X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_lock.c;h=08eae3553dc3971a2c952e098536eb3abe19c84f;hp=ec0727862e4d0b0a2555643de91e0c2025ecee8c;hb=43c01b9d94e783295edc2d1432ed3af0076a0e51;hpb=5ba5b49768995332e0c94ba70ae4d65aaa2f9972 diff --git a/lustre/ldlm/ldlm_lock.c b/lustre/ldlm/ldlm_lock.c index ec07278..08eae35 100644 --- a/lustre/ldlm/ldlm_lock.c +++ b/lustre/ldlm/ldlm_lock.c @@ -1,42 +1,59 @@ /* -*- 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 (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. + * Use is subject to license terms. + * + * Copyright (c) 2011 Whamcloud, Inc. + * + */ +/* + * This file is part of Lustre, http://www.lustre.org/ + * Lustre is a trademark of Sun Microsystems, Inc. + * + * lustre/ldlm/ldlm_lock.c + * + * Author: Peter Braam + * Author: Phil Schwan */ #define DEBUG_SUBSYSTEM S_LDLM #ifdef __KERNEL__ -# include -# include -# include +# include +# include #else # include -# include #endif -#include +#include #include "ldlm_internal.h" -//struct lustre_lock ldlm_everything_lock; - /* lock types */ char *ldlm_lockname[] = { [0] "--", @@ -46,8 +63,10 @@ char *ldlm_lockname[] = { [LCK_CW] "CW", [LCK_CR] "CR", [LCK_NL] "NL", - [LCK_GROUP] "GROUP" + [LCK_GROUP] "GROUP", + [LCK_COS] "COS" }; + char *ldlm_typename[] = { [LDLM_PLAIN] "PLN", [LDLM_EXTENT] "EXT", @@ -55,6 +74,48 @@ char *ldlm_typename[] = { [LDLM_IBITS] "IBT", }; +static ldlm_policy_wire_to_local_t ldlm_policy_wire_to_local[] = { + [LDLM_PLAIN - LDLM_MIN_TYPE] ldlm_plain_policy_wire_to_local, + [LDLM_EXTENT - LDLM_MIN_TYPE] ldlm_extent_policy_wire_to_local, + [LDLM_FLOCK - LDLM_MIN_TYPE] ldlm_flock_policy_wire_to_local, + [LDLM_IBITS - LDLM_MIN_TYPE] ldlm_ibits_policy_wire_to_local, +}; + +static ldlm_policy_local_to_wire_t ldlm_policy_local_to_wire[] = { + [LDLM_PLAIN - LDLM_MIN_TYPE] ldlm_plain_policy_local_to_wire, + [LDLM_EXTENT - LDLM_MIN_TYPE] ldlm_extent_policy_local_to_wire, + [LDLM_FLOCK - LDLM_MIN_TYPE] ldlm_flock_policy_local_to_wire, + [LDLM_IBITS - LDLM_MIN_TYPE] ldlm_ibits_policy_local_to_wire, +}; + +/** + * Converts lock policy from local format to on the wire lock_desc format + */ +void ldlm_convert_policy_to_wire(ldlm_type_t type, + const ldlm_policy_data_t *lpolicy, + ldlm_wire_policy_data_t *wpolicy) +{ + ldlm_policy_local_to_wire_t convert; + + convert = ldlm_policy_local_to_wire[type - LDLM_MIN_TYPE]; + + convert(lpolicy, wpolicy); +} + +/** + * Converts lock policy from on the wire lock_desc format to local format + */ +void ldlm_convert_policy_to_local(ldlm_type_t type, + const ldlm_wire_policy_data_t *wpolicy, + ldlm_policy_data_t *lpolicy) +{ + ldlm_policy_wire_to_local_t convert; + + convert = ldlm_policy_wire_to_local[type - LDLM_MIN_TYPE]; + + convert(wpolicy, lpolicy); +} + char *ldlm_it2str(int it) { switch (it) { @@ -80,8 +141,7 @@ char *ldlm_it2str(int it) } } -extern kmem_cache_t *ldlm_lock_slab; -struct lustre_lock ldlm_handle_lock; +extern cfs_mem_cache_t *ldlm_lock_slab; static ldlm_processing_policy ldlm_processing_policy_table[] = { [LDLM_PLAIN] ldlm_process_plain_lock, @@ -115,94 +175,166 @@ void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg) */ struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock) { - atomic_inc(&lock->l_refc); + cfs_atomic_inc(&lock->l_refc); return lock; } +static void ldlm_lock_free(struct ldlm_lock *lock, size_t size) +{ + LASSERT(size == sizeof(*lock)); + OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock)); +} + void ldlm_lock_put(struct ldlm_lock *lock) { ENTRY; - if (atomic_dec_and_test(&lock->l_refc)) { - struct ldlm_namespace *ns = lock->l_resource->lr_namespace; + LASSERT(lock->l_resource != LP_POISON); + LASSERT(cfs_atomic_read(&lock->l_refc) > 0); + if (cfs_atomic_dec_and_test(&lock->l_refc)) { + struct ldlm_resource *res; - l_lock(&ns->ns_lock); - LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing"); - LASSERT(lock->l_destroyed); - LASSERT(list_empty(&lock->l_res_link)); + LDLM_DEBUG(lock, + "final lock_put on destroyed lock, freeing it."); - spin_lock(&ns->ns_counter_lock); - ns->ns_locks--; - spin_unlock(&ns->ns_counter_lock); + res = lock->l_resource; + LASSERT(lock->l_destroyed); + LASSERT(cfs_list_empty(&lock->l_res_link)); + LASSERT(cfs_list_empty(&lock->l_pending_chain)); - ldlm_resource_putref(lock->l_resource); + lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats, + LDLM_NSS_LOCKS); + lu_ref_del(&res->lr_reference, "lock", lock); + ldlm_resource_putref(res); lock->l_resource = NULL; - if (lock->l_export) - class_export_put(lock->l_export); - - if (lock->l_parent) - LDLM_LOCK_PUT(lock->l_parent); + if (lock->l_export) { + class_export_lock_put(lock->l_export, lock); + lock->l_export = NULL; + } if (lock->l_lvb_data != NULL) OBD_FREE(lock->l_lvb_data, lock->l_lvb_len); - OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock)); - l_unlock(&ns->ns_lock); + ldlm_interval_free(ldlm_interval_detach(lock)); + lu_ref_fini(&lock->l_reference); + OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle, + ldlm_lock_free); } EXIT; } -void ldlm_lock_remove_from_lru(struct ldlm_lock *lock) +int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock) +{ + int rc = 0; + if (!cfs_list_empty(&lock->l_lru)) { + struct ldlm_namespace *ns = ldlm_lock_to_ns(lock); + + LASSERT(lock->l_resource->lr_type != LDLM_FLOCK); + cfs_list_del_init(&lock->l_lru); + if (lock->l_flags & LDLM_FL_SKIPPED) + lock->l_flags &= ~LDLM_FL_SKIPPED; + LASSERT(ns->ns_nr_unused > 0); + ns->ns_nr_unused--; + rc = 1; + } + return rc; +} + +int ldlm_lock_remove_from_lru(struct ldlm_lock *lock) +{ + struct ldlm_namespace *ns = ldlm_lock_to_ns(lock); + int rc; + + ENTRY; + if (lock->l_ns_srv) { + LASSERT(cfs_list_empty(&lock->l_lru)); + RETURN(0); + } + + cfs_spin_lock(&ns->ns_lock); + rc = ldlm_lock_remove_from_lru_nolock(lock); + cfs_spin_unlock(&ns->ns_lock); + EXIT; + return rc; +} + +void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock) +{ + struct ldlm_namespace *ns = ldlm_lock_to_ns(lock); + + lock->l_last_used = cfs_time_current(); + LASSERT(cfs_list_empty(&lock->l_lru)); + LASSERT(lock->l_resource->lr_type != LDLM_FLOCK); + cfs_list_add_tail(&lock->l_lru, &ns->ns_unused_list); + LASSERT(ns->ns_nr_unused >= 0); + ns->ns_nr_unused++; +} + +void ldlm_lock_add_to_lru(struct ldlm_lock *lock) +{ + struct ldlm_namespace *ns = ldlm_lock_to_ns(lock); + + ENTRY; + cfs_spin_lock(&ns->ns_lock); + ldlm_lock_add_to_lru_nolock(lock); + cfs_spin_unlock(&ns->ns_lock); + EXIT; +} + +void ldlm_lock_touch_in_lru(struct ldlm_lock *lock) { + struct ldlm_namespace *ns = ldlm_lock_to_ns(lock); + ENTRY; - l_lock(&lock->l_resource->lr_namespace->ns_lock); - if (!list_empty(&lock->l_lru)) { - list_del_init(&lock->l_lru); - lock->l_resource->lr_namespace->ns_nr_unused--; - LASSERT(lock->l_resource->lr_namespace->ns_nr_unused >= 0); + if (lock->l_ns_srv) { + LASSERT(cfs_list_empty(&lock->l_lru)); + EXIT; + return; } - l_unlock(&lock->l_resource->lr_namespace->ns_lock); + + cfs_spin_lock(&ns->ns_lock); + if (!cfs_list_empty(&lock->l_lru)) { + ldlm_lock_remove_from_lru_nolock(lock); + ldlm_lock_add_to_lru_nolock(lock); + } + cfs_spin_unlock(&ns->ns_lock); EXIT; } -/* This used to have a 'strict' flact, which recovery would use to mark an +/* This used to have a 'strict' flag, which recovery would use to mark an * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I * shall explain why it's gone: with the new hash table scheme, once you call * ldlm_lock_destroy, you can never drop your final references on this lock. * Because it's not in the hash table anymore. -phil */ -void ldlm_lock_destroy(struct ldlm_lock *lock) +int ldlm_lock_destroy_internal(struct ldlm_lock *lock) { ENTRY; - l_lock(&lock->l_resource->lr_namespace->ns_lock); - if (!list_empty(&lock->l_children)) { - LDLM_ERROR(lock, "still has children (%p)!", - lock->l_children.next); - ldlm_lock_dump(D_ERROR, lock, 0); - LBUG(); - } if (lock->l_readers || lock->l_writers) { LDLM_ERROR(lock, "lock still has references"); ldlm_lock_dump(D_ERROR, lock, 0); LBUG(); } - if (!list_empty(&lock->l_res_link)) { + if (!cfs_list_empty(&lock->l_res_link)) { LDLM_ERROR(lock, "lock still on resource"); ldlm_lock_dump(D_ERROR, lock, 0); LBUG(); } if (lock->l_destroyed) { - LASSERT(list_empty(&lock->l_lru)); - l_unlock(&lock->l_resource->lr_namespace->ns_lock); + LASSERT(cfs_list_empty(&lock->l_lru)); EXIT; - return; + return 0; } lock->l_destroyed = 1; - list_del_init(&lock->l_export_chain); + 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); + ldlm_lock_remove_from_lru(lock); class_handle_unhash(&lock->l_handle); @@ -216,9 +348,36 @@ void ldlm_lock_destroy(struct ldlm_lock *lock) if (lock->l_export && lock->l_completion_ast) lock->l_completion_ast(lock, 0); #endif + EXIT; + return 1; +} - l_unlock(&lock->l_resource->lr_namespace->ns_lock); - LDLM_LOCK_PUT(lock); +void ldlm_lock_destroy(struct ldlm_lock *lock) +{ + int first; + ENTRY; + lock_res_and_lock(lock); + first = ldlm_lock_destroy_internal(lock); + unlock_res_and_lock(lock); + + /* drop reference from hashtable only for first destroy */ + if (first) { + lu_ref_del(&lock->l_reference, "hash", lock); + LDLM_LOCK_RELEASE(lock); + } + EXIT; +} + +void ldlm_lock_destroy_nolock(struct ldlm_lock *lock) +{ + int first; + ENTRY; + first = ldlm_lock_destroy_internal(lock); + /* drop reference from hashtable only for first destroy */ + if (first) { + lu_ref_del(&lock->l_reference, "hash", lock); + LDLM_LOCK_RELEASE(lock); + } EXIT; } @@ -230,12 +389,10 @@ static void lock_handle_addref(void *lock) /* * usage: pass in a resource on which you have done ldlm_resource_get - * pass in a parent lock on which you have done a ldlm_lock_get - * after return, ldlm_*_put the resource and parent - * returns: lock with refcount 1 + * new lock will take over the refcount. + * returns: lock with refcount 2 - one for current caller and one for remote */ -static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent, - struct ldlm_resource *resource) +static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource) { struct ldlm_lock *lock; ENTRY; @@ -243,67 +400,101 @@ static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent, if (resource == NULL) LBUG(); - OBD_SLAB_ALLOC(lock, ldlm_lock_slab, SLAB_NOFS, sizeof(*lock)); + OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, CFS_ALLOC_IO); if (lock == NULL) RETURN(NULL); - lock->l_resource = ldlm_resource_getref(resource); - - atomic_set(&lock->l_refc, 2); - INIT_LIST_HEAD(&lock->l_children); - INIT_LIST_HEAD(&lock->l_res_link); - INIT_LIST_HEAD(&lock->l_lru); - INIT_LIST_HEAD(&lock->l_export_chain); - INIT_LIST_HEAD(&lock->l_pending_chain); - init_waitqueue_head(&lock->l_waitq); - - spin_lock(&resource->lr_namespace->ns_counter_lock); - resource->lr_namespace->ns_locks++; - spin_unlock(&resource->lr_namespace->ns_counter_lock); - - if (parent != NULL) { - l_lock(&parent->l_resource->lr_namespace->ns_lock); - lock->l_parent = LDLM_LOCK_GET(parent); - list_add(&lock->l_childof, &parent->l_children); - l_unlock(&parent->l_resource->lr_namespace->ns_lock); - } - - INIT_LIST_HEAD(&lock->l_handle.h_link); + cfs_spin_lock_init(&lock->l_lock); + lock->l_resource = resource; + lu_ref_add(&resource->lr_reference, "lock", lock); + + cfs_atomic_set(&lock->l_refc, 2); + CFS_INIT_LIST_HEAD(&lock->l_res_link); + CFS_INIT_LIST_HEAD(&lock->l_lru); + CFS_INIT_LIST_HEAD(&lock->l_pending_chain); + CFS_INIT_LIST_HEAD(&lock->l_bl_ast); + CFS_INIT_LIST_HEAD(&lock->l_cp_ast); + CFS_INIT_LIST_HEAD(&lock->l_rk_ast); + cfs_waitq_init(&lock->l_waitq); + lock->l_blocking_lock = NULL; + CFS_INIT_LIST_HEAD(&lock->l_sl_mode); + CFS_INIT_LIST_HEAD(&lock->l_sl_policy); + CFS_INIT_HLIST_NODE(&lock->l_exp_hash); + + lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats, + LDLM_NSS_LOCKS); + CFS_INIT_LIST_HEAD(&lock->l_handle.h_link); class_handle_hash(&lock->l_handle, lock_handle_addref); + lu_ref_init(&lock->l_reference); + lu_ref_add(&lock->l_reference, "hash", lock); + lock->l_callback_timeout = 0; + +#if LUSTRE_TRACKS_LOCK_EXP_REFS + CFS_INIT_LIST_HEAD(&lock->l_exp_refs_link); + lock->l_exp_refs_nr = 0; + lock->l_exp_refs_target = NULL; +#endif + RETURN(lock); } int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock, - struct ldlm_res_id new_resid) + const struct ldlm_res_id *new_resid) { struct ldlm_resource *oldres = lock->l_resource; + struct ldlm_resource *newres; + int type; ENTRY; - l_lock(&ns->ns_lock); - if (memcmp(&new_resid, &lock->l_resource->lr_name, + LASSERT(ns_is_client(ns)); + + lock_res_and_lock(lock); + if (memcmp(new_resid, &lock->l_resource->lr_name, sizeof(lock->l_resource->lr_name)) == 0) { /* Nothing to do */ - l_unlock(&ns->ns_lock); + unlock_res_and_lock(lock); RETURN(0); } - LASSERT(new_resid.name[0] != 0); + LASSERT(new_resid->name[0] != 0); /* This function assumes that the lock isn't on any lists */ - LASSERT(list_empty(&lock->l_res_link)); + LASSERT(cfs_list_empty(&lock->l_res_link)); - lock->l_resource = ldlm_resource_get(ns, NULL, new_resid, - lock->l_resource->lr_type, 1); - if (lock->l_resource == NULL) { - LBUG(); + type = oldres->lr_type; + unlock_res_and_lock(lock); + + newres = ldlm_resource_get(ns, NULL, new_resid, type, 1); + if (newres == NULL) RETURN(-ENOMEM); + + lu_ref_add(&newres->lr_reference, "lock", lock); + /* + * To flip the lock from the old to the new resource, lock, oldres and + * newres have to be locked. Resource spin-locks are nested within + * lock->l_lock, and are taken in the memory address order to avoid + * dead-locks. + */ + cfs_spin_lock(&lock->l_lock); + oldres = lock->l_resource; + if (oldres < newres) { + lock_res(oldres); + lock_res_nested(newres, LRT_NEW); + } else { + lock_res(newres); + lock_res_nested(oldres, LRT_NEW); } + LASSERT(memcmp(new_resid, &oldres->lr_name, + sizeof oldres->lr_name) != 0); + lock->l_resource = newres; + unlock_res(oldres); + unlock_res_and_lock(lock); /* ...and the flowers are still standing! */ + lu_ref_del(&oldres->lr_reference, "lock", lock); ldlm_resource_putref(oldres); - l_unlock(&ns->ns_lock); RETURN(0); } @@ -311,7 +502,7 @@ int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock, * HANDLES */ -void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh) +void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh) { lockh->cookie = lock->l_handle.h_cookie; } @@ -320,10 +511,10 @@ void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh) * Return NULL if flag already set */ -struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int flags) +struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle, + int flags) { - struct ldlm_namespace *ns; - struct ldlm_lock *lock = NULL, *retval = NULL; + struct ldlm_lock *lock; ENTRY; LASSERT(handle); @@ -332,92 +523,126 @@ struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int flags) if (lock == NULL) RETURN(NULL); - LASSERT(lock->l_resource != NULL); - ns = lock->l_resource->lr_namespace; - LASSERT(ns != NULL); - - l_lock(&ns->ns_lock); - /* It's unlikely but possible that someone marked the lock as * destroyed after we did handle2object on it */ - if (lock->l_destroyed) { + if (flags == 0 && !lock->l_destroyed) { + lu_ref_add(&lock->l_reference, "handle", cfs_current()); + RETURN(lock); + } + + lock_res_and_lock(lock); + + LASSERT(lock->l_resource != NULL); + + lu_ref_add_atomic(&lock->l_reference, "handle", cfs_current()); + if (unlikely(lock->l_destroyed)) { + unlock_res_and_lock(lock); CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock); LDLM_LOCK_PUT(lock); - GOTO(out, retval); + RETURN(NULL); } if (flags && (lock->l_flags & flags)) { + unlock_res_and_lock(lock); LDLM_LOCK_PUT(lock); - GOTO(out, retval); + RETURN(NULL); } if (flags) lock->l_flags |= flags; - retval = lock; - EXIT; - out: - l_unlock(&ns->ns_lock); - return retval; -} - -struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns, - struct lustre_handle *handle) -{ - struct ldlm_lock *retval = NULL; - - l_lock(&ns->ns_lock); - retval = __ldlm_handle2lock(handle, 0); - l_unlock(&ns->ns_lock); - - return retval; + unlock_res_and_lock(lock); + RETURN(lock); } void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc) { - ldlm_res2desc(lock->l_resource, &desc->l_resource); - desc->l_req_mode = lock->l_req_mode; - desc->l_granted_mode = lock->l_granted_mode; - memcpy(&desc->l_policy_data, &lock->l_policy_data, - sizeof(desc->l_policy_data)); -} - -void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new, - void *data, int datalen) -{ - struct ldlm_ast_work *w; - ENTRY; - - l_lock(&lock->l_resource->lr_namespace->ns_lock); - if (new && (lock->l_flags & LDLM_FL_AST_SENT)) - GOTO(out, 0); - - CDEBUG(D_OTHER, "lock %p incompatible; sending blocking AST.