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);
672 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
674 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
675 /* Upon umount process the AST fails because cannot be
676 * sent. This shouldn't lead to the client eviction.
677 * -ENODEV error is returned by ptl_send_rpc() for
678 * new request in such import.
679 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
680 * if imp_invalid is set or obd_no_recov.
681 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
682 * in ptlrpc_import_delay_req() as well with -EIO code.
683 * In all such cases errors are ignored.
685 LDLM_DEBUG(lock, "%s AST can't be sent due to a server"
686 " %s failure or umount process: rc = %d\n",
688 req->rq_import->imp_obd->obd_name, rc);
691 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
692 libcfs_nid2str(peer.nid),
693 req->rq_replied ? "returned error from" :
694 "failed to reply to",
695 ast_type, req, req->rq_xid,
696 (req->rq_repmsg != NULL) ?
697 lustre_msg_get_status(req->rq_repmsg) : 0,
699 ldlm_failed_ast(lock, rc, ast_type);
705 struct ldlm_resource *res = lock->l_resource;
708 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
709 libcfs_nid2str(peer.nid),
711 lustre_msg_get_status(req->rq_repmsg) : -1,
712 ast_type, req, req->rq_xid);
714 /* update lvbo to return proper attributes.
716 ldlm_resource_getref(res);
717 ldlm_res_lvbo_update(res, NULL, 1);
718 ldlm_resource_putref(res);
720 ldlm_lock_cancel(lock);
727 static int ldlm_cb_interpret(const struct lu_env *env,
728 struct ptlrpc_request *req, void *data, int rc)
730 struct ldlm_cb_async_args *ca = data;
731 struct ldlm_lock *lock = ca->ca_lock;
732 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
735 LASSERT(lock != NULL);
738 case LDLM_GL_CALLBACK:
739 /* Update the LVB from disk if the AST failed
740 * (this is a legal race)
742 * - Glimpse callback of local lock just returns
743 * -ELDLM_NO_LOCK_DATA.
744 * - Glimpse callback of remote lock might return
745 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
747 if (unlikely(arg->gl_interpret_reply)) {
748 rc = arg->gl_interpret_reply(env, req, data, rc);
749 } else if (rc == -ELDLM_NO_LOCK_DATA) {
750 LDLM_DEBUG(lock, "lost race - client has a lock but no "
752 ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
753 } else if (rc != 0) {
754 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
756 rc = ldlm_res_lvbo_update(lock->l_resource, req, 1);
759 case LDLM_BL_CALLBACK:
761 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
763 case LDLM_CP_CALLBACK:
765 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
768 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
773 /* release extra reference taken in ldlm_ast_fini() */
774 LDLM_LOCK_RELEASE(lock);
777 atomic_inc(&arg->restart);
782 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
784 struct ldlm_cb_async_args *ca = data;
785 struct ldlm_lock *lock = ca->ca_lock;
787 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
790 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
791 struct ldlm_cb_set_arg *arg,
792 struct ldlm_lock *lock,
798 if (unlikely(instant_cancel)) {
799 rc = ptl_send_rpc(req, 1);
800 ptlrpc_req_finished(req);
802 atomic_inc(&arg->restart);
805 ptlrpc_set_add_req(arg->set, req);
812 * Check if there are requests in the export request list which prevent
813 * the lock canceling and make these requests high priority ones.
815 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
817 struct ptlrpc_request *req;
820 if (lock->l_export == NULL) {
821 LDLM_DEBUG(lock, "client lock: no-op");
825 spin_lock_bh(&lock->l_export->exp_rpc_lock);
826 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
828 /* Do not process requests that were not yet added to there
829 * incoming queue or were already removed from there for
830 * processing. We evaluate ptlrpc_nrs_req_can_move() without
831 * holding svcpt->scp_req_lock, and then redo the check with
832 * the lock held once we need to obtain a reliable result.
834 if (ptlrpc_nrs_req_can_move(req) &&
835 req->rq_ops->hpreq_lock_match &&
836 req->rq_ops->hpreq_lock_match(req, lock))
837 ptlrpc_nrs_req_hp_move(req);
839 spin_unlock_bh(&lock->l_export->exp_rpc_lock);
844 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
845 * enqueued server lock conflicts with given one.
847 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
848 * to wait for client response.
850 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
851 struct ldlm_lock_desc *desc,
852 void *data, int flag)
854 struct ldlm_cb_async_args *ca;
855 struct ldlm_cb_set_arg *arg = data;
856 struct ldlm_request *body;
857 struct ptlrpc_request *req;
858 int instant_cancel = 0;
862 if (flag == LDLM_CB_CANCELING)
863 /* Don't need to do anything here. */
866 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
867 LDLM_DEBUG(lock, "dropping BL AST");
872 LASSERT(data != NULL);
873 if (lock->l_export->exp_obd->obd_recovering != 0)
874 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
876 ldlm_lock_reorder_req(lock);
878 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
879 &RQF_LDLM_BL_CALLBACK,
880 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
884 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
885 ca = ptlrpc_req_async_args(req);
886 ca->ca_set_arg = arg;
889 req->rq_interpret_reply = ldlm_cb_interpret;
891 lock_res_and_lock(lock);
892 if (ldlm_is_destroyed(lock)) {
893 /* What's the point? */
894 unlock_res_and_lock(lock);
895 ptlrpc_req_finished(req);
899 if (lock->l_granted_mode != lock->l_req_mode) {
900 /* this blocking AST will be communicated as part of the
901 * completion AST instead */
902 ldlm_add_blocked_lock(lock);
903 ldlm_set_waited(lock);
904 unlock_res_and_lock(lock);
906 ptlrpc_req_finished(req);
907 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
911 if (ldlm_is_cancel_on_block(lock))
914 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
915 body->lock_handle[0] = lock->l_remote_handle;
916 body->lock_desc = *desc;
917 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
919 LDLM_DEBUG(lock, "server preparing blocking AST");
921 ptlrpc_request_set_replen(req);
922 ldlm_set_cbpending(lock);
923 if (instant_cancel) {
924 unlock_res_and_lock(lock);
925 ldlm_lock_cancel(lock);
927 req->rq_no_resend = 1;
929 LASSERT(lock->l_granted_mode == lock->l_req_mode);
930 ldlm_add_waiting_lock(lock);
931 unlock_res_and_lock(lock);
933 /* Do not resend after lock callback timeout */
934 req->rq_delay_limit = ldlm_bl_timeout(lock);
935 req->rq_resend_cb = ldlm_update_resend;
938 req->rq_send_state = LUSTRE_IMP_FULL;
939 /* ptlrpc_request_alloc_pack already set timeout */
941 req->rq_timeout = ldlm_get_rq_timeout();
943 lock->l_last_activity = ktime_get_real_seconds();
945 if (lock->l_export && lock->l_export->exp_nid_stats &&
946 lock->l_export->exp_nid_stats->nid_ldlm_stats)
947 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
948 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
950 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
956 * ->l_completion_ast callback for a remote lock in server namespace.
