unsigned int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
module_param(ldlm_enqueue_min, uint, 0644);
MODULE_PARM_DESC(ldlm_enqueue_min, "lock enqueue timeout minimum");
+EXPORT_SYMBOL(ldlm_enqueue_min);
/* in client side, whether the cached locks will be canceled before replay */
unsigned int ldlm_cancel_unused_locks_before_replay = 1;
-static void interrupted_completion_wait(void *data)
-{
-}
-
struct lock_wait_data {
struct ldlm_lock *lwd_lock;
__u32 lwd_conn_cnt;
return sizeof(struct ldlm_request) + avail;
}
-int ldlm_expired_completion_wait(void *data)
+void ldlm_expired_completion_wait(struct lock_wait_data *lwd)
{
- struct lock_wait_data *lwd = data;
struct ldlm_lock *lock = lwd->lwd_lock;
struct obd_import *imp;
struct obd_device *obd;
LDLM_ERROR(lock,
"lock timed out (enqueued at %lld, %llds ago); not entering recovery in server code, just going back to sleep",
- (s64)lock->l_activity,
- (s64)(ktime_get_real_seconds() -
- lock->l_activity));
+ lock->l_activity,
+ ktime_get_real_seconds() - lock->l_activity);
if (ktime_get_seconds() > next_dump) {
last_dump = next_dump;
next_dump = ktime_get_seconds() + 300;
if (last_dump == 0)
libcfs_debug_dumplog();
}
- RETURN(0);
+ RETURN_EXIT;
}
obd = lock->l_conn_export->exp_obd;
ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
LDLM_ERROR(lock,
"lock timed out (enqueued at %lld, %llds ago), entering recovery for %s@%s",
- (s64)lock->l_activity,
- (s64)(ktime_get_real_seconds() - lock->l_activity),
+ lock->l_activity,
+ ktime_get_real_seconds() - lock->l_activity,
obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
- RETURN(0);
+ EXIT;
}
int is_granted_or_cancelled_nolock(struct ldlm_lock *lock)
* We use the same basis for both server side and client side functions
* from a single node.
*/
-static time64_t ldlm_cp_timeout(struct ldlm_lock *lock)
+static timeout_t ldlm_cp_timeout(struct ldlm_lock *lock)
{
- time64_t timeout;
+ timeout_t timeout;
if (AT_OFF)
return obd_timeout;
* doesn't respond reasonably, and then give us the lock.
*/
timeout = at_get(ldlm_lock_to_ns_at(lock));
- return max(3 * timeout, (time64_t) ldlm_enqueue_min);
+ return max(3 * timeout, (timeout_t)ldlm_enqueue_min);
}
/**
*/
static int ldlm_completion_tail(struct ldlm_lock *lock, void *data)
{
- time64_t delay;
int result = 0;
if (ldlm_is_destroyed(lock) || ldlm_is_failed(lock)) {
LDLM_DEBUG(lock, "client-side enqueue: granted");
} else {
/* Take into AT only CP RPC, not immediately granted locks */
- delay = ktime_get_real_seconds() - lock->l_activity;
- LDLM_DEBUG(lock, "client-side enqueue: granted after %llds",
- (s64)delay);
+ timeout_t delay = 0;
+
+ /* Discard negative timeouts. We should also limit the
+ * maximum value of the timeout
+ */
+ if (ktime_get_real_seconds() > lock->l_activity)
+ delay = ktime_get_real_seconds() - lock->l_activity;
+ LDLM_DEBUG(lock, "client-side enqueue: granted after %ds",
+ delay);
/* Update our time estimate */
at_measured(ldlm_lock_to_ns_at(lock), delay);
}
struct lock_wait_data lwd;
struct obd_device *obd;
struct obd_import *imp = NULL;
- struct l_wait_info lwi;
- time64_t timeout;
+ timeout_t timeout;
int rc = 0;
ENTRY;
lwd.lwd_lock = lock;
lock->l_activity = ktime_get_real_seconds();
- if (ldlm_is_no_timeout(lock)) {
- LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
- lwi = LWI_INTR(interrupted_completion_wait, &lwd);
- } else {
- lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
- ldlm_expired_completion_wait,
- interrupted_completion_wait, &lwd);
- }
-
if (imp != NULL) {
spin_lock(&imp->imp_lock);
lwd.lwd_conn_cnt = imp->imp_conn_cnt;
rc = -EINTR;
} else {
/* Go to sleep until the lock is granted or cancelled. */
- rc = l_wait_event(lock->l_waitq,
- is_granted_or_cancelled(lock), &lwi);
+ if (ldlm_is_no_timeout(lock)) {
+ LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
+ rc = l_wait_event_abortable(
+ lock->l_waitq,
+ is_granted_or_cancelled(lock));
+ } else {
+ if (wait_event_idle_timeout(
+ lock->l_waitq,
+ is_granted_or_cancelled(lock),
+ cfs_time_seconds(timeout)) == 0) {
+ ldlm_expired_completion_wait(&lwd);
+ rc = l_wait_event_abortable(
+ lock->l_waitq,
+ is_granted_or_cancelled(lock));
+ }
+ }
}
if (rc) {
}
}
+static bool ldlm_request_slot_needed(enum ldlm_type type)
+{
+ return type == LDLM_FLOCK || type == LDLM_IBITS;
+}
+
/**
* Finishing portion of client lock enqueue code.
