X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_lockd.c;h=178976b221d52624010068364916c0faf828ea1f;hp=c2a420d6269c0b099c433bf8dd959938796236a0;hb=5c883ea2748ae9e430a9cd863a9b630b2a74440a;hpb=a42a91c783903ea15ad902032166c7e312dad7ee diff --git a/lustre/ldlm/ldlm_lockd.c b/lustre/ldlm/ldlm_lockd.c index c2a420d..178976b 100644 --- a/lustre/ldlm/ldlm_lockd.c +++ b/lustre/ldlm/ldlm_lockd.c @@ -23,7 +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/ @@ -40,6 +40,7 @@ #include #include #include +#include #include #include #include "ldlm_internal.h" @@ -48,62 +49,61 @@ 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; + spinlock_t blp_lock; /* * blp_prio_list is used for callbacks that should be handled * as a priority. It is used for LDLM_FL_DISCARD_DATA requests. - * see bug 13843 + * see b=13843 */ - struct list_head blp_prio_list; + struct list_head blp_prio_list; /* * blp_list is used for all other callbacks which are likely * to take longer to process. */ - struct list_head blp_list; - - wait_queue_head_t blp_waitq; - struct completion blp_comp; - atomic_t blp_num_threads; - atomic_t blp_busy_threads; - int blp_min_threads; - int blp_max_threads; + struct list_head blp_list; + + wait_queue_head_t blp_waitq; + struct completion blp_comp; + atomic_t blp_num_threads; + atomic_t blp_busy_threads; + int blp_min_threads; + int blp_max_threads; }; struct ldlm_bl_work_item { @@ -123,7 +123,7 @@ 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. @@ -136,15 +136,24 @@ 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 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 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 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) { @@ -152,7 +161,7 @@ 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); @@ -163,35 +172,28 @@ 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 l_wait_info lwi = { 0 }; + struct list_head *expired = &expired_lock_list; 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, - have_expired_locks() || - expired_lock_thread.elt_state == ELT_TERMINATE, - &lwi); + wait_event_idle(expired_lock_wait_queue, + have_expired_locks() || + expired_lock_thread_state == ELT_TERMINATE); 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; @@ -216,32 +218,54 @@ static int expired_lock_main(void *arg) lock->l_export); lock->l_export = NULL; LDLM_ERROR(lock, "free export"); - /* release extra ref grabbed by + /* + * release extra ref grabbed by * ldlm_add_waiting_lock() or - * ldlm_failed_ast() */ + * ldlm_failed_ast() + */ LDLM_LOCK_RELEASE(lock); continue; } if (ldlm_is_destroyed(lock)) { - /* release the lock refcount where - * waiting_locks_callback() founds */ + /* + * release the lock refcount where + * waiting_locks_callback() founds + */ LDLM_LOCK_RELEASE(lock); continue; } 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() */ + /* + * release extra ref grabbed by ldlm_add_waiting_lock() + * or ldlm_failed_ast() + */ LDLM_LOCK_RELEASE(lock); spin_lock_bh(&waiting_locks_spinlock); @@ -253,18 +277,15 @@ 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. @@ -273,12 +294,13 @@ static int ldlm_lock_busy(struct ldlm_lock *lock) { struct ptlrpc_request *req; int match = 0; + ENTRY; 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) { @@ -287,83 +309,55 @@ 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; + struct ldlm_lock *lock; + int need_dump = 0; 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); + l_pending_chain); + if (lock->l_callback_timeout > ktime_get_seconds() || + lock->l_req_mode == LCK_GROUP) + break; - 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)); - - /* 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); + /* + * 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_move(&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; + time64_t now = ktime_get_seconds(); + time_t delta = 0; + 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); + if (lock->l_callback_timeout - now > 0) + delta = lock->l_callback_timeout - now; + mod_timer(&waiting_locks_timer, + jiffies + cfs_time_seconds(delta)); + } spin_unlock_bh(&waiting_locks_spinlock); } @@ -379,41 +373,46 @@ 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 = jiffies; + time64_t now = ktime_get_seconds(); + time64_t deadline; + time_t timeout; if (!list_empty(&lock->l_pending_chain)) - return 0; - - if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) || - 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))) - lock->l_callback_timeout = timeout; - - timeout_rounded = round_timeout(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 */ + return 0; + + if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) || + OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT)) + seconds = 1; + + deadline = now + seconds; + if (likely(deadline > lock->l_callback_timeout)) + lock->l_callback_timeout = deadline; + + timeout = clamp_t(time_t, lock->l_callback_timeout - now, + 0, seconds); + timeout_jiffies += cfs_time_seconds(timeout); + + 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) { spin_lock_bh(&lock->l_export->exp_bl_list_lock); if (list_empty(&lock->l_exp_list)) { - if (lock->l_granted_mode != lock->l_req_mode) + if (!ldlm_is_granted(lock)) list_add_tail(&lock->l_exp_list, &lock->l_export->exp_bl_list); else @@ -422,26 +421,29 @@ static void ldlm_add_blocked_lock(struct ldlm_lock *lock) } spin_unlock_bh(&lock->l_export->exp_bl_list_lock); - /* A blocked lock is added. Adjust the position in + /* + * A blocked lock is added. Adjust the position in * the stale list if the export is in the list. * If export is stale and not in the list - it is being * processed and will be placed on the right position - * on obd_stale_export_put(). */ + * on obd_stale_export_put(). + */ if (!list_empty(&lock->l_export->exp_stale_list)) 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)); LASSERT(!ldlm_is_cancel_on_block(lock)); - /* Do not put cross-MDT lock in the waiting list, since we - * will not evict it due to timeout for now */ + /* + * Do not put cross-MDT lock in the waiting list, since we + * will not evict it due to timeout for now + */ if (lock->l_export != NULL && (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS)) return 0; @@ -453,23 +455,25 @@ 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); + LDLM_ERROR(lock, "not waiting on destroyed lock (b=5653)"); + 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 - * waiting list */ + /* + * grab ref on the lock if it has been added to the + * waiting list + */ LDLM_LOCK_GET(lock); } spin_unlock_bh(&waiting_locks_spinlock); @@ -477,7 +481,7 @@ 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; @@ -497,36 +501,42 @@ static int __ldlm_del_waiting_lock(struct ldlm_lock *lock) struct list_head *list_next; if (list_empty(&lock->l_pending_chain)) - return 0; - - list_next = lock->l_pending_chain.next; - if (lock->l_pending_chain.prev == &waiting_locks_list) { - /* Removing the head of the list, adjust timer. */ - if (list_next == &waiting_locks_list) { - /* No more, just cancel. */ - cfs_timer_disarm(&waiting_locks_timer); - } else { - struct ldlm_lock *next; + return 0; + + list_next = lock->l_pending_chain.next; + if (lock->l_pending_chain.prev == &waiting_locks_list) { + /* Removing the head of the list, adjust timer. */ + if (list_next == &waiting_locks_list) { + /* No more, just cancel. */ + del_timer(&waiting_locks_timer); + } else { + time64_t now = ktime_get_seconds(); + struct ldlm_lock *next; + time_t delta = 0; + 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); + if (next->l_callback_timeout - now > 0) + delta = lock->l_callback_timeout - now; + + mod_timer(&waiting_locks_timer, + jiffies + cfs_time_seconds(delta)); + } + } list_del_init(&lock->l_pending_chain); - return 1; + return 1; } int ldlm_del_waiting_lock(struct ldlm_lock *lock) { - int ret; + int ret; - if (lock->l_export == NULL) { - /* We don't have a "waiting locks list" on clients. */ - CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock); - return 0; - } + if (lock->l_export == NULL) { + /* We don't have a "waiting locks list" on clients. */ + CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock); + return 0; + } spin_lock_bh(&waiting_locks_spinlock); ret = __ldlm_del_waiting_lock(lock); @@ -538,14 +548,16 @@ int ldlm_del_waiting_lock(struct ldlm_lock *lock) list_del_init(&lock->l_exp_list); spin_unlock_bh(&lock->l_export->exp_bl_list_lock); - if (ret) { - /* release lock ref if it has indeed been removed - * from a list */ - LDLM_LOCK_RELEASE(lock); - } + if (ret) { + /* + * release lock ref if it has indeed been removed + * from a list + */ + LDLM_LOCK_RELEASE(lock); + } - LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed"); - return ret; + LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed"); + return ret; } /** @@ -553,7 +565,7 @@ 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. */ @@ -575,8 +587,10 @@ int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout) return 0; } - /* we remove/add the lock to the waiting list, so no needs to - * release/take a lock reference */ + /* + * we remove/add the lock to the waiting list, so no needs to + * release/take a lock reference + */ __ldlm_del_waiting_lock(lock); __ldlm_add_waiting_lock(lock, timeout); spin_unlock_bh(&waiting_locks_spinlock); @@ -590,12 +604,12 @@ EXPORT_SYMBOL(ldlm_refresh_waiting_lock); int ldlm_del_waiting_lock(struct ldlm_lock *lock) { - RETURN(0); + 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); + RETURN(0); } #endif /* !HAVE_SERVER_SUPPORT */ @@ -611,19 +625,21 @@ 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; - /* Since these are non-updating timeouts, we should be conservative. + /* + * Since these are non-updating timeouts, we should be conservative. * Take more than usually, 150% * It would be nice to have some kind of "early reply" mechanism for - * lock callbacks too... */ + * 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); @@ -631,23 +647,25 @@ EXPORT_SYMBOL(ldlm_bl_timeout); * Perform lock cleanup if AST sending failed. */ static void ldlm_failed_ast(struct ldlm_lock *lock, int rc, - const char *ast_type) + const char *ast_type) { - LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due " - "to a lock %s callback time out: rc %d\n", - lock->l_export->exp_obd->obd_name, - obd_export_nid2str(lock->l_export), ast_type, rc); + LCONSOLE_ERROR_MSG(0x138, + "%s: A client on nid %s was evicted due to a lock %s callback time out: rc %d\n", + lock->l_export->exp_obd->obd_name, + obd_export_nid2str(lock->l_export), ast_type, rc); - if (obd_dump_on_timeout) - libcfs_debug_dumplog(); + if (obd_dump_on_timeout) + libcfs_debug_dumplog(); spin_lock_bh(&waiting_locks_spinlock); if (__ldlm_del_waiting_lock(lock) == 0) - /* the lock was not in any list, grab an extra ref before adding - * the lock to the expired list */ + /* + * 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); } @@ -658,23 +676,34 @@ static int ldlm_handle_ast_error(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 (req@%p x%llu) to liblustre client (nid %s) timeout, just cancelling lock", - ast_type, req, req->rq_xid, - libcfs_nid2str(peer.nid)); - ldlm_lock_cancel(lock); - rc = -ERESTART; - } else if (ldlm_is_cancel(lock)) { + 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 (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@%p x%llu status %d rc %d), evict it", @@ -700,10 +729,12 @@ static int ldlm_handle_ast_error(struct ldlm_lock *lock, lustre_msg_get_status(req->rq_repmsg) : -1, ast_type, req, req->rq_xid); if (res) { - /* update lvbo to return proper attributes. - * see bug 23174 */ + /* + * update lvbo to return proper attributes. + * see b=23174 + */ ldlm_resource_getref(res); - ldlm_res_lvbo_update(res, NULL, 1); + ldlm_lvbo_update(res, lock, NULL, 1); ldlm_resource_putref(res); } ldlm_lock_cancel(lock); @@ -714,18 +745,20 @@ 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; + + ENTRY; - LASSERT(lock != NULL); + LASSERT(lock != NULL); switch (arg->type) { case LDLM_GL_CALLBACK: - /* Update the LVB from disk if the AST failed + /* + * Update the LVB from disk if the AST failed * (this is a legal race) * * - Glimpse callback of local lock just returns @@ -733,14 +766,17 @@ 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) { - LDLM_DEBUG(lock, "lost race - client has a lock but no " - "inode"); - ldlm_res_lvbo_update(lock->l_resource, NULL, 1); + if (unlikely(arg->gl_interpret_reply)) { + rc = arg->gl_interpret_reply(NULL, req, args, rc); + } else if (rc == -ELDLM_NO_LOCK_DATA) { + LDLM_DEBUG(lock, + "lost race - client has a lock but no inode"); + ldlm_lvbo_update(lock->l_resource, lock, NULL, 1); } else if (rc != 0) { rc = ldlm_handle_ast_error(lock, req, rc, "glimpse"); } else { - rc = ldlm_res_lvbo_update(lock->l_resource, req, 1); + rc = ldlm_lvbo_update(lock->l_resource, + lock, req, 1); } break; case LDLM_BL_CALLBACK: @@ -758,7 +794,7 @@ static int ldlm_cb_interpret(const struct lu_env *env, } /* release extra reference taken in ldlm_ast_fini() */ - LDLM_LOCK_RELEASE(lock); + LDLM_LOCK_RELEASE(lock); if (rc == -ERESTART) atomic_inc(&arg->restart); @@ -768,8 +804,8 @@ 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)); } @@ -780,6 +816,7 @@ static inline int ldlm_ast_fini(struct ptlrpc_request *req, int instant_cancel) { int rc = 0; + ENTRY; if (unlikely(instant_cancel)) { @@ -802,6 +839,7 @@ static inline int ldlm_ast_fini(struct ptlrpc_request *req, static void ldlm_lock_reorder_req(struct ldlm_lock *lock) { struct ptlrpc_request *req; + ENTRY; if (lock->l_export == NULL) { @@ -809,10 +847,11 @@ 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 + /* + * Do not process requests that were not yet added to there * incoming queue or were already removed from there for * processing. We evaluate ptlrpc_nrs_req_can_move() without * holding svcpt->scp_req_lock, and then redo the check with @@ -823,7 +862,7 @@ 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; } @@ -835,45 +874,45 @@ static void ldlm_lock_reorder_req(struct ldlm_lock *lock) * to wait for client response. */ int ldlm_server_blocking_ast(struct ldlm_lock *lock, - struct ldlm_lock_desc *desc, - void *data, int flag) -{ - struct ldlm_cb_async_args *ca; - struct ldlm_cb_set_arg *arg = data; - struct ldlm_request *body; - struct ptlrpc_request *req; - int instant_cancel = 0; - int rc = 0; - ENTRY; + struct ldlm_lock_desc *desc, + void *data, int flag) +{ + struct ldlm_cb_async_args *ca; + struct ldlm_cb_set_arg *arg = data; + struct ldlm_request *body; + struct ptlrpc_request *req; + int instant_cancel = 0; + int rc = 0; + + ENTRY; - if (flag == LDLM_CB_CANCELING) - /* Don't need to do anything here. */ - RETURN(0); + if (flag == LDLM_CB_CANCELING) + /* 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) - LDLM_ERROR(lock, "BUG 6063: lock collide during recovery"); + LASSERT(lock); + LASSERT(data != NULL); + if (lock->l_export->exp_obd->obd_recovering != 0) + LDLM_ERROR(lock, "BUG 6063: lock collide during recovery"); - ldlm_lock_reorder_req(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; + ca = ptlrpc_req_async_args(ca, 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)) { @@ -883,9 +922,11 @@ int ldlm_server_blocking_ast(struct ldlm_lock *lock, RETURN(0); } - if (lock->l_granted_mode != lock->l_req_mode) { - /* this blocking AST will be communicated as part of the - * completion AST instead */ + if (!ldlm_is_granted(lock)) { + /* + * this blocking AST will be communicated as part of the + * completion AST instead + */ ldlm_add_blocked_lock(lock); ldlm_set_waited(lock); unlock_res_and_lock(lock); @@ -896,16 +937,16 @@ int ldlm_server_blocking_ast(struct ldlm_lock *lock, } if (ldlm_is_cancel_on_block(lock)) - instant_cancel = 1; + instant_cancel = 1; - body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); - body->lock_handle[0] = lock->l_remote_handle; - body->lock_desc = *desc; + body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); + body->lock_handle[0] = lock->l_remote_handle; + body->lock_desc = *desc; body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK); - LDLM_DEBUG(lock, "server preparing blocking AST"); + LDLM_DEBUG(lock, "server preparing blocking AST"); - ptlrpc_request_set_replen(req); + ptlrpc_request_set_replen(req); ldlm_set_cbpending(lock); if (instant_cancel) { unlock_res_and_lock(lock); @@ -913,8 +954,8 @@ 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); + LASSERT(ldlm_is_granted(lock)); + ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock)); unlock_res_and_lock(lock); /* Do not resend after lock callback timeout */ @@ -922,21 +963,19 @@ int ldlm_server_blocking_ast(struct ldlm_lock *lock, req->rq_resend_cb = ldlm_update_resend; } - req->rq_send_state = LUSTRE_IMP_FULL; - /* ptlrpc_request_alloc_pack already set timeout */ - if (AT_OFF) - req->rq_timeout = ldlm_get_rq_timeout(); - - 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(); - 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_BL_CALLBACK - LDLM_FIRST_OPC); + 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_BL_CALLBACK - LDLM_FIRST_OPC); rc = ldlm_ast_fini(req, arg, lock, instant_cancel); - RETURN(rc); + RETURN(rc); } /** @@ -948,66 +987,69 @@ int ldlm_server_blocking_ast(struct ldlm_lock *lock, */ int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data) { - struct ldlm_cb_set_arg *arg = data; - struct ldlm_request *body; - struct ptlrpc_request *req; - struct ldlm_cb_async_args *ca; - int instant_cancel = 0; - int rc = 0; - int lvb_len; - ENTRY; + struct ldlm_cb_set_arg *arg = data; + struct ldlm_request *body; + struct ptlrpc_request *req; + struct ldlm_cb_async_args *ca; + int instant_cancel = 0; + int rc = 0; + int lvb_len; - LASSERT(lock != NULL); - LASSERT(data != NULL); + ENTRY; + + LASSERT(lock != NULL); + LASSERT(data != NULL); if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) { LDLM_DEBUG(lock, "dropping CP AST"); RETURN(0); } - req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse, - &RQF_LDLM_CP_CALLBACK); - if (req == NULL) - RETURN(-ENOMEM); + req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse, + &RQF_LDLM_CP_CALLBACK); + if (req == NULL) + RETURN(-ENOMEM); /* server namespace, doesn't need lock */ lvb_len = ldlm_lvbo_size(lock); - /* LU-3124 & LU-2187: to not return layout in completion AST because + /* + * LU-3124 & LU-2187: to not return layout in completion AST because * it may deadlock for LU-2187, or client may not have enough space * for large layout. The layout will be returned to client with an - * extra RPC to fetch xattr.lov */ + * extra RPC to fetch xattr.lov + */ if (ldlm_has_layout(lock)) 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; + ca = ptlrpc_req_async_args(ca, 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); + ldlm_lock2desc(lock, &body->lock_desc); if (lvb_len > 0) { void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB); - - lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len); + lvb_len = ldlm_lvbo_fill(lock, lvb, &lvb_len); if (lvb_len < 0) { - /* We