958 * Sends AST to the client notifying it of lock granting. If initial
959 * lock response was not sent yet, instead of sending another RPC, just
960 * mark the lock as granted and client will understand
962 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
964 struct ldlm_cb_set_arg *arg = data;
965 struct ldlm_request *body;
966 struct ptlrpc_request *req;
967 struct ldlm_cb_async_args *ca;
968 int instant_cancel = 0;
973 LASSERT(lock != NULL);
974 LASSERT(data != NULL);
976 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
977 LDLM_DEBUG(lock, "dropping CP AST");
981 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
982 &RQF_LDLM_CP_CALLBACK);
986 /* server namespace, doesn't need lock */
987 lvb_len = ldlm_lvbo_size(lock);
988 /* LU-3124 & LU-2187: to not return layout in completion AST because
989 * it may deadlock for LU-2187, or client may not have enough space
990 * for large layout. The layout will be returned to client with an
991 * extra RPC to fetch xattr.lov */
992 if (ldlm_has_layout(lock))
995 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
996 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
998 ptlrpc_request_free(req);
1002 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1003 ca = ptlrpc_req_async_args(req);
1004 ca->ca_set_arg = arg;
1007 req->rq_interpret_reply = ldlm_cb_interpret;
1008 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1010 body->lock_handle[0] = lock->l_remote_handle;
1011 body->lock_flags = ldlm_flags_to_wire(flags);
1012 ldlm_lock2desc(lock, &body->lock_desc);
1014 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1016 lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len);
1018 /* We still need to send the RPC to wake up the blocked
1019 * enqueue thread on the client.
1021 * Consider old client, there is no better way to notify
1022 * the failure, just zero-sized the LVB, then the client
1023 * will fail out as "-EPROTO". */
1024 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1028 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1033 lock->l_last_activity = ktime_get_real_seconds();
1035 LDLM_DEBUG(lock, "server preparing completion AST");
1037 ptlrpc_request_set_replen(req);
1039 req->rq_send_state = LUSTRE_IMP_FULL;
1040 /* ptlrpc_request_pack already set timeout */
1042 req->rq_timeout = ldlm_get_rq_timeout();
1044 /* We only send real blocking ASTs after the lock is granted */
1045 lock_res_and_lock(lock);
1046 if (ldlm_is_ast_sent(lock)) {
1047 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1048 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1049 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1052 /* We might get here prior to ldlm_handle_enqueue setting
1053 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1054 * into waiting list, but this is safe and similar code in
1055 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1056 * that would not only cancel the lock, but will also remove
1057 * it from waiting list */
1058 if (ldlm_is_cancel_on_block(lock)) {
1059 unlock_res_and_lock(lock);
1060 ldlm_lock_cancel(lock);
1063 req->rq_no_resend = 1;
1065 lock_res_and_lock(lock);
1067 /* start the lock-timeout clock */
1068 ldlm_add_waiting_lock(lock);
1069 /* Do not resend after lock callback timeout */
1070 req->rq_delay_limit = ldlm_bl_timeout(lock);
1071 req->rq_resend_cb = ldlm_update_resend;
1074 unlock_res_and_lock(lock);
1076 if (lock->l_export && lock->l_export->exp_nid_stats &&
1077 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1078 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1079 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1081 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1083 RETURN(lvb_len < 0 ? lvb_len : rc);
1087 * Server side ->l_glimpse_ast handler for client locks.
1089 * Sends glimpse AST to the client and waits for reply. Then updates
1090 * lvbo with the result.
1092 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1094 struct ldlm_cb_set_arg *arg = data;
1095 struct ldlm_request *body;
1096 struct ptlrpc_request *req;
1097 struct ldlm_cb_async_args *ca;
1099 struct req_format *req_fmt;
1102 LASSERT(lock != NULL);
1104 if (arg->gl_desc != NULL)
1105 /* There is a glimpse descriptor to pack */
1106 req_fmt = &RQF_LDLM_GL_DESC_CALLBACK;
1108 req_fmt = &RQF_LDLM_GL_CALLBACK;
1110 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1111 req_fmt, LUSTRE_DLM_VERSION,
1117 if (arg->gl_desc != NULL) {
1118 /* copy the GL descriptor */
1119 union ldlm_gl_desc *desc;
1120 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1121 *desc = *arg->gl_desc;
1124 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1125 body->lock_handle[0] = lock->l_remote_handle;
1126 ldlm_lock2desc(lock, &body->lock_desc);
1128 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1129 ca = ptlrpc_req_async_args(req);
1130 ca->ca_set_arg = arg;
1133 /* server namespace, doesn't need lock */
1134 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1135 ldlm_lvbo_size(lock));
1136 ptlrpc_request_set_replen(req);
1138 req->rq_send_state = LUSTRE_IMP_FULL;
1139 /* ptlrpc_request_alloc_pack already set timeout */
1141 req->rq_timeout = ldlm_get_rq_timeout();
1143 lock->l_last_activity = ktime_get_real_seconds();
1145 req->rq_interpret_reply = ldlm_cb_interpret;
1147 if (lock->l_export && lock->l_export->exp_nid_stats &&
1148 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1149 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1150 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1152 rc = ldlm_ast_fini(req, arg, lock, 0);
1157 int ldlm_glimpse_locks(struct ldlm_resource *res,
1158 struct list_head *gl_work_list)
1163 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1165 if (rc == -ERESTART)
1166 ldlm_reprocess_all(res);
1170 EXPORT_SYMBOL(ldlm_glimpse_locks);
1172 /* return LDLM lock associated with a lock callback request */
1173 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1175 struct ldlm_cb_async_args *ca;
1176 struct ldlm_lock *lock;
1179 ca = ptlrpc_req_async_args(req);
1182 RETURN(ERR_PTR(-EFAULT));
1186 EXPORT_SYMBOL(ldlm_request_lock);
1188 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
1189 struct lprocfs_stats *srv_stats)
1191 int lock_type = 0, op = 0;
1193 lock_type = dlm_req->lock_desc.l_resource.lr_type;
1195 switch (lock_type) {
1197 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
1200 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
1201 op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
1203 op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
1206 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
1209 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
1217 lprocfs_counter_incr(srv_stats, op);
1223 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1224 * service threads to carry out client lock enqueueing requests.
1226 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1227 struct ptlrpc_request *req,
1228 const struct ldlm_request *dlm_req,
1229 const struct ldlm_callback_suite *cbs)
1231 struct ldlm_reply *dlm_rep;
1233 enum ldlm_error err = ELDLM_OK;
1234 struct ldlm_lock *lock = NULL;
1235 void *cookie = NULL;
1237 struct ldlm_resource *res = NULL;
1240 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1242 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1243 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1245 LASSERT(req->rq_export);
1247 if (ptlrpc_req2svc(req)->srv_stats != NULL)
1248 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1250 if (req->rq_export && req->rq_export->exp_nid_stats &&
1251 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1252 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1253 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1255 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1256 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1257 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1258 dlm_req->lock_desc.l_resource.lr_type);
1259 GOTO(out, rc = -EFAULT);
1262 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1263 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1264 dlm_req->lock_desc.l_req_mode &
1265 (dlm_req->lock_desc.l_req_mode-1))) {
1266 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1267 dlm_req->lock_desc.l_req_mode);
1268 GOTO(out, rc = -EFAULT);
1271 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) {
1272 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1274 DEBUG_REQ(D_ERROR, req,
1275 "PLAIN lock request from IBITS client?");
1276 GOTO(out, rc = -EPROTO);
1278 } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1280 DEBUG_REQ(D_ERROR, req,
1281 "IBITS lock request from unaware client?");
1282 GOTO(out, rc = -EPROTO);
1285 if (unlikely((flags & LDLM_FL_REPLAY) ||
1286 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1287 /* Find an existing lock in the per-export lock hash */
1288 /* In the function below, .hs_keycmp resolves to
1289 * ldlm_export_lock_keycmp() */
1290 /* coverity[overrun-buffer-val] */
1291 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1292 (void *)&dlm_req->lock_handle[0]);
1294 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie %#llx",
1295 lock->l_handle.h_cookie);
1296 flags |= LDLM_FL_RESENT;
1297 GOTO(existing_lock, rc = 0);
1300 if (ldlm_reclaim_full()) {
1301 DEBUG_REQ(D_DLMTRACE, req, "Too many granted locks, "
1302 "reject current enqueue request and let the "
1303 "client retry later.\n");
1304 GOTO(out, rc = -EINPROGRESS);
1308 /* The lock's callback data might be set in the policy function */
1309 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1310 dlm_req->lock_desc.l_resource.lr_type,
1311 dlm_req->lock_desc.l_req_mode,
1312 cbs, NULL, 0, LVB_T_NONE);
1319 lock->l_remote_handle = dlm_req->lock_handle[0];
1320 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1322 /* Initialize resource lvb but not for a lock being replayed since
1323 * Client already got lvb sent in this case.
