X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_resource.c;h=b86c66d216d031418365ce7296ad1095301113d3;hp=eb9cb22e9a31a9d224ec350a8123fd889c72753b;hb=a1052417b78bc18898161c4ed44d79de4a1a2f23;hpb=9cc7e09f24b6128ab871ff91e35cc8f6a46955e1 diff --git a/lustre/ldlm/ldlm_resource.c b/lustre/ldlm/ldlm_resource.c index eb9cb22..b86c66d 100644 --- a/lustre/ldlm/ldlm_resource.c +++ b/lustre/ldlm/ldlm_resource.c @@ -1,6 +1,4 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: - * +/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. @@ -28,6 +26,8 @@ /* * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. + * + * Copyright (c) 2010, 2013, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -50,16 +50,21 @@ #include #include "ldlm_internal.h" -cfs_mem_cache_t *ldlm_resource_slab, *ldlm_lock_slab; +struct kmem_cache *ldlm_resource_slab, *ldlm_lock_slab; -cfs_atomic_t ldlm_srv_namespace_nr = CFS_ATOMIC_INIT(0); -cfs_atomic_t ldlm_cli_namespace_nr = CFS_ATOMIC_INIT(0); +int ldlm_srv_namespace_nr = 0; +int ldlm_cli_namespace_nr = 0; -cfs_semaphore_t ldlm_srv_namespace_lock; +struct mutex ldlm_srv_namespace_lock; CFS_LIST_HEAD(ldlm_srv_namespace_list); -cfs_semaphore_t ldlm_cli_namespace_lock; -CFS_LIST_HEAD(ldlm_cli_namespace_list); +struct mutex ldlm_cli_namespace_lock; +/* Client Namespaces that have active resources in them. + * Once all resources go away, ldlm_poold moves such namespaces to the + * inactive list */ +CFS_LIST_HEAD(ldlm_cli_active_namespace_list); +/* Client namespaces that don't have any locks in them */ +CFS_LIST_HEAD(ldlm_cli_inactive_namespace_list); cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL; cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL; @@ -67,6 +72,10 @@ cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL; extern unsigned int ldlm_cancel_unused_locks_before_replay; +/* during debug dump certain amount of granted locks for one resource to avoid + * DDOS. */ +unsigned int ldlm_dump_granted_max = 256; + #ifdef LPROCFS static int ldlm_proc_dump_ns(struct file *file, const char *buffer, unsigned long count, void *data) @@ -81,6 +90,9 @@ int ldlm_proc_setup(void) int rc; struct lprocfs_vars list[] = { { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL }, + { "dump_granted_max", + lprocfs_rd_uint, lprocfs_wr_uint, + &ldlm_dump_granted_max, NULL }, { "cancel_unused_locks_before_replay", lprocfs_rd_uint, lprocfs_wr_uint, &ldlm_cancel_unused_locks_before_replay, NULL }, @@ -185,7 +197,7 @@ static int lprocfs_wr_lru_size(struct file *file, const char *buffer, int lru_resize; dummy[MAX_STRING_SIZE] = '\0'; - if (cfs_copy_from_user(dummy, buffer, MAX_STRING_SIZE)) + if (copy_from_user(dummy, buffer, MAX_STRING_SIZE)) return -EFAULT; if (strncmp(dummy, "clear", 5) == 0) { @@ -196,8 +208,8 @@ static int lprocfs_wr_lru_size(struct file *file, const char *buffer, int canceled, unused = ns->ns_nr_unused; /* Try to cancel all @ns_nr_unused locks. */ - canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC, - LDLM_CANCEL_PASSED); + canceled = ldlm_cancel_lru(ns, unused, 0, + LDLM_CANCEL_PASSED); if (canceled < unused) { CDEBUG(D_DLMTRACE, "not all requested locks are canceled, " @@ -208,7 +220,7 @@ static int lprocfs_wr_lru_size(struct file *file, const char *buffer, } else { tmp = ns->ns_max_unused; ns->ns_max_unused = 0; - ldlm_cancel_lru(ns, 0, LDLM_SYNC, LDLM_CANCEL_PASSED); + ldlm_cancel_lru(ns, 0, 0, LDLM_CANCEL_PASSED); ns->ns_max_unused = tmp; } return count; @@ -233,7 +245,7 @@ static int lprocfs_wr_lru_size(struct file *file, const char *buffer, "changing namespace %s unused locks from %u to %u\n", ldlm_ns_name(ns), ns->ns_nr_unused, (unsigned int)tmp); - ldlm_cancel_lru(ns, tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED); + ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_CANCEL_PASSED); if (!lru_resize) { CDEBUG(D_DLMTRACE, @@ -247,10 +259,10 @@ static int lprocfs_wr_lru_size(struct file *file, const char *buffer, ldlm_ns_name(ns), ns->ns_max_unused, (unsigned int)tmp); ns->ns_max_unused = (unsigned