*
ENTRY;
+ if (ldlm_request_slot_needed(type))
+ obd_put_request_slot(&req->rq_import->imp_obd->u.cli);
+
+ ptlrpc_put_mod_rpc_slot(req);
+
if (req && req->rq_svc_thread)
env = req->rq_svc_thread->t_env;
if ((*flags) & LDLM_FL_AST_SENT) {
lock_res_and_lock(lock);
+ ldlm_bl_desc2lock(&reply->lock_desc, lock);
lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
unlock_res_and_lock(lock);
LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
struct req_capsule *pill = &req->rq_pill;
struct ldlm_request *dlm = NULL;
- struct list_head head = LIST_HEAD_INIT(head);
+ LIST_HEAD(head);
enum ldlm_lru_flags lru_flags;
int avail, to_free, pack = 0;
int rc;
LDLM_GLIMPSE_ENQUEUE);
}
+ /* It is important to obtain modify RPC slot first (if applicable), so
+ * that threads that are waiting for a modify RPC slot are not polluting
+ * our rpcs in flight counter. */
+
+ if (einfo->ei_enq_slot)
+ ptlrpc_get_mod_rpc_slot(req);
+
+ if (ldlm_request_slot_needed(einfo->ei_type)) {
+ rc = obd_get_request_slot(&req->rq_import->imp_obd->u.cli);
+ if (rc) {
+ if (einfo->ei_enq_slot)
+ ptlrpc_put_mod_rpc_slot(req);
+ failed_lock_cleanup(ns, lock, einfo->ei_mode);
+ LDLM_LOCK_RELEASE(lock);
+ GOTO(out, rc);
+ }
+ }
+
if (async) {
LASSERT(reqp != NULL);
RETURN(0);
else
rc = err;
+out:
if (!req_passed_in && req != NULL) {
ptlrpc_req_finished(req);
if (reqp)
EXPORT_SYMBOL(ldlm_cli_enqueue);
/**
- * Client-side lock convert reply handling.
- *
- * Finish client lock converting, checks for concurrent converts
- * and clear 'converting' flag so lock can be placed back into LRU.
- */
-static int lock_convert_interpret(const struct lu_env *env,
- struct ptlrpc_request *req,
- void *args, int rc)
-{
- struct ldlm_async_args *aa = args;
- struct ldlm_lock *lock;
- struct ldlm_reply *reply;
-
- ENTRY;
-
- lock = ldlm_handle2lock(&aa->lock_handle);
- if (!lock) {
- LDLM_DEBUG_NOLOCK("convert ACK for unknown local cookie %#llx",
- aa->lock_handle.cookie);
- RETURN(-ESTALE);
- }
-
- LDLM_DEBUG(lock, "CONVERTED lock:");
-
- if (rc != ELDLM_OK)
- GOTO(out, rc);
-
- reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
- if (reply == NULL)
- GOTO(out, rc = -EPROTO);
-
- if (reply->lock_handle.cookie != aa->lock_handle.cookie) {
- LDLM_ERROR(lock,
- "convert ACK with wrong lock cookie %#llx but cookie %#llx from server %s id %s\n",
- aa->lock_handle.cookie, reply->lock_handle.cookie,
- req->rq_export->exp_client_uuid.uuid,
- libcfs_id2str(req->rq_peer));
- GOTO(out, rc = ELDLM_NO_LOCK_DATA);
- }
-
- lock_res_and_lock(lock);
- /*
- * Lock convert is sent for any new bits to drop, the converting flag
- * is dropped when ibits on server are the same as on client. Meanwhile
- * that can be so that more later convert will be replied first with
- * and clear converting flag, so in case of such race just exit here.
