Re: [RFC PATCH 13/22] x86/fpu/xstate: Expand dynamic user state area on first use

From: Dave Hansen
Date: Tue Nov 03 2020 - 16:41:39 EST


On 11/3/20 1:32 PM, Bae, Chang Seok wrote:
> On Wed, 2020-10-14 at 09:29 -0700, Dave Hansen wrote:
>> On 10/14/20 9:10 AM, Andy Lutomirski wrote:
>>> I was suggesting something a little bit different. We'd keep XMM,
>>> YMM, ZMM, etc state stored exactly the way we do now and, for
>>> AMX-using tasks, we would save the AMX state in an entirely separate
>>> buffer. This way the pain of having a variable xstate layout is
>>> confined just to AMX tasks.
>> OK, got it.
>>
>> So, we'd either need a second set of XSAVE/XRSTORs, or "manual" copying
>> of the registers out to memory. We can preserve the modified
>> optimization if we're careful about ordering, but only for *ONE* of the
>> XSAVE buffers (if we use two).
> For what is worth,
>
> If using two buffers, the buffer for saving the tile data also needs space
> for the legacy states.

Just to be clear, you're talking about the 512-byte 'struct
fxregs_state' which is the first member of 'struct xregs_state', right?

Basically, any two-buffer model wastes 576 bytes of space (legacy fxsave
plus xsave header) for each task that uses two buffers. This provides
an overall space savings unless there are something like 16x more
TMUL-using tasks than non-TMUL-using tasks.

We could eliminate the 576 bytes of waste, but at the *performance* cost
of having a permanently (non-task_struct-embedded) out-of-line XSAVE
buffer for all tasks.

> The AMX state is stored at the end of the XSAVE buffer (at least for now).
> So, the layout (in terms of offsets of non-AMX states) won't be changed at
> run-time.

I don't know what point you are trying to make here.