[PATCH 1/1] x86_64: fix KASan shadow region page tables

[PATCH 1/1] x86_64: fix KASan shadow region page tables

[Alexander Popov]

Physical addresses in KASan shadow region page tables need fixup:
kernel halts without it if phys_base is not zero.

Signed-off-by: Alexander Popov <alpopov@ptsecurity.com>
---
 arch/x86/kernel/head_64.S | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

diff --git a/arch/x86/kernel/head_64.S b/arch/x86/kernel/head_64.S
index ae6588b..f7711bb 100644
--- a/arch/x86/kernel/head_64.S
+++ b/arch/x86/kernel/head_64.S
@@ -95,6 +95,20 @@ startup_64:
 
 	addq	%rbp, level2_fixmap_pgt + (506*8)(%rip)
 
+#ifdef CONFIG_KASAN
+	xor	%rax, %rax
+	leaq	kasan_zero_pud(%rip), %rbx
+	leaq	kasan_zero_pmd(%rip), %rcx
+	leaq	kasan_zero_pte(%rip), %rdx
+1:
+	addq	%rbp, (%rbx,%rax,8)
+	addq	%rbp, (%rcx,%rax,8)
+	addq	%rbp, (%rdx,%rax,8)
+	inc	%rax
+	cmp	$512, %rax
+	jne	1b
+#endif
+
 	/*
 	 * Set up the identity mapping for the switchover.  These
 	 * entries should *NOT* have the global bit set!  This also
-- 
1.9.1

Re: [PATCH 1/1] x86_64: fix KASan shadow region page tables

[Andrey Ryabinin]

On 05/22/2015 07:03 PM, Alexander Popov wrote:
> Physical addresses in KASan shadow region page tables need fixup:
> kernel halts without it if phys_base is not zero.
> 

Indeed.
Since we have to patch page tables anyway, compile-time created kasan
page tables become pointless.
So, I'd suggest to create them in runtime. This could be done in plain C
in kasan_map_early_shadow().

Re: [PATCH 1/1] x86_64: fix KASan shadow region page tables

[Alexander Popov]

On 25.05.2015 11:12, Andrey Ryabinin wrote:
> On 05/22/2015 07:03 PM, Alexander Popov wrote:
>> Physical addresses in KASan shadow region page tables need fixup:
>> kernel halts without it if phys_base is not zero.
>>
> 
> Indeed.
> Since we have to patch page tables anyway, compile-time created kasan
> page tables become pointless.
> So, I'd suggest to create them in runtime. This could be done in plain C
> in kasan_map_early_shadow().
> 

Thanks for your reply, Andrey.

Creating kasan_zero_pud, kasan_zero_pmd and kasan_zero_pte in
kasan_map_early_shadow() doesn't look handy for me because this function
is called twice in x86_64_start_kernel().

What do you think about leaving the initialization of KASan shadow region
page tables in arch/x86/kernel/head_64.S and calling something like
kasan_fixup_early_shadow() before calling kasan_map_early_shadow()
for the first time?

Best regards,
Alexander

Re: [PATCH 1/1] x86_64: fix KASan shadow region page tables

[Andrey Ryabinin]

On 05/25/2015 04:43 PM, Alexander Popov wrote:
> On 25.05.2015 11:12, Andrey Ryabinin wrote:
>> On 05/22/2015 07:03 PM, Alexander Popov wrote:
>>> Physical addresses in KASan shadow region page tables need fixup:
>>> kernel halts without it if phys_base is not zero.
>>>
>>
>> Indeed.
>> Since we have to patch page tables anyway, compile-time created kasan
>> page tables become pointless.
>> So, I'd suggest to create them in runtime. This could be done in plain C
>> in kasan_map_early_shadow().
>>
> 
> Thanks for your reply, Andrey.
> 
> Creating kasan_zero_pud, kasan_zero_pmd and kasan_zero_pte in
> kasan_map_early_shadow() doesn't look handy for me because this function
> is called twice in x86_64_start_kernel().
> 
> What do you think about leaving the initialization of KASan shadow region
> page tables in arch/x86/kernel/head_64.S and calling something like
> kasan_fixup_early_shadow() before calling kasan_map_early_shadow()
> for the first time?
> 

Why do we have initialize + fixup later if we could simply initialize page tables properly?

Just move everything into kasan_init_64.c
kasan_zero_p* could be a normal C arrays (with __page_aligned_bss attribute).
Initialize these page tables in something like kasan_early_init() before kasan_map_early_shadow()
call.


> Best regards,
> Alexander
> 
> 

Re: [PATCH 1/1] x86_64: fix KASan shadow region page tables

[Alexander Popov]

On 26.05.2015 11:52, Andrey Ryabinin wrote:
> On 05/25/2015 04:43 PM, Alexander Popov wrote:
>> What do you think about leaving the initialization of KASan shadow region
>> page tables in arch/x86/kernel/head_64.S and calling something like
>> kasan_fixup_early_shadow() before calling kasan_map_early_shadow()
>> for the first time?
> 
> Why do we have initialize + fixup later if we could simply initialize page tables properly?
> 
> Just move everything into kasan_init_64.c
> kasan_zero_p* could be a normal C arrays (with __page_aligned_bss attribute).
> Initialize these page tables in something like kasan_early_init() before kasan_map_early_shadow()
> call.

Thanks, I'll do that and return with the second version.

Best regards,
Alexander