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_MSG(0x138,
716 "%s: A client on nid %s was evicted due to a lock %s callback time out: rc %d\n",
717 lock->l_export->exp_obd->obd_name,
718 obd_export_nid2str(lock->l_export), ast_type, rc);
720 if (obd_dump_on_timeout)
721 libcfs_debug_dumplog();
722 spin_lock_bh(&waiting_locks_spinlock);
723 if (__ldlm_del_waiting_lock(lock) == 0)
725 * the lock was not in any list, grab an extra ref before adding
726 * the lock to the expired list
729 /* differentiate it from expired locks */
730 lock->l_callback_timestamp = 0;
731 list_add(&lock->l_pending_chain, &expired_lock_list);
732 wake_up(&expired_lock_wait_queue);
733 spin_unlock_bh(&waiting_locks_spinlock);
737 * Perform lock cleanup if AST reply came with error.
739 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
740 struct ptlrpc_request *req, int rc,
741 const char *ast_type)
743 struct lnet_processid *peer = &req->rq_import->imp_connection->c_peer;
745 if (!req->rq_replied || (rc && rc != -EINVAL)) {
746 if (ldlm_is_cancel(lock)) {
748 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
749 ast_type, req, req->rq_xid,
750 libcfs_nidstr(&peer->nid));
751 ldlm_lock_cancel(lock);
753 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
755 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
757 * Upon umount process the AST fails because cannot be
758 * sent. This shouldn't lead to the client eviction.
759 * -ENODEV error is returned by ptl_send_rpc() for
760 * new request in such import.
761 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
762 * if imp_invalid is set or obd_no_recov.
763 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
764 * in ptlrpc_import_delay_req() as well with -EIO code.
765 * In all such cases errors are ignored.
768 "%s AST can't be sent due to a server %s failure or umount process: rc = %d",
770 req->rq_import->imp_obd->obd_name, rc);
773 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
774 libcfs_nidstr(&peer->nid),
775 req->rq_replied ? "returned error from" :
776 "failed to reply to",
777 ast_type, req, req->rq_xid,
778 (req->rq_repmsg != NULL) ?
779 lustre_msg_get_status(req->rq_repmsg) : 0,
781 ldlm_failed_ast(lock, rc, ast_type);
787 struct ldlm_resource *res = lock->l_resource;
790 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
791 libcfs_nidstr(&peer->nid),
793 lustre_msg_get_status(req->rq_repmsg) : -1,
794 ast_type, req, req->rq_xid);
797 * update lvbo to return proper attributes.
800 ldlm_resource_getref(res);
801 ldlm_lvbo_update(res, lock, NULL, 1);
802 ldlm_resource_putref(res);
804 ldlm_lock_cancel(lock);
811 static int ldlm_cb_interpret(const struct lu_env *env,
812 struct ptlrpc_request *req, void *args, int rc)
814 struct ldlm_cb_async_args *ca = args;
815 struct ldlm_lock *lock = ca->ca_lock;
816 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
820 LASSERT(lock != NULL);
823 case LDLM_GL_CALLBACK:
825 * Update the LVB from disk if the AST failed
826 * (this is a legal race)
828 * - Glimpse callback of local lock just returns
829 * -ELDLM_NO_LOCK_DATA.
830 * - Glimpse callback of remote lock might return
831 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
833 if (unlikely(arg->gl_interpret_reply)) {
834 rc = arg->gl_interpret_reply(NULL, req, args, rc);
835 } else if (rc == -ELDLM_NO_LOCK_DATA) {
837 "lost race - client has a lock but no inode");
838 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
839 } else if (rc != 0) {
840 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
842 rc = ldlm_lvbo_update(lock->l_resource,
846 case LDLM_BL_CALLBACK:
848 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
850 case LDLM_CP_CALLBACK:
852 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
855 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
860 /* release extra reference taken in ldlm_ast_fini() */
861 LDLM_LOCK_RELEASE(lock);
864 atomic_inc(&arg->restart);
869 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
871 struct ldlm_cb_async_args *ca = data;
872 struct ldlm_lock *lock = ca->ca_lock;
874 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
877 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
878 struct ldlm_cb_set_arg *arg,
879 struct ldlm_lock *lock,
886 if (unlikely(instant_cancel)) {
887 rc = ptl_send_rpc(req, 1);
888 ptlrpc_req_finished(req);
890 atomic_inc(&arg->restart);
893 ptlrpc_set_add_req(arg->set, req);
900 * Check if there are requests in the export request list which prevent
901 * the lock canceling and make these requests high priority ones.
903 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
905 struct ptlrpc_request *req;
909 if (lock->l_export == NULL) {
910 LDLM_DEBUG(lock, "client lock: no-op");
914 spin_lock(&lock->l_export->exp_rpc_lock);
915 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
918 * Do not process requests that were not yet added to there
919 * incoming queue or were already removed from there for
920 * processing. We evaluate ptlrpc_nrs_req_can_move() without
921 * holding svcpt->scp_req_lock, and then redo the check with
922 * the lock held once we need to obtain a reliable result.
924 if (ptlrpc_nrs_req_can_move(req) &&
925 req->rq_ops->hpreq_lock_match &&
926 req->rq_ops->hpreq_lock_match(req, lock))
927 ptlrpc_nrs_req_hp_move(req);
929 spin_unlock(&lock->l_export->exp_rpc_lock);
934 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
935 * enqueued server lock conflicts with given one.
937 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
938 * to wait for client response.
940 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
941 struct ldlm_lock_desc *desc,
942 void *data, int flag)
944 struct ldlm_cb_async_args *ca;
945 struct ldlm_cb_set_arg *arg = data;
946 struct ldlm_request *body;
947 struct ptlrpc_request *req;
948 int instant_cancel = 0;
950 struct obd_device *obd;
954 if (flag == LDLM_CB_CANCELING)
955 /* Don't need to do anything here. */
958 if (CFS_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
959 LDLM_DEBUG(lock, "dropping BL AST");
964 LASSERT(data != NULL);
966 obd = lock->l_export->exp_obd;
967 if (obd->obd_recovering != 0)
968 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
970 ldlm_lock_reorder_req(lock);
972 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
973 &RQF_LDLM_BL_CALLBACK,
974 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
978 ca = ptlrpc_req_async_args(ca, req);
979 ca->ca_set_arg = arg;
982 req->rq_interpret_reply = ldlm_cb_interpret;
984 lock_res_and_lock(lock);
985 if (ldlm_is_destroyed(lock)) {
986 /* What's the point? */
987 unlock_res_and_lock(lock);
988 ptlrpc_req_finished(req);
992 if (!ldlm_is_granted(lock)) {
994 * this blocking AST will be communicated as part of the
995 * completion AST instead
997 ldlm_add_blocked_lock(lock);
998 ldlm_set_waited(lock);
999 unlock_res_and_lock(lock);
1001 ptlrpc_req_finished(req);
1002 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
1006 if (ldlm_is_cancel_on_block(lock))
1009 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1010 body->lock_handle[0] = lock->l_remote_handle;
1011 body->lock_handle[1].cookie = lock->l_handle.h_cookie;
1012 body->lock_desc = *desc;
1013 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1016 LDLM_DEBUG(lock, "server preparing blocking AST");
1018 ptlrpc_request_set_replen(req);
1019 ldlm_set_cbpending(lock);
1020 if (instant_cancel) {
1021 unlock_res_and_lock(lock);
1022 ldlm_lock_cancel(lock);
1024 req->rq_no_resend = 1;
1026 LASSERT(ldlm_is_granted(lock));
1027 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1028 unlock_res_and_lock(lock);
1030 /* Do not resend after lock callback timeout */
1031 req->rq_delay_limit = ldlm_bl_timeout(lock);
1032 req->rq_resend_cb = ldlm_update_resend;
1035 req->rq_send_state = LUSTRE_IMP_FULL;
1036 /* ptlrpc_request_alloc_pack already set timeout */
1037 if (obd_at_off(obd))
1038 req->rq_timeout = ldlm_get_rq_timeout();
1040 if (lock->l_export && lock->l_export->exp_nid_stats &&
1041 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1042 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1043 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
1045 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1051 * ->l_completion_ast callback for a remote lock in server namespace.
