Messages in this thread | | | Date | Thu, 27 Aug 2009 14:19:48 +0300 | From | "Michael S. Tsirkin" <> | Subject | Re: [PATCHv4 2/2] vhost_net: a kernel-level virtio server |
| |
On Thu, Aug 27, 2009 at 08:40:02PM +0930, Rusty Russell wrote: > On Thu, 27 Aug 2009 08:15:18 pm Michael S. Tsirkin wrote: > > On Tue, Aug 25, 2009 at 09:40:40PM +0930, Rusty Russell wrote: > > > Also, see other fixes to the lguest launcher since then which might > > > be relevant to this code: > > > lguest: get more serious about wmb() in example Launcher code > > > > Heh, this just gets one step closer to a real wmb. I just used the > > correct code from linux, so I think nothing needs to be done in vhost. > > Apropos this change in lguest: why is a compiler barrier sufficient? The > > comment says devices are run in separate threads (presumably from > > guest?), if so don't you need to tell CPU that there's a barrier as > > well? > > Yep, but x86 only :)
Okay, writes are ordered them. But mb() orders reads as well, so it'll have to have a real one IMO, even in userspace?
> The kernel uses a real insn if XMM/XMM2, but I don't > know if userspace needs that. I just use compiler barriers. > > Thanks, > Rusty.
| |