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/
31 * lustre/ldlm/ldlm_lockd.c
33 * Author: Peter Braam <braam@clusterfs.com>
34 * Author: Phil Schwan <phil@clusterfs.com>
37 #define DEBUG_SUBSYSTEM S_LDLM
39 #include <linux/kthread.h>
40 #include <linux/list.h>
41 #include <libcfs/libcfs.h>
42 #include <libcfs/linux/linux-mem.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 timeout_t instead of time64_t
77 static inline timeout_t ldlm_get_rq_timeout(void)
80 timeout_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;
111 struct ldlm_bl_work_item {
112 struct list_head blwi_entry;
113 struct ldlm_namespace *blwi_ns;
114 struct ldlm_lock_desc blwi_ld;
115 struct ldlm_lock *blwi_lock;
116 struct list_head blwi_head;
118 struct completion blwi_comp;
119 enum ldlm_cancel_flags blwi_flags;
120 int blwi_mem_pressure;
123 #ifdef HAVE_SERVER_SUPPORT
126 * Protects both waiting_locks_list and expired_lock_thread.
128 static DEFINE_SPINLOCK(waiting_locks_spinlock); /* BH lock (timer) */
131 * List for contended locks.
133 * As soon as a lock is contended, it gets placed on this list and
134 * expected time to get a response is filled in the lock. A special
135 * thread walks the list looking for locks that should be released and
136 * schedules client evictions for those that have not been released in
139 * All access to it should be under waiting_locks_spinlock.
141 static LIST_HEAD(waiting_locks_list);
142 static void waiting_locks_callback(TIMER_DATA_TYPE unused);
143 static CFS_DEFINE_TIMER(waiting_locks_timer, waiting_locks_callback, 0, 0);
151 static DECLARE_WAIT_QUEUE_HEAD(expired_lock_wait_queue);
152 static enum elt_state expired_lock_thread_state = ELT_STOPPED;
153 static int expired_lock_dump;
154 static LIST_HEAD(expired_lock_list);
156 static int ldlm_lock_busy(struct ldlm_lock *lock);
157 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t timeout);
158 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t timeout);
160 static inline int have_expired_locks(void)
165 spin_lock_bh(&waiting_locks_spinlock);
166 need_to_run = !list_empty(&expired_lock_list);
167 spin_unlock_bh(&waiting_locks_spinlock);
173 * Check expired lock list for expired locks and time them out.
175 static int expired_lock_main(void *arg)
177 struct list_head *expired = &expired_lock_list;
182 expired_lock_thread_state = ELT_READY;
183 wake_up(&expired_lock_wait_queue);
186 wait_event_idle(expired_lock_wait_queue,
187 have_expired_locks() ||
188 expired_lock_thread_state == ELT_TERMINATE);
190 spin_lock_bh(&waiting_locks_spinlock);
191 if (expired_lock_dump) {
192 spin_unlock_bh(&waiting_locks_spinlock);
194 /* from waiting_locks_callback, but not in timer */
195 libcfs_debug_dumplog();
197 spin_lock_bh(&waiting_locks_spinlock);
198 expired_lock_dump = 0;
203 while (!list_empty(expired)) {
204 struct obd_export *export;
205 struct ldlm_lock *lock;
207 lock = list_first_entry(expired, struct ldlm_lock,
209 if ((void *)lock < LP_POISON + PAGE_SIZE &&
210 (void *)lock >= LP_POISON) {
211 spin_unlock_bh(&waiting_locks_spinlock);
212 CERROR("free lock on elt list %p\n", lock);
215 list_del_init(&lock->l_pending_chain);
216 if ((void *)lock->l_export <
217 LP_POISON + PAGE_SIZE &&
218 (void *)lock->l_export >= LP_POISON) {
219 CERROR("lock with free export on elt list %p\n",
221 lock->l_export = NULL;
222 LDLM_ERROR(lock, "free export");
224 * release extra ref grabbed by
225 * ldlm_add_waiting_lock() or
228 LDLM_LOCK_RELEASE(lock);
232 if (ldlm_is_destroyed(lock)) {
234 * release the lock refcount where
235 * waiting_locks_callback() founds
237 LDLM_LOCK_RELEASE(lock);
240 export = class_export_lock_get(lock->l_export, lock);
241 spin_unlock_bh(&waiting_locks_spinlock);
243 /* Check if we need to prolong timeout */
244 if (!CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
245 lock->l_callback_timestamp != 0 && /* not AST err */
246 ldlm_lock_busy(lock)) {
247 LDLM_DEBUG(lock, "prolong the busy lock");
248 lock_res_and_lock(lock);
249 ldlm_add_waiting_lock(lock,
250 ldlm_bl_timeout(lock) >> 1);
251 unlock_res_and_lock(lock);
253 spin_lock_bh(&export->exp_bl_list_lock);
254 list_del_init(&lock->l_exp_list);
255 spin_unlock_bh(&export->exp_bl_list_lock);
258 "lock callback timer expired after %llds: evicting client at %s ",
259 ktime_get_seconds() -
261 obd_export_nid2str(export));
262 ldlm_lock_to_ns(lock)->ns_timeouts++;
263 if (do_dump_on_eviction(export->exp_obd))
265 class_fail_export(export);
267 class_export_lock_put(export, lock);
269 * release extra ref grabbed by ldlm_add_waiting_lock()
270 * or ldlm_failed_ast()
272 LDLM_LOCK_RELEASE(lock);
274 spin_lock_bh(&waiting_locks_spinlock);
276 spin_unlock_bh(&waiting_locks_spinlock);
279 CERROR("dump the log upon eviction\n");
280 libcfs_debug_dumplog();
283 if (expired_lock_thread_state == ELT_TERMINATE)
287 expired_lock_thread_state = ELT_STOPPED;
288 wake_up(&expired_lock_wait_queue);
293 * Check if there is a request in the export request list
294 * which prevents the lock canceling.
296 static int ldlm_lock_busy(struct ldlm_lock *lock)
298 struct ptlrpc_request *req;
303 if (lock->l_export == NULL)
306 spin_lock(&lock->l_export->exp_rpc_lock);
307 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
309 if (req->rq_ops->hpreq_lock_match) {
310 match = req->rq_ops->hpreq_lock_match(req, lock);
315 spin_unlock(&lock->l_export->exp_rpc_lock);
319 /* This is called from within a timer interrupt and cannot schedule */
320 static void waiting_locks_callback(TIMER_DATA_TYPE unused)
322 struct ldlm_lock *lock;
325 spin_lock_bh(&waiting_locks_spinlock);
326 while (!list_empty(&waiting_locks_list)) {
327 lock = list_first_entry(&waiting_locks_list, struct ldlm_lock,
329 if (lock->l_callback_timestamp > ktime_get_seconds() ||
330 lock->l_req_mode == LCK_GROUP)
334 * no needs to take an extra ref on the lock since it was in
335 * the waiting_locks_list and ldlm_add_waiting_lock()
336 * already grabbed a ref
338 list_move(&lock->l_pending_chain, &expired_lock_list);
342 if (!list_empty(&expired_lock_list)) {
343 if (obd_dump_on_timeout && need_dump)
344 expired_lock_dump = __LINE__;
346 wake_up(&expired_lock_wait_queue);
350 * Make sure the timer will fire again if we have any locks
353 if (!list_empty(&waiting_locks_list)) {
354 time64_t now = ktime_get_seconds();
357 lock = list_first_entry(&waiting_locks_list, struct ldlm_lock,
359 if (lock->l_callback_timestamp - now > 0)
360 delta = lock->l_callback_timestamp - now;
361 mod_timer(&waiting_locks_timer,
362 jiffies + cfs_time_seconds(delta));
364 spin_unlock_bh(&waiting_locks_spinlock);
368 * Add lock to the list of contended locks.
370 * Indicate that we're waiting for a client to call us back cancelling a given
371 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
372 * timer to fire appropriately. (We round up to the next second, to avoid
373 * floods of timer firings during periods of high lock contention and traffic).
374 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
375 * if it has been added to the waiting list (1 is returned).
377 * Called with the namespace lock held.
379 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t delay)
381 unsigned long timeout_jiffies = jiffies;
385 lock->l_blast_sent = ktime_get_seconds();
386 if (!list_empty(&lock->l_pending_chain))
389 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
390 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
393 deadline = lock->l_blast_sent + delay;
394 if (likely(deadline > lock->l_callback_timestamp))
395 lock->l_callback_timestamp = deadline;
397 timeout = clamp_t(timeout_t,
398 lock->l_callback_timestamp - lock->l_blast_sent,
400 timeout_jiffies += cfs_time_seconds(timeout);
402 if (time_before(timeout_jiffies, waiting_locks_timer.expires) ||
403 !timer_pending(&waiting_locks_timer))
404 mod_timer(&waiting_locks_timer, timeout_jiffies);
407 * if the new lock has a shorter timeout than something earlier on
408 * the list, we'll wait the longer amount of time; no big deal.
411 list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
415 static void ldlm_add_blocked_lock(struct ldlm_lock *lock)
417 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
418 if (list_empty(&lock->l_exp_list)) {
419 if (!ldlm_is_granted(lock))
420 list_add_tail(&lock->l_exp_list,
421 &lock->l_export->exp_bl_list);
423 list_add(&lock->l_exp_list,
424 &lock->l_export->exp_bl_list);
426 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
429 * A blocked lock is added. Adjust the position in
430 * the stale list if the export is in the list.
431 * If export is stale and not in the list - it is being
432 * processed and will be placed on the right position
433 * on obd_stale_export_put().
435 if (!list_empty(&lock->l_export->exp_stale_list))
436 obd_stale_export_adjust(lock->l_export);
439 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t timeout)
441 struct obd_device *obd = NULL;
444 /* NB: must be called with hold of lock_res_and_lock() */
445 LASSERT(ldlm_is_res_locked(lock));
446 LASSERT(!ldlm_is_cancel_on_block(lock));
449 * Do not put cross-MDT lock in the waiting list, since we
450 * will not evict it due to timeout for now
452 if (lock->l_export != NULL) {
453 obd = lock->l_export->exp_obd;
455 if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS)
459 spin_lock_bh(&waiting_locks_spinlock);
460 if (ldlm_is_cancel(lock)) {
461 spin_unlock_bh(&waiting_locks_spinlock);
465 if (ldlm_is_destroyed(lock)) {
466 static time64_t next;
468 spin_unlock_bh(&waiting_locks_spinlock);
469 LDLM_ERROR(lock, "not waiting on destroyed lock (b=5653)");
470 if (ktime_get_seconds() > next) {
471 next = ktime_get_seconds() + 14400;
477 ldlm_set_waited(lock);
478 ret = __ldlm_add_waiting_lock(lock, timeout);
481 * grab ref on the lock if it has been added to the
486 spin_unlock_bh(&waiting_locks_spinlock);
489 ldlm_add_blocked_lock(lock);
491 at_off = obd_at_off(obd);
492 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)",
493 ret == 0 ? "not re-" : "", timeout,
494 at_off ? "off" : "on");
499 * Remove a lock from the pending list, likely because it had its cancellation
500 * callback arrive without incident. This adjusts the lock-timeout timer if
501 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
502 * As done by ldlm_del_waiting_lock(), the caller must release the lock
503 * reference when the lock is removed from any list (1 is returned).
