vgsp: mode 32 - sibling A promote before V6 constructs (compile-verified, untested)m32reorder

2 files changed, 140 insertions, 52 deletions

diff --git a/kernel/gsp_phys_booter.c b/kernel/gsp_phys_booter.c
index 98b85e8..c5de21e 100644
--- a/kernel/gsp_phys_booter.c
+++ b/kernel/gsp_phys_booter.c
@@ -666,6 +666,15 @@ bool gsp_phys_boot_requested(void)
  */
 #define GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT	31u
 /*
+ * Mode 32 (2026-10-06): mode 31 with one ordering change. The TSG-parented
+ * sibling A's alloc + 9-entry promote run immediately after V6's channel alloc
+ * and BEFORE V6's GR object constructs (0xc597 on the ladder channel and the
+ * 0xc5c0 compute object), inverting the mode-31 order whose A commit asserted
+ * GSP-RM (NOCAT 0x595f724/0x5162778). V6's own promote stays suppressed and
+ * V6 remains the execution channel.
+ */
+#define GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT_EARLY	32u
+/*
  * Mode-28 golden execution channel shape bits (gsp_phys_rmcall_golden_exec_shape,
  * default 7). Every combination keeps the captured flags 0x00200320 and the
  * FBMEM instance/RAMFC/mthd slices; the bits only choose the client-visible
@@ -779,7 +788,7 @@ bool gsp_phys_boot_requested(void)
  */
 #define GSP_PHYS_RMCALL_FWCLIENT_MASK_FECS_SYSMEM	(1u << 16)
 #define GSP_PHYS_RMCALL_GR_PROMOTE_MAX \
-	(GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT)
+	(GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT_EARLY)
 
 /*
  * Work-submit token notifier slot. The default index is
@@ -7643,6 +7652,8 @@ static const char *gsp_phys_rmcall_gr_promote_name(void)
 		return "mode 30: clean unwired TSG sibling A (promote + TURING_A first-init) + V6 execution";
 	case GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT:
 		return "mode 31: mode-25 D first-init + TSG sibling A 9-entry promote after V6 (explicit TSG commit) + V6 execution";
+	case GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT_EARLY:
+		return "mode 32: mode 31 with the TSG sibling A alloc+promote immediately after V6's alloc and before V6's GR constructs + V6 execution";
 	default:
 		return "off (PROMOTE_CTX never sent)";
 	}
@@ -7833,7 +7844,9 @@ static int gsp_phys_rmcall_alloc_channel(void __iomem *bar0)
 	     boot_stage.rmcall_gr_promote ==
 			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING ||
 	     boot_stage.rmcall_gr_promote ==
-			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT)) {
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT ||
+	     boot_stage.rmcall_gr_promote ==
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT_EARLY)) {
 		vars = &gsp_phys_rmcall_channel_variant_skipctx;
 		attempts = 1;
 	}
@@ -8296,7 +8309,9 @@ static int gsp_phys_rmcall_golden_replay(void __iomem *bar0)
 	     boot_stage.rmcall_gr_promote ==
 			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING ||
 	     boot_stage.rmcall_gr_promote ==
-			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT) &&
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT ||
+	     boot_stage.rmcall_gr_promote ==
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT_EARLY) &&
 	    gsp_phys_rmcall_engine_is_gr()) {
 		if (gsp_phys_rmcall_fwclient_mask_value() &
 		    GSP_PHYS_RMCALL_GOLDEN9_MASK_LADDER_VAS) {
@@ -8408,7 +8423,9 @@ static int gsp_phys_rmcall_golden_replay(void __iomem *bar0)
 	     boot_stage.rmcall_gr_promote ==
 			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING ||
 	     boot_stage.rmcall_gr_promote ==
-			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT) &&
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT ||
+	     boot_stage.rmcall_gr_promote ==
