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;
109 struct ldlm_bl_work_item {
110 struct list_head blwi_entry;
111 struct ldlm_namespace *blwi_ns;
112 struct ldlm_lock_desc blwi_ld;
113 struct ldlm_lock *blwi_lock;
114 struct list_head blwi_head;
116 struct completion blwi_comp;
117 enum ldlm_cancel_flags blwi_flags;
118 int blwi_mem_pressure;
121 #ifdef HAVE_SERVER_SUPPORT
124 * Protects both waiting_locks_list and expired_lock_thread.
126 static DEFINE_SPINLOCK(waiting_locks_spinlock); /* BH lock (timer) */
129 * List for contended locks.
131 * As soon as a lock is contended, it gets placed on this list and
132 * expected time to get a response is filled in the lock. A special
133 * thread walks the list looking for locks that should be released and
134 * schedules client evictions for those that have not been released in
137 * All access to it should be under waiting_locks_spinlock.
139 static LIST_HEAD(waiting_locks_list);
140 static void waiting_locks_callback(TIMER_DATA_TYPE unused);
141 static CFS_DEFINE_TIMER(waiting_locks_timer, waiting_locks_callback, 0, 0);
149 static DECLARE_WAIT_QUEUE_HEAD(expired_lock_wait_queue);
150 static enum elt_state expired_lock_thread_state = ELT_STOPPED;
151 static int expired_lock_dump;
152 static LIST_HEAD(expired_lock_list);
154 static int ldlm_lock_busy(struct ldlm_lock *lock);
155 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t timeout);
156 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t timeout);
158 static inline int have_expired_locks(void)
163 spin_lock_bh(&waiting_locks_spinlock);
164 need_to_run = !list_empty(&expired_lock_list);
165 spin_unlock_bh(&waiting_locks_spinlock);
171 * Check expired lock list for expired locks and time them out.
173 static int expired_lock_main(void *arg)
175 struct list_head *expired = &expired_lock_list;
180 expired_lock_thread_state = ELT_READY;
181 wake_up(&expired_lock_wait_queue);
184 wait_event_idle(expired_lock_wait_queue,
185 have_expired_locks() ||
186 expired_lock_thread_state == ELT_TERMINATE);
188 spin_lock_bh(&waiting_locks_spinlock);
189 if (expired_lock_dump) {
190 spin_unlock_bh(&waiting_locks_spinlock);
192 /* from waiting_locks_callback, but not in timer */
193 libcfs_debug_dumplog();
195 spin_lock_bh(&waiting_locks_spinlock);
196 expired_lock_dump = 0;
201 while (!list_empty(expired)) {
202 struct obd_export *export;
203 struct ldlm_lock *lock;
205 lock = list_entry(expired->next, struct ldlm_lock,
207 if ((void *)lock < LP_POISON + PAGE_SIZE &&
208 (void *)lock >= LP_POISON) {
209 spin_unlock_bh(&waiting_locks_spinlock);
210 CERROR("free lock on elt list %p\n", lock);
213 list_del_init(&lock->l_pending_chain);
214 if ((void *)lock->l_export <
215 LP_POISON + PAGE_SIZE &&
216 (void *)lock->l_export >= LP_POISON) {
217 CERROR("lock with free export on elt list %p\n",
219 lock->l_export = NULL;
220 LDLM_ERROR(lock, "free export");
222 * release extra ref grabbed by
223 * ldlm_add_waiting_lock() or
226 LDLM_LOCK_RELEASE(lock);
230 if (ldlm_is_destroyed(lock)) {
232 * release the lock refcount where
233 * waiting_locks_callback() founds
235 LDLM_LOCK_RELEASE(lock);
238 export = class_export_lock_get(lock->l_export, lock);
239 spin_unlock_bh(&waiting_locks_spinlock);
241 /* Check if we need to prolong timeout */
242 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
243 lock->l_callback_timestamp != 0 && /* not AST error */
244 ldlm_lock_busy(lock)) {
245 LDLM_DEBUG(lock, "prolong the busy lock");
246 lock_res_and_lock(lock);
247 ldlm_add_waiting_lock(lock,
248 ldlm_bl_timeout(lock) >> 1);
249 unlock_res_and_lock(lock);
251 spin_lock_bh(&export->exp_bl_list_lock);
252 list_del_init(&lock->l_exp_list);
253 spin_unlock_bh(&export->exp_bl_list_lock);
256 "lock callback timer expired after %llds: evicting client at %s ",
257 ktime_get_seconds() -
259 obd_export_nid2str(export));
260 ldlm_lock_to_ns(lock)->ns_timeouts++;
261 if (do_dump_on_eviction(export->exp_obd))
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);
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_timestamp > 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_timestamp - now > 0)
358 delta = lock->l_callback_timestamp - 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, timeout_t delay)
379 unsigned long timeout_jiffies = jiffies;
383 lock->l_blast_sent = 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 = lock->l_blast_sent + delay;
392 if (likely(deadline > lock->l_callback_timestamp))
393 lock->l_callback_timestamp = deadline;
395 timeout = clamp_t(timeout_t,
396 lock->l_callback_timestamp - lock->l_blast_sent,
398 timeout_jiffies += cfs_time_seconds(timeout);
400 if (time_before(timeout_jiffies, waiting_locks_timer.expires) ||
401 !timer_pending(&waiting_locks_timer))
402 mod_timer(&waiting_locks_timer, timeout_jiffies);
405 * if the new lock has a shorter timeout than something earlier on
406 * the list, we'll wait the longer amount of time; no big deal.
409 list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
413 static void ldlm_add_blocked_lock(struct ldlm_lock *lock)
415 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
416 if (list_empty(&lock->l_exp_list)) {
417 if (!ldlm_is_granted(lock))
418 list_add_tail(&lock->l_exp_list,
419 &lock->l_export->exp_bl_list);
421 list_add(&lock->l_exp_list,
422 &lock->l_export->exp_bl_list);
424 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
427 * A blocked lock is added. Adjust the position in
428 * the stale list if the export is in the list.
429 * If export is stale and not in the list - it is being
430 * processed and will be placed on the right position
431 * on obd_stale_export_put().
433 if (!list_empty(&lock->l_export->exp_stale_list))
434 obd_stale_export_adjust(lock->l_export);
437 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t timeout)
441 /* NB: must be called with hold of lock_res_and_lock() */
442 LASSERT(ldlm_is_res_locked(lock));
443 LASSERT(!ldlm_is_cancel_on_block(lock));
446 * Do not put cross-MDT lock in the waiting list, since we
447 * will not evict it due to timeout for now
449 if (lock->l_export != NULL &&
450 (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS))
453 spin_lock_bh(&waiting_locks_spinlock);
454 if (ldlm_is_cancel(lock)) {
455 spin_unlock_bh(&waiting_locks_spinlock);
459 if (ldlm_is_destroyed(lock)) {
460 static time64_t next;
462 spin_unlock_bh(&waiting_locks_spinlock);
463 LDLM_ERROR(lock, "not waiting on destroyed lock (b=5653)");
464 if (ktime_get_seconds() > next) {
465 next = ktime_get_seconds() + 14400;
466 libcfs_debug_dumpstack(NULL);
471 ldlm_set_waited(lock);
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: %d, 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_timestamp - now > 0)
521 delta = lock->l_callback_timestamp - 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, timeout_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 to %ds", timeout);
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, timeout_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 timeout_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_t(timeout_t, timeout + (timeout >> 1),
644 (timeout_t)ldlm_enqueue_min);
646 EXPORT_SYMBOL(ldlm_bl_timeout);
649 * Perform lock cleanup if AST sending failed.
651 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
652 const char *ast_type)
654 LCONSOLE_ERROR_MSG(0x138,
655 "%s: A client on nid %s was evicted due to a lock %s callback time out: rc %d\n",
656 lock->l_export->exp_obd->obd_name,
657 obd_export_nid2str(lock->l_export), ast_type, rc);
659 if (obd_dump_on_timeout)
660 libcfs_debug_dumplog();
661 spin_lock_bh(&waiting_locks_spinlock);
662 if (__ldlm_del_waiting_lock(lock) == 0)
664 * the lock was not in any list, grab an extra ref before adding
665 * the lock to the expired list
668 /* differentiate it from expired locks */
669 lock->l_callback_timestamp = 0;
670 list_add(&lock->l_pending_chain, &expired_lock_list);
671 wake_up(&expired_lock_wait_queue);
672 spin_unlock_bh(&waiting_locks_spinlock);
676 * Perform lock cleanup if AST reply came with error.
678 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
679 struct ptlrpc_request *req, int rc,
680 const char *ast_type)
682 struct lnet_processid *peer = &req->rq_import->imp_connection->c_peer;
684 if (!req->rq_replied || (rc && rc != -EINVAL)) {
685 if (ldlm_is_cancel(lock)) {
687 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
688 ast_type, req, req->rq_xid,
689 libcfs_nidstr(&peer->nid));
690 ldlm_lock_cancel(lock);
692 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
694 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
696 * Upon umount process the AST fails because cannot be
697 * sent. This shouldn't lead to the client eviction.
698 * -ENODEV error is returned by ptl_send_rpc() for
699 * new request in such import.
700 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
701 * if imp_invalid is set or obd_no_recov.
702 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
703 * in ptlrpc_import_delay_req() as well with -EIO code.
704 * In all such cases errors are ignored.
