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_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 DEFINE_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 /* timeout for initial callback (AST) reply (bz10399)
68 * Due to having to send a 32 bit time value over the
69 * wire return it as time_t instead of time64_t
71 static inline time_t ldlm_get_rq_timeout(void)
74 time_t timeout = min(ldlm_timeout, obd_timeout / 3);
76 return timeout < 1 ? 1 : timeout;
83 * blp_prio_list is used for callbacks that should be handled
84 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
87 struct list_head blp_prio_list;
90 * blp_list is used for all other callbacks which are likely
91 * to take longer to process.
93 struct list_head blp_list;
95 wait_queue_head_t blp_waitq;
96 struct completion blp_comp;
97 atomic_t blp_num_threads;
98 atomic_t blp_busy_threads;
103 struct ldlm_bl_work_item {
104 struct list_head blwi_entry;
105 struct ldlm_namespace *blwi_ns;
106 struct ldlm_lock_desc blwi_ld;
107 struct ldlm_lock *blwi_lock;
108 struct list_head blwi_head;
110 struct completion blwi_comp;
111 enum ldlm_cancel_flags blwi_flags;
112 int blwi_mem_pressure;
115 #ifdef HAVE_SERVER_SUPPORT
118 * Protects both waiting_locks_list and expired_lock_thread.
120 static DEFINE_SPINLOCK(waiting_locks_spinlock); /* BH lock (timer) */
123 * List for contended locks.
125 * As soon as a lock is contended, it gets placed on this list and
126 * expected time to get a response is filled in the lock. A special
127 * thread walks the list looking for locks that should be released and
128 * schedules client evictions for those that have not been released in
131 * All access to it should be under waiting_locks_spinlock.
133 static LIST_HEAD(waiting_locks_list);
134 static void waiting_locks_callback(unsigned long unused);
135 static DEFINE_TIMER(waiting_locks_timer, waiting_locks_callback, 0, 0);
143 static DECLARE_WAIT_QUEUE_HEAD(expired_lock_wait_queue);
144 static enum elt_state expired_lock_thread_state = ELT_STOPPED;
145 static int expired_lock_dump;
146 static LIST_HEAD(expired_lock_list);
148 static inline int have_expired_locks(void)
153 spin_lock_bh(&waiting_locks_spinlock);
154 need_to_run = !list_empty(&expired_lock_list);
155 spin_unlock_bh(&waiting_locks_spinlock);
161 * Check expired lock list for expired locks and time them out.
163 static int expired_lock_main(void *arg)
165 struct list_head *expired = &expired_lock_list;
166 struct l_wait_info lwi = { 0 };
171 expired_lock_thread_state = ELT_READY;
172 wake_up(&expired_lock_wait_queue);
175 l_wait_event(expired_lock_wait_queue,
176 have_expired_locks() ||
177 expired_lock_thread_state == ELT_TERMINATE,
180 spin_lock_bh(&waiting_locks_spinlock);
181 if (expired_lock_dump) {
182 spin_unlock_bh(&waiting_locks_spinlock);
184 /* from waiting_locks_callback, but not in timer */
185 libcfs_debug_dumplog();
187 spin_lock_bh(&waiting_locks_spinlock);
188 expired_lock_dump = 0;
193 while (!list_empty(expired)) {
194 struct obd_export *export;
195 struct ldlm_lock *lock;
197 lock = list_entry(expired->next, struct ldlm_lock,
199 if ((void *)lock < LP_POISON + PAGE_SIZE &&
200 (void *)lock >= LP_POISON) {
201 spin_unlock_bh(&waiting_locks_spinlock);
202 CERROR("free lock on elt list %p\n", lock);
205 list_del_init(&lock->l_pending_chain);
206 if ((void *)lock->l_export <
207 LP_POISON + PAGE_SIZE &&
208 (void *)lock->l_export >= LP_POISON) {
209 CERROR("lock with free export on elt list %p\n",
211 lock->l_export = NULL;
212 LDLM_ERROR(lock, "free export");
213 /* release extra ref grabbed by
214 * ldlm_add_waiting_lock() or
215 * ldlm_failed_ast() */
216 LDLM_LOCK_RELEASE(lock);
220 if (ldlm_is_destroyed(lock)) {
221 /* release the lock refcount where
222 * waiting_locks_callback() founds */
223 LDLM_LOCK_RELEASE(lock);
226 export = class_export_lock_get(lock->l_export, lock);
227 spin_unlock_bh(&waiting_locks_spinlock);
229 spin_lock_bh(&export->exp_bl_list_lock);
230 list_del_init(&lock->l_exp_list);
231 spin_unlock_bh(&export->exp_bl_list_lock);
234 class_fail_export(export);
235 class_export_lock_put(export, lock);
237 /* release extra ref grabbed by ldlm_add_waiting_lock()
238 * or ldlm_failed_ast() */
239 LDLM_LOCK_RELEASE(lock);
241 spin_lock_bh(&waiting_locks_spinlock);
243 spin_unlock_bh(&waiting_locks_spinlock);
245 if (do_dump && obd_dump_on_eviction) {
246 CERROR("dump the log upon eviction\n");
247 libcfs_debug_dumplog();
250 if (expired_lock_thread_state == ELT_TERMINATE)
254 expired_lock_thread_state = ELT_STOPPED;
255 wake_up(&expired_lock_wait_queue);
259 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
260 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds);
263 * Check if there is a request in the export request list
264 * which prevents the lock canceling.
266 static int ldlm_lock_busy(struct ldlm_lock *lock)
268 struct ptlrpc_request *req;
272 if (lock->l_export == NULL)
275 spin_lock_bh(&lock->l_export->exp_rpc_lock);
276 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
278 if (req->rq_ops->hpreq_lock_match) {
279 match = req->rq_ops->hpreq_lock_match(req, lock);
284 spin_unlock_bh(&lock->l_export->exp_rpc_lock);
288 /* This is called from within a timer interrupt and cannot schedule */
289 static void waiting_locks_callback(unsigned long unused)
291 struct ldlm_lock *lock;
294 spin_lock_bh(&waiting_locks_spinlock);
295 while (!list_empty(&waiting_locks_list)) {
296 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
298 if (lock->l_callback_timeout > ktime_get_seconds() ||
299 lock->l_req_mode == LCK_GROUP)
302 /* Check if we need to prolong timeout */
303 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
304 ldlm_lock_busy(lock)) {
307 if (lock->l_pending_chain.next == &waiting_locks_list)
312 spin_unlock_bh(&waiting_locks_spinlock);
313 LDLM_DEBUG(lock, "prolong the busy lock");
314 ldlm_refresh_waiting_lock(lock,
315 ldlm_bl_timeout(lock) >> 1);
316 spin_lock_bh(&waiting_locks_spinlock);
319 LDLM_LOCK_RELEASE(lock);
323 LDLM_LOCK_RELEASE(lock);
326 ldlm_lock_to_ns(lock)->ns_timeouts++;
327 LDLM_ERROR(lock, "lock callback timer expired after %llds: "
328 "evicting client at %s ",
329 ktime_get_real_seconds() - lock->l_last_activity,
331 lock->l_export->exp_connection->c_peer.nid));
333 /* no needs to take an extra ref on the lock since it was in
334 * the waiting_locks_list and ldlm_add_waiting_lock()
335 * already grabbed a ref */
336 list_del(&lock->l_pending_chain);
337 list_add(&lock->l_pending_chain, &expired_lock_list);
341 if (!list_empty(&expired_lock_list)) {
342 if (obd_dump_on_timeout && need_dump)
343 expired_lock_dump = __LINE__;
345 wake_up(&expired_lock_wait_queue);
349 * Make sure the timer will fire again if we have any locks
352 if (!list_empty(&waiting_locks_list)) {
353 unsigned long timeout_jiffies;
355 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
357 timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout);
358 mod_timer(&waiting_locks_timer, timeout_jiffies);
360 spin_unlock_bh(&waiting_locks_spinlock);
364 * Add lock to the list of contended locks.
366 * Indicate that we're waiting for a client to call us back cancelling a given
367 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
368 * timer to fire appropriately. (We round up to the next second, to avoid
369 * floods of timer firings during periods of high lock contention and traffic).
370 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
371 * if it has been added to the waiting list (1 is returned).
373 * Called with the namespace lock held.
375 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds)
377 unsigned long timeout_jiffies;
380 if (!list_empty(&lock->l_pending_chain))
383 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
384 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
387 timeout = ktime_get_seconds() + seconds;
388 if (likely(timeout > lock->l_callback_timeout))
389 lock->l_callback_timeout = timeout;
391 timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout);
393 if (time_before(timeout_jiffies, waiting_locks_timer.expires) ||
394 !timer_pending(&waiting_locks_timer))
395 mod_timer(&waiting_locks_timer, timeout_jiffies);
397 /* if the new lock has a shorter timeout than something earlier on
398 * the list, we'll wait the longer amount of time; no big deal.
