1. Patchew
  2. linux
  3. View series

[PATCH] x86/mce: Reduce number of machine checks taken during recovery

Tony Luck posted 1 patch 4 years, 6 months ago 
arch/x86/lib/copy_user_64.S | 7 +++++++
1 file changed, 7 insertions(+)
[PATCH] x86/mce: Reduce number of machine checks taken during recovery
Posted by Tony Luck 4 years, 6 months ago 
From: Youquan Song <youquan.song@intel.com>

A machine check recovery test case injects poison into memory and then
tries a write(2) system call with a buffer that begins 64 bytes before
the poison and includes the first poison byte:

$ sudo ./einj_mem_uc -f -m 64:65:0 copyin
0: copyin   vaddr = 0x7f13639cc400 paddr = 834fa0d400
./einj_mem_uc: short (64 bytes) write to temp file
page not present
Big surprise ... still running. Thought that would be fatal
Unusual number of MCEs seen: 4
Test passed

While it executes correctly (just the 64 non-poison bytes are written)
and recovers, four machine checks seems like a lot.

When any of the copy functions in arch/x86/lib/copy_user_64.S consume
poison and take a fault the fixup code copies the remaining byte count
from %ecx to %edx and unconditionally jumps to .Lcopy_user_handle_tail
to continue the copy in case any more bytes can be copied.

If the fault was #PF this may copy more bytes (because the page fault
handler might have fixed the fault). But when the fault is a machine
check the original copy code will have copied all the way to the poisoned
cache line. So .Lcopy_user_handle_tail will just take another machine
check for no good reason.

There are five distinct blocks of fixup code that branch to
.Lcopy_user_handle_tail, so add a check there to check the trap type
(in %eax) and simply return the count of remaining bytes.

As well as reducing the number of machine checks, this also allows
Skylake generation Xeons to recover some cases that currently fail. The
is because "rep movsb" is only recoverable when source and destination
are well aligned and the byte count is large. That useless call to
.Lcopy_user_handle_tail may violate one or more of these conditions and
generate a fatal machine check.

With this change applied the count of machine checks to recover is
reduced from four to three:

$ sudo ./einj_mem_uc -f -m 64:65:0 copyin
0: copyin   vaddr = 0x7f1e44845400 paddr = 55930b400
./einj_mem_uc: short (64 bytes) write to temp file
page not present
Big surprise ... still running. Thought that would be fatal
Unusual number of MCEs seen: 3
Test passed

[Tony: Added more details to commit message]

Signed-off-by: Youquan Song <youquan.song@intel.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
---
 arch/x86/lib/copy_user_64.S | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/arch/x86/lib/copy_user_64.S b/arch/x86/lib/copy_user_64.S
index 2797e630b9b1..8c53be99faa0 100644
--- a/arch/x86/lib/copy_user_64.S
+++ b/arch/x86/lib/copy_user_64.S
@@ -233,12 +233,19 @@ EXPORT_SYMBOL(copy_user_enhanced_fast_string)
  * eax uncopied bytes or 0 if successful.
  */
 SYM_CODE_START_LOCAL(.Lcopy_user_handle_tail)
+	cmp $X86_TRAP_MC,%eax
+	je 3f
 	movl %edx,%ecx
 1:	rep movsb
 2:	mov %ecx,%eax
 	ASM_CLAC
 	ret
 
+3:
+	movl %edx,%eax
+	ASM_CLAC
+	ret
+
 	_ASM_EXTABLE_CPY(1b, 2b)
 SYM_CODE_END(.Lcopy_user_handle_tail)
 
-- 
2.31.1
Re: [PATCH] x86/mce: Reduce number of machine checks taken during recovery
Posted by Peter Zijlstra 4 years, 6 months ago 
On Wed, Dec 15, 2021 at 02:20:16PM -0800, Tony Luck wrote:
> ---
>  arch/x86/lib/copy_user_64.S | 7 +++++++
>  1 file changed, 7 insertions(+)
> 
> diff --git a/arch/x86/lib/copy_user_64.S b/arch/x86/lib/copy_user_64.S
> index 2797e630b9b1..8c53be99faa0 100644
> --- a/arch/x86/lib/copy_user_64.S
> +++ b/arch/x86/lib/copy_user_64.S
> @@ -233,12 +233,19 @@ EXPORT_SYMBOL(copy_user_enhanced_fast_string)
>   * eax uncopied bytes or 0 if successful.
>   */
>  SYM_CODE_START_LOCAL(.Lcopy_user_handle_tail)
> +	cmp $X86_TRAP_MC,%eax
> +	je 3f
>  	movl %edx,%ecx
>  1:	rep movsb
>  2:	mov %ecx,%eax
>  	ASM_CLAC
>  	ret
>  
> +3:
> +	movl %edx,%eax
> +	ASM_CLAC
> +	ret
> +
>  	_ASM_EXTABLE_CPY(1b, 2b)
>  SYM_CODE_END(.Lcopy_user_handle_tail)

This patch no longer applies; however, you forgot to change the comment
on top about the calling convention, because now the function expects
rax to contain the trap number.