still need to send the RPC to wake up the blocked + /* + * We still need to send the RPC to wake up the blocked * enqueue thread on the client. * * Consider old client, there is no better way to notify * the failure, just zero-sized the LVB, then the client - * will fail out as "-EPROTO". */ + * will fail out as "-EPROTO". + */ req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0, RCL_CLIENT); instant_cancel = 1; @@ -1015,33 +1057,33 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data) req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len, RCL_CLIENT); } - } - - lock->l_last_activity = cfs_time_current_sec(); + } LDLM_DEBUG(lock, "server preparing completion AST"); - ptlrpc_request_set_replen(req); + ptlrpc_request_set_replen(req); - req->rq_send_state = LUSTRE_IMP_FULL; - /* ptlrpc_request_pack already set timeout */ - if (AT_OFF) - req->rq_timeout = ldlm_get_rq_timeout(); + req->rq_send_state = LUSTRE_IMP_FULL; + /* ptlrpc_request_pack already set timeout */ + if (AT_OFF) + req->rq_timeout = ldlm_get_rq_timeout(); - /* We only send real blocking ASTs after the lock is granted */ - lock_res_and_lock(lock); + /* We only send real blocking ASTs after the lock is granted */ + lock_res_and_lock(lock); if (ldlm_is_ast_sent(lock)) { body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT); /* Copy AST flags like LDLM_FL_DISCARD_DATA. */ body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK); - /* We might get here prior to ldlm_handle_enqueue setting - * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock - * into waiting list, but this is safe and similar code in - * ldlm_handle_enqueue will call ldlm_lock_cancel() still, - * that would not only cancel the lock, but will also remove - * it from waiting list */ + /* + * We might get here prior to ldlm_handle_enqueue setting + * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock + * into waiting list, but this is safe and similar code in + * ldlm_handle_enqueue will call ldlm_lock_cancel() still, + * that would not only cancel the lock, but will also remove + * it from waiting list + */ if (ldlm_is_cancel_on_block(lock)) { unlock_res_and_lock(lock); ldlm_lock_cancel(lock); @@ -1052,18 +1094,18 @@ 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; } - } - unlock_res_and_lock(lock); + } + unlock_res_and_lock(lock); - 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_CP_CALLBACK - LDLM_FIRST_OPC); + 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_CP_CALLBACK - LDLM_FIRST_OPC); rc = ldlm_ast_fini(req, arg, lock, instant_cancel); @@ -1078,79 +1120,82 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data) */ int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data) { - struct ldlm_cb_set_arg *arg = data; - struct ldlm_request *body; - struct ptlrpc_request *req; - struct ldlm_cb_async_args *ca; - int rc; - struct req_format *req_fmt; - ENTRY; + struct ldlm_cb_set_arg *arg = data; + struct ldlm_request *body; + struct ptlrpc_request *req; + struct ldlm_cb_async_args *ca; + int rc; + struct req_format *req_fmt; + + ENTRY; - LASSERT(lock != NULL); + LASSERT(lock != NULL); 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; - req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse, + req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse, req_fmt, LUSTRE_DLM_VERSION, LDLM_GL_CALLBACK); - if (req == NULL) - RETURN(-ENOMEM); + if (req == NULL) + RETURN(-ENOMEM); if (arg->gl_desc != NULL) { /* copy the GL descriptor */ union ldlm_gl_desc *desc; + desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC); *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 = ptlrpc_req_async_args(ca, 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) { - int rc; + int rc; + ENTRY; rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list, LDLM_WORK_GL_AST); if (rc == -ERESTART) - ldlm_reprocess_all(res); + ldlm_reprocess_all(res, NULL); RETURN(rc); } @@ -1159,11 +1204,12 @@ EXPORT_SYMBOL(ldlm_glimpse_locks); /* return LDLM lock associated with a lock callback request */ struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req) { - struct ldlm_cb_async_args *ca; - struct ldlm_lock *lock; + struct ldlm_cb_async_args *ca; + struct ldlm_lock *lock; + ENTRY; - ca = ptlrpc_req_async_args(req); + ca = ptlrpc_req_async_args(ca, req); lock = ca->ca_lock; if (lock == NULL) RETURN(ERR_PTR(-EFAULT)); @@ -1172,40 +1218,6 @@ 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. @@ -1222,6 +1234,8 @@ 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"); @@ -1231,63 +1245,53 @@ 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 && - req->rq_export->exp_nid_stats->nid_ldlm_stats) - lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats, - LDLM_ENQUEUE - LDLM_FIRST_OPC); - - if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE || - dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) { - DEBUG_REQ(D_ERROR, req, "invalid lock request type %d", - dlm_req->lock_desc.l_resource.lr_type); - GOTO(out, rc = -EFAULT); - } - - if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE || - dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE || - dlm_req->lock_desc.l_req_mode & - (dlm_req->lock_desc.l_req_mode-1))) { - DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d", - dlm_req->lock_desc.l_req_mode); - 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 (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_ENQUEUE - LDLM_FIRST_OPC); + + if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE || + dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) { + DEBUG_REQ(D_ERROR, req, "invalid lock request type %d", + dlm_req->lock_desc.l_resource.lr_type); + GOTO(out, rc = -EFAULT); + } + + if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE || + dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE || + dlm_req->lock_desc.l_req_mode & + (dlm_req->lock_desc.l_req_mode-1))) { + DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d", + dlm_req->lock_desc.l_req_mode); + GOTO(out, rc = -EFAULT); + } 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 */ - /* In the function below, .hs_keycmp resolves to - * ldlm_export_lock_keycmp() */ + /* Find an existing lock in the per-export lock hash */ + /* + * In the function below, .hs_keycmp resolves to + * ldlm_export_lock_keycmp() + */ /* coverity[overrun-buffer-val] */ - 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 %#llx", + 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 %#llx", lock->l_handle.h_cookie); flags |= LDLM_FL_RESENT; - GOTO(existing_lock, rc = 0); + GOTO(existing_lock, rc = 0); } } else { if (ldlm_reclaim_full()) { - DEBUG_REQ(D_DLMTRACE, req, "Too many granted locks, " - "reject current enqueue request and let the " - "client retry later.\n"); + DEBUG_REQ(D_DLMTRACE, req, + "Too many granted locks, reject current enqueue request and let the client retry later"); GOTO(out, rc = -EINPROGRESS); } } @@ -1303,10 +1307,11 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns, GOTO(out, rc); } - lock->l_remote_handle = dlm_req->lock_handle[0]; - LDLM_DEBUG(lock, "server-side enqueue handler, new lock created"); + lock->l_remote_handle = dlm_req->lock_handle[0]; + LDLM_DEBUG(lock, "server-side enqueue handler, new lock created"); - /* Initialize resource lvb but not for a lock being replayed since + /* + * Initialize resource lvb but not for a lock being replayed since * Client already got lvb sent in this case. * This must occur early since some policy methods assume resource * lvb is available (lr_lvb_data != NULL). @@ -1321,25 +1326,29 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns, } } - OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2); - /* Don't enqueue a lock onto the export if it is been disonnected - * due to eviction (bug 3822) or server umount (bug 24324). - * Cancel it now instead. */ - if (req->rq_export->exp_disconnected) { - LDLM_ERROR(lock, "lock on disconnected export %p", - req->rq_export); - GOTO(out, rc = -ENOTCONN); - } + OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2); + /* + * Don't enqueue a lock onto the export if it is been disonnected + * due to eviction (b=3822) or server umount (b=24324). + * Cancel it now instead. + */ + if (req->rq_export->exp_disconnected) { + LDLM_ERROR(lock, "lock on disconnected export %p", + req->rq_export); + GOTO(out, rc = -ENOTCONN); + } - lock->l_export = class_export_lock_get(req->rq_export, lock); - if (lock->l_export->exp_lock_hash) - cfs_hash_add(lock->l_export->exp_lock_hash, - &lock->l_remote_handle, - &lock->l_exp_hash); + lock->l_export = class_export_lock_get(req->rq_export, lock); + if (lock->l_export->exp_lock_hash) + cfs_hash_add(lock->l_export->exp_lock_hash, + &lock->l_remote_handle, + &lock->l_exp_hash); - /* Inherit the enqueue flags before the operation, because we do not + /* + * Inherit the enqueue flags before the operation, because we do not * keep the res lock on return and next operations (BL AST) may proceed - * without them. */ + * without them. + */ lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags & LDLM_FL_INHERIT_MASK); @@ -1349,156 +1358,176 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns, &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; + else if (dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) + lock->l_policy_data.l_inodebits.try_bits = + dlm_req->lock_desc.l_policy_data.l_inodebits.try_bits; 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. */ - cookie = req; - } else { - /* based on the assumption that lvb size never changes during - * resource life time otherwise it need resource->lr_lock's - * protection */ + cookie = req; + if (!