X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_lockd.c;h=04e3c8e93a7a2dedc1651d7ce3a5eb30b9f813dc;hp=346a6e7e08ffcd4a0335b18f2f323c0b0227e23a;hb=978428ed047602a311f016fe7f4cf7d21f5599b6;hpb=9fb46705ae86aa2c0ac29427f0ff24f923560eb7 diff --git a/lustre/ldlm/ldlm_lockd.c b/lustre/ldlm/ldlm_lockd.c index 346a6e7..04e3c8e 100644 --- a/lustre/ldlm/ldlm_lockd.c +++ b/lustre/ldlm/ldlm_lockd.c @@ -15,11 +15,7 @@ * * You should have received a copy of the GNU General Public License * version 2 along with this program; If not, see - * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf - * - * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, - * CA 95054 USA or visit www.sun.com if you need additional information or - * have any questions. + * http://www.gnu.org/licenses/gpl-2.0.html * * GPL HEADER END */ @@ -27,7 +23,7 @@ * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011, 2012, Whamcloud, Inc. + * Copyright (c) 2010, 2016, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -41,40 +37,34 @@ #define DEBUG_SUBSYSTEM S_LDLM -#ifdef __KERNEL__ -# include -#else -# include -#endif - +#include +#include +#include +#include #include #include -#include #include "ldlm_internal.h" static int ldlm_num_threads; -CFS_MODULE_PARM(ldlm_num_threads, "i", int, 0444, - "number of DLM service threads to start"); +module_param(ldlm_num_threads, int, 0444); +MODULE_PARM_DESC(ldlm_num_threads, "number of DLM service threads to start"); static char *ldlm_cpts; -CFS_MODULE_PARM(ldlm_cpts, "s", charp, 0444, - "CPU partitions ldlm threads should run on"); +module_param(ldlm_cpts, charp, 0444); +MODULE_PARM_DESC(ldlm_cpts, "CPU partitions ldlm threads should run on"); -extern cfs_mem_cache_t *ldlm_resource_slab; -extern cfs_mem_cache_t *ldlm_lock_slab; -static struct mutex ldlm_ref_mutex; +static DEFINE_MUTEX(ldlm_ref_mutex); static int ldlm_refcount; -struct ldlm_cb_async_args { - struct ldlm_cb_set_arg *ca_set_arg; - struct ldlm_lock *ca_lock; -}; +struct kobject *ldlm_kobj; +struct kset *ldlm_ns_kset; +struct kset *ldlm_svc_kset; /* LDLM state */ static struct ldlm_state *ldlm_state; -inline cfs_time_t round_timeout(cfs_time_t timeout) +static inline cfs_time_t round_timeout(cfs_time_t timeout) { return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1); } @@ -88,59 +78,74 @@ static inline unsigned int ldlm_get_rq_timeout(void) return timeout < 1 ? 1 : timeout; } -#define ELT_STOPPED 0 -#define ELT_READY 1 -#define ELT_TERMINATE 2 - struct ldlm_bl_pool { spinlock_t blp_lock; - /* - * blp_prio_list is used for callbacks that should be handled - * as a priority. It is used for LDLM_FL_DISCARD_DATA requests. - * see bug 13843 - */ - cfs_list_t blp_prio_list; - - /* - * blp_list is used for all other callbacks which are likely - * to take longer to process. - */ - cfs_list_t blp_list; - - cfs_waitq_t blp_waitq; - struct completion blp_comp; - cfs_atomic_t blp_num_threads; - cfs_atomic_t blp_busy_threads; - int blp_min_threads; - int blp_max_threads; + /* + * blp_prio_list is used for callbacks that should be handled + * as a priority. It is used for LDLM_FL_DISCARD_DATA requests. + * see bug 13843 + */ + struct list_head blp_prio_list; + + /* + * blp_list is used for all other callbacks which are likely + * to take longer to process. + */ + struct list_head blp_list; + + wait_queue_head_t blp_waitq; + struct completion blp_comp; + atomic_t blp_num_threads; + atomic_t blp_busy_threads; + int blp_min_threads; + int blp_max_threads; }; struct ldlm_bl_work_item { - cfs_list_t blwi_entry; - struct ldlm_namespace *blwi_ns; - struct ldlm_lock_desc blwi_ld; - struct ldlm_lock *blwi_lock; - cfs_list_t blwi_head; - int blwi_count; - struct completion blwi_comp; - int blwi_mode; - int blwi_mem_pressure; + struct list_head blwi_entry; + struct ldlm_namespace *blwi_ns; + struct ldlm_lock_desc blwi_ld; + struct ldlm_lock *blwi_lock; + struct list_head blwi_head; + int blwi_count; + struct completion blwi_comp; + enum ldlm_cancel_flags blwi_flags; + int blwi_mem_pressure; }; -#if defined(HAVE_SERVER_SUPPORT) && defined(__KERNEL__) +#ifdef HAVE_SERVER_SUPPORT -/* w_l_spinlock protects both waiting_locks_list and expired_lock_thread */ -static spinlock_t waiting_locks_spinlock; /* BH lock (timer) */ -static cfs_list_t waiting_locks_list; -static cfs_timer_t waiting_locks_timer; +/** + * Protects both waiting_locks_list and expired_lock_thread. + */ +static DEFINE_SPINLOCK(waiting_locks_spinlock); /* BH lock (timer) */ + +/** + * List for contended locks. + * + * As soon as a lock is contended, it gets placed on this list and + * expected time to get a response is filled in the lock. A special + * thread walks the list looking for locks that should be released and + * schedules client evictions for those that have not been released in + * time. + * + * All access to it should be under waiting_locks_spinlock. + */ +static LIST_HEAD(waiting_locks_list); +static void waiting_locks_callback(unsigned long unused); +static DEFINE_TIMER(waiting_locks_timer, waiting_locks_callback, 0, 0); + +enum elt_state { + ELT_STOPPED, + ELT_READY, + ELT_TERMINATE, +}; -static struct expired_lock_thread { - cfs_waitq_t elt_waitq; - int elt_state; - int elt_dump; - cfs_list_t elt_expired_locks; -} expired_lock_thread; +static DECLARE_WAIT_QUEUE_HEAD(expired_lock_wait_queue); +static enum elt_state expired_lock_thread_state = ELT_STOPPED; +static int expired_lock_dump; +static LIST_HEAD(expired_lock_list); static inline int have_expired_locks(void) { @@ -148,75 +153,73 @@ static inline int have_expired_locks(void) ENTRY; spin_lock_bh(&waiting_locks_spinlock); - need_to_run = !cfs_list_empty(&expired_lock_thread.elt_expired_locks); + need_to_run = !list_empty(&expired_lock_list); spin_unlock_bh(&waiting_locks_spinlock); RETURN(need_to_run); } +/** + * Check expired lock list for expired locks and time them out. + */ static int expired_lock_main(void *arg) { - cfs_list_t *expired = &expired_lock_thread.elt_expired_locks; - struct l_wait_info lwi = { 0 }; - int do_dump; + struct list_head *expired = &expired_lock_list; + struct l_wait_info lwi = { 0 }; + int do_dump; - ENTRY; - cfs_daemonize("ldlm_elt"); + ENTRY; - expired_lock_thread.elt_state = ELT_READY; - cfs_waitq_signal(&expired_lock_thread.elt_waitq); + expired_lock_thread_state = ELT_READY; + wake_up(&expired_lock_wait_queue); - while (1) { - l_wait_event(expired_lock_thread.elt_waitq, - have_expired_locks() || - expired_lock_thread.elt_state == ELT_TERMINATE, - &lwi); + while (1) { + l_wait_event(expired_lock_wait_queue, + have_expired_locks() || + expired_lock_thread_state == ELT_TERMINATE, + &lwi); spin_lock_bh(&waiting_locks_spinlock); - if (expired_lock_thread.elt_dump) { - struct libcfs_debug_msg_data msgdata = { - .msg_file = __FILE__, - .msg_fn = "waiting_locks_callback", - .msg_line = expired_lock_thread.elt_dump }; + if (expired_lock_dump) { spin_unlock_bh(&waiting_locks_spinlock); /* from waiting_locks_callback, but not in timer */ libcfs_debug_dumplog(); - libcfs_run_lbug_upcall(&msgdata); spin_lock_bh(&waiting_locks_spinlock); - expired_lock_thread.elt_dump = 0; - } + expired_lock_dump = 0; + } - do_dump = 0; + do_dump = 0; - while (!cfs_list_empty(expired)) { - struct obd_export *export; - struct ldlm_lock *lock; + while (!list_empty(expired)) { + struct obd_export *export; + struct ldlm_lock *lock; - lock = cfs_list_entry(expired->next, struct ldlm_lock, - l_pending_chain); - if ((void *)lock < LP_POISON + CFS_PAGE_SIZE && - (void *)lock >= LP_POISON) { + lock = list_entry(expired->next, struct ldlm_lock, + l_pending_chain); + if ((void *)lock < LP_POISON + PAGE_SIZE && + (void *)lock >= LP_POISON) { spin_unlock_bh(&waiting_locks_spinlock); - CERROR("free lock on elt list %p\n", lock); - LBUG(); - } - cfs_list_del_init(&lock->l_pending_chain); - if ((void *)lock->l_export < LP_POISON + CFS_PAGE_SIZE && - (void *)lock->l_export >= LP_POISON) { - CERROR("lock with free export on elt list %p\n", - lock->l_export); - lock->l_export = NULL; - LDLM_ERROR(lock, "free export"); - /* release extra ref grabbed by - * ldlm_add_waiting_lock() or - * ldlm_failed_ast() */ - LDLM_LOCK_RELEASE(lock); - continue; - } + CERROR("free lock on elt list %p\n", lock); + LBUG(); + } + list_del_init(&lock->l_pending_chain); + if ((void *)lock->l_export < + LP_POISON + PAGE_SIZE && + (void *)lock->l_export >= LP_POISON) { + CERROR("lock with free export on elt list %p\n", + lock->l_export); + lock->l_export = NULL; + LDLM_ERROR(lock, "free export"); + /* release extra ref grabbed by + * ldlm_add_waiting_lock() or + * ldlm_failed_ast() */ + LDLM_LOCK_RELEASE(lock); + continue; + } - if (lock->l_destroyed) { + if (ldlm_is_destroyed(lock)) { /* release the lock refcount where * waiting_locks_callback() founds */ LDLM_LOCK_RELEASE(lock); @@ -225,6 +228,10 @@ static int expired_lock_main(void *arg) export = class_export_lock_get(lock->l_export, lock); spin_unlock_bh(&waiting_locks_spinlock); + spin_lock_bh(&export->exp_bl_list_lock); + list_del_init(&lock->l_exp_list); + spin_unlock_bh(&export->exp_bl_list_lock); + do_dump++; class_fail_export(export); class_export_lock_put(export, lock); @@ -237,18 +244,18 @@ static int expired_lock_main(void *arg) } spin_unlock_bh(&waiting_locks_spinlock); - if (do_dump && obd_dump_on_eviction) { - CERROR("dump the log upon eviction\n"); - libcfs_debug_dumplog(); - } + if (do_dump && obd_dump_on_eviction) { + CERROR("dump the log upon eviction\n"); + libcfs_debug_dumplog(); + } - if (expired_lock_thread.elt_state == ELT_TERMINATE) - break; - } + if (expired_lock_thread_state == ELT_TERMINATE) + break; + } - expired_lock_thread.elt_state = ELT_STOPPED; - cfs_waitq_signal(&expired_lock_thread.elt_waitq); - RETURN(0); + expired_lock_thread_state = ELT_STOPPED; + wake_up(&expired_lock_wait_queue); + RETURN(0); } static int ldlm_add_waiting_lock(struct ldlm_lock *lock); @@ -268,7 +275,7 @@ static int ldlm_lock_busy(struct ldlm_lock *lock) return 0; spin_lock_bh(&lock->l_export->exp_rpc_lock); - cfs_list_for_each_entry(req, &lock->l_export->exp_hp_rpcs, + list_for_each_entry(req, &lock->l_export->exp_hp_rpcs, rq_exp_list) { if (req->rq_ops->hpreq_lock_match) { match = req->rq_ops->hpreq_lock_match(req, lock); @@ -287,61 +294,14 @@ static void waiting_locks_callback(unsigned long unused) int need_dump = 0; spin_lock_bh(&waiting_locks_spinlock); - while (!cfs_list_empty(&waiting_locks_list)) { - lock = cfs_list_entry(waiting_locks_list.next, struct ldlm_lock, + while (!list_empty(&waiting_locks_list)) { + lock = list_entry(waiting_locks_list.next, struct ldlm_lock, l_pending_chain); if (cfs_time_after(lock->l_callback_timeout, cfs_time_current()) || (lock->l_req_mode == LCK_GROUP)) break; - if (ptlrpc_check_suspend()) { - /* there is a case when we talk to one mds, holding - * lock from another mds. this way we easily can get - * here, if second mds is being recovered. so, we - * suspend timeouts. bug 6019 */ - - LDLM_ERROR(lock, "recharge timeout: %s@%s nid %s ", - lock->l_export->exp_client_uuid.uuid, - lock->l_export->exp_connection->c_remote_uuid.uuid, - libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid)); - - cfs_list_del_init(&lock->l_pending_chain); - if (lock->l_destroyed) { - /* relay the lock refcount decrease to - * expired lock thread */ - cfs_list_add(&lock->l_pending_chain, - &expired_lock_thread.elt_expired_locks); - } else { - __ldlm_add_waiting_lock(lock, - ldlm_get_enq_timeout(lock)); - } - continue; - } - - /* if timeout overlaps the activation time of suspended timeouts - * then extend it to give a chance for client to reconnect */ - if (cfs_time_before(cfs_time_sub(lock->l_callback_timeout, - cfs_time_seconds(obd_timeout)/2), - ptlrpc_suspend_wakeup_time())) { - LDLM_ERROR(lock, "extend timeout due to recovery: %s@%s nid %s ", - lock->l_export->exp_client_uuid.uuid, - lock->l_export->exp_connection->c_remote_uuid.uuid, - libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid)); - - cfs_list_del_init(&lock->l_pending_chain); - if (lock->l_destroyed) { - /* relay the lock refcount decrease to - * expired lock thread */ - cfs_list_add(&lock->l_pending_chain, - &expired_lock_thread.elt_expired_locks); - } else { - __ldlm_add_waiting_lock(lock, - ldlm_get_enq_timeout(lock)); - } - continue; - } - /* Check