X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_lockd.c;h=5d1bc0d2f119ee417e33d96563d410158dac2a55;hp=d009a1fe7a88afb10846e08930ebeca43938117c;hb=175ce732b881e99d77d3f5c11fd1d0d5f604218a;hpb=0bd27be7f20a671e7128f341a070838a2bd318dc diff --git a/lustre/ldlm/ldlm_lockd.c b/lustre/ldlm/ldlm_lockd.c index d009a1f..5d1bc0d 100644 --- a/lustre/ldlm/ldlm_lockd.c +++ b/lustre/ldlm/ldlm_lockd.c @@ -1,6 +1,4 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: - * +/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. @@ -29,8 +27,7 @@ * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011 Whamcloud, Inc. - * + * Copyright (c) 2010, 2013, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -42,31 +39,25 @@ * Author: Phil Schwan */ -#ifndef EXPORT_SYMTAB -# define EXPORT_SYMTAB -#endif #define DEBUG_SUBSYSTEM S_LDLM -#ifdef __KERNEL__ -# include -#else -# include -#endif - +#include #include #include #include #include "ldlm_internal.h" -#ifdef __KERNEL__ static int ldlm_num_threads; CFS_MODULE_PARM(ldlm_num_threads, "i", int, 0444, "number of DLM service threads to start"); -#endif -extern cfs_mem_cache_t *ldlm_resource_slab; -extern cfs_mem_cache_t *ldlm_lock_slab; -static cfs_semaphore_t ldlm_ref_sem; +static char *ldlm_cpts; +CFS_MODULE_PARM(ldlm_cpts, "s", charp, 0444, + "CPU partitions ldlm threads should run on"); + +extern struct kmem_cache *ldlm_resource_slab; +extern struct kmem_cache *ldlm_lock_slab; +static struct mutex ldlm_ref_mutex; static int ldlm_refcount; struct ldlm_cb_async_args { @@ -92,163 +83,188 @@ static inline unsigned int ldlm_get_rq_timeout(void) return timeout < 1 ? 1 : timeout; } -#ifdef __KERNEL__ -/* w_l_spinlock protects both waiting_locks_list and expired_lock_thread */ -static cfs_spinlock_t waiting_locks_spinlock; /* BH lock (timer) */ -static cfs_list_t waiting_locks_list; -static cfs_timer_t waiting_locks_timer; - -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; -#endif - #define ELT_STOPPED 0 #define ELT_READY 1 #define ELT_TERMINATE 2 struct ldlm_bl_pool { - cfs_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; - cfs_completion_t blp_comp; - cfs_atomic_t blp_num_threads; - cfs_atomic_t blp_busy_threads; - int blp_min_threads; - int blp_max_threads; + 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 + */ + 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 list_head blwi_entry; struct ldlm_namespace *blwi_ns; struct ldlm_lock_desc blwi_ld; struct ldlm_lock *blwi_lock; - cfs_list_t blwi_head; + struct list_head blwi_head; int blwi_count; - cfs_completion_t blwi_comp; - int blwi_mode; + struct completion blwi_comp; + ldlm_cancel_flags_t blwi_flags; int blwi_mem_pressure; }; -#ifdef __KERNEL__ +#ifdef HAVE_SERVER_SUPPORT + +/** + * Protects both waiting_locks_list and expired_lock_thread. + */ +static spinlock_t 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 struct list_head waiting_locks_list; +static struct timer_list waiting_locks_timer; + +static struct expired_lock_thread { + wait_queue_head_t elt_waitq; + int elt_state; + int elt_dump; + struct list_head elt_expired_locks; +} expired_lock_thread; static inline int have_expired_locks(void) { - int need_to_run; + int need_to_run; - ENTRY; - cfs_spin_lock_bh(&waiting_locks_spinlock); - need_to_run = !cfs_list_empty(&expired_lock_thread.elt_expired_locks); - cfs_spin_unlock_bh(&waiting_locks_spinlock); + ENTRY; + spin_lock_bh(&waiting_locks_spinlock); + need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks); + spin_unlock_bh(&waiting_locks_spinlock); - RETURN(need_to_run); + 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; - - ENTRY; - cfs_daemonize("ldlm_elt"); - - expired_lock_thread.elt_state = ELT_READY; - cfs_waitq_signal(&expired_lock_thread.elt_waitq); - - while (1) { - l_wait_event(expired_lock_thread.elt_waitq, - have_expired_locks() || - expired_lock_thread.elt_state == ELT_TERMINATE, - &lwi); - - cfs_spin_lock_bh(&waiting_locks_spinlock); - if (expired_lock_thread.elt_dump) { - cfs_spin_unlock_bh(&waiting_locks_spinlock); - - /* from waiting_locks_callback, but not in timer */ - libcfs_debug_dumplog(); - libcfs_run_lbug_upcall(__FILE__, - "waiting_locks_callback", - expired_lock_thread.elt_dump); - - cfs_spin_lock_bh(&waiting_locks_spinlock); - expired_lock_thread.elt_dump = 0; - } - - do_dump = 0; - - while (!cfs_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) { - cfs_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; - } - export = class_export_lock_get(lock->l_export, lock); - cfs_spin_unlock_bh(&waiting_locks_spinlock); - - do_dump++; - class_fail_export(export); - class_export_lock_put(export, lock); - - /* release extra ref grabbed by ldlm_add_waiting_lock() - * or ldlm_failed_ast() */ - LDLM_LOCK_RELEASE(lock); - - cfs_spin_lock_bh(&waiting_locks_spinlock); - } - cfs_spin_unlock_bh(&waiting_locks_spinlock); - - 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; - } - - expired_lock_thread.elt_state = ELT_STOPPED; - cfs_waitq_signal(&expired_lock_thread.elt_waitq); - RETURN(0); + struct list_head *expired = &expired_lock_thread.elt_expired_locks; + struct l_wait_info lwi = { 0 }; + int do_dump; + + ENTRY; + + expired_lock_thread.elt_state = ELT_READY; + wake_up(&expired_lock_thread.elt_waitq); + + while (1) { + l_wait_event(expired_lock_thread.elt_waitq, + have_expired_locks() || + expired_lock_thread.elt_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 }; + 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; + } + + do_dump = 0; + + while (!list_empty(expired)) { + struct obd_export *export; + struct ldlm_lock *lock; + + lock = list_entry(expired->next, struct ldlm_lock, + l_pending_chain); + if ((void *)lock < LP_POISON + PAGE_CACHE_SIZE && + (void *)lock >= LP_POISON) { + spin_unlock_bh(&waiting_locks_spinlock); + 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_CACHE_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 (ldlm_is_destroyed(lock)) { + /* release the lock refcount where + * waiting_locks_callback() founds */ + LDLM_LOCK_RELEASE(lock); + continue; + } + export = class_export_lock_get(lock->l_export, lock); + spin_unlock_bh(&waiting_locks_spinlock); + + do_dump++; + class_fail_export(export); + class_export_lock_put(export, lock); + + /* release extra ref grabbed by ldlm_add_waiting_lock() + * or ldlm_failed_ast() */ + LDLM_LOCK_RELEASE(lock); + + spin_lock_bh(&waiting_locks_spinlock); + } + spin_unlock_bh(&waiting_locks_spinlock); + + 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; + } + + expired_lock_thread.elt_state = ELT_STOPPED; + wake_up(&expired_lock_thread.elt_waitq); + RETURN(0); } static int ldlm_add_waiting_lock(struct ldlm_lock *lock); +static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds); /** * Check if there is a request in the export request list @@ -256,74 +272,41 @@ static int ldlm_add_waiting_lock(struct ldlm_lock *lock); */ static int ldlm_lock_busy(struct ldlm_lock *lock) { - struct ptlrpc_request *req; - int match = 0; - ENTRY; - - if (lock->l_export == NULL) - return 0; - - cfs_spin_lock_bh(&lock->l_export->exp_rpc_lock); - cfs_list_for_each_entry(req, &lock->l_export->exp_queued_rpc, - rq_exp_list) { - if (req->rq_ops->hpreq_lock_match) { - match = req->rq_ops->hpreq_lock_match(req, lock); - if (match) - break; - } - } - cfs_spin_unlock_bh(&lock->l_export->exp_rpc_lock); - RETURN(match); + struct ptlrpc_request *req; + int match = 0; + ENTRY; + + if (lock->l_export == NULL) + return 0; + + spin_lock_bh(&lock->l_export->exp_rpc_lock); + 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); + if (match) + break; + } + } + spin_unlock_bh(&lock->l_export->exp_rpc_lock); + RETURN(match); } /* This is called from within a timer interrupt and cannot schedule */ static void waiting_locks_callback(unsigned long unused) { - struct ldlm_lock *lock; + struct