int)tmp; - ldlm_cancel_lru(ns, 0, LDLM_ASYNC, LDLM_CANCEL_PASSED); + ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_CANCEL_PASSED); - /* Make sure that originally lru resize was supported before - * turning it on here. */ + /* Make sure that LRU resize was originally supported before + * turning it on here. */ if (lru_resize && (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) { CDEBUG(D_DLMTRACE, @@ -263,6 +275,33 @@ static int lprocfs_wr_lru_size(struct file *file, const char *buffer, return count; } +static int lprocfs_rd_elc(char *page, char **start, off_t off, + int count, int *eof, void *data) +{ + struct ldlm_namespace *ns = data; + unsigned int supp = ns_connect_cancelset(ns); + + return lprocfs_rd_uint(page, start, off, count, eof, &supp); +} + +static int lprocfs_wr_elc(struct file *file, const char *buffer, + unsigned long count, void *data) +{ + struct ldlm_namespace *ns = data; + unsigned int supp = -1; + int rc; + + rc = lprocfs_wr_uint(file, buffer, count, &supp); + if (rc < 0) + return rc; + + if (supp == 0) + ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET; + else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET) + ns->ns_connect_flags |= OBD_CONNECT_CANCELSET; + return count; +} + void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns) { struct proc_dir_entry *dir; @@ -331,6 +370,13 @@ int ldlm_namespace_proc_register(struct ldlm_namespace *ns) lock_vars[0].read_fptr = lprocfs_rd_uint; lock_vars[0].write_fptr = lprocfs_wr_uint; lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0); + + snprintf(lock_name, MAX_STRING_SIZE, "%s/early_lock_cancel", + ldlm_ns_name(ns)); + lock_vars[0].data = ns; + lock_vars[0].read_fptr = lprocfs_rd_elc; + lock_vars[0].write_fptr = lprocfs_wr_elc; + lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0); } else { snprintf(lock_name, MAX_STRING_SIZE, "%s/ctime_age_limit", ldlm_ns_name(ns)); @@ -400,19 +446,25 @@ static unsigned ldlm_res_hop_fid_hash(cfs_hash_t *hs, { const struct ldlm_res_id *id = key; struct lu_fid fid; - __u64 hash; + __u32 hash; + __u32 val; fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF]; - fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_OID_OFF]; - fid.f_ver = (__u32)id->name[LUSTRE_RES_ID_VER_OFF]; - - hash = fid_flatten(&fid); + fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF]; + fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32); + + hash = fid_flatten32(&fid); + hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */ + if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) { + val = id->name[LUSTRE_RES_ID_HSH_OFF]; + hash += (val >> 5) + (val << 11); + } else { + val = fid_oid(&fid); + } hash = cfs_hash_long(hash, hs->hs_bkt_bits); - /* ignore a few low bits */ - if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) - hash += id->name[LUSTRE_RES_ID_HSH_OFF] >> 5; - else - hash = hash >> 5; + /* give me another random factor */ + hash -= cfs_hash_long((unsigned long)hs, val % 11 + 3); + hash <<= hs->hs_cur_bits - hs->hs_bkt_bits; hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1); @@ -427,12 +479,6 @@ static void *ldlm_res_hop_key(cfs_hlist_node_t *hnode) return &res->lr_name; } -static int ldlm_res_eq(const struct ldlm_res_id *res0, - const struct ldlm_res_id *res1) -{ - return !memcmp(res0, res1, sizeof(*res0)); -} - static int ldlm_res_hop_keycmp(const void *key, cfs_hlist_node_t *hnode) { struct ldlm_resource *res; @@ -453,7 +499,6 @@ static void ldlm_res_hop_get_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode) res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash); ldlm_resource_getref(res); - LDLM_RESOURCE_ADDREF(res); } static void ldlm_res_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode) @@ -462,7 +507,6 @@ static void ldlm_res_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode) res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash); /* cfs_hash_for_each_nolock is the only chance we call it */ - LDLM_RESOURCE_DELREF(res); ldlm_resource_putref_locked(res); } @@ -471,7 +515,6 @@ static void ldlm_res_hop_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode) struct ldlm_resource *res; res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash); - LDLM_RESOURCE_DELREF(res); ldlm_resource_putref(res); } @@ -512,7 +555,7 @@ ldlm_ns_hash_def_t ldlm_ns_hash_defs[] = { .nsd_type = LDLM_NS_TYPE_MDC, .nsd_bkt_bits = 11, - .nsd_all_bits = 15, + .nsd_all_bits = 16, .nsd_hops = &ldlm_ns_fid_hash_ops, }, { @@ -550,6 +593,9 @@ ldlm_ns_hash_def_t ldlm_ns_hash_defs[] = }, }; +/** + * Create and initialize new empty namespace. + */ struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name, ldlm_side_t client, ldlm_appetite_t apt, @@ -611,7 +657,7 @@ struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name, CFS_INIT_LIST_HEAD(&ns->ns_list_chain); CFS_INIT_LIST_HEAD(&ns->ns_unused_list); - cfs_spin_lock_init(&ns->ns_lock); + spin_lock_init(&ns->ns_lock); cfs_atomic_set(&ns->ns_bref, 0); cfs_waitq_init(&ns->ns_waitq); @@ -619,6 +665,7 @@ struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name, ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS; ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS; + ns->ns_max_parallel_ast = LDLM_DEFAULT_PARALLEL_AST_LIMIT; ns->ns_nr_unused = 0; ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE; ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE; @@ -633,7 +680,7 @@ struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name, GOTO(out_hash, rc); } - idx = cfs_atomic_read(ldlm_namespace_nr(client)); + idx = ldlm_namespace_nr_read(client); rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client); if (rc) { CERROR("Can't initialize lock pool, rc %d\n", rc); @@ -653,24 +700,30 @@ out_ref: ldlm_put_ref(); RETURN(NULL); } +EXPORT_SYMBOL(ldlm_namespace_new); extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock); -/* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup. - * This is currently only used for recovery, and we make certain assumptions - * as a result--notably, that we shouldn't cancel locks with refs. -phil */ +/** + * Cancel and destroy all locks on a resource. + * + * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just + * clean up. This is currently only used for recovery, and we make + * certain assumptions as a result--notably, that we shouldn't cancel + * locks with refs. + */ static void cleanup_resource(struct ldlm_resource *res, cfs_list_t *q, - int flags) + __u64 flags) { - cfs_list_t *tmp; - int rc = 0, client = ns_is_client(ldlm_res_to_ns(res)); - int local_only = (flags & LDLM_FL_LOCAL_ONLY); + cfs_list_t *tmp; + int rc = 0, client = ns_is_client(ldlm_res_to_ns(res)); + bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY); do { struct ldlm_lock *lock = NULL; - /* first, we look for non-cleaned-yet lock - * all cleaned locks are marked by CLEANED flag */ + /* First, we look for non-cleaned-yet lock + * all cleaned locks are marked by CLEANED flag. */ lock_res(res); cfs_list_for_each(tmp, q) { lock = cfs_list_entry(tmp, struct ldlm_lock, @@ -690,7 +743,7 @@ static void cleanup_resource(struct ldlm_resource *res, cfs_list_t *q, } /* Set CBPENDING so nothing in the cancellation path - * can match this lock */ + * can match this lock. */ lock->l_flags |= LDLM_FL_CBPENDING; lock->l_flags |= LDLM_FL_FAILED; lock->l_flags |= flags; @@ -717,7 +770,7 @@ static void cleanup_resource(struct ldlm_resource *res, cfs_list_t *q, unlock_res(res); ldlm_lock2handle(lock, &lockh); - rc = ldlm_cli_cancel(&lockh); + rc = ldlm_cli_cancel(&lockh, LCF_ASYNC); if (rc) CERROR("ldlm_cli_cancel: %d\n", rc); } else { @@ -735,7 +788,7 @@ static int ldlm_resource_clean(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 flags = (int)(unsigned long)arg; + __u64 flags = *(__u64 *)arg; cleanup_resource(res, &res->lr_granted, flags); cleanup_resource(res, &res->lr_converting, flags); @@ -745,37 +798,46 @@ static int ldlm_resource_clean(cfs_hash_t *hs, cfs_hash_bd_t *bd, } static int ldlm_resource_complain(cfs_hash_t *hs, cfs_hash_bd_t *bd, - cfs_hlist_node_t *hnode, void *arg) + cfs_hlist_node_t *hnode, void *arg) { - struct ldlm_resource *res = cfs_hash_object(hs, hnode); - - CERROR("Namespace %s resource refcount nonzero " - "(%d) after lock cleanup; forcing " - "cleanup.