From 93d20d171c20491a96e5e85d7442a002f300619d Mon Sep 17 00:00:00 2001 From: Alex Zhuravlev Date: Fri, 10 Aug 2018 21:03:11 +0400 Subject: [PATCH 1/1] LU-11128 ptlrpc: new request vs disconnect race new request can race with disconnect-by-idle process. disconnect code detect this state and initiate a new connection. Change-Id: I6acc913c371f2ae63f27151edce457c2a13118c5 Signed-off-by: Alex Zhuravlev Reviewed-on: https://review.whamcloud.com/32980 Tested-by: Jenkins Tested-by: Maloo Reviewed-by: Mike Pershin Reviewed-by: Andreas Dilger Reviewed-by: Oleg Drokin --- lustre/ptlrpc/client.c | 16 ++++++++++------ lustre/ptlrpc/import.c | 31 ++++++++++++++++++++++++++++--- 2 files changed, 38 insertions(+), 9 deletions(-) diff --git a/lustre/ptlrpc/client.c b/lustre/ptlrpc/client.c index a3b8504..a167146 100644 --- a/lustre/ptlrpc/client.c +++ b/lustre/ptlrpc/client.c @@ -867,6 +867,12 @@ ptlrpc_request_alloc_internal(struct obd_import *imp, struct ptlrpc_request *request; int connect = 0; + request = __ptlrpc_request_alloc(imp, pool); + if (request == NULL) + return NULL; + + /* initiate connection if needed when the import has been + * referenced by the new request to avoid races with disconnect */ if (unlikely(imp->imp_state == LUSTRE_IMP_IDLE)) { int rc; CDEBUG_LIMIT(imp->imp_idle_debug, @@ -883,16 +889,14 @@ ptlrpc_request_alloc_internal(struct obd_import *imp, spin_unlock(&imp->imp_lock); if (connect) { rc = ptlrpc_connect_import(imp); - if (rc < 0) + if (rc < 0) { + ptlrpc_request_free(request); return NULL; + } ptlrpc_pinger_add_import(imp); } } - request = __ptlrpc_request_alloc(imp, pool); - if (request == NULL) - return NULL; - req_capsule_init(&request->rq_pill, request, RCL_CLIENT); req_capsule_set(&request->rq_pill, format); return request; @@ -2346,7 +2350,7 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set) /* wait until all complete, interrupted, or an in-flight * req times out */ CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n", - set, timeout); + set, timeout); if ((timeout == 0 && !signal_pending(current)) || set->set_allow_intr) diff --git a/lustre/ptlrpc/import.c b/lustre/ptlrpc/import.c index 241c1a7..8ca1dec 100644 --- a/lustre/ptlrpc/import.c +++ b/lustre/ptlrpc/import.c @@ -1665,13 +1665,38 @@ static int ptlrpc_disconnect_idle_interpret(const struct lu_env *env, void *data, int rc) { struct obd_import *imp = req->rq_import; + int connect = 0; + + DEBUG_REQ(D_HA, req, "inflight=%d, refcount=%d: rc = %d\n", + atomic_read(&imp->imp_inflight), + atomic_read(&imp->imp_refcount), rc); - LASSERT(imp->imp_state == LUSTRE_IMP_CONNECTING); spin_lock(&imp->imp_lock); - IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_IDLE); - memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle)); + /* DISCONNECT reply can be late and another connection can just + * be initiated. so we have to abort disconnection. */ + if (req->rq_import_generation == imp->imp_generation && + imp->imp_state != LUSTRE_IMP_CLOSED) { + LASSERTF(imp->imp_state == LUSTRE_IMP_CONNECTING, + "%s\n", ptlrpc_import_state_name(imp->imp_state)); + imp->imp_state = LUSTRE_IMP_IDLE; + memset(&imp->imp_remote_handle, 0, + sizeof(imp->imp_remote_handle)); + /* take our DISCONNECT into account */ + if (atomic_read(&imp->imp_inflight) > 1) { + imp->imp_generation++; + imp->imp_initiated_at = imp->imp_generation; + IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_NEW); + connect = 1; + } + } spin_unlock(&imp->imp_lock); + if (connect) { + rc = ptlrpc_connect_import(imp); + if (rc >= 0) + ptlrpc_pinger_add_import(imp); + } + return 0; } -- 1.8.3.1