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, 2017, 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 unsigned int ldlm_cpu_bind = 1;
53 module_param(ldlm_cpu_bind, uint, 0444);
54 MODULE_PARM_DESC(ldlm_cpu_bind,
55 "bind DLM service threads to particular CPU partitions");
57 static char *ldlm_cpts;
58 module_param(ldlm_cpts, charp, 0444);
59 MODULE_PARM_DESC(ldlm_cpts, "CPU partitions ldlm threads should run on");
61 static DEFINE_MUTEX(ldlm_ref_mutex);
62 static int ldlm_refcount;
64 struct kobject *ldlm_kobj;
65 struct kset *ldlm_ns_kset;
66 struct kset *ldlm_svc_kset;
70 static struct ldlm_state *ldlm_state;
73 * timeout for initial callback (AST) reply (bz10399)
74 * Due to having to send a 32 bit time value over the
75 * wire return it as time_t instead of time64_t
77 static inline time_t ldlm_get_rq_timeout(void)
80 time_t timeout = min(ldlm_timeout, obd_timeout / 3);
82 return timeout < 1 ? 1 : timeout;
89 * blp_prio_list is used for callbacks that should be handled
90 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
93 struct list_head blp_prio_list;
96 * blp_list is used for all other callbacks which are likely
97 * to take longer to process.
99 struct list_head blp_list;
101 wait_queue_head_t blp_waitq;
102 struct completion blp_comp;
103 atomic_t blp_num_threads;
104 atomic_t blp_busy_threads;
109 struct ldlm_bl_work_item {
110 struct list_head blwi_entry;
111 struct ldlm_namespace *blwi_ns;
112 struct ldlm_lock_desc blwi_ld;
113 struct ldlm_lock *blwi_lock;
114 struct list_head blwi_head;
116 struct completion blwi_comp;
117 enum ldlm_cancel_flags blwi_flags;
118 int blwi_mem_pressure;
121 #ifdef HAVE_SERVER_SUPPORT
124 * Protects both waiting_locks_list and expired_lock_thread.
126 static DEFINE_SPINLOCK(waiting_locks_spinlock); /* BH lock (timer) */
129 * List for contended locks.
131 * As soon as a lock is contended, it gets placed on this list and
132 * expected time to get a response is filled in the lock. A special
133 * thread walks the list looking for locks that should be released and
134 * schedules client evictions for those that have not been released in
137 * All access to it should be under waiting_locks_spinlock.
139 static LIST_HEAD(waiting_locks_list);
140 static void waiting_locks_callback(TIMER_DATA_TYPE unused);
141 static CFS_DEFINE_TIMER(waiting_locks_timer, waiting_locks_callback, 0, 0);
149 static DECLARE_WAIT_QUEUE_HEAD(expired_lock_wait_queue);
150 static enum elt_state expired_lock_thread_state = ELT_STOPPED;
151 static int expired_lock_dump;
152 static LIST_HEAD(expired_lock_list);
154 static int ldlm_lock_busy(struct ldlm_lock *lock);
155 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout);
156 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds);
158 static inline int have_expired_locks(void)
163 spin_lock_bh(&waiting_locks_spinlock);
164 need_to_run = !list_empty(&expired_lock_list);
165 spin_unlock_bh(&waiting_locks_spinlock);
171 * Check expired lock list for expired locks and time them out.
173 static int expired_lock_main(void *arg)
175 struct list_head *expired = &expired_lock_list;
180 expired_lock_thread_state = ELT_READY;
181 wake_up(&expired_lock_wait_queue);
184 wait_event_idle(expired_lock_wait_queue,
185 have_expired_locks() ||
186 expired_lock_thread_state == ELT_TERMINATE);
188 spin_lock_bh(&waiting_locks_spinlock);
189 if (expired_lock_dump) {
190 spin_unlock_bh(&waiting_locks_spinlock);
192 /* from waiting_locks_callback, but not in timer */
193 libcfs_debug_dumplog();
195 spin_lock_bh(&waiting_locks_spinlock);
196 expired_lock_dump = 0;
201 while (!list_empty(expired)) {
202 struct obd_export *export;
203 struct ldlm_lock *lock;
205 lock = list_entry(expired->next, struct ldlm_lock,
207 if ((void *)lock < LP_POISON + PAGE_SIZE &&
208 (void *)lock >= LP_POISON) {
209 spin_unlock_bh(&waiting_locks_spinlock);
210 CERROR("free lock on elt list %p\n", lock);
213 list_del_init(&lock->l_pending_chain);
214 if ((void *)lock->l_export <
215 LP_POISON + PAGE_SIZE &&
216 (void *)lock->l_export >= LP_POISON) {
217 CERROR("lock with free export on elt list %p\n",
219 lock->l_export = NULL;
220 LDLM_ERROR(lock, "free export");
222 * release extra ref grabbed by
223 * ldlm_add_waiting_lock() or
226 LDLM_LOCK_RELEASE(lock);
230 if (ldlm_is_destroyed(lock)) {
232 * release the lock refcount where
233 * waiting_locks_callback() founds
235 LDLM_LOCK_RELEASE(lock);
238 export = class_export_lock_get(lock->l_export, lock);
239 spin_unlock_bh(&waiting_locks_spinlock);
241 /* Check if we need to prolong timeout */
242 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
243 lock->l_callback_timeout != 0 && /* not AST error */
244 ldlm_lock_busy(lock)) {
245 LDLM_DEBUG(lock, "prolong the busy lock");
246 lock_res_and_lock(lock);
247 ldlm_add_waiting_lock(lock,
248 ldlm_bl_timeout(lock) >> 1);
249 unlock_res_and_lock(lock);
251 spin_lock_bh(&export->exp_bl_list_lock);
252 list_del_init(&lock->l_exp_list);
253 spin_unlock_bh(&export->exp_bl_list_lock);
256 "lock callback timer expired after %llds: evicting client at %s ",
257 ktime_get_real_seconds() -
259 obd_export_nid2str(export));
260 ldlm_lock_to_ns(lock)->ns_timeouts++;
262 class_fail_export(export);
264 class_export_lock_put(export, lock);
266 * release extra ref grabbed by ldlm_add_waiting_lock()
267 * or ldlm_failed_ast()
269 LDLM_LOCK_RELEASE(lock);
271 spin_lock_bh(&waiting_locks_spinlock);
273 spin_unlock_bh(&waiting_locks_spinlock);
275 if (do_dump && obd_dump_on_eviction) {
276 CERROR("dump the log upon eviction\n");
277 libcfs_debug_dumplog();
280 if (expired_lock_thread_state == ELT_TERMINATE)
284 expired_lock_thread_state = ELT_STOPPED;
285 wake_up(&expired_lock_wait_queue);
290 * Check if there is a request in the export request list
291 * which prevents the lock canceling.
293 static int ldlm_lock_busy(struct ldlm_lock *lock)
295 struct ptlrpc_request *req;
300 if (lock->l_export == NULL)
303 spin_lock(&lock->l_export->exp_rpc_lock);
304 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
306 if (req->rq_ops->hpreq_lock_match) {
307 match = req->rq_ops->hpreq_lock_match(req, lock);
312 spin_unlock(&lock->l_export->exp_rpc_lock);
316 /* This is called from within a timer interrupt and cannot schedule */
317 static void waiting_locks_callback(TIMER_DATA_TYPE unused)
319 struct ldlm_lock *lock;
322 spin_lock_bh(&waiting_locks_spinlock);
323 while (!list_empty(&waiting_locks_list)) {
324 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
326 if (lock->l_callback_timeout > ktime_get_seconds() ||
327 lock->l_req_mode == LCK_GROUP)
331 * no needs to take an extra ref on the lock since it was in
332 * the waiting_locks_list and ldlm_add_waiting_lock()
333 * already grabbed a ref
335 list_del(&lock->l_pending_chain);
336 list_add(&lock->l_pending_chain, &expired_lock_list);
340 if (!list_empty(&expired_lock_list)) {
341 if (obd_dump_on_timeout && need_dump)
342 expired_lock_dump = __LINE__;
344 wake_up(&expired_lock_wait_queue);
348 * Make sure the timer will fire again if we have any locks
351 if (!list_empty(&waiting_locks_list)) {
352 time64_t now = ktime_get_seconds();
355 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
357 if (lock->l_callback_timeout - now > 0)
358 delta = lock->l_callback_timeout - now;
359 mod_timer(&waiting_locks_timer,
360 jiffies + cfs_time_seconds(delta));
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, time64_t seconds)
379 unsigned long timeout_jiffies = jiffies;
380 time64_t now = ktime_get_seconds();
384 if (!list_empty(&lock->l_pending_chain))
387 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
388 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
391 deadline = now + seconds;
392 if (likely(deadline > lock->l_callback_timeout))
393 lock->l_callback_timeout = deadline;
395 timeout = clamp_t(time_t, lock->l_callback_timeout - now,
397 timeout_jiffies += cfs_time_seconds(timeout);
399 if (time_before(timeout_jiffies, waiting_locks_timer.expires) ||
400 !timer_pending(&waiting_locks_timer))
401 mod_timer(&waiting_locks_timer, timeout_jiffies);
404 * if the new lock has a shorter timeout than something earlier on
405 * the list, we'll wait the longer amount of time; no big deal.
408 list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
412 static void ldlm_add_blocked_lock(struct ldlm_lock *lock)
414 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
415 if (list_empty(&lock->l_exp_list)) {
416 if (!ldlm_is_granted(lock))
417 list_add_tail(&lock->l_exp_list,
418 &lock->l_export->exp_bl_list);
420 list_add(&lock->l_exp_list,
421 &lock->l_export->exp_bl_list);
423 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
426 * A blocked lock is added. Adjust the position in
427 * the stale list if the export is in the list.
428 * If export is stale and not in the list - it is being
429 * processed and will be placed on the right position
430 * on obd_stale_export_put().
