4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ldlm/ldlm_lockd.c
34 * Author: Peter Braam <braam@clusterfs.com>
35 * Author: Phil Schwan <phil@clusterfs.com>
38 #define DEBUG_SUBSYSTEM S_LDLM
40 #include <linux/kthread.h>
41 #include <linux/list.h>
42 #include <libcfs/libcfs.h>
43 #include <libcfs/linux/linux-mem.h>
44 #include <lustre_errno.h>
45 #include <lustre_dlm.h>
46 #include <obd_class.h>
47 #include "ldlm_internal.h"
49 static int ldlm_num_threads;
50 module_param(ldlm_num_threads, int, 0444);
51 MODULE_PARM_DESC(ldlm_num_threads, "number of DLM service threads to start");
53 static unsigned int ldlm_cpu_bind = 1;
54 module_param(ldlm_cpu_bind, uint, 0444);
55 MODULE_PARM_DESC(ldlm_cpu_bind,
56 "bind DLM service threads to particular CPU partitions");
58 static char *ldlm_cpts;
59 module_param(ldlm_cpts, charp, 0444);
60 MODULE_PARM_DESC(ldlm_cpts, "CPU partitions ldlm threads should run on");
62 static DEFINE_MUTEX(ldlm_ref_mutex);
63 static int ldlm_refcount;
65 struct kobject *ldlm_kobj;
66 struct kset *ldlm_ns_kset;
67 struct kset *ldlm_svc_kset;
71 static struct ldlm_state *ldlm_state;
74 * timeout for initial callback (AST) reply (bz10399)
75 * Due to having to send a 32 bit time value over the
76 * wire return it as timeout_t instead of time64_t
78 static inline timeout_t ldlm_get_rq_timeout(void)
81 timeout_t timeout = min(ldlm_timeout, obd_timeout / 3);
83 return timeout < 1 ? 1 : timeout;
90 * blp_prio_list is used for callbacks that should be handled
91 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
94 struct list_head blp_prio_list;
97 * blp_list is used for all other callbacks which are likely
98 * to take longer to process.
100 struct list_head blp_list;
102 wait_queue_head_t blp_waitq;
103 struct completion blp_comp;
104 atomic_t blp_num_threads;
105 atomic_t blp_busy_threads;
110 struct ldlm_bl_work_item {
111 struct list_head blwi_entry;
112 struct ldlm_namespace *blwi_ns;
113 struct ldlm_lock_desc blwi_ld;
114 struct ldlm_lock *blwi_lock;
115 struct list_head blwi_head;
117 struct completion blwi_comp;
118 enum ldlm_cancel_flags blwi_flags;
119 int blwi_mem_pressure;
122 #ifdef HAVE_SERVER_SUPPORT
125 * Protects both waiting_locks_list and expired_lock_thread.
127 static DEFINE_SPINLOCK(waiting_locks_spinlock); /* BH lock (timer) */
130 * List for contended locks.
132 * As soon as a lock is contended, it gets placed on this list and
133 * expected time to get a response is filled in the lock. A special
134 * thread walks the list looking for locks that should be released and
135 * schedules client evictions for those that have not been released in
138 * All access to it should be under waiting_locks_spinlock.
140 static LIST_HEAD(waiting_locks_list);
141 static void waiting_locks_callback(TIMER_DATA_TYPE unused);
142 static CFS_DEFINE_TIMER(waiting_locks_timer, waiting_locks_callback, 0, 0);
150 static DECLARE_WAIT_QUEUE_HEAD(expired_lock_wait_queue);
151 static enum elt_state expired_lock_thread_state = ELT_STOPPED;
152 static int expired_lock_dump;
153 static LIST_HEAD(expired_lock_list);
155 static int ldlm_lock_busy(struct ldlm_lock *lock);
156 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout);
157 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds);
159 static inline int have_expired_locks(void)
164 spin_lock_bh(&waiting_locks_spinlock);
165 need_to_run = !list_empty(&expired_lock_list);
166 spin_unlock_bh(&waiting_locks_spinlock);
172 * Check expired lock list for expired locks and time them out.
174 static int expired_lock_main(void *arg)
176 struct list_head *expired = &expired_lock_list;
181 expired_lock_thread_state = ELT_READY;
182 wake_up(&expired_lock_wait_queue);
185 wait_event_idle(expired_lock_wait_queue,
186 have_expired_locks() ||
187 expired_lock_thread_state == ELT_TERMINATE);
189 spin_lock_bh(&waiting_locks_spinlock);
190 if (expired_lock_dump) {
191 spin_unlock_bh(&waiting_locks_spinlock);
193 /* from waiting_locks_callback, but not in timer */
194 libcfs_debug_dumplog();
196 spin_lock_bh(&waiting_locks_spinlock);
197 expired_lock_dump = 0;
202 while (!list_empty(expired)) {
203 struct obd_export *export;
204 struct ldlm_lock *lock;
206 lock = list_entry(expired->next, struct ldlm_lock,
208 if ((void *)lock < LP_POISON + PAGE_SIZE &&
209 (void *)lock >= LP_POISON) {
210 spin_unlock_bh(&waiting_locks_spinlock);
211 CERROR("free lock on elt list %p\n", lock);
214 list_del_init(&lock->l_pending_chain);
215 if ((void *)lock->l_export <
216 LP_POISON + PAGE_SIZE &&
217 (void *)lock->l_export >= LP_POISON) {
218 CERROR("lock with free export on elt list %p\n",
220 lock->l_export = NULL;
221 LDLM_ERROR(lock, "free export");
223 * release extra ref grabbed by
224 * ldlm_add_waiting_lock() or
227 LDLM_LOCK_RELEASE(lock);
231 if (ldlm_is_destroyed(lock)) {
233 * release the lock refcount where
234 * waiting_locks_callback() founds
236 LDLM_LOCK_RELEASE(lock);
239 export = class_export_lock_get(lock->l_export, lock);
240 spin_unlock_bh(&waiting_locks_spinlock);
242 /* Check if we need to prolong timeout */
243 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
244 lock->l_callback_timeout != 0 && /* not AST error */
245 ldlm_lock_busy(lock)) {
246 LDLM_DEBUG(lock, "prolong the busy lock");
247 lock_res_and_lock(lock);
248 ldlm_add_waiting_lock(lock,
249 ldlm_bl_timeout(lock) >> 1);
250 unlock_res_and_lock(lock);
252 spin_lock_bh(&export->exp_bl_list_lock);
253 list_del_init(&lock->l_exp_list);
254 spin_unlock_bh(&export->exp_bl_list_lock);
257 "lock callback timer expired after %llds: evicting client at %s ",
258 ktime_get_real_seconds() -
260 obd_export_nid2str(export));
261 ldlm_lock_to_ns(lock)->ns_timeouts++;
263 class_fail_export(export);
265 class_export_lock_put(export, lock);
267 * release extra ref grabbed by ldlm_add_waiting_lock()
268 * or ldlm_failed_ast()
270 LDLM_LOCK_RELEASE(lock);
272 spin_lock_bh(&waiting_locks_spinlock);
274 spin_unlock_bh(&waiting_locks_spinlock);
276 if (do_dump && obd_dump_on_eviction) {
277 CERROR("dump the log upon eviction\n");
278 libcfs_debug_dumplog();
281 if (expired_lock_thread_state == ELT_TERMINATE)
285 expired_lock_thread_state = ELT_STOPPED;
286 wake_up(&expired_lock_wait_queue);
291 * Check if there is a request in the export request list
292 * which prevents the lock canceling.
294 static int ldlm_lock_busy(struct ldlm_lock *lock)
296 struct ptlrpc_request *req;
301 if (lock->l_export == NULL)
304 spin_lock(&lock->l_export->exp_rpc_lock);
305 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
307 if (req->rq_ops->hpreq_lock_match) {
308 match = req->rq_ops->hpreq_lock_match(req, lock);
313 spin_unlock(&lock->l_export->exp_rpc_lock);
317 /* This is called from within a timer interrupt and cannot schedule */
318 static void waiting_locks_callback(TIMER_DATA_TYPE unused)
320 struct ldlm_lock *lock;
323 spin_lock_bh(&waiting_locks_spinlock);
324 while (!list_empty(&waiting_locks_list)) {
325 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
327 if (lock->l_callback_timeout > ktime_get_seconds() ||
328 lock->l_req_mode == LCK_GROUP)
332 * no needs to take an extra ref on the lock since it was in
333 * the waiting_locks_list and ldlm_add_waiting_lock()
334 * already grabbed a ref
336 list_move(&lock->l_pending_chain, &expired_lock_list);
340 if (!list_empty(&expired_lock_list)) {
341 if (obd_dump_on_timeout && need_dump)
342 expired_lock_dump = __LINE__;
344 wake_up(&expired_lock_wait_queue);
348 * Make sure the timer will fire again if we have any locks
351 if (!list_empty(&waiting_locks_list)) {
352 time64_t now = ktime_get_seconds();
355 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
357 if (lock->l_callback_timeout - now > 0)
358 delta = lock->l_callback_timeout - now;
359 mod_timer(&waiting_locks_timer,
360 jiffies + cfs_time_seconds(delta));
362 spin_unlock_bh(&waiting_locks_spinlock);
366 * Add lock to the list of contended locks.
368 * Indicate that we're waiting for a client to call us back cancelling a given
369 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
370 * timer to fire appropriately. (We round up to the next second, to avoid
371 * floods of timer firings during periods of high lock contention and traffic).
372 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
373 * if it has been added to the waiting list (1 is returned).
375 * Called with the namespace lock held.
377 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds)
379 unsigned long timeout_jiffies = jiffies;
380 time64_t now = ktime_get_seconds();
384 if (!list_empty(&lock->l_pending_chain))
387 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
388 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
391 deadline = now + seconds;
392 if (likely(deadline > lock->l_callback_timeout))
393 lock->l_callback_timeout = deadline;
395 timeout = clamp_t(time_t, lock->l_callback_timeout - now,
397 timeout_jiffies += cfs_time_seconds(timeout);
399 if (time_before(timeout_jiffies, waiting_locks_timer.expires) ||
400 !timer_pending(&waiting_locks_timer))
401 mod_timer(&waiting_locks_timer, timeout_jiffies);
404 * if the new lock has a shorter timeout than something earlier on
405 * the list, we'll wait the longer amount of time; no big deal.
