vgsp-gr: CONT-114 golden MME slot overlay (sema_gl_mme_golden) - the capture lottery loses on partial boots (13/55 slots; r1405/6/10/11/12) and the called DRAW_IAB slot 34 gets keep-clamped+call-dropped, so no draw ever executes; this overlay serves called-but-uncaptured SA slots from the driver's own generated macro code (word-exact vs a live full capture: 34=0x265/75w, 39=0x29f/69w, 40=0x2b6/69w, 48=0x337/21w, 51=0x350/12w, 57=0x364/189w) via an untracked .gitignore'd build input kernel/golden-mme-tu104.h: slot_kept/slot_stubbed consult it, the window scan counts golden slots as covered (no stub pair, calls survive), and the pack builder overlays the golden words+SA pairs into ir_pack after the capture fill (mode 1 = uncovered only, force=2 = overwrite all); golden-filled slots are byte-shape-identical to captured ones (0x11c pairs + one 0x114 ONEINC chunk) so the IRAM gate and downstream consumers are unchangedHEADmaster

2 files changed, 170 insertions, 2 deletions

diff --git a/.gitignore b/.gitignore
index 6eb6509..4d20bc9 100644
--- a/.gitignore
+++ b/.gitignore
@@ -165,3 +165,4 @@ userspace/sdks/
 # KVMFR capture).
 /*.mp4
 /open-iov-*.mp4
+kernel/golden-mme-tu104.h
diff --git a/kernel/gsp_phys_booter.c b/kernel/gsp_phys_booter.c
index 2af384c..5a5412e 100644
--- a/kernel/gsp_phys_booter.c
+++ b/kernel/gsp_phys_booter.c
@@ -58,6 +58,15 @@
 #include "gsp_phys_booter_load_img.h"
 #include "gsp_phys_gsp_rm_boot_img.h"
 #include "gsp_phys_registry_img.h"
+/*
+ * SESSION 153-CONTINUED-114: golden MME slot library (untracked build
+ * input, .gitignore'd - see the file header for provenance). Used only
+ * when sema_gl_mme_golden=1: called-but-uncaptured slots are served
+ * from the driver's own generated macro code instead of being stubbed,
+ * so the DRAW_IAB call (slot 34) runs deterministically on every boot
+ * (the full-capture profile becomes a per-boot lottery otherwise).
+ */
+#include "golden-mme-tu104.h"
 #include "gsp_phys_registry_img_srvctx.h"
 
 static bool gsp_phys_boot_param;
@@ -18206,6 +18215,53 @@ MODULE_PARM_DESC(sema_gl_mme_keep_clamp,
 static u32 gsp_phys_gr_keep_clamp_log;
 
 /*
+ * SESSION 153-CONTINUED-114: golden MME slot overlay (see the include
+ * comment). The capture is a per-boot lottery (r1407/r1385 got the full
+ * 55-slot program; r1405/6/10/11/12 only 13 slots), and a called slot
+ * outside the capture is stubbed with its CALL runs dropped - which is
+ * why the draw never executes on partial boots. The golden library is
+ * the same driver-emitted microcode word-for-word (validated against a
+ * live full capture), so a golden-filled slot is byte-shape-identical
+ * to a captured one downstream (pack image, 0x11c pairs, IRAM gate).
+ *
+ * gsp_phys_gr_golden_slot: returns 1 and the slot's instruction entry +
+ * word pointer when the library can serve slot j (and the knob is on).
+ * gsp_phys_gr_golden_active: the knob as a bool, for the cheap sites.
+ */
+static unsigned int gsp_phys_rmcall_gr_mme_golden;
+module_param_named(sema_gl_mme_golden, gsp_phys_rmcall_gr_mme_golden,
+		   uint, 0644);
+MODULE_PARM_DESC(sema_gl_mme_golden,
+	"Golden MME slot overlay: serve called-but-uncaptured SA slots from the driver's own generated macro code (untracked kernel/golden-mme-tu104.h) instead of stubbing them, so CALL_MME_MACRO(DRAW_IAB etc.) runs deterministically (default 0; 1 = fill uncovered slots; 2 = force/overwrite all)");
+static unsigned int gsp_phys_rmcall_gr_mme_golden_force;
+module_param_named(sema_gl_mme_golden_force, gsp_phys_rmcall_gr_mme_golden_force,
+		   uint, 0644);
+MODULE_PARM_DESC(sema_gl_mme_golden_force,
+	"Force the golden overlay over captured slots too (A/B for a stale capture; default 0 = fill uncovered only)");
+
+static u32 gsp_phys_gr_golden_log;
+
+static int gsp_phys_gr_golden_slot(u32 j, u32 *entry, const u32 **words,
+				   u32 *nwords)
+{
+	const struct golden_mme_tu104_slot *gs;
+
+	if (!READ_ONCE(gsp_phys_rmcall_gr_mme_golden))
+		return 0;
+	if (j >= GOLDEN_MME_TU104_SLOTS || j >= 64u)
+		return 0;
+	gs = &golden_mme_tu104_sa[j];
+	if (!(gs->flags & GMME_F_PRESENT) || !gs->words || !gs->entry)
+		return 0;
+	if ((u64)gs->off + gs->words > GOLDEN_MME_TU104_WORDS)
+		return 0;
+	*entry = gs->entry;
+	*words = &golden_mme_tu104_words[gs->off];
+	*nwords = gs->words;
+	return 1;
+}
+
+/*
  * SESSION 153-CONTINUED-79: per-launch method-block word cap + prelude drain.
  * The r1172-r1175 A/B: every beef launch whose emitted block exceeded ~2000
  * words (3083-word full-SA prelude + window, 3315-3484 words) NEVER STARTED
@@ -19007,6 +19063,18 @@ static bool gsp_phys_gr_slot_kept(u32 j, u64 sa_mask, const u32 *sa_ent)
 		return true;
 	if ((sa_mask & (1ull << j)) && gsp_sema_mme_ir_covered(sa_ent[j]))
 		return true;
+	/*
+	 * SESSION 153-CONTINUED-114: a keep_mask slot the golden library
+	 * serves is kept without capture coverage - the pack overlay
+	 * carries its code (see the include comment).
+	 */
+	{
+		u32 gent, gn;
+		const u32 *gw;
+
+		if (gsp_phys_gr_golden_slot(j, &gent, &gw, &gn))
+			return true;
+	}
 	if (gsp_phys_gr_keep_clamp_log < 16u) {
 		gsp_phys_gr_keep_clamp_log++;
 		pr_warn("vgsp-gr: MME-KEEP-CLAMP slot=%u (keep_mask set but capture has no SA entry/coverage; stubbed + calls dropped, never run) entry=0x%x covered=%u\n",
@@ -19032,6 +19100,18 @@ static bool gsp_phys_gr_slot_stubbed(u32 j, u64 sa_mask, const u32 *sa_ent)
 		return true;
 	if (gsp_phys_gr_slot_kept(j, sa_mask, sa_ent))
 		return false;
+	/*
+	 * SESSION 153-CONTINUED-114: a slot the golden library serves is
+	 * never stubbed (its CALL runs stay in the window and the pack
+	 * carries its code; see the include comment).
+	 */
+	{
+		u32 gent, gn;
+		const u32 *gw;
+
+		if (gsp_phys_gr_golden_slot(j, &gent, &gw, &gn))
+			return false;
+	}
 	if (READ_ONCE(gsp_phys_rmcall_gr_sa_stub_all))
 		return true;
 	if (!(sa_mask & (1ull << j)))
@@ -19477,8 +19557,8 @@ static u32 gsp_phys_rmcall_build_gr_launch(u32 *pb,
 					 * (keep_clamp) - then it falls into
 					 * the stub + call-drop path below.
 					 */
-					if (gsp_phys_gr_slot_kept(j, sa_mask,
-								  sa_ent)) {
+if (gsp_phys_gr_slot_kept(j, sa_mask,
+							  sa_ent)) {
 						for (t = 0; t < sa_ncov; t++)
 							if (sa_cov[t] == j)
 								have = true;
@@ -19491,6 +19571,36 @@ static u32 gsp_phys_rmcall_build_gr_launch(u32 *pb,
 						continue;
 					}
 					/*
+					 * SESSION 153-CONTINUED-114: a called
+					 * slot outside the capture that the
+					 * golden library serves counts as
+					 * covered - the pack overlay emits its
+					 * SA pair + code, so no stub pair is
+					 * emitted and the CALL runs survive
+					 * the drop pass (see slot_stubbed).
+					 */
+					{
+						u32 gent, gn;
+						const u32 *gw;
+
+						if (gsp_phys_gr_golden_slot(
+							    j, &gent, &gw,
+							    &gn)) {
+							for (t = 0; t < sa_ncov;
+							     t++)
+								if (sa_cov[t] == j)
+									have = true;
+							if (!have &&
+							    sa_ncov < 40u) {
+								sa_cov[sa_ncov] = j;
+								sa_cov_ent[sa_ncov] =
+									gent;
+								sa_ncov++;
+							}
+							continue;
+						}
+					}
+					/*
 					 * SESSION 153-CONTINUED-68: a slot is only
 					 * usable when the capture holds its SA entry
 					 * AND the IR instruction that entry names
@@ -20107,6 +20217,63 @@ static u32 gsp_phys_rmcall_build_gr_launch(u32 *pb,
 					r += span;
 				}
 				/*
+				 * SESSION 153-CONTINUED-114: golden overlay.
+				 * After the capture fill, merge the golden
+				 * library's words into the packed image and
+				 * the local SA pair table for every slot the
+				 * capture cannot serve (mode 1), or every
+				 * slot (mode 2 / force). Written after the
+				 * capture fill so the golden words win any
+				 * overlap; a complete capture is untouched
+				 * in mode 1 (the overlay then only logs).
+				 */
+				if (READ_ONCE(gsp_phys_rmcall_gr_mme_golden)) {
+					u32 gj, goldn = 0;
+
+					for (gj = 0; gj < 64u; gj++) {
+						u32 gent, gn, q;
+						const u32 *gw;
+
+						if (!gsp_phys_gr_golden_slot(
+							    gj, &gent, &gw,
+							    &gn))
+							continue;
+						if (!READ_ONCE(gsp_phys_rmcall_gr_mme_golden_force)) {
+							/*
+							 * Mode 1: only slots the
+							 * capture cannot serve
+							 * (no SA entry, or the
+							 * entry's IR never
+							 * entered the replay).
+							 */
+							if ((sa_mask &
+							     (1ull << gj)) &&
+							    sa_ent[gj] &&
+							    gsp_sema_mme_ir_covered(
+								    sa_ent[gj]))
+								continue;
+						}
+						for (q = 0;
+						     q < gn &&
+						     gent * 3u + q < 4096u;
+						     q++)
+							gsp_phys_gr_mme_ir_pack[gent * 3u + q] =
+								gw[q];
+						if (gent * 3u + gn > maxa)
+							maxa = gent * 3u + gn;
+						sa_mask |= 1ull << gj;
+						sa_ent[gj] = gent;
+						goldn++;
+					}
+					if (goldn && gsp_phys_gr_golden_log < 8u) {
+						gsp_phys_gr_golden_log++;
+						pr_info("vgsp-gr: MME-GOLDEN overlay %u slot(s) maxa=%u h=0x%x push=0x%llx\n",
+							goldn, maxa,
+							req->h_object,
+							req->push_va);
+					}
+				}
+				/*
 				 * CONT-112: single-word A/B override of the
 				 * packed program (see the sema_gl_mme_irpatch
 				 * comment). Applied after the image is built so