It's also not obvious from the massive rambling on top that all
callsites were audited to make sure this is in fact true (it appears
so).
[PATCH v2] x86/mce: Reduce number of machine checks taken during recovery
Posted by Luck, Tony 4 years, 6 months ago 

From: Youquan Song <youquan.song@intel.com>

When any of the copy functions in arch/x86/lib/copy_user_64.S take a
fault, the fixup code copies the remaining byte count from %ecx to %edx
and unconditionally jumps to .Lcopy_user_handle_tail to continue the
copy in case any more bytes can be copied.

If the fault was #PF this may copy more bytes (because the page fault
handler might have fixed the fault). But when the fault is a machine
check the original copy code will have copied all the way to the poisoned
cache line. So .Lcopy_user_handle_tail will just take another machine
check for no good reason.

Every code path to .Lcopy_user_handle_tail comes from an exception fixup
path, so add a check there to check the trap type (in %eax) and simply
return the count of remaining bytes if the trap was a machine check.

Doing this reduces the number of machine checks taken during synthetic
tests from four to three.

As well as reducing the number of machine checks, this also allows
Skylake generation Xeons to recover some cases that currently fail. The
is because "rep movsb" is only recoverable when source and destination
are well aligned and the byte count is large. That useless call to
.Lcopy_user_handle_tail may violate one or more of these conditions and
generate a fatal machine check.

[Tony: Added more details to commit message]

Signed-off-by: Youquan Song <youquan.song@intel.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
---
 arch/x86/lib/copy_user_64.S | 9 +++++++++
 1 file changed, 9 insertions(+)

V2: Update based on comments from PeterZ
	- Rebase to tip x86/core branch
	- Add comment documenting eax is now part of calling convention
	  for .Lcopy_user_handle_tail
	- Trim commit comment to make it easier to see that I did
	  audit all the call sequences for .Lcopy_user_handle_tail

diff --git a/arch/x86/lib/copy_user_64.S b/arch/x86/lib/copy_user_64.S
index e6ac38587b40..26781cbe7e37 100644
--- a/arch/x86/lib/copy_user_64.S
+++ b/arch/x86/lib/copy_user_64.S
@@ -212,6 +212,7 @@ EXPORT_SYMBOL(copy_user_enhanced_fast_string)
  * Don't try to copy the tail if machine check happened
  *
  * Input:
+ * eax x86 trap number - set by fixup_excpetion()
  * rdi destination
  * rsi source
  * rdx count
@@ -220,12 +221,20 @@ EXPORT_SYMBOL(copy_user_enhanced_fast_string)
  * eax uncopied bytes or 0 if successful.
  */
 SYM_CODE_START_LOCAL(.Lcopy_user_handle_tail)
+	cmp $X86_TRAP_MC,%eax
+	je 3f
+
 	movl %edx,%ecx
 1:	rep movsb
 2:	mov %ecx,%eax
 	ASM_CLAC
 	RET
 
+3:
+	movl %edx,%eax
+	ASM_CLAC
+	RET
+
 	_ASM_EXTABLE_CPY(1b, 2b)
 
 .Lcopy_user_handle_align:

base-commit: 82a8954acd93ae95d6252fb93a3d210c8f71b093
-- 
2.31.1
Re: [PATCH v2] x86/mce: Reduce number of machine checks taken during recovery
Posted by Peter Zijlstra 4 years, 6 months ago 
On Thu, Dec 23, 2021 at 12:07:01PM -0800, Luck, Tony wrote:
> diff --git a/arch/x86/lib/copy_user_64.S b/arch/x86/lib/copy_user_64.S
> index e6ac38587b40..26781cbe7e37 100644
> --- a/arch/x86/lib/copy_user_64.S
> +++ b/arch/x86/lib/copy_user_64.S
> @@ -212,6 +212,7 @@ EXPORT_SYMBOL(copy_user_enhanced_fast_string)
>   * Don't try to copy the tail if machine check happened
>   *
>   * Input:
> + * eax x86 trap number - set by fixup_excpetion()

That's inaccurate, fixup_exception() (event if it's spelled correctly)
does not unconditionally set the trap number in RAX, that's only done by
ex_handler_fault() (or ex_handler_sgx()), which means all flows into
this function must pass through: EX_TYPE_FAULT, EX_TYPE_FAULT_MCE or
EX_TYPE_COPY.

