X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_resource.c;h=2174271ff73f13cf13b76d1c7267ddc51f605b8c;hp=b10c4b5bb7be902713d4ed5462c2b347aec9ba90;hb=702258c220a8d9224413f259af4a21dda9950278;hpb=a6527b28ba2082a445fbf584909a7791ce407ef3 diff --git a/lustre/ldlm/ldlm_resource.c b/lustre/ldlm/ldlm_resource.c index b10c4b5..2174271 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, 2012, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -46,6 +46,7 @@ # include #endif +#include #include #include "ldlm_internal.h" @@ -54,16 +55,22 @@ cfs_mem_cache_t *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); -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; +struct mutex ldlm_cli_namespace_lock; CFS_LIST_HEAD(ldlm_cli_namespace_list); cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL; cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL; 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) @@ -78,6 +85,12 @@ 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 }, { NULL }}; ENTRY; LASSERT(ldlm_ns_proc_dir == NULL); @@ -134,6 +147,31 @@ void ldlm_proc_cleanup(void) lprocfs_remove(&ldlm_type_proc_dir); } +static int lprocfs_rd_ns_resources(char *page, char **start, off_t off, + int count, int *eof, void *data) +{ + struct ldlm_namespace *ns = data; + __u64 res = 0; + cfs_hash_bd_t bd; + int i; + + /* result is not strictly consistant */ + cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i) + res += cfs_hash_bd_count_get(&bd); + return lprocfs_rd_u64(page, start, off, count, eof, &res); +} + +static int lprocfs_rd_ns_locks(char *page, char **start, off_t off, + int count, int *eof, void *data) +{ + struct ldlm_namespace *ns = data; + __u64 locks; + + locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS, + LPROCFS_FIELDS_FLAGS_SUM); + return lprocfs_rd_u64(page, start, off, count, eof, &locks); +} + static int lprocfs_rd_lru_size(char *page, char **start, off_t off, int count, int *eof, void *data) { @@ -160,7 +198,7 @@ static int lprocfs_wr_lru_size(struct file *file, const char *buffer, if (strncmp(dummy, "clear", 5) == 0) { CDEBUG(D_DLMTRACE, "dropping all unused locks from namespace %s\n", - ns->ns_name); + ldlm_ns_name(ns)); if (ns_connect_lru_resize(ns)) { int canceled, unused = ns->ns_nr_unused; @@ -200,29 +238,31 @@ static int lprocfs_wr_lru_size(struct file *file, const char *buffer, CDEBUG(D_DLMTRACE, "changing namespace %s unused locks from %u to %u\n", - ns->ns_name, ns->ns_nr_unused, (unsigned int)tmp); + ldlm_ns_name(ns), ns->ns_nr_unused, + (unsigned int)tmp); ldlm_cancel_lru(ns, tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED); if (!lru_resize) { CDEBUG(D_DLMTRACE, "disable lru_resize for namespace %s\n", - ns->ns_name); + ldlm_ns_name(ns)); ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE; } } else { CDEBUG(D_DLMTRACE, "changing namespace %s max_unused from %u to %u\n", - ns->ns_name, ns->ns_max_unused, (unsigned int)tmp); + 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); - /* 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, "enable lru_resize for namespace %s\n", - ns->ns_name); + ldlm_ns_name(ns)); ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE; } } @@ -230,150 +270,410 @@ static int lprocfs_wr_lru_size(struct file *file, const char *buffer, return count; } -void ldlm_proc_namespace(struct ldlm_namespace *ns) +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; + + dir = lprocfs_srch(ldlm_ns_proc_dir, ldlm_ns_name(ns)); + if (dir == NULL) { + CERROR("dlm namespace %s has no procfs dir?