432 if (!list_empty(&lock->l_export->exp_stale_list))
433 obd_stale_export_adjust(lock->l_export);
436 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
440 /* NB: must be called with hold of lock_res_and_lock() */
441 LASSERT(ldlm_is_res_locked(lock));
442 LASSERT(!ldlm_is_cancel_on_block(lock));
445 * Do not put cross-MDT lock in the waiting list, since we
446 * will not evict it due to timeout for now
448 if (lock->l_export != NULL &&
449 (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS))
452 spin_lock_bh(&waiting_locks_spinlock);
453 if (ldlm_is_cancel(lock)) {
454 spin_unlock_bh(&waiting_locks_spinlock);
458 if (ldlm_is_destroyed(lock)) {
459 static time64_t next;
461 spin_unlock_bh(&waiting_locks_spinlock);
462 LDLM_ERROR(lock, "not waiting on destroyed lock (b=5653)");
463 if (ktime_get_seconds() > next) {
464 next = ktime_get_seconds() + 14400;
465 libcfs_debug_dumpstack(NULL);
470 ldlm_set_waited(lock);
471 lock->l_blast_sent = ktime_get_real_seconds();
472 ret = __ldlm_add_waiting_lock(lock, timeout);
475 * grab ref on the lock if it has been added to the
480 spin_unlock_bh(&waiting_locks_spinlock);
483 ldlm_add_blocked_lock(lock);
485 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %lld, AT: %s)",
486 ret == 0 ? "not re-" : "", timeout,
487 AT_OFF ? "off" : "on");
492 * Remove a lock from the pending list, likely because it had its cancellation
493 * callback arrive without incident. This adjusts the lock-timeout timer if
494 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
495 * As done by ldlm_del_waiting_lock(), the caller must release the lock
496 * reference when the lock is removed from any list (1 is returned).
498 * Called with namespace lock held.
500 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
502 struct list_head *list_next;
504 if (list_empty(&lock->l_pending_chain))
507 list_next = lock->l_pending_chain.next;
508 if (lock->l_pending_chain.prev == &waiting_locks_list) {
509 /* Removing the head of the list, adjust timer. */
510 if (list_next == &waiting_locks_list) {
511 /* No more, just cancel. */
512 del_timer(&waiting_locks_timer);
514 time64_t now = ktime_get_seconds();
515 struct ldlm_lock *next;
518 next = list_entry(list_next, struct ldlm_lock,
520 if (next->l_callback_timeout - now > 0)
521 delta = lock->l_callback_timeout - now;
523 mod_timer(&waiting_locks_timer,
524 jiffies + cfs_time_seconds(delta));
527 list_del_init(&lock->l_pending_chain);
532 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
536 if (lock->l_export == NULL) {
537 /* We don't have a "waiting locks list" on clients. */
538 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
542 spin_lock_bh(&waiting_locks_spinlock);
543 ret = __ldlm_del_waiting_lock(lock);
544 ldlm_clear_waited(lock);
545 spin_unlock_bh(&waiting_locks_spinlock);
547 /* remove the lock out of export blocking list */
548 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
549 list_del_init(&lock->l_exp_list);
550 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
554 * release lock ref if it has indeed been removed
557 LDLM_LOCK_RELEASE(lock);
560 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
565 * Prolong the contended lock waiting time.
567 * Called with namespace lock held.
569 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
571 if (lock->l_export == NULL) {
572 /* We don't have a "waiting locks list" on clients. */
573 LDLM_DEBUG(lock, "client lock: no-op");
577 if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) {
578 /* We don't have a "waiting locks list" on OSP. */
579 LDLM_DEBUG(lock, "MDS-MDS lock: no-op");
583 spin_lock_bh(&waiting_locks_spinlock);
585 if (list_empty(&lock->l_pending_chain)) {
586 spin_unlock_bh(&waiting_locks_spinlock);
587 LDLM_DEBUG(lock, "wasn't waiting");
592 * we remove/add the lock to the waiting list, so no needs to
593 * release/take a lock reference
595 __ldlm_del_waiting_lock(lock);
596 __ldlm_add_waiting_lock(lock, timeout);
597 spin_unlock_bh(&waiting_locks_spinlock);
599 LDLM_DEBUG(lock, "refreshed");
602 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
604 #else /* HAVE_SERVER_SUPPORT */
606 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
611 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
616 #endif /* !HAVE_SERVER_SUPPORT */
618 #ifdef HAVE_SERVER_SUPPORT
621 * Calculate the per-export Blocking timeout (covering BL AST, data flush,
622 * lock cancel, and their replies). Used for lock callback timeout and AST
625 * \param[in] lock lock which is getting the blocking callback
627 * \retval timeout in seconds to wait for the client reply
629 time64_t ldlm_bl_timeout(struct ldlm_lock *lock)
634 return obd_timeout / 2;
637 * Since these are non-updating timeouts, we should be conservative.
638 * Take more than usually, 150%
639 * It would be nice to have some kind of "early reply" mechanism for
640 * lock callbacks too...
642 timeout = at_get(&lock->l_export->exp_bl_lock_at);
643 return max(timeout + (timeout >> 1), (time64_t)ldlm_enqueue_min);
645 EXPORT_SYMBOL(ldlm_bl_timeout);
648 * Perform lock cleanup if AST sending failed.
650 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
651 const char *ast_type)
653 LCONSOLE_ERROR_MSG(0x138,
654 "%s: A client on nid %s was evicted due to a lock %s callback time out: rc %d\n",
655 lock->l_export->exp_obd->obd_name,
656 obd_export_nid2str(lock->l_export), ast_type, rc);
658 if (obd_dump_on_timeout)
659 libcfs_debug_dumplog();
660 spin_lock_bh(&waiting_locks_spinlock);
661 if (__ldlm_del_waiting_lock(lock) == 0)
663 * the lock was not in any list, grab an extra ref before adding
664 * the lock to the expired list
667 lock->l_callback_timeout = 0; /* differentiate it from expired locks */
668 list_add(&lock->l_pending_chain, &expired_lock_list);
669 wake_up(&expired_lock_wait_queue);
670 spin_unlock_bh(&waiting_locks_spinlock);
674 * Perform lock cleanup if AST reply came with error.
676 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
677 struct ptlrpc_request *req, int rc,
678 const char *ast_type)
680 struct lnet_process_id peer = req->rq_import->imp_connection->c_peer;
682 if (!req->rq_replied || (rc && rc != -EINVAL)) {
683 if (ldlm_is_cancel(lock)) {
685 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
686 ast_type, req, req->rq_xid,
687 libcfs_nid2str(peer.nid));
688 ldlm_lock_cancel(lock);
690 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
692 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
694 * Upon umount process the AST fails because cannot be
695 * sent. This shouldn't lead to the client eviction.
696 * -ENODEV error is returned by ptl_send_rpc() for
697 * new request in such import.
698 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
699 * if imp_invalid is set or obd_no_recov.
700 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
701 * in ptlrpc_import_delay_req() as well with -EIO code.
702 * In all such cases errors are ignored.
705 "%s AST can't be sent due to a server %s failure or umount process: rc = %d\n",
707 req->rq_import->imp_obd->obd_name, rc);
710 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
711 libcfs_nid2str(peer.nid),
712 req->rq_replied ? "returned error from" :
713 "failed to reply to",
714 ast_type, req, req->rq_xid,
715 (req->rq_repmsg != NULL) ?
716 lustre_msg_get_status(req->rq_repmsg) : 0,
718 ldlm_failed_ast(lock, rc, ast_type);
724 struct ldlm_resource *res = lock->l_resource;
727 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
728 libcfs_nid2str(peer.nid),
730 lustre_msg_get_status(req->rq_repmsg) : -1,
731 ast_type, req, req->rq_xid);
734 * update lvbo to return proper attributes.
737 ldlm_resource_getref(res);
738 ldlm_lvbo_update(res, lock, NULL, 1);
739 ldlm_resource_putref(res);
741 ldlm_lock_cancel(lock);
748 static int ldlm_cb_interpret(const struct lu_env *env,
749 struct ptlrpc_request *req, void *args, int rc)
751 struct ldlm_cb_async_args *ca = args;
752 struct ldlm_lock *lock = ca->ca_lock;
753 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
757 LASSERT(lock != NULL);
760 case LDLM_GL_CALLBACK:
762 * Update the LVB from disk if the AST failed
763 * (this is a legal race)
765 * - Glimpse callback of local lock just returns
766 * -ELDLM_NO_LOCK_DATA.
767 * - Glimpse callback of remote lock might return
768 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
770 if (unlikely(arg->gl_interpret_reply)) {
771 rc = arg->gl_interpret_reply(NULL, req, args, rc);
772 } else if (rc == -ELDLM_NO_LOCK_DATA) {
774 "lost race - client has a lock but no inode");
775 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
776 } else if (rc != 0) {
777 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
779 rc = ldlm_lvbo_update(lock->l_resource,
783 case LDLM_BL_CALLBACK:
785 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
787 case LDLM_CP_CALLBACK:
789 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
792 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
797 /* release extra reference taken in ldlm_ast_fini() */
798 LDLM_LOCK_RELEASE(lock);
801 atomic_inc(&arg->restart);
806 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
808 struct ldlm_cb_async_args *ca = data;
809 struct ldlm_lock *lock = ca->ca_lock;
811 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
814 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
815 struct ldlm_cb_set_arg *arg,
816 struct ldlm_lock *lock,
823 if (unlikely(instant_cancel)) {
824 rc = ptl_send_rpc(req, 1);
825 ptlrpc_req_finished(req);
827 atomic_inc(&arg->restart);
830 ptlrpc_set_add_req(arg->set, req);
837 * Check if there are requests in the export request list which prevent
838 * the lock canceling and make these requests high priority ones.
840 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
842 struct ptlrpc_request *req;
846 if (lock->l_export == NULL) {
847 LDLM_DEBUG(lock, "client lock: no-op");
851 spin_lock(&lock->l_export->exp_rpc_lock);
852 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
855 * Do not process requests that were not yet added to there
856 * incoming queue or were already removed from there for
857 * processing. We evaluate ptlrpc_nrs_req_can_move() without
858 * holding svcpt->scp_req_lock, and then redo the check with
859 * the lock held once we need to obtain a reliable result.