ldlm_lock *lock; + int need_dump = 0; -repeat: - cfs_spin_lock_bh(&waiting_locks_spinlock); - while (!cfs_list_empty(&waiting_locks_list)) { - lock = cfs_list_entry(waiting_locks_list.next, struct ldlm_lock, + spin_lock_bh(&waiting_locks_spinlock); + 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); - cfs_spin_unlock_bh(&waiting_locks_spinlock); - ldlm_add_waiting_lock(lock); - goto repeat; - } - - /* 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); - cfs_spin_unlock_bh(&waiting_locks_spinlock); - ldlm_add_waiting_lock(lock); - goto repeat; - } - /* Check if we need to prolong timeout */ if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) && ldlm_lock_busy(lock)) { @@ -334,11 +317,11 @@ repeat: LDLM_LOCK_GET(lock); - cfs_spin_unlock_bh(&waiting_locks_spinlock); - LDLM_DEBUG(lock, "prolong the busy lock"); - ldlm_refresh_waiting_lock(lock, - ldlm_get_enq_timeout(lock)); - cfs_spin_lock_bh(&waiting_locks_spinlock); + spin_unlock_bh(&waiting_locks_spinlock); + LDLM_DEBUG(lock, "prolong the busy lock"); + ldlm_refresh_waiting_lock(lock, + ldlm_get_enq_timeout(lock)); + spin_lock_bh(&waiting_locks_spinlock); if (!cont) { LDLM_LOCK_RELEASE(lock); @@ -351,40 +334,43 @@ repeat: ldlm_lock_to_ns(lock)->ns_timeouts++; LDLM_ERROR(lock, "lock callback timer expired after %lds: " "evicting client at %s ", - cfs_time_current_sec()- lock->l_last_activity, + cfs_time_current_sec() - 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, + list_del(&lock->l_pending_chain); + list_add(&lock->l_pending_chain, &expired_lock_thread.elt_expired_locks); - } + need_dump = 1; + } - if (!cfs_list_empty(&expired_lock_thread.elt_expired_locks)) { - if (obd_dump_on_timeout) - expired_lock_thread.elt_dump = __LINE__; + if (!list_empty(&expired_lock_thread.elt_expired_locks)) { + if (obd_dump_on_timeout && need_dump) + expired_lock_thread.elt_dump = __LINE__; - cfs_waitq_signal(&expired_lock_thread.elt_waitq); - } + wake_up(&expired_lock_thread.elt_waitq); + } /* * 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); } - cfs_spin_unlock_bh(&waiting_locks_spinlock); + 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 @@ -399,7 +385,7 @@ 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) || @@ -420,21 +406,25 @@ static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds) /* 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 int ldlm_add_waiting_lock(struct ldlm_lock *lock) { - int ret; - int timeout = ldlm_get_enq_timeout(lock); + int ret; + int timeout = ldlm_get_enq_timeout(lock); + + /* NB: must be called with hold of lock_res_and_lock() */ + LASSERT(ldlm_is_res_locked(lock)); + ldlm_set_waited(lock); - LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)); + LASSERT(!ldlm_is_cancel_on_block(lock)); - cfs_spin_lock_bh(&waiting_locks_spinlock); - if (lock->l_destroyed) { - static cfs_time_t next; - cfs_spin_unlock_bh(&waiting_locks_spinlock); + spin_lock_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); @@ -444,19 +434,28 @@ static int ldlm_add_waiting_lock(struct ldlm_lock *lock) } ret = __ldlm_add_waiting_lock(lock, timeout); - if (ret) + if (ret) { /* grab ref on the lock if it has been added to the * waiting list */ LDLM_LOCK_GET(lock); - cfs_spin_unlock_bh(&waiting_locks_spinlock); + } + spin_unlock_bh(&waiting_locks_spinlock); - LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)", - ret == 0 ? "not re-" : "", timeout, - AT_OFF ? "off" : "on"); - return ret; + if (ret) { + spin_lock_bh(&lock->l_export->exp_bl_list_lock); + if (list_empty(&lock->l_exp_list)) + list_add(&lock->l_exp_list, + &lock->l_export->exp_bl_list); + spin_unlock_bh(&lock->l_export->exp_bl_list_lock); + } + + LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)", + ret == 0 ? "not re-" : "", timeout, + AT_OFF ? "off" : "on"); + 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. @@ -467,9 +466,9 @@ 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; @@ -480,13 +479,13 @@ static int __ldlm_del_waiting_lock(struct ldlm_lock *lock) cfs_timer_disarm(&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)); } } - cfs_list_del_init(&lock->l_pending_chain); + list_del_init(&lock->l_pending_chain); return 1; } @@ -501,55 +500,59 @@ int ldlm_del_waiting_lock(struct ldlm_lock *lock) return 0; } - cfs_spin_lock_bh(&waiting_locks_spinlock); - ret = __ldlm_del_waiting_lock(lock); - cfs_spin_unlock_bh(&waiting_locks_spinlock); - if (ret) + spin_lock_bh(&waiting_locks_spinlock); + ret = __ldlm_del_waiting_lock(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); + list_del_init(&lock->l_exp_list); + spin_unlock_bh(&lock->l_export->exp_bl_list_lock); + + if (ret) { /* release lock ref if it has indeed been removed * from a list */ LDLM_LOCK_RELEASE(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. */ int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout) { - if (lock->l_export == NULL) { - /* We don't have a "waiting locks list" on clients. */ - LDLM_DEBUG(lock, "client lock: no-op"); - return 0; - } - - cfs_spin_lock_bh(&waiting_locks_spinlock); - - if (cfs_list_empty(&lock->l_pending_chain)) { - cfs_spin_unlock_bh(&waiting_locks_spinlock); - LDLM_DEBUG(lock, "wasn't waiting"); - return 0; - } - - /* we remove/add the lock to the waiting list, so no needs to - * release/take a lock reference */ - __ldlm_del_waiting_lock(lock); - __ldlm_add_waiting_lock(lock, timeout); - cfs_spin_unlock_bh(&waiting_locks_spinlock); - - LDLM_DEBUG(lock, "refreshed"); - return 1; + if (lock->l_export == NULL) { + /* We don't have a "waiting locks list" on clients. */ + LDLM_DEBUG(lock, "client lock: no-op"); + return 0; + } + + spin_lock_bh(&waiting_locks_spinlock); + + if (list_empty(&lock->l_pending_chain)) { + spin_unlock_bh(&waiting_locks_spinlock); + LDLM_DEBUG(lock, "wasn't waiting"); + return 0; + } + + /* we remove/add the lock to the waiting list, so no needs to + * release/take a lock reference */ + __ldlm_del_waiting_lock(lock); + __ldlm_add_waiting_lock(lock, timeout); + spin_unlock_bh(&waiting_locks_spinlock); + + LDLM_DEBUG(lock, "refreshed"); + return 1; } -#else /* !__KERNEL__ */ +EXPORT_SYMBOL(ldlm_refresh_waiting_lock); -static int ldlm_add_waiting_lock(struct ldlm_lock *lock) -{ - LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)); - RETURN(1); -} +#else /* HAVE_SERVER_SUPPORT */ int ldlm_del_waiting_lock(struct ldlm_lock *lock) { @@ -560,8 +563,14 @@ int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout) { RETURN(0); } -#endif /* __KERNEL__ */ +#endif /* !HAVE_SERVER_SUPPORT */ + +#ifdef HAVE_SERVER_SUPPORT + +/** + * Perform lock cleanup if AST sending failed. + */ static void ldlm_failed_ast(struct ldlm_lock *lock, int rc, const char *ast_type) { @@ -572,21 +581,20 @@ static void ldlm_failed_ast(struct ldlm_lock *lock, int rc, if (obd_dump_on_timeout) libcfs_debug_dumplog(); -#ifdef __KERNEL__ - cfs_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); - cfs_spin_unlock_bh(&waiting_locks_spinlock); -#else - class_fail_export(lock->l_export); -#endif + 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); + list_add(&lock->l_pending_chain, + &expired_lock_thread.elt_expired_locks); + wake_up(&expired_lock_thread.elt_waitq); + spin_unlock_bh(&waiting_locks_spinlock); } +/** + * 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) @@ -601,7 +609,7 @@ static int ldlm_handle_ast_error(struct ldlm_lock *lock, libcfs_nid2str(peer.nid)); ldlm_lock_cancel(lock); rc = -ERESTART; - } else if (lock->l_flags & LDLM_FL_CANCEL) { + } else if (ldlm_is_cancel(lock)) { LDLM_DEBUG(lock, "%s AST timeout from nid %s, but " "cancel was received (AST reply lost?)", ast_type, libcfs_nid2str(peer.nid)); @@ -653,40 +661,69 @@ static int ldlm_cb_interpret(const struct lu_env *env, ENTRY; LASSERT(lock != NULL); - if (rc != 0) { - rc = ldlm_handle_ast_error(lock, req, rc, - arg->type == LDLM_BL_CALLBACK - ? "blocking" : "completion"); - if (rc == -ERESTART) - cfs_atomic_inc(&arg->restart); - } + + switch (arg->type) { + case LDLM_GL_CALLBACK: + /* Update the LVB from disk if the AST failed + * (this is a legal race) + * + * - Glimpse callback of local lock just returns + * -ELDLM_NO_LOCK_DATA. + * - Glimpse callback of remote lock might return + * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274 + */ + 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); + } else if (rc != 0) { + rc = ldlm_handle_ast_error(lock, req, rc, "glimpse"); + } else { + rc = ldlm_res_lvbo_update(lock->l_resource, req, 1); + } + break; + case LDLM_BL_CALLBACK: + if (rc != 0) + rc = ldlm_handle_ast_error(lock, req, rc, "blocking"); + break; + case LDLM_CP_CALLBACK: + if (rc != 0) + rc = ldlm_handle_ast_error(lock, req, rc, "completion"); + break; + default: + LDLM_ERROR(lock, "invalid opcode for lock callback %d", + arg->type); + LBUG(); + } + + /* release extra reference taken in ldlm_ast_fini() */ LDLM_LOCK_RELEASE(lock); - if (cfs_atomic_dec_return(&arg->rpcs) < arg->threshold) - cfs_waitq_signal(&arg->waitq); - RETURN(0); + if (rc == -ERESTART) + atomic_inc(&arg->restart); + + RETURN(0); } -static inline int ldlm_bl_and_cp_ast_tail(struct ptlrpc_request *req, - struct ldlm_cb_set_arg *arg, - struct ldlm_lock *lock, - int instant_cancel) +static inline int ldlm_ast_fini(struct ptlrpc_request *req, + struct ldlm_cb_set_arg *arg, + struct ldlm_lock *lock, + int instant_cancel) { - int rc = 0; - ENTRY; - - if (unlikely(instant_cancel)) { - rc = ptl_send_rpc(req, 1); - ptlrpc_req_finished(req); - if (rc == 0) - cfs_atomic_inc(&arg->restart); - } else { - LDLM_LOCK_GET(lock); - ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1); - cfs_atomic_inc(&arg->rpcs); - } - - RETURN(rc); + int rc = 0; + ENTRY; + + if (unlikely(instant_cancel)) { + rc = ptl_send_rpc(req, 1); + ptlrpc_req_finished(req); + if (rc == 0) + atomic_inc(&arg->restart); + } else { + LDLM_LOCK_GET(lock); + ptlrpc_set_add_req(arg->set, req); + } + + RETURN(rc); } /** @@ -695,30 +732,37 @@ static inline int ldlm_bl_and_cp_ast_tail(struct ptlrpc_request *req, */ static void ldlm_lock_reorder_req(struct ldlm_lock *lock) { - struct ptlrpc_request *req; - ENTRY; - - if (lock->l_export == NULL) { - LDLM_DEBUG(lock, "client lock: no-op"); - RETURN_EXIT; - } - - cfs_spin_lock_bh(&lock->l_export->exp_rpc_lock); - cfs_list_for_each_entry(req, &lock->l_export->exp_queued_rpc, - rq_exp_list) { - if (!req->rq_hp && req->rq_ops->hpreq_lock_match && - req->rq_ops->hpreq_lock_match(req, lock)) - ptlrpc_hpreq_reorder(req); - } - cfs_spin_unlock_bh(&lock->l_export->exp_rpc_lock); - EXIT; + struct ptlrpc_request *req; + ENTRY; + + if (lock->l_export == NULL) { + LDLM_DEBUG(lock, "client lock: no-op"); + RETURN_EXIT; + } + + spin_lock_bh(&lock->l_export->exp_rpc_lock); + 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, @@ -739,10 +783,8 @@ int ldlm_server_blocking_ast(struct ldlm_lock *lock, LASSERT(lock); LASSERT(data != NULL); - if (lock->l_export->exp_obd->obd_recovering != 0) { + if (lock->l_export->exp_obd->obd_recovering != 0) LDLM_ERROR(lock, "BUG 6063: lock collide during recovery"); - ldlm_lock_dump(D_ERROR, lock, 0); - } ldlm_lock_reorder_req(lock); @@ -760,59 +802,70 @@ int ldlm_server_blocking_ast(struct ldlm_lock *lock, req->rq_interpret_reply = ldlm_cb_interpret; req->rq_no_resend = 1; - lock_res(lock->l_resource); - if (lock->l_granted_mode != lock->l_req_mode) { - /* this blocking AST will be communicated as part of the - * completion AST instead */ - unlock_res(lock->l_resource); - ptlrpc_req_finished(req); - LDLM_DEBUG(lock, "lock not granted, not sending blocking AST"); - RETURN(0); - } - - if (lock->l_destroyed) { - /* What's the point? */ - unlock_res(lock->l_resource); - ptlrpc_req_finished(req); - RETURN(0); - } - - if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) + 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 */ + unlock_res_and_lock(lock); + + ptlrpc_req_finished(req); + LDLM_DEBUG(lock, "lock not granted, not sending blocking AST"); + RETURN(0); + } + + if (ldlm_is_destroyed(lock)) { + /* What's the point? */ + unlock_res_and_lock(lock); + ptlrpc_req_finished(req); + RETURN(0); + } + + 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 |= (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); - if (instant_cancel) { - unlock_res(lock->l_resource); - ldlm_lock_cancel(lock); - } else { - LASSERT(lock->l_granted_mode == lock->l_req_mode); - ldlm_add_waiting_lock(lock); - unlock_res(lock->l_resource); - } + if (instant_cancel) { + unlock_res_and_lock(lock); + ldlm_lock_cancel(lock); + } else { + LASSERT(lock->l_granted_mode == lock->l_req_mode); + ldlm_add_waiting_lock(lock); + unlock_res_and_lock(lock); + } req->rq_send_state = LUSTRE_IMP_FULL; /* ptlrpc_request_alloc_pack already set timeout */ if (AT_OFF) req->rq_timeout = ldlm_get_rq_timeout(); + lock->l_last_activity = cfs_time_current_sec(); + 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, LDLM_BL_CALLBACK - LDLM_FIRST_OPC); - rc = ldlm_bl_and_cp_ast_tail(req, arg, lock, instant_cancel); + rc = ldlm_ast_fini(req, arg, lock, instant_cancel); RETURN(rc); } +EXPORT_SYMBOL(ldlm_server_blocking_ast); -int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data) +/** + * ->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; @@ -821,6 +874,7 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data) long total_enqueue_wait; int instant_cancel = 0; int rc = 0; + int lvb_len; ENTRY; LASSERT(lock != NULL); @@ -829,17 +883,26 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data) total_enqueue_wait = cfs_time_sub(cfs_time_current_sec(), lock->l_last_activity); + if (OBD_FAIL_PRECHECK(OBD_FAIL_OST_LDLM_REPLY_NET)) { + LDLM_DEBUG(lock, "dropping CP AST"); + RETURN(0); + } + req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse, &RQF_LDLM_CP_CALLBACK); if (req == NULL) RETURN(-ENOMEM); - /* server namespace, doesn't need lock */ - if (lock->l_resource->lr_lvb_len) { - req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, - lock->l_resource->lr_lvb_len); - } + /* server namespace, doesn't need lock */ + lvb_len = ldlm_lvbo_size(lock); + /* 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); @@ -856,20 +919,33 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data) body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); body->lock_handle[0] = lock->l_remote_handle; - body->lock_flags = flags; + body->lock_flags = ldlm_flags_to_wire(flags); ldlm_lock2desc(lock, &body->lock_desc); - if (lock->l_resource->lr_lvb_len) { - void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB); - - lock_res(lock->l_resource); - memcpy(lvb, lock->l_resource->lr_lvb_data, - lock->l_resource->lr_lvb_len); - unlock_res(lock->l_resource); + if (lvb_len > 0) { + void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB); + + lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len); + 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 = cfs_time_current_sec(); + /* 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)) @@ -894,10 +970,11 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, int 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) { - body->lock_flags |= LDLM_FL_AST_SENT; - /* copy ast flags like LDLM_FL_DISCARD_DATA */ - body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS); + 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. */ + body->lock_flags |= ldlm_flags_to_wire(lock->l_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 @@ -905,7 +982,7 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, int 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) { + if (ldlm_is_cancel_on_block(lock)) { unlock_res_and_lock(lock); ldlm_lock_cancel(lock); instant_cancel = 1; @@ -922,77 +999,114 