408 list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
412 static void ldlm_add_blocked_lock(struct ldlm_lock *lock)
414 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
415 if (list_empty(&lock->l_exp_list)) {
416 if (!ldlm_is_granted(lock))
417 list_add_tail(&lock->l_exp_list,
418 &lock->l_export->exp_bl_list);
420 list_add(&lock->l_exp_list,
421 &lock->l_export->exp_bl_list);
423 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
426 * A blocked lock is added. Adjust the position in
427 * the stale list if the export is in the list.
428 * If export is stale and not in the list - it is being
429 * processed and will be placed on the right position
430 * on obd_stale_export_put().
432 if (!list_empty(&lock->l_export->exp_stale_list))
433 obd_stale_export_adjust(lock->l_export);
436 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
440 /* NB: must be called with hold of lock_res_and_lock() */
441 LASSERT(ldlm_is_res_locked(lock));
442 LASSERT(!ldlm_is_cancel_on_block(lock));
445 * Do not put cross-MDT lock in the waiting list, since we
446 * will not evict it due to timeout for now
448 if (lock->l_export != NULL &&
449 (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS))
452 spin_lock_bh(&waiting_locks_spinlock);
453 if (ldlm_is_cancel(lock)) {
454 spin_unlock_bh(&waiting_locks_spinlock);
458 if (ldlm_is_destroyed(lock)) {
459 static time64_t next;
461 spin_unlock_bh(&waiting_locks_spinlock);
462 LDLM_ERROR(lock, "not waiting on destroyed lock (b=5653)");
463 if (ktime_get_seconds() > next) {
464 next = ktime_get_seconds() + 14400;
465 libcfs_debug_dumpstack(NULL);
470 ldlm_set_waited(lock);
471 lock->l_blast_sent = ktime_get_real_seconds();
472 ret = __ldlm_add_waiting_lock(lock, timeout);
475 * grab ref on the lock if it has been added to the
480 spin_unlock_bh(&waiting_locks_spinlock);
483 ldlm_add_blocked_lock(lock);
485 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %lld, AT: %s)",
486 ret == 0 ? "not re-" : "", timeout,
487 AT_OFF ? "off" : "on");
492 * Remove a lock from the pending list, likely because it had its cancellation
493 * callback arrive without incident. This adjusts the lock-timeout timer if
494 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
495 * As done by ldlm_del_waiting_lock(), the caller must release the lock
496 * reference when the lock is removed from any list (1 is returned).
498 * Called with namespace lock held.
500 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
502 struct list_head *list_next;
504 if (list_empty(&lock->l_pending_chain))
507 list_next = lock->l_pending_chain.next;
508 if (lock->l_pending_chain.prev == &waiting_locks_list) {
509 /* Removing the head of the list, adjust timer. */
510 if (list_next == &waiting_locks_list) {
511 /* No more, just cancel. */
512 del_timer(&waiting_locks_timer);
514 time64_t now = ktime_get_seconds();
515 struct ldlm_lock *next;
518 next = list_entry(list_next, struct ldlm_lock,
520 if (next->l_callback_timeout - now > 0)
521 delta = lock->l_callback_timeout - now;
523 mod_timer(&waiting_locks_timer,
524 jiffies + cfs_time_seconds(delta));
527 list_del_init(&lock->l_pending_chain);
532 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
536 if (lock->l_export == NULL) {
537 /* We don't have a "waiting locks list" on clients. */
538 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
542 spin_lock_bh(&waiting_locks_spinlock);
543 ret = __ldlm_del_waiting_lock(lock);
544 ldlm_clear_waited(lock);
545 spin_unlock_bh(&waiting_locks_spinlock);
547 /* remove the lock out of export blocking list */
548 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
549 list_del_init(&lock->l_exp_list);
550 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
554 * release lock ref if it has indeed been removed
557 LDLM_LOCK_RELEASE(lock);
560 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
565 * Prolong the contended lock waiting time.
567 * Called with namespace lock held.
569 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
571 if (lock->l_export == NULL) {
572 /* We don't have a "waiting locks list" on clients. */
573 LDLM_DEBUG(lock, "client lock: no-op");
577 if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) {
578 /* We don't have a "waiting locks list" on OSP. */
579 LDLM_DEBUG(lock, "MDS-MDS lock: no-op");
583 spin_lock_bh(&waiting_locks_spinlock);
585 if (list_empty(&lock->l_pending_chain)) {
586 spin_unlock_bh(&waiting_locks_spinlock);
587 LDLM_DEBUG(lock, "wasn't waiting");
592 * we remove/add the lock to the waiting list, so no needs to
593 * release/take a lock reference
595 __ldlm_del_waiting_lock(lock);
596 __ldlm_add_waiting_lock(lock, timeout);
597 spin_unlock_bh(&waiting_locks_spinlock);
599 LDLM_DEBUG(lock, "refreshed");
602 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
604 #else /* HAVE_SERVER_SUPPORT */
606 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
611 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
616 #endif /* !HAVE_SERVER_SUPPORT */
618 #ifdef HAVE_SERVER_SUPPORT
621 * Calculate the per-export Blocking timeout (covering BL AST, data flush,
622 * lock cancel, and their replies). Used for lock callback timeout and AST
625 * \param[in] lock lock which is getting the blocking callback
627 * \retval timeout in seconds to wait for the client reply
629 time64_t ldlm_bl_timeout(struct ldlm_lock *lock)
634 return obd_timeout / 2;
637 * Since these are non-updating timeouts, we should be conservative.
638 * Take more than usually, 150%
639 * It would be nice to have some kind of "early reply" mechanism for
640 * lock callbacks too...
642 timeout = at_get(&lock->l_export->exp_bl_lock_at);
643 return max(timeout + (timeout >> 1), (time64_t)ldlm_enqueue_min);
645 EXPORT_SYMBOL(ldlm_bl_timeout);
648 * Perform lock cleanup if AST sending failed.
650 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
651 const char *ast_type)
653 LCONSOLE_ERROR_MSG(0x138,
654 "%s: A client on nid %s was evicted due to a lock %s callback time out: rc %d\n",
655 lock->l_export->exp_obd->obd_name,
656 obd_export_nid2str(lock->l_export), ast_type, rc);
658 if (obd_dump_on_timeout)
659 libcfs_debug_dumplog();
660 spin_lock_bh(&waiting_locks_spinlock);
661 if (__ldlm_del_waiting_lock(lock) == 0)
663 * the lock was not in any list, grab an extra ref before adding
664 * the lock to the expired list
667 lock->l_callback_timeout = 0; /* differentiate it from expired locks */
668 list_add(&lock->l_pending_chain, &expired_lock_list);
669 wake_up(&expired_lock_wait_queue);
670 spin_unlock_bh(&waiting_locks_spinlock);
674 * Perform lock cleanup if AST reply came with error.
676 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
677 struct ptlrpc_request *req, int rc,
678 const char *ast_type)
680 struct lnet_process_id peer = req->rq_import->imp_connection->c_peer;
682 if (!req->rq_replied || (rc && rc != -EINVAL)) {
683 if (ldlm_is_cancel(lock)) {
685 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
686 ast_type, req, req->rq_xid,
687 libcfs_nid2str(peer.nid));
688 ldlm_lock_cancel(lock);
690 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
692 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
694 * Upon umount process the AST fails because cannot be
695 * sent. This shouldn't lead to the client eviction.
696 * -ENODEV error is returned by ptl_send_rpc() for
697 * new request in such import.
698 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
699 * if imp_invalid is set or obd_no_recov.
700 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
701 * in ptlrpc_import_delay_req() as well with -EIO code.
702 * In all such cases errors are ignored.
705 "%s AST can't be sent due to a server %s failure or umount process: rc = %d\n",
707 req->rq_import->imp_obd->obd_name, rc);
710 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
711 libcfs_nid2str(peer.nid),
712 req->rq_replied ? "returned error from" :
713 "failed to reply to",
714 ast_type, req, req->rq_xid,
715 (req->rq_repmsg != NULL) ?
716 lustre_msg_get_status(req->rq_repmsg) : 0,
718 ldlm_failed_ast(lock, rc, ast_type);
724 struct ldlm_resource *res = lock->l_resource;
727 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
728 libcfs_nid2str(peer.nid),
730 lustre_msg_get_status(req->rq_repmsg) : -1,
731 ast_type, req, req->rq_xid);
734 * update lvbo to return proper attributes.
737 ldlm_resource_getref(res);
738 ldlm_lvbo_update(res, lock, NULL, 1);
739 ldlm_resource_putref(res);
741 ldlm_lock_cancel(lock);
748 static int ldlm_cb_interpret(const struct lu_env *env,
749 struct ptlrpc_request *req, void *args, int rc)
751 struct ldlm_cb_async_args *ca = args;
752 struct ldlm_lock *lock = ca->ca_lock;
753 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
757 LASSERT(lock != NULL);
760 case LDLM_GL_CALLBACK:
762 * Update the LVB from disk if the AST failed
763 * (this is a legal race)
765 * - Glimpse callback of local lock just returns
766 * -ELDLM_NO_LOCK_DATA.
767 * - Glimpse callback of remote lock might return
768 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
770 if (unlikely(arg->gl_interpret_reply)) {
771 rc = arg->gl_interpret_reply(NULL, req, args, rc);
772 } else if (rc == -ELDLM_NO_LOCK_DATA) {
774 "lost race - client has a lock but no inode");
775 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
776 } else if (rc != 0) {
777 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
779 rc = ldlm_lvbo_update(lock->l_resource,
783 case LDLM_BL_CALLBACK:
785 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
787 case LDLM_CP_CALLBACK:
789 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
792 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
797 /* release extra reference taken in ldlm_ast_fini() */
798 LDLM_LOCK_RELEASE(lock);
801 atomic_inc(&arg->restart);
806 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
808 struct ldlm_cb_async_args *ca = data;
809 struct ldlm_lock *lock = ca->ca_lock;
811 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
814 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
815 struct ldlm_cb_set_arg *arg,
816 struct ldlm_lock *lock,
823 if (unlikely(instant_cancel)) {
824 rc = ptl_send_rpc(req, 1);
825 ptlrpc_req_finished(req);
827 atomic_inc(&arg->restart);
830 ptlrpc_set_add_req(arg->set, req);
837 * Check if there are requests in the export request list which prevent
838 * the lock canceling and make these requests high priority ones.
