Re: [PATCH v4 08/26] x86/boot: Map memory explicitly

From: Evgeniy Baskov
Date: Wed Mar 08 2023 - 11:10:36 EST


On 2023-03-08 13:28, Ard Biesheuvel wrote:
On Wed, 8 Mar 2023 at 10:38, Ard Biesheuvel <ardb@xxxxxxxxxx> wrote:

On Thu, 15 Dec 2022 at 13:38, Evgeniy Baskov <baskov@xxxxxxxxx> wrote:
>
> Implicit mappings hide possible memory errors, e.g. allocations for
> ACPI tables were not included in boot page table size.
>
> Replace all implicit mappings from page fault handler with
> explicit mappings.
>

I agree with the motivation but this patch seems to break the boot
under SeaBIOS/QEMU, and I imagine other legacy BIOS boot scenarios as
well.

Naively, I would assume that there is simply a legacy BIOS region that
we fail to map here, but I am fairly clueless when it comes to non-EFI
x86 boot so take this with a grain of salt.


The below seems to help - not sure why exactly, but apparently legacy
BIOS needs the bootparams struct to be mapped writable?

I think I got too eager adding mappings to everything.
In the process_efi_entries() bootparams should already be mapped, so
I will just remove the call. And AFAIK bootparams is indeed gets
written to.


--- a/arch/x86/boot/compressed/kaslr.c
+++ b/arch/x86/boot/compressed/kaslr.c
@@ -31,6 +31,7 @@
#include <linux/ctype.h>
#include <generated/utsversion.h>
#include <generated/utsrelease.h>
+#include <asm/shared/pgtable.h>

#define _SETUP
#include <asm/setup.h> /* For COMMAND_LINE_SIZE */
@@ -688,7 +689,7 @@ process_efi_entries(unsigned long minimum,
unsigned long image_size)
u32 nr_desc;
int i;

- kernel_add_identity_map((unsigned long)e, (unsigned long)(e + 1), 0);
+ kernel_add_identity_map((unsigned long)e, (unsigned long)(e +
1), MAP_WRITE);

signature = (char *)&e->efi_loader_signature;
if (strncmp(signature, EFI32_LOADER_SIGNATURE, 4) &&

Thanks,
Evgeniy Baskov