On Tue, Dec 09, 2014 at 03:47:44PM +0000, Peter Maydell wrote: > On 9 December 2014 at 15:41, Laszlo Ersek <ler...@redhat.com> wrote: > > Again, this was the idea that Rich had in 2010 (see the links in the > > discussion thus far). It was rejected back then (which is why I didn't > > even try to resurrect it now), and Peter has now asked if anything has > > changed that would make that approach more acceptable now. > > > > "Avi and Gleb not hanging around so they can't block the approach now" > > might or might not be such a change; I don't know. But, I tend to trust > > their opinion, even if dates back to 2010 in this case. > > Yeah, this is about my point of view. A direct-write-to-memory > fw_cfg doesn't seem too unreasonable to me, but it would be nice > to see a solidly argued case from its proponents to counterbalance > the previous rejection.
The argument for is that it's essentially instantaneous. This is a big deal for libguestfs where even our new initrd (1.5 MB) takes a noticable fraction of a second to load, and we want to get our total start-up time down to 3 seconds to match x86. > > Anyway I'm certainly not opposed to performance, so if someone can > > thoroughly refute everything they said in that thread, go ahead. > > As far as I can tell the main line of argument was "if you're > using fw_cfg to transfer huge amounts of data you're doing > something wrong". > > Is ARM much higher overhead than x86 for these accesses? On a slightly related topic, virtio-mmio traps are slow: https://bugzilla.redhat.com/show_bug.cgi?id=1123555 This also kills libguestfs performance -- throughput this time, rather than start-up time. Rich. -- Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones Read my programming and virtualization blog: http://rwmj.wordpress.com virt-top is 'top' for virtual machines. Tiny program with many powerful monitoring features, net stats, disk stats, logging, etc. http://people.redhat.com/~rjones/virt-top