- * if lock has no converting bits then
- */
- if (!ldlm_is_converting(lock)) {
- LDLM_DEBUG(lock,
- "convert ACK for lock without converting flag, reply ibits %#llx",
- reply->lock_desc.l_policy_data.l_inodebits.bits);
- } else if (reply->lock_desc.l_policy_data.l_inodebits.bits !=
- lock->l_policy_data.l_inodebits.bits) {
- /*
- * Compare server returned lock ibits and local lock ibits
- * if they are the same we consider convertion is done,
- * otherwise we have more converts inflight and keep
- * converting flag.
- */
- LDLM_DEBUG(lock, "convert ACK with ibits %#llx\n",
- reply->lock_desc.l_policy_data.l_inodebits.bits);
- } else {
- ldlm_clear_converting(lock);
-
- /*
- * Concurrent BL AST may arrive and cause another convert
- * or cancel so just do nothing here if bl_ast is set,
- * finish with convert otherwise.
- */
- if (!ldlm_is_bl_ast(lock)) {
- struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
-
- /*
- * Drop cancel_bits since there are no more converts
- * and put lock into LRU if it is still not used and
- * is not there yet.
- */
- lock->l_policy_data.l_inodebits.cancel_bits = 0;
- if (!lock->l_readers && !lock->l_writers &&
- !ldlm_is_canceling(lock)) {
- spin_lock(&ns->ns_lock);
- /* there is check for list_empty() inside */
- ldlm_lock_remove_from_lru_nolock(lock);
- ldlm_lock_add_to_lru_nolock(lock);
- spin_unlock(&ns->ns_lock);
- }
- }
- }
- unlock_res_and_lock(lock);
-out:
- if (rc) {
- int flag;
-
- lock_res_and_lock(lock);
- if (ldlm_is_converting(lock)) {
- ldlm_clear_converting(lock);
- ldlm_set_cbpending(lock);
- ldlm_set_bl_ast(lock);
- lock->l_policy_data.l_inodebits.cancel_bits = 0;
- }
- unlock_res_and_lock(lock);
-
- /*
- * fallback to normal lock cancel. If rc means there is no
- * valid lock on server, do only local cancel
- */
- if (rc == ELDLM_NO_LOCK_DATA)
- flag = LCF_LOCAL;
- else
- flag = LCF_ASYNC;
-
- rc = ldlm_cli_cancel(&aa->lock_handle, flag);
- if (rc < 0)
- LDLM_DEBUG(lock, "failed to cancel lock: rc = %d\n",
- rc);
- }
- LDLM_LOCK_PUT(lock);
- RETURN(rc);
-}
-
-/**
* Client-side IBITS lock convert.
*
* Inform server that lock has been converted instead of canceling.
* is made asynchronous.
*
*/
-int ldlm_cli_convert(struct ldlm_lock *lock, __u32 *flags)
+int ldlm_cli_convert_req(struct ldlm_lock *lock, __u32 *flags, __u64 new_bits)
{
struct ldlm_request *body;
struct ptlrpc_request *req;
- struct ldlm_async_args *aa;
struct obd_export *exp = lock->l_conn_export;
ENTRY;
- if (exp == NULL) {
- LDLM_ERROR(lock, "convert must not be called on local locks.");
- RETURN(-EINVAL);
- }
+ LASSERT(exp != NULL);
/*
* this is better to check earlier and it is done so already,
body->lock_desc.l_req_mode = lock->l_req_mode;
body->lock_desc.l_granted_mode = lock->l_granted_mode;
- body->lock_desc.l_policy_data.l_inodebits.bits =
- lock->l_policy_data.l_inodebits.bits;
+ body->lock_desc.l_policy_data.l_inodebits.bits = new_bits;
body->lock_desc.l_policy_data.l_inodebits.cancel_bits = 0;
body->lock_flags = ldlm_flags_to_wire(*flags);
lprocfs_counter_incr(exp->exp_obd->obd_svc_stats,
LDLM_CONVERT - LDLM_FIRST_OPC);
- aa = ptlrpc_req_async_args(aa, req);
- ldlm_lock2handle(lock, &aa->lock_handle);
- req->rq_interpret_reply = lock_convert_interpret;
-
ptlrpcd_add_req(req);
RETURN(0);
}
bool local_only;
