Hi,
I was running some perf mem test for an upcoming patch when
I realize that precise=2 was broken on 3.18. It seems it never
(or extremely rarely) correct the off-by-one error, when until 3.18-rc4
it was 100% on the same program. So something was introduced
that broke the asm walker in perf_event_intel_ds.c.
Looking at the log of that file, I can see one change that could have
some impact:
Author: Dave Hansen <[email protected]>
6ba48ff x86: Remove arbitrary instruction size limit in instruction decoder
if I use a kernel without this fix (prior to that commit), then correction
works. Any kernel after fails. I have not investigated why but may you
have an idea.
To reproduce try using perf mem -t load rec my_load_test, then use
perf report to navigate to the assembly view, the samples should be
on load instructions, not on the instructions following them. If you use
perf mem -t load rec -vv you can verify that precise=2. So something
is not working anymore in the instruction decoder that the fixup routine
bails out.
Any clue?
On Mon, Dec 15, 2014 at 11:51:07PM -0500, Stephane Eranian wrote:
> Hi,
>
> I was running some perf mem test for an upcoming patch when
> I realize that precise=2 was broken on 3.18. It seems it never
> (or extremely rarely) correct the off-by-one error, when until 3.18-rc4
> it was 100% on the same program. So something was introduced
> that broke the asm walker in perf_event_intel_ds.c.
>
> Looking at the log of that file, I can see one change that could have
> some impact:
>
> Author: Dave Hansen <[email protected]>
> 6ba48ff x86: Remove arbitrary instruction size limit in instruction decoder
>
> if I use a kernel without this fix (prior to that commit), then correction
> works. Any kernel after fails. I have not investigated why but may you
> have an idea.
>
> To reproduce try using perf mem -t load rec my_load_test, then use
> perf report to navigate to the assembly view, the samples should be
> on load instructions, not on the instructions following them. If you use
> perf mem -t load rec -vv you can verify that precise=2. So something
> is not working anymore in the instruction decoder that the fixup routine
> bails out.
>
> Any clue?
This appears to have fixed it.
---
Subject: x86: Fix off-by-one in instruction decoder
Stephane reported that the PEBS fixup was broken by the recent commit to
the instruction decoder. The thing had an off-by-one which resulted in
not being able to decode the last instruction and always bail.
Reported-by: Stephane Eranian <[email protected]>
Fixes: 6ba48ff46f76 ("x86: Remove arbitrary instruction size limit in instruction decoder")
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
---
arch/x86/lib/insn.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/arch/x86/lib/insn.c b/arch/x86/lib/insn.c
index 2480978..1313ae6 100644
--- a/arch/x86/lib/insn.c
+++ b/arch/x86/lib/insn.c
@@ -28,7 +28,7 @@
/* Verify next sizeof(t) bytes can be on the same instruction */
#define validate_next(t, insn, n) \
- ((insn)->next_byte + sizeof(t) + n < (insn)->end_kaddr)
+ ((insn)->next_byte + sizeof(t) + n <= (insn)->end_kaddr)
#define __get_next(t, insn) \
({ t r = *(t*)insn->next_byte; insn->next_byte += sizeof(t); r; })
On Tue, Dec 16, 2014 at 5:46 AM, Peter Zijlstra <[email protected]> wrote:
> On Mon, Dec 15, 2014 at 11:51:07PM -0500, Stephane Eranian wrote:
>> Hi,
>>
>> I was running some perf mem test for an upcoming patch when
>> I realize that precise=2 was broken on 3.18. It seems it never
>> (or extremely rarely) correct the off-by-one error, when until 3.18-rc4
>> it was 100% on the same program. So something was introduced
>> that broke the asm walker in perf_event_intel_ds.c.
>>
>> Looking at the log of that file, I can see one change that could have
>> some impact:
>>
>> Author: Dave Hansen <[email protected]>
>> 6ba48ff x86: Remove arbitrary instruction size limit in instruction decoder
>>
>> if I use a kernel without this fix (prior to that commit), then correction
>> works. Any kernel after fails. I have not investigated why but may you
>> have an idea.
>>
>> To reproduce try using perf mem -t load rec my_load_test, then use
>> perf report to navigate to the assembly view, the samples should be
>> on load instructions, not on the instructions following them. If you use
>> perf mem -t load rec -vv you can verify that precise=2. So something
>> is not working anymore in the instruction decoder that the fixup routine
>> bails out.
>>
>> Any clue?
>
> This appears to have fixed it.
>
> ---
> Subject: x86: Fix off-by-one in instruction decoder
>
> Stephane reported that the PEBS fixup was broken by the recent commit to
> the instruction decoder. The thing had an off-by-one which resulted in
> not being able to decode the last instruction and always bail.
>
Works again now. Thanks for fixing this quickly.
Acked-by: Stephane Eranian <[email protected]>
> Reported-by: Stephane Eranian <[email protected]>
> Fixes: 6ba48ff46f76 ("x86: Remove arbitrary instruction size limit in instruction decoder")
> Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
> ---
> arch/x86/lib/insn.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/arch/x86/lib/insn.c b/arch/x86/lib/insn.c
> index 2480978..1313ae6 100644
> --- a/arch/x86/lib/insn.c
> +++ b/arch/x86/lib/insn.c
> @@ -28,7 +28,7 @@
>
> /* Verify next sizeof(t) bytes can be on the same instruction */
> #define validate_next(t, insn, n) \
> - ((insn)->next_byte + sizeof(t) + n < (insn)->end_kaddr)
> + ((insn)->next_byte + sizeof(t) + n <= (insn)->end_kaddr)
>
> #define __get_next(t, insn) \
> ({ t r = *(t*)insn->next_byte; insn->next_byte += sizeof(t); r; })