On Mon, Feb 05, 2024 at 05:11:59PM +0000, Dave Martin wrote:

On Sat, Feb 03, 2024 at 12:16:49PM +0000, Mark Brown wrote:

quoted

We could also teach the ptrace core about runtime discoverable regset sizes
but that would be a more invasive change and this is being observed in
practical systems.

This is not hard at all: see
27e64b4be4b8 ("regset: Add support for dynamically sized regsets") 

But since this is precisely what was ripped out, I guess adding it back
may be controversial (?)

Also just that people might want to backport and while it's not super
*hard* I tend to prefer to do something as minimal as possible as a fix,
the less we do the less the chances that we mess up.

quoted

We should probably also use the actual architectural limit for the
bitmasks we use in the VL enumeration code, though that's both a little
bit more involved and less immediately a problem.

Since these masks are 64 bytes each and rarely accessed, it seemed
pointless complexity to make them resizeable...

I was suggesting making them use the architectural maximum rather than
making them dynamic.

quoted

+#define ARCH_SVE_VQ_MAX 16
 #define SME_VQ_MAX	16

Ack, though part of the reason for not doing this was to discourage
people from allocating statically sized buffers in general.

I was going to do a patch adding a comment to the header noting that
this is not actually the architectural maximum since at present it's
a bit of a landmine, people who have some idea of the architecture
likely have a rough idea what sort of allocation size is needed for the
maximum SVE state and are likely to not double check the value provided
(I think that's what happened with the refactoring to remove the dynamic
sizing).  A comment in the header is still very missable but it'd be
something.

If the kernel is now juggling two #defines for the maximum vector size,
this feels like it may seed bitrot...

Ideally we'd just not have the existing define externally but it's there
and it's been used.