that we're
2288 * cancelling right now, because it's unused, or have an intent result
2289 * in the reply, so we might have to push the responsibility for sending
2290 * the reply down into the AST handlers, alas. */
2292 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2293 case LDLM_BL_CALLBACK:
2294 CDEBUG(D_INODE, "blocking ast\n");
2295 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2296 if (!ldlm_is_cancel_on_block(lock)) {
2297 rc = ldlm_callback_reply(req, 0);
2298 if (req->rq_no_reply || rc)
2299 ldlm_callback_errmsg(req, "Normal process", rc,
2300 &dlm_req->lock_handle[0]);
2302 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2303 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2305 case LDLM_CP_CALLBACK:
2306 CDEBUG(D_INODE, "completion ast\n");
2307 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2308 ldlm_callback_reply(req, 0);
2309 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2311 case LDLM_GL_CALLBACK:
2312 CDEBUG(D_INODE, "glimpse ast\n");
2313 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2314 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2317 LBUG(); /* checked above */
2323 #ifdef HAVE_SERVER_SUPPORT
2325 * Main handler for canceld thread.
2327 * Separated into its own thread to avoid deadlocks.
2329 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2334 /* Requests arrive in sender's byte order. The ptlrpc service
2335 * handler has already checked and, if necessary, byte-swapped the
2336 * incoming request message body, but I am responsible for the
2337 * message buffers. */
2339 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2341 if (req->rq_export == NULL) {
2342 struct ldlm_request *dlm_req;
2344 CERROR("%s from %s arrived at %lu with bad export cookie "
2346 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2347 libcfs_nid2str(req->rq_peer.nid),
2348 req->rq_arrival_time.tv_sec,
2349 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2351 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2352 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2353 dlm_req = req_capsule_client_get(&req->rq_pill,
2355 if (dlm_req != NULL)
2356 ldlm_lock_dump_handle(D_ERROR,
2357 &dlm_req->lock_handle[0]);
2359 ldlm_callback_reply(req, -ENOTCONN);
2363 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2365 /* XXX FIXME move this back to mds/handler.c, bug 249 */
2367 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2368 CDEBUG(D_INODE, "cancel\n");
2369 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2370 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2371 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2373 rc = ldlm_handle_cancel(req);
2378 CERROR("invalid opcode %d\n",
2379 lustre_msg_get_opc(req->rq_reqmsg));
2380 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2381 ldlm_callback_reply(req, -EINVAL);
2387 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2388 struct ldlm_lock *lock)
2390 struct ldlm_request *dlm_req;
2391 struct lustre_handle lockh;
2396 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2397 if (dlm_req == NULL)
2400 ldlm_lock2handle(lock, &lockh);
2401 for (i = 0; i < dlm_req->lock_count; i++) {
2402 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2404 DEBUG_REQ(D_RPCTRACE, req,
2405 "Prio raised by lock %#llx.", lockh.cookie);
2416 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2418 struct ldlm_request *dlm_req;
2423 /* no prolong in recovery */
2424 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2427 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2428 if (dlm_req == NULL)
2431 for (i = 0; i < dlm_req->lock_count; i++) {
2432 struct ldlm_lock *lock;
2434 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2438 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2440 LDLM_DEBUG(lock, "hpreq cancel lock");
2441 LDLM_LOCK_PUT(lock);
2450 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2451 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2452 .hpreq_check = ldlm_cancel_hpreq_check,
2456 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2460 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2462 if (req->rq_export == NULL)
2465 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2466 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2467 req->rq_ops = &ldlm_cancel_hpreq_ops;
2472 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2473 struct hlist_node *hnode, void *data)
2476 struct list_head *rpc_list = data;
2477 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2479 lock_res_and_lock(lock);
2481 if (lock->l_req_mode != lock->l_granted_mode) {
2482 unlock_res_and_lock(lock);
2486 LASSERT(lock->l_resource);
2487 if (lock->l_resource->lr_type != LDLM_IBITS &&
2488 lock->l_resource->lr_type != LDLM_PLAIN) {
2489 unlock_res_and_lock(lock);
2493 if (ldlm_is_ast_sent(lock)) {
2494 unlock_res_and_lock(lock);
