When a connection is reversed, some elements must be resetted prior to
reusing it. Most notably, connection must be removed from lists specific
on frontend/backend sides.
When reverse was performed for frontend to backend side, connection was
not removed via its <stopping_list> attach point. On previous releases,
this did not cause any issue. However, crashes start to occur recently,
probably due to the recent reorganization of connection list attach
points from the following patch.
commit
a96f1286a75246fef6db3e615fabdef1de927d83
BUG/MINOR: connection: rearrange union list members
To fix this, simply ensure that <stopping_list> detach is performed via
conn_reverse().
This patch must be backported up to 3.0 release.
(cherry picked from commit
27ff7ff296dec073e84cf03106a3e285181c24d8)
Signed-off-by: Christopher Faulet <cfaulet@haproxy.com>
(cherry picked from commit
f656fb45fe7e2873541db31308c010fbc78ff6e3)
Signed-off-by: Christopher Faulet <cfaulet@haproxy.com>
(cherry picked from commit
8faba2b9062ce4ade65cee5ec556c111b54c2390)
Signed-off-by: Christopher Faulet <cfaulet@haproxy.com>
struct server *srv = objt_server(conn->reverse.target);
BUG_ON(!srv);
+ LIST_DEL_INIT(&conn->stopping_list);
+
if (conn_backend_init(conn))
return 1;