707 "%s AST can't be sent due to a server %s failure or umount process: rc = %d\n",
709 req->rq_import->imp_obd->obd_name, rc);
712 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
713 libcfs_nidstr(&peer->nid),
714 req->rq_replied ? "returned error from" :
715 "failed to reply to",
716 ast_type, req, req->rq_xid,
717 (req->rq_repmsg != NULL) ?
718 lustre_msg_get_status(req->rq_repmsg) : 0,
720 ldlm_failed_ast(lock, rc, ast_type);
726 struct ldlm_resource *res = lock->l_resource;
729 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
730 libcfs_nidstr(&peer->nid),
732 lustre_msg_get_status(req->rq_repmsg) : -1,
733 ast_type, req, req->rq_xid);
736 * update lvbo to return proper attributes.
739 ldlm_resource_getref(res);
740 ldlm_lvbo_update(res, lock, NULL, 1);
741 ldlm_resource_putref(res);
743 ldlm_lock_cancel(lock);
750 static int ldlm_cb_interpret(const struct lu_env *env,
751 struct ptlrpc_request *req, void *args, int rc)
753 struct ldlm_cb_async_args *ca = args;
754 struct ldlm_lock *lock = ca->ca_lock;
755 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
759 LASSERT(lock != NULL);
762 case LDLM_GL_CALLBACK:
764 * Update the LVB from disk if the AST failed
765 * (this is a legal race)
767 * - Glimpse callback of local lock just returns
768 * -ELDLM_NO_LOCK_DATA.
769 * - Glimpse callback of remote lock might return
770 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
772 if (unlikely(arg->gl_interpret_reply)) {
773 rc = arg->gl_interpret_reply(NULL, req, args, rc);
774 } else if (rc == -ELDLM_NO_LOCK_DATA) {
776 "lost race - client has a lock but no inode");
777 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
778 } else if (rc != 0) {
779 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
781 rc = ldlm_lvbo_update(lock->l_resource,
785 case LDLM_BL_CALLBACK:
787 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
789 case LDLM_CP_CALLBACK:
791 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
794 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
799 /* release extra reference taken in ldlm_ast_fini() */
800 LDLM_LOCK_RELEASE(lock);
803 atomic_inc(&arg->restart);
808 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
810 struct ldlm_cb_async_args *ca = data;
811 struct ldlm_lock *lock = ca->ca_lock;
813 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
816 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
817 struct ldlm_cb_set_arg *arg,
818 struct ldlm_lock *lock,
825 if (unlikely(instant_cancel)) {
826 rc = ptl_send_rpc(req, 1);
827 ptlrpc_req_finished(req);
829 atomic_inc(&arg->restart);
832 ptlrpc_set_add_req(arg->set, req);
839 * Check if there are requests in the export request list which prevent
840 * the lock canceling and make these requests high priority ones.
842 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
844 struct ptlrpc_request *req;
848 if (lock->l_export == NULL) {
849 LDLM_DEBUG(lock, "client lock: no-op");
853 spin_lock(&lock->l_export->exp_rpc_lock);
854 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
857 * Do not process requests that were not yet added to there
858 * incoming queue or were already removed from there for
859 * processing. We evaluate ptlrpc_nrs_req_can_move() without
860 * holding svcpt->scp_req_lock, and then redo the check with
861 * the lock held once we need to obtain a reliable result.
863 if (ptlrpc_nrs_req_can_move(req) &&
864 req->rq_ops->hpreq_lock_match &&
865 req->rq_ops->hpreq_lock_match(req, lock))
866 ptlrpc_nrs_req_hp_move(req);
868 spin_unlock(&lock->l_export->exp_rpc_lock);
873 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
874 * enqueued server lock conflicts with given one.
876 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
877 * to wait for client response.
879 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
880 struct ldlm_lock_desc *desc,
881 void *data, int flag)
883 struct ldlm_cb_async_args *ca;
884 struct ldlm_cb_set_arg *arg = data;
885 struct ldlm_request *body;
886 struct ptlrpc_request *req;
887 int instant_cancel = 0;
892 if (flag == LDLM_CB_CANCELING)
893 /* Don't need to do anything here. */
896 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
897 LDLM_DEBUG(lock, "dropping BL AST");
902 LASSERT(data != NULL);
903 if (lock->l_export->exp_obd->obd_recovering != 0)
904 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
906 ldlm_lock_reorder_req(lock);
908 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
909 &RQF_LDLM_BL_CALLBACK,
910 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
914 ca = ptlrpc_req_async_args(ca, req);
915 ca->ca_set_arg = arg;
918 req->rq_interpret_reply = ldlm_cb_interpret;
920 lock_res_and_lock(lock);
921 if (ldlm_is_destroyed(lock)) {
922 /* What's the point? */
923 unlock_res_and_lock(lock);
924 ptlrpc_req_finished(req);
928 if (!ldlm_is_granted(lock)) {
930 * this blocking AST will be communicated as part of the
931 * completion AST instead
933 ldlm_add_blocked_lock(lock);
934 ldlm_set_waited(lock);
935 unlock_res_and_lock(lock);
937 ptlrpc_req_finished(req);
938 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
942 if (ldlm_is_cancel_on_block(lock))
945 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
946 body->lock_handle[0] = lock->l_remote_handle;
947 body->lock_handle[1].cookie = lock->l_handle.h_cookie;
948 body->lock_desc = *desc;
949 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
951 LDLM_DEBUG(lock, "server preparing blocking AST");
953 ptlrpc_request_set_replen(req);
954 ldlm_set_cbpending(lock);
955 if (instant_cancel) {
956 unlock_res_and_lock(lock);
957 ldlm_lock_cancel(lock);
959 req->rq_no_resend = 1;
961 LASSERT(ldlm_is_granted(lock));
962 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
963 unlock_res_and_lock(lock);
965 /* Do not resend after lock callback timeout */
966 req->rq_delay_limit = ldlm_bl_timeout(lock);
967 req->rq_resend_cb = ldlm_update_resend;
970 req->rq_send_state = LUSTRE_IMP_FULL;
971 /* ptlrpc_request_alloc_pack already set timeout */
973 req->rq_timeout = ldlm_get_rq_timeout();
975 if (lock->l_export && lock->l_export->exp_nid_stats &&
976 lock->l_export->exp_nid_stats->nid_ldlm_stats)
977 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
978 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
980 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
986 * ->l_completion_ast callback for a remote lock in server namespace.
988 * Sends AST to the client notifying it of lock granting. If initial
989 * lock response was not sent yet, instead of sending another RPC, just
990 * mark the lock as granted and client will understand
992 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
994 struct ldlm_cb_set_arg *arg = data;
995 struct ldlm_request *body;
996 struct ptlrpc_request *req;
997 struct ldlm_cb_async_args *ca;
998 int instant_cancel = 0;
1004 LASSERT(lock != NULL);
1005 LASSERT(data != NULL);
1007 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
1008 LDLM_DEBUG(lock, "dropping CP AST");
1012 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
1013 &RQF_LDLM_CP_CALLBACK);
1017 /* server namespace, doesn't need lock */
1018 lvb_len = ldlm_lvbo_size(lock);
1020 * LU-3124 & LU-2187: to not return layout in completion AST because
1021 * it may deadlock for LU-2187, or client may not have enough space
1022 * for large layout. The layout will be returned to client with an
1023 * extra RPC to fetch xattr.lov
1025 if (ldlm_has_layout(lock))
1028 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
1029 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
1031 ptlrpc_request_free(req);
1035 ca = ptlrpc_req_async_args(ca, req);
1036 ca->ca_set_arg = arg;
1039 req->rq_interpret_reply = ldlm_cb_interpret;
1040 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1042 body->lock_handle[0] = lock->l_remote_handle;
1043 body->lock_handle[1].cookie = lock->l_handle.h_cookie;
1044 body->lock_flags = ldlm_flags_to_wire(flags);
1045 ldlm_lock2desc(lock, &body->lock_desc);
1047 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1048 lvb_len = ldlm_lvbo_fill(lock, lvb, &lvb_len);
1051 * We still need to send the RPC to wake up the blocked
1052 * enqueue thread on the client.
1054 * Consider old client, there is no better way to notify
1055 * the failure, just zero-sized the LVB, then the client
1056 * will fail out as "-EPROTO".
1058 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1062 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1067 LDLM_DEBUG(lock, "server preparing completion AST");
1069 ptlrpc_request_set_replen(req);
1071 req->rq_send_state = LUSTRE_IMP_FULL;
1072 /* ptlrpc_request_pack already set timeout */
1074 req->rq_timeout = ldlm_get_rq_timeout();
1076 /* We only send real blocking ASTs after the lock is granted */
1077 lock_res_and_lock(lock);
1078 if (ldlm_is_ast_sent(lock)) {
1079 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1080 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1081 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1085 * We might get here prior to ldlm_handle_enqueue setting
1086 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1087 * into waiting list, but this is safe and similar code in
1088 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1089 * that would not only cancel the lock, but will also remove
1090 * it from waiting list
1092 if (ldlm_is_cancel_on_block(lock)) {
1093 unlock_res_and_lock(lock);
1094 ldlm_lock_cancel(lock);
1097 req->rq_no_resend = 1;
1099 lock_res_and_lock(lock);
1101 /* start the lock-timeout clock */
1102 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1103 /* Do not resend after lock callback timeout */
1104 req->rq_delay_limit = ldlm_bl_timeout(lock);
1105 req->rq_resend_cb = ldlm_update_resend;
1108 unlock_res_and_lock(lock);
1110 if (lock->l_export && lock->l_export->exp_nid_stats &&
1111 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1112 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1113 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1115 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1117 RETURN(lvb_len < 0 ? lvb_len : rc);
1121 * Server side ->l_glimpse_ast handler for client locks.