505 * Called with namespace lock held.
507 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
509 struct list_head *list_next;
511 if (list_empty(&lock->l_pending_chain))
514 list_next = lock->l_pending_chain.next;
515 if (lock->l_pending_chain.prev == &waiting_locks_list) {
516 /* Removing the head of the list, adjust timer. */
517 if (list_next == &waiting_locks_list) {
518 /* No more, just cancel. */
519 timer_delete(&waiting_locks_timer);
521 time64_t now = ktime_get_seconds();
522 struct ldlm_lock *next;
525 next = list_entry(list_next, struct ldlm_lock,
527 if (next->l_callback_timestamp - now > 0)
528 delta = lock->l_callback_timestamp - now;
530 mod_timer(&waiting_locks_timer,
531 jiffies + cfs_time_seconds(delta));
534 list_del_init(&lock->l_pending_chain);
539 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
543 if (lock->l_export == NULL) {
544 /* We don't have a "waiting locks list" on clients. */
545 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
549 spin_lock_bh(&waiting_locks_spinlock);
550 ret = __ldlm_del_waiting_lock(lock);
551 ldlm_clear_waited(lock);
552 spin_unlock_bh(&waiting_locks_spinlock);
554 /* remove the lock out of export blocking list */
555 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
556 list_del_init(&lock->l_exp_list);
557 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
561 * release lock ref if it has indeed been removed
564 LDLM_LOCK_RELEASE(lock);
567 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
572 * Prolong the contended lock waiting time.
574 * Called with namespace lock held.
576 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, timeout_t timeout)
578 if (lock->l_export == NULL) {
579 /* We don't have a "waiting locks list" on clients. */
580 LDLM_DEBUG(lock, "client lock: no-op");
584 if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) {
585 /* We don't have a "waiting locks list" on OSP. */
586 LDLM_DEBUG(lock, "MDS-MDS lock: no-op");
590 spin_lock_bh(&waiting_locks_spinlock);
592 if (list_empty(&lock->l_pending_chain)) {
593 spin_unlock_bh(&waiting_locks_spinlock);
594 LDLM_DEBUG(lock, "wasn't waiting");
599 * we remove/add the lock to the waiting list, so no needs to
600 * release/take a lock reference
602 __ldlm_del_waiting_lock(lock);
603 __ldlm_add_waiting_lock(lock, timeout);
604 spin_unlock_bh(&waiting_locks_spinlock);
606 LDLM_DEBUG(lock, "refreshed to %ds", timeout);
609 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
611 #else /* HAVE_SERVER_SUPPORT */
613 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
618 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, timeout_t timeout)
623 #endif /* !HAVE_SERVER_SUPPORT */
625 #ifdef HAVE_SERVER_SUPPORT
628 * Calculate the per-export Blocking timeout (covering BL AST, data flush,
629 * lock cancel, and their replies). Used for lock callback timeout and AST
632 * \param[in] lock lock which is getting the blocking callback
634 * \retval timeout in seconds to wait for the client reply
636 timeout_t ldlm_bl_timeout(struct ldlm_lock *lock)
639 struct obd_device *obd = lock->l_export->exp_obd;
642 return obd_timeout / 2;
645 * Since these are non-updating timeouts, we should be conservative.
646 * Take more than usually, 150%
647 * It would be nice to have some kind of "early reply" mechanism for
648 * lock callbacks too...
650 timeout = obd_at_get(obd, &lock->l_export->exp_bl_lock_at);
651 return max_t(timeout_t, timeout + (timeout >> 1),
652 (timeout_t)obd_get_ldlm_enqueue_min(obd));
654 EXPORT_SYMBOL(ldlm_bl_timeout);
657 * Calculate the per-export Blocking timeout by the given RPC (covering the
658 * reply to this RPC and the next RPC). The next RPC could be still not CANCEL,
659 * but having the lock refresh mechanism it is enough.
661 * Used for lock refresh timeout when we are in the middle of the process -
662 * BL AST is sent, CANCEL is ahead - it is still 1 reply for the current RPC
663 * and at least 1 RPC (which will trigger another refresh if it will be not
664 * CANCEL) - but more accurate than ldlm_bl_timeout as the timeout is taken
665 * from the RPC (i.e. the view of the client on the current AT) is taken into
668 * \param[in] req req which export needs the timeout calculation
670 * \retval timeout in seconds to wait for the next client's RPC
672 timeout_t ldlm_bl_timeout_by_rpc(struct ptlrpc_request *req)
674 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
675 timeout_t timeout, req_timeout, at_timeout, netl;
676 struct obd_device *obd = req->rq_export->exp_obd;
679 return obd_timeout / 2;
681 /* A blocked lock means somebody in the cluster is waiting, and we
682 * should not consider the worst ever case, consisting of a chain of
683 * failures on each step, however this timeout should survive a
684 * recovery of at least 1 failure, let this one to be the worst one:
685 * in case a server NID is dead first re-connect is done through the
686 * same router and also times out.
688 * Either this on the next RPC times out, take the max.
689 * Considering the current RPC, take just the left time.
691 netl = obd_at_get(obd,
692 &req->rq_export->exp_imp_reverse->imp_at.iat_net_latency);
693 req_timeout = req->rq_deadline - ktime_get_real_seconds() + netl;
694 at_timeout = at_est2timeout(obd_at_get(obd, &svcpt->scp_at_estimate))
696 req_timeout = max(req_timeout, at_timeout);
698 /* Take 1 re-connect failure and 1 re-connect success into account. */
699 timeout = at_timeout + INITIAL_CONNECT_TIMEOUT + netl + req_timeout;
701 /* Client's timeout is calculated as at_est2timeout(), let's be a bit
702 * more conservative than client
704 return max(timeout + (timeout >> 4),
705 (timeout_t)obd_get_ldlm_enqueue_min(obd));
707 EXPORT_SYMBOL(ldlm_bl_timeout_by_rpc);
710 * Perform lock cleanup if AST sending failed.
712 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
713 const char *ast_type)
715 LCONSOLE_ERROR("%s: A client on nid %s was evicted due to a lock %s callback time out: rc %d\n",
716 lock->l_export->exp_obd->obd_name,
717 obd_export_nid2str(lock->l_export), ast_type, rc);
719 if (obd_dump_on_timeout)
720 libcfs_debug_dumplog();
721 spin_lock_bh(&waiting_locks_spinlock);
722 if (__ldlm_del_waiting_lock(lock) == 0)
724 * the lock was not in any list, grab an extra ref before adding
725 * the lock to the expired list
728 /* differentiate it from expired locks */
729 lock->l_callback_timestamp = 0;
730 list_add(&lock->l_pending_chain, &expired_lock_list);
731 wake_up(&expired_lock_wait_queue);
732 spin_unlock_bh(&waiting_locks_spinlock);
736 * Perform lock cleanup if AST reply came with error.
738 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
739 struct ptlrpc_request *req, int rc,
740 const char *ast_type)
742 struct lnet_processid *peer = &req->rq_import->imp_connection->c_peer;
744 if (!req->rq_replied || (rc && rc != -EINVAL)) {
745 if (ldlm_is_cancel(lock)) {
747 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
748 ast_type, req, req->rq_xid,
749 libcfs_nidstr(&peer->nid));
750 ldlm_lock_cancel(lock);
752 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
754 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
756 * Upon umount process the AST fails because cannot be
757 * sent. This shouldn't lead to the client eviction.
758 * -ENODEV error is returned by ptl_send_rpc() for
759 * new request in such import.
760 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
761 * if imp_invalid is set or obd_no_recov.
762 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
763 * in ptlrpc_import_delay_req() as well with -EIO code.
764 * In all such cases errors are ignored.
767 "%s AST can't be sent due to a server %s failure or umount process: rc = %d",
769 req->rq_import->imp_obd->obd_name, rc);
772 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
773 libcfs_nidstr(&peer->nid),
774 req->rq_replied ? "returned error from" :
775 "failed to reply to",
776 ast_type, req, req->rq_xid,
777 (req->rq_repmsg != NULL) ?
778 lustre_msg_get_status(req->rq_repmsg) : 0,
780 ldlm_failed_ast(lock, rc, ast_type);
786 struct ldlm_resource *res = lock->l_resource;
789 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
790 libcfs_nidstr(&peer->nid),
792 lustre_msg_get_status(req->rq_repmsg) : -1,
793 ast_type, req, req->rq_xid);
796 * update lvbo to return proper attributes.
799 ldlm_resource_getref(res);
800 ldlm_lvbo_update(res, lock, NULL, 1);
801 ldlm_resource_putref(res);
803 ldlm_lock_cancel(lock);
810 static int ldlm_cb_interpret(const struct lu_env *env,
811 struct ptlrpc_request *req, void *args, int rc)
813 struct ldlm_cb_async_args *ca = args;
814 struct ldlm_lock *lock = ca->ca_lock;
815 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
819 LASSERT(lock != NULL);
822 case LDLM_GL_CALLBACK:
824 * Update the LVB from disk if the AST failed
825 * (this is a legal race)
827 * - Glimpse callback of local lock just returns
828 * -ELDLM_NO_LOCK_DATA.
829 * - Glimpse callback of remote lock might return
830 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
832 if (unlikely(arg->gl_interpret_reply)) {
833 rc = arg->gl_interpret_reply(NULL, req, args, rc);
834 } else if (rc == -ELDLM_NO_LOCK_DATA) {
836 "lost race - client has a lock but no inode");
837 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
838 } else if (rc != 0) {
839 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
841 rc = ldlm_lvbo_update(lock->l_resource,
845 case LDLM_BL_CALLBACK:
847 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
849 case LDLM_CP_CALLBACK:
851 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
854 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
859 /* release extra reference taken in ldlm_ast_fini() */
860 LDLM_LOCK_RELEASE(lock);
863 atomic_inc(&arg->restart);
868 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
870 struct ldlm_cb_async_args *ca = data;
871 struct ldlm_lock *lock = ca->ca_lock;
873 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
876 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
877 struct ldlm_cb_set_arg *arg,
878 struct ldlm_lock *lock,
885 if (unlikely(instant_cancel)) {
886 rc = ptl_send_rpc(req, 1);
889 atomic_inc(&arg->restart);
892 ptlrpc_set_add_req(arg->set, req);
899 * Check if there are requests in the export request list which prevent
900 * the lock canceling and make these requests high priority ones.