(flags & LDLM_FL_HAS_INTENT)) { + /* based on the assumption that lvb size never changes during + * resource life time otherwise it need resource->lr_lock's + * protection */ 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); + 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); - } + 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); - /* We never send a blocking AST until the lock is granted, but - * we can tell it right now */ - lock_res_and_lock(lock); + /* + * We never send a blocking AST until the lock is granted, but + * we can tell it right now + */ + lock_res_and_lock(lock); - /* Now take into account flags to be inherited from original lock - request both in reply to client and in our own lock flags. */ + /* + * Now take into account flags to be inherited from original lock + * request both in reply to client and in our own lock flags. + */ dlm_rep->lock_flags = ldlm_flags_to_wire(flags); lock->l_flags |= flags & LDLM_FL_INHERIT_MASK; - /* Don't move a pending lock onto the export if it has already been - * disconnected due to eviction (bug 5683) or server umount (bug 24324). - * Cancel it now instead. */ - if (unlikely(req->rq_export->exp_disconnected || - OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) { - LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export); - rc = -ENOTCONN; + /* + * Don't move a pending lock onto the export if it has already been + * disconnected due to eviction (b=5683) or server umount (b=24324). + * Cancel it now instead. + */ + if (unlikely(req->rq_export->exp_disconnected || + OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) { + 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) { - /* - * Only cancel lock if it was granted, because it would - * be destroyed immediately and would never be granted - * in the future, causing timeouts on client. Not - * granted lock will be cancelled immediately after - * sending completion AST. - */ - if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) { - 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 %#llx\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 (%llu)\n", - ldlm_it2str(it->opc), it->opc); - } + if (ldlm_is_granted(lock)) { + /* + * Only cancel lock if it was granted, because it would + * be destroyed immediately and would never be granted + * in the future, causing timeouts on client. Not + * granted lock will be cancelled immediately after + * sending completion AST. + */ + if (ldlm_is_cancel_on_block(lock)) { + unlock_res_and_lock(lock); + ldlm_lock_cancel(lock); + lock_res_and_lock(lock); + } else { + ldlm_add_waiting_lock(lock, + ldlm_bl_timeout(lock)); } - } - } + } + } + unlock_res_and_lock(lock); - unlock_res_and_lock(lock); + EXIT; +out: + req->rq_status = rc ?: err; /* return either error - b=11190 */ + if (!req->rq_packed_final) { + err = lustre_pack_reply(req, 1, NULL, NULL); + if (rc == 0) + rc = err; + } - EXIT; - out: - req->rq_status = rc ?: err; /* return either error - bug 11190 */ - if (!req->rq_packed_final) { - err = lustre_pack_reply(req, 1, NULL, NULL); - if (rc == 0) - rc = err; - } - - /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this - * ldlm_reprocess_all. If this moves, revisit that code. -phil */ + /* + * The LOCK_CHANGED code in ldlm_lock_enqueue depends on this + * ldlm_reprocess_all. If this moves, revisit that code. -phil + */ if (lock != NULL) { - 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( + LDLM_DEBUG(lock, + "server-side enqueue handler, sending reply (err=%d, rc=%d)", + err, rc); + + 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(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; } } @@ -1514,114 +1543,127 @@ 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, lock); - 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; + 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) -{ - struct ldlm_request *dlm_req; - struct ldlm_callback_suite cbs = { - .lcs_completion = completion_callback, - .lcs_blocking = blocking_callback, - .lcs_glimpse = glimpse_callback - }; - int rc; - - 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; +void ldlm_clear_blocking_lock(struct ldlm_lock *lock) +{ + if (lock->l_blocking_lock) { + LDLM_LOCK_RELEASE(lock->l_blocking_lock); + lock->l_blocking_lock = NULL; + } } -/** - * Main LDLM entry point for server code to process lock conversion requests. +/* 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(). */ -int ldlm_handle_convert0(struct ptlrpc_request *req, - const struct ldlm_request *dlm_req) -{ - struct ldlm_reply *dlm_rep; - struct ldlm_lock *lock; - int rc; - ENTRY; - - 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); - - 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) { - req->rq_status = LUSTRE_EINVAL; - } else { - void *res = NULL; - - LDLM_DEBUG(lock, "server-side convert handler START"); - - 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; - } - } - - 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"); - - RETURN(0); +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); } /** - * Old-style main LDLM entry point for server code to process lock conversion - * requests. + * Main LDLM entry point for server code to process lock conversion requests. */ -int ldlm_handle_convert(struct ptlrpc_request *req) +int ldlm_handle_convert0(struct ptlrpc_request *req, + const struct ldlm_request *dlm_req) { - int rc; - struct ldlm_request *dlm_req; + struct obd_export *exp = req->rq_export; + struct ldlm_reply *dlm_rep; + struct ldlm_lock *lock; + __u64 bits; + __u64 new_bits; + int 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; + ENTRY; + + 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); + + 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) { + LDLM_DEBUG_NOLOCK("server lock is canceled already"); + req->rq_status = ELDLM_NO_LOCK_DATA; + RETURN(0); + } + + LDLM_DEBUG(lock, "server-side convert handler START"); + + lock_res_and_lock(lock); + bits = lock->l_policy_data.l_inodebits.bits; + new_bits = dlm_req->lock_desc.l_policy_data.l_inodebits.bits; + + if (ldlm_is_cancel(lock)) { + LDLM_DEBUG(lock, "convert on canceled lock!"); + unlock_res_and_lock(lock); + GOTO(out_put, rc = ELDLM_NO_LOCK_DATA); + } + + if (dlm_req->lock_desc.l_req_mode != lock->l_granted_mode) { + LDLM_ERROR(lock, "lock mode differs!"); + unlock_res_and_lock(lock); + GOTO(out_put, rc = -EPROTO); + } + + if (bits == new_bits) { + /* + * This can be valid situation if CONVERT RPCs are + * re-ordered. Just finish silently + */ + LDLM_DEBUG(lock, "lock is converted already!"); + unlock_res_and_lock(lock); + } else { + if (ldlm_is_waited(lock)) + ldlm_del_waiting_lock(lock); + + ldlm_clear_cbpending(lock); + lock->l_policy_data.l_inodebits.cancel_bits = 0; + ldlm_inodebits_drop(lock, bits & ~new_bits); + + ldlm_clear_blocking_data(lock); + unlock_res_and_lock(lock); + + ldlm_reprocess_all(lock->l_resource, NULL); + } + + 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); + rc = ELDLM_OK; + EXIT; +out_put: + LDLM_DEBUG(lock, "server-side convert handler END, rc = %d", rc); + LDLM_LOCK_PUT(lock); + req->rq_status = rc; + return 0; } /** @@ -1634,72 +1676,89 @@ int ldlm_request_cancel(struct ptlrpc_request *req, const struct ldlm_request *dlm_req, int first, enum lustre_at_flags flags) { - struct ldlm_resource *res, *pres = NULL; - struct ldlm_lock *lock; - int i, count, done = 0; - ENTRY; + struct ldlm_resource *res, *pres = NULL; + struct ldlm_lock *lock; + int i, count, done = 0; + unsigned int size; + + ENTRY; - count = dlm_req->lock_count ? dlm_req->lock_count : 1; - if (first >= count) - RETURN(0); + size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT); + if (size <= offsetof(struct ldlm_request, lock_handle) || + (size - offsetof(struct ldlm_request, lock_handle)) / + sizeof(struct lustre_handle) < dlm_req->lock_count) + RETURN(0); + + count = dlm_req->lock_count ? dlm_req->lock_count : 1; + if (first >= count) + RETURN(0); if (count == 1 && dlm_req->lock_handle[0].cookie == 0) RETURN(0); - /* There is no lock on the server at the replay time, - * skip lock cancelling to make replay tests to pass. */ - if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) - RETURN(0); + /* + * There is no lock on the server at the replay time, + * skip lock cancelling to make replay tests to pass. + */ + if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) + RETURN(0); - LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, " - "starting at %d", count, first); + LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, starting at %d", + count, first); - for (i = first; i < count; i++) { - lock = ldlm_handle2lock(&dlm_req->lock_handle[i]); - if (!lock) { - LDLM_DEBUG_NOLOCK("server-side cancel handler stale " - "lock (cookie %llu)", - dlm_req->lock_handle[i].cookie); - continue; - } + for (i = first; i < count; i++) { + lock = ldlm_handle2lock(&dlm_req->lock_handle[i]); + if (!lock) { + /* below message checked in replay-single.sh test_36 */ + LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (cookie %llu)", + dlm_req->lock_handle[i].cookie); + continue; + } - res = lock->l_resource; - done++; + res = lock->l_resource; + done++; - /* This code is an optimization to only attempt lock + /* + * This code is an optimization to only attempt lock * granting on the resource (that could be CPU-expensive) - * after we are done cancelling lock in that resource. */ - if (res != pres) { - if (pres != NULL) { - ldlm_reprocess_all(pres); - 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 ((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); + * after we are done cancelling lock in that resource. + */ + if (res != pres) { + if (pres != NULL) { + ldlm_reprocess_all(pres, NULL); + LDLM_RESOURCE_DELREF(pres); + ldlm_resource_putref(pres); + } + if (res != NULL) { + ldlm_resource_getref(res); + LDLM_RESOURCE_ADDREF(res); + + if (!ldlm_is_discard_data(lock)) + ldlm_lvbo_update(res, lock, + NULL, 1); + } + pres = res; + } + + 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); - LDLM_LOCK_PUT(lock); - } - if (pres != NULL) { - ldlm_reprocess_all(pres); - LDLM_RESOURCE_DELREF(pres); - ldlm_resource_putref(pres); - } - LDLM_DEBUG_NOLOCK("server-side cancel handler END"); - RETURN(done); + ldlm_lock_cancel(lock); + LDLM_LOCK_PUT(lock); + } + if (pres != NULL) { + ldlm_reprocess_all(pres, NULL); + LDLM_RESOURCE_DELREF(pres); + ldlm_resource_putref(pres); + } + LDLM_DEBUG_NOLOCK("server-side cancel handler END"); + RETURN(done); } EXPORT_SYMBOL(ldlm_request_cancel); @@ -1710,68 +1769,132 @@ EXPORT_SYMBOL(ldlm_request_cancel); */ int ldlm_handle_cancel(struct ptlrpc_request *req) { - struct ldlm_request *dlm_req; - int rc; - ENTRY; + struct ldlm_request *dlm_req; + int rc; + + ENTRY; - 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"); - RETURN(-EFAULT); - } + 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"); + RETURN(-EFAULT); + } + + if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) < + offsetof(struct ldlm_request, lock_handle[1])) + RETURN(-EPROTO); - 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_CANCEL - LDLM_FIRST_OPC); + 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_CANCEL - LDLM_FIRST_OPC); - rc = req_capsule_server_pack(&req->rq_pill); - if (rc) - RETURN(rc); + rc = req_capsule_server_pack(&req->rq_pill); + if (rc) + RETURN(rc); if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS)) req->rq_status = LUSTRE_ESTALE; - RETURN(ptlrpc_reply(req)); + RETURN(ptlrpc_reply(req)); } #endif /* HAVE_SERVER_SUPPORT */ /** + * Server may pass additional information about blocking lock. + * For IBITS locks it is conflicting bits which can be used for + * lock convert instead of cancel. + */ +void ldlm_bl_desc2lock(const struct ldlm_lock_desc *ld, struct ldlm_lock *lock) +{ + struct ldlm_namespace *ns = ldlm_lock_to_ns(lock); + + check_res_locked(lock->l_resource); + if (ns_is_client(ns) && ld && + (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 server may replace lock resource keeping the same cookie, + * never use cancel bits from different resource, full cancel + * is to be used. + */ + if (ld->l_policy_data.l_inodebits.cancel_bits && + ldlm_res_eq(&ld->l_resource.lr_name, + &lock->l_resource->lr_name) && + !(ldlm_is_cbpending(lock) && + lock->l_policy_data.l_inodebits.cancel_bits == 0)) { + /* always combine conflicting ibits */ + lock->l_policy_data.l_inodebits.cancel_bits |= + ld->l_policy_data.l_inodebits.cancel_bits; + } else { + /* If cancel_bits are not obtained or + * if the lock is already CBPENDING and + * has no cancel_bits set + * - the full lock is to be cancelled + */ + lock->l_policy_data.l_inodebits.cancel_bits = 0; + } + } +} + +/** * Callback handler for receiving incoming blocking ASTs. * * This can only happen on client side. */ void ldlm_handle_bl_callback(struct ldlm_namespace *ns, - struct ldlm_lock_desc *ld, struct ldlm_lock *lock) + struct ldlm_lock_desc *ld, struct ldlm_lock *lock) { - int do_ast; - ENTRY; + int do_ast; - LDLM_DEBUG(lock, "client blocking AST callback handler"); + ENTRY; + + LDLM_DEBUG(lock, "client blocking AST callback handler"); - lock_res_and_lock(lock); + lock_res_and_lock(lock); + + /* get extra information from desc if any */ + ldlm_bl_desc2lock(ld, lock); ldlm_set_cbpending(lock); - if (ldlm_is_cancel_on_block(lock)) - ldlm_set_cancel(lock); + do_ast = (!lock->l_readers && !lock->l_writers); + unlock_res_and_lock(lock); - do_ast = (!lock->l_readers && !lock->l_writers); - unlock_res_and_lock(lock); + if (do_ast) { + CDEBUG(D_DLMTRACE, + "Lock %p already unused, calling callback (%p)\n", + lock, lock->l_blocking_ast); + if (lock->l_blocking_ast != NULL) + lock->l_blocking_ast(lock, ld, lock->l_ast_data, + LDLM_CB_BLOCKING); + } else { + CDEBUG(D_DLMTRACE, + "Lock %p is referenced, will be cancelled later\n", + lock); + } - if (do_ast) { - CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n", - lock, lock->l_blocking_ast); - if (lock->l_blocking_ast != NULL) - lock->l_blocking_ast(lock, ld, lock->l_ast_data, - LDLM_CB_BLOCKING); - } else { - CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n", - lock); - } + LDLM_DEBUG(lock, "client blocking callback handler END"); + LDLM_LOCK_RELEASE(lock); + EXIT; +} + +static int ldlm_callback_reply(struct ptlrpc_request *req, int rc) +{ + if (req->rq_no_reply) + return 0; - LDLM_DEBUG(lock, "client blocking callback handler END"); - LDLM_LOCK_RELEASE(lock); - EXIT; + req->rq_status = rc; + if (!req->rq_packed_final) { + rc = lustre_pack_reply(req, 1, NULL, NULL); + if (rc) + return rc; + } + return ptlrpc_reply(req); } /** @@ -1779,25 +1902,27 @@ void ldlm_handle_bl_callback(struct ldlm_namespace *ns, * * This only can happen on client side. */ -static void ldlm_handle_cp_callback(struct ptlrpc_request *req, +static int ldlm_handle_cp_callback(struct ptlrpc_request *req, struct ldlm_namespace *ns, struct ldlm_request *dlm_req, struct ldlm_lock *lock) { - struct list_head ast_list; + LIST_HEAD(ast_list); int lvb_len; int rc = 0; + ENTRY; LDLM_DEBUG(lock, "client completion callback handler START"); - 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); + + ldlm_callback_reply(req, 0); + while (to > 0) { - set_current_state(TASK_INTERRUPTIBLE); - schedule_timeout(to); - if (lock->l_granted_mode == lock->l_req_mode || + schedule_timeout_interruptible(to); + if (ldlm_is_granted(lock) || ldlm_is_destroyed(lock)) break; } @@ -1813,9 +1938,8 @@ static void ldlm_handle_cp_callback(struct ptlrpc_request *req, LASSERT(lock->l_lvb_data != NULL); if (unlikely(lock->l_lvb_len < lvb_len)) { - LDLM_ERROR(lock, "Replied LVB is larger than " - "expectation, expected = %d, " - "replied = %d", + LDLM_ERROR(lock, + "Replied LVB is larger than expectation, expected = %d, replied = %d", lock->l_lvb_len, lvb_len); GOTO(out, rc = -EINVAL); } @@ -1823,16 +1947,39 @@ 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_failed(lock)) { + unlock_res_and_lock(lock); + LDLM_LOCK_RELEASE(lock); + RETURN(-EINVAL); + } + if (ldlm_is_destroyed(lock) || - lock->l_granted_mode == lock->l_req_mode) { - /* bug 11300: the lock has already been granted */ + ldlm_is_granted(lock)) { + /* b=11300: the lock has already been granted */ unlock_res_and_lock(lock); LDLM_DEBUG(lock, "Double grant race happened"); GOTO(out, rc = 0); } - /* If we receive the completion AST before the actual enqueue returned, - * then we might need to switch lock modes, resources, or extents. */ + /* + * If we receive the completion AST before the actual enqueue returned, + * then we might need to switch lock modes, resources, or extents. + */ if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) { lock->l_req_mode = dlm_req->lock_desc.l_granted_mode; LDLM_DEBUG(lock, "completion AST, new lock mode"); @@ -1846,29 +1993,18 @@ 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); - } - - if (dlm_req->lock_flags & LDLM_FL_AST_SENT) { - /* BL_AST locks are not needed in LRU. - * Let ldlm_cancel_lru() be fast. */ - ldlm_lock_remove_from_lru(lock); + ldlm_resource_unlink_lock(lock); + + if (dlm_req->lock_flags & LDLM_FL_AST_SENT) { + /* + * BL_AST locks are not needed in LRU. + * Let ldlm_cancel_lru() be fast. + */ + ldlm_lock_remove_from_lru(lock); + ldlm_bl_desc2lock(&dlm_req->lock_desc, lock); lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST; - LDLM_DEBUG(lock, "completion AST includes blocking AST"); - } + LDLM_DEBUG(lock, "completion AST includes blocking AST"); + } if (lock->l_lvb_len > 0) { rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT, @@ -1879,19 +2015,21 @@ static void ldlm_handle_cp_callback(struct ptlrpc_request *req, } } - ldlm_grant_lock(lock, &ast_list); - unlock_res_and_lock(lock); + ldlm_grant_lock(lock, &ast_list); + unlock_res_and_lock(lock); - LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work"); + LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work"); - /* Let Enqueue to call osc_lock_upcall() and initialize - * l_ast_data */ - OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2); + /* + * Let Enqueue to call osc_lock_upcall() and initialize + * l_ast_data + */ + OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2); - ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST); + ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST); - LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)", - lock); + LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)", + lock); GOTO(out, rc); out: @@ -1902,6 +2040,8 @@ out: wake_up(&lock->l_waitq); } LDLM_LOCK_RELEASE(lock); + + return 0; } /** @@ -1912,61 +2052,57 @@ out: * long time. */ static void ldlm_handle_gl_callback(struct ptlrpc_request *req, - struct ldlm_namespace *ns, - struct ldlm_request *dlm_req, - struct ldlm_lock *lock) + struct ldlm_namespace *ns, + struct ldlm_request *dlm_req, + struct ldlm_lock *lock) { - int rc = -ENOSYS; - ENTRY; + struct ldlm_lock_desc *ld = &dlm_req->lock_desc; + int rc = -ENOSYS; - LDLM_DEBUG(lock, "client glimpse AST callback handler"); - - if (lock->l_glimpse_ast != NULL) - rc = lock->l_glimpse_ast(lock, req); + ENTRY; - if (req->rq_repmsg != NULL) { - ptlrpc_reply(req); - } else { - req->rq_status = rc; - ptlrpc_error(req); - } + LDLM_DEBUG(lock, "client glimpse AST callback handler"); - 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)))) { - unlock_res_and_lock(lock); - if (ldlm_bl_to_thread_lock(ns, NULL, lock)) - ldlm_handle_bl_callback(ns, NULL, lock); + if (lock->l_glimpse_ast != NULL) + rc = lock->l_glimpse_ast(lock, req); - EXIT; - return; - } - unlock_res_and_lock(lock); - LDLM_LOCK_RELEASE(lock); - EXIT; -} + if (req->rq_repmsg != NULL) { + ptlrpc_reply(req); + } else { + req->rq_status = rc; + ptlrpc_error(req); + } -static int ldlm_callback_reply(struct ptlrpc_request *req, int rc) -{ - if (req->rq_no_reply) - return 0; + lock_res_and_lock(lock); + if (lock->l_granted_mode == LCK_PW && + !lock->l_readers && !lock->l_writers && + ktime_after(ktime_get(), + ktime_add(lock->l_last_used, + ktime_set(ns->ns_dirty_age_limit, 0)))) { + unlock_res_and_lock(lock); - req->rq_status = rc; - if (!req->rq_packed_final) { - rc = lustre_pack_reply(req, 1, NULL, NULL); - if (rc) - return rc; - } - return ptlrpc_reply(req); + /* For MDS glimpse it is always DOM lock, set corresponding + * cancel_bits to perform lock convert if needed + */ + if (lock->l_resource->lr_type == LDLM_IBITS) + ld->l_policy_data.l_inodebits.cancel_bits = + MDS_INODELOCK_DOM; + if (ldlm_bl_to_thread_lock(ns, ld, lock)) + ldlm_handle_bl_callback(ns, ld, lock); + + EXIT; + return; + } + unlock_res_and_lock(lock); + LDLM_LOCK_RELEASE(lock); + EXIT; } static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi, enum ldlm_cancel_flags cancel_flags) { struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool; + ENTRY; spin_lock(&blp->blp_lock); @@ -1982,8 +2118,10 @@ static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi, wake_up(&blp->blp_waitq); - /* can not check blwi->blwi_flags as blwi could be already freed in - LCF_ASYNC mode */ + /* + * can not check blwi->blwi_flags as blwi could be already freed in + * LCF_ASYNC mode + */ if (!(cancel_flags & LCF_ASYNC)) wait_for_completion(&blwi->blwi_comp); @@ -2001,19 +2139,19 @@ static inline void init_blwi(struct ldlm_bl_work_item *blwi, INIT_LIST_HEAD(&blwi->blwi_head); if (memory_pressure_get()) - blwi->blwi_mem_pressure = 1; + blwi->blwi_mem_pressure = 1; - blwi->blwi_ns = ns; + blwi->blwi_ns = ns; blwi->blwi_flags = cancel_flags; - if (ld != NULL) - blwi->blwi_ld = *ld; - if (count) { + if (ld != NULL) + blwi->blwi_ld = *ld; + if (count) { list_add(&blwi->blwi_head, cancels); list_del_init(cancels); - blwi->blwi_count = count; - } else { - blwi->blwi_lock = lock; - } + blwi->blwi_count = count; + } else { + blwi->blwi_lock = lock; + } } /** @@ -2046,7 +2184,8 @@ static int ldlm_bl_to_thread(struct ldlm_namespace *ns, RETURN(__ldlm_bl_to_thread(blwi, cancel_flags)); } else { - /* if it is synchronous call do minimum mem alloc, as it could + /* + * if it is synchronous call do minimum mem alloc, as it could * be triggered from kernel shrinker */ struct ldlm_bl_work_item blwi; @@ -2080,90 +2219,94 @@ int ldlm_bl_thread_wakeup(void) /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */ static int ldlm_handle_setinfo(struct ptlrpc_request *req) { - struct obd_device *obd = req->rq_export->exp_obd; - char *key; - void *val; - int keylen, vallen; - int rc = -ENOSYS; - ENTRY; - - DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name); - - req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO); - - key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY); - if (key == NULL) { - DEBUG_REQ(D_IOCTL, req, "no set_info key"); - RETURN(-EFAULT); - } - keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY, - RCL_CLIENT); - val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL); - if (val == NULL) { - DEBUG_REQ(D_IOCTL, req, "no set_info val"); - RETURN(-EFAULT); - } - vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL, - RCL_CLIENT); - - /* We are responsible for swabbing contents of val */ - - if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) - /* Pass it on to mdc (the "export" in this case) */ - rc = obd_set_info_async(req->rq_svc_thread->t_env, - req->rq_export, - sizeof(KEY_HSM_COPYTOOL_SEND), - KEY_HSM_COPYTOOL_SEND, - vallen, val, NULL); - else - DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key); - - return rc; + struct obd_device *obd = req->rq_export->exp_obd; + char *key; + void *val; + int keylen, vallen; + int rc = -ENOSYS; + + ENTRY; + + DEBUG_REQ(D_HSM, req, "%s: handle setinfo", obd->obd_name); + + req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO); + + key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY); + if (key == NULL) { + DEBUG_REQ(D_IOCTL, req, "no set_info key"); + RETURN(-EFAULT); + } + keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY, + RCL_CLIENT); + val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL); + if (val == NULL) { + DEBUG_REQ(D_IOCTL, req, "no set_info val"); + RETURN(-EFAULT); + } + vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL, + RCL_CLIENT); + + /* We are responsible for swabbing contents of val */ + + if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) + /* Pass it on to mdc (the "export" in this case) */ + rc = obd_set_info_async(req->rq_svc_thread->t_env, + req->rq_export, + sizeof(KEY_HSM_COPYTOOL_SEND), + KEY_HSM_COPYTOOL_SEND, + vallen, val, NULL); + else + DEBUG_REQ(D_WARNING, req, "ignoring unknown key '%s'", key); + + return rc; } static inline void ldlm_callback_errmsg(struct ptlrpc_request *req, const char *msg, int rc, const struct lustre_handle *handle) { - DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req, - "%s: [nid %s] [rc %d] [lock %#llx]", - msg, libcfs_id2str(req->rq_peer), rc, - handle ? handle->cookie : 0); - if (req->rq_no_reply) - CWARN("No reply was sent, maybe cause bug 21636.\n"); - else if (rc) - CWARN("Send reply failed, maybe cause bug 21636.\n"); + DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req, + "%s, NID=%s lock=%#llx: rc = %d", + msg, libcfs_id2str(req->rq_peer), + handle ? handle->cookie : 0, rc); + if (req->rq_no_reply) + CWARN("No reply was sent, maybe cause b=21636.\n"); + else if (rc) + CWARN("Send reply failed, maybe cause b=21636.\n"); } /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */ static int ldlm_callback_handler(struct ptlrpc_request *req) { - struct ldlm_namespace *ns; - struct ldlm_request *dlm_req; - struct ldlm_lock *lock; - int rc; - ENTRY; + struct ldlm_namespace *ns; + struct ldlm_request *dlm_req; + struct ldlm_lock *lock; + int rc; - /* 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. */ + ENTRY; - /* do nothing for sec context finalize */ - if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI) - RETURN(0); + /* + * 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. + */ + + /* do nothing for sec context finalize */ + if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI) + RETURN(0); - req_capsule_init(&req->rq_pill, req, RCL_SERVER); + req_capsule_init(&req->rq_pill, req, RCL_SERVER); - if (req->rq_export == NULL) { - rc = ldlm_callback_reply(req, -ENOTCONN); - ldlm_callback_errmsg(req, "Operate on unconnected server", - rc, NULL); - RETURN(0); - } + if (req->rq_export == NULL) { + rc = ldlm_callback_reply(req, -ENOTCONN); + ldlm_callback_errmsg(req, "Operate on unconnected server", + rc, NULL); + RETURN(0); + } - LASSERT(req->rq_export != NULL); - LASSERT(req->rq_export->exp_obd != NULL); + LASSERT(req->rq_export != NULL); + LASSERT(req->rq_export->exp_obd != NULL); switch (lustre_msg_get_opc(req->rq_reqmsg)) { case LDLM_BL_CALLBACK: @@ -2181,94 +2324,71 @@ static int ldlm_callback_handler(struct ptlrpc_request *req) if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET)) RETURN(0); break; - case LDLM_SET_INFO: - 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)); - ldlm_callback_reply(req, -EPROTO); - RETURN(0); - } - - ns = req->rq_export->exp_obd->obd_namespace; - LASSERT(ns != NULL); - - 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) { - rc = ldlm_callback_reply(req, -EPROTO); - ldlm_callback_errmsg(req, "Operate without parameter", rc, - NULL); - RETURN(0); - } - - /* Force a known safe race, send a cancel to the server for a lock - * which the server has already started a blocking callback on. */ - if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) && - lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) { + case LDLM_SET_INFO: + rc = ldlm_handle_setinfo(req); + ldlm_callback_reply(req, rc); + RETURN(0); + default: + CERROR("unknown opcode %u\n", + lustre_msg_get_opc(req->rq_reqmsg)); + ldlm_callback_reply(req, -EPROTO); + RETURN(0); + } + + ns = req->rq_export->exp_obd->obd_namespace; + LASSERT(ns != NULL); + + 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) { + rc = ldlm_callback_reply(req, -EPROTO); + ldlm_callback_errmsg(req, "Operate without parameter", rc, + NULL); + RETURN(0); + } + + /* + * Force a known safe race, send a cancel to the server for a lock + * which the server has already started a blocking callback on. + */ + if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) && + lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) { rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0); - if (rc < 0) - CERROR("ldlm_cli_cancel: %d\n", rc); - } - - lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0); - if (!