2498 LASSERT(lock->l_blocking_ast);
2499 LASSERT(!lock->l_blocking_lock);
2501 ldlm_set_ast_sent(lock);
2502 if (lock->l_export && lock->l_export->exp_lock_hash) {
2503 /* NB: it's safe to call cfs_hash_del() even lock isn't
2504 * in exp_lock_hash. */
2505 /* In the function below, .hs_keycmp resolves to
2506 * ldlm_export_lock_keycmp() */
2507 /* coverity[overrun-buffer-val] */
2508 cfs_hash_del(lock->l_export->exp_lock_hash,
2509 &lock->l_remote_handle, &lock->l_exp_hash);
2512 list_add_tail(&lock->l_rk_ast, rpc_list);
2513 LDLM_LOCK_GET(lock);
2515 unlock_res_and_lock(lock);
2519 void ldlm_revoke_export_locks(struct obd_export *exp)
2521 struct list_head rpc_list;
2524 INIT_LIST_HEAD(&rpc_list);
2525 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2526 ldlm_revoke_lock_cb, &rpc_list, 0);
2527 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2528 LDLM_WORK_REVOKE_AST);
2532 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2533 #endif /* HAVE_SERVER_SUPPORT */
2535 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2536 struct ldlm_bl_work_item **p_blwi,
2537 struct obd_export **p_exp)
2539 struct ldlm_bl_work_item *blwi = NULL;
2540 static unsigned int num_bl = 0;
2541 static unsigned int num_stale;
2542 int num_th = atomic_read(&blp->blp_num_threads);
2544 *p_exp = obd_stale_export_get();
2546 spin_lock(&blp->blp_lock);
2547 if (*p_exp != NULL) {
2548 if (num_th == 1 || ++num_stale < num_th) {
2549 spin_unlock(&blp->blp_lock);
2556 /* process a request from the blp_list at least every blp_num_threads */
2557 if (!list_empty(&blp->blp_list) &&
2558 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2559 blwi = list_entry(blp->blp_list.next,
2560 struct ldlm_bl_work_item, blwi_entry);
2562 if (!list_empty(&blp->blp_prio_list))
2563 blwi = list_entry(blp->blp_prio_list.next,
2564 struct ldlm_bl_work_item,
2568 if (++num_bl >= num_th)
2570 list_del(&blwi->blwi_entry);
2572 spin_unlock(&blp->blp_lock);
2575 if (*p_exp != NULL && *p_blwi != NULL) {
2576 obd_stale_export_put(*p_exp);
2580 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2583 /* This only contains temporary data until the thread starts */
2584 struct ldlm_bl_thread_data {
2585 struct ldlm_bl_pool *bltd_blp;
2586 struct completion bltd_comp;
2590 static int ldlm_bl_thread_main(void *arg);
2592 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2594 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2595 struct task_struct *task;
2597 init_completion(&bltd.bltd_comp);
2599 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2600 if (bltd.bltd_num >= blp->blp_max_threads) {
2601 atomic_dec(&blp->blp_num_threads);
2605 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2607 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2608 atomic_dec(&blp->blp_num_threads);
2612 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2615 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2616 bltd.bltd_num, PTR_ERR(task));
2617 atomic_dec(&blp->blp_num_threads);
2618 return PTR_ERR(task);
2620 wait_for_completion(&bltd.bltd_comp);
2625 /* Not fatal if racy and have a few too many threads */
2626 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2627 struct ldlm_bl_work_item *blwi)
2629 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2632 if (atomic_read(&blp->blp_busy_threads) <
2633 atomic_read(&blp->blp_num_threads))
2636 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2637 blwi->blwi_mem_pressure))
2643 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2644 struct ldlm_bl_work_item *blwi)
2648 if (blwi->blwi_ns == NULL)
2649 /* added by ldlm_cleanup() */
2650 RETURN(LDLM_ITER_STOP);
2652 if (blwi->blwi_mem_pressure)
2653 memory_pressure_set();
2655 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2657 if (blwi->blwi_count) {
2659 /* The special case when we cancel locks in lru
2660 * asynchronously, we pass the list of locks here.
2661 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2662 * canceled locally yet. */
2663 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2666 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2669 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2672 if (blwi->blwi_mem_pressure)
2673 memory_pressure_clr();
2675 if (blwi->blwi_flags & LCF_ASYNC)
2676 OBD_FREE(blwi, sizeof(*blwi));
2678 complete(&blwi->blwi_comp);
2684 * Cancel stale locks on export. Cancel blocked locks first.
2685 * If the given export has blocked locks, the next in the list may have
2686 * them too, thus cancel not blocked locks only if the current export has
2689 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2690 struct obd_export *exp)
2695 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2697 num = ldlm_export_cancel_blocked_locks(exp);
2699 ldlm_export_cancel_locks(exp);
2701 obd_stale_export_put(exp);
2708 * Main blocking requests processing thread.
