X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_lockd.c;h=dd293e222764890f574240126f2a5d51545a82dc;hp=37feb38f8e4a0237cd605a82f45216fd39b87883;hb=82d014e71e14671e876055851a0d37e98b4cc079;hpb=97fe58c7ec2b5c88c4ecbc2e3597317792d985d6 diff --git a/lustre/ldlm/ldlm_lockd.c b/lustre/ldlm/ldlm_lockd.c index 37feb38..dd293e2 100644 --- a/lustre/ldlm/ldlm_lockd.c +++ b/lustre/ldlm/ldlm_lockd.c @@ -15,11 +15,7 @@ * * You should have received a copy of the GNU General Public License * version 2 along with this program; If not, see - * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf - * - * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, - * CA 95054 USA or visit www.sun.com if you need additional information or - * have any questions. + * http://www.gnu.org/licenses/gpl-2.0.html * * GPL HEADER END */ @@ -27,7 +23,7 @@ * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2010, 2015, Intel Corporation. + * Copyright (c) 2010, 2017, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -42,50 +38,49 @@ #define DEBUG_SUBSYSTEM S_LDLM #include +#include #include +#include #include #include -#include #include "ldlm_internal.h" static int ldlm_num_threads; module_param(ldlm_num_threads, int, 0444); MODULE_PARM_DESC(ldlm_num_threads, "number of DLM service threads to start"); +static unsigned int ldlm_cpu_bind = 1; +module_param(ldlm_cpu_bind, uint, 0444); +MODULE_PARM_DESC(ldlm_cpu_bind, + "bind DLM service threads to particular CPU partitions"); + static char *ldlm_cpts; module_param(ldlm_cpts, charp, 0444); MODULE_PARM_DESC(ldlm_cpts, "CPU partitions ldlm threads should run on"); -static struct mutex ldlm_ref_mutex; +static DEFINE_MUTEX(ldlm_ref_mutex); static int ldlm_refcount; -struct ldlm_cb_async_args { - struct ldlm_cb_set_arg *ca_set_arg; - struct ldlm_lock *ca_lock; -}; +struct kobject *ldlm_kobj; +struct kset *ldlm_ns_kset; +struct kset *ldlm_svc_kset; /* LDLM state */ static struct ldlm_state *ldlm_state; -static inline cfs_time_t round_timeout(cfs_time_t timeout) -{ - return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1); -} - -/* timeout for initial callback (AST) reply (bz10399) */ -static inline unsigned int ldlm_get_rq_timeout(void) +/* timeout for initial callback (AST) reply (bz10399) + * Due to having to send a 32 bit time value over the + * wire return it as time_t instead of time64_t + */ +static inline time_t ldlm_get_rq_timeout(void) { - /* Non-AT value */ - unsigned int timeout = min(ldlm_timeout, obd_timeout / 3); + /* Non-AT value */ + time_t timeout = min(ldlm_timeout, obd_timeout / 3); - return timeout < 1 ? 1 : timeout; + return timeout < 1 ? 1 : timeout; } -#define ELT_STOPPED 0 -#define ELT_READY 1 -#define ELT_TERMINATE 2 - struct ldlm_bl_pool { spinlock_t blp_lock; @@ -127,7 +122,7 @@ struct ldlm_bl_work_item { /** * Protects both waiting_locks_list and expired_lock_thread. */ -static spinlock_t waiting_locks_spinlock; /* BH lock (timer) */ +static DEFINE_SPINLOCK(waiting_locks_spinlock); /* BH lock (timer) */ /** * List for contended locks. @@ -140,15 +135,24 @@ static spinlock_t waiting_locks_spinlock; /* BH lock (timer) */ * * All access to it should be under waiting_locks_spinlock. */ -static struct list_head waiting_locks_list; -static struct timer_list waiting_locks_timer; +static LIST_HEAD(waiting_locks_list); +static void waiting_locks_callback(TIMER_DATA_TYPE unused); +static CFS_DEFINE_TIMER(waiting_locks_timer, waiting_locks_callback, 0, 0); + +enum elt_state { + ELT_STOPPED, + ELT_READY, + ELT_TERMINATE, +}; -static struct expired_lock_thread { - wait_queue_head_t elt_waitq; - int elt_state; - int elt_dump; - struct list_head elt_expired_locks; -} expired_lock_thread; +static DECLARE_WAIT_QUEUE_HEAD(expired_lock_wait_queue); +static enum elt_state expired_lock_thread_state = ELT_STOPPED; +static int expired_lock_dump; +static LIST_HEAD(expired_lock_list); + +static int ldlm_lock_busy(struct ldlm_lock *lock); +static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout); +static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds); static inline int have_expired_locks(void) { @@ -156,7 +160,7 @@ static inline int have_expired_locks(void) ENTRY; spin_lock_bh(&waiting_locks_spinlock); - need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks); + need_to_run = !list_empty(&expired_lock_list); spin_unlock_bh(&waiting_locks_spinlock); RETURN(need_to_run); @@ -167,35 +171,30 @@ static inline int have_expired_locks(void) */ static int expired_lock_main(void *arg) { - struct list_head *expired = &expired_lock_thread.elt_expired_locks; + struct list_head *expired = &expired_lock_list; struct l_wait_info lwi = { 0 }; int do_dump; ENTRY; - expired_lock_thread.elt_state = ELT_READY; - wake_up(&expired_lock_thread.elt_waitq); + expired_lock_thread_state = ELT_READY; + wake_up(&expired_lock_wait_queue); while (1) { - l_wait_event(expired_lock_thread.elt_waitq, + l_wait_event(expired_lock_wait_queue, have_expired_locks() || - expired_lock_thread.elt_state == ELT_TERMINATE, + expired_lock_thread_state == ELT_TERMINATE, &lwi); spin_lock_bh(&waiting_locks_spinlock); - if (expired_lock_thread.elt_dump) { - struct libcfs_debug_msg_data msgdata = { - .msg_file = __FILE__, - .msg_fn = "waiting_locks_callback", - .msg_line = expired_lock_thread.elt_dump }; + if (expired_lock_dump) { spin_unlock_bh(&waiting_locks_spinlock); /* from waiting_locks_callback, but not in timer */ libcfs_debug_dumplog(); - libcfs_run_lbug_upcall(&msgdata); spin_lock_bh(&waiting_locks_spinlock); - expired_lock_thread.elt_dump = 0; + expired_lock_dump = 0; } do_dump = 0; @@ -206,7 +205,7 @@ static int expired_lock_main(void *arg) lock = list_entry(expired->next, struct ldlm_lock, l_pending_chain); - if ((void *)lock < LP_POISON + PAGE_CACHE_SIZE && + if ((void *)lock < LP_POISON + PAGE_SIZE && (void *)lock >= LP_POISON) { spin_unlock_bh(&waiting_locks_spinlock); CERROR("free lock on elt list %p\n", lock); @@ -214,7 +213,7 @@ static int expired_lock_main(void *arg) } list_del_init(&lock->l_pending_chain); if ((void *)lock->l_export < - LP_POISON + PAGE_CACHE_SIZE && + LP_POISON + PAGE_SIZE && (void *)lock->l_export >= LP_POISON) { CERROR("lock with free export on elt list %p\n", lock->l_export); @@ -236,14 +235,30 @@ static int expired_lock_main(void *arg) export = class_export_lock_get(lock->l_export, lock); spin_unlock_bh(&waiting_locks_spinlock); - spin_lock_bh(&export->exp_bl_list_lock); - list_del_init(&lock->l_exp_list); - spin_unlock_bh(&export->exp_bl_list_lock); - - do_dump++; - class_fail_export(export); + /* Check if we need to prolong timeout */ + if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) && + lock->l_callback_timeout != 0 && /* not AST error */ + ldlm_lock_busy(lock)) { + LDLM_DEBUG(lock, "prolong the busy lock"); + lock_res_and_lock(lock); + ldlm_add_waiting_lock(lock, + ldlm_bl_timeout(lock) >> 1); + unlock_res_and_lock(lock); + } else { + spin_lock_bh(&export->exp_bl_list_lock); + list_del_init(&lock->l_exp_list); + spin_unlock_bh(&export->exp_bl_list_lock); + + LDLM_ERROR(lock, + "lock callback timer expired after %llds: evicting client at %s ", + ktime_get_real_seconds() - + lock->l_blast_sent, + obd_export_nid2str(export)); + ldlm_lock_to_ns(lock)->ns_timeouts++; + do_dump++; + class_fail_export(export); + } class_export_lock_put(export, lock); - /* release extra ref grabbed by ldlm_add_waiting_lock() * or ldlm_failed_ast() */ LDLM_LOCK_RELEASE(lock); @@ -257,18 +272,15 @@ static int expired_lock_main(void *arg) libcfs_debug_dumplog(); } - if (expired_lock_thread.elt_state == ELT_TERMINATE) + if (expired_lock_thread_state == ELT_TERMINATE) break; } - expired_lock_thread.elt_state = ELT_STOPPED; - wake_up(&expired_lock_thread.elt_waitq); + expired_lock_thread_state = ELT_STOPPED; + wake_up(&expired_lock_wait_queue); RETURN(0); } -static int ldlm_add_waiting_lock(struct ldlm_lock *lock); -static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds); - /** * Check if there is a request in the export request list * which prevents the lock canceling. @@ -282,7 +294,7 @@ static int ldlm_lock_busy(struct ldlm_lock *lock) if (lock->l_export == NULL) return 