\n", - ldlm_ns_name(ldlm_res_to_ns(res)), - cfs_atomic_read(&res->lr_refcount) - 1); - - CERROR("Resource: %p ("LPU64"/"LPU64"/"LPU64"/" - LPU64") (rc: %d)\n", res, - res->lr_name.name[0], res->lr_name.name[1], - res->lr_name.name[2], res->lr_name.name[3], - cfs_atomic_read(&res->lr_refcount) - 1); - return 0; + struct ldlm_resource *res = cfs_hash_object(hs, hnode); + + lock_res(res); + CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero " + "(%d) after lock cleanup; forcing cleanup.\n", + ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res, + cfs_atomic_read(&res->lr_refcount) - 1); + + ldlm_resource_dump(D_ERROR, res); + unlock_res(res); + return 0; } -int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags) +/** + * Cancel and destroy all locks in the namespace. + * + * Typically used during evictions when server notified client that it was + * evicted and all of its state needs to be destroyed. + * Also used during shutdown. + */ +int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags) { if (ns == NULL) { CDEBUG(D_INFO, "NULL ns, skipping cleanup\n"); return ELDLM_OK; } - cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean, - (void *)(unsigned long)flags); + cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean, &flags); cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain, NULL); return ELDLM_OK; } +EXPORT_SYMBOL(ldlm_namespace_cleanup); +/** + * Attempts to free namespace. + * + * Only used when namespace goes away, like during an unmount. + */ static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force) { ENTRY; @@ -821,12 +883,12 @@ force_wait: } /** - * Performs various cleanups for passed \a ns to make it drop refc and be ready - * for freeing. Waits for refc == 0. + * Performs various cleanups for passed \a ns to make it drop refc and be + * ready for freeing. Waits for refc == 0. * * The following is done: - * (0) Unregister \a ns from its list to make inaccessible for potential users - * like pools thread and others; + * (0) Unregister \a ns from its list to make inaccessible for potential + * users like pools thread and others; * (1) Clear all locks in \a ns. */ void ldlm_namespace_free_prior(struct ldlm_namespace *ns, @@ -840,9 +902,9 @@ void ldlm_namespace_free_prior(struct ldlm_namespace *ns, return; } - cfs_spin_lock(&ns->ns_lock); - ns->ns_stopping = 1; - cfs_spin_unlock(&ns->ns_lock); + spin_lock(&ns->ns_lock); + ns->ns_stopping = 1; + spin_unlock(&ns->ns_lock); /* * Can fail with -EINTR when force == 0 in which case try harder. @@ -865,9 +927,9 @@ void ldlm_namespace_free_prior(struct ldlm_namespace *ns, } /** - * Performs freeing memory structures related to \a ns. This is only done when - * ldlm_namespce_free_prior() successfully removed all resources referencing - * \a ns and its refc == 0. + * Performs freeing memory structures related to \a ns. This is only done + * when ldlm_namespce_free_prior() successfully removed all resources + * referencing \a ns and its refc == 0. */ void ldlm_namespace_free_post(struct ldlm_namespace *ns) { @@ -877,32 +939,26 @@ void ldlm_namespace_free_post(struct ldlm_namespace *ns) return; } - - /* - * Make sure that nobody can find this ns in its list. - */ - ldlm_namespace_unregister(ns, ns->ns_client); - /* - * Fini pool _before_ parent proc dir is removed. This is important as - * ldlm_pool_fini() removes own proc dir which is child to @dir. Removing - * it after @dir may cause oops. - */ - ldlm_pool_fini(&ns->ns_pool); - - ldlm_namespace_proc_unregister(ns); - cfs_hash_putref(ns->ns_rs_hash); - /* - * Namespace \a ns should be not on list in this time, otherwise this - * will cause issues realted to using freed \a ns in pools thread. - */ - LASSERT(cfs_list_empty(&ns->ns_list_chain)); - OBD_FREE_PTR(ns); - ldlm_put_ref(); - EXIT; + /* Make sure that nobody can find this ns