LDLM_DEBUG(lock, "client-side cancel");
+ OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL_LOCAL,
+ cfs_fail_val);
+
/* Set this flag to prevent others from getting new references*/
lock_res_and_lock(lock);
ldlm_set_cbpending(lock);
new_slv = lustre_msg_get_slv(req->rq_repmsg);
obd = req->rq_import->imp_obd;
+ read_lock(&obd->obd_pool_lock);
+ if (obd->obd_pool_slv == new_slv &&
+ obd->obd_pool_limit == new_limit) {
+ read_unlock(&obd->obd_pool_lock);
+ RETURN(0);
+ }
+ read_unlock(&obd->obd_pool_lock);
+
/*
* Set new SLV and limit in OBD fields to make them accessible
* to the pool thread. We do not access obd_namespace and pool
RETURN(0);
}
+int ldlm_cli_convert(struct ldlm_lock *lock,
+ enum ldlm_cancel_flags cancel_flags)
+{
+ int rc = -EINVAL;
+
+ LASSERT(!lock->l_readers && !lock->l_writers);
+ LDLM_DEBUG(lock, "client lock convert START");
+
+ if (lock->l_resource->lr_type == LDLM_IBITS) {
+ lock_res_and_lock(lock);
+ do {
+ rc = ldlm_cli_inodebits_convert(lock, cancel_flags);
+ } while (rc == -EAGAIN);
+ unlock_res_and_lock(lock);
+ }
+
+ LDLM_DEBUG(lock, "client lock convert END");
+ RETURN(rc);
+}
+EXPORT_SYMBOL(ldlm_cli_convert);
+
/**
* Client side lock cancel.
*
__u64 rc = 0;
struct ldlm_namespace *ns;
struct ldlm_lock *lock;
- struct list_head cancels = LIST_HEAD_INIT(cancels);
+ LIST_HEAD(cancels);
ENTRY;
RETURN(0);
}
- /* Convert lock bits instead of cancel for IBITS locks */
- if (cancel_flags & LCF_CONVERT) {
- LASSERT(lock->l_resource->lr_type == LDLM_IBITS);
- LASSERT(lock->l_policy_data.l_inodebits.cancel_bits != 0);
-
- rc = ldlm_cli_dropbits(lock,
- lock->l_policy_data.l_inodebits.cancel_bits);
- if (rc == 0) {
- LDLM_LOCK_RELEASE(lock);
- RETURN(0);
- }
- }
-
lock_res_and_lock(lock);
+ LASSERT(!ldlm_is_converting(lock));
+
/* Lock is being canceled and the caller doesn't want to wait */
if (ldlm_is_canceling(lock)) {
if (cancel_flags & LCF_ASYNC) {
RETURN(0);
}
- /*
- * Lock is being converted, cancel it immediately.
- * When convert will end, it releases lock and it will be gone.
- */
- if (ldlm_is_converting(lock)) {
- /* set back flags removed by convert */
- ldlm_set_cbpending(lock);
- ldlm_set_bl_ast(lock);
- }
-
ldlm_set_canceling(lock);
unlock_res_and_lock(lock);
int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
enum ldlm_cancel_flags cancel_flags)
{
- struct list_head head = LIST_HEAD_INIT(head);
+ LIST_HEAD(head);
struct ldlm_lock *lock, *next;
int left = 0, bl_ast = 0;
__u64 rc;
/* No locks which got blocking requests. */
LASSERT(!ldlm_is_bl_ast(lock));
- if (!ldlm_is_canceling(lock) &&
- !ldlm_is_converting(lock))
+ if (!ldlm_is_canceling(lock))
break;
/*
lock_res_and_lock(lock);
/* Check flags again under the lock. */
- if (ldlm_is_canceling(lock) || ldlm_is_converting(lock) ||
+ if (ldlm_is_canceling(lock) ||
ldlm_lock_remove_from_lru_check(lock, last_use) == 0) {
/*
* Another thread is removing lock from LRU, or
enum ldlm_cancel_flags cancel_flags,
enum ldlm_lru_flags lru_flags)
{
- struct list_head cancels = LIST_HEAD_INIT(cancels);
+ LIST_HEAD(cancels);
int count, rc;
ENTRY;
continue;
/*
- * If somebody is already doing CANCEL, or blocking AST came,
- * or lock is being converted then skip this lock.
+ * If somebody is already doing CANCEL, or blocking AST came
+ * then skip this lock.