\n", lock); + struct obd_export *exp = lock->l_export?:lock->l_conn_export; + /* INODEBITS_INTEROP: If the other side does not support + * inodebits, reply with a plain lock descriptor. + */ + if ((lock->l_resource->lr_type == LDLM_IBITS) && + (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) { + /* Make sure all the right bits are set in this lock we + are going to pass to client */ + LASSERTF(lock->l_policy_data.l_inodebits.bits == + (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE), + "Inappropriate inode lock bits during " + "conversion " LPU64 "\n", + lock->l_policy_data.l_inodebits.bits); + + ldlm_res2desc(lock->l_resource, &desc->l_resource); + desc->l_resource.lr_type = LDLM_PLAIN; + + /* Convert "new" lock mode to something old client can + understand */ + if ((lock->l_req_mode == LCK_CR) || + (lock->l_req_mode == LCK_CW)) + desc->l_req_mode = LCK_PR; + else + desc->l_req_mode = lock->l_req_mode; + if ((lock->l_granted_mode == LCK_CR) || + (lock->l_granted_mode == LCK_CW)) { + desc->l_granted_mode = LCK_PR; + } else { + /* We never grant PW/EX locks to clients */ + LASSERT((lock->l_granted_mode != LCK_PW) && + (lock->l_granted_mode != LCK_EX)); + desc->l_granted_mode = lock->l_granted_mode; + } - OBD_ALLOC(w, sizeof(*w)); - if (!w) { - LBUG(); - GOTO(out, 0); + /* We do not copy policy here, because there is no + policy for plain locks */ + } else { + ldlm_res2desc(lock->l_resource, &desc->l_resource); + desc->l_req_mode = lock->l_req_mode; + desc->l_granted_mode = lock->l_granted_mode; + ldlm_convert_policy_to_wire(lock->l_resource->lr_type, + &lock->l_policy_data, + &desc->l_policy_data); } +} - w->w_data = data; - w->w_datalen = datalen; - if (new) { +void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new, + cfs_list_t *work_list) +{ + if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) { LDLM_DEBUG(lock, "lock incompatible; sending blocking AST."); lock->l_flags |= LDLM_FL_AST_SENT; /* If the enqueuing client said so, tell the AST recipient to * discard dirty data, rather than writing back. */ if (new->l_flags & LDLM_AST_DISCARD_DATA) lock->l_flags |= LDLM_FL_DISCARD_DATA; - w->w_blocking = 1; - ldlm_lock2desc(new, &w->w_desc); + LASSERT(cfs_list_empty(&lock->l_bl_ast)); + cfs_list_add(&lock->l_bl_ast, work_list); + LDLM_LOCK_GET(lock); + LASSERT(lock->l_blocking_lock == NULL); + lock->l_blocking_lock = LDLM_LOCK_GET(new); + } +} + +void ldlm_add_cp_work_item(struct ldlm_lock *lock, cfs_list_t *work_list) +{ + if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) { + lock->l_flags |= LDLM_FL_CP_REQD; + LDLM_DEBUG(lock, "lock granted; sending completion AST."); + LASSERT(cfs_list_empty(&lock->l_cp_ast)); + cfs_list_add(&lock->l_cp_ast, work_list); + LDLM_LOCK_GET(lock); } +} - w->w_lock = LDLM_LOCK_GET(lock); - list_add(&w->w_list, lock->l_resource->lr_tmp); +/* must be called with lr_lock held */ +void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new, + cfs_list_t *work_list) +{ + ENTRY; + check_res_locked(lock->l_resource); + if (new) + ldlm_add_bl_work_item(lock, new, work_list); + else + ldlm_add_cp_work_item(lock, work_list); EXIT; - out: - l_unlock(&lock->l_resource->lr_namespace->ns_lock); } void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode) @@ -425,42 +650,96 @@ void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode) struct ldlm_lock *lock; lock = ldlm_handle2lock(lockh); + LASSERT(lock != NULL); ldlm_lock_addref_internal(lock, mode); LDLM_LOCK_PUT(lock); } -/* only called for local locks */ -void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode) +void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode) { - l_lock(&lock->l_resource->lr_namespace->ns_lock); ldlm_lock_remove_from_lru(lock); - if (mode & (LCK_NL | LCK_CR | LCK_PR)) + if (mode & (LCK_NL | LCK_CR | LCK_PR)) { lock->l_readers++; - if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) + lu_ref_add_atomic(&lock->l_reference, "reader", lock); + } + if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) { lock->l_writers++; - lock->l_last_used = jiffies; - l_unlock(&lock->l_resource->lr_namespace->ns_lock); + lu_ref_add_atomic(&lock->l_reference, "writer", lock); + } LDLM_LOCK_GET(lock); + lu_ref_add_atomic(&lock->l_reference, "user", lock); LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]); } -void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode) +/** + * Attempts to addref a lock, and fails if lock is already LDLM_FL_CBPENDING + * or destroyed. + * + * \retval 0 success, lock was addref-ed + * + * \retval -EAGAIN lock is being canceled. + */ +int ldlm_lock_addref_try(struct lustre_handle *lockh, __u32 mode) { - struct ldlm_namespace *ns; - ENTRY; + struct ldlm_lock *lock; + int result; + + result = -EAGAIN; + lock = ldlm_handle2lock(lockh); + if (lock != NULL) { + lock_res_and_lock(lock); + if (lock->l_readers != 0 || lock->l_writers != 0 || + !(lock->l_flags & LDLM_FL_CBPENDING)) { + ldlm_lock_addref_internal_nolock(lock, mode); + result = 0; + } + unlock_res_and_lock(lock); + LDLM_LOCK_PUT(lock); + } + return result; +} +/* only called for local locks */ +void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode) +{ + lock_res_and_lock(lock); + ldlm_lock_addref_internal_nolock(lock, mode); + unlock_res_and_lock(lock); +} + +/* only called in ldlm_flock_destroy and for local locks. + * * for LDLM_FLOCK type locks, l_blocking_ast is null, and + * * ldlm_lock_remove_from_lru() does nothing, it is safe + * * for ldlm_flock_destroy usage by dropping some code */ +void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode) +{ LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]); - ns = lock->l_resource->lr_namespace; - l_lock(&ns->ns_lock); if (mode & (LCK_NL | LCK_CR | LCK_PR)) { LASSERT(lock->l_readers > 0); + lu_ref_del(&lock->l_reference, "reader", lock); lock->l_readers--; } - if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) { + if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) { LASSERT(lock->l_writers > 0); + lu_ref_del(&lock->l_reference, "writer", lock); lock->l_writers--; } + lu_ref_del(&lock->l_reference, "user", lock); + LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */ +} + +void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode) +{ + struct ldlm_namespace *ns; + ENTRY; + + lock_res_and_lock(lock); + + ns = ldlm_lock_to_ns(lock); + + ldlm_lock_decref_internal_nolock(lock, mode); + if (lock->l_flags & LDLM_FL_LOCAL && !lock->l_readers && !lock->l_writers) { /* If this is a local lock on a server namespace and this was @@ -473,7 +752,7 @@ void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode) (lock->l_flags & LDLM_FL_CBPENDING)) { /* If we received a blocked AST and this was the last reference, * run the callback. */ - if (ns->ns_client == LDLM_NAMESPACE_SERVER && lock->l_export) + if (lock->l_ns_srv && lock->l_export) CERROR("FL_CBPENDING set on non-local lock--just a " "warning\n"); @@ -481,36 +760,42 @@ void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode) LDLM_LOCK_GET(lock); /* dropped by bl thread */ ldlm_lock_remove_from_lru(lock); -#ifdef __KERNEL__ - ldlm_bl_to_thread(ns, NULL, lock); - l_unlock(&ns->ns_lock); -#else - l_unlock(&ns->ns_lock); - ldlm_handle_bl_callback(ns, NULL, lock); -#endif - } else if (ns->ns_client == LDLM_NAMESPACE_CLIENT && - !lock->l_readers && !lock->l_writers) { + unlock_res_and_lock(lock); + + if (lock->l_flags & LDLM_FL_FAIL_LOC) + OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE); + + if ((lock->l_flags & LDLM_FL_ATOMIC_CB) || + ldlm_bl_to_thread_lock(ns, NULL, lock) != 0) + ldlm_handle_bl_callback(ns, NULL, lock); + } else if (ns_is_client(ns) && + !lock->l_readers && !lock->l_writers && + !(lock->l_flags & LDLM_FL_BL_AST)) { /* If this is a client-side namespace and this was the last * reference, put it on the LRU. */ - LASSERT(list_empty(&lock->l_lru)); - LASSERT(ns->ns_nr_unused >= 0); - list_add_tail(&lock->l_lru, &ns->ns_unused_list); - ns->ns_nr_unused++; - l_unlock(&ns->ns_lock); - ldlm_cancel_lru(ns, LDLM_ASYNC); + ldlm_lock_add_to_lru(lock); + unlock_res_and_lock(lock); + + if (lock->l_flags & LDLM_FL_FAIL_LOC) + OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE); + + /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE + * are not supported by the server, otherwise, it is done on + * enqueue. */ + if (!exp_connect_cancelset(lock->l_conn_export) && + !ns_connect_lru_resize(ns)) + ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0); } else { - l_unlock(&ns->ns_lock); + unlock_res_and_lock(lock); } - LDLM_LOCK_PUT(lock); /* matches the ldlm_lock_get in addref */ - EXIT; } void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode) { struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0); - LASSERT(lock != NULL); + LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie); ldlm_lock_decref_internal(lock, mode); LDLM_LOCK_PUT(lock); } @@ -525,49 +810,201 @@ void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode) LASSERT(lock != NULL); LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]); - l_lock(&lock->l_resource->lr_namespace->ns_lock); + lock_res_and_lock(lock); lock->l_flags |= LDLM_FL_CBPENDING; - l_unlock(&lock->l_resource->lr_namespace->ns_lock); + unlock_res_and_lock(lock); ldlm_lock_decref_internal(lock, mode); LDLM_LOCK_PUT(lock); } +struct sl_insert_point { + cfs_list_t *res_link; + cfs_list_t *mode_link; + cfs_list_t *policy_link; +}; + +/* + * search_granted_lock + * + * Description: + * Finds a position to insert the new lock. + * Parameters: + * queue [input]: the granted list where search acts on; + * req [input]: the lock whose position to be located; + * prev [output]: positions within 3 lists to insert @req to + * Return Value: + * filled @prev + * NOTE: called by + * - ldlm_grant_lock_with_skiplist + */ +static void search_granted_lock(cfs_list_t *queue, + struct ldlm_lock *req, + struct sl_insert_point *prev) +{ + cfs_list_t *tmp; + struct ldlm_lock *lock, *mode_end, *policy_end; + ENTRY; + + cfs_list_for_each(tmp, queue) { + lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link); + + mode_end = cfs_list_entry(lock->l_sl_mode.prev, + struct ldlm_lock, l_sl_mode); + + if (lock->l_req_mode != req->l_req_mode) { + /* jump to last lock of mode group */ + tmp = &mode_end->l_res_link; + continue; + } + + /* suitable mode group is found */ + if (lock->l_resource->lr_type == LDLM_PLAIN) { + /* insert point is last lock of the mode group */ + prev->res_link = &mode_end->l_res_link; + prev->mode_link = &mode_end->l_sl_mode; + prev->policy_link = &req->l_sl_policy; + EXIT; + return; + } else if (lock->l_resource->lr_type == LDLM_IBITS) { + for (;;) { + policy_end = + cfs_list_entry(lock->l_sl_policy.prev, + struct ldlm_lock, + l_sl_policy); + + if (lock->l_policy_data.l_inodebits.bits == + req->l_policy_data.l_inodebits.bits) { + /* insert point is last lock of + * the policy group */ + prev->res_link = + &policy_end->l_res_link; + prev->mode_link = + &policy_end->l_sl_mode; + prev->policy_link = + &policy_end->l_sl_policy; + EXIT; + return; + } + + if (policy_end == mode_end) + /* done with mode group */ + break; + + /* go to next policy group within mode group */ + tmp = policy_end->l_res_link.next; + lock = cfs_list_entry(tmp, struct ldlm_lock, + l_res_link); + } /* loop over policy groups within the mode group */ + + /* insert point is last lock of the mode group, + * new policy group is started */ + prev->res_link = &mode_end->l_res_link; + prev->mode_link = &mode_end->l_sl_mode; + prev->policy_link = &req->l_sl_policy; + EXIT; + return; + } else { + LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock"); + LBUG(); + } + } + + /* insert point is last lock on the queue, + * new mode group and new policy group are started */ + prev->res_link = queue->prev; + prev->mode_link = &req->l_sl_mode; + prev->policy_link = &req->l_sl_policy; + EXIT; + return; +} + +static void ldlm_granted_list_add_lock(struct ldlm_lock *lock, + struct sl_insert_point *prev) +{ + struct ldlm_resource *res = lock->l_resource; + ENTRY; + + check_res_locked(res); + + ldlm_resource_dump(D_INFO, res); + CDEBUG(D_OTHER, "About to add this lock:\n"); + ldlm_lock_dump(D_OTHER, lock, 0); + + if (lock->l_destroyed) { + CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n"); + return; + } + + LASSERT(cfs_list_empty(&lock->l_res_link)); + LASSERT(cfs_list_empty(&lock->l_sl_mode)); + LASSERT(cfs_list_empty(&lock->l_sl_policy)); + + cfs_list_add(&lock->l_res_link, prev->res_link); + cfs_list_add(&lock->l_sl_mode, prev->mode_link); + cfs_list_add(&lock->l_sl_policy, prev->policy_link); + + EXIT; +} + +static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock) +{ + struct sl_insert_point prev; + ENTRY; + + LASSERT(lock->l_req_mode == lock->l_granted_mode); + + search_granted_lock(&lock->l_resource->lr_granted, lock, &prev); + ldlm_granted_list_add_lock(lock, &prev); + EXIT; +} + /* NOTE: called by * - ldlm_lock_enqueue * - ldlm_reprocess_queue * - ldlm_lock_convert + * + * must be called with lr_lock held */ -void ldlm_grant_lock(struct ldlm_lock *lock, void *data, int datalen, - int run_ast) +void ldlm_grant_lock(struct ldlm_lock *lock, cfs_list_t *work_list) { struct ldlm_resource *res = lock->l_resource; ENTRY; - l_lock(&lock->l_resource->lr_namespace->ns_lock); + check_res_locked(res); + lock->l_granted_mode = lock->l_req_mode; - ldlm_resource_add_lock(res, &res->lr_granted, lock); + if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) + ldlm_grant_lock_with_skiplist(lock); + else if (res->lr_type == LDLM_EXTENT) + ldlm_extent_add_lock(res, lock); + else + ldlm_resource_add_lock(res, &res->lr_granted, lock); if (lock->l_granted_mode < res->lr_most_restr) res->lr_most_restr = lock->l_granted_mode; - if (run_ast && lock->l_completion_ast != NULL) - ldlm_add_ast_work_item(lock, NULL, data, datalen); + if (work_list && lock->l_completion_ast != NULL) + ldlm_add_ast_work_item(lock, NULL, work_list); - l_unlock(&lock->l_resource->lr_namespace->ns_lock); + ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock); EXIT; } /* returns a referenced lock or NULL. See the flag descriptions below, in the * comment above ldlm_lock_match */ -static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode, +static struct ldlm_lock *search_queue(cfs_list_t *queue, + ldlm_mode_t *mode, ldlm_policy_data_t *policy, - struct ldlm_lock *old_lock, int flags) + struct ldlm_lock *old_lock, + int flags, int unref) { struct ldlm_lock *lock; - struct list_head *tmp; + cfs_list_t *tmp; - list_for_each(tmp, queue) { - lock = list_entry(tmp, struct ldlm_lock, l_res_link); + cfs_list_for_each(tmp, queue) { + ldlm_mode_t match; + + lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link); if (lock == old_lock) break; @@ -575,18 +1012,19 @@ static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode, /* llite sometimes wants to match locks that will be * canceled when their users drop, but we allow it to match * if it passes in CBPENDING and the lock still has users. - * this is generally only going to be used by children + * this is generally only going to be used by children * whose parents already hold a lock so forward progress * can still happen. */ if (lock->l_flags & LDLM_FL_CBPENDING && !(flags & LDLM_FL_CBPENDING)) continue; - if (lock->l_flags & LDLM_FL_CBPENDING && + if (!unref && lock->l_flags & LDLM_FL_CBPENDING && lock->l_readers == 0 && lock->l_writers == 0) continue; - if (!(lock->l_req_mode & mode)) + if (!(lock->l_req_mode & *mode)) continue; + match = lock->l_req_mode; if (lock->l_resource->lr_type == LDLM_EXTENT && (lock->l_policy_data.l_extent.start > @@ -594,8 +1032,8 @@ static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode, lock->l_policy_data.l_extent.end < policy->l_extent.end)) continue; - if (lock->l_resource->lr_type == LDLM_EXTENT && - mode == LCK_GROUP && + if (unlikely(match == LCK_GROUP) && + lock->l_resource->lr_type == LDLM_EXTENT && lock->l_policy_data.l_extent.gid != policy->l_extent.gid) continue; @@ -607,29 +1045,38 @@ static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode, policy->l_inodebits.bits)) continue; - if (lock->l_destroyed) + if (!unref && + (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED))) continue; if ((flags & LDLM_FL_LOCAL_ONLY) && !(lock->l_flags & LDLM_FL_LOCAL)) continue; - if (flags & LDLM_FL_TEST_LOCK) + if (flags & LDLM_FL_TEST_LOCK) { LDLM_LOCK_GET(lock); - else - ldlm_lock_addref_internal(lock, mode); + ldlm_lock_touch_in_lru(lock); + } else { + ldlm_lock_addref_internal_nolock(lock, match); + } + *mode = match; return lock; } return NULL; } +void ldlm_lock_allow_match_locked(struct ldlm_lock *lock) +{ + lock->l_flags |= LDLM_FL_LVB_READY; + cfs_waitq_signal(&lock->l_waitq); +} + void ldlm_lock_allow_match(struct ldlm_lock *lock) { - l_lock(&lock->l_resource->lr_namespace->ns_lock); - lock->l_flags |= LDLM_FL_CAN_MATCH; - wake_up(&lock->l_waitq); - l_unlock(&lock->l_resource->lr_namespace->ns_lock); + lock_res_and_lock(lock); + ldlm_lock_allow_match_locked(lock); + unlock_res_and_lock(lock); } /* Can be called in two ways: @@ -651,11 +1098,15 @@ void ldlm_lock_allow_match(struct ldlm_lock *lock) * * Returns 1 if it finds an already-existing lock that is compatible; in this * case, lockh is filled in with a addref()ed lock + * + * we also check security context, if that failed we simply return 0 (to keep + * caller code unchanged), the context failure will be discovered by caller + * sometime later. */ -int ldlm_lock_match(struct ldlm_namespace *ns, int flags, - struct ldlm_res_id *res_id, __u32 type, - ldlm_policy_data_t *policy, ldlm_mode_t mode, - struct lustre_handle *lockh) +ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags, + const struct ldlm_res_id *res_id, ldlm_type_t type, + ldlm_policy_data_t *policy, ldlm_mode_t mode, + struct lustre_handle *lockh, int unref) { struct ldlm_resource *res; struct ldlm_lock *lock, *old_lock = NULL; @@ -666,124 +1117,160 @@ int ldlm_lock_match(struct ldlm_namespace *ns, int flags, old_lock = ldlm_handle2lock(lockh); LASSERT(old_lock); - ns = old_lock->l_resource->lr_namespace; + ns = ldlm_lock_to_ns(old_lock); res_id = &old_lock->l_resource->lr_name; type = old_lock->l_resource->lr_type; mode = old_lock->l_req_mode; } - res = ldlm_resource_get(ns, NULL, *res_id, type, 0); + res = ldlm_resource_get(ns, NULL, res_id, type, 0); if (res == NULL) { LASSERT(old_lock == NULL); RETURN(0); } - l_lock(&ns->ns_lock); + LDLM_RESOURCE_ADDREF(res); + lock_res(res); - lock = search_queue(&res->lr_granted, mode, policy, old_lock, flags); + lock = search_queue(&res->lr_granted, &mode, policy, old_lock, + flags, unref); if (lock != NULL) GOTO(out, rc = 1); if (flags & LDLM_FL_BLOCK_GRANTED) GOTO(out, rc = 0); - lock = search_queue(&res->lr_converting, mode, policy, old_lock, flags); + lock = search_queue(&res->lr_converting, &mode, policy, old_lock, + flags, unref); if (lock != NULL) GOTO(out, rc = 1); - lock = search_queue(&res->lr_waiting, mode, policy, old_lock, flags); + lock = search_queue(&res->lr_waiting, &mode, policy, old_lock, + flags, unref); if (lock != NULL) GOTO(out, rc = 1); EXIT; out: + unlock_res(res); + LDLM_RESOURCE_DELREF(res); ldlm_resource_putref(res); - l_unlock(&ns->ns_lock); if (lock) { ldlm_lock2handle(lock, lockh); - if (!(lock->l_flags & LDLM_FL_CAN_MATCH)) { + if ((flags & LDLM_FL_LVB_READY) && + (!(lock->l_flags & LDLM_FL_LVB_READY))) { struct l_wait_info lwi; - if (lock->l_completion_ast) - lock->l_completion_ast(lock, - LDLM_FL_WAIT_NOREPROC, - NULL); + if (lock->l_completion_ast) { + int err = lock->l_completion_ast(lock, + LDLM_FL_WAIT_NOREPROC, + NULL); + if (err) { + if (flags & LDLM_FL_TEST_LOCK) + LDLM_LOCK_RELEASE(lock); + else + ldlm_lock_decref_internal(lock, + mode); + rc = 0; + goto out2; + } + } - lwi = LWI_TIMEOUT_INTR(obd_timeout*HZ, NULL,NULL,NULL); + lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout), + NULL, LWI_ON_SIGNAL_NOOP, NULL); /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */ l_wait_event(lock->l_waitq, - (lock->l_flags & LDLM_FL_CAN_MATCH), &lwi); + (lock->l_flags & LDLM_FL_LVB_READY), &lwi); } } - if (rc) + out2: + if (rc) { LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")", - type == LDLM_PLAIN ? res_id->name[2] : - policy->l_extent.start, - type == LDLM_PLAIN ? res_id->name[3] : - policy->l_extent.end); - else if (!(flags & LDLM_FL_TEST_LOCK)) /* less verbose for test-only */ + (type == LDLM_PLAIN || type == LDLM_IBITS) ? + res_id->name[2] : policy->l_extent.start, + (type == LDLM_PLAIN || type == LDLM_IBITS) ? + res_id->name[3] : policy->l_extent.end); + + /* check user's security context */ + if (lock->l_conn_export && + sptlrpc_import_check_ctx( + class_exp2cliimp(lock->l_conn_export))) { + if (!(flags & LDLM_FL_TEST_LOCK)) + ldlm_lock_decref_internal(lock, mode); + rc = 0; + } + + if (flags & LDLM_FL_TEST_LOCK) + LDLM_LOCK_RELEASE(lock); + + } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/ LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res " LPU64"/"LPU64" ("LPU64" "LPU64")", ns, type, mode, res_id->name[0], res_id->name[1], - type == LDLM_PLAIN ? res_id->name[2] : - policy->l_extent.start, - type == LDLM_PLAIN ? res_id->name[3] : - policy->l_extent.end); - + (type == LDLM_PLAIN || type == LDLM_IBITS) ? + res_id->name[2] :policy->l_extent.start, + (type == LDLM_PLAIN || type == LDLM_IBITS) ? + res_id->name[3] : policy->l_extent.end); + } if (old_lock) LDLM_LOCK_PUT(old_lock); - if (flags & LDLM_FL_TEST_LOCK && rc) - LDLM_LOCK_PUT(lock); - return rc; + return rc ? mode : 0; } /* Returns a referenced lock */ struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns, - struct lustre_handle *parent_lock_handle, - struct ldlm_res_id res_id, __u32 type, + const struct ldlm_res_id *res_id, + ldlm_type_t type, ldlm_mode_t mode, - ldlm_blocking_callback blocking, - ldlm_completion_callback completion, - ldlm_glimpse_callback glimpse, + const struct ldlm_callback_suite *cbs, void *data, __u32 lvb_len) { - struct ldlm_resource *res, *parent_res = NULL; - struct ldlm_lock *lock, *parent_lock = NULL; + struct ldlm_lock *lock; + struct ldlm_resource *res; ENTRY; - if (parent_lock_handle) { - parent_lock = ldlm_handle2lock(parent_lock_handle); - if (parent_lock) - parent_res = parent_lock->l_resource; - } - - res = ldlm_resource_get(ns, parent_res, res_id, type, 1); + res = ldlm_resource_get(ns, NULL, res_id, type, 1); if (res == NULL) RETURN(NULL); - lock = ldlm_lock_new(parent_lock, res); - ldlm_resource_putref(res); - if (parent_lock != NULL) - LDLM_LOCK_PUT(parent_lock); + lock = ldlm_lock_new(res); if (lock == NULL) RETURN(NULL); lock->l_req_mode = mode; lock->l_ast_data = data; - lock->l_blocking_ast = blocking; - lock->l_completion_ast = completion; - lock->l_glimpse_ast = glimpse; + lock->l_pid = cfs_curproc_pid(); + lock->l_ns_srv = ns_is_server(ns); + if (cbs) { + lock->l_blocking_ast = cbs->lcs_blocking; + lock->l_completion_ast = cbs->lcs_completion; + lock->l_glimpse_ast = cbs->lcs_glimpse; + lock->l_weigh_ast = cbs->lcs_weigh; + } + + lock->l_tree_node = NULL; + /* if this is the extent lock, allocate the interval tree node */ + if (type == LDLM_EXTENT) { + if (ldlm_interval_alloc(lock) == NULL) + GOTO(out, 0); + } if (lvb_len) { lock->l_lvb_len = lvb_len; OBD_ALLOC(lock->l_lvb_data, lvb_len); - if (lock->l_lvb_data == NULL) { - OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock)); - RETURN(NULL); - } + if (lock->l_lvb_data == NULL) + GOTO(out, 0); } + if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK)) + GOTO(out, 0); + RETURN(lock); + +out: + ldlm_lock_destroy(lock); + LDLM_LOCK_RELEASE(lock); + return NULL; } ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns, @@ -792,11 +1279,13 @@ ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns, { struct ldlm_lock *lock = *lockp; struct ldlm_resource *res = lock->l_resource; - int local = res->lr_namespace->ns_client; + int local = ns_is_client(ldlm_res_to_ns(res)); ldlm_processing_policy policy; ldlm_error_t rc = ELDLM_OK; + struct ldlm_interval *node = NULL; ENTRY; + lock->l_last_activity = cfs_time_current_sec(); /* policies are not executed on the client or during replay */ if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT && !local && ns->ns_policy) { @@ -809,31 +1298,49 @@ ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns, * work here is done. */ if (lock != *lockp) { ldlm_lock_destroy(lock); - LDLM_LOCK_PUT(lock); + LDLM_LOCK_RELEASE(lock); } *flags |= LDLM_FL_LOCK_CHANGED; RETURN(0); - } else if (rc == ELDLM_LOCK_ABORTED || - (rc == 0 && (*flags & LDLM_FL_INTENT_ONLY))) { + } else if (rc != ELDLM_OK || + (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) { ldlm_lock_destroy(lock); RETURN(rc); } - LASSERT(rc == ELDLM_OK); } - l_lock(&ns->ns_lock); + /* For a replaying lock, it might be already in granted list. So + * unlinking the lock will cause the interval node to be freed, we + * have to allocate the interval node early otherwise we can't regrant + * this lock in the future. - jay */ + if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT) + OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO); + + lock_res_and_lock(lock); if (local && lock->l_req_mode == lock->l_granted_mode) { - /* The server returned a blocked lock, but it was granted before - * we got a chance to actually enqueue it. We don't need to do - * anything else. */ + /* The server returned a blocked lock, but it was granted + * before we got a chance to actually enqueue it. We don't + * need to do anything else. */ *flags &= ~(LDLM_FL_BLOCK_GRANTED | LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT); GOTO(out, ELDLM_OK); } + ldlm_resource_unlink_lock(lock); + if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) { + if (node == NULL) { + ldlm_lock_destroy_nolock(lock); + GOTO(out, rc = -ENOMEM); + } + + CFS_INIT_LIST_HEAD(&node->li_group); + ldlm_interval_attach(node, lock); + node = NULL; + } + /* Some flags from the enqueue want to make it into the AST, via the * lock's l_flags. */ - lock->l_flags |= (*flags & LDLM_AST_DISCARD_DATA); + lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA; /* This distinction between local lock trees is very important; a client * namespace only has information about locks taken by that client, and @@ -846,14 +1353,13 @@ ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns, * * FIXME (bug 268): Detect obvious lies by checking compatibility in * granted/converting queues. */ - ldlm_resource_unlink_lock(lock); if (local) { if (*flags & LDLM_FL_BLOCK_CONV) ldlm_resource_add_lock(res, &res->lr_converting, lock); else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED)) ldlm_resource_add_lock(res, &res->lr_waiting, lock); else - ldlm_grant_lock(lock, NULL, 0, 0); + ldlm_grant_lock(lock, NULL); GOTO(out, ELDLM_OK); } else if (*flags & LDLM_FL_REPLAY) { if (*flags & LDLM_FL_BLOCK_CONV) { @@ -863,41 +1369,46 @@ ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns, ldlm_resource_add_lock(res, &res->lr_waiting, lock); GOTO(out, ELDLM_OK); } else if (*flags & LDLM_FL_BLOCK_GRANTED) { - ldlm_grant_lock(lock, NULL, 0, 0); + ldlm_grant_lock(lock, NULL); GOTO(out, ELDLM_OK); } /* If no flags, fall through to normal enqueue path. */ } policy = ldlm_processing_policy_table[res->lr_type]; - policy(lock, flags, 1, &rc); - EXIT; + policy(lock, flags, 1, &rc, NULL); + GOTO(out, rc); out: - l_unlock(&ns->ns_lock); + unlock_res_and_lock(lock); + if (node) + OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node)); return rc; } /* Must be called with namespace taken: queue is waiting or converting. */ -int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue) +int ldlm_reprocess_queue(struct ldlm_resource *res, cfs_list_t *queue, + cfs_list_t *work_list) { - struct list_head *tmp, *pos; + cfs_list_t *tmp, *pos; ldlm_processing_policy policy; int flags; int rc = LDLM_ITER_CONTINUE; ldlm_error_t err; ENTRY; + check_res_locked(res); + policy = ldlm_processing_policy_table[res->lr_type]; LASSERT(policy); - list_for_each_safe(tmp, pos, queue) { + cfs_list_for_each_safe(tmp, pos, queue) { struct ldlm_lock *pending; - pending = list_entry(tmp, struct ldlm_lock, l_res_link); + pending = cfs_list_entry(tmp, struct ldlm_lock, l_res_link); CDEBUG(D_INFO, "Reprocessing lock %p\n", pending); flags = 0; - rc = policy(pending, &flags, 0, &err); + rc = policy(pending, &flags, 0, &err, work_list); if (rc != LDLM_ITER_CONTINUE) break; } @@ -905,51 +1416,172 @@ int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue) RETURN(rc); } -int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list) +/* Helper function for ldlm_run_ast_work(). + * + * Send an existing rpc set specified by @arg->set and then + * destroy it. Create new one if @do_create flag is set. */ +static int ldlm_deliver_cb_set(struct ldlm_cb_set_arg *arg, int do_create) { - struct list_head *tmp, *pos; - int rc, retval = 0; + int rc = 0; ENTRY; - l_check_no_ns_lock(ns); - - list_for_each_safe(tmp, pos, rpc_list) { - struct ldlm_ast_work *w = - list_entry(tmp, struct ldlm_ast_work, w_list); - - /* It's possible to receive a completion AST before we've set - * the l_completion_ast pointer: either because the AST arrived - * before the reply, or simply because there's a small race - * window between receiving the reply and finishing the local - * enqueue. (bug 842) - * - * This can't happen with the blocking_ast, however, because we - * will never call the local blocking_ast until we drop our - * reader/writer reference, which we won't do until we get the - * reply and finish enqueueing. */ - LASSERT(w->w_lock != NULL); - if (w->w_blocking) { - LASSERT(w->w_lock->l_blocking_ast != NULL); - rc = w->w_lock->l_blocking_ast - (w->w_lock, &w->w_desc, w->w_data, - LDLM_CB_BLOCKING); - } else if (w->w_lock->l_completion_ast != NULL) { - LASSERT(w->w_lock->l_completion_ast != NULL); - rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags, - w->w_data); - } else { - rc = 0; + if (arg->set) { + ptlrpc_set_wait(arg->set); + if (arg->type == LDLM_BL_CALLBACK) + OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2); + ptlrpc_set_destroy(arg->set); + arg->set = NULL; + arg->rpcs = 0; + } + + if (do_create) { + arg->set = ptlrpc_prep_set(); + if (arg->set == NULL) + rc = -ENOMEM; + } + + RETURN(rc); +} + +static int +ldlm_work_bl_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg) +{ + struct ldlm_lock_desc d; + struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock, + l_bl_ast); + int rc; + ENTRY; + + /* nobody should touch l_bl_ast */ + lock_res_and_lock(lock); + cfs_list_del_init(&lock->l_bl_ast); + + LASSERT(lock->l_flags & LDLM_FL_AST_SENT); + LASSERT(lock->l_bl_ast_run == 0); + LASSERT(lock->l_blocking_lock); + lock->l_bl_ast_run++; + unlock_res_and_lock(lock); + + ldlm_lock2desc(lock->l_blocking_lock, &d); + + rc = lock->l_blocking_ast(lock, &d, (void *)arg, + LDLM_CB_BLOCKING); + LDLM_LOCK_RELEASE(lock->l_blocking_lock); + lock->l_blocking_lock = NULL; + LDLM_LOCK_RELEASE(lock); + + RETURN(rc); +} + +static int +ldlm_work_cp_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg) +{ + struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock, l_cp_ast); + ldlm_completion_callback completion_callback; + int rc = 0; + ENTRY; + + /* It's possible to receive a completion AST before we've set + * the l_completion_ast pointer: either because the AST arrived + * before the reply, or simply because there's a small race + * window between receiving the reply and finishing the local + * enqueue. (bug 842) + * + * This can't happen with the blocking_ast, however, because we + * will never call the local blocking_ast until we drop our + * reader/writer reference, which we won't do until we get the + * reply and finish enqueueing. */ + + /* nobody should touch l_cp_ast */ + lock_res_and_lock(lock); + cfs_list_del_init(&lock->l_cp_ast); + LASSERT(lock->l_flags & LDLM_FL_CP_REQD); + /* save l_completion_ast since it can be changed by + * mds_intent_policy(), see bug 14225 */ + completion_callback = lock->l_completion_ast; + lock->l_flags &= ~LDLM_FL_CP_REQD; + unlock_res_and_lock(lock); + + if (completion_callback != NULL) + rc = completion_callback(lock, 0, (void *)arg); + LDLM_LOCK_RELEASE(lock); + + RETURN(rc); +} + +static int +ldlm_work_revoke_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg) +{ + struct ldlm_lock_desc desc; + struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock, + l_rk_ast); + int rc; + ENTRY; + + cfs_list_del_init(&lock->l_rk_ast); + + /* the desc just pretend to exclusive */ + ldlm_lock2desc(lock, &desc); + desc.l_req_mode = LCK_EX; + desc.l_granted_mode = 0; + + rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING); + LDLM_LOCK_RELEASE(lock); + + RETURN(rc); +} + +int ldlm_run_ast_work(struct ldlm_namespace *ns, cfs_list_t *rpc_list, + ldlm_desc_ast_t ast_type) +{ + struct ldlm_cb_set_arg arg = { 0 }; + cfs_list_t *tmp, *pos; + int (*work_ast_lock)(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg); + unsigned int max_ast_count; + int rc; + ENTRY; + + if (cfs_list_empty(rpc_list)) + RETURN(0); + + rc = ldlm_deliver_cb_set(&arg, 1); + if (rc != 0) + RETURN(rc); + + switch (ast_type) { + case LDLM_WORK_BL_AST: + arg.type = LDLM_BL_CALLBACK; + work_ast_lock = ldlm_work_bl_ast_lock; + break; + case LDLM_WORK_CP_AST: + arg.type = LDLM_CP_CALLBACK; + work_ast_lock = ldlm_work_cp_ast_lock; + break; + case LDLM_WORK_REVOKE_AST: + arg.type = LDLM_BL_CALLBACK; + work_ast_lock = ldlm_work_revoke_ast_lock; + break; + default: + LBUG(); + } + + max_ast_count = ns->ns_max_parallel_ast ? : UINT_MAX; + + cfs_list_for_each_safe(tmp, pos, rpc_list) { + (void)work_ast_lock(tmp, &arg); + if (arg.rpcs > max_ast_count) { + rc = ldlm_deliver_cb_set(&arg, 1); + if (rc != 0) + break; } - if (rc == -ERESTART) - retval = rc; - else if (rc) - CDEBUG(D_DLMTRACE, "Failed AST - should clean & " - "disconnect client\n"); - LDLM_LOCK_PUT(w->w_lock); - list_del(&w->w_list); - OBD_FREE(w, sizeof(*w)); } - RETURN(retval); + + (void)ldlm_deliver_cb_set(&arg, 0); + + if (rc == 0 && cfs_atomic_read(&arg.restart)) + rc = -ERESTART; + + RETURN(rc); } static int reprocess_one_queue(struct ldlm_resource *res, void *closure) @@ -958,58 +1590,51 @@ static int reprocess_one_queue(struct ldlm_resource *res, void *closure) return LDLM_ITER_CONTINUE; } +static int ldlm_reprocess_res(cfs_hash_t *hs, cfs_hash_bd_t *bd, + cfs_hlist_node_t *hnode, void *arg) +{ + struct ldlm_resource *res = cfs_hash_object(hs, hnode); + int rc; + + rc = reprocess_one_queue(res, arg); + + return rc == LDLM_ITER_STOP; +} + void ldlm_reprocess_all_ns(struct ldlm_namespace *ns) { - int i, rc; - - l_lock(&ns->ns_lock); - for (i = 0; i < RES_HASH_SIZE; i++) { - struct list_head *tmp, *next; - list_for_each_safe(tmp, next, &(ns->ns_hash[i])) { - struct ldlm_resource *res = - list_entry(tmp, struct ldlm_resource, lr_hash); - - ldlm_resource_getref(res); - l_unlock(&ns->ns_lock); - rc = reprocess_one_queue(res, NULL); - l_lock(&ns->ns_lock); - next = tmp->next; - ldlm_resource_putref(res); - if (rc == LDLM_ITER_STOP) - GOTO(out, rc); - } + ENTRY; + + if (ns != NULL) { + cfs_hash_for_each_nolock(ns->ns_rs_hash, + ldlm_reprocess_res, NULL); } - out: - l_unlock(&ns->ns_lock); EXIT; } void ldlm_reprocess_all(struct ldlm_resource *res) { - struct list_head rpc_list = LIST_HEAD_INIT(rpc_list); + CFS_LIST_HEAD(rpc_list); int rc; ENTRY; /* Local lock trees don't get reprocessed. */ - if (res->lr_namespace->ns_client) { + if (ns_is_client(ldlm_res_to_ns(res))) { EXIT; return; } restart: - l_lock(&res->lr_namespace->ns_lock); - res->lr_tmp = &rpc_list; - - rc = ldlm_reprocess_queue(res, &res->lr_converting); + lock_res(res); + rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list); if (rc == LDLM_ITER_CONTINUE) - ldlm_reprocess_queue(res, &res->lr_waiting); - - res->lr_tmp = NULL; - l_unlock(&res->lr_namespace->ns_lock); + ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list); + unlock_res(res); - rc = ldlm_run_ast_work(res->lr_namespace, &rpc_list); + rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list, + LDLM_WORK_CP_AST); if (rc == -ERESTART) { - LASSERT(list_empty(&rpc_list)); + LASSERT(cfs_list_empty(&rpc_list)); goto restart; } EXIT; @@ -1017,20 +1642,30 @@ void ldlm_reprocess_all(struct ldlm_resource *res) void ldlm_cancel_callback(struct ldlm_lock *lock) { - l_lock(&lock->l_resource->lr_namespace->ns_lock); + check_res_locked(lock->l_resource); if (!