guest-diag: semfix_watch read64 uncached by default (stale-cache pool-pin failure)

The watcher's resolve_sem_pool read sem->page->mem->vas through a
process-lifetime cache: a transient zero captured while UVM was still
initialising the structures was served forever, the pool pin never
landed, and the second-session client hung in cudaMallocManaged
(r88 evidence: 'pool unresolved (retry)' from the watcher while the
same walk succeeded from a fresh QMP client). Cache is now opt-in.

1 files changed, 8 insertions, 1 deletions

diff --git a/tools/guest-diag/semfix_watch.py b/tools/guest-diag/semfix_watch.py
index 049a612..5124805 100644
--- a/tools/guest-diag/semfix_watch.py
+++ b/tools/guest-diag/semfix_watch.py
@@ -123,7 +123,14 @@ class Qmp:
         cur = getattr(self, "_cpu", 0)
         return [cur] + [c for c in range(self.ncpu()) if c != cur]
 
-    def read64(self, addr, n=1, cache=True):
+    def read64(self, addr, n=1, cache=False):
+        # Cache is opt-in and must stay that way: every reader here watches
+        # structures the guest mutates (UVM init fills sem->page->mem->vas in
+        # place). Caching a transient zero poisoned resolve_sem_pool forever
+        # (the watcher logged "pool unresolved (retry)" while the same reads
+        # succeeded from a fresh QMP client) - the second-session pool pin
+        # never landed and the client hung in cudaMallocManaged. Default is
+        # therefore uncached; callers may opt in for within-pass scans only.
         key = (addr, n)
         if cache and key in self._cache:
             return self._cache[key]