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data) lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats, LDLM_CP_CALLBACK - LDLM_FIRST_OPC); - rc = ldlm_bl_and_cp_ast_tail(req, arg, lock, instant_cancel); + 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_resource *res = lock->l_resource; - struct ldlm_request *body; - struct ptlrpc_request *req; - int rc; + struct ldlm_cb_set_arg *arg = data; + struct ldlm_request *body; + struct ptlrpc_request *req; + struct ldlm_cb_async_args *ca; + int rc; + struct req_format *req_fmt; ENTRY; LASSERT(lock != NULL); + if (arg->gl_desc != NULL) + /* There is a glimpse descriptor to pack */ + req_fmt = &RQF_LDLM_GL_DESC_CALLBACK; + else + req_fmt = &RQF_LDLM_GL_CALLBACK; + req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse, - &RQF_LDLM_GL_CALLBACK, - LUSTRE_DLM_VERSION, LDLM_GL_CALLBACK); + req_fmt, LUSTRE_DLM_VERSION, + LDLM_GL_CALLBACK); if (req == NULL) RETURN(-ENOMEM); + if (arg->gl_desc != NULL) { + /* copy the GL descriptor */ + union ldlm_gl_desc *desc; + desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC); + *desc = *arg->gl_desc; + } + body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); body->lock_handle[0] = lock->l_remote_handle; ldlm_lock2desc(lock, &body->lock_desc); + CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args)); + ca = ptlrpc_req_async_args(req); + ca->ca_set_arg = arg; + ca->ca_lock = lock; + /* server namespace, doesn't need lock */ req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, - lock->l_resource->lr_lvb_len); - res = lock->l_resource; + ldlm_lvbo_size(lock)); ptlrpc_request_set_replen(req); - req->rq_send_state = LUSTRE_IMP_FULL; /* ptlrpc_request_alloc_pack already set timeout */ if (AT_OFF) req->rq_timeout = ldlm_get_rq_timeout(); + lock->l_last_activity = cfs_time_current_sec(); + + req->rq_interpret_reply = ldlm_cb_interpret; + 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, LDLM_GL_CALLBACK - LDLM_FIRST_OPC); - rc = ptlrpc_queue_wait(req); - /* Update the LVB from disk if the AST failed (this is a legal race) - * - * - Glimpse callback of local lock just return -ELDLM_NO_LOCK_DATA. - * - Glimpse callback of remote lock might return -ELDLM_NO_LOCK_DATA - * when inode is cleared. LU-274 - */ - if (rc == -ELDLM_NO_LOCK_DATA) { - LDLM_DEBUG(lock, "lost race - client has a lock but no inode"); - ldlm_res_lvbo_update(res, NULL, 1); - } else if (rc != 0) { - rc = ldlm_handle_ast_error(lock, req, rc, "glimpse"); - } else { - rc = ldlm_res_lvbo_update(res, req, 1); - } + rc = ldlm_ast_fini(req, arg, lock, 0); - ptlrpc_req_finished(req); - if (rc == -ERESTART) - ldlm_reprocess_all(res); + RETURN(rc); +} +EXPORT_SYMBOL(ldlm_server_glimpse_ast); - RETURN(rc); +int ldlm_glimpse_locks(struct ldlm_resource *res, + struct list_head *gl_work_list) +{ + int rc; + ENTRY; + + rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list, + LDLM_WORK_GL_AST); + if (rc == -ERESTART) + ldlm_reprocess_all(res); + + RETURN(rc); } +EXPORT_SYMBOL(ldlm_glimpse_locks); -#ifdef __KERNEL__ -extern unsigned long long lu_time_stamp_get(void); -#else -#define lu_time_stamp_get() time(NULL) -#endif +/* 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; + struct ldlm_lock *lock; + ENTRY; + + ca = ptlrpc_req_async_args(req); + lock = ca->ca_lock; + if (lock == NULL) + RETURN(ERR_PTR(-EFAULT)); + + RETURN(lock); +} +EXPORT_SYMBOL(ldlm_request_lock); static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req, struct lprocfs_stats *srv_stats) @@ -1028,9 +1142,9 @@ 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, @@ -1038,23 +1152,23 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns, const struct ldlm_callback_suite *cbs) { struct ldlm_reply *dlm_rep; - __u32 flags; + __u64 flags; ldlm_error_t 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_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF); - flags = dlm_req->lock_flags; + flags = ldlm_flags_from_wire(dlm_req->lock_flags); LASSERT(req->rq_export); - if (req->rq_rqbd->rqbd_service->srv_stats) - ldlm_svc_get_eopc(dlm_req, - req->rq_rqbd->rqbd_service->srv_stats); + if (ptlrpc_req2svc(req)->srv_stats != NULL) + ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats); if (req->rq_export && req->rq_export->exp_nid_stats && req->rq_export->exp_nid_stats->nid_ldlm_stats) @@ -1077,7 +1191,7 @@ 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, @@ -1098,7 +1212,7 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns, /* 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) && + if (!(exp_connect_flags(req->rq_export) & 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 = @@ -1108,30 +1222,52 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns, } #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() */ + /* coverity[overrun-buffer-val] */ 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); + 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); - - 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_ERROR(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). @@ -1148,6 +1284,11 @@ 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); existing_lock: if (flags & LDLM_FL_HAS_INTENT) { @@ -1158,11 +1299,8 @@ existing_lock: /* based on the assumption that lvb size never changes during * resource life time otherwise it need resource->lr_lock's * protection */ - if (lock->l_resource->lr_lvb_len) { - req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, - RCL_SERVER, - lock->l_resource->lr_lvb_len); - } + req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, + RCL_SERVER, ldlm_lvbo_size(lock)); if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR)) GOTO(out, rc = -ENOMEM); @@ -1173,19 +1311,21 @@ existing_lock: } if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN) - ldlm_convert_policy_to_local( + 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, (int *)&flags); - if (err) - GOTO(out, err); + err = ldlm_lock_enqueue(ns, &lock, cookie, &flags); + 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 = flags; ldlm_lock2desc(lock, &dlm_rep->lock_desc); ldlm_lock2handle(lock, &dlm_rep->lock_handle); @@ -1196,8 +1336,8 @@ existing_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 |= 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). @@ -1206,8 +1346,8 @@ 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) { - dlm_rep->lock_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) { /* * Only cancel lock if it was granted, because it would @@ -1229,7 +1369,7 @@ 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", @@ -1262,27 +1402,51 @@ 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) { - if (lock->l_resource->lr_lvb_len > 0) { - /* MDT path won't handle lr_lvb_data, so - * lock/unlock better be contained in the - * if block */ - void *lvb; - - lvb = req_capsule_server_get(&req->rq_pill, - &RMF_DLM_LVB); - LASSERTF(lvb != NULL, "req %p, lock %p\n", - req, lock); - lock_res(lock->l_resource); - memcpy(lvb, lock->l_resource->lr_lvb_data, - lock->l_resource->lr_lvb_len); - unlock_res(lock->l_resource); - } - } else { + 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; + + buf = req_capsule_server_get(&req->rq_pill, + &RMF_DLM_LVB); + LASSERTF(buf != NULL, "req %p, lock %p\n", + req, lock); + buflen = req_capsule_get_size(&req->rq_pill, + &RMF_DLM_LVB, 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; + } + } + } + + if (rc != 0) { lock_res_and_lock(lock); ldlm_resource_unlink_lock(lock); ldlm_lock_destroy_nolock(lock); @@ -1300,7 +1464,11 @@ 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, @@ -1323,7 +1491,11 @@ 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) { @@ -1346,7 +1518,7 @@ 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; @@ -1360,7 +1532,7 @@ 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; } } @@ -1374,7 +1546,12 @@ 