840 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
842 struct ptlrpc_request *req;
846 if (lock->l_export == NULL) {
847 LDLM_DEBUG(lock, "client lock: no-op");
851 spin_lock(&lock->l_export->exp_rpc_lock);
852 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
855 * Do not process requests that were not yet added to there
856 * incoming queue or were already removed from there for
857 * processing. We evaluate ptlrpc_nrs_req_can_move() without
858 * holding svcpt->scp_req_lock, and then redo the check with
859 * the lock held once we need to obtain a reliable result.
861 if (ptlrpc_nrs_req_can_move(req) &&
862 req->rq_ops->hpreq_lock_match &&
863 req->rq_ops->hpreq_lock_match(req, lock))
864 ptlrpc_nrs_req_hp_move(req);
866 spin_unlock(&lock->l_export->exp_rpc_lock);
871 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
872 * enqueued server lock conflicts with given one.
874 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
875 * to wait for client response.
877 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
878 struct ldlm_lock_desc *desc,
879 void *data, int flag)
881 struct ldlm_cb_async_args *ca;
882 struct ldlm_cb_set_arg *arg = data;
883 struct ldlm_request *body;
884 struct ptlrpc_request *req;
885 int instant_cancel = 0;
890 if (flag == LDLM_CB_CANCELING)
891 /* Don't need to do anything here. */
894 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
895 LDLM_DEBUG(lock, "dropping BL AST");
900 LASSERT(data != NULL);
901 if (lock->l_export->exp_obd->obd_recovering != 0)
902 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
904 ldlm_lock_reorder_req(lock);
906 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
907 &RQF_LDLM_BL_CALLBACK,
908 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
912 ca = ptlrpc_req_async_args(ca, req);
913 ca->ca_set_arg = arg;
916 req->rq_interpret_reply = ldlm_cb_interpret;
918 lock_res_and_lock(lock);
919 if (ldlm_is_destroyed(lock)) {
920 /* What's the point? */
921 unlock_res_and_lock(lock);
922 ptlrpc_req_finished(req);
926 if (!ldlm_is_granted(lock)) {
928 * this blocking AST will be communicated as part of the
929 * completion AST instead
931 ldlm_add_blocked_lock(lock);
932 ldlm_set_waited(lock);
933 unlock_res_and_lock(lock);
935 ptlrpc_req_finished(req);
936 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
940 if (ldlm_is_cancel_on_block(lock))
943 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
944 body->lock_handle[0] = lock->l_remote_handle;
945 body->lock_desc = *desc;
946 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
948 LDLM_DEBUG(lock, "server preparing blocking AST");
950 ptlrpc_request_set_replen(req);
951 ldlm_set_cbpending(lock);
952 if (instant_cancel) {
953 unlock_res_and_lock(lock);
954 ldlm_lock_cancel(lock);
956 req->rq_no_resend = 1;
958 LASSERT(ldlm_is_granted(lock));
959 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
960 unlock_res_and_lock(lock);
962 /* Do not resend after lock callback timeout */
963 req->rq_delay_limit = ldlm_bl_timeout(lock);
964 req->rq_resend_cb = ldlm_update_resend;
967 req->rq_send_state = LUSTRE_IMP_FULL;
968 /* ptlrpc_request_alloc_pack already set timeout */
970 req->rq_timeout = ldlm_get_rq_timeout();
972 if (lock->l_export && lock->l_export->exp_nid_stats &&
973 lock->l_export->exp_nid_stats->nid_ldlm_stats)
974 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
975 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
977 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
983 * ->l_completion_ast callback for a remote lock in server namespace.
985 * Sends AST to the client notifying it of lock granting. If initial
986 * lock response was not sent yet, instead of sending another RPC, just
987 * mark the lock as granted and client will understand
989 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
991 struct ldlm_cb_set_arg *arg = data;
992 struct ldlm_request *body;
993 struct ptlrpc_request *req;
994 struct ldlm_cb_async_args *ca;
995 int instant_cancel = 0;
1001 LASSERT(lock != NULL);
1002 LASSERT(data != NULL);
1004 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
1005 LDLM_DEBUG(lock, "dropping CP AST");
1009 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
1010 &RQF_LDLM_CP_CALLBACK);
1014 /* server namespace, doesn't need lock */
1015 lvb_len = ldlm_lvbo_size(lock);
1017 * LU-3124 & LU-2187: to not return layout in completion AST because
1018 * it may deadlock for LU-2187, or client may not have enough space
1019 * for large layout. The layout will be returned to client with an
1020 * extra RPC to fetch xattr.lov
1022 if (ldlm_has_layout(lock))
1025 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
1026 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
1028 ptlrpc_request_free(req);
1032 ca = ptlrpc_req_async_args(ca, req);
1033 ca->ca_set_arg = arg;
1036 req->rq_interpret_reply = ldlm_cb_interpret;
1037 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1039 body->lock_handle[0] = lock->l_remote_handle;
1040 body->lock_flags = ldlm_flags_to_wire(flags);
1041 ldlm_lock2desc(lock, &body->lock_desc);
1043 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1044 lvb_len = ldlm_lvbo_fill(lock, lvb, &lvb_len);
1047 * We still need to send the RPC to wake up the blocked
1048 * enqueue thread on the client.
1050 * Consider old client, there is no better way to notify
1051 * the failure, just zero-sized the LVB, then the client
1052 * will fail out as "-EPROTO".
1054 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1058 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1063 LDLM_DEBUG(lock, "server preparing completion AST");
1065 ptlrpc_request_set_replen(req);
1067 req->rq_send_state = LUSTRE_IMP_FULL;
1068 /* ptlrpc_request_pack already set timeout */
1070 req->rq_timeout = ldlm_get_rq_timeout();
1072 /* We only send real blocking ASTs after the lock is granted */
1073 lock_res_and_lock(lock);
1074 if (ldlm_is_ast_sent(lock)) {
1075 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1076 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1077 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1081 * We might get here prior to ldlm_handle_enqueue setting
1082 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1083 * into waiting list, but this is safe and similar code in
1084 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1085 * that would not only cancel the lock, but will also remove
1086 * it from waiting list
1088 if (ldlm_is_cancel_on_block(lock)) {
1089 unlock_res_and_lock(lock);
1090 ldlm_lock_cancel(lock);
1093 req->rq_no_resend = 1;
1095 lock_res_and_lock(lock);
1097 /* start the lock-timeout clock */
1098 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1099 /* Do not resend after lock callback timeout */
1100 req->rq_delay_limit = ldlm_bl_timeout(lock);
1101 req->rq_resend_cb = ldlm_update_resend;
1104 unlock_res_and_lock(lock);
1106 if (lock->l_export && lock->l_export->exp_nid_stats &&
1107 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1108 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1109 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1111 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1113 RETURN(lvb_len < 0 ? lvb_len : rc);
1117 * Server side ->l_glimpse_ast handler for client locks.
1119 * Sends glimpse AST to the client and waits for reply. Then updates
1120 * lvbo with the result.
1122 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1124 struct ldlm_cb_set_arg *arg = data;
1125 struct ldlm_request *body;
1126 struct ptlrpc_request *req;
1127 struct ldlm_cb_async_args *ca;
1129 struct req_format *req_fmt;
1133 LASSERT(lock != NULL);
1135 if (arg->gl_desc != NULL)
1136 /* There is a glimpse descriptor to pack */
1137 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1139 req_fmt = &RQF_LDLM_GL_CALLBACK;
1141 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1142 req_fmt, LUSTRE_DLM_VERSION,
1148 if (arg->gl_desc != NULL) {
1149 /* copy the GL descriptor */
1150 union ldlm_gl_desc *desc;
1152 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1153 *desc = *arg->gl_desc;
1156 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1157 body->lock_handle[0] = lock->l_remote_handle;
1158 ldlm_lock2desc(lock, &body->lock_desc);
1160 ca = ptlrpc_req_async_args(ca, req);
1161 ca->ca_set_arg = arg;
1164 /* server namespace, doesn't need lock */
1165 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1166 ldlm_lvbo_size(lock));
1167 ptlrpc_request_set_replen(req);
1169 req->rq_send_state = LUSTRE_IMP_FULL;
1170 /* ptlrpc_request_alloc_pack already set timeout */
1172 req->rq_timeout = ldlm_get_rq_timeout();
1174 req->rq_interpret_reply = ldlm_cb_interpret;
1176 if (lock->l_export && lock->l_export->exp_nid_stats) {
1177 struct nid_stat *nid_stats = lock->l_export->exp_nid_stats;
1179 lprocfs_counter_incr(nid_stats->nid_ldlm_stats,
1180 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1183 rc = ldlm_ast_fini(req, arg, lock, 0);
1187 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1189 int ldlm_glimpse_locks(struct ldlm_resource *res,
1190 struct list_head *gl_work_list)
1196 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1198 if (rc == -ERESTART)
1199 ldlm_reprocess_all(res, NULL);
1203 EXPORT_SYMBOL(ldlm_glimpse_locks);
1205 /* return LDLM lock associated with a lock callback request */
1206 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1208 struct ldlm_cb_async_args *ca;
1209 struct ldlm_lock *lock;
1213 ca = ptlrpc_req_async_args(ca, req);
1216 RETURN(ERR_PTR(-EFAULT));
1220 EXPORT_SYMBOL(ldlm_request_lock);
1223 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1224 * service threads to carry out client lock enqueueing requests.