1324 * This must occur early since some policy methods assume resource
1325 * lvb is available (lr_lvb_data != NULL).
1327 res = lock->l_resource;
1328 if (!(flags & LDLM_FL_REPLAY)) {
1329 /* non-replayed lock, delayed lvb init may need to be done */
1330 rc = ldlm_lvbo_init(res);
1332 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1337 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1338 /* Don't enqueue a lock onto the export if it is been disonnected
1339 * due to eviction (bug 3822) or server umount (bug 24324).
1340 * Cancel it now instead. */
1341 if (req->rq_export->exp_disconnected) {
1342 LDLM_ERROR(lock, "lock on disconnected export %p",
1344 GOTO(out, rc = -ENOTCONN);
1347 lock->l_export = class_export_lock_get(req->rq_export, lock);
1348 if (lock->l_export->exp_lock_hash)
1349 cfs_hash_add(lock->l_export->exp_lock_hash,
1350 &lock->l_remote_handle,
1353 /* Inherit the enqueue flags before the operation, because we do not
1354 * keep the res lock on return and next operations (BL AST) may proceed
1356 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1357 LDLM_FL_INHERIT_MASK);
1359 ldlm_convert_policy_to_local(req->rq_export,
1360 dlm_req->lock_desc.l_resource.lr_type,
1361 &dlm_req->lock_desc.l_policy_data,
1362 &lock->l_policy_data);
1363 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1364 lock->l_req_extent = lock->l_policy_data.l_extent;
1368 if (flags & LDLM_FL_HAS_INTENT) {
1369 /* In this case, the reply buffer is allocated deep in
1370 * local_lock_enqueue by the policy function. */
1373 /* based on the assumption that lvb size never changes during
1374 * resource life time otherwise it need resource->lr_lock's
1376 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1377 RCL_SERVER, ldlm_lvbo_size(lock));
1379 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1380 GOTO(out, rc = -ENOMEM);
1382 rc = req_capsule_server_pack(&req->rq_pill);
1387 err = ldlm_lock_enqueue(ns, &lock, cookie, &flags);
1394 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1396 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1397 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1399 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1400 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1402 /* We never send a blocking AST until the lock is granted, but
1403 * we can tell it right now */
1404 lock_res_and_lock(lock);
1406 /* Now take into account flags to be inherited from original lock
1407 request both in reply to client and in our own lock flags. */
1408 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1409 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1411 /* Don't move a pending lock onto the export if it has already been
1412 * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1413 * Cancel it now instead. */
1414 if (unlikely(req->rq_export->exp_disconnected ||
1415 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1416 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1418 } else if (ldlm_is_ast_sent(lock)) {
1419 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1420 if (lock->l_granted_mode == lock->l_req_mode) {
1422 * Only cancel lock if it was granted, because it would
1423 * be destroyed immediately and would never be granted
1424 * in the future, causing timeouts on client. Not
1425 * granted lock will be cancelled immediately after
1426 * sending completion AST.
1428 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1429 unlock_res_and_lock(lock);
1430 ldlm_lock_cancel(lock);
1431 lock_res_and_lock(lock);
1433 ldlm_add_waiting_lock(lock);
1436 /* Make sure we never ever grant usual metadata locks to liblustre
1438 if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1439 dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1440 req->rq_export->exp_libclient) {
1441 if (unlikely(!ldlm_is_cancel_on_block(lock) ||
1442 !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1443 CERROR("Granting sync lock to libclient. "
1444 "req fl %d, rep fl %d, lock fl %#llx\n",
1445 dlm_req->lock_flags, dlm_rep->lock_flags,
1447 LDLM_ERROR(lock, "sync lock");
1448 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1449 struct ldlm_intent *it;
1451 it = req_capsule_client_get(&req->rq_pill,
1454 CERROR("This is intent %s (%llu)\n",
1455 ldlm_it2str(it->opc), it->opc);
1461 unlock_res_and_lock(lock);
1465 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1466 if (!req->rq_packed_final) {
1467 err = lustre_pack_reply(req, 1, NULL, NULL);
1472 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1473 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1475 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1476 "(err=%d, rc=%d)", err, rc);
1479 if (req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1481 ldlm_lvbo_size(lock) > 0) {
1485 buf = req_capsule_server_get(&req->rq_pill,
1487 LASSERTF(buf != NULL, "req %p, lock %p\n",
1489 buflen = req_capsule_get_size(&req->rq_pill,
1490 &RMF_DLM_LVB, RCL_SERVER);
1491 /* non-replayed lock, delayed lvb init may
1492 * need to be occur now */
1493 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1494 buflen = ldlm_lvbo_fill(lock, buf,
1500 buflen, RCL_SERVER);
1503 } else if (flags & LDLM_FL_REPLAY) {
1504 /* no LVB resend upon replay */
1518 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1519 if (lock->l_export) {
1520 ldlm_lock_cancel(lock);
1522 lock_res_and_lock(lock);
1523 ldlm_resource_unlink_lock(lock);
1524 ldlm_lock_destroy_nolock(lock);
1525 unlock_res_and_lock(lock);
1530 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1531 ldlm_reprocess_all(lock->l_resource);
1533 LDLM_LOCK_RELEASE(lock);
1536 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1543 * Old-style LDLM main entry point for server code enqueue.
1545 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1546 ldlm_completion_callback completion_callback,
1547 ldlm_blocking_callback blocking_callback,
1548 ldlm_glimpse_callback glimpse_callback)
1550 struct ldlm_request *dlm_req;
1551 struct ldlm_callback_suite cbs = {
1552 .lcs_completion = completion_callback,
1553 .lcs_blocking = blocking_callback,
1554 .lcs_glimpse = glimpse_callback
1558 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1559 if (dlm_req != NULL) {
1560 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1561 req, dlm_req, &cbs);
1569 * Main LDLM entry point for server code to process lock conversion requests.
