vgsp-gr r612: refuse the defer cap only for a DEGENERATE argument fill - stalled decoy candidates must not kill a launch whose real operands resolved

r586's fail-closed refused any launch still carrying unresolved argument
pages at the 15 s cap. streams2 as the first client of a fresh boot (r611)
shows the cost: its two launches carry candidates 0x783299000000 and
0x783297000000 that stall in EVERY source from the first fill (no pin, no
walk, no ledger; 'stalled VA' listed, never resolving in the run) while 969
real pages resolve, the r543 arg-defer is pre-empted by the CE watermark
defers, and at the cap both kernels are refused - STREAMS bad=65534 with the
engine never running.

Refuse only a degenerate fill (resolved < dropped - the r584 matmul shape, 2
resolved / 18 dropped, where the engine has no usable window). A
non-degenerate fill at the cap forwards historically and loudly
(fwd_gr_argfw witness): waiting longer cannot resolve a candidate stalled in
every source, and by this point every copy-related guard (dstpend, CE
watermark) has already had its chance and returned above when it fired.

1 files changed, 37 insertions, 1 deletions

diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c
index 00d9b56..c0be131 100644
--- a/kernel/gsp_sema_service.c
+++ b/kernel/gsp_sema_service.c
@@ -1132,6 +1132,10 @@ struct sema_svc {
 	u64			fwd_gr_argdefer;	/* r543: held for a partial
 						 * argument set (unresolved arg) */
 	u32			fwd_gr_argdefer_log;
+	u64			fwd_gr_argfw;	/* r612: non-degenerate fill
+						 * forwarded at the cap (stalled
+						 * candidates are decoys) */
+	u32			fwd_gr_argfw_log;
 	u64			fwd_gr_cewm;	/* r590: held on the CE ring's
 						 * live submission watermark */
 	u64			fwd_gr_cewm_cap;	/* r590: watermark still
@@ -21797,9 +21801,26 @@ static bool svc_gr_request(struct nv_vgpu_device *vgpu,
 	 * no-write). Refuse the forward, author the completion in software,
 	 * consume; the dropped pages were already demanded through the pin
 	 * channel (r585) so the retry path had its chance first.
+	 *
+	 * r612: refuse only a DEGENERATE fill - one whose dropped set is at
+	 * least as large as the resolved set (r584's 2 resolved / 18
+	 * dropped). A launch with a full resolved window (969 resolved /
+	 * 9 dropped) whose unresolved candidates are still unresolved at
+	 * the absolute cap has them stalled in EVERY source for the whole
+	 * window; they are the param page's decoy/foreign pointers (the
+	 * r611 streams2 first-client capture: candidates 0x783299000000 /
+	 * 0x783297000000, `stalled VA` from the first fill, no pin, no
+	 * walk, no ledger, never resolving in the run) - waiting longer
+	 * cannot help, and refusing converts a launch whose real operands
+	 * all resolved into a guaranteed no-write client failure (streams2
+	 * bad=65534). At the cap the copy-related guards (dstpend, CE
+	 * watermark) have already cleared or the refusal branches above
+	 * win first, so a non-degenerate fill forwards historically,
+	 * loudly witnessed.
 	 */
 	if (svc_fwd_req.gr_arg_partial && !svc_gr_data_stalled &&
-	    svc_gr_defer_expired(svc_gr_cand.push_va)) {
+	    svc_gr_defer_expired(svc_gr_cand.push_va) &&
+	    svc_fwd_req.gr_n_data < svc_fwd_req.gr_dropped) {
 		svc.fwd_gr_refused++;
 		svc.fwd_gr_closed++;
 		if (svc_gr_cand.rel_va) {
@@ -21822,6 +21843,21 @@ static bool svc_gr_request(struct nv_vgpu_device *vgpu,
 				 * candidate is owned and sw-completed */
 	}
 	/*
+	 * r612: see the refusal note above - a non-degenerate fill at the cap
+	 * forwards historically (witnessed once per budget slot).
+	 */
+	if (svc_fwd_req.gr_arg_partial && !svc_gr_data_stalled &&
+	    svc_gr_defer_expired(svc_gr_cand.push_va)) {
+		svc.fwd_gr_argfw++;
+		if (svc.fwd_gr_argfw_log < 32) {
+			svc.fwd_gr_argfw_log++;
+			pr_info("vgsp-work: GR forward at defer cap h=0x%x push=0x%llx (partial argument set: %u dropped of %u resolved; stalled candidates never resolve, forwarding historically)\n",
+				ch->h_object, svc_gr_cand.push_va,
+				svc_fwd_req.gr_dropped,
+				svc_fwd_req.gr_n_data);
+		}
+	}
+	/*
 	 * m150: the fill found high argument pointers that no source can
 	 * translate (the second-client zeros class). Do NOT forward a launch
 	 * whose kernel would dereference unresolved pointers: leave the