0; - spin_lock_bh(&lock->l_export->exp_rpc_lock); + spin_lock(&lock->l_export->exp_rpc_lock); list_for_each_entry(req, &lock->l_export->exp_hp_rpcs, rq_exp_list) { if (req->rq_ops->hpreq_lock_match) { @@ -291,12 +303,12 @@ static int ldlm_lock_busy(struct ldlm_lock *lock) break; } } - spin_unlock_bh(&lock->l_export->exp_rpc_lock); + spin_unlock(&lock->l_export->exp_rpc_lock); RETURN(match); } /* This is called from within a timer interrupt and cannot schedule */ -static void waiting_locks_callback(unsigned long unused) +static void waiting_locks_callback(TIMER_DATA_TYPE unused) { struct ldlm_lock *lock; int need_dump = 0; @@ -304,70 +316,38 @@ static void waiting_locks_callback(unsigned long unused) spin_lock_bh(&waiting_locks_spinlock); while (!list_empty(&waiting_locks_list)) { lock = list_entry(waiting_locks_list.next, struct ldlm_lock, - l_pending_chain); - if (cfs_time_after(lock->l_callback_timeout, - cfs_time_current()) || - (lock->l_req_mode == LCK_GROUP)) - break; - - /* Check if we need to prolong timeout */ - if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) && - ldlm_lock_busy(lock)) { - int cont = 1; - - if (lock->l_pending_chain.next == &waiting_locks_list) - cont = 0; - - LDLM_LOCK_GET(lock); - - spin_unlock_bh(&waiting_locks_spinlock); - LDLM_DEBUG(lock, "prolong the busy lock"); - ldlm_refresh_waiting_lock(lock, - ldlm_bl_timeout(lock) >> 1); - spin_lock_bh(&waiting_locks_spinlock); - - if (!cont) { - LDLM_LOCK_RELEASE(lock); - break; - } - - LDLM_LOCK_RELEASE(lock); - continue; - } - ldlm_lock_to_ns(lock)->ns_timeouts++; - LDLM_ERROR(lock, "lock callback timer expired after %lds: " - "evicting client at %s ", - cfs_time_current_sec() - lock->l_last_activity, - libcfs_nid2str( - lock->l_export->exp_connection->c_peer.nid)); + l_pending_chain); + if (lock->l_callback_timeout > ktime_get_seconds() || + lock->l_req_mode == LCK_GROUP) + break; /* no needs to take an extra ref on the lock since it was in * the waiting_locks_list and ldlm_add_waiting_lock() * already grabbed a ref */ list_del(&lock->l_pending_chain); - list_add(&lock->l_pending_chain, - &expired_lock_thread.elt_expired_locks); + list_add(&lock->l_pending_chain, &expired_lock_list); need_dump = 1; } - if (!list_empty(&expired_lock_thread.elt_expired_locks)) { + if (!list_empty(&expired_lock_list)) { if (obd_dump_on_timeout && need_dump) - expired_lock_thread.elt_dump = __LINE__; + expired_lock_dump = __LINE__; - wake_up(&expired_lock_thread.elt_waitq); + wake_up(&expired_lock_wait_queue); } - /* - * Make sure the timer will fire again if we have any locks - * left. - */ + /* + * Make sure the timer will fire again if we have any locks + * left. + */ if (!list_empty(&waiting_locks_list)) { - cfs_time_t timeout_rounded; + unsigned long timeout_jiffies; + lock = list_entry(waiting_locks_list.next, struct ldlm_lock, - l_pending_chain); - timeout_rounded = (cfs_time_t)round_timeout(lock->l_callback_timeout); - cfs_timer_arm(&waiting_locks_timer, timeout_rounded); - } + l_pending_chain); + timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout); + mod_timer(&waiting_locks_timer, timeout_jiffies); + } spin_unlock_bh(&waiting_locks_spinlock); } @@ -383,10 +363,10 @@ static void waiting_locks_callback(unsigned long unused) * * Called with the namespace lock held. */ -static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds) +static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds) { - cfs_time_t timeout; - cfs_time_t timeout_rounded; + unsigned long timeout_jiffies; + time64_t timeout; if (!list_empty(&lock->l_pending_chain)) return 0; @@ -395,22 +375,22 @@ static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds) OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT)) seconds = 1; - timeout = cfs_time_shift(seconds); - if (likely(cfs_time_after(timeout, lock->l_callback_timeout))) + timeout = ktime_get_seconds() + seconds; + if (likely(timeout > lock->l_callback_timeout)) lock->l_callback_timeout = timeout; - timeout_rounded = round_timeout(lock->l_callback_timeout); + timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout); - if (cfs_time_before(timeout_rounded, - cfs_timer_deadline(&waiting_locks_timer)) || - !cfs_timer_is_armed(&waiting_locks_timer)) { - cfs_timer_arm(&waiting_locks_timer, timeout_rounded); - } - /* if the new lock has a shorter timeout than something earlier on - the list, we'll wait the longer amount of time; no big deal. */ - /* FIFO */ + if (time_before(timeout_jiffies, waiting_locks_timer.expires) || + !timer_pending(&waiting_locks_timer)) + mod_timer(&waiting_locks_timer, timeout_jiffies); + + /* if the new lock has a shorter timeout than something earlier on + * the list, we'll wait the longer amount of time; no big deal. + */ + /* FIFO */ list_add_tail(&lock->l_pending_chain, &waiting_locks_list); - return 1; + return 1; } static void ldlm_add_blocked_lock(struct ldlm_lock *lock) @@ -435,10 +415,9 @@ static void ldlm_add_blocked_lock(struct ldlm_lock *lock) obd_stale_export_adjust(lock->l_export); } -static int ldlm_add_waiting_lock(struct ldlm_lock *lock) +static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout) { int ret; - int timeout = ldlm_bl_timeout(lock); /* NB: must be called with hold of lock_res_and_lock() */ LASSERT(ldlm_is_res_locked(lock)); @@ -457,19 +436,19 @@ static int ldlm_add_waiting_lock(struct ldlm_lock *lock) } if (ldlm_is_destroyed(lock)) { - static cfs_time_t next; + static time64_t next; spin_unlock_bh(&waiting_locks_spinlock); LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)"); - if (cfs_time_after(cfs_time_current(), next)) { - next = cfs_time_shift(14400); + if (ktime_get_seconds() > next) { + next = ktime_get_seconds() + 14400; libcfs_debug_dumpstack(NULL); } return 0; } ldlm_set_waited(lock); - lock->l_last_activity = cfs_time_current_sec(); + lock->l_blast_sent = ktime_get_real_seconds(); ret = __ldlm_add_waiting_lock(lock, timeout); if (ret) { /* grab ref on the lock if it has been added to the @@ -481,7 +460,7 @@ static int ldlm_add_waiting_lock(struct ldlm_lock *lock) if (ret) ldlm_add_blocked_lock(lock); - LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)", + LDLM_DEBUG(lock, "%sadding to wait list(timeout: %lld, AT: %s)", ret == 0 ? "not re-" : "", timeout, AT_OFF ? "off" : "on"); return ret; @@ -508,13 +487,14 @@ static int __ldlm_del_waiting_lock(struct ldlm_lock *lock) /* Removing the head of the list, adjust timer. */ if (list_next == &waiting_locks_list) { /* No more, just cancel. */ - cfs_timer_disarm(&waiting_locks_timer); + del_timer(&waiting_locks_timer); } else { struct ldlm_lock *next; + next = list_entry(list_next, struct ldlm_lock, - l_pending_chain); - cfs_timer_arm(&waiting_locks_timer, - round_timeout(next->l_callback_timeout)); + l_pending_chain); + mod_timer(&waiting_locks_timer, + cfs_time_seconds(next->l_callback_timeout)); } } list_del_init(&lock->l_pending_chain); @@ -557,7 +537,7 @@ int ldlm_del_waiting_lock(struct ldlm_lock *lock) * * Called with namespace lock held. */ -int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout) +int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout) { if (lock->l_export == NULL) { /* We don't have a "waiting locks list" on clients. */ @@ -597,7 +577,7 @@ int ldlm_del_waiting_lock(struct ldlm_lock *lock) RETURN(0); } -int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout) +int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout) { RETURN(0); } @@ -615,9 +595,9 @@ int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout) * * \retval timeout in seconds to wait for the client reply */ -unsigned int ldlm_bl_timeout(struct ldlm_lock *lock) +time64_t ldlm_bl_timeout(struct ldlm_lock *lock) { - unsigned int timeout; + time64_t timeout; if (AT_OFF) return obd_timeout / 2; @@ -627,7 +607,7 @@ unsigned int ldlm_bl_timeout(struct ldlm_lock *lock) * It would be nice to have some kind of "early reply" mechanism for * lock callbacks too... */ timeout = at_get(&lock->l_export->exp_bl_lock_at); - return max(timeout + (timeout >> 1), ldlm_enqueue_min); + return max(timeout + (timeout >> 1), (time64_t)ldlm_enqueue_min); } EXPORT_SYMBOL(ldlm_bl_timeout); @@ -649,41 +629,54 @@ static void ldlm_failed_ast(struct ldlm_lock *lock, int rc, /* the lock was not in any list, grab