Re: ofppc - panic pmap_pvo_find_va returning NULL for kernel pmap

[Radoslaw Kujawa <radoslaw.kujawa%c0ff33.net@localhost>]

Hi list.

On 6/30/26 00:40, Radosław Kujawa wrote:

> The check should not assume that given VA >= PMAP_DIRECT_MAPPED_LEN . Note that PMAP_DIRECT_MAPPED_LEN is calculated according to the formula in oea/pmap.h (line 128) :
>
> SEGMENT_LENGTH (256MB) * USER_SR
>
> Why does the panic not trigger on macppc and other PowerPC ports? By coincidence that depends on ordering of USER_SR and KERNEL_SR utilized by a given port:
>
> macppc, evbppc etc. use USER_SR 12 and KERNEL_SR 13, 14, so segment-mappend KVA lands between 0xD0000000 - 0xF0000000, and PMAP_DIRECT_MAPPED_LEN as calculated is 0xC0000000 .
> ofppc uses USER_SR 14 and KERNEL_SR 10, 11, so segment-mapped KVA lands between 0xA0000000 - 0xC0000000, and PMAP_DIRECT_MAPPED_LEN as calculated is 0xE0000000 .

After discussion with other developers (thanks Martin, Nick), it was 
concluded that cleanest solution is to fix the PMAP_DIRECT_MAPPED_LEN 
calculation itself, so that PMAP_DIRECT_MAPPED_LEN reflects true range 
of identity-mapped addresses and is independent of KERNEL_SR / USER_SR 
ordering on a given port.

The following fix was committed today.

Index: src/sys/arch/powerpc/include/oea/pmap.h
diff -u src/sys/arch/powerpc/include/oea/pmap.h:1.39 
src/sys/arch/powerpc/include/oea/pmap.h:1.40
--- src/sys/arch/powerpc/include/oea/pmap.h:1.39	Fri Dec 15 09:42:33 2023
+++ src/sys/arch/powerpc/include/oea/pmap.h	Tue Jun 30 17:21:18 2026
@@ -1,4 +1,4 @@
-/*	$NetBSD: pmap.h,v 1.39 2023/12/15 09:42:33 rin Exp $	*/
+/*	$NetBSD: pmap.h,v 1.40 2026/06/30 17:21:18 rkujawa Exp $	*/

 /*-
  * Copyright (C) 1995, 1996 Wolfgang Solfrank.
@@ -122,11 +122,13 @@ __BEGIN_DECLS
 #include <sys/systm.h>

 /*
- * For OEA and OEA64_BRIDGE, we guarantee that pa below USER_ADDR
- * (== 3GB < VM_MIN_KERNEL_ADDRESS) is direct-mapped.
+ * Physical memory below PMAP_DIRECT_MAPPED_LEN is direct-mapped
+ * (pa == va). Direct region covers the segments below BOTH
+ * the user copyin window (USER_SR) and the kernel HTAB window
+ * (KERNEL_SR), so it can never overlap.
  */
 #if defined(PPC_OEA) || defined(PPC_OEA64_BRIDGE)
-#define	PMAP_DIRECT_MAPPED_SR	(USER_SR - 1)
+#define	PMAP_DIRECT_MAPPED_SR	(MIN(USER_SR, KERNEL_SR) - 1)
 #define	PMAP_DIRECT_MAPPED_LEN \
     ((vaddr_t)SEGMENT_LENGTH * (PMAP_DIRECT_MAPPED_SR + 1))
 #endif


Best regards,
Radoslaw