X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_lock.c;h=e83d3966b947444b3b49ee0a80054289ccb82535;hp=a361cabc258a52c1c161a6061bbbd3385879ae88;hb=ec295cad998d4d5e7fc915491da5ac646dbc6af6;hpb=44f345e83c48935f4efc4750804f192a728a630e diff --git a/lustre/ldlm/ldlm_lock.c b/lustre/ldlm/ldlm_lock.c index a361cab..e83d396 100644 --- a/lustre/ldlm/ldlm_lock.c +++ b/lustre/ldlm/ldlm_lock.c @@ -1,36 +1,48 @@ /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- * vim:expandtab:shiftwidth=8:tabstop=8: * - * Copyright (c) 2002 Cluster File Systems, Inc. + * Copyright (c) 2002, 2003 Cluster File Systems, Inc. * Author: Peter Braam * Author: Phil Schwan * - * This file is part of Lustre, http://www.lustre.org. + * This file is part of the Lustre file system, http://www.lustre.org + * Lustre is a trademark of Cluster File Systems, Inc. * - * 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. + * You may have signed or agreed to another license before downloading + * this software. If so, you are bound by the terms and conditions + * of that agreement, and the following does not apply to you. See the + * LICENSE file included with this distribution for more information. * - * 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. + * If you did not agree to a different license, then this copy of Lustre + * is open source 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. * - * 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. + * In either case, 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 + * license text for more details. */ #define DEBUG_SUBSYSTEM S_LDLM -#include -#include -#include -#include -#include -#include +#ifdef __KERNEL__ +# include +# include +#else +# include +# include +#endif -//struct lustre_lock ldlm_everything_lock; +#include +#include "ldlm_internal.h" + +/* lock's skip list pointers fix mode */ +#define LDLM_JOIN_NONE 0 +#define LDLM_MODE_JOIN_RIGHT 1 +#define LDLM_MODE_JOIN_LEFT (1 << 1) +#define LDLM_POLICY_JOIN_RIGHT (1 << 2) +#define LDLM_POLICY_JOIN_LEFT (1 << 3) /* lock types */ char *ldlm_lockname[] = { @@ -40,12 +52,15 @@ char *ldlm_lockname[] = { [LCK_PR] "PR", [LCK_CW] "CW", [LCK_CR] "CR", - [LCK_NL] "NL" + [LCK_NL] "NL", + [LCK_GROUP] "GROUP" }; + char *ldlm_typename[] = { [LDLM_PLAIN] "PLN", [LDLM_EXTENT] "EXT", - [LDLM_MDSINTENT] "INT" + [LDLM_FLOCK] "FLK", + [LDLM_IBITS] "IBT", }; char *ldlm_it2str(int it) @@ -57,65 +72,41 @@ char *ldlm_it2str(int it) return "creat"; case (IT_OPEN | IT_CREAT): return "open|creat"; - case IT_MKDIR: - return "mkdir"; - case IT_LINK: - return "link"; - case IT_LINK2: - return "link2"; - case IT_SYMLINK: - return "symlink"; - case IT_UNLINK: - return "unlink"; - case IT_RMDIR: - return "rmdir"; - case IT_RENAME: - return "rename"; - case IT_RENAME2: - return "rename2"; case IT_READDIR: return "readdir"; case IT_GETATTR: return "getattr"; - case IT_SETATTR: - return "setattr"; - case IT_READLINK: - return "readlink"; - case IT_MKNOD: - return "mknod"; case IT_LOOKUP: return "lookup"; + case IT_UNLINK: + return "unlink"; + case IT_GETXATTR: + return "getxattr"; default: CERROR("Unknown intent %d\n", it); return "UNKNOWN"; } } -extern kmem_cache_t *ldlm_lock_slab; +extern cfs_mem_cache_t *ldlm_lock_slab; -static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b); - -ldlm_res_compat ldlm_res_compat_table[] = { - [LDLM_PLAIN] ldlm_plain_compat, - [LDLM_EXTENT] ldlm_extent_compat, - [LDLM_MDSINTENT] ldlm_plain_compat -}; - -ldlm_res_policy ldlm_res_policy_table[] = { - [LDLM_PLAIN] NULL, - [LDLM_EXTENT] ldlm_extent_policy, - [LDLM_MDSINTENT] NULL +static ldlm_processing_policy ldlm_processing_policy_table[] = { + [LDLM_PLAIN] ldlm_process_plain_lock, + [LDLM_EXTENT] ldlm_process_extent_lock, +#ifdef __KERNEL__ + [LDLM_FLOCK] ldlm_process_flock_lock, +#endif + [LDLM_IBITS] ldlm_process_inodebits_lock, }; -void ldlm_register_intent(int (*arg) (struct ldlm_lock * lock, void *req_cookie, - ldlm_mode_t mode, void *data)) +ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res) { - ldlm_res_policy_table[LDLM_MDSINTENT] = arg; + return ldlm_processing_policy_table[res->lr_type]; } -void ldlm_unregister_intent(void) +void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg) { - ldlm_res_policy_table[LDLM_MDSINTENT] = NULL; + ns->ns_policy = arg; } /* @@ -131,104 +122,199 @@ void ldlm_unregister_intent(void) */ struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock) { - l_lock(&lock->l_resource->lr_namespace->ns_lock); - lock->l_refc++; - ldlm_resource_getref(lock->l_resource); - l_unlock(&lock->l_resource->lr_namespace->ns_lock); + 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) { - struct ldlm_namespace *ns = lock->l_resource->lr_namespace; ENTRY; - l_lock(&ns->ns_lock); - lock->l_refc--; - //LDLM_DEBUG(lock, "after refc--"); - if (lock->l_refc < 0) - LBUG(); + LASSERT(lock->l_resource != LP_POISON); + LASSERT(atomic_read(&lock->l_refc) > 0); + if (atomic_dec_and_test(&lock->l_refc)) { + struct ldlm_resource *res; - if (ldlm_resource_put(lock->l_resource)) { - LASSERT(lock->l_refc == 0); + LDLM_DEBUG(lock, + "final lock_put on destroyed lock, freeing it."); + + res = lock->l_resource; + LASSERT(lock->l_destroyed); + LASSERT(list_empty(&lock->l_res_link)); + LASSERT(list_empty(&lock->l_pending_chain)); + + atomic_dec(&res->lr_namespace->ns_locks); + ldlm_resource_putref(res); lock->l_resource = NULL; - } - if (lock->l_parent) - LDLM_LOCK_PUT(lock->l_parent); + if (lock->l_export) { + class_export_put(lock->l_export); + lock->l_export = NULL; + } - if (lock->l_refc == 0 && (lock->l_flags & LDLM_FL_DESTROYED)) { - l_unlock(&ns->ns_lock); - LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing"); + if (lock->l_lvb_data != NULL) + OBD_FREE(lock->l_lvb_data, lock->l_lvb_len); - //spin_lock(&ldlm_handle_lock); - spin_lock(&ns->ns_counter_lock); - ns->ns_locks--; - spin_unlock(&ns->ns_counter_lock); + OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle, + ldlm_lock_free); + } - lock->l_resource = NULL; - lock->l_random = DEAD_HANDLE_MAGIC; - if (lock->l_export && lock->l_export->exp_connection) - ptlrpc_put_connection(lock->l_export->exp_connection); - kmem_cache_free(ldlm_lock_slab, lock); - //spin_unlock(&ldlm_handle_lock); - CDEBUG(D_MALLOC, "kfreed 'lock': %d at %p (tot 0).