in its list. */ + ldlm_namespace_unregister(ns, ns->ns_client); + /* Fini pool _before_ parent proc dir is removed. This is important as + * ldlm_pool_fini() removes own proc dir which is child to @dir. + * Removing it after @dir may cause oops. */ + ldlm_pool_fini(&ns->ns_pool); + + ldlm_namespace_proc_unregister(ns); + cfs_hash_putref(ns->ns_rs_hash); + /* Namespace \a ns should be not on list at this time, otherwise + * this will cause issues related to using freed \a ns in poold + * thread. */ + LASSERT(cfs_list_empty(&ns->ns_list_chain)); + OBD_FREE_PTR(ns); + ldlm_put_ref(); + EXIT; } - -/* Cleanup the resource, and free namespace. +/** + * Cleanup the resource, and free namespace. * bug 12864: * Deadlock issue: * proc1: destroy import @@ -926,68 +982,87 @@ void ldlm_namespace_free(struct ldlm_namespace *ns, ldlm_namespace_free_prior(ns, imp, force); ldlm_namespace_free_post(ns); } +EXPORT_SYMBOL(ldlm_namespace_free); void ldlm_namespace_get(struct ldlm_namespace *ns) { cfs_atomic_inc(&ns->ns_bref); } +EXPORT_SYMBOL(ldlm_namespace_get); + +/* This is only for callers that care about refcount */ +int ldlm_namespace_get_return(struct ldlm_namespace *ns) +{ + return cfs_atomic_inc_return(&ns->ns_bref); +} void ldlm_namespace_put(struct ldlm_namespace *ns) { - if (cfs_atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) { - cfs_waitq_signal(&ns->ns_waitq); - cfs_spin_unlock(&ns->ns_lock); - } + if (cfs_atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) { + cfs_waitq_signal(&ns->ns_waitq); + spin_unlock(&ns->ns_lock); + } } +EXPORT_SYMBOL(ldlm_namespace_put); -/* Register @ns in the list of namespaces */ +/** Register \a ns in the list of namespaces */ void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client) { - cfs_mutex_down(ldlm_namespace_lock(client)); - LASSERT(cfs_list_empty(&ns->ns_list_chain)); - cfs_list_add(&ns->ns_list_chain, ldlm_namespace_list(client)); - cfs_atomic_inc(ldlm_namespace_nr(client)); - cfs_mutex_up(ldlm_namespace_lock(client)); + mutex_lock(ldlm_namespace_lock(client)); + LASSERT(cfs_list_empty(&ns->ns_list_chain)); + cfs_list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client)); + ldlm_namespace_nr_inc(client); + mutex_unlock(ldlm_namespace_lock(client)); } -/* Unregister @ns from the list of namespaces */ +/** Unregister \a ns from the list of namespaces. */ void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client) { - cfs_mutex_down(ldlm_namespace_lock(client)); - LASSERT(!cfs_list_empty(&ns->ns_list_chain)); - /* - * Some asserts and possibly other parts of code still using - * list_empty(&ns->ns_list_chain). This is why it is important - * to use list_del_init() here. - */ - cfs_list_del_init(&ns->ns_list_chain); - cfs_atomic_dec(ldlm_namespace_nr(client)); - cfs_mutex_up(ldlm_namespace_lock(client)); + mutex_lock(ldlm_namespace_lock(client)); + LASSERT(!cfs_list_empty(&ns->ns_list_chain)); + /* Some asserts and possibly other parts of the code are still + * using list_empty(&ns->ns_list_chain). This is why it is + * important to use list_del_init() here. */ + cfs_list_del_init(&ns->ns_list_chain); + ldlm_namespace_nr_dec(client); + mutex_unlock(ldlm_namespace_lock(client)); +} + +/** Should be called with ldlm_namespace_lock(client) taken. */ +void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns, + ldlm_side_t client) +{ + LASSERT(!cfs_list_empty(&ns->ns_list_chain)); + LASSERT_MUTEX_LOCKED(ldlm_namespace_lock(client)); + cfs_list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client)); } -/* Should be called under ldlm_namespace_lock(client) taken */ -void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client) +/** Should be called with ldlm_namespace_lock(client) taken. */ +void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns, + ldlm_side_t client) { - LASSERT(!cfs_list_empty(&ns->ns_list_chain)); - LASSERT_SEM_LOCKED(ldlm_namespace_lock(client)); - cfs_list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client)); + LASSERT(!cfs_list_empty(&ns->ns_list_chain)); + LASSERT_MUTEX_LOCKED(ldlm_namespace_lock(client)); + cfs_list_move_tail(&ns->ns_list_chain, + ldlm_namespace_inactive_list(client)); } -/* Should