X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_lockd.c;h=83b08d0063f34e8e10d407344df181eddd2cf9f3;hp=30a0b1fdf8bccf4a3e5f7c6f13216960d7c50dd7;hb=0cb008aa46d283c03aeae79fd3ab36c828da2174;hpb=8e387d33e24971c7025b6adb1d7f99f557c45227 diff --git a/lustre/ldlm/ldlm_lockd.c b/lustre/ldlm/ldlm_lockd.c index 30a0b1f..83b08d0 100644 --- a/lustre/ldlm/ldlm_lockd.c +++ b/lustre/ldlm/ldlm_lockd.c @@ -1,24 +1,42 @@ /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- * vim:expandtab:shiftwidth=8:tabstop=8: * - * Copyright (C) 2002, 2003 Cluster File Systems, Inc. - * Author: Peter Braam - * Author: Phil Schwan + * GPL HEADER START * - * This file is part of Lustre, http://www.lustre.org. + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * - * Lustre is free software; you can redistribute it and/or - * modify it under the terms of version 2 of the GNU General Public - * License as published by the Free Software Foundation. + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 only, + * as published by the Free Software Foundation. * - * Lustre is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * GNU General Public License for more details. + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * General Public License version 2 for more details (a copy is included + * in the LICENSE file that accompanied this code). * - * You should have received a copy of the GNU General Public License - * along with Lustre; if not, write to the Free Software - * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + * You should have received a copy of the GNU General Public License + * version 2 along with this program; If not, see + * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf + * + * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, + * CA 95054 USA or visit www.sun.com if you need additional information or + * have any questions. + * + * GPL HEADER END + */ +/* + * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Use is subject to license terms. + */ +/* + * This file is part of Lustre, http://www.lustre.org/ + * Lustre is a trademark of Sun Microsystems, Inc. + * + * lustre/ldlm/ldlm_lockd.c + * + * Author: Peter Braam + * Author: Phil Schwan */ #ifndef EXPORT_SYMTAB @@ -27,48 +45,56 @@ #define DEBUG_SUBSYSTEM S_LDLM #ifdef __KERNEL__ -# include -# include -# include -# include +# include #else # include #endif -#include -#include +#include +#include +#include #include "ldlm_internal.h" -extern kmem_cache_t *ldlm_resource_slab; -extern kmem_cache_t *ldlm_lock_slab; -extern struct lustre_lock ldlm_handle_lock; -extern struct list_head ldlm_namespace_list; -extern int (*mds_reint_p)(int offset, struct ptlrpc_request *req); -extern int (*mds_getattr_name_p)(int offset, struct ptlrpc_request *req); +#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 struct semaphore ldlm_ref_sem; +static int ldlm_refcount; -static DECLARE_MUTEX(ldlm_ref_sem); -static int ldlm_refcount = 0; +/* LDLM state */ -/* LDLM state */ +static struct ldlm_state *ldlm_state; -static struct ldlm_state *ldlm ; +inline cfs_time_t round_timeout(cfs_time_t timeout) +{ + return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1); +} -inline unsigned long round_timeout(unsigned long timeout) +/* timeout for initial callback (AST) reply (bz10399) */ +static inline unsigned int ldlm_get_rq_timeout(void) { - return ((timeout / HZ) + 1) * HZ; + /* Non-AT value */ + unsigned int timeout = min(ldlm_timeout, obd_timeout / 3); + + return timeout < 1 ? 1 : timeout; } #ifdef __KERNEL__ -/* XXX should this be per-ldlm? */ +/* w_l_spinlock protects both waiting_locks_list and expired_lock_thread */ +static spinlock_t waiting_locks_spinlock; /* BH lock (timer) */ static struct list_head waiting_locks_list; -static spinlock_t waiting_locks_spinlock; -static struct timer_list waiting_locks_timer; +static cfs_timer_t waiting_locks_timer; static struct expired_lock_thread { - wait_queue_head_t elt_waitq; + cfs_waitq_t elt_waitq; int elt_state; + int elt_dump; struct list_head elt_expired_locks; - spinlock_t elt_lock; } expired_lock_thread; #endif @@ -78,10 +104,26 @@ static struct expired_lock_thread { struct ldlm_bl_pool { spinlock_t blp_lock; + + /* + * blp_prio_list is used for callbacks that should be handled + * as a priority. It is used for LDLM_FL_DISCARD_DATA requests. + * see bug 13843 + */ + 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; - atomic_t blp_num_threads; + + cfs_waitq_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 { @@ -89,6 +131,8 @@ struct ldlm_bl_work_item { struct ldlm_namespace *blwi_ns; struct ldlm_lock_desc blwi_ld; struct ldlm_lock *blwi_lock; + struct list_head blwi_head; + int blwi_count; }; #ifdef __KERNEL__ @@ -97,9 +141,10 @@ static inline int have_expired_locks(void) { int need_to_run; - spin_lock_bh(&expired_lock_thread.elt_lock); + ENTRY; + spin_lock_bh(&waiting_locks_spinlock); need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks); - spin_unlock_bh(&expired_lock_thread.elt_lock); + spin_unlock_bh(&waiting_locks_spinlock); RETURN(need_to_run); } @@ -108,21 +153,13 @@ static int expired_lock_main(void *arg) { struct list_head *expired = &expired_lock_thread.elt_expired_locks; struct l_wait_info lwi = { 0 }; - unsigned long flags; + int do_dump; ENTRY; - lock_kernel(); - kportal_daemonize("ldlm_elt"); - - SIGNAL_MASK_LOCK(current, flags); - sigfillset(¤t->blocked); - RECALC_SIGPENDING; - SIGNAL_MASK_UNLOCK(current, flags); - - unlock_kernel(); + cfs_daemonize("ldlm_elt"); expired_lock_thread.elt_state = ELT_READY; - wake_up(&expired_lock_thread.elt_waitq); + cfs_waitq_signal(&expired_lock_thread.elt_waitq); while (1) { l_wait_event(expired_lock_thread.elt_waitq, @@ -130,58 +167,148 @@ static int expired_lock_main(void *arg) expired_lock_thread.elt_state == ELT_TERMINATE, &lwi); - spin_lock_bh(&expired_lock_thread.elt_lock); + spin_lock_bh(&waiting_locks_spinlock); + if (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(__FILE__, + "waiting_locks_callback", + expired_lock_thread.elt_dump); + + spin_lock_bh(&waiting_locks_spinlock); + expired_lock_thread.elt_dump = 0; + } + + do_dump = 0; + while (!list_empty(expired)) { - struct ldlm_lock *lock = list_entry(expired->next, - struct ldlm_lock, - l_pending_chain); - spin_unlock_bh(&expired_lock_thread.elt_lock); + struct obd_export *export; + struct ldlm_lock *lock; - ptlrpc_fail_export(lock->l_export); + lock = list_entry(expired->next, struct ldlm_lock, + l_pending_chain); + if ((void *)lock < LP_POISON + CFS_PAGE_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 + 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"); + continue; + } + export = class_export_get(lock->l_export); + spin_unlock_bh(&waiting_locks_spinlock); + + do_dump++; + class_fail_export(export); + class_export_put(export); + spin_lock_bh(&waiting_locks_spinlock); + } + spin_unlock_bh(&waiting_locks_spinlock); - spin_lock_bh(&expired_lock_thread.elt_lock); + if (do_dump && obd_dump_on_eviction) { + CERROR("dump the log upon eviction\n"); + libcfs_debug_dumplog(); } - spin_unlock_bh(&expired_lock_thread.elt_lock); if (expired_lock_thread.elt_state == ELT_TERMINATE) break; } expired_lock_thread.elt_state = ELT_STOPPED; - wake_up(&expired_lock_thread.elt_waitq); + cfs_waitq_signal(&expired_lock_thread.elt_waitq); RETURN(0); } +static int ldlm_add_waiting_lock(struct ldlm_lock *lock); + +/* This is called from within a timer interrupt and cannot schedule */ static void waiting_locks_callback(unsigned long unused) { - struct ldlm_lock *lock; - char str[PTL_NALFMT_SIZE]; + struct ldlm_lock *lock, *last = NULL; +repeat: 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 (lock->l_callback_timeout > jiffies) + if (cfs_time_after(lock->l_callback_timeout, cfs_time_current()) || + (lock->l_req_mode == LCK_GROUP)) break; - LDLM_ERROR(lock, "lock callback timer expired: evicting client " - "%s@%s nid "LPX64" (%s) ", - lock->l_export->exp_client_uuid.uuid, - lock->l_export->exp_connection->c_remote_uuid.uuid, - lock->l_export->exp_connection->c_peer.peer_nid, - portals_nid2str(lock->l_export->exp_connection->c_peer.peer_ni->pni_number, - lock->l_export->exp_connection->c_peer.peer_nid, - str)); + 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)); + + list_del_init(&lock->l_pending_chain); + 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)); + + list_del_init(&lock->l_pending_chain); + spin_unlock_bh(&waiting_locks_spinlock); + ldlm_add_waiting_lock(lock); + goto repeat; + } + + LDLM_ERROR(lock, "lock callback timer expired after %lds: " + "evicting client at %s ", + cfs_time_current_sec()- lock->l_enqueued_time.tv_sec, + libcfs_nid2str( + lock->l_export->exp_connection->c_peer.nid)); + + last = lock; - spin_lock_bh(&expired_lock_thread.elt_lock); list_del(&lock->l_pending_chain); list_add(&lock->l_pending_chain, &expired_lock_thread.elt_expired_locks); - spin_unlock_bh(&expired_lock_thread.elt_lock); - wake_up(&expired_lock_thread.elt_waitq); } + if (!list_empty(&expired_lock_thread.elt_expired_locks)) { + if (obd_dump_on_timeout) + expired_lock_thread.elt_dump = __LINE__; + + cfs_waitq_signal(&expired_lock_thread.elt_waitq); + } + + /* + * Make sure the timer will fire again if we have any locks + * left. + */ + if (!list_empty(&waiting_locks_list)) { + cfs_time_t timeout_rounded; + 