(lock->l_flags & LDLM_FL_CANCEL)) { lock->l_flags |= LDLM_FL_CANCEL; if (lock->l_blocking_ast) { - l_unlock(&lock->l_resource->lr_namespace->ns_lock); - // l_check_no_ns_lock(lock->l_resource->lr_namespace); + // l_check_no_ns_lock(ns); + unlock_res_and_lock(lock); lock->l_blocking_ast(lock, NULL, lock->l_ast_data, LDLM_CB_CANCELING); - return; + lock_res_and_lock(lock); } else { LDLM_DEBUG(lock, "no blocking ast"); } } - l_unlock(&lock->l_resource->lr_namespace->ns_lock); + lock->l_flags |= LDLM_FL_BL_DONE; +} + +void ldlm_unlink_lock_skiplist(struct ldlm_lock *req) +{ + if (req->l_resource->lr_type != LDLM_PLAIN && + req->l_resource->lr_type != LDLM_IBITS) + return; + + cfs_list_del_init(&req->l_sl_policy); + cfs_list_del_init(&req->l_sl_mode); } void ldlm_lock_cancel(struct ldlm_lock *lock) @@ -1039,15 +1674,11 @@ void ldlm_lock_cancel(struct ldlm_lock *lock) struct ldlm_namespace *ns; ENTRY; - /* There's no race between calling this and taking the ns lock below; - * a lock can only be put on the waiting list once, because it can only - * issue a blocking AST once. */ - ldlm_del_waiting_lock(lock); + lock_res_and_lock(lock); res = lock->l_resource; - ns = res->lr_namespace; + ns = ldlm_res_to_ns(res); - l_lock(&ns->ns_lock); /* Please do not, no matter how tempting, remove this LBUG without * talking to me first. -phik */ if (lock->l_readers || lock->l_writers) { @@ -1055,104 +1686,212 @@ void ldlm_lock_cancel(struct ldlm_lock *lock) LBUG(); } + ldlm_del_waiting_lock(lock); + + /* Releases cancel callback. */ ldlm_cancel_callback(lock); + /* Yes, second time, just in case it was added again while we were + running with no res lock in ldlm_cancel_callback */ + ldlm_del_waiting_lock(lock); ldlm_resource_unlink_lock(lock); - ldlm_lock_destroy(lock); - l_unlock(&ns->ns_lock); + ldlm_lock_destroy_nolock(lock); + + if (lock->l_granted_mode == lock->l_req_mode) + ldlm_pool_del(&ns->ns_pool, lock); + + /* Make sure we will not be called again for same lock what is possible + * if not to zero out lock->l_granted_mode */ + lock->l_granted_mode = LCK_MINMODE; + unlock_res_and_lock(lock); + EXIT; } int ldlm_lock_set_data(struct lustre_handle *lockh, void *data) { struct ldlm_lock *lock = ldlm_handle2lock(lockh); + int rc = -EINVAL; ENTRY; - if (lock == NULL) - RETURN(-EINVAL); + if (lock) { + if (lock->l_ast_data == NULL) + lock->l_ast_data = data; + if (lock->l_ast_data == data) + rc = 0; + LDLM_LOCK_PUT(lock); + } + RETURN(rc); +} +EXPORT_SYMBOL(ldlm_lock_set_data); - lock->l_ast_data = data; - LDLM_LOCK_PUT(lock); - RETURN(0); +int ldlm_cancel_locks_for_export_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd, + cfs_hlist_node_t *hnode, void *data) + +{ + struct obd_export *exp = data; + struct ldlm_lock *lock = cfs_hash_object(hs, hnode); + struct ldlm_resource *res; + + res = ldlm_resource_getref(lock->l_resource); + LDLM_LOCK_GET(lock); + + LDLM_DEBUG(lock, "export %p", exp); + ldlm_res_lvbo_update(res, NULL, 1); + ldlm_lock_cancel(lock); + ldlm_reprocess_all(res); + ldlm_resource_putref(res); + LDLM_LOCK_RELEASE(lock); + return 0; } void ldlm_cancel_locks_for_export(struct obd_export *exp) { - struct ldlm_namespace *ns = exp->exp_obd->obd_namespace; - struct ldlm_lock *lock; - struct ldlm_resource *res; + cfs_hash_for_each_empty(exp->exp_lock_hash, + ldlm_cancel_locks_for_export_cb, exp); +} - l_lock(&ns->ns_lock); - while(!list_empty(&exp->exp_ldlm_data.led_held_locks)) { - lock = list_entry(exp->exp_ldlm_data.led_held_locks.next, - struct ldlm_lock, l_export_chain); - res = ldlm_resource_getref(lock->l_resource); - LDLM_DEBUG(lock, "export %p", exp); - ldlm_lock_cancel(lock); - l_unlock(&ns->ns_lock); - ldlm_reprocess_all(res); - ldlm_resource_putref(res); - l_lock(&ns->ns_lock); - } - l_unlock(&ns->ns_lock); +/** + * Downgrade an exclusive lock. + * + * A fast variant of ldlm_lock_convert for convertion of exclusive + * locks. The convertion is always successful. + * + * \param lock A lock to convert + * \param new_mode new lock mode + */ +void ldlm_lock_downgrade(struct ldlm_lock *lock, int new_mode) +{ + ENTRY; + + LASSERT(lock->l_granted_mode & (LCK_PW | LCK_EX)); + LASSERT(new_mode == LCK_COS); + + lock_res_and_lock(lock); + ldlm_resource_unlink_lock(lock); + /* + * Remove the lock from pool as it will be added again in + * ldlm_grant_lock() called below. + */ + ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock); + + lock->l_req_mode = new_mode; + ldlm_grant_lock(lock, NULL); + unlock_res_and_lock(lock); + ldlm_reprocess_all(lock->l_resource); + + EXIT; } struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode, - int *flags) + __u32 *flags) { - struct list_head rpc_list = LIST_HEAD_INIT(rpc_list); + CFS_LIST_HEAD(rpc_list); struct ldlm_resource *res; struct ldlm_namespace *ns; int granted = 0; + int old_mode, rc; + struct sl_insert_point prev; + ldlm_error_t err; + struct ldlm_interval *node; ENTRY; - LBUG(); + if (new_mode == lock->l_granted_mode) { // No changes? Just return. + *flags |= LDLM_FL_BLOCK_GRANTED; + RETURN(lock->l_resource); + } - res = lock->l_resource; - ns = res->lr_namespace; + /* I can't check the type of lock here because the bitlock of lock + * is not held here, so do the allocation blindly. -jay */ + OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO); + if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */ + RETURN(NULL); + + LASSERTF((new_mode == LCK_PW && lock->l_granted_mode == LCK_PR), + "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode); + + lock_res_and_lock(lock); - l_lock(&ns->ns_lock); + res = lock->l_resource; + ns = ldlm_res_to_ns(res); + old_mode = lock->l_req_mode; lock->l_req_mode = new_mode; - ldlm_resource_unlink_lock(lock); + if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) { + /* remember the lock position where the lock might be + * added back to the granted list later and also + * remember the join mode for skiplist fixing. */ + prev.res_link = lock->l_res_link.prev; + prev.mode_link = lock->l_sl_mode.prev; + prev.policy_link = lock->l_sl_policy.prev; + ldlm_resource_unlink_lock(lock); + } else { + ldlm_resource_unlink_lock(lock); + if (res->lr_type == LDLM_EXTENT) { + /* FIXME: ugly code, I have to attach the lock to a + * interval node again since perhaps it will be granted + * soon */ + CFS_INIT_LIST_HEAD(&node->li_group); + ldlm_interval_attach(node, lock); + node = NULL; + } + } + + /* + * Remove old lock from the pool before adding the lock with new + * mode below in ->policy() + */ + ldlm_pool_del(&ns->ns_pool, lock); /* If this is a local resource, put it on the appropriate list. */ - if (res->lr_namespace->ns_client) { + if (ns_is_client(ldlm_res_to_ns(res))) { if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) { ldlm_resource_add_lock(res, &res->lr_converting, lock); } else { /* This should never happen, because of the way the * server handles conversions. */ + LDLM_ERROR(lock, "Erroneous flags %d on local lock\n", + *flags); LBUG(); - res->lr_tmp = &rpc_list; - ldlm_grant_lock(lock, NULL, 0, 0); - res->lr_tmp = NULL; + ldlm_grant_lock(lock, &rpc_list); granted = 1; - /* FIXME: completion handling not with ns_lock held ! */ + /* FIXME: completion handling not with lr_lock held ! */ if (lock->l_completion_ast) lock->l_completion_ast(lock, 0, NULL); } } else { - /* FIXME: We should try the conversion right away and possibly - * return success without the need for an extra AST */ - ldlm_resource_add_lock(res, &res->lr_converting, lock); - *flags |= LDLM_FL_BLOCK_CONV; + int pflags = 0; + ldlm_processing_policy policy; + policy = ldlm_processing_policy_table[res->lr_type]; + rc = policy(lock, &pflags, 0, &err, &rpc_list); + if (rc == LDLM_ITER_STOP) { + lock->l_req_mode = old_mode; + if (res->lr_type == LDLM_EXTENT) + ldlm_extent_add_lock(res, lock); + else + ldlm_granted_list_add_lock(lock, &prev); + + res = NULL; + } else { + *flags |= LDLM_FL_BLOCK_GRANTED; + granted = 1; + } } - - l_unlock(&ns->ns_lock); + unlock_res_and_lock(lock); if (granted) - ldlm_run_ast_work(ns, &rpc_list); + ldlm_run_ast_work(ns, &rpc_list, LDLM_WORK_CP_AST); + if (node) + OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node)); RETURN(res); } void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos) { - char str[PTL_NALFMT_SIZE]; struct obd_device *obd = NULL; - if (!