\n", + ldlm_ns_name(ns)); + } else { + lprocfs_remove(&dir); + } + + if (ns->ns_stats != NULL) + lprocfs_free_stats(&ns->ns_stats); +} + +int ldlm_namespace_proc_register(struct ldlm_namespace *ns) { struct lprocfs_vars lock_vars[2]; char lock_name[MAX_STRING_SIZE + 1]; LASSERT(ns != NULL); - LASSERT(ns->ns_name != NULL); + LASSERT(ns->ns_rs_hash != NULL); + + ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0); + if (ns->ns_stats == NULL) + return -ENOMEM; + + lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS, + LPROCFS_CNTR_AVGMINMAX, "locks", "locks"); lock_name[MAX_STRING_SIZE] = '\0'; memset(lock_vars, 0, sizeof(lock_vars)); lock_vars[0].name = lock_name; - snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name); - lock_vars[0].data = &ns->ns_refcount; - lock_vars[0].read_fptr = lprocfs_rd_atomic; + snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", + ldlm_ns_name(ns)); + lock_vars[0].data = ns; + lock_vars[0].read_fptr = lprocfs_rd_ns_resources; lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0); - snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name); - lock_vars[0].data = &ns->ns_locks; - lock_vars[0].read_fptr = lprocfs_rd_atomic; + snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", + ldlm_ns_name(ns)); + lock_vars[0].data = ns; + lock_vars[0].read_fptr = lprocfs_rd_ns_locks; lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0); if (ns_is_client(ns)) { snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count", - ns->ns_name); + ldlm_ns_name(ns)); lock_vars[0].data = &ns->ns_nr_unused; lock_vars[0].read_fptr = lprocfs_rd_uint; lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0); snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size", - ns->ns_name); + ldlm_ns_name(ns)); lock_vars[0].data = ns; lock_vars[0].read_fptr = lprocfs_rd_lru_size; lock_vars[0].write_fptr = lprocfs_wr_lru_size; lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0); snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age", - ns->ns_name); + ldlm_ns_name(ns)); lock_vars[0].data = &ns->ns_max_age; 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", - ns->ns_name); + ldlm_ns_name(ns)); lock_vars[0].data = &ns->ns_ctime_age_limit; 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/lock_timeouts", - ns->ns_name); + ldlm_ns_name(ns)); lock_vars[0].data = &ns->ns_timeouts; lock_vars[0].read_fptr = lprocfs_rd_uint; lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0); snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes", - ns->ns_name); + ldlm_ns_name(ns)); lock_vars[0].data = &ns->ns_max_nolock_size; 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/contention_seconds", - ns->ns_name); + ldlm_ns_name(ns)); lock_vars[0].data = &ns->ns_contention_time; 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/contended_locks", - ns->ns_name); + ldlm_ns_name(ns)); lock_vars[0].data = &ns->ns_contended_locks; 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/max_parallel_ast", + ldlm_ns_name(ns)); + lock_vars[0].data = &ns->ns_max_parallel_ast; + 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); } + return 0; } #undef MAX_STRING_SIZE -#else -#define ldlm_proc_namespace(ns) do {} while (0) +#else /* LPROCFS */ + +#define ldlm_namespace_proc_unregister(ns) ({;}) +#define ldlm_namespace_proc_register(ns) ({0;}) + #endif /* LPROCFS */ +static unsigned ldlm_res_hop_hash(cfs_hash_t *hs, + const void *key, unsigned mask) +{ + const struct ldlm_res_id *id = key; + unsigned val = 0; + unsigned i; + + for (i = 0; i < RES_NAME_SIZE; i++) + val += id->name[i]; + return val & mask; +} + +static unsigned ldlm_res_hop_fid_hash(cfs_hash_t *hs, + const void *key, unsigned mask) +{ + const struct ldlm_res_id *id = key; + struct lu_fid fid; + __u32 hash; + __u32 val; + + fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF]; + 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); + /* 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); + + return hash & mask; +} + +static void *ldlm_res_hop_key(cfs_hlist_node_t *hnode) +{ + struct ldlm_resource *res; + + res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash); + return &res->lr_name; +} + +static int ldlm_res_hop_keycmp(const void *key, cfs_hlist_node_t *hnode) +{ + struct ldlm_resource *res; + + res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash); + return ldlm_res_eq((const