902 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
904 struct ptlrpc_request *req;
908 if (lock->l_export == NULL) {
909 LDLM_DEBUG(lock, "client lock: no-op");
913 spin_lock(&lock->l_export->exp_rpc_lock);
914 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
917 * Do not process requests that were not yet added to there
918 * incoming queue or were already removed from there for
919 * processing. We evaluate ptlrpc_nrs_req_can_move() without
920 * holding svcpt->scp_req_lock, and then redo the check with
921 * the lock held once we need to obtain a reliable result.
923 if (ptlrpc_nrs_req_can_move(req) &&
924 req->rq_ops->hpreq_lock_match &&
925 req->rq_ops->hpreq_lock_match(req, lock))
926 ptlrpc_nrs_req_hp_move(req);
928 spin_unlock(&lock->l_export->exp_rpc_lock);
933 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
934 * enqueued server lock conflicts with given one.
936 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
937 * to wait for client response.
939 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
940 struct ldlm_lock_desc *desc,
941 void *data, int flag)
943 struct ldlm_cb_async_args *ca;
944 struct ldlm_cb_set_arg *arg = data;
945 struct ldlm_request *body;
946 struct ptlrpc_request *req;
947 int instant_cancel = 0;
949 struct obd_device *obd;
953 if (flag == LDLM_CB_CANCELING)
954 /* Don't need to do anything here. */
957 if (CFS_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
958 LDLM_DEBUG(lock, "dropping BL AST");
963 LASSERT(data != NULL);
965 obd = lock->l_export->exp_obd;
966 if (obd->obd_recovering != 0)
967 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
969 ldlm_lock_reorder_req(lock);
971 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
972 &RQF_LDLM_BL_CALLBACK,
973 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
975 RETURN(PTR_ERR(req));
977 ca = ptlrpc_req_async_args(ca, req);
978 ca->ca_set_arg = arg;
981 req->rq_interpret_reply = ldlm_cb_interpret;
983 lock_res_and_lock(lock);
984 if (ldlm_is_destroyed(lock)) {
985 /* What's the point? */
986 unlock_res_and_lock(lock);
991 if (!ldlm_is_granted(lock)) {
993 * this blocking AST will be communicated as part of the
994 * completion AST instead
996 ldlm_add_blocked_lock(lock);
997 ldlm_set_waited(lock);
998 unlock_res_and_lock(lock);
1000 ptlrpc_req_put(req);
1001 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
1005 if (ldlm_is_cancel_on_block(lock))
1008 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1009 body->lock_handle[0] = lock->l_remote_handle;
1010 body->lock_handle[1].cookie = lock->l_handle.h_cookie;
1011 body->lock_desc = *desc;
1012 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1015 LDLM_DEBUG(lock, "server preparing blocking AST");
1017 ptlrpc_request_set_replen(req);
1018 ldlm_set_cbpending(lock);
1019 if (instant_cancel) {
1020 unlock_res_and_lock(lock);
1021 ldlm_lock_cancel(lock);
1023 req->rq_no_resend = 1;
1025 LASSERT(ldlm_is_granted(lock));
1026 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1027 unlock_res_and_lock(lock);
1029 /* Do not resend after lock callback timeout */
1030 req->rq_delay_limit = ldlm_bl_timeout(lock);
1031 req->rq_resend_cb = ldlm_update_resend;
1034 req->rq_send_state = LUSTRE_IMP_FULL;
1035 /* ptlrpc_request_alloc_pack already set timeout */
1036 if (obd_at_off(obd))
1037 req->rq_timeout = ldlm_get_rq_timeout();
1039 if (lock->l_export && lock->l_export->exp_nid_stats &&
1040 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1041 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1042 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
1044 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1050 * ->l_completion_ast callback for a remote lock in server namespace.
1052 * Sends AST to the client notifying it of lock granting. If initial
1053 * lock response was not sent yet, instead of sending another RPC, just
1054 * mark the lock as granted and client will understand
1056 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
1058 struct ldlm_cb_set_arg *arg = data;
1059 struct ldlm_request *body;
1060 struct ptlrpc_request *req;
1061 struct ldlm_cb_async_args *ca;
1062 int instant_cancel = 0;
1065 struct obd_device *obd;
1069 LASSERT(lock != NULL);
1070 LASSERT(data != NULL);
1072 if (CFS_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
1073 LDLM_DEBUG(lock, "dropping CP AST");
1077 obd = lock->l_export->exp_obd;
1078 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
1079 &RQF_LDLM_CP_CALLBACK);
1083 /* server namespace, doesn't need lock */
1084 lvb_len = ldlm_lvbo_size(lock);
1086 * LU-3124 & LU-2187: to not return layout in completion AST because
1087 * it may deadlock for LU-2187, or client may not have enough space
1088 * for large layout. The layout will be returned to client with an
1089 * extra RPC to fetch xattr.lov
1091 if (ldlm_has_layout(lock))
1094 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
1095 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
1097 ptlrpc_request_free(req);
1101 ca = ptlrpc_req_async_args(ca, req);
1102 ca->ca_set_arg = arg;
1105 req->rq_interpret_reply = ldlm_cb_interpret;
1106 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1108 body->lock_handle[0] = lock->l_remote_handle;
1109 body->lock_handle[1].cookie = lock->l_handle.h_cookie;
1110 body->lock_flags = ldlm_flags_to_wire(flags);
1111 ldlm_lock2desc(lock, &body->lock_desc);
1113 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1115 lvb_len = ldlm_lvbo_fill(lock, lvb, &lvb_len);
1118 * We still need to send the RPC to wake up the blocked
1119 * enqueue thread on the client.
1121 * Consider old client, there is no better way to notify
1122 * the failure, just zero-sized the LVB, then the client
1123 * will fail out as "-EPROTO".
1125 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1129 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1134 LDLM_DEBUG(lock, "server preparing completion AST");
1136 ptlrpc_request_set_replen(req);
1138 req->rq_send_state = LUSTRE_IMP_FULL;
1139 /* ptlrpc_request_pack already set timeout */
1140 if (obd_at_off(obd))
1141 req->rq_timeout = ldlm_get_rq_timeout();
1143 /* We only send real blocking ASTs after the lock is granted */
1144 lock_res_and_lock(lock);
1145 if (ldlm_is_ast_sent(lock)) {
1146 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1147 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1148 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1152 * We might get here prior to ldlm_handle_enqueue setting
1153 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1154 * into waiting list, but this is safe and similar code in
1155 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1156 * that would not only cancel the lock, but will also remove
1157 * it from waiting list
1159 if (ldlm_is_cancel_on_block(lock)) {
1160 unlock_res_and_lock(lock);
1161 ldlm_lock_cancel(lock);
1164 req->rq_no_resend = 1;
1166 lock_res_and_lock(lock);
1168 /* start the lock-timeout clock */
1169 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1170 /* Do not resend after lock callback timeout */
1171 req->rq_delay_limit = ldlm_bl_timeout(lock);
1172 req->rq_resend_cb = ldlm_update_resend;
1175 unlock_res_and_lock(lock);
1177 if (lock->l_export && lock->l_export->exp_nid_stats &&
1178 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1179 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1180 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1182 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1184 RETURN(lvb_len < 0 ? lvb_len : rc);
1186 EXPORT_SYMBOL(ldlm_server_completion_ast);
1189 * Server side ->l_glimpse_ast handler for client locks.
1191 * Sends glimpse AST to the client and waits for reply. Then updates
1192 * lvbo with the result.
1194 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1196 struct ldlm_cb_set_arg *arg = data;
1197 struct ldlm_request *body;
1198 struct ptlrpc_request *req;
1199 struct ldlm_cb_async_args *ca;
1201 struct req_format *req_fmt;
1202 struct obd_device *obd = lock->l_export->exp_obd;
1206 LASSERT(lock != NULL);
1208 if (arg->gl_desc != NULL)
1209 /* There is a glimpse descriptor to pack */
1210 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1212 req_fmt = &RQF_LDLM_GL_CALLBACK;
1214 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1215 req_fmt, LUSTRE_DLM_VERSION,
1219 RETURN(PTR_ERR(req));
1221 if (arg->gl_desc != NULL) {
1222 /* copy the GL descriptor */
1223 union ldlm_gl_desc *desc;
1225 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1226 *desc = *arg->gl_desc;
1229 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1230 body->lock_handle[0] = lock->l_remote_handle;
1231 ldlm_lock2desc(lock, &body->lock_desc);
1233 ca = ptlrpc_req_async_args(ca, req);
1234 ca->ca_set_arg = arg;
1237 /* server namespace, doesn't need lock */
1238 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1239 ldlm_lvbo_size(lock));
1240 ptlrpc_request_set_replen(req);
1242 req->rq_send_state = LUSTRE_IMP_FULL;
1243 /* ptlrpc_request_alloc_pack already set timeout */
1244 if (obd_at_off(obd))
1245 req->rq_timeout = ldlm_get_rq_timeout();
1247 req->rq_interpret_reply = ldlm_cb_interpret;
1249 if (lock->l_export && lock->l_export->exp_nid_stats) {
1250 struct nid_stat *nid_stats = lock->l_export->exp_nid_stats;
1252 lprocfs_counter_incr(nid_stats->nid_ldlm_stats,
1253 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1256 rc = ldlm_ast_fini(req, arg, lock, 0);
1260 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1262 int ldlm_glimpse_locks(struct ldlm_resource *res,
1263 struct list_head *gl_work_list)
1269 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1271 if (rc == -ERESTART)
1272 ldlm_reprocess_all(res, 0);
1276 EXPORT_SYMBOL(ldlm_glimpse_locks);
1278 /* return LDLM lock associated with a lock callback request */
1279 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1281 struct ldlm_cb_async_args *ca;
1282 struct ldlm_lock *lock;
1286 ca = ptlrpc_req_async_args(ca, req);
1289 RETURN(ERR_PTR(-EFAULT));
1293 EXPORT_SYMBOL(ldlm_request_lock);
1296 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1297 * service threads to carry out client lock enqueueing requests.