1053 * Sends AST to the client notifying it of lock granting. If initial
1054 * lock response was not sent yet, instead of sending another RPC, just
1055 * mark the lock as granted and client will understand
1057 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
1059 struct ldlm_cb_set_arg *arg = data;
1060 struct ldlm_request *body;
1061 struct ptlrpc_request *req;
1062 struct ldlm_cb_async_args *ca;
1063 int instant_cancel = 0;
1066 struct obd_device *obd;
1070 LASSERT(lock != NULL);
1071 LASSERT(data != NULL);
1073 if (CFS_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
1074 LDLM_DEBUG(lock, "dropping CP AST");
1078 obd = lock->l_export->exp_obd;
1079 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
1080 &RQF_LDLM_CP_CALLBACK);
1084 /* server namespace, doesn't need lock */
1085 lvb_len = ldlm_lvbo_size(lock);
1087 * LU-3124 & LU-2187: to not return layout in completion AST because
1088 * it may deadlock for LU-2187, or client may not have enough space
1089 * for large layout. The layout will be returned to client with an
1090 * extra RPC to fetch xattr.lov
1092 if (ldlm_has_layout(lock))
1095 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
1096 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
1098 ptlrpc_request_free(req);
1102 ca = ptlrpc_req_async_args(ca, req);
1103 ca->ca_set_arg = arg;
1106 req->rq_interpret_reply = ldlm_cb_interpret;
1107 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1109 body->lock_handle[0] = lock->l_remote_handle;
1110 body->lock_handle[1].cookie = lock->l_handle.h_cookie;
1111 body->lock_flags = ldlm_flags_to_wire(flags);
1112 ldlm_lock2desc(lock, &body->lock_desc);
1114 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1116 lvb_len = ldlm_lvbo_fill(lock, lvb, &lvb_len);
1119 * We still need to send the RPC to wake up the blocked
1120 * enqueue thread on the client.
1122 * Consider old client, there is no better way to notify
1123 * the failure, just zero-sized the LVB, then the client
1124 * will fail out as "-EPROTO".
1126 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1130 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1135 LDLM_DEBUG(lock, "server preparing completion AST");
1137 ptlrpc_request_set_replen(req);
1139 req->rq_send_state = LUSTRE_IMP_FULL;
1140 /* ptlrpc_request_pack already set timeout */
1141 if (obd_at_off(obd))
1142 req->rq_timeout = ldlm_get_rq_timeout();
1144 /* We only send real blocking ASTs after the lock is granted */
1145 lock_res_and_lock(lock);
1146 if (ldlm_is_ast_sent(lock)) {
1147 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1148 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1149 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1153 * We might get here prior to ldlm_handle_enqueue setting
1154 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1155 * into waiting list, but this is safe and similar code in
1156 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1157 * that would not only cancel the lock, but will also remove
1158 * it from waiting list
1160 if (ldlm_is_cancel_on_block(lock)) {
1161 unlock_res_and_lock(lock);
1162 ldlm_lock_cancel(lock);
1165 req->rq_no_resend = 1;
1167 lock_res_and_lock(lock);
1169 /* start the lock-timeout clock */
1170 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1171 /* Do not resend after lock callback timeout */
1172 req->rq_delay_limit = ldlm_bl_timeout(lock);
1173 req->rq_resend_cb = ldlm_update_resend;
1176 unlock_res_and_lock(lock);
1178 if (lock->l_export && lock->l_export->exp_nid_stats &&
1179 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1180 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1181 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1183 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1185 RETURN(lvb_len < 0 ? lvb_len : rc);
1187 EXPORT_SYMBOL(ldlm_server_completion_ast);
1190 * Server side ->l_glimpse_ast handler for client locks.
1192 * Sends glimpse AST to the client and waits for reply. Then updates
1193 * lvbo with the result.
1195 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1197 struct ldlm_cb_set_arg *arg = data;
1198 struct ldlm_request *body;
1199 struct ptlrpc_request *req;
1200 struct ldlm_cb_async_args *ca;
1202 struct req_format *req_fmt;
1203 struct obd_device *obd = lock->l_export->exp_obd;
1207 LASSERT(lock != NULL);
1209 if (arg->gl_desc != NULL)
1210 /* There is a glimpse descriptor to pack */
1211 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1213 req_fmt = &RQF_LDLM_GL_CALLBACK;
1215 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1216 req_fmt, LUSTRE_DLM_VERSION,
1222 if (arg->gl_desc != NULL) {
1223 /* copy the GL descriptor */
1224 union ldlm_gl_desc *desc;
1226 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1227 *desc = *arg->gl_desc;
1230 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1231 body->lock_handle[0] = lock->l_remote_handle;
1232 ldlm_lock2desc(lock, &body->lock_desc);
1234 ca = ptlrpc_req_async_args(ca, req);
1235 ca->ca_set_arg = arg;
1238 /* server namespace, doesn't need lock */
1239 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1240 ldlm_lvbo_size(lock));
1241 ptlrpc_request_set_replen(req);
1243 req->rq_send_state = LUSTRE_IMP_FULL;
1244 /* ptlrpc_request_alloc_pack already set timeout */
1245 if (obd_at_off(obd))
1246 req->rq_timeout = ldlm_get_rq_timeout();
1248 req->rq_interpret_reply = ldlm_cb_interpret;
1250 if (lock->l_export && lock->l_export->exp_nid_stats) {
1251 struct nid_stat *nid_stats = lock->l_export->exp_nid_stats;
1253 lprocfs_counter_incr(nid_stats->nid_ldlm_stats,
1254 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1257 rc = ldlm_ast_fini(req, arg, lock, 0);
1261 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1263 int ldlm_glimpse_locks(struct ldlm_resource *res,
1264 struct list_head *gl_work_list)
1270 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1272 if (rc == -ERESTART)
1273 ldlm_reprocess_all(res, 0);
1277 EXPORT_SYMBOL(ldlm_glimpse_locks);
1279 /* return LDLM lock associated with a lock callback request */
1280 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1282 struct ldlm_cb_async_args *ca;
1283 struct ldlm_lock *lock;
1287 ca = ptlrpc_req_async_args(ca, req);
1290 RETURN(ERR_PTR(-EFAULT));
1294 EXPORT_SYMBOL(ldlm_request_lock);
1297 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1298 * service threads to carry out client lock enqueueing requests.
1300 int ldlm_handle_enqueue(struct ldlm_namespace *ns,
1301 struct req_capsule *pill,
1302 const struct ldlm_request *dlm_req,
1303 const struct ldlm_callback_suite *cbs)
1305 struct ldlm_reply *dlm_rep;
1307 enum ldlm_error err = ELDLM_OK;
1308 struct ldlm_lock *lock = NULL;
1309 void *cookie = NULL;
1311 struct ldlm_resource *res = NULL;
1312 struct ptlrpc_request *req = pill->rc_req;
1313 const struct lu_env *env = req->rq_svc_thread->t_env;
1317 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1319 LASSERT(req && req->rq_export);
1321 if (req_capsule_ptlreq(pill))
1322 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF,
1325 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1327 /* for intent enqueue the stat will be updated inside intent policy */
1328 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1329 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1330 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1332 if (req->rq_export->exp_nid_stats &&
1333 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1334 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1335 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1337 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1338 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1339 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1340 dlm_req->lock_desc.l_resource.lr_type);
1341 GOTO(out, rc = -EFAULT);
1344 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1345 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1346 dlm_req->lock_desc.l_req_mode &
1347 (dlm_req->lock_desc.l_req_mode-1))) {
1348 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1349 dlm_req->lock_desc.l_req_mode);
1350 GOTO(out, rc = -EFAULT);
1353 if (unlikely((flags & LDLM_FL_REPLAY) ||
1354 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1355 /* Find an existing lock in the per-export lock hash */
1357 * In the function below, .hs_keycmp resolves to
1358 * ldlm_export_lock_keycmp()
1360 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1361 (void *)&dlm_req->lock_handle[0]);
1363 DEBUG_REQ(D_DLMTRACE, req,
1364 "found existing lock cookie %#llx",
1365 lock->l_handle.h_cookie);
1366 flags |= LDLM_FL_RESENT;
1367 GOTO(existing_lock, rc = 0);
1370 if (ldlm_reclaim_full()) {
1371 DEBUG_REQ(D_DLMTRACE, req,
1372 "Too many granted locks, reject current enqueue request and let the client retry later");
1373 GOTO(out, rc = -EINPROGRESS);
1377 /* The lock's callback data might be set in the policy function */
1378 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1379 dlm_req->lock_desc.l_resource.lr_type,
1380 dlm_req->lock_desc.l_req_mode,
1381 cbs, NULL, 0, LVB_T_NONE);
1388 lock->l_remote_handle = dlm_req->lock_handle[0];
1389 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1392 * Initialize resource lvb but not for a lock being replayed since
1393 * Client already got lvb sent in this case.
1394 * This must occur early since some policy methods assume resource
1395 * lvb is available (lr_lvb_data != NULL).
1397 res = lock->l_resource;
1398 if (!(flags & LDLM_FL_REPLAY)) {
1399 /* non-replayed lock, delayed lvb init may need to be done */
1400 rc = ldlm_lvbo_init(res);
1402 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1407 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1409 * Don't enqueue a lock onto the export if it is been disonnected
1410 * due to eviction (b=3822) or server umount (b=24324).
1411 * Cancel it now instead.