be called under ldlm_namespace_lock(client) taken */ +/** Should be called with ldlm_namespace_lock(client) taken. */ struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client) { - LASSERT_SEM_LOCKED(ldlm_namespace_lock(client)); + LASSERT_MUTEX_LOCKED(ldlm_namespace_lock(client)); LASSERT(!cfs_list_empty(ldlm_namespace_list(client))); return container_of(ldlm_namespace_list(client)->next, struct ldlm_namespace, ns_list_chain); } +/** Create and initialize new resource. */ static struct ldlm_resource *ldlm_resource_new(void) { struct ldlm_resource *res; int idx; - OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, CFS_ALLOC_IO); + OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, __GFP_IO); if (res == NULL) return NULL; @@ -995,7 +1070,7 @@ static struct ldlm_resource *ldlm_resource_new(void) CFS_INIT_LIST_HEAD(&res->lr_converting); CFS_INIT_LIST_HEAD(&res->lr_waiting); - /* initialize interval trees for each lock mode*/ + /* Initialize interval trees for each lock mode. */ for (idx = 0; idx < LCK_MODE_NUM; idx++) { res->lr_itree[idx].lit_size = 0; res->lr_itree[idx].lit_mode = 1 << idx; @@ -1003,19 +1078,23 @@ static struct ldlm_resource *ldlm_resource_new(void) } cfs_atomic_set(&res->lr_refcount, 1); - cfs_spin_lock_init(&res->lr_lock); - lu_ref_init(&res->lr_reference); + spin_lock_init(&res->lr_lock); + lu_ref_init(&res->lr_reference); - /* one who creates the resource must unlock - * the semaphore after lvb initialization */ - cfs_init_mutex_locked(&res->lr_lvb_sem); + /* The creator of the resource must unlock the mutex after LVB + * initialization. */ + mutex_init(&res->lr_lvb_mutex); + mutex_lock(&res->lr_lvb_mutex); - return res; + return res; } -/* Args: unlocked namespace - * * Locks: takes and releases NS hash-lock and res->lr_lock - * * Returns: referenced, unlocked ldlm_resource or NULL */ +/** + * Return a reference to resource with given name, creating it if necessary. + * Args: namespace with ns_lock unlocked + * Locks: takes and releases NS hash-lock and res->lr_lock + * Returns: referenced, unlocked ldlm_resource or NULL + */ struct ldlm_resource * ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent, const struct ldlm_res_id *name, ldlm_type_t type, int create) @@ -1024,6 +1103,7 @@ ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent, struct ldlm_resource *res; cfs_hash_bd_t bd; __u64 version; + int ns_refcount = 0; LASSERT(ns != NULL); LASSERT(parent == NULL); @@ -1035,11 +1115,16 @@ ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent, if (hnode != NULL) { cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0); res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash); - /* synchronize WRT resource creation */ + /* Synchronize with regard to resource creation. */ if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) { - cfs_down(&res->lr_lvb_sem); - cfs_up(&res->lr_lvb_sem); + mutex_lock(&res->lr_lvb_mutex); + mutex_unlock(&res->lr_lvb_mutex); } + + if (unlikely(res->lr_lvb_len < 0)) { + ldlm_resource_putref(res); + res = NULL; + } return res; } @@ -1065,24 +1150,31 @@ ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent, cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name); if (hnode != NULL) { - /* someone won the race and added the resource before */ - cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1); - /* clean lu_ref for failed resource */ - lu_ref_fini(&res->lr_reference); - OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res); - - res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash); - /* synchronize WRT resource creation */ - if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) { - cfs_down(&res->lr_lvb_sem); - cfs_up(&res->lr_lvb_sem); - } - return res; - } - /* we won! let's add the resource */ + /* Someone won the race and already added the resource. */ + cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1); + /* Clean