401 list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
405 static void ldlm_add_blocked_lock(struct ldlm_lock *lock)
407 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
408 if (list_empty(&lock->l_exp_list)) {
409 if (lock->l_granted_mode != lock->l_req_mode)
410 list_add_tail(&lock->l_exp_list,
411 &lock->l_export->exp_bl_list);
413 list_add(&lock->l_exp_list,
414 &lock->l_export->exp_bl_list);
416 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
418 /* A blocked lock is added. Adjust the position in
419 * the stale list if the export is in the list.
420 * If export is stale and not in the list - it is being
421 * processed and will be placed on the right position
422 * on obd_stale_export_put(). */
423 if (!list_empty(&lock->l_export->exp_stale_list))
424 obd_stale_export_adjust(lock->l_export);
427 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
429 time64_t timeout = ldlm_bl_timeout(lock);
432 /* NB: must be called with hold of lock_res_and_lock() */
433 LASSERT(ldlm_is_res_locked(lock));
434 LASSERT(!ldlm_is_cancel_on_block(lock));
436 /* Do not put cross-MDT lock in the waiting list, since we
437 * will not evict it due to timeout for now */
438 if (lock->l_export != NULL &&
439 (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS))
442 spin_lock_bh(&waiting_locks_spinlock);
443 if (ldlm_is_cancel(lock)) {
444 spin_unlock_bh(&waiting_locks_spinlock);
448 if (ldlm_is_destroyed(lock)) {
449 static time64_t next;
451 spin_unlock_bh(&waiting_locks_spinlock);
452 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
453 if (ktime_get_seconds() > next) {
454 next = ktime_get_seconds() + 14400;
455 libcfs_debug_dumpstack(NULL);
460 ldlm_set_waited(lock);
461 lock->l_last_activity = ktime_get_real_seconds();
462 ret = __ldlm_add_waiting_lock(lock, timeout);
464 /* grab ref on the lock if it has been added to the
468 spin_unlock_bh(&waiting_locks_spinlock);
471 ldlm_add_blocked_lock(lock);
473 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %lld, AT: %s)",
474 ret == 0 ? "not re-" : "", timeout,
475 AT_OFF ? "off" : "on");
480 * Remove a lock from the pending list, likely because it had its cancellation
481 * callback arrive without incident. This adjusts the lock-timeout timer if
482 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
483 * As done by ldlm_del_waiting_lock(), the caller must release the lock
484 * reference when the lock is removed from any list (1 is returned).
486 * Called with namespace lock held.
488 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
490 struct list_head *list_next;
492 if (list_empty(&lock->l_pending_chain))
495 list_next = lock->l_pending_chain.next;
496 if (lock->l_pending_chain.prev == &waiting_locks_list) {
497 /* Removing the head of the list, adjust timer. */
498 if (list_next == &waiting_locks_list) {
499 /* No more, just cancel. */
500 del_timer(&waiting_locks_timer);
502 struct ldlm_lock *next;
504 next = list_entry(list_next, struct ldlm_lock,
506 mod_timer(&waiting_locks_timer,
507 cfs_time_seconds(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, time64_t 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, time64_t 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 time64_t ldlm_bl_timeout(struct ldlm_lock *lock)
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), (time64_t)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, &expired_lock_list);
643 wake_up(&expired_lock_wait_queue);
644 spin_unlock_bh(&waiting_locks_spinlock);
648 * Perform lock cleanup if AST reply came with error.
650 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
651 struct ptlrpc_request *req, int rc,
652 const char *ast_type)
654 struct lnet_process_id peer = req->rq_import->imp_connection->c_peer;
656 if (!req->rq_replied || (rc && rc != -EINVAL)) {
657 if (ldlm_is_cancel(lock)) {
659 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
660 ast_type, req, req->rq_xid,
661 libcfs_nid2str(peer.nid));
662 ldlm_lock_cancel(lock);
664 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
666 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
667 /* Upon umount process the AST fails because cannot be
668 * sent. This shouldn't lead to the client eviction.
669 * -ENODEV error is returned by ptl_send_rpc() for
670 * new request in such import.
671 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
672 * if imp_invalid is set or obd_no_recov.
673 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
674 * in ptlrpc_import_delay_req() as well with -EIO code.
675 * In all such cases errors are ignored.
677 LDLM_DEBUG(lock, "%s AST can't be sent due to a server"
678 " %s failure or umount process: rc = %d\n",
680 req->rq_import->imp_obd->obd_name, rc);
683 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
684 libcfs_nid2str(peer.nid),
685 req->rq_replied ? "returned error from" :
686 "failed to reply to",
687 ast_type, req, req->rq_xid,
688 (req->rq_repmsg != NULL) ?
689 lustre_msg_get_status(req->rq_repmsg) : 0,
691 ldlm_failed_ast(lock, rc, ast_type);
697 struct ldlm_resource *res = lock->l_resource;
700 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
701 libcfs_nid2str(peer.nid),
703 lustre_msg_get_status(req->rq_repmsg) : -1,
704 ast_type, req, req->rq_xid);
706 /* update lvbo to return proper attributes.
708 ldlm_resource_getref(res);
709 ldlm_res_lvbo_update(res, NULL, 1);
710 ldlm_resource_putref(res);
712 ldlm_lock_cancel(lock);
719 static int ldlm_cb_interpret(const struct lu_env *env,
720 struct ptlrpc_request *req, void *data, int rc)
722 struct ldlm_cb_async_args *ca = data;
723 struct ldlm_lock *lock = ca->ca_lock;
724 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
727 LASSERT(lock != NULL);
730 case LDLM_GL_CALLBACK:
731 /* Update the LVB from disk if the AST failed
732 * (this is a legal race)
734 * - Glimpse callback of local lock just returns
735 * -ELDLM_NO_LOCK_DATA.
736 * - Glimpse callback of remote lock might return
737 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
739 if (unlikely(arg->gl_interpret_reply)) {
740 rc = arg->gl_interpret_reply(env, req, data, rc);
741 } else if (rc == -ELDLM_NO_LOCK_DATA) {
742 LDLM_DEBUG(lock, "lost race - client has a lock but no "
744 ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
745 } else if (rc != 0) {
746 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
748 rc = ldlm_res_lvbo_update(lock->l_resource, req, 1);
751 case LDLM_BL_CALLBACK:
753 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
755 case LDLM_CP_CALLBACK:
757 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
760 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
765 /* release extra reference taken in ldlm_ast_fini() */
766 LDLM_LOCK_RELEASE(lock);
769 atomic_inc(&arg->restart);
774 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
776 struct ldlm_cb_async_args *ca = data;
777 struct ldlm_lock *lock = ca->ca_lock;
779 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
782 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
783 struct ldlm_cb_set_arg *arg,
784 struct ldlm_lock *lock,
790 if (unlikely(instant_cancel)) {
791 rc = ptl_send_rpc(req, 1);
792 ptlrpc_req_finished(req);
794 atomic_inc(&arg->restart);
797 ptlrpc_set_add_req(arg->set, req);
804 * Check if there are requests in the export request list which prevent
805 * the lock canceling and make these requests high priority ones.
807 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
809 struct ptlrpc_request *req;
812 if (lock->l_export == NULL) {
813 LDLM_DEBUG(lock, "client lock: no-op");
817 spin_lock_bh(&lock->l_export->exp_rpc_lock);
818 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
820 /* Do not process requests that were not yet added to there
821 * incoming queue or were already removed from there for
822 * processing. We evaluate ptlrpc_nrs_req_can_move() without
823 * holding svcpt->scp_req_lock, and then redo the check with
824 * the lock held once we need to obtain a reliable result.
826 if (ptlrpc_nrs_req_can_move(req) &&
827 req->rq_ops->hpreq_lock_match &&
828 req->rq_ops->hpreq_lock_match(req, lock))
829 ptlrpc_nrs_req_hp_move(req);
831 spin_unlock_bh(&lock->l_export->exp_rpc_lock);
836 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
837 * enqueued server lock conflicts with given one.
839 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
840 * to wait for client response.