if we need to prolong timeout */ if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) && ldlm_lock_busy(lock)) { @@ -355,7 +315,7 @@ static void waiting_locks_callback(unsigned long unused) spin_unlock_bh(&waiting_locks_spinlock); LDLM_DEBUG(lock, "prolong the busy lock"); ldlm_refresh_waiting_lock(lock, - ldlm_get_enq_timeout(lock)); + ldlm_bl_timeout(lock) >> 1); spin_lock_bh(&waiting_locks_spinlock); if (!cont) { @@ -367,43 +327,44 @@ static void waiting_locks_callback(unsigned long unused) continue; } ldlm_lock_to_ns(lock)->ns_timeouts++; - LDLM_ERROR(lock, "lock callback timer expired after %lds: " + LDLM_ERROR(lock, "lock callback timer expired after %llds: " "evicting client at %s ", - cfs_time_current_sec()- lock->l_last_activity, + ktime_get_real_seconds() - lock->l_last_activity, libcfs_nid2str( lock->l_export->exp_connection->c_peer.nid)); /* no needs to take an extra ref on the lock since it was in * the waiting_locks_list and ldlm_add_waiting_lock() * already grabbed a ref */ - cfs_list_del(&lock->l_pending_chain); - cfs_list_add(&lock->l_pending_chain, - &expired_lock_thread.elt_expired_locks); + list_del(&lock->l_pending_chain); + list_add(&lock->l_pending_chain, &expired_lock_list); need_dump = 1; } - if (!cfs_list_empty(&expired_lock_thread.elt_expired_locks)) { + if (!list_empty(&expired_lock_list)) { if (obd_dump_on_timeout && need_dump) - expired_lock_thread.elt_dump = __LINE__; + expired_lock_dump = __LINE__; - cfs_waitq_signal(&expired_lock_thread.elt_waitq); + wake_up(&expired_lock_wait_queue); } /* * Make sure the timer will fire again if we have any locks * left. */ - if (!cfs_list_empty(&waiting_locks_list)) { + if (!list_empty(&waiting_locks_list)) { cfs_time_t timeout_rounded; - lock = cfs_list_entry(waiting_locks_list.next, struct ldlm_lock, + lock = list_entry(waiting_locks_list.next, struct ldlm_lock, l_pending_chain); timeout_rounded = (cfs_time_t)round_timeout(lock->l_callback_timeout); - cfs_timer_arm(&waiting_locks_timer, timeout_rounded); + mod_timer(&waiting_locks_timer, timeout_rounded); } spin_unlock_bh(&waiting_locks_spinlock); } -/* +/** + * Add lock to the list of contended locks. + * * Indicate that we're waiting for a client to call us back cancelling a given * lock. We add it to the pending-callback chain, and schedule the lock-timeout * timer to fire appropriately. (We round up to the next second, to avoid @@ -418,7 +379,7 @@ static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds) cfs_time_t timeout; cfs_time_t timeout_rounded; - if (!cfs_list_empty(&lock->l_pending_chain)) + if (!list_empty(&lock->l_pending_chain)) return 0; if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) || @@ -431,56 +392,84 @@ static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds) timeout_rounded = round_timeout(lock->l_callback_timeout); - if (cfs_time_before(timeout_rounded, - cfs_timer_deadline(&waiting_locks_timer)) || - !cfs_timer_is_armed(&waiting_locks_timer)) { - cfs_timer_arm(&waiting_locks_timer, timeout_rounded); + if (cfs_time_before(timeout_rounded, waiting_locks_timer.expires) || + !timer_pending(&waiting_locks_timer)) { + mod_timer(&waiting_locks_timer, timeout_rounded); } /* if the new lock has a shorter timeout than something earlier on the list, we'll wait the longer amount of time; no big deal. */ /* FIFO */ - cfs_list_add_tail(&lock->l_pending_chain, &waiting_locks_list); + list_add_tail(&lock->l_pending_chain, &waiting_locks_list); return 1; } +static void ldlm_add_blocked_lock(struct ldlm_lock *lock) +{ + spin_lock_bh(&lock->l_export->exp_bl_list_lock); + if (list_empty(&lock->l_exp_list)) { + if (lock->l_granted_mode != lock->l_req_mode) + list_add_tail(&lock->l_exp_list, + &lock->l_export->exp_bl_list); + else + list_add(&lock->l_exp_list, + &lock->l_export->exp_bl_list); + } + spin_unlock_bh(&lock->l_export->exp_bl_list_lock); + + /* A blocked lock is added. Adjust the position in + * the stale list if the export is in the list. + * If export is stale and not in the list - it is being + * processed and will be placed on the right position + * on obd_stale_export_put(). */ + if (!list_empty(&lock->l_export->exp_stale_list)) + obd_stale_export_adjust(lock->l_export); +} + static int ldlm_add_waiting_lock(struct ldlm_lock *lock) { int ret; - int timeout = ldlm_get_enq_timeout(lock); + int timeout = ldlm_bl_timeout(lock); /* NB: must be called with hold of lock_res_and_lock() */ - LASSERT(lock->l_res_locked); - lock->l_waited = 1; + LASSERT(ldlm_is_res_locked(lock)); + LASSERT(!ldlm_is_cancel_on_block(lock)); - LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)); + /* Do not put cross-MDT lock in the waiting list, since we + * will not evict it due to timeout for now */ + if (lock->l_export != NULL && + (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS)) + return 0; spin_lock_bh(&waiting_locks_spinlock); - if (lock->l_destroyed) { - static cfs_time_t next; + if (ldlm_is_cancel(lock)) { spin_unlock_bh(&waiting_locks_spinlock); - LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)"); - if (cfs_time_after(cfs_time_current(), next)) { - next = cfs_time_shift(14400); - libcfs_debug_dumpstack(NULL); - } - return 0; - } + return 0; + } - ret = __ldlm_add_waiting_lock(lock, timeout); - if (ret) { - /* grab ref on the lock if it has been added to the - * waiting list */ - LDLM_LOCK_GET(lock); - } - spin_unlock_bh(&waiting_locks_spinlock); + if (ldlm_is_destroyed(lock)) { + static cfs_time_t next; + spin_unlock_bh(&waiting_locks_spinlock); + LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)"); + if (cfs_time_after(cfs_time_current(), next)) { + next = cfs_time_shift(14400); + libcfs_debug_dumpstack(NULL); + } + return 0; + } + + ldlm_set_waited(lock); + lock->l_last_activity = ktime_get_real_seconds(); + ret = __ldlm_add_waiting_lock(lock, timeout); if (ret) { - spin_lock_bh(&lock->l_export->exp_bl_list_lock); - if (cfs_list_empty(&lock->l_exp_list)) - cfs_list_add(&lock->l_exp_list, - &lock->l_export->exp_bl_list); - spin_unlock_bh(&lock->l_export->exp_bl_list_lock); + /* grab ref on the lock if it has been added to the + * waiting list */ + LDLM_LOCK_GET(lock); } + spin_unlock_bh(&waiting_locks_spinlock); + + if (ret) + ldlm_add_blocked_lock(lock); LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)", ret == 0 ? "not re-" : "", timeout, @@ -488,7 +477,7 @@ static int ldlm_add_waiting_lock(struct ldlm_lock *lock) return ret; } -/* +/** * Remove a lock from the pending list, likely because it had its cancellation * callback arrive without incident. This adjusts the lock-timeout timer if * needed. Returns 0 if the lock wasn't pending after all, 1 if it was. @@ -499,9 +488,9 @@ static int ldlm_add_waiting_lock(struct ldlm_lock *lock) */ static int __ldlm_del_waiting_lock(struct ldlm_lock *lock) { - cfs_list_t *list_next; + struct list_head *list_next; - if (cfs_list_empty(&lock->l_pending_chain)) + if (list_empty(&lock->l_pending_chain)) return 0; list_next = lock->l_pending_chain.next; @@ -509,16 +498,16 @@ static int __ldlm_del_waiting_lock(struct ldlm_lock *lock) /* Removing the head of the list, adjust timer. */ if (list_next == &waiting_locks_list) { /* No more, just cancel. */ - cfs_timer_disarm(&waiting_locks_timer); + del_timer(&waiting_locks_timer); } else { struct ldlm_lock *next; - next = cfs_list_entry(list_next, struct ldlm_lock, + next = list_entry(list_next, struct ldlm_lock, l_pending_chain); - cfs_timer_arm(&waiting_locks_timer, - round_timeout(next->l_callback_timeout)); + mod_timer(&waiting_locks_timer, + round_timeout(next->l_callback_timeout)); } } - cfs_list_del_init(&lock->l_pending_chain); + list_del_init(&lock->l_pending_chain); return 1; } @@ -535,11 +524,12 @@ int ldlm_del_waiting_lock(struct ldlm_lock *lock) spin_lock_bh(&waiting_locks_spinlock); ret = __ldlm_del_waiting_lock(lock); + ldlm_clear_waited(lock); spin_unlock_bh(&waiting_locks_spinlock); /* remove the lock out of export blocking list */ spin_lock_bh(&lock->l_export->exp_bl_list_lock); - cfs_list_del_init(&lock->l_exp_list); + list_del_init(&lock->l_exp_list); spin_unlock_bh(&lock->l_export->exp_bl_list_lock); if (ret) { @@ -551,10 +541,9 @@ int ldlm_del_waiting_lock(struct ldlm_lock *lock) LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed"); return ret; } -EXPORT_SYMBOL(ldlm_del_waiting_lock); -/* - * Prolong the lock +/** + * Prolong the contended lock waiting time. * * Called with namespace lock held. */ @@ -566,9 +555,15 @@ int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout) return 0; } + if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) { + /* We don't have a "waiting locks list" on OSP. */ + LDLM_DEBUG(lock, "MDS-MDS lock: no-op"); + return 0; + } + spin_lock_bh(&waiting_locks_spinlock); - if (cfs_list_empty(&lock->l_pending_chain)) { + if (list_empty(&lock->l_pending_chain)) { spin_unlock_bh(&waiting_locks_spinlock); LDLM_DEBUG(lock, "wasn't waiting"); return 0; @@ -585,7 +580,7 @@ int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout) } EXPORT_SYMBOL(ldlm_refresh_waiting_lock); -#else /* !HAVE_SERVER_SUPPORT || !__KERNEL__ */ +#else /* HAVE_SERVER_SUPPORT */ int ldlm_del_waiting_lock(struct ldlm_lock *lock) { @@ -597,19 +592,38 @@ int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout) RETURN(0); } -# ifdef HAVE_SERVER_SUPPORT -static int ldlm_add_waiting_lock(struct ldlm_lock *lock) +#endif /* !HAVE_SERVER_SUPPORT */ + +#ifdef HAVE_SERVER_SUPPORT + +/** + * Calculate the per-export Blocking timeout (covering BL AST, data flush, + * lock cancel, and their replies). Used for lock callback timeout and AST + * re-send period. + * + * \param[in] lock lock which is getting the blocking callback + * + * \retval timeout in seconds to wait for the client reply + */ +unsigned int ldlm_bl_timeout(struct ldlm_lock *lock) { - LASSERT(lock->l_res_locked); - LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)); - RETURN(1); -} + unsigned int timeout; -# endif -#endif /* HAVE_SERVER_SUPPORT && __KERNEL__ */ + if (AT_OFF) + return obd_timeout / 2; -#ifdef HAVE_SERVER_SUPPORT + /* Since these are non-updating timeouts, we should be conservative. + * Take more than usually, 150% + * It would be nice to have some kind of "early reply" mechanism for + * lock callbacks too... */ + timeout = at_get(&lock->l_export->exp_bl_lock_at); + return max(timeout + (timeout >> 1), ldlm_enqueue_min); +} +EXPORT_SYMBOL(ldlm_bl_timeout); +/** + * Perform lock cleanup if AST sending failed. + */ static void ldlm_failed_ast(struct ldlm_lock *lock, int rc, const char *ast_type) { @@ -620,76 +634,93 @@ static void ldlm_failed_ast(struct ldlm_lock *lock, int rc, if (obd_dump_on_timeout) libcfs_debug_dumplog(); -#ifdef __KERNEL__ spin_lock_bh(&waiting_locks_spinlock); if (__ldlm_del_waiting_lock(lock) == 0) /* the lock was not in any list, grab an extra ref before adding * the lock to the expired list */ LDLM_LOCK_GET(lock); - cfs_list_add(&lock->l_pending_chain, - &expired_lock_thread.elt_expired_locks); - cfs_waitq_signal(&expired_lock_thread.elt_waitq); + list_add(&lock->l_pending_chain, &expired_lock_list); + wake_up(&expired_lock_wait_queue); spin_unlock_bh(&waiting_locks_spinlock); -#else - class_fail_export(lock->l_export); -#endif } +/** + * Perform lock cleanup if AST reply came with error. + */ static int ldlm_handle_ast_error(struct ldlm_lock *lock, - struct ptlrpc_request *req, int rc, - const char *ast_type) -{ - lnet_process_id_t peer = req->rq_import->imp_connection->c_peer; - - if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) { - LASSERT(lock->l_export); - if (lock->l_export->exp_libclient) { - LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)" - " timeout, just cancelling lock", ast_type, - libcfs_nid2str(peer.nid)); - ldlm_lock_cancel(lock); - rc = -ERESTART; - } else if (lock->l_flags & LDLM_FL_CANCEL) { - LDLM_DEBUG(lock, "%s AST timeout from nid %s, but " - "cancel was received (AST reply lost?)", - ast_type, libcfs_nid2str(peer.nid)); - ldlm_lock_cancel(lock); - rc = -ERESTART; - } else { - ldlm_del_waiting_lock(lock); - ldlm_failed_ast(lock, rc, ast_type); - } - } else if (rc) { - if (rc == -EINVAL) { - struct ldlm_resource *res = lock->l_resource; - LDLM_DEBUG(lock, "client (nid %s) returned %d" - " from %s AST - normal