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 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;
85 * blp_prio_list is used for callbacks that should be handled
86 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
89 struct list_head blp_prio_list;
92 * blp_list is used for all other callbacks which are likely
93 * to take longer to process.
95 struct list_head blp_list;
97 wait_queue_head_t blp_waitq;
98 struct completion blp_comp;
99 atomic_t blp_num_threads;
100 atomic_t blp_busy_threads;
105 struct ldlm_bl_work_item {
106 struct list_head blwi_entry;
107 struct ldlm_namespace *blwi_ns;
108 struct ldlm_lock_desc blwi_ld;
109 struct ldlm_lock *blwi_lock;
110 struct list_head blwi_head;
112 struct completion blwi_comp;
113 enum ldlm_cancel_flags blwi_flags;
114 int blwi_mem_pressure;
117 #ifdef HAVE_SERVER_SUPPORT
120 * Protects both waiting_locks_list and expired_lock_thread.
122 static DEFINE_SPINLOCK(waiting_locks_spinlock); /* BH lock (timer) */
125 * List for contended locks.
127 * As soon as a lock is contended, it gets placed on this list and
128 * expected time to get a response is filled in the lock. A special
129 * thread walks the list looking for locks that should be released and
130 * schedules client evictions for those that have not been released in
133 * All access to it should be under waiting_locks_spinlock.
135 static LIST_HEAD(waiting_locks_list);
136 static void waiting_locks_callback(unsigned long unused);
137 static DEFINE_TIMER(waiting_locks_timer, waiting_locks_callback, 0, 0);
145 static DECLARE_WAIT_QUEUE_HEAD(expired_lock_wait_queue);
146 static enum elt_state expired_lock_thread_state = ELT_STOPPED;
147 static int expired_lock_dump;
148 static LIST_HEAD(expired_lock_list);
150 static inline int have_expired_locks(void)
155 spin_lock_bh(&waiting_locks_spinlock);
156 need_to_run = !list_empty(&expired_lock_list);
157 spin_unlock_bh(&waiting_locks_spinlock);
163 * Check expired lock list for expired locks and time them out.
165 static int expired_lock_main(void *arg)
167 struct list_head *expired = &expired_lock_list;
168 struct l_wait_info lwi = { 0 };
173 expired_lock_thread_state = ELT_READY;
174 wake_up(&expired_lock_wait_queue);
177 l_wait_event(expired_lock_wait_queue,
178 have_expired_locks() ||
179 expired_lock_thread_state == ELT_TERMINATE,
182 spin_lock_bh(&waiting_locks_spinlock);
183 if (expired_lock_dump) {
184 spin_unlock_bh(&waiting_locks_spinlock);
186 /* from waiting_locks_callback, but not in timer */
187 libcfs_debug_dumplog();
189 spin_lock_bh(&waiting_locks_spinlock);
190 expired_lock_dump = 0;
195 while (!list_empty(expired)) {
196 struct obd_export *export;
197 struct ldlm_lock *lock;
199 lock = list_entry(expired->next, struct ldlm_lock,
201 if ((void *)lock < LP_POISON + PAGE_SIZE &&
202 (void *)lock >= LP_POISON) {
203 spin_unlock_bh(&waiting_locks_spinlock);
204 CERROR("free lock on elt list %p\n", lock);
207 list_del_init(&lock->l_pending_chain);
208 if ((void *)lock->l_export <
209 LP_POISON + PAGE_SIZE &&
210 (void *)lock->l_export >= LP_POISON) {
211 CERROR("lock with free export on elt list %p\n",
213 lock->l_export = NULL;
214 LDLM_ERROR(lock, "free export");
215 /* release extra ref grabbed by
216 * ldlm_add_waiting_lock() or
217 * ldlm_failed_ast() */
218 LDLM_LOCK_RELEASE(lock);
222 if (ldlm_is_destroyed(lock)) {
223 /* release the lock refcount where
224 * waiting_locks_callback() founds */
225 LDLM_LOCK_RELEASE(lock);
228 export = class_export_lock_get(lock->l_export, lock);
229 spin_unlock_bh(&waiting_locks_spinlock);
231 spin_lock_bh(&export->exp_bl_list_lock);
232 list_del_init(&lock->l_exp_list);
233 spin_unlock_bh(&export->exp_bl_list_lock);
236 class_fail_export(export);
237 class_export_lock_put(export, lock);
239 /* release extra ref grabbed by ldlm_add_waiting_lock()
240 * or ldlm_failed_ast() */
241 LDLM_LOCK_RELEASE(lock);
243 spin_lock_bh(&waiting_locks_spinlock);
245 spin_unlock_bh(&waiting_locks_spinlock);
247 if (do_dump && obd_dump_on_eviction) {
248 CERROR("dump the log upon eviction\n");
249 libcfs_debug_dumplog();
252 if (expired_lock_thread_state == ELT_TERMINATE)
256 expired_lock_thread_state = ELT_STOPPED;
257 wake_up(&expired_lock_wait_queue);
261 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
262 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds);
265 * Check if there is a request in the export request list
266 * which prevents the lock canceling.
268 static int ldlm_lock_busy(struct ldlm_lock *lock)
270 struct ptlrpc_request *req;
274 if (lock->l_export == NULL)
277 spin_lock_bh(&lock->l_export->exp_rpc_lock);
278 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
280 if (req->rq_ops->hpreq_lock_match) {
281 match = req->rq_ops->hpreq_lock_match(req, lock);
286 spin_unlock_bh(&lock->l_export->exp_rpc_lock);
290 /* This is called from within a timer interrupt and cannot schedule */
291 static void waiting_locks_callback(unsigned long unused)
293 struct ldlm_lock *lock;
296 spin_lock_bh(&waiting_locks_spinlock);
297 while (!list_empty(&waiting_locks_list)) {
298 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
300 if (cfs_time_after(lock->l_callback_timeout,
301 cfs_time_current()) ||
302 (lock->l_req_mode == LCK_GROUP))
305 /* Check if we need to prolong timeout */
306 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
307 ldlm_lock_busy(lock)) {
310 if (lock->l_pending_chain.next == &waiting_locks_list)
315 spin_unlock_bh(&waiting_locks_spinlock);
316 LDLM_DEBUG(lock, "prolong the busy lock");
317 ldlm_refresh_waiting_lock(lock,
318 ldlm_bl_timeout(lock) >> 1);
319 spin_lock_bh(&waiting_locks_spinlock);
322 LDLM_LOCK_RELEASE(lock);
326 LDLM_LOCK_RELEASE(lock);
329 ldlm_lock_to_ns(lock)->ns_timeouts++;
330 LDLM_ERROR(lock, "lock callback timer expired after %llds: "
331 "evicting client at %s ",
332 ktime_get_real_seconds() - lock->l_last_activity,
334 lock->l_export->exp_connection->c_peer.nid));
336 /* no needs to take an extra ref on the lock since it was in
337 * the waiting_locks_list and ldlm_add_waiting_lock()
338 * already grabbed a ref */
339 list_del(&lock->l_pending_chain);
340 list_add(&lock->l_pending_chain, &expired_lock_list);
344 if (!list_empty(&expired_lock_list)) {
345 if (obd_dump_on_timeout && need_dump)
346 expired_lock_dump = __LINE__;
348 wake_up(&expired_lock_wait_queue);
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, &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 (lock->l_export && lock->l_export->exp_libclient) {
659 "%s AST (req@%p x%llu) to liblustre client (nid %s) timeout, just cancelling lock",
660 ast_type, req, req->rq_xid,
661 libcfs_nid2str(peer.nid));
662 ldlm_lock_cancel(lock);
664 } else if (ldlm_is_cancel(lock)) {
666 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
667 ast_type, req, req->rq_xid,
668 libcfs_nid2str(peer.nid));
669 ldlm_lock_cancel(lock);
671 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
673 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
674 /* Upon umount process the AST fails because cannot be
675 * sent. This shouldn't lead to the client eviction.