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT_EARLY) &&
 	    gsp_phys_rmcall_engine_is_gr()) {
 		ch.flags = cpu_to_le32(GSP_PHYS_RMCALL_CHANNEL_FLAG_GOLDEN9);
 		ch.internal_flags = cpu_to_le32(
@@ -8614,7 +8631,9 @@ static int gsp_phys_rmcall_golden_replay(void __iomem *bar0)
 	     boot_stage.rmcall_gr_promote ==
 			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING ||
 	     boot_stage.rmcall_gr_promote ==
-			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT) &&
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT ||
+	     boot_stage.rmcall_gr_promote ==
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT_EARLY) &&
 	    gsp_phys_rmcall_engine_is_gr()) {
 		ret = gsp_phys_rmcall_golden9_promote(bar0);
 		if (ret)
@@ -9477,10 +9496,10 @@ static int gsp_phys_rmcall_gr_ctx_prepare(void __iomem *bar0)
 	 * Modes 11/15 back the context buffer with a server-owned FBMEM
 	 * allocation (VIDMEM aperture, the shape the KMD promote path
 	 * expects); mode 23 uses the same FBMEM object as the hVirtMemory
-	 * promote target; mode 25 and mode 31 use the FBMEM MAIN that made
-	 * the golden first-init construct succeed (m25c); every other mode
-	 * keeps the host sysmem chunk. The PT pages are always host sysmem
-	 * pages behind the sysmem-aperture PD0.
+	 * promote target; mode 25 and modes 31/32 use the FBMEM MAIN that
+	 * made the golden first-init construct succeed (m25c); every other
+	 * mode keeps the host sysmem chunk. The PT pages are always host
+	 * sysmem pages behind the sysmem-aperture PD0.
 	 */
 	if (boot_stage.rmcall_gr_promote ==
 			GSP_PHYS_RMCALL_GR_PROMOTE_OLDIFACE_FB ||
@@ -9491,7 +9510,9 @@ static int gsp_phys_rmcall_gr_ctx_prepare(void __iomem *bar0)
 	    boot_stage.rmcall_gr_promote ==
 			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN9 ||
 	    boot_stage.rmcall_gr_promote ==
-			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT) {
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT ||
+	    boot_stage.rmcall_gr_promote ==
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT_EARLY) {
 		u64 fb_base = 0;
 
 		ret = gsp_phys_rmcall_alloc_fb_mem(bar0,
@@ -9544,14 +9565,16 @@ static int gsp_phys_rmcall_gr_ctx_prepare(void __iomem *bar0)
 			"server FBMEM, VIDMEM aperture PTEs" :
 			"host sysmem, contiguous, zeroed");
 	/*
-	 * Mode-25 golden9e and mode-31 (whose D first-init replay uses the
-	 * same golden9 promote): prepare the FBMEM PATCH/FECS/PRIV/UNR
+	 * Mode-25 golden9e and modes 31/32 (whose D first-init replay uses
+	 * the same golden9 promote): prepare the FBMEM PATCH/FECS/PRIV/UNR
 	 * buffers now, before the PDB build maps their promoted VAs.
 	 */
 	if (boot_stage.rmcall_gr_promote ==
 			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN9 ||
 	    boot_stage.rmcall_gr_promote ==
-			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT) {
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT ||
+	    boot_stage.rmcall_gr_promote ==
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT_EARLY) {
 		ret = gsp_phys_rmcall_golden9_fb_aux_prepare(bar0);
 		if (ret) {
 			pr_err("vgsp-boot: rm-call: GOLDEN9E FBMEM aux prepare failed (%d); aborting the ladder with no retry\n",
@@ -9943,12 +9966,13 @@ static int gsp_phys_rmcall_golden9_promote(void __iomem *bar0)
 }
 
 /*
- * Mode 31: allocate the clean TSG-parented sibling A (captured shape, no
- * ring/USERD/HOBJ wiring) and run its 9-entry golden9 promote after V6's own
- * constructs, so the server's TSG-wide engine-context commit walk
- * (fifo_gv100.c:1480-1494) commits V6's instance pointer + RAMFC while V6's
- * own promote is suppressed. Two 9-entry promotes total (D and A); V6 gets
- * none.
