vgsp: CONT-97 - gsp_phys_ds_spy re-clears DS_REPORT_MASK/_2 every sample + maskholds/last_mask logging (W1C confirmed: pgraph-rw 0x405840<-0x20 reads back 00000000, no synthetic calibration possible); slots NAMED: c597 SetPipeline[j] only j=0..5, window binds j=1/2/5 ProgramAddressA=0x19a0900/0x19a0200/0x19a0000 byte-equal SPH_INFO stage VAs; DS_HWW_ESR bits 0..23 = SPH0..23 checks (RM DS_SPH<n>_ERR=0x6e+n) so j=9/10/11 binds cannot exist - {5,9,10,11} are per-SPH validation checks; pgraph-rw A/B: golden PATCH DS writes are TGA constraint logic ALREADY LIVE (0x700/0x800) not SPH-related; DS-spy captures: hits all badf1201 poison or zero (timing anomaly recorded); GATE1 black 3 boots ds=a4000e20 sph=0,019a0900,019a0200 0/900 pages; guest root capture 99.92% black; next: beef window's only call form is op=4 IMMD_DATA_METHOD on CallMme[0].Macro 0x3800 (MME-WINRUN off=a0 op=4 cnt=48 mthd=3800) never executed by MME (pcv=0; held windows retire, call-bearing park) while prelude resident - decode run vs MME call spec offline (mme_replay_slots81.txt) + re-emit client's own call shape (trigger 0x3800 + params 0x3804); if op=4 correct run pure-method window test keep_mask=0+hold=N to decide if replay needs MME at all - SESSION 153 r1326-r1327

1 files changed, 27 insertions, 2 deletions

diff --git a/kernel/gsp_phys_booter.c b/kernel/gsp_phys_booter.c
index 7c4a946..3d0aae2 100644
--- a/kernel/gsp_phys_booter.c
+++ b/kernel/gsp_phys_booter.c
@@ -31076,7 +31076,7 @@ static int gsp_phys_ds_spy_show(struct seq_file *m, void *unused)
 {
 	void __iomem *bar0 = gsp_phys_gpu_bar0();
 	u64 t0;
-	u32 hits = 0, samples = 0;
+	u32 hits = 0, samples = 0, maskholds = 0, mask0_last = 0, mask2_last = 0;
 	u32 ms = READ_ONCE(gsp_phys_ds_spy_ms);
 
 	if (!bar0)
@@ -31093,7 +31093,31 @@ static int gsp_phys_ds_spy_show(struct seq_file *m, void *unused)
 	t0 = ktime_get_ns();
 	while (ktime_get_ns() - t0 < (u64)ms * 1000000ull) {
 		u32 esr = readl(bar0 + GSP_PHYS_DS_ESR);
+		u32 m0 = readl(bar0 + GSP_PHYS_DS_REPORT_MASK);
+		u32 m2 = readl(bar0 + GSP_PHYS_DS_REPORT_MASK_2);
 
+		/*
+		 * SESSION 153-CONTINUED-97: hold the HWW report masks
+		 * cleared while sampling. A pending report POISONS the
+		 * whole PRI block (every read returns 0xbadf1201), which
+		 * hid the SPH stage snapshot at the validation instant
+		 * (r1325 spy2: all 16+ hits badf1201, masks back at
+		 * 00ffffff). The mask is (re)armed by the context load /
+		 * the RM setup, not only by our replay path, so re-clear
+		 * it every sample; the re-clear count is logged so the
+		 * capture's provenance is visible. NOTE (r1326): poison
+		 * still appeared on one pass despite 29 re-clears, so a
+		 * clean latch is not guaranteed and the store-side read
+		 * remains the authoritative fault-time datum.
+		 * Development-only diagnostic tap.
+		 */
+		if (m0 || m2) {
+			iowrite32(0u, bar0 + GSP_PHYS_DS_REPORT_MASK);
+			iowrite32(0u, bar0 + GSP_PHYS_DS_REPORT_MASK_2);
+			maskholds++;
+			mask0_last = m0;
+			mask2_last = m2;
+		}
 		samples++;
 		if (esr & 0x7fffffffu) {
 			u32 esr2 = readl(bar0 + GSP_PHYS_DS_ESR_2);
@@ -31122,7 +31146,8 @@ static int gsp_phys_ds_spy_show(struct seq_file *m, void *unused)
 			udelay(20);
 		}
 	}
-	seq_printf(m, "# DS-SPY end samples=%u hits=%u\n", samples, hits);
+	seq_printf(m, "# DS-SPY end samples=%u hits=%u maskholds=%u last_mask=%08x/%08x\n",
+		   samples, hits, maskholds, mask0_last, mask2_last);
 	return 0;
 }