1123 * Sends glimpse AST to the client and waits for reply. Then updates
1124 * lvbo with the result.
1126 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1128 struct ldlm_cb_set_arg *arg = data;
1129 struct ldlm_request *body;
1130 struct ptlrpc_request *req;
1131 struct ldlm_cb_async_args *ca;
1133 struct req_format *req_fmt;
1137 LASSERT(lock != NULL);
1139 if (arg->gl_desc != NULL)
1140 /* There is a glimpse descriptor to pack */
1141 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1143 req_fmt = &RQF_LDLM_GL_CALLBACK;
1145 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1146 req_fmt, LUSTRE_DLM_VERSION,
1152 if (arg->gl_desc != NULL) {
1153 /* copy the GL descriptor */
1154 union ldlm_gl_desc *desc;
1156 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1157 *desc = *arg->gl_desc;
1160 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1161 body->lock_handle[0] = lock->l_remote_handle;
1162 ldlm_lock2desc(lock, &body->lock_desc);
1164 ca = ptlrpc_req_async_args(ca, req);
1165 ca->ca_set_arg = arg;
1168 /* server namespace, doesn't need lock */
1169 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1170 ldlm_lvbo_size(lock));
1171 ptlrpc_request_set_replen(req);
1173 req->rq_send_state = LUSTRE_IMP_FULL;
1174 /* ptlrpc_request_alloc_pack already set timeout */
1176 req->rq_timeout = ldlm_get_rq_timeout();
1178 req->rq_interpret_reply = ldlm_cb_interpret;
1180 if (lock->l_export && lock->l_export->exp_nid_stats) {
1181 struct nid_stat *nid_stats = lock->l_export->exp_nid_stats;
1183 lprocfs_counter_incr(nid_stats->nid_ldlm_stats,
1184 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1187 rc = ldlm_ast_fini(req, arg, lock, 0);
1191 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1193 int ldlm_glimpse_locks(struct ldlm_resource *res,
1194 struct list_head *gl_work_list)
1200 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1202 if (rc == -ERESTART)
1203 ldlm_reprocess_all(res, 0);
1207 EXPORT_SYMBOL(ldlm_glimpse_locks);
1209 /* return LDLM lock associated with a lock callback request */
1210 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1212 struct ldlm_cb_async_args *ca;
1213 struct ldlm_lock *lock;
1217 ca = ptlrpc_req_async_args(ca, req);
1220 RETURN(ERR_PTR(-EFAULT));
1224 EXPORT_SYMBOL(ldlm_request_lock);
1227 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1228 * service threads to carry out client lock enqueueing requests.
1230 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1231 struct ptlrpc_request *req,
1232 const struct ldlm_request *dlm_req,
1233 const struct ldlm_callback_suite *cbs)
1235 struct ldlm_reply *dlm_rep;
1237 enum ldlm_error err = ELDLM_OK;
1238 struct ldlm_lock *lock = NULL;
1239 void *cookie = NULL;
1241 struct ldlm_resource *res = NULL;
1242 const struct lu_env *env = req->rq_svc_thread->t_env;
1246 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1248 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1249 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1251 LASSERT(req->rq_export);
1253 /* for intent enqueue the stat will be updated inside intent policy */
1254 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1255 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1256 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1258 if (req->rq_export && req->rq_export->exp_nid_stats &&
1259 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1260 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1261 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1263 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1264 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1265 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1266 dlm_req->lock_desc.l_resource.lr_type);
1267 GOTO(out, rc = -EFAULT);
1270 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1271 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1272 dlm_req->lock_desc.l_req_mode &
1273 (dlm_req->lock_desc.l_req_mode-1))) {
1274 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1275 dlm_req->lock_desc.l_req_mode);
1276 GOTO(out, rc = -EFAULT);
1279 if (unlikely((flags & LDLM_FL_REPLAY) ||
1280 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1281 /* Find an existing lock in the per-export lock hash */
1283 * In the function below, .hs_keycmp resolves to
1284 * ldlm_export_lock_keycmp()
1286 /* coverity[overrun-buffer-val] */
1287 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1288 (void *)&dlm_req->lock_handle[0]);
1290 DEBUG_REQ(D_DLMTRACE, req,
1291 "found existing lock cookie %#llx",
1292 lock->l_handle.h_cookie);
1293 flags |= LDLM_FL_RESENT;
1294 GOTO(existing_lock, rc = 0);
1297 if (ldlm_reclaim_full()) {
1298 DEBUG_REQ(D_DLMTRACE, req,
1299 "Too many granted locks, reject current enqueue request and let the client retry later");
1300 GOTO(out, rc = -EINPROGRESS);
1304 /* The lock's callback data might be set in the policy function */
1305 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1306 dlm_req->lock_desc.l_resource.lr_type,
1307 dlm_req->lock_desc.l_req_mode,
1308 cbs, NULL, 0, LVB_T_NONE);
1315 lock->l_remote_handle = dlm_req->lock_handle[0];
1316 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1319 * Initialize resource lvb but not for a lock being replayed since
1320 * Client already got lvb sent in this case.
1321 * This must occur early since some policy methods assume resource
1322 * lvb is available (lr_lvb_data != NULL).
1324 res = lock->l_resource;
1325 if (!(flags & LDLM_FL_REPLAY)) {
1326 /* non-replayed lock, delayed lvb init may need to be done */
1327 rc = ldlm_lvbo_init(res);
1329 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1334 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1336 * Don't enqueue a lock onto the export if it is been disonnected
1337 * due to eviction (b=3822) or server umount (b=24324).
1338 * Cancel it now instead.
1340 if (req->rq_export->exp_disconnected) {
1341 LDLM_ERROR(lock, "lock on disconnected export %p",
1343 GOTO(out, rc = -ENOTCONN);
1346 lock->l_export = class_export_lock_get(req->rq_export, lock);
1347 if (lock->l_export->exp_lock_hash)
1348 cfs_hash_add(lock->l_export->exp_lock_hash,
1349 &lock->l_remote_handle,
1353 * Inherit the enqueue flags before the operation, because we do not
1354 * keep the res lock on return and next operations (BL AST) may proceed
1357 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1358 LDLM_FL_INHERIT_MASK);
1360 ldlm_convert_policy_to_local(req->rq_export,
1361 dlm_req->lock_desc.l_resource.lr_type,
1362 &dlm_req->lock_desc.l_policy_data,
1363 &lock->l_policy_data);
1364 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT) {
1365 lock->l_req_extent = lock->l_policy_data.l_extent;
1366 } else if (dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) {
1367 lock->l_policy_data.l_inodebits.try_bits =
1368 dlm_req->lock_desc.l_policy_data.l_inodebits.try_bits;
1369 lock->l_policy_data.l_inodebits.li_gid =
1370 dlm_req->lock_desc.l_policy_data.l_inodebits.li_gid;
1375 if (!(flags & LDLM_FL_HAS_INTENT)) {
1376 /* based on the assumption that lvb size never changes during
1377 * resource life time otherwise it need resource->lr_lock's
1379 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1380 RCL_SERVER, ldlm_lvbo_size(lock));
1382 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1383 GOTO(out, rc = -ENOMEM);
1385 rc = req_capsule_server_pack(&req->rq_pill);
1390 err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1397 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1399 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1400 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1402 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1403 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1406 * We never send a blocking AST until the lock is granted, but
1407 * we can tell it right now
1409 lock_res_and_lock(lock);
1412 * Now take into account flags to be inherited from original lock
1413 * request both in reply to client and in our own lock flags.
1415 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1416 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1419 * Don't move a pending lock onto the export if it has already been
1420 * disconnected due to eviction (b=5683) or server umount (b=24324).
1421 * Cancel it now instead.
1423 if (unlikely(req->rq_export->exp_disconnected ||
1424 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1425 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1427 } else if (ldlm_is_ast_sent(lock)) {
1428 /* fill lock desc for possible lock convert */
1429 if (lock->l_blocking_lock &&
1430 lock->l_resource->lr_type == LDLM_IBITS) {
1431 struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1432 struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1435 "save blocking bits %llx in granted lock",
1436 bl_lock->l_policy_data.l_inodebits.bits);
1438 * If lock is blocked then save blocking ibits
1439 * in returned lock policy for the possible lock
1440 * convert on a client.
1442 rep_desc->l_policy_data.l_inodebits.cancel_bits =
1443 bl_lock->l_policy_data.l_inodebits.bits;
1445 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1446 if (ldlm_is_granted(lock)) {
1448 * Only cancel lock if it was granted, because it would
1449 * be destroyed immediately and would never be granted
1450 * in the future, causing timeouts on client. Not
1451 * granted lock will be cancelled immediately after
1452 * sending completion AST.