1226 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1227 struct ptlrpc_request *req,
1228 const struct ldlm_request *dlm_req,
1229 const struct ldlm_callback_suite *cbs)
1231 struct ldlm_reply *dlm_rep;
1233 enum ldlm_error err = ELDLM_OK;
1234 struct ldlm_lock *lock = NULL;
1235 void *cookie = NULL;
1237 struct ldlm_resource *res = NULL;
1238 const struct lu_env *env = req->rq_svc_thread->t_env;
1242 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1244 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1245 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1247 LASSERT(req->rq_export);
1249 /* for intent enqueue the stat will be updated inside intent policy */
1250 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1251 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1252 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1254 if (req->rq_export && req->rq_export->exp_nid_stats &&
1255 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1256 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1257 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1259 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1260 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1261 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1262 dlm_req->lock_desc.l_resource.lr_type);
1263 GOTO(out, rc = -EFAULT);
1266 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1267 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1268 dlm_req->lock_desc.l_req_mode &
1269 (dlm_req->lock_desc.l_req_mode-1))) {
1270 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1271 dlm_req->lock_desc.l_req_mode);
1272 GOTO(out, rc = -EFAULT);
1275 if (unlikely((flags & LDLM_FL_REPLAY) ||
1276 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1277 /* Find an existing lock in the per-export lock hash */
1279 * In the function below, .hs_keycmp resolves to
1280 * ldlm_export_lock_keycmp()
1282 /* coverity[overrun-buffer-val] */
1283 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1284 (void *)&dlm_req->lock_handle[0]);
1286 DEBUG_REQ(D_DLMTRACE, req,
1287 "found existing lock cookie %#llx",
1288 lock->l_handle.h_cookie);
1289 flags |= LDLM_FL_RESENT;
1290 GOTO(existing_lock, rc = 0);
1293 if (ldlm_reclaim_full()) {
1294 DEBUG_REQ(D_DLMTRACE, req,
1295 "Too many granted locks, reject current enqueue request and let the client retry later");
1296 GOTO(out, rc = -EINPROGRESS);
1300 /* The lock's callback data might be set in the policy function */
1301 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1302 dlm_req->lock_desc.l_resource.lr_type,
1303 dlm_req->lock_desc.l_req_mode,
1304 cbs, NULL, 0, LVB_T_NONE);
1311 lock->l_remote_handle = dlm_req->lock_handle[0];
1312 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1315 * Initialize resource lvb but not for a lock being replayed since
1316 * Client already got lvb sent in this case.
1317 * This must occur early since some policy methods assume resource
1318 * lvb is available (lr_lvb_data != NULL).
1320 res = lock->l_resource;
1321 if (!(flags & LDLM_FL_REPLAY)) {
1322 /* non-replayed lock, delayed lvb init may need to be done */
1323 rc = ldlm_lvbo_init(res);
1325 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1330 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1332 * Don't enqueue a lock onto the export if it is been disonnected
1333 * due to eviction (b=3822) or server umount (b=24324).
1334 * Cancel it now instead.
1336 if (req->rq_export->exp_disconnected) {
1337 LDLM_ERROR(lock, "lock on disconnected export %p",
1339 GOTO(out, rc = -ENOTCONN);
1342 lock->l_export = class_export_lock_get(req->rq_export, lock);
1343 if (lock->l_export->exp_lock_hash)
1344 cfs_hash_add(lock->l_export->exp_lock_hash,
1345 &lock->l_remote_handle,
1349 * Inherit the enqueue flags before the operation, because we do not
1350 * keep the res lock on return and next operations (BL AST) may proceed
1353 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1354 LDLM_FL_INHERIT_MASK);
1356 ldlm_convert_policy_to_local(req->rq_export,
1357 dlm_req->lock_desc.l_resource.lr_type,
1358 &dlm_req->lock_desc.l_policy_data,
1359 &lock->l_policy_data);
1360 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1361 lock->l_req_extent = lock->l_policy_data.l_extent;
1362 else if (dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS)
1363 lock->l_policy_data.l_inodebits.try_bits =
1364 dlm_req->lock_desc.l_policy_data.l_inodebits.try_bits;
1368 if (!(flags & LDLM_FL_HAS_INTENT)) {
1369 /* based on the assumption that lvb size never changes during
1370 * resource life time otherwise it need resource->lr_lock's
1372 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1373 RCL_SERVER, ldlm_lvbo_size(lock));
1375 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1376 GOTO(out, rc = -ENOMEM);
1378 rc = req_capsule_server_pack(&req->rq_pill);
1383 err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1390 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1392 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1393 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1395 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1396 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1399 * We never send a blocking AST until the lock is granted, but
1400 * we can tell it right now
1402 lock_res_and_lock(lock);
1405 * Now take into account flags to be inherited from original lock
1406 * request both in reply to client and in our own lock flags.
1408 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1409 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1412 * Don't move a pending lock onto the export if it has already been
1413 * disconnected due to eviction (b=5683) or server umount (b=24324).
1414 * Cancel it now instead.
1416 if (unlikely(req->rq_export->exp_disconnected ||
1417 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1418 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1420 } else if (ldlm_is_ast_sent(lock)) {
1421 /* fill lock desc for possible lock convert */
1422 if (lock->l_blocking_lock &&
1423 lock->l_resource->lr_type == LDLM_IBITS) {
1424 struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1425 struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1428 "save blocking bits %llx in granted lock",
1429 bl_lock->l_policy_data.l_inodebits.bits);
1431 * If lock is blocked then save blocking ibits
1432 * in returned lock policy for the possible lock
1433 * convert on a client.
1435 rep_desc->l_policy_data.l_inodebits.cancel_bits =
1436 bl_lock->l_policy_data.l_inodebits.bits;
1438 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1439 if (ldlm_is_granted(lock)) {
1441 * Only cancel lock if it was granted, because it would
1442 * be destroyed immediately and would never be granted
1443 * in the future, causing timeouts on client. Not
1444 * granted lock will be cancelled immediately after
1445 * sending completion AST.
1447 if (ldlm_is_cancel_on_block(lock)) {
1448 unlock_res_and_lock(lock);
1449 ldlm_lock_cancel(lock);
1450 lock_res_and_lock(lock);
1452 ldlm_add_waiting_lock(lock,
1453 ldlm_bl_timeout(lock));
1457 unlock_res_and_lock(lock);
1461 req->rq_status = rc ?: err; /* return either error - b=11190 */
1462 if (!req->rq_packed_final) {
1463 err = lustre_pack_reply(req, 1, NULL, NULL);
1469 * The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1470 * ldlm_reprocess_all. If this moves, revisit that code. -phil
1474 "server-side enqueue handler, sending reply (err=%d, rc=%d)",
1478 req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1480 ldlm_lvbo_size(lock) > 0) {
1485 buf = req_capsule_server_get(&req->rq_pill,
1487 LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock);
1488 buflen = req_capsule_get_size(&req->rq_pill,
1489 &RMF_DLM_LVB, RCL_SERVER);
1491 * non-replayed lock, delayed lvb init may
1492 * need to be occur now
1494 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1497 rc2 = ldlm_lvbo_fill(lock, buf, &buflen);
1499 req_capsule_shrink(&req->rq_pill,
1502 } else if (rc2 == -ERANGE) {
1503 rc2 = req_capsule_server_grow(
1505 &RMF_DLM_LVB, buflen);
1510 * if we can't grow the buffer,
1511 * it's ok to return empty lvb
1522 } else if (flags & LDLM_FL_REPLAY) {
1523 /* no LVB resend upon replay */
1525 req_capsule_shrink(&req->rq_pill,
1535 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1536 if (lock->l_export) {
1537 ldlm_lock_cancel(lock);
1539 lock_res_and_lock(lock);
1540 ldlm_resource_unlink_lock(lock);
1541 ldlm_lock_destroy_nolock(lock);
1542 unlock_res_and_lock(lock);
1547 if (!err && !ldlm_is_cbpending(lock) &&
1548 dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1549 ldlm_reprocess_all(lock->l_resource, lock);
1551 LDLM_LOCK_RELEASE(lock);
1554 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1561 * Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1562 * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1564 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1566 if (lock->l_blocking_lock) {
1567 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1568 lock->l_blocking_lock = NULL;
1572 /* A lock can be converted to new ibits or mode and should be considered
1573 * as new lock. Clear all states related to a previous blocking AST
1574 * processing so new conflicts will cause new blocking ASTs.
1576 * This is used during lock convert below and lock downgrade to COS mode in
1577 * ldlm_lock_mode_downgrade().
1579 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1581 ldlm_clear_ast_sent(lock);
1582 lock->l_bl_ast_run = 0;
1583 ldlm_clear_blocking_lock(lock);
1587 * Main LDLM entry point for server code to process lock conversion requests.
1589 int ldlm_handle_convert0(struct ptlrpc_request *req,
1590 const struct ldlm_request *dlm_req)
1592 struct obd_export *exp = req->rq_export;
1593 struct ldlm_reply *dlm_rep;
1594 struct ldlm_lock *lock;
1601 if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1602 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1603 LDLM_CONVERT - LDLM_FIRST_OPC);
1605 rc = req_capsule_server_pack(&req->rq_pill);
1609 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1610 dlm_rep->lock_flags = dlm_req->lock_flags;
1612 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1614 LDLM_DEBUG_NOLOCK("server lock is canceled already");
1615 req->rq_status = ELDLM_NO_LOCK_DATA;
1619 LDLM_DEBUG(lock, "server-side convert handler START");
1621 lock_res_and_lock(lock);
1622 bits = lock->l_policy_data.l_inodebits.bits;
1623 new_bits = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1625 if (ldlm_is_cancel(lock)) {
1626 LDLM_DEBUG(lock, "convert on canceled lock!");
1627 unlock_res_and_lock(lock);
1628 GOTO(out_put, rc = ELDLM_NO_LOCK_DATA);
1631 if (dlm_req->lock_desc.l_req_mode != lock->l_granted_mode) {
1632 LDLM_ERROR(lock, "lock mode differs!");
1633 unlock_res_and_lock(lock);
1634 GOTO(out_put, rc = -EPROTO);
1637 if (bits == new_bits) {
1639 * This can be valid situation if CONVERT RPCs are
1640 * re-ordered. Just finish silently
1642 LDLM_DEBUG(lock, "lock is converted already!");
1643 unlock_res_and_lock(lock);
1645 if (ldlm_is_waited(lock))
1646 ldlm_del_waiting_lock(lock);
1648 ldlm_clear_cbpending(lock);
1649 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1650 ldlm_inodebits_drop(lock, bits & ~new_bits);
1652 ldlm_clear_blocking_data(lock);
1653 unlock_res_and_lock(lock);
1655 ldlm_reprocess_all(lock->l_resource, NULL);
1658 dlm_rep->lock_handle = lock->l_remote_handle;
1659 ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1660 &dlm_rep->lock_desc.l_policy_data);
1664 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d", rc);
1665 LDLM_LOCK_PUT(lock);
1666 req->rq_status = rc;
1671 * Cancel all the locks whose handles are packed into ldlm_request
1673 * Called by server code expecting such combined cancel activity
1676 int ldlm_request_cancel(struct ptlrpc_request *req,
1677 const struct ldlm_request *dlm_req,
1678 int first, enum lustre_at_flags flags)
1680 struct ldlm_resource *res, *pres = NULL;
1681 struct ldlm_lock *lock;
1682 int i, count, done = 0;
1687 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
1688 if (size <= offsetof(struct ldlm_request, lock_handle) ||
1689 (size - offsetof(struct ldlm_request, lock_handle)) /
1690 sizeof(struct lustre_handle) < dlm_req->lock_count)
1693 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1697 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1701 * There is no lock on the server at the replay time,
1702 * skip lock cancelling to make replay tests to pass.