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); + } spin_unlock_bh(&waiting_locks_spinlock); } @@ -190,72 +317,137 @@ static void waiting_locks_callback(unsigned long unused) * 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 * floods of timer firings during periods of high lock contention and traffic). + * + * Called with the namespace lock held. */ -static int ldlm_add_waiting_lock(struct ldlm_lock *lock) +static int __ldlm_add_waiting_lock(struct ldlm_lock *lock) { - unsigned long timeout_rounded; + int timeout; + cfs_time_t timeout_rounded; - spin_lock_bh(&waiting_locks_spinlock); - if (!list_empty(&lock->l_pending_chain)) { - LDLM_DEBUG(lock, "not re-adding to wait list"); - spin_unlock_bh(&waiting_locks_spinlock); + if (!list_empty(&lock->l_pending_chain)) return 0; - } - LDLM_DEBUG(lock, "adding to wait list"); - lock->l_callback_timeout = jiffies + (obd_timeout * HZ / 2); + timeout = ldlm_get_enq_timeout(lock); + + lock->l_callback_timeout = cfs_time_shift(timeout); timeout_rounded = round_timeout(lock->l_callback_timeout); - if (timeout_rounded < waiting_locks_timer.expires || - !timer_pending(&waiting_locks_timer)) { - mod_timer(&waiting_locks_timer, timeout_rounded); + if (cfs_time_before(timeout_rounded, + cfs_timer_deadline(&waiting_locks_timer)) || + !cfs_timer_is_armed(&waiting_locks_timer)) { + cfs_timer_arm(&waiting_locks_timer, timeout_rounded); } + /* if the new lock has a shorter timeout than something earlier on + the list, we'll wait the longer amount of time; no big deal. */ list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */ - spin_unlock_bh(&waiting_locks_spinlock); return 1; } +static int ldlm_add_waiting_lock(struct ldlm_lock *lock) +{ + int ret; + + LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)); + + spin_lock_bh(&waiting_locks_spinlock); + if (lock->l_destroyed) { + static cfs_time_t next; + spin_unlock_bh(&waiting_locks_spinlock); + LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)"); + if (cfs_time_after(cfs_time_current(), next)) { + next = cfs_time_shift(14400); + libcfs_debug_dumpstack(NULL); + } + return 0; + } + + ret = __ldlm_add_waiting_lock(lock); + spin_unlock_bh(&waiting_locks_spinlock); + + LDLM_DEBUG(lock, "%sadding to wait list", + ret == 0 ? "not re-" : ""); + 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. + * + * Called with namespace lock held. */ -int ldlm_del_waiting_lock(struct ldlm_lock *lock) +int __ldlm_del_waiting_lock(struct ldlm_lock *lock) { struct list_head *list_next; - 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"); + if (list_empty(&lock->l_pending_chain)) return 0; - } list_next = lock->l_pending_chain.next; if (lock->l_pending_chain.prev == &waiting_locks_list) { /* Removing the head of the list, adjust timer. */ if (list_next == &waiting_locks_list) { /* No more, just cancel. */ - del_timer(&waiting_locks_timer); + cfs_timer_disarm(&waiting_locks_timer); } else { struct ldlm_lock *next; next = list_entry(list_next, struct ldlm_lock, l_pending_chain); - mod_timer(&waiting_locks_timer, - round_timeout(next->l_callback_timeout)); + cfs_timer_arm(&waiting_locks_timer, + round_timeout(next->l_callback_timeout)); } } list_del_init(&lock->l_pending_chain); + + return 1; +} + +int ldlm_del_waiting_lock(struct ldlm_lock *lock) +{ + int ret; + + 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); + ret = __ldlm_del_waiting_lock(lock); spin_unlock_bh(&waiting_locks_spinlock); - LDLM_DEBUG(lock, "removed"); + + LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed"); + return ret; +} + +/* + * Prolong the lock + * + * Called with namespace lock held. + */ +int ldlm_refresh_waiting_lock(struct ldlm_lock *lock) +{ + 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; + } + + __ldlm_del_waiting_lock(lock); + __ldlm_add_waiting_lock(lock); + spin_unlock_bh(&waiting_locks_spinlock); + + LDLM_DEBUG(lock, "refreshed"); return 1; } @@ -263,6 +455,7 @@ int ldlm_del_waiting_lock(struct ldlm_lock *lock) static int ldlm_add_waiting_lock(struct ldlm_lock *lock) { + LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)); RETURN(1); } @@ -271,349 +464,760 @@ int ldlm_del_waiting_lock(struct ldlm_lock *lock) RETURN(0); } +int ldlm_refresh_waiting_lock(struct ldlm_lock *lock) +{ + RETURN(0); +} #endif /* __KERNEL__ */ -static void ldlm_failed_ast(struct ldlm_lock *lock, int rc, char *ast_type) +static void ldlm_failed_ast(struct ldlm_lock *lock, int rc, + const char *ast_type) +{ + struct ptlrpc_connection *conn = lock->l_export->exp_connection; + char *str = libcfs_nid2str(conn->c_peer.nid); + + LCONSOLE_ERROR_MSG(0x138, "A client on nid %s was evicted from " + "service %s.\n", str, + lock->l_export->exp_obd->obd_name); + + LCONSOLE_ERROR_MSG(0x012, "Lock %s callback to %s timed out for " + "resource %d\n", ast_type, + obd_export_nid2str(lock->l_export), rc); + + if (obd_dump_on_timeout) + libcfs_debug_dumplog(); +#ifdef __KERNEL__ + spin_lock_bh(&waiting_locks_spinlock); + list_add(&lock->l_pending_chain, &expired_lock_thread.elt_expired_locks); + cfs_waitq_signal(&expired_lock_thread.elt_waitq); + spin_unlock_bh(&waiting_locks_spinlock); +#else + class_fail_export(lock->l_export); +#endif +} + +static int ldlm_handle_ast_error(struct ldlm_lock *lock, + struct ptlrpc_request *req, int rc, + const char *ast_type) +{ + lnet_process_id_t peer = req->rq_import->imp_connection->c_peer; + + if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) { + LASSERT(lock->l_export); + if (lock->l_export->exp_libclient) { + LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)" + " timeout, just cancelling lock", ast_type, + libcfs_nid2str(peer.nid)); + ldlm_lock_cancel(lock); + rc = -ERESTART; + } else if (lock->l_flags & LDLM_FL_CANCEL) { + LDLM_DEBUG(lock, "%s AST timeout from nid %s, but " + "cancel was received (AST reply lost?)", + ast_type, libcfs_nid2str(peer.nid)); + ldlm_lock_cancel(lock); + rc = -ERESTART; + } else { + ldlm_del_waiting_lock(lock); + ldlm_failed_ast(lock, rc, ast_type); + } + } else if (rc) { + if (rc == -EINVAL) + LDLM_DEBUG(lock, "client (nid %s) returned %d" + " from %s AST - normal race", + libcfs_nid2str(peer.nid), + req->rq_repmsg ? + lustre_msg_get_status(req->rq_repmsg) : -1, + ast_type); + else + LDLM_ERROR(lock, "client (nid %s) returned %d " + "from %s AST", libcfs_nid2str(peer.nid), + (req->rq_repmsg != NULL) ? + lustre_msg_get_status(req->rq_repmsg) : 0, + ast_type); + ldlm_lock_cancel(lock); + /* Server-side AST functions are called from ldlm_reprocess_all, + * which needs to be told to please restart its reprocessing. */ + rc = -ERESTART; + } + + return rc; +} + +static int ldlm_cb_interpret(const struct lu_env *env, + struct ptlrpc_request *req, void *data, int rc) +{ + struct ldlm_cb_set_arg *arg; + struct ldlm_lock *lock; + ENTRY; + + LASSERT(data != NULL); + + arg = req->rq_async_args.pointer_arg[0]; + lock = req->rq_async_args.pointer_arg[1]; + LASSERT(lock != NULL); + if (rc != 0) { + /* If client canceled the lock but the cancel has not + * been recieved yet, we need to update lvbo to have the + * proper attributes cached. */ + if (rc == -EINVAL && arg->type == LDLM_BL_CALLBACK) + ldlm_res_lvbo_update(lock->l_resource, NULL, + 0, 1); + rc = ldlm_handle_ast_error(lock, req, rc, + arg->type == LDLM_BL_CALLBACK + ? "blocking" : "completion"); + } + + LDLM_LOCK_RELEASE(lock); + + if (rc == -ERESTART) + atomic_set(&arg->restart, 1); + + RETURN(0); +} + +static inline int ldlm_bl_and_cp_ast_fini(struct ptlrpc_request *req, + struct ldlm_cb_set_arg *arg, + struct ldlm_lock *lock, + int instant_cancel) { - const struct ptlrpc_connection *conn = lock->l_export->exp_connection; - char str[PTL_NALFMT_SIZE]; - - CERROR("%s AST failed (%d) for res "LPU64"/"LPU64 - ", mode %s: evicting client %s@%s NID "LPX64" (%s)\n", - ast_type, rc, - lock->l_resource->lr_name.name[0], - lock->l_resource->lr_name.name[1], - ldlm_lockname[lock->l_granted_mode], - lock->l_export->exp_client_uuid.uuid, - conn->c_remote_uuid.uuid, conn->c_peer.peer_nid, - portals_nid2str(conn->c_peer.peer_ni->pni_number, - conn->c_peer.peer_nid, str)); - ptlrpc_fail_export(lock->l_export); + int rc = 0; + ENTRY; + + if (unlikely(instant_cancel)) { + rc = ptl_send_rpc(req, 1); + ptlrpc_req_finished(req); + if (rc == 0) + /* If we cancelled the lock, we need to restart + * ldlm_reprocess_queue */ + atomic_set(&arg->restart, 1); + } else { + LDLM_LOCK_GET(lock); + ptlrpc_set_add_req(arg->set, req); + } + + RETURN(rc); } +/* + * ->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 + * to wait for client response. + */ int ldlm_server_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc, void *data, int flag) { - struct ldlm_request *body; - struct ptlrpc_request *req; - int rc = 0, size = sizeof(*body); + struct ldlm_cb_set_arg *arg = data; + struct ldlm_request *body; + struct ptlrpc_request *req; + int instant_cancel = 0; + int rc = 0; ENTRY; - if (flag == LDLM_CB_CANCELING) { + if (flag == LDLM_CB_CANCELING) /* Don't need to do anything here. */ RETURN(0); - } LASSERT(lock); + LASSERT(data != NULL); + 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); + } - l_lock(&lock->l_resource->lr_namespace->ns_lock); + req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse, + &RQF_LDLM_BL_CALLBACK, + LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK); + if (req == NULL) + RETURN(-ENOMEM); + + req->rq_async_args.pointer_arg[0] = arg; + req->rq_async_args.pointer_arg[1] = 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 */ - l_unlock(&lock->l_resource->lr_namespace->ns_lock); - LDLM_DEBUG(lock, "lock not granted, not sending blocking AST"); RETURN(0); + 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? */ - l_unlock(&lock->l_resource->lr_namespace->ns_lock); + unlock_res(lock->l_resource); + ptlrpc_req_finished(req); RETURN(0); } -#if 0 - if (LTIME_S(CURRENT_TIME) - lock->l_export->exp_last_request_time > 30){ - ldlm_failed_ast(lock, -ETIMEDOUT, "Not-attempted blocking"); - RETURN(-ETIMEDOUT); - } -#endif - - req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse, - LDLM_BL_CALLBACK, 1, &size, NULL); - if (!req) - RETURN(-ENOMEM); + if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) + instant_cancel = 1; - body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body)); - memcpy(&body->lock_handle1, &lock->l_remote_handle, - sizeof(body->lock_handle1)); - memcpy(&body->lock_desc, desc, sizeof(*desc)); + 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); LDLM_DEBUG(lock, "server preparing blocking AST"); - req->rq_replen = lustre_msg_size(0, NULL); - if (lock->l_granted_mode == lock->l_req_mode) + 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); - l_unlock(&lock->l_resource->lr_namespace->ns_lock); + unlock_res(lock->l_resource); + } req->rq_send_state = LUSTRE_IMP_FULL; - req->rq_timeout = 2; /* 2 second timeout for initial AST reply */ - rc = ptlrpc_queue_wait(req); - if (rc == -ETIMEDOUT || rc == -EINTR) { -#ifdef __KERNEL__ - ldlm_del_waiting_lock(lock); - ldlm_failed_ast(lock, rc, "blocking"); -#else - /* XXX - * Here we treat all clients as liblustre. When BLOCKING AST - * timeout we don't evicting the client and only cancel - * the lock. - * restore to orignial implementation later!!! - * XXX - */ - CERROR("BLOCKING AST to client (nid "LPU64") timeout, " - "simply cancel lock 0x%p\n", - req->rq_connection->c_peer.peer_nid, lock); - ldlm_lock_cancel(lock); - rc = -ERESTART; -#endif - } else if (rc) { - if (rc == -EINVAL) - CDEBUG(D_DLMTRACE, "client (nid "LPU64") returned %d " - "from blocking AST for lock %p--normal race\n", - req->rq_connection->c_peer.peer_nid, - req->rq_repmsg->status, lock); - else if (rc == -ENOTCONN) - CDEBUG(D_DLMTRACE, "client (nid "LPU64") returned %d " - "from blocking AST for lock %p--this client was " - "probably rebooted while it held a lock, nothing" - " serious\n",req->rq_connection->c_peer.peer_nid, - req->rq_repmsg->status, lock); - else - CDEBUG(D_ERROR, "client (nid "LPU64") returned %d " - "from blocking AST for lock %p\n", - req->rq_connection->c_peer.peer_nid, - (req->rq_repmsg != NULL)? - req->rq_repmsg->status : 0, - lock); - LDLM_DEBUG(lock, "client sent rc %d rq_status %d from blocking " - "AST", rc, req->rq_status); - ldlm_lock_cancel(lock); - /* Server-side AST functions are called from ldlm_reprocess_all, - * which needs to be told to please restart its reprocessing. */ - rc = -ERESTART; - } + /* ptlrpc_prep_req already set timeout */ + if (AT_OFF) + req->rq_timeout = ldlm_get_rq_timeout(); - ptlrpc_req_finished(req); + 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); - RETURN(rc); -} + rc = ldlm_bl_and_cp_ast_fini(req, arg, lock, instant_cancel); -/* XXX copied from ptlrpc/service.c */ -static long timeval_sub(struct timeval *large, struct timeval *small) -{ - return (large->tv_sec - small->tv_sec) * 1000000 + - (large->tv_usec - small->tv_usec); + RETURN(rc); } int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data) { - struct ldlm_request *body; - struct ptlrpc_request *req; - struct timeval granted_time; - long total_enqueue_wait; - int rc = 0, size = sizeof(*body); + struct ldlm_cb_set_arg *arg = data; + struct ldlm_request *body; + struct ptlrpc_request *req; + struct timeval granted_time; + long total_enqueue_wait; + int instant_cancel = 0; + int rc = 0; ENTRY; - if (lock == NULL) { - LBUG(); - RETURN(-EINVAL); - } + LASSERT(lock != NULL); + LASSERT(data != NULL); do_gettimeofday(&granted_time); - total_enqueue_wait = timeval_sub(&granted_time, &lock->l_enqueued_time); + total_enqueue_wait = cfs_timeval_sub(&granted_time, + &lock->l_enqueued_time, NULL); - if (total_enqueue_wait / 1000000 > obd_timeout) - LDLM_ERROR(lock, "enqueue wait took %ldus", total_enqueue_wait); + if (total_enqueue_wait / ONE_MILLION > obd_timeout) + /* non-fatal with AT - change to LDLM_DEBUG? */ + LDLM_ERROR(lock, "enqueue wait took %luus from "CFS_TIME_T, + total_enqueue_wait, lock->l_enqueued_time.tv_sec); - req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse, - LDLM_CP_CALLBACK, 1, &size, NULL); - if (!req) + req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse, + &RQF_LDLM_CP_CALLBACK); + if (req == NULL) RETURN(-ENOMEM); - body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body)); - memcpy(&body->lock_handle1, &lock->l_remote_handle, - sizeof(body->lock_handle1)); + lock_res_and_lock(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); + unlock_res_and_lock(lock); + + rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK); + if (rc) { + ptlrpc_request_free(req); + RETURN(rc); + } + + req->rq_async_args.pointer_arg[0] = arg; + req->rq_async_args.pointer_arg[1] = lock; + req->rq_interpret_reply = ldlm_cb_interpret; + req->rq_no_resend = 1; + body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); + + body->lock_handle[0] = lock->l_remote_handle; body->lock_flags = 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_and_lock(lock); + memcpy(lvb, lock->l_resource->lr_lvb_data, + lock->l_resource->lr_lvb_len); + unlock_res_and_lock(lock); + } LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)", total_enqueue_wait); - req->rq_replen = lustre_msg_size(0, NULL); + + /* Server-side enqueue wait time estimate, used in + __ldlm_add_waiting_lock to set future enqueue timers */ + at_add(&lock->l_resource->lr_namespace->ns_at_estimate, + total_enqueue_wait / ONE_MILLION); + + ptlrpc_request_set_replen(req); req->rq_send_state = LUSTRE_IMP_FULL; - req->rq_timeout = 2; /* 2 second timeout for initial AST reply */ + /* ptlrpc_prep_req already set timeout */ + if (AT_OFF) + req->rq_timeout = ldlm_get_rq_timeout(); /* We only send real blocking ASTs after the lock is granted */ - l_lock(&lock->l_resource->lr_namespace->ns_lock); + lock_res_and_lock(lock); if (lock->l_flags & LDLM_FL_AST_SENT) { body->lock_flags |= LDLM_FL_AST_SENT; - ldlm_add_waiting_lock(lock); /* start the lock-timeout clock */ + + /* We might get here prior to ldlm_handle_enqueue setting + * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock + * into waiting list, but this is safe and similar code in + * ldlm_handle_enqueue will call ldlm_lock_cancel() still, + * that would not only cancel the lock, but will also remove + * it from waiting list */ + if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) { + unlock_res_and_lock(lock); + ldlm_lock_cancel(lock); + instant_cancel = 1; + lock_res_and_lock(lock); + } else { + /* start the lock-timeout clock */ + ldlm_add_waiting_lock(lock); + } } - l_unlock(&lock->l_resource->lr_namespace->ns_lock); + unlock_res_and_lock(lock); + + 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_CP_CALLBACK - LDLM_FIRST_OPC); + + rc = ldlm_bl_and_cp_ast_fini(req, arg, lock, instant_cancel); + + RETURN(rc); +} + +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; + ENTRY; + + LASSERT(lock != NULL); + + req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse, + &RQF_LDLM_GL_CALLBACK, + LUSTRE_DLM_VERSION, LDLM_GL_CALLBACK); + + if (req == NULL) + RETURN(-ENOMEM); + + 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); + + lock_res_and_lock(lock); + req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, + lock->l_resource->lr_lvb_len); + unlock_res_and_lock(lock); + res = lock->l_resource; + ptlrpc_request_set_replen(req); + + + req->rq_send_state = LUSTRE_IMP_FULL; + /* ptlrpc_prep_req already set timeout */ + if (AT_OFF) + req->rq_timeout = ldlm_get_rq_timeout(); + + 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); - if (rc == -ETIMEDOUT || rc == -EINTR) { - ldlm_del_waiting_lock(lock); - ldlm_failed_ast(lock, rc, "completion"); - } else if (rc) { - LDLM_ERROR(lock, "client sent rc %d rq_status %d from " - "completion AST\n", rc, req->rq_status); - ldlm_lock_cancel(lock); - /* Server-side AST functions are called from ldlm_reprocess_all, - * which needs to be told to please restart its reprocessing. */ - rc = -ERESTART; - } + if (rc == -ELDLM_NO_LOCK_DATA) + LDLM_DEBUG(lock, "lost race - client has a lock but no inode"); + else if (rc != 0) + rc = ldlm_handle_ast_error(lock, req, rc, "glimpse"); + else + rc = ldlm_res_lvbo_update(res, req->rq_repmsg, + REPLY_REC_OFF, 1); ptlrpc_req_finished(req); + if (rc == -ERESTART) + ldlm_reprocess_all(res); RETURN(rc); } -int ldlm_handle_enqueue(struct ptlrpc_request *req, - ldlm_completion_callback completion_callback, - ldlm_blocking_callback blocking_callback) +#ifdef __KERNEL__ +extern unsigned long long lu_time_stamp_get(void); +#else +#define lu_time_stamp_get() time(NULL) +#endif + +static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req, + struct lprocfs_stats *srv_stats) +{ + int lock_type = 0, op = 0; + + lock_type = dlm_req->lock_desc.l_resource.lr_type; + + switch (lock_type) { + case LDLM_PLAIN: + op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE; + break; + case LDLM_EXTENT: + if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) + op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE; + else + op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE; + break; + case LDLM_FLOCK: + op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE; + break; + case LDLM_IBITS: + op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE; + break; + default: + op = 0; + break; + } + + if (op) + lprocfs_counter_incr(srv_stats, op); + + return ; +} + +/* + * Main server-side entry point into LDLM. This is called by ptlrpc service + * threads to carry out client lock enqueueing requests. + */ +int ldlm_handle_enqueue0(struct ldlm_namespace *ns, + struct ptlrpc_request *req, + const struct ldlm_request *dlm_req, + const struct ldlm_callback_suite *cbs) { - struct obd_device *obddev = req->rq_export->exp_obd; struct ldlm_reply *dlm_rep; - struct ldlm_request *dlm_req; - int rc, size = sizeof(*dlm_rep), cookielen = 0; __u32 flags; - ldlm_error_t err; + ldlm_error_t err = ELDLM_OK; struct ldlm_lock *lock = NULL; void *cookie = NULL; + int rc = 0; ENTRY; LDLM_DEBUG_NOLOCK("server-side enqueue handler START"); - dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req), - lustre_swab_ldlm_request); - if (dlm_req == NULL) { - CERROR ("Can't unpack dlm_req\n"); - RETURN (-EFAULT); + ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF); + flags = 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 (req->rq_export && req->rq_export->exp_nid_stats && + req->rq_export->exp_nid_stats->nid_ldlm_stats) + lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats, + LDLM_ENQUEUE - LDLM_FIRST_OPC); + + if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE || + dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) { + DEBUG_REQ(D_ERROR, req, "invalid lock request type %d", + dlm_req->lock_desc.l_resource.lr_type); + GOTO(out, rc = -EFAULT); } - flags = dlm_req->lock_flags; - if (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN && - (flags & LDLM_FL_HAS_INTENT)) { - /* In this case, the reply buffer is allocated deep in - * local_lock_enqueue by the policy function. */ - cookie = req; - cookielen = sizeof(*req); - } else { - rc = lustre_pack_reply(req, 1, &size, NULL); - if (rc) { - CERROR("out of memory\n"); - RETURN(-ENOMEM); + if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE || + dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE || + dlm_req->lock_desc.l_req_mode & + (dlm_req->lock_desc.l_req_mode-1))) { + DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d", + dlm_req->lock_desc.l_req_mode); + GOTO(out, rc = -EFAULT); + } + + if (req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) { + if (unlikely(dlm_req->lock_desc.l_resource.lr_type == + LDLM_PLAIN)) { + DEBUG_REQ(D_ERROR, req, + "PLAIN lock request from IBITS client?"); + GOTO(out, rc = -EPROTO); } - if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN) { - cookie = &dlm_req->lock_desc.l_policy_data; - cookielen = sizeof(ldlm_policy_data_t); + } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type == + LDLM_IBITS)) { + DEBUG_REQ(D_ERROR, req, + "IBITS lock request from unaware client?"); + GOTO(out, rc = -EPROTO); + } + +#if 0 + /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check + against server's _CONNECT_SUPPORTED flags? (I don't want to use + ibits for mgc/mgs) */ + + /* INODEBITS_INTEROP: Perform conversion from plain lock to + * inodebits lock if client does not support them. */ + if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) && + (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) { + dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS; + dlm_req->lock_desc.l_policy_data.l_inodebits.bits = + MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE; + if (dlm_req->lock_desc.l_req_mode == LCK_PR) + dlm_req->lock_desc.l_req_mode = LCK_CR; + } +#endif + + if (unlikely(flags & LDLM_FL_REPLAY)) { + /* Find an existing lock in the per-export lock hash */ + lock = lustre_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); + GOTO(existing_lock, rc = 0); } } /* The lock's callback data might be set in the policy function */ - lock = ldlm_lock_create(obddev->obd_namespace, - &dlm_req->lock_handle2, - dlm_req->lock_desc.l_resource.lr_name, + 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, - blocking_callback, completion_callback, NULL); + cbs, NULL, 0); + if (!lock) - GOTO(out, err = -ENOMEM); + GOTO(out, rc = -ENOMEM); do_gettimeofday(&lock->l_enqueued_time); - memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1, - sizeof(lock->l_remote_handle)); + lock->l_remote_handle = dlm_req->lock_handle[0]; LDLM_DEBUG(lock, "server-side enqueue handler, new lock created"); - LASSERT(req->rq_export); + OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2); + /* Don't enqueue a lock onto the export if it has already + * been evicted. Cancel it now instead. (bug 3822) */ + if (req->rq_export->exp_failed) { + LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export); + GOTO(out, rc = -ENOTCONN); + } lock->l_export = class_export_get(req->rq_export); - l_lock(&lock->l_resource->lr_namespace->ns_lock); - list_add(&lock->l_export_chain, - &lock->l_export->exp_ldlm_data.led_held_locks); - l_unlock(&lock->l_resource->lr_namespace->ns_lock); - err = ldlm_lock_enqueue(obddev->obd_namespace, &lock, cookie, cookielen, - &flags); + if (lock->l_export->exp_lock_hash) + lustre_hash_add(lock->l_export->exp_lock_hash, + &lock->l_remote_handle, + &lock->l_exp_hash); + +existing_lock: + + if (flags & LDLM_FL_HAS_INTENT) { + /* In this case, the reply buffer is allocated deep in + * local_lock_enqueue by the policy function. */ + cookie = req; + } else { + lock_res_and_lock(lock); + 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); + } + unlock_res_and_lock(lock); + + if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR)) + GOTO(out, rc = -ENOMEM); + + rc = req_capsule_server_pack(&req->rq_pill); + if (rc) + GOTO(out, rc); + } + + if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN) + lock->l_policy_data = dlm_req->lock_desc.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); - dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep)); + 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); - if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN) { - memcpy(&dlm_rep->lock_policy_data, &lock->l_policy_data, - cookielen); - } - if (dlm_rep->lock_flags & LDLM_FL_LOCK_CHANGED) { - memcpy(&dlm_rep->lock_resource_name, &lock->l_resource->lr_name, - sizeof(dlm_rep->lock_resource_name)); - dlm_rep->lock_mode = lock->l_req_mode; - } /* We never send a blocking AST until the lock is granted, but * we can tell it right now */ - l_lock(&lock->l_resource->lr_namespace->ns_lock); - if (lock->l_flags & LDLM_FL_AST_SENT) + lock_res_and_lock(lock); + + /* Now take into account flags to be inherited from original lock + request both in reply to client and in our own lock flags. */ + dlm_rep->lock_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS; + lock->l_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS; + + /* Don't move a pending lock onto the export if it has already + * been evicted. Cancel it now instead. (bug 5683) */ + if (unlikely(req->rq_export->exp_failed || + 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; - l_unlock(&lock->l_resource->lr_namespace->ns_lock); + if (lock->l_granted_mode == lock->l_req_mode) { + /* + * Only cancel lock if it was granted, because it would + * be destroyed immediatelly and would never be granted + * in the future, causing timeouts on client. Not + * granted lock will be cancelled immediatelly after + * sending completion AST. + */ + if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) { + unlock_res_and_lock(lock); + ldlm_lock_cancel(lock); + lock_res_and_lock(lock); + } else + ldlm_add_waiting_lock(lock); + } + } + /* Make sure we never ever grant usual metadata locks to liblustre + clients */ + 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) || + !