diff mbox series

[v5,09/27] x86/boot: Remove mapping from page fault handler

Message ID dab948690a74db1bb75d95aa7e0362deeca6dbf4.1678785672.git.baskov@ispras.ru
State New
Headers
Series x86_64: Improvements at compressed kernel stage |

Commit Message

Evgeniy Baskov March 14, 2023, 10:13 a.m. UTC 
  After every implicit mapping is removed, this code is no longer needed.

Remove memory mapping from page fault handler to ensure that there are
no hidden invalid memory accesses.

Tested-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Evgeniy Baskov <baskov@ispras.ru>
---
 arch/x86/boot/compressed/ident_map_64.c | 26 ++++++++++---------------
 1 file changed, 10 insertions(+), 16 deletions(-)

Comments

Andy Lutomirski March 14, 2023, 8:33 p.m. UTC | #1 
On Tue, Mar 14, 2023, at 3:13 AM, Evgeniy Baskov wrote:
> After every implicit mapping is removed, this code is no longer needed.
>
> Remove memory mapping from page fault handler to ensure that there are
> no hidden invalid memory accesses.

This patch is *by far* the scariest of the bunch in my boot.  And it violates a basic principle of kernel development: it's better to run in degraded mode than to fail outright unless running in degraded mode is dangerous for some reason.

And this boot code is not actually meaningfully exposed to attack.  Anyone who can get the boot code to consume garbage likely *already* controls the system, including anything that we might write to TPM or any other verification mechanism.

So I think this should log an error, set a flag to make sure we print an even louder error after full boot, but still add the mapping and keep trying.

--Andy

>
> Tested-by: Mario Limonciello <mario.limonciello@amd.com>
> Signed-off-by: Evgeniy Baskov <baskov@ispras.ru>
> ---
>  arch/x86/boot/compressed/ident_map_64.c | 26 ++++++++++---------------
>  1 file changed, 10 insertions(+), 16 deletions(-)
>
> diff --git a/arch/x86/boot/compressed/ident_map_64.c 
> b/arch/x86/boot/compressed/ident_map_64.c
> index eb28ce9812c5..378f99b1d7e8 100644
> --- a/arch/x86/boot/compressed/ident_map_64.c
> +++ b/arch/x86/boot/compressed/ident_map_64.c
> @@ -393,27 +393,21 @@ void do_boot_page_fault(struct pt_regs *regs, 
> unsigned long error_code)
>  {
>  	unsigned long address = native_read_cr2();
>  	unsigned long end;
> -	bool ghcb_fault;
> +	char *msg;
> 
> -	ghcb_fault = sev_es_check_ghcb_fault(address);
> +	if (sev_es_check_ghcb_fault(address))
> +		msg = "Page-fault on GHCB page:";
> +	else
> +		msg = "Unexpected page-fault:";
> 
>  	address   &= PMD_MASK;
>  	end        = address + PMD_SIZE;
> 
>  	/*
> -	 * Check for unexpected error codes. Unexpected are:
> -	 *	- Faults on present pages
> -	 *	- User faults
> -	 *	- Reserved bits set
> -	 */
> -	if (error_code & (X86_PF_PROT | X86_PF_USER | X86_PF_RSVD))
> -		do_pf_error("Unexpected page-fault:", error_code, address, regs->ip);
> -	else if (ghcb_fault)
> -		do_pf_error("Page-fault on GHCB page:", error_code, address, regs->ip);
> -
> -	/*
> -	 * Error code is sane - now identity map the 2M region around
> -	 * the faulting address.
> +	 * Since all memory allocations are made explicit
> +	 * now, every page fault at this stage is an
> +	 * error and the error handler is there only
> +	 * for debug purposes.
>  	 */
> -	kernel_add_identity_map(address, end, MAP_WRITE);
> +	do_pf_error(msg, error_code, address, regs->ip);
>  }
> -- 
> 2.39.2
Evgeniy Baskov March 15, 2023, 1:25 p.m. UTC | #2 
On 2023-03-14 23:33, Andy Lutomirski wrote:
> On Tue, Mar 14, 2023, at 3:13 AM, Evgeniy Baskov wrote:
>> After every implicit mapping is removed, this code is no longer 
>> needed.
>> 
>> Remove memory mapping from page fault handler to ensure that there are
>> no hidden invalid memory accesses.
> 
> This patch is *by far* the scariest of the bunch in my boot.  And it
> violates a basic principle of kernel development: it's better to run
> in degraded mode than to fail outright unless running in degraded mode
> is dangerous for some reason.
> 
> And this boot code is not actually meaningfully exposed to attack.
> Anyone who can get the boot code to consume garbage likely *already*
> controls the system, including anything that we might write to TPM or
> any other verification mechanism.
> 
> So I think this should log an error, set a flag to make sure we print
> an even louder error after full boot, but still add the mapping and
> keep trying.