861 if (ptlrpc_nrs_req_can_move(req) &&
862 req->rq_ops->hpreq_lock_match &&
863 req->rq_ops->hpreq_lock_match(req, lock))
864 ptlrpc_nrs_req_hp_move(req);
866 spin_unlock(&lock->l_export->exp_rpc_lock);
871 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
872 * enqueued server lock conflicts with given one.
874 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
875 * to wait for client response.
877 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
878 struct ldlm_lock_desc *desc,
879 void *data, int flag)
881 struct ldlm_cb_async_args *ca;
882 struct ldlm_cb_set_arg *arg = data;
883 struct ldlm_request *body;
884 struct ptlrpc_request *req;
885 int instant_cancel = 0;
890 if (flag == LDLM_CB_CANCELING)
891 /* Don't need to do anything here. */
894 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
895 LDLM_DEBUG(lock, "dropping BL AST");
900 LASSERT(data != NULL);
901 if (lock->l_export->exp_obd->obd_recovering != 0)
902 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
904 ldlm_lock_reorder_req(lock);
906 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
907 &RQF_LDLM_BL_CALLBACK,
908 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
912 ca = ptlrpc_req_async_args(ca, req);
913 ca->ca_set_arg = arg;
916 req->rq_interpret_reply = ldlm_cb_interpret;
918 lock_res_and_lock(lock);
919 if (ldlm_is_destroyed(lock)) {
920 /* What's the point? */
921 unlock_res_and_lock(lock);
922 ptlrpc_req_finished(req);
926 if (!ldlm_is_granted(lock)) {
928 * this blocking AST will be communicated as part of the
929 * completion AST instead
931 ldlm_add_blocked_lock(lock);
932 ldlm_set_waited(lock);
933 unlock_res_and_lock(lock);
935 ptlrpc_req_finished(req);
936 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
940 if (ldlm_is_cancel_on_block(lock))
943 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
944 body->lock_handle[0] = lock->l_remote_handle;
945 body->lock_desc = *desc;
946 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
948 LDLM_DEBUG(lock, "server preparing blocking AST");
950 ptlrpc_request_set_replen(req);
951 ldlm_set_cbpending(lock);
952 if (instant_cancel) {
953 unlock_res_and_lock(lock);
954 ldlm_lock_cancel(lock);
956 req->rq_no_resend = 1;
958 LASSERT(ldlm_is_granted(lock));
959 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
960 unlock_res_and_lock(lock);
962 /* Do not resend after lock callback timeout */
963 req->rq_delay_limit = ldlm_bl_timeout(lock);
964 req->rq_resend_cb = ldlm_update_resend;
967 req->rq_send_state = LUSTRE_IMP_FULL;
968 /* ptlrpc_request_alloc_pack already set timeout */
970 req->rq_timeout = ldlm_get_rq_timeout();
972 if (lock->l_export && lock->l_export->exp_nid_stats &&
973 lock->l_export->exp_nid_stats->nid_ldlm_stats)
974 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
975 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
977 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
983 * ->l_completion_ast callback for a remote lock in server namespace.
985 * Sends AST to the client notifying it of lock granting. If initial
986 * lock response was not sent yet, instead of sending another RPC, just
987 * mark the lock as granted and client will understand
989 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
991 struct ldlm_cb_set_arg *arg = data;
992 struct ldlm_request *body;
993 struct ptlrpc_request *req;
994 struct ldlm_cb_async_args *ca;
995 int instant_cancel = 0;
1001 LASSERT(lock != NULL);
1002 LASSERT(data != NULL);
1004 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
1005 LDLM_DEBUG(lock, "dropping CP AST");
1009 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
1010 &RQF_LDLM_CP_CALLBACK);
1014 /* server namespace, doesn't need lock */
1015 lvb_len = ldlm_lvbo_size(lock);
1017 * LU-3124 & LU-2187: to not return layout in completion AST because
1018 * it may deadlock for LU-2187, or client may not have enough space
1019 * for large layout. The layout will be returned to client with an
1020 * extra RPC to fetch xattr.lov
1022 if (ldlm_has_layout(lock))
1025 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
1026 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
1028 ptlrpc_request_free(req);
1032 ca = ptlrpc_req_async_args(ca, req);
1033 ca->ca_set_arg = arg;
1036 req->rq_interpret_reply = ldlm_cb_interpret;
1037 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1039 body->lock_handle[0] = lock->l_remote_handle;
1040 body->lock_flags = ldlm_flags_to_wire(flags);
1041 ldlm_lock2desc(lock, &body->lock_desc);
1043 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1044 lvb_len = ldlm_lvbo_fill(lock, lvb, &lvb_len);
1047 * We still need to send the RPC to wake up the blocked
1048 * enqueue thread on the client.
1050 * Consider old client, there is no better way to notify
1051 * the failure, just zero-sized the LVB, then the client
1052 * will fail out as "-EPROTO".
1054 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1058 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1063 LDLM_DEBUG(lock, "server preparing completion AST");
1065 ptlrpc_request_set_replen(req);
1067 req->rq_send_state = LUSTRE_IMP_FULL;
1068 /* ptlrpc_request_pack already set timeout */
1070 req->rq_timeout = ldlm_get_rq_timeout();
1072 /* We only send real blocking ASTs after the lock is granted */
1073 lock_res_and_lock(lock);
1074 if (ldlm_is_ast_sent(lock)) {
1075 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1076 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1077 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1081 * We might get here prior to ldlm_handle_enqueue setting
1082 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1083 * into waiting list, but this is safe and similar code in
1084 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1085 * that would not only cancel the lock, but will also remove
1086 * it from waiting list
1088 if (ldlm_is_cancel_on_block(lock)) {
1089 unlock_res_and_lock(lock);
1090 ldlm_lock_cancel(lock);
1093 req->rq_no_resend = 1;
1095 lock_res_and_lock(lock);
1097 /* start the lock-timeout clock */
1098 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1099 /* Do not resend after lock callback timeout */
1100 req->rq_delay_limit = ldlm_bl_timeout(lock);
1101 req->rq_resend_cb = ldlm_update_resend;
1104 unlock_res_and_lock(lock);
1106 if (lock->l_export && lock->l_export->exp_nid_stats &&
1107 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1108 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1109 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1111 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1113 RETURN(lvb_len < 0 ? lvb_len : rc);
1117 * Server side ->l_glimpse_ast handler for client locks.
1119 * Sends glimpse AST to the client and waits for reply. Then updates
1120 * lvbo with the result.
1122 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1124 struct ldlm_cb_set_arg *arg = data;
1125 struct ldlm_request *body;
1126 struct ptlrpc_request *req;
1127 struct ldlm_cb_async_args *ca;
1129 struct req_format *req_fmt;
1133 LASSERT(lock != NULL);
1135 if (arg->gl_desc != NULL)
1136 /* There is a glimpse descriptor to pack */
1137 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1139 req_fmt = &RQF_LDLM_GL_CALLBACK;
1141 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1142 req_fmt, LUSTRE_DLM_VERSION,
1148 if (arg->gl_desc != NULL) {
1149 /* copy the GL descriptor */
1150 union ldlm_gl_desc *desc;
1152 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1153 *desc = *arg->gl_desc;
1156 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1157 body->lock_handle[0] = lock->l_remote_handle;
1158 ldlm_lock2desc(lock, &body->lock_desc);
1160 ca = ptlrpc_req_async_args(ca, req);
1161 ca->ca_set_arg = arg;
1164 /* server namespace, doesn't need lock */
1165 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1166 ldlm_lvbo_size(lock));
1167 ptlrpc_request_set_replen(req);
1169 req->rq_send_state = LUSTRE_IMP_FULL;
1170 /* ptlrpc_request_alloc_pack already set timeout */
1172 req->rq_timeout = ldlm_get_rq_timeout();
1174 req->rq_interpret_reply = ldlm_cb_interpret;
1176 if (lock->l_export && lock->l_export->exp_nid_stats) {
1177 struct nid_stat *nid_stats = lock->l_export->exp_nid_stats;
1179 lprocfs_counter_incr(nid_stats->nid_ldlm_stats,
1180 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1183 rc = ldlm_ast_fini(req, arg, lock, 0);
1187 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1189 int ldlm_glimpse_locks(struct ldlm_resource *res,
1190 struct list_head *gl_work_list)
1196 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1198 if (rc == -ERESTART)
1199 ldlm_reprocess_all(res, NULL);
1203 EXPORT_SYMBOL(ldlm_glimpse_locks);
1205 /* return LDLM lock associated with a lock callback request */
1206 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1208 struct ldlm_cb_async_args *ca;
1209 struct ldlm_lock *lock;
1213 ca = ptlrpc_req_async_args(ca, req);
1216 RETURN(ERR_PTR(-EFAULT));
1220 EXPORT_SYMBOL(ldlm_request_lock);
1223 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1224 * service threads to carry out client lock enqueueing requests.
1226 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1227 struct ptlrpc_request *req,
1228 const struct ldlm_request *dlm_req,
1229 const struct ldlm_callback_suite *cbs)
1231 struct ldlm_reply *dlm_rep;
1233 enum ldlm_error err = ELDLM_OK;
1234 struct ldlm_lock *lock = NULL;
1235 void *cookie = NULL;
1237 struct ldlm_resource *res = NULL;
1238 const struct lu_env *env = req->rq_svc_thread->t_env;
1242 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1244 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1245 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1247 LASSERT(req->rq_export);
1249 /* for intent enqueue the stat will be updated inside intent policy */
1250 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1251 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1252 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1254 if (req->rq_export && req->rq_export->exp_nid_stats &&
1255 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1256 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1257 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1259 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1260 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1261 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1262 dlm_req->lock_desc.l_resource.lr_type);
1263 GOTO(out, rc = -EFAULT);
1266 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1267 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1268 dlm_req->lock_desc.l_req_mode &
1269 (dlm_req->lock_desc.l_req_mode-1))) {
1270 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1271 dlm_req->lock_desc.l_req_mode);
1272 GOTO(out, rc = -EFAULT);
1275 if (unlikely((flags & LDLM_FL_REPLAY) ||
1276 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1277 /* Find an existing lock in the per-export lock hash */
1279 * In the function below, .hs_keycmp resolves to
1280 * 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,
1287 "found existing lock cookie %#llx",
1288 lock->l_handle.h_cookie);
1289 flags |= LDLM_FL_RESENT;
1290 GOTO(existing_lock, rc = 0);
1293 if (ldlm_reclaim_full()) {
1294 DEBUG_REQ(D_DLMTRACE, req,
1295 "Too many granted locks, reject current enqueue request and let the client retry later");
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");
1315 * Initialize resource lvb but not for a lock being replayed since
1316 * Client already got lvb sent in this case.