1299 int ldlm_handle_enqueue(struct ldlm_namespace *ns,
1300 struct req_capsule *pill,
1301 const struct ldlm_request *dlm_req,
1302 const struct ldlm_callback_suite *cbs)
1304 struct ldlm_reply *dlm_rep;
1306 enum ldlm_error err = ELDLM_OK;
1307 struct ldlm_lock *lock = NULL;
1308 void *cookie = NULL;
1310 struct ldlm_resource *res = NULL;
1311 struct ptlrpc_request *req = pill->rc_req;
1312 const struct lu_env *env = req->rq_svc_thread->t_env;
1316 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1318 LASSERT(req && req->rq_export);
1320 if (req_capsule_ptlreq(pill))
1321 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF,
1324 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1326 /* for intent enqueue the stat will be updated inside intent policy */
1327 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1328 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1329 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1331 if (req->rq_export->exp_nid_stats &&
1332 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1333 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1334 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1336 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1337 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1338 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1339 dlm_req->lock_desc.l_resource.lr_type);
1340 GOTO(out, rc = -EFAULT);
1343 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1344 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1345 dlm_req->lock_desc.l_req_mode &
1346 (dlm_req->lock_desc.l_req_mode-1))) {
1347 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1348 dlm_req->lock_desc.l_req_mode);
1349 GOTO(out, rc = -EFAULT);
1352 if (unlikely((flags & LDLM_FL_REPLAY) ||
1353 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1354 /* Find an existing lock in the per-export lock hash */
1356 * In the function below, .hs_keycmp resolves to
1357 * ldlm_export_lock_keycmp()
1359 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1360 (void *)&dlm_req->lock_handle[0]);
1362 DEBUG_REQ(D_DLMTRACE, req,
1363 "found existing lock cookie %#llx",
1364 lock->l_handle.h_cookie);
1365 flags |= LDLM_FL_RESENT;
1366 GOTO(existing_lock, rc = 0);
1369 if (ldlm_reclaim_full()) {
1370 DEBUG_REQ(D_WARNING | D_RPCTRACE, req,
1371 "Too many granted locks, reject current enqueue request and let the client retry later");
1372 GOTO(out, rc = -EINPROGRESS);
1376 /* The lock's callback data might be set in the policy function */
1377 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1378 dlm_req->lock_desc.l_resource.lr_type,
1379 dlm_req->lock_desc.l_req_mode,
1380 cbs, NULL, 0, LVB_T_NONE);
1387 lock->l_remote_handle = dlm_req->lock_handle[0];
1388 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1391 * Initialize resource lvb but not for a lock being replayed since
1392 * Client already got lvb sent in this case.
1393 * This must occur early since some policy methods assume resource
1394 * lvb is available (lr_lvb_data != NULL).
1396 res = lock->l_resource;
1397 if (!(flags & LDLM_FL_REPLAY)) {
1398 /* non-replayed lock, delayed lvb init may need to be done */
1399 rc = ldlm_lvbo_init(res);
1401 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1406 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1408 * Don't enqueue a lock onto the export if it is been disonnected
1409 * due to eviction (b=3822) or server umount (b=24324).
1410 * Cancel it now instead.
1412 if (req->rq_export->exp_disconnected) {
1413 LDLM_ERROR(lock, "lock on disconnected export %p",
1415 GOTO(out, rc = -ENOTCONN);
1418 lock->l_export = class_export_lock_get(req->rq_export, lock);
1419 if (lock->l_export->exp_lock_hash)
1420 cfs_hash_add(lock->l_export->exp_lock_hash,
1421 &lock->l_remote_handle,
1425 * Inherit the enqueue flags before the operation, because we do not
1426 * keep the res lock on return and next operations (BL AST) may proceed
1429 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1430 LDLM_FL_INHERIT_MASK);
1432 ldlm_convert_policy_to_local(req->rq_export,
1433 dlm_req->lock_desc.l_resource.lr_type,
1434 &dlm_req->lock_desc.l_policy_data,
1435 &lock->l_policy_data);
1436 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT) {
1437 lock->l_req_extent = lock->l_policy_data.l_extent;
1438 } else if (dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) {
1439 lock->l_policy_data.l_inodebits.try_bits =
1440 dlm_req->lock_desc.l_policy_data.l_inodebits.try_bits;
1441 lock->l_policy_data.l_inodebits.li_gid =
1442 dlm_req->lock_desc.l_policy_data.l_inodebits.li_gid;
1447 if (!(flags & LDLM_FL_HAS_INTENT)) {
1448 /* based on the assumption that lvb size never changes during
1449 * resource life time otherwise it need resource->lr_lock's
1452 req_capsule_set_size(pill, &RMF_DLM_LVB,
1453 RCL_SERVER, ldlm_lvbo_size(lock));
1455 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1456 GOTO(out, rc = -ENOMEM);
1458 rc = req_capsule_server_pack(pill);
1463 err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1470 dlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
1472 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1473 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1475 if (lock->l_resource->lr_type == LDLM_EXTENT)
1476 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1479 * We never send a blocking AST until the lock is granted, but
1480 * we can tell it right now
1482 lock_res_and_lock(lock);
1485 * Now take into account flags to be inherited from original lock
1486 * request both in reply to client and in our own lock flags.
1488 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1489 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1492 * Don't move a pending lock onto the export if it has already been
1493 * disconnected due to eviction (b=5683) or server umount (b=24324).
1494 * Cancel it now instead.
1496 if (unlikely(req->rq_export->exp_disconnected ||
1497 CFS_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1498 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1500 } else if (ldlm_is_ast_sent(lock)) {
1501 /* fill lock desc for possible lock convert */
1502 if (lock->l_blocking_lock &&
1503 lock->l_resource->lr_type == LDLM_IBITS) {
1504 struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1505 struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1508 "save blocking bits %llx in granted lock",
1509 bl_lock->l_policy_data.l_inodebits.bits);
1511 * If lock is blocked then save blocking ibits
1512 * in returned lock policy for the possible lock
1513 * convert on a client.
1515 rep_desc->l_policy_data.l_inodebits.cancel_bits =
1516 bl_lock->l_policy_data.l_inodebits.bits;
1518 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1519 if (ldlm_is_granted(lock)) {
1521 * Only cancel lock if it was granted, because it would
1522 * be destroyed immediately and would never be granted
1523 * in the future, causing timeouts on client. Not
1524 * granted lock will be cancelled immediately after
1525 * sending completion AST.
1527 if (ldlm_is_cancel_on_block(lock)) {
1528 unlock_res_and_lock(lock);
1529 ldlm_lock_cancel(lock);
1530 lock_res_and_lock(lock);
1532 ldlm_add_waiting_lock(lock,
1533 ldlm_bl_timeout(lock));
1537 unlock_res_and_lock(lock);
1541 if (req_capsule_ptlreq(pill)) {
1542 req->rq_status = rc ?: err; /* return either error - b=11190 */
1543 if (!req->rq_packed_final) {
1544 int rc1 = lustre_pack_reply(req, 1, NULL, NULL);
1552 * The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1553 * ldlm_reprocess_all. If this moves, revisit that code. -phil
1557 "server-side enqueue handler, sending reply (err=%d, rc=%d)",
1561 req_capsule_has_field(pill, &RMF_DLM_LVB,
1563 ldlm_lvbo_size(lock) > 0) {
1568 buf = req_capsule_server_get(pill, &RMF_DLM_LVB);
1569 LASSERTF(buf != NULL, "req %px, lock %px\n", req, lock);
1570 buflen = req_capsule_get_size(pill, &RMF_DLM_LVB,
1573 * non-replayed lock, delayed lvb init may
1574 * need to be occur now
1576 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1579 rc2 = ldlm_lvbo_fill(lock, buf, &buflen);
1581 req_capsule_shrink(pill, &RMF_DLM_LVB,
1583 } else if (rc2 == -ERANGE) {
1584 rc2 = req_capsule_server_grow(
1591 * if we can't grow the buffer,
1592 * it's ok to return empty lvb
1596 pill, &RMF_DLM_LVB, 0,
1602 } else if (flags & LDLM_FL_REPLAY) {
1603 /* no LVB resend upon replay */
1605 req_capsule_shrink(pill, &RMF_DLM_LVB,
1614 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1615 if (lock->l_export) {
1616 ldlm_lock_cancel(lock);
1618 lock_res_and_lock(lock);
1619 ldlm_resource_unlink_lock(lock);
1620 ldlm_lock_destroy_nolock(lock);
1621 unlock_res_and_lock(lock);
1623 ldlm_reprocess_all(lock->l_resource,
1624 lock->l_policy_data.l_inodebits.bits);
1627 if (!err && !ldlm_is_cbpending(lock) &&
1628 dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1629 ldlm_reprocess_all(lock->l_resource,
1630 lock->l_policy_data.l_inodebits.bits);
1632 LDLM_LOCK_RELEASE(lock);
1635 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1640 EXPORT_SYMBOL(ldlm_handle_enqueue);
1643 * Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1644 * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1646 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1648 if (lock->l_blocking_lock) {
1649 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1650 lock->l_blocking_lock = NULL;
1654 /* A lock can be converted to new ibits or mode and should be considered
1655 * as new lock. Clear all states related to a previous blocking AST
1656 * processing so new conflicts will cause new blocking ASTs.
1658 * This is used during lock convert below and lock downgrade to COS mode in
1659 * ldlm_lock_mode_downgrade().
1661 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1663 ldlm_clear_ast_sent(lock);
1664 lock->l_bl_ast_run = 0;
1665 ldlm_clear_blocking_lock(lock);
1668 /* Main LDLM entry point for server code to process lock conversion requests */
1669 int ldlm_handle_convert0(struct ptlrpc_request *req,
1670 const struct ldlm_request *dlm_req)
1672 struct obd_export *exp = req->rq_export;
1673 struct ldlm_reply *dlm_rep;
1674 struct ldlm_lock *lock;
1681 if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1682 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1683 LDLM_CONVERT - LDLM_FIRST_OPC);
1685 rc = req_capsule_server_pack(&req->rq_pill);
1689 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1690 dlm_rep->lock_flags = dlm_req->lock_flags;
1692 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1694 LDLM_DEBUG_NOLOCK("server lock is canceled already");
1695 req->rq_status = ELDLM_NO_LOCK_DATA;
1699 LDLM_DEBUG(lock, "server-side convert handler START");
1701 lock_res_and_lock(lock);
1702 bits = lock->l_policy_data.l_inodebits.bits;
1703 new_bits = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1705 if (ldlm_is_cancel(lock)) {
1706 LDLM_DEBUG(lock, "convert on canceled lock!");
1707 unlock_res_and_lock(lock);
1708 GOTO(out_put, rc = ELDLM_NO_LOCK_DATA);
1711 if (dlm_req->lock_desc.l_req_mode != lock->l_granted_mode) {
1712 LDLM_ERROR(lock, "lock mode differs!");
1713 unlock_res_and_lock(lock);
1714 GOTO(out_put, rc = -EPROTO);
1717 if (bits == new_bits) {
1719 * This can be valid situation if CONVERT RPCs are
1720 * re-ordered. Just finish silently
1722 LDLM_DEBUG(lock, "lock is converted already!");
1723 unlock_res_and_lock(lock);
1725 if (ldlm_is_waited(lock))
1726 ldlm_del_waiting_lock(lock);
1728 ldlm_clear_cbpending(lock);
1729 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1730 ldlm_inodebits_drop(lock, bits & ~new_bits);
1732 ldlm_clear_blocking_data(lock);
1733 unlock_res_and_lock(lock);
1735 /* All old bits should be reprocessed to send new BL AST if
1736 * it wasn't sent earlier due to LDLM_FL_AST_SENT bit set.