\n", - sizeof(*lock), lock); - } else - l_unlock(&ns->ns_lock); + EXIT; +} + +int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock) +{ + int rc = 0; + if (!list_empty(&lock->l_lru)) { + struct ldlm_namespace *ns = lock->l_resource->lr_namespace; + LASSERT(lock->l_resource->lr_type != LDLM_FLOCK); + list_del_init(&lock->l_lru); + ns->ns_nr_unused--; + LASSERT(ns->ns_nr_unused >= 0); + rc = 1; + } + return rc; +} +int ldlm_lock_remove_from_lru(struct ldlm_lock *lock) +{ + struct ldlm_namespace *ns = lock->l_resource->lr_namespace; + int rc; + ENTRY; + spin_lock(&ns->ns_unused_lock); + rc = ldlm_lock_remove_from_lru_nolock(lock); + spin_unlock(&ns->ns_unused_lock); EXIT; + return rc; } -void ldlm_lock_destroy(struct ldlm_lock *lock) +void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock) +{ + struct ldlm_namespace *ns = lock->l_resource->lr_namespace; + lock->l_last_used = cfs_time_current(); + LASSERT(list_empty(&lock->l_lru)); + 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 = lock->l_resource->lr_namespace; ENTRY; - l_lock(&lock->l_resource->lr_namespace->ns_lock); + spin_lock(&ns->ns_unused_lock); + ldlm_lock_add_to_lru_nolock(lock); + spin_unlock(&ns->ns_unused_lock); + EXIT; +} - if (!list_empty(&lock->l_children)) { - LDLM_DEBUG(lock, "still has children (%p)!", - lock->l_children.next); - ldlm_lock_dump(lock); - LBUG(); +void ldlm_lock_touch_in_lru(struct ldlm_lock *lock) +{ + struct ldlm_namespace *ns = lock->l_resource->lr_namespace; + ENTRY; + spin_lock(&ns->ns_unused_lock); + if (!list_empty(&lock->l_lru)) { + ldlm_lock_remove_from_lru_nolock(lock); + ldlm_lock_add_to_lru_nolock(lock); } + spin_unlock(&ns->ns_unused_lock); + EXIT; +} + +/* 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 */ +int ldlm_lock_destroy_internal(struct ldlm_lock *lock) +{ + ENTRY; + if (lock->l_readers || lock->l_writers) { - LDLM_DEBUG(lock, "lock still has references"); - ldlm_lock_dump(lock); + LDLM_ERROR(lock, "lock still has references"); + ldlm_lock_dump(D_ERROR, lock, 0); + LBUG(); } if (!list_empty(&lock->l_res_link)) { - ldlm_lock_dump(lock); + LDLM_ERROR(lock, "lock still on resource"); + ldlm_lock_dump(D_ERROR, lock, 0); LBUG(); } - if (lock->l_flags & LDLM_FL_DESTROYED) { - l_unlock(&lock->l_resource->lr_namespace->ns_lock); + if (lock->l_destroyed) { + LASSERT(list_empty(&lock->l_lru)); EXIT; - return; + return 0; } + lock->l_destroyed = 1; - list_del(&lock->l_export_chain); - lock->l_export = NULL; - lock->l_flags |= LDLM_FL_DESTROYED; + if (lock->l_export) + spin_lock(&lock->l_export->exp_ldlm_data.led_lock); + list_del_init(&lock->l_export_chain); + if (lock->l_export) + spin_unlock(&lock->l_export->exp_ldlm_data.led_lock); + + ldlm_lock_remove_from_lru(lock); + class_handle_unhash(&lock->l_handle); +#if 0 /* Wake anyone waiting for this lock */ /* FIXME: I should probably add yet another flag, instead of using * l_export to only call this on clients */ + if (lock->l_export) + class_export_put(lock->l_export); + lock->l_export = NULL; 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) + LDLM_LOCK_PUT(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) + LDLM_LOCK_PUT(lock); EXIT; } +/* this is called by portals_handle2object with the handle lock taken */ +static void lock_handle_addref(void *lock) +{ + LDLM_LOCK_GET((struct ldlm_lock *)lock); +} + /* - usage: pass in a resource on which you have done get - pass in a parent lock on which you have done a get - do not put the resource or the parent - returns: lock with refcount 1 -*/ -static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent, - struct ldlm_resource *resource) + * 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 2 - one for current caller and one for remote + */ +static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource) { struct ldlm_lock *lock; ENTRY; @@ -236,77 +322,77 @@ static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent, if (resource == NULL) LBUG(); - lock = kmem_cache_alloc(ldlm_lock_slab, SLAB_KERNEL); + OBD_SLAB_ALLOC(lock, ldlm_lock_slab, CFS_ALLOC_IO, sizeof(*lock)); if (lock == NULL) RETURN(NULL); - memset(lock, 0, sizeof(*lock)); - get_random_bytes(&lock->l_random, sizeof(__u64)); - - lock->l_resource = resource; - /* this refcount matches the one of the resource passed - in which is not being put away */ - lock->l_refc = 1; - INIT_LIST_HEAD(&lock->l_children); - INIT_LIST_HEAD(&lock->l_res_link); - 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 = parent; - list_add(&lock->l_childof, &parent->l_children); - l_unlock(&parent->l_resource->lr_namespace->ns_lock); - } - - CDEBUG(D_MALLOC, "kmalloced 'lock': %d at " - "%p (tot %d).\n", sizeof(*lock), lock, 1); - /* this is the extra refcount, to prevent the lock from evaporating */ - LDLM_LOCK_GET(lock); + spin_lock_init(&lock->l_lock); + lock->l_resource = ldlm_resource_getref(resource); + + 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_export_chain); + 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_waitq_init(&lock->l_waitq); + lock->l_blocking_lock = NULL; + lock->l_sl_mode.prev = NULL; + lock->l_sl_mode.next = NULL; + lock->l_sl_policy.prev = NULL; + lock->l_sl_policy.next = NULL; + + atomic_inc(&resource->lr_namespace->ns_locks); + CFS_INIT_LIST_HEAD(&lock->l_handle.h_link); + class_handle_hash(&lock->l_handle, lock_handle_addref); + RETURN(lock); } -int ldlm_lock_change_resource(struct ldlm_lock *lock, __u64 new_resid[3]) +int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock, + const struct ldlm_res_id *new_resid) { - struct ldlm_namespace *ns = lock->l_resource->lr_namespace; struct ldlm_resource *oldres = lock->l_resource; - int type, i; + 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); } - type = lock->l_resource->lr_type; - if (new_resid[0] == 0) - LBUG(); - lock->l_resource = ldlm_resource_get(ns, NULL, new_resid, type, 1); - if (lock->l_resource == NULL) { + LASSERT(new_resid->name[0] != 0); + + /* This function assumes that the lock isn't on any lists */ + LASSERT(list_empty(&lock->l_res_link)); + + type = oldres->lr_type; + unlock_res_and_lock(lock); + + newres = ldlm_resource_get(ns, NULL, new_resid, type, 1); + if (newres == NULL) { LBUG(); RETURN(-ENOMEM); } - /* move references over */ - for (i = 0; i < lock->l_refc; i++) { - int rc; - ldlm_resource_getref(lock->l_resource); - rc = ldlm_resource_put(oldres); - if (rc == 1 && i != lock->l_refc - 1) - LBUG(); - } - /* compensate for the initial get above.. */ - ldlm_resource_put(lock->l_resource); + lock_res_and_lock(lock); + LASSERT(memcmp(new_resid, &lock->l_resource->lr_name, + sizeof(lock->l_resource->lr_name)) != 0); + lock_res(newres); + lock->l_resource = newres; + unlock_res(oldres); + unlock_res_and_lock(lock); + + /* ...and the flowers are still standing! */ + ldlm_resource_putref(oldres); - l_unlock(&ns->ns_lock); RETURN(0); } @@ -314,103 +400,153 @@ int ldlm_lock_change_resource(struct ldlm_lock *lock, __u64 new_resid[3]) * 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->addr = (__u64) (unsigned long)lock; - lockh->cookie = lock->l_random; + lockh->cookie = lock->l_handle.h_cookie; } -struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, - int strict) +/* if flags: atomically