+ * Modes 31/32: allocate the clean TSG-parented sibling A (captured shape, no
+ * ring/USERD/HOBJ wiring) and run its 9-entry golden9 promote, so the
+ * server's TSG-wide engine-context commit walk (fifo_gv100.c:1480-1494)
+ * commits V6's instance pointer + RAMFC while V6's own promote is suppressed.
+ * Two 9-entry promotes total (D and A); V6 gets none. Mode 31 runs this after
+ * V6's own constructs; mode 32 runs it immediately after V6's channel alloc
+ * and before V6's constructs.
  */
 static int gsp_phys_rmcall_golden_sibling_alloc(void __iomem *bar0)
 {
@@ -9979,11 +10003,17 @@ static int gsp_phys_rmcall_golden_sibling_alloc(void __iomem *bar0)
 			GSP_PHYS_RMCALL_HANDLE_GOLDEN_SIBLING,
 			GSP_PHYS_RMCALL_CLASS_CHANNEL, &ch, sizeof(ch),
 			false, NULL,
-			"mode 31 GOLDEN SIBLING A (TSG commit carrier)",
+			boot_stage.rmcall_gr_promote ==
+				GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT_EARLY ?
+				"mode 32 GOLDEN SIBLING A (pre-construct commit carrier)" :
+				"mode 31 GOLDEN SIBLING A (TSG commit carrier)",
 			true, &rpc, &status);
 	if (ret)
 		return ret;
-	gsp_phys_rmcall_ctx_diag_decode("post-V6-constructs pre-A-commit");
+	gsp_phys_rmcall_ctx_diag_decode(boot_stage.rmcall_gr_promote ==
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT_EARLY ?
+		"post-V6-alloc pre-A-commit" :
+		"post-V6-constructs pre-A-commit");
 	ret = gsp_phys_rmcall_golden9_promote_on(bar0,
 			GSP_PHYS_RMCALL_HANDLE_GOLDEN_SIBLING);
 	if (ret)
@@ -9997,6 +10027,32 @@ static int gsp_phys_rmcall_golden_sibling_alloc(void __iomem *bar0)
 }
 
 /*
+ * Modes 31/32 gating: sibling A consumes D's golden-replay first-init state
+ * and is only meaningful on the physical GR backend, so the golden and gpu
+ * gates are mandatory before the helper sends anything; abort with no retry
+ * otherwise (SA-CHK F1). The call sites stay mode-specific: mode 31 after
+ * V6's GR constructs, mode 32 immediately after V6's channel alloc.
+ */
+static int gsp_phys_rmcall_golden_sibling_commit(void __iomem *bar0)
+{
+	u32 mode = boot_stage.rmcall_gr_promote;
+	int ret;
+
+	if (!gsp_phys_rmcall_golden_enabled() || !gsp_phys_gpu_enabled()) {
+		pr_err("vgsp-boot: rm-call: mode %u requires gsp_phys_rmcall_golden=1 and gsp_phys_gpu=1 (the golden replay did not run); sibling A is not allocated and no promote is sent; the ladder aborts with no retry\n",
+		       mode);
+		return -EINVAL;
+	}
+	ret = gsp_phys_rmcall_golden_sibling_alloc(bar0);
+	if (ret) {
+		pr_err("vgsp-boot: rm-call: mode %u sibling A commit failed (%d); no V6 promote follows; the ladder aborts with no retry\n",
+		       mode, ret);
+		return ret;
+	}
+	return 0;
+}
+
+/*
  * Fill one MAIN promote entry with the prepared PA/VA and the snapshot's
  * entry size. initialize/nonmapped carry the bisection shape: modes 1/3/5
  * entry 0 and mode 2 step 1 use a full field set; mode 2 step 2 repeats the
@@ -10182,14 +10238,15 @@ static int gsp_phys_rmcall_gr_ctx_promote(void __iomem *bar0)
 	     mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_EXEC ||
 	     mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_EXEC_LAST ||
 	     mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING ||
-	     mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT) &&