842 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
843 struct ldlm_lock_desc *desc,
844 void *data, int flag)
846 struct ldlm_cb_async_args *ca;
847 struct ldlm_cb_set_arg *arg = data;
848 struct ldlm_request *body;
849 struct ptlrpc_request *req;
850 int instant_cancel = 0;
854 if (flag == LDLM_CB_CANCELING)
855 /* Don't need to do anything here. */
858 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
859 LDLM_DEBUG(lock, "dropping BL AST");
864 LASSERT(data != NULL);
865 if (lock->l_export->exp_obd->obd_recovering != 0)
866 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
868 ldlm_lock_reorder_req(lock);
870 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
871 &RQF_LDLM_BL_CALLBACK,
872 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
876 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
877 ca = ptlrpc_req_async_args(req);
878 ca->ca_set_arg = arg;
881 req->rq_interpret_reply = ldlm_cb_interpret;
883 lock_res_and_lock(lock);
884 if (ldlm_is_destroyed(lock)) {
885 /* What's the point? */
886 unlock_res_and_lock(lock);
887 ptlrpc_req_finished(req);
891 if (lock->l_granted_mode != lock->l_req_mode) {
892 /* this blocking AST will be communicated as part of the
893 * completion AST instead */
894 ldlm_add_blocked_lock(lock);
895 ldlm_set_waited(lock);
896 unlock_res_and_lock(lock);
898 ptlrpc_req_finished(req);
899 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
903 if (ldlm_is_cancel_on_block(lock))
906 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
907 body->lock_handle[0] = lock->l_remote_handle;
908 body->lock_desc = *desc;
909 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
911 LDLM_DEBUG(lock, "server preparing blocking AST");
913 ptlrpc_request_set_replen(req);
914 ldlm_set_cbpending(lock);
915 if (instant_cancel) {
916 unlock_res_and_lock(lock);
917 ldlm_lock_cancel(lock);
919 req->rq_no_resend = 1;
921 LASSERT(lock->l_granted_mode == lock->l_req_mode);
922 ldlm_add_waiting_lock(lock);
923 unlock_res_and_lock(lock);
925 /* Do not resend after lock callback timeout */
926 req->rq_delay_limit = ldlm_bl_timeout(lock);
927 req->rq_resend_cb = ldlm_update_resend;
930 req->rq_send_state = LUSTRE_IMP_FULL;
931 /* ptlrpc_request_alloc_pack already set timeout */
933 req->rq_timeout = ldlm_get_rq_timeout();
935 lock->l_last_activity = ktime_get_real_seconds();
937 if (lock->l_export && lock->l_export->exp_nid_stats &&
938 lock->l_export->exp_nid_stats->nid_ldlm_stats)
939 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
940 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
942 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
948 * ->l_completion_ast callback for a remote lock in server namespace.
950 * Sends AST to the client notifying it of lock granting. If initial
951 * lock response was not sent yet, instead of sending another RPC, just
952 * mark the lock as granted and client will understand
954 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
956 struct ldlm_cb_set_arg *arg = data;
957 struct ldlm_request *body;
958 struct ptlrpc_request *req;
959 struct ldlm_cb_async_args *ca;
960 int instant_cancel = 0;
965 LASSERT(lock != NULL);
966 LASSERT(data != NULL);
968 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
969 LDLM_DEBUG(lock, "dropping CP AST");
973 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
974 &RQF_LDLM_CP_CALLBACK);
978 /* server namespace, doesn't need lock */
979 lvb_len = ldlm_lvbo_size(lock);
980 /* LU-3124 & LU-2187: to not return layout in completion AST because
981 * it may deadlock for LU-2187, or client may not have enough space
982 * for large layout. The layout will be returned to client with an
983 * extra RPC to fetch xattr.lov */
984 if (ldlm_has_layout(lock))
987 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
988 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
990 ptlrpc_request_free(req);
994 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
995 ca = ptlrpc_req_async_args(req);
996 ca->ca_set_arg = arg;
999 req->rq_interpret_reply = ldlm_cb_interpret;
1000 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1002 body->lock_handle[0] = lock->l_remote_handle;
1003 body->lock_flags = ldlm_flags_to_wire(flags);
1004 ldlm_lock2desc(lock, &body->lock_desc);
1006 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1008 lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len);
1010 /* We still need to send the RPC to wake up the blocked
1011 * enqueue thread on the client.
1013 * Consider old client, there is no better way to notify
1014 * the failure, just zero-sized the LVB, then the client
1015 * will fail out as "-EPROTO". */
1016 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1020 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1025 lock->l_last_activity = ktime_get_real_seconds();
1027 LDLM_DEBUG(lock, "server preparing completion AST");
1029 ptlrpc_request_set_replen(req);
1031 req->rq_send_state = LUSTRE_IMP_FULL;
1032 /* ptlrpc_request_pack already set timeout */
1034 req->rq_timeout = ldlm_get_rq_timeout();
1036 /* We only send real blocking ASTs after the lock is granted */
1037 lock_res_and_lock(lock);
1038 if (ldlm_is_ast_sent(lock)) {
1039 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1040 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1041 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1044 /* We might get here prior to ldlm_handle_enqueue setting
1045 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1046 * into waiting list, but this is safe and similar code in
1047 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1048 * that would not only cancel the lock, but will also remove
1049 * it from waiting list */
1050 if (ldlm_is_cancel_on_block(lock)) {
1051 unlock_res_and_lock(lock);
1052 ldlm_lock_cancel(lock);
1055 req->rq_no_resend = 1;
1057 lock_res_and_lock(lock);
1059 /* start the lock-timeout clock */
1060 ldlm_add_waiting_lock(lock);
1061 /* Do not resend after lock callback timeout */
1062 req->rq_delay_limit = ldlm_bl_timeout(lock);
1063 req->rq_resend_cb = ldlm_update_resend;
1066 unlock_res_and_lock(lock);
1068 if (lock->l_export && lock->l_export->exp_nid_stats &&
1069 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1070 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1071 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1073 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1075 RETURN(lvb_len < 0 ? lvb_len : rc);
1079 * Server side ->l_glimpse_ast handler for client locks.
1081 * Sends glimpse AST to the client and waits for reply. Then updates
1082 * lvbo with the result.
1084 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1086 struct ldlm_cb_set_arg *arg = data;
1087 struct ldlm_request *body;
1088 struct ptlrpc_request *req;
1089 struct ldlm_cb_async_args *ca;
1091 struct req_format *req_fmt;
1094 LASSERT(lock != NULL);
1096 if (arg->gl_desc != NULL)
1097 /* There is a glimpse descriptor to pack */
1098 req_fmt = &RQF_LDLM_GL_DESC_CALLBACK;
1100 req_fmt = &RQF_LDLM_GL_CALLBACK;
1102 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1103 req_fmt, LUSTRE_DLM_VERSION,
1109 if (arg->gl_desc != NULL) {
1110 /* copy the GL descriptor */
1111 union ldlm_gl_desc *desc;
1112 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1113 *desc = *arg->gl_desc;
1116 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1117 body->lock_handle[0] = lock->l_remote_handle;
1118 ldlm_lock2desc(lock, &body->lock_desc);
1120 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1121 ca = ptlrpc_req_async_args(req);
1122 ca->ca_set_arg = arg;
1125 /* server namespace, doesn't need lock */
1126 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1127 ldlm_lvbo_size(lock));
1128 ptlrpc_request_set_replen(req);
1130 req->rq_send_state = LUSTRE_IMP_FULL;
1131 /* ptlrpc_request_alloc_pack already set timeout */
1133 req->rq_timeout = ldlm_get_rq_timeout();
1135 lock->l_last_activity = ktime_get_real_seconds();
1137 req->rq_interpret_reply = ldlm_cb_interpret;
1139 if (lock->l_export && lock->l_export->exp_nid_stats &&
1140 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1141 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1142 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1144 rc = ldlm_ast_fini(req, arg, lock, 0);
1149 int ldlm_glimpse_locks(struct ldlm_resource *res,
1150 struct list_head *gl_work_list)
1155 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1157 if (rc == -ERESTART)
1158 ldlm_reprocess_all(res);
1162 EXPORT_SYMBOL(ldlm_glimpse_locks);
1164 /* return LDLM lock associated with a lock callback request */
1165 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1167 struct ldlm_cb_async_args *ca;
1168 struct ldlm_lock *lock;
1171 ca = ptlrpc_req_async_args(req);
1174 RETURN(ERR_PTR(-EFAULT));
1178 EXPORT_SYMBOL(ldlm_request_lock);
1180 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
1181 struct lprocfs_stats *srv_stats)
1183 int lock_type = 0, op = 0;
1185 lock_type = dlm_req->lock_desc.l_resource.lr_type;
1187 switch (lock_type) {
1189 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
1192 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
1193 op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
1195 op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
1198 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
1201 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
1209 lprocfs_counter_incr(srv_stats, op);
1215 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1216 * service threads to carry out client lock enqueueing requests.