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; @@ -1389,8 +1566,14 @@ 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) { @@ -1403,6 +1586,9 @@ int ldlm_request_cancel(struct ptlrpc_request *req, if (first >= count) RETURN(0); + if (count == 1 && dlm_req->lock_handle[0].cookie == 0) + RETURN(0); + /* There is no lock on the server at the replay time, * skip lock cancelling to make replay tests to pass. */ if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) @@ -1423,6 +1609,9 @@ 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); @@ -1447,7 +1636,13 @@ int ldlm_request_cancel(struct ptlrpc_request *req, LDLM_DEBUG_NOLOCK("server-side cancel handler END"); RETURN(done); } +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; @@ -1470,11 +1665,18 @@ int ldlm_handle_cancel(struct ptlrpc_request *req) RETURN(rc); if (!ldlm_request_cancel(req, dlm_req, 0)) - req->rq_status = ESTALE; + 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) { @@ -1484,10 +1686,10 @@ 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); @@ -1508,89 +1710,126 @@ 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) { - CFS_LIST_HEAD(ast_list); - ENTRY; - - 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; - } - } - - 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; - } - - /* 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( - 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"); - } + struct list_head ast_list; + int lvb_len; + int rc = 0; + ENTRY; + + LDLM_DEBUG(lock, "client completion callback handler START"); + + 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) { + schedule_timeout_and_set_state( + TASK_INTERRUPTIBLE, 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) { + 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); + + 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); + } + } else if (ldlm_has_layout(lock)) { /* for layout lock, lvb has + * variable length */ + void *lvb_data; + + OBD_ALLOC_LARGE(lvb_data, lvb_len); + if (lvb_data == NULL) { + LDLM_ERROR(lock, "No memory: %d.\n", lvb_len); + GOTO(out, rc = -ENOMEM); + } + + lock_res_and_lock(lock); + LASSERT(lock->l_lvb_data == NULL); + lock->l_lvb_type = LVB_T_LAYOUT; + lock->l_lvb_data = lvb_data; + lock->l_lvb_len = lvb_len; + unlock_res_and_lock(lock); + } + } + + 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 (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); @@ -1605,10 +1844,25 @@ 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, @@ -1661,106 +1915,112 @@ 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, + ldlm_cancel_flags_t cancel_flags) { - struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool; - ENTRY; - - cfs_spin_lock(&blp->blp_lock); - if (blwi->blwi_lock && blwi->blwi_lock->l_flags & LDLM_FL_DISCARD_DATA) { - /* add LDLM_FL_DISCARD_DATA requests to the priority list */ - cfs_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); - } - cfs_spin_unlock(&blp->blp_lock); - - cfs_waitq_signal(&blp->blp_waitq); - - /* can not use blwi->blwi_mode as blwi could be already freed in - LDLM_ASYNC mode */ - if (mode == LDLM_SYNC) - cfs_wait_for_completion(&blwi->blwi_comp); - - RETURN(0); + struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool; + ENTRY; + + spin_lock(&blp->blp_lock); + if (blwi->blwi_lock && + ldlm_is_discard_data(blwi->blwi_lock)) { + /* add LDLM_FL_DISCARD_DATA requests to the priority list */ + list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list); + } else { + /* other blocking callbacks are added to the regular list */ + list_add_tail(&blwi->blwi_entry, &blp->blp_list); + } + spin_unlock(&blp->blp_lock); + + wake_up(&blp->blp_waitq); + + /* 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, + ldlm_cancel_flags_t cancel_flags) { - cfs_init_completion(&blwi->blwi_comp); - CFS_INIT_LIST_HEAD(&blwi->blwi_head); + init_completion(&blwi->blwi_comp); + 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, + ldlm_cancel_flags_t cancel_flags) { - ENTRY; - - 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); - - RETURN(__ldlm_bl_to_thread(blwi, LDLM_ASYNC)); - } + ENTRY; + + if (cancels && count == 0) + RETURN(0); + + if (cancel_flags & LCF_ASYNC) { + struct ldlm_bl_work_item *blwi; + + 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, + ldlm_cancel_flags_t 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); } /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */ @@ -1796,7 +2056,8 @@ static int ldlm_handle_setinfo(struct ptlrpc_request *req) if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) /* Pass it on to mdc (the "export" in this case) */ - rc = obd_set_info_async(req->rq_export, + rc = obd_set_info_async(req->rq_svc_thread->t_env, + req->rq_export, sizeof(KEY_HSM_COPYTOOL_SEND), KEY_HSM_COPYTOOL_SEND, vallen, val, NULL); @@ -1820,6 +2081,23 @@ static inline void ldlm_callback_errmsg(struct ptlrpc_request *req, CWARN("Send reply failed, maybe cause bug 21636.\n"); } +static int ldlm_handle_qc_callback(struct ptlrpc_request *req) +{ + struct obd_quotactl *oqctl; + struct client_obd *cli = &req->rq_export->exp_obd->u.cli; + + oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL); + if (oqctl == NULL) { + CERROR("Can't unpack obd_quotactl\n"); + RETURN(-EPROTO); + } + + oqctl->qc_stat = ptlrpc_status_ntoh(oqctl->qc_stat); + + cli->cl_qchk_stat = oqctl->qc_stat; + return 0; +} + /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */ static int ldlm_callback_handler(struct ptlrpc_request *req) { @@ -1850,51 +2128,28 @@ static int ldlm_callback_handler(struct ptlrpc_request *req) LASSERT(req->rq_export != NULL); LASSERT(req->rq_export->exp_obd != NULL); - switch (lustre_msg_get_opc(req->rq_reqmsg)) { - case LDLM_BL_CALLBACK: - if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK)) - RETURN(0); - break; - case LDLM_CP_CALLBACK: - if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK)) - RETURN(0); - break; - case LDLM_GL_CALLBACK: - if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK)) - RETURN(0); - break; + switch (lustre_msg_get_opc(req->rq_reqmsg)) { + case LDLM_BL_CALLBACK: + if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) + RETURN(0); + break; + case LDLM_CP_CALLBACK: + if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET)) + RETURN(0); + break; + case LDLM_GL_CALLBACK: + if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET)) + RETURN(0); + break; case LDLM_SET_INFO: 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 OBD_QC_CALLBACK: - req_capsule_set(&req->rq_pill, &RQF_QC_CALLBACK); - if (OBD_FAIL_CHECK(OBD_FAIL_OBD_QC_CALLBACK_NET)) - RETURN(0); - rc = target_handle_qc_callback(req); - ldlm_callback_reply(req, rc); - RETURN(0); - case QUOTA_DQACQ: - case QUOTA_DQREL: - /* reply in handler */ - req_capsule_set(&req->rq_pill, &RQF_MDS_QUOTA_DQACQ); - rc = target_handle_dqacq_callback(req); - 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)) RETURN(0); - rc = llog_origin_handle_create(req); + rc = llog_origin_handle_open(req); ldlm_callback_reply(req, rc); RETURN(0); case LLOG_ORIGIN_HANDLE_NEXT_BLOCK: @@ -1919,6 +2174,13 @@ static int ldlm_callback_handler(struct ptlrpc_request *req) rc = llog_origin_handle_close(req); ldlm_callback_reply(req, rc); RETURN(0); + case OBD_QC_CALLBACK: + req_capsule_set(&req->rq_pill, &RQF_QC_CALLBACK); + if (OBD_FAIL_CHECK(OBD_FAIL_OBD_QC_CALLBACK_NET)) + RETURN(0); + rc = ldlm_handle_qc_callback(req); + ldlm_callback_reply(req, rc); + RETURN(0); default: CERROR("unknown opcode %u\n", lustre_msg_get_opc(req->rq_reqmsg)); @@ -1943,7 +2205,7 @@ 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); } @@ -1958,21 +2220,21 @@ 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 |= (dlm_req->lock_flags & LDLM_AST_FLAGS); + lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags & + 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 (((lock->l_flags & LDLM_FL_CANCELING) && - (lock->l_flags & LDLM_FL_BL_DONE)) || - (lock->l_flags & LDLM_FL_FAILED)) { + if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) || + ldlm_is_failed(lock)) { LDLM_DEBUG(lock, "callback on lock " LPX64" - lock disappeared\n", dlm_req->lock_handle[0].cookie); @@ -1983,10 +2245,10 @@ static int ldlm_callback_handler(struct ptlrpc_request *req) &dlm_req->lock_handle[0]); RETURN(0); } - /* 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_BL_AST; + ldlm_set_bl_ast(lock); } unlock_res_and_lock(lock); @@ -2003,7 +2265,7 @@ 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, @@ -2030,6 +2292,12 @@ static int ldlm_callback_handler(struct ptlrpc_request *req) RETURN(0); } +#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; @@ -2070,21 +2338,13 @@ static int ldlm_cancel_handler(struct ptlrpc_request *req) case LDLM_CANCEL: req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL); CDEBUG(D_INODE, "cancel\n"); - if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL)) - RETURN(0); + if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) || + CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND)) + 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)); @@ -2095,11 +2355,96 @@ static int ldlm_cancel_handler(struct ptlrpc_request *req) RETURN(0); } +static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req, + struct ldlm_lock *lock) +{ + struct ldlm_request *dlm_req; + struct lustre_handle lockh; + int rc = 0; + int i; + ENTRY; + + dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); + if (dlm_req == NULL) + RETURN(0); + + ldlm_lock2handle(lock, &lockh); + for (i = 0; i < dlm_req->lock_count; i++) { + if (lustre_handle_equal(&dlm_req->lock_handle[i], + &lockh)) { + DEBUG_REQ(D_RPCTRACE, req, + "Prio raised by lock "LPX64".", lockh.cookie); + + rc = 1; + break; + } + } + + RETURN(rc); + +} + +static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req) +{ + struct ldlm_request *dlm_req; + int rc = 0; + int i; + ENTRY; + + /* no prolong in recovery */ + if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) + RETURN(0); + + dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); + if (dlm_req == NULL) + RETURN(-EFAULT); + + for (i = 0; i < dlm_req->lock_count; i++) { + struct ldlm_lock *lock; + + lock = ldlm_handle2lock(&dlm_req->lock_handle[i]); + if (lock == NULL) + continue; + + rc = ldlm_is_ast_sent(lock) ? 1 : 0; + if (rc) + LDLM_DEBUG(lock, "hpreq cancel lock"); + LDLM_LOCK_PUT(lock); + + if (rc) + break; + } + + RETURN(rc); +} + +static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = { + .hpreq_lock_match = ldlm_cancel_hpreq_lock_match, + .hpreq_check = ldlm_cancel_hpreq_check, + .hpreq_fini = NULL, +}; + +static int ldlm_hpreq_handler(struct ptlrpc_request *req) +{ + ENTRY; + + req_capsule_init(&req->rq_pill, req, RCL_SERVER); + + if (req->rq_export == NULL) + RETURN(0); + + if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) { + req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL); + req->rq_ops = &ldlm_cancel_hpreq_ops; + } + RETURN(0); +} + int ldlm_revoke_lock_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd, - cfs_hlist_node_t *hnode, void *data) + 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); @@ -2116,7 +2461,7 @@ 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; } @@ -2124,12 +2469,18 @@ 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; - if (lock->l_export && lock->l_export->exp_lock_hash && - !cfs_hlist_unhashed(&lock->l_exp_hash)) - cfs_hash_del(lock->l_export->exp_lock_hash, - &lock->l_remote_handle, &lock->l_exp_hash); - cfs_list_add_tail(&lock->l_rk_ast, rpc_list); + 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. */ + /* In the function below, .hs_keycmp resolves to + * ldlm_export_lock_keycmp() */ + /* coverity[overrun-buffer-val] */ + cfs_hash_del(lock->l_export->exp_lock_hash, + &lock->l_remote_handle, &lock->l_exp_hash); + } + + list_add_tail(&lock->l_rk_ast, rpc_list); LDLM_LOCK_GET(lock); unlock_res_and_lock(lock); @@ -2138,10 +2489,10 @@ 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; + struct list_head rpc_list; ENTRY; - CFS_INIT_LIST_HEAD(&rpc_list); + 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, @@ -2149,147 +2500,150 @@ void ldlm_revoke_export_locks(struct obd_export *exp) 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) { - struct ldlm_bl_work_item *blwi = NULL; - static unsigned int num_bl = 0; - - cfs_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); - } - cfs_spin_unlock(&blp->blp_lock); - - return blwi; + struct ldlm_bl_work_item *blwi = NULL; + static unsigned int num_bl = 0; + + spin_lock(&blp->blp_lock); + /* 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 >= atomic_read(&blp->blp_num_threads)) + num_bl = 0; + list_del(&blwi->blwi_entry); + } + spin_unlock(&blp->blp_lock); + + return blwi; } /* 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; - cfs_completion_t bltd_comp; - int bltd_num; + char bltd_name[CFS_CURPROC_COMM_MAX]; + struct ldlm_bl_pool *bltd_blp; + struct completion bltd_comp; + int bltd_num; }; static int ldlm_bl_thread_main(void *arg); static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp) { - struct ldlm_bl_thread_data bltd = { .bltd_blp = blp }; - int rc; - - cfs_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; - } - cfs_wait_for_completion(&bltd.bltd_comp); - - return 0; + struct ldlm_bl_thread_data bltd = { .bltd_blp = blp }; + struct task_struct *task; + + init_completion(&bltd.bltd_comp); + bltd.bltd_num = atomic_read(&blp->blp_num_threads); + snprintf(bltd.bltd_name, sizeof(bltd.bltd_name) - 1, + "ldlm_bl_%02d", bltd.bltd_num); + task = kthread_run(ldlm_bl_thread_main, &bltd, bltd.bltd_name); + if (IS_ERR(task)) { + CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n", + atomic_read(&blp->blp_num_threads), PTR_ERR(task)); + return PTR_ERR(task); + } + wait_for_completion(&bltd.bltd_comp); + + return 0; } +/** + * 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; - { - struct ldlm_bl_thread_data *bltd = arg; - - blp = bltd->bltd_blp; - - bltd->bltd_num = - cfs_atomic_inc_return(&blp->blp_num_threads) - 1; - cfs_atomic_inc(&blp->blp_busy_threads); + blp = bltd->bltd_blp; - snprintf(bltd->bltd_name, sizeof(bltd->bltd_name) - 1, - "ldlm_bl_%02d", bltd->bltd_num); - cfs_daemonize(bltd->bltd_name); + atomic_inc(&blp->blp_num_threads); + atomic_inc(&blp->blp_busy_threads); - cfs_complete(&bltd->bltd_comp); - /* cannot use bltd after this, it is only on caller's stack */ - } + 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; + while (1) { + struct l_wait_info lwi = { 0 }; + struct ldlm_bl_work_item *blwi = NULL; + int busy; - blwi = ldlm_bl_get_work(blp); + blwi = ldlm_bl_get_work(blp); - if (blwi == NULL) { - int busy; + if (blwi == NULL) { + atomic_dec(&blp->blp_busy_threads); + l_wait_event_exclusive(blp->blp_waitq, + (blwi = ldlm_bl_get_work(blp)) != NULL, + &lwi); + busy = atomic_inc_return(&blp->blp_busy_threads); + } else { + busy = atomic_read(&blp->blp_busy_threads); + } - 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); + if (blwi->blwi_ns == NULL) + /* added by ldlm_cleanup() */ + break; - if (blwi->blwi_ns == NULL) - /* added by ldlm_cleanup() */ - break; + /* Not fatal if racy and have a few too many threads */ + if (unlikely(busy < blp->blp_max_threads && + busy >= atomic_read(&blp->blp_num_threads) && + !blwi->blwi_mem_pressure)) + /* discard the return value, we tried */ + ldlm_bl_thread_start(blp); - /* 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); - } else { - if (blwi->blwi_ns == NULL) - /* added by ldlm_cleanup() */ - break; - } if (blwi->blwi_mem_pressure) - cfs_memory_pressure_set(); + 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 + /* 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); + 