*/
- if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock) ||
- ldlm_is_converting(lock))
+ if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock))
continue;
if (lockmode_compat(lock->l_granted_mode, mode))
/* See CBPENDING comment in ldlm_cancel_lru */
lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
lock_flags;
-
LASSERT(list_empty(&lock->l_bl_ast));
list_add(&lock->l_bl_ast, cancels);
LDLM_LOCK_GET(lock);
enum ldlm_cancel_flags flags, void *opaque)
{
struct ldlm_resource *res;
- struct list_head cancels = LIST_HEAD_INIT(cancels);
+ LIST_HEAD(cancels);
int count;
int rc;
ENTRY;
atomic_dec(&req->rq_import->imp_replay_inflight);
+ wake_up(&req->rq_import->imp_replay_waitq);
+
if (rc != ELDLM_OK)
GOTO(out, rc);
LDLM_DEBUG(lock, "replaying lock:");
- atomic_inc(&req->rq_import->imp_replay_inflight);
+ atomic_inc(&imp->imp_replay_inflight);
aa = ptlrpc_req_async_args(aa, req);
aa->lock_handle = body->lock_handle[0];
req->rq_interpret_reply = replay_lock_interpret;
static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
{
int canceled;
- struct list_head cancels = LIST_HEAD_INIT(cancels);
+ LIST_HEAD(cancels);
CDEBUG(D_DLMTRACE,
"Dropping as many unused locks as possible before replay for namespace %s (%d)\n",
canceled, ldlm_ns_name(ns));
}
-int ldlm_replay_locks(struct obd_import *imp)
+static int lock_can_replay(struct obd_import *imp)
+{
+ struct client_obd *cli = &imp->imp_obd->u.cli;
+
+ CDEBUG(D_HA, "check lock replay limit, inflights = %u(%u)\n",
+ atomic_read(&imp->imp_replay_inflight) - 1,
+ cli->cl_max_rpcs_in_flight);
+
+ /* +1 due to ldlm_lock_replay() increment */
+ return atomic_read(&imp->imp_replay_inflight) <
+ 1 + min_t(u32, cli->cl_max_rpcs_in_flight, 8);
+}
+
+int __ldlm_replay_locks(struct obd_import *imp, bool rate_limit)
{
struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
- struct list_head list = LIST_HEAD_INIT(list);
+ LIST_HEAD(list);
struct ldlm_lock *lock, *next;
int rc = 0;
ENTRY;
- LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
+ LASSERT(atomic_read(&imp->imp_replay_inflight) == 1);
/* don't replay locks if import failed recovery */
if (imp->imp_vbr_failed)
RETURN(0);
- /* ensure this doesn't fall to 0 before all have been queued */
- atomic_inc(&imp->imp_replay_inflight);
-
if (ldlm_cancel_unused_locks_before_replay)
ldlm_cancel_unused_locks_for_replay(ns);
}
rc = replay_one_lock(imp, lock);
LDLM_LOCK_RELEASE(lock);
+
+ if (rate_limit)
+ wait_event_idle_exclusive(imp->imp_replay_waitq,
+ lock_can_replay(imp));
}
+ RETURN(rc);
+}
+
+/**
+ * Lock replay uses rate control and can sleep waiting so
+ * must be in separate thread from ptlrpcd itself
+ */
+static int ldlm_lock_replay_thread(void *data)
+{
+ struct obd_import *imp = data;
+
+ unshare_fs_struct();
+
+ CDEBUG(D_HA, "lock replay thread %s to %s@%s\n",
+ imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
+ imp->imp_connection->c_remote_uuid.uuid);
+
+ __ldlm_replay_locks(imp, true);
atomic_dec(&imp->imp_replay_inflight);
+ ptlrpc_import_recovery_state_machine(imp);
+ class_import_put(imp);
- RETURN(rc);
+ return 0;
+}
+
+int ldlm_replay_locks(struct obd_import *imp)
+{
+ struct task_struct *task;
+ int rc = 0;
+
+ class_import_get(imp);
+ /* ensure this doesn't fall to 0 before all have been queued */
+ atomic_inc(&imp->imp_replay_inflight);
+
+ task = kthread_run(ldlm_lock_replay_thread, imp, "ldlm_lock_replay");
+ if (IS_ERR(task)) {
+ rc = PTR_ERR(task);
+ CDEBUG(D_HA, "can't start lock replay thread: rc = %d\n", rc);
+
+ /* run lock replay without rate control */
+ rc = __ldlm_replay_locks(imp, false);
+ atomic_dec(&imp->imp_replay_inflight);
+ class_import_put(imp);
+ }
+
+ return rc;
}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff