vgsp: mode 31 - D-golden-first TSG-sibling ordering (compile-verified, untested)m31

1 files changed, 139 insertions, 33 deletions

diff --git a/kernel/gsp_phys_booter.c b/kernel/gsp_phys_booter.c
index 11f205c..dc8e641 100644
--- a/kernel/gsp_phys_booter.c
+++ b/kernel/gsp_phys_booter.c
@@ -660,6 +660,12 @@ bool gsp_phys_boot_requested(void)
  */
 #define GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING	30u
 /*
+ * Mode 31 (2026-10-06): mode-25 D first-init + a clean TSG sibling A whose
+ * 9-entry promote runs after V6 exists, so fifoCommitRamfcEngCtxForTsg
+ * commits V6's instance pointer + RAMFC while V6's own promote is suppressed.
+ */
+#define GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT	31u
+/*
  * 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
@@ -773,7 +779,7 @@ 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)
+	(GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT)
 
 /*
  * Work-submit token notifier slot. The default index is
@@ -1027,6 +1033,12 @@ bool gsp_phys_boot_requested(void)
 #define GSP_PHYS_RMCALL_GOLDEN_FB_USERD		0x1000u
 #define GSP_PHYS_RMCALL_GOLDEN_FB_RAMFC		0x2000u
 #define GSP_PHYS_RMCALL_GOLDEN_FB_MTHD		0x3000u
+/*
+ * Mode 31: the TSG-parented sibling A (the explicit TSG-commit carrier) and
+ * its FBMEM instance/RAMFC/mthd memdesc block.
+ */
+#define GSP_PHYS_RMCALL_HANDLE_GOLDEN_SIBLING	0xcaf00084u
+#define GSP_PHYS_RMCALL_HANDLE_GOLDEN_SIBLING_FB	0xcaf00088u
 