676 * -ENODEV error is returned by ptl_send_rpc() for
677 * new request in such import.
678 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
679 * if imp_invalid is set or obd_no_recov.
680 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
681 * in ptlrpc_import_delay_req() as well with -EIO code.
682 * In all such cases errors are ignored.
684 LDLM_DEBUG(lock, "%s AST can't be sent due to a server"
685 " %s failure or umount process: rc = %d\n",
687 req->rq_import->imp_obd->obd_name, rc);
690 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
691 libcfs_nid2str(peer.nid),
692 req->rq_replied ? "returned error from" :
693 "failed to reply to",
694 ast_type, req, req->rq_xid,
695 (req->rq_repmsg != NULL) ?
696 lustre_msg_get_status(req->rq_repmsg) : 0,
698 ldlm_failed_ast(lock, rc, ast_type);
704 struct ldlm_resource *res = lock->l_resource;
707 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
708 libcfs_nid2str(peer.nid),
710 lustre_msg_get_status(req->rq_repmsg) : -1,
711 ast_type, req, req->rq_xid);
713 /* update lvbo to return proper attributes.
715 ldlm_resource_getref(res);
716 ldlm_lvbo_update(res, lock, NULL, 1);
717 ldlm_resource_putref(res);
719 ldlm_lock_cancel(lock);
726 static int ldlm_cb_interpret(const struct lu_env *env,
727 struct ptlrpc_request *req, void *data, int rc)
729 struct ldlm_cb_async_args *ca = data;
730 struct ldlm_lock *lock = ca->ca_lock;
731 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
734 LASSERT(lock != NULL);
737 case LDLM_GL_CALLBACK:
738 /* Update the LVB from disk if the AST failed
739 * (this is a legal race)
741 * - Glimpse callback of local lock just returns
742 * -ELDLM_NO_LOCK_DATA.
743 * - Glimpse callback of remote lock might return
744 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
746 if (unlikely(arg->gl_interpret_reply)) {
747 rc = arg->gl_interpret_reply(env, req, data, rc);
748 } else if (rc == -ELDLM_NO_LOCK_DATA) {
749 LDLM_DEBUG(lock, "lost race - client has a lock but no "
751 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
752 } else if (rc != 0) {
753 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
755 rc = ldlm_lvbo_update(lock->l_resource, lock, req, 1);
758 case LDLM_BL_CALLBACK:
760 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
762 case LDLM_CP_CALLBACK:
764 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
767 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
772 /* release extra reference taken in ldlm_ast_fini() */
773 LDLM_LOCK_RELEASE(lock);
776 atomic_inc(&arg->restart);
781 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
783 struct ldlm_cb_async_args *ca = data;
784 struct ldlm_lock *lock = ca->ca_lock;
786 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
789 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
790 struct ldlm_cb_set_arg *arg,
791 struct ldlm_lock *lock,
797 if (unlikely(instant_cancel)) {
798 rc = ptl_send_rpc(req, 1);
799 ptlrpc_req_finished(req);
801 atomic_inc(&arg->restart);
804 ptlrpc_set_add_req(arg->set, req);
811 * Check if there are requests in the export request list which prevent
812 * the lock canceling and make these requests high priority ones.
814 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
816 struct ptlrpc_request *req;
819 if (lock->l_export == NULL) {
820 LDLM_DEBUG(lock, "client lock: no-op");
824 spin_lock_bh(&lock->l_export->exp_rpc_lock);
825 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
827 /* Do not process requests that were not yet added to there
828 * incoming queue or were already removed from there for
829 * processing. We evaluate ptlrpc_nrs_req_can_move() without
830 * holding svcpt->scp_req_lock, and then redo the check with
831 * the lock held once we need to obtain a reliable result.
833 if (ptlrpc_nrs_req_can_move(req) &&
834 req->rq_ops->hpreq_lock_match &&
835 req->rq_ops->hpreq_lock_match(req, lock))
836 ptlrpc_nrs_req_hp_move(req);
838 spin_unlock_bh(&lock->l_export->exp_rpc_lock);
843 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
844 * enqueued server lock conflicts with given one.
846 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
847 * to wait for client response.
849 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
850 struct ldlm_lock_desc *desc,
851 void *data, int flag)
853 struct ldlm_cb_async_args *ca;
854 struct ldlm_cb_set_arg *arg = data;
855 struct ldlm_request *body;
856 struct ptlrpc_request *req;
857 int instant_cancel = 0;
861 if (flag == LDLM_CB_CANCELING)
862 /* Don't need to do anything here. */
865 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
866 LDLM_DEBUG(lock, "dropping BL AST");
871 LASSERT(data != NULL);
872 if (lock->l_export->exp_obd->obd_recovering != 0)
873 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
875 ldlm_lock_reorder_req(lock);
877 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
878 &RQF_LDLM_BL_CALLBACK,
879 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
883 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
884 ca = ptlrpc_req_async_args(req);
885 ca->ca_set_arg = arg;
888 req->rq_interpret_reply = ldlm_cb_interpret;
890 lock_res_and_lock(lock);
891 if (ldlm_is_destroyed(lock)) {
892 /* What's the point? */
893 unlock_res_and_lock(lock);
894 ptlrpc_req_finished(req);
898 if (lock->l_granted_mode != lock->l_req_mode) {
899 /* this blocking AST will be communicated as part of the
900 * completion AST instead */
901 ldlm_add_blocked_lock(lock);
902 ldlm_set_waited(lock);
903 unlock_res_and_lock(lock);
905 ptlrpc_req_finished(req);
906 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
910 if (ldlm_is_cancel_on_block(lock))
913 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
914 body->lock_handle[0] = lock->l_remote_handle;
915 body->lock_desc = *desc;
916 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
918 LDLM_DEBUG(lock, "server preparing blocking AST");
920 ptlrpc_request_set_replen(req);
921 ldlm_set_cbpending(lock);
922 if (instant_cancel) {
923 unlock_res_and_lock(lock);
924 ldlm_lock_cancel(lock);
926 req->rq_no_resend = 1;
928 LASSERT(lock->l_granted_mode == lock->l_req_mode);
929 ldlm_add_waiting_lock(lock);
930 unlock_res_and_lock(lock);
932 /* Do not resend after lock callback timeout */
933 req->rq_delay_limit = ldlm_bl_timeout(lock);
934 req->rq_resend_cb = ldlm_update_resend;
937 req->rq_send_state = LUSTRE_IMP_FULL;
938 /* ptlrpc_request_alloc_pack already set timeout */
940 req->rq_timeout = ldlm_get_rq_timeout();
942 lock->l_last_activity = ktime_get_real_seconds();
944 if (lock->l_export && lock->l_export->exp_nid_stats &&
945 lock->l_export->exp_nid_stats->nid_ldlm_stats)
946 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
947 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
949 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
955 * ->l_completion_ast callback for a remote lock in server namespace.
957 * Sends AST to the client notifying it of lock granting. If initial
958 * lock response was not sent yet, instead of sending another RPC, just
959 * mark the lock as granted and client will understand
961 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
963 struct ldlm_cb_set_arg *arg = data;
964 struct ldlm_request *body;
965 struct ptlrpc_request *req;
966 struct ldlm_cb_async_args *ca;
967 int instant_cancel = 0;
972 LASSERT(lock != NULL);
973 LASSERT(data != NULL);
975 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
976 LDLM_DEBUG(lock, "dropping CP AST");
980 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
981 &RQF_LDLM_CP_CALLBACK);
985 /* server namespace, doesn't need lock */
986 lvb_len = ldlm_lvbo_size(lock);
987 /* LU-3124 & LU-2187: to not return layout in completion AST because
988 * it may deadlock for LU-2187, or client may not have enough space
989 * for large layout. The layout will be returned to client with an
990 * extra RPC to fetch xattr.lov */
991 if (ldlm_has_layout(lock))
994 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
995 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
997 ptlrpc_request_free(req);
1001 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1002 ca = ptlrpc_req_async_args(req);
1003 ca->ca_set_arg = arg;
1006 req->rq_interpret_reply = ldlm_cb_interpret;
1007 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1009 body->lock_handle[0] = lock->l_remote_handle;
1010 body->lock_flags = ldlm_flags_to_wire(flags);
1011 ldlm_lock2desc(lock, &body->lock_desc);
1013 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1015 lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len);
1017 /* We still need to send the RPC to wake up the blocked
1018 * enqueue thread on the client.
1020 * Consider old client, there is no better way to notify
1021 * the failure, just zero-sized the LVB, then the client
1022 * will fail out as "-EPROTO". */
1023 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1027 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1032 lock->l_last_activity = ktime_get_real_seconds();
1034 LDLM_DEBUG(lock, "server preparing completion AST");
1036 ptlrpc_request_set_replen(req);
1038 req->rq_send_state = LUSTRE_IMP_FULL;
1039 /* ptlrpc_request_pack already set timeout */
1041 req->rq_timeout = ldlm_get_rq_timeout();
1043 /* We only send real blocking ASTs after the lock is granted */
1044 lock_res_and_lock(lock);
1045 if (ldlm_is_ast_sent(lock)) {
1046 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1047 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1048 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1051 /* We might get here prior to ldlm_handle_enqueue setting
1052 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1053 * into waiting list, but this is safe and similar code in
1054 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1055 * that would not only cancel the lock, but will also remove
1056 * it from waiting list */
1057 if (ldlm_is_cancel_on_block(lock)) {
1058 unlock_res_and_lock(lock);
1059 ldlm_lock_cancel(lock);
1062 req->rq_no_resend = 1;
1064 lock_res_and_lock(lock);
1066 /* start the lock-timeout clock */
1067 ldlm_add_waiting_lock(lock);
1068 /* Do not resend after lock callback timeout */
1069 req->rq_delay_limit = ldlm_bl_timeout(lock);
1070 req->rq_resend_cb = ldlm_update_resend;
1073 unlock_res_and_lock(lock);
1075 if (lock->l_export && lock->l_export->exp_nid_stats &&
1076 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1077 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1078 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1080 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1082 RETURN(lvb_len < 0 ? lvb_len : rc);
1086 * Server side ->l_glimpse_ast handler for client locks.
1088 * Sends glimpse AST to the client and waits for reply. Then updates
1089 * lvbo with the result.
1091 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1093 struct ldlm_cb_set_arg *arg = data;
1094 struct ldlm_request *body;
1095 struct ptlrpc_request *req;
1096 struct ldlm_cb_async_args *ca;
1098 struct req_format *req_fmt;
1101 LASSERT(lock != NULL);
1103 if (arg->gl_desc != NULL)
1104 /* There is a glimpse descriptor to pack */
1105 req_fmt = &RQF_LDLM_GL_DESC_CALLBACK;
1107 req_fmt = &RQF_LDLM_GL_CALLBACK;
1109 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1110 req_fmt, LUSTRE_DLM_VERSION,
1116 if (arg->gl_desc != NULL) {
1117 /* copy the GL descriptor */
1118 union ldlm_gl_desc *desc;
1119 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1120 *desc = *arg->gl_desc;
1123 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1124 body->lock_handle[0] = lock->l_remote_handle;
1125 ldlm_lock2desc(lock, &body->lock_desc);
1127 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1128 ca = ptlrpc_req_async_args(req);
1129 ca->ca_set_arg = arg;
1132 /* server namespace, doesn't need lock */
1133 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1134 ldlm_lvbo_size(lock));
1135 ptlrpc_request_set_replen(req);
1137 req->rq_send_state = LUSTRE_IMP_FULL;
1138 /* ptlrpc_request_alloc_pack already set timeout */
1140 req->rq_timeout = ldlm_get_rq_timeout();
1142 lock->l_last_activity = ktime_get_real_seconds();
1144 req->rq_interpret_reply = ldlm_cb_interpret;
1146 if (lock->l_export && lock->l_export->exp_nid_stats &&
1147 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1148 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1149 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1151 rc = ldlm_ast_fini(req, arg, lock, 0);
1155 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
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);
1189 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1190 * service threads to carry out client lock enqueueing requests.
1192 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1193 struct ptlrpc_request *req,
1194 const struct ldlm_request *dlm_req,
1195 const struct ldlm_callback_suite *cbs)
1197 struct ldlm_reply *dlm_rep;
1199 enum ldlm_error err = ELDLM_OK;
1200 struct ldlm_lock *lock = NULL;
1201 void *cookie = NULL;
1203 struct ldlm_resource *res = NULL;
1206 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1208 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1209 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1211 LASSERT(req->rq_export);
1213 /* for intent enqueue the stat will be updated inside intent policy */
1214 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1215 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1216 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1218 if (req->rq_export && req->rq_export->exp_nid_stats &&
1219 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1220 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1221 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1223 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1224 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1225 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1226 dlm_req->lock_desc.l_resource.lr_type);
1227 GOTO(out, rc = -EFAULT);
1230 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1231 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1232 dlm_req->lock_desc.l_req_mode &
1233 (dlm_req->lock_desc.l_req_mode-1))) {
1234 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1235 dlm_req->lock_desc.l_req_mode);
1236 GOTO(out, rc = -EFAULT);
1239 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) {
1240 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1242 DEBUG_REQ(D_ERROR, req,
1243 "PLAIN lock request from IBITS client?");
1244 GOTO(out, rc = -EPROTO);
1246 } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1248 DEBUG_REQ(D_ERROR, req,
1249 "IBITS lock request from unaware client?");
1250 GOTO(out, rc = -EPROTO);
1253 if (unlikely((flags & LDLM_FL_REPLAY) ||
1254 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1255 /* Find an existing lock in the per-export lock hash */
1256 /* In the function below, .hs_keycmp resolves to
1257 * ldlm_export_lock_keycmp() */
1258 /* coverity[overrun-buffer-val] */
1259 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1260 (void *)&dlm_req->lock_handle[0]);
1262 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie %#llx",
1263 lock->l_handle.h_cookie);
1264 flags |= LDLM_FL_RESENT;
1265 GOTO(existing_lock, rc = 0);
1268 if (ldlm_reclaim_full()) {
1269 DEBUG_REQ(D_DLMTRACE, req, "Too many granted locks, "
1270 "reject current enqueue request and let the "
1271 "client retry later.\n");
1272 GOTO(out, rc = -EINPROGRESS);
1276 /* The lock's callback data might be set in the policy function */
1277 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1278 dlm_req->lock_desc.l_resource.lr_type,
1279 dlm_req->lock_desc.l_req_mode,
1280 cbs, NULL, 0, LVB_T_NONE);
1287 lock->l_remote_handle = dlm_req->lock_handle[0];
1288 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1290 /* Initialize resource lvb but not for a lock being replayed since
1291 * Client already got lvb sent in this case.
1292 * This must occur early since some policy methods assume resource
1293 * lvb is available (lr_lvb_data != NULL).
1295 res = lock->l_resource;
1296 if (!(flags & LDLM_FL_REPLAY)) {
1297 /* non-replayed lock, delayed lvb init may need to be done */
1298 rc = ldlm_lvbo_init(res);
1300 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1305 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1306 /* Don't enqueue a lock onto the export if it is been disonnected
1307 * due to eviction (bug 3822) or server umount (bug 24324).
1308 * Cancel it now instead. */
1309 if (req->rq_export->exp_disconnected) {
1310 LDLM_ERROR(lock, "lock on disconnected export %p",
1312 GOTO(out, rc = -ENOTCONN);
1315 lock->l_export = class_export_lock_get(req->rq_export, lock);
1316 if (lock->l_export->exp_lock_hash)
1317 cfs_hash_add(lock->l_export->exp_lock_hash,
1318 &lock->l_remote_handle,
1321 /* Inherit the enqueue flags before the operation, because we do not
1322 * keep the res lock on return and next operations (BL AST) may proceed
1324 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1325 LDLM_FL_INHERIT_MASK);
1327 ldlm_convert_policy_to_local(req->rq_export,
1328 dlm_req->lock_desc.l_resource.lr_type,
1329 &dlm_req->lock_desc.l_policy_data,
1330 &lock->l_policy_data);
1331 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1332 lock->l_req_extent = lock->l_policy_data.l_extent;
1335 if (flags & LDLM_FL_HAS_INTENT) {
1336 /* In this case, the reply buffer is allocated deep in
1337 * local_lock_enqueue by the policy function. */
1340 /* based on the assumption that lvb size never changes during
1341 * resource life time otherwise it need resource->lr_lock's
1343 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1344 RCL_SERVER, ldlm_lvbo_size(lock));
1346 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1347 GOTO(out, rc = -ENOMEM);
1349 rc = req_capsule_server_pack(&req->rq_pill);
1354 err = ldlm_lock_enqueue(ns, &lock, cookie, &flags);
1361 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1363 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1364 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1366 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1367 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1369 /* We never send a blocking AST until the lock is granted, but
1370 * we can tell it right now */
1371 lock_res_and_lock(lock);
1373 /* Now take into account flags to be inherited from original lock
1374 request both in reply to client and in our own lock flags. */
1375 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1376 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1378 /* Don't move a pending lock onto the export if it has already been
1379 * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1380 * Cancel it now instead. */
1381 if (unlikely(req->rq_export->exp_disconnected ||
1382 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1383 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1385 } else if (ldlm_is_ast_sent(lock)) {
1386 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1387 if (lock->l_granted_mode == lock->l_req_mode) {
1389 * Only cancel lock if it was granted, because it would
1390 * be destroyed immediately and would never be granted
1391 * in the future, causing timeouts on client. Not
1392 * granted lock will be cancelled immediately after
1393 * sending completion AST.
1395 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1396 unlock_res_and_lock(lock);
1397 ldlm_lock_cancel(lock);
1398 lock_res_and_lock(lock);
1400 ldlm_add_waiting_lock(lock);
1403 /* Make sure we never ever grant usual metadata locks to liblustre
1405 if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1406 dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1407 req->rq_export->exp_libclient) {
1408 if (unlikely(!ldlm_is_cancel_on_block(lock) ||
1409 !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1410 CERROR("Granting sync lock to libclient. "
1411 "req fl %d, rep fl %d, lock fl %#llx\n",
1412 dlm_req->lock_flags, dlm_rep->lock_flags,
1414 LDLM_ERROR(lock, "sync lock");
1415 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1416 struct ldlm_intent *it;
1418 it = req_capsule_client_get(&req->rq_pill,
1421 CERROR("This is intent %s (%llu)\n",
1422 ldlm_it2str(it->opc), it->opc);
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);
1662 if (!ldlm_is_discard_data(lock))
1663 ldlm_lvbo_update(res, lock, NULL, 1);
1668 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock)) {
1669 time64_t delay = ktime_get_real_seconds() -
1670 lock->l_last_activity;
1671 LDLM_DEBUG(lock, "server cancels blocked lock after %llds",
1673 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1675 ldlm_lock_cancel(lock);
1676 LDLM_LOCK_PUT(lock);
1679 ldlm_reprocess_all(pres);
1680 LDLM_RESOURCE_DELREF(pres);
1681 ldlm_resource_putref(pres);
1683 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1686 EXPORT_SYMBOL(ldlm_request_cancel);
1689 * Main LDLM entry point for server code to cancel locks.
1691 * Typically gets called from service handler on LDLM_CANCEL opc.
1693 int ldlm_handle_cancel(struct ptlrpc_request *req)
1695 struct ldlm_request *dlm_req;
1699 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1700 if (dlm_req == NULL) {
1701 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1705 if (req->rq_export && req->rq_export->exp_nid_stats &&
1706 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1707 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1708 LDLM_CANCEL - LDLM_FIRST_OPC);
1710 rc = req_capsule_server_pack(&req->rq_pill);
1714 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1715 req->rq_status = LUSTRE_ESTALE;
1717 RETURN(ptlrpc_reply(req));
1719 #endif /* HAVE_SERVER_SUPPORT */
1722 * Callback handler for receiving incoming blocking ASTs.
1724 * This can only happen on client side.
1726 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1727 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1732 LDLM_DEBUG(lock, "client blocking AST callback handler");
1734 lock_res_and_lock(lock);
1735 ldlm_set_cbpending(lock);
1737 if (ldlm_is_cancel_on_block(lock))
1738 ldlm_set_cancel(lock);
1740 do_ast = (!lock->l_readers && !lock->l_writers);
1741 unlock_res_and_lock(lock);
1744 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1745 lock, lock->l_blocking_ast);
1746 if (lock->l_blocking_ast != NULL)
1747 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1750 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1754 LDLM_DEBUG(lock, "client blocking callback handler END");
1755 LDLM_LOCK_RELEASE(lock);
1760 * Callback handler for receiving incoming completion ASTs.
1762 * This only can happen on client side.
1764 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1765 struct ldlm_namespace *ns,
1766 struct ldlm_request *dlm_req,
1767 struct ldlm_lock *lock)
1769 struct list_head ast_list;
1774 LDLM_DEBUG(lock, "client completion callback handler START");
1776 INIT_LIST_HEAD(&ast_list);
1777 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1778 int to = cfs_time_seconds(1);
1780 set_current_state(TASK_INTERRUPTIBLE);
1781 schedule_timeout(to);
1782 if (lock->l_granted_mode == lock->l_req_mode ||
1783 ldlm_is_destroyed(lock))
1788 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1790 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1791 GOTO(out, rc = lvb_len);
1792 } else if (lvb_len > 0) {
1793 if (lock->l_lvb_len > 0) {
1794 /* for extent lock, lvb contains ost_lvb{}. */
1795 LASSERT(lock->l_lvb_data != NULL);
1797 if (unlikely(lock->l_lvb_len < lvb_len)) {
1798 LDLM_ERROR(lock, "Replied LVB is larger than "
1799 "expectation, expected = %d, "
1801 lock->l_lvb_len, lvb_len);
1802 GOTO(out, rc = -EINVAL);
1807 lock_res_and_lock(lock);
1808 if (ldlm_is_destroyed(lock) ||
1809 lock->l_granted_mode == lock->l_req_mode) {
1810 /* bug 11300: the lock has already been granted */
1811 unlock_res_and_lock(lock);
1812 LDLM_DEBUG(lock, "Double grant race happened");
1816 /* If we receive the completion AST before the actual enqueue returned,
1817 * then we might need to switch lock modes, resources, or extents. */
1818 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1819 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1820 LDLM_DEBUG(lock, "completion AST, new lock mode");
1823 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1824 ldlm_convert_policy_to_local(req->rq_export,
1825 dlm_req->lock_desc.l_resource.lr_type,
1826 &dlm_req->lock_desc.l_policy_data,
1827 &lock->l_policy_data);
1828 LDLM_DEBUG(lock, "completion AST, new policy data");
1831 ldlm_resource_unlink_lock(lock);
1832 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1833 &lock->l_resource->lr_name,
1834 sizeof(lock->l_resource->lr_name)) != 0) {
1835 unlock_res_and_lock(lock);
1836 rc = ldlm_lock_change_resource(ns, lock,
1837 &dlm_req->lock_desc.l_resource.lr_name);
1839 LDLM_ERROR(lock, "Failed to allocate resource");
1842 LDLM_DEBUG(lock, "completion AST, new resource");
1843 CERROR("change resource!\n");
1844 lock_res_and_lock(lock);
1847 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1848 /* BL_AST locks are not needed in LRU.
1849 * Let ldlm_cancel_lru() be fast. */
1850 ldlm_lock_remove_from_lru(lock);
1851 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1852 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1855 if (lock->l_lvb_len > 0) {
1856 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1857 lock->l_lvb_data, lvb_len);
1859 unlock_res_and_lock(lock);
1864 ldlm_grant_lock(lock, &ast_list);
1865 unlock_res_and_lock(lock);
1867 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1869 /* Let Enqueue to call osc_lock_upcall() and initialize
1871 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1873 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1875 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1881 lock_res_and_lock(lock);
1882 ldlm_set_failed(lock);
1883 unlock_res_and_lock(lock);
1884 wake_up(&lock->l_waitq);
1886 LDLM_LOCK_RELEASE(lock);
1890 * Callback handler for receiving incoming glimpse ASTs.
1892 * This only can happen on client side. After handling the glimpse AST
1893 * we also consider dropping the lock here if it is unused locally for a
1896 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1897 struct ldlm_namespace *ns,
1898 struct ldlm_request *dlm_req,
1899 struct ldlm_lock *lock)
1904 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1906 if (lock->l_glimpse_ast != NULL)
1907 rc = lock->l_glimpse_ast(lock, req);
1909 if (req->rq_repmsg != NULL) {
1912 req->rq_status = rc;
1916 lock_res_and_lock(lock);
1917 if (lock->l_granted_mode == LCK_PW &&
1918 !lock->l_readers && !lock->l_writers &&
1919 ktime_after(ktime_get(),
1920 ktime_add(lock->l_last_used,
1921 ktime_set(10, 0)))) {
1922 unlock_res_and_lock(lock);
1923 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1924 ldlm_handle_bl_callback(ns, NULL, lock);
1929 unlock_res_and_lock(lock);
1930 LDLM_LOCK_RELEASE(lock);
1934 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1936 if (req->rq_no_reply)
1939 req->rq_status = rc;
1940 if (!req->rq_packed_final) {
1941 rc = lustre_pack_reply(req, 1, NULL, NULL);
1945 return ptlrpc_reply(req);
1948 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
1949 enum ldlm_cancel_flags cancel_flags)
1951 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1954 spin_lock(&blp->blp_lock);
1955 if (blwi->blwi_lock &&
1956 ldlm_is_discard_data(blwi->blwi_lock)) {
1957 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1958 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1960 /* other blocking callbacks are added to the regular list */
1961 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1963 spin_unlock(&blp->blp_lock);
1965 wake_up(&blp->blp_waitq);
1967 /* can not check blwi->blwi_flags as blwi could be already freed in
1969 if (!(cancel_flags & LCF_ASYNC))
1970 wait_for_completion(&blwi->blwi_comp);
1975 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1976 struct ldlm_namespace *ns,
1977 struct ldlm_lock_desc *ld,
1978 struct list_head *cancels, int count,
1979 struct ldlm_lock *lock,
1980 enum ldlm_cancel_flags cancel_flags)
1982 init_completion(&blwi->blwi_comp);
1983 INIT_LIST_HEAD(&blwi->blwi_head);
1985 if (memory_pressure_get())
1986 blwi->blwi_mem_pressure = 1;
1989 blwi->blwi_flags = cancel_flags;
1991 blwi->blwi_ld = *ld;
1993 list_add(&blwi->blwi_head, cancels);
1994 list_del_init(cancels);
1995 blwi->blwi_count = count;
1997 blwi->blwi_lock = lock;
2002 * Queues a list of locks \a cancels containing \a count locks
2003 * for later processing by a blocking thread. If \a count is zero,
2004 * then the lock referenced as \a lock is queued instead.
2006 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2007 * If list addition fails an error is returned and caller is supposed to
2008 * call ->l_blocking_ast itself.
2010 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2011 struct ldlm_lock_desc *ld,
2012 struct ldlm_lock *lock,
2013 struct list_head *cancels, int count,
2014 enum ldlm_cancel_flags cancel_flags)
2018 if (cancels && count == 0)
2021 if (cancel_flags & LCF_ASYNC) {
2022 struct ldlm_bl_work_item *blwi;
2024 OBD_ALLOC(blwi, sizeof(*blwi));
2027 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2029 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2031 /* if it is synchronous call do minimum mem alloc, as it could
2032 * be triggered from kernel shrinker
2034 struct ldlm_bl_work_item blwi;
2036 memset(&blwi, 0, sizeof(blwi));
2037 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2038 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2043 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2044 struct ldlm_lock *lock)
2046 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2049 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2050 struct list_head *cancels, int count,
2051 enum ldlm_cancel_flags cancel_flags)
2053 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2056 int ldlm_bl_thread_wakeup(void)
2058 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2062 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2063 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2065 struct obd_device *obd = req->rq_export->exp_obd;
2072 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2074 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2076 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2078 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2081 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2083 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2085 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2088 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2091 /* We are responsible for swabbing contents of val */
2093 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2094 /* Pass it on to mdc (the "export" in this case) */
2095 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2097 sizeof(KEY_HSM_COPYTOOL_SEND),
2098 KEY_HSM_COPYTOOL_SEND,
2101 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2106 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2107 const char *msg, int rc,
2108 const struct lustre_handle *handle)
2110 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2111 "%s: [nid %s] [rc %d] [lock %#llx]",
2112 msg, libcfs_id2str(req->rq_peer), rc,
2113 handle ? handle->cookie : 0);
2114 if (req->rq_no_reply)
2115 CWARN("No reply was sent, maybe cause bug 21636.\n");
2117 CWARN("Send reply failed, maybe cause bug 21636.\n");
2120 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2121 static int ldlm_callback_handler(struct ptlrpc_request *req)
2123 struct ldlm_namespace *ns;
2124 struct ldlm_request *dlm_req;
2125 struct ldlm_lock *lock;
2129 /* Requests arrive in sender's byte order. The ptlrpc service
2130 * handler has already checked and, if necessary, byte-swapped the
2131 * incoming request message body, but I am responsible for the
2132 * message buffers. */
2134 /* do nothing for sec context finalize */
2135 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2138 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2140 if (req->rq_export == NULL) {
2141 rc = ldlm_callback_reply(req, -ENOTCONN);
2142 ldlm_callback_errmsg(req, "Operate on unconnected server",
2147 LASSERT(req->rq_export != NULL);
2148 LASSERT(req->rq_export->exp_obd != NULL);
2150 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2151 case LDLM_BL_CALLBACK:
2152 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2154 ldlm_callback_reply(req, -(int)cfs_fail_err);
2158 case LDLM_CP_CALLBACK:
2159 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2162 case LDLM_GL_CALLBACK:
2163 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2167 rc = ldlm_handle_setinfo(req);
2168 ldlm_callback_reply(req, rc);
2170 case LLOG_ORIGIN_HANDLE_CREATE:
2171 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
2172 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2174 rc = llog_origin_handle_open(req);
2175 ldlm_callback_reply(req, rc);
2177 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2178 req_capsule_set(&req->rq_pill,
2179 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
2180 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2182 rc = llog_origin_handle_next_block(req);
2183 ldlm_callback_reply(req, rc);
2185 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2186 req_capsule_set(&req->rq_pill,
2187 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
2188 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2190 rc = llog_origin_handle_read_header(req);
2191 ldlm_callback_reply(req, rc);
2193 case LLOG_ORIGIN_HANDLE_CLOSE:
2194 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2196 rc = llog_origin_handle_close(req);
2197 ldlm_callback_reply(req, rc);
2200 CERROR("unknown opcode %u\n",
2201 lustre_msg_get_opc(req->rq_reqmsg));
2202 ldlm_callback_reply(req, -EPROTO);
2206 ns = req->rq_export->exp_obd->obd_namespace;
2207 LASSERT(ns != NULL);
2209 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2211 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2212 if (dlm_req == NULL) {
2213 rc = ldlm_callback_reply(req, -EPROTO);
2214 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2219 /* Force a known safe race, send a cancel to the server for a lock
2220 * which the server has already started a blocking callback on. */
2221 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2222 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2223 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2225 CERROR("ldlm_cli_cancel: %d\n", rc);
2228 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2230 CDEBUG(D_DLMTRACE, "callback on lock %#llx - lock "
2231 "disappeared\n", dlm_req->lock_handle[0].cookie);
2232 rc = ldlm_callback_reply(req, -EINVAL);
2233 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2234 &dlm_req->lock_handle[0]);
2238 if (ldlm_is_fail_loc(lock) &&
2239 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2240 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2242 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2243 lock_res_and_lock(lock);
2244 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2246 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2247 /* If somebody cancels lock and cache is already dropped,
2248 * or lock is failed before cp_ast received on client,
2249 * we can tell the server we have no lock. Otherwise, we
2250 * should send cancel after dropping the cache. */
2251 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2252 ldlm_is_failed(lock)) {
2253 LDLM_DEBUG(lock, "callback on lock %llx - lock disappeared",
2254 dlm_req->lock_handle[0].cookie);
2255 unlock_res_and_lock(lock);
2256 LDLM_LOCK_RELEASE(lock);
2257 rc = ldlm_callback_reply(req, -EINVAL);
2258 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2259 &dlm_req->lock_handle[0]);
2262 /* BL_AST locks are not needed in LRU.