lu_ref for failed resource. */ + lu_ref_fini(&res->lr_reference); + /* We have taken lr_lvb_mutex. Drop it. */ + mutex_unlock(&res->lr_lvb_mutex); + OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res); + + res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash); + /* Synchronize with regard to resource creation. */ + if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) { + mutex_lock(&res->lr_lvb_mutex); + mutex_unlock(&res->lr_lvb_mutex); + } + + if (unlikely(res->lr_lvb_len < 0)) { + ldlm_resource_putref(res); + res = NULL; + } + return res; + } + /* We won! Let's add the resource. */ cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash); - if (cfs_hash_bd_count_get(&bd) == 1) - ldlm_namespace_get(ns); + if (cfs_hash_bd_count_get(&bd) == 1) + ns_refcount = ldlm_namespace_get_return(ns); cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1); if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) { @@ -1090,15 +1182,37 @@ ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent, OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2); rc = ns->ns_lvbo->lvbo_init(res); - if (rc) - CERROR("lvbo_init failed for resource " - LPU64": rc %d\n", name->name[0], rc); - /* we create resource with locked lr_lvb_sem */ - cfs_up(&res->lr_lvb_sem); - } - - return res; + if (rc < 0) { + CERROR("%s: lvbo_init failed for resource "LPX64":" + LPX64": rc = %d\n", ns->ns_obd->obd_name, + name->name[0], name->name[1], rc); + if (res->lr_lvb_data) { + OBD_FREE(res->lr_lvb_data, res->lr_lvb_len); + res->lr_lvb_data = NULL; + } + res->lr_lvb_len = rc; + mutex_unlock(&res->lr_lvb_mutex); + ldlm_resource_putref(res); + return NULL; + } + } + + /* We create resource with locked lr_lvb_mutex. */ + mutex_unlock(&res->lr_lvb_mutex); + + /* Let's see if we happened to be the very first resource in this + * namespace. If so, and this is a client namespace, we need to move + * the namespace into the active namespaces list to be patrolled by + * the ldlm_poold. */ + if (ns_is_client(ns) && ns_refcount == 1) { + mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT)); + ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT); + mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT)); + } + + return res; } +EXPORT_SYMBOL(ldlm_resource_get); struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res) { @@ -1158,6 +1272,7 @@ int ldlm_resource_putref(struct ldlm_resource *res) } return 0; } +EXPORT_SYMBOL(ldlm_resource_putref); /* Returns 1 if the resource was freed, 0 if it remains. */ int ldlm_resource_putref_locked(struct ldlm_resource *res) @@ -1190,25 +1305,31 @@ int ldlm_resource_putref_locked(struct ldlm_resource *res) return 0; } +/** + * Add a lock into a given resource into specified lock list. + */ void ldlm_resource_add_lock(struct ldlm_resource *res, cfs_list_t *head, struct ldlm_lock *lock) { - check_res_locked(res); + 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); + LDLM_DEBUG(lock, "About to add this lock:\n"); - if (lock->l_destroyed) { - CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n"); - return; - } + if (lock->l_flags & LDLM_FL_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_res_link)); - cfs_list_add_tail(&lock->l_res_link, head); + cfs_list_add_tail(&lock->l_res_link, head); } +/** + * Insert a lock into resource after specified lock. + * + * Obtain resource description from the lock we are inserting after. + */ void ldlm_resource_insert_lock_after(struct ldlm_lock *original, struct ldlm_lock *new) { @@ -1217,13 +1338,12 @@ void ldlm_resource_insert_lock_after(struct ldlm_lock *original, check_res_locked(res); ldlm_resource_dump(D_INFO, res); - CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original); - ldlm_lock_dump(D_OTHER, new, 0); + LDLM_DEBUG(new, "About to insert this lock after %p:\n", original); - if (new->l_destroyed) { - CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n"); - goto out; - } + if (new->l_flags & LDLM_FL_DESTROYED) { + CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n"); + goto