vgsp: relfix PART 3 follow-up - param-page remainder scan runs as a non-arg pass (is_arg/hold semantics unchanged)

1 files changed, 39 insertions, 21 deletions

diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c
index 3887dbb..88c519d 100644
--- a/kernel/gsp_sema_service.c
+++ b/kernel/gsp_sema_service.c
@@ -14635,18 +14635,45 @@ static void svc_gr_data_fill(struct nv_vgpu_device *vgpu,
 			if (w == n_cand && n_cand < ARRAY_SIZE(wins))
 				wins[n_cand++] = wv;
 		}
-		/*
-		 * relfix PART 3 (fill candidate superset): the rewrite scans
-		 * the WHOLE authored param page (0x0..0xfff) and re-resolves
-		 * every device-VA-shaped word it finds; any word this scan
-		 * does not collect resurfaces as a RELOC-GAP (r217: the two
-		 * words the late path fixed were param/program descriptor
-		 * fields below 0x150 / past 0x800 the argument scan never
-		 * looked at). Collect the page remainder with the same shape
-		 * test, after the argument region so the operand pages keep
-		 * budget priority (m120/m121). Launch-internal pages remain
-		 * filtered at resolution time below (R3 hygiene).
-		 */
+		/* m144: also scan the fresh inline-captured parameter block,
+		 * so a reused/stale guest page cannot hide the kernel's device
+		 * pointers. */
+		svc_gr_scan_inline(wins, &n_cand, pp->va);
+	}
+	/* m150: everything collected so far came from the c[0x0] argument
+	 * page (or its inline bytes) - these are the pointers the kernel
+	 * actually dereferences. */
+	for (w = 0; w < n_cand; w++)
+		is_arg[w] = true;
+	/*
+	 * relfix PART 3 (fill candidate superset): the rewrite scans the WHOLE
+	 * authored param page (0x0..0xfff) and re-resolves every
+	 * device-VA-shaped word it finds; any word this fill does not collect
+	 * resurfaces as a RELOC-GAP (r217: the two words the late path fixed
+	 * were param/program descriptor fields below 0x150 / past 0x800 that
+	 * the argument scan never looked at). Collect the page remainder here,
+	 * AFTER the argument marking, so these page words keep is_arg=false -
+	 * the per-argument hold and diagnostic semantics stay exactly as
+	 * before. Budget priority is also unchanged: the argument candidates
+	 * above are resolved first. Launch-internal pages remain filtered at
+	 * resolution time (R3 hygiene below).
+	 */
+	for (i = 0; i < svc_fwd_req.gr_n_pages; i++) {
+		const struct svc_gr_page *pp = &svc_fwd_req.gr_pages[i];
+		u64 hpa = 0;
+		const u8 *page;
+		u32 k;
+
+		if (pp->qmd_word != 32u || !pp->pa)
+			continue;
+		kind = (pp->ap == SVC_APERTURE_FB) ? SVC_APERTURE_FB :
+						     SVC_APERTURE_SYS;
+		if (svc_fwd_operand_hpa(vgpu, kind, pp->pa, 0,
+					SEMA_PAGE_SIZE, &hpa, &off) || off)
+			continue;
+		page = (const u8 *)__va(hpa);
+		if (!page)
+			continue;
 		for (k = 0; k < SEMA_PAGE_SIZE / 8; k++) {
 			u64 v, wv;
 
@@ -14663,16 +14690,7 @@ static void svc_gr_data_fill(struct nv_vgpu_device *vgpu,
 			if (w == n_cand && n_cand < ARRAY_SIZE(wins))
 				wins[n_cand++] = wv;
 		}
-		/* m144: also scan the fresh inline-captured parameter block,
-		 * so a reused/stale guest page cannot hide the kernel's device
-		 * pointers. */
-		svc_gr_scan_inline(wins, &n_cand, pp->va);
 	}
-	/* m150: everything collected so far came from the c[0x0] argument
-	 * page (or its inline bytes) - these are the pointers the kernel
-	 * actually dereferences. */
-	for (w = 0; w < n_cand; w++)
-		is_arg[w] = true;
 	for (i = 0; i < svc_fwd_req.gr_n_pages; i++) {
 		const struct svc_gr_page *pp = &svc_fwd_req.gr_pages[i];
 		u64 hpa = 0;