1218 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1219 struct ptlrpc_request *req,
1220 const struct ldlm_request *dlm_req,
1221 const struct ldlm_callback_suite *cbs)
1223 struct ldlm_reply *dlm_rep;
1225 enum ldlm_error err = ELDLM_OK;
1226 struct ldlm_lock *lock = NULL;
1227 void *cookie = NULL;
1229 struct ldlm_resource *res = NULL;
1232 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1234 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1235 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1237 LASSERT(req->rq_export);
1239 if (ptlrpc_req2svc(req)->srv_stats != NULL)
1240 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1242 if (req->rq_export && req->rq_export->exp_nid_stats &&
1243 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1244 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1245 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1247 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1248 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1249 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1250 dlm_req->lock_desc.l_resource.lr_type);
1251 GOTO(out, rc = -EFAULT);
1254 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1255 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1256 dlm_req->lock_desc.l_req_mode &
1257 (dlm_req->lock_desc.l_req_mode-1))) {
1258 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1259 dlm_req->lock_desc.l_req_mode);
1260 GOTO(out, rc = -EFAULT);
1263 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) {
1264 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1266 DEBUG_REQ(D_ERROR, req,
1267 "PLAIN lock request from IBITS client?");
1268 GOTO(out, rc = -EPROTO);
1270 } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1272 DEBUG_REQ(D_ERROR, req,
1273 "IBITS lock request from unaware client?");
1274 GOTO(out, rc = -EPROTO);
1277 if (unlikely((flags & LDLM_FL_REPLAY) ||
1278 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1279 /* Find an existing lock in the per-export lock hash */
1280 /* In the function below, .hs_keycmp resolves to
1281 * ldlm_export_lock_keycmp() */
1282 /* coverity[overrun-buffer-val] */
1283 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1284 (void *)&dlm_req->lock_handle[0]);
1286 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie %#llx",
1287 lock->l_handle.h_cookie);
1288 flags |= LDLM_FL_RESENT;
1289 GOTO(existing_lock, rc = 0);
1292 if (ldlm_reclaim_full()) {
1293 DEBUG_REQ(D_DLMTRACE, req, "Too many granted locks, "
1294 "reject current enqueue request and let the "
1295 "client retry later.\n");
1296 GOTO(out, rc = -EINPROGRESS);
1300 /* The lock's callback data might be set in the policy function */
1301 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1302 dlm_req->lock_desc.l_resource.lr_type,
1303 dlm_req->lock_desc.l_req_mode,
1304 cbs, NULL, 0, LVB_T_NONE);
1311 lock->l_remote_handle = dlm_req->lock_handle[0];
1312 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1314 /* Initialize resource lvb but not for a lock being replayed since
1315 * Client already got lvb sent in this case.
1316 * This must occur early since some policy methods assume resource
1317 * lvb is available (lr_lvb_data != NULL).
1319 res = lock->l_resource;
1320 if (!(flags & LDLM_FL_REPLAY)) {
1321 /* non-replayed lock, delayed lvb init may need to be done */
1322 rc = ldlm_lvbo_init(res);
1324 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1329 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1330 /* Don't enqueue a lock onto the export if it is been disonnected
1331 * due to eviction (bug 3822) or server umount (bug 24324).
1332 * Cancel it now instead. */
1333 if (req->rq_export->exp_disconnected) {
1334 LDLM_ERROR(lock, "lock on disconnected export %p",
1336 GOTO(out, rc = -ENOTCONN);
1339 lock->l_export = class_export_lock_get(req->rq_export, lock);
1340 if (lock->l_export->exp_lock_hash)
1341 cfs_hash_add(lock->l_export->exp_lock_hash,
1342 &lock->l_remote_handle,
1345 /* Inherit the enqueue flags before the operation, because we do not
1346 * keep the res lock on return and next operations (BL AST) may proceed
1348 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1349 LDLM_FL_INHERIT_MASK);
1351 ldlm_convert_policy_to_local(req->rq_export,
1352 dlm_req->lock_desc.l_resource.lr_type,
1353 &dlm_req->lock_desc.l_policy_data,
1354 &lock->l_policy_data);
1355 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1356 lock->l_req_extent = lock->l_policy_data.l_extent;
1360 if (flags & LDLM_FL_HAS_INTENT) {
1361 /* In this case, the reply buffer is allocated deep in
1362 * local_lock_enqueue by the policy function. */
1365 /* based on the assumption that lvb size never changes during
1366 * resource life time otherwise it need resource->lr_lock's
1368 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1369 RCL_SERVER, ldlm_lvbo_size(lock));
1371 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1372 GOTO(out, rc = -ENOMEM);
1374 rc = req_capsule_server_pack(&req->rq_pill);
1379 err = ldlm_lock_enqueue(ns, &lock, cookie, &flags);
1386 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1388 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1389 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1391 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1392 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1394 /* We never send a blocking AST until the lock is granted, but
1395 * we can tell it right now */
1396 lock_res_and_lock(lock);
1398 /* Now take into account flags to be inherited from original lock
1399 request both in reply to client and in our own lock flags. */
1400 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1401 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1403 /* Don't move a pending lock onto the export if it has already been
1404 * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1405 * Cancel it now instead. */
1406 if (unlikely(req->rq_export->exp_disconnected ||
1407 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1408 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1410 } else if (ldlm_is_ast_sent(lock)) {
1411 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1412 if (lock->l_granted_mode == lock->l_req_mode) {
1414 * Only cancel lock if it was granted, because it would
1415 * be destroyed immediately and would never be granted
1416 * in the future, causing timeouts on client. Not
1417 * granted lock will be cancelled immediately after
1418 * sending completion AST.
1420 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1421 unlock_res_and_lock(lock);
1422 ldlm_lock_cancel(lock);
1423 lock_res_and_lock(lock);
1425 ldlm_add_waiting_lock(lock);
1428 unlock_res_and_lock(lock);
1432 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1433 if (!req->rq_packed_final) {
1434 err = lustre_pack_reply(req, 1, NULL, NULL);
1439 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1440 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1442 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1443 "(err=%d, rc=%d)", err, rc);
1446 if (req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1448 ldlm_lvbo_size(lock) > 0) {
1452 buf = req_capsule_server_get(&req->rq_pill,
1454 LASSERTF(buf != NULL, "req %p, lock %p\n",
1456 buflen = req_capsule_get_size(&req->rq_pill,
1457 &RMF_DLM_LVB, RCL_SERVER);
1458 /* non-replayed lock, delayed lvb init may
1459 * need to be occur now */
1460 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1461 buflen = ldlm_lvbo_fill(lock, buf,
1467 buflen, RCL_SERVER);
1470 } else if (flags & LDLM_FL_REPLAY) {
1471 /* no LVB resend upon replay */
1485 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1486 if (lock->l_export) {
1487 ldlm_lock_cancel(lock);
1489 lock_res_and_lock(lock);
1490 ldlm_resource_unlink_lock(lock);
1491 ldlm_lock_destroy_nolock(lock);
1492 unlock_res_and_lock(lock);
1497 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1498 ldlm_reprocess_all(lock->l_resource);
1500 LDLM_LOCK_RELEASE(lock);
1503 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1510 * Old-style LDLM main entry point for server code enqueue.
1512 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1513 ldlm_completion_callback completion_callback,
1514 ldlm_blocking_callback blocking_callback,
1515 ldlm_glimpse_callback glimpse_callback)
1517 struct ldlm_request *dlm_req;
1518 struct ldlm_callback_suite cbs = {
1519 .lcs_completion = completion_callback,
1520 .lcs_blocking = blocking_callback,
1521 .lcs_glimpse = glimpse_callback
1525 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1526 if (dlm_req != NULL) {
1527 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1528 req, dlm_req, &cbs);
1536 * Main LDLM entry point for server code to process lock conversion requests.
1538 int ldlm_handle_convert0(struct ptlrpc_request *req,
1539 const struct ldlm_request *dlm_req)
1541 struct ldlm_reply *dlm_rep;
1542 struct ldlm_lock *lock;
1546 if (req->rq_export && req->rq_export->exp_nid_stats &&
1547 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1548 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1549 LDLM_CONVERT - LDLM_FIRST_OPC);
1551 rc = req_capsule_server_pack(&req->rq_pill);
1555 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1556 dlm_rep->lock_flags = dlm_req->lock_flags;
1558 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1560 req->rq_status = LUSTRE_EINVAL;
1564 LDLM_DEBUG(lock, "server-side convert handler START");
1566 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1567 &dlm_rep->lock_flags);
1569 if (ldlm_del_waiting_lock(lock))
1570 LDLM_DEBUG(lock, "converted waiting lock");
1573 req->rq_status = LUSTRE_EDEADLK;
1578 if (!req->rq_status)
1579 ldlm_reprocess_all(lock->l_resource);
1580 LDLM_DEBUG(lock, "server-side convert handler END");
1581 LDLM_LOCK_PUT(lock);
1583 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1589 * Old-style main LDLM entry point for server code to process lock conversion
1592 int ldlm_handle_convert(struct ptlrpc_request *req)
1595 struct ldlm_request *dlm_req;
1597 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1598 if (dlm_req != NULL) {
1599 rc = ldlm_handle_convert0(req, dlm_req);
1601 CERROR ("Can't unpack dlm_req\n");
1608 * Cancel all the locks whose handles are packed into ldlm_request
1610 * Called by server code expecting such combined cancel activity
1613 int ldlm_request_cancel(struct ptlrpc_request *req,
1614 const struct ldlm_request *dlm_req,
1615 int first, enum lustre_at_flags flags)
1617 struct ldlm_resource *res, *pres = NULL;
1618 struct ldlm_lock *lock;
1619 int i, count, done = 0;
1622 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1626 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1629 /* There is no lock on the server at the replay time,
1630 * skip lock cancelling to make replay tests to pass. */
1631 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1634 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1635 "starting at %d", count, first);
1637 for (i = first; i < count; i++) {
1638 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1640 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1641 "lock (cookie %llu)",
1642 dlm_req->lock_handle[i].cookie);
1646 res = lock->l_resource;
1649 /* This code is an optimization to only attempt lock
1650 * granting on the resource (that could be CPU-expensive)
1651 * after we are done cancelling lock in that resource. */
1654 ldlm_reprocess_all(pres);
1655 LDLM_RESOURCE_DELREF(pres);
1656 ldlm_resource_putref(pres);
1659 ldlm_resource_getref(res);
1660 LDLM_RESOURCE_ADDREF(res);
1661 ldlm_res_lvbo_update(res, NULL, 1);
1666 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock)) {
1667 time64_t delay = ktime_get_real_seconds() -
1668 lock->l_last_activity;
1669 LDLM_DEBUG(lock, "server cancels blocked lock after %llds",
1671 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1673 ldlm_lock_cancel(lock);
1674 LDLM_LOCK_PUT(lock);
1677 ldlm_reprocess_all(pres);
1678 LDLM_RESOURCE_DELREF(pres);
1679 ldlm_resource_putref(pres);
1681 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1684 EXPORT_SYMBOL(ldlm_request_cancel);
1687 * Main LDLM entry point for server code to cancel locks.