out; + } LASSERT(cfs_list_empty(&new->l_res_link)); @@ -1242,6 +1362,7 @@ void ldlm_resource_unlink_lock(struct ldlm_lock *lock) ldlm_extent_unlink_lock(lock); cfs_list_del_init(&lock->l_res_link); } +EXPORT_SYMBOL(ldlm_resource_unlink_lock); void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc) { @@ -1249,6 +1370,10 @@ void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc) desc->lr_name = res->lr_name; } +/** + * Print information about all locks in all namespaces on this node to debug + * log. + */ void ldlm_dump_all_namespaces(ldlm_side_t client, int level) { cfs_list_t *tmp; @@ -1256,7 +1381,7 @@ void ldlm_dump_all_namespaces(ldlm_side_t client, int level) if (!((libcfs_debug | D_ERROR) & level)) return; - cfs_mutex_down(ldlm_namespace_lock(client)); + mutex_lock(ldlm_namespace_lock(client)); cfs_list_for_each(tmp, ldlm_namespace_list(client)) { struct ldlm_namespace *ns; @@ -1264,8 +1389,9 @@ void ldlm_dump_all_namespaces(ldlm_side_t client, int level) ldlm_namespace_dump(level, ns); } - cfs_mutex_up(ldlm_namespace_lock(client)); + mutex_unlock(ldlm_namespace_lock(client)); } +EXPORT_SYMBOL(ldlm_dump_all_namespaces); static int ldlm_res_hash_dump(cfs_hash_t *hs, cfs_hash_bd_t *bd, cfs_hlist_node_t *hnode, void *arg) @@ -1280,6 +1406,10 @@ static int ldlm_res_hash_dump(cfs_hash_t *hs, cfs_hash_bd_t *bd, return 0; } +/** + * Print information about all locks in this namespace on this node to debug + * log. + */ void ldlm_namespace_dump(int level, struct ldlm_namespace *ns) { if (!((libcfs_debug | D_ERROR) & level)) @@ -1295,54 +1425,49 @@ void ldlm_namespace_dump(int level, struct ldlm_namespace *ns) cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_res_hash_dump, (void *)(unsigned long)level); - cfs_spin_lock(&ns->ns_lock); - ns->ns_next_dump = cfs_time_shift(10); - cfs_spin_unlock(&ns->ns_lock); + spin_lock(&ns->ns_lock); + ns->ns_next_dump = cfs_time_shift(10); + spin_unlock(&ns->ns_lock); } +EXPORT_SYMBOL(ldlm_namespace_dump); +/** + * Print information about all locks in this resource to debug log. + */ void ldlm_resource_dump(int level, struct ldlm_resource *res) { - cfs_list_t *tmp; - int pos; + struct ldlm_lock *lock; + unsigned int granted = 0; - CLASSERT(RES_NAME_SIZE == 4); + CLASSERT(RES_NAME_SIZE == 4); - if (!((libcfs_debug | D_ERROR) & level)) - return; + if (!((libcfs_debug | D_ERROR) & level)) + return; - CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64 - ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1], - res->lr_name.name[2], res->lr_name.name[3], - cfs_atomic_read(&res->lr_refcount)); + CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n", + PLDLMRES(res), res, cfs_atomic_read(&res->lr_refcount)); if (!cfs_list_empty(&res->lr_granted)) { - pos = 0; - CDEBUG(level, "Granted locks:\n"); - cfs_list_for_each(tmp, &res->lr_granted) { - struct ldlm_lock *lock; - lock = cfs_list_entry(tmp, struct ldlm_lock, - l_res_link); - ldlm_lock_dump(level, lock, ++pos); + CDEBUG(level, "Granted locks (in reverse order):\n"); + cfs_list_for_each_entry_reverse(lock, &res->lr_granted, + l_res_link) { + LDLM_DEBUG_LIMIT(level, lock, "###"); + if (!(level & D_CANTMASK) && + ++granted > ldlm_dump_granted_max) { + CDEBUG(level, "only dump %d granted locks to " + "avoid DDOS.\n", granted); + break; + } } } if (!cfs_list_empty(&res->lr_converting)) { - pos = 0; CDEBUG(level, "Converting locks:\n"); - cfs_list_for_each(tmp, &res->lr_converting) { - struct ldlm_lock *lock; - lock = cfs_list_entry(tmp, struct ldlm_lock, - l_res_link); - ldlm_lock_dump(level, lock, ++pos); - } + cfs_list_for_each_entry(lock, &res->lr_converting, l_res_link) + LDLM_DEBUG_LIMIT(level, lock, "###"); } if (!cfs_list_empty(&res->lr_waiting)) { - pos = 0; CDEBUG(level, "Waiting locks:\n"); - cfs_list_for_each(tmp, &res->lr_waiting) { - struct ldlm_lock *lock; - lock = cfs_list_entry(tmp, struct ldlm_lock, - l_res_link); - ldlm_lock_dump(level, lock, ++pos); - } + cfs_list_for_each_entry(lock, &res->lr_waiting, l_res_link) + LDLM_DEBUG_LIMIT(level, lock, "###"); } }