1454 if (ldlm_is_cancel_on_block(lock)) {
1455 unlock_res_and_lock(lock);
1456 ldlm_lock_cancel(lock);
1457 lock_res_and_lock(lock);
1459 ldlm_add_waiting_lock(lock,
1460 ldlm_bl_timeout(lock));
1464 unlock_res_and_lock(lock);
1468 req->rq_status = rc ?: err; /* return either error - b=11190 */
1469 if (!req->rq_packed_final) {
1470 int rc1 = lustre_pack_reply(req, 1, NULL, NULL);
1476 * The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1477 * ldlm_reprocess_all. If this moves, revisit that code. -phil
1481 "server-side enqueue handler, sending reply (err=%d, rc=%d)",
1485 req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1487 ldlm_lvbo_size(lock) > 0) {
1492 buf = req_capsule_server_get(&req->rq_pill,
1494 LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock);
1495 buflen = req_capsule_get_size(&req->rq_pill,
1496 &RMF_DLM_LVB, RCL_SERVER);
1498 * non-replayed lock, delayed lvb init may
1499 * need to be occur now
1501 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1504 rc2 = ldlm_lvbo_fill(lock, buf, &buflen);
1506 req_capsule_shrink(&req->rq_pill,
1509 } else if (rc2 == -ERANGE) {
1510 rc2 = req_capsule_server_grow(
1512 &RMF_DLM_LVB, buflen);
1517 * if we can't grow the buffer,
1518 * it's ok to return empty lvb
1529 } else if (flags & LDLM_FL_REPLAY) {
1530 /* no LVB resend upon replay */
1532 req_capsule_shrink(&req->rq_pill,
1542 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1543 if (lock->l_export) {
1544 ldlm_lock_cancel(lock);
1546 lock_res_and_lock(lock);
1547 ldlm_resource_unlink_lock(lock);
1548 ldlm_lock_destroy_nolock(lock);
1549 unlock_res_and_lock(lock);
1551 ldlm_reprocess_all(lock->l_resource, lock->l_policy_data.l_inodebits.bits);
1554 if (!err && !ldlm_is_cbpending(lock) &&
1555 dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1556 ldlm_reprocess_all(lock->l_resource,
1557 lock->l_policy_data.l_inodebits.bits);
1559 LDLM_LOCK_RELEASE(lock);
1562 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1569 * Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1570 * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1572 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1574 if (lock->l_blocking_lock) {
1575 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1576 lock->l_blocking_lock = NULL;
1580 /* A lock can be converted to new ibits or mode and should be considered
1581 * as new lock. Clear all states related to a previous blocking AST
1582 * processing so new conflicts will cause new blocking ASTs.
1584 * This is used during lock convert below and lock downgrade to COS mode in
1585 * ldlm_lock_mode_downgrade().
1587 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1589 ldlm_clear_ast_sent(lock);
1590 lock->l_bl_ast_run = 0;
1591 ldlm_clear_blocking_lock(lock);
1595 * Main LDLM entry point for server code to process lock conversion requests.
1597 int ldlm_handle_convert0(struct ptlrpc_request *req,
1598 const struct ldlm_request *dlm_req)
1600 struct obd_export *exp = req->rq_export;
1601 struct ldlm_reply *dlm_rep;
1602 struct ldlm_lock *lock;
1609 if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1610 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1611 LDLM_CONVERT - LDLM_FIRST_OPC);
1613 rc = req_capsule_server_pack(&req->rq_pill);
1617 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1618 dlm_rep->lock_flags = dlm_req->lock_flags;
1620 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1622 LDLM_DEBUG_NOLOCK("server lock is canceled already");
1623 req->rq_status = ELDLM_NO_LOCK_DATA;
1627 LDLM_DEBUG(lock, "server-side convert handler START");
1629 lock_res_and_lock(lock);
1630 bits = lock->l_policy_data.l_inodebits.bits;
1631 new_bits = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1633 if (ldlm_is_cancel(lock)) {
1634 LDLM_DEBUG(lock, "convert on canceled lock!");
1635 unlock_res_and_lock(lock);
1636 GOTO(out_put, rc = ELDLM_NO_LOCK_DATA);
1639 if (dlm_req->lock_desc.l_req_mode != lock->l_granted_mode) {
1640 LDLM_ERROR(lock, "lock mode differs!");
1641 unlock_res_and_lock(lock);
1642 GOTO(out_put, rc = -EPROTO);
1645 if (bits == new_bits) {
1647 * This can be valid situation if CONVERT RPCs are
1648 * re-ordered. Just finish silently
1650 LDLM_DEBUG(lock, "lock is converted already!");
1651 unlock_res_and_lock(lock);
1653 if (ldlm_is_waited(lock))
1654 ldlm_del_waiting_lock(lock);
1656 ldlm_clear_cbpending(lock);
1657 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1658 ldlm_inodebits_drop(lock, bits & ~new_bits);
1660 ldlm_clear_blocking_data(lock);
1661 unlock_res_and_lock(lock);
1663 /* All old bits should be reprocessed to send new BL AST if
1664 * it wasn't sent earlier due to LDLM_FL_AST_SENT bit set.
1666 ldlm_reprocess_all(lock->l_resource, bits);
1669 dlm_rep->lock_handle = lock->l_remote_handle;
1670 ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1671 &dlm_rep->lock_desc.l_policy_data);
1675 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d", rc);
1676 LDLM_LOCK_PUT(lock);
1677 req->rq_status = rc;
1682 * Cancel all the locks whose handles are packed into ldlm_request
1684 * Called by server code expecting such combined cancel activity
1687 int ldlm_request_cancel(struct ptlrpc_request *req,
1688 const struct ldlm_request *dlm_req,
1689 int first, enum lustre_at_flags flags)
1691 struct ldlm_resource *res, *pres = NULL;
1692 struct ldlm_lock *lock;
1693 int i, count, done = 0;
1698 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
1699 if (size <= offsetof(struct ldlm_request, lock_handle) ||
1700 (size - offsetof(struct ldlm_request, lock_handle)) /
1701 sizeof(struct lustre_handle) < dlm_req->lock_count)
1704 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1708 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1712 * There is no lock on the server at the replay time,
1713 * skip lock cancelling to make replay tests to pass.
1715 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1718 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, starting at %d",
1721 for (i = first; i < count; i++) {
1722 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1724 /* below message checked in replay-single.sh test_36 */
1725 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (cookie %llu)",
1726 dlm_req->lock_handle[i].cookie);
1730 res = lock->l_resource;
1734 * This code is an optimization to only attempt lock
1735 * granting on the resource (that could be CPU-expensive)
1736 * after we are done cancelling lock in that resource.
1740 ldlm_reprocess_all(pres, 0);
1741 LDLM_RESOURCE_DELREF(pres);
1742 ldlm_resource_putref(pres);
1745 ldlm_resource_getref(res);
1746 LDLM_RESOURCE_ADDREF(res);
1748 if (!ldlm_is_discard_data(lock))
1749 ldlm_lvbo_update(res, lock,
1755 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1756 lock->l_blast_sent != 0) {
1757 timeout_t delay = 0;
1759 if (ktime_get_seconds() > lock->l_blast_sent)
1760 delay = ktime_get_seconds() -
1763 "server cancels blocked lock after %ds",
1765 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1767 ldlm_lock_cancel(lock);
1768 LDLM_LOCK_PUT(lock);
1771 ldlm_reprocess_all(pres, 0);
1772 LDLM_RESOURCE_DELREF(pres);
1773 ldlm_resource_putref(pres);
1775 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1778 EXPORT_SYMBOL(ldlm_request_cancel);
1781 * Main LDLM entry point for server code to cancel locks.
1783 * Typically gets called from service handler on LDLM_CANCEL opc.
1785 int ldlm_handle_cancel(struct ptlrpc_request *req)
1787 struct ldlm_request *dlm_req;
1792 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1793 if (dlm_req == NULL) {
1794 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1798 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) <
1799 offsetof(struct ldlm_request, lock_handle[1]))
1802 if (req->rq_export && req->rq_export->exp_nid_stats &&
1803 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1804 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1805 LDLM_CANCEL - LDLM_FIRST_OPC);
1807 rc = req_capsule_server_pack(&req->rq_pill);
1811 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1812 req->rq_status = LUSTRE_ESTALE;
1814 RETURN(ptlrpc_reply(req));
1816 #endif /* HAVE_SERVER_SUPPORT */
1819 * Server may pass additional information about blocking lock.
1820 * For IBITS locks it is conflicting bits which can be used for
1821 * lock convert instead of cancel.
1823 void ldlm_bl_desc2lock(const struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1825 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
1827 check_res_locked(lock->l_resource);
1828 if (ns_is_client(ns) && ld &&
1829 (lock->l_resource->lr_type == LDLM_IBITS)) {
1831 * Lock description contains policy of blocking lock,
1832 * and its cancel_bits is used to pass conflicting bits.
1833 * NOTE: ld can be NULL or can be not NULL but zeroed if
1834 * passed from ldlm_bl_thread_blwi(), check below used bits
1835 * in ld to make sure it is valid description.
1837 * If server may replace lock resource keeping the same cookie,
1838 * never use cancel bits from different resource, full cancel
1841 if (ld->l_policy_data.l_inodebits.cancel_bits &&
1842 ldlm_res_eq(&ld->l_resource.lr_name,
1843 &lock->l_resource->lr_name) &&
1844 !(ldlm_is_cbpending(lock) &&
1845 lock->l_policy_data.l_inodebits.cancel_bits == 0)) {
1846 /* always combine conflicting ibits */
1847 lock->l_policy_data.l_inodebits.cancel_bits |=
1848 ld->l_policy_data.l_inodebits.cancel_bits;
1850 /* If cancel_bits are not obtained or
1851 * if the lock is already CBPENDING and
1852 * has no cancel_bits set
1853 * - the full lock is to be cancelled
1855 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1861 * Callback handler for receiving incoming blocking ASTs.