1317 * This must occur early since some policy methods assume resource
1318 * lvb is available (lr_lvb_data != NULL).
1320 res = lock->l_resource;
1321 if (!(flags & LDLM_FL_REPLAY)) {
1322 /* non-replayed lock, delayed lvb init may need to be done */
1323 rc = ldlm_lvbo_init(res);
1325 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1330 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1332 * Don't enqueue a lock onto the export if it is been disonnected
1333 * due to eviction (b=3822) or server umount (b=24324).
1334 * Cancel it now instead.
1336 if (req->rq_export->exp_disconnected) {
1337 LDLM_ERROR(lock, "lock on disconnected export %p",
1339 GOTO(out, rc = -ENOTCONN);
1342 lock->l_export = class_export_lock_get(req->rq_export, lock);
1343 if (lock->l_export->exp_lock_hash)
1344 cfs_hash_add(lock->l_export->exp_lock_hash,
1345 &lock->l_remote_handle,
1349 * Inherit the enqueue flags before the operation, because we do not
1350 * keep the res lock on return and next operations (BL AST) may proceed
1353 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1354 LDLM_FL_INHERIT_MASK);
1356 ldlm_convert_policy_to_local(req->rq_export,
1357 dlm_req->lock_desc.l_resource.lr_type,
1358 &dlm_req->lock_desc.l_policy_data,
1359 &lock->l_policy_data);
1360 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1361 lock->l_req_extent = lock->l_policy_data.l_extent;
1362 else if (dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS)
1363 lock->l_policy_data.l_inodebits.try_bits =
1364 dlm_req->lock_desc.l_policy_data.l_inodebits.try_bits;
1368 if (!(flags & LDLM_FL_HAS_INTENT)) {
1369 /* based on the assumption that lvb size never changes during
1370 * resource life time otherwise it need resource->lr_lock's
1372 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1373 RCL_SERVER, ldlm_lvbo_size(lock));
1375 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1376 GOTO(out, rc = -ENOMEM);
1378 rc = req_capsule_server_pack(&req->rq_pill);
1383 err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1390 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1392 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1393 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1395 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1396 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1399 * We never send a blocking AST until the lock is granted, but
1400 * we can tell it right now
1402 lock_res_and_lock(lock);
1405 * Now take into account flags to be inherited from original lock
1406 * request both in reply to client and in our own lock flags.
1408 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1409 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1412 * Don't move a pending lock onto the export if it has already been
1413 * disconnected due to eviction (b=5683) or server umount (b=24324).
1414 * Cancel it now instead.
1416 if (unlikely(req->rq_export->exp_disconnected ||
1417 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1418 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1420 } else if (ldlm_is_ast_sent(lock)) {
1421 /* fill lock desc for possible lock convert */
1422 if (lock->l_blocking_lock &&
1423 lock->l_resource->lr_type == LDLM_IBITS) {
1424 struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1425 struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1428 "save blocking bits %llx in granted lock",
1429 bl_lock->l_policy_data.l_inodebits.bits);
1431 * If lock is blocked then save blocking ibits
1432 * in returned lock policy for the possible lock
1433 * convert on a client.
1435 rep_desc->l_policy_data.l_inodebits.cancel_bits =
1436 bl_lock->l_policy_data.l_inodebits.bits;
1438 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1439 if (ldlm_is_granted(lock)) {
1441 * Only cancel lock if it was granted, because it would
1442 * be destroyed immediately and would never be granted
1443 * in the future, causing timeouts on client. Not
1444 * granted lock will be cancelled immediately after
1445 * sending completion AST.
1447 if (ldlm_is_cancel_on_block(lock)) {
1448 unlock_res_and_lock(lock);
1449 ldlm_lock_cancel(lock);
1450 lock_res_and_lock(lock);
1452 ldlm_add_waiting_lock(lock,
1453 ldlm_bl_timeout(lock));
1457 unlock_res_and_lock(lock);
1461 req->rq_status = rc ?: err; /* return either error - b=11190 */
1462 if (!req->rq_packed_final) {
1463 err = lustre_pack_reply(req, 1, NULL, NULL);
1469 * The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1470 * ldlm_reprocess_all. If this moves, revisit that code. -phil
1474 "server-side enqueue handler, sending reply (err=%d, rc=%d)",
1478 req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1480 ldlm_lvbo_size(lock) > 0) {
1485 buf = req_capsule_server_get(&req->rq_pill,
1487 LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock);
1488 buflen = req_capsule_get_size(&req->rq_pill,
1489 &RMF_DLM_LVB, RCL_SERVER);
1491 * non-replayed lock, delayed lvb init may
1492 * need to be occur now
1494 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1497 rc2 = ldlm_lvbo_fill(lock, buf, &buflen);
1499 req_capsule_shrink(&req->rq_pill,
1502 } else if (rc2 == -ERANGE) {
1503 rc2 = req_capsule_server_grow(
1505 &RMF_DLM_LVB, buflen);
1510 * if we can't grow the buffer,
1511 * it's ok to return empty lvb
1522 } else if (flags & LDLM_FL_REPLAY) {
1523 /* no LVB resend upon replay */
1525 req_capsule_shrink(&req->rq_pill,
1535 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1536 if (lock->l_export) {
1537 ldlm_lock_cancel(lock);
1539 lock_res_and_lock(lock);
1540 ldlm_resource_unlink_lock(lock);
1541 ldlm_lock_destroy_nolock(lock);
1542 unlock_res_and_lock(lock);
1547 if (!err && !ldlm_is_cbpending(lock) &&
1548 dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1549 ldlm_reprocess_all(lock->l_resource, lock);
1551 LDLM_LOCK_RELEASE(lock);
1554 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1561 * Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1562 * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1564 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1566 if (lock->l_blocking_lock) {
1567 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1568 lock->l_blocking_lock = NULL;
1572 /* A lock can be converted to new ibits or mode and should be considered
1573 * as new lock. Clear all states related to a previous blocking AST
1574 * processing so new conflicts will cause new blocking ASTs.
1576 * This is used during lock convert below and lock downgrade to COS mode in
1577 * ldlm_lock_mode_downgrade().
1579 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1581 ldlm_clear_ast_sent(lock);
1582 lock->l_bl_ast_run = 0;
1583 ldlm_clear_blocking_lock(lock);
1587 * Main LDLM entry point for server code to process lock conversion requests.
1589 int ldlm_handle_convert0(struct ptlrpc_request *req,
1590 const struct ldlm_request *dlm_req)
1592 struct obd_export *exp = req->rq_export;
1593 struct ldlm_reply *dlm_rep;
1594 struct ldlm_lock *lock;
1601 if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1602 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1603 LDLM_CONVERT - LDLM_FIRST_OPC);
1605 rc = req_capsule_server_pack(&req->rq_pill);
1609 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1610 dlm_rep->lock_flags = dlm_req->lock_flags;
1612 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1614 LDLM_DEBUG_NOLOCK("server lock is canceled already");
1615 req->rq_status = ELDLM_NO_LOCK_DATA;
1619 LDLM_DEBUG(lock, "server-side convert handler START");
1621 lock_res_and_lock(lock);
1622 bits = lock->l_policy_data.l_inodebits.bits;
1623 new_bits = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1625 if (ldlm_is_cancel(lock)) {
1626 LDLM_DEBUG(lock, "convert on canceled lock!");
1627 unlock_res_and_lock(lock);
1628 GOTO(out_put, rc = ELDLM_NO_LOCK_DATA);
1631 if (dlm_req->lock_desc.l_req_mode != lock->l_granted_mode) {
1632 LDLM_ERROR(lock, "lock mode differs!");
1633 unlock_res_and_lock(lock);
1634 GOTO(out_put, rc = -EPROTO);
1637 if (bits == new_bits) {
1639 * This can be valid situation if CONVERT RPCs are
1640 * re-ordered. Just finish silently
1642 LDLM_DEBUG(lock, "lock is converted already!");
1643 unlock_res_and_lock(lock);
1645 if (ldlm_is_waited(lock))
1646 ldlm_del_waiting_lock(lock);
1648 ldlm_clear_cbpending(lock);
1649 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1650 ldlm_inodebits_drop(lock, bits & ~new_bits);
1652 ldlm_clear_blocking_data(lock);
1653 unlock_res_and_lock(lock);
1655 ldlm_reprocess_all(lock->l_resource, NULL);
1658 dlm_rep->lock_handle = lock->l_remote_handle;
1659 ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1660 &dlm_rep->lock_desc.l_policy_data);
1664 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d", rc);
1665 LDLM_LOCK_PUT(lock);
1666 req->rq_status = rc;
1671 * Cancel all the locks whose handles are packed into ldlm_request
1673 * Called by server code expecting such combined cancel activity
1676 int ldlm_request_cancel(struct ptlrpc_request *req,
1677 const struct ldlm_request *dlm_req,
1678 int first, enum lustre_at_flags flags)
1680 struct ldlm_resource *res, *pres = NULL;
1681 struct ldlm_lock *lock;
1682 int i, count, done = 0;
1687 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
1688 if (size <= offsetof(struct ldlm_request, lock_handle) ||
1689 (size - offsetof(struct ldlm_request, lock_handle)) /
1690 sizeof(struct lustre_handle) < dlm_req->lock_count)
1693 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1697 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1701 * There is no lock on the server at the replay time,
1702 * skip lock cancelling to make replay tests to pass.
