vgsp-gr r597: emit the captured prelude's SEM_EXECUTE ACQUIRE groups ahead of the rebuilt window (default off)

The verbatim prelude replay breaks the launch (r599: qmd_rel=0 after an
8.9 s stall), so the acquire cannot be taken by replaying the prelude.
This takes only the wait method groups - C46F addr 0x5c count>=5 with
EXECUTE&7 != 1, and C5C0 SET_REPORT_SEMAPHORE A..D with a non-release op -
and copies them verbatim before the proven init+window+PCAS+release
block, leaving the window untouched.

Gate: gsp_phys_rmcall_gr_acq (default 0). Live A/B on r601: the mail
launch's prelude carries ZERO acquire groups (the C5C0 prelude is inline
parameter data + release only), so emitted=0 and the VADD race is
unchanged - the driver assumes its relocation copies precede the launch
without an engine-visible wait.

1 files changed, 19 insertions, 0 deletions

diff --git a/kernel/gsp_phys_gpu.c b/kernel/gsp_phys_gpu.c
index 74062dd..e9c0c5a 100644
--- a/kernel/gsp_phys_gpu.c
+++ b/kernel/gsp_phys_gpu.c
@@ -617,6 +617,25 @@ int gsp_phys_rmcall_gr_prelude_value(void)
 EXPORT_SYMBOL_GPL(gsp_phys_rmcall_gr_prelude_value);
 
 /*
+ * r597: emit the SEM_EXECUTE ACQUIRE groups found in the captured guest
+ * prelude ahead of the rebuilt inline-QMD window (default 0). The verbatim
+ * prelude replay breaks the launch (qmd_rel=0 after an 8.9 s stall, r599),
+ * so the acquire cannot be taken by replaying the prelude; this gate takes
+ * only the wait method groups and leaves the proven window in place.
+ */
+static int gsp_phys_rmcall_gr_acq_param;
+module_param_named(gsp_phys_rmcall_gr_acq, gsp_phys_rmcall_gr_acq_param,
+		   int, 0644);
+MODULE_PARM_DESC(gsp_phys_rmcall_gr_acq,
+		 "Emit the captured prelude's SEM_EXECUTE ACQUIRE groups ahead of the rebuilt GR window (default 0)");
+
+int gsp_phys_rmcall_gr_acq_value(void)
+{
+	return gsp_phys_rmcall_gr_acq_param;
+}
+EXPORT_SYMBOL_GPL(gsp_phys_rmcall_gr_acq_value);
+
+/*
  * QMD virtualization control gate (default 0 after m26 falsified the
  * hypothesis: on Turing the context init value is 0, native CUDA never
  * writes 0x29c, and the nonzero enable alone does not schedule). When set,