1738 ldlm_reprocess_all(lock->l_resource, bits);
1741 dlm_rep->lock_handle = lock->l_remote_handle;
1742 ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1743 &dlm_rep->lock_desc.l_policy_data);
1747 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d", rc);
1748 LDLM_LOCK_PUT(lock);
1749 req->rq_status = rc;
1754 * Cancel all the locks whose handles are packed into ldlm_request
1756 * Called by server code expecting such combined cancel activity
1759 int ldlm_request_cancel(struct ptlrpc_request *req,
1760 const struct ldlm_request *dlm_req,
1761 int first, enum lustre_at_flags flags)
1763 struct ldlm_resource *res, *pres = NULL;
1764 struct ldlm_lock *lock;
1765 int i, count, done = 0;
1770 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
1771 if (size <= sizeof(*dlm_req) || dlm_req->lock_count >
1772 (size - sizeof(*dlm_req)) / sizeof(struct lustre_handle))
1775 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1779 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1783 * There is no lock on the server at the replay time,
1784 * skip lock cancelling to make replay tests to pass.
1786 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1789 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, starting at %d",
1792 for (i = first; i < count; i++) {
1793 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1795 /* below message checked in replay-single.sh test_36 */
1796 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (cookie %llu)",
1797 dlm_req->lock_handle[i].cookie);
1801 res = lock->l_resource;
1805 * This code is an optimization to only attempt lock
1806 * granting on the resource (that could be CPU-expensive)
1807 * after we are done cancelling lock in that resource.
1811 ldlm_reprocess_all(pres, 0);
1812 ldlm_resource_putref(pres);
1815 ldlm_resource_getref(res);
1817 if (!ldlm_is_discard_data(lock))
1818 ldlm_lvbo_update(res, lock,
1824 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1825 lock->l_blast_sent != 0) {
1826 timeout_t delay = 0;
1828 if (ktime_get_seconds() > lock->l_blast_sent)
1829 delay = ktime_get_seconds() -
1832 "server cancels blocked lock after %ds",
1834 obd_at_measure(lock->l_export->exp_obd,
1835 &lock->l_export->exp_bl_lock_at,
1838 ldlm_lock_cancel(lock);
1839 LDLM_LOCK_PUT(lock);
1842 ldlm_reprocess_all(pres, 0);
1843 ldlm_resource_putref(pres);
1845 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1848 EXPORT_SYMBOL(ldlm_request_cancel);
1851 * Main LDLM entry point for server code to cancel locks.
1853 * Typically gets called from service handler on LDLM_CANCEL opc.
1855 int ldlm_handle_cancel(struct ptlrpc_request *req)
1857 struct ldlm_request *dlm_req;
1862 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1863 if (dlm_req == NULL) {
1864 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1868 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) <
1869 offsetof(struct ldlm_request, lock_handle[1]))
1872 if (req->rq_export && req->rq_export->exp_nid_stats &&
1873 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1874 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1875 LDLM_CANCEL - LDLM_FIRST_OPC);
1877 rc = req_capsule_server_pack(&req->rq_pill);
1881 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1882 req->rq_status = LUSTRE_ESTALE;
1884 RETURN(ptlrpc_reply(req));
1886 #endif /* HAVE_SERVER_SUPPORT */
1889 * Server may pass additional information about blocking lock.
1890 * For IBITS locks it is conflicting bits which can be used for
1891 * lock convert instead of cancel.
1893 void ldlm_bl_desc2lock(const struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1895 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
1897 check_res_locked(lock->l_resource);
1898 if (ns_is_client(ns) && ld &&
1899 (lock->l_resource->lr_type == LDLM_IBITS)) {
1901 * Lock description contains policy of blocking lock,
1902 * and its cancel_bits is used to pass conflicting bits.
1903 * NOTE: ld can be NULL or can be not NULL but zeroed if
1904 * passed from ldlm_bl_thread_blwi(), check below used bits
1905 * in ld to make sure it is valid description.
1907 * If server may replace lock resource keeping the same cookie,
1908 * never use cancel bits from different resource, full cancel
1911 if (ld->l_policy_data.l_inodebits.cancel_bits &&
1912 ldlm_res_eq(&ld->l_resource.lr_name,
1913 &lock->l_resource->lr_name) &&
1914 !(ldlm_is_cbpending(lock) &&
1915 lock->l_policy_data.l_inodebits.cancel_bits == 0)) {
1916 /* always combine conflicting ibits */
1917 lock->l_policy_data.l_inodebits.cancel_bits |=
1918 ld->l_policy_data.l_inodebits.cancel_bits;
1920 /* If cancel_bits are not obtained or
1921 * if the lock is already CBPENDING and
1922 * has no cancel_bits set
1923 * - the full lock is to be cancelled
1925 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1931 * Callback handler for receiving incoming blocking ASTs.
1933 * This can only happen on client side.
1935 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1936 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1942 LDLM_DEBUG(lock, "client blocking AST callback handler");
1944 lock_res_and_lock(lock);
1946 /* get extra information from desc if any */
1947 ldlm_bl_desc2lock(ld, lock);
1948 ldlm_set_cbpending(lock);
1950 do_ast = (!lock->l_readers && !lock->l_writers);
1951 unlock_res_and_lock(lock);
1955 "Lock %p already unused, calling callback (%p)\n",
1956 lock, lock->l_blocking_ast);
1957 if (lock->l_blocking_ast != NULL)
1958 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1962 "Lock %p is referenced, will be cancelled later\n",
1966 LDLM_DEBUG(lock, "client blocking callback handler END");
1967 LDLM_LOCK_RELEASE(lock);
1971 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1973 if (req->rq_no_reply)
1976 req->rq_status = rc;
1977 if (!req->rq_packed_final) {
1978 rc = lustre_pack_reply(req, 1, NULL, NULL);
1982 return ptlrpc_reply(req);
1986 * Callback handler for receiving incoming completion ASTs.
1988 * This only can happen on client side.
1990 static int ldlm_handle_cp_callback(struct ptlrpc_request *req,
1991 struct ldlm_namespace *ns,
1992 struct ldlm_request *dlm_req,
1993 struct ldlm_lock *lock)
1995 LIST_HEAD(ast_list);
2001 LDLM_DEBUG(lock, "client completion callback handler START");
2003 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
2004 long to = cfs_time_seconds(1);
2006 ldlm_callback_reply(req, 0);
2009 to = schedule_timeout_interruptible(to);
2010 if (ldlm_is_granted(lock) ||
2011 ldlm_is_destroyed(lock))
2016 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
2018 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
2019 GOTO(out, rc = lvb_len);
2020 } else if (lvb_len > 0) {
2021 if (lock->l_lvb_len > 0) {
2022 /* for extent lock, lvb contains ost_lvb{}. */
2023 LASSERT(lock->l_lvb_data != NULL);
2025 if (unlikely(lock->l_lvb_len < lvb_len)) {
2027 "Replied LVB is larger than expectation, expected = %d, replied = %d",
2028 lock->l_lvb_len, lvb_len);
2029 GOTO(out, rc = -EINVAL);
2034 lock_res_and_lock(lock);
2036 if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
2037 &lock->l_resource->lr_name)) {
2038 ldlm_resource_unlink_lock(lock);
2039 unlock_res_and_lock(lock);
2040 rc = ldlm_lock_change_resource(ns, lock,
2041 &dlm_req->lock_desc.l_resource.lr_name);
2043 LDLM_ERROR(lock, "Failed to allocate resource");
2046 LDLM_DEBUG(lock, "completion AST, new resource");
2047 lock_res_and_lock(lock);
2050 if (ldlm_is_failed(lock)) {
2051 unlock_res_and_lock(lock);
2052 LDLM_LOCK_RELEASE(lock);
2056 if (ldlm_is_destroyed(lock) ||
2057 ldlm_is_granted(lock)) {
2058 /* b=11300: the lock has already been granted */
2059 unlock_res_and_lock(lock);
2060 LDLM_DEBUG(lock, "Double grant race happened");
2065 * If we receive the completion AST before the actual enqueue returned,
2066 * then we might need to switch lock modes, resources, or extents.
2068 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
2069 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
2070 LDLM_DEBUG(lock, "completion AST, new lock mode");
2073 if (lock->l_resource->lr_type != LDLM_PLAIN) {
2074 ldlm_convert_policy_to_local(req->rq_export,
2075 dlm_req->lock_desc.l_resource.lr_type,
2076 &dlm_req->lock_desc.l_policy_data,
2077 &lock->l_policy_data);
2078 LDLM_DEBUG(lock, "completion AST, new policy data");
2081 ldlm_resource_unlink_lock(lock);
2083 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
2085 * BL_AST locks are not needed in LRU.
2086 * Let ldlm_cancel_lru() be fast.
2088 ldlm_lock_remove_from_lru(lock);
2089 ldlm_bl_desc2lock(&dlm_req->lock_desc, lock);
2090 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
2091 LDLM_DEBUG(lock, "completion AST includes blocking AST");
2094 if (lock->l_lvb_len > 0) {
2095 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
2096 lock->l_lvb_data, lvb_len);
2098 unlock_res_and_lock(lock);
2103 ldlm_grant_lock(lock, &ast_list);
2104 unlock_res_and_lock(lock);
2106 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
2109 * Let Enqueue to call osc_lock_upcall() and initialize
2112 CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
2114 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
2116 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
2122 lock_res_and_lock(lock);
2123 ldlm_set_failed(lock);
2124 unlock_res_and_lock(lock);
2125 wake_up(&lock->l_waitq);
2127 LDLM_LOCK_RELEASE(lock);
2133 * Callback handler for receiving incoming glimpse ASTs.