1704 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1707 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, starting at %d",
1710 for (i = first; i < count; i++) {
1711 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1713 /* below message checked in replay-single.sh test_36 */
1714 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (cookie %llu)",
1715 dlm_req->lock_handle[i].cookie);
1719 res = lock->l_resource;
1723 * This code is an optimization to only attempt lock
1724 * granting on the resource (that could be CPU-expensive)
1725 * after we are done cancelling lock in that resource.
1729 ldlm_reprocess_all(pres, NULL);
1730 LDLM_RESOURCE_DELREF(pres);
1731 ldlm_resource_putref(pres);
1734 ldlm_resource_getref(res);
1735 LDLM_RESOURCE_ADDREF(res);
1737 if (!ldlm_is_discard_data(lock))
1738 ldlm_lvbo_update(res, lock,
1744 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1745 lock->l_blast_sent != 0) {
1746 time64_t delay = ktime_get_real_seconds() -
1749 "server cancels blocked lock after %llds",
1751 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1753 ldlm_lock_cancel(lock);
1754 LDLM_LOCK_PUT(lock);
1757 ldlm_reprocess_all(pres, NULL);
1758 LDLM_RESOURCE_DELREF(pres);
1759 ldlm_resource_putref(pres);
1761 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1764 EXPORT_SYMBOL(ldlm_request_cancel);
1767 * Main LDLM entry point for server code to cancel locks.
1769 * Typically gets called from service handler on LDLM_CANCEL opc.
1771 int ldlm_handle_cancel(struct ptlrpc_request *req)
1773 struct ldlm_request *dlm_req;
1778 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1779 if (dlm_req == NULL) {
1780 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1784 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) <
1785 offsetof(struct ldlm_request, lock_handle[1]))
1788 if (req->rq_export && req->rq_export->exp_nid_stats &&
1789 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1790 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1791 LDLM_CANCEL - LDLM_FIRST_OPC);
1793 rc = req_capsule_server_pack(&req->rq_pill);
1797 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1798 req->rq_status = LUSTRE_ESTALE;
1800 RETURN(ptlrpc_reply(req));
1802 #endif /* HAVE_SERVER_SUPPORT */
1805 * Server may pass additional information about blocking lock.
1806 * For IBITS locks it is conflicting bits which can be used for
1807 * lock convert instead of cancel.
1809 void ldlm_bl_desc2lock(const struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1811 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
1813 check_res_locked(lock->l_resource);
1814 if (ns_is_client(ns) && ld &&
1815 (lock->l_resource->lr_type == LDLM_IBITS)) {
1817 * Lock description contains policy of blocking lock,
1818 * and its cancel_bits is used to pass conflicting bits.
1819 * NOTE: ld can be NULL or can be not NULL but zeroed if
1820 * passed from ldlm_bl_thread_blwi(), check below used bits
1821 * in ld to make sure it is valid description.
1823 * If server may replace lock resource keeping the same cookie,
1824 * never use cancel bits from different resource, full cancel
1827 if (ld->l_policy_data.l_inodebits.cancel_bits &&
1828 ldlm_res_eq(&ld->l_resource.lr_name,
1829 &lock->l_resource->lr_name) &&
1830 !(ldlm_is_cbpending(lock) &&
1831 lock->l_policy_data.l_inodebits.cancel_bits == 0)) {
1832 /* always combine conflicting ibits */
1833 lock->l_policy_data.l_inodebits.cancel_bits |=
1834 ld->l_policy_data.l_inodebits.cancel_bits;
1836 /* If cancel_bits are not obtained or
1837 * if the lock is already CBPENDING and
1838 * has no cancel_bits set
1839 * - the full lock is to be cancelled
1841 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1847 * Callback handler for receiving incoming blocking ASTs.
1849 * This can only happen on client side.
1851 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1852 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1858 LDLM_DEBUG(lock, "client blocking AST callback handler");
1860 lock_res_and_lock(lock);
1862 /* get extra information from desc if any */
1863 ldlm_bl_desc2lock(ld, lock);
1864 ldlm_set_cbpending(lock);
1866 do_ast = (!lock->l_readers && !lock->l_writers);
1867 unlock_res_and_lock(lock);
1871 "Lock %p already unused, calling callback (%p)\n",
1872 lock, lock->l_blocking_ast);
1873 if (lock->l_blocking_ast != NULL)
1874 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1878 "Lock %p is referenced, will be cancelled later\n",
1882 LDLM_DEBUG(lock, "client blocking callback handler END");
1883 LDLM_LOCK_RELEASE(lock);
1887 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1889 if (req->rq_no_reply)
1892 req->rq_status = rc;
1893 if (!req->rq_packed_final) {
1894 rc = lustre_pack_reply(req, 1, NULL, NULL);
1898 return ptlrpc_reply(req);
1902 * Callback handler for receiving incoming completion ASTs.
1904 * This only can happen on client side.
1906 static int ldlm_handle_cp_callback(struct ptlrpc_request *req,
1907 struct ldlm_namespace *ns,
1908 struct ldlm_request *dlm_req,
1909 struct ldlm_lock *lock)
1911 LIST_HEAD(ast_list);
1917 LDLM_DEBUG(lock, "client completion callback handler START");
1919 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1920 long to = cfs_time_seconds(1);
1922 ldlm_callback_reply(req, 0);
1925 schedule_timeout_interruptible(to);
1926 if (ldlm_is_granted(lock) ||
1927 ldlm_is_destroyed(lock))
1932 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1934 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1935 GOTO(out, rc = lvb_len);
1936 } else if (lvb_len > 0) {
1937 if (lock->l_lvb_len > 0) {
1938 /* for extent lock, lvb contains ost_lvb{}. */
1939 LASSERT(lock->l_lvb_data != NULL);
1941 if (unlikely(lock->l_lvb_len < lvb_len)) {
1943 "Replied LVB is larger than expectation, expected = %d, replied = %d",
1944 lock->l_lvb_len, lvb_len);
1945 GOTO(out, rc = -EINVAL);
1950 lock_res_and_lock(lock);
1952 if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
1953 &lock->l_resource->lr_name)) {
1954 ldlm_resource_unlink_lock(lock);
1955 unlock_res_and_lock(lock);
1956 rc = ldlm_lock_change_resource(ns, lock,
1957 &dlm_req->lock_desc.l_resource.lr_name);
1959 LDLM_ERROR(lock, "Failed to allocate resource");
1962 LDLM_DEBUG(lock, "completion AST, new resource");
1963 lock_res_and_lock(lock);
1966 if (ldlm_is_failed(lock)) {
1967 unlock_res_and_lock(lock);
1968 LDLM_LOCK_RELEASE(lock);
1972 if (ldlm_is_destroyed(lock) ||
1973 ldlm_is_granted(lock)) {
1974 /* b=11300: the lock has already been granted */
1975 unlock_res_and_lock(lock);
1976 LDLM_DEBUG(lock, "Double grant race happened");
1981 * If we receive the completion AST before the actual enqueue returned,
1982 * then we might need to switch lock modes, resources, or extents.
1984 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1985 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1986 LDLM_DEBUG(lock, "completion AST, new lock mode");
1989 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1990 ldlm_convert_policy_to_local(req->rq_export,
1991 dlm_req->lock_desc.l_resource.lr_type,
1992 &dlm_req->lock_desc.l_policy_data,
1993 &lock->l_policy_data);
1994 LDLM_DEBUG(lock, "completion AST, new policy data");
1997 ldlm_resource_unlink_lock(lock);
1999 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
2001 * BL_AST locks are not needed in LRU.
2002 * Let ldlm_cancel_lru() be fast.
2004 ldlm_lock_remove_from_lru(lock);
2005 ldlm_bl_desc2lock(&dlm_req->lock_desc, lock);
2006 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
2007 LDLM_DEBUG(lock, "completion AST includes blocking AST");
2010 if (lock->l_lvb_len > 0) {
2011 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
2012 lock->l_lvb_data, lvb_len);
2014 unlock_res_and_lock(lock);
2019 ldlm_grant_lock(lock, &ast_list);
2020 unlock_res_and_lock(lock);
2022 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
2025 * Let Enqueue to call osc_lock_upcall() and initialize
2028 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
2030 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
2032 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
2038 lock_res_and_lock(lock);
2039 ldlm_set_failed(lock);
2040 unlock_res_and_lock(lock);
2041 wake_up(&lock->l_waitq);
2043 LDLM_LOCK_RELEASE(lock);
2049 * Callback handler for receiving incoming glimpse ASTs.