1689 * Typically gets called from service handler on LDLM_CANCEL opc.
1691 int ldlm_handle_cancel(struct ptlrpc_request *req)
1693 struct ldlm_request *dlm_req;
1697 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1698 if (dlm_req == NULL) {
1699 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1703 if (req->rq_export && req->rq_export->exp_nid_stats &&
1704 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1705 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1706 LDLM_CANCEL - LDLM_FIRST_OPC);
1708 rc = req_capsule_server_pack(&req->rq_pill);
1712 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1713 req->rq_status = LUSTRE_ESTALE;
1715 RETURN(ptlrpc_reply(req));
1717 #endif /* HAVE_SERVER_SUPPORT */
1720 * Callback handler for receiving incoming blocking ASTs.
1722 * This can only happen on client side.
1724 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1725 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1730 LDLM_DEBUG(lock, "client blocking AST callback handler");
1732 lock_res_and_lock(lock);
1733 ldlm_set_cbpending(lock);
1735 if (ldlm_is_cancel_on_block(lock))
1736 ldlm_set_cancel(lock);
1738 do_ast = (!lock->l_readers && !lock->l_writers);
1739 unlock_res_and_lock(lock);
1742 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1743 lock, lock->l_blocking_ast);
1744 if (lock->l_blocking_ast != NULL)
1745 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1748 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1752 LDLM_DEBUG(lock, "client blocking callback handler END");
1753 LDLM_LOCK_RELEASE(lock);
1758 * Callback handler for receiving incoming completion ASTs.
1760 * This only can happen on client side.
1762 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1763 struct ldlm_namespace *ns,
1764 struct ldlm_request *dlm_req,
1765 struct ldlm_lock *lock)
1767 struct list_head ast_list;
1772 LDLM_DEBUG(lock, "client completion callback handler START");
1774 INIT_LIST_HEAD(&ast_list);
1775 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1776 long to = cfs_time_seconds(1);
1779 set_current_state(TASK_INTERRUPTIBLE);
1780 schedule_timeout(to);
1781 if (lock->l_granted_mode == lock->l_req_mode ||
1782 ldlm_is_destroyed(lock))
1787 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1789 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1790 GOTO(out, rc = lvb_len);
1791 } else if (lvb_len > 0) {
1792 if (lock->l_lvb_len > 0) {
1793 /* for extent lock, lvb contains ost_lvb{}. */
1794 LASSERT(lock->l_lvb_data != NULL);
1796 if (unlikely(lock->l_lvb_len < lvb_len)) {
1797 LDLM_ERROR(lock, "Replied LVB is larger than "
1798 "expectation, expected = %d, "
1800 lock->l_lvb_len, lvb_len);
1801 GOTO(out, rc = -EINVAL);
1806 lock_res_and_lock(lock);
1807 if (ldlm_is_destroyed(lock) ||
1808 lock->l_granted_mode == lock->l_req_mode) {
1809 /* bug 11300: the lock has already been granted */
1810 unlock_res_and_lock(lock);
1811 LDLM_DEBUG(lock, "Double grant race happened");
1815 /* If we receive the completion AST before the actual enqueue returned,
1816 * then we might need to switch lock modes, resources, or extents. */
1817 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1818 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1819 LDLM_DEBUG(lock, "completion AST, new lock mode");
1822 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1823 ldlm_convert_policy_to_local(req->rq_export,
1824 dlm_req->lock_desc.l_resource.lr_type,
1825 &dlm_req->lock_desc.l_policy_data,
1826 &lock->l_policy_data);
1827 LDLM_DEBUG(lock, "completion AST, new policy data");
1830 ldlm_resource_unlink_lock(lock);
1831 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1832 &lock->l_resource->lr_name,
1833 sizeof(lock->l_resource->lr_name)) != 0) {
1834 unlock_res_and_lock(lock);
1835 rc = ldlm_lock_change_resource(ns, lock,
1836 &dlm_req->lock_desc.l_resource.lr_name);
1838 LDLM_ERROR(lock, "Failed to allocate resource");
1841 LDLM_DEBUG(lock, "completion AST, new resource");
1842 CERROR("change resource!\n");
1843 lock_res_and_lock(lock);
1846 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1847 /* BL_AST locks are not needed in LRU.
1848 * Let ldlm_cancel_lru() be fast. */
1849 ldlm_lock_remove_from_lru(lock);
1850 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1851 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1854 if (lock->l_lvb_len > 0) {
1855 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1856 lock->l_lvb_data, lvb_len);
1858 unlock_res_and_lock(lock);
1863 ldlm_grant_lock(lock, &ast_list);
1864 unlock_res_and_lock(lock);
1866 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1868 /* Let Enqueue to call osc_lock_upcall() and initialize
1870 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1872 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1874 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1880 lock_res_and_lock(lock);
1881 ldlm_set_failed(lock);
1882 unlock_res_and_lock(lock);
1883 wake_up(&lock->l_waitq);
1885 LDLM_LOCK_RELEASE(lock);
1889 * Callback handler for receiving incoming glimpse ASTs.
1891 * This only can happen on client side. After handling the glimpse AST
1892 * we also consider dropping the lock here if it is unused locally for a
1895 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1896 struct ldlm_namespace *ns,
1897 struct ldlm_request *dlm_req,
1898 struct ldlm_lock *lock)
1903 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1905 if (lock->l_glimpse_ast != NULL)
1906 rc = lock->l_glimpse_ast(lock, req);
1908 if (req->rq_repmsg != NULL) {
1911 req->rq_status = rc;
1915 lock_res_and_lock(lock);
1916 if (lock->l_granted_mode == LCK_PW &&
1917 !lock->l_readers && !lock->l_writers &&
1918 ktime_after(ktime_get(),
1919 ktime_add(lock->l_last_used,
1920 ktime_set(10, 0)))) {
1921 unlock_res_and_lock(lock);
1922 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1923 ldlm_handle_bl_callback(ns, NULL, lock);
1928 unlock_res_and_lock(lock);
1929 LDLM_LOCK_RELEASE(lock);
1933 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1935 if (req->rq_no_reply)
1938 req->rq_status = rc;
1939 if (!req->rq_packed_final) {
1940 rc = lustre_pack_reply(req, 1, NULL, NULL);
1944 return ptlrpc_reply(req);
1947 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
1948 enum ldlm_cancel_flags cancel_flags)
1950 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1953 spin_lock(&blp->blp_lock);
1954 if (blwi->blwi_lock &&
1955 ldlm_is_discard_data(blwi->blwi_lock)) {
1956 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1957 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1959 /* other blocking callbacks are added to the regular list */
1960 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1962 spin_unlock(&blp->blp_lock);
1964 wake_up(&blp->blp_waitq);
1966 /* can not check blwi->blwi_flags as blwi could be already freed in
1968 if (!(cancel_flags & LCF_ASYNC))
1969 wait_for_completion(&blwi->blwi_comp);
1974 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1975 struct ldlm_namespace *ns,
1976 struct ldlm_lock_desc *ld,
1977 struct list_head *cancels, int count,
1978 struct ldlm_lock *lock,
1979 enum ldlm_cancel_flags cancel_flags)
1981 init_completion(&blwi->blwi_comp);
1982 INIT_LIST_HEAD(&blwi->blwi_head);
1984 if (memory_pressure_get())
1985 blwi->blwi_mem_pressure = 1;
1988 blwi->blwi_flags = cancel_flags;
1990 blwi->blwi_ld = *ld;
1992 list_add(&blwi->blwi_head, cancels);
1993 list_del_init(cancels);
1994 blwi->blwi_count = count;
1996 blwi->blwi_lock = lock;
2001 * Queues a list of locks \a cancels containing \a count locks
2002 * for later processing by a blocking thread. If \a count is zero,
2003 * then the lock referenced as \a lock is queued instead.
2005 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2006 * If list addition fails an error is returned and caller is supposed to
2007 * call ->l_blocking_ast itself.
