Rusty Russell [off-list ref] writes:
Anthony Liguori [off-list ref] writes:
quoted
Rusty Russell [off-list ref] writes:
quoted
"Michael S. Tsirkin" [off-list ref] writes:
quoted
Thinking about Sasha's patches, we can reduce ring usage
for virtio net small packets dramatically if we put
virtio net header inline with the data.
This can be done for free in case guest net stack allocated
extra head room for the packet, and I don't see
why would this have any downsides.
I've been wanting to do this for the longest time... but...
quoted
Even though with my recent patches qemu
no longer requires header to be the first s/g element,
we need a new feature bit to detect this.
A trivial qemu patch will be sent separately.
There's a reason I haven't done this. I really, really dislike "my
implemention isn't broken" feature bits. We could have an infinite
number of them, for each bug in each device.
This is a bug in the specification.
The QEMU implementation pre-dates the specification. All of the actual
implementations of virtio relied on the semantics of s/g elements and
still do.
lguest fix is pending in my queue. lkvm and qemu are broken; lkvm isn't
ever going to be merged, so I'm not sure what its status is? But I'm
determined to fix qemu, and hence my torture patch to make sure this
doesn't creep in again.
There are even more implementations out there and I'd wager they all
rely on framing.
quoted
What's in the specification really doesn't matter when it doesn't agree
with all of the existing implementations.
Users use implementations, not specifications. The specification really
ought to be changed here.
I'm sorely tempted, except that we're losing a real optimization because
of this :(
What optimizations? What Michael is proposing is still achievable with
a device feature. Are there other optimizations that can be achieved by
changing framing that we can't achieve with feature bits?
As I mentioned in another note, bad framing decisions can cause
performance issues too...
The specification has long contained the footnote:
The current qemu device implementations mistakenly insist that
the first descriptor cover the header in these cases exactly, so
a cautious driver should arrange it so.
I seem to recall this being a compromise between you and I.. I think
I objected strongly to this back when you first wrote the spec and you
added this to appease me ;-)
Regards,
Anthony Liguori
I'd like to tie this caveat to the PCI capability change, so this note
will move to the appendix with the old PCI layout.
Cheers,
Rusty.