[Qemu-devel] [PATCH] linux-user: provide frame information in x86-64 safe_syscall

[Qemu-devel] [PATCH] linux-user: provide frame information in x86-64 safe_syscall

[Peter Maydell]

Use cfi directives in the x86-64 safe_syscall to allow gdb to get
backtraces right from within it. (In particular this will be
quite a common situation if the user interrupts QEMU while it's
in a blocked safe-syscall: at the point of the syscall insn RBP
is in use for something else, and so gdb can't find the frame then
without assistance.)

Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
---
The requirements for frame information annotations seem to be a bit
of an undocumented black art, but I think I have these right. At
any rate, gdb now gives correct backtraces from all points in
the routine as far as I can see. Review appreciated...


 linux-user/host/x86_64/safe-syscall.inc.S | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/linux-user/host/x86_64/safe-syscall.inc.S b/linux-user/host/x86_64/safe-syscall.inc.S
index dde434c..bbb1eca 100644
--- a/linux-user/host/x86_64/safe-syscall.inc.S
+++ b/linux-user/host/x86_64/safe-syscall.inc.S
@@ -24,6 +24,7 @@
          * -1-and-errno-set convention is done by the calling wrapper.
          */
 safe_syscall_base:
+        .cfi_startproc
         /* This saves a frame pointer and aligns the stack for the syscall.
          * (It's unclear if the syscall ABI has the same stack alignment
          * requirements as the userspace function call ABI, but better safe than
@@ -31,6 +32,8 @@ safe_syscall_base:
          * does not list any ABI differences regarding stack alignment.)
          */
         push    %rbp
+        .cfi_def_cfa_offset 16
+        .cfi_offset rbp,-16
 
         /* The syscall calling convention isn't the same as the
          * C one:
@@ -70,12 +73,19 @@ safe_syscall_start:
 safe_syscall_end:
         /* code path for having successfully executed the syscall */
         pop     %rbp
+        .cfi_remember_state
+        .cfi_def_cfa_offset 8
+        .cfi_restore ebp
         ret
 
 return_ERESTARTSYS:
         /* code path when we didn't execute the syscall */
+        .cfi_restore_state
         mov     $-TARGET_ERESTARTSYS, %rax
         pop     %rbp
+        .cfi_def_cfa_offset 8
+        .cfi_restore ebp
         ret
+        .cfi_endproc
 
         .size   safe_syscall_base, .-safe_syscall_base
-- 
1.9.1

Re: [Qemu-devel] [PATCH] linux-user: provide frame information in x86-64 safe_syscall

[Richard Henderson]

On 05/27/2016 08:06 AM, Peter Maydell wrote:
> @@ -31,6 +32,8 @@ safe_syscall_base:
>           * does not list any ABI differences regarding stack alignment.)
>           */
>          push    %rbp
> +        .cfi_def_cfa_offset 16
> +        .cfi_offset rbp,-16

While this is correct, there are two other directives that make it easier to 
describe changes without having to compute globally correct constants.  Here 
they would be:

	.cfi_adjust_cfa_offset 8

Add 8 to the offset, i.e. decrement the SP by 8.

	.cfi_rel_offset rbp, 0

Save rbp at the current top-of-stack.

The assembler will compute the absolute values from these relative values. 
Using them makes it easy to see from a narrow window of code that the 
annotations are correct.

Otherwise,

Reviewed-by: Richard Henderson <rth@twiddle.net>


r~

Re: [Qemu-devel] [PATCH] linux-user: provide frame information in x86-64 safe_syscall

[Peter Maydell]

On 27 May 2016 at 17:21, Richard Henderson <rth@twiddle.net> wrote:
> On 05/27/2016 08:06 AM, Peter Maydell wrote:
>>
>> @@ -31,6 +32,8 @@ safe_syscall_base:
>>           * does not list any ABI differences regarding stack alignment.)
>>           */
>>          push    %rbp
>> +        .cfi_def_cfa_offset 16
>> +        .cfi_offset rbp,-16
>
>
> While this is correct, there are two other directives that make it easier to
> describe changes without having to compute globally correct constants.  Here
> they would be:
>
>         .cfi_adjust_cfa_offset 8
>
> Add 8 to the offset, i.e. decrement the SP by 8.

Presumably .cfi_startproc sets the initial offset to 8?
(It's not documented that it does so, which is I think partly why
I preferred to use a directive that definitely set the offset
to the right thing.)

>         .cfi_rel_offset rbp, 0
>
> Save rbp at the current top-of-stack.
>
> The assembler will compute the absolute values from these relative values.
> Using them makes it easy to see from a narrow window of code that the
> annotations are correct.
>
> Otherwise,
>
> Reviewed-by: Richard Henderson <rth@twiddle.net>

Thanks. I'll spin a v2 with your changes in it next week.

-- PMM

Re: [Qemu-devel] [PATCH] linux-user: provide frame information in x86-64 safe_syscall

[Peter Maydell]

On 27 May 2016 at 16:06, Peter Maydell <peter.maydell@linaro.org> wrote:
>  return_ERESTARTSYS:
>          /* code path when we didn't execute the syscall */
> +        .cfi_restore_state
>          mov     $-TARGET_ERESTARTSYS, %rax
>          pop     %rbp
> +        .cfi_def_cfa_offset 8
> +        .cfi_restore ebp

These should read '.cfi_restore rbp'. I have no idea how that got through
testing since it doesn't compile at all...

thanks
-- PMM

Re: [Qemu-devel] [PATCH] linux-user: provide frame information in x86-64 safe_syscall

[Richard Henderson]

On 05/27/2016 09:34 AM, Peter Maydell wrote:
> On 27 May 2016 at 17:21, Richard Henderson <rth@twiddle.net> wrote:
>> On 05/27/2016 08:06 AM, Peter Maydell wrote:
>>>
>>> @@ -31,6 +32,8 @@ safe_syscall_base:
>>>           * does not list any ABI differences regarding stack alignment.)
>>>           */
>>>          push    %rbp
>>> +        .cfi_def_cfa_offset 16
>>> +        .cfi_offset rbp,-16
>>
>>
>> While this is correct, there are two other directives that make it easier to
>> describe changes without having to compute globally correct constants.  Here
>> they would be:
>>
>>         .cfi_adjust_cfa_offset 8
>>
>> Add 8 to the offset, i.e. decrement the SP by 8.
>
> Presumably .cfi_startproc sets the initial offset to 8?
> (It's not documented that it does so, which is I think partly why
> I preferred to use a directive that definitely set the offset
> to the right thing.)

It is documented to set up the normal no-instructions-executed call frame. 
Which in the case of x86, does have a non-zero offset.

There is a ".cfi_startproc simple" that begins a frame with no opcodes at all.


r~