race", - libcfs_nid2str(peer.nid), - req->rq_repmsg ? - lustre_msg_get_status(req->rq_repmsg) : -1, - ast_type); - if (res) { - /* update lvbo to return proper attributes. - * see bug 23174 */ - ldlm_resource_getref(res); - ldlm_res_lvbo_update(res, NULL, 1); - ldlm_resource_putref(res); - } + struct ptlrpc_request *req, int rc, + const char *ast_type) +{ + struct lnet_process_id peer = req->rq_import->imp_connection->c_peer; + + if (!req->rq_replied || (rc && rc != -EINVAL)) { + if (lock->l_export && lock->l_export->exp_libclient) { + LDLM_DEBUG(lock, + "%s AST (req@%p x%llu) to liblustre client (nid %s) timeout, just cancelling lock", + ast_type, req, req->rq_xid, + libcfs_nid2str(peer.nid)); + ldlm_lock_cancel(lock); + rc = -ERESTART; + } else if (ldlm_is_cancel(lock)) { + LDLM_DEBUG(lock, + "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)", + ast_type, req, req->rq_xid, + libcfs_nid2str(peer.nid)); + ldlm_lock_cancel(lock); + rc = -ERESTART; + } else if (rc == -ENODEV || rc == -ESHUTDOWN || + (rc == -EIO && + req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) { + /* Upon umount process the AST fails because cannot be + * sent. This shouldn't lead to the client eviction. + * -ENODEV error is returned by ptl_send_rpc() for + * new request in such import. + * -SHUTDOWN is returned by ptlrpc_import_delay_req() + * if imp_invalid is set or obd_no_recov. + * Meanwhile there is also check for LUSTRE_IMP_CLOSED + * in ptlrpc_import_delay_req() as well with -EIO code. + * In all such cases errors are ignored. + */ + LDLM_DEBUG(lock, "%s AST can't be sent due to a server" + " %s failure or umount process: rc = %d\n", + ast_type, + req->rq_import->imp_obd->obd_name, rc); + } else { + LDLM_ERROR(lock, + "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it", + libcfs_nid2str(peer.nid), + req->rq_replied ? "returned error from" : + "failed to reply to", + ast_type, req, req->rq_xid, + (req->rq_repmsg != NULL) ? + lustre_msg_get_status(req->rq_repmsg) : 0, + rc); + ldlm_failed_ast(lock, rc, ast_type); + } + return rc; + } - } else { - LDLM_ERROR(lock, "client (nid %s) returned %d " - "from %s AST", libcfs_nid2str(peer.nid), - (req->rq_repmsg != NULL) ? - lustre_msg_get_status(req->rq_repmsg) : 0, - ast_type); - } - ldlm_lock_cancel(lock); - /* Server-side AST functions are called from ldlm_reprocess_all, - * which needs to be told to please restart its reprocessing. */ - rc = -ERESTART; - } + if (rc == -EINVAL) { + struct ldlm_resource *res = lock->l_resource; + + LDLM_DEBUG(lock, + "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race", + libcfs_nid2str(peer.nid), + req->rq_repmsg ? + lustre_msg_get_status(req->rq_repmsg) : -1, + ast_type, req, req->rq_xid); + if (res) { + /* update lvbo to return proper attributes. + * see bug 23174 */ + ldlm_resource_getref(res); + ldlm_lvbo_update(res, lock, NULL, 1); + ldlm_resource_putref(res); + } + ldlm_lock_cancel(lock); + rc = -ERESTART; + } - return rc; + return rc; } static int ldlm_cb_interpret(const struct lu_env *env, @@ -712,14 +743,16 @@ static int ldlm_cb_interpret(const struct lu_env *env, * - Glimpse callback of remote lock might return * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274 */ - if (rc == -ELDLM_NO_LOCK_DATA) { + if (unlikely(arg->gl_interpret_reply)) { + rc = arg->gl_interpret_reply(env, req, data, rc); + } else if (rc == -ELDLM_NO_LOCK_DATA) { LDLM_DEBUG(lock, "lost race - client has a lock but no " "inode"); - ldlm_res_lvbo_update(lock->l_resource, NULL, 1); + ldlm_lvbo_update(lock->l_resource, lock, NULL, 1); } else if (rc != 0) { rc = ldlm_handle_ast_error(lock, req, rc, "glimpse"); } else { - rc = ldlm_res_lvbo_update(lock->l_resource, req, 1); + rc = ldlm_lvbo_update(lock->l_resource, lock, req, 1); } break; case LDLM_BL_CALLBACK: @@ -740,9 +773,17 @@ static int ldlm_cb_interpret(const struct lu_env *env, LDLM_LOCK_RELEASE(lock); if (rc == -ERESTART) - cfs_atomic_inc(&arg->restart); + atomic_inc(&arg->restart); - RETURN(0); + RETURN(0); +} + +static void ldlm_update_resend(struct ptlrpc_request *req, void *data) +{ + struct ldlm_cb_async_args *ca = data; + struct ldlm_lock *lock = ca->ca_lock; + + ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock)); } static inline int ldlm_ast_fini(struct ptlrpc_request *req, @@ -757,7 +798,7 @@ static inline int ldlm_ast_fini(struct ptlrpc_request *req, rc = ptl_send_rpc(req, 1); ptlrpc_req_finished(req); if (rc == 0) - cfs_atomic_inc(&arg->restart); + atomic_inc(&arg->restart); } else { LDLM_LOCK_GET(lock); ptlrpc_set_add_req(arg->set, req); @@ -772,34 +813,37 @@ static inline int ldlm_ast_fini(struct ptlrpc_request *req, */ static void ldlm_lock_reorder_req(struct ldlm_lock *lock) { - struct ptlrpc_request *req; - ENTRY; + struct ptlrpc_request *req; + ENTRY; - if (lock->l_export == NULL) { - LDLM_DEBUG(lock, "client lock: no-op"); - RETURN_EXIT; - } + if (lock->l_export == NULL) { + LDLM_DEBUG(lock, "client lock: no-op"); + RETURN_EXIT; + } spin_lock_bh(&lock->l_export->exp_rpc_lock); - cfs_list_for_each_entry(req, &lock->l_export->exp_hp_rpcs, - rq_exp_list) { - /* Do not process requests that were not yet added to there - * incoming queue or were already removed from there for - * processing */ - if (!req->rq_hp && !cfs_list_empty(&req->rq_list) && - req->rq_ops->hpreq_lock_match && - req->rq_ops->hpreq_lock_match(req, lock)) - ptlrpc_hpreq_reorder(req); - } + list_for_each_entry(req, &lock->l_export->exp_hp_rpcs, + rq_exp_list) { + /* Do not process requests that were not yet added to there + * incoming queue or were already removed from there for + * processing. We evaluate ptlrpc_nrs_req_can_move() without + * holding svcpt->scp_req_lock, and then redo the check with + * the lock held once we need to obtain a reliable result. + */ + if (ptlrpc_nrs_req_can_move(req) && + req->rq_ops->hpreq_lock_match && + req->rq_ops->hpreq_lock_match(req, lock)) + ptlrpc_nrs_req_hp_move(req); + } spin_unlock_bh(&lock->l_export->exp_rpc_lock); EXIT; } -/* +/** * ->l_blocking_ast() method for server-side locks. This is invoked when newly * enqueued server lock conflicts with given one. * - * Sends blocking ast rpc to the client owning that lock; arms timeout timer + * Sends blocking AST RPC to the client owning that lock; arms timeout timer * to wait for client response. */ int ldlm_server_blocking_ast(struct ldlm_lock *lock, @@ -818,6 +862,11 @@ int ldlm_server_blocking_ast(struct ldlm_lock *lock, /* Don't need to do anything here. */ RETURN(0); + if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) { + LDLM_DEBUG(lock, "dropping BL AST"); + RETURN(0); + } + LASSERT(lock); LASSERT(data != NULL); if (lock->l_export->exp_obd->obd_recovering != 0) @@ -837,44 +886,52 @@ int ldlm_server_blocking_ast(struct ldlm_lock *lock, ca->ca_lock = lock; req->rq_interpret_reply = ldlm_cb_interpret; - req->rq_no_resend = 1; lock_res_and_lock(lock); - if (lock->l_granted_mode != lock->l_req_mode) { - /* this blocking AST will be communicated as part of the - * completion AST instead */ + if (ldlm_is_destroyed(lock)) { + /* What's the point? */ unlock_res_and_lock(lock); - ptlrpc_req_finished(req); - LDLM_DEBUG(lock, "lock not granted, not sending blocking AST"); RETURN(0); } - if (lock->l_destroyed) { - /* What's the point? */ + if (lock->l_granted_mode != lock->l_req_mode) { + /* this blocking AST will be communicated as part of the + * completion AST instead */ + ldlm_add_blocked_lock(lock); + ldlm_set_waited(lock); unlock_res_and_lock(lock); + ptlrpc_req_finished(req); + LDLM_DEBUG(lock, "lock not granted, not sending blocking AST"); RETURN(0); } - if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) + if (ldlm_is_cancel_on_block(lock)) instant_cancel = 1; body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); body->lock_handle[0] = lock->l_remote_handle; body->lock_desc = *desc; - body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_AST_FLAGS); + body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK); LDLM_DEBUG(lock, "server preparing blocking AST"); ptlrpc_request_set_replen(req); + ldlm_set_cbpending(lock); if (instant_cancel) { unlock_res_and_lock(lock); ldlm_lock_cancel(lock); + + req->rq_no_resend = 1; } else { LASSERT(lock->l_granted_mode == lock->l_req_mode); ldlm_add_waiting_lock(lock); unlock_res_and_lock(lock); + + /* Do not resend after lock callback timeout */ + req->rq_delay_limit = ldlm_bl_timeout(lock); + req->rq_resend_cb = ldlm_update_resend; } req->rq_send_state = LUSTRE_IMP_FULL; @@ -882,6 +939,8 @@ int ldlm_server_blocking_ast(struct ldlm_lock *lock, if (AT_OFF) req->rq_timeout = ldlm_get_rq_timeout(); + lock->l_last_activity = ktime_get_real_seconds(); + if (lock->l_export && lock->l_export->exp_nid_stats && lock->l_export->exp_nid_stats->nid_ldlm_stats) lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats, @@ -891,15 +950,20 @@ int ldlm_server_blocking_ast(struct ldlm_lock *lock, RETURN(rc); } -EXPORT_SYMBOL(ldlm_server_blocking_ast); +/** + * ->l_completion_ast callback for a remote lock in server namespace. + * + * Sends AST to the client notifying it of lock granting. If initial + * lock response was not sent yet, instead of sending another RPC, just + * mark the lock as granted and client will understand + */ int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data) { struct ldlm_cb_set_arg *arg = data; struct ldlm_request *body; struct ptlrpc_request *req; struct ldlm_cb_async_args *ca; - long total_enqueue_wait; int instant_cancel = 0; int rc = 0; int lvb_len; @@ -908,8 +972,10 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data) LASSERT(lock != NULL); LASSERT(data != NULL); - total_enqueue_wait = cfs_time_sub(cfs_time_current_sec(), - lock->l_last_activity); + if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) { + LDLM_DEBUG(lock, "dropping CP AST"); + RETURN(0); + } req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse, &RQF_LDLM_CP_CALLBACK); @@ -918,10 +984,14 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data) /* server namespace, doesn't need lock */ lvb_len = ldlm_lvbo_size(lock); - if (lvb_len > 0) - req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, - lvb_len); - + /* LU-3124 & LU-2187: to not return layout in completion AST because + * it may deadlock for LU-2187, or client may not have enough space + * for large layout. The layout will be returned to client with an + * extra RPC to fetch xattr.lov */ + if (ldlm_has_layout(lock)) + lvb_len = 0; + + req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len); rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK); if (rc) { ptlrpc_request_free(req); @@ -934,7 +1004,6 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data) ca->ca_lock = lock; req->rq_interpret_reply = ldlm_cb_interpret; - req->rq_no_resend = 1; body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); body->lock_handle[0] = lock->l_remote_handle; @@ -944,27 +1013,25 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data) void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB); lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len); - req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, - lvb_len, RCL_CLIENT); + if (lvb_len < 0) { + /* We still need to send the RPC to wake up the blocked + * enqueue thread on the client. + * + * Consider old client, there is no better way to notify + * the failure, just zero-sized the LVB, then the client + * will fail out as "-EPROTO". */ + req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0, + RCL_CLIENT); + instant_cancel = 1; + } else { + req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len, + RCL_CLIENT); + } } - LDLM_DEBUG(lock, "server preparing completion AST (after %lds wait)", - total_enqueue_wait); + lock->l_last_activity = ktime_get_real_seconds(); - /* Server-side enqueue wait time estimate, used in - __ldlm_add_waiting_lock to set future enqueue timers */ - if (total_enqueue_wait < ldlm_get_enq_timeout(lock)) - at_measured(ldlm_lock_to_ns_at(lock), - total_enqueue_wait); - else - /* bz18618. Don't add lock enqueue time we spend waiting for a - previous callback to fail. Locks waiting legitimately will - get extended by ldlm_refresh_waiting_lock regardless of the - estimate, so it's okay to underestimate here. */ - LDLM_DEBUG(lock, "lock