struct ldlm_res_id *)key, + (const struct ldlm_res_id *)&res->lr_name); +} + +static void *ldlm_res_hop_object(cfs_hlist_node_t *hnode) +{ + return cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash); +} + +static void ldlm_res_hop_get_locked(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_getref(res); +} + +static void ldlm_res_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode) +{ + struct ldlm_resource *res; + + res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash); + /* cfs_hash_for_each_nolock is the only chance we call it */ + ldlm_resource_putref_locked(res); +} + +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_putref(res); +} + +cfs_hash_ops_t ldlm_ns_hash_ops = { + .hs_hash = ldlm_res_hop_hash, + .hs_key = ldlm_res_hop_key, + .hs_keycmp = ldlm_res_hop_keycmp, + .hs_keycpy = NULL, + .hs_object = ldlm_res_hop_object, + .hs_get = ldlm_res_hop_get_locked, + .hs_put_locked = ldlm_res_hop_put_locked, + .hs_put = ldlm_res_hop_put +}; + +cfs_hash_ops_t ldlm_ns_fid_hash_ops = { + .hs_hash = ldlm_res_hop_fid_hash, + .hs_key = ldlm_res_hop_key, + .hs_keycmp = ldlm_res_hop_keycmp, + .hs_keycpy = NULL, + .hs_object = ldlm_res_hop_object, + .hs_get = ldlm_res_hop_get_locked, + .hs_put_locked = ldlm_res_hop_put_locked, + .hs_put = ldlm_res_hop_put +}; + +typedef struct { + ldlm_ns_type_t nsd_type; + /** hash bucket bits */ + unsigned nsd_bkt_bits; + /** hash bits */ + unsigned nsd_all_bits; + /** hash operations */ + cfs_hash_ops_t *nsd_hops; +} ldlm_ns_hash_def_t; + +ldlm_ns_hash_def_t ldlm_ns_hash_defs[] = +{ + { + .nsd_type = LDLM_NS_TYPE_MDC, + .nsd_bkt_bits = 11, + .nsd_all_bits = 16, + .nsd_hops = &ldlm_ns_fid_hash_ops, + }, + { + .nsd_type = LDLM_NS_TYPE_MDT, + .nsd_bkt_bits = 14, + .nsd_all_bits = 21, + .nsd_hops = &ldlm_ns_fid_hash_ops, + }, + { + .nsd_type = LDLM_NS_TYPE_OSC, + .nsd_bkt_bits = 8, + .nsd_all_bits = 12, + .nsd_hops = &ldlm_ns_hash_ops, + }, + { + .nsd_type = LDLM_NS_TYPE_OST, + .nsd_bkt_bits = 11, + .nsd_all_bits = 17, + .nsd_hops = &ldlm_ns_hash_ops, + }, + { + .nsd_type = LDLM_NS_TYPE_MGC, + .nsd_bkt_bits = 4, + .nsd_all_bits = 4, + .nsd_hops = &ldlm_ns_hash_ops, + }, + { + .nsd_type = LDLM_NS_TYPE_MGT, + .nsd_bkt_bits = 4, + .nsd_all_bits = 4, + .nsd_hops = &ldlm_ns_hash_ops, + }, + { + .nsd_type = LDLM_NS_TYPE_UNKNOWN, + }, +}; + +/** + * 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) + ldlm_side_t client, + ldlm_appetite_t apt, + ldlm_ns_type_t ns_type) { struct ldlm_namespace *ns = NULL; - cfs_list_t *bucket; - int rc, idx, namelen; + struct ldlm_ns_bucket *nsb; + ldlm_ns_hash_def_t *nsd; + cfs_hash_bd_t bd; + int idx; + int rc; ENTRY; + LASSERT(obd != NULL); + rc = ldlm_get_ref(); if (rc) { CERROR("ldlm_get_ref failed: %d\n", rc); RETURN(NULL); } + for (idx = 0;;idx++) { + nsd = &ldlm_ns_hash_defs[idx]; + if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) { + CERROR("Unknown type %d for ns %s\n", ns_type, name); + GOTO(out_ref, NULL); + } + + if (nsd->nsd_type == ns_type) + break; + } + OBD_ALLOC_PTR(ns); if (!ns) GOTO(out_ref, NULL); - OBD_VMALLOC(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE); - if (!ns->ns_hash) + ns->ns_rs_hash = cfs_hash_create(name, + nsd->nsd_all_bits, nsd->nsd_all_bits, + nsd->nsd_bkt_bits, sizeof(*nsb), + CFS_HASH_MIN_THETA, + CFS_HASH_MAX_THETA, + nsd->nsd_hops, + CFS_HASH_DEPTH | + CFS_HASH_BIGNAME | + CFS_HASH_SPIN_BKTLOCK | + CFS_HASH_NO_ITEMREF); + if (ns->ns_rs_hash == NULL) GOTO(out_ns, NULL); - ns->ns_appetite = apt; - - LASSERT(obd != NULL); - ns->ns_obd = obd; - - namelen = strlen(name); - OBD_ALLOC(ns->ns_name, namelen + 1); - if (!ns->ns_name) - GOTO(out_hash, NULL); + cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) { + nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd); + at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0); + nsb->nsb_namespace = ns; + } - strcpy(ns->ns_name, name); + ns->ns_obd = obd; + ns->ns_appetite = apt; + ns->ns_client = client; - CFS_INIT_LIST_HEAD(&ns->ns_root_list); CFS_INIT_LIST_HEAD(&ns->ns_list_chain); - ns->ns_refcount = 0; - ns->ns_client = client; - cfs_spin_lock_init(&ns->ns_hash_lock); - cfs_atomic_set(&ns->ns_locks, 0); - ns->ns_resources = 0; + CFS_INIT_LIST_HEAD(&ns->ns_unused_list); + spin_lock_init(&ns->ns_lock); + cfs_atomic_set(&ns->ns_bref, 0); cfs_waitq_init(&ns->ns_waitq); - ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES; - ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS; - ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS; - for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash; - bucket--) - CFS_INIT_LIST_HEAD(bucket); + ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES; + ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS; + ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS; - CFS_INIT_LIST_HEAD(&ns->ns_unused_list); - ns->ns_nr_unused = 0; - ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE; - ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE; - ns->ns_ctime_age_limit = LDLM_CTIME_AGE_LIMIT; - ns->ns_timeouts = 0; - cfs_spin_lock_init(&ns->ns_unused_lock); + 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; + ns->ns_ctime_age_limit = LDLM_CTIME_AGE_LIMIT; + ns->ns_timeouts = 0; ns->ns_orig_connect_flags = 0; - ns->ns_connect_flags = 0; - ldlm_proc_namespace(ns); + ns->ns_connect_flags = 0; + ns->ns_stopping = 0; + rc = ldlm_namespace_proc_register(ns); + if (rc != 0) { + CERROR("Can't initialize ns proc, rc %d\n", rc); + GOTO(out_hash, rc); + } idx = cfs_atomic_read(ldlm_namespace_nr(client)); rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client); @@ -382,43 +682,43 @@ struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name, GOTO(out_proc, rc); } - at_init(&ns->ns_at_estimate, ldlm_enqueue_min, 0); - ldlm_namespace_register(ns, client); RETURN(ns); out_proc: + ldlm_namespace_proc_unregister(ns); ldlm_namespace_cleanup(ns, 0); - OBD_FREE(ns->ns_name, namelen + 1); out_hash: - OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE); + cfs_hash_putref(ns->ns_rs_hash); out_ns: OBD_FREE_PTR(ns); 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. * - * Called with the ns_lock held. */ + * 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(res->lr_namespace); - int local_only = (flags & LDLM_FL_LOCAL_ONLY); - ENTRY; - + 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, @@ -438,7 +738,7 @@ 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; @@ -477,65 +777,69 @@ static void cleanup_resource(struct ldlm_resource *res, cfs_list_t *q, } LDLM_LOCK_RELEASE(lock); } while (1); +} - EXIT; +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); + __u64 flags = *(__u64 *)arg; + + cleanup_resource(res, &res->lr_granted, flags); + cleanup_resource(res, &res->lr_converting, flags); + cleanup_resource(res, &res->lr_waiting, flags); + + return 0; } -int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags) +static int ldlm_resource_complain(cfs_hash_t *hs, cfs_hash_bd_t *bd, + cfs_hlist_node_t *hnode, void *arg) { - cfs_list_t *tmp; - int i; + struct ldlm_resource *res = cfs_hash_object(hs, hnode); + + lock_res(res); + 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); + + ldlm_resource_dump(D_ERROR, res); + unlock_res(res); + return 0; +} + +/** + * 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; } - for (i = 0; i < RES_HASH_SIZE; i++) { - cfs_spin_lock(&ns->ns_hash_lock); - tmp = ns->ns_hash[i].next; - while (tmp != &(ns->ns_hash[i])) { - struct ldlm_resource *res; - res = cfs_list_entry(tmp, struct ldlm_resource, - lr_hash); - ldlm_resource_getref(res); - cfs_spin_unlock(&ns->ns_hash_lock); - LDLM_RESOURCE_ADDREF(res); - - cleanup_resource(res, &res->lr_granted, flags); - cleanup_resource(res, &res->lr_converting, flags); - cleanup_resource(res, &res->lr_waiting, flags); - - cfs_spin_lock(&ns->ns_hash_lock); - tmp = tmp->next; - - /* XXX: former stuff caused issues in case of race - * between ldlm_namespace_cleanup() and lockd() when - * client gets blocking ast when lock gets distracted by - * server. This is 1_4 branch solution, let's see how - * will it behave. */ - LDLM_RESOURCE_DELREF(res); - if (!ldlm_resource_putref_locked(res)) { - CERROR("Namespace %s resource refcount nonzero " - "(%d) after lock cleanup; forcing " - "cleanup.