1571 int ldlm_handle_convert0(struct ptlrpc_request *req,
1572 const struct ldlm_request *dlm_req)
1574 struct ldlm_reply *dlm_rep;
1575 struct ldlm_lock *lock;
1579 if (req->rq_export && req->rq_export->exp_nid_stats &&
1580 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1581 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1582 LDLM_CONVERT - LDLM_FIRST_OPC);
1584 rc = req_capsule_server_pack(&req->rq_pill);
1588 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1589 dlm_rep->lock_flags = dlm_req->lock_flags;
1591 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1593 req->rq_status = LUSTRE_EINVAL;
1597 LDLM_DEBUG(lock, "server-side convert handler START");
1599 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1600 &dlm_rep->lock_flags);
1602 if (ldlm_del_waiting_lock(lock))
1603 LDLM_DEBUG(lock, "converted waiting lock");
1606 req->rq_status = LUSTRE_EDEADLK;
1611 if (!req->rq_status)
1612 ldlm_reprocess_all(lock->l_resource);
1613 LDLM_DEBUG(lock, "server-side convert handler END");
1614 LDLM_LOCK_PUT(lock);
1616 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1622 * Old-style main LDLM entry point for server code to process lock conversion
1625 int ldlm_handle_convert(struct ptlrpc_request *req)
1628 struct ldlm_request *dlm_req;
1630 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1631 if (dlm_req != NULL) {
1632 rc = ldlm_handle_convert0(req, dlm_req);
1634 CERROR ("Can't unpack dlm_req\n");
1641 * Cancel all the locks whose handles are packed into ldlm_request
1643 * Called by server code expecting such combined cancel activity
1646 int ldlm_request_cancel(struct ptlrpc_request *req,
1647 const struct ldlm_request *dlm_req,
1648 int first, enum lustre_at_flags flags)
1650 struct ldlm_resource *res, *pres = NULL;
1651 struct ldlm_lock *lock;
1652 int i, count, done = 0;
1655 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1659 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1662 /* There is no lock on the server at the replay time,
1663 * skip lock cancelling to make replay tests to pass. */
1664 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1667 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1668 "starting at %d", count, first);
1670 for (i = first; i < count; i++) {
1671 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1673 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1674 "lock (cookie %llu)",
1675 dlm_req->lock_handle[i].cookie);
1679 res = lock->l_resource;
1682 /* This code is an optimization to only attempt lock
1683 * granting on the resource (that could be CPU-expensive)
1684 * after we are done cancelling lock in that resource. */
1687 ldlm_reprocess_all(pres);
1688 LDLM_RESOURCE_DELREF(pres);
1689 ldlm_resource_putref(pres);
1692 ldlm_resource_getref(res);
1693 LDLM_RESOURCE_ADDREF(res);
1694 ldlm_res_lvbo_update(res, NULL, 1);
1699 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock)) {
1700 time64_t delay = ktime_get_real_seconds() -
1701 lock->l_last_activity;
1702 LDLM_DEBUG(lock, "server cancels blocked lock after %llds",
1704 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1706 ldlm_lock_cancel(lock);
1707 LDLM_LOCK_PUT(lock);
1710 ldlm_reprocess_all(pres);
1711 LDLM_RESOURCE_DELREF(pres);
1712 ldlm_resource_putref(pres);
1714 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1717 EXPORT_SYMBOL(ldlm_request_cancel);
1720 * Main LDLM entry point for server code to cancel locks.
1722 * Typically gets called from service handler on LDLM_CANCEL opc.
1724 int ldlm_handle_cancel(struct ptlrpc_request *req)
1726 struct ldlm_request *dlm_req;
1730 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1731 if (dlm_req == NULL) {
1732 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1736 if (req->rq_export && req->rq_export->exp_nid_stats &&
1737 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1738 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1739 LDLM_CANCEL - LDLM_FIRST_OPC);
1741 rc = req_capsule_server_pack(&req->rq_pill);
1745 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1746 req->rq_status = LUSTRE_ESTALE;
1748 RETURN(ptlrpc_reply(req));
1750 #endif /* HAVE_SERVER_SUPPORT */
1753 * Callback handler for receiving incoming blocking ASTs.
1755 * This can only happen on client side.
1757 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1758 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1763 LDLM_DEBUG(lock, "client blocking AST callback handler");
1765 lock_res_and_lock(lock);
1766 ldlm_set_cbpending(lock);
1768 if (ldlm_is_cancel_on_block(lock))
1769 ldlm_set_cancel(lock);
1771 do_ast = (!lock->l_readers && !lock->l_writers);
1772 unlock_res_and_lock(lock);
1775 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1776 lock, lock->l_blocking_ast);
1777 if (lock->l_blocking_ast != NULL)
1778 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1781 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1785 LDLM_DEBUG(lock, "client blocking callback handler END");
1786 LDLM_LOCK_RELEASE(lock);
1791 * Callback handler for receiving incoming completion ASTs.
1793 * This only can happen on client side.
1795 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1796 struct ldlm_namespace *ns,
1797 struct ldlm_request *dlm_req,
1798 struct ldlm_lock *lock)
1800 struct list_head ast_list;
1805 LDLM_DEBUG(lock, "client completion callback handler START");
1807 INIT_LIST_HEAD(&ast_list);
1808 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1809 int to = cfs_time_seconds(1);
1811 set_current_state(TASK_INTERRUPTIBLE);
1812 schedule_timeout(to);
1813 if (lock->l_granted_mode == lock->l_req_mode ||
1814 ldlm_is_destroyed(lock))
1819 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1821 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1822 GOTO(out, rc = lvb_len);
1823 } else if (lvb_len > 0) {
1824 if (lock->l_lvb_len > 0) {
1825 /* for extent lock, lvb contains ost_lvb{}. */
1826 LASSERT(lock->l_lvb_data != NULL);
1828 if (unlikely(lock->l_lvb_len < lvb_len)) {
1829 LDLM_ERROR(lock, "Replied LVB is larger than "
1830 "expectation, expected = %d, "
1832 lock->l_lvb_len, lvb_len);
1833 GOTO(out, rc = -EINVAL);
1838 lock_res_and_lock(lock);
1839 if (ldlm_is_destroyed(lock) ||
1840 lock->l_granted_mode == lock->l_req_mode) {
1841 /* bug 11300: the lock has already been granted */
1842 unlock_res_and_lock(lock);
1843 LDLM_DEBUG(lock, "Double grant race happened");
1847 /* If we receive the completion AST before the actual enqueue returned,
1848 * then we might need to switch lock modes, resources, or extents. */
1849 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1850 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1851 LDLM_DEBUG(lock, "completion AST, new lock mode");
1854 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1855 ldlm_convert_policy_to_local(req->rq_export,
1856 dlm_req->lock_desc.l_resource.lr_type,
1857 &dlm_req->lock_desc.l_policy_data,
1858 &lock->l_policy_data);
1859 LDLM_DEBUG(lock, "completion AST, new policy data");
1862 ldlm_resource_unlink_lock(lock);
1863 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1864 &lock->l_resource->lr_name,
1865 sizeof(lock->l_resource->lr_name)) != 0) {
1866 unlock_res_and_lock(lock);
1867 rc = ldlm_lock_change_resource(ns, lock,
1868 &dlm_req->lock_desc.l_resource.lr_name);
1870 LDLM_ERROR(lock, "Failed to allocate resource");
1873 LDLM_DEBUG(lock, "completion AST, new resource");
1874 CERROR("change resource!\n");
1875 lock_res_and_lock(lock);
1878 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1879 /* BL_AST locks are not needed in LRU.
1880 * Let ldlm_cancel_lru() be fast. */
1881 ldlm_lock_remove_from_lru(lock);
1882 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1883 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1886 if (lock->l_lvb_len > 0) {
1887 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1888 lock->l_lvb_data, lvb_len);
1890 unlock_res_and_lock(lock);
1895 ldlm_grant_lock(lock, &ast_list);
1896 unlock_res_and_lock(lock);
1898 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1900 /* Let Enqueue to call osc_lock_upcall() and initialize
1902 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1904 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1906 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1912 lock_res_and_lock(lock);
1913 ldlm_set_failed(lock);
1914 unlock_res_and_lock(lock);
1915 wake_up(&lock->l_waitq);
1917 LDLM_LOCK_RELEASE(lock);
1921 * Callback handler for receiving incoming glimpse ASTs.
