X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_resource.c;h=2f74265e74519c0e01487de75831e23d68b65017;hb=60b3f96ff1fb98d056fe4bc5c1abf402fbda46ef;hp=297da2ebb2567d88eac1f116c53f82031d47a553;hpb=df8ddaba8a5595e474cb9abceeb6a6fb17073ed9;p=fs%2Flustre-release.git diff --git a/lustre/ldlm/ldlm_resource.c b/lustre/ldlm/ldlm_resource.c index 297da2e..2f74265 100644 --- a/lustre/ldlm/ldlm_resource.c +++ b/lustre/ldlm/ldlm_resource.c @@ -1,186 +1,618 @@ /* -*- 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: Phil Schwan + * Author: Peter Braam * - * This code is issued under the GNU General Public License. - * See the file COPYING in this distribution + * This file is part of the Lustre file system, http://www.lustre.org + * Lustre is a trademark of Cluster File Systems, Inc. * - * by Cluster File Systems, Inc. + * 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. + * + * 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. + * + * 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 +#ifdef __KERNEL__ +# include +#else +# include +#endif + +#include +#include "ldlm_internal.h" + +cfs_mem_cache_t *ldlm_resource_slab, *ldlm_lock_slab; -#include +atomic_t ldlm_srv_namespace_nr = ATOMIC_INIT(0); +atomic_t ldlm_cli_namespace_nr = ATOMIC_INIT(0); -kmem_cache_t *ldlm_resource_slab, *ldlm_lock_slab; +struct semaphore ldlm_srv_namespace_lock; +struct list_head ldlm_srv_namespace_list = + CFS_LIST_HEAD_INIT(ldlm_srv_namespace_list); -spinlock_t ldlm_namespace_lock = SPIN_LOCK_UNLOCKED; -struct list_head ldlm_namespace_list = LIST_HEAD_INIT(ldlm_namespace_list); -static struct proc_dir_entry *ldlm_ns_proc_dir = NULL; +struct semaphore ldlm_cli_namespace_lock; +struct list_head ldlm_cli_namespace_list = + CFS_LIST_HEAD_INIT(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; + +#ifdef LPROCFS +static int ldlm_proc_dump_ns(struct file *file, const char *buffer, + unsigned long count, void *data) +{ + ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE); + ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE); + RETURN(count); +} -int ldlm_proc_setup(struct obd_device *obd) +int ldlm_proc_setup(void) { + int rc; + struct lprocfs_vars list[] = { + { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL }, + { NULL }}; ENTRY; + LASSERT(ldlm_ns_proc_dir == NULL); + + ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME, + proc_lustre_root, + NULL, NULL); + if (IS_ERR(ldlm_type_proc_dir)) { + CERROR("LProcFS failed in ldlm-init\n"); + rc = PTR_ERR(ldlm_type_proc_dir); + GOTO(err, rc); + } - if (obd->obd_proc_entry == NULL) - RETURN(-EINVAL); + ldlm_ns_proc_dir = lprocfs_register("namespaces", + ldlm_type_proc_dir, + NULL, NULL); + if (IS_ERR(ldlm_ns_proc_dir)) { + CERROR("LProcFS failed in ldlm-init\n"); + rc = PTR_ERR(ldlm_ns_proc_dir); + GOTO(err_type, rc); + } - ldlm_ns_proc_dir = proc_mkdir("namespaces", obd->obd_proc_entry); - if (ldlm_ns_proc_dir == NULL) { - CERROR("Couldn't create /proc/lustre/ldlm/namespaces\n"); - RETURN(-EPERM); + ldlm_svc_proc_dir = lprocfs_register("services", + ldlm_type_proc_dir, + NULL, NULL); + if (IS_ERR(ldlm_svc_proc_dir)) { + CERROR("LProcFS failed in ldlm-init\n"); + rc = PTR_ERR(ldlm_svc_proc_dir); + GOTO(err_ns, rc); } + + rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL); + RETURN(0); + +err_ns: + lprocfs_remove(&ldlm_ns_proc_dir); +err_type: + lprocfs_remove(&ldlm_type_proc_dir); +err: + ldlm_svc_proc_dir = NULL; + RETURN(rc); } -void ldlm_proc_cleanup(struct obd_device *obd) +void ldlm_proc_cleanup(void) { - proc_lustre_remove_obd_entry("namespaces", obd); + if (ldlm_svc_proc_dir) + lprocfs_remove(&ldlm_svc_proc_dir); + + if (ldlm_ns_proc_dir) + lprocfs_remove(&ldlm_ns_proc_dir); + + if (ldlm_type_proc_dir) + lprocfs_remove(&ldlm_type_proc_dir); } -struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 