2051 * This only can happen on client side. After handling the glimpse AST
2052 * we also consider dropping the lock here if it is unused locally for a
2055 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
2056 struct ldlm_namespace *ns,
2057 struct ldlm_request *dlm_req,
2058 struct ldlm_lock *lock)
2060 struct ldlm_lock_desc *ld = &dlm_req->lock_desc;
2065 LDLM_DEBUG(lock, "client glimpse AST callback handler");
2067 if (lock->l_glimpse_ast != NULL)
2068 rc = lock->l_glimpse_ast(lock, req);
2070 if (req->rq_repmsg != NULL) {
2073 req->rq_status = rc;
2077 lock_res_and_lock(lock);
2078 if (lock->l_granted_mode == LCK_PW &&
2079 !lock->l_readers && !lock->l_writers &&
2080 ktime_after(ktime_get(),
2081 ktime_add(lock->l_last_used,
2082 ktime_set(ns->ns_dirty_age_limit, 0)))) {
2083 unlock_res_and_lock(lock);
2085 /* For MDS glimpse it is always DOM lock, set corresponding
2086 * cancel_bits to perform lock convert if needed
2088 if (lock->l_resource->lr_type == LDLM_IBITS)
2089 ld->l_policy_data.l_inodebits.cancel_bits =
2091 if (ldlm_bl_to_thread_lock(ns, ld, lock))
2092 ldlm_handle_bl_callback(ns, ld, lock);
2097 unlock_res_and_lock(lock);
2098 LDLM_LOCK_RELEASE(lock);
2102 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
2103 enum ldlm_cancel_flags cancel_flags)
2105 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2109 spin_lock(&blp->blp_lock);
2110 if (blwi->blwi_lock &&
2111 ldlm_is_discard_data(blwi->blwi_lock)) {
2112 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
2113 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
2115 /* other blocking callbacks are added to the regular list */
2116 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
2118 spin_unlock(&blp->blp_lock);
2120 wake_up(&blp->blp_waitq);
2123 * can not check blwi->blwi_flags as blwi could be already freed in
2126 if (!(cancel_flags & LCF_ASYNC))
2127 wait_for_completion(&blwi->blwi_comp);
2132 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2133 struct ldlm_namespace *ns,
2134 struct ldlm_lock_desc *ld,
2135 struct list_head *cancels, int count,
2136 struct ldlm_lock *lock,
2137 enum ldlm_cancel_flags cancel_flags)
2139 init_completion(&blwi->blwi_comp);
2140 INIT_LIST_HEAD(&blwi->blwi_head);
2142 if (memory_pressure_get())
2143 blwi->blwi_mem_pressure = 1;
2146 blwi->blwi_flags = cancel_flags;
2148 blwi->blwi_ld = *ld;
2150 list_splice_init(cancels, &blwi->blwi_head);
2151 blwi->blwi_count = count;
2153 blwi->blwi_lock = lock;
2158 * Queues a list of locks \a cancels containing \a count locks
2159 * for later processing by a blocking thread. If \a count is zero,
2160 * then the lock referenced as \a lock is queued instead.
2162 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2163 * If list addition fails an error is returned and caller is supposed to
2164 * call ->l_blocking_ast itself.
2166 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2167 struct ldlm_lock_desc *ld,
2168 struct ldlm_lock *lock,
2169 struct list_head *cancels, int count,
2170 enum ldlm_cancel_flags cancel_flags)
2174 if (cancels && count == 0)
2177 if (cancel_flags & LCF_ASYNC) {
2178 struct ldlm_bl_work_item *blwi;
2180 OBD_ALLOC(blwi, sizeof(*blwi));
2183 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2185 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2188 * if it is synchronous call do minimum mem alloc, as it could
2189 * be triggered from kernel shrinker
2191 struct ldlm_bl_work_item blwi;
2193 memset(&blwi, 0, sizeof(blwi));
2194 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2195 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2200 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2201 struct ldlm_lock *lock)
2203 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2206 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2207 struct list_head *cancels, int count,
2208 enum ldlm_cancel_flags cancel_flags)
2210 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2213 int ldlm_bl_thread_wakeup(void)
2215 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2219 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2220 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2222 struct obd_device *obd = req->rq_export->exp_obd;
2230 DEBUG_REQ(D_HSM, req, "%s: handle setinfo", obd->obd_name);
2232 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2234 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2236 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2239 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2241 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2243 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2246 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2249 /* We are responsible for swabbing contents of val */
2251 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2252 /* Pass it on to mdc (the "export" in this case) */
2253 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2255 sizeof(KEY_HSM_COPYTOOL_SEND),
2256 KEY_HSM_COPYTOOL_SEND,
2259 DEBUG_REQ(D_WARNING, req, "ignoring unknown key '%s'", key);
2264 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2265 const char *msg, int rc,
2266 const struct lustre_handle *handle)
2268 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2269 "%s, NID=%s lock=%#llx: rc = %d",
2270 msg, libcfs_id2str(req->rq_peer),
2271 handle ? handle->cookie : 0, rc);
2272 if (req->rq_no_reply)
2273 CWARN("No reply was sent, maybe cause b=21636.\n");
2275 CWARN("Send reply failed, maybe cause b=21636.\n");
2278 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2279 static int ldlm_callback_handler(struct ptlrpc_request *req)
2281 struct ldlm_namespace *ns;
2282 struct ldlm_request *dlm_req;
2283 struct ldlm_lock *lock;
2289 * Requests arrive in sender's byte order. The ptlrpc service
2290 * handler has already checked and, if necessary, byte-swapped the
2291 * incoming request message body, but I am responsible for the
2295 /* do nothing for sec context finalize */
2296 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2299 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2301 if (req->rq_export == NULL) {
2302 rc = ldlm_callback_reply(req, -ENOTCONN);
2303 ldlm_callback_errmsg(req, "Operate on unconnected server",
2308 LASSERT(req->rq_export != NULL);
2309 LASSERT(req->rq_export->exp_obd != NULL);
2311 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2312 case LDLM_BL_CALLBACK:
2313 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2315 ldlm_callback_reply(req, -(int)cfs_fail_err);
2319 case LDLM_CP_CALLBACK:
2320 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2323 case LDLM_GL_CALLBACK:
2324 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2328 rc = ldlm_handle_setinfo(req);
2329 ldlm_callback_reply(req, rc);
2332 CERROR("unknown opcode %u\n",
2333 lustre_msg_get_opc(req->rq_reqmsg));
2334 ldlm_callback_reply(req, -EPROTO);
2338 ns = req->rq_export->exp_obd->obd_namespace;
2339 LASSERT(ns != NULL);
2341 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2343 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2344 if (dlm_req == NULL) {
2345 rc = ldlm_callback_reply(req, -EPROTO);
2346 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2352 * Force a known safe race, send a cancel to the server for a lock
2353 * which the server has already started a blocking callback on.
2355 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2356 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2357 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2359 CERROR("ldlm_cli_cancel: %d\n", rc);
2362 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2365 "callback on lock %#llx - lock disappeared\n",
2366 dlm_req->lock_handle[0].cookie);
2367 rc = ldlm_callback_reply(req, -EINVAL);
2368 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2369 &dlm_req->lock_handle[0]);
2373 if (ldlm_is_fail_loc(lock) &&
2374 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2375 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2377 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2378 lock_res_and_lock(lock);
2379 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2381 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2383 * If somebody cancels lock and cache is already dropped,
2384 * or lock is failed before cp_ast received on client,
2385 * we can tell the server we have no lock. Otherwise, we
2386 * should send cancel after dropping the cache.
2388 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2389 ldlm_is_failed(lock)) {
2391 "callback on lock %llx - lock disappeared",
2392 dlm_req->lock_handle[0].cookie);
2393 unlock_res_and_lock(lock);
2394 LDLM_LOCK_RELEASE(lock);
2395 rc = ldlm_callback_reply(req, -EINVAL);
2396 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2397 &dlm_req->lock_handle[0]);
2401 * BL_AST locks are not needed in LRU.
2402 * Let ldlm_cancel_lru() be fast.
2404 ldlm_lock_remove_from_lru(lock);
2405 ldlm_set_bl_ast(lock);
2407 unlock_res_and_lock(lock);
2410 * We want the ost thread to get this reply so that it can respond
2411 * to ost requests (write cache writeback) that might be triggered
2414 * But we'd also like to be able to indicate in the reply that we're
2415 * cancelling right now, because it's unused, or have an intent result
2416 * in the reply, so we might have to push the responsibility for sending
2417 * the reply down into the AST handlers, alas.
2420 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2421 case LDLM_BL_CALLBACK:
2422 CDEBUG(D_INODE, "blocking ast\n");
2423 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2424 if (!ldlm_is_cancel_on_block(lock)) {
2425 rc = ldlm_callback_reply(req, 0);
2426 if (req->rq_no_reply || rc)
2427 ldlm_callback_errmsg(req, "Normal process", rc,
2428 &dlm_req->lock_handle[0]);
2430 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2431 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2433 case LDLM_CP_CALLBACK:
2434 CDEBUG(D_INODE, "completion ast\n");
2435 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2436 rc = ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2437 if (!OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE))
2438 ldlm_callback_reply(req, rc);
2440 case LDLM_GL_CALLBACK:
2441 CDEBUG(D_INODE, "glimpse ast\n");
2442 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2443 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2446 LBUG(); /* checked above */
2452 #ifdef HAVE_SERVER_SUPPORT
2454 * Main handler for canceld thread.
2456 * Separated into its own thread to avoid deadlocks.