2009 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2010 struct ldlm_lock_desc *ld,
2011 struct ldlm_lock *lock,
2012 struct list_head *cancels, int count,
2013 enum ldlm_cancel_flags cancel_flags)
2017 if (cancels && count == 0)
2020 if (cancel_flags & LCF_ASYNC) {
2021 struct ldlm_bl_work_item *blwi;
2023 OBD_ALLOC(blwi, sizeof(*blwi));
2026 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2028 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2030 /* if it is synchronous call do minimum mem alloc, as it could
2031 * be triggered from kernel shrinker
2033 struct ldlm_bl_work_item blwi;
2035 memset(&blwi, 0, sizeof(blwi));
2036 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2037 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2042 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2043 struct ldlm_lock *lock)
2045 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2048 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2049 struct list_head *cancels, int count,
2050 enum ldlm_cancel_flags cancel_flags)
2052 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2055 int ldlm_bl_thread_wakeup(void)
2057 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2061 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2062 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2064 struct obd_device *obd = req->rq_export->exp_obd;
2071 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2073 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2075 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2077 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2080 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2082 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2084 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2087 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2090 /* We are responsible for swabbing contents of val */
2092 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2093 /* Pass it on to mdc (the "export" in this case) */
2094 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2096 sizeof(KEY_HSM_COPYTOOL_SEND),
2097 KEY_HSM_COPYTOOL_SEND,
2100 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2105 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2106 const char *msg, int rc,
2107 const struct lustre_handle *handle)
2109 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2110 "%s: [nid %s] [rc %d] [lock %#llx]",
2111 msg, libcfs_id2str(req->rq_peer), rc,
2112 handle ? handle->cookie : 0);
2113 if (req->rq_no_reply)
2114 CWARN("No reply was sent, maybe cause bug 21636.\n");
2116 CWARN("Send reply failed, maybe cause bug 21636.\n");
2119 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2120 static int ldlm_callback_handler(struct ptlrpc_request *req)
2122 struct ldlm_namespace *ns;
2123 struct ldlm_request *dlm_req;
2124 struct ldlm_lock *lock;
2128 /* Requests arrive in sender's byte order. The ptlrpc service
2129 * handler has already checked and, if necessary, byte-swapped the
2130 * incoming request message body, but I am responsible for the
2131 * message buffers. */
2133 /* do nothing for sec context finalize */
2134 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2137 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2139 if (req->rq_export == NULL) {
2140 rc = ldlm_callback_reply(req, -ENOTCONN);
2141 ldlm_callback_errmsg(req, "Operate on unconnected server",
2146 LASSERT(req->rq_export != NULL);
2147 LASSERT(req->rq_export->exp_obd != NULL);
2149 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2150 case LDLM_BL_CALLBACK:
2151 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2153 ldlm_callback_reply(req, -(int)cfs_fail_err);
2157 case LDLM_CP_CALLBACK:
2158 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2161 case LDLM_GL_CALLBACK:
2162 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2166 rc = ldlm_handle_setinfo(req);
2167 ldlm_callback_reply(req, rc);
2169 case LLOG_ORIGIN_HANDLE_CREATE:
2170 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
2171 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2173 rc = llog_origin_handle_open(req);
2174 ldlm_callback_reply(req, rc);
2176 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2177 req_capsule_set(&req->rq_pill,
2178 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
2179 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2181 rc = llog_origin_handle_next_block(req);
2182 ldlm_callback_reply(req, rc);
2184 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2185 req_capsule_set(&req->rq_pill,
2186 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
2187 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2189 rc = llog_origin_handle_read_header(req);
2190 ldlm_callback_reply(req, rc);
2192 case LLOG_ORIGIN_HANDLE_CLOSE:
2193 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2195 rc = llog_origin_handle_close(req);
2196 ldlm_callback_reply(req, rc);
2199 CERROR("unknown opcode %u\n",
2200 lustre_msg_get_opc(req->rq_reqmsg));
2201 ldlm_callback_reply(req, -EPROTO);
2205 ns = req->rq_export->exp_obd->obd_namespace;
2206 LASSERT(ns != NULL);
2208 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2210 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2211 if (dlm_req == NULL) {
2212 rc = ldlm_callback_reply(req, -EPROTO);
2213 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2218 /* Force a known safe race, send a cancel to the server for a lock
2219 * which the server has already started a blocking callback on. */
2220 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2221 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2222 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2224 CERROR("ldlm_cli_cancel: %d\n", rc);
2227 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2229 CDEBUG(D_DLMTRACE, "callback on lock %#llx - lock "
2230 "disappeared\n", dlm_req->lock_handle[0].cookie);
2231 rc = ldlm_callback_reply(req, -EINVAL);
2232 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2233 &dlm_req->lock_handle[0]);
2237 if (ldlm_is_fail_loc(lock) &&
2238 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2239 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2241 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2242 lock_res_and_lock(lock);
2243 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2245 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2246 /* If somebody cancels lock and cache is already dropped,
2247 * or lock is failed before cp_ast received on client,
2248 * we can tell the server we have no lock. Otherwise, we
2249 * should send cancel after dropping the cache. */
2250 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2251 ldlm_is_failed(lock)) {
2252 LDLM_DEBUG(lock, "callback on lock %llx - lock disappeared",
2253 dlm_req->lock_handle[0].cookie);
2254 unlock_res_and_lock(lock);
2255 LDLM_LOCK_RELEASE(lock);
2256 rc = ldlm_callback_reply(req, -EINVAL);
2257 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2258 &dlm_req->lock_handle[0]);
2261 /* BL_AST locks are not needed in LRU.
2262 * Let ldlm_cancel_lru() be fast. */
2263 ldlm_lock_remove_from_lru(lock);
2264 ldlm_set_bl_ast(lock);
2266 unlock_res_and_lock(lock);
2268 /* We want the ost thread to get this reply so that it can respond
2269 * to ost requests (write cache writeback) that might be triggered
2272 * But we'd also like to be able to indicate in the reply that we're
2273 * cancelling right now, because it's unused, or have an intent result
2274 * in the reply, so we might have to push the responsibility for sending
2275 * the reply down into the AST handlers, alas. */
2277 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2278 case LDLM_BL_CALLBACK:
2279 CDEBUG(D_INODE, "blocking ast\n");
2280 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2281 if (!ldlm_is_cancel_on_block(lock)) {
2282 rc = ldlm_callback_reply(req, 0);
2283 if (req->rq_no_reply || rc)
2284 ldlm_callback_errmsg(req, "Normal process", rc,
2285 &dlm_req->lock_handle[0]);
2287 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2288 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2290 case LDLM_CP_CALLBACK:
2291 CDEBUG(D_INODE, "completion ast\n");
2292 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2293 ldlm_callback_reply(req, 0);
2294 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2296 case LDLM_GL_CALLBACK:
2297 CDEBUG(D_INODE, "glimpse ast\n");
2298 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2299 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2302 LBUG(); /* checked above */
2308 #ifdef HAVE_SERVER_SUPPORT
2310 * Main handler for canceld thread.
2312 * Separated into its own thread to avoid deadlocks.