+	     mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT ||
+	     mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT_EARLY) &&
 	    (!gsp_phys_rmcall_golden_enabled() || !gsp_phys_gpu_enabled())) {
 		pr_err("vgsp-boot: rm-call: mode %u requires gsp_phys_rmcall_golden=1 and gsp_phys_gpu=1 (the golden replay did not run); no PROMOTE_CTX is sent and the ladder aborts with no retry\n",
 		       mode);
 		return -EINVAL;
 	}
 	/*
-	 * Mode 30/31: the sibling A's 9-entry promote carries the group
+	 * Mode 30/31/32: the sibling A's 9-entry promote carries the group
 	 * context commit; nothing is sent on the V6 channel. A multi-entry
 	 * promote containing MAIN is refused for a SKIP-ctx channel
 	 * (graphics_context.c: bKMDVirtualContext -> NV_ERR_INVALID_ARGUMENT)
@@ -10197,11 +10254,14 @@ static int gsp_phys_rmcall_gr_ctx_promote(void __iomem *bar0)
 	 * shapes never establish, so a V6 promote would abort stage 3 on 0x1f.
 	 */
 	if (mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING ||
-	    mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT) {
+	    mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT ||
+	    mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT_EARLY) {
 		if (mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING)
 			pr_info("vgsp-boot: rm-call: mode 30: no PROMOTE_CTX on V6; the sibling A carries the group context commit (config_group ppEngCtxDesc), V6 inherits via the TSG commit/channel setup\n");
-		else
+		else if (mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT)
 			pr_info("vgsp-boot: rm-call: mode 31: no PROMOTE_CTX on V6; sibling A's 9-entry promote ran after V6's alloc and the TSG loop (fifo_gv100.c:1480-1494) already committed V6\n");
+		else
+			pr_info("vgsp-boot: rm-call: mode 32: no PROMOTE_CTX on V6; sibling A's 9-entry promote ran before V6's GR constructs and the TSG loop (fifo_gv100.c:1480-1494) may have committed V6\n");
 		boot_stage.rmcall_gr_promote_step = 1;
 		return 0;
 	}
@@ -12039,7 +12099,9 @@ static int gsp_phys_rmcall_stage3(void __iomem *bar0)
 			     boot_stage.rmcall_gr_promote ==
 					GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING ||
 			     boot_stage.rmcall_gr_promote ==
-					GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT) &&
+					GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT ||
+			     boot_stage.rmcall_gr_promote ==
+					GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT_EARLY) &&
 			    gsp_phys_rmcall_engine_is_gr()) {
 				pr_err("vgsp-boot: rm-call: stage 3 aborted: mode-%u golden replay failed (%d); the golden construct is atomic and is not retried\n",
 				       boot_stage.rmcall_gr_promote, ret);
@@ -12084,26 +12146,57 @@ static int gsp_phys_rmcall_stage3(void __iomem *bar0)
 		return ret;
 
 	/*
-	 * Mode 30/31 readback point 1: V6's instance/RAMFC right after the
+	 * Mode 32: allocate the TSG-parented sibling A and run its 9-entry
+	 * promote immediately after V6's channel alloc and BEFORE V6's GR
+	 * object constructs (the 0xc597 ladder construct and the 0xc5c0
+	 * compute object below), inverting the mode-31 order whose A commit
+	 * asserted GSP-RM (NOCAT 0x595f724/0x5162778). V6's own promote is
+	 * suppressed in gsp_phys_rmcall_gr_ctx_promote and V6 stays the
+	 * execution channel.