\n", - ns->ns_name, - cfs_atomic_read(&res->lr_refcount)); - 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)); - } - } - cfs_spin_unlock(&ns->ns_hash_lock); - } - + 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; @@ -543,50 +847,50 @@ static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force) /* At shutdown time, don't call the cancellation callback */ ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0); - if (ns->ns_refcount > 0) { + if (cfs_atomic_read(&ns->ns_bref) > 0) { struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL); int rc; CDEBUG(D_DLMTRACE, "dlm namespace %s free waiting on refcount %d\n", - ns->ns_name, ns->ns_refcount); + ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref)); force_wait: if (force) lwi = LWI_TIMEOUT(obd_timeout * CFS_HZ / 4, NULL, NULL); rc = l_wait_event(ns->ns_waitq, - ns->ns_refcount == 0, &lwi); + cfs_atomic_read(&ns->ns_bref) == 0, &lwi); /* Forced cleanups should be able to reclaim all references, * so it's safe to wait forever... we can't leak locks... */ if (force && rc == -ETIMEDOUT) { LCONSOLE_ERROR("Forced cleanup waiting for %s " "namespace with %d resources in use, " - "(rc=%d)\n", ns->ns_name, - ns->ns_refcount, rc); + "(rc=%d)\n", ldlm_ns_name(ns), + cfs_atomic_read(&ns->ns_bref), rc); GOTO(force_wait, rc); } - if (ns->ns_refcount) { + if (cfs_atomic_read(&ns->ns_bref)) { LCONSOLE_ERROR("Cleanup waiting for %s namespace " "with %d resources in use, (rc=%d)\n", - ns->ns_name, - ns->ns_refcount, rc); + ldlm_ns_name(ns), + cfs_atomic_read(&ns->ns_bref), rc); RETURN(ELDLM_NAMESPACE_EXISTS); } - CDEBUG(D_DLMTRACE, - "dlm namespace %s free done waiting\n", ns->ns_name); + CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n", + ldlm_ns_name(ns)); } RETURN(ELDLM_OK); } /** - * 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, @@ -600,6 +904,9 @@ void ldlm_namespace_free_prior(struct ldlm_namespace *ns, return; } + 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. @@ -622,9 +929,9 @@ 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) { @@ -634,46 +941,26 @@ 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); - -#ifdef LPROCFS - { - struct proc_dir_entry *dir; - dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name); - if (dir == NULL) { - CERROR("dlm namespace %s has no procfs dir?\n", - ns->ns_name); - } else { - lprocfs_remove(&dir); - } - } -#endif - - OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE); - OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1); - - /* - * 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 @@ -697,90 +984,64 @@ void ldlm_namespace_free(struct ldlm_namespace *ns, ldlm_namespace_free_prior(ns, imp, force); ldlm_namespace_free_post(ns); } - - -void ldlm_namespace_get_locked(struct ldlm_namespace *ns) -{ - ns->ns_refcount++; -} +EXPORT_SYMBOL(ldlm_namespace_free); void ldlm_namespace_get(struct ldlm_namespace *ns) { - cfs_spin_lock(&ns->ns_hash_lock); - ldlm_namespace_get_locked(ns); - cfs_spin_unlock(&ns->ns_hash_lock); + cfs_atomic_inc(&ns->ns_bref); } +EXPORT_SYMBOL(ldlm_namespace_get); -void ldlm_namespace_put_locked(struct ldlm_namespace *ns, int wakeup) +void ldlm_namespace_put(struct ldlm_namespace *ns) { - LASSERT(ns->ns_refcount > 0); - ns->ns_refcount--; - if (ns->ns_refcount == 0 && wakeup) - cfs_waitq_signal(&ns->ns_waitq); + 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); -void ldlm_namespace_put(struct ldlm_namespace *ns, int wakeup) -{ - cfs_spin_lock(&ns->ns_hash_lock); - ldlm_namespace_put_locked(ns, wakeup); - cfs_spin_unlock(&ns->ns_hash_lock); -} - -/* 