client) +static int lprocfs_rd_lru_size(char *page, char **start, off_t off, + int count, int *eof, void *data) { - struct ldlm_namespace *ns = NULL; - struct list_head *bucket; + struct ldlm_namespace *ns = data; + __u32 *nr = &ns->ns_max_unused; - OBD_ALLOC(ns, sizeof(*ns)); - if (!ns) { - LBUG(); - GOTO(out, NULL); + if (ns_connect_lru_resize(ns)) + nr = &ns->ns_nr_unused; + return lprocfs_rd_uint(page, start, off, count, eof, nr); +} + +static int lprocfs_wr_lru_size(struct file *file, const char *buffer, + unsigned long count, void *data) +{ + struct ldlm_namespace *ns = data; + char dummy[MAX_STRING_SIZE + 1], *end; + unsigned long tmp; + int lru_resize; + + dummy[MAX_STRING_SIZE] = '\0'; + if (copy_from_user(dummy, buffer, MAX_STRING_SIZE)) + return -EFAULT; + + if (count == 6 && memcmp(dummy, "clear", 5) == 0) { + CDEBUG(D_DLMTRACE, + "dropping all unused locks from namespace %s\n", + ns->ns_name); + if (ns_connect_lru_resize(ns)) { + 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); + if (canceled < unused) { + CERROR("not all requested locks are canceled, " + "requested: %d, canceled: %d\n", unused, + canceled); + return -EINVAL; + } + } else { + tmp = ns->ns_max_unused; + ns->ns_max_unused = 0; + ldlm_cancel_lru(ns, 0, LDLM_SYNC, LDLM_CANCEL_PASSED); + ns->ns_max_unused = tmp; + } + return count; } - ns->ns_hash = vmalloc(sizeof(*ns->ns_hash) * RES_HASH_SIZE); - if (!ns->ns_hash) { - LBUG(); - GOTO(out, ns); + tmp = simple_strtoul(dummy, &end, 0); + if (dummy == end) { + CERROR("invalid value written\n"); + return -EINVAL; + } + lru_resize = (tmp == 0); + + if (ns_connect_lru_resize(ns)) { + if (!lru_resize) + ns->ns_max_unused = (unsigned int)tmp; + + if (tmp > ns->ns_nr_unused) + tmp = ns->ns_nr_unused; + tmp = ns->ns_nr_unused - tmp; + + CDEBUG(D_DLMTRACE, "changing namespace %s unused locks from %u to %u\n", + ns->ns_name, ns->ns_nr_unused, (unsigned int)tmp); + ldlm_cancel_lru(ns, (unsigned int)tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED); + + if (!lru_resize) { + CDEBUG(D_DLMTRACE, "disable lru_resize for namespace %s\n", + ns->ns_name); + 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); + 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. */ + 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); + ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE; + } } - OBD_ALLOC(ns->ns_name, strlen(name) + 1); - if (!ns->ns_name) { - LBUG(); - GOTO(out, ns); + return count; +} + +void ldlm_proc_namespace(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); + + 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; + 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; + 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); + 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); + 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/shrink_thumb", + ns->ns_name); + lock_vars[0].data = 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/lru_max_age", + ns->ns_name); + 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); } - strcpy(ns->ns_name, name); +} +#undef MAX_STRING_SIZE +#else +#define ldlm_proc_namespace(ns) do {} while (0) +#endif /* LPROCFS */ - ptlrpc_init_client(NULL, NULL, LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL, - &ns->ns_rpc_client); +struct ldlm_namespace *ldlm_namespace_new(char *name, ldlm_side_t client, + ldlm_appetite_t apt) +{ + struct ldlm_namespace *ns = NULL; + struct list_head *bucket; + int rc, idx, namelen; + ENTRY; - INIT_LIST_HEAD(&ns->ns_root_list); - l_lock_init(&ns->ns_lock); + rc = ldlm_get_ref(); + if (rc) { + CERROR("ldlm_get_ref failed: %d\n", rc); + RETURN(NULL); + } + + 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) + GOTO(out_ns, NULL); + + ns->ns_shrink_thumb = LDLM_LOCK_SHRINK_THUMB; + ns->ns_appetite = apt; + namelen = strlen(name); + OBD_ALLOC(ns->ns_name, namelen + 1); + if (!ns->ns_name) + GOTO(out_hash, NULL); + + strcpy(ns->ns_name, name); + + CFS_INIT_LIST_HEAD(&ns->ns_root_list); ns->ns_refcount = 0; ns->ns_client = client; + spin_lock_init(&ns->ns_hash_lock); + atomic_set(&ns->ns_locks, 0); + ns->ns_resources = 0; + cfs_waitq_init(&ns->ns_waitq); for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash; bucket--) - INIT_LIST_HEAD(bucket); + CFS_INIT_LIST_HEAD(bucket); + + 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; + spin_lock_init(&ns->ns_unused_lock); + ns->ns_orig_connect_flags = 0; + ns->ns_connect_flags = 0; + ldlm_proc_namespace(ns); + + idx = atomic_read(ldlm_namespace_nr(client)); + rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client); + if (rc) { + CERROR("Can't initialize lock pool, rc %d\n", rc); + GOTO(out_proc, rc); + } - spin_lock(&ldlm_namespace_lock); - list_add(&ns->ns_list_chain, &ldlm_namespace_list); - ns->ns_proc_dir = proc_mkdir(ns->ns_name, ldlm_ns_proc_dir); - if (ns->ns_proc_dir == NULL) - CERROR("Unable to create proc directory for namespace.\n"); - spin_unlock(&ldlm_namespace_lock); + mutex_down(ldlm_namespace_lock(client)); + list_add(&ns->ns_list_chain, ldlm_namespace_list(client)); + atomic_inc(ldlm_namespace_nr(client)); + mutex_up(ldlm_namespace_lock(client)); RETURN(ns); - - out: - if (ns && ns->ns_hash) - vfree(ns->ns_hash); - if (ns && ns->ns_name) - OBD_FREE(ns->ns_name, strlen(name) + 1); - if (ns) - OBD_FREE(ns, sizeof(*ns)); - return NULL; +out_proc: + 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); +out_ns: + OBD_FREE_PTR(ns); +out_ref: + ldlm_put_ref(0); + RETURN(NULL); } extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock); -static void cleanup_resource(struct ldlm_resource *res, struct list_head *q) +/* 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 + * + * Called with the ns_lock held. */ +static void cleanup_resource(struct ldlm_resource *res, struct list_head *q, + int flags) { - struct list_head *tmp, *pos; - int rc = 0, client = res->lr_namespace->ns_client; + struct list_head *tmp; + int rc = 0, client = ns_is_client(res->lr_namespace); + int local_only = (flags & LDLM_FL_LOCAL_ONLY); ENTRY; - list_for_each_safe(tmp, pos, q) { - struct ldlm_lock *lock; - lock = list_entry(tmp, struct ldlm_lock, l_res_link); - LDLM_LOCK_GET(lock); + + do { + struct ldlm_lock *lock = NULL; + + /* first, we look for non-cleaned-yet lock + * all cleaned locks are marked by CLEANED flag */ + lock_res(res); + list_for_each(tmp, q) { + lock = list_entry(tmp, struct ldlm_lock, l_res_link); + if (lock->l_flags & LDLM_FL_CLEANED) { + lock = NULL; + continue; + } + LDLM_LOCK_GET(lock); + lock->l_flags |= LDLM_FL_CLEANED; + break; + } + + if (lock == NULL) { + unlock_res(res); + break; + } + + /* Set CBPENDING so nothing in the cancellation path + * can match this lock */ + lock->l_flags |= LDLM_FL_CBPENDING; + lock->l_flags |= LDLM_FL_FAILED; + lock->l_flags |= flags; + + /* ... without sending a CANCEL message for local_only. */ + if (local_only) + lock->l_flags |= LDLM_FL_LOCAL_ONLY; + + if (local_only && (lock->l_readers || lock->l_writers)) { + /* This is a little bit gross, but much better than the + * alternative: pretend that we got a blocking AST from + * the server, so that when the lock is decref'd, it + * will go away ... */ + unlock_res(res); + LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY"); + if (lock->l_completion_ast) + lock->l_completion_ast(lock, 0, NULL); + LDLM_LOCK_PUT(lock); + continue; + } if (client) { struct lustre_handle lockh; + + unlock_res(res); ldlm_lock2handle(lock, &lockh); - /* can we get away without a connh here? */ rc = ldlm_cli_cancel(&lockh); - if (rc != ELDLM_OK) { - /* It failed remotely, but we'll force it to - * cleanup locally. */ + if (rc) CERROR("ldlm_cli_cancel: %d\n", rc); - ldlm_lock_cancel(lock); - } } else { - CERROR("Freeing lock %p still held by client node.