+	 */
+	if (gsp_phys_rmcall_engine_is_gr() &&
+	    boot_stage.rmcall_gr_promote ==
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT_EARLY) {
+		ret = gsp_phys_rmcall_golden_sibling_commit(bar0);
+		if (ret)
+			return ret;
+	}
+
+	/*
+	 * Mode 30/31/32 readback point 1: V6's instance/RAMFC right after the
 	 * channel alloc (mode 30: A's replay contains the 9-entry promote and
 	 * A's TURING_A first-init by now; mode 31: D's replay ran and the
-	 * sibling A commit plus V6's constructs are still ahead), so the
-	 * commit effect can be separated from the later BIND/GPFIFO_SCHEDULE.
-	 * The decode line prints the decisive fields (instance WFI pointer
-	 * pair + both ENG_CTX_VALID target words).
+	 * sibling A commit plus V6's constructs are still ahead; mode 32: the
+	 * sibling A commit already ran and V6's constructs are still ahead),
+	 * so the commit effect can be separated from the later
+	 * BIND/GPFIFO_SCHEDULE. The decode line prints the decisive fields
+	 * (instance WFI pointer pair + both ENG_CTX_VALID target words).
 	 */
 	if (gsp_phys_rmcall_engine_is_gr() &&
 	    (boot_stage.rmcall_gr_promote ==
 			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING ||
 	     boot_stage.rmcall_gr_promote ==
-			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT)) {
-		gsp_phys_rmcall_ctx_diag_dump(
-			"INSTANCE (post-V6-alloc pre-A-commit)",
-			&boot_stage.rmcall_chunks[GSP_PHYS_RMCALL_BUF_INSTANCE]);
-		gsp_phys_rmcall_ctx_diag_dump(
-			"RAMFC (post-V6-alloc pre-A-commit)",
-			&boot_stage.rmcall_chunks[GSP_PHYS_RMCALL_BUF_RAMFC]);
-		gsp_phys_rmcall_ctx_diag_decode("post-V6-alloc pre-A-commit");
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT ||
+	     boot_stage.rmcall_gr_promote ==
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT_EARLY)) {
+		if (boot_stage.rmcall_gr_promote ==
+				GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT_EARLY) {
+			gsp_phys_rmcall_ctx_diag_dump(
+				"INSTANCE (post-A-commit pre-V6-constructs)",
+				&boot_stage.rmcall_chunks[GSP_PHYS_RMCALL_BUF_INSTANCE]);
+			gsp_phys_rmcall_ctx_diag_dump(
+				"RAMFC (post-A-commit pre-V6-constructs)",
+				&boot_stage.rmcall_chunks[GSP_PHYS_RMCALL_BUF_RAMFC]);
+			gsp_phys_rmcall_ctx_diag_decode("post-A-commit pre-V6-constructs");
+		} else {
+			gsp_phys_rmcall_ctx_diag_dump(
+				"INSTANCE (post-V6-alloc pre-A-commit)",
+				&boot_stage.rmcall_chunks[GSP_PHYS_RMCALL_BUF_INSTANCE]);
+			gsp_phys_rmcall_ctx_diag_dump(
+				"RAMFC (post-V6-alloc pre-A-commit)",
+				&boot_stage.rmcall_chunks[GSP_PHYS_RMCALL_BUF_RAMFC]);
+			gsp_phys_rmcall_ctx_diag_decode("post-V6-alloc pre-A-commit");
+		}
 	}
 
 	/*
@@ -12178,20 +12271,15 @@ static int gsp_phys_rmcall_stage3(void __iomem *bar0)
 		 * 9-entry promote here, after V6's own constructs, so the
 		 * server's TSG-wide commit walk sees V6 in its group; V6's
 		 * own promote is suppressed in gsp_phys_rmcall_gr_ctx_promote.
+		 * Mode 32 runs the same commit right after V6's channel alloc
+		 * instead (before the constructs); the shared wrapper carries
+		 * the golden/gpu gate and the abort policy.
