Fix xen driver refcounting.
The commit cb51aa48 "Fix up connection reference counting." changed the driver closing and virConnectPtr unref-logic in virConnectClose(). Before this commit virConnectClose() closed all drivers of the given virConnectPtr and virUnrefConnect()'ed it afterwards. After this commit the driver-closing is done in virUnrefConnect() if and only if the ref-count of the virConnectPtr dropped to zero. This change in execution order leads to a virConnectPtr leak, at least for connections to Xen. The relevant call sequences: virConnectOpen() -> xenUnifiedOpen() ... ... xenInotifyOpen() -> virConnectRef(conn) ... xenStoreOpen() -> xenStoreAddWatch() -> conn->refs++ virConnectClose() -> xenUnifiedClose() ... ... xenInotifyClose() -> virUnrefConnect(conn) ... xenStoreClose() -> xenStoreRemoveWatch() -> virUnrefConnect(conn) Before the commit this additional virConnectRef/virUnrefConnect calls where no problem, because virConnectClose() closed the drivers explicitly and the additional refs added by the Xen subdrivers were removed properly. After the commit this additional refs result in a virConnectPtr leak (including a leak of the hypercall file handle; that's how I noticed this problem), because now the drivers are only close if and only if the ref-count drops to zero, but this cannot happen anymore, because the additional refs from the Xen subdrivers would only be removed if the drivers get closed, but that doesn't happen because the ref-count cannot drop to zero. The fix for this problem is simple: remove the virConnectRef/virUnrefConnect calls from the Xen subdrivers (see attached patch). Maybe someone could explain why the Xen Inotify and Xen Store driver do this extra ref-counting, but none of the other Xen subdrivers. It seems unnecessary to me and can be removed without problems. Signed-off-by: NChris Lalancette <clalance@redhat.com>
Showing
想要评论请 注册 或 登录