1413 if (req->rq_export->exp_disconnected) {
1414 LDLM_ERROR(lock, "lock on disconnected export %p",
1416 GOTO(out, rc = -ENOTCONN);
1419 lock->l_export = class_export_lock_get(req->rq_export, lock);
1420 if (lock->l_export->exp_lock_hash)
1421 cfs_hash_add(lock->l_export->exp_lock_hash,
1422 &lock->l_remote_handle,
1426 * Inherit the enqueue flags before the operation, because we do not
1427 * keep the res lock on return and next operations (BL AST) may proceed
1430 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1431 LDLM_FL_INHERIT_MASK);
1433 ldlm_convert_policy_to_local(req->rq_export,
1434 dlm_req->lock_desc.l_resource.lr_type,
1435 &dlm_req->lock_desc.l_policy_data,
1436 &lock->l_policy_data);
1437 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT) {
1438 lock->l_req_extent = lock->l_policy_data.l_extent;
1439 } else if (dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) {
1440 lock->l_policy_data.l_inodebits.try_bits =
1441 dlm_req->lock_desc.l_policy_data.l_inodebits.try_bits;
1442 lock->l_policy_data.l_inodebits.li_gid =
1443 dlm_req->lock_desc.l_policy_data.l_inodebits.li_gid;
1448 if (!(flags & LDLM_FL_HAS_INTENT)) {
1449 /* based on the assumption that lvb size never changes during
1450 * resource life time otherwise it need resource->lr_lock's
1453 req_capsule_set_size(pill, &RMF_DLM_LVB,
1454 RCL_SERVER, ldlm_lvbo_size(lock));
1456 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1457 GOTO(out, rc = -ENOMEM);
1459 rc = req_capsule_server_pack(pill);
1464 err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1471 dlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
1473 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1474 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1476 if (lock->l_resource->lr_type == LDLM_EXTENT)
1477 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1480 * We never send a blocking AST until the lock is granted, but
1481 * we can tell it right now
1483 lock_res_and_lock(lock);
1486 * Now take into account flags to be inherited from original lock
1487 * request both in reply to client and in our own lock flags.
1489 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1490 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1493 * Don't move a pending lock onto the export if it has already been
1494 * disconnected due to eviction (b=5683) or server umount (b=24324).
1495 * Cancel it now instead.
1497 if (unlikely(req->rq_export->exp_disconnected ||
1498 CFS_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1499 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1501 } else if (ldlm_is_ast_sent(lock)) {
1502 /* fill lock desc for possible lock convert */
1503 if (lock->l_blocking_lock &&
1504 lock->l_resource->lr_type == LDLM_IBITS) {
1505 struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1506 struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1509 "save blocking bits %llx in granted lock",
1510 bl_lock->l_policy_data.l_inodebits.bits);
1512 * If lock is blocked then save blocking ibits
1513 * in returned lock policy for the possible lock
1514 * convert on a client.
1516 rep_desc->l_policy_data.l_inodebits.cancel_bits =
1517 bl_lock->l_policy_data.l_inodebits.bits;
1519 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1520 if (ldlm_is_granted(lock)) {
1522 * Only cancel lock if it was granted, because it would
1523 * be destroyed immediately and would never be granted
1524 * in the future, causing timeouts on client. Not
1525 * granted lock will be cancelled immediately after
1526 * sending completion AST.
1528 if (ldlm_is_cancel_on_block(lock)) {
1529 unlock_res_and_lock(lock);
1530 ldlm_lock_cancel(lock);
1531 lock_res_and_lock(lock);
1533 ldlm_add_waiting_lock(lock,
1534 ldlm_bl_timeout(lock));
1538 unlock_res_and_lock(lock);
1542 if (req_capsule_ptlreq(pill)) {
1543 req->rq_status = rc ?: err; /* return either error - b=11190 */
1544 if (!req->rq_packed_final) {
1545 int rc1 = lustre_pack_reply(req, 1, NULL, NULL);
1553 * The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1554 * ldlm_reprocess_all. If this moves, revisit that code. -phil
1558 "server-side enqueue handler, sending reply (err=%d, rc=%d)",
1562 req_capsule_has_field(pill, &RMF_DLM_LVB,
1564 ldlm_lvbo_size(lock) > 0) {
1569 buf = req_capsule_server_get(pill, &RMF_DLM_LVB);
1570 LASSERTF(buf != NULL, "req %px, lock %px\n", req, lock);
1571 buflen = req_capsule_get_size(pill, &RMF_DLM_LVB,
1574 * non-replayed lock, delayed lvb init may
1575 * need to be occur now
1577 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1580 rc2 = ldlm_lvbo_fill(lock, buf, &buflen);
1582 req_capsule_shrink(pill, &RMF_DLM_LVB,
1584 } else if (rc2 == -ERANGE) {
1585 rc2 = req_capsule_server_grow(
1592 * if we can't grow the buffer,
1593 * it's ok to return empty lvb
1597 pill, &RMF_DLM_LVB, 0,
1603 } else if (flags & LDLM_FL_REPLAY) {
1604 /* no LVB resend upon replay */
1606 req_capsule_shrink(pill, &RMF_DLM_LVB,
1615 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1616 if (lock->l_export) {
1617 ldlm_lock_cancel(lock);
1619 lock_res_and_lock(lock);
1620 ldlm_resource_unlink_lock(lock);
1621 ldlm_lock_destroy_nolock(lock);
1622 unlock_res_and_lock(lock);
1624 ldlm_reprocess_all(lock->l_resource,
1625 lock->l_policy_data.l_inodebits.bits);
1628 if (!err && !ldlm_is_cbpending(lock) &&
1629 dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1630 ldlm_reprocess_all(lock->l_resource,
1631 lock->l_policy_data.l_inodebits.bits);
1633 LDLM_LOCK_RELEASE(lock);
1636 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1641 EXPORT_SYMBOL(ldlm_handle_enqueue);
1644 * Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1645 * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1647 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1649 if (lock->l_blocking_lock) {
1650 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1651 lock->l_blocking_lock = NULL;
1655 /* A lock can be converted to new ibits or mode and should be considered
1656 * as new lock. Clear all states related to a previous blocking AST
1657 * processing so new conflicts will cause new blocking ASTs.
1659 * This is used during lock convert below and lock downgrade to COS mode in
1660 * ldlm_lock_mode_downgrade().
1662 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1664 ldlm_clear_ast_sent(lock);
1665 lock->l_bl_ast_run = 0;
1666 ldlm_clear_blocking_lock(lock);
1669 /* Main LDLM entry point for server code to process lock conversion requests */
1670 int ldlm_handle_convert0(struct ptlrpc_request *req,
1671 const struct ldlm_request *dlm_req)
1673 struct obd_export *exp = req->rq_export;
1674 struct ldlm_reply *dlm_rep;
1675 struct ldlm_lock *lock;
1682 if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1683 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1684 LDLM_CONVERT - LDLM_FIRST_OPC);
1686 rc = req_capsule_server_pack(&req->rq_pill);
1690 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1691 dlm_rep->lock_flags = dlm_req->lock_flags;
1693 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1695 LDLM_DEBUG_NOLOCK("server lock is canceled already");
1696 req->rq_status = ELDLM_NO_LOCK_DATA;
1700 LDLM_DEBUG(lock, "server-side convert handler START");
1702 lock_res_and_lock(lock);
1703 bits = lock->l_policy_data.l_inodebits.bits;
1704 new_bits = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1706 if (ldlm_is_cancel(lock)) {
1707 LDLM_DEBUG(lock, "convert on canceled lock!");
1708 unlock_res_and_lock(lock);
1709 GOTO(out_put, rc = ELDLM_NO_LOCK_DATA);
1712 if (dlm_req->lock_desc.l_req_mode != lock->l_granted_mode) {
1713 LDLM_ERROR(lock, "lock mode differs!");
1714 unlock_res_and_lock(lock);
1715 GOTO(out_put, rc = -EPROTO);
1718 if (bits == new_bits) {
1720 * This can be valid situation if CONVERT RPCs are
1721 * re-ordered. Just finish silently
1723 LDLM_DEBUG(lock, "lock is converted already!");
1724 unlock_res_and_lock(lock);
1726 if (ldlm_is_waited(lock))
1727 ldlm_del_waiting_lock(lock);
1729 ldlm_clear_cbpending(lock);
1730 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1731 ldlm_inodebits_drop(lock, bits & ~new_bits);
1733 ldlm_clear_blocking_data(lock);
1734 unlock_res_and_lock(lock);
1736 /* All old bits should be reprocessed to send new BL AST if
1737 * it wasn't sent earlier due to LDLM_FL_AST_SENT bit set.
