Fix clients possibly not destroyed after being switched to RDMA

It could in theory lead to hung operations on RDMA client disconnections.

Workflow (probable):
* OSD switches an outbound connection to PEER_RDMA and removes the FD from epoll
* But a receive request in io_uring is still active
* Receive response is never handled, refs remains > 0
* The connection is dropped, but osd_client_t is not destroyed because refs > 0
* Outbound ops are not canceled, suboperations hang, primary operations hang too
This commit is contained in:
Vitaliy Filippov
2026-03-17 01:04:17 +03:00
parent fdea595913
commit 44c895dc30
4 changed files with 11 additions and 19 deletions
-8
View File
@@ -739,14 +739,6 @@ void osd_messenger_t::check_peer_config(osd_client_t *cl)
fprintf(stderr, "Connected to OSD %ju using RDMA\n", cl->osd_num);
}
cl->peer_state = PEER_RDMA;
tfd->set_fd_handler(cl->peer_fd, false, [this](int peer_fd, int epoll_events)
{
// Do not miss the disconnection!
if (epoll_events & EPOLLRDHUP)
{
handle_peer_epoll(peer_fd, epoll_events);
}
});
// Add the initial receive request
init_recv_rdma(cl);
}
+4
View File
@@ -696,6 +696,10 @@ void osd_messenger_t::handle_rdma_events(msgr_rdma_context_t *rdma_context)
continue;
}
osd_client_t *cl = cl_it->second;
if (cl->peer_state == PEER_STOPPED)
{
continue;
}
auto rc = cl->rdma_conn;
if (wc[i].status != IBV_WC_SUCCESS)
{
+6 -1
View File
@@ -9,7 +9,8 @@ void osd_messenger_t::read_requests()
{
int peer_fd = read_ready_clients[i];
auto cl_it = clients.find(peer_fd);
if (cl_it == clients.end() || !cl_it->second || cl_it->second->read_msg.msg_iovlen)
if (cl_it == clients.end() || !cl_it->second || cl_it->second->read_msg.msg_iovlen ||
cl_it->second->peer_state == PEER_RDMA || cl_it->second->peer_state == PEER_RDMA_CONNECTING)
{
continue;
}
@@ -75,6 +76,10 @@ bool osd_messenger_t::handle_read(int result, osd_client_t *cl)
int peer_fd = cl->peer_fd;
cl->read_msg.msg_iovlen = 0;
cl->refs--;
if (cl->peer_state == PEER_RDMA)
{
return true;
}
if (cl->peer_state == PEER_STOPPED)
{
if (cl->refs <= 0)
+1 -10
View File
@@ -251,7 +251,7 @@ void osd_messenger_t::send_replies()
{
int peer_fd = write_ready_clients[i];
auto cl_it = clients.find(peer_fd);
if (cl_it != clients.end() && !try_send(cl_it->second))
if (cl_it != clients.end() && cl_it->second->peer_state != PEER_RDMA && !try_send(cl_it->second))
{
write_ready_clients.erase(write_ready_clients.begin(), write_ready_clients.begin() + i);
return;
@@ -349,21 +349,12 @@ void osd_messenger_t::handle_send(int result, bool prev, bool more, osd_client_t
#ifdef WITH_RDMA
if (cl->rdma_conn && !cl->outbox.size() && cl->peer_state == PEER_RDMA_CONNECTING)
{
// FIXME: Do something better than just forgetting the FD
// FIXME: Ignore pings during RDMA state transition
if (log_level > 0)
{
fprintf(stderr, "Successfully connected with client %d using RDMA\n", cl->peer_fd);
}
cl->peer_state = PEER_RDMA;
tfd->set_fd_handler(cl->peer_fd, false, [this](int peer_fd, int epoll_events)
{
// Do not miss the disconnection!
if (epoll_events & EPOLLRDHUP)
{
handle_peer_epoll(peer_fd, epoll_events);
}
});
// Add the initial receive request
init_recv_rdma(cl);
}