 /*
  * Golden-image replay sizes (kernel_graphics.c:2058-2060, 2273-2276): the
@@ -7629,6 +7641,8 @@ static const char *gsp_phys_rmcall_gr_promote_name(void)
 		return "mode 29: construct-last golden execution lane (mode 28 minus the TURING_COMPUTE_A object construct on the wired golden channel)";
 	case GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING:
 		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";
 	default:
 		return "off (PROMOTE_CTX never sent)";
 	}
@@ -7817,7 +7831,9 @@ static int gsp_phys_rmcall_alloc_channel(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)) {
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING ||
+	     boot_stage.rmcall_gr_promote ==
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT)) {
 		vars = &gsp_phys_rmcall_channel_variant_skipctx;
 		attempts = 1;
 	}
@@ -8278,7 +8294,9 @@ static int gsp_phys_rmcall_golden_replay(void __iomem *bar0)
 	     boot_stage.rmcall_gr_promote ==
 			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_EXEC_LAST ||
 	     boot_stage.rmcall_gr_promote ==
-			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING) &&
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING ||
+	     boot_stage.rmcall_gr_promote ==
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT) &&
 	    gsp_phys_rmcall_engine_is_gr()) {
 		if (gsp_phys_rmcall_fwclient_mask_value() &
 		    GSP_PHYS_RMCALL_GOLDEN9_MASK_LADDER_VAS) {
@@ -8388,7 +8406,9 @@ static int gsp_phys_rmcall_golden_replay(void __iomem *bar0)
 	     boot_stage.rmcall_gr_promote ==
 			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_EXEC_LAST ||
 	     boot_stage.rmcall_gr_promote ==
-			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING) &&
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING ||
+	     boot_stage.rmcall_gr_promote ==
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT) &&
 	    gsp_phys_rmcall_engine_is_gr()) {
 		ch.flags = cpu_to_le32(GSP_PHYS_RMCALL_CHANNEL_FLAG_GOLDEN9);
 		ch.internal_flags = cpu_to_le32(
@@ -8592,7 +8612,9 @@ static int gsp_phys_rmcall_golden_replay(void __iomem *bar0)
 	     boot_stage.rmcall_gr_promote ==
 			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_EXEC_LAST ||
 	     boot_stage.rmcall_gr_promote ==
-			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING) &&
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING ||
+	     boot_stage.rmcall_gr_promote ==
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT) &&
 	    gsp_phys_rmcall_engine_is_gr()) {
 		ret = gsp_phys_rmcall_golden9_promote(bar0);
 		if (ret)
@@ -9914,6 +9936,59 @@ 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.
+ */
+static int gsp_phys_rmcall_golden_sibling_alloc(void __iomem *bar0)
+{
+	struct gsp_phys_rmcall_channel_params ch;
+	u32 rpc = 0, status = 0;
+	u64 fb = 0;
+	int ret;
+
+	memset(&ch, 0, sizeof(ch));
+	ch.h_va_space = cpu_to_le32(0);
+	ch.gp_fifo_entries =
+		cpu_to_le32(GSP_PHYS_RMCALL_GOLDEN_GPFIFO_ENTRIES);
+	ch.engine_type = cpu_to_le32(GSP_PHYS_RMCALL_ENGINE_GR0);
+	ch.flags = cpu_to_le32(GSP_PHYS_RMCALL_CHANNEL_FLAG_GOLDEN9);
+	ch.internal_flags = cpu_to_le32(
+		GSP_PHYS_RMCALL_CHANNEL_INTERNALFLAGS_PRIV_KERNEL);
+	ret = gsp_phys_rmcall_alloc_fb_mem(bar0,
+			GSP_PHYS_RMCALL_HANDLE_GOLDEN_SIBLING_FB, 0x4000, &fb,
+			"GOLDEN SIBLING A memdescs");
+	if (ret)
+		return ret;
+	ch.instance_mem = gsp_phys_rmcall_mem_desc_fb(fb + 0x0000, 0x1000);
+	ch.ramfc_mem = gsp_phys_rmcall_mem_desc_fb(fb + 0x2000, 0x1000);
+	ch.mthdbuf_mem = gsp_phys_rmcall_mem_desc_fb(fb + 0x3000, 0x1000);
+	ret = gsp_phys_rmcall_alloc(bar0, boot_stage.rmcall_client,
+			boot_stage.rmcall_tsg,
+			GSP_PHYS_RMCALL_HANDLE_GOLDEN_SIBLING,
+			GSP_PHYS_RMCALL_CLASS_CHANNEL, &ch, sizeof(ch),
+			false, NULL,
+			"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");
+	ret = gsp_phys_rmcall_golden9_promote_on(bar0,
+			GSP_PHYS_RMCALL_HANDLE_GOLDEN_SIBLING);
+	if (ret)
+		return ret;
+	gsp_phys_rmcall_ctx_diag_dump("INSTANCE (post-A-commit pre-BIND)",
+		&boot_stage.rmcall_chunks[GSP_PHYS_RMCALL_BUF_INSTANCE]);
+	gsp_phys_rmcall_ctx_diag_dump("RAMFC (post-A-commit pre-BIND)",
+		&boot_stage.rmcall_chunks[GSP_PHYS_RMCALL_BUF_RAMFC]);
+	gsp_phys_rmcall_ctx_diag_decode("post-A-commit pre-BIND");
+	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
@@ -10098,22 +10173,27 @@ static int gsp_phys_rmcall_gr_ctx_promote(void __iomem *bar0)
 	     mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN9_V6_ONLY ||
 	     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 ||
+	     mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT) &&
 	    (!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: 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
+	 * Mode 30/31: 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)
 	 * and the single-MAIN exemption needs the pCtxPhysMemDesc the FWCLIENT
 	 * shapes never establish, so a V6 promote would abort stage 3 on 0x1f.
 	 */
-	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");
+	if (mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING ||
+	    mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT) {
+		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
+			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");
 		boot_stage.rmcall_gr_promote_step = 1;
 		return 0;
 	}
@@ -11949,7 +12029,9 @@ static int gsp_phys_rmcall_stage3(void __iomem *bar0)
 			     boot_stage.rmcall_gr_promote ==
 					GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_EXEC_LAST ||
 			     boot_stage.rmcall_gr_promote ==
-					GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING) &&
+					GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING ||
+			     boot_stage.rmcall_gr_promote ==
+					GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT) &&
 			    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);
@@ -11994,22 +12076,26 @@ static int gsp_phys_rmcall_stage3(void __iomem *bar0)
 		return ret;
 
 	/*
-	 * Mode 30 readback point 1: V6's instance/RAMFC right after the channel
-	 * alloc (A's replay contains the 9-entry promote and A's TURING_A
-	 * first-init by now), so A's commit effect can be separated from the
-	 * later BIND/GPFIFO_SCHEDULE. The decode line prints the two decisive
-	 * fields (instance WFI pointer pair + RAMFC ENG_CTX_VALID).
+	 * Mode 30/31 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).
 	 */
 	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 ||
+	     boot_stage.rmcall_gr_promote ==
+			GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT)) {
 		gsp_phys_rmcall_ctx_diag_dump(
-			"INSTANCE (post-A-commit pre-BIND)",
+			"INSTANCE (post-V6-alloc pre-A-commit)",
 			&boot_stage.rmcall_chunks[GSP_PHYS_RMCALL_BUF_INSTANCE]);
 		gsp_phys_rmcall_ctx_diag_dump(
-			"RAMFC (post-A-commit pre-BIND)",
+			"RAMFC (post-V6-alloc pre-A-commit)",
 			&boot_stage.rmcall_chunks[GSP_PHYS_RMCALL_BUF_RAMFC]);
-		gsp_phys_rmcall_ctx_diag_decode("post-A-commit pre-BIND");
+		gsp_phys_rmcall_ctx_diag_decode("post-V6-alloc pre-A-commit");
 	}
 