2314 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2319 /* Requests arrive in sender's byte order. The ptlrpc service
2320 * handler has already checked and, if necessary, byte-swapped the
2321 * incoming request message body, but I am responsible for the
2322 * message buffers. */
2324 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2326 if (req->rq_export == NULL) {
2327 struct ldlm_request *dlm_req;
2329 CERROR("%s from %s arrived at %lu with bad export cookie "
2331 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2332 libcfs_nid2str(req->rq_peer.nid),
2333 req->rq_arrival_time.tv_sec,
2334 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2336 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2337 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2338 dlm_req = req_capsule_client_get(&req->rq_pill,
2340 if (dlm_req != NULL)
2341 ldlm_lock_dump_handle(D_ERROR,
2342 &dlm_req->lock_handle[0]);
2344 ldlm_callback_reply(req, -ENOTCONN);
2348 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2350 /* XXX FIXME move this back to mds/handler.c, bug 249 */
2352 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2353 CDEBUG(D_INODE, "cancel\n");
2354 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2355 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2356 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2358 rc = ldlm_handle_cancel(req);
2363 CERROR("invalid opcode %d\n",
2364 lustre_msg_get_opc(req->rq_reqmsg));
2365 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2366 ldlm_callback_reply(req, -EINVAL);
2372 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2373 struct ldlm_lock *lock)
2375 struct ldlm_request *dlm_req;
2376 struct lustre_handle lockh;
2381 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2382 if (dlm_req == NULL)
2385 ldlm_lock2handle(lock, &lockh);
2386 for (i = 0; i < dlm_req->lock_count; i++) {
2387 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2389 DEBUG_REQ(D_RPCTRACE, req,
2390 "Prio raised by lock %#llx.", lockh.cookie);
2401 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2403 struct ldlm_request *dlm_req;
2408 /* no prolong in recovery */
2409 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2412 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2413 if (dlm_req == NULL)
2416 for (i = 0; i < dlm_req->lock_count; i++) {
2417 struct ldlm_lock *lock;
2419 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2423 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2425 LDLM_DEBUG(lock, "hpreq cancel lock");
2426 LDLM_LOCK_PUT(lock);
2435 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2436 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2437 .hpreq_check = ldlm_cancel_hpreq_check,
2441 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2445 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2447 if (req->rq_export == NULL)
2450 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2451 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2452 req->rq_ops = &ldlm_cancel_hpreq_ops;
2457 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2458 struct hlist_node *hnode, void *data)
2461 struct list_head *rpc_list = data;
2462 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2464 lock_res_and_lock(lock);
2466 if (lock->l_req_mode != lock->l_granted_mode) {
2467 unlock_res_and_lock(lock);
2471 LASSERT(lock->l_resource);
2472 if (lock->l_resource->lr_type != LDLM_IBITS &&
2473 lock->l_resource->lr_type != LDLM_PLAIN) {
2474 unlock_res_and_lock(lock);
2478 if (ldlm_is_ast_sent(lock)) {
2479 unlock_res_and_lock(lock);
2483 LASSERT(lock->l_blocking_ast);
2484 LASSERT(!lock->l_blocking_lock);
2486 ldlm_set_ast_sent(lock);
2487 if (lock->l_export && lock->l_export->exp_lock_hash) {
2488 /* NB: it's safe to call cfs_hash_del() even lock isn't
2489 * in exp_lock_hash. */
2490 /* In the function below, .hs_keycmp resolves to
2491 * ldlm_export_lock_keycmp() */
2492 /* coverity[overrun-buffer-val] */
2493 cfs_hash_del(lock->l_export->exp_lock_hash,
2494 &lock->l_remote_handle, &lock->l_exp_hash);
2497 list_add_tail(&lock->l_rk_ast, rpc_list);
2498 LDLM_LOCK_GET(lock);
2500 unlock_res_and_lock(lock);
2504 void ldlm_revoke_export_locks(struct obd_export *exp)
2506 struct list_head rpc_list;
2509 INIT_LIST_HEAD(&rpc_list);
2510 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2511 ldlm_revoke_lock_cb, &rpc_list, 0);
2512 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2513 LDLM_WORK_REVOKE_AST);
2517 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2518 #endif /* HAVE_SERVER_SUPPORT */
2520 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2521 struct ldlm_bl_work_item **p_blwi,
2522 struct obd_export **p_exp)
2524 struct ldlm_bl_work_item *blwi = NULL;
2525 static unsigned int num_bl = 0;
2526 static unsigned int num_stale;
2527 int num_th = atomic_read(&blp->blp_num_threads);
2529 *p_exp = obd_stale_export_get();
2531 spin_lock(&blp->blp_lock);
2532 if (*p_exp != NULL) {
2533 if (num_th == 1 || ++num_stale < num_th) {
2534 spin_unlock(&blp->blp_lock);
2541 /* process a request from the blp_list at least every blp_num_threads */
2542 if (!list_empty(&blp->blp_list) &&
2543 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2544 blwi = list_entry(blp->blp_list.next,
2545 struct ldlm_bl_work_item, blwi_entry);
2547 if (!list_empty(&blp->blp_prio_list))
2548 blwi = list_entry(blp->blp_prio_list.next,
2549 struct ldlm_bl_work_item,
2553 if (++num_bl >= num_th)
2555 list_del(&blwi->blwi_entry);
2557 spin_unlock(&blp->blp_lock);
2560 if (*p_exp != NULL && *p_blwi != NULL) {
2561 obd_stale_export_put(*p_exp);
2565 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2568 /* This only contains temporary data until the thread starts */
2569 struct ldlm_bl_thread_data {
2570 struct ldlm_bl_pool *bltd_blp;
2571 struct completion bltd_comp;
2575 static int ldlm_bl_thread_main(void *arg);
2577 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2579 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2580 struct task_struct *task;
2582 init_completion(&bltd.bltd_comp);
2584 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2585 if (bltd.bltd_num >= blp->blp_max_threads) {
2586 atomic_dec(&blp->blp_num_threads);
2590 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2592 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2593 atomic_dec(&blp->blp_num_threads);
2597 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2600 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2601 bltd.bltd_num, PTR_ERR(task));
2602 atomic_dec(&blp->blp_num_threads);
2603 return PTR_ERR(task);
2605 wait_for_completion(&bltd.bltd_comp);
2610 /* Not fatal if racy and have a few too many threads */
2611 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2612 struct ldlm_bl_work_item *blwi)
2614 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2617 if (atomic_read(&blp->blp_busy_threads) <
2618 atomic_read(&blp->blp_num_threads))
2621 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2622 blwi->blwi_mem_pressure))
2628 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2629 struct ldlm_bl_work_item *blwi)
2633 if (blwi->blwi_ns == NULL)
2634 /* added by ldlm_cleanup() */
2635 RETURN(LDLM_ITER_STOP);
2637 if (blwi->blwi_mem_pressure)
2638 memory_pressure_set();
2640 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2642 if (blwi->blwi_count) {
2644 /* The special case when we cancel locks in lru
2645 * asynchronously, we pass the list of locks here.
2646 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2647 * canceled locally yet. */
2648 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2651 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2654 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2657 if (blwi->blwi_mem_pressure)
2658 memory_pressure_clr();
2660 if (blwi->blwi_flags & LCF_ASYNC)
2661 OBD_FREE(blwi, sizeof(*blwi));
2663 complete(&blwi->blwi_comp);
2669 * Cancel stale locks on export. Cancel blocked locks first.
2670 * If the given export has blocked locks, the next in the list may have
2671 * them too, thus cancel not blocked locks only if the current export has
2674 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2675 struct obd_export *exp)
2680 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2682 num = ldlm_export_cancel_blocked_locks(exp);
2684 ldlm_export_cancel_locks(exp);
2686 obd_stale_export_put(exp);
2693 * Main blocking requests processing thread.
2695 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2696 * This thread in the end ends up doing actual call to ->l_blocking_ast
2699 static int ldlm_bl_thread_main(void *arg)
2701 struct ldlm_bl_pool *blp;
2702 struct ldlm_bl_thread_data *bltd = arg;
2705 blp = bltd->bltd_blp;
2707 complete(&bltd->bltd_comp);
2708 /* cannot use bltd after this, it is only on caller's stack */
2711 struct l_wait_info lwi = { 0 };
2712 struct ldlm_bl_work_item *blwi = NULL;
2713 struct obd_export *exp = NULL;
2716 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2719 l_wait_event_exclusive(blp->blp_waitq,
2720 ldlm_bl_get_work(blp, &blwi,
2723 atomic_inc(&blp->blp_busy_threads);
2725 if (ldlm_bl_thread_need_create(blp, blwi))
2726 /* discard the return value, we tried */
2727 ldlm_bl_thread_start(blp, true);
2730 rc = ldlm_bl_thread_exports(blp, exp);
2732 rc = ldlm_bl_thread_blwi(blp, blwi);
2734 atomic_dec(&blp->blp_busy_threads);
2736 if (rc == LDLM_ITER_STOP)
2739 /* If there are many namespaces, we will not sleep waiting for
2740 * work, and must do a cond_resched to avoid holding the CPU
2745 atomic_dec(&blp->blp_num_threads);
2746 complete(&blp->blp_comp);
2751 static int ldlm_setup(void);
2752 static int ldlm_cleanup(void);
2754 int ldlm_get_ref(void)
2758 mutex_lock(&ldlm_ref_mutex);
2759 if (++ldlm_refcount == 1) {
2764 mutex_unlock(&ldlm_ref_mutex);
2769 void ldlm_put_ref(void)
2772 mutex_lock(&ldlm_ref_mutex);
2773 if (ldlm_refcount == 1) {
2774 int rc = ldlm_cleanup();
2776 CERROR("ldlm_cleanup failed: %d\n", rc);
2782 mutex_unlock(&ldlm_ref_mutex);
2788 * Export handle<->lock hash operations.