completed after %lus; estimate was %ds. " - "It is likely that a previous callback timed out.", - total_enqueue_wait, - at_get(ldlm_lock_to_ns_at(lock))); + LDLM_DEBUG(lock, "server preparing completion AST"); ptlrpc_request_set_replen(req); @@ -975,11 +1042,11 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data) /* We only send real blocking ASTs after the lock is granted */ lock_res_and_lock(lock); - if (lock->l_flags & LDLM_FL_AST_SENT) { + if (ldlm_is_ast_sent(lock)) { body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT); - /* copy ast flags like LDLM_FL_DISCARD_DATA */ + /* Copy AST flags like LDLM_FL_DISCARD_DATA. */ body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & - LDLM_AST_FLAGS); + LDLM_FL_AST_MASK); /* We might get here prior to ldlm_handle_enqueue setting * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock @@ -987,15 +1054,21 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data) * ldlm_handle_enqueue will call ldlm_lock_cancel() still, * that would not only cancel the lock, but will also remove * it from waiting list */ - if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) { - unlock_res_and_lock(lock); - ldlm_lock_cancel(lock); - instant_cancel = 1; - lock_res_and_lock(lock); - } else { - /* start the lock-timeout clock */ - ldlm_add_waiting_lock(lock); - } + if (ldlm_is_cancel_on_block(lock)) { + unlock_res_and_lock(lock); + ldlm_lock_cancel(lock); + + instant_cancel = 1; + req->rq_no_resend = 1; + + lock_res_and_lock(lock); + } else { + /* start the lock-timeout clock */ + ldlm_add_waiting_lock(lock); + /* Do not resend after lock callback timeout */ + req->rq_delay_limit = ldlm_bl_timeout(lock); + req->rq_resend_cb = ldlm_update_resend; + } } unlock_res_and_lock(lock); @@ -1006,10 +1079,15 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data) rc = ldlm_ast_fini(req, arg, lock, instant_cancel); - RETURN(rc); + RETURN(lvb_len < 0 ? lvb_len : rc); } -EXPORT_SYMBOL(ldlm_server_completion_ast); +/** + * Server side ->l_glimpse_ast handler for client locks. + * + * Sends glimpse AST to the client and waits for reply. Then updates + * lvbo with the result. + */ int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data) { struct ldlm_cb_set_arg *arg = data; @@ -1061,6 +1139,8 @@ int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data) if (AT_OFF) req->rq_timeout = ldlm_get_rq_timeout(); + lock->l_last_activity = ktime_get_real_seconds(); + req->rq_interpret_reply = ldlm_cb_interpret; if (lock->l_export && lock->l_export->exp_nid_stats && @@ -1072,9 +1152,9 @@ int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data) RETURN(rc); } -EXPORT_SYMBOL(ldlm_server_glimpse_ast); -int ldlm_glimpse_locks(struct ldlm_resource *res, cfs_list_t *gl_work_list) +int ldlm_glimpse_locks(struct ldlm_resource *res, + struct list_head *gl_work_list) { int rc; ENTRY; @@ -1088,7 +1168,7 @@ int ldlm_glimpse_locks(struct ldlm_resource *res, cfs_list_t *gl_work_list) } EXPORT_SYMBOL(ldlm_glimpse_locks); -/* return ldlm lock associated with a lock callback request */ +/* return LDLM lock associated with a lock callback request */ struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req) { struct ldlm_cb_async_args *ca; @@ -1138,29 +1218,30 @@ static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req, return; } -/* - * Main server-side entry point into LDLM. This is called by ptlrpc service - * threads to carry out client lock enqueueing requests. +/** + * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc + * service threads to carry out client lock enqueueing requests. */ int ldlm_handle_enqueue0(struct ldlm_namespace *ns, - struct ptlrpc_request *req, - const struct ldlm_request *dlm_req, - const struct ldlm_callback_suite *cbs) + struct ptlrpc_request *req, + const struct ldlm_request *dlm_req, + const struct ldlm_callback_suite *cbs) { - struct ldlm_reply *dlm_rep; + struct ldlm_reply *dlm_rep; __u64 flags; - ldlm_error_t err = ELDLM_OK; - struct ldlm_lock *lock = NULL; - void *cookie = NULL; - int rc = 0; - ENTRY; + enum ldlm_error err = ELDLM_OK; + struct ldlm_lock *lock = NULL; + void *cookie = NULL; + int rc = 0; + struct ldlm_resource *res = NULL; + ENTRY; - LDLM_DEBUG_NOLOCK("server-side enqueue handler START"); + LDLM_DEBUG_NOLOCK("server-side enqueue handler START"); - ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF); + ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP); flags = ldlm_flags_from_wire(dlm_req->lock_flags); - LASSERT(req->rq_export); + LASSERT(req->rq_export); if (ptlrpc_req2svc(req)->srv_stats != NULL) ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats); @@ -1186,7 +1267,7 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns, GOTO(out, rc = -EFAULT); } - if (req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) { + if (exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) { if (unlikely(dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) { DEBUG_REQ(D_ERROR, req, @@ -1200,24 +1281,8 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns, GOTO(out, rc = -EPROTO); } -#if 0 - /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check - against server's _CONNECT_SUPPORTED flags? (I don't want to use - ibits for mgc/mgs) */ - - /* INODEBITS_INTEROP: Perform conversion from plain lock to - * inodebits lock if client does not support them. */ - if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) && - (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) { - dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS; - dlm_req->lock_desc.l_policy_data.l_inodebits.bits = - MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE; - if (dlm_req->lock_desc.l_req_mode == LCK_PR) - dlm_req->lock_desc.l_req_mode = LCK_CR; - } -#endif - - if (unlikely(flags & LDLM_FL_REPLAY)) { + if (unlikely((flags & LDLM_FL_REPLAY) || + (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) { /* Find an existing lock in the per-export lock hash */ /* In the function below, .hs_keycmp resolves to * ldlm_export_lock_keycmp() */ @@ -1225,25 +1290,49 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns, lock = cfs_hash_lookup(req->rq_export->exp_lock_hash, (void *)&dlm_req->lock_handle[0]); if (lock != NULL) { - DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie " - LPX64, lock->l_handle.h_cookie); + DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie %#llx", + lock->l_handle.h_cookie); + flags |= LDLM_FL_RESENT; GOTO(existing_lock, rc = 0); - } - } - - /* The lock's callback data might be set in the policy function */ - lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name, - dlm_req->lock_desc.l_resource.lr_type, - dlm_req->lock_desc.l_req_mode, - cbs, NULL, 0); + } + } else { + if (ldlm_reclaim_full()) { + DEBUG_REQ(D_DLMTRACE, req, "Too many granted locks, " + "reject current enqueue request and let the " + "client retry later.\n"); + GOTO(out, rc = -EINPROGRESS); + } + } - if (!lock) - GOTO(out, rc = -ENOMEM); + /* The lock's callback data might be set in the policy function */ + lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name, + dlm_req->lock_desc.l_resource.lr_type, + dlm_req->lock_desc.l_req_mode, + cbs, NULL, 0, LVB_T_NONE); + if (IS_ERR(lock)) { + rc = PTR_ERR(lock); + lock = NULL; + GOTO(out, rc); + } - lock->l_last_activity = cfs_time_current_sec(); lock->l_remote_handle = dlm_req->lock_handle[0]; LDLM_DEBUG(lock, "server-side enqueue handler, new lock created"); + /* Initialize resource lvb but not for a lock being replayed since + * Client already got lvb sent in this case. + * This must occur early since some policy methods assume resource + * lvb is available (lr_lvb_data != NULL). + */ + res = lock->l_resource; + if (!(flags & LDLM_FL_REPLAY)) { + /* non-replayed lock, delayed lvb init may need to be done */ + rc = ldlm_lvbo_init(res); + if (rc < 0) { + LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc); + GOTO(out, rc); + } + } + OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2); /* Don't enqueue a lock onto the export if it is been disonnected * due to eviction (bug 3822) or server umount (bug 24324). @@ -1260,6 +1349,19 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns, &lock->l_remote_handle, &lock->l_exp_hash); + /* Inherit the enqueue flags before the operation, because we do not + * keep the res lock on return and next operations (BL AST) may proceed + * without them. */ + lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags & + LDLM_FL_INHERIT_MASK); + + ldlm_convert_policy_to_local(req->rq_export, + dlm_req->lock_desc.l_resource.lr_type, + &dlm_req->lock_desc.l_policy_data, + &lock->l_policy_data); + if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT) + lock->l_req_extent = lock->l_policy_data.l_extent; + existing_lock: if (flags & LDLM_FL_HAS_INTENT) { @@ -1281,33 +1383,29 @@ existing_lock: GOTO(out, rc); } - if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN) - ldlm_convert_policy_to_local(req->rq_export, - dlm_req->lock_desc.l_resource.lr_type, - &dlm_req->lock_desc.l_policy_data, - &lock->l_policy_data); - if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT) - lock->l_req_extent = lock->l_policy_data.l_extent; - err = ldlm_lock_enqueue(ns, &lock, cookie, &flags); - if (err) - GOTO(out, err); + if (err) { + if ((int)err < 0) + rc = (int)err; + GOTO(out, err); + } dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP); - dlm_rep->lock_flags = ldlm_flags_to_wire(flags); ldlm_lock2desc(lock, &dlm_rep->lock_desc); ldlm_lock2handle(lock, &dlm_rep->lock_handle); + if (lock && lock->l_resource->lr_type == LDLM_EXTENT) + OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6); + /* We never send a blocking AST until the lock is granted, but * we can tell it right now */ lock_res_and_lock(lock); /* Now take into account flags to be inherited from original lock request both in reply to client and in our own lock flags. */ - dlm_rep->lock_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS; - lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags & - LDLM_INHERIT_FLAGS); + dlm_rep->lock_flags = ldlm_flags_to_wire(flags); + lock->l_flags |= flags & LDLM_FL_INHERIT_MASK; /* Don't move a pending lock onto the export if it has already been * disconnected due to eviction (bug 5683) or server umount (bug 24324). @@ -1316,7 +1414,7 @@ existing_lock: OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) { LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export); rc = -ENOTCONN; - } else if (lock->l_flags & LDLM_FL_AST_SENT) { + } else if (ldlm_is_ast_sent(lock)) { dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT); if (lock->l_granted_mode == lock->l_req_mode) { /* @@ -1339,23 +1437,23 @@ existing_lock: if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN || dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) && req->rq_export->exp_libclient) { - if (unlikely(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) || + if (unlikely(!ldlm_is_cancel_on_block(lock) || !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){ CERROR("Granting sync lock to libclient. " - "req fl %d, rep fl %d, lock fl "LPX64"\n", + "req fl %d, rep fl %d, lock fl %#llx\n", dlm_req->lock_flags, dlm_rep->lock_flags, lock->l_flags); LDLM_ERROR(lock, "sync lock"); - if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) { - struct ldlm_intent *it; - - it = req_capsule_client_get(&req->rq_pill, - &RMF_LDLM_INTENT); - if (it != NULL) { - CERROR("This is intent %s ("LPU64")\n", - ldlm_it2str(it->opc), it->opc); - } - } + if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) { + struct ldlm_intent *it; + + it = req_capsule_client_get(&req->rq_pill, + &RMF_LDLM_INTENT); + if (it != NULL) { + CERROR("This is intent %s (%llu)\n", + ldlm_it2str(it->opc), it->opc); + } + } } } @@ -1372,14 +1470,14 @@ existing_lock: /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this * ldlm_reprocess_all. If this moves, revisit that code. -phil */ - if (lock) { - LDLM_DEBUG(lock, "server-side enqueue handler, sending reply" - "(err=%d, rc=%d)", err, rc); - - if (rc == 0) { - int lvb_len = ldlm_lvbo_size(lock); - - if (lvb_len > 0) { + if (lock != NULL) { + LDLM_DEBUG(lock, "server-side enqueue handler, sending reply" + "(err=%d, rc=%d)", err, rc); + + if (rc == 0) { + if (req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB, + RCL_SERVER) && + ldlm_lvbo_size(lock) > 0) { void *buf; int buflen; @@ -1389,16 +1487,44 @@ existing_lock: req, lock); buflen = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER); - buflen = ldlm_lvbo_fill(lock, buf, buflen); - req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, - buflen, RCL_SERVER); + /* non-replayed lock, delayed lvb init may + * need to be occur now */ + if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) { + buflen = ldlm_lvbo_fill(lock, buf, + buflen); + if (buflen >= 0) + req_capsule_shrink( + &req->rq_pill, + &RMF_DLM_LVB, + buflen, RCL_SERVER); + else + rc = buflen; + } else if (flags & LDLM_FL_REPLAY) { + /* no LVB resend upon replay */ + if (buflen > 0) + req_capsule_shrink( + &req->rq_pill, + &RMF_DLM_LVB, + 0, RCL_SERVER); + else + rc = buflen; + } else { + rc = buflen; + } } - } else { - lock_res_and_lock(lock); - ldlm_resource_unlink_lock(lock); - ldlm_lock_destroy_nolock(lock); - unlock_res_and_lock(lock); - } + } + + if (rc != 0 && !