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, - &dlm_req->lock_handle[0]); - RETURN(0); - } + if (rc < 0) + CERROR("ldlm_cli_cancel: %d\n", rc); + } + + lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0); + if (!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, + &dlm_req->lock_handle[0]); + RETURN(0); + } if (ldlm_is_fail_loc(lock) && - lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) - OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE); + lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) + OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE); - /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */ - lock_res_and_lock(lock); + /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */ + lock_res_and_lock(lock); lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags & LDLM_FL_AST_MASK); if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) { - /* If somebody cancels lock and cache is already dropped, + /* + * If somebody cancels lock and cache is already dropped, * or lock is failed before cp_ast received on client, * we can tell the server we have no lock. Otherwise, we - * should send cancel after dropping the cache. */ + * 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 %llx - 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); @@ -2277,51 +2397,56 @@ static int ldlm_callback_handler(struct ptlrpc_request *req) &dlm_req->lock_handle[0]); RETURN(0); } - /* BL_AST locks are not needed in LRU. - * Let ldlm_cancel_lru() be fast. */ + /* + * BL_AST locks are not needed in LRU. + * Let ldlm_cancel_lru() be fast. + */ ldlm_lock_remove_from_lru(lock); ldlm_set_bl_ast(lock); } - unlock_res_and_lock(lock); - - /* We want the ost thread to get this reply so that it can respond - * to ost requests (write cache writeback) that might be triggered - * in the callback. - * - * But we'd also like to be able to indicate in the reply that we're - * cancelling right now, because it's unused, or have an intent result - * in the reply, so we might have to push the responsibility for sending - * the reply down into the AST handlers, alas. */ - - switch (lustre_msg_get_opc(req->rq_reqmsg)) { - case LDLM_BL_CALLBACK: - CDEBUG(D_INODE, "blocking ast\n"); - req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK); + unlock_res_and_lock(lock); + + /* + * We want the ost thread to get this reply so that it can respond + * to ost requests (write cache writeback) that might be triggered + * in the callback. + * + * But we'd also like to be able to indicate in the reply that we're + * cancelling right now, because it's unused, or have an intent result + * in the reply, so we might have to push the responsibility for sending + * the reply down into the AST handlers, alas. + */ + + switch (lustre_msg_get_opc(req->rq_reqmsg)) { + case LDLM_BL_CALLBACK: + CDEBUG(D_INODE, "blocking ast\n"); + req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK); if (!ldlm_is_cancel_on_block(lock)) { - rc = ldlm_callback_reply(req, 0); - if (req->rq_no_reply || rc) - ldlm_callback_errmsg(req, "Normal process", rc, - &dlm_req->lock_handle[0]); - } - if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock)) - ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock); - break; - case LDLM_CP_CALLBACK: - CDEBUG(D_INODE, "completion ast\n"); - req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK); - ldlm_callback_reply(req, 0); - ldlm_handle_cp_callback(req, ns, dlm_req, lock); - break; - case LDLM_GL_CALLBACK: - CDEBUG(D_INODE, "glimpse ast\n"); - req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK); - ldlm_handle_gl_callback(req, ns, dlm_req, lock); - break; - default: - LBUG(); /* checked above */ - } - - RETURN(0); + rc = ldlm_callback_reply(req, 0); + if (req->rq_no_reply || rc) + ldlm_callback_errmsg(req, "Normal process", rc, + &dlm_req->lock_handle[0]); + } + if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock)) + ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock); + break; + case LDLM_CP_CALLBACK: + CDEBUG(D_INODE, "completion ast\n"); + req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK); + rc = ldlm_handle_cp_callback(req, ns, dlm_req, lock); + if (!OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) + ldlm_callback_reply(req, rc); + break; + case LDLM_GL_CALLBACK: + CDEBUG(D_INODE, "glimpse ast\n"); + req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK); + ldlm_handle_gl_callback(req, ns, dlm_req, lock); + break; + default: + LBUG(); /* checked above */ + } + + RETURN(0); } #ifdef HAVE_SERVER_SUPPORT @@ -2332,200 +2457,230 @@ 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); - - 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); - } - - 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"); + 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); + + if (req->rq_export == NULL) { + struct ldlm_request *dlm_req; + + CERROR("%s from %s arrived at %llu with bad export cookie %llu\n", + ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)), + libcfs_nid2str(req->rq_peer.nid), + (unsigned long long)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); + } + + switch (lustre_msg_get_opc(req->rq_reqmsg)) { + /* XXX FIXME move this back to mds/handler.c, b=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; + + req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT); + CDEBUG(D_INODE, "convert\n"); - RETURN(0); + 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 %#llx.", 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; + unsigned int size; + + ENTRY; - /* no prolong in recovery */ - if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) - RETURN(0); + /* no prolong in recovery */ + if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) + RETURN(0); - dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); - if (dlm_req == NULL) - RETURN(-EFAULT); + dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); + if (dlm_req == NULL) + RETURN(-EFAULT); - for (i = 0; i < dlm_req->lock_count; i++) { - struct ldlm_lock *lock; + size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT); + if (size <= offsetof(struct ldlm_request, lock_handle) || + (size - offsetof(struct ldlm_request, lock_handle)) / + sizeof(struct lustre_handle) < dlm_req->lock_count) + RETURN(-EPROTO); - 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, struct hlist_node *hnode, void *data) { - struct list_head *rpc_list = data; - struct ldlm_lock *lock = cfs_hash_object(hs, hnode); + struct list_head *rpc_list = data; + struct ldlm_lock *lock = cfs_hash_object(hs, hnode); - lock_res_and_lock(lock); + lock_res_and_lock(lock); - if (lock->l_req_mode != lock->l_granted_mode) { - unlock_res_and_lock(lock); - return 0; - } + if (!ldlm_is_granted(lock)) { + unlock_res_and_lock(lock); + return 0; + } - LASSERT(lock->l_resource); - if (lock->l_resource->lr_type != LDLM_IBITS && - lock->l_resource->lr_type != LDLM_PLAIN) { - unlock_res_and_lock(lock); - return 0; - } + LASSERT(lock->l_resource); + if (lock->l_resource->lr_type != LDLM_IBITS && + lock->l_resource->lr_type != LDLM_PLAIN) { + unlock_res_and_lock(lock); + return 0; + } if (ldlm_is_ast_sent(lock)) { - unlock_res_and_lock(lock); - return 0; - } + unlock_res_and_lock(lock); + return 0; + } - LASSERT(lock->l_blocking_ast); - LASSERT(!lock->l_blocking_lock); + LASSERT(lock->l_blocking_ast); + LASSERT(!lock->l_blocking_lock); ldlm_set_ast_sent(lock); - if (lock->l_export && lock->l_export->exp_lock_hash) { - /* NB: it's safe to call cfs_hash_del() even lock isn't - * in exp_lock_hash. */ - /* In the function below, .hs_keycmp resolves to - * ldlm_export_lock_keycmp() */ + if (lock->l_export && lock->l_export->exp_lock_hash) { + /* + * NB: it's safe to call cfs_hash_del() even lock isn't + * in exp_lock_hash. + */ + /* + * In the function below, .hs_keycmp resolves to + * ldlm_export_lock_keycmp() + */ /* coverity[overrun-buffer-val] */ cfs_hash_del(lock->l_export->exp_lock_hash, &lock->l_remote_handle, &lock->l_exp_hash); } list_add_tail(&lock->l_rk_ast, rpc_list); - LDLM_LOCK_GET(lock); + LDLM_LOCK_GET(lock); - unlock_res_and_lock(lock); - return 0; + unlock_res_and_lock(lock); + return 0; } void ldlm_revoke_export_locks(struct obd_export *exp) { - struct list_head rpc_list; + LIST_HEAD(rpc_list); + ENTRY; - INIT_LIST_HEAD(&rpc_list); 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, @@ -2541,7 +2696,7 @@ static int ldlm_bl_get_work(struct ldlm_bl_pool *blp, struct obd_export **p_exp) { struct ldlm_bl_work_item *blwi = NULL; - static unsigned int num_bl = 0; + static unsigned int num_bl; static unsigned int num_stale; int num_th = atomic_read(&blp->blp_num_threads); @@ -2552,9 +2707,8 @@ static int ldlm_bl_get_work(struct ldlm_bl_pool *blp, if (num_th == 1 || ++num_stale < num_th) { spin_unlock(&blp->blp_lock); return 1; - } else { - num_stale = 0; } + num_stale = 0; } /* process a request from the blp_list at least every blp_num_threads */ @@ -2660,10 +2814,12 @@ static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp, if (blwi->blwi_count) { int count; - /* The special case when we cancel locks in lru + /* + * The special case when we cancel locks in lru * asynchronously, we pass the list of locks here. * Thus locks are marked LDLM_FL_CANCELING, but NOT - * canceled locally yet. */ + * canceled locally yet. + */ count = ldlm_cli_cancel_list_local(&blwi->blwi_head, blwi->blwi_count, LCF_BL_AST); @@ -2694,6 +2850,7 @@ static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp, struct obd_export *exp) { int num; + ENTRY; OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4); @@ -2717,9 +2874,22 @@ static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp, */ static int ldlm_bl_thread_main(void *arg) { - struct ldlm_bl_pool *blp; + struct lu_env *env; + struct ldlm_bl_pool *blp; struct ldlm_bl_thread_data *bltd = arg; - ENTRY; + int rc; + + ENTRY; + + OBD_ALLOC_PTR(env); + if (!env) + RETURN(-ENOMEM); + rc = lu_env_init(env, LCT_DT_THREAD); + if (rc) + GOTO(out_env, rc); + rc = lu_env_add(env); + if (rc) + GOTO(out_env_fini, rc); blp = bltd->bltd_blp; @@ -2727,7 +2897,6 @@ static int ldlm_bl_thread_main(void *arg) /* cannot use bltd after this, it is only on caller's stack */ while (1) { - struct l_wait_info lwi = { 0 }; struct ldlm_bl_work_item *blwi = NULL; struct obd_export *exp = NULL; int rc; @@ -2735,10 +2904,9 @@ static int ldlm_bl_thread_main(void *arg) rc = ldlm_bl_get_work(blp, &blwi, &exp); if (rc == 0) - l_wait_event_exclusive(blp->blp_waitq, - ldlm_bl_get_work(blp, &blwi, - &exp), - &lwi); + wait_event_idle_exclusive(blp->blp_waitq, + ldlm_bl_get_work(blp, &blwi, + &exp)); atomic_inc(&blp->blp_busy_threads); if (ldlm_bl_thread_need_create(blp, blwi)) @@ -2754,11 +2922,24 @@ 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); complete(&blp->blp_comp); - RETURN(0); + + lu_env_remove(env); +out_env_fini: + lu_env_fini(env); +out_env: + OBD_FREE_PTR(env); + RETURN(rc); } @@ -2767,68 +2948,70 @@ static int ldlm_cleanup(void); int ldlm_get_ref(void) { - int rc = 0; - ENTRY; + int rc = 0; + + ENTRY; mutex_lock(&ldlm_ref_mutex); - if (++ldlm_refcount == 1) { - rc = ldlm_setup(); - if (rc) - ldlm_refcount--; - } + if (++ldlm_refcount == 1) { + rc = ldlm_setup(); + if (rc) + ldlm_refcount--; + } mutex_unlock(&ldlm_ref_mutex); - RETURN(rc); + RETURN(rc); } void ldlm_put_ref(void) { - ENTRY; + ENTRY; mutex_lock(&ldlm_ref_mutex); - if (ldlm_refcount == 1) { - int rc = ldlm_cleanup(); - if (rc) - CERROR("ldlm_cleanup failed: %d\n", rc); - else - ldlm_refcount--; - } else { - ldlm_refcount--; - } + if (ldlm_refcount == 1) { + int rc = ldlm_cleanup(); + + if (rc) + CERROR("ldlm_cleanup failed: %d\n", rc); + else + ldlm_refcount--; + } else { + ldlm_refcount--; + } mutex_unlock(&ldlm_ref_mutex); - EXIT; + EXIT; } /* * Export handle<->lock hash operations. */ static unsigned -ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned mask) +ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned int mask) { - return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask); + return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask); } static void * ldlm_export_lock_key(struct hlist_node *hnode) { - struct ldlm_lock *lock; + struct ldlm_lock *lock; lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash); - return &lock->l_remote_handle; + return &lock->l_remote_handle; } static void ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key) { - struct ldlm_lock *lock; + struct ldlm_lock *lock; lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash); - lock->l_remote_handle = *(struct lustre_handle *)key; + lock->l_remote_handle = *(struct lustre_handle *)key; } static int ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode) { - return lustre_handle_equal(ldlm_export_lock_key(hnode), key); + return lustre_handle_equal(ldlm_export_lock_key(hnode), key); } static void * @@ -2840,55 +3023,56 @@ ldlm_export_lock_object(struct hlist_node *hnode) static void ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode) { - struct ldlm_lock *lock; + struct ldlm_lock *lock; lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash); - LDLM_LOCK_GET(lock); + LDLM_LOCK_GET(lock); } static void ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode) { - struct ldlm_lock *lock; + struct ldlm_lock *lock; lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash); - LDLM_LOCK_RELEASE(lock); + LDLM_LOCK_RELEASE(lock); } static struct cfs_hash_ops ldlm_export_lock_ops = { - .hs_hash = ldlm_export_lock_hash, - .hs_key = ldlm_export_lock_key, - .hs_keycmp = ldlm_export_lock_keycmp, - .hs_keycpy = ldlm_export_lock_keycpy, - .hs_object = ldlm_export_lock_object, - .hs_get = ldlm_export_lock_get, - .hs_put = ldlm_export_lock_put, - .hs_put_locked = ldlm_export_lock_put, + .hs_hash = ldlm_export_lock_hash, + .hs_key = ldlm_export_lock_key, + .hs_keycmp = ldlm_export_lock_keycmp, + .hs_keycpy = ldlm_export_lock_keycpy, + .hs_object = ldlm_export_lock_object, + .hs_get = ldlm_export_lock_get, + .hs_put = ldlm_export_lock_put, + .hs_put_locked = ldlm_export_lock_put, }; int ldlm_init_export(struct obd_export *exp) { int rc; - ENTRY; - exp->exp_lock_hash = - cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid), - HASH_EXP_LOCK_CUR_BITS, - HASH_EXP_LOCK_MAX_BITS, - HASH_EXP_LOCK_BKT_BITS, 0, - CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA, - &ldlm_export_lock_ops, - CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY | - CFS_HASH_NBLK_CHANGE); + ENTRY; + + exp->exp_lock_hash = + cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid), + HASH_EXP_LOCK_CUR_BITS, + HASH_EXP_LOCK_MAX_BITS, + HASH_EXP_LOCK_BKT_BITS, 0, + CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA, + &ldlm_export_lock_ops, + CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY | + CFS_HASH_NBLK_CHANGE); - if (!exp->exp_lock_hash) - RETURN(-ENOMEM); + if (!exp->exp_lock_hash) + RETURN(-ENOMEM); rc = ldlm_init_flock_export(exp); if (rc) GOTO(err, rc); - RETURN(0); + RETURN(0); err: ldlm_destroy_export(exp); RETURN(rc); @@ -2897,15 +3081,49 @@ EXPORT_SYMBOL(ldlm_init_export); void ldlm_destroy_export(struct obd_export *exp) { - ENTRY; - cfs_hash_putref(exp->exp_lock_hash); - exp->exp_lock_hash = NULL; + ENTRY; + cfs_hash_putref(exp->exp_lock_hash); + exp->exp_lock_hash = NULL; ldlm_destroy_flock_export(exp); - EXIT; + EXIT; } 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; @@ -2916,20 +3134,34 @@ static int ldlm_setup(void) int i; int rc = 0; - ENTRY; + ENTRY; + + if (ldlm_state != NULL) + RETURN(-EALREADY); + + OBD_ALLOC(ldlm_state, sizeof(*ldlm_state)); + if (ldlm_state == NULL) + RETURN(-ENOMEM); - if (ldlm_state != NULL) - RETURN(-EALREADY); + ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj); + if (!ldlm_kobj) + GOTO(out, -ENOMEM); - OBD_ALLOC(ldlm_state, sizeof(*ldlm_state)); - if (ldlm_state == NULL) - RETURN(-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); -#ifdef CONFIG_PROC_FS - rc = ldlm_proc_setup(); - if (rc != 0) + rc = ldlm_debugfs_setup(); + if (rc != 0) GOTO(out, rc); -#endif /* CONFIG_PROC_FS */ memset(&conf, 0, sizeof(conf)); conf = (typeof(conf)) { @@ -2950,18 +3182,20 @@ 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); @@ -2990,13 +3224,14 @@ 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, @@ -3004,7 +3239,8 @@ 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); @@ -3041,14 +3277,6 @@ 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); @@ -3056,8 +3284,8 @@ 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(); @@ -3080,15 +3308,15 @@ static int ldlm_setup(void) static int ldlm_cleanup(void) { - ENTRY; + ENTRY; if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) || !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) { - CERROR("ldlm still has namespaces; clean these up first.\n"); - ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE); - ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE); - RETURN(-EBUSY); - } + CERROR("ldlm still has namespaces; clean these up first.\n"); + ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE); + ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE); + RETURN(-EBUSY); + } ldlm_reclaim_cleanup(); ldlm_pools_fini(); @@ -3119,33 +3347,34 @@ 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 - OBD_FREE(ldlm_state, sizeof(*ldlm_state)); - ldlm_state = NULL; + OBD_FREE(ldlm_state, sizeof(*ldlm_state)); + ldlm_state = NULL; - RETURN(0); + RETURN(0); } 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); @@ -3154,12 +3383,12 @@ 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; ldlm_interval_slab = kmem_cache_create("interval_node", - sizeof(struct ldlm_interval), + sizeof(struct ldlm_interval), 0, SLAB_HWCACHE_ALIGN, NULL); if (ldlm_interval_slab == NULL) goto out_lock; @@ -3170,11 +3399,30 @@ int ldlm_init(void) if (ldlm_interval_tree_slab == NULL) goto out_interval; +#ifdef HAVE_SERVER_SUPPORT + ldlm_inodebits_slab = kmem_cache_create("ldlm_ibits_node", + sizeof(struct ldlm_ibits_node), + 0, SLAB_HWCACHE_ALIGN, NULL); + if (ldlm_inodebits_slab == NULL) + goto out_interval_tree; + + 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_inodebits; +#endif + #if LUSTRE_TRACKS_LOCK_EXP_REFS class_export_dump_hook = ldlm_dump_export_locks; #endif return 0; - +#ifdef HAVE_SERVER_SUPPORT +out_inodebits: + kmem_cache_destroy(ldlm_inodebits_slab); +out_interval_tree: + kmem_cache_destroy(ldlm_interval_tree_slab); +#endif out_interval: kmem_cache_destroy(ldlm_interval_slab); out_lock: @@ -3190,11 +3438,17 @@ void ldlm_exit(void) if (ldlm_refcount) CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount); kmem_cache_destroy(ldlm_resource_slab); - /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call - * synchronize_rcu() to wait a grace period elapsed, so that - * ldlm_lock_free() get a chance to be called. */ - synchronize_rcu(); + /* + * ldlm_lock_put() use RCU to call ldlm_lock_free, so need call + * rcu_barrier() to wait all outstanding RCU callbacks to complete, + * so that ldlm_lock_free() get a chance to be called. + */ + rcu_barrier(); 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_inodebits_slab); + kmem_cache_destroy(ldlm_glimpse_work_kmem); +#endif }