2263 * Let ldlm_cancel_lru() be fast. */
2264 ldlm_lock_remove_from_lru(lock);
2265 ldlm_set_bl_ast(lock);
2267 unlock_res_and_lock(lock);
2269 /* We want the ost thread to get this reply so that it can respond
2270 * to ost requests (write cache writeback) that might be triggered
2273 * But we'd also like to be able to indicate in the reply that we're
2274 * cancelling right now, because it's unused, or have an intent result
2275 * in the reply, so we might have to push the responsibility for sending
2276 * the reply down into the AST handlers, alas. */
2278 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2279 case LDLM_BL_CALLBACK:
2280 CDEBUG(D_INODE, "blocking ast\n");
2281 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2282 if (!ldlm_is_cancel_on_block(lock)) {
2283 rc = ldlm_callback_reply(req, 0);
2284 if (req->rq_no_reply || rc)
2285 ldlm_callback_errmsg(req, "Normal process", rc,
2286 &dlm_req->lock_handle[0]);
2288 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2289 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2291 case LDLM_CP_CALLBACK:
2292 CDEBUG(D_INODE, "completion ast\n");
2293 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2294 ldlm_callback_reply(req, 0);
2295 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2297 case LDLM_GL_CALLBACK:
2298 CDEBUG(D_INODE, "glimpse ast\n");
2299 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2300 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2303 LBUG(); /* checked above */
2309 #ifdef HAVE_SERVER_SUPPORT
2311 * Main handler for canceld thread.
2313 * Separated into its own thread to avoid deadlocks.
2315 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2320 /* Requests arrive in sender's byte order. The ptlrpc service
2321 * handler has already checked and, if necessary, byte-swapped the
2322 * incoming request message body, but I am responsible for the
2323 * message buffers. */
2325 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2327 if (req->rq_export == NULL) {
2328 struct ldlm_request *dlm_req;
2330 CERROR("%s from %s arrived at %lu with bad export cookie "
2332 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2333 libcfs_nid2str(req->rq_peer.nid),
2334 req->rq_arrival_time.tv_sec,
2335 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2337 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2338 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2339 dlm_req = req_capsule_client_get(&req->rq_pill,
2341 if (dlm_req != NULL)
2342 ldlm_lock_dump_handle(D_ERROR,
2343 &dlm_req->lock_handle[0]);
2345 ldlm_callback_reply(req, -ENOTCONN);
2349 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2351 /* XXX FIXME move this back to mds/handler.c, bug 249 */
2353 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2354 CDEBUG(D_INODE, "cancel\n");
2355 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2356 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2357 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2359 rc = ldlm_handle_cancel(req);
2364 CERROR("invalid opcode %d\n",
2365 lustre_msg_get_opc(req->rq_reqmsg));
2366 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2367 ldlm_callback_reply(req, -EINVAL);
2373 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2374 struct ldlm_lock *lock)
2376 struct ldlm_request *dlm_req;
2377 struct lustre_handle lockh;
2382 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2383 if (dlm_req == NULL)
2386 ldlm_lock2handle(lock, &lockh);
2387 for (i = 0; i < dlm_req->lock_count; i++) {
2388 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2390 DEBUG_REQ(D_RPCTRACE, req,
2391 "Prio raised by lock %#llx.", lockh.cookie);
2402 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2404 struct ldlm_request *dlm_req;
2409 /* no prolong in recovery */
2410 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2413 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2414 if (dlm_req == NULL)
2417 for (i = 0; i < dlm_req->lock_count; i++) {
2418 struct ldlm_lock *lock;
2420 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2424 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2426 LDLM_DEBUG(lock, "hpreq cancel lock");
2427 LDLM_LOCK_PUT(lock);
2436 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2437 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2438 .hpreq_check = ldlm_cancel_hpreq_check,
2442 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2446 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2448 if (req->rq_export == NULL)
2451 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2452 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2453 req->rq_ops = &ldlm_cancel_hpreq_ops;
2458 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2459 struct hlist_node *hnode, void *data)
2462 struct list_head *rpc_list = data;
2463 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2465 lock_res_and_lock(lock);
2467 if (lock->l_req_mode != lock->l_granted_mode) {
2468 unlock_res_and_lock(lock);
2472 LASSERT(lock->l_resource);
2473 if (lock->l_resource->lr_type != LDLM_IBITS &&
2474 lock->l_resource->lr_type != LDLM_PLAIN) {
2475 unlock_res_and_lock(lock);
2479 if (ldlm_is_ast_sent(lock)) {
2480 unlock_res_and_lock(lock);
2484 LASSERT(lock->l_blocking_ast);
2485 LASSERT(!lock->l_blocking_lock);
2487 ldlm_set_ast_sent(lock);
2488 if (lock->l_export && lock->l_export->exp_lock_hash) {
2489 /* NB: it's safe to call cfs_hash_del() even lock isn't
2490 * in exp_lock_hash. */
2491 /* In the function below, .hs_keycmp resolves to
2492 * ldlm_export_lock_keycmp() */
2493 /* coverity[overrun-buffer-val] */
2494 cfs_hash_del(lock->l_export->exp_lock_hash,
2495 &lock->l_remote_handle, &lock->l_exp_hash);
2498 list_add_tail(&lock->l_rk_ast, rpc_list);
2499 LDLM_LOCK_GET(lock);
2501 unlock_res_and_lock(lock);
2505 void ldlm_revoke_export_locks(struct obd_export *exp)
2507 struct list_head rpc_list;
2510 INIT_LIST_HEAD(&rpc_list);
2511 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2512 ldlm_revoke_lock_cb, &rpc_list, 0);
2513 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2514 LDLM_WORK_REVOKE_AST);
2518 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2519 #endif /* HAVE_SERVER_SUPPORT */
2521 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2522 struct ldlm_bl_work_item **p_blwi,
2523 struct obd_export **p_exp)
2525 struct ldlm_bl_work_item *blwi = NULL;
2526 static unsigned int num_bl = 0;
2527 static unsigned int num_stale;
2528 int num_th = atomic_read(&blp->blp_num_threads);
2530 *p_exp = obd_stale_export_get();
2532 spin_lock(&blp->blp_lock);
2533 if (*p_exp != NULL) {
2534 if (num_th == 1 || ++num_stale < num_th) {
2535 spin_unlock(&blp->blp_lock);
2542 /* process a request from the blp_list at least every blp_num_threads */
2543 if (!list_empty(&blp->blp_list) &&
2544 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2545 blwi = list_entry(blp->blp_list.next,
2546 struct ldlm_bl_work_item, blwi_entry);
2548 if (!list_empty(&blp->blp_prio_list))
2549 blwi = list_entry(blp->blp_prio_list.next,
2550 struct ldlm_bl_work_item,
2554 if (++num_bl >= num_th)
2556 list_del(&blwi->blwi_entry);
2558 spin_unlock(&blp->blp_lock);
2561 if (*p_exp != NULL && *p_blwi != NULL) {
2562 obd_stale_export_put(*p_exp);
2566 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2569 /* This only contains temporary data until the thread starts */
2570 struct ldlm_bl_thread_data {
2571 struct ldlm_bl_pool *bltd_blp;
2572 struct completion bltd_comp;
2576 static int ldlm_bl_thread_main(void *arg);
2578 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2580 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2581 struct task_struct *task;
2583 init_completion(&bltd.bltd_comp);
2585 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2586 if (bltd.bltd_num >= blp->blp_max_threads) {
2587 atomic_dec(&blp->blp_num_threads);
2591 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2593 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2594 atomic_dec(&blp->blp_num_threads);
2598 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2601 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2602 bltd.bltd_num, PTR_ERR(task));
2603 atomic_dec(&blp->blp_num_threads);
2604 return PTR_ERR(task);
2606 wait_for_completion(&bltd.bltd_comp);
2611 /* Not fatal if racy and have a few too many threads */
2612 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2613 struct ldlm_bl_work_item *blwi)
2615 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2618 if (atomic_read(&blp->blp_busy_threads) <
2619 atomic_read(&blp->blp_num_threads))
2622 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2623 blwi->blwi_mem_pressure))
2629 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2630 struct ldlm_bl_work_item *blwi)
2634 if (blwi->blwi_ns == NULL)
2635 /* added by ldlm_cleanup() */
2636 RETURN(LDLM_ITER_STOP);
2638 if (blwi->blwi_mem_pressure)
2639 memory_pressure_set();
2641 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2643 if (blwi->blwi_count) {
2645 /* The special case when we cancel locks in lru
2646 * asynchronously, we pass the list of locks here.