1704 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1707 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, starting at %d",
1710 for (i = first; i < count; i++) {
1711 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1713 /* below message checked in replay-single.sh test_36 */
1714 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (cookie %llu)",
1715 dlm_req->lock_handle[i].cookie);
1719 res = lock->l_resource;
1723 * This code is an optimization to only attempt lock
1724 * granting on the resource (that could be CPU-expensive)
1725 * after we are done cancelling lock in that resource.
1729 ldlm_reprocess_all(pres, NULL);
1730 LDLM_RESOURCE_DELREF(pres);
1731 ldlm_resource_putref(pres);
1734 ldlm_resource_getref(res);
1735 LDLM_RESOURCE_ADDREF(res);
1737 if (!ldlm_is_discard_data(lock))
1738 ldlm_lvbo_update(res, lock,
1744 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1745 lock->l_blast_sent != 0) {
1746 time64_t delay = ktime_get_real_seconds() -
1749 "server cancels blocked lock after %llds",
1751 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1753 ldlm_lock_cancel(lock);
1754 LDLM_LOCK_PUT(lock);
1757 ldlm_reprocess_all(pres, NULL);
1758 LDLM_RESOURCE_DELREF(pres);
1759 ldlm_resource_putref(pres);
1761 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1764 EXPORT_SYMBOL(ldlm_request_cancel);
1767 * Main LDLM entry point for server code to cancel locks.
1769 * Typically gets called from service handler on LDLM_CANCEL opc.
1771 int ldlm_handle_cancel(struct ptlrpc_request *req)
1773 struct ldlm_request *dlm_req;
1778 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1779 if (dlm_req == NULL) {
1780 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1784 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) <
1785 offsetof(struct ldlm_request, lock_handle[1]))
1788 if (req->rq_export && req->rq_export->exp_nid_stats &&
1789 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1790 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1791 LDLM_CANCEL - LDLM_FIRST_OPC);
1793 rc = req_capsule_server_pack(&req->rq_pill);
1797 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1798 req->rq_status = LUSTRE_ESTALE;
1800 RETURN(ptlrpc_reply(req));
1802 #endif /* HAVE_SERVER_SUPPORT */
1805 * Server may pass additional information about blocking lock.
1806 * For IBITS locks it is conflicting bits which can be used for
1807 * lock convert instead of cancel.
1809 void ldlm_bl_desc2lock(const struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1811 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
1813 check_res_locked(lock->l_resource);
1814 if (ns_is_client(ns) && ld &&
1815 (lock->l_resource->lr_type == LDLM_IBITS)) {
1817 * Lock description contains policy of blocking lock,
1818 * and its cancel_bits is used to pass conflicting bits.
1819 * NOTE: ld can be NULL or can be not NULL but zeroed if
1820 * passed from ldlm_bl_thread_blwi(), check below used bits
1821 * in ld to make sure it is valid description.
1823 * If server may replace lock resource keeping the same cookie,
1824 * never use cancel bits from different resource, full cancel
1827 if (ld->l_policy_data.l_inodebits.cancel_bits &&
1828 ldlm_res_eq(&ld->l_resource.lr_name,
1829 &lock->l_resource->lr_name) &&
1830 !(ldlm_is_cbpending(lock) &&
1831 lock->l_policy_data.l_inodebits.cancel_bits == 0)) {
1832 /* always combine conflicting ibits */
1833 lock->l_policy_data.l_inodebits.cancel_bits |=
1834 ld->l_policy_data.l_inodebits.cancel_bits;
1836 /* If cancel_bits are not obtained or
1837 * if the lock is already CBPENDING and
1838 * has no cancel_bits set
1839 * - the full lock is to be cancelled
1841 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1847 * Callback handler for receiving incoming blocking ASTs.
1849 * This can only happen on client side.
1851 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1852 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1858 LDLM_DEBUG(lock, "client blocking AST callback handler");
1860 lock_res_and_lock(lock);
1862 /* get extra information from desc if any */
1863 ldlm_bl_desc2lock(ld, lock);
1864 ldlm_set_cbpending(lock);
1866 do_ast = (!lock->l_readers && !lock->l_writers);
1867 unlock_res_and_lock(lock);
1871 "Lock %p already unused, calling callback (%p)\n",
1872 lock, lock->l_blocking_ast);
1873 if (lock->l_blocking_ast != NULL)
1874 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1878 "Lock %p is referenced, will be cancelled later\n",
1882 LDLM_DEBUG(lock, "client blocking callback handler END");
1883 LDLM_LOCK_RELEASE(lock);
1887 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1889 if (req->rq_no_reply)
1892 req->rq_status = rc;
1893 if (!req->rq_packed_final) {
1894 rc = lustre_pack_reply(req, 1, NULL, NULL);
1898 return ptlrpc_reply(req);
1902 * Callback handler for receiving incoming completion ASTs.
1904 * This only can happen on client side.
1906 static int ldlm_handle_cp_callback(struct ptlrpc_request *req,
1907 struct ldlm_namespace *ns,
1908 struct ldlm_request *dlm_req,
1909 struct ldlm_lock *lock)
1911 struct list_head ast_list;
1917 LDLM_DEBUG(lock, "client completion callback handler START");
1919 INIT_LIST_HEAD(&ast_list);
1920 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1921 long to = cfs_time_seconds(1);
1923 ldlm_callback_reply(req, 0);
1926 set_current_state(TASK_INTERRUPTIBLE);
1927 schedule_timeout(to);
1928 if (ldlm_is_granted(lock) ||
1929 ldlm_is_destroyed(lock))
1934 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1936 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1937 GOTO(out, rc = lvb_len);
1938 } else if (lvb_len > 0) {
1939 if (lock->l_lvb_len > 0) {
1940 /* for extent lock, lvb contains ost_lvb{}. */
1941 LASSERT(lock->l_lvb_data != NULL);
1943 if (unlikely(lock->l_lvb_len < lvb_len)) {
1945 "Replied LVB is larger than expectation, expected = %d, replied = %d",
1946 lock->l_lvb_len, lvb_len);
1947 GOTO(out, rc = -EINVAL);
1952 lock_res_and_lock(lock);
1954 if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
1955 &lock->l_resource->lr_name)) {
1956 ldlm_resource_unlink_lock(lock);
1957 unlock_res_and_lock(lock);
1958 rc = ldlm_lock_change_resource(ns, lock,
1959 &dlm_req->lock_desc.l_resource.lr_name);
1961 LDLM_ERROR(lock, "Failed to allocate resource");
1964 LDLM_DEBUG(lock, "completion AST, new resource");
1965 lock_res_and_lock(lock);
1968 if (ldlm_is_failed(lock)) {
1969 unlock_res_and_lock(lock);
1970 LDLM_LOCK_RELEASE(lock);
1974 if (ldlm_is_destroyed(lock) ||
1975 ldlm_is_granted(lock)) {
1976 /* b=11300: the lock has already been granted */
1977 unlock_res_and_lock(lock);
1978 LDLM_DEBUG(lock, "Double grant race happened");
1983 * If we receive the completion AST before the actual enqueue returned,
1984 * then we might need to switch lock modes, resources, or extents.
1986 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1987 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1988 LDLM_DEBUG(lock, "completion AST, new lock mode");
1991 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1992 ldlm_convert_policy_to_local(req->rq_export,
1993 dlm_req->lock_desc.l_resource.lr_type,
1994 &dlm_req->lock_desc.l_policy_data,
1995 &lock->l_policy_data);
1996 LDLM_DEBUG(lock, "completion AST, new policy data");
1999 ldlm_resource_unlink_lock(lock);
2001 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
2003 * BL_AST locks are not needed in LRU.
2004 * Let ldlm_cancel_lru() be fast.
2006 ldlm_lock_remove_from_lru(lock);
2007 ldlm_bl_desc2lock(&dlm_req->lock_desc, lock);
2008 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
2009 LDLM_DEBUG(lock, "completion AST includes blocking AST");
2012 if (lock->l_lvb_len > 0) {
2013 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
2014 lock->l_lvb_data, lvb_len);
2016 unlock_res_and_lock(lock);
2021 ldlm_grant_lock(lock, &ast_list);
2022 unlock_res_and_lock(lock);
2024 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
2027 * Let Enqueue to call osc_lock_upcall() and initialize
2030 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
2032 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
2034 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
2040 lock_res_and_lock(lock);
2041 ldlm_set_failed(lock);
2042 unlock_res_and_lock(lock);
2043 wake_up(&lock->l_waitq);
2045 LDLM_LOCK_RELEASE(lock);
2051 * Callback handler for receiving incoming glimpse ASTs.