\n", - lock); - ldlm_lock_dump(lock); - ldlm_resource_unlink_lock(lock); + unlock_res(res); + LDLM_DEBUG(lock, "Freeing a lock still held by a " + "client node"); ldlm_lock_destroy(lock); } LDLM_LOCK_PUT(lock); - } + } while (1); + + EXIT; } -int ldlm_namespace_free(struct ldlm_namespace *ns) +int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags) { - struct list_head *tmp, *pos; + struct list_head *tmp; int i; + if (ns == NULL) { + CDEBUG(D_INFO, "NULL ns, skipping cleanup\n"); + return ELDLM_OK; + } + + for (i = 0; i < RES_HASH_SIZE; i++) { + spin_lock(&ns->ns_hash_lock); + tmp = ns->ns_hash[i].next; + while (tmp != &(ns->ns_hash[i])) { + struct ldlm_resource *res; + res = list_entry(tmp, struct ldlm_resource, lr_hash); + ldlm_resource_getref(res); + spin_unlock(&ns->ns_hash_lock); + + cleanup_resource(res, &res->lr_granted, flags); + cleanup_resource(res, &res->lr_converting, flags); + cleanup_resource(res, &res->lr_waiting, flags); + + 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. */ + if (!ldlm_resource_putref_locked(res)) + CDEBUG(D_INFO, + "Namespace %s resource refcount nonzero " + "(%d) after lock cleanup; forcing cleanup.\n", + ns->ns_name, atomic_read(&res->lr_refcount)); + } + spin_unlock(&ns->ns_hash_lock); + } + + return ELDLM_OK; +} + +/* Cleanup, but also free, the namespace */ +int ldlm_namespace_free_prior(struct ldlm_namespace *ns) +{ + ENTRY; if (!ns) RETURN(ELDLM_OK); - spin_lock(&ldlm_namespace_lock); + mutex_down(ldlm_namespace_lock(ns->ns_client)); list_del(&ns->ns_list_chain); - remove_proc_entry(ns->ns_name, ldlm_ns_proc_dir); - spin_unlock(&ldlm_namespace_lock); + atomic_dec(ldlm_namespace_nr(ns->ns_client)); + ldlm_pool_fini(&ns->ns_pool); + mutex_up(ldlm_namespace_lock(ns->ns_client)); + + /* At shutdown time, don't call the cancellation callback */ + ldlm_namespace_cleanup(ns, 0); + + if (ns->ns_refcount > 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); + rc = l_wait_event(ns->ns_waitq, + ns->ns_refcount == 0, &lwi); + if (ns->ns_refcount) + LCONSOLE_ERROR_MSG(0x139, "Lock manager: wait for %s " + "namespace cleanup aborted with %d " + "resources in use. (%d)\nI'm going " + "to try to clean up anyway, but I " + "might need a reboot of this node.\n", + ns->ns_name, (int) ns->ns_refcount, + rc); + CDEBUG(D_DLMTRACE, + "dlm namespace %s free done waiting\n", ns->ns_name); + } - l_lock(&ns->ns_lock); + RETURN(ELDLM_OK); +} - for (i = 0; i < RES_HASH_SIZE; i++) { - list_for_each_safe(tmp, pos, &(ns->ns_hash[i])) { - struct ldlm_resource *res; - res = list_entry(tmp, struct ldlm_resource, lr_hash); - ldlm_resource_getref(res); +int ldlm_namespace_free_post(struct ldlm_namespace *ns, int force) +{ + ENTRY; + if (!ns) + RETURN(ELDLM_OK); - cleanup_resource(res, &res->lr_granted); - cleanup_resource(res, &res->lr_converting); - cleanup_resource(res, &res->lr_waiting); - - if (!ldlm_resource_put(res)) { - CERROR("Resource refcount nonzero (%d) after " - "lock cleanup; forcing cleanup.