Commit 1958b5fc4010 ("x86/boot: Add early boot support when running
with SEV active") is causing kexec becomes sometimes unstable even if
SEV is not active. kexec reboot won't start a second kernel bypassing
BIOS boot process, instead, the system got reset.
That's because, in get_sev_encryption_bit function, we are using
32-bit RIP-relative addressing to read the value of enc_bit, but
kexec may alloc the early boot up code to a higher location, which
is beyond 32-bit addressing limit. Some garbage will be read and
get_sev_encryption_bit will return the wrong value, which leads to
wrong memory page flag.
This patch removes the use of enc_bit, as currently, enc_bit's only
purpose is to avoid duplicated encryption bit reading, but the overhead
of reading encryption bit is so tiny, so no need to cache that.
Fixes: 1958b5fc4010 ("x86/boot: Add early boot support when running with SEV active")
Suggested-by: Borislav Petkov <[email protected]>
Signed-off-by: Kairui Song <[email protected]>
---
arch/x86/boot/compressed/mem_encrypt.S | 19 -------------------
1 file changed, 19 deletions(-)
diff --git a/arch/x86/boot/compressed/mem_encrypt.S b/arch/x86/boot/compressed/mem_encrypt.S
index eaa843a52907..a480356e0ed8 100644
--- a/arch/x86/boot/compressed/mem_encrypt.S
+++ b/arch/x86/boot/compressed/mem_encrypt.S
@@ -25,20 +25,6 @@ ENTRY(get_sev_encryption_bit)
push %ebx
push %ecx
push %edx
- push %edi
-
- /*
- * RIP-relative addressing is needed to access the encryption bit
- * variable. Since we are running in 32-bit mode we need this call/pop
- * sequence to get the proper relative addressing.
- */
- call 1f
-1: popl %edi
- subl $1b, %edi
-
- movl enc_bit(%edi), %eax
- cmpl $0, %eax
- jge .Lsev_exit
/* Check if running under a hypervisor */
movl $1, %eax
@@ -69,15 +55,12 @@ ENTRY(get_sev_encryption_bit)
movl %ebx, %eax
andl $0x3f, %eax /* Return the encryption bit location */
- movl %eax, enc_bit(%edi)
jmp .Lsev_exit
.Lno_sev:
xor %eax, %eax
- movl %eax, enc_bit(%edi)
.Lsev_exit:
- pop %edi
pop %edx
pop %ecx
pop %ebx
@@ -113,8 +96,6 @@ ENTRY(set_sev_encryption_mask)
ENDPROC(set_sev_encryption_mask)
.data
-enc_bit:
- .int 0xffffffff
#ifdef CONFIG_AMD_MEM_ENCRYPT
.balign 8
--
2.17.1
On 09/27/2018 07:38 AM, Kairui Song wrote:
> Commit 1958b5fc4010 ("x86/boot: Add early boot support when running
> with SEV active") is causing kexec becomes sometimes unstable even if
> SEV is not active. kexec reboot won't start a second kernel bypassing
> BIOS boot process, instead, the system got reset.
>
> That's because, in get_sev_encryption_bit function, we are using
> 32-bit RIP-relative addressing to read the value of enc_bit, but
> kexec may alloc the early boot up code to a higher location, which
> is beyond 32-bit addressing limit. Some garbage will be read and
> get_sev_encryption_bit will return the wrong value, which leads to
> wrong memory page flag.
>
> This patch removes the use of enc_bit, as currently, enc_bit's only
> purpose is to avoid duplicated encryption bit reading, but the overhead
> of reading encryption bit is so tiny, so no need to cache that.
>
> Fixes: 1958b5fc4010 ("x86/boot: Add early boot support when running with SEV active")
> Suggested-by: Borislav Petkov <[email protected]>
> Signed-off-by: Kairui Song <[email protected]>
Reviewed-by: Tom Lendacky <[email protected]>
> ---
> arch/x86/boot/compressed/mem_encrypt.S | 19 -------------------
> 1 file changed, 19 deletions(-)
>
> diff --git a/arch/x86/boot/compressed/mem_encrypt.S b/arch/x86/boot/compressed/mem_encrypt.S
> index eaa843a52907..a480356e0ed8 100644
> --- a/arch/x86/boot/compressed/mem_encrypt.S
> +++ b/arch/x86/boot/compressed/mem_encrypt.S
> @@ -25,20 +25,6 @@ ENTRY(get_sev_encryption_bit)
> push %ebx
> push %ecx
> push %edx
> - push %edi
> -
> - /*
> - * RIP-relative addressing is needed to access the encryption bit
> - * variable. Since we are running in 32-bit mode we need this call/pop
> - * sequence to get the proper relative addressing.