an extra ref before adding * the lock to the expired list */ LDLM_LOCK_GET(lock); - list_add(&lock->l_pending_chain, - &expired_lock_thread.elt_expired_locks); - wake_up(&expired_lock_thread.elt_waitq); + lock->l_callback_timeout = 0; /* differentiate it from expired locks */ + list_add(&lock->l_pending_chain, &expired_lock_list); + wake_up(&expired_lock_wait_queue); spin_unlock_bh(&waiting_locks_spinlock); } /** * Perform lock cleanup if AST reply came with error. */ -static int ldlm_handle_ast_error(struct ldlm_lock *lock, - struct ptlrpc_request *req, int rc, - const char *ast_type) +static int ldlm_handle_ast_error(const struct lu_env *env, + struct ldlm_lock *lock, + struct ptlrpc_request *req, int rc, + const char *ast_type) { - lnet_process_id_t peer = req->rq_import->imp_connection->c_peer; + struct lnet_process_id peer = req->rq_import->imp_connection->c_peer; if (!req->rq_replied || (rc && rc != -EINVAL)) { - if (lock->l_export && lock->l_export->exp_libclient) { - LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)" - " timeout, just cancelling lock", ast_type, + if (ldlm_is_cancel(lock)) { + LDLM_DEBUG(lock, + "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)", + ast_type, req, req->rq_xid, libcfs_nid2str(peer.nid)); ldlm_lock_cancel(lock); rc = -ERESTART; - } else if (ldlm_is_cancel(lock)) { - LDLM_DEBUG(lock, "%s AST timeout from nid %s, but " - "cancel was received (AST reply lost?)", - ast_type, libcfs_nid2str(peer.nid)); - ldlm_lock_cancel(lock); - rc = -ERESTART; + } else if (rc == -ENODEV || rc == -ESHUTDOWN || + (rc == -EIO && + req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) { + /* Upon umount process the AST fails because cannot be + * sent. This shouldn't lead to the client eviction. + * -ENODEV error is returned by ptl_send_rpc() for + * new request in such import. + * -SHUTDOWN is returned by ptlrpc_import_delay_req() + * if imp_invalid is set or obd_no_recov. + * Meanwhile there is also check for LUSTRE_IMP_CLOSED + * in ptlrpc_import_delay_req() as well with -EIO code. + * In all such cases errors are ignored. + */ + LDLM_DEBUG(lock, "%s AST can't be sent due to a server" + " %s failure or umount process: rc = %d\n", + ast_type, + req->rq_import->imp_obd->obd_name, rc); } else { - LDLM_ERROR(lock, "client (nid %s) %s %s AST " - "(req status %d rc %d), evict it", + LDLM_ERROR(lock, + "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it", libcfs_nid2str(peer.nid), req->rq_replied ? "returned error from" : "failed to reply to", - ast_type, + ast_type, req, req->rq_xid, (req->rq_repmsg != NULL) ? lustre_msg_get_status(req->rq_repmsg) : 0, rc); @@ -695,17 +688,17 @@ static int ldlm_handle_ast_error(struct ldlm_lock *lock, if (rc == -EINVAL) { struct ldlm_resource *res = lock->l_resource; - LDLM_DEBUG(lock, "client (nid %s) returned %d" - " from %s AST - normal race", + LDLM_DEBUG(lock, + "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race", libcfs_nid2str(peer.nid), req->rq_repmsg ? lustre_msg_get_status(req->rq_repmsg) : -1, - ast_type); + ast_type, req, req->rq_xid); if (res) { /* update lvbo to return proper attributes. * see bug 23174 */ ldlm_resource_getref(res); - ldlm_res_lvbo_update(res, NULL, 1); + ldlm_lvbo_update(env, res, lock, NULL, 1); ldlm_resource_putref(res); } ldlm_lock_cancel(lock); @@ -716,14 +709,15 @@ static int ldlm_handle_ast_error(struct ldlm_lock *lock, } static int ldlm_cb_interpret(const struct lu_env *env, - struct ptlrpc_request *req, void *data, int rc) + struct ptlrpc_request *req, void *args, int rc) { - struct ldlm_cb_async_args *ca = data; - struct ldlm_lock *lock = ca->ca_lock; - struct ldlm_cb_set_arg *arg = ca->ca_set_arg; - ENTRY; + struct ldlm_cb_async_args *ca = args; + struct ldlm_lock *lock = ca->ca_lock; + struct ldlm_cb_set_arg *arg = ca->ca_set_arg; - LASSERT(lock != NULL); + ENTRY; + + LASSERT(lock != NULL); switch (arg->type) { case LDLM_GL_CALLBACK: @@ -735,23 +729,29 @@ static int ldlm_cb_interpret(const struct lu_env *env, * - Glimpse callback of remote lock might return * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274 */ - if (rc == -ELDLM_NO_LOCK_DATA) { + if (unlikely(arg->gl_interpret_reply)) { + rc = arg->gl_interpret_reply(env, req, args, rc); + } else if (rc == -ELDLM_NO_LOCK_DATA) { LDLM_DEBUG(lock, "lost race - client has a lock but no " "inode"); - ldlm_res_lvbo_update(lock->l_resource, NULL, 1); + ldlm_lvbo_update(env, lock->l_resource, lock, NULL, 1); } else if (rc != 0) { - rc = ldlm_handle_ast_error(lock, req, rc, "glimpse"); + rc = ldlm_handle_ast_error(env, lock, req, + rc, "glimpse"); } else { - rc = ldlm_res_lvbo_update(lock->l_resource, req, 1); + rc = ldlm_lvbo_update(env, lock->l_resource, + lock, req, 1); } break; case LDLM_BL_CALLBACK: if (rc != 0) - rc = ldlm_handle_ast_error(lock, req, rc, "blocking"); + rc = ldlm_handle_ast_error(env, lock, req, + rc, "blocking"); break; case LDLM_CP_CALLBACK: if (rc != 0) - rc = ldlm_handle_ast_error(lock, req, rc, "completion"); + rc = ldlm_handle_ast_error(env, lock, req, + rc, "completion"); break; default: LDLM_ERROR(lock, "invalid opcode for lock callback %d", @@ -770,8 +770,8 @@ static int ldlm_cb_interpret(const struct lu_env *env, static void ldlm_update_resend(struct ptlrpc_request *req, void *data) { - struct ldlm_cb_async_args *ca = data; - struct ldlm_lock *lock = ca->ca_lock; + struct ldlm_cb_async_args *ca = data; + struct ldlm_lock *lock = ca->ca_lock; ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock)); } @@ -811,7 +811,7 @@ static void ldlm_lock_reorder_req(struct ldlm_lock *lock) RETURN_EXIT; } - spin_lock_bh(&lock->l_export->exp_rpc_lock); + spin_lock(&lock->l_export->exp_rpc_lock); list_for_each_entry(req, &lock->l_export->exp_hp_rpcs, rq_exp_list) { /* Do not process requests that were not yet added to there @@ -825,7 +825,7 @@ static void ldlm_lock_reorder_req(struct ldlm_lock *lock) req->rq_ops->hpreq_lock_match(req, lock)) ptlrpc_nrs_req_hp_move(req); } - spin_unlock_bh(&lock->l_export->exp_rpc_lock); + spin_unlock(&lock->l_export->exp_rpc_lock); EXIT; } @@ -852,6 +852,11 @@ int ldlm_server_blocking_ast(struct ldlm_lock *lock, /* Don't need to do anything here. */ RETURN(0); + if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) { + LDLM_DEBUG(lock, "dropping BL AST"); + RETURN(0); + } + LASSERT(lock); LASSERT(data != NULL); if (lock->l_export->exp_obd->obd_recovering != 0) @@ -859,18 +864,18 @@ int ldlm_server_blocking_ast(struct ldlm_lock *lock, ldlm_lock_reorder_req(lock); - req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse, - &RQF_LDLM_BL_CALLBACK, - LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK); - if (req == NULL) - RETURN(-ENOMEM); + req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse, + &RQF_LDLM_BL_CALLBACK, + LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK); + if (req == NULL) + RETURN(-ENOMEM); - CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args)); - ca = ptlrpc_req_async_args(req); - ca->ca_set_arg = arg; - ca->ca_lock = lock; + CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args)); + ca = ptlrpc_req_async_args(req); + ca->ca_set_arg = arg; + ca->ca_lock = lock; - req->rq_interpret_reply = ldlm_cb_interpret; + req->rq_interpret_reply = ldlm_cb_interpret; lock_res_and_lock(lock); if (ldlm_is_destroyed(lock)) { @@ -911,7 +916,7 @@ int ldlm_server_blocking_ast(struct ldlm_lock *lock, req->rq_no_resend = 1; } else { LASSERT(lock->l_granted_mode == lock->l_req_mode); - ldlm_add_waiting_lock(lock); + ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock)); unlock_res_and_lock(lock); /* Do not resend after lock callback timeout */ @@ -924,8 +929,6 @@ int ldlm_server_blocking_ast(struct ldlm_lock *lock, if (AT_OFF) req->rq_timeout = ldlm_get_rq_timeout(); - lock->l_last_activity = cfs_time_current_sec(); - if (lock->l_export && lock->l_export->exp_nid_stats && lock->l_export->exp_nid_stats->nid_ldlm_stats) lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats, @@ -977,27 +980,31 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data) lvb_len = 0; req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len); - rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK); - if (rc) { - ptlrpc_request_free(req); - RETURN(rc); - } + rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK); + if (rc) { + ptlrpc_request_free(req); + RETURN(rc); + } - CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args)); - ca = ptlrpc_req_async_args(req); - ca->ca_set_arg = arg; - ca->ca_lock = lock; + CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args)); + ca = ptlrpc_req_async_args(req); + ca->ca_set_arg = arg; + ca->ca_lock = lock; - req->rq_interpret_reply = ldlm_cb_interpret; - body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); + req->rq_interpret_reply = ldlm_cb_interpret; + body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); - body->lock_handle[0] = lock->l_remote_handle; + body->lock_handle[0] = lock->l_remote_handle; body->lock_flags = ldlm_flags_to_wire(flags); ldlm_lock2desc(lock, &body->lock_desc); if (lvb_len > 0) { void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB); + const struct lu_env *env = NULL; + + if (req->rq_svc_thread) + env = req->rq_svc_thread->t_env; - lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len); + lvb_len = ldlm_lvbo_fill(env, lock, lvb, &lvb_len); if (lvb_len < 0) { /* We still need to send the RPC to wake up the blocked * enqueue thread on the client. @@ -1014,8 +1021,6 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data) } } - lock->l_last_activity = cfs_time_current_sec(); - LDLM_DEBUG(lock, "server preparing completion AST"); ptlrpc_request_set_replen(req); @@ -1049,7 +1054,7 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data) lock_res_and_lock(lock); } else { /* start the lock-timeout clock */ - ldlm_add_waiting_lock(lock); + ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock)); /* Do not resend after lock callback timeout */ req->rq_delay_limit = ldlm_bl_timeout(lock); req->rq_resend_cb = ldlm_update_resend; @@ -1087,7 +1092,7 @@ int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data) if (arg->gl_desc != NULL) /* There is a glimpse descriptor to pack */ - req_fmt = &RQF_LDLM_GL_DESC_CALLBACK; + req_fmt = &RQF_LDLM_GL_CALLBACK_DESC; else req_fmt = &RQF_LDLM_GL_CALLBACK; @@ -1105,38 +1110,39 @@ int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data) *desc = *arg->gl_desc; } - body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); - body->lock_handle[0] = lock->l_remote_handle; - ldlm_lock2desc(lock, &body->lock_desc); + body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); + body->lock_handle[0] = lock->l_remote_handle; + ldlm_lock2desc(lock, &body->lock_desc); CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args)); ca = ptlrpc_req_async_args(req); ca->ca_set_arg = arg; ca->ca_lock = lock; - /* server namespace, doesn't need lock */ - req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, - ldlm_lvbo_size(lock)); - ptlrpc_request_set_replen(req); - - req->rq_send_state = LUSTRE_IMP_FULL; - /* ptlrpc_request_alloc_pack already set timeout */ - if (AT_OFF) - req->rq_timeout = ldlm_get_rq_timeout(); + /* server namespace, doesn't need lock */ + req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, + ldlm_lvbo_size(lock)); + ptlrpc_request_set_replen(req); - lock->l_last_activity = cfs_time_current_sec(); + req->rq_send_state = LUSTRE_IMP_FULL; + /* ptlrpc_request_alloc_pack already set timeout */ + if (AT_OFF) + req->rq_timeout = ldlm_get_rq_timeout(); req->rq_interpret_reply = ldlm_cb_interpret; - if (lock->l_export && lock->l_export->exp_nid_stats && - lock->l_export->exp_nid_stats->nid_ldlm_stats) - lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats, - LDLM_GL_CALLBACK - LDLM_FIRST_OPC); + if (lock->l_export && lock->l_export->exp_nid_stats) { + struct nid_stat *nid_stats = lock->l_export->exp_nid_stats; + + lprocfs_counter_incr(nid_stats->nid_ldlm_stats, + LDLM_GL_CALLBACK - LDLM_FIRST_OPC); + } rc = ldlm_ast_fini(req, arg, lock, 0); RETURN(rc); } +EXPORT_SYMBOL(ldlm_server_glimpse_ast); int ldlm_glimpse_locks(struct ldlm_resource *res, struct list_head *gl_work_list) @@ -1169,40 +1175,6 @@ struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req) } EXPORT_SYMBOL(ldlm_request_lock); -static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req, - struct lprocfs_stats *srv_stats) -{ - int lock_type = 0, op = 0; - - lock_type = dlm_req->lock_desc.l_resource.lr_type; - - switch (lock_type) { - case LDLM_PLAIN: - op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE; - break; - case LDLM_EXTENT: - if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) - op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE; - else - op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE; - break; - case LDLM_FLOCK: - op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE; - break; - case LDLM_IBITS: - op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE; - break; - default: - op = 0; - break; - } - - if (op) - lprocfs_counter_incr(srv_stats, op); - - return; -} - /** * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc * service threads to carry out client lock enqueueing requests. @@ -1219,6 +1191,7 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns, void *cookie = NULL; int rc = 0; struct ldlm_resource *res = NULL; + const struct lu_env *env = req->rq_svc_thread->t_env; ENTRY; LDLM_DEBUG_NOLOCK("server-side enqueue handler START"); @@ -1228,7 +1201,9 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns, LASSERT(req->rq_export); - if (ptlrpc_req2svc(req)->srv_stats != NULL) + /* for intent enqueue the stat will be updated inside intent policy */ + if (ptlrpc_req2svc(req)->srv_stats != NULL && + !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT)) ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats); if (req->rq_export && req->rq_export->exp_nid_stats && @@ -1252,37 +1227,6 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns, GOTO(out, rc = -EFAULT); } - if (exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) { - if (unlikely(dlm_req->lock_desc.l_resource.lr_type == - LDLM_PLAIN)) { - DEBUG_REQ(D_ERROR, req, - "PLAIN lock request from IBITS client?"); - GOTO(out, rc = -EPROTO); - } - } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type == - LDLM_IBITS)) { - DEBUG_REQ(D_ERROR, req, - "IBITS lock request from unaware client?"); - GOTO(out, rc = -EPROTO); - } - -#if 0 - /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check - against server's _CONNECT_SUPPORTED flags? (I don't want to use - ibits for mgc/mgs) */ - - /* INODEBITS_INTEROP: Perform conversion from plain lock to - * inodebits lock if client does not support them. */ - if (!(exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) && - (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) { - dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS; - dlm_req->lock_desc.l_policy_data.l_inodebits.bits = - MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE; - if (dlm_req->lock_desc.l_req_mode == LCK_PR) - dlm_req->lock_desc.l_req_mode = LCK_CR; - } -#endif - if (unlikely((flags & LDLM_FL_REPLAY) || (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) { /* Find an existing lock in the per-export lock hash */ @@ -1292,8 +1236,8 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns, lock = cfs_hash_lookup(req->rq_export->exp_lock_hash, (void *)&dlm_req->lock_handle[0]); if (lock != NULL) { - DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie " - LPX64, lock->l_handle.h_cookie); + DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie %#llx", + lock->l_handle.h_cookie); flags |= LDLM_FL_RESENT; GOTO(existing_lock, rc = 0); } @@ -1328,7 +1272,7 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns, res = lock->l_resource; if (!