\n", - atomic_read(&res->lr_refcount)); - ldlm_resource_dump(res); - atomic_set(&res->lr_refcount, 1); - ldlm_resource_put(res); - } +#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); } } - - vfree(ns->ns_hash /* , sizeof(struct list_head) * RES_HASH_SIZE */); - ptlrpc_cleanup_client(&ns->ns_rpc_client); +#endif + OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE); OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1); - OBD_FREE(ns, sizeof(*ns)); + OBD_FREE_PTR(ns); + ldlm_put_ref(force); + RETURN(ELDLM_OK); +} + +/* Cleanup the resource, and free namespace. + * bug 12864: + * Deadlock issue: + * proc1: destroy import + * class_disconnect_export(grab cl_sem) -> + * -> ldlm_namespace_free -> + * -> lprocfs_remove(grab _lprocfs_lock). + * proc2: read proc info + * lprocfs_fops_read(grab _lprocfs_lock) -> + * -> osc_rd_active, etc(grab cl_sem). + * + * So that I have to split the ldlm_namespace_free into two parts - the first + * part ldlm_namespace_free_prior is used to cleanup the resource which is + * being used; the 2nd part ldlm_namespace_free_post is used to unregister the + * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem + * held. + */ +int ldlm_namespace_free(struct ldlm_namespace *ns, int force) +{ + ldlm_namespace_free_prior(ns); + ldlm_namespace_free_post(ns, force); return ELDLM_OK; } -static __u32 ldlm_hash_fn(struct ldlm_resource *parent, __u64 *name) + +void ldlm_namespace_get_nolock(struct ldlm_namespace *ns) +{ + LASSERT(ns->ns_refcount >= 0); + ns->ns_refcount++; +} + +void ldlm_namespace_get(struct ldlm_namespace *ns) +{ + spin_lock(&ns->ns_hash_lock); + ldlm_namespace_get_nolock(ns); + spin_unlock(&ns->ns_hash_lock); +} + +void ldlm_namespace_put_nolock(struct ldlm_namespace *ns, int wakeup) +{ + LASSERT(ns->ns_refcount > 0); + ns->ns_refcount--; + if (ns->ns_refcount == 0 && wakeup) + wake_up(&ns->ns_waitq); +} + +void ldlm_namespace_put(struct ldlm_namespace *ns, int wakeup) +{ + spin_lock(&ns->ns_hash_lock); + ldlm_namespace_put_nolock(ns, wakeup); + spin_unlock(&ns->ns_hash_lock); +} + +/* Should be called under ldlm_namespace_lock(client) taken */ +void ldlm_namespace_move(struct ldlm_namespace *ns, ldlm_side_t client) +{ + LASSERT(!list_empty(&ns->ns_list_chain)); + LASSERT_SEM_LOCKED(ldlm_namespace_lock(client)); + list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client)); +} + +/* Should be called under ldlm_namespace_lock(client) taken */ +struct ldlm_namespace *ldlm_namespace_first(ldlm_side_t client) +{ + LASSERT_SEM_LOCKED(ldlm_namespace_lock(client)); + LASSERT(!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[i]; + hash += name->name[i]; hash += (__u32)((unsigned long)parent >> 4); @@ -191,61 +623,109 @@ static struct ldlm_resource *ldlm_resource_new(void) { struct ldlm_resource *res; - res = kmem_cache_alloc(ldlm_resource_slab, SLAB_KERNEL); - if (res == NULL) { - LBUG(); + OBD_SLAB_ALLOC(res, ldlm_resource_slab, CFS_ALLOC_IO, sizeof *res); + if (res == NULL) return NULL; - } - memset(res, 0, sizeof(*res)); - INIT_LIST_HEAD(&res->lr_children); - INIT_LIST_HEAD(&res->lr_childof); - INIT_LIST_HEAD(&res->lr_granted); - INIT_LIST_HEAD(&res->lr_converting); - INIT_LIST_HEAD(&res->lr_waiting); + 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); atomic_set(&res->lr_refcount, 1); + spin_lock_init(&res->lr_lock); + + /* one who creates the resource must unlock + * the semaphore after lvb initialization */ + init_MUTEX_LOCKED(&res->lr_lvb_sem); 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) +{ + struct list_head *bucket, *tmp; + struct ldlm_resource *res; + + LASSERT_SPIN_LOCKED(&ns->ns_hash_lock); + bucket = ns->ns_hash + hash; + + list_for_each(tmp, bucket) { + res = list_entry(tmp, struct ldlm_resource, lr_hash); + if (memcmp(&res->lr_name, name, sizeof(res->lr_name)) == 0) + return res; + } + + return NULL; +} + /* Args: locked namespace * Returns: newly-allocated, referenced, unlocked resource */ -static struct ldlm_resource *ldlm_resource_add(struct ldlm_namespace *ns, - struct ldlm_resource *parent, - __u64 *name, __u32 type) +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) { struct list_head *bucket; - struct ldlm_resource *res; + struct ldlm_resource *res, *old_res; ENTRY; - if (type < LDLM_MIN_TYPE || type > LDLM_MAX_TYPE) { - LBUG(); - RETURN(NULL); - } + LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE, + "type: %d\n", type); res = ldlm_resource_new(); - if (!res) { - LBUG(); + if (!res) RETURN(NULL); - } - memcpy(res->lr_name, name, sizeof(res->lr_name)); + res->lr_name = *name; res->lr_namespace = ns; - ns->ns_refcount++; - res->lr_type = type; res->lr_most_restr = LCK_NL; - bucket = ns->ns_hash + ldlm_hash_fn(parent, name); + 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); + spin_unlock(&ns->ns_hash_lock); + OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res); + /* synchronize WRT resource creation */ + if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) { + down(&old_res->lr_lvb_sem); + up(&old_res->lr_lvb_sem); + } + RETURN(old_res); + } + + /* we won! let's add the resource */ + bucket = ns->ns_hash + hash; list_add(&res->lr_hash, bucket); + ns->ns_resources++; + ldlm_namespace_get_nolock(ns); - if (parent == NULL) + if (parent == NULL) { list_add(&res->lr_childof, &ns->ns_root_list); - else { + } else { res->lr_parent = parent; list_add(&res->lr_childof, &parent->lr_children); } + spin_unlock(&ns->ns_hash_lock); + + if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) { + int rc; + + 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 */ + up(&res->lr_lvb_sem); + } RETURN(res); } @@ -253,89 +733,130 @@ static struct ldlm_resource *ldlm_resource_add(struct ldlm_namespace *ns, /* 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, - __u64 *name, __u32 type, int create) +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) { - struct list_head *bucket; - struct list_head *tmp = bucket; + __u32 hash = ldlm_hash_fn(parent, name); struct ldlm_resource *res = NULL; ENTRY; - if (ns == NULL || ns->ns_hash == NULL) { - LBUG(); - RETURN(NULL); - } - - l_lock(&ns->ns_lock); - bucket = ns->ns_hash + ldlm_hash_fn(parent, name); - - list_for_each(tmp, bucket) { - struct ldlm_resource *chk; - chk = list_entry(tmp, struct ldlm_resource, lr_hash); - - if (memcmp(chk->lr_name, name, sizeof(chk->lr_name)) == 0) { - res = chk; - atomic_inc(&res->lr_refcount); - EXIT; - break; + LASSERT(ns != NULL); + LASSERT(ns->ns_hash != NULL); + LASSERT(name->name[0] != 0); + + spin_lock(&ns->ns_hash_lock); + res = ldlm_resource_find(ns, name, hash); + if (res) { + ldlm_resource_getref(res); + spin_unlock(&ns->ns_hash_lock); + /* synchronize WRT resource creation */ + if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) { + down(&res->lr_lvb_sem); + up(&res->lr_lvb_sem); } + RETURN(res); } + spin_unlock(&ns->ns_hash_lock); - if (res == NULL && create) - res = ldlm_resource_add(ns, parent, name, type); - l_unlock(&ns->ns_lock); + if (create == 0) + RETURN(NULL); + res = ldlm_resource_add(ns, parent, name, hash, type); RETURN(res); } struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res) { + LASSERT(res != NULL); + LASSERT(res != LP_POISON); atomic_inc(&res->lr_refcount); + CDEBUG(D_INFO, "getref res: %p count: %d\n", res, + atomic_read(&res->lr_refcount)); return res; } -/* Returns 1 if the resource was freed, 0 if it remains. */ -int ldlm_resource_put(struct ldlm_resource *res) +void __ldlm_resource_putref_final(struct ldlm_resource *res) { - int rc = 0; + struct ldlm_namespace *ns = res->lr_namespace; - if (atomic_dec_and_test(&res->lr_refcount)) { - struct ldlm_namespace *ns = res->lr_namespace; - ENTRY; + LASSERT_SPIN_LOCKED(&ns->ns_hash_lock); - l_lock(&ns->ns_lock); + if (!list_empty(&res->lr_granted)) { + ldlm_resource_dump(D_ERROR, res); + LBUG(); + } - if (atomic_read(&res->lr_refcount) != 0) { - /* We lost the race. */ - l_unlock(&ns->ns_lock); - goto out; - } + if (!list_empty(&res->lr_converting)) { + ldlm_resource_dump(D_ERROR, res); + LBUG(); + } - if (!list_empty(&res->lr_granted)) - LBUG(); + if (!list_empty(&res->lr_waiting)) { + ldlm_resource_dump(D_ERROR, res); + LBUG(); + } - if (!list_empty(&res->lr_converting)) - LBUG(); + if (!list_empty(&res->lr_children)) { + ldlm_resource_dump(D_ERROR, res); + LBUG(); + } - if (!list_empty(&res->lr_waiting)) - 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_nolock(ns, 0); + list_del_init(&res->lr_hash); + list_del_init(&res->lr_childof); - if (!list_empty(&res->lr_children)) - LBUG(); + ns->ns_resources--; + if (ns->ns_resources == 0) + wake_up(&ns->ns_waitq); +} - ns->ns_refcount--; - list_del(&res->lr_hash); - list_del(&res->lr_childof); +/* 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; - kmem_cache_free(ldlm_resource_slab, res); - l_unlock(&ns->ns_lock); + CDEBUG(D_INFO, "putref res: %p count: %d\n", res, + atomic_read(&res->lr_refcount) - 1); + LASSERTF(atomic_read(&res->lr_refcount) > 0, "%d", + atomic_read(&res->lr_refcount)); + LASSERTF(atomic_read(&res->lr_refcount) < LI_POISON, "%d", + atomic_read(&res->lr_refcount)); + + if (atomic_dec_and_lock(&res->lr_refcount, &ns->ns_hash_lock)) { + __ldlm_resource_putref_final(res); + spin_unlock(&ns->ns_hash_lock); + if (res->lr_lvb_data) + OBD_FREE(res->lr_lvb_data, res->lr_lvb_len); + OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res); + rc = 1; + } + + RETURN(rc); +} + +/* Returns 1 if the resource was freed, 0 if it remains. */ +int ldlm_resource_putref_locked(struct ldlm_resource *res) +{ + int rc = 0; + ENTRY; + + CDEBUG(D_INFO, "putref res: %p count: %d\n", res, + atomic_read(&res->lr_refcount) - 1); + LASSERT(atomic_read(&res->lr_refcount) > 0); + LASSERT(atomic_read(&res->lr_refcount) < LI_POISON); + + LASSERT(atomic_read(&res->lr_refcount) >= 0); + if (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); + OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res); rc = 1; - } else { - ENTRY; - out: - if (atomic_read(&res->lr_refcount) < 0) - LBUG(); } RETURN(rc); @@ -344,103 +865,150 @@ int ldlm_resource_put(struct ldlm_resource *res) void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head, struct ldlm_lock *lock) { - l_lock(&res->lr_namespace->ns_lock); + check_res_locked(res); - ldlm_resource_dump(res); - ldlm_lock_dump(lock); + ldlm_resource_dump(D_OTHER, res); + CDEBUG(D_OTHER, "About to add this lock:\n"); + ldlm_lock_dump(D_OTHER, lock, 0); - if (!list_empty(&lock->l_res_link)) - LBUG(); + if (lock->l_destroyed) { + CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n"); + return; + } + + LASSERT(list_empty(&lock->l_res_link)); + + list_add_tail(&lock->l_res_link, head); +} + +void ldlm_resource_insert_lock_after(struct ldlm_lock *original, + struct ldlm_lock *new) +{ + struct ldlm_resource *res = original->l_resource; + + check_res_locked(res); + + ldlm_resource_dump(D_OTHER, res); + CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original); + ldlm_lock_dump(D_OTHER, new, 0); + + if (new->l_destroyed) { + CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n"); + goto out; + } - list_add(&lock->l_res_link, head); - l_unlock(&res->lr_namespace->ns_lock); + LASSERT(list_empty(&new->l_res_link)); + + list_add(&new->l_res_link, &original->l_res_link); + out:; } void ldlm_resource_unlink_lock(struct ldlm_lock *lock) { - l_lock(&lock->l_resource->lr_namespace->ns_lock); + check_res_locked(lock->l_resource); + ldlm_unlink_lock_skiplist(lock); list_del_init(&lock->l_res_link); - l_unlock(&lock->l_resource->lr_namespace->ns_lock); } void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc) { desc->lr_type = res->lr_type; - memcpy(desc->lr_name, res->lr_name, sizeof(desc->lr_name)); - memcpy(desc->lr_version, res->lr_version, sizeof(desc->lr_version)); + desc->lr_name = res->lr_name; } -void ldlm_dump_all_namespaces(void) +void ldlm_dump_all_namespaces(ldlm_side_t client, int level) { struct list_head *tmp; - spin_lock(&ldlm_namespace_lock); + if (!