Re: [Xen-devel] [PATCHv1] x86: don't schedule when handling #NM exception

[Sarah Newman <srn@prgmr.com>]

Can you please review my patch first?  It's only enabled when absolutely required.

On 03/16/2014 08:33 PM, H. Peter Anvin wrote:
> No, the right thing is to unf*ck the Xen braindamage and use eagerfpu as a workaround for the legacy hypervisor versions.
> 
> GFP_ATOMIC -> SIGKILL is definitely a NAK.
> 
> On March 16, 2014 8:13:05 PM PDT, Sarah Newman <srn@prgmr.com> wrote:
>> On 03/10/2014 10:15 AM, David Vrabel wrote:
>>> On 10/03/14 16:40, H. Peter Anvin wrote:
>>>> On 03/10/2014 09:17 AM, David Vrabel wrote:
>>>>> math_state_restore() is called from the #NM exception handler.  It
>> may
>>>>> do a GFP_KERNEL allocation (in init_fpu()) which may schedule.
>>>>>
>>>>> Change this allocation to GFP_ATOMIC, but leave all the other
>> callers
>>>>> of init_fpu() or fpu_alloc() using GFP_KERNEL.
>>>>
>>>> And what the [Finnish] do you do if GFP_ATOMIC fails?
>>>
>>> The same thing it used to do -- kill the task with SIGKILL.  I
>> haven't
>>> changed this behaviour.
>>>
>>>> Sarah's patchset switches Xen PV to use eagerfpu unconditionally,
>> which
>>>> removes the dependency on #NM and is the right thing to do.
>>>
>>> Ok. I'll wait for this series and not pursue this patch any further.
>>
>> Sorry, this got swallowed by my mail filter.
>>
>> I did some more testing and I think eagerfpu is going to noticeably
>> slow things down. When I ran
>> "time sysbench --num-threads=64 --test=threads run" I saw on the order
>> of 15% more time spent in
>> system mode and this seemed consistent over different runs.
>>
>> As for GFP_ATOMIC, unfortunately I don't know a sanctioned test here so
>> I rolled my own. This test
>> sequentially allocated math-using processes in the background until it
>> could not any more.  On a
>> 64MB instance, I saw 10% fewer processes allocated with GFP_ATOMIC
>> compared to GFP_KERNEL when I
>> continually allocated new processes up to OOM conditions (256 vs 228.) 
>> A similar test on a
>> different RFS and a kernel using GFP_NOWAIT showed pretty much no
>> difference in how many processes I
>> could allocate. This doesn't seem too bad unless there is some kind of
>> fragmentation over time which
>> would cause worse performance.
>>
>> Since performance degradation applies at all times and not just under
>> extreme conditions, I think
>> the lesser evil will actually be GFP_ATOMIC.  But it's not necessary to
>> always use GFP_ATOMIC, only
>> under certain conditions - IE when the xen PVABI forces us to.
>>
>> Patches will be supplied shortly.
>