tools: prompt re-assert after recv chunks, heartbeat inside long QMP loops, novel-only activity accounting (tools-fix a4840952)

2 files changed, 29 insertions, 3 deletions

diff --git a/tools/guest-diag/pin_gr_stream.py b/tools/guest-diag/pin_gr_stream.py
index 2f78f2a..6ba51ee 100644
--- a/tools/guest-diag/pin_gr_stream.py
+++ b/tools/guest-diag/pin_gr_stream.py
@@ -328,10 +328,11 @@ def main():
                     chunk = conn.recv(1 << 20)
                 except socket.timeout:
                     if runs:
+                        novel = _novel_pages(runs, seen)
                         pages = _flush(runs, seen)
                         pin_total += pages
                         pins_since_client += pages
-                        if pages:
+                        if novel:
                             _activity()
                         runs = []
                     _maybe_reassert_request(seen)
@@ -433,12 +434,18 @@ def main():
                                      int(m.group(2), 16),
                                      int(m.group(3)) if m.group(3) else 1))
                 if runs:
+                    novel = _novel_pages(runs, seen)
                     pages = _flush(runs, seen)
                     pin_total += pages
                     pins_since_client += pages
-                    if pages:
+                    if novel:
                         _activity()
                     runs = []
+                # Poll the semfix re-assert request on every recv cycle, not
+                # only on recv-timeouts: sustained scanner traffic starves the
+                # timeout branch, so the request fired post-mortem in r203
+                # (119 s late, `seen` already cleared by #nopid).
+                _maybe_reassert_request(seen)
         except Exception as e:
             # A transient error on one connection must not kill the process:
             # log it, drop the connection and go back to accept().
@@ -482,6 +489,17 @@ def _coalesce_seen(seen):
     return runs
 
 
+def _novel_pages(runs, seen):
+    """Count pages in runs whose translation is new or changed.
+
+    Progress accounting for the stall watchdog: the guest re-streams the CE
+    window and UVM pool every couple of seconds even when nothing moves, so a
+    pure repeat of an already-known pair must not count as pin activity
+    (r203: those repeats masked the watchdog through every hang)."""
+    return sum(1 for va, pa, n in runs for i in range(n)
+               if seen.get(va + i * 0x1000) != pa + i * 0x1000)
+
+
 def _flush(runs, seen, full=False):
     """Inject run lines (one batched sudo pass), return number of pages pinned.
 
diff --git a/tools/guest-diag/semfix_uvmt.py b/tools/guest-diag/semfix_uvmt.py
index f28efd3..11c191c 100755
--- a/tools/guest-diag/semfix_uvmt.py
+++ b/tools/guest-diag/semfix_uvmt.py
@@ -671,7 +671,12 @@ def main():
         for sem in sorted(sems - known_sems):
             known_sems.add(sem)
 
-        for sem in sorted(known_sems):
+        for hb_i, sem in enumerate(sorted(known_sems)):
+            if hb_i % 16 == 0:
+                # A full pass can run 60-90 s (thousands of QMP reads); keep
+                # the heartbeat fresh inside it (r203: 79 s gap -> supervisor
+                # restarted the tool mid-pass).
+                _hb()
             now = time.time()
             if now < skip_until.get(sem, 0):
                 continue
@@ -793,6 +798,7 @@ def main():
                 qmp_next_try = now + 5.0
 
         for ch in sorted(chans - seen_ch):
+            _hb()  # a resolve can block for tens of seconds
             d = {}
             try:
                 res = resolve_from_channel(q, ch, d)
@@ -807,6 +813,7 @@ def main():
             else:
                 log("semfix: ch=0x%x pushbuf miss [%s]" % (ch, diag_str(d)))
         for pb in sorted(pbs - seen_pb):
+            _hb()  # a resolve can block for tens of seconds
             d = {}
             try:
                 res = resolve_from_pushbuffer(q, pb, d)
@@ -833,6 +840,7 @@ def main():
         if loop % 8 == 0:
             for st in list(pinned_ranges.values()):
                 if not st["done"] and st["tries"] > 0:
+                    _hb()  # pin_range_once can walk thousands of QMP reads
                     pin_range_once(*st["args"])
 
         # Dead-pid detection and the pre-capture stall watchdog. The guest is