RE: [RFC PATCH 2/3] x86/io: implement 256-bit IO read and write

From: Linus Torvalds

Sent: 22 March 2018 01:27
On Tue, Mar 20, 2018 at 7:42 AM, Alexander Duyck
[off-list ref] wrote:

quoted

Instead of framing this as an enhanced version of the read/write ops
why not look at replacing or extending something like the
memcpy_fromio or memcpy_toio operations?

Yes, doing something like "memcpy_fromio_avx()" is much more
palatable, in that it works like the crypto functions do - if you do
big chunks, the "kernel_fpu_begin/end()" isn't nearly the issue it can
be otherwise.

Note that we definitely have seen hardware that *depends* on the
regular memcpy_fromio()" not doing big reads. I don't know how
hardware people screw it up, but it's clearly possible.

I wonder if that hardware works with the current kernel on recent cpus?
I bet it doesn't like the byte accesses that get generated either.

So it really needs to be an explicitly named function that basically a
driver can use to say "my hardware really likes big aligned accesses"
and explicitly ask for some AVX version if possible.

For x86 being able to request a copy done as 'rep movsx' (for each x)
would be useful.
For io copies the cost of the memory access is probably much smaller
that the io access, so really fancy copies are unlikely make much
difference unless the width of the io access changes.

	David