1 files changed, 148 insertions, 8 deletions

diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c
index 419a54a..2d41bb0 100644
--- a/kernel/gsp_sema_service.c
+++ b/kernel/gsp_sema_service.c
@@ -4717,6 +4717,65 @@ static bool svc_fill_target_is_pt(struct nv_vgpu_device *vgpu, u64 addr)
 	return hits >= SVC_PT_MIN;
 }
 
+/*
+ * r391 per-page destination resolution for a VIRTUAL fill (the r390
+ * guest-panic class). A device VA's pages are independent guest allocations
+ * - the pin stream shows split runs (va=0x7c36c7400000 -> 29 pages at
+ * 0x131b63000, the next run at 0x131a80000) - so resolving only the fill's
+ * first page and writing `total` bytes contiguously past it writes over
+ * whatever memory follows the first run. On r390 a 383620-byte zero fill
+ * did exactly that: the contiguous write covered the guest kernel's
+ * page-table pages (0x131ba1000/0x131ba2000/0x131bab000), the direct map
+ * grew holes, and the guest died with "Fatal exception in interrupt"
+ * (unmap_page_range / cpuacct / psi faults). svc_copy has always translated
+ * virtual operands per 4 KB page; the fill must do the same.
+ *
+ * Authority order mirrors the top-level resolution (and svc_copy's per-page
+ * dst): a fresh pin (guest pagemap truth) wins and repairs the landing
+ * table, then the landing table (the page a previous write of this VA
+ * landed on), then the resolver chain. A page with no resolution returns
+ * -ENOENT: the caller fails the fill loudly instead of writing through a
+ * contiguous guess.
+ */
+static int svc_fill_dst_page_pa(struct nv_vgpu_device *vgpu, u64 va,
+				u64 *pa_out, u32 *ap_out, u32 *prov_out)
+{
+	u64 page = va & ~(u64)(SEMA_PAGE_SIZE - 1);
+	u64 pa = 0, lpa = 0;
+	u32 aper = 0, lap = 0;
+
+	if (!svc_pin_lookup_any(page, &pa, &aper)) {
+		*prov_out = 1;
+		if (!svc_ledger_lookup(page, &lpa, &lap) &&
+		    (((lpa & ~(u64)(SEMA_PAGE_SIZE - 1)) != page) ||
+		     (u32)lap != (u32)aper)) {
+			svc_ledger_pin_repairs++;
+			if (svc_ledger_pin_repairs <= 32)
+				pr_info("sema-ledger: PIN-WINS fill-page va=0x%llx ledger=0x%llx/%u pin=0x%llx/%u\n",
+					page,
+					lpa & ~(u64)(SEMA_PAGE_SIZE - 1),
+					lap, pa, aper);
+			svc_ledger_store(page, pa, aper);
+		}
+		*pa_out = pa;
+		*ap_out = aper;
+		return 0;
+	}
+	if (!svc_ledger_lookup(page, &pa, &aper)) {
+		*prov_out = 2;
+		*pa_out = pa;
+		*ap_out = aper;
+		return 0;
+	}
+	if (svc_op_translate(vgpu, page, &pa, &aper)) {
+		*prov_out = 3;
+		*pa_out = pa;
+		*ap_out = aper;
+		return 0;
+	}
+	return -ENOENT;
+}
+
 static int svc_fill(struct nv_vgpu_device *vgpu, u64 addr, u64 value, u32 elem,
 		    u32 count, u32 phys, u32 aperture, u64 *done_out)
 {
@@ -4909,20 +4968,29 @@ static int svc_fill(struct nv_vgpu_device *vgpu, u64 addr, u64 value, u32 elem,
 	/*
 	 * C2 (data class) + m180 unification: a large virtual fill is a
 	 * write/read pair with the memset's readers like any H2D copy -
-	 * record its landing pages (SYS and VID alike) so the engine fill
-	 * and the D2H read of this VA use exactly the page the fill wrote.
+	 * record its landing pages so the engine fill and the D2H read of
+	 * this VA use exactly the page the fill wrote. Only the VID
+	 * (shadow) case is contiguous; a SYS landing is recorded per page
+	 * in the loop below (r391: the contiguous `addr + po` store also
+	 * poisoned the table with pages the fill never wrote).
 	 */
 	if (!phys && !shadow && va0 >= SVC_VAS_FLOOR &&
-	    total >= (64u << 10)) {
-		u32 cap = (aperture == CE_APERTURE_VID) ?
-			SVC_APERTURE_FB : SVC_APERTURE_SYS;
+	    total >= (64u << 10) && aperture == CE_APERTURE_VID) {
 		u64 po;
 
 		for (po = 0; po < total; po += SEMA_PAGE_SIZE)
-			svc_ledger_store(va0 + po, addr + po, cap);
+			svc_ledger_store(va0 + po, addr + po,
+					 SVC_APERTURE_FB);
 	}
 