1863 * This can only happen on client side.
1865 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1866 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1872 LDLM_DEBUG(lock, "client blocking AST callback handler");
1874 lock_res_and_lock(lock);
1876 /* get extra information from desc if any */
1877 ldlm_bl_desc2lock(ld, lock);
1878 ldlm_set_cbpending(lock);
1880 do_ast = (!lock->l_readers && !lock->l_writers);
1881 unlock_res_and_lock(lock);
1885 "Lock %p already unused, calling callback (%p)\n",
1886 lock, lock->l_blocking_ast);
1887 if (lock->l_blocking_ast != NULL)
1888 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1892 "Lock %p is referenced, will be cancelled later\n",
1896 LDLM_DEBUG(lock, "client blocking callback handler END");
1897 LDLM_LOCK_RELEASE(lock);
1901 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1903 if (req->rq_no_reply)
1906 req->rq_status = rc;
1907 if (!req->rq_packed_final) {
1908 rc = lustre_pack_reply(req, 1, NULL, NULL);
1912 return ptlrpc_reply(req);
1916 * Callback handler for receiving incoming completion ASTs.
1918 * This only can happen on client side.
1920 static int ldlm_handle_cp_callback(struct ptlrpc_request *req,
1921 struct ldlm_namespace *ns,
1922 struct ldlm_request *dlm_req,
1923 struct ldlm_lock *lock)
1925 LIST_HEAD(ast_list);
1931 LDLM_DEBUG(lock, "client completion callback handler START");
1933 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1934 long to = cfs_time_seconds(1);
1936 ldlm_callback_reply(req, 0);
1939 to = schedule_timeout_interruptible(to);
1940 if (ldlm_is_granted(lock) ||
1941 ldlm_is_destroyed(lock))
1946 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1948 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1949 GOTO(out, rc = lvb_len);
1950 } else if (lvb_len > 0) {
1951 if (lock->l_lvb_len > 0) {
1952 /* for extent lock, lvb contains ost_lvb{}. */
1953 LASSERT(lock->l_lvb_data != NULL);
1955 if (unlikely(lock->l_lvb_len < lvb_len)) {
1957 "Replied LVB is larger than expectation, expected = %d, replied = %d",
1958 lock->l_lvb_len, lvb_len);
1959 GOTO(out, rc = -EINVAL);
1964 lock_res_and_lock(lock);
1966 if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
1967 &lock->l_resource->lr_name)) {
1968 ldlm_resource_unlink_lock(lock);
1969 unlock_res_and_lock(lock);
1970 rc = ldlm_lock_change_resource(ns, lock,
1971 &dlm_req->lock_desc.l_resource.lr_name);
1973 LDLM_ERROR(lock, "Failed to allocate resource");
1976 LDLM_DEBUG(lock, "completion AST, new resource");
1977 lock_res_and_lock(lock);
1980 if (ldlm_is_failed(lock)) {
1981 unlock_res_and_lock(lock);
1982 LDLM_LOCK_RELEASE(lock);
1986 if (ldlm_is_destroyed(lock) ||
1987 ldlm_is_granted(lock)) {
1988 /* b=11300: the lock has already been granted */
1989 unlock_res_and_lock(lock);
1990 LDLM_DEBUG(lock, "Double grant race happened");
1995 * If we receive the completion AST before the actual enqueue returned,
1996 * then we might need to switch lock modes, resources, or extents.
1998 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1999 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
2000 LDLM_DEBUG(lock, "completion AST, new lock mode");
2003 if (lock->l_resource->lr_type != LDLM_PLAIN) {
2004 ldlm_convert_policy_to_local(req->rq_export,
2005 dlm_req->lock_desc.l_resource.lr_type,
2006 &dlm_req->lock_desc.l_policy_data,
2007 &lock->l_policy_data);
2008 LDLM_DEBUG(lock, "completion AST, new policy data");
2011 ldlm_resource_unlink_lock(lock);
2013 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
2015 * BL_AST locks are not needed in LRU.
2016 * Let ldlm_cancel_lru() be fast.
2018 ldlm_lock_remove_from_lru(lock);
2019 ldlm_bl_desc2lock(&dlm_req->lock_desc, lock);
2020 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
2021 LDLM_DEBUG(lock, "completion AST includes blocking AST");
2024 if (lock->l_lvb_len > 0) {
2025 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
2026 lock->l_lvb_data, lvb_len);
2028 unlock_res_and_lock(lock);
2033 ldlm_grant_lock(lock, &ast_list);
2034 unlock_res_and_lock(lock);
2036 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
2039 * Let Enqueue to call osc_lock_upcall() and initialize
2042 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
2044 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
2046 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
2052 lock_res_and_lock(lock);
2053 ldlm_set_failed(lock);
2054 unlock_res_and_lock(lock);
2055 wake_up(&lock->l_waitq);
2057 LDLM_LOCK_RELEASE(lock);
2063 * Callback handler for receiving incoming glimpse ASTs.
2065 * This only can happen on client side. After handling the glimpse AST
2066 * we also consider dropping the lock here if it is unused locally for a
2069 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
2070 struct ldlm_namespace *ns,
2071 struct ldlm_request *dlm_req,
2072 struct ldlm_lock *lock)
2074 struct ldlm_lock_desc *ld = &dlm_req->lock_desc;
2079 LDLM_DEBUG(lock, "client glimpse AST callback handler");
2081 if (lock->l_glimpse_ast != NULL)
2082 rc = lock->l_glimpse_ast(lock, req);
2084 if (req->rq_repmsg != NULL) {
2087 req->rq_status = rc;
2091 lock_res_and_lock(lock);
2092 if (lock->l_granted_mode == LCK_PW &&
2093 !lock->l_readers && !lock->l_writers &&
2094 ktime_after(ktime_get(),
2095 ktime_add(lock->l_last_used, ns->ns_dirty_age_limit))) {
2096 unlock_res_and_lock(lock);
2098 /* For MDS glimpse it is always DOM lock, set corresponding
2099 * cancel_bits to perform lock convert if needed
2101 if (lock->l_resource->lr_type == LDLM_IBITS)
2102 ld->l_policy_data.l_inodebits.cancel_bits =
2104 if (ldlm_bl_to_thread_lock(ns, ld, lock))
2105 ldlm_handle_bl_callback(ns, ld, lock);
2110 unlock_res_and_lock(lock);
2111 LDLM_LOCK_RELEASE(lock);
2115 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
2116 enum ldlm_cancel_flags cancel_flags)
2118 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2122 spin_lock(&blp->blp_lock);
2123 if (blwi->blwi_lock &&
2124 ldlm_is_discard_data(blwi->blwi_lock)) {
2125 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
2126 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
2128 /* other blocking callbacks are added to the regular list */
2129 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
2131 spin_unlock(&blp->blp_lock);
2133 wake_up(&blp->blp_waitq);
2136 * can not check blwi->blwi_flags as blwi could be already freed in
2139 if (!(cancel_flags & LCF_ASYNC))
2140 wait_for_completion(&blwi->blwi_comp);
2145 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2146 struct ldlm_namespace *ns,
2147 struct ldlm_lock_desc *ld,
2148 struct list_head *cancels, int count,
2149 struct ldlm_lock *lock,
2150 enum ldlm_cancel_flags cancel_flags)
2152 init_completion(&blwi->blwi_comp);
2153 INIT_LIST_HEAD(&blwi->blwi_head);
2155 if (current->flags & PF_MEMALLOC)
2156 blwi->blwi_mem_pressure = 1;
2159 blwi->blwi_flags = cancel_flags;
2161 blwi->blwi_ld = *ld;
2163 list_splice_init(cancels, &blwi->blwi_head);
2164 blwi->blwi_count = count;
2166 blwi->blwi_lock = lock;
2171 * Queues a list of locks \a cancels containing \a count locks
2172 * for later processing by a blocking thread. If \a count is zero,
2173 * then the lock referenced as \a lock is queued instead.
2175 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2176 * If list addition fails an error is returned and caller is supposed to
2177 * call ->l_blocking_ast itself.