1923 * This only can happen on client side. After handling the glimpse AST
1924 * we also consider dropping the lock here if it is unused locally for a
1927 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1928 struct ldlm_namespace *ns,
1929 struct ldlm_request *dlm_req,
1930 struct ldlm_lock *lock)
1935 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1937 if (lock->l_glimpse_ast != NULL)
1938 rc = lock->l_glimpse_ast(lock, req);
1940 if (req->rq_repmsg != NULL) {
1943 req->rq_status = rc;
1947 lock_res_and_lock(lock);
1948 if (lock->l_granted_mode == LCK_PW &&
1949 !lock->l_readers && !lock->l_writers &&
1950 cfs_time_after(cfs_time_current(),
1951 cfs_time_add(lock->l_last_used,
1952 cfs_time_seconds(10)))) {
1953 unlock_res_and_lock(lock);
1954 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1955 ldlm_handle_bl_callback(ns, NULL, lock);
1960 unlock_res_and_lock(lock);
1961 LDLM_LOCK_RELEASE(lock);
1965 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1967 if (req->rq_no_reply)
1970 req->rq_status = rc;
1971 if (!req->rq_packed_final) {
1972 rc = lustre_pack_reply(req, 1, NULL, NULL);
1976 return ptlrpc_reply(req);
1979 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
1980 enum ldlm_cancel_flags cancel_flags)
1982 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1985 spin_lock(&blp->blp_lock);
1986 if (blwi->blwi_lock &&
1987 ldlm_is_discard_data(blwi->blwi_lock)) {
1988 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1989 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1991 /* other blocking callbacks are added to the regular list */
1992 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1994 spin_unlock(&blp->blp_lock);
1996 wake_up(&blp->blp_waitq);
1998 /* can not check blwi->blwi_flags as blwi could be already freed in
2000 if (!(cancel_flags & LCF_ASYNC))
2001 wait_for_completion(&blwi->blwi_comp);
2006 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2007 struct ldlm_namespace *ns,
2008 struct ldlm_lock_desc *ld,
2009 struct list_head *cancels, int count,
2010 struct ldlm_lock *lock,
2011 enum ldlm_cancel_flags cancel_flags)
2013 init_completion(&blwi->blwi_comp);
2014 INIT_LIST_HEAD(&blwi->blwi_head);
2016 if (memory_pressure_get())
2017 blwi->blwi_mem_pressure = 1;
2020 blwi->blwi_flags = cancel_flags;
2022 blwi->blwi_ld = *ld;
2024 list_add(&blwi->blwi_head, cancels);
2025 list_del_init(cancels);
2026 blwi->blwi_count = count;
2028 blwi->blwi_lock = lock;
2033 * Queues a list of locks \a cancels containing \a count locks
2034 * for later processing by a blocking thread. If \a count is zero,
2035 * then the lock referenced as \a lock is queued instead.
2037 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2038 * If list addition fails an error is returned and caller is supposed to
2039 * call ->l_blocking_ast itself.
2041 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2042 struct ldlm_lock_desc *ld,
2043 struct ldlm_lock *lock,
2044 struct list_head *cancels, int count,
2045 enum ldlm_cancel_flags cancel_flags)
2049 if (cancels && count == 0)
2052 if (cancel_flags & LCF_ASYNC) {
2053 struct ldlm_bl_work_item *blwi;
2055 OBD_ALLOC(blwi, sizeof(*blwi));
2058 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2060 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2062 /* if it is synchronous call do minimum mem alloc, as it could
2063 * be triggered from kernel shrinker
2065 struct ldlm_bl_work_item blwi;
2067 memset(&blwi, 0, sizeof(blwi));
2068 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2069 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2074 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2075 struct ldlm_lock *lock)
2077 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2080 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2081 struct list_head *cancels, int count,
2082 enum ldlm_cancel_flags cancel_flags)
2084 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2087 int ldlm_bl_thread_wakeup(void)
2089 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2093 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2094 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2096 struct obd_device *obd = req->rq_export->exp_obd;
2103 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2105 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2107 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2109 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2112 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2114 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2116 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2119 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2122 /* We are responsible for swabbing contents of val */
2124 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2125 /* Pass it on to mdc (the "export" in this case) */
2126 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2128 sizeof(KEY_HSM_COPYTOOL_SEND),
2129 KEY_HSM_COPYTOOL_SEND,
2132 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2137 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2138 const char *msg, int rc,
2139 const struct lustre_handle *handle)
2141 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2142 "%s: [nid %s] [rc %d] [lock %#llx]",
2143 msg, libcfs_id2str(req->rq_peer), rc,
2144 handle ? handle->cookie : 0);
2145 if (req->rq_no_reply)
2146 CWARN("No reply was sent, maybe cause bug 21636.\n");
2148 CWARN("Send reply failed, maybe cause bug 21636.\n");
2151 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2152 static int ldlm_callback_handler(struct ptlrpc_request *req)
2154 struct ldlm_namespace *ns;
2155 struct ldlm_request *dlm_req;
2156 struct ldlm_lock *lock;
2160 /* Requests arrive in sender's byte order. The ptlrpc service
2161 * handler has already checked and, if necessary, byte-swapped the
2162 * incoming request message body, but I am responsible for the
2163 * message buffers. */
2165 /* do nothing for sec context finalize */
2166 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2169 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2171 if (req->rq_export == NULL) {
2172 rc = ldlm_callback_reply(req, -ENOTCONN);
2173 ldlm_callback_errmsg(req, "Operate on unconnected server",
2178 LASSERT(req->rq_export != NULL);
2179 LASSERT(req->rq_export->exp_obd != NULL);
2181 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2182 case LDLM_BL_CALLBACK:
2183 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2185 ldlm_callback_reply(req, -(int)cfs_fail_err);
2189 case LDLM_CP_CALLBACK:
2190 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2193 case LDLM_GL_CALLBACK:
2194 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2198 rc = ldlm_handle_setinfo(req);
2199 ldlm_callback_reply(req, rc);
2201 case LLOG_ORIGIN_HANDLE_CREATE:
2202 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
2203 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2205 rc = llog_origin_handle_open(req);
2206 ldlm_callback_reply(req, rc);
2208 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2209 req_capsule_set(&req->rq_pill,
2210 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
2211 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2213 rc = llog_origin_handle_next_block(req);
2214 ldlm_callback_reply(req, rc);
2216 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2217 req_capsule_set(&req->rq_pill,
2218 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
2219 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2221 rc = llog_origin_handle_read_header(req);
2222 ldlm_callback_reply(req, rc);
2224 case LLOG_ORIGIN_HANDLE_CLOSE:
2225 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2227 rc = llog_origin_handle_close(req);
2228 ldlm_callback_reply(req, rc);
2231 CERROR("unknown opcode %u\n",
2232 lustre_msg_get_opc(req->rq_reqmsg));
2233 ldlm_callback_reply(req, -EPROTO);
2237 ns = req->rq_export->exp_obd->obd_namespace;
2238 LASSERT(ns != NULL);
2240 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2242 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2243 if (dlm_req == NULL) {
2244 rc = ldlm_callback_reply(req, -EPROTO);
2245 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2250 /* Force a known safe race, send a cancel to the server for a lock
2251 * which the server has already started a blocking callback on. */
2252 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2253 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2254 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2256 CERROR("ldlm_cli_cancel: %d\n", rc);
2259 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2261 CDEBUG(D_DLMTRACE, "callback on lock %#llx - lock "
2262 "disappeared\n", dlm_req->lock_handle[0].cookie);
2263 rc = ldlm_callback_reply(req, -EINVAL);
2264 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2265 &dlm_req->lock_handle[0]);
2269 if (ldlm_is_fail_loc(lock) &&
2270 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2271 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2273 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2274 lock_res_and_lock(lock);
2275 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2277 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2278 /* If somebody cancels lock and cache is already dropped,
2279 * or lock is failed before cp_ast received on client,
2280 * we can tell the server we have no lock. Otherwise, we
2281 * should send cancel after dropping the cache. */
2282 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2283 ldlm_is_failed(lock)) {
2284 LDLM_DEBUG(lock, "callback on lock %llx - lock disappeared",
2285 dlm_req->lock_handle[0].cookie);
2286 unlock_res_and_lock(lock);
2287 LDLM_LOCK_RELEASE(lock);
2288 rc = ldlm_callback_reply(req, -EINVAL);
2289 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2290 &dlm_req->lock_handle[0]);
2293 /* BL_AST locks are not needed in LRU.
