xen/pvcalls-back: fix permanently masked event channel

There is a sequence of events that can lead to a permanently masked
event channel, because xen_irq_lateeoi() is newer called. This happens
when a backend receives spurious write event from a frontend. In this
case pvcalls_conn_back_write() returns early and it does not clears the
map->write counter. As map->write > 0, pvcalls_back_ioworker() returns
without calling xen_irq_lateeoi(). This leaves the event channel in
masked state, a backend does not receive any new events from a
frontend and the whole communication stops.

Move atomic_set(&map->write, 0) to the very beginning of
pvcalls_conn_back_write() to fix this issue.

Signed-off-by: Volodymyr Babchuk <volodymyr_babchuk@epam.com>
Reported-by: Oleksii Moisieiev <oleksii_moisieiev@epam.com>
Reviewed-by: Juergen Gross <jgross@suse.com>
Link: https://lore.kernel.org/r/20230119211037.1234931-1-volodymyr_babchuk@epam.com
Signed-off-by: Juergen Gross <jgross@suse.com>

diff --git a/drivers/xen/pvcalls-back.c b/drivers/xen/pvcalls-back.c
index 0d4f8f4f494888898c759d0f14ddc28f594cdabc..7a464a541d913f7beff4d1fa714ceb63fa4ec52a 100644 (file)
--- a/drivers/xen/pvcalls-back.c
+++ b/drivers/xen/pvcalls-back.c
@@ -173,6 +173,8 @@ static bool pvcalls_conn_back_write(struct sock_mapping *map)
        RING_IDX cons, prod, size, array_size;
        int ret;
 
+       atomic_set(&map->write, 0);
+
        cons = intf->out_cons;
        prod = intf->out_prod;
        /* read the indexes before dealing with the data */
@@ -197,7 +199,6 @@ static bool pvcalls_conn_back_write(struct sock_mapping *map)
                iov_iter_kvec(&msg.msg_iter, ITER_SOURCE, vec, 2, size);
        }
 
-       atomic_set(&map->write, 0);
        ret = inet_sendmsg(map->sock, &msg, size);
        if (ret == -EAGAIN) {
                atomic_inc(&map->write);