2647 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2648 * canceled locally yet. */
2649 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2652 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2655 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2658 if (blwi->blwi_mem_pressure)
2659 memory_pressure_clr();
2661 if (blwi->blwi_flags & LCF_ASYNC)
2662 OBD_FREE(blwi, sizeof(*blwi));
2664 complete(&blwi->blwi_comp);
2670 * Cancel stale locks on export. Cancel blocked locks first.
2671 * If the given export has blocked locks, the next in the list may have
2672 * them too, thus cancel not blocked locks only if the current export has
2675 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2676 struct obd_export *exp)
2681 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2683 num = ldlm_export_cancel_blocked_locks(exp);
2685 ldlm_export_cancel_locks(exp);
2687 obd_stale_export_put(exp);
2694 * Main blocking requests processing thread.
2696 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2697 * This thread in the end ends up doing actual call to ->l_blocking_ast
2700 static int ldlm_bl_thread_main(void *arg)
2702 struct ldlm_bl_pool *blp;
2703 struct ldlm_bl_thread_data *bltd = arg;
2706 blp = bltd->bltd_blp;
2708 complete(&bltd->bltd_comp);
2709 /* cannot use bltd after this, it is only on caller's stack */
2712 struct l_wait_info lwi = { 0 };
2713 struct ldlm_bl_work_item *blwi = NULL;
2714 struct obd_export *exp = NULL;
2717 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2720 l_wait_event_exclusive(blp->blp_waitq,
2721 ldlm_bl_get_work(blp, &blwi,
2724 atomic_inc(&blp->blp_busy_threads);
2726 if (ldlm_bl_thread_need_create(blp, blwi))
2727 /* discard the return value, we tried */
2728 ldlm_bl_thread_start(blp, true);
2731 rc = ldlm_bl_thread_exports(blp, exp);
2733 rc = ldlm_bl_thread_blwi(blp, blwi);
2735 atomic_dec(&blp->blp_busy_threads);
2737 if (rc == LDLM_ITER_STOP)
2740 /* If there are many namespaces, we will not sleep waiting for
2741 * work, and must do a cond_resched to avoid holding the CPU
2746 atomic_dec(&blp->blp_num_threads);
2747 complete(&blp->blp_comp);
2752 static int ldlm_setup(void);
2753 static int ldlm_cleanup(void);
2755 int ldlm_get_ref(void)
2759 mutex_lock(&ldlm_ref_mutex);
2760 if (++ldlm_refcount == 1) {
2765 mutex_unlock(&ldlm_ref_mutex);
2770 void ldlm_put_ref(void)
2773 mutex_lock(&ldlm_ref_mutex);
2774 if (ldlm_refcount == 1) {
2775 int rc = ldlm_cleanup();
2777 CERROR("ldlm_cleanup failed: %d\n", rc);
2783 mutex_unlock(&ldlm_ref_mutex);
2789 * Export handle<->lock hash operations.
2792 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
2794 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2798 ldlm_export_lock_key(struct hlist_node *hnode)
2800 struct ldlm_lock *lock;
2802 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2803 return &lock->l_remote_handle;
2807 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
2809 struct ldlm_lock *lock;
2811 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2812 lock->l_remote_handle = *(struct lustre_handle *)key;
2816 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
2818 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2822 ldlm_export_lock_object(struct hlist_node *hnode)
2824 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2828 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
2830 struct ldlm_lock *lock;
2832 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2833 LDLM_LOCK_GET(lock);
2837 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
2839 struct ldlm_lock *lock;
2841 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2842 LDLM_LOCK_RELEASE(lock);
2845 static struct cfs_hash_ops ldlm_export_lock_ops = {
2846 .hs_hash = ldlm_export_lock_hash,
2847 .hs_key = ldlm_export_lock_key,
2848 .hs_keycmp = ldlm_export_lock_keycmp,
2849 .hs_keycpy = ldlm_export_lock_keycpy,
2850 .hs_object = ldlm_export_lock_object,
2851 .hs_get = ldlm_export_lock_get,
2852 .hs_put = ldlm_export_lock_put,
2853 .hs_put_locked = ldlm_export_lock_put,
2856 int ldlm_init_export(struct obd_export *exp)
2861 exp->exp_lock_hash =
2862 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2863 HASH_EXP_LOCK_CUR_BITS,
2864 HASH_EXP_LOCK_MAX_BITS,
2865 HASH_EXP_LOCK_BKT_BITS, 0,
2866 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2867 &ldlm_export_lock_ops,
2868 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2869 CFS_HASH_NBLK_CHANGE);
2871 if (!exp->exp_lock_hash)
2874 rc = ldlm_init_flock_export(exp);
2880 ldlm_destroy_export(exp);
2883 EXPORT_SYMBOL(ldlm_init_export);
2885 void ldlm_destroy_export(struct obd_export *exp)
2888 cfs_hash_putref(exp->exp_lock_hash);
2889 exp->exp_lock_hash = NULL;
2891 ldlm_destroy_flock_export(exp);
2894 EXPORT_SYMBOL(ldlm_destroy_export);
2896 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
2897 struct attribute *attr,
2900 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
2903 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
2904 struct attribute *attr,
2911 rc = kstrtoul(buffer, 10, &val);
2915 ldlm_cancel_unused_locks_before_replay = val;
2919 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
2921 static struct attribute *ldlm_attrs[] = {
2922 &lustre_attr_cancel_unused_locks_before_replay.attr,
2926 static struct attribute_group ldlm_attr_group = {
2927 .attrs = ldlm_attrs,
2930 static int ldlm_setup(void)
2932 static struct ptlrpc_service_conf conf;
2933 struct ldlm_bl_pool *blp = NULL;
2934 #ifdef HAVE_SERVER_SUPPORT
2935 struct task_struct *task;
2936 #endif /* HAVE_SERVER_SUPPORT */
2942 if (ldlm_state != NULL)
2945 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2946 if (ldlm_state == NULL)
2949 ldlm_kobj = kobject_create_and_add("ldlm", lustre_kobj);
2953 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
2957 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
2961 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
2965 #ifdef CONFIG_PROC_FS
2966 rc = ldlm_proc_setup();
2969 #endif /* CONFIG_PROC_FS */
2971 memset(&conf, 0, sizeof(conf));
2972 conf = (typeof(conf)) {
2973 .psc_name = "ldlm_cbd",
2974 .psc_watchdog_factor = 2,
2976 .bc_nbufs = LDLM_CLIENT_NBUFS,
2977 .bc_buf_size = LDLM_BUFSIZE,
2978 .bc_req_max_size = LDLM_MAXREQSIZE,
2979 .bc_rep_max_size = LDLM_MAXREPSIZE,
2980 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
