[PATCH 2/2] xen/xenbus: check otherend_id only after it has been initialized

Marek Marczykowski-Górecki <[email protected]> Mon, 3 Aug 2026 05:08:02 +0200
Newsgroups org.xenproject.lists.xen-devel,org.kernel.vger.linux-kernel
Message-ID <[email protected]>
When device just got initialized (for example on module load), the
otherend_id field is initialized only after
xenbus_read_otherend_details() gets called. If xenstore watch triggers
xenbus_dev_changed() before that, it might consider still zeroed
otherend_id field (not matching actual xenstore content) as a sign of
device state reset. It can happen because xenstore watch are handled in
another thread (xenwatch), which can run in parallel to the initial
device probe running at module load. In that case, it would call
device_unregister(), which would deadlock against device probe from
module init.

Fix this by considering dev->otherend_id change only after dev->otherend
is set (which happen after otherend_id is initialized).

Fixes: e2dcf9065536 "xen/xenbus: better handle backend crash"
Signed-off-by: Marek Marczykowski-Górecki <[email protected]>
---
 drivers/xen/xenbus/xenbus_probe.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/xen/xenbus/xenbus_probe.c b/drivers/xen/xenbus/xenbus_probe.c
index a259c8f0fff4..b42d8d2e5a33 100644
--- a/drivers/xen/xenbus/xenbus_probe.c
+++ b/drivers/xen/xenbus/xenbus_probe.c
@@ -680,7 +680,8 @@ void xenbus_dev_changed(const char *node, struct xen_bus_type *bus)
 							    dev->otherend_id);
 
 		if (state == XenbusStateInitialising &&
-		    (state != dev->state || backend != dev->otherend_id)) {
+		    (state != dev->state ||
+		     (dev->otherend && backend != dev->otherend_id))) {
 			/*
 			 * State has been reset, assume the old one vanished
 			 * and new one needs to be probed.
-- 
2.54.0