guest-diag: seed channel-event sets from the trace ring on client switch

clear_trace() on a pid switch dropped exactly the one-shot uvmt/uvch
events of the new client when they landed in the switch instant; the
pipe reader then never saw them and the whole session ran with an empty
sem set (r126: vecadd_loop sem=0 for minutes while its CE entries paid
the 8 s grace). Scan the ring tail once per switch and seed the sets
instead - stale kptrs that are also re-learned resolve to no content and
are deferred by the candidate scoring.

1 files changed, 41 insertions, 9 deletions

diff --git a/tools/guest-diag/guest_pin_stream.py b/tools/guest-diag/guest_pin_stream.py
index b031733..c0cf485 100644
--- a/tools/guest-diag/guest_pin_stream.py
+++ b/tools/guest-diag/guest_pin_stream.py
@@ -807,15 +807,41 @@ def resolve_channel_pins(kc, ch, out):
 
 
 
-def clear_trace():
-    """Drop stale ring events so the fallback tail reader cannot re-learn a
-    previous client's freed sems after a client switch. The new client's
-    events are delivered fresh through the consuming trace_pipe reader."""
+def seed_events_from_trace(seen_sems, seen_pbs, seen_chs):
+    """One-shot tail scan of the trace ring after a pid switch / start.
+
+    The uvmt/uvbp/uvch kprobe events are one-shot per object, and the
+    trace_pipe reader only sees events that arrive after it attaches: a
+    client whose channels are registered in the instant around the scanner's
+    pid switch would otherwise never appear in the sets (observed r126:
+    vecadd_loop ran for minutes with sem=0 while its CE entries paid the
+    8 s grace). The ring keeps recent events, so scan its tail once and add
+    every distinct kptr. Stale previous-client pointers that are also
+    re-learned are harmless: their chains no longer resolve to content
+    (the mem candidates are scored and deferred).
+    """
     try:
-        with open(TRACE, "w") as f:
-            f.write("")
-    except OSError:
-        pass
+        out = subprocess.run(["tail", "-c", "1048576", TRACE],
+                             capture_output=True, timeout=10).stdout
+    except Exception:
+        return 0
+    n = 0
+    for m in re.finditer(rb"sem=0x([0-9a-fA-F]+)", out):
+        v = int(m.group(1), 16)
+        if v not in seen_sems:
+            seen_sems.add(v)
+            n += 1
+    for m in re.finditer(rb"pb=0x([0-9a-fA-F]+)", out):
+        v = int(m.group(1), 16)
+        if v not in seen_pbs:
+            seen_pbs.add(v)
+            n += 1
+    for m in re.finditer(rb"ch=0x([0-9a-fA-F]+)", out):
+        v = int(m.group(1), 16)
+        if v not in seen_chs:
+            seen_chs.add(v)
+            n += 1
+    return n
 
 
 class TraceReader:
@@ -1190,10 +1216,16 @@ def main():
             # emitted as if they were the new client's counters.
             CHAN_PUTS.clear()
             PUT_LAST.clear()
-            clear_trace()
             tracer.reset()
             RESOLVED_KC_PB.clear()
             KC_PB_STATE.clear()
+            # One-shot: recover the new client's one-shot channel events
+            # from the trace ring (they can fall in the switch instant and
+            # be invisible to the pipe reader; the old clear_trace() made
+            # that permanent - part of the r126 stall class).
+            _seeded = seed_events_from_trace(seen_sems, seen_pbs, seen_chs)
+            if _seeded:
+                dbg("switch seed: +%d event kptr(s)" % _seeded)
             if sock is not None:
                 try:
                     sock.sendall(("#pid %d\n" % pid).encode())