From f07da6a78ae3189740ac72cd99783c4f3d835749 Mon Sep 17 00:00:00 2001
From: Arthur Hanson Rasmusson <arthur@vgpu.io>
Date: Mon, 2 Nov 2026 12:17:10 -0500
Subject: vgsp-diag: re-assert GL setup-push pins end-to-end (strict window
 re-emit + host GL-family exception)

The r1397 cycle located and content-validated the 60 216 B GL setup push
(the MME 0x114 carrier: 15-page window at a client rw-s mapping, 3728 code
words / 74 SA pairs) and the strict client-mem parse emitted all 15 pins -
yet the host entry stayed unreadable (src=255) on every re-read: pages
0x400a000..0x400e000 never reached the pin table (no injection line, no
re-resolution), so the per-page read always failed on a missing tail page,
and the loose family-learn could additionally overwrite a validated page
with a wrong one (weak 0x200-byte prefix learn).

Two one-line-class changes, mirroring the proven CE-window rationale
('a pin batch that raced in must be allowed to land again even when the
pair is unchanged'):

* guest_pin_stream: a strict-window success stashes its content-validated
  pairs (_strict_reemit) and re-emits them every ~4 s, so a page dropped
  in burst delivery heals without a new resolution and without defeating
  the done latch (r1397: the missing pages were never retried).

* pin_gr_stream: re-asserts GL setup-push family pairs (0x4000000-0x4160000,
  0x4200000-0x4300000, 0xcc80000-0xcc90000, 0x6f00000-0x6f10000) like the
  CE window, instead of dropping unchanged repeats via the seen dedup.

SESSION 153 r1397b; follows the r1397 offline brief Change 1 (pagemap
window fallback, kept).
---
 tools/guest-diag/guest_pin_stream.py | 23 +++++++++++++++++++++++
 tools/guest-diag/pin_gr_stream.py    | 14 +++++++++++++-
 2 files changed, 36 insertions(+), 1 deletion(-)

diff --git a/tools/guest-diag/guest_pin_stream.py b/tools/guest-diag/guest_pin_stream.py
index 3222a90..d8566c2 100644
--- a/tools/guest-diag/guest_pin_stream.py
+++ b/tools/guest-diag/guest_pin_stream.py
@@ -316,6 +316,17 @@ _area_off = {}
 # zero pages included); cleared when the area disappears (context rebuild).
 _gl_fam = {}
 
+# SESSION 153-r1397b: strict-validated window pins are re-emitted every few
+# seconds. The one-shot emission loses pages: a burst can be dropped under
+# host load, and the `done` latch then means the lost pages are never
+# retried - the 60216 B GL setup push needs all 15 pages, so one missing
+# page keeps the whole entry unreadable (r1397: pages 0x4007000..0x4009000
+# re-landed via the loose path but 0x400a000..0x400e000 never did; every
+# 60216 B read answered src=255 unreadable=1 and the entry was skipped).
+# The pairs are content-validated, so a repeat is idempotent; the host
+# re-asserts GL-family pairs on receipt. epage -> [pairs, next_ts].
+_strict_reemit = {}
+
 
 def resolve_va_by_content(kc, va, out):
     """Demand pin: find the page carrying GPU VA 'va' by content match.
@@ -2581,6 +2592,12 @@ def main():
                                 got = got2
                         if ok and got:
                             lines.extend(got)
+                            # SESSION 153-r1397b: keep re-emitting this
+                            # content-validated window so a page lost in
+                            # burst delivery heals without waiting for a
+                            # new resolution (see _strict_reemit).
+                            _strict_reemit[epage] = [list(got),
+                                                     time.time() + 4.0]
                             for g in got:
                                 try:
                                     _want_state["done"].add(
@@ -2592,6 +2609,12 @@ def main():
                         else:
                             _want_state["strict_pending"].setdefault(
                                 epage, now_d)
+                    _rr_now = time.time()
+                    for _epg in list(_strict_reemit):
+                        _pairs, _nxt = _strict_reemit[_epg]
+                        if _rr_now >= _nxt:
+                            lines.extend(_pairs)
+                            _strict_reemit[_epg][1] = _rr_now + 4.0
                     # Loose VAs in a family whose strict anchor is still
                     # pending are deferred to the strict path only.
                     pend_bases = set()
diff --git a/tools/guest-diag/pin_gr_stream.py b/tools/guest-diag/pin_gr_stream.py
index 35c2ea3..b2ba9b8 100644
--- a/tools/guest-diag/pin_gr_stream.py
+++ b/tools/guest-diag/pin_gr_stream.py
@@ -675,7 +675,19 @@ def _flush(runs, seen, full=False):
         # r145 stall: sema-probe "NO ENTRY" for the live push pages) and the
         # pool/tracker page (pending release starved, every client's first
         # seconds).
-        ce_win = 0x120000000 <= va < 0x122000000
+        #
+        # SESSION 153-r1397b: the GL setup-push families get the same
+        # treatment. The guest re-emits a strict-validated 60 216 B window
+        # every ~4 s because a burst can be partially dropped under host
+        # load and the `done` latch otherwise never retries the lost pages
+        # (r1397: pages 0x400a000..0x400e000 never landed; every 60216 B
+        # entry read answered src=255 and the push was skipped). Without
+        # this exception the unchanged re-emits were dropped here.
+        gl_win = ((0x4000000 <= va < 0x4160000) or
+                  (0x4200000 <= va < 0x4300000) or
+                  (0xcc80000 <= va < 0xcc90000) or
+                  (0x6f00000 <= va < 0x6f10000))
+        ce_win = 0x120000000 <= va < 0x122000000 or gl_win
         if not full and not ce_win and seen.get(va) == pa and n == 1:
             continue
         todo[(va, pa, n)] = True
-- 
cgit v1.2.3