(flags & LDLM_FL_REPLAY)) { /* non-replayed lock, delayed lvb init may need to be done */ - rc = ldlm_lvbo_init(res); + rc = ldlm_lvbo_init(env, res); if (rc < 0) { LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc); GOTO(out, rc); @@ -1356,8 +1300,15 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns, * without them. */ lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags & LDLM_FL_INHERIT_MASK); -existing_lock: + ldlm_convert_policy_to_local(req->rq_export, + dlm_req->lock_desc.l_resource.lr_type, + &dlm_req->lock_desc.l_policy_data, + &lock->l_policy_data); + if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT) + lock->l_req_extent = lock->l_policy_data.l_extent; + +existing_lock: if (flags & LDLM_FL_HAS_INTENT) { /* In this case, the reply buffer is allocated deep in * local_lock_enqueue by the policy function. */ @@ -1369,33 +1320,25 @@ existing_lock: req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, ldlm_lvbo_size(lock)); - if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR)) - GOTO(out, rc = -ENOMEM); - - rc = req_capsule_server_pack(&req->rq_pill); - if (rc) - GOTO(out, rc); - } + if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR)) + GOTO(out, rc = -ENOMEM); - if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN) - ldlm_convert_policy_to_local(req->rq_export, - dlm_req->lock_desc.l_resource.lr_type, - &dlm_req->lock_desc.l_policy_data, - &lock->l_policy_data); - if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT) - lock->l_req_extent = lock->l_policy_data.l_extent; + rc = req_capsule_server_pack(&req->rq_pill); + if (rc) + GOTO(out, rc); + } - err = ldlm_lock_enqueue(ns, &lock, cookie, &flags); + err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags); if (err) { if ((int)err < 0) rc = (int)err; GOTO(out, err); } - dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP); + dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP); - ldlm_lock2desc(lock, &dlm_rep->lock_desc); - ldlm_lock2handle(lock, &dlm_rep->lock_handle); + ldlm_lock2desc(lock, &dlm_rep->lock_desc); + ldlm_lock2handle(lock, &dlm_rep->lock_handle); if (lock && lock->l_resource->lr_type == LDLM_EXTENT) OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6); @@ -1417,6 +1360,22 @@ existing_lock: LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export); rc = -ENOTCONN; } else if (ldlm_is_ast_sent(lock)) { + /* fill lock desc for possible lock convert */ + if (lock->l_blocking_lock && + lock->l_resource->lr_type == LDLM_IBITS) { + struct ldlm_lock *bl_lock = lock->l_blocking_lock; + struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc; + + LDLM_DEBUG(lock, + "save blocking bits %llx in granted lock", + bl_lock->l_policy_data.l_inodebits.bits); + /* If lock is blocked then save blocking ibits + * in returned lock policy for the possible lock + * convert on a client. + */ + rep_desc->l_policy_data.l_inodebits.cancel_bits = + bl_lock->l_policy_data.l_inodebits.bits; + } dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT); if (lock->l_granted_mode == lock->l_req_mode) { /* @@ -1430,38 +1389,15 @@ existing_lock: unlock_res_and_lock(lock); ldlm_lock_cancel(lock); lock_res_and_lock(lock); - } else - ldlm_add_waiting_lock(lock); - } - } - /* Make sure we never ever grant usual metadata locks to liblustre - clients */ - if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN || - dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) && - req->rq_export->exp_libclient) { - if (unlikely(!ldlm_is_cancel_on_block(lock) || - !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){ - CERROR("Granting sync lock to libclient. " - "req fl %d, rep fl %d, lock fl "LPX64"\n", - dlm_req->lock_flags, dlm_rep->lock_flags, - lock->l_flags); - LDLM_ERROR(lock, "sync lock"); - if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) { - struct ldlm_intent *it; - - it = req_capsule_client_get(&req->rq_pill, - &RMF_LDLM_INTENT); - if (it != NULL) { - CERROR("This is intent %s ("LPU64")\n", - ldlm_it2str(it->opc), it->opc); - } + } else { + ldlm_add_waiting_lock(lock, + ldlm_bl_timeout(lock)); } } } + unlock_res_and_lock(lock); - unlock_res_and_lock(lock); - - EXIT; + EXIT; out: req->rq_status = rc ?: err; /* return either error - bug 11190 */ if (!req->rq_packed_final) { @@ -1476,43 +1412,59 @@ existing_lock: LDLM_DEBUG(lock, "server-side enqueue handler, sending reply" "(err=%d, rc=%d)", err, rc); - if (rc == 0) { - if (req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB, - RCL_SERVER) && - ldlm_lvbo_size(lock) > 0) { - void *buf; - int buflen; - - buf = req_capsule_server_get(&req->rq_pill, - &RMF_DLM_LVB); - LASSERTF(buf != NULL, "req %p, lock %p\n", - req, lock); - buflen = req_capsule_get_size(&req->rq_pill, - &RMF_DLM_LVB, RCL_SERVER); - /* non-replayed lock, delayed lvb init may - * need to be occur now */ - if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) { - buflen = ldlm_lvbo_fill(lock, buf, - buflen); - if (buflen >= 0) - req_capsule_shrink( + if (rc == 0 && + req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB, + RCL_SERVER) && + ldlm_lvbo_size(lock) > 0) { + void *buf; + int buflen; + +retry: + buf = req_capsule_server_get(&req->rq_pill, + &RMF_DLM_LVB); + LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock); + buflen = req_capsule_get_size(&req->rq_pill, + &RMF_DLM_LVB, RCL_SERVER); + /* non-replayed lock, delayed lvb init may + * need to be occur now + */ + if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) { + int rc2; + + rc2 = ldlm_lvbo_fill(env, lock, buf, &buflen); + if (rc2 >= 0) { + req_capsule_shrink(&req->rq_pill, + &RMF_DLM_LVB, + rc2, RCL_SERVER); + } else if (rc2 == -ERANGE) { + rc2 = req_capsule_server_grow( &req->rq_pill, - &RMF_DLM_LVB, - buflen, RCL_SERVER); - else - rc = buflen; - } else if (flags & LDLM_FL_REPLAY) { - /* no LVB resend upon replay */ - if (buflen > 0) + &RMF_DLM_LVB, buflen); + if (!rc2) { + goto retry; + } else { + /* if we can't grow the buffer, + * it's ok to return empty lvb + * to client. + */ req_capsule_shrink( &req->rq_pill, - &RMF_DLM_LVB, - 0, RCL_SERVER); - else - rc = buflen; + &RMF_DLM_LVB, 0, + RCL_SERVER); + } } else { - rc = buflen; + rc = rc2; } + } else if (flags & LDLM_FL_REPLAY) { + /* no LVB resend upon replay */ + if (buflen > 0) + req_capsule_shrink(&req->rq_pill, + &RMF_DLM_LVB, + 0, RCL_SERVER); + else + rc = buflen; + } else { + rc = buflen; } } @@ -1528,114 +1480,123 @@ existing_lock: } } - if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK) - ldlm_reprocess_all(lock->l_resource); + if (!err && !ldlm_is_cbpending(lock) && + dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK) + ldlm_reprocess_all(lock->l_resource); - LDLM_LOCK_RELEASE(lock); - } + LDLM_LOCK_RELEASE(lock); + } - LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)", - lock, rc); + LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)", + lock, rc); return rc; } -/** - * Old-style LDLM main entry point for server code enqueue. +/* Clear the blocking lock, the race is possible between ldlm_handle_convert0() + * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL. */ -int ldlm_handle_enqueue(struct ptlrpc_request *req, - ldlm_completion_callback completion_callback, - ldlm_blocking_callback blocking_callback, - ldlm_glimpse_callback glimpse_callback) +void ldlm_clear_blocking_lock(struct ldlm_lock *lock) { - struct ldlm_request *dlm_req; - struct ldlm_callback_suite cbs = { - .lcs_completion = completion_callback, - .lcs_blocking = blocking_callback, - .lcs_glimpse = glimpse_callback - }; - int rc; + if (lock->l_blocking_lock) { + LDLM_LOCK_RELEASE(lock->l_blocking_lock); + lock->l_blocking_lock = NULL; + } +} - dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); - if (dlm_req != NULL) { - rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace, - req, dlm_req, &cbs); - } else { - rc = -EFAULT; - } - return rc; +/* A lock can be converted to new ibits or mode and should be considered + * as new lock. Clear all states related to a previous blocking AST + * processing so new conflicts will cause new blocking ASTs. + * + * This is used during lock convert below and lock downgrade to COS mode in + * ldlm_lock_mode_downgrade(). + */ +void ldlm_clear_blocking_data(struct ldlm_lock *lock) +{ + ldlm_clear_ast_sent(lock); + lock->l_bl_ast_run = 0; + ldlm_clear_blocking_lock(lock); } /** * Main LDLM entry point for server code to process lock conversion requests. */ int ldlm_handle_convert0(struct ptlrpc_request *req, - const struct ldlm_request *dlm_req) + const struct ldlm_request *dlm_req) { - struct ldlm_reply *dlm_rep; - struct ldlm_lock *lock; - int rc; - ENTRY; + struct obd_export *exp = req->rq_export; + struct ldlm_reply *dlm_rep; + struct ldlm_lock *lock; + int rc; - if (req->rq_export && req->rq_export->exp_nid_stats && - req->rq_export->exp_nid_stats->nid_ldlm_stats) - lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats, - LDLM_CONVERT - LDLM_FIRST_OPC); + ENTRY; - rc = req_capsule_server_pack(&req->rq_pill); - if (rc) - RETURN(rc); + if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats) + lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats, + LDLM_CONVERT - LDLM_FIRST_OPC); - dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP); - dlm_rep->lock_flags = dlm_req->lock_flags; + rc = req_capsule_server_pack(&req->rq_pill); + if (rc) + RETURN(rc); + + dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP); + dlm_rep->lock_flags = dlm_req->lock_flags; + + lock = ldlm_handle2lock(&dlm_req->lock_handle[0]); + if (lock) { + __u64 bits; + __u64 new; + + bits = lock->l_policy_data.l_inodebits.bits; + new = dlm_req->lock_desc.l_policy_data.l_inodebits.bits; + LDLM_DEBUG(lock, "server-side convert handler START"); + + if (ldlm_is_cancel(lock)) { + LDLM_ERROR(lock, "convert on canceled lock!"); + rc = ELDLM_NO_LOCK_DATA; + } else if (dlm_req->lock_desc.l_req_mode != + lock->l_granted_mode) { + LDLM_ERROR(lock, "lock mode differs!"); + rc = ELDLM_NO_LOCK_DATA; + } else if (bits == new) { + /* This can be valid situation if CONVERT RPCs are + * re-ordered. Just finish silently */ + LDLM_DEBUG(lock, "lock is converted already!"); + rc = ELDLM_OK; + } else { + lock_res_and_lock(lock); + if (ldlm_is_waited(lock)) + ldlm_del_waiting_lock(lock); - lock = ldlm_handle2lock(&dlm_req->lock_handle[0]); - if (!lock) { - req->rq_status = LUSTRE_EINVAL; - } else { - void *res = NULL; + ldlm_clear_cbpending(lock); + lock->l_policy_data.l_inodebits.cancel_bits = 0; + ldlm_inodebits_drop(lock, bits & ~new); - LDLM_DEBUG(lock, "server-side convert handler START"); + ldlm_clear_blocking_data(lock); + unlock_res_and_lock(lock); - res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode, - &dlm_rep->lock_flags); - if (res) { - if (ldlm_del_waiting_lock(lock)) - LDLM_DEBUG(lock, "converted waiting lock"); - req->rq_status = 0; - } else { - req->rq_status = LUSTRE_EDEADLK; - } - } + ldlm_reprocess_all(lock->l_resource); + rc = ELDLM_OK; + } - if (lock) { - if (!req->rq_status) - ldlm_reprocess_all(lock->l_resource); - LDLM_DEBUG(lock, "server-side convert handler END"); - LDLM_LOCK_PUT(lock); - } else - LDLM_DEBUG_NOLOCK("server-side convert handler END"); + if (rc == ELDLM_OK) { + dlm_rep->lock_handle = lock->l_remote_handle; + ldlm_ibits_policy_local_to_wire(&lock->l_policy_data, + &dlm_rep->lock_desc.l_policy_data); + } - RETURN(0); -} + LDLM_DEBUG(lock, "server-side convert handler END, rc = %d", + rc); + LDLM_LOCK_PUT(lock); + } else { + rc = ELDLM_NO_LOCK_DATA; + LDLM_DEBUG_NOLOCK("server-side convert handler END, rc = %d", + rc); + } -/** - * Old-style main LDLM entry point for server code to process lock conversion - * requests. - */ -int ldlm_handle_convert(struct ptlrpc_request *req) -{ - int rc; - struct ldlm_request *dlm_req; + req->rq_status = rc; - dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); - if (dlm_req != NULL) { - rc = ldlm_handle_convert0(req, dlm_req); - } else { - CERROR ("Can't unpack dlm_req\n"); - rc = -EFAULT; - } - return rc; + RETURN(0); } /** @@ -1648,6 +1609,7 @@ int ldlm_request_cancel(struct ptlrpc_request *req, const struct ldlm_request *dlm_req, int first, enum lustre_at_flags flags) { + const struct lu_env *env = req->rq_svc_thread->t_env; struct ldlm_resource *res, *pres = NULL; struct ldlm_lock *lock; int i, count, done = 0; @@ -1672,7 +1634,7 @@ int ldlm_request_cancel(struct ptlrpc_request *req, lock = ldlm_handle2lock(&dlm_req->lock_handle[i]); if (!lock) { LDLM_DEBUG_NOLOCK("server-side cancel handler stale " - "lock (cookie "LPU64")", + "lock (cookie %llu)", dlm_req->lock_handle[i].cookie); continue; } @@ -1689,19 +1651,23 @@ int ldlm_request_cancel(struct ptlrpc_request *req, LDLM_RESOURCE_DELREF(pres); ldlm_resource_putref(pres); } - if (res != NULL) { - ldlm_resource_getref(res); - LDLM_RESOURCE_ADDREF(res); - ldlm_res_lvbo_update(res, NULL, 1); - } - pres = res; - } + if (res != NULL) { + ldlm_resource_getref(res); + LDLM_RESOURCE_ADDREF(res); + + if (!ldlm_is_discard_data(lock)) + ldlm_lvbo_update(env, res, lock, + NULL, 1); + } + pres = res; + } - if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock)) { - long delay = cfs_time_sub(cfs_time_current_sec(), - lock->l_last_activity); - LDLM_DEBUG(lock, "server cancels blocked lock after " - CFS_DURATION_T"s", delay); + if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) && + lock->l_blast_sent != 0) { + time64_t delay = ktime_get_real_seconds() - + lock->l_blast_sent; + LDLM_DEBUG(lock, "server cancels blocked lock after %llds", + (s64)delay); at_measured(&lock->l_export->exp_bl_lock_at, delay); } ldlm_lock_cancel(lock); @@ -1764,6 +1730,24 @@ void ldlm_handle_bl_callback(struct ldlm_namespace *ns, LDLM_DEBUG(lock, "client blocking AST callback handler"); lock_res_and_lock(lock); + + /* set bits to cancel for this lock for possible lock convert */ + if (ns_is_client(ns) && (lock->l_resource->lr_type == LDLM_IBITS)) { + /* Lock description contains policy of blocking lock, + * and its cancel_bits is used to pass conflicting bits. + * NOTE: ld can be NULL or can be not NULL but zeroed if + * passed from ldlm_bl_thread_blwi(), check below used bits + * in ld to make sure it is valid description. + */ + if (ld && ld->l_policy_data.l_inodebits.bits) + lock->l_policy_data.l_inodebits.cancel_bits = + ld->l_policy_data.l_inodebits.cancel_bits; + /* if there is no valid ld and lock is cbpending already + * then cancel_bits should be kept, otherwise it is zeroed. + */ + else if (!ldlm_is_cbpending(lock)) + lock->l_policy_data.l_inodebits.cancel_bits = 0; + } ldlm_set_cbpending(lock); if (ldlm_is_cancel_on_block(lock)) @@ -1807,7 +1791,8 @@ static void ldlm_handle_cp_callback(struct ptlrpc_request *req, INIT_LIST_HEAD(&ast_list); if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) { - int to = cfs_time_seconds(1); + long to = cfs_time_seconds(1); + while (to > 0) { set_current_state(TASK_INTERRUPTIBLE); schedule_timeout(to); @@ -1837,6 +1822,21 @@ static void ldlm_handle_cp_callback(struct ptlrpc_request *req, } lock_res_and_lock(lock); + + if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name, + &lock->l_resource->lr_name)) { + ldlm_resource_unlink_lock(lock); + unlock_res_and_lock(lock); + rc = ldlm_lock_change_resource(ns, lock, + &dlm_req->lock_desc.l_resource.lr_name); + if (rc < 0) { + LDLM_ERROR(lock, "Failed to allocate resource"); + GOTO(out, rc); + } + LDLM_DEBUG(lock, "completion AST, new resource"); + lock_res_and_lock(lock); + } + if (ldlm_is_destroyed(lock) || lock->l_granted_mode == lock->l_req_mode) { /* bug 11300: the lock has already been granted */ @@ -1860,21 +1860,7 @@ static void ldlm_handle_cp_callback(struct ptlrpc_request *req, LDLM_DEBUG(lock, "completion AST, new policy data"); } - ldlm_resource_unlink_lock(lock); - if (memcmp(&dlm_req->lock_desc.l_resource.lr_name, - &lock->l_resource->lr_name, - sizeof(lock->l_resource->lr_name)) != 0) { - unlock_res_and_lock(lock); - rc = ldlm_lock_change_resource(ns, lock, - &dlm_req->lock_desc.l_resource.lr_name); - if (rc < 0) { - LDLM_ERROR(lock, "Failed to allocate resource"); - GOTO(out, rc); - } - LDLM_DEBUG(lock, "completion AST, new resource"); - CERROR("change resource!