vgsp-sema: limit the failed-op hold to large ops (>= 256 KB) (ISSUE-0002)

r72 showed the broad rule holding entries on small failed ops too: two
opdef holds on h=0x5c000059 (entries 82/84) were never resolved and withheld
their releases for the whole session, while c2 hit the pre-existing GR launch
forward timeout. The scrub class this rule exists for is always large (r58/
r71: 1 MiB / 297 KiB / 608 KiB), so only failures of ops >= 256 KB mark the
pass op-pending; small failing ops keep the old consume-and-apply behaviour.

1 files changed, 16 insertions, 14 deletions

diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c
index 77ae4b4..fa76162 100644
--- a/kernel/gsp_sema_service.c
+++ b/kernel/gsp_sema_service.c
@@ -4614,21 +4614,23 @@ static void svc_run_ops(struct nv_vgpu_device *vgpu, struct svc_memset_op *ops,
 		else
 			svc.memsets_fail++;
 		/*
-		 * ISSUE-0002 (r71): ANY failed op leaves work owed on this
-		 * entry, not only an unresolvable operand. The pma scrub
-		 * memsets resolve a bogus out-of-bounds vidmem destination
-		 * through a walk before the scanner pins the real sysmem
-		 * page; svc_copy fails them with -EINVAL, and the old code
-		 * consumed the entry and applied the release anyway, leaving
-		 * the client frozen in scrubCheckAndWaitForSize. Mark the
-		 * pass as op-pending so the pump holds the entry and
-		 * withholds its releases (bounded by SVC_OPDEF_HOLD_MS,
-		 * exactly like the -ENOENT path): the retried decode checks
-		 * the injected pins first and runs the memset once the real
-		 * translation lands. On expiry the release is flushed and the
-		 * op stays skipped (the pre-r62 fallback).
+		 * ISSUE-0002 (r71/r72): a failed LARGE op leaves work owed on
+		 * this entry. The pma scrub memsets (>= 256 KB) resolve a bogus
+		 * out-of-bounds vidmem destination through a walk before the
+		 * scanner pins the real sysmem pages; svc_copy fails them with
+		 * -EINVAL, and the old code consumed the entry and applied the
+		 * release anyway, leaving the client frozen in
+		 * scrubCheckAndWaitForSize. Mark the pass as op-pending so the
+		 * pump holds the entry and withholds its releases (bounded by
+		 * SVC_OPDEF_HOLD_MS): the retried decode checks the injected
+		 * pins first and runs the memset once the real translation
+		 * lands. Deliberately limited to >= 256 KB: a small failing op
+		 * (observed r72 c2) must not withhold a release that gates a GR
+		 * launch with a 5 s poll deadline. On expiry the release is
+		 * flushed and the op stays skipped (the pre-r62 fallback).
 		 */
-		if (rc != 0 && rc != SVC_COPY_FORWARDED)
+		if (rc != 0 && rc != SVC_COPY_FORWARDED &&
+		    m->bytes >= (256u << 10))
 			svc.op_tr_pending = 1;
 		if (gsp_sema_dbg && m->bytes == 256 && svc.sem256_log < 64) {
 			svc.sem256_log++;