vgsp-sema r620: r619's device-VA sem-write witness range used the 1 TiB SVC_VAS_FLOOR and silently excluded the 0x200000000-0x20fffffff VAS window (the release words it exists for) - use the window base so the r618-style write-target divergence is actually logged

1 files changed, 9 insertions, 7 deletions

diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c
index 3f50439..e76cb7a 100644
--- a/kernel/gsp_sema_service.c
+++ b/kernel/gsp_sema_service.c
@@ -4242,16 +4242,18 @@ static int svc_sem_write(struct nv_vgpu_device *vgpu, u32 h, u64 va, u32 val)
 	 * the read-back, so the second session's translations can be
 	 * compared with the first session's for the same VA. Bounded.
 	 *
-	 * r619 diag: the same proof for NON-pool device-VA writes (>= the VAS
-	 * floor; e.g. 0x20480fef0 push releases, 0x204827f7c QMD releases).
-	 * The r618 hung math client read 0 at 0x20480fef0 while the host's
-	 * deferred releases logged `cur` 97..105 on its own target - a
-	 * write-target divergence that was invisible because this range
-	 * never printed `pa`. Bounded per boot (512).
+	 * r619 diag: the same proof for NON-pool device-VA writes (the VAS
+	 * window >= 0x200000000; e.g. 0x20480fef0 push releases, 0x204827f7c
+	 * QMD releases). The r618 hung math client read 0 at 0x20480fef0
+	 * while the host's deferred releases logged `cur` 97..105 on its own
+	 * target - a write-target divergence that was invisible because this
+	 * range never printed `pa`. r620: the first cut used the 1 TiB
+	 * SVC_VAS_FLOOR and excluded this very window (no lines ever);
+	 * bounded per boot (512).
 	 */
 	if (((va >= 0x79240000000ULL && va < 0x79260000000ULL &&
 	      svc.sem_sw_log < 8192) ||
-	     (va >= SVC_VAS_FLOOR && !is_pool && sem_sw_dev_log < 512))) {
+	     (va >= 0x200000000ULL && !is_pool && sem_sw_dev_log < 512))) {
 		if (va >= 0x79240000000ULL && va < 0x79260000000ULL)
 			svc.sem_sw_log++;
 		else