1739 ldlm_reprocess_all(lock->l_resource, bits);
1742 dlm_rep->lock_handle = lock->l_remote_handle;
1743 ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1744 &dlm_rep->lock_desc.l_policy_data);
1748 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d", rc);
1749 LDLM_LOCK_PUT(lock);
1750 req->rq_status = rc;
1755 * Cancel all the locks whose handles are packed into ldlm_request
1757 * Called by server code expecting such combined cancel activity
1760 int ldlm_request_cancel(struct ptlrpc_request *req,
1761 const struct ldlm_request *dlm_req,
1762 int first, enum lustre_at_flags flags)
1764 struct ldlm_resource *res, *pres = NULL;
1765 struct ldlm_lock *lock;
1766 int i, count, done = 0;
1771 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
1772 if (size <= offsetof(struct ldlm_request, lock_handle) ||
1773 (size - offsetof(struct ldlm_request, lock_handle)) /
1774 sizeof(struct lustre_handle) < dlm_req->lock_count)
1777 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1781 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1785 * There is no lock on the server at the replay time,
1786 * skip lock cancelling to make replay tests to pass.
1788 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1791 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, starting at %d",
1794 for (i = first; i < count; i++) {
1795 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1797 /* below message checked in replay-single.sh test_36 */
1798 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (cookie %llu)",
1799 dlm_req->lock_handle[i].cookie);
1803 res = lock->l_resource;
1807 * This code is an optimization to only attempt lock
1808 * granting on the resource (that could be CPU-expensive)
1809 * after we are done cancelling lock in that resource.
1813 ldlm_reprocess_all(pres, 0);
1814 LDLM_RESOURCE_DELREF(pres);
1815 ldlm_resource_putref(pres);
1818 ldlm_resource_getref(res);
1819 LDLM_RESOURCE_ADDREF(res);
1821 if (!ldlm_is_discard_data(lock))
1822 ldlm_lvbo_update(res, lock,
1828 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1829 lock->l_blast_sent != 0) {
1830 timeout_t delay = 0;
1832 if (ktime_get_seconds() > lock->l_blast_sent)
1833 delay = ktime_get_seconds() -
1836 "server cancels blocked lock after %ds",
1838 obd_at_measure(lock->l_export->exp_obd,
1839 &lock->l_export->exp_bl_lock_at,
1842 ldlm_lock_cancel(lock);
1843 LDLM_LOCK_PUT(lock);
1846 ldlm_reprocess_all(pres, 0);
1847 LDLM_RESOURCE_DELREF(pres);
1848 ldlm_resource_putref(pres);
1850 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1853 EXPORT_SYMBOL(ldlm_request_cancel);
1856 * Main LDLM entry point for server code to cancel locks.
1858 * Typically gets called from service handler on LDLM_CANCEL opc.
1860 int ldlm_handle_cancel(struct ptlrpc_request *req)
1862 struct ldlm_request *dlm_req;
1867 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1868 if (dlm_req == NULL) {
1869 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1873 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) <
1874 offsetof(struct ldlm_request, lock_handle[1]))
1877 if (req->rq_export && req->rq_export->exp_nid_stats &&
1878 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1879 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1880 LDLM_CANCEL - LDLM_FIRST_OPC);
1882 rc = req_capsule_server_pack(&req->rq_pill);
1886 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1887 req->rq_status = LUSTRE_ESTALE;
1889 RETURN(ptlrpc_reply(req));
1891 #endif /* HAVE_SERVER_SUPPORT */
1894 * Server may pass additional information about blocking lock.
1895 * For IBITS locks it is conflicting bits which can be used for
1896 * lock convert instead of cancel.
1898 void ldlm_bl_desc2lock(const struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1900 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
1902 check_res_locked(lock->l_resource);
1903 if (ns_is_client(ns) && ld &&
1904 (lock->l_resource->lr_type == LDLM_IBITS)) {
1906 * Lock description contains policy of blocking lock,
1907 * and its cancel_bits is used to pass conflicting bits.
1908 * NOTE: ld can be NULL or can be not NULL but zeroed if
1909 * passed from ldlm_bl_thread_blwi(), check below used bits
1910 * in ld to make sure it is valid description.
1912 * If server may replace lock resource keeping the same cookie,
1913 * never use cancel bits from different resource, full cancel
1916 if (ld->l_policy_data.l_inodebits.cancel_bits &&
1917 ldlm_res_eq(&ld->l_resource.lr_name,
1918 &lock->l_resource->lr_name) &&
1919 !(ldlm_is_cbpending(lock) &&
1920 lock->l_policy_data.l_inodebits.cancel_bits == 0)) {
1921 /* always combine conflicting ibits */
1922 lock->l_policy_data.l_inodebits.cancel_bits |=
1923 ld->l_policy_data.l_inodebits.cancel_bits;
1925 /* If cancel_bits are not obtained or
1926 * if the lock is already CBPENDING and
1927 * has no cancel_bits set
1928 * - the full lock is to be cancelled
1930 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1936 * Callback handler for receiving incoming blocking ASTs.
1938 * This can only happen on client side.
1940 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1941 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1947 LDLM_DEBUG(lock, "client blocking AST callback handler");
1949 lock_res_and_lock(lock);
1951 /* get extra information from desc if any */
1952 ldlm_bl_desc2lock(ld, lock);
1953 ldlm_set_cbpending(lock);
1955 do_ast = (!lock->l_readers && !lock->l_writers);
1956 unlock_res_and_lock(lock);
1960 "Lock %p already unused, calling callback (%p)\n",
1961 lock, lock->l_blocking_ast);
1962 if (lock->l_blocking_ast != NULL)
1963 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1967 "Lock %p is referenced, will be cancelled later\n",
1971 LDLM_DEBUG(lock, "client blocking callback handler END");
1972 LDLM_LOCK_RELEASE(lock);
1976 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1978 if (req->rq_no_reply)
1981 req->rq_status = rc;
1982 if (!req->rq_packed_final) {
1983 rc = lustre_pack_reply(req, 1, NULL, NULL);
1987 return ptlrpc_reply(req);
1991 * Callback handler for receiving incoming completion ASTs.
1993 * This only can happen on client side.
1995 static int ldlm_handle_cp_callback(struct ptlrpc_request *req,
1996 struct ldlm_namespace *ns,
1997 struct ldlm_request *dlm_req,
1998 struct ldlm_lock *lock)
2000 LIST_HEAD(ast_list);
2006 LDLM_DEBUG(lock, "client completion callback handler START");
2008 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
2009 long to = cfs_time_seconds(1);
2011 ldlm_callback_reply(req, 0);
2014 to = schedule_timeout_interruptible(to);
2015 if (ldlm_is_granted(lock) ||
2016 ldlm_is_destroyed(lock))
2021 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
2023 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
2024 GOTO(out, rc = lvb_len);
2025 } else if (lvb_len > 0) {
2026 if (lock->l_lvb_len > 0) {
2027 /* for extent lock, lvb contains ost_lvb{}. */
2028 LASSERT(lock->l_lvb_data != NULL);
2030 if (unlikely(lock->l_lvb_len < lvb_len)) {
2032 "Replied LVB is larger than expectation, expected = %d, replied = %d",
2033 lock->l_lvb_len, lvb_len);
2034 GOTO(out, rc = -EINVAL);
2039 lock_res_and_lock(lock);
2041 if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
2042 &lock->l_resource->lr_name)) {
2043 ldlm_resource_unlink_lock(lock);
2044 unlock_res_and_lock(lock);
2045 rc = ldlm_lock_change_resource(ns, lock,
2046 &dlm_req->lock_desc.l_resource.lr_name);
2048 LDLM_ERROR(lock, "Failed to allocate resource");
2051 LDLM_DEBUG(lock, "completion AST, new resource");
2052 lock_res_and_lock(lock);
2055 if (ldlm_is_failed(lock)) {
2056 unlock_res_and_lock(lock);
2057 LDLM_LOCK_RELEASE(lock);
2061 if (ldlm_is_destroyed(lock) ||
2062 ldlm_is_granted(lock)) {
2063 /* b=11300: the lock has already been granted */
2064 unlock_res_and_lock(lock);
2065 LDLM_DEBUG(lock, "Double grant race happened");
2070 * If we receive the completion AST before the actual enqueue returned,
2071 * then we might need to switch lock modes, resources, or extents.
2073 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
2074 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
2075 LDLM_DEBUG(lock, "completion AST, new lock mode");
2078 if (lock->l_resource->lr_type != LDLM_PLAIN) {
2079 ldlm_convert_policy_to_local(req->rq_export,
2080 dlm_req->lock_desc.l_resource.lr_type,
2081 &dlm_req->lock_desc.l_policy_data,
2082 &lock->l_policy_data);
2083 LDLM_DEBUG(lock, "completion AST, new policy data");
2086 ldlm_resource_unlink_lock(lock);
2088 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
2090 * BL_AST locks are not needed in LRU.