(flags & LDLM_FL_RESENT)) { + if (lock->l_export) { + ldlm_lock_cancel(lock); + } else { + lock_res_and_lock(lock); + ldlm_resource_unlink_lock(lock); + ldlm_lock_destroy_nolock(lock); + unlock_res_and_lock(lock); + + } + } if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK) ldlm_reprocess_all(lock->l_resource); @@ -1411,8 +1537,10 @@ existing_lock: return rc; } -EXPORT_SYMBOL(ldlm_handle_enqueue0); +/** + * Old-style LDLM main entry point for server code enqueue. + */ int ldlm_handle_enqueue(struct ptlrpc_request *req, ldlm_completion_callback completion_callback, ldlm_blocking_callback blocking_callback, @@ -1435,8 +1563,10 @@ int ldlm_handle_enqueue(struct ptlrpc_request *req, } return rc; } -EXPORT_SYMBOL(ldlm_handle_enqueue); +/** + * Main LDLM entry point for server code to process lock conversion requests. + */ int ldlm_handle_convert0(struct ptlrpc_request *req, const struct ldlm_request *dlm_req) { @@ -1459,13 +1589,12 @@ int ldlm_handle_convert0(struct ptlrpc_request *req, lock = ldlm_handle2lock(&dlm_req->lock_handle[0]); if (!lock) { - req->rq_status = EINVAL; + req->rq_status = LUSTRE_EINVAL; } else { void *res = NULL; LDLM_DEBUG(lock, "server-side convert handler START"); - lock->l_last_activity = cfs_time_current_sec(); res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode, &dlm_rep->lock_flags); if (res) { @@ -1473,7 +1602,7 @@ int ldlm_handle_convert0(struct ptlrpc_request *req, LDLM_DEBUG(lock, "converted waiting lock"); req->rq_status = 0; } else { - req->rq_status = EDEADLOCK; + req->rq_status = LUSTRE_EDEADLK; } } @@ -1487,8 +1616,11 @@ int ldlm_handle_convert0(struct ptlrpc_request *req, RETURN(0); } -EXPORT_SYMBOL(ldlm_handle_convert0); +/** + * Old-style main LDLM entry point for server code to process lock conversion + * requests. + */ int ldlm_handle_convert(struct ptlrpc_request *req) { int rc; @@ -1503,11 +1635,16 @@ int ldlm_handle_convert(struct ptlrpc_request *req) } return rc; } -EXPORT_SYMBOL(ldlm_handle_convert); -/* Cancel all the locks whos handles are packed into ldlm_request */ +/** + * Cancel all the locks whose handles are packed into ldlm_request + * + * Called by server code expecting such combined cancel activity + * requests. + */ int ldlm_request_cancel(struct ptlrpc_request *req, - const struct ldlm_request *dlm_req, int first) + const struct ldlm_request *dlm_req, + int first, enum lustre_at_flags flags) { struct ldlm_resource *res, *pres = NULL; struct ldlm_lock *lock; @@ -1533,7 +1670,7 @@ int ldlm_request_cancel(struct ptlrpc_request *req, lock = ldlm_handle2lock(&dlm_req->lock_handle[i]); if (!lock) { LDLM_DEBUG_NOLOCK("server-side cancel handler stale " - "lock (cookie "LPU64")", + "lock (cookie %llu)", dlm_req->lock_handle[i].cookie); continue; } @@ -1541,6 +1678,9 @@ int ldlm_request_cancel(struct ptlrpc_request *req, res = lock->l_resource; done++; + /* This code is an optimization to only attempt lock + * granting on the resource (that could be CPU-expensive) + * after we are done cancelling lock in that resource. */ if (res != pres) { if (pres != NULL) { ldlm_reprocess_all(pres); @@ -1550,10 +1690,20 @@ int ldlm_request_cancel(struct ptlrpc_request *req, if (res != NULL) { ldlm_resource_getref(res); LDLM_RESOURCE_ADDREF(res); - ldlm_res_lvbo_update(res, NULL, 1); + + if (!ldlm_is_discard_data(lock)) + ldlm_lvbo_update(res, lock, NULL, 1); } pres = res; } + + if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock)) { + time64_t delay = ktime_get_real_seconds() - + lock->l_last_activity; + LDLM_DEBUG(lock, "server cancels blocked lock after %llds", + (s64)delay); + at_measured(&lock->l_export->exp_bl_lock_at, delay); + } ldlm_lock_cancel(lock); LDLM_LOCK_PUT(lock); } @@ -1567,6 +1717,11 @@ int ldlm_request_cancel(struct ptlrpc_request *req, } EXPORT_SYMBOL(ldlm_request_cancel); +/** + * Main LDLM entry point for server code to cancel locks. + * + * Typically gets called from service handler on LDLM_CANCEL opc. + */ int ldlm_handle_cancel(struct ptlrpc_request *req) { struct ldlm_request *dlm_req; @@ -1588,14 +1743,18 @@ int ldlm_handle_cancel(struct ptlrpc_request *req) if (rc) RETURN(rc); - if (!ldlm_request_cancel(req, dlm_req, 0)) - req->rq_status = ESTALE; + if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS)) + req->rq_status = LUSTRE_ESTALE; RETURN(ptlrpc_reply(req)); } -EXPORT_SYMBOL(ldlm_handle_cancel); #endif /* HAVE_SERVER_SUPPORT */ +/** + * Callback handler for receiving incoming blocking ASTs. + * + * This can only happen on client side. + */ void ldlm_handle_bl_callback(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld, struct ldlm_lock *lock) { @@ -1605,10 +1764,10 @@ void ldlm_handle_bl_callback(struct ldlm_namespace *ns, LDLM_DEBUG(lock, "client blocking AST callback handler"); lock_res_and_lock(lock); - lock->l_flags |= LDLM_FL_CBPENDING; + ldlm_set_cbpending(lock); - if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) - lock->l_flags |= LDLM_FL_CANCEL; + if (ldlm_is_cancel_on_block(lock)) + ldlm_set_cancel(lock); do_ast = (!lock->l_readers && !lock->l_writers); unlock_res_and_lock(lock); @@ -1629,117 +1788,110 @@ void ldlm_handle_bl_callback(struct ldlm_namespace *ns, EXIT; } +/** + * Callback handler for receiving incoming completion ASTs. + * + * This only can happen on client side. + */ static void ldlm_handle_cp_callback(struct ptlrpc_request *req, struct ldlm_namespace *ns, struct ldlm_request *dlm_req, struct ldlm_lock *lock) { + struct list_head ast_list; int lvb_len; - CFS_LIST_HEAD(ast_list); - ENTRY; + int rc = 0; + ENTRY; - LDLM_DEBUG(lock, "client completion callback handler START"); + LDLM_DEBUG(lock, "client completion callback handler START"); - if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) { - int to = cfs_time_seconds(1); - while (to > 0) { - cfs_schedule_timeout_and_set_state( - CFS_TASK_INTERRUPTIBLE, to); - if (lock->l_granted_mode == lock->l_req_mode || - lock->l_destroyed) - break; - } - } + INIT_LIST_HEAD(&ast_list); + if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) { + int to = cfs_time_seconds(1); + while (to > 0) { + set_current_state(TASK_INTERRUPTIBLE); + schedule_timeout(to); + if (lock->l_granted_mode == lock->l_req_mode || + ldlm_is_destroyed(lock)) + break; + } + } lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT); - if (lvb_len > 0) { + if (lvb_len < 0) { + LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len); + GOTO(out, rc = lvb_len); + } else if (lvb_len > 0) { if (lock->l_lvb_len > 0) { /* for extent lock, lvb contains ost_lvb{}. */ LASSERT(lock->l_lvb_data != NULL); - LASSERTF(lock->l_lvb_len == lvb_len, - "preallocated %d, actual %d.\n", - lock->l_lvb_len, lvb_len); - } else { /* for layout lock, lvb has variable length */ - void *lvb_data; - OBD_ALLOC(lvb_data, lvb_len); - if (lvb_data == NULL) - LDLM_ERROR(lock, "no memory.\n"); - - lock_res_and_lock(lock); - if (lvb_data == NULL) { - lock->l_flags |= LDLM_FL_FAILED; - } else { - LASSERT(lock->l_lvb_data == NULL); - lock->l_lvb_data = lvb_data; - lock->l_lvb_len = lvb_len; + if (unlikely(lock->l_lvb_len < lvb_len)) { + LDLM_ERROR(lock, "Replied LVB is larger than " + "expectation, expected = %d, " + "replied = %d", + lock->l_lvb_len, lvb_len); + GOTO(out, rc = -EINVAL); } - unlock_res_and_lock(lock); } } - lock_res_and_lock(lock); - if (lock->l_destroyed || - lock->l_granted_mode == lock->l_req_mode) { - /* bug 11300: the lock has already been granted */ - unlock_res_and_lock(lock); - LDLM_DEBUG(lock, "Double grant race happened"); - LDLM_LOCK_RELEASE(lock); - EXIT; - return; - } + lock_res_and_lock(lock); + if (ldlm_is_destroyed(lock) || + lock->l_granted_mode == lock->l_req_mode) { + /* bug 11300: the lock has already been granted */ + unlock_res_and_lock(lock); + LDLM_DEBUG(lock, "Double grant race happened"); + GOTO(out, rc = 0); + } - /* If we receive the completion AST before the actual enqueue returned, - * then we might need to switch lock modes, resources, or extents. */ - if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) { - lock->l_req_mode = dlm_req->lock_desc.l_granted_mode; - LDLM_DEBUG(lock, "completion AST, new lock mode"); - } + /* If we receive the completion AST before the actual enqueue returned, + * then we might need to switch lock modes, resources, or extents. */ + if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) { + lock->l_req_mode = dlm_req->lock_desc.l_granted_mode; + LDLM_DEBUG(lock, "completion AST, new lock mode"); + } - if (lock->l_resource->lr_type != LDLM_PLAIN) { - ldlm_convert_policy_to_local(req->rq_export, - dlm_req->lock_desc.l_resource.lr_type, - &dlm_req->lock_desc.l_policy_data, - &lock->l_policy_data); - LDLM_DEBUG(lock, "completion AST, new policy data"); - } + if (lock->l_resource->lr_type != LDLM_PLAIN) { + ldlm_convert_policy_to_local(req->rq_export, + dlm_req->lock_desc.l_resource.lr_type, + &dlm_req->lock_desc.l_policy_data, + &lock->l_policy_data); + LDLM_DEBUG(lock, "completion AST, new policy data"); + } ldlm_resource_unlink_lock(lock); if (memcmp(&dlm_req->lock_desc.l_resource.lr_name, &lock->l_resource->lr_name, sizeof(lock->l_resource->lr_name)) != 0) { unlock_res_and_lock(lock); - if (ldlm_lock_change_resource(ns, lock, - &dlm_req->lock_desc.l_resource.lr_name) != 0) { - LDLM_ERROR(lock, "Failed to allocate resource"); - LDLM_LOCK_RELEASE(lock); - EXIT; - return; - } + rc = ldlm_lock_change_resource(ns, lock, + &dlm_req->lock_desc.l_resource.lr_name); + if (rc < 0) { + LDLM_ERROR(lock, "Failed to allocate resource"); + GOTO(out, rc); + } LDLM_DEBUG(lock, "completion AST, new resource"); CERROR("change resource!