Boris might fix up your comment if he applies I suppose..

>   * rdi destination
>   * rsi source
>   * rdx count
> @@ -220,12 +221,20 @@ EXPORT_SYMBOL(copy_user_enhanced_fast_string)
>   * eax uncopied bytes or 0 if successful.
>   */
>  SYM_CODE_START_LOCAL(.Lcopy_user_handle_tail)
> +	cmp $X86_TRAP_MC,%eax
> +	je 3f
> +
>  	movl %edx,%ecx
>  1:	rep movsb
>  2:	mov %ecx,%eax
>  	ASM_CLAC
>  	RET
>  
> +3:
> +	movl %edx,%eax
> +	ASM_CLAC
> +	RET
> +
>  	_ASM_EXTABLE_CPY(1b, 2b)
>  
>  .Lcopy_user_handle_align:
> 
> base-commit: 82a8954acd93ae95d6252fb93a3d210c8f71b093
> -- 
> 2.31.1
>
[tip: ras/core] x86/mce: Reduce number of machine checks taken during recovery
Posted by tip-bot2 for Youquan Song 4 years, 6 months ago 
The following commit has been merged into the ras/core branch of tip:

Commit-ID:     3376136300a00df9a864b88fa969177d6c3be8e5
Gitweb:        https://git.kernel.org/tip/3376136300a00df9a864b88fa969177d6c3be8e5
Author:        Youquan Song <youquan.song@intel.com>
AuthorDate:    Thu, 23 Dec 2021 12:07:01 -08:00
Committer:     Borislav Petkov <bp@suse.de>
CommitterDate: Fri, 31 Dec 2021 18:22:32 +01:00

x86/mce: Reduce number of machine checks taken during recovery

When any of the copy functions in arch/x86/lib/copy_user_64.S take a
fault, the fixup code copies the remaining byte count from %ecx to %edx
and unconditionally jumps to .Lcopy_user_handle_tail to continue the
copy in case any more bytes can be copied.

If the fault was #PF this may copy more bytes (because the page fault
handler might have fixed the fault). But when the fault is a machine
check the original copy code will have copied all the way to the poisoned
cache line. So .Lcopy_user_handle_tail will just take another machine
check for no good reason.

Every code path to .Lcopy_user_handle_tail comes from an exception fixup
path, so add a check there to check the trap type (in %eax) and simply
return the count of remaining bytes if the trap was a machine check.

Doing this reduces the number of machine checks taken during synthetic
tests from four to three.

As well as reducing the number of machine checks, this also allows
Skylake generation Xeons to recover some cases that currently fail. The
is because REP; MOVSB is only recoverable when source and destination
are well aligned and the byte count is large. That useless call to
.Lcopy_user_handle_tail may violate one or more of these conditions and
generate a fatal machine check.

  [ Tony: Add more details to commit message. ]
  [ bp: Fixup comment.
    Also, another tip patchset which is adding straight-line speculation
    mitigation changes the "ret" instruction to an all-caps macro "RET".
    But, since gas is case-insensitive, use "RET" in the newly added asm block
    already in order to simplify tip branch merging on its way upstream.
  ]

Signed-off-by: Youquan Song <youquan.song@intel.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Link: https://lore.kernel.org/r/YcTW5dh8yTGucDd+@agluck-desk2.amr.corp.intel.com
---
 arch/x86/lib/copy_user_64.S |  9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/arch/x86/lib/copy_user_64.S b/arch/x86/lib/copy_user_64.S
index 2797e63..e73b76e 100644
--- a/arch/x86/lib/copy_user_64.S
+++ b/arch/x86/lib/copy_user_64.S
@@ -225,6 +225,7 @@ EXPORT_SYMBOL(copy_user_enhanced_fast_string)
  * Don't try to copy the tail if machine check happened
  *
  * Input:
+ * eax trap number written by ex_handler_copy()
  * rdi destination
  * rsi source
  * rdx count
@@ -233,12 +234,20 @@ EXPORT_SYMBOL(copy_user_enhanced_fast_string)
  * eax uncopied bytes or 0 if successful.
  */
 SYM_CODE_START_LOCAL(.Lcopy_user_handle_tail)
+	cmp $X86_TRAP_MC,%eax
+	je 3f
+
 	movl %edx,%ecx
 1:	rep movsb
 2:	mov %ecx,%eax
 	ASM_CLAC
 	ret
 
+3:
+	movl %edx,%eax
+	ASM_CLAC
+	RET
+
 	_ASM_EXTABLE_CPY(1b, 2b)
 SYM_CODE_END(.Lcopy_user_handle_tail)

Patchew 2.1-devel-ea86937

© 2016 - 2026 Red Hat, Inc.