2091 * Let ldlm_cancel_lru() be fast.
2093 ldlm_lock_remove_from_lru(lock);
2094 ldlm_bl_desc2lock(&dlm_req->lock_desc, lock);
2095 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
2096 LDLM_DEBUG(lock, "completion AST includes blocking AST");
2099 if (lock->l_lvb_len > 0) {
2100 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
2101 lock->l_lvb_data, lvb_len);
2103 unlock_res_and_lock(lock);
2108 ldlm_grant_lock(lock, &ast_list);
2109 unlock_res_and_lock(lock);
2111 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
2114 * Let Enqueue to call osc_lock_upcall() and initialize
2117 CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
2119 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
2121 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
2127 lock_res_and_lock(lock);
2128 ldlm_set_failed(lock);
2129 unlock_res_and_lock(lock);
2130 wake_up(&lock->l_waitq);
2132 LDLM_LOCK_RELEASE(lock);
2138 * Callback handler for receiving incoming glimpse ASTs.
2140 * This only can happen on client side. After handling the glimpse AST
2141 * we also consider dropping the lock here if it is unused locally for a
2144 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
2145 struct ldlm_namespace *ns,
2146 struct ldlm_request *dlm_req,
2147 struct ldlm_lock *lock)
2149 struct ldlm_lock_desc *ld = &dlm_req->lock_desc;
2154 LDLM_DEBUG(lock, "client glimpse AST callback handler");
2156 if (lock->l_glimpse_ast != NULL)
2157 rc = lock->l_glimpse_ast(lock, req);
2159 if (req->rq_repmsg != NULL) {
2162 req->rq_status = rc;
2166 lock_res_and_lock(lock);
2167 if (lock->l_granted_mode == LCK_PW &&
2168 !lock->l_readers && !lock->l_writers &&
2169 ktime_after(ktime_get(),
2170 ktime_add(lock->l_last_used, ns->ns_dirty_age_limit))) {
2171 unlock_res_and_lock(lock);
2173 /* For MDS glimpse it is always DOM lock, set corresponding
2174 * cancel_bits to perform lock convert if needed
2176 if (lock->l_resource->lr_type == LDLM_IBITS)
2177 ld->l_policy_data.l_inodebits.cancel_bits =
2179 if (ldlm_bl_to_thread_lock(ns, ld, lock))
2180 ldlm_handle_bl_callback(ns, ld, lock);
2185 unlock_res_and_lock(lock);
2186 LDLM_LOCK_RELEASE(lock);
2190 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
2191 enum ldlm_cancel_flags cancel_flags)
2193 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2194 char *prio = "regular";
2199 spin_lock(&blp->blp_lock);
2200 /* cannot access blwi after added to list and lock is dropped */
2201 count = blwi->blwi_lock ? 1 : blwi->blwi_count;
2203 /* if the server is waiting on a lock to be cancelled (bl_ast), this is
2204 * an urgent request and should go in the priority queue so it doesn't
2205 * get stuck behind non-priority work (eg, lru size management)
2207 * We also prioritize discard_data, which is for eviction handling
2209 if (blwi->blwi_lock &&
2210 (ldlm_is_discard_data(blwi->blwi_lock) ||
2211 ldlm_is_bl_ast(blwi->blwi_lock))) {
2212 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
2215 /* other blocking callbacks are added to the regular list */
2216 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
2218 blp->blp_total_locks += count;
2219 blp->blp_total_blwis++;
2220 spin_unlock(&blp->blp_lock);
2222 wake_up(&blp->blp_waitq);
2224 /* unlocked read of blp values is intentional - OK for debug */
2226 "added %d/%d locks to %s blp list, %d blwis in pool\n",
2227 count, blp->blp_total_locks, prio, blp->blp_total_blwis);
2230 * can not check blwi->blwi_flags as blwi could be already freed in
2233 if (!(cancel_flags & LCF_ASYNC))
2234 wait_for_completion(&blwi->blwi_comp);
2239 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2240 struct ldlm_namespace *ns,
2241 struct ldlm_lock_desc *ld,
2242 struct list_head *cancels, int count,
2243 struct ldlm_lock *lock,
2244 enum ldlm_cancel_flags cancel_flags)
2246 init_completion(&blwi->blwi_comp);
2247 INIT_LIST_HEAD(&blwi->blwi_head);
2249 if (current->flags & PF_MEMALLOC)
2250 blwi->blwi_mem_pressure = 1;
2253 blwi->blwi_flags = cancel_flags;
2255 blwi->blwi_ld = *ld;
2257 list_splice_init(cancels, &blwi->blwi_head);
2258 blwi->blwi_count = count;
2260 blwi->blwi_lock = lock;
2265 * Queues a list of locks \a cancels containing \a count locks
2266 * for later processing by a blocking thread. If \a count is zero,
2267 * then the lock referenced as \a lock is queued instead.
2269 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2270 * If list addition fails an error is returned and caller is supposed to
2271 * call ->l_blocking_ast itself.
2273 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2274 struct ldlm_lock_desc *ld,
2275 struct ldlm_lock *lock,
2276 struct list_head *cancels, int count,
2277 enum ldlm_cancel_flags cancel_flags)
2281 if (cancels && count == 0)
2284 if (cancel_flags & LCF_ASYNC) {
2285 struct ldlm_bl_work_item *blwi;
2287 OBD_ALLOC(blwi, sizeof(*blwi));
2290 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2292 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2295 * if it is synchronous call do minimum mem alloc, as it could
2296 * be triggered from kernel shrinker
2298 struct ldlm_bl_work_item blwi;
2300 memset(&blwi, 0, sizeof(blwi));
2301 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2302 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2307 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2308 struct ldlm_lock *lock)
2310 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2313 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2314 struct list_head *cancels, int count,
2315 enum ldlm_cancel_flags cancel_flags)
2317 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2320 int ldlm_bl_to_thread_ns(struct ldlm_namespace *ns)
2322 return ldlm_bl_to_thread(ns, NULL, NULL, NULL, 0, LCF_ASYNC);
2325 int ldlm_bl_thread_wakeup(void)
2327 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2331 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2332 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2334 struct obd_device *obd = req->rq_export->exp_obd;
2342 DEBUG_REQ(D_HSM, req, "%s: handle setinfo", obd->obd_name);
2344 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2346 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2348 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2351 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2353 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2355 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2358 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2361 /* We are responsible for swabbing contents of val */
2363 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2364 /* Pass it on to mdc (the "export" in this case) */
2365 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2367 sizeof(KEY_HSM_COPYTOOL_SEND),
2368 KEY_HSM_COPYTOOL_SEND,
2371 DEBUG_REQ(D_WARNING, req, "ignoring unknown key '%s'", key);
2376 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2377 const char *msg, int rc,
2378 const struct lustre_handle *handle)
2380 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2381 "%s, NID=%s lock=%#llx: rc = %d",
2382 msg, libcfs_idstr(&req->rq_peer),
2383 handle ? handle->cookie : 0, rc);
2384 if (req->rq_no_reply)
2385 CWARN("No reply was sent, maybe cause b=21636.\n");
2387 CWARN("Send reply failed, maybe cause b=21636.\n");
2390 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2391 static int ldlm_callback_handler(struct ptlrpc_request *req)
2393 struct ldlm_namespace *ns;
2394 struct ldlm_request *dlm_req;
2395 struct ldlm_lock *lock;
2401 * Requests arrive in sender's byte order. The ptlrpc service
2402 * handler has already checked and, if necessary, byte-swapped the
2403 * incoming request message body, but I am responsible for the
2407 /* do nothing for sec context finalize */
2408 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2411 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2413 if (req->rq_export == NULL) {
2414 rc = ldlm_callback_reply(req, -ENOTCONN);
2415 ldlm_callback_errmsg(req, "Operate on unconnected server",
2420 LASSERT(req->rq_export != NULL);
2421 LASSERT(req->rq_export->exp_obd != NULL);
2423 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2424 case LDLM_BL_CALLBACK:
2425 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2427 ldlm_callback_reply(req, -(int)cfs_fail_err);
2431 case LDLM_CP_CALLBACK:
2432 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2435 case LDLM_GL_CALLBACK:
2436 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2440 rc = ldlm_handle_setinfo(req);
2441 ldlm_callback_reply(req, rc);
2444 CERROR("unknown opcode %u\n",
2445 lustre_msg_get_opc(req->rq_reqmsg));
2446 ldlm_callback_reply(req, -EPROTO);
2450 ns = req->rq_export->exp_obd->obd_namespace;
2451 LASSERT(ns != NULL);
2453 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2455 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2456 if (dlm_req == NULL) {
2457 rc = ldlm_callback_reply(req, -EPROTO);
2458 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2464 * Force a known safe race, send a cancel to the server for a lock
2465 * which the server has already started a blocking callback on.