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_list(client)); + cfs_atomic_inc(ldlm_namespace_nr(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); + cfs_atomic_dec(ldlm_namespace_nr(client)); + mutex_unlock(ldlm_namespace_lock(client)); } -/* Should be called under ldlm_namespace_lock(client) taken */ +/** Should be called with ldlm_namespace_lock(client) taken. */ void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client) { LASSERT(!cfs_list_empty(&ns->ns_list_chain)); - LASSERT_SEM_LOCKED(ldlm_namespace_lock(client)); + 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 */ +/** 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); } -static __u32 ldlm_hash_fn(struct ldlm_resource *parent, - const struct ldlm_res_id *name) -{ - __u32 hash = 0; - int i; - - for (i = 0; i < RES_NAME_SIZE; i++) - hash += name->name[i]; - - hash += (__u32)((unsigned long)parent >> 4); - - return (hash & RES_HASH_MASK); -} +/** Create and initialize new resource. */ static struct ldlm_resource *ldlm_resource_new(void) { struct ldlm_resource *res; @@ -790,15 +1051,11 @@ static struct ldlm_resource *ldlm_resource_new(void) if (res == NULL) return NULL; - memset(res, 0, sizeof(*res)); - - CFS_INIT_LIST_HEAD(&res->lr_children); - CFS_INIT_LIST_HEAD(&res->lr_childof); CFS_INIT_LIST_HEAD(&res->lr_granted); 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; @@ -806,89 +1063,94 @@ 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; } -/* must be called with hash lock held */ -static struct ldlm_resource * -ldlm_resource_find(struct ldlm_namespace *ns, const struct ldlm_res_id *name, - __u32 hash) +/** + * 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) { - cfs_list_t *bucket, *tmp; + cfs_hlist_node_t *hnode; struct ldlm_resource *res; + cfs_hash_bd_t bd; + __u64 version; - LASSERT_SPIN_LOCKED(&ns->ns_hash_lock); - bucket = ns->ns_hash + hash; + LASSERT(ns != NULL); + LASSERT(parent == NULL); + LASSERT(ns->ns_rs_hash != NULL); + LASSERT(name->name[0] != 0); - cfs_list_for_each(tmp, bucket) { - res = cfs_list_entry(tmp, struct ldlm_resource, lr_hash); - if (memcmp(&res->lr_name, name, sizeof(res->lr_name)) == 0) - return res; + cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0); + hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name); + if (hnode != NULL) { + cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0); + 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); + } + return res; } - return NULL; -} + version = cfs_hash_bd_version_get(&bd); + cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0); -/* Args: locked namespace - * Returns: newly-allocated, referenced, unlocked resource */ -static struct ldlm_resource * -ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent, - const struct ldlm_res_id *name, __u32 hash, ldlm_type_t type) -{ - cfs_list_t *bucket; - struct ldlm_resource *res, *old_res; - ENTRY; + if (create == 0) + return NULL; LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE, "type: %d\n", type); - res = ldlm_resource_new(); if (!res) - RETURN(NULL); + return NULL; - res->lr_name = *name; - res->lr_namespace = ns; - res->lr_type = type; + res->lr_ns_bucket = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd); + res->lr_name = *name; + res->lr_type = type; res->lr_most_restr = LCK_NL; - cfs_spin_lock(&ns->ns_hash_lock); - old_res = ldlm_resource_find(ns, name, hash); - if (old_res) { - /* someone won the race and added the resource before */ - ldlm_resource_getref(old_res); - cfs_spin_unlock(&ns->ns_hash_lock); - /* clean lu_ref for failed resource */ - lu_ref_fini(&res->lr_reference); - OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res); - /* synchronize WRT resource creation */ - if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) { - cfs_down(&old_res->lr_lvb_sem); - cfs_up(&old_res->lr_lvb_sem); - } - RETURN(old_res); - } - - /* we won! let's add the resource */ - bucket = ns->ns_hash + hash; - cfs_list_add(&res->lr_hash, bucket); - ns->ns_resources++; - ldlm_namespace_get_locked(ns); - - if (parent == NULL) { - cfs_list_add(&res->lr_childof, &ns->ns_root_list); - } else { - res->lr_parent = parent; - cfs_list_add(&res->lr_childof, &parent->lr_children); - } - cfs_spin_unlock(&ns->ns_hash_lock); - + cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1); + hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL : + cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name); + + if (hnode != NULL) { + /* 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); + } + 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); + + cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1); if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) { int rc; @@ -897,48 +1159,14 @@ ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent, 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); -} + } -/* Args: unlocked namespace - * Locks: takes and releases ns->ns_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) -{ - __u32 hash = ldlm_hash_fn(parent, name); - struct ldlm_resource *res = NULL; - ENTRY; - - LASSERT(ns != NULL); - LASSERT(ns->ns_hash != NULL); - LASSERT(name->name[0] != 0); + /* We create resource with locked lr_lvb_mutex. */ + mutex_unlock(&res->lr_lvb_mutex); - cfs_spin_lock(&ns->ns_hash_lock); - res = ldlm_resource_find(ns, name, hash); - if (res) { - ldlm_resource_getref(res); - cfs_spin_unlock(&ns->ns_hash_lock); - /* 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); - } - cfs_spin_unlock(&ns->ns_hash_lock); - - if (create == 0) - RETURN(NULL); - - res = ldlm_resource_add(ns, parent, name, hash, type); - RETURN(res); + return res; } +EXPORT_SYMBOL(ldlm_resource_get); struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res) { @@ -950,11 +1178,10 @@ struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res) return res; } -void __ldlm_resource_putref_final(struct ldlm_resource *res) +static void __ldlm_resource_putref_final(cfs_hash_bd_t *bd, + struct ldlm_resource *res) { - struct ldlm_namespace *ns = res->lr_namespace; - - LASSERT_SPIN_LOCKED(&ns->ns_hash_lock); + struct ldlm_ns_bucket *nsb = res->lr_ns_bucket; if (!cfs_list_empty(&res->lr_granted)) { ldlm_resource_dump(D_ERROR, res); @@ -971,80 +1198,76 @@ void __ldlm_resource_putref_final(struct ldlm_resource *res) LBUG(); } - if (!cfs_list_empty(&res->lr_children)) { - ldlm_resource_dump(D_ERROR, res); - LBUG(); - } - - /* Pass 0 here to not wake ->ns_waitq up yet, we will do it few - * lines below when all children are freed. */ - ldlm_namespace_put_locked(ns, 0); - cfs_list_del_init(&res->lr_hash); - cfs_list_del_init(&res->lr_childof); + cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash, + bd, &res->lr_hash); lu_ref_fini(&res->lr_reference); - - ns->ns_resources--; - if (ns->ns_resources == 0) - cfs_waitq_signal(&ns->ns_waitq); + if (cfs_hash_bd_count_get(bd) == 0) + ldlm_namespace_put(nsb->nsb_namespace); } /* Returns 1 if the resource was freed, 0 if it remains. */ int ldlm_resource_putref(struct ldlm_resource *res) { - struct ldlm_namespace *ns = res->lr_namespace; - int rc = 0; - ENTRY; - - CDEBUG(D_INFO, "putref res: %p count: %d\n", res, - cfs_atomic_read(&res->lr_refcount) - 1); - LASSERTF(cfs_atomic_read(&res->lr_refcount) > 0, "%d", - cfs_atomic_read(&res->lr_refcount)); - LASSERTF(cfs_atomic_read(&res->lr_refcount) < LI_POISON, "%d", - cfs_atomic_read(&res->lr_refcount)); - - if (cfs_atomic_dec_and_lock(&res->lr_refcount, &ns->ns_hash_lock)) { - __ldlm_resource_putref_final(res); - cfs_spin_unlock(&ns->ns_hash_lock); - if (res->lr_lvb_data) - OBD_FREE(res->lr_lvb_data, res->lr_lvb_len); + struct ldlm_namespace *ns = ldlm_res_to_ns(res); + cfs_hash_bd_t bd; + + LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON); + CDEBUG(D_INFO, "putref res: %p count: %d\n", + res, cfs_atomic_read(&res->lr_refcount) - 1); + + cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd); + if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) { + __ldlm_resource_putref_final(&bd, res); + cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1); + if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free) + ns->ns_lvbo->lvbo_free(res); OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res); - rc = 1; + return 1; } - - RETURN(rc); + 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) { - int rc = 0; - ENTRY; + struct ldlm_namespace *ns = ldlm_res_to_ns(res); - CDEBUG(D_INFO, "putref res: %p count: %d\n", res, - cfs_atomic_read(&res->lr_refcount) - 1); - LASSERT(cfs_atomic_read(&res->lr_refcount) > 0); - LASSERT(cfs_atomic_read(&res->lr_refcount) < LI_POISON); + LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON); + CDEBUG(D_INFO, "putref res: %p count: %d\n", + res, cfs_atomic_read(&res->lr_refcount) - 1); - LASSERT(cfs_atomic_read(&res->lr_refcount) >= 0); if (cfs_atomic_dec_and_test(&res->lr_refcount)) { - __ldlm_resource_putref_final(res); - if (res->lr_lvb_data) - OBD_FREE(res->lr_lvb_data, res->lr_lvb_len); + cfs_hash_bd_t bd; + + cfs_hash_bd_get(ldlm_res_to_ns(res)->ns_rs_hash, + &res->lr_name, &bd); + __ldlm_resource_putref_final(&bd, res); + cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1); + /* NB: ns_rs_hash is created with CFS_HASH_NO_ITEMREF, + * so we should never be here while calling cfs_hash_del, + * cfs_hash_for_each_nolock is the only case we can get + * here, which is safe to release cfs_hash_bd_lock. + */ + if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free) + ns->ns_lvbo->lvbo_free(res); OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res); - rc = 1; - } - RETURN(rc); + cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1); + return 1; + } + 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); - 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"); @@ -1056,6 +1279,11 @@ void ldlm_resource_add_lock(struct ldlm_resource *res, cfs_list_t *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) { @@ -1064,8 +1292,7 @@ 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"); @@ -1089,6 +1316,7 @@ 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) { @@ -1096,6 +1324,10 @@ 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; @@ -1103,7 +1335,7 @@ 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; @@ -1111,50 +1343,55 @@ 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); -void ldlm_namespace_dump(int level, struct ldlm_namespace *ns) +static int ldlm_res_hash_dump(cfs_hash_t *hs, cfs_hash_bd_t *bd, + cfs_hlist_node_t *hnode, void *arg) { - cfs_list_t *tmp; + struct ldlm_resource *res = cfs_hash_object(hs, hnode); + int level = (int)(unsigned long)arg; + lock_res(res); + ldlm_resource_dump(level, res); + unlock_res(res); + + 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)) return; CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n", - ns->ns_name, ns->ns_refcount, + ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref), ns_is_client(ns) ? "client" : "server"); if (cfs_time_before(cfs_time_current(), ns->ns_next_dump)) return; - cfs_spin_lock(&ns->ns_hash_lock); - tmp = ns->ns_root_list.next; - while (tmp != &ns->ns_root_list) { - struct ldlm_resource *res; - res = cfs_list_entry(tmp, struct ldlm_resource, lr_childof); - - ldlm_resource_getref(res); - cfs_spin_unlock(&ns->ns_hash_lock); - LDLM_RESOURCE_ADDREF(res); - - lock_res(res); - ldlm_resource_dump(level, res); - unlock_res(res); - - LDLM_RESOURCE_DELREF(res); - cfs_spin_lock(&ns->ns_hash_lock); - tmp = tmp->next; - ldlm_resource_putref_locked(res); - } - ns->ns_next_dump = cfs_time_shift(10); - cfs_spin_unlock(&ns->ns_hash_lock); + cfs_hash_for_each_nolock(ns->ns_rs_hash, + ldlm_res_hash_dump, + (void *)(unsigned long)level); + 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); @@ -1167,33 +1404,26 @@ void ldlm_resource_dump(int level, struct ldlm_resource *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, "###"); } }