2179 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2180 struct ldlm_lock_desc *ld,
2181 struct ldlm_lock *lock,
2182 struct list_head *cancels, int count,
2183 enum ldlm_cancel_flags cancel_flags)
2187 if (cancels && count == 0)
2190 if (cancel_flags & LCF_ASYNC) {
2191 struct ldlm_bl_work_item *blwi;
2193 OBD_ALLOC(blwi, sizeof(*blwi));
2196 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2198 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2201 * if it is synchronous call do minimum mem alloc, as it could
2202 * be triggered from kernel shrinker
2204 struct ldlm_bl_work_item blwi;
2206 memset(&blwi, 0, sizeof(blwi));
2207 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2208 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2213 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2214 struct ldlm_lock *lock)
2216 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2219 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2220 struct list_head *cancels, int count,
2221 enum ldlm_cancel_flags cancel_flags)
2223 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2226 int ldlm_bl_to_thread_ns(struct ldlm_namespace *ns)
2228 return ldlm_bl_to_thread(ns, NULL, NULL, NULL, 0, LCF_ASYNC);
2231 int ldlm_bl_thread_wakeup(void)
2233 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2237 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2238 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2240 struct obd_device *obd = req->rq_export->exp_obd;
2248 DEBUG_REQ(D_HSM, req, "%s: handle setinfo", obd->obd_name);
2250 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2252 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2254 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2257 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2259 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2261 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2264 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2267 /* We are responsible for swabbing contents of val */
2269 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2270 /* Pass it on to mdc (the "export" in this case) */
2271 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2273 sizeof(KEY_HSM_COPYTOOL_SEND),
2274 KEY_HSM_COPYTOOL_SEND,
2277 DEBUG_REQ(D_WARNING, req, "ignoring unknown key '%s'", key);
2282 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2283 const char *msg, int rc,
2284 const struct lustre_handle *handle)
2286 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2287 "%s, NID=%s lock=%#llx: rc = %d",
2288 msg, libcfs_id2str(req->rq_peer),
2289 handle ? handle->cookie : 0, rc);
2290 if (req->rq_no_reply)
2291 CWARN("No reply was sent, maybe cause b=21636.\n");
2293 CWARN("Send reply failed, maybe cause b=21636.\n");
2296 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2297 static int ldlm_callback_handler(struct ptlrpc_request *req)
2299 struct ldlm_namespace *ns;
2300 struct ldlm_request *dlm_req;
2301 struct ldlm_lock *lock;
2307 * Requests arrive in sender's byte order. The ptlrpc service
2308 * handler has already checked and, if necessary, byte-swapped the
2309 * incoming request message body, but I am responsible for the
2313 /* do nothing for sec context finalize */
2314 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2317 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2319 if (req->rq_export == NULL) {
2320 rc = ldlm_callback_reply(req, -ENOTCONN);
2321 ldlm_callback_errmsg(req, "Operate on unconnected server",
2326 LASSERT(req->rq_export != NULL);
2327 LASSERT(req->rq_export->exp_obd != NULL);
2329 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2330 case LDLM_BL_CALLBACK:
2331 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2333 ldlm_callback_reply(req, -(int)cfs_fail_err);
2337 case LDLM_CP_CALLBACK:
2338 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2341 case LDLM_GL_CALLBACK:
2342 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2346 rc = ldlm_handle_setinfo(req);
2347 ldlm_callback_reply(req, rc);
2350 CERROR("unknown opcode %u\n",
2351 lustre_msg_get_opc(req->rq_reqmsg));
2352 ldlm_callback_reply(req, -EPROTO);
2356 ns = req->rq_export->exp_obd->obd_namespace;
2357 LASSERT(ns != NULL);
2359 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2361 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2362 if (dlm_req == NULL) {
2363 rc = ldlm_callback_reply(req, -EPROTO);
2364 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2370 * Force a known safe race, send a cancel to the server for a lock
2371 * which the server has already started a blocking callback on.
2373 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2374 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2375 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2377 CERROR("ldlm_cli_cancel: %d\n", rc);
2380 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2383 "callback on lock %#llx - lock disappeared\n",
2384 dlm_req->lock_handle[0].cookie);
2385 rc = ldlm_callback_reply(req, -EINVAL);
2386 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2387 &dlm_req->lock_handle[0]);
2391 if (ldlm_is_fail_loc(lock) &&
2392 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2393 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2395 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2396 lock_res_and_lock(lock);
2397 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2399 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2401 * If somebody cancels lock and cache is already dropped,
2402 * or lock is failed before cp_ast received on client,
2403 * we can tell the server we have no lock. Otherwise, we
2404 * should send cancel after dropping the cache.
2406 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2407 ldlm_is_failed(lock)) {
2409 "callback on lock %llx - lock disappeared",
2410 dlm_req->lock_handle[0].cookie);
2411 unlock_res_and_lock(lock);
2412 LDLM_LOCK_RELEASE(lock);
2413 rc = ldlm_callback_reply(req, -EINVAL);
2414 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2415 &dlm_req->lock_handle[0]);
2419 * BL_AST locks are not needed in LRU.
2420 * Let ldlm_cancel_lru() be fast.
2422 ldlm_lock_remove_from_lru(lock);
2423 ldlm_set_bl_ast(lock);
2425 if (lock->l_remote_handle.cookie == 0)
2426 lock->l_remote_handle = dlm_req->lock_handle[1];
2427 unlock_res_and_lock(lock);
2430 * We want the ost thread to get this reply so that it can respond
2431 * to ost requests (write cache writeback) that might be triggered
2434 * But we'd also like to be able to indicate in the reply that we're
2435 * cancelling right now, because it's unused, or have an intent result
2436 * in the reply, so we might have to push the responsibility for sending
2437 * the reply down into the AST handlers, alas.
2440 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2441 case LDLM_BL_CALLBACK:
2442 CDEBUG(D_INODE, "blocking ast\n");
2443 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2444 if (!ldlm_is_cancel_on_block(lock)) {
2445 rc = ldlm_callback_reply(req, 0);
2446 if (req->rq_no_reply || rc)
2447 ldlm_callback_errmsg(req, "Normal process", rc,
2448 &dlm_req->lock_handle[0]);
2450 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2451 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2453 case LDLM_CP_CALLBACK:
2454 CDEBUG(D_INODE, "completion ast\n");
2455 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2456 rc = ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2457 if (!OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE))
2458 ldlm_callback_reply(req, rc);
2460 case LDLM_GL_CALLBACK:
2461 CDEBUG(D_INODE, "glimpse ast\n");
2462 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2463 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2466 LBUG(); /* checked above */
2472 #ifdef HAVE_SERVER_SUPPORT
2474 * Main handler for canceld thread.
2476 * Separated into its own thread to avoid deadlocks.
2478 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2485 * Requests arrive in sender's byte order. The ptlrpc service
2486 * handler has already checked and, if necessary, byte-swapped the
2487 * incoming request message body, but I am responsible for the
2491 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2493 if (req->rq_export == NULL) {
2494 struct ldlm_request *dlm_req;
2496 CERROR("%s from %s arrived at %llu with bad export cookie %llu\n",
2497 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2498 libcfs_nid2str(req->rq_peer.nid),
2499 (unsigned long long)req->rq_arrival_time.tv_sec,
2500 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2502 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2503 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2504 dlm_req = req_capsule_client_get(&req->rq_pill,
2506 if (dlm_req != NULL)
2507 ldlm_lock_dump_handle(D_ERROR,
2508 &dlm_req->lock_handle[0]);
2510 ldlm_callback_reply(req, -ENOTCONN);
2514 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2515 /* XXX FIXME move this back to mds/handler.c, b=249 */
2517 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2518 CDEBUG(D_INODE, "cancel\n");
2519 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2520 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2521 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2523 rc = ldlm_handle_cancel(req);
2527 struct ldlm_request *dlm_req;
2529 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2530 CDEBUG(D_INODE, "convert\n");
2532 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2533 if (dlm_req == NULL) {
2534 CDEBUG(D_INFO, "bad request buffer for cancel\n");
2535 rc = ldlm_callback_reply(req, -EPROTO);