2981 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
2984 .tc_thr_name = "ldlm_cb",
2985 .tc_thr_factor = LDLM_THR_FACTOR,
2986 .tc_nthrs_init = LDLM_NTHRS_INIT,
2987 .tc_nthrs_base = LDLM_NTHRS_BASE,
2988 .tc_nthrs_max = LDLM_NTHRS_MAX,
2989 .tc_nthrs_user = ldlm_num_threads,
2990 .tc_cpu_affinity = 1,
2991 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
2994 .cc_pattern = ldlm_cpts,
2997 .so_req_handler = ldlm_callback_handler,
3000 ldlm_state->ldlm_cb_service = \
3001 ptlrpc_register_service(&conf, ldlm_svc_kset,
3003 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3004 CERROR("failed to start service\n");
3005 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3006 ldlm_state->ldlm_cb_service = NULL;
3010 #ifdef HAVE_SERVER_SUPPORT
3011 memset(&conf, 0, sizeof(conf));
3012 conf = (typeof(conf)) {
3013 .psc_name = "ldlm_canceld",
3014 .psc_watchdog_factor = 6,
3016 .bc_nbufs = LDLM_SERVER_NBUFS,
3017 .bc_buf_size = LDLM_BUFSIZE,
3018 .bc_req_max_size = LDLM_MAXREQSIZE,
3019 .bc_rep_max_size = LDLM_MAXREPSIZE,
3020 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3021 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3025 .tc_thr_name = "ldlm_cn",
3026 .tc_thr_factor = LDLM_THR_FACTOR,
3027 .tc_nthrs_init = LDLM_NTHRS_INIT,
3028 .tc_nthrs_base = LDLM_NTHRS_BASE,
3029 .tc_nthrs_max = LDLM_NTHRS_MAX,
3030 .tc_nthrs_user = ldlm_num_threads,
3031 .tc_cpu_affinity = 1,
3032 .tc_ctx_tags = LCT_MD_THREAD | \
3037 .cc_pattern = ldlm_cpts,
3040 .so_req_handler = ldlm_cancel_handler,
3041 .so_hpreq_handler = ldlm_hpreq_handler,
3044 ldlm_state->ldlm_cancel_service = \
3045 ptlrpc_register_service(&conf, ldlm_svc_kset,
3047 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3048 CERROR("failed to start service\n");
3049 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3050 ldlm_state->ldlm_cancel_service = NULL;
3053 #endif /* HAVE_SERVER_SUPPORT */
3055 OBD_ALLOC(blp, sizeof(*blp));
3057 GOTO(out, rc = -ENOMEM);
3058 ldlm_state->ldlm_bl_pool = blp;
3060 spin_lock_init(&blp->blp_lock);
3061 INIT_LIST_HEAD(&blp->blp_list);
3062 INIT_LIST_HEAD(&blp->blp_prio_list);
3063 init_waitqueue_head(&blp->blp_waitq);
3064 atomic_set(&blp->blp_num_threads, 0);
3065 atomic_set(&blp->blp_busy_threads, 0);
3067 if (ldlm_num_threads == 0) {
3068 blp->blp_min_threads = LDLM_NTHRS_INIT;
3069 blp->blp_max_threads = LDLM_NTHRS_MAX;
3071 blp->blp_min_threads = blp->blp_max_threads = \
3072 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3076 for (i = 0; i < blp->blp_min_threads; i++) {
3077 rc = ldlm_bl_thread_start(blp, false);
3082 #ifdef HAVE_SERVER_SUPPORT
3083 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3086 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3090 wait_event(expired_lock_wait_queue,
3091 expired_lock_thread_state == ELT_READY);
3092 #endif /* HAVE_SERVER_SUPPORT */
3094 rc = ldlm_pools_init();
3096 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3100 rc = ldlm_reclaim_setup();
3102 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3112 static int ldlm_cleanup(void)
3116 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3117 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3118 CERROR("ldlm still has namespaces; clean these up first.\n");
3119 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3120 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3124 ldlm_reclaim_cleanup();
3127 if (ldlm_state->ldlm_bl_pool != NULL) {
3128 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3130 while (atomic_read(&blp->blp_num_threads) > 0) {
3131 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3133 init_completion(&blp->blp_comp);
3135 spin_lock(&blp->blp_lock);
3136 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3137 wake_up(&blp->blp_waitq);
3138 spin_unlock(&blp->blp_lock);
3140 wait_for_completion(&blp->blp_comp);
3143 OBD_FREE(blp, sizeof(*blp));
3146 if (ldlm_state->ldlm_cb_service != NULL)
3147 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3148 #ifdef HAVE_SERVER_SUPPORT
3149 if (ldlm_state->ldlm_cancel_service != NULL)
3150 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3154 kset_unregister(ldlm_ns_kset);
3156 kset_unregister(ldlm_svc_kset);
3158 kobject_put(ldlm_kobj);
3160 ldlm_proc_cleanup();
3162 #ifdef HAVE_SERVER_SUPPORT
3163 if (expired_lock_thread_state != ELT_STOPPED) {
3164 expired_lock_thread_state = ELT_TERMINATE;
3165 wake_up(&expired_lock_wait_queue);
3166 wait_event(expired_lock_wait_queue,
3167 expired_lock_thread_state == ELT_STOPPED);
3171 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3179 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3180 sizeof(struct ldlm_resource), 0,
3181 SLAB_HWCACHE_ALIGN, NULL);
3182 if (ldlm_resource_slab == NULL)
3185 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3186 sizeof(struct ldlm_lock), 0,
3187 SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU, NULL);
3188 if (ldlm_lock_slab == NULL)
3191 ldlm_interval_slab = kmem_cache_create("interval_node",
3192 sizeof(struct ldlm_interval),
3193 0, SLAB_HWCACHE_ALIGN, NULL);
3194 if (ldlm_interval_slab == NULL)
3197 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3198 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3199 0, SLAB_HWCACHE_ALIGN, NULL);
3200 if (ldlm_interval_tree_slab == NULL)
3203 #ifdef HAVE_SERVER_SUPPORT
3204 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3205 sizeof(struct ldlm_glimpse_work),
3207 if (ldlm_glimpse_work_kmem == NULL)
3208 goto out_interval_tree;
3211 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3212 class_export_dump_hook = ldlm_dump_export_locks;
3215 #ifdef HAVE_SERVER_SUPPORT
3217 kmem_cache_destroy(ldlm_interval_tree_slab);
3220 kmem_cache_destroy(ldlm_interval_slab);
3222 kmem_cache_destroy(ldlm_lock_slab);
3224 kmem_cache_destroy(ldlm_resource_slab);
3229 void ldlm_exit(void)
3232 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3233 kmem_cache_destroy(ldlm_resource_slab);
3234 /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3235 * synchronize_rcu() to wait a grace period elapsed, so that
3236 * ldlm_lock_free() get a chance to be called. */
3238 kmem_cache_destroy(ldlm_lock_slab);
3239 kmem_cache_destroy(ldlm_interval_slab);
3240 kmem_cache_destroy(ldlm_interval_tree_slab);
3241 #ifdef HAVE_SERVER_SUPPORT
3242 kmem_cache_destroy(ldlm_glimpse_work_kmem);