2294 * Let ldlm_cancel_lru() be fast. */
2295 ldlm_lock_remove_from_lru(lock);
2296 ldlm_set_bl_ast(lock);
2298 unlock_res_and_lock(lock);
2300 /* We want the ost thread to get this reply so that it can respond
2301 * to ost requests (write cache writeback) that might be triggered
2304 * But we'd also like to be able to indicate in the reply that we're
2305 * cancelling right now, because it's unused, or have an intent result
2306 * in the reply, so we might have to push the responsibility for sending
2307 * the reply down into the AST handlers, alas. */
2309 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2310 case LDLM_BL_CALLBACK:
2311 CDEBUG(D_INODE, "blocking ast\n");
2312 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2313 if (!ldlm_is_cancel_on_block(lock)) {
2314 rc = ldlm_callback_reply(req, 0);
2315 if (req->rq_no_reply || rc)
2316 ldlm_callback_errmsg(req, "Normal process", rc,
2317 &dlm_req->lock_handle[0]);
2319 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2320 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2322 case LDLM_CP_CALLBACK:
2323 CDEBUG(D_INODE, "completion ast\n");
2324 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2325 ldlm_callback_reply(req, 0);
2326 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2328 case LDLM_GL_CALLBACK:
2329 CDEBUG(D_INODE, "glimpse ast\n");
2330 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2331 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2334 LBUG(); /* checked above */
2340 #ifdef HAVE_SERVER_SUPPORT
2342 * Main handler for canceld thread.
2344 * Separated into its own thread to avoid deadlocks.
2346 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2351 /* Requests arrive in sender's byte order. The ptlrpc service
2352 * handler has already checked and, if necessary, byte-swapped the
2353 * incoming request message body, but I am responsible for the
2354 * message buffers. */
2356 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2358 if (req->rq_export == NULL) {
2359 struct ldlm_request *dlm_req;
2361 CERROR("%s from %s arrived at %lu with bad export cookie "
2363 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2364 libcfs_nid2str(req->rq_peer.nid),
2365 req->rq_arrival_time.tv_sec,
2366 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2368 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2369 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2370 dlm_req = req_capsule_client_get(&req->rq_pill,
2372 if (dlm_req != NULL)
2373 ldlm_lock_dump_handle(D_ERROR,
2374 &dlm_req->lock_handle[0]);
2376 ldlm_callback_reply(req, -ENOTCONN);
2380 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2382 /* XXX FIXME move this back to mds/handler.c, bug 249 */
2384 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2385 CDEBUG(D_INODE, "cancel\n");
2386 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2387 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2388 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2390 rc = ldlm_handle_cancel(req);
2395 CERROR("invalid opcode %d\n",
2396 lustre_msg_get_opc(req->rq_reqmsg));
2397 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2398 ldlm_callback_reply(req, -EINVAL);
2404 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2405 struct ldlm_lock *lock)
2407 struct ldlm_request *dlm_req;
2408 struct lustre_handle lockh;
2413 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2414 if (dlm_req == NULL)
2417 ldlm_lock2handle(lock, &lockh);
2418 for (i = 0; i < dlm_req->lock_count; i++) {
2419 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2421 DEBUG_REQ(D_RPCTRACE, req,
2422 "Prio raised by lock %#llx.", lockh.cookie);
2433 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2435 struct ldlm_request *dlm_req;
2440 /* no prolong in recovery */
2441 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2444 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2445 if (dlm_req == NULL)
2448 for (i = 0; i < dlm_req->lock_count; i++) {
2449 struct ldlm_lock *lock;
2451 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2455 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2457 LDLM_DEBUG(lock, "hpreq cancel lock");
2458 LDLM_LOCK_PUT(lock);
2467 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2468 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2469 .hpreq_check = ldlm_cancel_hpreq_check,
2473 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2477 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2479 if (req->rq_export == NULL)
2482 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2483 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2484 req->rq_ops = &ldlm_cancel_hpreq_ops;
2489 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2490 struct hlist_node *hnode, void *data)
2493 struct list_head *rpc_list = data;
2494 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2496 lock_res_and_lock(lock);
2498 if (lock->l_req_mode != lock->l_granted_mode) {
2499 unlock_res_and_lock(lock);
2503 LASSERT(lock->l_resource);
2504 if (lock->l_resource->lr_type != LDLM_IBITS &&
2505 lock->l_resource->lr_type != LDLM_PLAIN) {
2506 unlock_res_and_lock(lock);
2510 if (ldlm_is_ast_sent(lock)) {
2511 unlock_res_and_lock(lock);
2515 LASSERT(lock->l_blocking_ast);
2516 LASSERT(!lock->l_blocking_lock);
2518 ldlm_set_ast_sent(lock);
2519 if (lock->l_export && lock->l_export->exp_lock_hash) {
2520 /* NB: it's safe to call cfs_hash_del() even lock isn't
2521 * in exp_lock_hash. */
2522 /* In the function below, .hs_keycmp resolves to
2523 * ldlm_export_lock_keycmp() */
2524 /* coverity[overrun-buffer-val] */
2525 cfs_hash_del(lock->l_export->exp_lock_hash,
2526 &lock->l_remote_handle, &lock->l_exp_hash);
2529 list_add_tail(&lock->l_rk_ast, rpc_list);
2530 LDLM_LOCK_GET(lock);
2532 unlock_res_and_lock(lock);
2536 void ldlm_revoke_export_locks(struct obd_export *exp)
2538 struct list_head rpc_list;
2541 INIT_LIST_HEAD(&rpc_list);
2542 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2543 ldlm_revoke_lock_cb, &rpc_list, 0);
2544 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2545 LDLM_WORK_REVOKE_AST);
2549 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2550 #endif /* HAVE_SERVER_SUPPORT */
2552 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2553 struct ldlm_bl_work_item **p_blwi,
2554 struct obd_export **p_exp)
2556 struct ldlm_bl_work_item *blwi = NULL;
2557 static unsigned int num_bl = 0;
2558 static unsigned int num_stale;
2559 int num_th = atomic_read(&blp->blp_num_threads);
2561 *p_exp = obd_stale_export_get();
2563 spin_lock(&blp->blp_lock);
2564 if (*p_exp != NULL) {
2565 if (num_th == 1 || ++num_stale < num_th) {
2566 spin_unlock(&blp->blp_lock);
2573 /* process a request from the blp_list at least every blp_num_threads */
2574 if (!list_empty(&blp->blp_list) &&
2575 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2576 blwi = list_entry(blp->blp_list.next,
2577 struct ldlm_bl_work_item, blwi_entry);
2579 if (!list_empty(&blp->blp_prio_list))
2580 blwi = list_entry(blp->blp_prio_list.next,
2581 struct ldlm_bl_work_item,
2585 if (++num_bl >= num_th)
2587 list_del(&blwi->blwi_entry);
2589 spin_unlock(&blp->blp_lock);
2592 if (*p_exp != NULL && *p_blwi != NULL) {
2593 obd_stale_export_put(*p_exp);
2597 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2600 /* This only contains temporary data until the thread starts */
2601 struct ldlm_bl_thread_data {
2602 struct ldlm_bl_pool *bltd_blp;
2603 struct completion bltd_comp;
2607 static int ldlm_bl_thread_main(void *arg);
2609 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2611 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2612 struct task_struct *task;
2614 init_completion(&bltd.bltd_comp);
2616 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2617 if (bltd.bltd_num >= blp->blp_max_threads) {
2618 atomic_dec(&blp->blp_num_threads);
2622 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2624 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2625 atomic_dec(&blp->blp_num_threads);
2629 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2632 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2633 bltd.bltd_num, PTR_ERR(task));
2634 atomic_dec(&blp->blp_num_threads);
2635 return PTR_ERR(task);
2637 wait_for_completion(&bltd.bltd_comp);
2642 /* Not fatal if racy and have a few too many threads */
2643 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2644 struct ldlm_bl_work_item *blwi)
2646 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2649 if (atomic_read(&blp->blp_busy_threads) <
2650 atomic_read(&blp->blp_num_threads))
2653 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2654 blwi->blwi_mem_pressure))
2660 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2661 struct ldlm_bl_work_item *blwi)
2665 if (blwi->blwi_ns == NULL)
2666 /* added by ldlm_cleanup() */
2667 RETURN(LDLM_ITER_STOP);
2669 if (blwi->blwi_mem_pressure)
2670 memory_pressure_set();
2672 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2674 if (blwi->blwi_count) {
2676 /* The special case when we cancel locks in lru
2677 * asynchronously, we pass the list of locks here.
