vgsp: CONT-103 - TWO ROOT CAUSES FOUND: (1) '(engine fetch NONE)' was an ARTIFACT - qmd_dst_mem never set for stream3d (all three sites !stream3d-guarded 25707/25733/25741) so the sentinel test was vacuous; GP_GET proves the engine fetches (FETCH-WIT h=0xbeef0100 slot=4 gpget 4->5 engine-fetched); (2) DEADLOCK: calls-kept launch was 2126-2142 words (1613-word prelude re-emit) > CONT-79 ~2050-wall so it never started (r1345/47/48 begin grfifo=00000001 events=0 timeout) - FIXES: sema_gl_mme_prelude_live_skip (default 1; skip forced 1613-word prelude re-emission once the generation's final part was carried by an engine-executed launch; live-gen set at poll break only when GP_GET advanced, cleared on timeout) + sema_gl_mme_nop_pair (default 0; elide program-pair runs method space [0x2054,0x2158]) + FETCH-WIT instrumentation (poll break/timeout gpget A->B engine-fetched, gpget/gpput in GR LAUNCH submitted, honest 'engine fetch n/a (no QMD deposit...sentinel vacuous)' wording); r1356 fix1: words=500, launch starts (FETCH-WIT timeout slot=6 gpget 6->7) then traps CONT-101 pair trap=80002058 data=019a0900; r1357 fix1+2: MME-PAIR-DROP 25 word(s) window=386->361, words=475, 0x2058 gone, FE advances to CALL runs: trap=80102390 data=00000005 (mmegen=1) park grfifo 0f0f1602->00000001 = FIRST MME-GENERATED METHOD WRITE AT THE WALL; GATE1 black 0/900 pages - next: FE rejects first MME-generated method 0x2390 (mmegen=1 data=5): identify 0x2390 role (macro-port family; check MME-WINRUN/3D win dumps + mmeref + direct lane h=0x1000013) and why generated write traps (CONT-80: bare DATA/macro-port write without open macro context parks; generated port may need macro context conserved); keep mode stays OFF - SESSION 153 r1356-r1357

1 files changed, 162 insertions, 9 deletions

diff --git a/kernel/gsp_phys_booter.c b/kernel/gsp_phys_booter.c
index 663f276..eada9ae 100644
--- a/kernel/gsp_phys_booter.c
+++ b/kernel/gsp_phys_booter.c
@@ -17980,6 +17980,47 @@ static u32 gsp_phys_gr_mme_pre_emit_h;
 static u32 gsp_phys_gr_mme_pre_emit_off;
 
 /*
+ * CONT-103: prelude residency witness + calls-kept block shrink.
+ * The CONT-79 split-emission re-carries the whole prelude prefix on EVERY
+ * launch while keep_mask is set (idempotent re-emission, CONT-72), so the
+ * first calls-kept launch is 1613 prelude words + the 410-word window =
+ * 2126-2142 words - over the ~2050-word fetch wall (CONT-79: oversized
+ * blocks never start, PBDMA parks). The deadlock: the calls can only run
+ * after the program is resident, and the launch that would run them is too
+ * big to start. The re-emission is only needed while the program is NOT
+ * known resident; the launch that carries the final part is the witness
+ * (its poll break with GP_GET advanced proves the engine consumed it).
+ * live_gen records the generation whose carrier executed; while it equals
+ * the capture generation the drain restart is skipped and the calls-kept
+ * launch ships window-only (~530 words, well under the wall).
+ * sema_gl_mme_prelude_live_skip=0 restores the unconditional re-emission
+ * for a one-flip A/B.
+ */
+static u32 gsp_phys_gr_mme_pending_gen;	/* final part carried by the last
+					 * launch; consumed at its poll */
+static u32 gsp_phys_gr_mme_live_gen;	/* generation whose carrier executed */
+static bool gsp_phys_rmcall_gr_mme_live_skip = true;
+module_param_named(sema_gl_mme_prelude_live_skip,
+		   gsp_phys_rmcall_gr_mme_live_skip, bool, 0644);
+MODULE_PARM_DESC(sema_gl_mme_prelude_live_skip,
+	"Skip the forced per-launch prelude re-emission once the generation's final part has been carried by a launch the engine executed (default 1; 0 = always re-emit, pre-CONT-103 behavior)");
+
+/*
+ * CONT-103: elide the program-address pair runs (method space
+ * [0x2054, 0x2158]) from a calls-kept window. With the prelude-live skip the
+ * calls-kept launch starts, and the engine then traps 0x80002058 on the
+ * relocated (shadow-pool) pair operand and parks - the CONT-101 trap class.
+ * The binds persist on the channel from the earlier calls-held launches, so
+ * dropping the re-binds removes the trapping write; the calls and the MME
+ * program stay. Off by default: the A/B against the trapping state.
+ */
+static bool gsp_phys_rmcall_gr_mme_nop_pair;
+module_param_named(sema_gl_mme_nop_pair, gsp_phys_rmcall_gr_mme_nop_pair,
+		   bool, 0644);
+MODULE_PARM_DESC(sema_gl_mme_nop_pair,
+	"Elide program-address pair runs (0x2054/0x2094/0x2154 families) from a calls-kept 3D window (default 0; the engine traps 0x2058 on the relocated shadow operand)");
+
+/*
  * SESSION 153-CONTINUED-79: parameter-conserving NOP for kept parking slots.
  * The r1177 bisect: keep={0,2,57} (full 3089-word program resident, split
  * emission) parks the PBDMA on the emitted block (MME-FAST events=0, grfifo
@@ -18945,7 +18986,11 @@ static u32 gsp_phys_rmcall_build_gr_launch(u32 *pb,
 					READ_ONCE(gsp_phys_gr_mme_pre_emit_h) ||
 				    (READ_ONCE(gsp_phys_gr_mme_pre_emit_off) >=
 					 pl_n &&
-				     (READ_ONCE(gsp_phys_rmcall_gr_keep_mask) ||
+				     ((READ_ONCE(gsp_phys_rmcall_gr_keep_mask) &&
+				       !(READ_ONCE(
+						 gsp_phys_rmcall_gr_mme_live_skip) &&
+					 READ_ONCE(gsp_phys_gr_mme_live_gen) ==
+					 pl_gen)) ||
 				      pl_gen != READ_ONCE(
 						gsp_phys_gr_mme_prelude_gen)))) {
 					WRITE_ONCE(gsp_phys_gr_mme_pre_emit_gen,
@@ -19446,6 +19491,10 @@ static u32 gsp_phys_rmcall_build_gr_launch(u32 *pb,
 				WRITE_ONCE(gsp_phys_gr_mme_prelude_gen, g0);
 				WRITE_ONCE(gsp_phys_gr_mme_prelude_h,
 					   req->h_object);
+				/* CONT-103: this launch carries the final
+				 * part; its poll outcome (executed or not)
+				 * decides whether the program is live. */
+				WRITE_ONCE(gsp_phys_gr_mme_pending_gen, g0);
 				pr_info("vgsp-gr: MME-PRELUDE emitted gen=%u words=%u window=%u h=0x%x push=0x%llx sdm_drop=%u cap=%u\n",
 					g0, pl_n, win3d_n, req->h_object,
 					req->push_va, sdrop, block_cap);
@@ -19813,6 +19862,54 @@ static u32 gsp_phys_rmcall_build_gr_launch(u32 *pb,
 					held, pl_gen, req->h_object,
 					win3d_n, w, req->push_va);
 			p += w;
+		} else if (READ_ONCE(gsp_phys_rmcall_gr_mme_nop_pair)) {
+			/*
+			 * CONT-103: the calls-kept launch now starts (the
+			 * prelude-live skip shrank it under the ~2050-word
+			 * wall) and the engine decodes the window, but the
+			 * FE then TRAPS on the relocated program-address
+			 * pair write (trap=0x80002058 data=<shadow VA>,
+			 * CONT-101: the relocated 0x2054-pair operand) and
+			 * parks the front-end (grfifo 00000001). The binds
+			 * persist on the channel from the earlier
+			 * calls-held launches of the same capture, so
+			 * eliding the pair re-binds removes the trapping
+			 * write while keeping the program bound. Drops any
+			 * method run overlapping the program-pair space
+			 * [0x2054, 0x2158] (0x2054/0x2058, 0x2094/0x2098,
+			 * 0x2154/0x2158 - the client window's own pairs and
+			 * the synthesized SPH binds alike; this walk runs on
+			 * the final window, prepends included).
+			 */
+			u32 r = 0, w = 0, elided = 0;
+
+			while (r < win3d_n) {
+				u32 h = win3d[r];
+				u32 op = (h >> 29) & 7u;
+				u32 cnt = (h >> 16) & 0xfffu;
+				u32 span = 1u, m0, mend;
+
+				if ((op == 1u || op == 3u || op == 5u) &&
+				    cnt && r + 1u + cnt <= win3d_n)
+					span = 1u + cnt;
+				m0 = (h & 0xfffu) << 2;
+				mend = m0 + 4u * (span - 1u);
+				if ((op == 0u || op == 1u || op == 3u ||
+				     op == 4u || op == 5u) &&
+				    m0 <= 0x2158u && mend >= 0x2054u) {
+					elided += span;
+					r += span;
+					continue;
+				}
+				memcpy(&p[w], &win3d[r], (size_t)span * 4u);
+				w += span;
+				r += span;
+			}
+			if (elided)
+				pr_info("vgsp-gr: MME-PAIR-DROP %u word(s) of program-pair runs elided (calls-kept launch; binds persist from the held-call launches) h=0x%x push=0x%llx window=%u->%u\n",
+					elided, req->h_object, req->push_va,
+					win3d_n, w);
+			p += w;
 		} else {
 			memcpy(p, win3d, (size_t)win3d_n * 4);
 			p += win3d_n;
@@ -21767,7 +21864,8 @@ static bool gsp_phys_gr_win0_retry_rearm(u32 *qmd,
  */
 static noinline void
 gsp_phys_gr_poll_break_log(u32 w0, u32 want_qmd, u32 w1, u32 since_us,
-			   const struct gsp_phys_rmcall_ce_submit *sub)
+			   const struct gsp_phys_rmcall_ce_submit *sub,
+			   u32 h_object)
 {
 	struct gsp_phys_boot_chunk *ud_c =
 		&boot_stage.rmcall_chunks[GSP_PHYS_RMCALL_BUF_USERD];
@@ -21796,11 +21894,53 @@ gsp_phys_gr_poll_break_log(u32 w0, u32 want_qmd, u32 w1, u32 since_us,
 		(u32)ktime_to_us(ktime_get()) - since_us,
 		usr_gpget, sub->gpget, sub->gpput,
 		sub->slot, re0, re1, sub->e0, sub->e1);
+	/*
+	 * CONT-103: engine-fetch witness + prelude-liveness bookkeeping.
+	 * GP_GET advancing past the published slot is the PBDMA consuming
+	 * this launch's GPFIFO entry - the fetch witness the QMD sentinel
+	 * cannot provide on the stream3d lane (no QMD deposit there). A
+	 * break with GP_GET frozen is a stale/software-authored release.
+	 */
+	pr_info("vgsp-gr: FETCH-WIT h=0x%x slot=%u gpget %u->%u %s\n",
+		h_object, sub->slot, sub->gpget, usr_gpget,
+		usr_gpget != sub->gpget ? "engine-fetched" :
+					  "GP_GET-frozen (software ack)");
+	if (usr_gpget != sub->gpget &&
+	    READ_ONCE(gsp_phys_gr_mme_pending_gen)) {
+		WRITE_ONCE(gsp_phys_gr_mme_live_gen,
+			   READ_ONCE(gsp_phys_gr_mme_pending_gen));
+		WRITE_ONCE(gsp_phys_gr_mme_pending_gen, 0);
+	}
 	/* r425: the healthy-launch SKED/FECS baseline for the wedge compare. */
 	gsp_phys_gr_sked_wit(gsp_phys_gpu_bar0(), "post-poll");
 }
 
 /*
+ * CONT-103: fetch-witness for a launch that did NOT break - GP_GET frozen
+ * means the PBDMA never consumed the entry (the >2050-word fetch wall /
+ * parked front-end class), so the carrier's generation is not live and the
+ * next launch must re-emit the prelude. noinline: keeps the submit frame.
+ */
+static noinline void
+gsp_phys_gr_fetch_miss_log(u32 h_object,
+			   const struct gsp_phys_rmcall_ce_submit *sub,
+			   const char *tag)
+{
+	struct gsp_phys_boot_chunk *ud_c =
+		&boot_stage.rmcall_chunks[GSP_PHYS_RMCALL_BUF_USERD];
+	u8 *ud = ud_c->pages ? gsp_phys_chunk_page_va(ud_c, 0) : NULL;
+	u32 usr_gpget = ud ?
+		get_unaligned_le32(ud + gsp_phys_rmcall_userd_ctl_off() +
+				   GSP_PHYS_RMCALL_USERD_GPGET_OFF) : 0;
+
+	pr_info("vgsp-gr: FETCH-WIT %s h=0x%x slot=%u gpget %u->%u %s\n",
+		tag, h_object, sub->slot, sub->gpget, usr_gpget,
+		usr_gpget != sub->gpget ? "engine-fetched" :
+					  "GP_GET-frozen (NOT fetched: parked or over-cap block)");
+	WRITE_ONCE(gsp_phys_gr_mme_pending_gen, 0);
+}
+
+/*
  * SESSION 153-CONTINUED-72: GR-STATE retire readback. The keep-mask A/B
  * must answer "does the kept macro execute at all" before chasing the
  * 0/903 stores. The FE decode-state EX records (0x40a630..3c) name the
@@ -26746,9 +26886,9 @@ gr_page_unmapped:
 			break;
 		}
 	}
-	pr_info("vgsp-gr: GR LAUNCH submitted h=0x%x push=0x%llx len=%u pages=%u shadow=%u qmd_va=0x%llx slot=%u words=%u e0=0x%08x e1=0x%08x prelude=%d qmdvirt=%d token=0x%x; polling qmd_rel=%u ch_rel=0x%08x (<=%u ms)%s\n",
+	pr_info("vgsp-gr: GR LAUNCH submitted h=0x%x push=0x%llx len=%u pages=%u shadow=%u qmd_va=0x%llx slot=%u gpget=%u gpput=%u words=%u e0=0x%08x e1=0x%08x prelude=%d qmdvirt=%d token=0x%x; polling qmd_rel=%u ch_rel=0x%08x (<=%u ms)%s\n",
 		req->h_object, req->push_va, req->push_len, req->n_pages,
-		n_shadow, req->qmd_va, sub.slot, words,
+		n_shadow, req->qmd_va, sub.slot, sub.gpget, sub.gpput, words,
 		sub.e0, sub.e1,
 		gsp_phys_rmcall_gr_prelude_value(),
 		gsp_phys_rmcall_gr_qmdvirt_value(),
@@ -27118,7 +27258,8 @@ gr_page_unmapped:
 		if ((!qmd_rel || w0 == want_qmd) && w1 == want_ch) {
 			/* F4-a: first poll break + USERD witnesses. */
 			gsp_phys_gr_poll_break_log(w0, want_qmd, w1,
-						   poll_start_us, &sub);
+						   poll_start_us, &sub,
+						   req->h_object);
 			/*
 			 * SESSION 153-CONTINUED-72: GR-STATE readback at the
 			 * retire (fed MME-generated-write records + latches).
@@ -27142,6 +27283,8 @@ gr_page_unmapped:
 			       w0, want_qmd, w1, want_ch, req->h_object,
 			       req->push_va);
 			gsp_phys_mme_regs_log(bar0, "wait-max");
+			gsp_phys_gr_fetch_miss_log(req->h_object, &sub,
+						   "wait-max");
 			ret = -ETIMEDOUT;
 			goto out;
 		}
@@ -27154,6 +27297,8 @@ gr_page_unmapped:
 			       poll_ms, w0, want_qmd, w1,
 			       want_ch, req->h_object, req->push_va);
 			gsp_phys_mme_regs_log(bar0, "timeout");
+			gsp_phys_gr_fetch_miss_log(req->h_object, &sub,
+						   "timeout");
 			/*
 			 * F4-b: one-shot deposit readback at the FIRST timeout
 			 * (the old post-mortem only ran after attempt 3, which
@@ -27702,9 +27847,16 @@ gr_page_unmapped:
 		if (wrote == 0 && n_tracked) {
 			bool ff = gsp_phys_gr_qmd_fetched(qmd_dst_mem);
 
-			pr_info("vgsp-gr: first-attempt witness: engine fetch %s (qmd sentinel %s) tracked=%u\n",
-				ff ? "OK" : "NONE",
-				ff ? "overwritten" : "intact", n_tracked);
+			/* CONT-103: no QMD deposit on the stream3d lane, so the
+			 * sentinel is vacuous there; the GP_GET FETCH-WIT is the
+			 * fetch witness for that lane. */
+			if (qmd_dst_mem)
+				pr_info("vgsp-gr: first-attempt witness: engine fetch %s (qmd sentinel %s) tracked=%u\n",
+					ff ? "OK" : "NONE",
+					ff ? "overwritten" : "intact", n_tracked);
+			else
+				pr_info("vgsp-gr: first-attempt witness: engine fetch n/a (no QMD deposit on this lane - sentinel vacuous; see FETCH-WIT) tracked=%u\n",
+					n_tracked);
 		}
 		/*
 		 * r427 idempotent fast-accept: a zero-write launch whose identity
@@ -28141,7 +28293,8 @@ gr_page_unmapped:
 				fetched ? "AMBIGUOUS" : "CONFIRMED",
 				req->h_object, req->push_va, req->n_data,
 				n_tracked, nw_attempts, want_qmd,
-				fetched ? "OK" : "NONE",
+				qmd_dst_mem ? (fetched ? "OK" : "NONE") :
+					      "n/a(noQMD-stream3d)",
 				READ_ONCE(gsp_phys_gr_nw_streak) + 1u,
 				gsp_phys_gr_wedge_quiet,
 				gsp_phys_gr_wedge_heals);