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) - cfs_memory_pressure_clr(); - - if (blwi->blwi_mode == LDLM_ASYNC) - OBD_FREE(blwi, sizeof(*blwi)); - else - cfs_complete(&blwi->blwi_comp); - } - - cfs_atomic_dec(&blp->blp_busy_threads); - cfs_atomic_dec(&blp->blp_num_threads); - cfs_complete(&blp->blp_comp); - RETURN(0); + memory_pressure_clr(); + + if (blwi->blwi_flags & LCF_ASYNC) + OBD_FREE(blwi, sizeof(*blwi)); + else + complete(&blwi->blwi_comp); + } + + atomic_dec(&blp->blp_busy_threads); + atomic_dec(&blp->blp_num_threads); + complete(&blp->blp_comp); + RETURN(0); } -#endif static int ldlm_setup(void); static int ldlm_cleanup(void); @@ -2298,21 +2652,22 @@ int ldlm_get_ref(void) { int rc = 0; ENTRY; - cfs_mutex_down(&ldlm_ref_sem); + mutex_lock(&ldlm_ref_mutex); if (++ldlm_refcount == 1) { rc = ldlm_setup(); if (rc) ldlm_refcount--; } - cfs_mutex_up(&ldlm_ref_sem); + mutex_unlock(&ldlm_ref_mutex); RETURN(rc); } +EXPORT_SYMBOL(ldlm_get_ref); void ldlm_put_ref(void) { ENTRY; - cfs_mutex_down(&ldlm_ref_sem); + mutex_lock(&ldlm_ref_mutex); if (ldlm_refcount == 1) { int rc = ldlm_cleanup(); if (rc) @@ -2322,10 +2677,11 @@ void ldlm_put_ref(void) } else { ldlm_refcount--; } - cfs_mutex_up(&ldlm_ref_sem); + mutex_unlock(&ldlm_ref_mutex); EXIT; } +EXPORT_SYMBOL(ldlm_put_ref); /* * Export handle<->lock hash operations. @@ -2337,50 +2693,50 @@ ldlm_export_lock_hash(cfs_hash_t *hs, const void *key, unsigned 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(cfs_hash_t *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(cfs_hash_t *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); } @@ -2397,6 +2753,7 @@ static cfs_hash_ops_t ldlm_export_lock_ops = { int ldlm_init_export(struct obd_export *exp) { + int rc; ENTRY; exp->exp_lock_hash = @@ -2412,7 +2769,14 @@ 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); @@ -2421,19 +2785,22 @@ void ldlm_destroy_export(struct obd_export *exp) ENTRY; cfs_hash_putref(exp->exp_lock_hash); exp->exp_lock_hash = NULL; + + ldlm_destroy_flock_export(exp); EXIT; } EXPORT_SYMBOL(ldlm_destroy_export); static int ldlm_setup(void) { - struct ldlm_bl_pool *blp; - int rc = 0; - int ldlm_min_threads = LDLM_THREADS_AUTO_MIN; - int ldlm_max_threads = LDLM_THREADS_AUTO_MAX; -#ifdef __KERNEL__ - int i; -#endif + static struct ptlrpc_service_conf conf; + 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) @@ -2446,166 +2813,199 @@ static int ldlm_setup(void) #ifdef LPROCFS rc = ldlm_proc_setup(); if (rc != 0) - GOTO(out_free, rc); -#endif - -#ifdef __KERNEL__ - if (ldlm_num_threads) { - /* If ldlm_num_threads is set, it is the min and the max. */ - if (ldlm_num_threads > LDLM_THREADS_AUTO_MAX) - ldlm_num_threads = LDLM_THREADS_AUTO_MAX; - if (ldlm_num_threads < LDLM_THREADS_AUTO_MIN) - ldlm_num_threads = LDLM_THREADS_AUTO_MIN; - ldlm_min_threads = ldlm_max_threads = ldlm_num_threads; - } -#endif - - ldlm_state->ldlm_cb_service = - ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE, - LDLM_MAXREPSIZE, LDLM_CB_REQUEST_PORTAL, - LDLM_CB_REPLY_PORTAL, 2, - ldlm_callback_handler, "ldlm_cbd", - ldlm_svc_proc_dir, NULL, - ldlm_min_threads, ldlm_max_threads, - "ldlm_cb", - LCT_MD_THREAD|LCT_DT_THREAD, NULL); - - if (!ldlm_state->ldlm_cb_service) { - CERROR("failed to start service\n"); - GOTO(out_proc, rc = -ENOMEM); - } - - ldlm_state->ldlm_cancel_service = - ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE, - LDLM_MAXREPSIZE, LDLM_CANCEL_REQUEST_PORTAL, - LDLM_CANCEL_REPLY_PORTAL, 6, - ldlm_cancel_handler, "ldlm_canceld", - ldlm_svc_proc_dir, NULL, - ldlm_min_threads, ldlm_max_threads, - "ldlm_cn", - LCT_MD_THREAD|LCT_DT_THREAD|LCT_CL_THREAD, - NULL); - - if (!ldlm_state->ldlm_cancel_service) { - CERROR("failed to start service\n"); - GOTO(out_proc, rc = -ENOMEM); - } + GOTO(out, rc); +#endif /* LPROCFS */ + + memset(&conf, 0, sizeof(conf)); + conf = (typeof(conf)) { + .psc_name = "ldlm_cbd", + .psc_watchdog_factor = 2, + .psc_buf = { + .bc_nbufs = LDLM_CLIENT_NBUFS, + .bc_buf_size = LDLM_BUFSIZE, + .bc_req_max_size = LDLM_MAXREQSIZE, + .bc_rep_max_size = LDLM_MAXREPSIZE, + .bc_req_portal = LDLM_CB_REQUEST_PORTAL, + .bc_rep_portal = LDLM_CB_REPLY_PORTAL, + }, + .psc_thr = { + .tc_thr_name = "ldlm_cb", + .tc_thr_factor = LDLM_THR_FACTOR, + .tc_nthrs_init = LDLM_NTHRS_INIT, + .tc_nthrs_base = LDLM_NTHRS_BASE, + .tc_nthrs_max = LDLM_NTHRS_MAX, + .tc_nthrs_user = ldlm_num_threads, + .tc_cpu_affinity = 1, + .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD, + }, + .psc_cpt = { + .cc_pattern = ldlm_cpts, + }, + .psc_ops = { + .so_req_handler = ldlm_callback_handler, + }, + }; + ldlm_state->ldlm_cb_service = \ + ptlrpc_register_service(&conf, 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); + ldlm_state->ldlm_cb_service = NULL; + GOTO(out, rc); + } + +#ifdef HAVE_SERVER_SUPPORT + memset(&conf, 0, sizeof(conf)); + conf = (typeof(conf)) { + .psc_name = "ldlm_canceld", + .psc_watchdog_factor = 6, + .psc_buf = { + .bc_nbufs = LDLM_SERVER_NBUFS, + .bc_buf_size = LDLM_BUFSIZE, + .bc_req_max_size = LDLM_MAXREQSIZE, + .bc_rep_max_size = LDLM_MAXREPSIZE, + .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL, + .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL, + + }, + .psc_thr = { + .tc_thr_name = "ldlm_cn", + .tc_thr_factor = LDLM_THR_FACTOR, + .tc_nthrs_init = LDLM_NTHRS_INIT, + .tc_nthrs_base = LDLM_NTHRS_BASE, + .tc_nthrs_max = LDLM_NTHRS_MAX, + .tc_nthrs_user = ldlm_num_threads, + .tc_cpu_affinity = 1, + .tc_ctx_tags = LCT_MD_THREAD | \ + LCT_DT_THREAD | \ + LCT_CL_THREAD, + }, + .psc_cpt = { + .cc_pattern = ldlm_cpts, + }, + .psc_ops = { + .so_req_handler = ldlm_cancel_handler, + .so_hpreq_handler = ldlm_hpreq_handler, + }, + }; + ldlm_state->ldlm_cancel_service = \ + ptlrpc_register_service(&conf, 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 /* HAVE_SERVER_SUPPORT */ + + OBD_ALLOC(blp, sizeof(*blp)); + if (blp == NULL) + GOTO(out, rc = -ENOMEM); + ldlm_state->ldlm_bl_pool = blp; + + spin_lock_init(&blp->blp_lock); + 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); + + if (ldlm_num_threads == 0) { + blp->blp_min_threads = LDLM_NTHRS_INIT; + blp->blp_max_threads = LDLM_NTHRS_MAX; + } else { + blp->blp_min_threads = blp->blp_max_threads = \ + min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT, + ldlm_num_threads)); + } + + for (i = 0; i < blp->blp_min_threads; i++) { + rc = ldlm_bl_thread_start(blp); + if (rc < 0) + GOTO(out, rc); + } + +#ifdef HAVE_SERVER_SUPPORT + INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks); + expired_lock_thread.elt_state = ELT_STOPPED; + init_waitqueue_head(&expired_lock_thread.elt_waitq); + + INIT_LIST_HEAD(&waiting_locks_list); + spin_lock_init(&waiting_locks_spinlock); + cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0); + + 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); + } + + wait_event(expired_lock_thread.elt_waitq, + expired_lock_thread.elt_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); + } + RETURN(0); - OBD_ALLOC(blp, sizeof(*blp)); - if (blp == NULL) - GOTO(out_proc, rc = -ENOMEM); - ldlm_state->ldlm_bl_pool = blp; - - cfs_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); - blp->blp_min_threads = ldlm_min_threads; - blp->blp_max_threads = ldlm_max_threads; - -#ifdef __KERNEL__ - for (i = 0; i < blp->blp_min_threads; i++) { - rc = ldlm_bl_thread_start(blp); - if (rc < 0) - GOTO(out_thread, rc); - } - - rc = ptlrpc_start_threads(ldlm_state->ldlm_cancel_service); - if (rc) - GOTO(out_thread, rc); - - rc = ptlrpc_start_threads(ldlm_state->ldlm_cb_service); - if (rc) - GOTO(out_thread, rc); - - 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); - cfs_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) { - CERROR("Cannot start ldlm expired-lock thread: %d\n", rc); - GOTO(out_thread, rc); - } - - cfs_wait_event(expired_lock_thread.elt_waitq, - expired_lock_thread.elt_state == ELT_READY); -#endif - -#ifdef __KERNEL__ - rc = ldlm_pools_init(); - if (rc) - GOTO(out_thread, rc); -#endif - RETURN(0); - -#ifdef __KERNEL__ - out_thread: - ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service); - ptlrpc_unregister_service(ldlm_state->ldlm_cb_service); -#endif - - out_proc: -#ifdef LPROCFS - ldlm_proc_cleanup(); - out_free: -#endif - OBD_FREE(ldlm_state, sizeof(*ldlm_state)); - ldlm_state = NULL; - return rc; + out: + ldlm_cleanup(); + RETURN(rc); } static int ldlm_cleanup(void) { -#ifdef __KERNEL__ - struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool; -#endif 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(); -#endif -#ifdef __KERNEL__ - while (cfs_atomic_read(&blp->blp_num_threads) > 0) { - struct ldlm_bl_work_item blwi = { .blwi_ns = NULL }; + if (ldlm_state->ldlm_bl_pool != NULL) { + struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool; - cfs_init_completion(&blp->blp_comp); + while (atomic_read(&blp->blp_num_threads) > 0) { + struct ldlm_bl_work_item blwi = { .blwi_ns = NULL }; - cfs_spin_lock(&blp->blp_lock); - cfs_list_add_tail(&blwi.blwi_entry, &blp->blp_list); - cfs_waitq_signal(&blp->blp_waitq); - cfs_spin_unlock(&blp->blp_lock); + init_completion(&blp->blp_comp); - cfs_wait_for_completion(&blp->blp_comp); - } - OBD_FREE(blp, sizeof(*blp)); - - ptlrpc_unregister_service(ldlm_state->ldlm_cb_service); - ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service); - ldlm_proc_cleanup(); - - 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); -#else - ptlrpc_unregister_service(ldlm_state->ldlm_cb_service); - ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service); + spin_lock(&blp->blp_lock); + 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); + } + + OBD_FREE(blp, sizeof(*blp)); + } + + if (ldlm_state->ldlm_cb_service != NULL) + ptlrpc_unregister_service(ldlm_state->ldlm_cb_service); +#ifdef HAVE_SERVER_SUPPORT + if (ldlm_state->ldlm_cancel_service != NULL) + ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service); +#endif + + ldlm_proc_cleanup(); + +#ifdef HAVE_SERVER_SUPPORT + if (expired_lock_thread.elt_state != ELT_STOPPED) { + expired_lock_thread.elt_state = ELT_TERMINATE; + wake_up(&expired_lock_thread.elt_waitq); + wait_event(expired_lock_thread.elt_waitq, + expired_lock_thread.elt_state == ELT_STOPPED); + } #endif OBD_FREE(ldlm_state, sizeof(*ldlm_state)); @@ -2614,31 +3014,36 @@ static int ldlm_cleanup(void) RETURN(0); } -int __init ldlm_init(void) +int ldlm_init(void) { - cfs_init_mutex(&ldlm_ref_sem); - cfs_init_mutex(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER)); - cfs_init_mutex(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_lock_slab = cfs_mem_cache_create("ldlm_locks", - sizeof(struct ldlm_lock), 0, - CFS_SLAB_HWCACHE_ALIGN | CFS_SLAB_DESTROY_BY_RCU); - if (ldlm_lock_slab == NULL) { - cfs_mem_cache_destroy(ldlm_resource_slab); - return -ENOMEM; - } - - ldlm_interval_slab = cfs_mem_cache_create("interval_node", + mutex_init(&ldlm_ref_mutex); + mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER)); + mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT)); + + INIT_LIST_HEAD(&ldlm_srv_namespace_list); + INIT_LIST_HEAD(&ldlm_cli_active_namespace_list); + INIT_LIST_HEAD(&ldlm_cli_inactive_namespace_list); + + 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 = kmem_cache_create("ldlm_locks", + sizeof(struct ldlm_lock), 0, + SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU, NULL); + if (ldlm_lock_slab == NULL) { + kmem_cache_destroy(ldlm_resource_slab); + return -ENOMEM; + } + + ldlm_interval_slab = kmem_cache_create("interval_node", sizeof(struct ldlm_interval), - 0, CFS_SLAB_HWCACHE_ALIGN); + 0, SLAB_HWCACHE_ALIGN, NULL); if (ldlm_interval_slab == NULL) { - cfs_mem_cache_destroy(ldlm_resource_slab); - cfs_mem_cache_destroy(ldlm_lock_slab); + kmem_cache_destroy(ldlm_resource_slab); + kmem_cache_destroy(ldlm_lock_slab); return -ENOMEM; } #if LUSTRE_TRACKS_LOCK_EXP_REFS @@ -2647,124 +3052,15 @@ int __init ldlm_init(void) return 0; } -void __exit ldlm_exit(void) +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"); -} - -/* ldlm_extent.c */ -EXPORT_SYMBOL(ldlm_extent_shift_kms); - -/* ldlm_lock.c */ -EXPORT_SYMBOL(ldlm_get_processing_policy); -EXPORT_SYMBOL(ldlm_lock2desc); -EXPORT_SYMBOL(ldlm_register_intent); -EXPORT_SYMBOL(ldlm_lockname); -EXPORT_SYMBOL(ldlm_typename); -EXPORT_SYMBOL(ldlm_lock2handle); -EXPORT_SYMBOL(__ldlm_handle2lock); -EXPORT_SYMBOL(ldlm_lock_get); -EXPORT_SYMBOL(ldlm_lock_put); -EXPORT_SYMBOL(ldlm_lock_match); -EXPORT_SYMBOL(ldlm_lock_cancel); -EXPORT_SYMBOL(ldlm_lock_addref); -EXPORT_SYMBOL(ldlm_lock_addref_try); -EXPORT_SYMBOL(ldlm_lock_decref); -EXPORT_SYMBOL(ldlm_lock_decref_and_cancel); -EXPORT_SYMBOL(ldlm_lock_change_resource); -EXPORT_SYMBOL(ldlm_it2str); -EXPORT_SYMBOL(ldlm_lock_dump); -EXPORT_SYMBOL(ldlm_lock_dump_handle); -EXPORT_SYMBOL(ldlm_reprocess_all_ns); -EXPORT_SYMBOL(ldlm_lock_allow_match_locked); -EXPORT_SYMBOL(ldlm_lock_allow_match); -EXPORT_SYMBOL(ldlm_lock_downgrade); -EXPORT_SYMBOL(ldlm_lock_convert); - -/* ldlm_request.c */ -EXPORT_SYMBOL(ldlm_completion_ast_async); -EXPORT_SYMBOL(ldlm_blocking_ast_nocheck); -EXPORT_SYMBOL(ldlm_completion_ast); -EXPORT_SYMBOL(ldlm_blocking_ast); -EXPORT_SYMBOL(ldlm_glimpse_ast); -EXPORT_SYMBOL(ldlm_expired_completion_wait); -EXPORT_SYMBOL(ldlm_prep_enqueue_req); -EXPORT_SYMBOL(ldlm_prep_elc_req); -EXPORT_SYMBOL(ldlm_cli_convert); -EXPORT_SYMBOL(ldlm_cli_enqueue); -EXPORT_SYMBOL(ldlm_cli_enqueue_fini); -EXPORT_SYMBOL(ldlm_cli_enqueue_local); -EXPORT_SYMBOL(ldlm_cli_cancel); -EXPORT_SYMBOL(ldlm_cli_cancel_unused); -EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource); -EXPORT_SYMBOL(ldlm_cli_cancel_req); -EXPORT_SYMBOL(ldlm_replay_locks); -EXPORT_SYMBOL(ldlm_resource_foreach); -EXPORT_SYMBOL(ldlm_namespace_foreach); -EXPORT_SYMBOL(ldlm_resource_iterate); -EXPORT_SYMBOL(ldlm_cancel_resource_local); -EXPORT_SYMBOL(ldlm_cli_cancel_list_local); -EXPORT_SYMBOL(ldlm_cli_cancel_list); - -/* ldlm_lockd.c */ -EXPORT_SYMBOL(ldlm_server_blocking_ast); -EXPORT_SYMBOL(ldlm_server_completion_ast); -EXPORT_SYMBOL(ldlm_server_glimpse_ast); -EXPORT_SYMBOL(ldlm_handle_enqueue); -EXPORT_SYMBOL(ldlm_handle_enqueue0); -EXPORT_SYMBOL(ldlm_handle_cancel); -EXPORT_SYMBOL(ldlm_request_cancel); -EXPORT_SYMBOL(ldlm_handle_convert); -EXPORT_SYMBOL(ldlm_handle_convert0); -EXPORT_SYMBOL(ldlm_del_waiting_lock); -EXPORT_SYMBOL(ldlm_get_ref); -EXPORT_SYMBOL(ldlm_put_ref); -EXPORT_SYMBOL(ldlm_refresh_waiting_lock); -EXPORT_SYMBOL(ldlm_revoke_export_locks); - -/* ldlm_resource.c */ -EXPORT_SYMBOL(ldlm_namespace_new); -EXPORT_SYMBOL(ldlm_namespace_cleanup); -EXPORT_SYMBOL(ldlm_namespace_free); -EXPORT_SYMBOL(ldlm_namespace_dump); -EXPORT_SYMBOL(ldlm_dump_all_namespaces); -EXPORT_SYMBOL(ldlm_resource_get); -EXPORT_SYMBOL(ldlm_resource_putref); -EXPORT_SYMBOL(ldlm_resource_unlink_lock); - -/* ldlm_lib.c */ -EXPORT_SYMBOL(client_import_add_conn); -EXPORT_SYMBOL(client_import_del_conn); -EXPORT_SYMBOL(client_obd_setup); -EXPORT_SYMBOL(client_obd_cleanup); -EXPORT_SYMBOL(client_connect_import); -EXPORT_SYMBOL(client_disconnect_export); -EXPORT_SYMBOL(server_disconnect_export); -EXPORT_SYMBOL(target_stop_recovery_thread); -EXPORT_SYMBOL(target_handle_connect); -EXPORT_SYMBOL(target_cleanup_recovery); -EXPORT_SYMBOL(target_destroy_export); -EXPORT_SYMBOL(target_cancel_recovery_timer); -EXPORT_SYMBOL(target_send_reply); -EXPORT_SYMBOL(target_queue_recovery_request); -EXPORT_SYMBOL(target_handle_ping); -EXPORT_SYMBOL(target_pack_pool_reply); -EXPORT_SYMBOL(target_handle_disconnect); - -/* l_lock.c */ -EXPORT_SYMBOL(lock_res_and_lock); -EXPORT_SYMBOL(unlock_res_and_lock); + 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); +}