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){ + CERROR("Granting sync lock to libclient. " + "req fl %d, rep fl %d, lock fl %d\n", + dlm_req->lock_flags, dlm_rep->lock_flags, + lock->l_flags); + LDLM_ERROR(lock, "sync lock"); + if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) { + struct ldlm_intent *it; + + it = req_capsule_client_get(&req->rq_pill, + &RMF_LDLM_INTENT); + if (it != NULL) { + CERROR("This is intent %s ("LPU64")\n", + ldlm_it2str(it->opc), it->opc); + } + } + } + } + + unlock_res_and_lock(lock); EXIT; out: - req->rq_status = err; + req->rq_status = rc ?: err; /* return either error - bug 11190 */ + if (!req->rq_packed_final) { + err = lustre_pack_reply(req, 1, NULL, NULL); + if (rc == 0) + rc = err; + } /* 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)", err); + "(err=%d, rc=%d)", err, rc); + + lock_res_and_lock(lock); + if (rc == 0) { + if (lock->l_resource->lr_lvb_len > 0) { + void *lvb; + + lvb = req_capsule_server_get(&req->rq_pill, + &RMF_DLM_LVB); + LASSERTF(lvb != NULL, "req %p, lock %p\n", + req, lock); + + memcpy(lvb, lock->l_resource->lr_lvb_data, + lock->l_resource->lr_lvb_len); + } + } else { + ldlm_resource_unlink_lock(lock); + ldlm_lock_destroy_nolock(lock); + } + unlock_res_and_lock(lock); + if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK) ldlm_reprocess_all(lock->l_resource); - LDLM_LOCK_PUT(lock); + + LDLM_LOCK_RELEASE(lock); } - LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p)", lock); - return 0; + LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)", + lock, rc); + + return rc; } -int ldlm_handle_convert(struct ptlrpc_request *req) +int ldlm_handle_enqueue(struct ptlrpc_request *req, + ldlm_completion_callback completion_callback, + ldlm_blocking_callback blocking_callback, + ldlm_glimpse_callback glimpse_callback) { struct ldlm_request *dlm_req; + struct ldlm_callback_suite cbs = { + .lcs_completion = completion_callback, + .lcs_blocking = blocking_callback, + .lcs_glimpse = glimpse_callback + }; + int rc; + + dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); + if (dlm_req != NULL) { + rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace, + req, dlm_req, &cbs); + } else { + rc = -EFAULT; + } + return rc; +} + +int ldlm_handle_convert0(struct ptlrpc_request *req, + const struct ldlm_request *dlm_req) +{ struct ldlm_reply *dlm_rep; struct ldlm_lock *lock; - int rc, size = sizeof(*dlm_rep); + int rc; ENTRY; - dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req), - lustre_swab_ldlm_request); - if (dlm_req == NULL) { - CERROR ("Can't unpack dlm_req\n"); - RETURN (-EFAULT); - } + if (req->rq_export && req->rq_export->exp_nid_stats && + req->rq_export->exp_nid_stats->nid_ldlm_stats) + lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats, + LDLM_CONVERT - LDLM_FIRST_OPC); - rc = lustre_pack_reply(req, 1, &size, NULL); - if (rc) { - CERROR("out of memory\n"); - RETURN(-ENOMEM); - } - dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep)); + rc = req_capsule_server_pack(&req->rq_pill); + if (rc) + RETURN(rc); + + dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP); dlm_rep->lock_flags = dlm_req->lock_flags; - lock = ldlm_handle2lock(&dlm_req->lock_handle1); + lock = ldlm_handle2lock(&dlm_req->lock_handle[0]); if (!lock) { req->rq_status = EINVAL; } else { + void *res = NULL; + LDLM_DEBUG(lock, "server-side convert handler START"); - ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode, - &dlm_rep->lock_flags); - if (ldlm_del_waiting_lock(lock)) - CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock); - req->rq_status = 0; + + do_gettimeofday(&lock->l_enqueued_time); + res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode, + &dlm_rep->lock_flags); + if (res) { + if (ldlm_del_waiting_lock(lock)) + LDLM_DEBUG(lock, "converted waiting lock"); + req->rq_status = 0; + } else { + req->rq_status = EDEADLOCK; + } } if (lock) { - ldlm_reprocess_all(lock->l_resource); + if (!req->rq_status) + ldlm_reprocess_all(lock->l_resource); LDLM_DEBUG(lock, "server-side convert handler END"); LDLM_LOCK_PUT(lock); } else @@ -622,89 +1226,139 @@ int ldlm_handle_convert(struct ptlrpc_request *req) RETURN(0); } -int ldlm_handle_cancel(struct ptlrpc_request *req) +int ldlm_handle_convert(struct ptlrpc_request *req) { + int rc; struct ldlm_request *dlm_req; + + dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); + if (dlm_req != NULL) { + rc = ldlm_handle_convert0(req, dlm_req); + } else { + CERROR ("Can't unpack dlm_req\n"); + rc = -EFAULT; + } + return rc; +} + +/* Cancel all the locks whos handles are packed into ldlm_request */ +int ldlm_request_cancel(struct ptlrpc_request *req, + const struct ldlm_request *dlm_req, int first) +{ + struct ldlm_resource *res, *pres = NULL; struct ldlm_lock *lock; - char str[PTL_NALFMT_SIZE]; + int i, count, done = 0; + ENTRY; + + count = dlm_req->lock_count ? dlm_req->lock_count : 1; + if (first >= count) + 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) + RETURN(0); + + LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, " + "starting at %d", count, first); + + for (i = first; i < count; i++) { + lock = ldlm_handle2lock(&dlm_req->lock_handle[i]); + if (!lock) { + LDLM_DEBUG_NOLOCK("server-side cancel handler stale " + "lock (cookie "LPU64")", + dlm_req->lock_handle[i].cookie); + continue; + } + + res = lock->l_resource; + done++; + + if (res != pres) { + if (pres != NULL) { + ldlm_reprocess_all(pres); + LDLM_RESOURCE_DELREF(pres); + ldlm_resource_putref(pres); + } + if (res != NULL) { + ldlm_resource_getref(res); + LDLM_RESOURCE_ADDREF(res); + ldlm_res_lvbo_update(res, NULL, 0, 1); + } + pres = res; + } + ldlm_lock_cancel(lock); + LDLM_LOCK_PUT(lock); + } + if (pres != NULL) { + ldlm_reprocess_all(pres); + LDLM_RESOURCE_DELREF(pres); + ldlm_resource_putref(pres); + } + LDLM_DEBUG_NOLOCK("server-side cancel handler END"); + RETURN(done); +} + +int ldlm_handle_cancel(struct ptlrpc_request *req) +{ + struct ldlm_request *dlm_req; int rc; ENTRY; - dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req), - lustre_swab_ldlm_request); + dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); if (dlm_req == NULL) { - CERROR("bad request buffer for cancel\n"); + CDEBUG(D_INFO, "bad request buffer for cancel\n"); RETURN(-EFAULT); } - rc = lustre_pack_reply(req, 0, NULL, NULL); - if (rc) { - CERROR("out of memory\n"); - RETURN(-ENOMEM); - } + if (req->rq_export && req->rq_export->exp_nid_stats && + req->rq_export->exp_nid_stats->nid_ldlm_stats) + lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats, + LDLM_CANCEL - LDLM_FIRST_OPC); - lock = ldlm_handle2lock(&dlm_req->lock_handle1); - if (!lock) { - CERROR("received cancel for unknown lock cookie "LPX64 - " from nid "LPX64" (%s)\n", dlm_req->lock_handle1.cookie, - req->rq_connection->c_peer.peer_nid, - portals_nid2str(req->rq_connection->c_peer.peer_ni->pni_number, - req->rq_connection->c_peer.peer_nid, str)); - LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock " - "(cookie "LPU64")", - dlm_req->lock_handle1.cookie); + rc = req_capsule_server_pack(&req->rq_pill); + if (rc) + RETURN(rc); + + if (!ldlm_request_cancel(req, dlm_req, 0)) req->rq_status = ESTALE; - } else { - LDLM_DEBUG(lock, "server-side cancel handler START"); - ldlm_lock_cancel(lock); - if (ldlm_del_waiting_lock(lock)) - CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock); - req->rq_status = 0; - } if (ptlrpc_reply(req) != 0) LBUG(); - if (lock) { - ldlm_reprocess_all(lock->l_resource); - LDLM_DEBUG(lock, "server-side cancel handler END"); - LDLM_LOCK_PUT(lock); - } - RETURN(0); } -static void ldlm_handle_bl_callback(struct ldlm_namespace *ns, - struct ldlm_lock_desc *ld, - struct ldlm_lock *lock) +void ldlm_handle_bl_callback(struct ldlm_namespace *ns, + struct ldlm_lock_desc *ld, struct ldlm_lock *lock) { int do_ast; ENTRY; - l_lock(&ns->ns_lock); LDLM_DEBUG(lock, "client blocking AST callback handler START"); + lock_res_and_lock(lock); lock->l_flags |= LDLM_FL_CBPENDING; + + if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) + lock->l_flags |= LDLM_FL_CANCEL; + do_ast = (!lock->l_readers && !lock->l_writers); + unlock_res_and_lock(lock); if (do_ast) { LDLM_DEBUG(lock, "already unused, calling " "callback (%p)", lock->l_blocking_ast); - if (lock->l_blocking_ast != NULL) { - l_unlock(&ns->ns_lock); - l_check_no_ns_lock(ns); + if (lock->l_blocking_ast != NULL) lock->l_blocking_ast(lock, ld, lock->l_ast_data, LDLM_CB_BLOCKING); - l_lock(&ns->ns_lock); - } } else { LDLM_DEBUG(lock, "Lock still has references, will be" " cancelled later"); } LDLM_DEBUG(lock, "client blocking callback handler END"); - l_unlock(&ns->ns_lock); - LDLM_LOCK_PUT(lock); + LDLM_LOCK_RELEASE(lock); EXIT; } @@ -713,11 +1367,31 @@ static void ldlm_handle_cp_callback(struct ptlrpc_request *req, struct ldlm_request *dlm_req, struct ldlm_lock *lock) { - LIST_HEAD(ast_list); + CFS_LIST_HEAD(ast_list); ENTRY; - l_lock(&ns->ns_lock); - LDLM_DEBUG(lock, "client completion callback handler START"); + LDLM_DEBUG(lock, "client completion callback handler START"); + + if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) { + int to = cfs_time_seconds(1); + while (to > 0) { + cfs_schedule_timeout(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. */ @@ -725,78 +1399,182 @@ static void ldlm_handle_cp_callback(struct ptlrpc_request *req, 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) - memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data, - sizeof(lock->l_policy_data)); + + if (lock->l_resource->lr_type != LDLM_PLAIN) { + lock->l_policy_data = dlm_req->lock_desc.