 	/*
@@ -12079,6 +12165,21 @@ static int gsp_phys_rmcall_stage3(void __iomem *bar0)
 				boot_stage.rmcall_ctx_size);
 			return -EOPNOTSUPP;
 		}
+		/*
+		 * Mode 31: allocate the TSG-parented sibling A and run its
+		 * 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.
+		 */
+		if (boot_stage.rmcall_gr_promote ==
+				GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT) {
+			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);
+				return ret;
+			}
+		}
 		ret = gsp_phys_rmcall_gr_ctx_promote(bar0);
 		if (ret) {
 			pr_err("vgsp-boot: rm-call: stage 3 aborted: GPU_PROMOTE_CTX failed (%d); the GR context is not committed and BIND/SCHEDULE/submit are not attempted (no retry; UNLOAD + rmmod remains the fail-safe)\n",
@@ -14694,12 +14795,14 @@ static void gsp_phys_rmcall_ctx_diag_dump(const char *what,
 }
 
 /*
- * One-line decode of the two fields that decide whether V6's engine context is
+ * One-line decode of the fields that decide whether V6's engine context is
  * live: the instance block's NV_RAMIN_ENGINE_WFI_PTR pair (dword 132 bits
- * 31:12 = PA[31:12], dword 133 bits 7:0 = PA[39:32]; fifo_gv100.c) and RAMFC
- * NV_RAMFC_TARGET (dword 43, NV_PPBDMA_TARGET_ENG_CTX_VALID bit 16). Host-view
- * LE chunks, same as the dump above; used at the mode-30 post-replay and the
- * stage-4 post-BIND points so A's commit can be told apart from BIND.
+ * 31:12 = PA[31:12], dword 133 bits 7:0 = PA[39:32]; fifo_gv100.c) and the
+ * NV_RAMFC_TARGET dword (43, NV_PPBDMA_TARGET_ENG_CTX_VALID bit 16) in BOTH
+ * the instance and RAMFC chunks - which of the two the PBDMA actually reads
+ * is an open question (m30b saw the bit only in the instance view). Host-view
+ * LE chunks, same as the dump above; used at the mode-30/31 pre- and
+ * post-commit points and the stage-4 post-BIND point.
  */
 static void gsp_phys_rmcall_ctx_diag_decode(const char *tag)
 {
@@ -14709,8 +14812,8 @@ static void gsp_phys_rmcall_ctx_diag_decode(const char *tag)
 		&boot_stage.rmcall_chunks[GSP_PHYS_RMCALL_BUF_RAMFC];
 	const __le32 *iv, *rv;
 	u64 ctx_pa;
-	u32 target;
-	bool valid;
+	u32 inst_target, ramfc_target;
+	bool inst_valid, ramfc_valid;
 
 	if (!inst->pages || !ramfc->pages) {
 		pr_info("vgsp-boot: rm-call: V6 ctx (%s): instance/RAMFC chunk absent\n",
@@ -14721,10 +14824,13 @@ static void gsp_phys_rmcall_ctx_diag_decode(const char *tag)
 	rv = gsp_phys_chunk_page_va(ramfc, 0);
 	ctx_pa = ((u64)(le32_to_cpu(iv[133]) & 0xffu) << 32) |
 		 (le32_to_cpu(iv[132]) & 0xfffff000u);
-	target = le32_to_cpu(rv[43]);
-	valid = !!(target & (1u << 16));
-	pr_info("vgsp-boot: rm-call: V6 ctx (%s): inst_wfi=0x%llx RAMFC_TARGET=0x%08x ENG_CTX_VALID=%u\n",
-		tag, ctx_pa, target, valid);
+	inst_target = le32_to_cpu(iv[43]);
+	ramfc_target = le32_to_cpu(rv[43]);
+	inst_valid = !!(inst_target & (1u << 16));
+	ramfc_valid = !!(ramfc_target & (1u << 16));
+	pr_info("vgsp-boot: rm-call: V6 ctx (%s): inst_wfi=0x%llx INST_TARGET=0x%08x RAMFC_TARGET=0x%08x ENG_CTX_VALID(inst=%u,ramfc=%u)\n",
+		tag, ctx_pa, inst_target, ramfc_target, inst_valid,
+		ramfc_valid);
 }
 
 static int gsp_phys_rmcall_stage4(void __iomem *bar0)