Re: [PATCH 2/3] x86_64: Define 128-bit memory-mapped I/O operations

On Wed, 2012-08-22 at 11:28 -0700, Linus Torvalds wrote:

On Wed, Aug 22, 2012 at 11:11 AM, Ben Hutchings
[off-list ref] wrote:

quoted

Right, I think it's been made pretty clear that it's going to be
dependent on more than just architecture.

Well, it's entirely possible that the 128-bit case will work correctly
on all x86-64 hardware out there on PCIe.

I just can't guarantee it, because we certainly have had issues with
hw doing odd things before. But maybe PCIe really is well-specified
enough, and maybe nobody has done a odd PCIe bridges, and maybe every
time some 128-bit access is split, the bus in question still always
remembers the original 128-bit size in the transaction. It's not at
all impossible. I just wouldn't *guarantee* it.

Even then, everything works out OK if the particular MMIO write I'm
concerned about *is* split - just as long as the resulting operations
are in ascending address order.  This is not true on all systems if we
enable write-combining (without a fence in the middle, which defeats the
purpose), and I think hpa was saying that it may not be the case with
SSE writes either.

And to some degree, for high-end server-only hardware in particular,
it really *is* acceptable to say "If you have odd hardware, odd things
will happen". So for this particular driver, maybe the right approach
is simply to say "we require that your fabric works right". And see if
anybody ever complains.

Maybe.  At the moment reordering tends to cause the hardware to complain
that we sent an invalid sequence of DMA descriptors, but that's only
because we're not being as smart as we could about using TX push.  I
don't want to run the risk of sending out corrupted packets (with
offloaded checksums, so they're not that obviously invalid) on the wire.

Ben.

The 100ns may be worth those kinds of "you'd better not have old/crap
hardware" decisions. It's not acceptable for some drivers (a driver
for some consumer ATA chip might not want to make that kind of choice,
and say "whatever, we'll be really conservative), but "Quod licet
Jovi, non licet bovi".

The fact that something might not be *guaranteed* to always work
doesn't necessarily mean that it is always the wrong thing to do..

-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.