Good point. This patch can be dropped and replaced by the loud warning,
since it is not required for the functioning of the rest of the series
it is here mainly to indicate bugs in the kernel rather than for the
increased protection. But I would not expect anything in the working
systems, I made my best to map all the things explicitly. And since
no code but the extraction code is supposed to be run (interrupts
are disabled and we are not using any UEFI services there), this should
be practically save to remove the implicit mapping.

And at least this allowed me to find out about the insufficient size of
the boot page tables which did not account for the ACPI and UEFI
mappings. (patch 4 "x86/boot: Increase boot page table size")

If this patch is dropped now, I can send the follow up patch later
adding the warning.

Thanks,
Evgeniy Baskov

> 
> --Andy
> 
>> 
>> Tested-by: Mario Limonciello <mario.limonciello@amd.com>
>> Signed-off-by: Evgeniy Baskov <baskov@ispras.ru>
>> ---
>>  arch/x86/boot/compressed/ident_map_64.c | 26 
>> ++++++++++---------------
>>  1 file changed, 10 insertions(+), 16 deletions(-)
>> 
...
diff mbox series

Patch

diff --git a/arch/x86/boot/compressed/ident_map_64.c b/arch/x86/boot/compressed/ident_map_64.c
index eb28ce9812c5..378f99b1d7e8 100644
--- a/arch/x86/boot/compressed/ident_map_64.c
+++ b/arch/x86/boot/compressed/ident_map_64.c
@@ -393,27 +393,21 @@  void do_boot_page_fault(struct pt_regs *regs, unsigned long error_code)
 {
 	unsigned long address = native_read_cr2();
 	unsigned long end;
-	bool ghcb_fault;
+	char *msg;
 
-	ghcb_fault = sev_es_check_ghcb_fault(address);
+	if (sev_es_check_ghcb_fault(address))
+		msg = "Page-fault on GHCB page:";
+	else
+		msg = "Unexpected page-fault:";
 
 	address   &= PMD_MASK;
 	end        = address + PMD_SIZE;
 
 	/*
-	 * Check for unexpected error codes. Unexpected are:
-	 *	- Faults on present pages
-	 *	- User faults
-	 *	- Reserved bits set
-	 */
-	if (error_code & (X86_PF_PROT | X86_PF_USER | X86_PF_RSVD))
-		do_pf_error("Unexpected page-fault:", error_code, address, regs->ip);
-	else if (ghcb_fault)
-		do_pf_error("Page-fault on GHCB page:", error_code, address, regs->ip);
-
-	/*
-	 * Error code is sane - now identity map the 2M region around
-	 * the faulting address.
+	 * Since all memory allocations are made explicit
+	 * now, every page fault at this stage is an
+	 * error and the error handler is there only
+	 * for debug purposes.
 	 */
-	kernel_add_identity_map(address, end, MAP_WRITE);
+	do_pf_error(msg, error_code, address, regs->ip);
 }