Re: ABI change for device drivers using future AVX instruction set

["H. Peter Anvin" <hpa@zytor.com>]

Andi Kleen wrote:
>>>   
>> If you use xsave, I don't see how this is different to the user fpu save
>> area.
> 
> For once there's no clear error handling path for allocation failures
> on the (arbitarily sized) xsave state. On user code that can be barely
> tolerated, but for the kernel it would be deadly.
> 

Well, that's relatively easily dealt with... you'd have to allocate that 
state save area explicitly in kernel_fpu_begin(), and it would be up to 
the callers of that function to handle the resulting sleep and/or 
allocation failure -- we could even make kernel_fpu_begin() take a GFP_* 
flag.

Now, there are a few possibilities what to do with said state area.  One 
is to make it associated with the task; if used for something as RAID, 
this would mean pretty soon *all* (or nearly all) tasks have such a 
state area, and we might as well go back to allocating them preemptively 
on thread creation.  The other, of course, is to destroy it in 
kernel_fpu_end().  This may cause quite a bit of allocation/deallocation 
overhead, but perhaps this would be a decent use of a quicklist.

	-hpa