2458 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2465 * Requests arrive in sender's byte order. The ptlrpc service
2466 * handler has already checked and, if necessary, byte-swapped the
2467 * incoming request message body, but I am responsible for the
2471 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2473 if (req->rq_export == NULL) {
2474 struct ldlm_request *dlm_req;
2476 CERROR("%s from %s arrived at %llu with bad export cookie %llu\n",
2477 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2478 libcfs_nid2str(req->rq_peer.nid),
2479 (unsigned long long)req->rq_arrival_time.tv_sec,
2480 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2482 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2483 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2484 dlm_req = req_capsule_client_get(&req->rq_pill,
2486 if (dlm_req != NULL)
2487 ldlm_lock_dump_handle(D_ERROR,
2488 &dlm_req->lock_handle[0]);
2490 ldlm_callback_reply(req, -ENOTCONN);
2494 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2495 /* XXX FIXME move this back to mds/handler.c, b=249 */
2497 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2498 CDEBUG(D_INODE, "cancel\n");
2499 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2500 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2501 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2503 rc = ldlm_handle_cancel(req);
2507 struct ldlm_request *dlm_req;
2509 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2510 CDEBUG(D_INODE, "convert\n");
2512 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2513 if (dlm_req == NULL) {
2514 CDEBUG(D_INFO, "bad request buffer for cancel\n");
2515 rc = ldlm_callback_reply(req, -EPROTO);
2517 req->rq_status = ldlm_handle_convert0(req, dlm_req);
2518 rc = ptlrpc_reply(req);
2523 CERROR("invalid opcode %d\n",
2524 lustre_msg_get_opc(req->rq_reqmsg));
2525 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2526 rc = ldlm_callback_reply(req, -EINVAL);
2532 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2533 struct ldlm_lock *lock)
2535 struct ldlm_request *dlm_req;
2536 struct lustre_handle lockh;
2542 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2543 if (dlm_req == NULL)
2546 ldlm_lock2handle(lock, &lockh);
2547 for (i = 0; i < dlm_req->lock_count; i++) {
2548 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2550 DEBUG_REQ(D_RPCTRACE, req,
2551 "Prio raised by lock %#llx", lockh.cookie);
2560 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2562 struct ldlm_request *dlm_req;
2569 /* no prolong in recovery */
2570 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2573 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2574 if (dlm_req == NULL)
2577 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
2578 if (size <= offsetof(struct ldlm_request, lock_handle) ||
2579 (size - offsetof(struct ldlm_request, lock_handle)) /
2580 sizeof(struct lustre_handle) < dlm_req->lock_count)
2583 for (i = 0; i < dlm_req->lock_count; i++) {
2584 struct ldlm_lock *lock;
2586 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2590 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2592 LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2593 LDLM_LOCK_PUT(lock);
2602 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2603 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2604 .hpreq_check = ldlm_cancel_hpreq_check,
2608 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2612 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2614 if (req->rq_export == NULL)
2617 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2618 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2619 req->rq_ops = &ldlm_cancel_hpreq_ops;
2620 } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2621 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2622 req->rq_ops = &ldlm_cancel_hpreq_ops;
2627 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2628 struct hlist_node *hnode, void *data)
2631 struct list_head *rpc_list = data;
2632 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2634 lock_res_and_lock(lock);
2636 if (!ldlm_is_granted(lock)) {
2637 unlock_res_and_lock(lock);
2641 LASSERT(lock->l_resource);
2642 if (lock->l_resource->lr_type != LDLM_IBITS &&
2643 lock->l_resource->lr_type != LDLM_PLAIN) {
2644 unlock_res_and_lock(lock);
2648 if (ldlm_is_ast_sent(lock)) {
2649 unlock_res_and_lock(lock);
2653 LASSERT(lock->l_blocking_ast);
2654 LASSERT(!lock->l_blocking_lock);
2656 ldlm_set_ast_sent(lock);
2657 if (lock->l_export && lock->l_export->exp_lock_hash) {
2659 * NB: it's safe to call cfs_hash_del() even lock isn't
2663 * In the function below, .hs_keycmp resolves to
2664 * ldlm_export_lock_keycmp()
2666 /* coverity[overrun-buffer-val] */
2667 cfs_hash_del(lock->l_export->exp_lock_hash,
2668 &lock->l_remote_handle, &lock->l_exp_hash);
2671 list_add_tail(&lock->l_rk_ast, rpc_list);
2672 LDLM_LOCK_GET(lock);
2674 unlock_res_and_lock(lock);
2678 void ldlm_revoke_export_locks(struct obd_export *exp)
2680 LIST_HEAD(rpc_list);
2684 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2685 ldlm_revoke_lock_cb, &rpc_list, 0);
2686 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2687 LDLM_WORK_REVOKE_AST);
2691 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2692 #endif /* HAVE_SERVER_SUPPORT */
2694 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2695 struct ldlm_bl_work_item **p_blwi,
2696 struct obd_export **p_exp)
2698 struct ldlm_bl_work_item *blwi = NULL;
2699 static unsigned int num_bl;
2700 static unsigned int num_stale;
2701 int num_th = atomic_read(&blp->blp_num_threads);
2703 *p_exp = obd_stale_export_get();
2705 spin_lock(&blp->blp_lock);
2706 if (*p_exp != NULL) {
2707 if (num_th == 1 || ++num_stale < num_th) {
2708 spin_unlock(&blp->blp_lock);
2714 /* process a request from the blp_list at least every blp_num_threads */
2715 if (!list_empty(&blp->blp_list) &&
2716 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2717 blwi = list_entry(blp->blp_list.next,
2718 struct ldlm_bl_work_item, blwi_entry);
2720 if (!list_empty(&blp->blp_prio_list))
2721 blwi = list_entry(blp->blp_prio_list.next,
2722 struct ldlm_bl_work_item,
2726 if (++num_bl >= num_th)
2728 list_del(&blwi->blwi_entry);
2730 spin_unlock(&blp->blp_lock);
2733 if (*p_exp != NULL && *p_blwi != NULL) {
2734 obd_stale_export_put(*p_exp);
2738 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2741 /* This only contains temporary data until the thread starts */
2742 struct ldlm_bl_thread_data {
2743 struct ldlm_bl_pool *bltd_blp;
2744 struct completion bltd_comp;
2748 static int ldlm_bl_thread_main(void *arg);
2750 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2752 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2753 struct task_struct *task;
2755 init_completion(&bltd.bltd_comp);
2757 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2758 if (bltd.bltd_num >= blp->blp_max_threads) {
2759 atomic_dec(&blp->blp_num_threads);
2763 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2765 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2766 atomic_dec(&blp->blp_num_threads);
2770 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2773 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2774 bltd.bltd_num, PTR_ERR(task));
2775 atomic_dec(&blp->blp_num_threads);
2776 return PTR_ERR(task);
2778 wait_for_completion(&bltd.bltd_comp);
2783 /* Not fatal if racy and have a few too many threads */
2784 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2785 struct ldlm_bl_work_item *blwi)
2787 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2790 if (atomic_read(&blp->blp_busy_threads) <
2791 atomic_read(&blp->blp_num_threads))
2794 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2795 blwi->blwi_mem_pressure))
2801 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2802 struct ldlm_bl_work_item *blwi)
2806 if (blwi->blwi_ns == NULL)
2807 /* added by ldlm_cleanup() */
2808 RETURN(LDLM_ITER_STOP);
2810 if (blwi->blwi_mem_pressure)
2811 memory_pressure_set();
2813 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2815 if (blwi->blwi_count) {
2818 * The special case when we cancel locks in lru
2819 * asynchronously, we pass the list of locks here.
2820 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2821 * canceled locally yet.
2823 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2826 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2829 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2832 if (blwi->blwi_mem_pressure)
2833 memory_pressure_clr();
2835 if (blwi->blwi_flags & LCF_ASYNC)
2836 OBD_FREE(blwi, sizeof(*blwi));
2838 complete(&blwi->blwi_comp);
2844 * Cancel stale locks on export. Cancel blocked locks first.
2845 * If the given export has blocked locks, the next in the list may have
2846 * them too, thus cancel not blocked locks only if the current export has
2849 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2850 struct obd_export *exp)
2856 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2858 num = ldlm_export_cancel_blocked_locks(exp);
2860 ldlm_export_cancel_locks(exp);
2862 obd_stale_export_put(exp);
2869 * Main blocking requests processing thread.
2871 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2872 * This thread in the end ends up doing actual call to ->l_blocking_ast
2875 static int ldlm_bl_thread_main(void *arg)
2878 struct ldlm_bl_pool *blp;
2879 struct ldlm_bl_thread_data *bltd = arg;
2887 rc = lu_env_init(env, LCT_DT_THREAD);
2890 rc = lu_env_add(env);
2892 GOTO(out_env_fini, rc);
2894 blp = bltd->bltd_blp;
2896 complete(&bltd->bltd_comp);
2897 /* cannot use bltd after this, it is only on caller's stack */
2900 struct ldlm_bl_work_item *blwi = NULL;
2901 struct obd_export *exp = NULL;
2904 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2907 wait_event_idle_exclusive(blp->blp_waitq,
2908 ldlm_bl_get_work(blp, &blwi,
2910 atomic_inc(&blp->blp_busy_threads);
2912 if (ldlm_bl_thread_need_create(blp, blwi))
2913 /* discard the return value, we tried */
2914 ldlm_bl_thread_start(blp, true);
2917 rc = ldlm_bl_thread_exports(blp, exp);
2919 rc = ldlm_bl_thread_blwi(blp, blwi);
2921 atomic_dec(&blp->blp_busy_threads);
2923 if (rc == LDLM_ITER_STOP)
2927 * If there are many namespaces, we will not sleep waiting for
2928 * work, and must do a cond_resched to avoid holding the CPU
2934 atomic_dec(&blp->blp_num_threads);
2935 complete(&blp->blp_comp);
2946 static int ldlm_setup(void);
2947 static int ldlm_cleanup(void);
2949 int ldlm_get_ref(void)
2954 mutex_lock(&ldlm_ref_mutex);
2955 if (++ldlm_refcount == 1) {
2960 mutex_unlock(&ldlm_ref_mutex);
2965 void ldlm_put_ref(void)
2968 mutex_lock(&ldlm_ref_mutex);
2969 if (ldlm_refcount == 1) {
2970 int rc = ldlm_cleanup();
2973 CERROR("ldlm_cleanup failed: %d\n", rc);
2979 mutex_unlock(&ldlm_ref_mutex);
2985 * Export handle<->lock hash operations.