2710 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2711 * This thread in the end ends up doing actual call to ->l_blocking_ast
2714 static int ldlm_bl_thread_main(void *arg)
2716 struct ldlm_bl_pool *blp;
2717 struct ldlm_bl_thread_data *bltd = arg;
2720 blp = bltd->bltd_blp;
2722 complete(&bltd->bltd_comp);
2723 /* cannot use bltd after this, it is only on caller's stack */
2726 struct l_wait_info lwi = { 0 };
2727 struct ldlm_bl_work_item *blwi = NULL;
2728 struct obd_export *exp = NULL;
2731 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2734 l_wait_event_exclusive(blp->blp_waitq,
2735 ldlm_bl_get_work(blp, &blwi,
2738 atomic_inc(&blp->blp_busy_threads);
2740 if (ldlm_bl_thread_need_create(blp, blwi))
2741 /* discard the return value, we tried */
2742 ldlm_bl_thread_start(blp, true);
2745 rc = ldlm_bl_thread_exports(blp, exp);
2747 rc = ldlm_bl_thread_blwi(blp, blwi);
2749 atomic_dec(&blp->blp_busy_threads);
2751 if (rc == LDLM_ITER_STOP)
2755 atomic_dec(&blp->blp_num_threads);
2756 complete(&blp->blp_comp);
2761 static int ldlm_setup(void);
2762 static int ldlm_cleanup(void);
2764 int ldlm_get_ref(void)
2768 mutex_lock(&ldlm_ref_mutex);
2769 if (++ldlm_refcount == 1) {
2774 mutex_unlock(&ldlm_ref_mutex);
2779 void ldlm_put_ref(void)
2782 mutex_lock(&ldlm_ref_mutex);
2783 if (ldlm_refcount == 1) {
2784 int rc = ldlm_cleanup();
2786 CERROR("ldlm_cleanup failed: %d\n", rc);
2792 mutex_unlock(&ldlm_ref_mutex);
2798 * Export handle<->lock hash operations.
2801 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
2803 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2807 ldlm_export_lock_key(struct hlist_node *hnode)
2809 struct ldlm_lock *lock;
2811 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2812 return &lock->l_remote_handle;
2816 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
2818 struct ldlm_lock *lock;
2820 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2821 lock->l_remote_handle = *(struct lustre_handle *)key;
2825 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
2827 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2831 ldlm_export_lock_object(struct hlist_node *hnode)
2833 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2837 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
2839 struct ldlm_lock *lock;
2841 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2842 LDLM_LOCK_GET(lock);
2846 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
2848 struct ldlm_lock *lock;
2850 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2851 LDLM_LOCK_RELEASE(lock);
2854 static struct cfs_hash_ops ldlm_export_lock_ops = {
2855 .hs_hash = ldlm_export_lock_hash,
2856 .hs_key = ldlm_export_lock_key,
2857 .hs_keycmp = ldlm_export_lock_keycmp,
2858 .hs_keycpy = ldlm_export_lock_keycpy,
2859 .hs_object = ldlm_export_lock_object,
2860 .hs_get = ldlm_export_lock_get,
2861 .hs_put = ldlm_export_lock_put,
2862 .hs_put_locked = ldlm_export_lock_put,
2865 int ldlm_init_export(struct obd_export *exp)
2870 exp->exp_lock_hash =
2871 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2872 HASH_EXP_LOCK_CUR_BITS,
2873 HASH_EXP_LOCK_MAX_BITS,
2874 HASH_EXP_LOCK_BKT_BITS, 0,
2875 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2876 &ldlm_export_lock_ops,
2877 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2878 CFS_HASH_NBLK_CHANGE);
2880 if (!exp->exp_lock_hash)
2883 rc = ldlm_init_flock_export(exp);
2889 ldlm_destroy_export(exp);
2892 EXPORT_SYMBOL(ldlm_init_export);
2894 void ldlm_destroy_export(struct obd_export *exp)
2897 cfs_hash_putref(exp->exp_lock_hash);
2898 exp->exp_lock_hash = NULL;
2900 ldlm_destroy_flock_export(exp);
2903 EXPORT_SYMBOL(ldlm_destroy_export);
2905 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
2906 struct attribute *attr,
2909 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
2912 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
2913 struct attribute *attr,
2920 rc = kstrtoul(buffer, 10, &val);