((libcfs_debug | D_ERROR) & level)) + return; + + mutex_down(ldlm_namespace_lock(client)); - list_for_each(tmp, &ldlm_namespace_list) { + list_for_each(tmp, ldlm_namespace_list(client)) { struct ldlm_namespace *ns; ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain); - ldlm_namespace_dump(ns); + ldlm_namespace_dump(level, ns); } - spin_unlock(&ldlm_namespace_lock); + mutex_up(ldlm_namespace_lock(client)); } -void ldlm_namespace_dump(struct ldlm_namespace *ns) +void ldlm_namespace_dump(int level, struct ldlm_namespace *ns) { struct list_head *tmp; - l_lock(&ns->ns_lock); - CDEBUG(D_OTHER, "--- Namespace: %s (rc: %d, client: %d)\n", ns->ns_name, - ns->ns_refcount, ns->ns_client); + if (!((libcfs_debug | D_ERROR) & level)) + return; + + CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n", + ns->ns_name, ns->ns_refcount, + ns_is_client(ns) ? "client" : "server"); + + if (cfs_time_before(cfs_time_current(), ns->ns_next_dump)) + return; - list_for_each(tmp, &ns->ns_root_list) { + spin_lock(&ns->ns_hash_lock); + tmp = ns->ns_root_list.next; + while (tmp != &ns->ns_root_list) { struct ldlm_resource *res; res = list_entry(tmp, struct ldlm_resource, lr_childof); - /* Once we have resources with children, this should really dump - * them recursively. */ - ldlm_resource_dump(res); + ldlm_resource_getref(res); + spin_unlock(&ns->ns_hash_lock); + + lock_res(res); + ldlm_resource_dump(level, res); + unlock_res(res); + + spin_lock(&ns->ns_hash_lock); + tmp = tmp->next; + ldlm_resource_putref_locked(res); } - l_unlock(&ns->ns_lock); + ns->ns_next_dump = cfs_time_shift(10); + spin_unlock(&ns->ns_hash_lock); } -void ldlm_resource_dump(struct ldlm_resource *res) +void ldlm_resource_dump(int level, struct ldlm_resource *res) { struct list_head *tmp; - char name[256]; + int pos; - if (RES_NAME_SIZE != 3) - LBUG(); + CLASSERT(RES_NAME_SIZE == 4); - snprintf(name, sizeof(name), "%Lx %Lx %Lx", - (unsigned long long)res->lr_name[0], - (unsigned long long)res->lr_name[1], - (unsigned long long)res->lr_name[2]); + if (!((libcfs_debug | D_ERROR) & level)) + return; - CDEBUG(D_OTHER, "--- Resource: %p (%s) (rc: %d)\n", res, name, + 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], atomic_read(&res->lr_refcount)); - CDEBUG(D_OTHER, "Namespace: %p (%s)\n", res->lr_namespace, - res->lr_namespace->ns_name); - CDEBUG(D_OTHER, "Parent: %p, root: %p\n", res->lr_parent, res->lr_root); - - CDEBUG(D_OTHER, "Granted locks:\n"); - list_for_each(tmp, &res->lr_granted) { - struct ldlm_lock *lock; - lock = list_entry(tmp, struct ldlm_lock, l_res_link); - ldlm_lock_dump(lock); - } - CDEBUG(D_OTHER, "Converting locks:\n"); - list_for_each(tmp, &res->lr_converting) { - struct ldlm_lock *lock; - lock = list_entry(tmp, struct ldlm_lock, l_res_link); - ldlm_lock_dump(lock); + if (!list_empty(&res->lr_granted)) { + pos = 0; + CDEBUG(level, "Granted locks:\n"); + list_for_each(tmp, &res->lr_granted) { + struct ldlm_lock *lock; + lock = list_entry(tmp, struct ldlm_lock, l_res_link); + ldlm_lock_dump(level, lock, ++pos); + } } - - CDEBUG(D_OTHER, "Waiting locks:\n"); - list_for_each(tmp, &res->lr_waiting) { - struct ldlm_lock *lock; - lock = list_entry(tmp, struct ldlm_lock, l_res_link); - ldlm_lock_dump(lock); + if (!list_empty(&res->lr_converting)) { + pos = 0; + CDEBUG(level, "Converting locks:\n"); + list_for_each(tmp, &res->lr_converting) { + struct ldlm_lock *lock; + lock = list_entry(tmp, struct ldlm_lock, l_res_link); + ldlm_lock_dump(level, lock, ++pos); + } + } + if (!list_empty(&res->lr_waiting)) { + pos = 0; + CDEBUG(level, "Waiting locks:\n"); + list_for_each(tmp, &res->lr_waiting) { + struct ldlm_lock *lock; + lock = list_entry(tmp, struct ldlm_lock, l_res_link); + ldlm_lock_dump(level, lock, ++pos); + } } }