\n"); lock_res_and_lock(lock); } if (dlm_req->lock_flags & LDLM_FL_AST_SENT) { - /* BL_AST locks are not needed in lru. - * let ldlm_cancel_lru() be fast. */ + /* BL_AST locks are not needed in LRU. + * Let ldlm_cancel_lru() be fast. */ ldlm_lock_remove_from_lru(lock); - lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST; + lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST; LDLM_DEBUG(lock, "completion AST includes blocking AST"); } - if (lock->l_lvb_len) { - if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, - RCL_CLIENT) < lock->l_lvb_len) { - LDLM_ERROR(lock, "completion AST did not contain " - "expected LVB!"); - } else { - void *lvb = req_capsule_client_get(&req->rq_pill, - &RMF_DLM_LVB); - memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len); - } - } + if (lock->l_lvb_len > 0) { + rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT, + lock->l_lvb_data, lvb_len); + if (rc < 0) { + unlock_res_and_lock(lock); + GOTO(out, rc); + } + } ldlm_grant_lock(lock, &ast_list); unlock_res_and_lock(lock); @@ -1754,10 +1906,25 @@ static void ldlm_handle_cp_callback(struct ptlrpc_request *req, LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)", lock); - LDLM_LOCK_RELEASE(lock); - EXIT; + GOTO(out, rc); + +out: + if (rc < 0) { + lock_res_and_lock(lock); + ldlm_set_failed(lock); + unlock_res_and_lock(lock); + wake_up(&lock->l_waitq); + } + LDLM_LOCK_RELEASE(lock); } +/** + * Callback handler for receiving incoming glimpse ASTs. + * + * This only can happen on client side. After handling the glimpse AST + * we also consider dropping the lock here if it is unused locally for a + * long time. + */ static void ldlm_handle_gl_callback(struct ptlrpc_request *req, struct ldlm_namespace *ns, struct ldlm_request *dlm_req, @@ -1781,9 +1948,9 @@ static void ldlm_handle_gl_callback(struct ptlrpc_request *req, lock_res_and_lock(lock); if (lock->l_granted_mode == LCK_PW && !lock->l_readers && !lock->l_writers && - cfs_time_after(cfs_time_current(), - cfs_time_add(lock->l_last_used, - cfs_time_seconds(10)))) { + ktime_after(ktime_get(), + ktime_add(lock->l_last_used, + ktime_set(10, 0)))) { unlock_res_and_lock(lock); if (ldlm_bl_to_thread_lock(ns, NULL, lock)) ldlm_handle_bl_callback(ns, NULL, lock); @@ -1810,107 +1977,118 @@ static int ldlm_callback_reply(struct ptlrpc_request *req, int rc) return ptlrpc_reply(req); } -#ifdef __KERNEL__ -static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi, int mode) +static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi, + enum ldlm_cancel_flags cancel_flags) { struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool; ENTRY; spin_lock(&blp->blp_lock); if (blwi->blwi_lock && - blwi->blwi_lock->l_flags & LDLM_FL_DISCARD_DATA) { + ldlm_is_discard_data(blwi->blwi_lock)) { /* add LDLM_FL_DISCARD_DATA requests to the priority list */ - cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list); + list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list); } else { /* other blocking callbacks are added to the regular list */ - cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_list); + list_add_tail(&blwi->blwi_entry, &blp->blp_list); } spin_unlock(&blp->blp_lock); - cfs_waitq_signal(&blp->blp_waitq); + wake_up(&blp->blp_waitq); - /* can not use blwi->blwi_mode as blwi could be already freed in - LDLM_ASYNC mode */ - if (mode == LDLM_SYNC) + /* can not check blwi->blwi_flags as blwi could be already freed in + LCF_ASYNC mode */ + if (!(cancel_flags & LCF_ASYNC)) wait_for_completion(&blwi->blwi_comp); RETURN(0); } static inline void init_blwi(struct ldlm_bl_work_item *blwi, - struct ldlm_namespace *ns, - struct ldlm_lock_desc *ld, - cfs_list_t *cancels, int count, - struct ldlm_lock *lock, - int mode) + struct ldlm_namespace *ns, + struct ldlm_lock_desc *ld, + struct list_head *cancels, int count, + struct ldlm_lock *lock, + enum ldlm_cancel_flags cancel_flags) { init_completion(&blwi->blwi_comp); - CFS_INIT_LIST_HEAD(&blwi->blwi_head); + INIT_LIST_HEAD(&blwi->blwi_head); - if (cfs_memory_pressure_get()) + if (memory_pressure_get()) blwi->blwi_mem_pressure = 1; blwi->blwi_ns = ns; - blwi->blwi_mode = mode; + blwi->blwi_flags = cancel_flags; if (ld != NULL) blwi->blwi_ld = *ld; if (count) { - cfs_list_add(&blwi->blwi_head, cancels); - cfs_list_del_init(cancels); + list_add(&blwi->blwi_head, cancels); + list_del_init(cancels); blwi->blwi_count = count; } else { blwi->blwi_lock = lock; } } +/** + * Queues a list of locks \a cancels containing \a count locks + * for later processing by a blocking thread. If \a count is zero, + * then the lock referenced as \a lock is queued instead. + * + * The blocking thread would then call ->l_blocking_ast callback in the lock. + * If list addition fails an error is returned and caller is supposed to + * call ->l_blocking_ast itself. + */ static int ldlm_bl_to_thread(struct ldlm_namespace *ns, - struct ldlm_lock_desc *ld, struct ldlm_lock *lock, - cfs_list_t *cancels, int count, int mode) + struct ldlm_lock_desc *ld, + struct ldlm_lock *lock, + struct list_head *cancels, int count, + enum ldlm_cancel_flags cancel_flags) { - ENTRY; + ENTRY; - if (cancels && count == 0) - RETURN(0); + if (cancels && count == 0) + RETURN(0); - if (mode == LDLM_SYNC) { - /* if it is synchronous call do minimum mem alloc, as it could - * be triggered from kernel shrinker - */ - struct ldlm_bl_work_item blwi; - memset(&blwi, 0, sizeof(blwi)); - init_blwi(&blwi, ns, ld, cancels, count, lock, LDLM_SYNC); - RETURN(__ldlm_bl_to_thread(&blwi, LDLM_SYNC)); - } else { - struct ldlm_bl_work_item *blwi; - OBD_ALLOC(blwi, sizeof(*blwi)); - if (blwi == NULL) - RETURN(-ENOMEM); - init_blwi(blwi, ns, ld, cancels, count, lock, LDLM_ASYNC); + if (cancel_flags & LCF_ASYNC) { + struct ldlm_bl_work_item *blwi; - RETURN(__ldlm_bl_to_thread(blwi, LDLM_ASYNC)); - } + OBD_ALLOC(blwi, sizeof(*blwi)); + if (blwi == NULL) + RETURN(-ENOMEM); + init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags); + + RETURN(__ldlm_bl_to_thread(blwi, cancel_flags)); + } else { + /* if it is synchronous call do minimum mem alloc, as it could + * be triggered from kernel shrinker + */ + struct ldlm_bl_work_item blwi; + + memset(&blwi, 0, sizeof(blwi)); + init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags); + RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags)); + } } -#endif int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld, - struct ldlm_lock *lock) + struct ldlm_lock *lock) { -#ifdef __KERNEL__ - RETURN(ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LDLM_ASYNC)); -#else - RETURN(-ENOSYS); -#endif + return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC); } int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld, - cfs_list_t *cancels, int count, int mode) + struct list_head *cancels, int count, + enum ldlm_cancel_flags cancel_flags) { -#ifdef __KERNEL__ - RETURN(ldlm_bl_to_thread(ns, ld, NULL, cancels, count, mode)); -#else - RETURN(-ENOSYS); -#endif + return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags); +} + +int ldlm_bl_thread_wakeup(void) +{ + wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq); + return 0; } /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */ @@ -1958,11 +2136,11 @@ static int ldlm_handle_setinfo(struct ptlrpc_request *req) } static inline void ldlm_callback_errmsg(struct ptlrpc_request *req, - const char *msg, int rc, - struct lustre_handle *handle) + const char *msg, int rc, + const struct lustre_handle *handle) { DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req, - "%s: [nid %s] [rc %d] [lock "LPX64"]", + "%s: [nid %s] [rc %d] [lock %#llx]", msg, libcfs_id2str(req->rq_peer), rc, handle ? handle->cookie : 0); if (req->rq_no_reply) @@ -2003,8 +2181,11 @@ static int ldlm_callback_handler(struct ptlrpc_request *req) switch (lustre_msg_get_opc(req->rq_reqmsg)) { case LDLM_BL_CALLBACK: - if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) + if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) { + if (cfs_fail_err) + ldlm_callback_reply(req, -(int)cfs_fail_err); RETURN(0); + } break; case LDLM_CP_CALLBACK: if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET)) @@ -2018,16 +2199,6 @@ static int ldlm_callback_handler(struct ptlrpc_request *req) rc = ldlm_handle_setinfo(req); ldlm_callback_reply(req, rc); RETURN(0); - case OBD_LOG_CANCEL: /* remove this eventually - for 1.4.0 compat */ - CERROR("shouldn't be handling OBD_LOG_CANCEL on DLM thread\n"); - req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL); - if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET)) - RETURN(0); - rc = llog_origin_handle_cancel(req); - if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP)) - RETURN(0); - ldlm_callback_reply(req, rc); - RETURN(0); case LLOG_ORIGIN_HANDLE_CREATE: req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE); if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET)) @@ -2081,14 +2252,14 @@ static int ldlm_callback_handler(struct ptlrpc_request *req) * which the server has already started a blocking callback on. */ if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) && lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) { - rc = ldlm_cli_cancel(&dlm_req->lock_handle[0]); + rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0); if (rc < 0) CERROR("ldlm_cli_cancel: %d\n", rc); } lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0); if (!lock) { - CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock " + CDEBUG(D_DLMTRACE, "callback on lock %#llx - lock " "disappeared\n", dlm_req->lock_handle[0].cookie); rc = ldlm_callback_reply(req, -EINVAL); ldlm_callback_errmsg(req, "Operate with invalid parameter", rc, @@ -2096,37 +2267,35 @@ static int ldlm_callback_handler(struct ptlrpc_request *req) RETURN(0); } - if ((lock->l_flags & LDLM_FL_FAIL_LOC) && + if (ldlm_is_fail_loc(lock) && lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE); /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */ lock_res_and_lock(lock); lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags & - LDLM_AST_FLAGS); - if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) { - /* If somebody cancels lock and cache is already dropped, - * or lock is failed before cp_ast received on client, - * we can tell the server we have no lock. Otherwise, we - * should send cancel after dropping the cache. */ - if (((lock->l_flags & LDLM_FL_CANCELING) && - (lock->l_flags & LDLM_FL_BL_DONE)) || - (lock->l_flags & LDLM_FL_FAILED)) { - LDLM_DEBUG(lock, "callback on lock " - LPX64" - lock disappeared\n", - dlm_req->lock_handle[0].cookie); - unlock_res_and_lock(lock); - LDLM_LOCK_RELEASE(lock); - rc = ldlm_callback_reply(req, -EINVAL); - ldlm_callback_errmsg(req, "Operate on stale lock", rc, - &dlm_req->lock_handle[0]); - RETURN(0); - } - /* BL_AST locks are not needed in lru. - * let ldlm_cancel_lru() be fast. */ - ldlm_lock_remove_from_lru(lock); - lock->l_flags |= LDLM_FL_BL_AST; - } + LDLM_FL_AST_MASK); + if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) { + /* If somebody cancels lock and cache is already dropped, + * or lock is failed before cp_ast received on client, + * we can tell the server we have no lock. Otherwise, we + * should send cancel after dropping the cache. */ + if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) || + ldlm_is_failed(lock)) { + LDLM_DEBUG(lock, "callback on lock %llx - lock disappeared", + dlm_req->lock_handle[0].cookie); + unlock_res_and_lock(lock); + LDLM_LOCK_RELEASE(lock); + rc = ldlm_callback_reply(req, -EINVAL); + ldlm_callback_errmsg(req, "Operate on stale lock", rc, + &dlm_req->lock_handle[0]); + RETURN(0); + } + /* BL_AST locks are not needed in LRU. + * Let ldlm_cancel_lru() be fast. */ + ldlm_lock_remove_from_lru(lock); + ldlm_set_bl_ast(lock); + } unlock_res_and_lock(lock); /* We want the ost thread to get this reply so that it can respond @@ -2142,7 +2311,7 @@ static int ldlm_callback_handler(struct ptlrpc_request *req) case LDLM_BL_CALLBACK: CDEBUG(D_INODE, "blocking ast\n"); req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK); - if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)) { + if (!ldlm_is_cancel_on_block(lock)) { rc = ldlm_callback_reply(req, 0); if (req->rq_no_reply || rc) ldlm_callback_errmsg(req, "Normal process", rc, @@ -2170,6 +2339,11 @@ static int ldlm_callback_handler(struct ptlrpc_request *req) } #ifdef HAVE_SERVER_SUPPORT +/** + * Main handler for canceld thread. + * + * Separated into its own thread to avoid deadlocks. + */ static int ldlm_cancel_handler(struct ptlrpc_request *req) { int rc; @@ -2186,7 +2360,7 @@ static int ldlm_cancel_handler(struct ptlrpc_request *req) struct ldlm_request *dlm_req; CERROR("%s from %s arrived at %lu with bad export cookie " - LPU64"\n", + "%llu\n", ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)), libcfs_nid2str(req->rq_peer.nid), req->rq_arrival_time.tv_sec, @@ -2211,21 +2385,13 @@ static int ldlm_cancel_handler(struct ptlrpc_request *req) req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL); CDEBUG(D_INODE, "cancel\n"); if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) || - CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND)) + CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) || + CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT)) RETURN(0); rc = ldlm_handle_cancel(req); if (rc) break; RETURN(0); - case OBD_LOG_CANCEL: - req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL); - if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET)) - RETURN(0); - rc = llog_origin_handle_cancel(req); - if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP)) - RETURN(0); - ldlm_callback_reply(req, rc); - RETURN(0); default: CERROR("invalid opcode %d\n", lustre_msg_get_opc(req->rq_reqmsg)); @@ -2254,7 +2420,7 @@ static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req, if (lustre_handle_equal(&dlm_req->lock_handle[i], &lockh)) { DEBUG_REQ(D_RPCTRACE, req, - "Prio raised by lock "LPX64".", lockh.cookie); + "Prio raised by lock %#llx.", lockh.cookie); rc = 1; break; @@ -2287,7 +2453,7 @@ static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req) if (lock == NULL) continue; - rc = !!