2924 ldlm_cancel_unused_locks_before_replay = val;
2928 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
2930 static struct attribute *ldlm_attrs[] = {
2931 &lustre_attr_cancel_unused_locks_before_replay.attr,
2935 static struct attribute_group ldlm_attr_group = {
2936 .attrs = ldlm_attrs,
2939 static int ldlm_setup(void)
2941 static struct ptlrpc_service_conf conf;
2942 struct ldlm_bl_pool *blp = NULL;
2943 #ifdef HAVE_SERVER_SUPPORT
2944 struct task_struct *task;
2945 #endif /* HAVE_SERVER_SUPPORT */
2951 if (ldlm_state != NULL)
2954 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2955 if (ldlm_state == NULL)
2958 ldlm_kobj = kobject_create_and_add("ldlm", lustre_kobj);
2962 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
2966 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
2970 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
2974 #ifdef CONFIG_PROC_FS
2975 rc = ldlm_proc_setup();
2978 #endif /* CONFIG_PROC_FS */
2980 memset(&conf, 0, sizeof(conf));
2981 conf = (typeof(conf)) {
2982 .psc_name = "ldlm_cbd",
2983 .psc_watchdog_factor = 2,
2985 .bc_nbufs = LDLM_CLIENT_NBUFS,
2986 .bc_buf_size = LDLM_BUFSIZE,
2987 .bc_req_max_size = LDLM_MAXREQSIZE,
2988 .bc_rep_max_size = LDLM_MAXREPSIZE,
2989 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
2990 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
2993 .tc_thr_name = "ldlm_cb",
2994 .tc_thr_factor = LDLM_THR_FACTOR,
2995 .tc_nthrs_init = LDLM_NTHRS_INIT,
2996 .tc_nthrs_base = LDLM_NTHRS_BASE,
2997 .tc_nthrs_max = LDLM_NTHRS_MAX,
2998 .tc_nthrs_user = ldlm_num_threads,
2999 .tc_cpu_affinity = 1,
3000 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
3003 .cc_pattern = ldlm_cpts,
3006 .so_req_handler = ldlm_callback_handler,
3009 ldlm_state->ldlm_cb_service = \
3010 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
3011 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3012 CERROR("failed to start service\n");
3013 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3014 ldlm_state->ldlm_cb_service = NULL;
3018 #ifdef HAVE_SERVER_SUPPORT
3019 memset(&conf, 0, sizeof(conf));
3020 conf = (typeof(conf)) {
3021 .psc_name = "ldlm_canceld",
3022 .psc_watchdog_factor = 6,
3024 .bc_nbufs = LDLM_SERVER_NBUFS,
3025 .bc_buf_size = LDLM_BUFSIZE,
3026 .bc_req_max_size = LDLM_MAXREQSIZE,
3027 .bc_rep_max_size = LDLM_MAXREPSIZE,
3028 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3029 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3033 .tc_thr_name = "ldlm_cn",
3034 .tc_thr_factor = LDLM_THR_FACTOR,
3035 .tc_nthrs_init = LDLM_NTHRS_INIT,
3036 .tc_nthrs_base = LDLM_NTHRS_BASE,
3037 .tc_nthrs_max = LDLM_NTHRS_MAX,
3038 .tc_nthrs_user = ldlm_num_threads,
3039 .tc_cpu_affinity = 1,
3040 .tc_ctx_tags = LCT_MD_THREAD | \
3045 .cc_pattern = ldlm_cpts,
3048 .so_req_handler = ldlm_cancel_handler,
3049 .so_hpreq_handler = ldlm_hpreq_handler,
3052 ldlm_state->ldlm_cancel_service = \
3053 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
3054 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3055 CERROR("failed to start service\n");
3056 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3057 ldlm_state->ldlm_cancel_service = NULL;
3060 #endif /* HAVE_SERVER_SUPPORT */
3062 OBD_ALLOC(blp, sizeof(*blp));
3064 GOTO(out, rc = -ENOMEM);
3065 ldlm_state->ldlm_bl_pool = blp;
3067 spin_lock_init(&blp->blp_lock);
3068 INIT_LIST_HEAD(&blp->blp_list);
3069 INIT_LIST_HEAD(&blp->blp_prio_list);
3070 init_waitqueue_head(&blp->blp_waitq);
3071 atomic_set(&blp->blp_num_threads, 0);
3072 atomic_set(&blp->blp_busy_threads, 0);
3074 if (ldlm_num_threads == 0) {
3075 blp->blp_min_threads = LDLM_NTHRS_INIT;
3076 blp->blp_max_threads = LDLM_NTHRS_MAX;
3078 blp->blp_min_threads = blp->blp_max_threads = \
3079 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3083 for (i = 0; i < blp->blp_min_threads; i++) {
3084 rc = ldlm_bl_thread_start(blp, false);
3089 #ifdef HAVE_SERVER_SUPPORT
3090 INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