2135 * This only can happen on client side. After handling the glimpse AST
2136 * we also consider dropping the lock here if it is unused locally for a
2139 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
2140 struct ldlm_namespace *ns,
2141 struct ldlm_request *dlm_req,
2142 struct ldlm_lock *lock)
2144 struct ldlm_lock_desc *ld = &dlm_req->lock_desc;
2149 LDLM_DEBUG(lock, "client glimpse AST callback handler");
2151 if (lock->l_glimpse_ast != NULL)
2152 rc = lock->l_glimpse_ast(lock, req);
2154 if (req->rq_repmsg != NULL) {
2157 req->rq_status = rc;
2161 lock_res_and_lock(lock);
2162 if (lock->l_granted_mode == LCK_PW &&
2163 !lock->l_readers && !lock->l_writers &&
2164 ktime_after(ktime_get(),
2165 ktime_add(lock->l_last_used, ns->ns_dirty_age_limit))) {
2166 unlock_res_and_lock(lock);
2168 /* For MDS glimpse it is always DOM lock, set corresponding
2169 * cancel_bits to perform lock convert if needed
2171 if (lock->l_resource->lr_type == LDLM_IBITS)
2172 ld->l_policy_data.l_inodebits.cancel_bits =
2174 if (ldlm_bl_to_thread_lock(ns, ld, lock))
2175 ldlm_handle_bl_callback(ns, ld, lock);
2180 unlock_res_and_lock(lock);
2181 LDLM_LOCK_RELEASE(lock);
2185 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
2186 enum ldlm_cancel_flags cancel_flags)
2188 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2189 char *prio = "regular";
2194 spin_lock(&blp->blp_lock);
2195 /* cannot access blwi after added to list and lock is dropped */
2196 count = blwi->blwi_lock ? 1 : blwi->blwi_count;
2198 /* if the server is waiting on a lock to be cancelled (bl_ast), this is
2199 * an urgent request and should go in the priority queue so it doesn't
2200 * get stuck behind non-priority work (eg, lru size management)
2202 * We also prioritize discard_data, which is for eviction handling
2204 if (blwi->blwi_lock &&
2205 (ldlm_is_discard_data(blwi->blwi_lock) ||
2206 ldlm_is_bl_ast(blwi->blwi_lock))) {
2207 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
2210 /* other blocking callbacks are added to the regular list */
2211 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
2213 blp->blp_total_locks += count;
2214 blp->blp_total_blwis++;
2215 spin_unlock(&blp->blp_lock);
2217 wake_up(&blp->blp_waitq);
2219 /* unlocked read of blp values is intentional - OK for debug */
2221 "added %d/%d locks to %s blp list, %d blwis in pool\n",
2222 count, blp->blp_total_locks, prio, blp->blp_total_blwis);
2225 * can not check blwi->blwi_flags as blwi could be already freed in
2228 if (!(cancel_flags & LCF_ASYNC))
2229 wait_for_completion(&blwi->blwi_comp);
2234 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2235 struct ldlm_namespace *ns,
2236 struct ldlm_lock_desc *ld,
2237 struct list_head *cancels, int count,
2238 struct ldlm_lock *lock,
2239 enum ldlm_cancel_flags cancel_flags)
2241 init_completion(&blwi->blwi_comp);
2242 INIT_LIST_HEAD(&blwi->blwi_head);
2244 if (current->flags & PF_MEMALLOC)
2245 blwi->blwi_mem_pressure = 1;
2248 blwi->blwi_flags = cancel_flags;
2250 blwi->blwi_ld = *ld;
2252 list_splice_init(cancels, &blwi->blwi_head);
2253 blwi->blwi_count = count;
2255 blwi->blwi_lock = lock;
2260 * Queues a list of locks \a cancels containing \a count locks
2261 * for later processing by a blocking thread. If \a count is zero,
2262 * then the lock referenced as \a lock is queued instead.
2264 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2265 * If list addition fails an error is returned and caller is supposed to
2266 * call ->l_blocking_ast itself.
2268 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2269 struct ldlm_lock_desc *ld,
2270 struct ldlm_lock *lock,
2271 struct list_head *cancels, int count,
2272 enum ldlm_cancel_flags cancel_flags)
2276 if (cancels && count == 0)
2279 if (cancel_flags & LCF_ASYNC) {
2280 struct ldlm_bl_work_item *blwi;
2282 OBD_ALLOC(blwi, sizeof(*blwi));
2285 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2287 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2290 * if it is synchronous call do minimum mem alloc, as it could
2291 * be triggered from kernel shrinker
2293 struct ldlm_bl_work_item blwi;
2295 memset(&blwi, 0, sizeof(blwi));
2296 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2297 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2302 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2303 struct ldlm_lock *lock)
2305 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2308 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2309 struct list_head *cancels, int count,
2310 enum ldlm_cancel_flags cancel_flags)
2312 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2315 int ldlm_bl_to_thread_ns(struct ldlm_namespace *ns)
2317 return ldlm_bl_to_thread(ns, NULL, NULL, NULL, 0, LCF_ASYNC);
2320 int ldlm_bl_thread_wakeup(void)
2322 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2326 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2327 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2329 struct obd_device *obd = req->rq_export->exp_obd;
2337 DEBUG_REQ(D_HSM, req, "%s: handle setinfo", obd->obd_name);
2339 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2341 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2343 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2346 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2348 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2350 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2353 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2356 /* We are responsible for swabbing contents of val */
2358 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2359 /* Pass it on to mdc (the "export" in this case) */
2360 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2362 sizeof(KEY_HSM_COPYTOOL_SEND),
2363 KEY_HSM_COPYTOOL_SEND,
2366 DEBUG_REQ(D_WARNING, req, "ignoring unknown key '%s'", key);
2371 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2372 const char *msg, int rc,
2373 const struct lustre_handle *handle)
2375 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2376 "%s, NID=%s lock=%#llx: rc = %d",
2377 msg, libcfs_idstr(&req->rq_peer),
2378 handle ? handle->cookie : 0, rc);
2379 if (req->rq_no_reply)
2380 CWARN("No reply was sent, maybe cause b=21636.\n");
2382 CWARN("Send reply failed, maybe cause b=21636.\n");
2385 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2386 static int ldlm_callback_handler(struct ptlrpc_request *req)
2388 struct ldlm_namespace *ns;
2389 struct ldlm_request *dlm_req;
2390 struct ldlm_lock *lock;
2396 * Requests arrive in sender's byte order. The ptlrpc service
2397 * handler has already checked and, if necessary, byte-swapped the
2398 * incoming request message body, but I am responsible for the
2402 /* do nothing for sec context finalize */
2403 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2406 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2408 if (req->rq_export == NULL) {
2409 rc = ldlm_callback_reply(req, -ENOTCONN);
2410 ldlm_callback_errmsg(req, "Operate on unconnected server",
2415 LASSERT(req->rq_export != NULL);
2416 LASSERT(req->rq_export->exp_obd != NULL);
2418 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2419 case LDLM_BL_CALLBACK:
2420 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2422 ldlm_callback_reply(req, -(int)cfs_fail_err);
2426 case LDLM_CP_CALLBACK:
2427 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2430 case LDLM_GL_CALLBACK:
2431 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2435 rc = ldlm_handle_setinfo(req);
2436 ldlm_callback_reply(req, rc);
2439 CERROR("unknown opcode %u\n",
2440 lustre_msg_get_opc(req->rq_reqmsg));
2441 ldlm_callback_reply(req, -EPROTO);
2445 ns = req->rq_export->exp_obd->obd_namespace;
2446 LASSERT(ns != NULL);
2448 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2450 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2451 if (dlm_req == NULL) {
2452 rc = ldlm_callback_reply(req, -EPROTO);
2453 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2459 * Force a known safe race, send a cancel to the server for a lock
2460 * which the server has already started a blocking callback on.
2462 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2463 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2464 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2466 CERROR("ldlm_cli_cancel: %d\n", rc);
2469 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2472 "callback on lock %#llx - lock disappeared\n",
2473 dlm_req->lock_handle[0].cookie);
2474 rc = ldlm_callback_reply(req, -EINVAL);
2475 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2476 &dlm_req->lock_handle[0]);
2480 if (ldlm_is_fail_loc(lock) &&
2481 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2482 CFS_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2484 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2485 lock_res_and_lock(lock);
2486 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2488 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2490 * If somebody cancels lock and cache is already dropped,
2491 * or lock is failed before cp_ast received on client,
2492 * we can tell the server we have no lock. Otherwise, we
2493 * should send cancel after dropping the cache.
2495 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2496 ldlm_is_failed(lock)) {
2498 "callback on lock %llx - lock disappeared",
2499 dlm_req->lock_handle[0].cookie);
2500 unlock_res_and_lock(lock);
2501 LDLM_LOCK_RELEASE(lock);
2502 rc = ldlm_callback_reply(req, -EINVAL);
2503 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2504 &dlm_req->lock_handle[0]);
2508 * BL_AST locks are not needed in LRU.
2509 * Let ldlm_cancel_lru() be fast.
2511 ldlm_lock_remove_from_lru(lock);
2512 ldlm_set_bl_ast(lock);
2514 if (lock->l_remote_handle.cookie == 0)
2515 lock->l_remote_handle = dlm_req->lock_handle[1];
2516 unlock_res_and_lock(lock);
2519 * We want the ost thread to get this reply so that it can respond
2520 * to ost requests (write cache writeback) that might be triggered
2523 * But we'd also like to be able to indicate in the reply that we're
2524 * cancelling right now, because it's unused, or have an intent result
2525 * in the reply, so we might have to push the responsibility for sending
2526 * the reply down into the AST handlers, alas.
2529 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2530 case LDLM_BL_CALLBACK:
2531 LDLM_DEBUG(lock, "blocking ast");
2532 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2533 if (!ldlm_is_cancel_on_block(lock)) {
2534 rc = ldlm_callback_reply(req, 0);
2535 if (req->rq_no_reply || rc)
2536 ldlm_callback_errmsg(req, "Normal process", rc,
2537 &dlm_req->lock_handle[0]);
2539 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2540 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2542 case LDLM_CP_CALLBACK:
2543 LDLM_DEBUG(lock, "completion ast");
2544 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2545 rc = ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2546 if (!CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE))
2547 ldlm_callback_reply(req, rc);
2549 case LDLM_GL_CALLBACK:
2550 LDLM_DEBUG(lock, "glimpse ast");
2551 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2552 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2555 LBUG(); /* checked above */
2561 #ifdef HAVE_SERVER_SUPPORT
2563 * Main handler for canceld thread.
2565 * Separated into its own thread to avoid deadlocks.