2053 * This only can happen on client side. After handling the glimpse AST
2054 * we also consider dropping the lock here if it is unused locally for a
2057 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
2058 struct ldlm_namespace *ns,
2059 struct ldlm_request *dlm_req,
2060 struct ldlm_lock *lock)
2062 struct ldlm_lock_desc *ld = &dlm_req->lock_desc;
2067 LDLM_DEBUG(lock, "client glimpse AST callback handler");
2069 if (lock->l_glimpse_ast != NULL)
2070 rc = lock->l_glimpse_ast(lock, req);
2072 if (req->rq_repmsg != NULL) {
2075 req->rq_status = rc;
2079 lock_res_and_lock(lock);
2080 if (lock->l_granted_mode == LCK_PW &&
2081 !lock->l_readers && !lock->l_writers &&
2082 ktime_after(ktime_get(),
2083 ktime_add(lock->l_last_used,
2084 ktime_set(ns->ns_dirty_age_limit, 0)))) {
2085 unlock_res_and_lock(lock);
2087 /* For MDS glimpse it is always DOM lock, set corresponding
2088 * cancel_bits to perform lock convert if needed
2090 if (lock->l_resource->lr_type == LDLM_IBITS)
2091 ld->l_policy_data.l_inodebits.cancel_bits =
2093 if (ldlm_bl_to_thread_lock(ns, ld, lock))
2094 ldlm_handle_bl_callback(ns, ld, lock);
2099 unlock_res_and_lock(lock);
2100 LDLM_LOCK_RELEASE(lock);
2104 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
2105 enum ldlm_cancel_flags cancel_flags)
2107 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2111 spin_lock(&blp->blp_lock);
2112 if (blwi->blwi_lock &&
2113 ldlm_is_discard_data(blwi->blwi_lock)) {
2114 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
2115 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
2117 /* other blocking callbacks are added to the regular list */
2118 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
2120 spin_unlock(&blp->blp_lock);
2122 wake_up(&blp->blp_waitq);
2125 * can not check blwi->blwi_flags as blwi could be already freed in
2128 if (!(cancel_flags & LCF_ASYNC))
2129 wait_for_completion(&blwi->blwi_comp);
2134 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2135 struct ldlm_namespace *ns,
2136 struct ldlm_lock_desc *ld,
2137 struct list_head *cancels, int count,
2138 struct ldlm_lock *lock,
2139 enum ldlm_cancel_flags cancel_flags)
2141 init_completion(&blwi->blwi_comp);
2142 INIT_LIST_HEAD(&blwi->blwi_head);
2144 if (memory_pressure_get())
2145 blwi->blwi_mem_pressure = 1;
2148 blwi->blwi_flags = cancel_flags;
2150 blwi->blwi_ld = *ld;
2152 list_add(&blwi->blwi_head, cancels);
2153 list_del_init(cancels);
2154 blwi->blwi_count = count;
2156 blwi->blwi_lock = lock;
2161 * Queues a list of locks \a cancels containing \a count locks
2162 * for later processing by a blocking thread. If \a count is zero,
2163 * then the lock referenced as \a lock is queued instead.
2165 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2166 * If list addition fails an error is returned and caller is supposed to
2167 * call ->l_blocking_ast itself.
2169 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2170 struct ldlm_lock_desc *ld,
2171 struct ldlm_lock *lock,
2172 struct list_head *cancels, int count,
2173 enum ldlm_cancel_flags cancel_flags)
2177 if (cancels && count == 0)
2180 if (cancel_flags & LCF_ASYNC) {
2181 struct ldlm_bl_work_item *blwi;
2183 OBD_ALLOC(blwi, sizeof(*blwi));
2186 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2188 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2191 * if it is synchronous call do minimum mem alloc, as it could
2192 * be triggered from kernel shrinker
2194 struct ldlm_bl_work_item blwi;
2196 memset(&blwi, 0, sizeof(blwi));
2197 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2198 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2203 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2204 struct ldlm_lock *lock)
2206 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2209 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2210 struct list_head *cancels, int count,
2211 enum ldlm_cancel_flags cancel_flags)
2213 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2216 int ldlm_bl_thread_wakeup(void)
2218 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2222 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2223 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2225 struct obd_device *obd = req->rq_export->exp_obd;
2233 DEBUG_REQ(D_HSM, req, "%s: handle setinfo", obd->obd_name);
2235 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2237 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2239 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2242 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2244 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2246 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2249 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2252 /* We are responsible for swabbing contents of val */
2254 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2255 /* Pass it on to mdc (the "export" in this case) */
2256 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2258 sizeof(KEY_HSM_COPYTOOL_SEND),
2259 KEY_HSM_COPYTOOL_SEND,
2262 DEBUG_REQ(D_WARNING, req, "ignoring unknown key '%s'", key);
2267 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2268 const char *msg, int rc,
2269 const struct lustre_handle *handle)
2271 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2272 "%s, NID=%s lock=%#llx: rc = %d",
2273 msg, libcfs_id2str(req->rq_peer),
2274 handle ? handle->cookie : 0, rc);
2275 if (req->rq_no_reply)
2276 CWARN("No reply was sent, maybe cause b=21636.\n");
2278 CWARN("Send reply failed, maybe cause b=21636.\n");
2281 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2282 static int ldlm_callback_handler(struct ptlrpc_request *req)
2284 struct ldlm_namespace *ns;
2285 struct ldlm_request *dlm_req;
2286 struct ldlm_lock *lock;
2292 * Requests arrive in sender's byte order. The ptlrpc service
2293 * handler has already checked and, if necessary, byte-swapped the
2294 * incoming request message body, but I am responsible for the
2298 /* do nothing for sec context finalize */
2299 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2302 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2304 if (req->rq_export == NULL) {
2305 rc = ldlm_callback_reply(req, -ENOTCONN);
2306 ldlm_callback_errmsg(req, "Operate on unconnected server",
2311 LASSERT(req->rq_export != NULL);
2312 LASSERT(req->rq_export->exp_obd != NULL);
2314 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2315 case LDLM_BL_CALLBACK:
2316 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2318 ldlm_callback_reply(req, -(int)cfs_fail_err);
2322 case LDLM_CP_CALLBACK:
2323 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2326 case LDLM_GL_CALLBACK:
2327 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2331 rc = ldlm_handle_setinfo(req);
2332 ldlm_callback_reply(req, rc);
2335 CERROR("unknown opcode %u\n",
2336 lustre_msg_get_opc(req->rq_reqmsg));
2337 ldlm_callback_reply(req, -EPROTO);
2341 ns = req->rq_export->exp_obd->obd_namespace;
2342 LASSERT(ns != NULL);
2344 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2346 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2347 if (dlm_req == NULL) {
2348 rc = ldlm_callback_reply(req, -EPROTO);
2349 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2355 * Force a known safe race, send a cancel to the server for a lock
2356 * which the server has already started a blocking callback on.
2358 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2359 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2360 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2362 CERROR("ldlm_cli_cancel: %d\n", rc);
2365 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2368 "callback on lock %#llx - lock disappeared\n",
2369 dlm_req->lock_handle[0].cookie);
2370 rc = ldlm_callback_reply(req, -EINVAL);
2371 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2372 &dlm_req->lock_handle[0]);
2376 if (ldlm_is_fail_loc(lock) &&
2377 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2378 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2380 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2381 lock_res_and_lock(lock);
2382 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2384 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2386 * If somebody cancels lock and cache is already dropped,
2387 * or lock is failed before cp_ast received on client,
2388 * we can tell the server we have no lock. Otherwise, we
2389 * should send cancel after dropping the cache.
2391 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2392 ldlm_is_failed(lock)) {
2394 "callback on lock %llx - lock disappeared",
2395 dlm_req->lock_handle[0].cookie);
2396 unlock_res_and_lock(lock);
2397 LDLM_LOCK_RELEASE(lock);
2398 rc = ldlm_callback_reply(req, -EINVAL);
2399 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2400 &dlm_req->lock_handle[0]);
2404 * BL_AST locks are not needed in LRU.
2405 * Let ldlm_cancel_lru() be fast.
2407 ldlm_lock_remove_from_lru(lock);
2408 ldlm_set_bl_ast(lock);
2410 unlock_res_and_lock(lock);
2413 * We want the ost thread to get this reply so that it can respond
2414 * to ost requests (write cache writeback) that might be triggered
2417 * But we'd also like to be able to indicate in the reply that we're
2418 * cancelling right now, because it's unused, or have an intent result
2419 * in the reply, so we might have to push the responsibility for sending
2420 * the reply down into the AST handlers, alas.
2423 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2424 case LDLM_BL_CALLBACK:
2425 CDEBUG(D_INODE, "blocking ast\n");
2426 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2427 if (!ldlm_is_cancel_on_block(lock)) {
2428 rc = ldlm_callback_reply(req, 0);
2429 if (req->rq_no_reply || rc)
2430 ldlm_callback_errmsg(req, "Normal process", rc,
2431 &dlm_req->lock_handle[0]);
2433 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2434 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2436 case LDLM_CP_CALLBACK:
2437 CDEBUG(D_INODE, "completion ast\n");
2438 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2439 rc = ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2440 if (!OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE))
2441 ldlm_callback_reply(req, rc);
2443 case LDLM_GL_CALLBACK:
2444 CDEBUG(D_INODE, "glimpse ast\n");
2445 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2446 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2449 LBUG(); /* checked above */
2455 #ifdef HAVE_SERVER_SUPPORT
2457 * Main handler for canceld thread.
2459 * Separated into its own thread to avoid deadlocks.
