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>
---
 drivers/xen/pvcalls-back.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/xen/pvcalls-back.c b/drivers/xen/pvcalls-back.c
index a7d293fa8d14..60f5cd70d770 100644
--- 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, READ, vec, 2, size);
 	}
 
-	atomic_set(&map->write, 0);
 	ret = inet_sendmsg(map->sock, &msg, size);
 	if (ret == -EAGAIN) {
 		atomic_inc(&map->write);
-- 
2.38.1

On 19.01.23 22:11, Volodymyr Babchuk wrote:
> 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>

Pushed to: xen/tip.git for-linus-6.3


Juergen

On 19.01.23 22:11, Volodymyr Babchuk wrote:
> 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>


Juergen