get the lock and set the flags. + * Return NULL if flag already set + */ + +struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle, + int flags) { - struct ldlm_lock *lock = NULL, *retval = NULL; + struct ldlm_namespace *ns; + struct ldlm_lock *lock, *retval = NULL; ENTRY; - if (!handle || !handle->addr) + LASSERT(handle); + + lock = class_handle2object(handle->cookie); + if (lock == NULL) RETURN(NULL); - //spin_lock(&ldlm_handle_lock); - lock = (struct ldlm_lock *)(unsigned long)(handle->addr); - if (!kmem_cache_validate(ldlm_lock_slab, (void *)lock)) { - CERROR("bogus lock %p\n", lock); - GOTO(out2, retval); - } + LASSERT(lock->l_resource != NULL); + ns = lock->l_resource->lr_namespace; + LASSERT(ns != NULL); - if (lock->l_random != handle->cookie) { - CERROR("bogus cookie: lock %p has "LPX64" vs. handle "LPX64"\n", - lock, lock->l_random, handle->cookie); - GOTO(out2, NULL); - } - if (!lock->l_resource) { - CERROR("trying to lock bogus resource: lock %p\n", lock); - LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock); - GOTO(out2, retval); - } - if (!lock->l_resource->lr_namespace) { - CERROR("trying to lock bogus namespace: lock %p\n", lock); - LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock); - GOTO(out2, retval); + lock_res_and_lock(lock); + + /* It's unlikely but possible that someone marked the lock as + * destroyed after we did handle2object on it */ + if (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); } - l_lock(&lock->l_resource->lr_namespace->ns_lock); - if (strict && lock->l_flags & LDLM_FL_DESTROYED) { - CERROR("lock already destroyed: lock %p\n", lock); - LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock); - GOTO(out, NULL); + if (flags && (lock->l_flags & flags)) { + unlock_res_and_lock(lock); + LDLM_LOCK_PUT(lock); + GOTO(out, retval); } - retval = LDLM_LOCK_GET(lock); - if (!retval) - CERROR("lock disappeared below us!!! %p\n", lock); + if (flags) + lock->l_flags |= flags; + + unlock_res_and_lock(lock); + retval = lock; EXIT; out: - l_unlock(&lock->l_resource->lr_namespace->ns_lock); - out2: - //spin_unlock(&ldlm_handle_lock); return retval; } -static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b) +struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns, + const struct lustre_handle *handle) { - return lockmode_compat(a->l_req_mode, b->l_req_mode); + struct ldlm_lock *retval = NULL; + retval = __ldlm_handle2lock(handle, 0); + return retval; } 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_extent, &lock->l_extent, sizeof(desc->l_extent)); - memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version)); + 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))) { + struct ldlm_resource res = *lock->l_resource; + + /* 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); + res.lr_type = LDLM_PLAIN; + ldlm_res2desc(&res, &desc->l_resource); + /* 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; + } + + /* 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; + desc->l_policy_data = lock->l_policy_data; + } } -static void ldlm_add_ast_work_item(struct ldlm_lock *lock, - struct ldlm_lock *new) +void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new, + struct list_head *work_list) { - 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); - - OBD_ALLOC(w, sizeof(*w)); - if (!w) { - LBUG(); - GOTO(out, 0); + 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; + LASSERT(list_empty(&lock->l_bl_ast)); + 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); } +} - if (new) { - lock->l_flags |= LDLM_FL_AST_SENT; - w->w_blocking = 1; - ldlm_lock2desc(new, &w->w_desc); +void ldlm_add_cp_work_item(struct ldlm_lock *lock, struct list_head *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(list_empty(&lock->l_cp_ast)); + 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); - out: - l_unlock(&lock->l_resource->lr_namespace->ns_lock); - return; +/* must be called with lr_lock held */ +void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new, + struct list_head *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; } void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode) @@ -418,417 +554,933 @@ 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); - if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR) + ldlm_lock_remove_from_lru(lock); + if (mode & (LCK_NL | LCK_CR | LCK_PR)) lock->l_readers++; - else + if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) lock->l_writers++; - l_unlock(&lock->l_resource->lr_namespace->ns_lock); LDLM_LOCK_GET(lock); LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]); } -/* Args: unlocked lock */ -void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode) +/* only called for local locks */ +void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode) { - struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0); + lock_res_and_lock(lock); + ldlm_lock_addref_internal_nolock(lock, mode); + unlock_res_and_lock(lock); +} + +void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode) +{ + struct ldlm_namespace *ns; ENTRY; - if (lock == NULL) - LBUG(); + lock_res_and_lock(lock); + + ns = lock->l_resource->lr_namespace; LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]); - l_lock(&lock->l_resource->lr_namespace->ns_lock); - if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR) + if (mode & (LCK_NL | LCK_CR | LCK_PR)) { + LASSERT(lock->l_readers > 0); lock->l_readers--; - else + } + if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) { + LASSERT(lock->l_writers > 0); lock->l_writers--; + } + + 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 + * the last reference, cancel the lock. */ + CDEBUG(D_INFO, "forcing cancel of local lock\n"); + lock->l_flags |= LDLM_FL_CBPENDING; + } - /* If we received a blocked AST and this was the last reference, - * run the callback. */ if (!lock->l_readers && !lock->l_writers && (lock->l_flags & LDLM_FL_CBPENDING)) { - if (!lock->l_resource->lr_namespace->ns_client && - lock->l_export) + /* If we received a blocked AST and this was the last reference, + * run the callback. */ + if (ns_is_server(ns) && lock->l_export) CERROR("FL_CBPENDING set on non-local lock--just a " "warning\n"); LDLM_DEBUG(lock, "final decref done on cbpending lock"); - l_unlock(&lock->l_resource->lr_namespace->ns_lock); - /* FIXME: need a real 'desc' here */ - lock->l_blocking_ast(lock, NULL, lock->l_data, - lock->l_data_len, LDLM_CB_BLOCKING); - } else - l_unlock(&lock->l_resource->lr_namespace->ns_lock); + LDLM_LOCK_GET(lock); /* dropped by bl thread */ + ldlm_lock_remove_from_lru(lock); + unlock_res_and_lock(lock); + 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_NO_LRU)) { + /* If this is a client-side namespace and this was the last + * reference, put