\n"); - lock_res_and_lock(lock); - } + ldlm_resource_unlink_lock(lock); if (dlm_req->lock_flags & LDLM_FL_AST_SENT) { /* BL_AST locks are not needed in LRU. @@ -1948,9 +1934,9 @@ static void ldlm_handle_gl_callback(struct ptlrpc_request *req, lock_res_and_lock(lock); if (lock->l_granted_mode == LCK_PW && !lock->l_readers && !lock->l_writers && - cfs_time_after(cfs_time_current(), - cfs_time_add(lock->l_last_used, - cfs_time_seconds(10)))) { + ktime_after(ktime_get(), + ktime_add(lock->l_last_used, + ktime_set(ns->ns_dirty_age_limit, 0)))) { unlock_res_and_lock(lock); if (ldlm_bl_to_thread_lock(ns, NULL, lock)) ldlm_handle_bl_callback(ns, NULL, lock); @@ -2140,7 +2126,7 @@ static inline void ldlm_callback_errmsg(struct ptlrpc_request *req, const struct lustre_handle *handle) { DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req, - "%s: [nid %s] [rc %d] [lock "LPX64"]", + "%s: [nid %s] [rc %d] [lock %#llx]", msg, libcfs_id2str(req->rq_peer), rc, handle ? handle->cookie : 0); if (req->rq_no_reply) @@ -2199,35 +2185,6 @@ static int ldlm_callback_handler(struct ptlrpc_request *req) rc = ldlm_handle_setinfo(req); ldlm_callback_reply(req, rc); RETURN(0); - case LLOG_ORIGIN_HANDLE_CREATE: - req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE); - if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET)) - RETURN(0); - rc = llog_origin_handle_open(req); - ldlm_callback_reply(req, rc); - RETURN(0); - case LLOG_ORIGIN_HANDLE_NEXT_BLOCK: - req_capsule_set(&req->rq_pill, - &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK); - if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET)) - RETURN(0); - rc = llog_origin_handle_next_block(req); - ldlm_callback_reply(req, rc); - RETURN(0); - case LLOG_ORIGIN_HANDLE_READ_HEADER: - req_capsule_set(&req->rq_pill, - &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER); - if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET)) - RETURN(0); - rc = llog_origin_handle_read_header(req); - ldlm_callback_reply(req, rc); - RETURN(0); - case LLOG_ORIGIN_HANDLE_CLOSE: - if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET)) - RETURN(0); - rc = llog_origin_handle_close(req); - ldlm_callback_reply(req, rc); - RETURN(0); default: CERROR("unknown opcode %u\n", lustre_msg_get_opc(req->rq_reqmsg)); @@ -2259,7 +2216,7 @@ static int ldlm_callback_handler(struct ptlrpc_request *req) lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0); if (!lock) { - CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock " + CDEBUG(D_DLMTRACE, "callback on lock %#llx - lock " "disappeared\n", dlm_req->lock_handle[0].cookie); rc = ldlm_callback_reply(req, -EINVAL); ldlm_callback_errmsg(req, "Operate with invalid parameter", rc, @@ -2282,8 +2239,7 @@ static int ldlm_callback_handler(struct ptlrpc_request *req) * should send cancel after dropping the cache. */ if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) || ldlm_is_failed(lock)) { - LDLM_DEBUG(lock, "callback on lock " - LPX64" - lock disappeared", + LDLM_DEBUG(lock, "callback on lock %llx - lock disappeared", dlm_req->lock_handle[0].cookie); unlock_res_and_lock(lock); LDLM_LOCK_RELEASE(lock); @@ -2347,145 +2303,163 @@ static int ldlm_callback_handler(struct ptlrpc_request *req) */ static int ldlm_cancel_handler(struct ptlrpc_request *req) { - int rc; - ENTRY; - - /* Requests arrive in sender's byte order. The ptlrpc service - * handler has already checked and, if necessary, byte-swapped the - * incoming request message body, but I am responsible for the - * message buffers. */ - - req_capsule_init(&req->rq_pill, req, RCL_SERVER); + int rc; - if (req->rq_export == NULL) { - struct ldlm_request *dlm_req; - - CERROR("%s from %s arrived at %lu with bad export cookie " - LPU64"\n", - ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)), - libcfs_nid2str(req->rq_peer.nid), - req->rq_arrival_time.tv_sec, - lustre_msg_get_handle(req->rq_reqmsg)->cookie); - - if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) { - req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK); - dlm_req = req_capsule_client_get(&req->rq_pill, - &RMF_DLM_REQ); - if (dlm_req != NULL) - ldlm_lock_dump_handle(D_ERROR, - &dlm_req->lock_handle[0]); - } - ldlm_callback_reply(req, -ENOTCONN); - RETURN(0); - } + ENTRY; - switch (lustre_msg_get_opc(req->rq_reqmsg)) { + /* Requests arrive in sender's byte order. The ptlrpc service + * handler has already checked and, if necessary, byte-swapped the + * incoming request message body, but I am responsible for the + * message buffers. */ + + req_capsule_init(&req->rq_pill, req, RCL_SERVER); + + if (req->rq_export == NULL) { + struct ldlm_request *dlm_req; + + CERROR("%s from %s arrived at %lu with bad export cookie " + "%llu\n", + ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)), + libcfs_nid2str(req->rq_peer.nid), + req->rq_arrival_time.tv_sec, + lustre_msg_get_handle(req->rq_reqmsg)->cookie); + + if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) { + req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK); + dlm_req = req_capsule_client_get(&req->rq_pill, + &RMF_DLM_REQ); + if (dlm_req != NULL) + ldlm_lock_dump_handle(D_ERROR, + &dlm_req->lock_handle[0]); + } + ldlm_callback_reply(req, -ENOTCONN); + RETURN(0); + } - /* XXX FIXME move this back to mds/handler.c, bug 249 */ - case LDLM_CANCEL: - req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL); - CDEBUG(D_INODE, "cancel\n"); + switch (lustre_msg_get_opc(req->rq_reqmsg)) { + /* XXX FIXME move this back to mds/handler.c, bug 249 */ + case LDLM_CANCEL: + req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL); + CDEBUG(D_INODE, "cancel\n"); if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) || CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) || CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT)) RETURN(0); - rc = ldlm_handle_cancel(req); - if (rc) - break; - RETURN(0); - default: - CERROR("invalid opcode %d\n", - lustre_msg_get_opc(req->rq_reqmsg)); - req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK); - ldlm_callback_reply(req, -EINVAL); - } + rc = ldlm_handle_cancel(req); + break; + case LDLM_CONVERT: + { + struct ldlm_request *dlm_req; - RETURN(0); + req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT); + CDEBUG(D_INODE, "convert\n"); + + dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); + if (dlm_req == NULL) { + CDEBUG(D_INFO, "bad request buffer for cancel\n"); + rc = ldlm_callback_reply(req, -EPROTO); + } else { + req->rq_status = ldlm_handle_convert0(req, dlm_req); + rc = ptlrpc_reply(req); + } + break; + } + default: + CERROR("invalid opcode %d\n", + lustre_msg_get_opc(req->rq_reqmsg)); + req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK); + rc = ldlm_callback_reply(req, -EINVAL); + } + + RETURN(rc); } static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req, - struct ldlm_lock *lock) + struct ldlm_lock *lock) { - struct ldlm_request *dlm_req; - struct lustre_handle lockh; - int rc = 0; - int i; - ENTRY; - - dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); - if (dlm_req == NULL) - RETURN(0); + struct ldlm_request *dlm_req; + struct lustre_handle lockh; + int rc = 0; + int i; - ldlm_lock2handle(lock, &lockh); - for (i = 0; i < dlm_req->lock_count; i++) { - if (lustre_handle_equal(&dlm_req->lock_handle[i], - &lockh)) { - DEBUG_REQ(D_RPCTRACE, req, - "Prio raised by lock "LPX64".", lockh.cookie); + ENTRY; - rc = 1; - break; - } - } + dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); + if (dlm_req == NULL) + RETURN(0); - RETURN(rc); + ldlm_lock2handle(lock, &lockh); + for (i = 0; i < dlm_req->lock_count; i++) { + if (lustre_handle_equal(&dlm_req->lock_handle[i], + &lockh)) { + DEBUG_REQ(D_RPCTRACE, req, + "Prio raised by lock %#llx.", lockh.cookie); + rc = 1; + break; + } + } + RETURN(rc); } static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req) { - struct ldlm_request *dlm_req; - int rc = 0; - int i; - ENTRY; + struct ldlm_request *dlm_req; + int rc = 0; + int i; - /* no prolong in recovery */ - if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) - RETURN(0); + ENTRY; - dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); - if (dlm_req == NULL) - RETURN(-EFAULT); + /* no prolong in recovery */ + if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) + RETURN(0); - for (i = 0; i < dlm_req->lock_count; i++) { - struct ldlm_lock *lock; + dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); + if (dlm_req == NULL) + RETURN(-EFAULT); - lock = ldlm_handle2lock(&dlm_req->lock_handle[i]); - if (lock == NULL) - continue; + for (i = 0; i < dlm_req->lock_count; i++) { + struct ldlm_lock *lock; + + lock = ldlm_handle2lock(&dlm_req->lock_handle[i]); + if (lock == NULL) + continue; rc = ldlm_is_ast_sent(lock) ? 