2461 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2468 * Requests arrive in sender's byte order. The ptlrpc service
2469 * handler has already checked and, if necessary, byte-swapped the
2470 * incoming request message body, but I am responsible for the
2474 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2476 if (req->rq_export == NULL) {
2477 struct ldlm_request *dlm_req;
2479 CERROR("%s from %s arrived at %llu with bad export cookie %llu\n",
2480 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2481 libcfs_nid2str(req->rq_peer.nid),
2482 (unsigned long long)req->rq_arrival_time.tv_sec,
2483 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2485 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2486 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2487 dlm_req = req_capsule_client_get(&req->rq_pill,
2489 if (dlm_req != NULL)
2490 ldlm_lock_dump_handle(D_ERROR,
2491 &dlm_req->lock_handle[0]);
2493 ldlm_callback_reply(req, -ENOTCONN);
2497 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2498 /* XXX FIXME move this back to mds/handler.c, b=249 */
2500 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2501 CDEBUG(D_INODE, "cancel\n");
2502 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2503 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2504 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2506 rc = ldlm_handle_cancel(req);
2510 struct ldlm_request *dlm_req;
2512 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2513 CDEBUG(D_INODE, "convert\n");
2515 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2516 if (dlm_req == NULL) {
2517 CDEBUG(D_INFO, "bad request buffer for cancel\n");
2518 rc = ldlm_callback_reply(req, -EPROTO);
2520 req->rq_status = ldlm_handle_convert0(req, dlm_req);
2521 rc = ptlrpc_reply(req);
2526 CERROR("invalid opcode %d\n",
2527 lustre_msg_get_opc(req->rq_reqmsg));
2528 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2529 rc = ldlm_callback_reply(req, -EINVAL);
2535 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2536 struct ldlm_lock *lock)
2538 struct ldlm_request *dlm_req;
2539 struct lustre_handle lockh;
2545 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2546 if (dlm_req == NULL)
2549 ldlm_lock2handle(lock, &lockh);
2550 for (i = 0; i < dlm_req->lock_count; i++) {
2551 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2553 DEBUG_REQ(D_RPCTRACE, req,
2554 "Prio raised by lock %#llx", lockh.cookie);
2563 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2565 struct ldlm_request *dlm_req;
2572 /* no prolong in recovery */
2573 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2576 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2577 if (dlm_req == NULL)
2580 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
2581 if (size <= offsetof(struct ldlm_request, lock_handle) ||
2582 (size - offsetof(struct ldlm_request, lock_handle)) /
2583 sizeof(struct lustre_handle) < dlm_req->lock_count)
2586 for (i = 0; i < dlm_req->lock_count; i++) {
2587 struct ldlm_lock *lock;
2589 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2593 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2595 LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2596 LDLM_LOCK_PUT(lock);
2605 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2606 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2607 .hpreq_check = ldlm_cancel_hpreq_check,
2611 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2615 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2617 if (req->rq_export == NULL)
2620 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2621 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2622 req->rq_ops = &ldlm_cancel_hpreq_ops;
2623 } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2624 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2625 req->rq_ops = &ldlm_cancel_hpreq_ops;
2630 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2631 struct hlist_node *hnode, void *data)
2634 struct list_head *rpc_list = data;
2635 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2637 lock_res_and_lock(lock);
2639 if (!ldlm_is_granted(lock)) {
2640 unlock_res_and_lock(lock);
2644 LASSERT(lock->l_resource);
2645 if (lock->l_resource->lr_type != LDLM_IBITS &&
2646 lock->l_resource->lr_type != LDLM_PLAIN) {
2647 unlock_res_and_lock(lock);
2651 if (ldlm_is_ast_sent(lock)) {
2652 unlock_res_and_lock(lock);
2656 LASSERT(lock->l_blocking_ast);
2657 LASSERT(!lock->l_blocking_lock);
2659 ldlm_set_ast_sent(lock);
2660 if (lock->l_export && lock->l_export->exp_lock_hash) {
2662 * NB: it's safe to call cfs_hash_del() even lock isn't
2666 * In the function below, .hs_keycmp resolves to
2667 * ldlm_export_lock_keycmp()
2669 /* coverity[overrun-buffer-val] */
2670 cfs_hash_del(lock->l_export->exp_lock_hash,
2671 &lock->l_remote_handle, &lock->l_exp_hash);
2674 list_add_tail(&lock->l_rk_ast, rpc_list);
2675 LDLM_LOCK_GET(lock);
2677 unlock_res_and_lock(lock);
2681 void ldlm_revoke_export_locks(struct obd_export *exp)
2683 struct list_head rpc_list;
2687 INIT_LIST_HEAD(&rpc_list);
2688 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2689 ldlm_revoke_lock_cb, &rpc_list, 0);
2690 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2691 LDLM_WORK_REVOKE_AST);
2695 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2696 #endif /* HAVE_SERVER_SUPPORT */
2698 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2699 struct ldlm_bl_work_item **p_blwi,
2700 struct obd_export **p_exp)
2702 struct ldlm_bl_work_item *blwi = NULL;
2703 static unsigned int num_bl;
2704 static unsigned int num_stale;
2705 int num_th = atomic_read(&blp->blp_num_threads);
2707 *p_exp = obd_stale_export_get();
2709 spin_lock(&blp->blp_lock);
2710 if (*p_exp != NULL) {
2711 if (num_th == 1 || ++num_stale < num_th) {
2712 spin_unlock(&blp->blp_lock);
2718 /* process a request from the blp_list at least every blp_num_threads */
2719 if (!list_empty(&blp->blp_list) &&
2720 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2721 blwi = list_entry(blp->blp_list.next,
2722 struct ldlm_bl_work_item, blwi_entry);
2724 if (!list_empty(&blp->blp_prio_list))
2725 blwi = list_entry(blp->blp_prio_list.next,
2726 struct ldlm_bl_work_item,
2730 if (++num_bl >= num_th)
2732 list_del(&blwi->blwi_entry);
2734 spin_unlock(&blp->blp_lock);
2737 if (*p_exp != NULL && *p_blwi != NULL) {
2738 obd_stale_export_put(*p_exp);
2742 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2745 /* This only contains temporary data until the thread starts */
2746 struct ldlm_bl_thread_data {
2747 struct ldlm_bl_pool *bltd_blp;
2748 struct completion bltd_comp;
2752 static int ldlm_bl_thread_main(void *arg);
2754 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2756 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2757 struct task_struct *task;
2759 init_completion(&bltd.bltd_comp);
2761 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2762 if (bltd.bltd_num >= blp->blp_max_threads) {
2763 atomic_dec(&blp->blp_num_threads);
2767 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2769 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2770 atomic_dec(&blp->blp_num_threads);
2774 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2777 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2778 bltd.bltd_num, PTR_ERR(task));
2779 atomic_dec(&blp->blp_num_threads);
2780 return PTR_ERR(task);
2782 wait_for_completion(&bltd.bltd_comp);
2787 /* Not fatal if racy and have a few too many threads */
2788 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2789 struct ldlm_bl_work_item *blwi)
2791 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2794 if (atomic_read(&blp->blp_busy_threads) <
2795 atomic_read(&blp->blp_num_threads))
2798 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2799 blwi->blwi_mem_pressure))
2805 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2806 struct ldlm_bl_work_item *blwi)
2810 if (blwi->blwi_ns == NULL)
2811 /* added by ldlm_cleanup() */
2812 RETURN(LDLM_ITER_STOP);
2814 if (blwi->blwi_mem_pressure)
2815 memory_pressure_set();
2817 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2819 if (blwi->blwi_count) {
2822 * The special case when we cancel locks in lru
2823 * asynchronously, we pass the list of locks here.
2824 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2825 * canceled locally yet.
2827 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2830 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2833 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2836 if (blwi->blwi_mem_pressure)
2837 memory_pressure_clr();
2839 if (blwi->blwi_flags & LCF_ASYNC)
2840 OBD_FREE(blwi, sizeof(*blwi));
2842 complete(&blwi->blwi_comp);
2848 * Cancel stale locks on export. Cancel blocked locks first.
2849 * If the given export has blocked locks, the next in the list may have
2850 * them too, thus cancel not blocked locks only if the current export has
2853 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2854 struct obd_export *exp)
2860 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2862 num = ldlm_export_cancel_blocked_locks(exp);
2864 ldlm_export_cancel_locks(exp);
2866 obd_stale_export_put(exp);
2873 * Main blocking requests processing thread.
2875 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2876 * This thread in the end ends up doing actual call to ->l_blocking_ast
2879 static int ldlm_bl_thread_main(void *arg)
2882 struct ldlm_bl_pool *blp;
2883 struct ldlm_bl_thread_data *bltd = arg;
2891 rc = lu_env_init(env, LCT_DT_THREAD);
2894 rc = lu_env_add(env);
2896 GOTO(out_env_fini, rc);
2898 blp = bltd->bltd_blp;
2900 complete(&bltd->bltd_comp);
2901 /* cannot use bltd after this, it is only on caller's stack */
2904 struct ldlm_bl_work_item *blwi = NULL;
2905 struct obd_export *exp = NULL;
2908 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2911 wait_event_idle_exclusive(blp->blp_waitq,
2912 ldlm_bl_get_work(blp, &blwi,
2914 atomic_inc(&blp->blp_busy_threads);
2916 if (ldlm_bl_thread_need_create(blp, blwi))
2917 /* discard the return value, we tried */
2918 ldlm_bl_thread_start(blp, true);
2921 rc = ldlm_bl_thread_exports(blp, exp);
2923 rc = ldlm_bl_thread_blwi(blp, blwi);
2925 atomic_dec(&blp->blp_busy_threads);
2927 if (rc == LDLM_ITER_STOP)
2931 * If there are many namespaces, we will not sleep waiting for
2932 * work, and must do a cond_resched to avoid holding the CPU
2938 atomic_dec(&blp->blp_num_threads);
2939 complete(&blp->blp_comp);
2950 static int ldlm_setup(void);
2951 static int ldlm_cleanup(void);
2953 int ldlm_get_ref(void)
2958 mutex_lock(&ldlm_ref_mutex);
2959 if (++ldlm_refcount == 1) {
2964 mutex_unlock(&ldlm_ref_mutex);
2969 void ldlm_put_ref(void)
2972 mutex_lock(&ldlm_ref_mutex);
2973 if (ldlm_refcount == 1) {
2974 int rc = ldlm_cleanup();
2977 CERROR("ldlm_cleanup failed: %d\n", rc);
2983 mutex_unlock(&ldlm_ref_mutex);
2989 * Export handle<->lock hash operations.