 		 */
 		if (boot_stage.rmcall_gr_promote ==
 				GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT) {
-			if (!gsp_phys_rmcall_golden_enabled() ||
-			    !gsp_phys_gpu_enabled()) {
-				pr_err("vgsp-boot: rm-call: mode 31 requires gsp_phys_rmcall_golden=1 and gsp_phys_gpu=1 (the golden replay did not run); sibling A is not allocated and no promote is sent; the ladder aborts with no retry\n");
-				return -EINVAL;
-			}
-			ret = gsp_phys_rmcall_golden_sibling_alloc(bar0);
-			if (ret) {
-				pr_err("vgsp-boot: rm-call: mode 31 sibling A commit failed (%d); no V6 promote follows; the ladder aborts with no retry\n",
-				       ret);
+			ret = gsp_phys_rmcall_golden_sibling_commit(bar0);
+			if (ret)
 				return ret;
-			}
 		}
 		ret = gsp_phys_rmcall_gr_ctx_promote(bar0);
 		if (ret) {
diff --git a/kernel/gsp_phys_gpu.c b/kernel/gsp_phys_gpu.c
index 4848f4b..bfeaa10 100644
--- a/kernel/gsp_phys_gpu.c
+++ b/kernel/gsp_phys_gpu.c
@@ -551,8 +551,8 @@ static int gsp_phys_rmcall_gr_promote_set(const char *val,
 		return ret;
 
 	if (gsp_phys_rmcall_gr_promote_param < 0 ||
-	    gsp_phys_rmcall_gr_promote_param > 31) {
-		pr_warn("vgsp: gsp_phys_rmcall_gr_promote=%d is outside 0..31; clamping to 0 (no PROMOTE_CTX)\n",
+	    gsp_phys_rmcall_gr_promote_param > 32) {
+		pr_warn("vgsp: gsp_phys_rmcall_gr_promote=%d is outside 0..32; clamping to 0 (no PROMOTE_CTX)\n",
 			gsp_phys_rmcall_gr_promote_param);
 		gsp_phys_rmcall_gr_promote_param = 0;
 	}
@@ -571,7 +571,7 @@ module_param_cb(gsp_phys_rmcall_gr_promote,
 		&gsp_phys_rmcall_gr_promote_ops,
 		&gsp_phys_rmcall_gr_promote_param, 0644);
 MODULE_PARM_DESC(gsp_phys_rmcall_gr_promote,
-		 "GR context-promote bisection (default 0: never send PROMOTE_CTX; 1: single entry init+nonmapped=0 size=raw+subctx, the p6h1b shape; 2: two-phase PA-only nonmapped then VA-only with the same fields; 3: single entry size=raw; 4: two-phase with a zero-field VA-only step; 5: batched entryCount=2 PA-only init + zero-field VA-only; 6: no-entry hVirtMemory, NV01_MEMORY_VIRTUAL (class 0x70) on the VAS at the context window + PROMOTE_CTX entryCount=0; 7: no-entry bare virtAddress+size, entryCount=0; 8: single MAIN entry like 1 on the KMD-managed channel (V6 SKIP_CTXBUFFER_ALLOC), server creates the virtual context memdesc itself; 9: same V6 channel with the stock client's zero-field VA-only MAIN entry, bInitialize=0; 10: same V6 channel, GPU_INITIALIZE_CTX (0x2080012d) installs the physical context memdesc then PROMOTE_CTX entryCount=0 bare virtAddress+size; 11: mode 10 with a server FBMEM context buffer, VIDMEM physAttr 0x4; 12: V5 UVM/vGPU-faithful PA-only MAIN entry (bInitialize=1 bNonmapped=1, no VA) and continue to BIND/SCHEDULE/submit; 13: mode 12 plus a PATCH PA-only promote first, sized from GET_ENGINE_CONTEXT_PROPERTIES engineId 0x10; 14: KMD channel (V6) + PATCH PA-only then the single-entry MAIN init+VA (the flow that books the context VA and commits ENG_CTX_VALID); 15: mode 14 with a server FBMEM MAIN context buffer and VIDMEM physAttr 0x4; 16: mode 14 plus a PM ctxsw PA-only promote first; 17: no client PROMOTE_CTX, paired with gsp_phys_srv_ctxbuf=1 so the GSP-RM server-manages and commits the context; 18: mode 8 shape plus a client FERMI_CONTEXT_SHARE_A on the TSG (hContextShare) whose construct allocates the GraphicsSubcontext; 19: the stock guest's exact 9-entry promote