2537 req->rq_status = ldlm_handle_convert0(req, dlm_req);
2538 rc = ptlrpc_reply(req);
2543 CERROR("invalid opcode %d\n",
2544 lustre_msg_get_opc(req->rq_reqmsg));
2545 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2546 rc = ldlm_callback_reply(req, -EINVAL);
2552 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2553 struct ldlm_lock *lock)
2555 struct ldlm_request *dlm_req;
2556 struct lustre_handle lockh;
2562 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2563 if (dlm_req == NULL)
2566 ldlm_lock2handle(lock, &lockh);
2567 for (i = 0; i < dlm_req->lock_count; i++) {
2568 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2570 DEBUG_REQ(D_RPCTRACE, req,
2571 "Prio raised by lock %#llx", lockh.cookie);
2580 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2582 struct ldlm_request *dlm_req;
2589 /* no prolong in recovery */
2590 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2593 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2594 if (dlm_req == NULL)
2597 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
2598 if (size <= offsetof(struct ldlm_request, lock_handle) ||
2599 (size - offsetof(struct ldlm_request, lock_handle)) /
2600 sizeof(struct lustre_handle) < dlm_req->lock_count)
2603 for (i = 0; i < dlm_req->lock_count; i++) {
2604 struct ldlm_lock *lock;
2606 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2610 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2612 LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2613 LDLM_LOCK_PUT(lock);
2622 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2623 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2624 .hpreq_check = ldlm_cancel_hpreq_check,
2628 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2632 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2634 if (req->rq_export == NULL)
2637 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2638 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2639 req->rq_ops = &ldlm_cancel_hpreq_ops;
2640 } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2641 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2642 req->rq_ops = &ldlm_cancel_hpreq_ops;
2647 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2648 struct hlist_node *hnode, void *data)
2651 struct list_head *rpc_list = data;
2652 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2654 lock_res_and_lock(lock);
2656 if (!ldlm_is_granted(lock)) {
2657 unlock_res_and_lock(lock);
2661 LASSERT(lock->l_resource);
2662 if (lock->l_resource->lr_type != LDLM_IBITS &&
2663 lock->l_resource->lr_type != LDLM_PLAIN) {
2664 unlock_res_and_lock(lock);
2668 if (ldlm_is_ast_sent(lock)) {
2669 unlock_res_and_lock(lock);
2673 LASSERT(lock->l_blocking_ast);
2674 LASSERT(!lock->l_blocking_lock);
2676 ldlm_set_ast_sent(lock);
2677 if (lock->l_export && lock->l_export->exp_lock_hash) {
2679 * NB: it's safe to call cfs_hash_del() even lock isn't
2683 * In the function below, .hs_keycmp resolves to
2684 * ldlm_export_lock_keycmp()
2686 /* coverity[overrun-buffer-val] */
2687 cfs_hash_del(lock->l_export->exp_lock_hash,
2688 &lock->l_remote_handle, &lock->l_exp_hash);
2691 list_add_tail(&lock->l_rk_ast, rpc_list);
2692 LDLM_LOCK_GET(lock);
2694 unlock_res_and_lock(lock);
2698 void ldlm_revoke_export_locks(struct obd_export *exp)
2701 LIST_HEAD(rpc_list);
2704 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2705 ldlm_revoke_lock_cb, &rpc_list, 0);
2706 rc = ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2707 LDLM_WORK_REVOKE_AST);
2709 if (rc == -ERESTART)
2710 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2714 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2715 #endif /* HAVE_SERVER_SUPPORT */
2717 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2718 struct ldlm_bl_work_item **p_blwi,
2719 struct obd_export **p_exp)
2721 struct ldlm_bl_work_item *blwi = NULL;
2722 static unsigned int num_bl;
2723 static unsigned int num_stale;
2724 int num_th = atomic_read(&blp->blp_num_threads);
2726 *p_exp = obd_stale_export_get();
2728 spin_lock(&blp->blp_lock);
2729 if (*p_exp != NULL) {
2730 if (num_th == 1 || ++num_stale < num_th) {
2731 spin_unlock(&blp->blp_lock);
2737 /* process a request from the blp_list at least every blp_num_threads */
2738 if (!list_empty(&blp->blp_list) &&
2739 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2740 blwi = list_entry(blp->blp_list.next,
2741 struct ldlm_bl_work_item, blwi_entry);
2743 if (!list_empty(&blp->blp_prio_list))
2744 blwi = list_entry(blp->blp_prio_list.next,
2745 struct ldlm_bl_work_item,
2749 if (++num_bl >= num_th)
2751 list_del(&blwi->blwi_entry);
2753 spin_unlock(&blp->blp_lock);
2756 if (*p_exp != NULL && *p_blwi != NULL) {
2757 obd_stale_export_put(*p_exp);
2761 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2764 /* This only contains temporary data until the thread starts */
2765 struct ldlm_bl_thread_data {
2766 struct ldlm_bl_pool *bltd_blp;
2767 struct completion bltd_comp;
2771 static int ldlm_bl_thread_main(void *arg);
2773 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2775 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2776 struct task_struct *task;
2778 init_completion(&bltd.bltd_comp);
2780 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2781 if (bltd.bltd_num >= blp->blp_max_threads) {
2782 atomic_dec(&blp->blp_num_threads);
2786 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2788 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2789 atomic_dec(&blp->blp_num_threads);
2793 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2796 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2797 bltd.bltd_num, PTR_ERR(task));
2798 atomic_dec(&blp->blp_num_threads);
2799 return PTR_ERR(task);
2801 wait_for_completion(&bltd.bltd_comp);
2806 /* Not fatal if racy and have a few too many threads */
2807 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2808 struct ldlm_bl_work_item *blwi)
2810 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2813 if (atomic_read(&blp->blp_busy_threads) <
2814 atomic_read(&blp->blp_num_threads))
2817 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2818 blwi->blwi_mem_pressure))
2824 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2825 struct ldlm_bl_work_item *blwi)
2827 /* '1' for consistency with code that checks !mpflag to restore */
2828 unsigned int mpflags = 1;
2832 if (blwi->blwi_ns == NULL)
2833 /* added by ldlm_cleanup() */
2834 RETURN(LDLM_ITER_STOP);
2836 if (blwi->blwi_mem_pressure)
2837 mpflags = memalloc_noreclaim_save();
2839 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2841 if (blwi->blwi_count) {
2844 * The special case when we cancel locks in lru
2845 * asynchronously, we pass the list of locks here.
2846 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2847 * canceled locally yet.
2849 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2852 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2854 } else if (blwi->blwi_lock) {
2855 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2858 ldlm_pool_recalc(&blwi->blwi_ns->ns_pool, true);
2859 spin_lock(&blwi->blwi_ns->ns_lock);
2860 blwi->blwi_ns->ns_rpc_recalc = 0;
2861 spin_unlock(&blwi->blwi_ns->ns_lock);
2862 ldlm_namespace_put(blwi->blwi_ns);
2865 if (blwi->blwi_mem_pressure)
2866 memalloc_noreclaim_restore(mpflags);
2868 if (blwi->blwi_flags & LCF_ASYNC)
2869 OBD_FREE(blwi, sizeof(*blwi));
2871 complete(&blwi->blwi_comp);
2877 * Cancel stale locks on export. Cancel blocked locks first.
2878 * If the given export has blocked locks, the next in the list may have
2879 * them too, thus cancel not blocked locks only if the current export has
2882 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2883 struct obd_export *exp)
2889 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2891 num = ldlm_export_cancel_blocked_locks(exp);
2893 ldlm_export_cancel_locks(exp);
2895 obd_stale_export_put(exp);
2902 * Main blocking requests processing thread.
2904 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2905 * This thread in the end ends up doing actual call to ->l_blocking_ast
2908 static int ldlm_bl_thread_main(void *arg)
2911 struct ldlm_bl_pool *blp;
2912 struct ldlm_bl_thread_data *bltd = arg;
2920 rc = lu_env_init(env, LCT_DT_THREAD);
2923 rc = lu_env_add(env);
2925 GOTO(out_env_fini, rc);
2927 blp = bltd->bltd_blp;
2929 complete(&bltd->bltd_comp);
2930 /* cannot use bltd after this, it is only on caller's stack */
2933 struct ldlm_bl_work_item *blwi = NULL;
2934 struct obd_export *exp = NULL;
2937 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2940 wait_event_idle_exclusive(blp->blp_waitq,
2941 ldlm_bl_get_work(blp, &blwi,
2943 atomic_inc(&blp->blp_busy_threads);
2945 if (ldlm_bl_thread_need_create(blp, blwi))
2946 /* discard the return value, we tried */
2947 ldlm_bl_thread_start(blp, true);
2950 rc = ldlm_bl_thread_exports(blp, exp);
2952 rc = ldlm_bl_thread_blwi(blp, blwi);
2954 atomic_dec(&blp->blp_busy_threads);
2956 if (rc == LDLM_ITER_STOP)
2960 * If there are many namespaces, we will not sleep waiting for
2961 * work, and must do a cond_resched to avoid holding the CPU
2967 atomic_dec(&blp->blp_num_threads);
2968 complete(&blp->blp_comp);
2979 static int ldlm_setup(void);
2980 static int ldlm_cleanup(void);
2982 int ldlm_get_ref(void)
2987 mutex_lock(&ldlm_ref_mutex);
2988 if (++ldlm_refcount == 1) {
2993 mutex_unlock(&ldlm_ref_mutex);
2998 void ldlm_put_ref(void)
3001 mutex_lock(&ldlm_ref_mutex);
3002 if (ldlm_refcount == 1) {
3003 int rc = ldlm_cleanup();
3006 CERROR("ldlm_cleanup failed: %d\n", rc);
3012 mutex_unlock(&ldlm_ref_mutex);
3018 * Export handle<->lock hash operations.
