On Wed, Feb 03, 2021 at 04:10:24AM +0000, Bae, Chang Seok wrote:
> Okay, how about:
> “
> This alignment bit is set if the state is saved on a 64B-aligned address in
> the compacted format buffer.
> "
I'd prefer:
/*
* True if the buffer of the corresponding XFEATURE is located on the next 64
* byte boundary. Otherwise, it follows the preceding component immediately.
*/
static bool xstate_aligns[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = false };
> The threshold here could be more than that. But the intention is a heads-up to
> (re-)consider (a) a new allocation mechanism and (b) to shrink the memory
> allocation.
>
> Also, the AMX state size is limited to (a bit less than) 64KB and it was
> discussed that vmalloc() will be okay with AMX [2].
So if nothing is going to grow over 64K, why are we even talking about this?
> Maybe it is possible to backtrack this allocation failure out of #NM handling.
> But the tracepoint can provide a clear context, although limited to those
> using it.
Yes, add it when it is really needed. Not slapping it proactively and
hoping for any potential usage.
> Indeed, this is the most preferred way on one hand. But there was a change to
> the current allocation approach by Ingo about 6 years ago [3].
Yah, there's that. :-\
I guess it needs to stay embedded. Oh well.
I guess you can diminish the confusion by doing this:
struct fpu {
...
union fpregs_state *state;
union fpregs_state __default_state;
};
and tasks will have
state = &__default_state;
set up by default in fpu__copy() etc.
AMX tasks will simply change the pointer to the vmalloc'ed xstate
buffer. This way at least the pointer will be a single one and the task
alloc code will simply reroute it instead of having two things to pay
attention to.
Thx.
--
Regards/Gruss,
Boris.
SUSE Software Solutions Germany GmbH, GF: Felix Imendörffer, HRB 36809, AG
Nürnberg