(lock->l_flags & LDLM_FL_AST_SENT); + rc = ldlm_is_ast_sent(lock) ? 1 : 0; if (rc) LDLM_DEBUG(lock, "hpreq cancel lock"); LDLM_LOCK_PUT(lock); @@ -2321,11 +2487,11 @@ static int ldlm_hpreq_handler(struct ptlrpc_request *req) RETURN(0); } -int ldlm_revoke_lock_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd, - cfs_hlist_node_t *hnode, void *data) +static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd, + struct hlist_node *hnode, void *data) { - cfs_list_t *rpc_list = data; + struct list_head *rpc_list = data; struct ldlm_lock *lock = cfs_hash_object(hs, hnode); lock_res_and_lock(lock); @@ -2342,7 +2508,7 @@ int ldlm_revoke_lock_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd, return 0; } - if (lock->l_flags & LDLM_FL_AST_SENT) { + if (ldlm_is_ast_sent(lock)) { unlock_res_and_lock(lock); return 0; } @@ -2350,7 +2516,7 @@ int ldlm_revoke_lock_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd, LASSERT(lock->l_blocking_ast); LASSERT(!lock->l_blocking_lock); - lock->l_flags |= LDLM_FL_AST_SENT; + ldlm_set_ast_sent(lock); if (lock->l_export && lock->l_export->exp_lock_hash) { /* NB: it's safe to call cfs_hash_del() even lock isn't * in exp_lock_hash. */ @@ -2361,7 +2527,7 @@ int ldlm_revoke_lock_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd, &lock->l_remote_handle, &lock->l_exp_hash); } - cfs_list_add_tail(&lock->l_rk_ast, rpc_list); + list_add_tail(&lock->l_rk_ast, rpc_list); LDLM_LOCK_GET(lock); unlock_res_and_lock(lock); @@ -2370,51 +2536,70 @@ int ldlm_revoke_lock_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd, void ldlm_revoke_export_locks(struct obd_export *exp) { - cfs_list_t rpc_list; - ENTRY; + struct list_head rpc_list; + ENTRY; - CFS_INIT_LIST_HEAD(&rpc_list); - cfs_hash_for_each_empty(exp->exp_lock_hash, - ldlm_revoke_lock_cb, &rpc_list); - ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list, - LDLM_WORK_REVOKE_AST); + INIT_LIST_HEAD(&rpc_list); + cfs_hash_for_each_nolock(exp->exp_lock_hash, + ldlm_revoke_lock_cb, &rpc_list, 0); + ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list, + LDLM_WORK_REVOKE_AST); - EXIT; + EXIT; } EXPORT_SYMBOL(ldlm_revoke_export_locks); #endif /* HAVE_SERVER_SUPPORT */ -#ifdef __KERNEL__ -static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp) +static int ldlm_bl_get_work(struct ldlm_bl_pool *blp, + struct ldlm_bl_work_item **p_blwi, + struct obd_export **p_exp) { struct ldlm_bl_work_item *blwi = NULL; static unsigned int num_bl = 0; + static unsigned int num_stale; + int num_th = atomic_read(&blp->blp_num_threads); + + *p_exp = obd_stale_export_get(); spin_lock(&blp->blp_lock); - /* process a request from the blp_list at least every blp_num_threads */ - if (!cfs_list_empty(&blp->blp_list) && - (cfs_list_empty(&blp->blp_prio_list) || num_bl == 0)) - blwi = cfs_list_entry(blp->blp_list.next, - struct ldlm_bl_work_item, blwi_entry); - else - if (!cfs_list_empty(&blp->blp_prio_list)) - blwi = cfs_list_entry(blp->blp_prio_list.next, - struct ldlm_bl_work_item, - blwi_entry); - - if (blwi) { - if (++num_bl >= cfs_atomic_read(&blp->blp_num_threads)) - num_bl = 0; - cfs_list_del(&blwi->blwi_entry); - } + if (*p_exp != NULL) { + if (num_th == 1 || ++num_stale < num_th) { + spin_unlock(&blp->blp_lock); + return 1; + } else { + num_stale = 0; + } + } + + /* process a request from the blp_list at least every blp_num_threads */ + if (!list_empty(&blp->blp_list) && + (list_empty(&blp->blp_prio_list) || num_bl == 0)) + blwi = list_entry(blp->blp_list.next, + struct ldlm_bl_work_item, blwi_entry); + else + if (!list_empty(&blp->blp_prio_list)) + blwi = list_entry(blp->blp_prio_list.next, + struct ldlm_bl_work_item, + blwi_entry); + + if (blwi) { + if (++num_bl >= num_th) + num_bl = 0; + list_del(&blwi->blwi_entry); + } spin_unlock(&blp->blp_lock); + *p_blwi = blwi; - return blwi; + if (*p_exp != NULL && *p_blwi != NULL) { + obd_stale_export_put(*p_exp); + *p_exp = NULL; + } + + return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0; } /* This only contains temporary data until the thread starts */ struct ldlm_bl_thread_data { - char bltd_name[CFS_CURPROC_COMM_MAX]; struct ldlm_bl_pool *bltd_blp; struct completion bltd_comp; int bltd_num; @@ -2422,106 +2607,179 @@ struct ldlm_bl_thread_data { static int ldlm_bl_thread_main(void *arg); -static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp) +static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy) { struct ldlm_bl_thread_data bltd = { .bltd_blp = blp }; - int rc; + struct task_struct *task; init_completion(&bltd.bltd_comp); - rc = cfs_create_thread(ldlm_bl_thread_main, &bltd, 0); - if (rc < 0) { - CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %d\n", - cfs_atomic_read(&blp->blp_num_threads), rc); - return rc; + + bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads); + if (bltd.bltd_num >= blp->blp_max_threads) { + atomic_dec(&blp->blp_num_threads); + return 0; + } + + LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num); + if (check_busy && + atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) { + atomic_dec(&blp->blp_num_threads); + return 0; + } + + task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d", + bltd.bltd_num); + if (IS_ERR(task)) { + CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n", + bltd.bltd_num, PTR_ERR(task)); + atomic_dec(&blp->blp_num_threads); + return PTR_ERR(task); } wait_for_completion(&bltd.bltd_comp); return 0; } -static int ldlm_bl_thread_main(void *arg) +/* Not fatal if racy and have a few too many threads */ +static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp, + struct ldlm_bl_work_item *blwi) { - struct ldlm_bl_pool *blp; - ENTRY; + if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads) + return 0; - { - struct ldlm_bl_thread_data *bltd = arg; + if (atomic_read(&blp->blp_busy_threads) < + atomic_read(&blp->blp_num_threads)) + return 0; - blp = bltd->bltd_blp; + if (blwi != NULL && (blwi->blwi_ns == NULL || + blwi->blwi_mem_pressure)) + return 0; - bltd->bltd_num = - cfs_atomic_inc_return(&blp->blp_num_threads) - 1; - cfs_atomic_inc(&blp->blp_busy_threads); + return 1; +} - snprintf(bltd->bltd_name, sizeof(bltd->bltd_name) - 1, - "ldlm_bl_%02d", bltd->bltd_num); - cfs_daemonize(bltd->bltd_name); +static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp, + struct ldlm_bl_work_item *blwi) +{ + ENTRY; - complete(&bltd->bltd_comp); - /* cannot use bltd after this, it is only on caller's stack */ - } + if (blwi->blwi_ns == NULL) + /* added by ldlm_cleanup() */ + RETURN(LDLM_ITER_STOP); + + if (blwi->blwi_mem_pressure) + memory_pressure_set(); + + OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4); + + if (blwi->blwi_count) { + int count; + /* The special case when we cancel locks in lru + * asynchronously, we pass the list of locks here. + * Thus locks are marked LDLM_FL_CANCELING, but NOT + * canceled locally yet. */ + count = ldlm_cli_cancel_list_local(&blwi->blwi_head, + blwi->blwi_count, + LCF_BL_AST); + ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL, + blwi->blwi_flags); + } else { + ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld, + blwi->blwi_lock); + } + if (blwi->blwi_mem_pressure) + memory_pressure_clr(); - while (1) { - struct l_wait_info lwi = { 0 }; - struct ldlm_bl_work_item *blwi = NULL; - int busy; + if (blwi->blwi_flags & LCF_ASYNC) + OBD_FREE(blwi, sizeof(*blwi)); + else + complete(&blwi->blwi_comp); - blwi = ldlm_bl_get_work(blp); + RETURN(0); +} - if (blwi == NULL) { - cfs_atomic_dec(&blp->blp_busy_threads); - l_wait_event_exclusive(blp->blp_waitq, - (blwi = ldlm_bl_get_work(blp)) != NULL, - &lwi); - busy = cfs_atomic_inc_return(&blp->blp_busy_threads); - } else { - busy = cfs_atomic_read(&blp->blp_busy_threads); - } +/** + * Cancel stale locks on export. Cancel blocked locks first. + * If the given export has blocked locks, the next in the list may have + * them too, thus cancel not blocked locks only if the current export has + * no blocked locks. + **/ +static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp, + struct obd_export *exp) +{ + int num; + ENTRY; - if (blwi->blwi_ns == NULL) - /* added by ldlm_cleanup() */ - break; + OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4); - /* Not fatal if racy and have a few too many threads */ - if (unlikely(busy < blp->blp_max_threads && - busy >= cfs_atomic_read(&blp->blp_num_threads) && - !blwi->blwi_mem_pressure)) - /* discard the return value, we tried */ - ldlm_bl_thread_start(blp); - - if (blwi->blwi_mem_pressure) - cfs_memory_pressure_set(); - - if (blwi->blwi_count) { - int count; - /* The special case when we cancel locks in lru - * asynchronously, we pass the list of locks here. - * Thus locks are marked LDLM_FL_CANCELING, but NOT - * canceled locally yet. */ - count = ldlm_cli_cancel_list_local(&blwi->blwi_head, - blwi->blwi_count, - LCF_BL_AST); - ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL, 0); - } else { - ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld, - blwi->blwi_lock); - } - if (blwi->blwi_mem_pressure) - cfs_memory_pressure_clr(); + num = ldlm_export_cancel_blocked_locks(exp); + if (num == 0) + ldlm_export_cancel_locks(exp); - if (blwi->blwi_mode == LDLM_ASYNC) - OBD_FREE(blwi, sizeof(*blwi)); - else - complete(&blwi->blwi_comp); - } + obd_stale_export_put(exp); + + RETURN(0); +} - cfs_atomic_dec(&blp->blp_busy_threads); - cfs_atomic_dec(&blp->blp_num_threads); + +/** + * Main blocking requests processing thread. + * + * Callers put locks into its queue by calling ldlm_bl_to_thread. + * This thread in the end ends up doing actual call to ->l_blocking_ast + * for queued locks. + */ +static int ldlm_bl_thread_main(void *arg) +{ + struct ldlm_bl_pool *blp; + struct ldlm_bl_thread_data *bltd = arg; + ENTRY; + + blp = bltd->bltd_blp; + + complete(&bltd->bltd_comp); + /* cannot use bltd after this, it is only on caller's stack */ + + while (1) { + struct l_wait_info lwi = { 0 }; + struct ldlm_bl_work_item *blwi = NULL; + struct obd_export *exp = NULL; + int rc; + + rc = ldlm_bl_get_work(blp, &blwi, &exp); + + if (rc == 0) + l_wait_event_exclusive(blp->blp_waitq, + ldlm_bl_get_work(blp, &blwi, + &exp), + &lwi); + atomic_inc(&blp->blp_busy_threads); + + if (ldlm_bl_thread_need_create(blp, blwi)) + /* discard the return value, we tried */ + ldlm_bl_thread_start(blp, true); + + if (exp) + rc = ldlm_bl_thread_exports(blp, exp); + else if (blwi) + rc = ldlm_bl_thread_blwi(blp, blwi); + + atomic_dec(&blp->blp_busy_threads); + + if (rc == LDLM_ITER_STOP) + break; + + /* If there are many namespaces, we will not sleep waiting for + * work, and must do a cond_resched to avoid holding the CPU + * for too long */ + cond_resched(); + } + + atomic_dec(&blp->blp_num_threads); complete(&blp->blp_comp); - RETURN(0); + RETURN(0); } -#endif static int ldlm_setup(void); static int ldlm_cleanup(void); @@ -2540,7 +2798,6 @@ int ldlm_get_ref(void) RETURN(rc); } -EXPORT_SYMBOL(ldlm_get_ref); void ldlm_put_ref(void) { @@ -2559,66 +2816,65 @@ void ldlm_put_ref(void) EXIT; } -EXPORT_SYMBOL(ldlm_put_ref); /* * Export handle<->lock hash operations. */ static unsigned -ldlm_export_lock_hash(cfs_hash_t *hs, const void *key, unsigned mask) +ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned mask) { return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask); } static void * -ldlm_export_lock_key(cfs_hlist_node_t *hnode) +ldlm_export_lock_key(struct hlist_node *hnode) { struct ldlm_lock *lock; - lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash); + lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash); return &lock->l_remote_handle; } static void -ldlm_export_lock_keycpy(cfs_hlist_node_t *hnode, void *key) +ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key) { struct ldlm_lock *lock; - lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash); + lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash); lock->l_remote_handle = *(struct lustre_handle *)key; } static int -ldlm_export_lock_keycmp(const void *key, cfs_hlist_node_t *hnode) +ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode) { return lustre_handle_equal(ldlm_export_lock_key(hnode), key); } static void * -ldlm_export_lock_object(cfs_hlist_node_t *hnode) +ldlm_export_lock_object(struct hlist_node *hnode) { - return cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash); + return hlist_entry(hnode, struct ldlm_lock, l_exp_hash); } static void -ldlm_export_lock_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode) +ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode) { struct ldlm_lock *lock; - lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash); + lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash); LDLM_LOCK_GET(lock); } static void -ldlm_export_lock_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode) +ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode) { struct ldlm_lock *lock; - lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash); + lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash); LDLM_LOCK_RELEASE(lock); } -static cfs_hash_ops_t ldlm_export_lock_ops = { +static struct cfs_hash_ops ldlm_export_lock_ops = { .hs_hash = ldlm_export_lock_hash, .hs_key = ldlm_export_lock_key, .hs_keycmp = ldlm_export_lock_keycmp, @@ -2631,6 +2887,7 @@ static cfs_hash_ops_t ldlm_export_lock_ops = { int ldlm_init_export(struct obd_export *exp) { + int rc; ENTRY; exp->exp_lock_hash = @@ -2646,7 +2903,14 @@ int ldlm_init_export(struct obd_export *exp) if (!exp->exp_lock_hash) RETURN(-ENOMEM); + rc = ldlm_init_flock_export(exp); + if (rc) + GOTO(err, rc); + RETURN(0); +err: + ldlm_destroy_export(exp); + RETURN(rc); } EXPORT_SYMBOL(ldlm_init_export); @@ -2661,14 +2925,50 @@ void ldlm_destroy_export(struct obd_export *exp) } EXPORT_SYMBOL(ldlm_destroy_export); +static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj, + struct attribute *attr, + char *buf) +{ + return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay); +} + +static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj, + struct attribute *attr, + const char *buffer, + size_t count) +{ + int rc; + unsigned long val; + + rc = kstrtoul(buffer, 10, &val); + if (rc) + return rc; + + ldlm_cancel_unused_locks_before_replay = val; + + return count; +} +LUSTRE_RW_ATTR(cancel_unused_locks_before_replay); + +static struct attribute *ldlm_attrs[] = { + &lustre_attr_cancel_unused_locks_before_replay.attr, + NULL, +}; + +static struct attribute_group ldlm_attr_group = { + .attrs = ldlm_attrs, +}; + static int ldlm_setup(void) { static struct ptlrpc_service_conf conf; - struct ldlm_bl_pool *blp = NULL; - int rc = 0; -#ifdef __KERNEL__ - int i; -#endif + struct ldlm_bl_pool *blp = NULL; +#ifdef HAVE_SERVER_SUPPORT + struct task_struct *task; +#endif /* HAVE_SERVER_SUPPORT */ + int i; + int rc = 0; + ENTRY; if (ldlm_state != NULL) @@ -2678,18 +2978,34 @@ static int ldlm_setup(void) if (ldlm_state == NULL) RETURN(-ENOMEM); -#ifdef LPROCFS - rc = ldlm_proc_setup(); - if (rc != 0) + ldlm_kobj = kobject_create_and_add("ldlm", lustre_kobj); + if (!ldlm_kobj) + GOTO(out, -ENOMEM); + + rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group); + if (rc) GOTO(out, rc); -#endif + + ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj); + if (!ldlm_ns_kset) + GOTO(out, -ENOMEM); + + ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj); + if (!ldlm_svc_kset) + GOTO(out, -ENOMEM); + +#ifdef CONFIG_PROC_FS + rc = ldlm_proc_setup(); + if (rc != 0) + GOTO(out, rc); +#endif /* CONFIG_PROC_FS */ memset(&conf, 0, sizeof(conf)); conf = (typeof(conf)) { .psc_name = "ldlm_cbd", .psc_watchdog_factor = 2, .psc_buf = { - .bc_nbufs = LDLM_NBUFS, + .bc_nbufs = LDLM_CLIENT_NBUFS, .bc_buf_size = LDLM_BUFSIZE, .bc_req_max_size = LDLM_MAXREQSIZE, .bc_rep_max_size = LDLM_MAXREPSIZE, @@ -2714,7 +3030,8 @@ static int ldlm_setup(void) }, }; ldlm_state->ldlm_cb_service = \ - ptlrpc_register_service(&conf, ldlm_svc_proc_dir); + ptlrpc_register_service(&conf, ldlm_svc_kset, + ldlm_svc_proc_dir); if (IS_ERR(ldlm_state->ldlm_cb_service)) { CERROR("failed to start service\n"); rc = PTR_ERR(ldlm_state->ldlm_cb_service); @@ -2728,7 +3045,7 @@ static int ldlm_setup(void) .psc_name = "ldlm_canceld", .psc_watchdog_factor = 6, .psc_buf = { - .bc_nbufs = LDLM_NBUFS, + .bc_nbufs = LDLM_SERVER_NBUFS, .bc_buf_size = LDLM_BUFSIZE, .bc_req_max_size = LDLM_MAXREQSIZE, .bc_rep_max_size = LDLM_MAXREPSIZE, @@ -2757,14 +3074,15 @@ static int ldlm_setup(void) }, }; ldlm_state->ldlm_cancel_service = \ - ptlrpc_register_service(&conf, ldlm_svc_proc_dir); + ptlrpc_register_service(&conf, ldlm_svc_kset, + ldlm_svc_proc_dir); if (IS_ERR(ldlm_state->ldlm_cancel_service)) { CERROR("failed to start service\n"); rc = PTR_ERR(ldlm_state->ldlm_cancel_service); ldlm_state->ldlm_cancel_service = NULL; GOTO(out, rc); } -#endif +#endif /* HAVE_SERVER_SUPPORT */ OBD_ALLOC(blp, sizeof(*blp)); if (blp == NULL) @@ -2772,13 +3090,12 @@ static int ldlm_setup(void) ldlm_state->ldlm_bl_pool = blp; spin_lock_init(&blp->blp_lock); - CFS_INIT_LIST_HEAD(&blp->blp_list); - CFS_INIT_LIST_HEAD(&blp->blp_prio_list); - cfs_waitq_init(&blp->blp_waitq); - cfs_atomic_set(&blp->blp_num_threads, 0); - cfs_atomic_set(&blp->blp_busy_threads, 0); + INIT_LIST_HEAD(&blp->blp_list); + INIT_LIST_HEAD(&blp->blp_prio_list); + init_waitqueue_head(&blp->blp_waitq); + atomic_set(&blp->blp_num_threads, 0); + atomic_set(&blp->blp_busy_threads, 0); -#ifdef __KERNEL__ if (ldlm_num_threads == 0) { blp->blp_min_threads = LDLM_NTHRS_INIT; blp->blp_max_threads = LDLM_NTHRS_MAX; @@ -2789,36 +3106,34 @@ static int ldlm_setup(void) } for (i = 0; i < blp->blp_min_threads; i++) { - rc = ldlm_bl_thread_start(blp); + rc = ldlm_bl_thread_start(blp, false); if (rc < 0) GOTO(out, rc); } -# ifdef HAVE_SERVER_SUPPORT - CFS_INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks); - expired_lock_thread.elt_state = ELT_STOPPED; - cfs_waitq_init(&expired_lock_thread.elt_waitq); - - CFS_INIT_LIST_HEAD(&waiting_locks_list); - spin_lock_init(&waiting_locks_spinlock); - cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0); - - rc = cfs_create_thread(expired_lock_main, NULL, CFS_DAEMON_FLAGS); - if (rc < 0) { +#ifdef HAVE_SERVER_SUPPORT + task = kthread_run(expired_lock_main, NULL, "ldlm_elt"); + if (IS_ERR(task)) { + rc = PTR_ERR(task); CERROR("Cannot start ldlm expired-lock thread: %d\n", rc); GOTO(out, rc); } - cfs_wait_event(expired_lock_thread.elt_waitq, - expired_lock_thread.elt_state == ELT_READY); -# endif /* HAVE_SERVER_SUPPORT */ + wait_event(expired_lock_wait_queue, + expired_lock_thread_state == ELT_READY); +#endif /* HAVE_SERVER_SUPPORT */ rc = ldlm_pools_init(); if (rc) { CERROR("Failed to initialize LDLM pools: %d\n", rc); GOTO(out, rc); } -#endif + + rc = ldlm_reclaim_setup(); + if (rc) { + CERROR("Failed to setup reclaim thread: rc = %d\n", rc); + GOTO(out, rc); + } RETURN(0); out: @@ -2830,28 +3145,28 @@ static int ldlm_cleanup(void) { ENTRY; - if (!cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) || - !cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) { + if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) || + !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) { CERROR("ldlm still has namespaces; clean these up first.\n"); ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE); ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE); RETURN(-EBUSY); } -#ifdef __KERNEL__ - ldlm_pools_fini(); + ldlm_reclaim_cleanup(); + ldlm_pools_fini(); if (ldlm_state->ldlm_bl_pool != NULL) { struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool; - while (cfs_atomic_read(&blp->blp_num_threads) > 0) { + while (atomic_read(&blp->blp_num_threads) > 0) { struct ldlm_bl_work_item blwi = { .blwi_ns = NULL }; init_completion(&blp->blp_comp); spin_lock(&blp->blp_lock); - cfs_list_add_tail(&blwi.blwi_entry, &blp->blp_list); - cfs_waitq_signal(&blp->blp_waitq); + list_add_tail(&blwi.blwi_entry, &blp->blp_list); + wake_up(&blp->blp_waitq); spin_unlock(&blp->blp_lock); wait_for_completion(&blp->blp_comp); @@ -2859,27 +3174,31 @@ static int ldlm_cleanup(void) OBD_FREE(blp, sizeof(*blp)); } -#endif /* __KERNEL__ */ if (ldlm_state->ldlm_cb_service != NULL) ptlrpc_unregister_service(ldlm_state->ldlm_cb_service); -# ifdef HAVE_SERVER_SUPPORT +#ifdef HAVE_SERVER_SUPPORT if (ldlm_state->ldlm_cancel_service != NULL) ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service); -# endif +#endif + + if (ldlm_ns_kset) + kset_unregister(ldlm_ns_kset); + if (ldlm_svc_kset) + kset_unregister(ldlm_svc_kset); + if (ldlm_kobj) + kobject_put(ldlm_kobj); -#ifdef __KERNEL__ ldlm_proc_cleanup(); -# ifdef HAVE_SERVER_SUPPORT - if (expired_lock_thread.elt_state != ELT_STOPPED) { - expired_lock_thread.elt_state = ELT_TERMINATE; - cfs_waitq_signal(&expired_lock_thread.elt_waitq); - cfs_wait_event(expired_lock_thread.elt_waitq, - expired_lock_thread.elt_state == ELT_STOPPED); +#ifdef HAVE_SERVER_SUPPORT + if (expired_lock_thread_state != ELT_STOPPED) { + expired_lock_thread_state = ELT_TERMINATE; + wake_up(&expired_lock_wait_queue); + wait_event(expired_lock_wait_queue, + expired_lock_thread_state == ELT_STOPPED); } -# endif -#endif /* __KERNEL__ */ +#endif OBD_FREE(ldlm_state, sizeof(*ldlm_state)); ldlm_state = NULL; @@ -2889,52 +3208,55 @@ static int ldlm_cleanup(void) int ldlm_init(void) { - mutex_init(&ldlm_ref_mutex); - mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER)); - mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT)); - ldlm_resource_slab = cfs_mem_cache_create("ldlm_resources", - sizeof(struct ldlm_resource), 0, - CFS_SLAB_HWCACHE_ALIGN); - if (ldlm_resource_slab == NULL) - return -ENOMEM; + ldlm_resource_slab = kmem_cache_create("ldlm_resources", + sizeof(struct ldlm_resource), 0, + SLAB_HWCACHE_ALIGN, NULL); + if (ldlm_resource_slab == NULL) + return -ENOMEM; - ldlm_lock_slab = cfs_mem_cache_create("ldlm_locks", + ldlm_lock_slab = kmem_cache_create("ldlm_locks", sizeof(struct ldlm_lock), 0, - CFS_SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU); - if (ldlm_lock_slab == NULL) { - cfs_mem_cache_destroy(ldlm_resource_slab); - return -ENOMEM; - } + SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU, NULL); + if (ldlm_lock_slab == NULL) + goto out_resource; - ldlm_interval_slab = cfs_mem_cache_create("interval_node", + ldlm_interval_slab = kmem_cache_create("interval_node", sizeof(struct ldlm_interval), - 0, CFS_SLAB_HWCACHE_ALIGN); - if (ldlm_interval_slab == NULL) { - cfs_mem_cache_destroy(ldlm_resource_slab); - cfs_mem_cache_destroy(ldlm_lock_slab); - return -ENOMEM; - } + 0, SLAB_HWCACHE_ALIGN, NULL); + if (ldlm_interval_slab == NULL) + goto out_lock; + + ldlm_interval_tree_slab = kmem_cache_create("interval_tree", + sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM, + 0, SLAB_HWCACHE_ALIGN, NULL); + if (ldlm_interval_tree_slab == NULL) + goto out_interval; + #if LUSTRE_TRACKS_LOCK_EXP_REFS - class_export_dump_hook = ldlm_dump_export_locks; + class_export_dump_hook = ldlm_dump_export_locks; #endif - return 0; + return 0; + +out_interval: + kmem_cache_destroy(ldlm_interval_slab); +out_lock: + kmem_cache_destroy(ldlm_lock_slab); +out_resource: + kmem_cache_destroy(ldlm_resource_slab); + + return -ENOMEM; } void ldlm_exit(void) { - int rc; - if (ldlm_refcount) - CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount); - rc = cfs_mem_cache_destroy(ldlm_resource_slab); - LASSERTF(rc == 0, "couldn't free ldlm resource slab\n"); -#ifdef __KERNEL__ - /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call - * synchronize_rcu() to wait a grace period elapsed, so that - * ldlm_lock_free() get a chance to be called. */ - synchronize_rcu(); -#endif - rc = cfs_mem_cache_destroy(ldlm_lock_slab); - LASSERTF(rc == 0, "couldn't free ldlm lock slab\n"); - rc = cfs_mem_cache_destroy(ldlm_interval_slab); - LASSERTF(rc == 0, "couldn't free interval node slab\n"); + if (ldlm_refcount) + CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount); + kmem_cache_destroy(ldlm_resource_slab); + /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call + * synchronize_rcu() to wait a grace period elapsed, so that + * ldlm_lock_free() get a chance to be called. */ + synchronize_rcu(); + kmem_cache_destroy(ldlm_lock_slab); + kmem_cache_destroy(ldlm_interval_slab); + kmem_cache_destroy(ldlm_interval_tree_slab); }