2467 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2468 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2469 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2471 CERROR("ldlm_cli_cancel: %d\n", rc);
2474 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2477 "callback on lock %#llx - lock disappeared\n",
2478 dlm_req->lock_handle[0].cookie);
2479 rc = ldlm_callback_reply(req, -EINVAL);
2480 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2481 &dlm_req->lock_handle[0]);
2485 if (ldlm_is_fail_loc(lock) &&
2486 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2487 CFS_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2489 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2490 lock_res_and_lock(lock);
2491 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2493 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2495 * If somebody cancels lock and cache is already dropped,
2496 * or lock is failed before cp_ast received on client,
2497 * we can tell the server we have no lock. Otherwise, we
2498 * should send cancel after dropping the cache.
2500 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2501 ldlm_is_failed(lock)) {
2503 "callback on lock %llx - lock disappeared",
2504 dlm_req->lock_handle[0].cookie);
2505 unlock_res_and_lock(lock);
2506 LDLM_LOCK_RELEASE(lock);
2507 rc = ldlm_callback_reply(req, -EINVAL);
2508 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2509 &dlm_req->lock_handle[0]);
2513 * BL_AST locks are not needed in LRU.
2514 * Let ldlm_cancel_lru() be fast.
2516 ldlm_lock_remove_from_lru(lock);
2517 ldlm_set_bl_ast(lock);
2519 if (lock->l_remote_handle.cookie == 0)
2520 lock->l_remote_handle = dlm_req->lock_handle[1];
2521 unlock_res_and_lock(lock);
2524 * We want the ost thread to get this reply so that it can respond
2525 * to ost requests (write cache writeback) that might be triggered
2528 * But we'd also like to be able to indicate in the reply that we're
2529 * cancelling right now, because it's unused, or have an intent result
2530 * in the reply, so we might have to push the responsibility for sending
2531 * the reply down into the AST handlers, alas.
2534 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2535 case LDLM_BL_CALLBACK:
2536 LDLM_DEBUG(lock, "blocking ast");
2537 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2538 if (!ldlm_is_cancel_on_block(lock)) {
2539 rc = ldlm_callback_reply(req, 0);
2540 if (req->rq_no_reply || rc)
2541 ldlm_callback_errmsg(req, "Normal process", rc,
2542 &dlm_req->lock_handle[0]);
2544 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2545 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2547 case LDLM_CP_CALLBACK:
2548 LDLM_DEBUG(lock, "completion ast");
2549 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2550 rc = ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2551 if (!CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE))
2552 ldlm_callback_reply(req, rc);
2554 case LDLM_GL_CALLBACK:
2555 LDLM_DEBUG(lock, "glimpse ast");
2556 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2557 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2560 LBUG(); /* checked above */
2566 #ifdef HAVE_SERVER_SUPPORT
2568 * Main handler for canceld thread.
2570 * Separated into its own thread to avoid deadlocks.
2572 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2579 * Requests arrive in sender's byte order. The ptlrpc service
2580 * handler has already checked and, if necessary, byte-swapped the
2581 * incoming request message body, but I am responsible for the
2585 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2587 if (req->rq_export == NULL) {
2588 struct ldlm_request *dlm_req;
2590 CERROR("%s from %s arrived at %llu with bad export cookie %llu\n",
2591 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2592 libcfs_nidstr(&req->rq_peer.nid),
2593 (unsigned long long)req->rq_arrival_time.tv_sec,
2594 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2596 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2597 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2598 dlm_req = req_capsule_client_get(&req->rq_pill,
2600 if (dlm_req != NULL)
2601 ldlm_lock_dump_handle(D_ERROR,
2602 &dlm_req->lock_handle[0]);
2604 ldlm_callback_reply(req, -ENOTCONN);
2608 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2609 /* XXX FIXME move this back to mds/handler.c, b=249 */
2611 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2612 CDEBUG(D_INODE, "cancel\n");
2613 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2614 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2615 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2617 rc = ldlm_handle_cancel(req);
2621 struct ldlm_request *dlm_req;
2623 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2624 CDEBUG(D_INODE, "convert\n");
2626 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2627 if (dlm_req == NULL) {
2628 CDEBUG(D_INFO, "bad request buffer for cancel\n");
2629 rc = ldlm_callback_reply(req, -EPROTO);
2631 req->rq_status = ldlm_handle_convert0(req, dlm_req);
2632 rc = ptlrpc_reply(req);
2637 CERROR("invalid opcode %d\n",
2638 lustre_msg_get_opc(req->rq_reqmsg));
2639 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2640 rc = ldlm_callback_reply(req, -EINVAL);
2646 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2647 struct ldlm_lock *lock)
2649 struct ldlm_request *dlm_req;
2650 struct lustre_handle lockh;
2656 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2657 if (dlm_req == NULL)
2660 ldlm_lock2handle(lock, &lockh);
2661 for (i = 0; i < dlm_req->lock_count; i++) {
2662 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2664 DEBUG_REQ(D_RPCTRACE, req,
2665 "Prio raised by lock %#llx", lockh.cookie);
2674 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2676 struct ldlm_request *dlm_req;
2683 /* no prolong in recovery */
2684 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2687 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2688 if (dlm_req == NULL)
2691 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
2692 if (size <= offsetof(struct ldlm_request, lock_handle) ||
2693 (size - offsetof(struct ldlm_request, lock_handle)) /
2694 sizeof(struct lustre_handle) < dlm_req->lock_count)
2697 for (i = 0; i < dlm_req->lock_count; i++) {
2698 struct ldlm_lock *lock;
2700 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2704 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2706 LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2707 LDLM_LOCK_PUT(lock);
2716 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2717 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2718 .hpreq_check = ldlm_cancel_hpreq_check,
2722 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2726 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2728 if (req->rq_export == NULL)
2731 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2732 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2733 req->rq_ops = &ldlm_cancel_hpreq_ops;
2734 } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2735 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2736 req->rq_ops = &ldlm_cancel_hpreq_ops;
2741 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2742 struct hlist_node *hnode, void *data)
2745 struct list_head *rpc_list = data;
2746 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2748 lock_res_and_lock(lock);
2750 if (!ldlm_is_granted(lock)) {
2751 unlock_res_and_lock(lock);
2755 LASSERT(lock->l_resource);
2756 if (lock->l_resource->lr_type != LDLM_IBITS &&
2757 lock->l_resource->lr_type != LDLM_PLAIN) {
2758 unlock_res_and_lock(lock);
2762 if (ldlm_is_ast_sent(lock)) {
2763 unlock_res_and_lock(lock);
2767 LASSERT(lock->l_blocking_ast);
2768 LASSERT(!lock->l_blocking_lock);
2770 ldlm_set_ast_sent(lock);
2771 if (lock->l_export && lock->l_export->exp_lock_hash) {
2773 * NB: it's safe to call cfs_hash_del() even lock isn't
2777 * In the function below, .hs_keycmp resolves to
2778 * ldlm_export_lock_keycmp()
2780 cfs_hash_del(lock->l_export->exp_lock_hash,
2781 &lock->l_remote_handle, &lock->l_exp_hash);
2784 list_add_tail(&lock->l_rk_ast, rpc_list);
2785 LDLM_LOCK_GET(lock);
2787 unlock_res_and_lock(lock);
2791 void ldlm_revoke_export_locks(struct obd_export *exp)
2794 LIST_HEAD(rpc_list);
2798 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2799 ldlm_revoke_lock_cb, &rpc_list, 0);
2800 rc = ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2801 LDLM_WORK_REVOKE_AST);
2803 if (rc == -ERESTART)
2804 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2808 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2809 #endif /* HAVE_SERVER_SUPPORT */
2811 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2812 struct ldlm_bl_work_item **p_blwi,
2813 struct obd_export **p_exp)
2815 struct ldlm_bl_work_item *blwi = NULL;
2816 static unsigned int num_bl;
2817 static unsigned int num_stale;
2818 int num_th = atomic_read(&blp->blp_num_threads);
2820 *p_exp = obd_stale_export_get();
2822 spin_lock(&blp->blp_lock);
2823 if (*p_exp != NULL) {
2824 if (num_th == 1 || ++num_stale < num_th) {
2825 spin_unlock(&blp->blp_lock);
2831 /* process a request from the blp_list at least every blp_num_threads */
2832 if (!list_empty(&blp->blp_list) &&
2833 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2834 blwi = list_first_entry(&blp->blp_list,
2835 struct ldlm_bl_work_item, blwi_entry);
2837 if (!list_empty(&blp->blp_prio_list))
2838 blwi = list_first_entry(&blp->blp_prio_list,
2839 struct ldlm_bl_work_item,
2843 if (++num_bl >= num_th)
2845 list_del(&blwi->blwi_entry);
2846 blp->blp_total_locks -= blwi->blwi_lock ? 1 : blwi->blwi_count;
2847 blp->blp_total_blwis--;
2849 spin_unlock(&blp->blp_lock);
2852 /* intentional unlocked read of blp values - OK for debug */
2855 "Got %d locks of %d total in blp. (%d blwis in pool)\n",
2856 blwi->blwi_lock ? 1 : blwi->blwi_count,
2857 blp->blp_total_locks, blp->blp_total_blwis);
2860 "No blwi found in queue (no bl locks in queue)\n");
2863 if (*p_exp != NULL && *p_blwi != NULL) {
2864 obd_stale_export_put(*p_exp);
2868 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2871 /* This only contains temporary data until the thread starts */
2872 struct ldlm_bl_thread_data {
2873 struct ldlm_bl_pool *bltd_blp;
2874 struct completion bltd_comp;
2878 static int ldlm_bl_thread_main(void *arg);
2880 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2882 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2883 struct task_struct *task;
2885 init_completion(&bltd.bltd_comp);
2887 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2888 if (bltd.bltd_num >= blp->blp_max_threads) {
2889 atomic_dec(&blp->blp_num_threads);
2893 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2895 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2896 atomic_dec(&blp->blp_num_threads);
2900 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2903 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2904 bltd.bltd_num, PTR_ERR(task));
2905 atomic_dec(&blp->blp_num_threads);
2906 return PTR_ERR(task);
2908 wait_for_completion(&bltd.bltd_comp);
2913 /* Not fatal if racy and have a few too many threads */
2914 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2915 struct ldlm_bl_work_item *blwi)
2917 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2920 if (atomic_read(&blp->blp_busy_threads) <
2921 atomic_read(&blp->blp_num_threads))
2924 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2925 blwi->blwi_mem_pressure))
2931 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2932 struct ldlm_bl_work_item *blwi)
2934 /* '1' for consistency with code that checks !mpflag to restore */
2935 unsigned int mpflags = 1;
2939 if (blwi->blwi_ns == NULL)
2940 /* added by ldlm_cleanup() */
2941 RETURN(LDLM_ITER_STOP);
2943 if (blwi->blwi_mem_pressure)
2944 mpflags = memalloc_noreclaim_save();
2946 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2948 if (blwi->blwi_count) {
2951 * The special case when we cancel locks in lru
2952 * asynchronously, we pass the list of locks here.