set (MAIN 2 MB PA+VA, PATCH, global buffers, FECS event, private access maps) on V5; 20: mode 6's no-entry hVirtMemory promote on the KMD-managed V6 channel (SKIP_CTXBUFFER_ALLOC), untested - the combination the V5 refusal (0x1f) points at; 21: mode 7's no-entry bare (virtAddress, size) promote on the V6 KMD-managed channel; 22: GPU_INITIALIZE_CTX then mode 20's no-entry hVirtMemory promote; 23: the hVirtMemory no-entry promote backed by the server-owned FBMEM context object (GSP_RM_ALLOC class 0x40) on the server-managed shared V5 channel - a physical memdesc reaches the per-buffer commit without the ADDR_VIRTUAL vaList gate; 24: the reference FWCLIENT non-3D set {MAIN, PM, PATCH, FECS_EVENT, PRIV_ACCESS_MAP, UNRESTRICTED_PRIV_ACCESS_MAP} as one control on the V6 KMD-managed channel, combined PA+VA entries where the ladder has a VA (MAIN PDB-mapped; PATCH/FECS/UNRESTRICTED booking-only), PM and PRIV_ACCESS_MAP init-only (no VA source), unsourceable buffers omitted and logged); 25: the real golden-client construct on the golden replay handles (captured channel flags 0x00200320 + single atomic 9-entry PROMOTE_CTX on the golden channel), then the V6 FWCLIENT follow-on promote on the golden-initialized GR-instance state; requires gsp_phys_rmcall_golden=1 and gsp_phys_rmcall_engine=1; 26/27: golden9 promote on the V6 execution channel (dead ends); 28: the golden execution lane - the golden channel gets a client GPFIFO ring (VA_GPFIFO), a sysmem host-writable USERD and the ladder TSG as parent, keeps the 9-entry promote, and the ladder switches execution to it (V6 stays unscheduled); 29: mode 28 minus the TURING_COMPUTE_A (0xc5c0) object construct on the client-wired golden channel (every GR object construct there deadlocks RM), i.e. the construct-last golden execution lane; 30: the TSG-sibling split: a clean unwired golden sibling A (captured shape, TSG-parented) runs the 9-entry promote and its own TURING_A first-init while V6 stays the execution channel (no swap) with its FWCLIENT promote suppressed; 31: mode-25's device-parented golden first-init (9-entry promote + TURING_A) followed by a clean TSG-parented sibling A whose 9-entry promote runs after V6's own constructs, so the TSG-wide server commit sets V6's instance pointer + RAMFC while V6 stays the execution channel and its own promote is suppressed; values outside 0..31 clamp to 0 with a warning");
+		 "GR context-promote bisection (default 0: never send PROMOTE_CTX; 1: single entry init+nonmapped=0 size=raw+subctx, the p6h1b shape; 2: two-phase PA-only nonmapped then VA-only with the same fields; 3: single entry size=raw; 4: two-phase with a zero-field VA-only step; 5: batched entryCount=2 PA-only init + zero-field VA-only; 6: no-entry hVirtMemory, NV01_MEMORY_VIRTUAL (class 0x70) on the VAS at the context window + PROMOTE_CTX entryCount=0; 7: no-entry bare virtAddress+size, entryCount=0; 8: single MAIN entry like 1 on the KMD-managed channel (V6 SKIP_CTXBUFFER_ALLOC), server creates the virtual context memdesc itself; 9: same V6 channel with the stock client's zero-field VA-only MAIN entry, bInitialize=0; 10: same V6 channel, GPU_INITIALIZE_CTX (0x2080012d) installs the physical context memdesc then PROMOTE_CTX entryCount=0 bare virtAddress+size; 11: mode 10 with a server FBMEM context buffer, VIDMEM physAttr 0x4; 12: V5 UVM/vGPU-faithful PA-only MAIN