2567 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2574 * Requests arrive in sender's byte order. The ptlrpc service
2575 * handler has already checked and, if necessary, byte-swapped the
2576 * incoming request message body, but I am responsible for the
2580 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2582 if (req->rq_export == NULL) {
2583 struct ldlm_request *dlm_req;
2585 CERROR("%s from %s arrived at %llu with bad export cookie %llu\n",
2586 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2587 libcfs_nidstr(&req->rq_peer.nid),
2588 (unsigned long long)req->rq_arrival_time.tv_sec,
2589 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2591 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2592 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2593 dlm_req = req_capsule_client_get(&req->rq_pill,
2595 if (dlm_req != NULL)
2596 ldlm_lock_dump_handle(D_ERROR,
2597 &dlm_req->lock_handle[0]);
2599 ldlm_callback_reply(req, -ENOTCONN);
2603 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2604 /* XXX FIXME move this back to mds/handler.c, b=249 */
2606 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2607 CDEBUG(D_INODE, "cancel\n");
2608 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2609 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2610 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2612 rc = ldlm_handle_cancel(req);
2616 struct ldlm_request *dlm_req;
2618 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2619 CDEBUG(D_INODE, "convert\n");
2621 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2622 if (dlm_req == NULL) {
2623 CDEBUG(D_INFO, "bad request buffer for cancel\n");
2624 rc = ldlm_callback_reply(req, -EPROTO);
2626 req->rq_status = ldlm_handle_convert0(req, dlm_req);
2627 rc = ptlrpc_reply(req);
2632 CERROR("invalid opcode %d\n",
2633 lustre_msg_get_opc(req->rq_reqmsg));
2634 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2635 rc = ldlm_callback_reply(req, -EINVAL);
2641 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2642 struct ldlm_lock *lock)
2644 struct ldlm_request *dlm_req;
2645 struct lustre_handle lockh;
2651 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2652 if (dlm_req == NULL)
2655 ldlm_lock2handle(lock, &lockh);
2656 for (i = 0; i < dlm_req->lock_count; i++) {
2657 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2659 DEBUG_REQ(D_RPCTRACE, req,
2660 "Prio raised by lock %#llx", lockh.cookie);
2669 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2671 struct ldlm_request *dlm_req;
2678 /* no prolong in recovery */
2679 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2682 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2683 if (dlm_req == NULL)
2686 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
2687 if (size <= offsetof(struct ldlm_request, lock_handle) ||
2688 (size - offsetof(struct ldlm_request, lock_handle)) /
2689 sizeof(struct lustre_handle) < dlm_req->lock_count)
2692 for (i = 0; i < dlm_req->lock_count; i++) {
2693 struct ldlm_lock *lock;
2695 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2699 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2701 LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2702 LDLM_LOCK_PUT(lock);
2711 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2712 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2713 .hpreq_check = ldlm_cancel_hpreq_check,
2717 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2721 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2723 if (req->rq_export == NULL)
2726 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2727 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2728 req->rq_ops = &ldlm_cancel_hpreq_ops;
2729 } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2730 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2731 req->rq_ops = &ldlm_cancel_hpreq_ops;
2736 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2737 struct hlist_node *hnode, void *data)
2740 struct list_head *rpc_list = data;
2741 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2743 lock_res_and_lock(lock);
2745 if (!ldlm_is_granted(lock)) {
2746 unlock_res_and_lock(lock);
2750 LASSERT(lock->l_resource);
2751 if (lock->l_resource->lr_type != LDLM_IBITS &&
2752 lock->l_resource->lr_type != LDLM_PLAIN) {
2753 unlock_res_and_lock(lock);
2757 if (ldlm_is_ast_sent(lock)) {
2758 unlock_res_and_lock(lock);
2762 LASSERT(lock->l_blocking_ast);
2763 LASSERT(!lock->l_blocking_lock);
2765 ldlm_set_ast_sent(lock);
2766 if (lock->l_export && lock->l_export->exp_lock_hash) {
2768 * NB: it's safe to call cfs_hash_del() even lock isn't
2772 * In the function below, .hs_keycmp resolves to
2773 * ldlm_export_lock_keycmp()
2775 cfs_hash_del(lock->l_export->exp_lock_hash,
2776 &lock->l_remote_handle, &lock->l_exp_hash);
2779 list_add_tail(&lock->l_rk_ast, rpc_list);
2780 LDLM_LOCK_GET(lock);
2782 unlock_res_and_lock(lock);
2786 void ldlm_revoke_export_locks(struct obd_export *exp)
2789 LIST_HEAD(rpc_list);
2793 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2794 ldlm_revoke_lock_cb, &rpc_list, 0);
2795 rc = ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2796 LDLM_WORK_REVOKE_AST);
2798 if (rc == -ERESTART)
2799 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2803 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2804 #endif /* HAVE_SERVER_SUPPORT */
2806 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2807 struct ldlm_bl_work_item **p_blwi,
2808 struct obd_export **p_exp)
2810 struct ldlm_bl_work_item *blwi = NULL;
2811 static unsigned int num_bl;
2812 static unsigned int num_stale;
2813 int num_th = atomic_read(&blp->blp_num_threads);
2815 *p_exp = obd_stale_export_get();
2817 spin_lock(&blp->blp_lock);
2818 if (*p_exp != NULL) {
2819 if (num_th == 1 || ++num_stale < num_th) {
2820 spin_unlock(&blp->blp_lock);
2826 /* process a request from the blp_list at least every blp_num_threads */
2827 if (!list_empty(&blp->blp_list) &&
2828 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2829 blwi = list_first_entry(&blp->blp_list,
2830 struct ldlm_bl_work_item, blwi_entry);
2832 if (!list_empty(&blp->blp_prio_list))
2833 blwi = list_first_entry(&blp->blp_prio_list,
2834 struct ldlm_bl_work_item,
2838 if (++num_bl >= num_th)
2840 list_del(&blwi->blwi_entry);
2841 blp->blp_total_locks -= blwi->blwi_lock ? 1 : blwi->blwi_count;
2842 blp->blp_total_blwis--;
2844 spin_unlock(&blp->blp_lock);
2847 /* intentional unlocked read of blp values - OK for debug */
2850 "Got %d locks of %d total in blp. (%d blwis in pool)\n",
2851 blwi->blwi_lock ? 1 : blwi->blwi_count,
2852 blp->blp_total_locks, blp->blp_total_blwis);
2855 "No blwi found in queue (no bl locks in queue)\n");
2858 if (*p_exp != NULL && *p_blwi != NULL) {
2859 obd_stale_export_put(*p_exp);
2863 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2866 /* This only contains temporary data until the thread starts */
2867 struct ldlm_bl_thread_data {
2868 struct ldlm_bl_pool *bltd_blp;
2869 struct completion bltd_comp;
2873 static int ldlm_bl_thread_main(void *arg);
2875 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2877 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2878 struct task_struct *task;
2880 init_completion(&bltd.bltd_comp);
2882 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2883 if (bltd.bltd_num >= blp->blp_max_threads) {
2884 atomic_dec(&blp->blp_num_threads);
2888 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2890 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2891 atomic_dec(&blp->blp_num_threads);
2895 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2898 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2899 bltd.bltd_num, PTR_ERR(task));
2900 atomic_dec(&blp->blp_num_threads);
2901 return PTR_ERR(task);
2903 wait_for_completion(&bltd.bltd_comp);
2908 /* Not fatal if racy and have a few too many threads */
2909 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2910 struct ldlm_bl_work_item *blwi)
2912 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2915 if (atomic_read(&blp->blp_busy_threads) <
2916 atomic_read(&blp->blp_num_threads))
2919 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2920 blwi->blwi_mem_pressure))
2926 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2927 struct ldlm_bl_work_item *blwi)
2929 /* '1' for consistency with code that checks !mpflag to restore */
2930 unsigned int mpflags = 1;
2934 if (blwi->blwi_ns == NULL)
2935 /* added by ldlm_cleanup() */
2936 RETURN(LDLM_ITER_STOP);
2938 if (blwi->blwi_mem_pressure)
2939 mpflags = memalloc_noreclaim_save();
2941 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2943 if (blwi->blwi_count) {
2946 * The special case when we cancel locks in lru
2947 * asynchronously, we pass the list of locks here.
2948 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2949 * canceled locally yet.
2951 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2954 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2956 } else if (blwi->blwi_lock) {
2957 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2960 ldlm_pool_recalc(&blwi->blwi_ns->ns_pool, true);
2961 spin_lock(&blwi->blwi_ns->ns_lock);
2962 blwi->blwi_ns->ns_rpc_recalc = 0;
2963 spin_unlock(&blwi->blwi_ns->ns_lock);
2964 ldlm_namespace_put(blwi->blwi_ns);
2967 if (blwi->blwi_mem_pressure)
2968 memalloc_noreclaim_restore(mpflags);
2970 if (blwi->blwi_flags & LCF_ASYNC)
2971 OBD_FREE(blwi, sizeof(*blwi));
2973 complete(&blwi->blwi_comp);
2979 * Cancel stale locks on export. Cancel blocked locks first.
2980 * If the given export has blocked locks, the next in the list may have
2981 * them too, thus cancel not blocked locks only if the current export has
2984 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2985 struct obd_export *exp)
2991 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2993 num = ldlm_export_cancel_blocked_locks(exp);
2995 ldlm_export_cancel_locks(exp);
2997 obd_stale_export_put(exp);
3004 * Main blocking requests processing thread.
3006 * Callers put locks into its queue by calling ldlm_bl_to_thread.
