1 files changed, 43 insertions, 4 deletions
diff --git a/tools/guest-diag/guest_pin_stream.py b/tools/guest-diag/guest_pin_stream.py
index 97c68f2..885a87e 100644
--- a/tools/guest-diag/guest_pin_stream.py
+++ b/tools/guest-diag/guest_pin_stream.py
@@ -639,6 +639,16 @@ def _kc_pb_fresh(kc, sem):
         return False
     if not st.get("page_va"):
         return True
+    if not st.get("page_pa"):
+        # Unknown PA at store time: the manual page-table walk could not
+        # translate the vmap page (it may have been read before the kernel
+        # page tables were walkable). Treat the record as fresh instead of
+        # failing every check - a permanently-failing freshness test made
+        # the scanner drop the latch and re-resolve in a tight loop, and
+        # each failed re-resolution left the host with no pins at all
+        # (observed r131: "kc-chain stale: re-resolving" every second for
+        # five sems while smiley sat in cudaMalloc).
+        return True
     pa = kc.va_to_pa(st["page_va"])
     if pa is None or pa != st.get("page_pa"):
         return False
@@ -1333,12 +1343,41 @@ def main():
                     except Exception:
                         _fresh = True
                     if not _fresh:
-                        _st = KC_PB_STATE.pop(sem, None)
+                        # Re-resolve, but never lose the latch on a failed
+                        # attempt: a dropped latch means the host keeps no
+                        # pins for this channel at all, which is worse than
+                        # a possibly-stale pin (the host content-checks the
+                        # first page and now rejects all-FF continuation
+                        # chunks, so stale pages are bounded). On failure
+                        # keep the old state and defer the next attempt.
+                        _st = KC_PB_STATE.get(sem)
                         if _st and _st.get("mem"):
                             RESOLVED_KC_PB.discard(_st["mem"])
-                        resolved_kc_pb.discard(sem)
-                        retry_at.pop(("pk", sem), None)
-                        dbg("kc-chain stale: re-resolving sem=0x%x" % sem)
+                        got = []
+                        try:
+                            _found2 = resolve_pb_chain(kc, sem, got)
+                        except Exception:
+                            _found2 = False
+                            got = []
+                        if _found2 and got:
+                            _new = KC_PB_STATE.get(sem)
+                            _newmem = _new.get("mem") if _new else None
+                            dbg("kc-chain stale: re-resolved sem=0x%x "
+                                "new-mem=%s runs=%d"
+                                % (sem, "yes" if _newmem !=
+                                   (_st.get("mem") if _st else None)
+                                   else "no", len(got)))
+                            lines.extend(got)
+                            added += len(got)
+                            resolved_kc_pb.add(sem)
+                        else:
+                            if _st and _st.get("mem"):
+                                RESOLVED_KC_PB.add(_st["mem"])
+                                if sem not in KC_PB_STATE:
+                                    KC_PB_STATE[sem] = _st
+                            kc_refresh_at[sem] = now_r + 5.0
+                            dbg("kc-chain stale: keep old latch sem=0x%x"
+                                % sem)
                 # CE push-buffer chain (second-client stall fix): resolve
                 # pool -> manager -> pushbuffer -> rm_mem directly, so the
                 # push-buffer pages never depend on the racy one-shot uvbp