2678 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2679 * canceled locally yet. */
2680 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2683 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2686 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2689 if (blwi->blwi_mem_pressure)
2690 memory_pressure_clr();
2692 if (blwi->blwi_flags & LCF_ASYNC)
2693 OBD_FREE(blwi, sizeof(*blwi));
2695 complete(&blwi->blwi_comp);
2701 * Cancel stale locks on export. Cancel blocked locks first.
2702 * If the given export has blocked locks, the next in the list may have
2703 * them too, thus cancel not blocked locks only if the current export has
2706 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2707 struct obd_export *exp)
2712 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2714 num = ldlm_export_cancel_blocked_locks(exp);
2716 ldlm_export_cancel_locks(exp);
2718 obd_stale_export_put(exp);
2725 * Main blocking requests processing thread.
2727 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2728 * This thread in the end ends up doing actual call to ->l_blocking_ast
2731 static int ldlm_bl_thread_main(void *arg)
2733 struct ldlm_bl_pool *blp;
2734 struct ldlm_bl_thread_data *bltd = arg;
2737 blp = bltd->bltd_blp;
2739 complete(&bltd->bltd_comp);
2740 /* cannot use bltd after this, it is only on caller's stack */
2743 struct l_wait_info lwi = { 0 };
2744 struct ldlm_bl_work_item *blwi = NULL;
2745 struct obd_export *exp = NULL;
2748 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2751 l_wait_event_exclusive(blp->blp_waitq,
2752 ldlm_bl_get_work(blp, &blwi,
2755 atomic_inc(&blp->blp_busy_threads);
2757 if (ldlm_bl_thread_need_create(blp, blwi))
2758 /* discard the return value, we tried */
2759 ldlm_bl_thread_start(blp, true);
2762 rc = ldlm_bl_thread_exports(blp, exp);
2764 rc = ldlm_bl_thread_blwi(blp, blwi);
2766 atomic_dec(&blp->blp_busy_threads);
2768 if (rc == LDLM_ITER_STOP)
2771 /* If there are many namespaces, we will not sleep waiting for
2772 * work, and must do a cond_resched to avoid holding the CPU
2777 atomic_dec(&blp->blp_num_threads);
2778 complete(&blp->blp_comp);
2783 static int ldlm_setup(void);
2784 static int ldlm_cleanup(void);
2786 int ldlm_get_ref(void)
2790 mutex_lock(&ldlm_ref_mutex);
2791 if (++ldlm_refcount == 1) {
2796 mutex_unlock(&ldlm_ref_mutex);
2801 void ldlm_put_ref(void)
2804 mutex_lock(&ldlm_ref_mutex);
2805 if (ldlm_refcount == 1) {
2806 int rc = ldlm_cleanup();
2808 CERROR("ldlm_cleanup failed: %d\n", rc);
2814 mutex_unlock(&ldlm_ref_mutex);
2820 * Export handle<->lock hash operations.
2823 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
2825 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2829 ldlm_export_lock_key(struct hlist_node *hnode)
2831 struct ldlm_lock *lock;
2833 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2834 return &lock->l_remote_handle;
2838 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
2840 struct ldlm_lock *lock;
2842 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2843 lock->l_remote_handle = *(struct lustre_handle *)key;
2847 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
2849 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2853 ldlm_export_lock_object(struct hlist_node *hnode)
2855 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2859 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
2861 struct ldlm_lock *lock;
2863 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2864 LDLM_LOCK_GET(lock);
2868 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
2870 struct ldlm_lock *lock;
2872 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2873 LDLM_LOCK_RELEASE(lock);
2876 static struct cfs_hash_ops ldlm_export_lock_ops = {
2877 .hs_hash = ldlm_export_lock_hash,
2878 .hs_key = ldlm_export_lock_key,
2879 .hs_keycmp = ldlm_export_lock_keycmp,
2880 .hs_keycpy = ldlm_export_lock_keycpy,
2881 .hs_object = ldlm_export_lock_object,
2882 .hs_get = ldlm_export_lock_get,
2883 .hs_put = ldlm_export_lock_put,
2884 .hs_put_locked = ldlm_export_lock_put,
2887 int ldlm_init_export(struct obd_export *exp)
2892 exp->exp_lock_hash =
2893 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2894 HASH_EXP_LOCK_CUR_BITS,
2895 HASH_EXP_LOCK_MAX_BITS,
2896 HASH_EXP_LOCK_BKT_BITS, 0,
2897 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2898 &ldlm_export_lock_ops,
2899 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2900 CFS_HASH_NBLK_CHANGE);
2902 if (!exp->exp_lock_hash)
2905 rc = ldlm_init_flock_export(exp);
2911 ldlm_destroy_export(exp);
2914 EXPORT_SYMBOL(ldlm_init_export);
2916 void ldlm_destroy_export(struct obd_export *exp)
2919 cfs_hash_putref(exp->exp_lock_hash);
2920 exp->exp_lock_hash = NULL;
2922 ldlm_destroy_flock_export(exp);
2925 EXPORT_SYMBOL(ldlm_destroy_export);
2927 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
2928 struct attribute *attr,
2931 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
2934 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
2935 struct attribute *attr,
2942 rc = kstrtoul(buffer, 10, &val);
2946 ldlm_cancel_unused_locks_before_replay = val;
2950 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
2952 static struct attribute *ldlm_attrs[] = {
2953 &lustre_attr_cancel_unused_locks_before_replay.attr,
2957 static struct attribute_group ldlm_attr_group = {
2958 .attrs = ldlm_attrs,
2961 static int ldlm_setup(void)
2963 static struct ptlrpc_service_conf conf;
2964 struct ldlm_bl_pool *blp = NULL;
2965 #ifdef HAVE_SERVER_SUPPORT
2966 struct task_struct *task;
2967 #endif /* HAVE_SERVER_SUPPORT */
2973 if (ldlm_state != NULL)
2976 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2977 if (ldlm_state == NULL)
2980 ldlm_kobj = kobject_create_and_add("ldlm", lustre_kobj);
2984 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
2988 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
2992 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
2996 #ifdef CONFIG_PROC_FS
2997 rc = ldlm_proc_setup();
3000 #endif /* CONFIG_PROC_FS */
3002 memset(&conf, 0, sizeof(conf));
3003 conf = (typeof(conf)) {
3004 .psc_name = "ldlm_cbd",
3005 .psc_watchdog_factor = 2,
3007 .bc_nbufs = LDLM_CLIENT_NBUFS,
3008 .bc_buf_size = LDLM_BUFSIZE,
3009 .bc_req_max_size = LDLM_MAXREQSIZE,
3010 .bc_rep_max_size = LDLM_MAXREPSIZE,
3011 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
3012 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
3015 .tc_thr_name = "ldlm_cb",
3016 .tc_thr_factor = LDLM_THR_FACTOR,
3017 .tc_nthrs_init = LDLM_NTHRS_INIT,