3091 expired_lock_thread.elt_state = ELT_STOPPED;
3092 init_waitqueue_head(&expired_lock_thread.elt_waitq);
3094 INIT_LIST_HEAD(&waiting_locks_list);
3095 spin_lock_init(&waiting_locks_spinlock);
3096 setup_timer(&waiting_locks_timer, waiting_locks_callback, 0);
3098 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3101 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3105 wait_event(expired_lock_thread.elt_waitq,
3106 expired_lock_thread.elt_state == ELT_READY);
3107 #endif /* HAVE_SERVER_SUPPORT */
3109 rc = ldlm_pools_init();
3111 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3115 rc = ldlm_reclaim_setup();
3117 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3127 static int ldlm_cleanup(void)
3131 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3132 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3133 CERROR("ldlm still has namespaces; clean these up first.\n");
3134 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3135 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3139 ldlm_reclaim_cleanup();
3142 if (ldlm_state->ldlm_bl_pool != NULL) {
3143 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3145 while (atomic_read(&blp->blp_num_threads) > 0) {
3146 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3148 init_completion(&blp->blp_comp);
3150 spin_lock(&blp->blp_lock);
3151 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3152 wake_up(&blp->blp_waitq);
3153 spin_unlock(&blp->blp_lock);
3155 wait_for_completion(&blp->blp_comp);
3158 OBD_FREE(blp, sizeof(*blp));
3161 if (ldlm_state->ldlm_cb_service != NULL)
3162 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3163 #ifdef HAVE_SERVER_SUPPORT
3164 if (ldlm_state->ldlm_cancel_service != NULL)
3165 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3169 kset_unregister(ldlm_ns_kset);
3171 kset_unregister(ldlm_svc_kset);
3173 kobject_put(ldlm_kobj);
3175 ldlm_proc_cleanup();
3177 #ifdef HAVE_SERVER_SUPPORT
3178 if (expired_lock_thread.elt_state != ELT_STOPPED) {
3179 expired_lock_thread.elt_state = ELT_TERMINATE;
3180 wake_up(&expired_lock_thread.elt_waitq);
3181 wait_event(expired_lock_thread.elt_waitq,
3182 expired_lock_thread.elt_state == ELT_STOPPED);
3186 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3194 mutex_init(&ldlm_ref_mutex);
3195 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
3196 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
3198 INIT_LIST_HEAD(&ldlm_srv_namespace_list);
3199 INIT_LIST_HEAD(&ldlm_cli_active_namespace_list);
3200 INIT_LIST_HEAD(&ldlm_cli_inactive_namespace_list);
3202 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3203 sizeof(struct ldlm_resource), 0,
3204 SLAB_HWCACHE_ALIGN, NULL);
3205 if (ldlm_resource_slab == NULL)
3208 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3209 sizeof(struct ldlm_lock), 0,
3210 SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU, NULL);
3211 if (ldlm_lock_slab == NULL)
3214 ldlm_interval_slab = kmem_cache_create("interval_node",
3215 sizeof(struct ldlm_interval),
3216 0, SLAB_HWCACHE_ALIGN, NULL);
3217 if (ldlm_interval_slab == NULL)
3220 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3221 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3222 0, SLAB_HWCACHE_ALIGN, NULL);
3223 if (ldlm_interval_tree_slab == NULL)
3226 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3227 class_export_dump_hook = ldlm_dump_export_locks;
3232 kmem_cache_destroy(ldlm_interval_slab);
3234 kmem_cache_destroy(ldlm_lock_slab);
3236 kmem_cache_destroy(ldlm_resource_slab);
3241 void ldlm_exit(void)
3244 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3245 kmem_cache_destroy(ldlm_resource_slab);
3246 /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3247 * synchronize_rcu() to wait a grace period elapsed, so that
3248 * ldlm_lock_free() get a chance to be called. */
3250 kmem_cache_destroy(ldlm_lock_slab);
3251 kmem_cache_destroy(ldlm_interval_slab);
3252 kmem_cache_destroy(ldlm_interval_tree_slab);