entry (bInitialize=1 bNonmapped=1, no VA) and continue to BIND/SCHEDULE/submit; 13: mode 12 plus a PATCH PA-only promote first, sized from GET_ENGINE_CONTEXT_PROPERTIES engineId 0x10; 14: KMD channel (V6) + PATCH PA-only then the single-entry MAIN init+VA (the flow that books the context VA and commits ENG_CTX_VALID); 15: mode 14 with a server FBMEM MAIN context buffer and VIDMEM physAttr 0x4; 16: mode 14 plus a PM ctxsw PA-only promote first; 17: no client PROMOTE_CTX, paired with gsp_phys_srv_ctxbuf=1 so the GSP-RM server-manages and commits the context; 18: mode 8 shape plus a client FERMI_CONTEXT_SHARE_A on the TSG (hContextShare) whose construct allocates the GraphicsSubcontext; 19: the stock guest's exact 9-entry promote set (MAIN 2 MB PA+VA, PATCH, global buffers, FECS event, private access maps) on V5; 20: mode 6's no-entry hVirtMemory promote on the KMD-managed V6 channel (SKIP_CTXBUFFER_ALLOC), untested - the combination the V5 refusal (0x1f) points at; 21: mode 7's no-entry bare (virtAddress, size) promote on the V6 KMD-managed channel; 22: GPU_INITIALIZE_CTX then mode 20's no-entry hVirtMemory promote; 23: the hVirtMemory no-entry promote backed by the server-owned FBMEM context object (GSP_RM_ALLOC class 0x40) on the server-managed shared V5 channel - a physical memdesc reaches the per-buffer commit without the ADDR_VIRTUAL vaList gate; 24: the reference FWCLIENT non-3D set {MAIN, PM, PATCH, FECS_EVENT, PRIV_ACCESS_MAP, UNRESTRICTED_PRIV_ACCESS_MAP} as one control on the V6 KMD-managed channel, combined PA+VA entries where the ladder has a VA (MAIN PDB-mapped; PATCH/FECS/UNRESTRICTED booking-only), PM and PRIV_ACCESS_MAP init-only (no VA source), unsourceable buffers omitted and logged); 25: the real golden-client construct on the golden replay handles (captured channel flags 0x00200320 + single atomic 9-entry PROMOTE_CTX on the golden channel), then the V6 FWCLIENT follow-on promote on the golden-initialized GR-instance state; requires gsp_phys_rmcall_golden=1 and gsp_phys_rmcall_engine=1; 26/27: golden9 promote on the V6 execution channel (dead ends); 28: the golden execution lane - the golden channel gets a client GPFIFO ring (VA_GPFIFO), a sysmem host-writable USERD and the ladder TSG as parent, keeps the 9-entry promote, and the ladder switches execution to it (V6 stays unscheduled); 29: mode 28 minus the TURING_COMPUTE_A (0xc5c0) object construct on the client-wired golden channel (every GR object construct there deadlocks RM), i.e. the construct-last golden execution lane; 30: the TSG-sibling split: a clean unwired golden sibling A (captured shape, TSG-parented) runs the 9-entry promote and its own TURING_A first-init while V6 stays the execution channel (no swap) with its FWCLIENT promote suppressed; 31: mode-25's device-parented golden first-init (9-entry promote + TURING_A) followed by a clean TSG-parented sibling A whose 9-entry promote runs after V6's own constructs, so the TSG-wide server commit sets V6's instance pointer + RAMFC while V6 stays the execution channel and its own promote is suppressed; 32: mode 31 with the sibling A's alloc + 9-entry promote moved immediately after V6's channel alloc and before V6's GR object constructs (the m29 order whose A commit asserted is inverted); values outside 0..32 clamp to 0 with a warning");
 
 int gsp_phys_rmcall_gr_promote_value(void)
 {