On 07/07/2016 15:21, Benjamin Herrenschmidt wrote:
> On Thu, 2016-07-07 at 15:12 +0200, Laurent Dufour wrote:
>>
>> Basically, CRIU checkpoints the process register's state through the
>> ptrace API, and it restores it through a signal frame at restart time.
>> This is quite odd but that the way it works on all the CRIU's supported
>> architectures.
>>
>> Obviously everything is done from/in user space, so the sigframe
>> building too.
>> Since we can't know from user space if the thread has used or not the
>> Altivec/VSX registers, since we can't rely on the MSR bits, we always
>> dump these registers.
>
> Right, however is that an issue ? These days with glibc using V{M,S}X
> for things like memcpy I would think there is little to gain in trying
> to avoid dumping them.
>
>>> Alternately, when restoring, can you setup the sigframe with the Altivec/VSX
>>> fields populated, and the kernel will then load them, regardless of whether
>>> they were actually used or not prior to the checkpoint?
>>
>> In the case of Altivec/VSX fields, we currently force the kernel to
>> retrieve them from the signal frame by setting MSR_VEC/MSR_VSX so
>> restore_sigcontext() will copy them to the kernel thread's state.
>
> Yup, that's the way to go.
>
>> However this doesn't touch to used_vsr and used_vr which may remain at 0.
>
> That would be a kernel bug.
>
>> Most of the time this is fine, but in the case a thread which has really
>> used those registers is catching a signal just after the restore and
>> before it has touched to these registers again (and so set used_vsr/vr),
>> these registers will not be pushed in the newly built signal frame since
>> setup_sigcontext() check for used_vsr/vr before pushing the registers on
>> the stack.
>> This may be an issue in the case the thread wants to changed those
>> registers (don't ask me why :)) in the stacked signal frame from the
>> signal handler since they will not be there...
>>
>> Being able to get and set the used_vr and used_vsr thread's variables,
>> fixes this issue.
>
> I think the right fix is that if a restore_sigcontext() has the MSR bits set,
> it should set the corresponding used_* flag.
>
> Or is there a reason why that won't work ?
I got your point and I agree that most of the time now, the Altivec/VSX
registers are used by libc. In that case is there still a need for the
lazy Altivec/VSX registers dump in the signal frame ?
I'm fine with your proposal, except that every restarted process will
have the used_vr/used_vsx turned on after the restart since we can't
check if these registers were used or not at checkpoint time.
But that may be a minor point...
Cheers,
Laurent.