1 : 0; - if (rc) - LDLM_DEBUG(lock, "hpreq cancel lock"); - LDLM_LOCK_PUT(lock); + if (rc) + LDLM_DEBUG(lock, "hpreq cancel/convert lock"); + LDLM_LOCK_PUT(lock); - if (rc) - break; - } + if (rc) + break; + } - RETURN(rc); + RETURN(rc); } static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = { - .hpreq_lock_match = ldlm_cancel_hpreq_lock_match, + .hpreq_lock_match = ldlm_cancel_hpreq_lock_match, .hpreq_check = ldlm_cancel_hpreq_check, .hpreq_fini = NULL, }; static int ldlm_hpreq_handler(struct ptlrpc_request *req) { - ENTRY; + ENTRY; - req_capsule_init(&req->rq_pill, req, RCL_SERVER); + req_capsule_init(&req->rq_pill, req, RCL_SERVER); - if (req->rq_export == NULL) - RETURN(0); + if (req->rq_export == NULL) + RETURN(0); - if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) { - req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL); - req->rq_ops = &ldlm_cancel_hpreq_ops; - } - RETURN(0); + if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) { + req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL); + req->rq_ops = &ldlm_cancel_hpreq_ops; + } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) { + req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT); + req->rq_ops = &ldlm_cancel_hpreq_ops; + } + RETURN(0); } static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd, @@ -2538,15 +2512,15 @@ static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd, void ldlm_revoke_export_locks(struct obd_export *exp) { struct list_head rpc_list; - ENTRY; + ENTRY; INIT_LIST_HEAD(&rpc_list); - cfs_hash_for_each_empty(exp->exp_lock_hash, - ldlm_revoke_lock_cb, &rpc_list); - ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list, - LDLM_WORK_REVOKE_AST); + cfs_hash_for_each_nolock(exp->exp_lock_hash, + ldlm_revoke_lock_cb, &rpc_list, 0); + ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list, + LDLM_WORK_REVOKE_AST); - EXIT; + EXIT; } EXPORT_SYMBOL(ldlm_revoke_export_locks); #endif /* HAVE_SERVER_SUPPORT */ @@ -2769,6 +2743,11 @@ static int ldlm_bl_thread_main(void *arg) if (rc == LDLM_ITER_STOP) break; + + /* If there are many namespaces, we will not sleep waiting for + * work, and must do a cond_resched to avoid holding the CPU + * for too long */ + cond_resched(); } atomic_dec(&blp->blp_num_threads); @@ -2921,6 +2900,40 @@ void ldlm_destroy_export(struct obd_export *exp) } EXPORT_SYMBOL(ldlm_destroy_export); +static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj, + struct attribute *attr, + char *buf) +{ + return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay); +} + +static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj, + struct attribute *attr, + const char *buffer, + size_t count) +{ + int rc; + unsigned long val; + + rc = kstrtoul(buffer, 10, &val); + if (rc) + return rc; + + ldlm_cancel_unused_locks_before_replay = val; + + return count; +} +LUSTRE_RW_ATTR(cancel_unused_locks_before_replay); + +static struct attribute *ldlm_attrs[] = { + &lustre_attr_cancel_unused_locks_before_replay.attr, + NULL, +}; + +static struct attribute_group ldlm_attr_group = { + .attrs = ldlm_attrs, +}; + static int ldlm_setup(void) { static struct ptlrpc_service_conf conf; @@ -2940,11 +2953,25 @@ static int ldlm_setup(void) if (ldlm_state == NULL) RETURN(-ENOMEM); -#ifdef CONFIG_PROC_FS - rc = ldlm_proc_setup(); - if (rc != 0) + ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj); + if (!ldlm_kobj) + GOTO(out, -ENOMEM); + + rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group); + if (rc) + GOTO(out, rc); + + ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj); + if (!ldlm_ns_kset) + GOTO(out, -ENOMEM); + + ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj); + if (!ldlm_svc_kset) + GOTO(out, -ENOMEM); + + rc = ldlm_debugfs_setup(); + if (rc != 0) GOTO(out, rc); -#endif /* CONFIG_PROC_FS */ memset(&conf, 0, sizeof(conf)); conf = (typeof(conf)) { @@ -2965,18 +2992,20 @@ static int ldlm_setup(void) .tc_nthrs_base = LDLM_NTHRS_BASE, .tc_nthrs_max = LDLM_NTHRS_MAX, .tc_nthrs_user = ldlm_num_threads, - .tc_cpu_affinity = 1, + .tc_cpu_bind = ldlm_cpu_bind, .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD, }, .psc_cpt = { .cc_pattern = ldlm_cpts, + .cc_affinity = true, }, .psc_ops = { .so_req_handler = ldlm_callback_handler, }, }; ldlm_state->ldlm_cb_service = \ - ptlrpc_register_service(&conf, ldlm_svc_proc_dir); + ptlrpc_register_service(&conf, ldlm_svc_kset, + ldlm_svc_debugfs_dir); if (IS_ERR(ldlm_state->ldlm_cb_service)) { CERROR("failed to start service\n"); rc = PTR_ERR(ldlm_state->ldlm_cb_service); @@ -3005,13 +3034,14 @@ static int ldlm_setup(void) .tc_nthrs_base = LDLM_NTHRS_BASE, .tc_nthrs_max = LDLM_NTHRS_MAX, .tc_nthrs_user = ldlm_num_threads, - .tc_cpu_affinity = 1, + .tc_cpu_bind = ldlm_cpu_bind, .tc_ctx_tags = LCT_MD_THREAD | \ LCT_DT_THREAD | \ LCT_CL_THREAD, }, .psc_cpt = { .cc_pattern = ldlm_cpts, + .cc_affinity = true, }, .psc_ops = { .so_req_handler = ldlm_cancel_handler, @@ -3019,7 +3049,8 @@ static int ldlm_setup(void) }, }; ldlm_state->ldlm_cancel_service = \ - ptlrpc_register_service(&conf, ldlm_svc_proc_dir); + ptlrpc_register_service(&conf, ldlm_svc_kset, + ldlm_svc_debugfs_dir); if (IS_ERR(ldlm_state->ldlm_cancel_service)) { CERROR("failed to start service\n"); rc = PTR_ERR(ldlm_state->ldlm_cancel_service); @@ -3056,14 +3087,6 @@ static int ldlm_setup(void) } #ifdef HAVE_SERVER_SUPPORT - INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks); - expired_lock_thread.elt_state = ELT_STOPPED; - init_waitqueue_head(&expired_lock_thread.elt_waitq); - - INIT_LIST_HEAD(&waiting_locks_list); - spin_lock_init(&waiting_locks_spinlock); - cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, NULL); - task = kthread_run(expired_lock_main, NULL, "ldlm_elt"); if (IS_ERR(task)) { rc = PTR_ERR(task); @@ -3071,8 +3094,8 @@ static int ldlm_setup(void) GOTO(out, rc); } - wait_event(expired_lock_thread.elt_waitq, - expired_lock_thread.elt_state == ELT_READY); + wait_event(expired_lock_wait_queue, + expired_lock_thread_state == ELT_READY); #endif /* HAVE_SERVER_SUPPORT */ rc = ldlm_pools_init(); @@ -3134,14 +3157,23 @@ static int ldlm_cleanup(void) ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service); #endif - ldlm_proc_cleanup(); + if (ldlm_ns_kset) + kset_unregister(ldlm_ns_kset); + if (ldlm_svc_kset) + kset_unregister(ldlm_svc_kset); + if (ldlm_kobj) { + sysfs_remove_group(ldlm_kobj, &ldlm_attr_group); + kobject_put(ldlm_kobj); + } + + ldlm_debugfs_cleanup(); #ifdef HAVE_SERVER_SUPPORT - if (expired_lock_thread.elt_state != ELT_STOPPED) { - expired_lock_thread.elt_state = ELT_TERMINATE; - wake_up(&expired_lock_thread.elt_waitq); - wait_event(expired_lock_thread.elt_waitq, - expired_lock_thread.elt_state == ELT_STOPPED); + if (expired_lock_thread_state != ELT_STOPPED) { + expired_lock_thread_state = ELT_TERMINATE; + wake_up(&expired_lock_wait_queue); + wait_event(expired_lock_wait_queue, + expired_lock_thread_state == ELT_STOPPED); } #endif @@ -3153,14 +3185,6 @@ static int ldlm_cleanup(void) int ldlm_init(void) { - mutex_init(&ldlm_ref_mutex); - mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER)); - mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT)); - - INIT_LIST_HEAD(&ldlm_srv_namespace_list); - INIT_LIST_HEAD(&ldlm_cli_active_namespace_list); - INIT_LIST_HEAD(&ldlm_cli_inactive_namespace_list); - ldlm_resource_slab = kmem_cache_create("ldlm_resources", sizeof(struct ldlm_resource), 0, SLAB_HWCACHE_ALIGN, NULL); @@ -3169,7 +3193,7 @@ int ldlm_init(void) ldlm_lock_slab = kmem_cache_create("ldlm_locks", sizeof(struct ldlm_lock), 0, - SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU, NULL); + SLAB_HWCACHE_ALIGN, NULL); if (ldlm_lock_slab == NULL) goto out_resource; @@ -3185,11 +3209,22 @@ int ldlm_init(void) if (ldlm_interval_tree_slab == NULL) goto out_interval; +#ifdef HAVE_SERVER_SUPPORT + ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem", + sizeof(struct ldlm_glimpse_work), + 0, 0, NULL); + if (ldlm_glimpse_work_kmem == NULL) + goto out_interval_tree; +#endif + #if LUSTRE_TRACKS_LOCK_EXP_REFS class_export_dump_hook = ldlm_dump_export_locks; #endif return 0; - +#ifdef HAVE_SERVER_SUPPORT +out_interval_tree: + kmem_cache_destroy(ldlm_interval_tree_slab); +#endif out_interval: kmem_cache_destroy(ldlm_interval_slab); out_lock: @@ -3212,4 +3247,7 @@ void ldlm_exit(void) kmem_cache_destroy(ldlm_lock_slab); kmem_cache_destroy(ldlm_interval_slab); kmem_cache_destroy(ldlm_interval_tree_slab); +#ifdef HAVE_SERVER_SUPPORT + kmem_cache_destroy(ldlm_glimpse_work_kmem); +#endif }