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) { - ldlm_lock_change_resource(ns, lock, - dlm_req->lock_desc.l_resource.lr_name); + 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; + } 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) { - lock->l_flags |= LDLM_FL_CBPENDING; + /* 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; LDLM_DEBUG(lock, "completion AST includes blocking AST"); } - lock->l_resource->lr_tmp = &ast_list; - ldlm_grant_lock(lock, req, sizeof(*req), 1); - lock->l_resource->lr_tmp = NULL; + 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_swab_get(&req->rq_pill, + &RMF_DLM_LVB, + (void *)lock->l_lvb_swabber); + memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len); + } + } + + ldlm_grant_lock(lock, &ast_list); + unlock_res_and_lock(lock); + LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work"); - l_unlock(&ns->ns_lock); - LDLM_LOCK_PUT(lock); - ldlm_run_ast_work(ns, &ast_list); + ldlm_run_ast_work(&ast_list, LDLM_WORK_CP_AST); LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)", lock); + LDLM_LOCK_RELEASE(lock); + EXIT; +} + +static void ldlm_handle_gl_callback(struct ptlrpc_request *req, + struct ldlm_namespace *ns, + struct ldlm_request *dlm_req, + struct ldlm_lock *lock) +{ + int rc = -ENOSYS; + ENTRY; + + LDLM_DEBUG(lock, "client glimpse AST callback handler"); + + if (lock->l_glimpse_ast != NULL) + rc = lock->l_glimpse_ast(lock, req); + + if (req->rq_repmsg != NULL) { + ptlrpc_reply(req); + } else { + req->rq_status = rc; + ptlrpc_error(req); + } + + lock_res_and_lock(lock); + if (lock->l_granted_mode == LCK_PW && + !lock->l_readers && !lock->l_writers && + cfs_time_after(cfs_time_current(), + cfs_time_add(lock->l_last_used, + cfs_time_seconds(10)))) { + unlock_res_and_lock(lock); + if (ldlm_bl_to_thread_lock(ns, NULL, lock)) + ldlm_handle_bl_callback(ns, NULL, lock); + + EXIT; + return; + } + unlock_res_and_lock(lock); + LDLM_LOCK_RELEASE(lock); EXIT; } static int ldlm_callback_reply(struct ptlrpc_request *req, int rc) { + if (req->rq_no_reply) + return 0; + req->rq_status = rc; - rc = lustre_pack_reply(req, 0, NULL, NULL); - if (rc) - return rc; + if (!req->rq_packed_final) { + rc = lustre_pack_reply(req, 1, NULL, NULL); + if (rc) + return rc; + } return ptlrpc_reply(req); } #ifdef __KERNEL__ -static int ldlm_bl_to_thread(struct ldlm_state *ldlm, struct ldlm_namespace *ns, - struct ldlm_lock_desc *ld, struct ldlm_lock *lock) +static int ldlm_bl_to_thread(struct ldlm_namespace *ns, + struct ldlm_lock_desc *ld, struct ldlm_lock *lock, + struct list_head *cancels, int count) { - struct ldlm_bl_pool *blp = ldlm->ldlm_bl_pool; + struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool; struct ldlm_bl_work_item *blwi; ENTRY; + if (cancels && count == 0) + RETURN(0); + OBD_ALLOC(blwi, sizeof(*blwi)); if (blwi == NULL) RETURN(-ENOMEM); blwi->blwi_ns = ns; - blwi->blwi_ld = *ld; - blwi->blwi_lock = lock; - + if (ld != NULL) + blwi->blwi_ld = *ld; + if (count) { + list_add(&blwi->blwi_head, cancels); + list_del_init(cancels); + blwi->blwi_count = count; + } else { + blwi->blwi_lock = lock; + } spin_lock(&blp->blp_lock); - list_add_tail(&blwi->blwi_entry, &blp->blp_list); - wake_up(&blp->blp_waitq); + if (lock && lock->l_flags & LDLM_FL_DISCARD_DATA) { + /* 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); + } + cfs_waitq_signal(&blp->blp_waitq); spin_unlock(&blp->blp_lock); RETURN(0); } #endif +int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld, + struct ldlm_lock *lock) +{ +#ifdef __KERNEL__ + RETURN(ldlm_bl_to_thread(ns, ld, lock, NULL, 0)); +#else + RETURN(-ENOSYS); +#endif +} + +int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld, + struct list_head *cancels, int count) +{ +#ifdef __KERNEL__ + RETURN(ldlm_bl_to_thread(ns, ld, NULL, cancels, count)); +#else + RETURN(-ENOSYS); +#endif +} + +/* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */ static int ldlm_callback_handler(struct ptlrpc_request *req) { struct ldlm_namespace *ns; struct ldlm_request *dlm_req; struct ldlm_lock *lock; - char str[PTL_NALFMT_SIZE]; int rc; ENTRY; @@ -805,106 +1583,157 @@ static int ldlm_callback_handler(struct ptlrpc_request *req) * incoming request message body, but I am responsible for the * message buffers. */ + /* do nothing for sec context finalize */ + if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI) + RETURN(0); + + req_capsule_init(&req->rq_pill, req, RCL_SERVER); + if (req->rq_export == NULL) { struct ldlm_request *dlm_req; - CDEBUG(D_RPCTRACE, "operation %d from nid "LPX64" (%s) with bad " - "export cookie "LPX64" (ptl req %d/rep %d); this is " + CDEBUG(D_RPCTRACE, "operation %d from %s with bad " + "export cookie "LPX64"; this is " "normal if this node rebooted with a lock held\n", - req->rq_reqmsg->opc, req->rq_connection->c_peer.peer_nid, - portals_nid2str(req->rq_connection->c_peer.peer_ni->pni_number, - req->rq_connection->c_peer.peer_nid, str), - req->rq_reqmsg->handle.cookie, - req->rq_request_portal, req->rq_reply_portal); - - dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req), - lustre_swab_ldlm_request); + lustre_msg_get_opc(req->rq_reqmsg), + libcfs_id2str(req->rq_peer), + lustre_msg_get_handle(req->rq_reqmsg)->cookie); + + req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK); + dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); if (dlm_req != NULL) CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n", - dlm_req->lock_handle1.cookie); + dlm_req->lock_handle[0].cookie); ldlm_callback_reply(req, -ENOTCONN); RETURN(0); } - if (req->rq_reqmsg->opc == LDLM_BL_CALLBACK) { - OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0); - } else if (req->rq_reqmsg->opc == LDLM_CP_CALLBACK) { - OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0); - } else if (req->rq_reqmsg->opc == OBD_LOG_CANCEL) { - OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0); - } else if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_CREATE) { - OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0); - } else if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_NEXT_BLOCK) { - OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0); - } else if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_READ_HEADER) { - OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0); - } else if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_CLOSE) { - OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0); - } else { - ldlm_callback_reply(req, -EPROTO); - RETURN(0); - } - LASSERT(req->rq_export != NULL); LASSERT(req->rq_export->exp_obd != NULL); -#ifdef ENABLE_ORPHANS - /* FIXME - how to send reply */ - if (req->rq_reqmsg->opc == OBD_LOG_CANCEL) { - int rc = llog_origin_handle_cancel(req); + 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; + case OBD_LOG_CANCEL: /* remove this eventually - for 1.4.0 compat */ + 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); - } - - if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_CREATE) { - int rc = llog_origin_handle_create(req); - req->rq_status = rc; - ptlrpc_reply(req); + 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); - } - - if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_NEXT_BLOCK) { - int rc = llog_origin_handle_next_block(req); - req->rq_status = rc; - ptlrpc_reply(req); + 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); - } - - if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_READ_HEADER) { - int rc = llog_origin_handle_read_header(req); - req->rq_status = rc; - ptlrpc_reply(req); + 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); + ldlm_callback_reply(req, rc); RETURN(0); - } - - if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_CLOSE) { - int rc = llog_origin_handle_close(req); + case LLOG_ORIGIN_HANDLE_NEXT_BLOCK: + req_capsule_set(&req->rq_pill, + &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK); + if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET)) + RETURN(0); + rc = llog_origin_handle_next_block(req); + ldlm_callback_reply(req, rc); + RETURN(0); + case LLOG_ORIGIN_HANDLE_READ_HEADER: + req_capsule_set(&req->rq_pill, + &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER); + if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET)) + RETURN(0); + rc = llog_origin_handle_read_header(req); + ldlm_callback_reply(req, rc); + RETURN(0); + case LLOG_ORIGIN_HANDLE_CLOSE: + if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET)) + RETURN(0); + rc = llog_origin_handle_close(req); ldlm_callback_reply(req, rc); RETURN(0); + default: + CERROR("unknown opcode %u\n", + lustre_msg_get_opc(req->rq_reqmsg)); + ldlm_callback_reply(req, -EPROTO); + RETURN(0); } -#endif + ns = req->rq_export->exp_obd->obd_namespace; LASSERT(ns != NULL); - dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req), - lustre_swab_ldlm_request); + req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK); + + dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); if (dlm_req == NULL) { - CERROR ("can't unpack dlm_req\n"); - ldlm_callback_reply (req, -EPROTO); - RETURN (0); + ldlm_callback_reply(req, -EPROTO); + RETURN(0); + } + + /* Force a known safe race, send a cancel to the server for a lock + * 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]); + if (rc < 0) + CERROR("ldlm_cli_cancel: %d\n", rc); } - lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle1); + lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0); if (!lock) { - CDEBUG(D_INODE, "callback on lock "LPX64" - lock disappeared\n", - dlm_req->lock_handle1.cookie); + CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock " + "disappeared\n", dlm_req->lock_handle[0].cookie); ldlm_callback_reply(req, -EINVAL); RETURN(0); } /* 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); + if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) { + /* If somebody cancels locks and cache is already droped, + * 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)) { + LDLM_DEBUG(lock, "callback on lock " + LPX64" - lock disappeared\n", + dlm_req->lock_handle[0].cookie); + unlock_res_and_lock(lock); + LDLM_LOCK_RELEASE(lock); + ldlm_callback_reply(req, -EINVAL); + RETURN(0); + } + /* BL_AST locks are not needed in lru. + * let ldlm_cancel_lru() be fast. */ + ldlm_lock_remove_from_lru(lock); + lock->l_flags |= LDLM_FL_BL_AST; + } + unlock_res_and_lock(lock); /* We want the ost thread to get this reply so that it can respond * to ost requests (write cache writeback) that might be triggered @@ -914,25 +1743,27 @@ static int ldlm_callback_handler(struct ptlrpc_request *req) * cancelling right now, because it's unused, or have an intent result * in the reply, so we might have to push the responsibility for sending * the reply down into the AST handlers, alas. */ - if (req->rq_reqmsg->opc != LDLM_BL_CALLBACK) - ldlm_callback_reply(req, 0); - switch (req->rq_reqmsg->opc) { + switch (lustre_msg_get_opc(req->rq_reqmsg)) { case LDLM_BL_CALLBACK: CDEBUG(D_INODE, "blocking ast\n"); -#ifdef __KERNEL__ - rc = ldlm_bl_to_thread(ldlm, ns, &dlm_req->lock_desc, lock); - ldlm_callback_reply(req, rc); -#else - rc = 0; - ldlm_callback_reply(req, rc); - ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock); -#endif + req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK); + if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)) + ldlm_callback_reply(req, 0); + if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock)) + ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock); break; case LDLM_CP_CALLBACK: CDEBUG(D_INODE, "completion ast\n"); + req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK); + ldlm_callback_reply(req, 0); ldlm_handle_cp_callback(req, ns, dlm_req, lock); break; + case LDLM_GL_CALLBACK: + CDEBUG(D_INODE, "glimpse ast\n"); + req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK); + ldlm_handle_gl_callback(req, ns, dlm_req, lock); + break; default: LBUG(); /* checked above */ } @@ -950,48 +1781,126 @@ static int ldlm_cancel_handler(struct ptlrpc_request *req) * incoming request message body, but I am responsible for the * message buffers. */ + req_capsule_init(&req->rq_pill, req, RCL_SERVER); + if (req->rq_export == NULL) { struct ldlm_request *dlm_req; - CERROR("operation %d with bad export (ptl req %d/rep %d)\n", - req->rq_reqmsg->opc, req->rq_request_portal, - req->rq_reply_portal); - CERROR("--> export cookie: "LPX64"\n", - req->rq_reqmsg->handle.cookie); - dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req), - lustre_swab_ldlm_request); + + CERROR("operation %d from %s with bad export cookie "LPU64"\n", + lustre_msg_get_opc(req->rq_reqmsg), + libcfs_id2str(req->rq_peer), + lustre_msg_get_handle(req->rq_reqmsg)->cookie); + + req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK); + dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); if (dlm_req != NULL) - ldlm_lock_dump_handle(D_ERROR, &dlm_req->lock_handle1); - RETURN(-ENOTCONN); + ldlm_lock_dump_handle(D_ERROR, + &dlm_req->lock_handle[0]); + ldlm_callback_reply(req, -ENOTCONN); + RETURN(0); } - switch (req->rq_reqmsg->opc) { + switch (lustre_msg_get_opc(req->rq_reqmsg)) { /* XXX FIXME move this back to mds/handler.c, bug 249 */ case LDLM_CANCEL: + req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL); CDEBUG(D_INODE, "cancel\n"); - OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0); + if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL)) + 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", req->rq_reqmsg->opc); - RETURN(-EINVAL); + CERROR("invalid opcode %d\n", + lustre_msg_get_opc(req->rq_reqmsg)); + req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK); + ldlm_callback_reply(req, -EINVAL); } RETURN(0); } +void ldlm_revoke_lock_cb(void *obj, void *data) +{ + struct list_head *rpc_list = data; + struct ldlm_lock *lock = obj; + + lock_res_and_lock(lock); + + if (lock->l_req_mode != lock->l_granted_mode) { + unlock_res_and_lock(lock); + return; + } + + LASSERT(lock->l_resource); + if (lock->l_resource->lr_type != LDLM_IBITS && + lock->l_resource->lr_type != LDLM_PLAIN) { + unlock_res_and_lock(lock); + return; + } + + if (lock->l_flags & LDLM_FL_AST_SENT) { + unlock_res_and_lock(lock); + return; + } + + 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) + lustre_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); +} + +void ldlm_revoke_export_locks(struct obd_export *exp) +{ + struct list_head rpc_list; + ENTRY; + + CFS_INIT_LIST_HEAD(&rpc_list); + lustre_hash_for_each_empty(exp->exp_lock_hash, + ldlm_revoke_lock_cb, &rpc_list); + ldlm_run_ast_work(&rpc_list, LDLM_WORK_REVOKE_AST); + + EXIT; +} + #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; spin_lock(&blp->blp_lock); - if (!list_empty(&blp->blp_list)) { - blwi = list_entry(blp->blp_list.next, struct ldlm_bl_work_item, - blwi_entry); + /* 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); @@ -999,49 +1908,99 @@ static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp) return blwi; } +/* This only contains temporary data until the thread starts */ struct ldlm_bl_thread_data { - 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; + + init_completion(&bltd.bltd_comp); + rc = cfs_kernel_thread(ldlm_bl_thread_main, &bltd, 0); + if (rc < 0) { + CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %d\n", + atomic_read(&blp->blp_num_threads), rc); + return rc; + } + wait_for_completion(&bltd.bltd_comp); + + return 0; +} + static int ldlm_bl_thread_main(void *arg) { - struct ldlm_bl_thread_data *bltd = arg; - struct ldlm_bl_pool *blp = bltd->bltd_blp; - unsigned long flags; + struct ldlm_bl_pool *blp; ENTRY; - /* XXX boiler-plate */ { - char name[sizeof(current->comm)]; - snprintf(name, sizeof(name) - 1, "ldlm_bl_%02d", - bltd->bltd_num); - kportal_daemonize(name); - } - SIGNAL_MASK_LOCK(current, flags); - sigfillset(¤t->blocked); - RECALC_SIGPENDING; - SIGNAL_MASK_UNLOCK(current, flags); + struct ldlm_bl_thread_data *bltd = arg; - atomic_inc(&blp->blp_num_threads); - complete(&blp->blp_comp); + blp = bltd->bltd_blp; + + bltd->bltd_num = atomic_inc_return(&blp->blp_num_threads) - 1; + atomic_inc(&blp->blp_busy_threads); + + snprintf(bltd->bltd_name, sizeof(bltd->bltd_name) - 1, + "ldlm_bl_%02d", bltd->bltd_num); + cfs_daemonize(bltd->bltd_name); + + complete(&bltd->bltd_comp); + /* cannot use bltd after this, it is only on caller's stack */ + } - while(1) { + while (1) { struct l_wait_info lwi = { 0 }; struct ldlm_bl_work_item *blwi = NULL; - l_wait_event_exclusive(blp->blp_waitq, - (blwi = ldlm_bl_get_work(blp)) != NULL, - &lwi); + blwi = ldlm_bl_get_work(blp); - if (blwi->blwi_ns == NULL) - break; + if (blwi == NULL) { + int busy; + + 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); - ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld, - blwi->blwi_lock); + 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))) + /* 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_count) { + /* The special case when we cancel locks in lru + * asynchronously, we pass the list of locks here. + * Thus lock is marked LDLM_FL_CANCELING, and already + * canceled locally. */ + ldlm_cli_cancel_list(&blwi->blwi_head, + blwi->blwi_count, NULL, 0); + } else { + ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld, + blwi->blwi_lock); + } OBD_FREE(blwi, sizeof(*blwi)); } + atomic_dec(&blp->blp_busy_threads); atomic_dec(&blp->blp_num_threads); complete(&blp->blp_comp); RETURN(0); @@ -1050,27 +2009,29 @@ static int ldlm_bl_thread_main(void *arg) #endif static int ldlm_setup(void); -static int ldlm_cleanup(int force); +static int ldlm_cleanup(void); int ldlm_get_ref(void) { int rc = 0; - down(&ldlm_ref_sem); + ENTRY; + mutex_down(&ldlm_ref_sem); if (++ldlm_refcount == 1) { rc = ldlm_setup(); if (rc) ldlm_refcount--; } - up(&ldlm_ref_sem); + mutex_up(&ldlm_ref_sem); RETURN(rc); } -void ldlm_put_ref(int force) +void ldlm_put_ref(void) { - down(&ldlm_ref_sem); + ENTRY; + mutex_down(&ldlm_ref_sem); if (ldlm_refcount == 1) { - int rc = ldlm_cleanup(force); + int rc = ldlm_cleanup(); if (rc) CERROR("ldlm_cleanup failed: %d\n", rc); else @@ -1078,53 +2039,154 @@ void ldlm_put_ref(int force) } else { ldlm_refcount--; } - up(&ldlm_ref_sem); + mutex_up(&ldlm_ref_sem); + + EXIT; +} + +/* + * Export handle<->lock hash operations. + */ +static unsigned +ldlm_export_lock_hash(lustre_hash_t *lh, void *key, unsigned mask) +{ + return lh_u64_hash(((struct lustre_handle *)key)->cookie, mask); +} + +static void * +ldlm_export_lock_key(struct hlist_node *hnode) +{ + struct ldlm_lock *lock; + ENTRY; + + lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash); + RETURN(&lock->l_remote_handle); +} + +static int +ldlm_export_lock_compare(void *key, struct hlist_node *hnode) +{ + ENTRY; + RETURN(lustre_handle_equal(ldlm_export_lock_key(hnode), key)); +} + +static void * +ldlm_export_lock_get(struct hlist_node *hnode) +{ + struct ldlm_lock *lock; + ENTRY; + + lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash); + LDLM_LOCK_GET(lock); + + RETURN(lock); +} + +static void * +ldlm_export_lock_put(struct hlist_node *hnode) +{ + struct ldlm_lock *lock; + ENTRY; + + lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash); + LDLM_LOCK_RELEASE(lock); + + RETURN(lock); +} + +static lustre_hash_ops_t ldlm_export_lock_ops = { + .lh_hash = ldlm_export_lock_hash, + .lh_key = ldlm_export_lock_key, + .lh_compare = ldlm_export_lock_compare, + .lh_get = ldlm_export_lock_get, + .lh_put = ldlm_export_lock_put +}; + +int ldlm_init_export(struct obd_export *exp) +{ + ENTRY; + + exp->exp_lock_hash = + lustre_hash_init(obd_uuid2str(&exp->exp_client_uuid), + 128, 65536, &ldlm_export_lock_ops, LH_REHASH); + + if (!exp->exp_lock_hash) + RETURN(-ENOMEM); + RETURN(0); +} +EXPORT_SYMBOL(ldlm_init_export); + +void ldlm_destroy_export(struct obd_export *exp) +{ + ENTRY; + lustre_hash_exit(exp->exp_lock_hash); + exp->exp_lock_hash = NULL; 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 ENTRY; - if (ldlm != NULL) + if (ldlm_state != NULL) RETURN(-EALREADY); - OBD_ALLOC(ldlm, sizeof(*ldlm)); - if (ldlm == NULL) + OBD_ALLOC(ldlm_state, sizeof(*ldlm_state)); + if (ldlm_state == NULL) RETURN(-ENOMEM); -#ifdef __KERNEL__ +#ifdef LPROCFS rc = ldlm_proc_setup(); if (rc != 0) GOTO(out_free, rc); #endif - ldlm->ldlm_cb_service = - ptlrpc_init_svc(LDLM_NEVENTS, LDLM_NBUFS, LDLM_BUFSIZE, - LDLM_MAXREQSIZE, LDLM_CB_REQUEST_PORTAL, - LDLM_CB_REPLY_PORTAL, - ldlm_callback_handler, "ldlm_cbd", - ldlm_svc_proc_dir); +#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 - if (!ldlm->ldlm_cb_service) { + ldlm_state->ldlm_cb_service = + ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE, + LDLM_MAXREPSIZE, LDLM_CB_REQUEST_PORTAL, + LDLM_CB_REPLY_PORTAL, 1800, + ldlm_callback_handler, "ldlm_cbd", + ldlm_svc_proc_dir, NULL, + ldlm_min_threads, ldlm_max_threads, + "ldlm_cb", + LCT_MD_THREAD|LCT_DT_THREAD); + + if (!ldlm_state->ldlm_cb_service) { CERROR("failed to start service\n"); GOTO(out_proc, rc = -ENOMEM); } - ldlm->ldlm_cancel_service = - ptlrpc_init_svc(LDLM_NEVENTS, LDLM_NBUFS, LDLM_BUFSIZE, - LDLM_MAXREQSIZE, LDLM_CANCEL_REQUEST_PORTAL, - LDLM_CANCEL_REPLY_PORTAL, - ldlm_cancel_handler, "ldlm_canceld", - ldlm_svc_proc_dir); - - if (!ldlm->ldlm_cancel_service) { + ldlm_state->ldlm_cancel_service = + ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE, + LDLM_MAXREPSIZE, LDLM_CANCEL_REQUEST_PORTAL, + LDLM_CANCEL_REPLY_PORTAL, 6000, + 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); + + if (!ldlm_state->ldlm_cancel_service) { CERROR("failed to start service\n"); GOTO(out_proc, rc = -ENOMEM); } @@ -1132,50 +2194,41 @@ static int ldlm_setup(void) OBD_ALLOC(blp, sizeof(*blp)); if (blp == NULL) GOTO(out_proc, rc = -ENOMEM); - ldlm->ldlm_bl_pool = blp; + ldlm_state->ldlm_bl_pool = blp; - atomic_set(&blp->blp_num_threads, 0); - init_waitqueue_head(&blp->blp_waitq); spin_lock_init(&blp->blp_lock); - - INIT_LIST_HEAD(&blp->blp_list); + CFS_INIT_LIST_HEAD(&blp->blp_list); + CFS_INIT_LIST_HEAD(&blp->blp_prio_list); + cfs_waitq_init(&blp->blp_waitq); + atomic_set(&blp->blp_num_threads, 0); + 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 < LDLM_NUM_THREADS; i++) { - struct ldlm_bl_thread_data bltd = { - .bltd_num = i, - .bltd_blp = blp, - }; - init_completion(&blp->blp_comp); - rc = kernel_thread(ldlm_bl_thread_main, &bltd, 0); - if (rc < 0) { - CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc); - LBUG(); + for (i = 0; i < blp->blp_min_threads; i++) { + rc = ldlm_bl_thread_start(blp); + if (rc < 0) GOTO(out_thread, rc); - } - wait_for_completion(&blp->blp_comp); } - rc = ptlrpc_start_n_threads(NULL, ldlm->ldlm_cancel_service, - LDLM_NUM_THREADS, "ldlm_cn"); - if (rc) { - LBUG(); + rc = ptlrpc_start_threads(NULL, ldlm_state->ldlm_cancel_service); + if (rc) GOTO(out_thread, rc); - } - rc = ptlrpc_start_n_threads(NULL, ldlm->ldlm_cb_service, - LDLM_NUM_THREADS, "ldlm_cb"); - if (rc) { - LBUG(); + rc = ptlrpc_start_threads(NULL, ldlm_state->ldlm_cb_service); + if (rc) GOTO(out_thread, rc); - } - INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks); - spin_lock_init(&expired_lock_thread.elt_lock); + CFS_INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks); expired_lock_thread.elt_state = ELT_STOPPED; - init_waitqueue_head(&expired_lock_thread.elt_waitq); + cfs_waitq_init(&expired_lock_thread.elt_waitq); - rc = kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FS); + CFS_INIT_LIST_HEAD(&waiting_locks_list); + spin_lock_init(&waiting_locks_spinlock); + cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0); + + rc = cfs_kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FILES); if (rc < 0) { CERROR("Cannot start ldlm expired-lock thread: %d\n", rc); GOTO(out_thread, rc); @@ -1183,46 +2236,51 @@ static int ldlm_setup(void) wait_event(expired_lock_thread.elt_waitq, expired_lock_thread.elt_state == ELT_READY); - - INIT_LIST_HEAD(&waiting_locks_list); - spin_lock_init(&waiting_locks_spinlock); - waiting_locks_timer.function = waiting_locks_callback; - waiting_locks_timer.data = 0; - init_timer(&waiting_locks_timer); #endif +#ifdef __KERNEL__ + rc = ldlm_pools_init(); + if (rc) + GOTO(out_thread, rc); +#endif RETURN(0); #ifdef __KERNEL__ out_thread: - ptlrpc_unregister_service(ldlm->ldlm_cancel_service); - ptlrpc_unregister_service(ldlm->ldlm_cb_service); + ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service); + ptlrpc_unregister_service(ldlm_state->ldlm_cb_service); #endif out_proc: -#ifdef __KERNEL__ +#ifdef LPROCFS ldlm_proc_cleanup(); out_free: #endif - OBD_FREE(ldlm, sizeof(*ldlm)); - ldlm = NULL; + OBD_FREE(ldlm_state, sizeof(*ldlm_state)); + ldlm_state = NULL; return rc; } -static int ldlm_cleanup(int force) +static int ldlm_cleanup(void) { #ifdef __KERNEL__ - struct ldlm_bl_pool *blp = ldlm->ldlm_bl_pool; + struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool; #endif ENTRY; - if (!list_empty(&ldlm_namespace_list)) { + 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_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 (atomic_read(&blp->blp_num_threads) > 0) { struct ldlm_bl_work_item blwi = { .blwi_ns = NULL }; @@ -1230,78 +2288,95 @@ static int ldlm_cleanup(int force) spin_lock(&blp->blp_lock); list_add_tail(&blwi.blwi_entry, &blp->blp_list); - wake_up(&blp->blp_waitq); + cfs_waitq_signal(&blp->blp_waitq); spin_unlock(&blp->blp_lock); wait_for_completion(&blp->blp_comp); } OBD_FREE(blp, sizeof(*blp)); - ptlrpc_stop_all_threads(ldlm->ldlm_cb_service); - ptlrpc_unregister_service(ldlm->ldlm_cb_service); - ptlrpc_stop_all_threads(ldlm->ldlm_cancel_service); - ptlrpc_unregister_service(ldlm->ldlm_cancel_service); + 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; - wake_up(&expired_lock_thread.elt_waitq); + cfs_waitq_signal(&expired_lock_thread.elt_waitq); 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); #endif - OBD_FREE(ldlm, sizeof(*ldlm)); - ldlm = NULL; + OBD_FREE(ldlm_state, sizeof(*ldlm_state)); + ldlm_state = NULL; RETURN(0); } int __init ldlm_init(void) { - ldlm_resource_slab = kmem_cache_create("ldlm_resources", + init_mutex(&ldlm_ref_sem); + init_mutex(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER)); + init_mutex(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT)); + ldlm_resource_slab = cfs_mem_cache_create("ldlm_resources", sizeof(struct ldlm_resource), 0, - SLAB_HWCACHE_ALIGN, NULL, NULL); + SLAB_HWCACHE_ALIGN); if (ldlm_resource_slab == NULL) return -ENOMEM; - ldlm_lock_slab = kmem_cache_create("ldlm_locks", + ldlm_lock_slab = cfs_mem_cache_create("ldlm_locks", sizeof(struct ldlm_lock), 0, - SLAB_HWCACHE_ALIGN, NULL, NULL); + SLAB_HWCACHE_ALIGN); if (ldlm_lock_slab == NULL) { - kmem_cache_destroy(ldlm_resource_slab); + cfs_mem_cache_destroy(ldlm_resource_slab); return -ENOMEM; } - l_lock_init(&ldlm_handle_lock); + ldlm_interval_slab = cfs_mem_cache_create("interval_node", + sizeof(struct ldlm_interval), + 0, SLAB_HWCACHE_ALIGN); + if (ldlm_interval_slab == NULL) { + cfs_mem_cache_destroy(ldlm_resource_slab); + cfs_mem_cache_destroy(ldlm_lock_slab); + return -ENOMEM; + } return 0; } void __exit ldlm_exit(void) { - if ( ldlm_refcount ) + int rc; + if (ldlm_refcount) CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount); - if (kmem_cache_destroy(ldlm_resource_slab) != 0) - CERROR("couldn't free ldlm resource slab\n"); - if (kmem_cache_destroy(ldlm_lock_slab) != 0) - CERROR("couldn't free ldlm lock slab\n"); + rc = cfs_mem_cache_destroy(ldlm_resource_slab); + LASSERTF(rc == 0, "couldn't free ldlm resource slab\n"); + 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_flock.c */ -EXPORT_SYMBOL(ldlm_flock_completion_ast); +/* 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_unregister_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_fast_match); +EXPORT_SYMBOL(ldlm_lock_fast_release); 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); @@ -1311,59 +2386,77 @@ EXPORT_SYMBOL(ldlm_lock_dump); EXPORT_SYMBOL(ldlm_lock_dump_handle); EXPORT_SYMBOL(ldlm_cancel_locks_for_export); EXPORT_SYMBOL(ldlm_reprocess_all_ns); +EXPORT_SYMBOL(ldlm_lock_allow_match); /* ldlm_request.c */ +EXPORT_SYMBOL(ldlm_completion_ast_async); 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_namespace_foreach_res); -EXPORT_SYMBOL(ldlm_change_cbdata); +EXPORT_SYMBOL(ldlm_resource_iterate); +EXPORT_SYMBOL(ldlm_cancel_resource_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); - -#if 0 -/* ldlm_test.c */ -EXPORT_SYMBOL(ldlm_test); -EXPORT_SYMBOL(ldlm_regression_start); -EXPORT_SYMBOL(ldlm_regression_stop); -#endif +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); - -/* l_lock.c */ -EXPORT_SYMBOL(l_lock); -EXPORT_SYMBOL(l_unlock); +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(target_abort_recovery); +EXPORT_SYMBOL(target_start_recovery_thread); +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); -EXPORT_SYMBOL(target_queue_final_reply); -EXPORT_SYMBOL(ldlm_put_lock_into_req); + +/* l_lock.c */ +EXPORT_SYMBOL(lock_res_and_lock); +EXPORT_SYMBOL(unlock_res_and_lock);