2988 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned int mask)
2990 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2994 ldlm_export_lock_key(struct hlist_node *hnode)
2996 struct ldlm_lock *lock;
2998 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2999 return &lock->l_remote_handle;
3003 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
3005 struct ldlm_lock *lock;
3007 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3008 lock->l_remote_handle = *(struct lustre_handle *)key;
3012 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
3014 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
3018 ldlm_export_lock_object(struct hlist_node *hnode)
3020 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3024 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
3026 struct ldlm_lock *lock;
3028 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3029 LDLM_LOCK_GET(lock);
3033 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
3035 struct ldlm_lock *lock;
3037 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3038 LDLM_LOCK_RELEASE(lock);
3041 static struct cfs_hash_ops ldlm_export_lock_ops = {
3042 .hs_hash = ldlm_export_lock_hash,
3043 .hs_key = ldlm_export_lock_key,
3044 .hs_keycmp = ldlm_export_lock_keycmp,
3045 .hs_keycpy = ldlm_export_lock_keycpy,
3046 .hs_object = ldlm_export_lock_object,
3047 .hs_get = ldlm_export_lock_get,
3048 .hs_put = ldlm_export_lock_put,
3049 .hs_put_locked = ldlm_export_lock_put,
3052 int ldlm_init_export(struct obd_export *exp)
3058 exp->exp_lock_hash =
3059 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
3060 HASH_EXP_LOCK_CUR_BITS,
3061 HASH_EXP_LOCK_MAX_BITS,
3062 HASH_EXP_LOCK_BKT_BITS, 0,
3063 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
3064 &ldlm_export_lock_ops,
3065 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
3066 CFS_HASH_NBLK_CHANGE);
3068 if (!exp->exp_lock_hash)
3071 rc = ldlm_init_flock_export(exp);
3077 ldlm_destroy_export(exp);
3080 EXPORT_SYMBOL(ldlm_init_export);
3082 void ldlm_destroy_export(struct obd_export *exp)
3085 cfs_hash_putref(exp->exp_lock_hash);
3086 exp->exp_lock_hash = NULL;
3088 ldlm_destroy_flock_export(exp);
3091 EXPORT_SYMBOL(ldlm_destroy_export);
3093 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
3094 struct attribute *attr,
3097 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
3100 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
3101 struct attribute *attr,
3108 rc = kstrtoul(buffer, 10, &val);
3112 ldlm_cancel_unused_locks_before_replay = val;
3116 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
3118 static struct attribute *ldlm_attrs[] = {
3119 &lustre_attr_cancel_unused_locks_before_replay.attr,
3123 static struct attribute_group ldlm_attr_group = {
3124 .attrs = ldlm_attrs,
3127 static int ldlm_setup(void)
3129 static struct ptlrpc_service_conf conf;
3130 struct ldlm_bl_pool *blp = NULL;
3131 #ifdef HAVE_SERVER_SUPPORT
3132 struct task_struct *task;
3133 #endif /* HAVE_SERVER_SUPPORT */
3139 if (ldlm_state != NULL)
3142 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
3143 if (ldlm_state == NULL)
3146 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
3150 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
3154 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
3158 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
3162 rc = ldlm_debugfs_setup();
3166 memset(&conf, 0, sizeof(conf));
3167 conf = (typeof(conf)) {
3168 .psc_name = "ldlm_cbd",
3169 .psc_watchdog_factor = 2,
3171 .bc_nbufs = LDLM_CLIENT_NBUFS,
3172 .bc_buf_size = LDLM_BUFSIZE,
3173 .bc_req_max_size = LDLM_MAXREQSIZE,
3174 .bc_rep_max_size = LDLM_MAXREPSIZE,
3175 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
3176 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
3179 .tc_thr_name = "ldlm_cb",
3180 .tc_thr_factor = LDLM_THR_FACTOR,
3181 .tc_nthrs_init = LDLM_NTHRS_INIT,
3182 .tc_nthrs_base = LDLM_NTHRS_BASE,
3183 .tc_nthrs_max = LDLM_NTHRS_MAX,
3184 .tc_nthrs_user = ldlm_num_threads,
3185 .tc_cpu_bind = ldlm_cpu_bind,
3186 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
3189 .cc_pattern = ldlm_cpts,
3190 .cc_affinity = true,
3193 .so_req_handler = ldlm_callback_handler,
3196 ldlm_state->ldlm_cb_service = \
3197 ptlrpc_register_service(&conf, ldlm_svc_kset,
3198 ldlm_svc_debugfs_dir);
3199 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3200 CERROR("failed to start service\n");
3201 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3202 ldlm_state->ldlm_cb_service = NULL;
3206 #ifdef HAVE_SERVER_SUPPORT
3207 memset(&conf, 0, sizeof(conf));
3208 conf = (typeof(conf)) {
3209 .psc_name = "ldlm_canceld",
3210 .psc_watchdog_factor = 6,
3212 .bc_nbufs = LDLM_SERVER_NBUFS,
3213 .bc_buf_size = LDLM_BUFSIZE,
3214 .bc_req_max_size = LDLM_MAXREQSIZE,
3215 .bc_rep_max_size = LDLM_MAXREPSIZE,
3216 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3217 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3221 .tc_thr_name = "ldlm_cn",
3222 .tc_thr_factor = LDLM_THR_FACTOR,
3223 .tc_nthrs_init = LDLM_NTHRS_INIT,
3224 .tc_nthrs_base = LDLM_NTHRS_BASE,
3225 .tc_nthrs_max = LDLM_NTHRS_MAX,
3226 .tc_nthrs_user = ldlm_num_threads,
3227 .tc_cpu_bind = ldlm_cpu_bind,
3228 .tc_ctx_tags = LCT_MD_THREAD | \
3233 .cc_pattern = ldlm_cpts,
3234 .cc_affinity = true,
3237 .so_req_handler = ldlm_cancel_handler,
3238 .so_hpreq_handler = ldlm_hpreq_handler,
3241 ldlm_state->ldlm_cancel_service = \
3242 ptlrpc_register_service(&conf, ldlm_svc_kset,
3243 ldlm_svc_debugfs_dir);
3244 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3245 CERROR("failed to start service\n");
3246 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3247 ldlm_state->ldlm_cancel_service = NULL;
3250 #endif /* HAVE_SERVER_SUPPORT */
3252 OBD_ALLOC(blp, sizeof(*blp));
3254 GOTO(out, rc = -ENOMEM);
3255 ldlm_state->ldlm_bl_pool = blp;
3257 spin_lock_init(&blp->blp_lock);
3258 INIT_LIST_HEAD(&blp->blp_list);
3259 INIT_LIST_HEAD(&blp->blp_prio_list);
3260 init_waitqueue_head(&blp->blp_waitq);
3261 atomic_set(&blp->blp_num_threads, 0);
3262 atomic_set(&blp->blp_busy_threads, 0);
3264 if (ldlm_num_threads == 0) {
3265 blp->blp_min_threads = LDLM_NTHRS_INIT;
3266 blp->blp_max_threads = LDLM_NTHRS_MAX;
3268 blp->blp_min_threads = blp->blp_max_threads = \
3269 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3273 for (i = 0; i < blp->blp_min_threads; i++) {
3274 rc = ldlm_bl_thread_start(blp, false);
3279 #ifdef HAVE_SERVER_SUPPORT
3280 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3283 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3287 wait_event(expired_lock_wait_queue,
3288 expired_lock_thread_state == ELT_READY);
3289 #endif /* HAVE_SERVER_SUPPORT */
3291 rc = ldlm_pools_init();
3293 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3297 rc = ldlm_reclaim_setup();
3299 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3309 static int ldlm_cleanup(void)
3313 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3314 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3315 CERROR("ldlm still has namespaces; clean these up first.\n");
3316 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3317 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3321 ldlm_reclaim_cleanup();
3324 if (ldlm_state->ldlm_bl_pool != NULL) {
3325 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3327 while (atomic_read(&blp->blp_num_threads) > 0) {
3328 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3330 init_completion(&blp->blp_comp);
3332 spin_lock(&blp->blp_lock);
3333 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3334 wake_up(&blp->blp_waitq);
3335 spin_unlock(&blp->blp_lock);
3337 wait_for_completion(&blp->blp_comp);
3340 OBD_FREE(blp, sizeof(*blp));
3343 if (ldlm_state->ldlm_cb_service != NULL)
3344 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3345 #ifdef HAVE_SERVER_SUPPORT
3346 if (ldlm_state->ldlm_cancel_service != NULL)
3347 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3351 kset_unregister(ldlm_ns_kset);
3353 kset_unregister(ldlm_svc_kset);
3355 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3356 kobject_put(ldlm_kobj);
3359 ldlm_debugfs_cleanup();
3361 #ifdef HAVE_SERVER_SUPPORT
3362 if (expired_lock_thread_state != ELT_STOPPED) {
3363 expired_lock_thread_state = ELT_TERMINATE;
3364 wake_up(&expired_lock_wait_queue);
3365 wait_event(expired_lock_wait_queue,
3366 expired_lock_thread_state == ELT_STOPPED);
3370 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3376 void ldlm_resource_init_once(void *p)
3379 * It is import to initialise the spinlock only once,
3380 * as ldlm_lock_change_resource() could try to lock
3381 * the resource *after* it has been freed and possibly
3382 * reused. SLAB_TYPESAFE_BY_RCU ensures the memory won't
3383 * be freed while the lock is being taken, but we need to
3384 * ensure that it doesn't get reinitialized either.
3386 struct ldlm_resource *res = p;
3388 memset(res, 0, sizeof(*res));
3389 mutex_init(&res->lr_lvb_mutex);
3390 spin_lock_init(&res->lr_lock);
3395 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3396 sizeof(struct ldlm_resource), 0,
3397 SLAB_TYPESAFE_BY_RCU |
3399 ldlm_resource_init_once);
3400 if (ldlm_resource_slab == NULL)
3403 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3404 sizeof(struct ldlm_lock), 0,
3405 SLAB_HWCACHE_ALIGN, NULL);
3406 if (ldlm_lock_slab == NULL)
3409 ldlm_interval_slab = kmem_cache_create("interval_node",
3410 sizeof(struct ldlm_interval),
3411 0, SLAB_HWCACHE_ALIGN, NULL);
3412 if (ldlm_interval_slab == NULL)
3415 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3416 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3417 0, SLAB_HWCACHE_ALIGN, NULL);
3418 if (ldlm_interval_tree_slab == NULL)
3421 #ifdef HAVE_SERVER_SUPPORT
3422 ldlm_inodebits_slab = kmem_cache_create("ldlm_ibits_node",
3423 sizeof(struct ldlm_ibits_node),
3424 0, SLAB_HWCACHE_ALIGN, NULL);
3425 if (ldlm_inodebits_slab == NULL)
3426 goto out_interval_tree;
3428 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3429 sizeof(struct ldlm_glimpse_work),
3431 if (ldlm_glimpse_work_kmem == NULL)
3435 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3436 class_export_dump_hook = ldlm_dump_export_locks;
3439 #ifdef HAVE_SERVER_SUPPORT
3441 kmem_cache_destroy(ldlm_inodebits_slab);
3443 kmem_cache_destroy(ldlm_interval_tree_slab);
3446 kmem_cache_destroy(ldlm_interval_slab);
3448 kmem_cache_destroy(ldlm_lock_slab);
3450 kmem_cache_destroy(ldlm_resource_slab);
3455 void ldlm_exit(void)
3458 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3459 /* These two lines should not be needed, but appear to fix
3460 * a crash on RHEL7. The slab_cache sometimes gets freed before the
3461 * last slab is rcu_freed, and that can cause kmem_freepages()
3462 * to free too many pages and trip a BUG
3464 kmem_cache_shrink(ldlm_resource_slab);
3466 kmem_cache_destroy(ldlm_resource_slab);
3468 * ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3469 * rcu_barrier() to wait all outstanding RCU callbacks to complete,
3470 * so that ldlm_lock_free() get a chance to be called.
3473 kmem_cache_destroy(ldlm_lock_slab);
3474 kmem_cache_destroy(ldlm_interval_slab);
3475 kmem_cache_destroy(ldlm_interval_tree_slab);
3476 #ifdef HAVE_SERVER_SUPPORT
3477 kmem_cache_destroy(ldlm_inodebits_slab);
3478 kmem_cache_destroy(ldlm_glimpse_work_kmem);