3021 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned int mask)
3023 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
3027 ldlm_export_lock_key(struct hlist_node *hnode)
3029 struct ldlm_lock *lock;
3031 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3032 return &lock->l_remote_handle;
3036 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
3038 struct ldlm_lock *lock;
3040 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3041 lock->l_remote_handle = *(struct lustre_handle *)key;
3045 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
3047 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
3051 ldlm_export_lock_object(struct hlist_node *hnode)
3053 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3057 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
3059 struct ldlm_lock *lock;
3061 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3062 LDLM_LOCK_GET(lock);
3066 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
3068 struct ldlm_lock *lock;
3070 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3071 LDLM_LOCK_RELEASE(lock);
3074 static struct cfs_hash_ops ldlm_export_lock_ops = {
3075 .hs_hash = ldlm_export_lock_hash,
3076 .hs_key = ldlm_export_lock_key,
3077 .hs_keycmp = ldlm_export_lock_keycmp,
3078 .hs_keycpy = ldlm_export_lock_keycpy,
3079 .hs_object = ldlm_export_lock_object,
3080 .hs_get = ldlm_export_lock_get,
3081 .hs_put = ldlm_export_lock_put,
3082 .hs_put_locked = ldlm_export_lock_put,
3085 int ldlm_init_export(struct obd_export *exp)
3091 exp->exp_lock_hash =
3092 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
3093 HASH_EXP_LOCK_CUR_BITS,
3094 HASH_EXP_LOCK_MAX_BITS,
3095 HASH_EXP_LOCK_BKT_BITS, 0,
3096 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
3097 &ldlm_export_lock_ops,
3098 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
3099 CFS_HASH_NBLK_CHANGE);
3101 if (!exp->exp_lock_hash)
3104 rc = ldlm_init_flock_export(exp);
3110 ldlm_destroy_export(exp);
3113 EXPORT_SYMBOL(ldlm_init_export);
3115 void ldlm_destroy_export(struct obd_export *exp)
3118 cfs_hash_putref(exp->exp_lock_hash);
3119 exp->exp_lock_hash = NULL;
3121 ldlm_destroy_flock_export(exp);
3124 EXPORT_SYMBOL(ldlm_destroy_export);
3126 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
3127 struct attribute *attr,
3130 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
3133 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
3134 struct attribute *attr,
3141 rc = kstrtoul(buffer, 10, &val);
3145 ldlm_cancel_unused_locks_before_replay = val;
3149 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
3151 static struct attribute *ldlm_attrs[] = {
3152 &lustre_attr_cancel_unused_locks_before_replay.attr,
3156 static struct attribute_group ldlm_attr_group = {
3157 .attrs = ldlm_attrs,
3160 static int ldlm_setup(void)
3162 static struct ptlrpc_service_conf conf;
3163 struct ldlm_bl_pool *blp = NULL;
3164 #ifdef HAVE_SERVER_SUPPORT
3165 struct task_struct *task;
3166 #endif /* HAVE_SERVER_SUPPORT */
3172 if (ldlm_state != NULL)
3175 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
3176 if (ldlm_state == NULL)
3179 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
3183 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
3187 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
3191 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
3195 rc = ldlm_debugfs_setup();
3199 memset(&conf, 0, sizeof(conf));
3200 conf = (typeof(conf)) {
3201 .psc_name = "ldlm_cbd",
3202 .psc_watchdog_factor = 2,
3204 .bc_nbufs = LDLM_CLIENT_NBUFS,
3205 .bc_buf_size = LDLM_BUFSIZE,
3206 .bc_req_max_size = LDLM_MAXREQSIZE,
3207 .bc_rep_max_size = LDLM_MAXREPSIZE,
3208 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
3209 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
3212 .tc_thr_name = "ldlm_cb",
3213 .tc_thr_factor = LDLM_THR_FACTOR,
3214 .tc_nthrs_init = LDLM_NTHRS_INIT,
3215 .tc_nthrs_base = LDLM_NTHRS_BASE,
3216 .tc_nthrs_max = LDLM_NTHRS_MAX,
3217 .tc_nthrs_user = ldlm_num_threads,
3218 .tc_cpu_bind = ldlm_cpu_bind,
3219 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
3222 .cc_pattern = ldlm_cpts,
3223 .cc_affinity = true,
3226 .so_req_handler = ldlm_callback_handler,
3229 ldlm_state->ldlm_cb_service = \
3230 ptlrpc_register_service(&conf, ldlm_svc_kset,
3231 ldlm_svc_debugfs_dir);
3232 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3233 CERROR("failed to start service\n");
3234 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3235 ldlm_state->ldlm_cb_service = NULL;
3239 #ifdef HAVE_SERVER_SUPPORT
3240 memset(&conf, 0, sizeof(conf));
3241 conf = (typeof(conf)) {
3242 .psc_name = "ldlm_canceld",
3243 .psc_watchdog_factor = 6,
3245 .bc_nbufs = LDLM_SERVER_NBUFS,
3246 .bc_buf_size = LDLM_BUFSIZE,
3247 .bc_req_max_size = LDLM_MAXREQSIZE,
3248 .bc_rep_max_size = LDLM_MAXREPSIZE,
3249 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3250 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3254 .tc_thr_name = "ldlm_cn",
3255 .tc_thr_factor = LDLM_THR_FACTOR,
3256 .tc_nthrs_init = LDLM_NTHRS_INIT,
3257 .tc_nthrs_base = LDLM_NTHRS_BASE,
3258 .tc_nthrs_max = LDLM_NTHRS_MAX,
3259 .tc_nthrs_user = ldlm_num_threads,
3260 .tc_cpu_bind = ldlm_cpu_bind,
3261 .tc_ctx_tags = LCT_MD_THREAD | \
3266 .cc_pattern = ldlm_cpts,
3267 .cc_affinity = true,
3270 .so_req_handler = ldlm_cancel_handler,
3271 .so_hpreq_handler = ldlm_hpreq_handler,
3274 ldlm_state->ldlm_cancel_service = \
3275 ptlrpc_register_service(&conf, ldlm_svc_kset,
3276 ldlm_svc_debugfs_dir);
3277 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3278 CERROR("failed to start service\n");
3279 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3280 ldlm_state->ldlm_cancel_service = NULL;
3283 #endif /* HAVE_SERVER_SUPPORT */
3285 OBD_ALLOC(blp, sizeof(*blp));
3287 GOTO(out, rc = -ENOMEM);
3288 ldlm_state->ldlm_bl_pool = blp;
3290 spin_lock_init(&blp->blp_lock);
3291 INIT_LIST_HEAD(&blp->blp_list);
3292 INIT_LIST_HEAD(&blp->blp_prio_list);
3293 init_waitqueue_head(&blp->blp_waitq);
3294 atomic_set(&blp->blp_num_threads, 0);
3295 atomic_set(&blp->blp_busy_threads, 0);
3297 if (ldlm_num_threads == 0) {
3298 blp->blp_min_threads = LDLM_NTHRS_INIT;
3299 blp->blp_max_threads = LDLM_NTHRS_MAX;
3301 blp->blp_min_threads = blp->blp_max_threads = \
3302 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3306 for (i = 0; i < blp->blp_min_threads; i++) {
3307 rc = ldlm_bl_thread_start(blp, false);
3312 #ifdef HAVE_SERVER_SUPPORT
3313 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3316 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3320 wait_event(expired_lock_wait_queue,
3321 expired_lock_thread_state == ELT_READY);
3322 #endif /* HAVE_SERVER_SUPPORT */
3324 rc = ldlm_pools_init();
3326 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3330 rc = ldlm_reclaim_setup();
3332 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3342 static int ldlm_cleanup(void)
3346 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3347 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3348 CERROR("ldlm still has namespaces; clean these up first.\n");
3349 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3350 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3354 ldlm_reclaim_cleanup();
3357 if (ldlm_state->ldlm_bl_pool != NULL) {
3358 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3360 while (atomic_read(&blp->blp_num_threads) > 0) {
3361 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3363 init_completion(&blp->blp_comp);
3365 spin_lock(&blp->blp_lock);
3366 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3367 wake_up(&blp->blp_waitq);
3368 spin_unlock(&blp->blp_lock);
3370 wait_for_completion(&blp->blp_comp);
3373 OBD_FREE(blp, sizeof(*blp));
3376 if (ldlm_state->ldlm_cb_service != NULL)
3377 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3378 #ifdef HAVE_SERVER_SUPPORT
3379 if (ldlm_state->ldlm_cancel_service != NULL)
3380 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3384 kset_unregister(ldlm_ns_kset);
3386 kset_unregister(ldlm_svc_kset);
3388 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3389 kobject_put(ldlm_kobj);
3392 ldlm_debugfs_cleanup();
3394 #ifdef HAVE_SERVER_SUPPORT
3395 if (expired_lock_thread_state != ELT_STOPPED) {
3396 expired_lock_thread_state = ELT_TERMINATE;
3397 wake_up(&expired_lock_wait_queue);
3398 wait_event(expired_lock_wait_queue,
3399 expired_lock_thread_state == ELT_STOPPED);
3403 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3411 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3412 sizeof(struct ldlm_resource), 0,
3413 SLAB_HWCACHE_ALIGN, NULL);
3414 if (ldlm_resource_slab == NULL)
3417 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3418 sizeof(struct ldlm_lock), 0,
3419 SLAB_HWCACHE_ALIGN, NULL);
3420 if (ldlm_lock_slab == NULL)
3423 ldlm_interval_slab = kmem_cache_create("interval_node",
3424 sizeof(struct ldlm_interval),
3425 0, SLAB_HWCACHE_ALIGN, NULL);
3426 if (ldlm_interval_slab == NULL)
3429 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3430 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3431 0, SLAB_HWCACHE_ALIGN, NULL);
3432 if (ldlm_interval_tree_slab == NULL)
3435 #ifdef HAVE_SERVER_SUPPORT
3436 ldlm_inodebits_slab = kmem_cache_create("ldlm_ibits_node",
3437 sizeof(struct ldlm_ibits_node),
3438 0, SLAB_HWCACHE_ALIGN, NULL);
3439 if (ldlm_inodebits_slab == NULL)
3440 goto out_interval_tree;
3442 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3443 sizeof(struct ldlm_glimpse_work),
3445 if (ldlm_glimpse_work_kmem == NULL)
3449 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3450 class_export_dump_hook = ldlm_dump_export_locks;
3453 #ifdef HAVE_SERVER_SUPPORT
3455 kmem_cache_destroy(ldlm_inodebits_slab);
3457 kmem_cache_destroy(ldlm_interval_tree_slab);
3460 kmem_cache_destroy(ldlm_interval_slab);
3462 kmem_cache_destroy(ldlm_lock_slab);
3464 kmem_cache_destroy(ldlm_resource_slab);
3469 void ldlm_exit(void)
3472 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3474 kmem_cache_destroy(ldlm_resource_slab);
3476 * ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3477 * rcu_barrier() to wait all outstanding RCU callbacks to complete,
3478 * so that ldlm_lock_free() get a chance to be called.
3481 kmem_cache_destroy(ldlm_lock_slab);
3482 kmem_cache_destroy(ldlm_interval_slab);
3483 kmem_cache_destroy(ldlm_interval_tree_slab);
3484 #ifdef HAVE_SERVER_SUPPORT
3485 kmem_cache_destroy(ldlm_inodebits_slab);
3486 kmem_cache_destroy(ldlm_glimpse_work_kmem);