2992 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned int mask)
2994 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2998 ldlm_export_lock_key(struct hlist_node *hnode)
3000 struct ldlm_lock *lock;
3002 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3003 return &lock->l_remote_handle;
3007 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
3009 struct ldlm_lock *lock;
3011 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3012 lock->l_remote_handle = *(struct lustre_handle *)key;
3016 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
3018 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
3022 ldlm_export_lock_object(struct hlist_node *hnode)
3024 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3028 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
3030 struct ldlm_lock *lock;
3032 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3033 LDLM_LOCK_GET(lock);
3037 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
3039 struct ldlm_lock *lock;
3041 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3042 LDLM_LOCK_RELEASE(lock);
3045 static struct cfs_hash_ops ldlm_export_lock_ops = {
3046 .hs_hash = ldlm_export_lock_hash,
3047 .hs_key = ldlm_export_lock_key,
3048 .hs_keycmp = ldlm_export_lock_keycmp,
3049 .hs_keycpy = ldlm_export_lock_keycpy,
3050 .hs_object = ldlm_export_lock_object,
3051 .hs_get = ldlm_export_lock_get,
3052 .hs_put = ldlm_export_lock_put,
3053 .hs_put_locked = ldlm_export_lock_put,
3056 int ldlm_init_export(struct obd_export *exp)
3062 exp->exp_lock_hash =
3063 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
3064 HASH_EXP_LOCK_CUR_BITS,
3065 HASH_EXP_LOCK_MAX_BITS,
3066 HASH_EXP_LOCK_BKT_BITS, 0,
3067 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
3068 &ldlm_export_lock_ops,
3069 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
3070 CFS_HASH_NBLK_CHANGE);
3072 if (!exp->exp_lock_hash)
3075 rc = ldlm_init_flock_export(exp);
3081 ldlm_destroy_export(exp);
3084 EXPORT_SYMBOL(ldlm_init_export);
3086 void ldlm_destroy_export(struct obd_export *exp)
3089 cfs_hash_putref(exp->exp_lock_hash);
3090 exp->exp_lock_hash = NULL;
3092 ldlm_destroy_flock_export(exp);
3095 EXPORT_SYMBOL(ldlm_destroy_export);
3097 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
3098 struct attribute *attr,
3101 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
3104 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
3105 struct attribute *attr,
3112 rc = kstrtoul(buffer, 10, &val);
3116 ldlm_cancel_unused_locks_before_replay = val;
3120 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
3122 static struct attribute *ldlm_attrs[] = {
3123 &lustre_attr_cancel_unused_locks_before_replay.attr,
3127 static struct attribute_group ldlm_attr_group = {
3128 .attrs = ldlm_attrs,
3131 static int ldlm_setup(void)
3133 static struct ptlrpc_service_conf conf;
3134 struct ldlm_bl_pool *blp = NULL;
3135 #ifdef HAVE_SERVER_SUPPORT
3136 struct task_struct *task;
3137 #endif /* HAVE_SERVER_SUPPORT */
3143 if (ldlm_state != NULL)
3146 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
3147 if (ldlm_state == NULL)
3150 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
3154 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
3158 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
3162 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
3166 rc = ldlm_debugfs_setup();
3170 memset(&conf, 0, sizeof(conf));
3171 conf = (typeof(conf)) {
3172 .psc_name = "ldlm_cbd",
3173 .psc_watchdog_factor = 2,
3175 .bc_nbufs = LDLM_CLIENT_NBUFS,
3176 .bc_buf_size = LDLM_BUFSIZE,
3177 .bc_req_max_size = LDLM_MAXREQSIZE,
3178 .bc_rep_max_size = LDLM_MAXREPSIZE,
3179 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
3180 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
3183 .tc_thr_name = "ldlm_cb",
3184 .tc_thr_factor = LDLM_THR_FACTOR,
3185 .tc_nthrs_init = LDLM_NTHRS_INIT,
3186 .tc_nthrs_base = LDLM_NTHRS_BASE,
3187 .tc_nthrs_max = LDLM_NTHRS_MAX,
3188 .tc_nthrs_user = ldlm_num_threads,
3189 .tc_cpu_bind = ldlm_cpu_bind,
3190 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
3193 .cc_pattern = ldlm_cpts,
3194 .cc_affinity = true,
3197 .so_req_handler = ldlm_callback_handler,
3200 ldlm_state->ldlm_cb_service = \
3201 ptlrpc_register_service(&conf, ldlm_svc_kset,
3202 ldlm_svc_debugfs_dir);
3203 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3204 CERROR("failed to start service\n");
3205 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3206 ldlm_state->ldlm_cb_service = NULL;
3210 #ifdef HAVE_SERVER_SUPPORT
3211 memset(&conf, 0, sizeof(conf));
3212 conf = (typeof(conf)) {
3213 .psc_name = "ldlm_canceld",
3214 .psc_watchdog_factor = 6,
3216 .bc_nbufs = LDLM_SERVER_NBUFS,
3217 .bc_buf_size = LDLM_BUFSIZE,
3218 .bc_req_max_size = LDLM_MAXREQSIZE,
3219 .bc_rep_max_size = LDLM_MAXREPSIZE,
3220 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3221 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3225 .tc_thr_name = "ldlm_cn",
3226 .tc_thr_factor = LDLM_THR_FACTOR,
3227 .tc_nthrs_init = LDLM_NTHRS_INIT,
3228 .tc_nthrs_base = LDLM_NTHRS_BASE,
3229 .tc_nthrs_max = LDLM_NTHRS_MAX,
3230 .tc_nthrs_user = ldlm_num_threads,
3231 .tc_cpu_bind = ldlm_cpu_bind,
3232 .tc_ctx_tags = LCT_MD_THREAD | \
3237 .cc_pattern = ldlm_cpts,
3238 .cc_affinity = true,
3241 .so_req_handler = ldlm_cancel_handler,
3242 .so_hpreq_handler = ldlm_hpreq_handler,
3245 ldlm_state->ldlm_cancel_service = \
3246 ptlrpc_register_service(&conf, ldlm_svc_kset,
3247 ldlm_svc_debugfs_dir);
3248 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3249 CERROR("failed to start service\n");
3250 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3251 ldlm_state->ldlm_cancel_service = NULL;
3254 #endif /* HAVE_SERVER_SUPPORT */
3256 OBD_ALLOC(blp, sizeof(*blp));
3258 GOTO(out, rc = -ENOMEM);
3259 ldlm_state->ldlm_bl_pool = blp;
3261 spin_lock_init(&blp->blp_lock);
3262 INIT_LIST_HEAD(&blp->blp_list);
3263 INIT_LIST_HEAD(&blp->blp_prio_list);
3264 init_waitqueue_head(&blp->blp_waitq);
3265 atomic_set(&blp->blp_num_threads, 0);
3266 atomic_set(&blp->blp_busy_threads, 0);
3268 if (ldlm_num_threads == 0) {
3269 blp->blp_min_threads = LDLM_NTHRS_INIT;
3270 blp->blp_max_threads = LDLM_NTHRS_MAX;
3272 blp->blp_min_threads = blp->blp_max_threads = \
3273 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3277 for (i = 0; i < blp->blp_min_threads; i++) {
3278 rc = ldlm_bl_thread_start(blp, false);
3283 #ifdef HAVE_SERVER_SUPPORT
3284 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3287 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3291 wait_event(expired_lock_wait_queue,
3292 expired_lock_thread_state == ELT_READY);
3293 #endif /* HAVE_SERVER_SUPPORT */
3295 rc = ldlm_pools_init();
3297 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3301 rc = ldlm_reclaim_setup();
3303 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3313 static int ldlm_cleanup(void)
3317 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3318 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3319 CERROR("ldlm still has namespaces; clean these up first.\n");
3320 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3321 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3325 ldlm_reclaim_cleanup();
3328 if (ldlm_state->ldlm_bl_pool != NULL) {
3329 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3331 while (atomic_read(&blp->blp_num_threads) > 0) {
3332 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3334 init_completion(&blp->blp_comp);
3336 spin_lock(&blp->blp_lock);
3337 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3338 wake_up(&blp->blp_waitq);
3339 spin_unlock(&blp->blp_lock);
3341 wait_for_completion(&blp->blp_comp);
3344 OBD_FREE(blp, sizeof(*blp));
3347 if (ldlm_state->ldlm_cb_service != NULL)
3348 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3349 #ifdef HAVE_SERVER_SUPPORT
3350 if (ldlm_state->ldlm_cancel_service != NULL)
3351 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3355 kset_unregister(ldlm_ns_kset);
3357 kset_unregister(ldlm_svc_kset);
3359 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3360 kobject_put(ldlm_kobj);
3363 ldlm_debugfs_cleanup();
3365 #ifdef HAVE_SERVER_SUPPORT
3366 if (expired_lock_thread_state != ELT_STOPPED) {
3367 expired_lock_thread_state = ELT_TERMINATE;
3368 wake_up(&expired_lock_wait_queue);
3369 wait_event(expired_lock_wait_queue,
3370 expired_lock_thread_state == ELT_STOPPED);
3374 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3382 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3383 sizeof(struct ldlm_resource), 0,
3384 SLAB_HWCACHE_ALIGN, NULL);
3385 if (ldlm_resource_slab == NULL)
3388 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3389 sizeof(struct ldlm_lock), 0,
3390 SLAB_HWCACHE_ALIGN, NULL);
3391 if (ldlm_lock_slab == NULL)
3394 ldlm_interval_slab = kmem_cache_create("interval_node",
3395 sizeof(struct ldlm_interval),
3396 0, SLAB_HWCACHE_ALIGN, NULL);
3397 if (ldlm_interval_slab == NULL)
3400 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3401 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3402 0, SLAB_HWCACHE_ALIGN, NULL);
3403 if (ldlm_interval_tree_slab == NULL)
3406 #ifdef HAVE_SERVER_SUPPORT
3407 ldlm_inodebits_slab = kmem_cache_create("ldlm_ibits_node",
3408 sizeof(struct ldlm_ibits_node),
3409 0, SLAB_HWCACHE_ALIGN, NULL);
3410 if (ldlm_inodebits_slab == NULL)
3411 goto out_interval_tree;
3413 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3414 sizeof(struct ldlm_glimpse_work),
3416 if (ldlm_glimpse_work_kmem == NULL)
3420 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3421 class_export_dump_hook = ldlm_dump_export_locks;
3424 #ifdef HAVE_SERVER_SUPPORT
3426 kmem_cache_destroy(ldlm_inodebits_slab);
3428 kmem_cache_destroy(ldlm_interval_tree_slab);
3431 kmem_cache_destroy(ldlm_interval_slab);
3433 kmem_cache_destroy(ldlm_lock_slab);
3435 kmem_cache_destroy(ldlm_resource_slab);
3440 void ldlm_exit(void)
3443 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3444 kmem_cache_destroy(ldlm_resource_slab);
3446 * ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3447 * rcu_barrier() to wait all outstanding RCU callbacks to complete,
3448 * so that ldlm_lock_free() get a chance to be called.
3451 kmem_cache_destroy(ldlm_lock_slab);
3452 kmem_cache_destroy(ldlm_interval_slab);
3453 kmem_cache_destroy(ldlm_interval_tree_slab);
3454 #ifdef HAVE_SERVER_SUPPORT
3455 kmem_cache_destroy(ldlm_inodebits_slab);
3456 kmem_cache_destroy(ldlm_glimpse_work_kmem);