((portal_debug | D_ERROR) & level)) + if (!((libcfs_debug | D_ERROR) & level)) return; if (!lock) { @@ -1160,32 +1899,33 @@ void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos) return; } - CDEBUG(level, " -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d)\n", - lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc), - pos); + CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n", + lock, lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc), + pos, lock->l_pid); if (lock->l_conn_export != NULL) obd = lock->l_conn_export->exp_obd; if (lock->l_export && lock->l_export->exp_connection) { - CDEBUG(level, " Node: NID %s on %s (rhandle: "LPX64")\n", - ptlrpc_peernid2str(&lock->l_export->exp_connection->c_peer, str), - lock->l_export->exp_connection->c_peer.peer_ni->pni_name, - lock->l_remote_handle.cookie); + CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n", + libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid), + lock->l_remote_handle.cookie); } else if (obd == NULL) { CDEBUG(level, " Node: local\n"); } else { struct obd_import *imp = obd->u.cli.cl_import; - CDEBUG(level, " Node: NID %s on %s (rhandle: "LPX64")\n", - ptlrpc_peernid2str(&imp->imp_connection->c_peer, str), - imp->imp_connection->c_peer.peer_ni->pni_name, + CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n", + libcfs_nid2str(imp->imp_connection->c_peer.nid), lock->l_remote_handle.cookie); } - CDEBUG(level, " Resource: %p ("LPU64"/"LPU64")\n", lock->l_resource, - lock->l_resource->lr_name.name[0], - lock->l_resource->lr_name.name[1]); + CDEBUG(level, " Resource: %p ("LPU64"/"LPU64"/"LPU64")\n", + lock->l_resource, + lock->l_resource->lr_name.name[0], + lock->l_resource->lr_name.name[1], + lock->l_resource->lr_name.name[2]); CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, " - "write: %d\n", ldlm_lockname[lock->l_req_mode], + "write: %d flags: "LPX64"\n", ldlm_lockname[lock->l_req_mode], ldlm_lockname[lock->l_granted_mode], - atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers); + cfs_atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers, + lock->l_flags); if (lock->l_resource->lr_type == LDLM_EXTENT) CDEBUG(level, " Extent: "LPU64" -> "LPU64 " (req "LPU64"-"LPU64")\n", @@ -1193,12 +1933,12 @@ void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos) lock->l_policy_data.l_extent.end, lock->l_req_extent.start, lock->l_req_extent.end); else if (lock->l_resource->lr_type == LDLM_FLOCK) - CDEBUG(level, " Pid: "LPU64" Extent: "LPU64" -> "LPU64"\n", + CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n", lock->l_policy_data.l_flock.pid, lock->l_policy_data.l_flock.start, lock->l_policy_data.l_flock.end); - else if (lock->l_resource->lr_type == LDLM_IBITS) - CDEBUG(level, " Bits: "LPX64"\n", + else if (lock->l_resource->lr_type == LDLM_IBITS) + CDEBUG(level, " Bits: "LPX64"\n", lock->l_policy_data.l_inodebits.bits); } @@ -1206,6 +1946,9 @@ void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh) { struct ldlm_lock *lock; + if (!((libcfs_debug | D_ERROR) & level)) + return; + lock = ldlm_handle2lock(lockh); if (lock == NULL) return; @@ -1214,3 +1957,131 @@ void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh) LDLM_LOCK_PUT(lock); } + +void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level, + struct libcfs_debug_msg_data *data, const char *fmt, + ...) +{ + va_list args; + cfs_debug_limit_state_t *cdls = data->msg_cdls; + + va_start(args, fmt); + + if (lock->l_resource == NULL) { + libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file, + data->msg_fn, data->msg_line, fmt, args, + " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s " + "res: \?\? rrc=\?\? type: \?\?\? flags: "LPX64" remote: " + LPX64" expref: %d pid: %u timeout: %lu\n", lock, + lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc), + lock->l_readers, lock->l_writers, + ldlm_lockname[lock->l_granted_mode], + ldlm_lockname[lock->l_req_mode], + lock->l_flags, lock->l_remote_handle.cookie, + lock->l_export ? + cfs_atomic_read(&lock->l_export->exp_refcount) : -99, + lock->l_pid, lock->l_callback_timeout); + va_end(args); + return; + } + + switch (lock->l_resource->lr_type) { + case LDLM_EXTENT: + libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file, + data->msg_fn, data->msg_line, fmt, args, + " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s " + "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64 + "] (req "LPU64"->"LPU64") flags: "LPX64" remote: "LPX64 + " expref: %d pid: %u timeout %lu\n", + ldlm_lock_to_ns_name(lock), lock, + lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc), + lock->l_readers, lock->l_writers, + ldlm_lockname[lock->l_granted_mode], + ldlm_lockname[lock->l_req_mode], + lock->l_resource->lr_name.name[0], + lock->l_resource->lr_name.name[1], + cfs_atomic_read(&lock->l_resource->lr_refcount), + ldlm_typename[lock->l_resource->lr_type], + lock->l_policy_data.l_extent.start, + lock->l_policy_data.l_extent.end, + lock->l_req_extent.start, lock->l_req_extent.end, + lock->l_flags, lock->l_remote_handle.cookie, + lock->l_export ? + cfs_atomic_read(&lock->l_export->exp_refcount) : -99, + lock->l_pid, lock->l_callback_timeout); + break; + + case LDLM_FLOCK: + libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file, + data->msg_fn, data->msg_line, fmt, args, + " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s " + "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d " + "["LPU64"->"LPU64"] flags: "LPX64" remote: "LPX64 + " expref: %d pid: %u timeout: %lu\n", + ldlm_lock_to_ns_name(lock), lock, + lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc), + lock->l_readers, lock->l_writers, + ldlm_lockname[lock->l_granted_mode], + ldlm_lockname[lock->l_req_mode], + lock->l_resource->lr_name.name[0], + lock->l_resource->lr_name.name[1], + cfs_atomic_read(&lock->l_resource->lr_refcount), + ldlm_typename[lock->l_resource->lr_type], + lock->l_policy_data.l_flock.pid, + lock->l_policy_data.l_flock.start, + lock->l_policy_data.l_flock.end, + lock->l_flags, lock->l_remote_handle.cookie, + lock->l_export ? + cfs_atomic_read(&lock->l_export->exp_refcount) : -99, + lock->l_pid, lock->l_callback_timeout); + break; + + case LDLM_IBITS: + libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file, + data->msg_fn, data->msg_line, fmt, args, + " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s " + "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s " + "flags: "LPX64" remote: "LPX64" expref: %d " + "pid: %u timeout: %lu\n", + ldlm_lock_to_ns_name(lock), + lock, lock->l_handle.h_cookie, + cfs_atomic_read (&lock->l_refc), + lock->l_readers, lock->l_writers, + ldlm_lockname[lock->l_granted_mode], + ldlm_lockname[lock->l_req_mode], + lock->l_resource->lr_name.name[0], + lock->l_resource->lr_name.name[1], + lock->l_policy_data.l_inodebits.bits, + cfs_atomic_read(&lock->l_resource->lr_refcount), + ldlm_typename[lock->l_resource->lr_type], + lock->l_flags, lock->l_remote_handle.cookie, + lock->l_export ? + cfs_atomic_read(&lock->l_export->exp_refcount) : -99, + lock->l_pid, lock->l_callback_timeout); + break; + + default: + libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file, + data->msg_fn, data->msg_line, fmt, args, + " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s " + "res: "LPU64"/"LPU64" rrc: %d type: %s flags: "LPX64" " + "remote: "LPX64" expref: %d pid: %u timeout %lu\n", + ldlm_lock_to_ns_name(lock), + lock, lock->l_handle.h_cookie, + cfs_atomic_read (&lock->l_refc), + lock->l_readers, lock->l_writers, + ldlm_lockname[lock->l_granted_mode], + ldlm_lockname[lock->l_req_mode], + lock->l_resource->lr_name.name[0], + lock->l_resource->lr_name.name[1], + cfs_atomic_read(&lock->l_resource->lr_refcount), + ldlm_typename[lock->l_resource->lr_type], + lock->l_flags, lock->l_remote_handle.cookie, + lock->l_export ? + cfs_atomic_read(&lock->l_export->exp_refcount) : -99, + lock->l_pid, lock->l_callback_timeout); + break; + } + va_end(args); +} +EXPORT_SYMBOL(_ldlm_lock_debug);