On Thu, Jan 24, 2019 at 4:36 AM Jarkko Sakkinen
[off-list ref] wrote:

quoted

Is it just that this particular hardware always happened to trigger
the ERMS case (ie "rep movsb")?

This is the particular snippet in question:

memcpy_fromio(buf, priv->rsp, 6);
expected = be32_to_cpup((__be32 *) &buf[2]);
if (expected > count || expected < 6)
        return -EIO;

Ok, strange.

So what *used* to happen is that the memcpy_fromio() would just expand
as a "memcpy()", and in this case, gcc would then inline the memcpy().
In fact, gcc does it as a 4-byte access and a two-byte access from
what I can tell.

Which is actually exactly the same as memcpy_fromio() should do, just
using a different code sequence.

memcpy_fromio(&buf[6], &priv->rsp[6], expected - 6);

This one gets turned into an out-of-line "memcpy()" in the old world
order, which depending on size will do different things, but might be
a "rep movsb". Or it might be the software expansion that does
overlapping accesses and/or backwards copies.

In the new world order, it's the "memcpy_fromio()" that willdo first
4-byte accesses for the main bulk of the copy, and then end up with a
two-byte and single-byte move to pad out the end.

I guess it did in the first memcpy_fromio operation since it is less
than a quad word, right? Not sure why the 2nd memcpy_fromio() operation
has worked, though.

The first one seems to do the same thing now as it used to do, so I
don't *think* it should have mattered.

The second one looks like it is unaligned (offset 6) and doing the
4-byte io reads would fail if that device needs aligned accesses. The
old memcpy() *might* have done it with a "rep movsb" that would just
work (?).

                 Linus