3018 .tc_nthrs_base = LDLM_NTHRS_BASE,
3019 .tc_nthrs_max = LDLM_NTHRS_MAX,
3020 .tc_nthrs_user = ldlm_num_threads,
3021 .tc_cpu_affinity = 1,
3022 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
3025 .cc_pattern = ldlm_cpts,
3028 .so_req_handler = ldlm_callback_handler,
3031 ldlm_state->ldlm_cb_service = \
3032 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
3033 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3034 CERROR("failed to start service\n");
3035 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3036 ldlm_state->ldlm_cb_service = NULL;
3040 #ifdef HAVE_SERVER_SUPPORT
3041 memset(&conf, 0, sizeof(conf));
3042 conf = (typeof(conf)) {
3043 .psc_name = "ldlm_canceld",
3044 .psc_watchdog_factor = 6,
3046 .bc_nbufs = LDLM_SERVER_NBUFS,
3047 .bc_buf_size = LDLM_BUFSIZE,
3048 .bc_req_max_size = LDLM_MAXREQSIZE,
3049 .bc_rep_max_size = LDLM_MAXREPSIZE,
3050 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3051 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3055 .tc_thr_name = "ldlm_cn",
3056 .tc_thr_factor = LDLM_THR_FACTOR,
3057 .tc_nthrs_init = LDLM_NTHRS_INIT,
3058 .tc_nthrs_base = LDLM_NTHRS_BASE,
3059 .tc_nthrs_max = LDLM_NTHRS_MAX,
3060 .tc_nthrs_user = ldlm_num_threads,
3061 .tc_cpu_affinity = 1,
3062 .tc_ctx_tags = LCT_MD_THREAD | \
3067 .cc_pattern = ldlm_cpts,
3070 .so_req_handler = ldlm_cancel_handler,
3071 .so_hpreq_handler = ldlm_hpreq_handler,
3074 ldlm_state->ldlm_cancel_service = \
3075 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
3076 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3077 CERROR("failed to start service\n");
3078 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3079 ldlm_state->ldlm_cancel_service = NULL;
3082 #endif /* HAVE_SERVER_SUPPORT */
3084 OBD_ALLOC(blp, sizeof(*blp));
3086 GOTO(out, rc = -ENOMEM);
3087 ldlm_state->ldlm_bl_pool = blp;
3089 spin_lock_init(&blp->blp_lock);
3090 INIT_LIST_HEAD(&blp->blp_list);
3091 INIT_LIST_HEAD(&blp->blp_prio_list);
3092 init_waitqueue_head(&blp->blp_waitq);
3093 atomic_set(&blp->blp_num_threads, 0);
3094 atomic_set(&blp->blp_busy_threads, 0);
3096 if (ldlm_num_threads == 0) {
3097 blp->blp_min_threads = LDLM_NTHRS_INIT;
3098 blp->blp_max_threads = LDLM_NTHRS_MAX;
3100 blp->blp_min_threads = blp->blp_max_threads = \
3101 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3105 for (i = 0; i < blp->blp_min_threads; i++) {
3106 rc = ldlm_bl_thread_start(blp, false);
3111 #ifdef HAVE_SERVER_SUPPORT
3112 INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
3113 expired_lock_thread.elt_state = ELT_STOPPED;
3114 init_waitqueue_head(&expired_lock_thread.elt_waitq);
3116 INIT_LIST_HEAD(&waiting_locks_list);
3117 spin_lock_init(&waiting_locks_spinlock);
3118 setup_timer(&waiting_locks_timer, waiting_locks_callback, 0);
3120 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3123 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3127 wait_event(expired_lock_thread.elt_waitq,
3128 expired_lock_thread.elt_state == ELT_READY);
3129 #endif /* HAVE_SERVER_SUPPORT */
3131 rc = ldlm_pools_init();
3133 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3137 rc = ldlm_reclaim_setup();
3139 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3149 static int ldlm_cleanup(void)
3153 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3154 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3155 CERROR("ldlm still has namespaces; clean these up first.\n");
3156 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3157 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3161 ldlm_reclaim_cleanup();
3164 if (ldlm_state->ldlm_bl_pool != NULL) {
3165 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3167 while (atomic_read(&blp->blp_num_threads) > 0) {
3168 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3170 init_completion(&blp->blp_comp);
3172 spin_lock(&blp->blp_lock);
3173 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3174 wake_up(&blp->blp_waitq);
3175 spin_unlock(&blp->blp_lock);
3177 wait_for_completion(&blp->blp_comp);
3180 OBD_FREE(blp, sizeof(*blp));
3183 if (ldlm_state->ldlm_cb_service != NULL)
3184 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3185 #ifdef HAVE_SERVER_SUPPORT
3186 if (ldlm_state->ldlm_cancel_service != NULL)
3187 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3191 kset_unregister(ldlm_ns_kset);
3193 kset_unregister(ldlm_svc_kset);
3195 kobject_put(ldlm_kobj);
3197 ldlm_proc_cleanup();
3199 #ifdef HAVE_SERVER_SUPPORT
3200 if (expired_lock_thread.elt_state != ELT_STOPPED) {
3201 expired_lock_thread.elt_state = ELT_TERMINATE;
3202 wake_up(&expired_lock_thread.elt_waitq);
3203 wait_event(expired_lock_thread.elt_waitq,
3204 expired_lock_thread.elt_state == ELT_STOPPED);
3208 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3216 mutex_init(&ldlm_ref_mutex);
3217 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
3218 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
3220 INIT_LIST_HEAD(&ldlm_srv_namespace_list);
3221 INIT_LIST_HEAD(&ldlm_cli_active_namespace_list);
3222 INIT_LIST_HEAD(&ldlm_cli_inactive_namespace_list);
3224 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3225 sizeof(struct ldlm_resource), 0,
3226 SLAB_HWCACHE_ALIGN, NULL);
3227 if (ldlm_resource_slab == NULL)
3230 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3231 sizeof(struct ldlm_lock), 0,
3232 SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU, NULL);
3233 if (ldlm_lock_slab == NULL)
3236 ldlm_interval_slab = kmem_cache_create("interval_node",
3237 sizeof(struct ldlm_interval),
3238 0, SLAB_HWCACHE_ALIGN, NULL);
3239 if (ldlm_interval_slab == NULL)
3242 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3243 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3244 0, SLAB_HWCACHE_ALIGN, NULL);
3245 if (ldlm_interval_tree_slab == NULL)
3248 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3249 class_export_dump_hook = ldlm_dump_export_locks;
3254 kmem_cache_destroy(ldlm_interval_slab);
3256 kmem_cache_destroy(ldlm_lock_slab);
3258 kmem_cache_destroy(ldlm_resource_slab);
3263 void ldlm_exit(void)
3266 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3267 kmem_cache_destroy(ldlm_resource_slab);
3268 /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3269 * synchronize_rcu() to wait a grace period elapsed, so that
3270 * ldlm_lock_free() get a chance to be called. */
3272 kmem_cache_destroy(ldlm_lock_slab);
3273 kmem_cache_destroy(ldlm_interval_slab);
3274 kmem_cache_destroy(ldlm_interval_tree_slab);