2791 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
2793 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2797 ldlm_export_lock_key(struct hlist_node *hnode)
2799 struct ldlm_lock *lock;
2801 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2802 return &lock->l_remote_handle;
2806 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
2808 struct ldlm_lock *lock;
2810 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2811 lock->l_remote_handle = *(struct lustre_handle *)key;
2815 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
2817 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2821 ldlm_export_lock_object(struct hlist_node *hnode)
2823 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2827 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
2829 struct ldlm_lock *lock;
2831 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2832 LDLM_LOCK_GET(lock);
2836 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
2838 struct ldlm_lock *lock;
2840 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2841 LDLM_LOCK_RELEASE(lock);
2844 static struct cfs_hash_ops ldlm_export_lock_ops = {
2845 .hs_hash = ldlm_export_lock_hash,
2846 .hs_key = ldlm_export_lock_key,
2847 .hs_keycmp = ldlm_export_lock_keycmp,
2848 .hs_keycpy = ldlm_export_lock_keycpy,
2849 .hs_object = ldlm_export_lock_object,
2850 .hs_get = ldlm_export_lock_get,
2851 .hs_put = ldlm_export_lock_put,
2852 .hs_put_locked = ldlm_export_lock_put,
2855 int ldlm_init_export(struct obd_export *exp)
2860 exp->exp_lock_hash =
2861 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2862 HASH_EXP_LOCK_CUR_BITS,
2863 HASH_EXP_LOCK_MAX_BITS,
2864 HASH_EXP_LOCK_BKT_BITS, 0,
2865 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2866 &ldlm_export_lock_ops,
2867 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2868 CFS_HASH_NBLK_CHANGE);
2870 if (!exp->exp_lock_hash)
2873 rc = ldlm_init_flock_export(exp);
2879 ldlm_destroy_export(exp);
2882 EXPORT_SYMBOL(ldlm_init_export);
2884 void ldlm_destroy_export(struct obd_export *exp)
2887 cfs_hash_putref(exp->exp_lock_hash);
2888 exp->exp_lock_hash = NULL;
2890 ldlm_destroy_flock_export(exp);
2893 EXPORT_SYMBOL(ldlm_destroy_export);
2895 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
2896 struct attribute *attr,
2899 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
2902 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
2903 struct attribute *attr,
2910 rc = kstrtoul(buffer, 10, &val);
2914 ldlm_cancel_unused_locks_before_replay = val;
2918 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
2920 static struct attribute *ldlm_attrs[] = {
2921 &lustre_attr_cancel_unused_locks_before_replay.attr,
2925 static struct attribute_group ldlm_attr_group = {
2926 .attrs = ldlm_attrs,
2929 static int ldlm_setup(void)
2931 static struct ptlrpc_service_conf conf;
2932 struct ldlm_bl_pool *blp = NULL;
2933 #ifdef HAVE_SERVER_SUPPORT
2934 struct task_struct *task;
2935 #endif /* HAVE_SERVER_SUPPORT */
2941 if (ldlm_state != NULL)
2944 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2945 if (ldlm_state == NULL)
2948 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
2952 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
2956 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
2960 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
2964 #ifdef CONFIG_PROC_FS
2965 rc = ldlm_proc_setup();
2968 #endif /* CONFIG_PROC_FS */
2970 memset(&conf, 0, sizeof(conf));
2971 conf = (typeof(conf)) {
2972 .psc_name = "ldlm_cbd",
2973 .psc_watchdog_factor = 2,
2975 .bc_nbufs = LDLM_CLIENT_NBUFS,
2976 .bc_buf_size = LDLM_BUFSIZE,
2977 .bc_req_max_size = LDLM_MAXREQSIZE,
2978 .bc_rep_max_size = LDLM_MAXREPSIZE,
2979 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
2980 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
2983 .tc_thr_name = "ldlm_cb",
2984 .tc_thr_factor = LDLM_THR_FACTOR,
2985 .tc_nthrs_init = LDLM_NTHRS_INIT,
2986 .tc_nthrs_base = LDLM_NTHRS_BASE,
2987 .tc_nthrs_max = LDLM_NTHRS_MAX,
2988 .tc_nthrs_user = ldlm_num_threads,
2989 .tc_cpu_affinity = 1,
2990 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
2993 .cc_pattern = ldlm_cpts,
2996 .so_req_handler = ldlm_callback_handler,
2999 ldlm_state->ldlm_cb_service = \
3000 ptlrpc_register_service(&conf, ldlm_svc_kset,
3002 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3003 CERROR("failed to start service\n");
3004 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3005 ldlm_state->ldlm_cb_service = NULL;
3009 #ifdef HAVE_SERVER_SUPPORT
3010 memset(&conf, 0, sizeof(conf));
3011 conf = (typeof(conf)) {
3012 .psc_name = "ldlm_canceld",
3013 .psc_watchdog_factor = 6,
3015 .bc_nbufs = LDLM_SERVER_NBUFS,
3016 .bc_buf_size = LDLM_BUFSIZE,
3017 .bc_req_max_size = LDLM_MAXREQSIZE,
3018 .bc_rep_max_size = LDLM_MAXREPSIZE,
3019 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3020 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3024 .tc_thr_name = "ldlm_cn",
3025 .tc_thr_factor = LDLM_THR_FACTOR,
3026 .tc_nthrs_init = LDLM_NTHRS_INIT,
3027 .tc_nthrs_base = LDLM_NTHRS_BASE,
3028 .tc_nthrs_max = LDLM_NTHRS_MAX,
3029 .tc_nthrs_user = ldlm_num_threads,
3030 .tc_cpu_affinity = 1,
3031 .tc_ctx_tags = LCT_MD_THREAD | \
3036 .cc_pattern = ldlm_cpts,
3039 .so_req_handler = ldlm_cancel_handler,
3040 .so_hpreq_handler = ldlm_hpreq_handler,
3043 ldlm_state->ldlm_cancel_service = \
3044 ptlrpc_register_service(&conf, ldlm_svc_kset,
3046 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3047 CERROR("failed to start service\n");
3048 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3049 ldlm_state->ldlm_cancel_service = NULL;
3052 #endif /* HAVE_SERVER_SUPPORT */
3054 OBD_ALLOC(blp, sizeof(*blp));
3056 GOTO(out, rc = -ENOMEM);
3057 ldlm_state->ldlm_bl_pool = blp;
3059 spin_lock_init(&blp->blp_lock);
3060 INIT_LIST_HEAD(&blp->blp_list);
3061 INIT_LIST_HEAD(&blp->blp_prio_list);
3062 init_waitqueue_head(&blp->blp_waitq);
3063 atomic_set(&blp->blp_num_threads, 0);
3064 atomic_set(&blp->blp_busy_threads, 0);
3066 if (ldlm_num_threads == 0) {
3067 blp->blp_min_threads = LDLM_NTHRS_INIT;
3068 blp->blp_max_threads = LDLM_NTHRS_MAX;
3070 blp->blp_min_threads = blp->blp_max_threads = \
3071 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3075 for (i = 0; i < blp->blp_min_threads; i++) {
3076 rc = ldlm_bl_thread_start(blp, false);
3081 #ifdef HAVE_SERVER_SUPPORT
3082 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3085 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3089 wait_event(expired_lock_wait_queue,
3090 expired_lock_thread_state == ELT_READY);
3091 #endif /* HAVE_SERVER_SUPPORT */
3093 rc = ldlm_pools_init();
3095 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3099 rc = ldlm_reclaim_setup();
3101 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3111 static int ldlm_cleanup(void)
3115 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3116 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3117 CERROR("ldlm still has namespaces; clean these up first.\n");
3118 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3119 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3123 ldlm_reclaim_cleanup();
3126 if (ldlm_state->ldlm_bl_pool != NULL) {
3127 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3129 while (atomic_read(&blp->blp_num_threads) > 0) {
3130 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3132 init_completion(&blp->blp_comp);
3134 spin_lock(&blp->blp_lock);
3135 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3136 wake_up(&blp->blp_waitq);
3137 spin_unlock(&blp->blp_lock);
3139 wait_for_completion(&blp->blp_comp);
3142 OBD_FREE(blp, sizeof(*blp));
3145 if (ldlm_state->ldlm_cb_service != NULL)
3146 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3147 #ifdef HAVE_SERVER_SUPPORT
3148 if (ldlm_state->ldlm_cancel_service != NULL)
3149 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3153 kset_unregister(ldlm_ns_kset);
3155 kset_unregister(ldlm_svc_kset);
3157 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3158 kobject_put(ldlm_kobj);
3161 ldlm_proc_cleanup();
3163 #ifdef HAVE_SERVER_SUPPORT
3164 if (expired_lock_thread_state != ELT_STOPPED) {
3165 expired_lock_thread_state = ELT_TERMINATE;
3166 wake_up(&expired_lock_wait_queue);
3167 wait_event(expired_lock_wait_queue,
3168 expired_lock_thread_state == ELT_STOPPED);
3172 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3180 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3181 sizeof(struct ldlm_resource), 0,
3182 SLAB_HWCACHE_ALIGN, NULL);
3183 if (ldlm_resource_slab == NULL)
3186 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3187 sizeof(struct ldlm_lock), 0,
3188 SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU, NULL);
3189 if (ldlm_lock_slab == NULL)
3192 ldlm_interval_slab = kmem_cache_create("interval_node",
3193 sizeof(struct ldlm_interval),
3194 0, SLAB_HWCACHE_ALIGN, NULL);
3195 if (ldlm_interval_slab == NULL)
3198 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3199 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3200 0, SLAB_HWCACHE_ALIGN, NULL);
3201 if (ldlm_interval_tree_slab == NULL)
3204 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3205 class_export_dump_hook = ldlm_dump_export_locks;
3210 kmem_cache_destroy(ldlm_interval_slab);
3212 kmem_cache_destroy(ldlm_lock_slab);
3214 kmem_cache_destroy(ldlm_resource_slab);
3219 void ldlm_exit(void)
3222 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3223 kmem_cache_destroy(ldlm_resource_slab);
3224 /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3225 * synchronize_rcu() to wait a grace period elapsed, so that
3226 * ldlm_lock_free() get a chance to be called. */
3228 kmem_cache_destroy(ldlm_lock_slab);
3229 kmem_cache_destroy(ldlm_interval_slab);
3230 kmem_cache_destroy(ldlm_interval_tree_slab);