it on the LRU. */ + ldlm_lock_add_to_lru(lock); + unlock_res_and_lock(lock); + /* 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) && + !exp_connect_lru_resize(lock->l_conn_export)) + ldlm_cancel_lru(ns, 0, LDLM_ASYNC); + } else { + unlock_res_and_lock(lock); + } LDLM_LOCK_PUT(lock); /* matches the ldlm_lock_get in addref */ - LDLM_LOCK_PUT(lock); /* matches the handle2lock above */ EXIT; } -static int ldlm_lock_compat_list(struct ldlm_lock *lock, int send_cbs, - struct list_head *queue) +void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode) { - struct list_head *tmp, *pos; - int rc = 1; + struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0); + LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie); + ldlm_lock_decref_internal(lock, mode); + LDLM_LOCK_PUT(lock); +} - list_for_each_safe(tmp, pos, queue) { - struct ldlm_lock *child; - ldlm_res_compat compat; +/* This will drop a lock reference and mark it for destruction, but will not + * necessarily cancel the lock before returning. */ +void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode) +{ + struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0); + ENTRY; - child = list_entry(tmp, struct ldlm_lock, l_res_link); - if (lock == child) - continue; + LASSERT(lock != NULL); + + LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]); + lock_res_and_lock(lock); + lock->l_flags |= LDLM_FL_CBPENDING; + unlock_res_and_lock(lock); + ldlm_lock_decref_internal(lock, mode); + LDLM_LOCK_PUT(lock); +} + +/* + * 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; + * lockp [output]: the position where the lock should be inserted before, or + * NULL indicating @req should be appended to @queue. + * Return Values: + * Bit-masks combination of following values indicating in which way the + * lock need to be inserted. + * - LDLM_JOIN_NONE: noting about skip list needs to be fixed; + * - LDLM_MODE_JOIN_RIGHT: @req needs join right becoming the head of a + * mode group; + * - LDLM_POLICY_JOIN_RIGHT: @req needs join right becoming the head of + * a policy group. + * NOTE: called by + * - ldlm_grant_lock_with_skiplist + */ +static int search_granted_lock(struct list_head *queue, + struct ldlm_lock *req, + struct ldlm_lock **lockp) +{ + struct list_head *tmp, *tmp_tail; + struct ldlm_lock *lock, *mode_head_lock; + int rc = LDLM_JOIN_NONE; + ENTRY; + + list_for_each(tmp, queue) { + lock = list_entry(tmp, struct ldlm_lock, l_res_link); - compat = ldlm_res_compat_table[child->l_resource->lr_type]; - if (compat && compat(child, lock)) { - CDEBUG(D_OTHER, "compat function succeded, next.\n"); + if (lock->l_req_mode != req->l_req_mode) { + if (LDLM_SL_HEAD(&lock->l_sl_mode)) + tmp = &list_entry(lock->l_sl_mode.next, + struct ldlm_lock, + l_sl_mode)->l_res_link; continue; } - if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) { - CDEBUG(D_OTHER, "lock modes are compatible, next.\n"); - continue; + + /* found the same mode group */ + if (lock->l_resource->lr_type == LDLM_PLAIN) { + *lockp = lock; + rc = LDLM_MODE_JOIN_RIGHT; + GOTO(out, rc); + } else if (lock->l_resource->lr_type == LDLM_IBITS) { + tmp_tail = tmp; + if (LDLM_SL_HEAD(&lock->l_sl_mode)) + tmp_tail = &list_entry(lock->l_sl_mode.next, + struct ldlm_lock, + l_sl_mode)->l_res_link; + mode_head_lock = lock; + for (;;) { + if (lock->l_policy_data.l_inodebits.bits == + req->l_policy_data.l_inodebits.bits) { + /* matched policy lock is found */ + *lockp = lock; + rc |= LDLM_POLICY_JOIN_RIGHT; + + /* if the policy group head is also a + * mode group head or a single mode + * group lock */ + if (LDLM_SL_HEAD(&lock->l_sl_mode) || + (tmp == tmp_tail && + LDLM_SL_EMPTY(&lock->l_sl_mode))) + rc |= LDLM_MODE_JOIN_RIGHT; + GOTO(out, rc); + } + + if (LDLM_SL_HEAD(&lock->l_sl_policy)) + tmp = &list_entry(lock->l_sl_policy.next, + struct ldlm_lock, + l_sl_policy)->l_res_link; + + if (tmp == tmp_tail) + break; + else + tmp = tmp->next; + lock = list_entry(tmp, struct ldlm_lock, + l_res_link); + } /* for all locks in the matched mode group */ + + /* no matched policy group is found, insert before + * the mode group head lock */ + *lockp = mode_head_lock; + rc = LDLM_MODE_JOIN_RIGHT; + GOTO(out, rc); + } else { + LDLM_ERROR(lock, "is not LDLM_PLAIN or LDLM_IBITS lock"); + LBUG(); } + } - rc = 0; + /* no matched mode group is found, append to the end */ + *lockp = NULL; + rc = LDLM_JOIN_NONE; + EXIT; +out: + return rc; +} - if (send_cbs && child->l_blocking_ast != NULL) { - CDEBUG(D_OTHER, "lock %p incompatible; sending " - "blocking AST.\n", child); - ldlm_add_ast_work_item(child, lock); +static void ldlm_granted_list_add_lock(struct ldlm_lock *lock, + struct ldlm_lock *lockp, + int join) +{ + struct ldlm_resource *res = lock->l_resource; + ENTRY; + + LASSERT(lockp || join == LDLM_JOIN_NONE); + + check_res_locked(res); + + ldlm_resource_dump(D_OTHER, 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(list_empty(&lock->l_res_link)); + + if (!lockp) + list_add_tail(&lock->l_res_link, &lock->l_resource->lr_granted); + else if ((join & LDLM_MODE_JOIN_LEFT) || (join & LDLM_POLICY_JOIN_LEFT)) + list_add(&lock->l_res_link, &lockp->l_res_link); + else + list_add_tail(&lock->l_res_link, &lockp->l_res_link); + + /* fix skip lists */ + if (join & LDLM_MODE_JOIN_RIGHT) { + LASSERT(! LDLM_SL_TAIL(&lockp->l_sl_mode)); + if (LDLM_SL_EMPTY(&lockp->l_sl_mode)) { + lock->l_sl_mode.next = &lockp->l_sl_mode; + lockp->l_sl_mode.prev = &lock->l_sl_mode; + } else if (LDLM_SL_HEAD(&lockp->l_sl_mode)) { + lock->l_sl_mode.next = lockp->l_sl_mode.next; + lockp->l_sl_mode.next = NULL; + lock->l_sl_mode.next->prev = &lock->l_sl_mode; + } + } else if (join & LDLM_MODE_JOIN_LEFT) { + LASSERT(! LDLM_SL_HEAD(&lockp->l_sl_mode)); + if (LDLM_SL_EMPTY(&lockp->l_sl_mode)) { + lock->l_sl_mode.prev = &lockp->l_sl_mode; + lockp->l_sl_mode.next = &lock->l_sl_mode; + } else if (LDLM_SL_TAIL(&lockp->l_sl_mode)) { + lock->l_sl_mode.prev = lockp->l_sl_mode.prev; + lockp->l_sl_mode.prev = NULL; + lock->l_sl_mode.prev->next = &lock->l_sl_mode; + } + } + + if (join & LDLM_POLICY_JOIN_RIGHT) { + LASSERT(! LDLM_SL_TAIL(&lockp->l_sl_policy)); + if (LDLM_SL_EMPTY(&lockp->l_sl_policy)) { + lock->l_sl_policy.next = &lockp->l_sl_policy; + lockp->l_sl_policy.prev = &lock->l_sl_policy; + } else if (LDLM_SL_HEAD(&lockp->l_sl_policy)) { + lock->l_sl_policy.next = lockp->l_sl_policy.next; + lockp->l_sl_policy.next = NULL; + lock->l_sl_policy.next->prev = &lock->l_sl_policy; + } + } else if (join & LDLM_POLICY_JOIN_LEFT) { + LASSERT(! LDLM_SL_HEAD(&lockp->l_sl_policy)); + if (LDLM_SL_EMPTY(&lockp->l_sl_policy)) { + lock->l_sl_policy.prev = &lockp->l_sl_policy; + lockp->l_sl_policy.next = &lock->l_sl_policy; + } else if (LDLM_SL_TAIL(&lockp->l_sl_policy)) { + lock->l_sl_policy.prev = lockp->l_sl_policy.prev; + lockp->l_sl_policy.prev = NULL; + lock->l_sl_policy.prev->next = &lock->l_sl_policy; } } - return rc; + EXIT; } -static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs) +static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock) { - int rc; + int join = LDLM_JOIN_NONE; + struct ldlm_lock *lockp = NULL; ENTRY; - l_lock(&lock->l_resource->lr_namespace->ns_lock); - rc = ldlm_lock_compat_list(lock, send_cbs, - &lock->l_resource->lr_granted); - /* FIXME: should we be sending ASTs to converting? */ - if (rc) - rc = ldlm_lock_compat_list - (lock, send_cbs, &lock->l_resource->lr_converting); + LASSERT(lock->l_req_mode == lock->l_granted_mode); - l_unlock(&lock->l_resource->lr_namespace->ns_lock); - RETURN(rc); + join = search_granted_lock(&lock->l_resource->lr_granted, lock, &lockp); + ldlm_granted_list_add_lock(lock, lockp, join); + EXIT; } /* NOTE: called by - - ldlm_handle_enqueuque - resource -*/ -void ldlm_grant_lock(struct ldlm_lock *lock) + * - ldlm_lock_enqueue + * - ldlm_reprocess_queue + * - ldlm_lock_convert + * + * must be called with lr_lock held + */ +void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list) { struct ldlm_resource *res = lock->l_resource; ENTRY; - l_lock(&lock->l_resource->lr_namespace->ns_lock); - ldlm_resource_add_lock(res, &res->lr_granted, lock); + check_res_locked(res); + lock->l_granted_mode = lock->l_req_mode; + if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) + ldlm_grant_lock_with_skiplist(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 (lock->l_completion_ast) { - ldlm_add_ast_work_item(lock, NULL); - } - l_unlock(&lock->l_resource->lr_namespace->ns_lock); + if (work_list && lock->l_completion_ast != NULL) + ldlm_add_ast_work_item(lock, NULL, work_list); + + ldlm_pool_add(&res->lr_namespace->ns_pool, lock); EXIT; } -/* returns a referenced lock or NULL */ -static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode, - struct ldlm_extent *extent) +/* 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, + ldlm_policy_data_t *policy, + struct ldlm_lock *old_lock, int flags) { struct ldlm_lock *lock; struct list_head *tmp; list_for_each(tmp, queue) { + ldlm_mode_t match; + lock = list_entry(tmp, struct ldlm_lock, l_res_link); - if (lock->l_flags & (LDLM_FL_CBPENDING | LDLM_FL_DESTROYED)) + if (lock == old_lock) + break; + + /* 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 + * 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 && + lock->l_readers == 0 && lock->l_writers == 0) continue; - /* lock_convert() takes the resource lock, so we're sure that - * req_mode, lr_type, and l_cookie won't change beneath us */ - 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_extent.start > extent->start || - lock->l_extent.end < extent->end)) + (lock->l_policy_data.l_extent.start > + policy->l_extent.start || + lock->l_policy_data.l_extent.end < policy->l_extent.end)) + continue; + + if (unlikely(match == LCK_GROUP) && + lock->l_resource->lr_type == LDLM_EXTENT && + lock->l_policy_data.l_extent.gid != policy->l_extent.gid) + continue; + + /* We match if we have existing lock with same or wider set + of bits. */ + if (lock->l_resource->lr_type == LDLM_IBITS && + ((lock->l_policy_data.l_inodebits.bits & + policy->l_inodebits.bits) != + policy->l_inodebits.bits)) + continue; + + if (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED)) + continue; + + if ((flags & LDLM_FL_LOCAL_ONLY) && + !(lock->l_flags & LDLM_FL_LOCAL)) continue; - ldlm_lock_addref_internal(lock, mode); + if (flags & LDLM_FL_TEST_LOCK) { + LDLM_LOCK_GET(lock); + ldlm_lock_touch_in_lru(lock); + } else { + ldlm_lock_addref_internal_nolock(lock, match); + } + *mode = match; return lock; } return NULL; } -/* Must be called with no resource or lock locks held. +void ldlm_lock_allow_match(struct ldlm_lock *lock) +{ + lock_res_and_lock(lock); + lock->l_flags |= LDLM_FL_LVB_READY; + cfs_waitq_signal(&lock->l_waitq); + unlock_res_and_lock(lock); +} + +/* Can be called in two ways: + * + * If 'ns' is NULL, then lockh describes an existing lock that we want to look + * for a duplicate of. + * + * Otherwise, all of the fields must be filled in, to match against. + * + * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the + * server (ie, connh is NULL) + * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted + * list will be considered + * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked + * to be canceled can still be matched as long as they still have reader + * or writer refernces + * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock, + * just tell us if we would have matched. * * Returns 1 if it finds an already-existing lock that is compatible; in this * case, lockh is filled in with a addref()ed lock -*/ -int ldlm_lock_match(struct ldlm_namespace *ns, __u64 * res_id, __u32 type, - void *cookie, int cookielen, ldlm_mode_t mode, - struct lustre_handle *lockh) + * + * 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. + */ +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) { struct ldlm_resource *res; - struct ldlm_lock *lock; + struct ldlm_lock *lock, *old_lock = NULL; int rc = 0; ENTRY; + if (ns == NULL) { + old_lock = ldlm_handle2lock(lockh); + LASSERT(old_lock); + + ns = old_lock->l_resource->lr_namespace; + 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); - if (res == NULL) + if (res == NULL) { + LASSERT(old_lock == NULL); RETURN(0); + } - ns = res->lr_namespace; - l_lock(&ns->ns_lock); + lock_res(res); - if ((lock = search_queue(&res->lr_granted, mode, cookie))) + lock = search_queue(&res->lr_granted, &mode, policy, old_lock, flags); + if (lock != NULL) GOTO(out, rc = 1); - if ((lock = search_queue(&res->lr_converting, mode, cookie))) + if (flags & LDLM_FL_BLOCK_GRANTED) + GOTO(out, rc = 0); + lock = search_queue(&res->lr_converting, &mode, policy, old_lock, flags); + if (lock != NULL) GOTO(out, rc = 1); - if ((lock = search_queue(&res->lr_waiting, mode, cookie))) + lock = search_queue(&res->lr_waiting, &mode, policy, old_lock, flags); + if (lock != NULL) GOTO(out, rc = 1); EXIT; - out: - ldlm_resource_put(res); - l_unlock(&ns->ns_lock); + out: + unlock_res(res); + ldlm_resource_putref(res); if (lock) { ldlm_lock2handle(lock, lockh); - if (lock->l_completion_ast) - lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC); + if ((flags & LDLM_FL_LVB_READY) && + (!(lock->l_flags & LDLM_FL_LVB_READY))) { + struct l_wait_info lwi; + 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_PUT(lock); + else + ldlm_lock_decref_internal(lock, mode); + rc = 0; + goto out2; + } + } + + 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_LVB_READY), &lwi); + } } - if (rc) - LDLM_DEBUG(lock, "matched"); - else - LDLM_DEBUG_NOLOCK("not matched"); - return rc; + out2: + if (rc) { + LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")", + (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_PUT(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 || 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); + + return rc ? mode : 0; } /* Returns a referenced lock */ struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns, - struct lustre_handle *parent_lock_handle, - __u64 * res_id, __u32 type, - ldlm_mode_t mode, void *data, __u32 data_len) + 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, + void *data, __u32 lvb_len) { - struct ldlm_resource *res, *parent_res = NULL; - struct ldlm_lock *lock, *parent_lock; - - parent_lock = ldlm_handle2lock(parent_lock_handle); - if (parent_lock) - parent_res = parent_lock->l_resource; + struct ldlm_lock *lock; + struct ldlm_resource *res; + ENTRY; - 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); - if (lock == NULL) { - ldlm_resource_put(res); + lock = ldlm_lock_new(res); + ldlm_resource_putref(res); + + if (lock == NULL) RETURN(NULL); - } lock->l_req_mode = mode; - lock->l_data = data; - lock->l_data_len = data_len; + 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(); + + 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); + } + } - return lock; + RETURN(lock); } -/* Must be called with lock->l_lock and lock->l_resource->lr_lock not held */ -ldlm_error_t ldlm_lock_enqueue(struct ldlm_lock * lock, - void *cookie, int cookie_len, - int *flags, - ldlm_completion_callback completion, - ldlm_blocking_callback blocking) +ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns, + struct ldlm_lock **lockp, + void *cookie, int *flags) { - struct ldlm_resource *res; - int local; - ldlm_res_policy policy; + struct ldlm_lock *lock = *lockp; + struct ldlm_resource *res = lock->l_resource; + int local = ns_is_client(res->lr_namespace); + ldlm_processing_policy policy; + ldlm_error_t rc = ELDLM_OK; ENTRY; - res = lock->l_resource; - lock->l_blocking_ast = blocking; - - if (res->lr_type == LDLM_EXTENT) - memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent)); - - /* policies are not executed on the client */ - local = res->lr_namespace->ns_client; - if (!local && (policy = ldlm_res_policy_table[res->lr_type])) { - int rc; - rc = policy(lock, cookie, lock->l_req_mode, NULL); - - if (rc == ELDLM_LOCK_CHANGED) { - res = lock->l_resource; + do_gettimeofday(&lock->l_enqueued_time); + /* 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) { + rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags, + NULL); + if (rc == ELDLM_LOCK_REPLACED) { + /* The lock that was returned has already been granted, + * and placed into lockp. If it's not the same as the + * one we passed in, then destroy the old one and our + * work here is done. */ + if (lock != *lockp) { + ldlm_lock_destroy(lock); + LDLM_LOCK_PUT(lock); + } *flags |= LDLM_FL_LOCK_CHANGED; - } else if (rc == ELDLM_LOCK_ABORTED) { + RETURN(0); + } else if (rc != ELDLM_OK || + (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) { ldlm_lock_destroy(lock); RETURN(rc); } } - lock->l_cookie = cookie; - lock->l_cookie_len = cookie_len; - - l_lock(&res->lr_namespace->ns_lock); + 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. */ - *flags &= ~(LDLM_FL_BLOCK_GRANTED | - LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT); + *flags &= ~(LDLM_FL_BLOCK_GRANTED | + LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT); GOTO(out, ELDLM_OK); } + /* 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; + /* This distinction between local lock trees is very important; a client * namespace only has information about locks taken by that client, and * thus doesn't have enough information to decide for itself if it can * be granted (below). In this case, we do exactly what the server - * tells us to do, as dictated by the 'flags' */ + * tells us to do, as dictated by the 'flags'. + * + * We do exactly the same thing during recovery, when the server is + * more or less trusting the clients not to lie. + * + * 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.prev, - lock); + 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.prev, lock); + ldlm_resource_add_lock(res, &res->lr_waiting, lock); else - ldlm_grant_lock(lock); - GOTO(out, ELDLM_OK); - } - - /* FIXME: We may want to optimize by checking lr_most_restr */ - if (!list_empty(&res->lr_converting)) { - ldlm_resource_add_lock(res, res->lr_waiting.prev, lock); - *flags |= LDLM_FL_BLOCK_CONV; - GOTO(out, ELDLM_OK); - } - if (!list_empty(&res->lr_waiting)) { - ldlm_resource_add_lock(res, res->lr_waiting.prev, lock); - *flags |= LDLM_FL_BLOCK_WAIT; - GOTO(out, ELDLM_OK); - } - if (!ldlm_lock_compat(lock, 0)) { - ldlm_resource_add_lock(res, res->lr_waiting.prev, lock); - *flags |= LDLM_FL_BLOCK_GRANTED; + ldlm_grant_lock(lock, NULL); GOTO(out, ELDLM_OK); + } else if (*flags & LDLM_FL_REPLAY) { + if (*flags & LDLM_FL_BLOCK_CONV) { + ldlm_resource_add_lock(res, &res->lr_converting, lock); + GOTO(out, ELDLM_OK); + } else if (*flags & LDLM_FL_BLOCK_WAIT) { + 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); + GOTO(out, ELDLM_OK); + } + /* If no flags, fall through to normal enqueue path. */ } - ldlm_grant_lock(lock); - EXIT; - out: - l_unlock(&res->lr_namespace->ns_lock); - /* Don't set 'completion_ast' until here so that if the lock is granted - * immediately we don't do an unnecessary completion call. */ - lock->l_completion_ast = completion; - return ELDLM_OK; + policy = ldlm_processing_policy_table[res->lr_type]; + policy(lock, flags, 1, &rc, NULL); + GOTO(out, rc); +out: + unlock_res_and_lock(lock); + return rc; } /* Must be called with namespace taken: queue is waiting or converting. */ -static int ldlm_reprocess_queue(struct ldlm_resource *res, - struct list_head *queue) +int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue, + struct list_head *work_list) { struct list_head *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) { struct ldlm_lock *pending; pending = list_entry(tmp, struct ldlm_lock, l_res_link); CDEBUG(D_INFO, "Reprocessing lock %p\n", pending); - if (!ldlm_lock_compat(pending, 1)) - RETURN(1); - - list_del_init(&pending->l_res_link); - ldlm_grant_lock(pending); + flags = 0; + rc = policy(pending, &flags, 0, &err, work_list); + if (rc != LDLM_ITER_CONTINUE) + break; } - RETURN(0); + RETURN(rc); } -void ldlm_run_ast_work(struct list_head *rpc_list) +int ldlm_run_bl_ast_work(struct list_head *rpc_list) { struct list_head *tmp, *pos; - int rc; + struct ldlm_lock_desc d; + int rc = 0, retval = 0; ENTRY; list_for_each_safe(tmp, pos, rpc_list) { - struct ldlm_ast_work *w = - list_entry(tmp, struct ldlm_ast_work, w_list); + struct ldlm_lock *lock = + list_entry(tmp, struct ldlm_lock, l_bl_ast); + + /* nobody should touch l_bl_ast */ + lock_res_and_lock(lock); + 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); + + LDLM_LOCK_PUT(lock->l_blocking_lock); + lock->l_blocking_lock = NULL; + rc = lock->l_blocking_ast(lock, &d, NULL, LDLM_CB_BLOCKING); + + if (rc == -ERESTART) + retval = rc; + else if (rc) + CDEBUG(D_DLMTRACE, "Failed AST - should clean & " + "disconnect client\n"); + LDLM_LOCK_PUT(lock); + } + RETURN(retval); +} - if (w->w_blocking) - rc = w->w_lock->l_blocking_ast - (w->w_lock, &w->w_desc, w->w_data, - w->w_datalen, LDLM_CB_BLOCKING); - else - rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags); - if (rc) - CERROR("Failed AST - should clean & disconnect " - "client\n"); - LDLM_LOCK_PUT(w->w_lock); - list_del(&w->w_list); - OBD_FREE(w, sizeof(*w)); +int ldlm_run_cp_ast_work(struct list_head *rpc_list) +{ + struct list_head *tmp, *pos; + int rc = 0, retval = 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. */ + + list_for_each_safe(tmp, pos, rpc_list) { + struct ldlm_lock *lock = + list_entry(tmp, struct ldlm_lock, l_cp_ast); + + /* nobody should touch l_cp_ast */ + lock_res_and_lock(lock); + list_del_init(&lock->l_cp_ast); + LASSERT(lock->l_flags & LDLM_FL_CP_REQD); + lock->l_flags &= ~LDLM_FL_CP_REQD; + unlock_res_and_lock(lock); + + if (lock->l_completion_ast != NULL) + rc = lock->l_completion_ast(lock, 0, 0); + if (rc == -ERESTART) + retval = rc; + else if (rc) + CDEBUG(D_DLMTRACE, "Failed AST - should clean & " + "disconnect client\n"); + LDLM_LOCK_PUT(lock); } + RETURN(retval); +} + +static int reprocess_one_queue(struct ldlm_resource *res, void *closure) +{ + ldlm_reprocess_all(res); + return LDLM_ITER_CONTINUE; +} + +void ldlm_reprocess_all_ns(struct ldlm_namespace *ns) +{ + struct list_head *tmp; + int i, rc; + + if (ns == NULL) + return; + + ENTRY; + spin_lock(&ns->ns_hash_lock); + for (i = 0; i < RES_HASH_SIZE; i++) { + tmp = ns->ns_hash[i].next; + while (tmp != &(ns->ns_hash[i])) { + struct ldlm_resource *res = + list_entry(tmp, struct ldlm_resource, lr_hash); + + ldlm_resource_getref(res); + spin_unlock(&ns->ns_hash_lock); + + rc = reprocess_one_queue(res, NULL); + + spin_lock(&ns->ns_hash_lock); + tmp = tmp->next; + ldlm_resource_putref_locked(res); + + if (rc == LDLM_ITER_STOP) + GOTO(out, rc); + } + } + out: + spin_unlock(&ns->ns_hash_lock); EXIT; } -/* Must be called with resource->lr_lock not taken. */ void ldlm_reprocess_all(struct ldlm_resource *res) { struct list_head rpc_list = LIST_HEAD_INIT(rpc_list); + int rc; ENTRY; /* Local lock trees don't get reprocessed. */ - if (res->lr_namespace->ns_client) { + if (ns_is_client(res->lr_namespace)) { EXIT; return; } - l_lock(&res->lr_namespace->ns_lock); - res->lr_tmp = &rpc_list; - - ldlm_reprocess_queue(res, &res->lr_converting); - if (list_empty(&res->lr_converting)) - ldlm_reprocess_queue(res, &res->lr_waiting); - - res->lr_tmp = NULL; - l_unlock(&res->lr_namespace->ns_lock); - - ldlm_run_ast_work(&rpc_list); + restart: + 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, &rpc_list); + unlock_res(res); + + rc = ldlm_run_cp_ast_work(&rpc_list); + if (rc == -ERESTART) { + LASSERT(list_empty(&rpc_list)); + goto restart; + } EXIT; } 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; - lock->l_blocking_ast(lock, NULL, lock->l_data, - lock->l_data_len, LDLM_CB_CANCELING); + if (lock->l_blocking_ast) { + // l_check_no_ns_lock(ns); + unlock_res_and_lock(lock); + lock->l_blocking_ast(lock, NULL, lock->l_ast_data, + LDLM_CB_CANCELING); + lock_res_and_lock(lock); + } else { + LDLM_DEBUG(lock, "no blocking ast"); + } + } + lock->l_flags |= LDLM_FL_BL_DONE; +} + +void ldlm_unlink_lock_skiplist(struct ldlm_lock *req) +{ + struct ldlm_lock *lock; + + if (req->l_resource->lr_type != LDLM_PLAIN && + req->l_resource->lr_type != LDLM_IBITS) + return; + + if (LDLM_SL_HEAD(&req->l_sl_mode)) { + lock = list_entry(req->l_res_link.next, struct ldlm_lock, + l_res_link); + if (req->l_sl_mode.next == &lock->l_sl_mode) { + lock->l_sl_mode.prev = NULL; + } else { + lock->l_sl_mode.next = req->l_sl_mode.next; + lock->l_sl_mode.next->prev = &lock->l_sl_mode; + } + req->l_sl_mode.next = NULL; + } else if (LDLM_SL_TAIL(&req->l_sl_mode)) { + lock = list_entry(req->l_res_link.prev, struct ldlm_lock, + l_res_link); + if (req->l_sl_mode.prev == &lock->l_sl_mode) { + lock->l_sl_mode.next = NULL; + } else { + lock->l_sl_mode.prev = req->l_sl_mode.prev; + lock->l_sl_mode.prev->next = &lock->l_sl_mode; + } + req->l_sl_mode.prev = NULL; + } + + if (LDLM_SL_HEAD(&req->l_sl_policy)) { + lock = list_entry(req->l_res_link.next, struct ldlm_lock, + l_res_link); + if (req->l_sl_policy.next == &lock->l_sl_policy) { + lock->l_sl_policy.prev = NULL; + } else { + lock->l_sl_policy.next = req->l_sl_policy.next; + lock->l_sl_policy.next->prev = &lock->l_sl_policy; + } + req->l_sl_policy.next = NULL; + } else if (LDLM_SL_TAIL(&req->l_sl_policy)) { + lock = list_entry(req->l_res_link.prev, struct ldlm_lock, + l_res_link); + if (req->l_sl_policy.prev == &lock->l_sl_policy) { + lock->l_sl_policy.next = NULL; + } else { + lock->l_sl_policy.prev = req->l_sl_policy.prev; + lock->l_sl_policy.prev->next = &lock->l_sl_policy; + } + req->l_sl_policy.prev = NULL; } - l_unlock(&lock->l_resource->lr_namespace->ns_lock); } void ldlm_lock_cancel(struct ldlm_lock *lock) @@ -837,37 +1489,77 @@ void ldlm_lock_cancel(struct ldlm_lock *lock) struct ldlm_namespace *ns; ENTRY; + lock_res_and_lock(lock); + res = lock->l_resource; ns = res->lr_namespace; - l_lock(&ns->ns_lock); - if (lock->l_readers || lock->l_writers) - CDEBUG(D_INFO, "lock still has references (%d readers, %d " - "writers)\n", lock->l_readers, lock->l_writers); + /* Please do not, no matter how tempting, remove this LBUG without + * talking to me first. -phik */ + if (lock->l_readers || lock->l_writers) { + LDLM_ERROR(lock, "lock still has references"); + LBUG(); + } + ldlm_del_waiting_lock(lock); + + /* Releases cancel callback. */ ldlm_cancel_callback(lock); - ldlm_del_waiting_lock(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 = 0; + unlock_res_and_lock(lock); + EXIT; } +int ldlm_lock_set_data(struct lustre_handle *lockh, void *data) +{ + struct ldlm_lock *lock = ldlm_handle2lock(lockh); + ENTRY; + + if (lock == NULL) + RETURN(-EINVAL); + + lock->l_ast_data = data; + LDLM_LOCK_PUT(lock); + RETURN(0); +} + void ldlm_cancel_locks_for_export(struct obd_export *exp) { - struct list_head *iter, *n; /* MUST BE CALLED "n"! */ + struct ldlm_lock *lock; + struct ldlm_resource *res; - list_for_each_safe(iter, n, &exp->exp_ldlm_data.led_held_locks) { - struct ldlm_lock *lock; - struct ldlm_resource *res; - lock = list_entry(iter, struct ldlm_lock, l_export_chain); + spin_lock(&exp->exp_ldlm_data.led_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_LOCK_GET(lock); + spin_unlock(&exp->exp_ldlm_data.led_lock); + LDLM_DEBUG(lock, "export %p", exp); + ldlm_res_lvbo_update(res, NULL, 0, 1); + ldlm_lock_cancel(lock); ldlm_reprocess_all(res); - ldlm_resource_put(res); + + ldlm_resource_putref(res); + LDLM_LOCK_PUT(lock); + spin_lock(&exp->exp_ldlm_data.led_lock); } + spin_unlock(&exp->exp_ldlm_data.led_lock); } struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode, @@ -877,83 +1569,288 @@ struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode, struct ldlm_resource *res; struct ldlm_namespace *ns; int granted = 0; + int old_mode, rc; + struct ldlm_lock *mark_lock = NULL; + int join = LDLM_JOIN_NONE; + ldlm_error_t err; ENTRY; + if (new_mode == lock->l_granted_mode) { // No changes? Just return. + *flags |= LDLM_FL_BLOCK_GRANTED; + RETURN(lock->l_resource); + } + + 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); + res = lock->l_resource; ns = res->lr_namespace; - l_lock(&ns->ns_lock); - + old_mode = lock->l_req_mode; lock->l_req_mode = new_mode; + 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. */ + if (LDLM_SL_HEAD(&lock->l_sl_mode)) + join = LDLM_MODE_JOIN_RIGHT; + else if (LDLM_SL_TAIL(&lock->l_sl_mode)) + join = LDLM_MODE_JOIN_LEFT; + if (LDLM_SL_HEAD(&lock->l_sl_policy)) + join |= LDLM_POLICY_JOIN_RIGHT; + else if (LDLM_SL_TAIL(&lock->l_sl_policy)) + join |= LDLM_POLICY_JOIN_LEFT; + + LASSERT(!((join & LDLM_MODE_JOIN_RIGHT) && + (join & LDLM_POLICY_JOIN_LEFT))); + LASSERT(!((join & LDLM_MODE_JOIN_LEFT) && + (join & LDLM_POLICY_JOIN_RIGHT))); + + if ((join & LDLM_MODE_JOIN_LEFT) || + (join & LDLM_POLICY_JOIN_LEFT)) + mark_lock = list_entry(lock->l_res_link.prev, + struct ldlm_lock, l_res_link); + else if (lock->l_res_link.next != &res->lr_granted) + mark_lock = list_entry(lock->l_res_link.next, + struct ldlm_lock, l_res_link); + } ldlm_resource_unlink_lock(lock); /* If this is a local resource, put it on the appropriate list. */ - if (res->lr_namespace->ns_client) { - if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) - ldlm_resource_add_lock(res, res->lr_converting.prev, - lock); - else { + if (ns_is_client(res->lr_namespace)) { + 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); - res->lr_tmp = NULL; + ldlm_grant_lock(lock, &rpc_list); granted = 1; /* FIXME: completion handling not with ns_lock held ! */ if (lock->l_completion_ast) - lock->l_completion_ast(lock, 0); + 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.prev, 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; + ldlm_granted_list_add_lock(lock, mark_lock, join); + 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(&rpc_list); + ldlm_run_cp_ast_work(&rpc_list); RETURN(res); } -void ldlm_lock_dump(struct ldlm_lock *lock) +void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos) { - char ver[128]; + struct obd_device *obd = NULL; - if (!(portal_debug & D_OTHER)) + if (!((libcfs_debug | D_ERROR) & level)) return; - if (RES_VERSION_SIZE != 4) - LBUG(); - if (!lock) { - CDEBUG(D_OTHER, " NULL LDLM lock\n"); + CDEBUG(level, " NULL LDLM lock\n"); return; } - snprintf(ver, sizeof(ver), "%x %x %x %x", - lock->l_version[0], lock->l_version[1], - lock->l_version[2], lock->l_version[3]); - - CDEBUG(D_OTHER, " -- Lock dump: %p (%s)\n", lock, ver); - if (lock->l_export && lock->l_export->exp_connection) - CDEBUG(D_OTHER, " Node: NID %x (rhandle: "LPX64")\n", - lock->l_export->exp_connection->c_peer.peer_nid, - lock->l_remote_handle.addr); - else - CDEBUG(D_OTHER, " Node: local\n"); - CDEBUG(D_OTHER, " Parent: %p\n", lock->l_parent); - CDEBUG(D_OTHER, " Resource: %p ("LPD64")\n", lock->l_resource, - lock->l_resource->lr_name[0]); - CDEBUG(D_OTHER, " Requested mode: %d, granted mode: %d\n", - (int)lock->l_req_mode, (int)lock->l_granted_mode); - CDEBUG(D_OTHER, " Readers: %u ; Writers; %u\n", - lock->l_readers, lock->l_writers); + CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n", + lock, lock->l_handle.h_cookie, 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 (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 (rhandle: "LPX64")\n", + libcfs_nid2str(imp->imp_connection->c_peer.nid), + lock->l_remote_handle.cookie); + } + 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 flags: %#x\n", ldlm_lockname[lock->l_req_mode], + ldlm_lockname[lock->l_granted_mode], + atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers, + lock->l_flags); if (lock->l_resource->lr_type == LDLM_EXTENT) - CDEBUG(D_OTHER, " Extent: %Lu -> %Lu\n", - (unsigned long long)lock->l_extent.start, - (unsigned long long)lock->l_extent.end); + CDEBUG(level, " Extent: "LPU64" -> "LPU64 + " (req "LPU64"-"LPU64")\n", + lock->l_policy_data.l_extent.start, + 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: %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", + lock->l_policy_data.l_inodebits.bits); +} + +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; + + ldlm_lock_dump(D_OTHER, lock, 0); + + 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: %x remote: " + LPX64" expref: %d pid: %u\n", lock, + lock->l_handle.h_cookie, 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 ? + atomic_read(&lock->l_export->exp_refcount) : -99, + lock->l_pid); + 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: %x remote: "LPX64 + " expref: %d pid: %u\n", + lock->l_resource->lr_namespace->ns_name, lock, + lock->l_handle.h_cookie, 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], + 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 ? + atomic_read(&lock->l_export->exp_refcount) : -99, + lock->l_pid); + 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: %x remote: "LPX64 + " expref: %d pid: %u\n", + lock->l_resource->lr_namespace->ns_name, lock, + lock->l_handle.h_cookie, 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], + 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 ? + atomic_read(&lock->l_export->exp_refcount) : -99, + lock->l_pid); + 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: %x remote: "LPX64" expref: %d " + "pid %u\n", + lock->l_resource->lr_namespace->ns_name, + lock, lock->l_handle.h_cookie, + 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, + 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 ? + atomic_read(&lock->l_export->exp_refcount) : -99, + lock->l_pid); + 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: %x " + "remote: "LPX64" expref: %d pid: %u\n", + lock->l_resource->lr_namespace->ns_name, + lock, lock->l_handle.h_cookie, + 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], + 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 ? + atomic_read(&lock->l_export->exp_refcount) : -99, + lock->l_pid); + break; + } + va_end(args); } +EXPORT_SYMBOL(_ldlm_lock_debug);