2953 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2954 * canceled locally yet.
2956 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2959 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2961 } else if (blwi->blwi_lock) {
2962 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2965 ldlm_pool_recalc(&blwi->blwi_ns->ns_pool, true);
2966 spin_lock(&blwi->blwi_ns->ns_lock);
2967 blwi->blwi_ns->ns_rpc_recalc = 0;
2968 spin_unlock(&blwi->blwi_ns->ns_lock);
2969 ldlm_namespace_put(blwi->blwi_ns);
2972 if (blwi->blwi_mem_pressure)
2973 memalloc_noreclaim_restore(mpflags);
2975 if (blwi->blwi_flags & LCF_ASYNC)
2976 OBD_FREE(blwi, sizeof(*blwi));
2978 complete(&blwi->blwi_comp);
2984 * Cancel stale locks on export. Cancel blocked locks first.
2985 * If the given export has blocked locks, the next in the list may have
2986 * them too, thus cancel not blocked locks only if the current export has
2989 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2990 struct obd_export *exp)
2996 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2998 num = ldlm_export_cancel_blocked_locks(exp);
3000 ldlm_export_cancel_locks(exp);
3002 obd_stale_export_put(exp);
3009 * Main blocking requests processing thread.
3011 * Callers put locks into its queue by calling ldlm_bl_to_thread.
3012 * This thread in the end ends up doing actual call to ->l_blocking_ast
3015 static int ldlm_bl_thread_main(void *arg)
3018 struct ldlm_bl_pool *blp;
3019 struct ldlm_bl_thread_data *bltd = arg;
3027 rc = lu_env_init(env, LCT_DT_THREAD);
3030 rc = lu_env_add(env);
3032 GOTO(out_env_fini, rc);
3034 blp = bltd->bltd_blp;
3036 complete(&bltd->bltd_comp);
3037 /* cannot use bltd after this, it is only on caller's stack */
3040 struct ldlm_bl_work_item *blwi = NULL;
3041 struct obd_export *exp = NULL;
3044 rc = ldlm_bl_get_work(blp, &blwi, &exp);
3047 wait_event_idle_exclusive(blp->blp_waitq,
3048 ldlm_bl_get_work(blp, &blwi,
3050 atomic_inc(&blp->blp_busy_threads);
3052 if (ldlm_bl_thread_need_create(blp, blwi))
3053 /* discard the return value, we tried */
3054 ldlm_bl_thread_start(blp, true);
3057 rc = ldlm_bl_thread_exports(blp, exp);
3059 rc = ldlm_bl_thread_blwi(blp, blwi);
3061 atomic_dec(&blp->blp_busy_threads);
3063 if (rc == LDLM_ITER_STOP)
3067 * If there are many namespaces, we will not sleep waiting for
3068 * work, and must do a cond_resched to avoid holding the CPU
3074 atomic_dec(&blp->blp_num_threads);
3075 complete(&blp->blp_comp);
3086 static int ldlm_setup(void);
3087 static int ldlm_cleanup(void);
3089 int ldlm_get_ref(void)
3094 mutex_lock(&ldlm_ref_mutex);
3095 if (++ldlm_refcount == 1) {
3100 mutex_unlock(&ldlm_ref_mutex);
3105 void ldlm_put_ref(void)
3108 mutex_lock(&ldlm_ref_mutex);
3109 if (ldlm_refcount == 1) {
3110 int rc = ldlm_cleanup();
3113 CERROR("ldlm_cleanup failed: %d\n", rc);
3119 mutex_unlock(&ldlm_ref_mutex);
3124 /* Export handle<->lock hash operations. */
3126 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key,
3127 const unsigned int bits)
3129 return cfs_hash_64(((struct lustre_handle *)key)->cookie, bits);
3133 ldlm_export_lock_key(struct hlist_node *hnode)
3135 struct ldlm_lock *lock;
3137 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3138 return &lock->l_remote_handle;
3142 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
3144 struct ldlm_lock *lock;
3146 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3147 lock->l_remote_handle = *(struct lustre_handle *)key;
3151 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
3153 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
3157 ldlm_export_lock_object(struct hlist_node *hnode)
3159 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3163 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
3165 struct ldlm_lock *lock;
3167 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3168 LDLM_LOCK_GET(lock);
3172 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
3174 struct ldlm_lock *lock;
3176 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3177 LDLM_LOCK_RELEASE(lock);
3180 static struct cfs_hash_ops ldlm_export_lock_ops = {
3181 .hs_hash = ldlm_export_lock_hash,
3182 .hs_key = ldlm_export_lock_key,
3183 .hs_keycmp = ldlm_export_lock_keycmp,
3184 .hs_keycpy = ldlm_export_lock_keycpy,
3185 .hs_object = ldlm_export_lock_object,
3186 .hs_get = ldlm_export_lock_get,
3187 .hs_put = ldlm_export_lock_put,
3188 .hs_put_locked = ldlm_export_lock_put,
3191 int ldlm_init_export(struct obd_export *exp)
3197 exp->exp_lock_hash =
3198 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
3199 HASH_EXP_LOCK_CUR_BITS,
3200 HASH_EXP_LOCK_MAX_BITS,
3201 HASH_EXP_LOCK_BKT_BITS, 0,
3202 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
3203 &ldlm_export_lock_ops,
3204 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
3205 CFS_HASH_NBLK_CHANGE);
3207 if (!exp->exp_lock_hash)
3210 rc = ldlm_init_flock_export(exp);
3216 ldlm_destroy_export(exp);
3219 EXPORT_SYMBOL(ldlm_init_export);
3221 void ldlm_destroy_export(struct obd_export *exp)
3224 cfs_hash_putref(exp->exp_lock_hash);
3225 exp->exp_lock_hash = NULL;
3227 ldlm_destroy_flock_export(exp);
3230 EXPORT_SYMBOL(ldlm_destroy_export);
3232 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
3233 struct attribute *attr,
3236 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
3239 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
3240 struct attribute *attr,
3247 rc = kstrtoul(buffer, 10, &val);
3251 ldlm_cancel_unused_locks_before_replay = val;
3255 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
3257 static struct attribute *ldlm_attrs[] = {
3258 &lustre_attr_cancel_unused_locks_before_replay.attr,
3262 static struct attribute_group ldlm_attr_group = {
3263 .attrs = ldlm_attrs,
3266 static int ldlm_setup(void)
3268 static struct ptlrpc_service_conf conf;
3269 struct ldlm_bl_pool *blp = NULL;
3270 #ifdef HAVE_SERVER_SUPPORT
3271 struct task_struct *task;
3272 #endif /* HAVE_SERVER_SUPPORT */
3278 if (ldlm_state != NULL)
3281 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
3282 if (ldlm_state == NULL)
3285 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
3289 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
3293 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
3297 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
3301 rc = ldlm_debugfs_setup();
3305 memset(&conf, 0, sizeof(conf));
3306 conf = (typeof(conf)) {
3307 .psc_name = "ldlm_cbd",
3308 .psc_watchdog_factor = 2,
3310 .bc_nbufs = LDLM_CLIENT_NBUFS,
3311 .bc_buf_size = LDLM_BUFSIZE,
3312 .bc_req_max_size = LDLM_MAXREQSIZE,
3313 .bc_rep_max_size = LDLM_MAXREPSIZE,
3314 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
3315 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