3007 * This thread in the end ends up doing actual call to ->l_blocking_ast
3010 static int ldlm_bl_thread_main(void *arg)
3013 struct ldlm_bl_pool *blp;
3014 struct ldlm_bl_thread_data *bltd = arg;
3022 rc = lu_env_init(env, LCT_DT_THREAD);
3025 rc = lu_env_add(env);
3027 GOTO(out_env_fini, rc);
3029 blp = bltd->bltd_blp;
3031 complete(&bltd->bltd_comp);
3032 /* cannot use bltd after this, it is only on caller's stack */
3035 struct ldlm_bl_work_item *blwi = NULL;
3036 struct obd_export *exp = NULL;
3039 rc = ldlm_bl_get_work(blp, &blwi, &exp);
3042 wait_event_idle_exclusive(blp->blp_waitq,
3043 ldlm_bl_get_work(blp, &blwi,
3045 atomic_inc(&blp->blp_busy_threads);
3047 if (ldlm_bl_thread_need_create(blp, blwi))
3048 /* discard the return value, we tried */
3049 ldlm_bl_thread_start(blp, true);
3052 rc = ldlm_bl_thread_exports(blp, exp);
3054 rc = ldlm_bl_thread_blwi(blp, blwi);
3056 atomic_dec(&blp->blp_busy_threads);
3058 if (rc == LDLM_ITER_STOP)
3062 * If there are many namespaces, we will not sleep waiting for
3063 * work, and must do a cond_resched to avoid holding the CPU
3069 atomic_dec(&blp->blp_num_threads);
3070 complete(&blp->blp_comp);
3081 static int ldlm_setup(void);
3082 static int ldlm_cleanup(void);
3084 int ldlm_get_ref(void)
3089 mutex_lock(&ldlm_ref_mutex);
3090 if (++ldlm_refcount == 1) {
3095 mutex_unlock(&ldlm_ref_mutex);
3100 void ldlm_put_ref(void)
3103 mutex_lock(&ldlm_ref_mutex);
3104 if (ldlm_refcount == 1) {
3105 int rc = ldlm_cleanup();
3108 CERROR("ldlm_cleanup failed: %d\n", rc);
3114 mutex_unlock(&ldlm_ref_mutex);
3119 /* Export handle<->lock hash operations. */
3121 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key,
3122 const unsigned int bits)
3124 return cfs_hash_64(((struct lustre_handle *)key)->cookie, bits);
3128 ldlm_export_lock_key(struct hlist_node *hnode)
3130 struct ldlm_lock *lock;
3132 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3133 return &lock->l_remote_handle;
3137 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
3139 struct ldlm_lock *lock;
3141 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3142 lock->l_remote_handle = *(struct lustre_handle *)key;
3146 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
3148 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
3152 ldlm_export_lock_object(struct hlist_node *hnode)
3154 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3158 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
3160 struct ldlm_lock *lock;
3162 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3163 LDLM_LOCK_GET(lock);
3167 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
3169 struct ldlm_lock *lock;
3171 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3172 LDLM_LOCK_RELEASE(lock);
3175 static struct cfs_hash_ops ldlm_export_lock_ops = {
3176 .hs_hash = ldlm_export_lock_hash,
3177 .hs_key = ldlm_export_lock_key,
3178 .hs_keycmp = ldlm_export_lock_keycmp,
3179 .hs_keycpy = ldlm_export_lock_keycpy,
3180 .hs_object = ldlm_export_lock_object,
3181 .hs_get = ldlm_export_lock_get,
3182 .hs_put = ldlm_export_lock_put,
3183 .hs_put_locked = ldlm_export_lock_put,
3186 int ldlm_init_export(struct obd_export *exp)
3192 exp->exp_lock_hash =
3193 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
3194 HASH_EXP_LOCK_CUR_BITS,
3195 HASH_EXP_LOCK_MAX_BITS,
3196 HASH_EXP_LOCK_BKT_BITS, 0,
3197 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
3198 &ldlm_export_lock_ops,
3199 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
3200 CFS_HASH_NBLK_CHANGE);
3202 if (!exp->exp_lock_hash)
3205 rc = ldlm_init_flock_export(exp);
3211 ldlm_destroy_export(exp);
3214 EXPORT_SYMBOL(ldlm_init_export);
3216 void ldlm_destroy_export(struct obd_export *exp)
3219 cfs_hash_putref(exp->exp_lock_hash);
3220 exp->exp_lock_hash = NULL;
3222 ldlm_destroy_flock_export(exp);
3225 EXPORT_SYMBOL(ldlm_destroy_export);
3227 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
3228 struct attribute *attr,
3231 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
3234 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
3235 struct attribute *attr,
3242 rc = kstrtoul(buffer, 10, &val);
3246 ldlm_cancel_unused_locks_before_replay = val;
3250 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
3252 static struct attribute *ldlm_attrs[] = {
3253 &lustre_attr_cancel_unused_locks_before_replay.attr,
3257 static struct attribute_group ldlm_attr_group = {
3258 .attrs = ldlm_attrs,
3261 static int ldlm_setup(void)
3263 static struct ptlrpc_service_conf conf;
3264 struct ldlm_bl_pool *blp = NULL;
3265 #ifdef HAVE_SERVER_SUPPORT
3266 struct task_struct *task;
3267 #endif /* HAVE_SERVER_SUPPORT */
3273 if (ldlm_state != NULL)
3276 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
3277 if (ldlm_state == NULL)
3280 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
3284 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
3288 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
3292 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
3296 rc = ldlm_debugfs_setup();
3300 memset(&conf, 0, sizeof(conf));
3301 conf = (typeof(conf)) {
3302 .psc_name = "ldlm_cbd",
3303 .psc_watchdog_factor = 2,
3305 .bc_nbufs = LDLM_CLIENT_NBUFS,
3306 .bc_buf_size = LDLM_BUFSIZE,
3307 .bc_req_max_size = LDLM_MAXREQSIZE,
3308 .bc_rep_max_size = LDLM_MAXREPSIZE,
3309 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
3310 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
3313 .tc_thr_name = "ldlm_cb",
3314 .tc_thr_factor = LDLM_THR_FACTOR,
3315 .tc_nthrs_init = LDLM_NTHRS_INIT,
3316 .tc_nthrs_base = LDLM_NTHRS_BASE,
3317 .tc_nthrs_max = LDLM_NTHRS_MAX,
3318 .tc_nthrs_user = ldlm_num_threads,
3319 .tc_cpu_bind = ldlm_cpu_bind,
3320 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
3323 .cc_pattern = ldlm_cpts,
3324 .cc_affinity = true,
3327 .so_req_handler = ldlm_callback_handler,
3330 ldlm_state->ldlm_cb_service = ptlrpc_register_service(&conf,
3332 ldlm_svc_debugfs_dir);
3333 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3334 CERROR("failed to start service\n");
3335 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3336 ldlm_state->ldlm_cb_service = NULL;
3340 #ifdef HAVE_SERVER_SUPPORT
3341 memset(&conf, 0, sizeof(conf));
3342 conf = (typeof(conf)) {
3343 .psc_name = "ldlm_canceld",
3344 .psc_watchdog_factor = 6,
3346 .bc_nbufs = LDLM_SERVER_NBUFS,
3347 .bc_buf_size = LDLM_BUFSIZE,
3348 .bc_req_max_size = LDLM_MAXREQSIZE,
3349 .bc_rep_max_size = LDLM_MAXREPSIZE,
3350 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3351 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3355 .tc_thr_name = "ldlm_cn",
3356 .tc_thr_factor = LDLM_THR_FACTOR,
3357 .tc_nthrs_init = LDLM_NTHRS_INIT,
3358 .tc_nthrs_base = LDLM_NTHRS_BASE,
3359 .tc_nthrs_max = LDLM_NTHRS_MAX,
3360 .tc_nthrs_user = ldlm_num_threads,
3361 .tc_cpu_bind = ldlm_cpu_bind,
3362 .tc_ctx_tags = LCT_MD_THREAD |
3367 .cc_pattern = ldlm_cpts,
3368 .cc_affinity = true,
3371 .so_req_handler = ldlm_cancel_handler,
3372 .so_hpreq_handler = ldlm_hpreq_handler,
3375 ldlm_state->ldlm_cancel_service =
3376 ptlrpc_register_service(&conf, ldlm_svc_kset,
3377 ldlm_svc_debugfs_dir);
3378 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3379 CERROR("failed to start service\n");
3380 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3381 ldlm_state->ldlm_cancel_service = NULL;
3384 #endif /* HAVE_SERVER_SUPPORT */
3386 OBD_ALLOC(blp, sizeof(*blp));
3388 GOTO(out, rc = -ENOMEM);
3389 ldlm_state->ldlm_bl_pool = blp;
3391 spin_lock_init(&blp->blp_lock);
3392 INIT_LIST_HEAD(&blp->blp_list);
3393 INIT_LIST_HEAD(&blp->blp_prio_list);
3394 init_waitqueue_head(&blp->blp_waitq);
3395 atomic_set(&blp->blp_num_threads, 0);
3396 atomic_set(&blp->blp_busy_threads, 0);
3397 blp->blp_total_locks = 0;
3398 blp->blp_total_blwis = 0;
3400 if (ldlm_num_threads == 0) {
3401 blp->blp_min_threads = LDLM_NTHRS_INIT;
3402 blp->blp_max_threads = LDLM_NTHRS_MAX;
3404 blp->blp_min_threads = blp->blp_max_threads =
3405 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3409 for (i = 0; i < blp->blp_min_threads; i++) {
3410 rc = ldlm_bl_thread_start(blp, false);
3415 #ifdef HAVE_SERVER_SUPPORT
3416 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3419 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3423 wait_event(expired_lock_wait_queue,
3424 expired_lock_thread_state == ELT_READY);
3425 #endif /* HAVE_SERVER_SUPPORT */
3427 rc = ldlm_pools_init();
3429 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3433 rc = ldlm_reclaim_setup();
3435 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3445 static int ldlm_cleanup(void)
3449 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3450 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3451 CERROR("ldlm still has namespaces; clean these up first.\n");
3452 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3453 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3457 ldlm_reclaim_cleanup();
3460 if (ldlm_state->ldlm_bl_pool != NULL) {
3461 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3463 while (atomic_read(&blp->blp_num_threads) > 0) {
3464 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3466 init_completion(&blp->blp_comp);
3468 spin_lock(&blp->blp_lock);
3469 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3470 wake_up(&blp->blp_waitq);
3471 spin_unlock(&blp->blp_lock);
3473 wait_for_completion(&blp->blp_comp);
3476 OBD_FREE(blp, sizeof(*blp));
3479 if (ldlm_state->ldlm_cb_service != NULL)
3480 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3481 #ifdef HAVE_SERVER_SUPPORT
3482 if (ldlm_state->ldlm_cancel_service != NULL)
3483 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3487 kset_unregister(ldlm_ns_kset);
3489 kset_unregister(ldlm_svc_kset);
3491 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3492 kobject_put(ldlm_kobj);
3495 ldlm_debugfs_cleanup();
3497 #ifdef HAVE_SERVER_SUPPORT
3498 if (expired_lock_thread_state != ELT_STOPPED) {
3499 expired_lock_thread_state = ELT_TERMINATE;
3500 wake_up(&expired_lock_wait_queue);
3501 wait_event(expired_lock_wait_queue,
3502 expired_lock_thread_state == ELT_STOPPED);
3506 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3514 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3515 sizeof(struct ldlm_resource), 0,
3516 SLAB_HWCACHE_ALIGN, NULL);
3517 if (ldlm_resource_slab == NULL)
3520 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3521 sizeof(struct ldlm_lock), 0,
3522 SLAB_HWCACHE_ALIGN, NULL);
3523 if (ldlm_lock_slab == NULL)
3526 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3527 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3528 0, SLAB_HWCACHE_ALIGN, NULL);
3529 if (ldlm_interval_tree_slab == NULL)
3532 #ifdef HAVE_SERVER_SUPPORT
3533 ldlm_inodebits_slab = kmem_cache_create("ldlm_ibits_node",
3534 sizeof(struct ldlm_ibits_node),
3535 0, SLAB_HWCACHE_ALIGN, NULL);
3536 if (ldlm_inodebits_slab == NULL)
3537 goto out_interval_tree;
3539 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3540 sizeof(struct ldlm_glimpse_work),
3542 if (ldlm_glimpse_work_kmem == NULL)
3546 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3547 class_export_dump_hook = ldlm_dump_export_locks;
3550 #ifdef HAVE_SERVER_SUPPORT
3552 kmem_cache_destroy(ldlm_inodebits_slab);
3554 kmem_cache_destroy(ldlm_interval_tree_slab);
3557 kmem_cache_destroy(ldlm_lock_slab);
3559 kmem_cache_destroy(ldlm_resource_slab);
3564 void ldlm_exit(void)
3567 CERROR("ldlm_refcount is %d in %s\n", ldlm_refcount, __func__);
3569 kmem_cache_destroy(ldlm_resource_slab);
3571 * ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3572 * rcu_barrier() to wait all outstanding RCU callbacks to complete,
3573 * so that ldlm_lock_free() get a chance to be called.
3576 kmem_cache_destroy(ldlm_lock_slab);
3577 kmem_cache_destroy(ldlm_interval_tree_slab);
3578 #ifdef HAVE_SERVER_SUPPORT
3579 kmem_cache_destroy(ldlm_inodebits_slab);
3580 kmem_cache_destroy(ldlm_glimpse_work_kmem);