> - */
> - call 1f
> -1: popl %edi
> - subl $1b, %edi
> -
> - movl enc_bit(%edi), %eax
> - cmpl $0, %eax
> - jge .Lsev_exit
>
> /* Check if running under a hypervisor */
> movl $1, %eax
> @@ -69,15 +55,12 @@ ENTRY(get_sev_encryption_bit)
>
> movl %ebx, %eax
> andl $0x3f, %eax /* Return the encryption bit location */
> - movl %eax, enc_bit(%edi)
> jmp .Lsev_exit
>
> .Lno_sev:
> xor %eax, %eax
> - movl %eax, enc_bit(%edi)
>
> .Lsev_exit:
> - pop %edi
> pop %edx
> pop %ecx
> pop %ebx
> @@ -113,8 +96,6 @@ ENTRY(set_sev_encryption_mask)
> ENDPROC(set_sev_encryption_mask)
>
> .data
> -enc_bit:
> - .int 0xffffffff
>
> #ifdef CONFIG_AMD_MEM_ENCRYPT
> .balign 8
>
Commit-ID: bdec8d7fa55e6f5314ed72e5a0b435d90ff90548
Gitweb: https://git.kernel.org/tip/bdec8d7fa55e6f5314ed72e5a0b435d90ff90548
Author: Kairui Song <[email protected]>
AuthorDate: Thu, 27 Sep 2018 20:38:45 +0800
Committer: Borislav Petkov <[email protected]>
CommitDate: Thu, 27 Sep 2018 19:35:03 +0200
x86/boot: Fix kexec booting failure in the SEV bit detection code
Commit
1958b5fc4010 ("x86/boot: Add early boot support when running with SEV active")
can occasionally cause system resets when kexec-ing a second kernel even
if SEV is not active.
That's because get_sev_encryption_bit() uses 32-bit rIP-relative
addressing to read the value of enc_bit - a variable which caches a
previously detected encryption bit position - but kexec may allocate
the early boot code to a higher location, beyond the 32-bit addressing
limit.
In this case, garbage will be read and get_sev_encryption_bit() will
return the wrong value, leading to accessing memory with the wrong
encryption setting.
Therefore, remove enc_bit, and thus get rid of the need to do 32-bit
rIP-relative addressing in the first place.
[ bp: massage commit message heavily. ]
Fixes: 1958b5fc4010 ("x86/boot: Add early boot support when running with SEV active")
Suggested-by: Borislav Petkov <[email protected]>
Signed-off-by: Kairui Song <[email protected]>
Signed-off-by: Borislav Petkov <[email protected]>
Reviewed-by: Tom Lendacky <[email protected]>
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Link: https://lkml.kernel.org/r/[email protected]
---
arch/x86/boot/compressed/mem_encrypt.S | 19 -------------------
1 file changed, 19 deletions(-)
diff --git a/arch/x86/boot/compressed/mem_encrypt.S b/arch/x86/boot/compressed/mem_encrypt.S
index eaa843a52907..a480356e0ed8 100644
--- a/arch/x86/boot/compressed/mem_encrypt.S
+++ b/arch/x86/boot/compressed/mem_encrypt.S
@@ -25,20 +25,6 @@ ENTRY(get_sev_encryption_bit)
push %ebx
push %ecx
push %edx
- push %edi
-
- /*
- * RIP-relative addressing is needed to access the encryption bit
- * variable. Since we are running in 32-bit mode we need this call/pop
- * sequence to get the proper relative addressing.
- */
- call 1f
-1: popl %edi
- subl $1b, %edi
-
- movl enc_bit(%edi), %eax
- cmpl $0, %eax
- jge .Lsev_exit
/* Check if running under a hypervisor */
movl $1, %eax
@@ -69,15 +55,12 @@ ENTRY(get_sev_encryption_bit)
movl %ebx, %eax
andl $0x3f, %eax /* Return the encryption bit location */
- movl %eax, enc_bit(%edi)
jmp .Lsev_exit
.Lno_sev:
xor %eax, %eax
- movl %eax, enc_bit(%edi)
.Lsev_exit:
- pop %edi
pop %edx
pop %ecx
pop %ebx
@@ -113,8 +96,6 @@ ENTRY(set_sev_encryption_mask)
ENDPROC(set_sev_encryption_mask)
.data
-enc_bit:
- .int 0xffffffff
#ifdef CONFIG_AMD_MEM_ENCRYPT
.balign 8