3318 .tc_thr_name = "ldlm_cb",
3319 .tc_thr_factor = LDLM_THR_FACTOR,
3320 .tc_nthrs_init = LDLM_NTHRS_INIT,
3321 .tc_nthrs_base = LDLM_NTHRS_BASE,
3322 .tc_nthrs_max = LDLM_NTHRS_MAX,
3323 .tc_nthrs_user = ldlm_num_threads,
3324 .tc_cpu_bind = ldlm_cpu_bind,
3325 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
3328 .cc_pattern = ldlm_cpts,
3329 .cc_affinity = true,
3332 .so_req_handler = ldlm_callback_handler,
3335 ldlm_state->ldlm_cb_service = ptlrpc_register_service(&conf,
3337 ldlm_svc_debugfs_dir);
3338 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3339 CERROR("failed to start service\n");
3340 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3341 ldlm_state->ldlm_cb_service = NULL;
3345 #ifdef HAVE_SERVER_SUPPORT
3346 memset(&conf, 0, sizeof(conf));
3347 conf = (typeof(conf)) {
3348 .psc_name = "ldlm_canceld",
3349 .psc_watchdog_factor = 6,
3351 .bc_nbufs = LDLM_SERVER_NBUFS,
3352 .bc_buf_size = LDLM_BUFSIZE,
3353 .bc_req_max_size = LDLM_MAXREQSIZE,
3354 .bc_rep_max_size = LDLM_MAXREPSIZE,
3355 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3356 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3360 .tc_thr_name = "ldlm_cn",
3361 .tc_thr_factor = LDLM_THR_FACTOR,
3362 .tc_nthrs_init = LDLM_NTHRS_INIT,
3363 .tc_nthrs_base = LDLM_NTHRS_BASE,
3364 .tc_nthrs_max = LDLM_NTHRS_MAX,
3365 .tc_nthrs_user = ldlm_num_threads,
3366 .tc_cpu_bind = ldlm_cpu_bind,
3367 .tc_ctx_tags = LCT_MD_THREAD |
3372 .cc_pattern = ldlm_cpts,
3373 .cc_affinity = true,
3376 .so_req_handler = ldlm_cancel_handler,
3377 .so_hpreq_handler = ldlm_hpreq_handler,
3380 ldlm_state->ldlm_cancel_service =
3381 ptlrpc_register_service(&conf, ldlm_svc_kset,
3382 ldlm_svc_debugfs_dir);
3383 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3384 CERROR("failed to start service\n");
3385 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3386 ldlm_state->ldlm_cancel_service = NULL;
3389 #endif /* HAVE_SERVER_SUPPORT */
3391 OBD_ALLOC(blp, sizeof(*blp));
3393 GOTO(out, rc = -ENOMEM);
3394 ldlm_state->ldlm_bl_pool = blp;
3396 spin_lock_init(&blp->blp_lock);
3397 INIT_LIST_HEAD(&blp->blp_list);
3398 INIT_LIST_HEAD(&blp->blp_prio_list);
3399 init_waitqueue_head(&blp->blp_waitq);
3400 atomic_set(&blp->blp_num_threads, 0);
3401 atomic_set(&blp->blp_busy_threads, 0);
3402 blp->blp_total_locks = 0;
3403 blp->blp_total_blwis = 0;
3405 if (ldlm_num_threads == 0) {
3406 blp->blp_min_threads = LDLM_NTHRS_INIT;
3407 blp->blp_max_threads = LDLM_NTHRS_MAX;
3409 blp->blp_min_threads = blp->blp_max_threads =
3410 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3414 for (i = 0; i < blp->blp_min_threads; i++) {
3415 rc = ldlm_bl_thread_start(blp, false);
3420 #ifdef HAVE_SERVER_SUPPORT
3421 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3424 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3428 wait_event(expired_lock_wait_queue,
3429 expired_lock_thread_state == ELT_READY);
3430 #endif /* HAVE_SERVER_SUPPORT */
3432 rc = ldlm_pools_init();
3434 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3438 rc = ldlm_reclaim_setup();
3440 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3450 static int ldlm_cleanup(void)
3454 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3455 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3456 CERROR("ldlm still has namespaces; clean these up first.\n");
3457 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3458 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3462 ldlm_reclaim_cleanup();
3465 if (ldlm_state->ldlm_bl_pool != NULL) {
3466 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3468 while (atomic_read(&blp->blp_num_threads) > 0) {
3469 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3471 init_completion(&blp->blp_comp);
3473 spin_lock(&blp->blp_lock);
3474 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3475 wake_up(&blp->blp_waitq);
3476 spin_unlock(&blp->blp_lock);
3478 wait_for_completion(&blp->blp_comp);
3481 OBD_FREE(blp, sizeof(*blp));
3484 if (ldlm_state->ldlm_cb_service != NULL)
3485 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3486 #ifdef HAVE_SERVER_SUPPORT
3487 if (ldlm_state->ldlm_cancel_service != NULL)
3488 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3492 kset_unregister(ldlm_ns_kset);
3494 kset_unregister(ldlm_svc_kset);
3496 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3497 kobject_put(ldlm_kobj);
3500 ldlm_debugfs_cleanup();
3502 #ifdef HAVE_SERVER_SUPPORT
3503 if (expired_lock_thread_state != ELT_STOPPED) {
3504 expired_lock_thread_state = ELT_TERMINATE;
3505 wake_up(&expired_lock_wait_queue);
3506 wait_event(expired_lock_wait_queue,
3507 expired_lock_thread_state == ELT_STOPPED);
3511 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3519 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3520 sizeof(struct ldlm_resource), 0,
3521 SLAB_HWCACHE_ALIGN, NULL);
3522 if (ldlm_resource_slab == NULL)
3525 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3526 sizeof(struct ldlm_lock), 0,
3527 SLAB_HWCACHE_ALIGN, NULL);
3528 if (ldlm_lock_slab == NULL)
3531 ldlm_interval_slab = kmem_cache_create("interval_node",
3532 sizeof(struct ldlm_interval),
3533 0, SLAB_HWCACHE_ALIGN, NULL);
3534 if (ldlm_interval_slab == NULL)
3537 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3538 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3539 0, SLAB_HWCACHE_ALIGN, NULL);
3540 if (ldlm_interval_tree_slab == NULL)
3543 #ifdef HAVE_SERVER_SUPPORT
3544 ldlm_inodebits_slab = kmem_cache_create("ldlm_ibits_node",
3545 sizeof(struct ldlm_ibits_node),
3546 0, SLAB_HWCACHE_ALIGN, NULL);
3547 if (ldlm_inodebits_slab == NULL)
3548 goto out_interval_tree;
3550 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3551 sizeof(struct ldlm_glimpse_work),
3553 if (ldlm_glimpse_work_kmem == NULL)
3557 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3558 class_export_dump_hook = ldlm_dump_export_locks;
3561 #ifdef HAVE_SERVER_SUPPORT
3563 kmem_cache_destroy(ldlm_inodebits_slab);
3565 kmem_cache_destroy(ldlm_interval_tree_slab);
3568 kmem_cache_destroy(ldlm_interval_slab);
3570 kmem_cache_destroy(ldlm_lock_slab);
3572 kmem_cache_destroy(ldlm_resource_slab);
3577 void ldlm_exit(void)
3580 CERROR("ldlm_refcount is %d in %s\n", ldlm_refcount, __func__);
3582 kmem_cache_destroy(ldlm_resource_slab);
3584 * ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3585 * rcu_barrier() to wait all outstanding RCU callbacks to complete,
3586 * so that ldlm_lock_free() get a chance to be called.
3589 kmem_cache_destroy(ldlm_lock_slab);
3590 kmem_cache_destroy(ldlm_interval_slab);
3591 kmem_cache_destroy(ldlm_interval_tree_slab);
3592 #ifdef HAVE_SERVER_SUPPORT
3593 kmem_cache_destroy(ldlm_inodebits_slab);
3594 kmem_cache_destroy(ldlm_glimpse_work_kmem);