-	if (aperture == CE_APERTURE_VID) {
+	/*
+	 * r391: a virtual device destination is translated per 4 KB page
+	 * (see svc_fill_dst_page_pa) - never written through the first
+	 * page's PA contiguously. The whole-range shadow write below only
+	 * applies to a resolution that is contiguous by construction.
+	 */
+	if (aperture == CE_APERTURE_VID &&
+	    !(va0 >= SVC_VAS_FLOOR && !phys && !shadow)) {
 		u8 *base = vgpu->bars[1].vaddr;
 		size_t size = svc_fb_shadow_size(vgpu);
 
@@ -4942,14 +5010,77 @@ static int svc_fill(struct nv_vgpu_device *vgpu, u64 addr, u64 value, u32 elem,
 	for (o = 0; o < total; ) {
 		size_t chunk = total - o;
 		size_t i;
+		u64 wpa = addr;
+		u32 wap = aperture;
+		bool per_page = !phys && !shadow && va0 >= SVC_VAS_FLOOR;
 
 		if (chunk > sizeof(bounce))
 			chunk = sizeof(bounce);
+		/*
+		 * r391: resolve THIS page's own landing (see
+		 * svc_fill_dst_page_pa) and clamp the chunk to the page
+		 * boundary - a virtual VA's pages are independent guest
+		 * allocations and must never be written through the first
+		 * page's PA contiguously.
+		 */
+		if (per_page) {
+			u32 off = (u32)((va0 + o) & (SEMA_PAGE_SIZE - 1));
+			u64 ppa = 0;
+			u32 pap = 0, prov = 0;
+			int prc;
+
+			chunk = min_t(size_t, chunk, SEMA_PAGE_SIZE - off);
+			prc = svc_fill_dst_page_pa(vgpu, va0 + o, &ppa,
+						   &pap, &prov);
+			if (prc) {
+				if (done_out)
+					*done_out = (u64)o;
+				pr_err_ratelimited("vgsp: FILL-TRUNC va=0x%llx want=0x%llx done=0x%llx path=sema-fill (page 0x%llx unresolved)\n",
+						   va0, (u64)total, (u64)o,
+						   (va0 + o) &
+						   ~(u64)(SEMA_PAGE_SIZE - 1));
+				return -ENOENT;
+			}
+			wap = (pap == SVC_APERTURE_FB) ?
+				CE_APERTURE_VID : CE_APERTURE_SYS;
+			wpa = (ppa & ~(u64)(SEMA_PAGE_SIZE - 1)) + off;
+			/*
+			 * r391 net: the r384 page-table-shape refusal
+			 * applied to EVERY page of a virtual zero fill.
+			 * The r390 fill's first page was the client's
+			 * buffer, the later pages were the guest's PTE
+			 * pages - the first-page-only check could not see
+			 * them. A page-table-shaped guest-RAM page is never
+			 * a legitimate zero-fill target: skip it loudly
+			 * instead of zeroing it. Pin-backed resolutions
+			 * are exempt (the guest's pagemap is truth there).
+			 */
+			if (wap == CE_APERTURE_SYS && value == 0 && prov != 1 &&
+			    svc_fill_target_is_pt(vgpu,
+						  wpa & ~(u64)(SEMA_PAGE_SIZE - 1))) {
+				svc_fill_pt_refused++;
+				if (svc_fill_pt_refused <= 32)
+					pr_err("vgsp-sema: FILL-PT-REFUSED va=0x%llx pa=0x%llx prov=%u total=%zu (zero fill page is a page table; page skipped)\n",
+					       va0 + o, wpa, prov, total);
+				o += chunk;
+				continue;
+			}
+		}
 		for (i = 0; i + 8 <= chunk; i += 8)
 			memcpy(bounce + i, pat, 8);
 		if (i < chunk)
 			memcpy(bounce + i, pat, chunk - i);
-		if (svc_write(vgpu, addr + o, bounce, chunk)) {
+		if (wap == CE_APERTURE_VID) {
+			u8 *base = vgpu->bars[1].vaddr;
+			size_t size = svc_fb_shadow_size(vgpu);
+
+			if (!base || wpa + chunk > size) {
+				if (done_out)
+					*done_out = (u64)o;
+				return -EINVAL;
+			}
+			memcpy(base + wpa, bounce, chunk);
+		} else if (svc_write(vgpu, wpa, bounce, chunk)) {
 			if (done_out)
 				*done_out = (u64)o;
 			if (total >= SVC_COPY_AUDIT_MIN)
@@ -4957,6 +5088,15 @@ static int svc_fill(struct nv_vgpu_device *vgpu, u64 addr, u64 value, u32 elem,
 						   va0, (u64)total, (u64)o, -EIO);
 			return -EIO;
 		}
+		/*
+		 * r391: record this page's landing (the readers follow the
+		 * table) - SYS and VID alike, exactly like the copy path.
+		 */
+		if (per_page)
+			svc_ledger_store(va0 + o,
+					 wpa & ~(u64)(SEMA_PAGE_SIZE - 1),
+					 (wap == CE_APERTURE_VID) ?
+					 SVC_APERTURE_FB : SVC_APERTURE_SYS);
 		o += chunk;
 	}
 	if (done_out)