On Mon, Dec 07, 2009 at 12:24:02PM +0100, Gerd Hoffmann wrote: > On 12/07/09 11:59, Michael S. Tsirkin wrote: >>> The motivation was to move them away from the ioapic, to reduce irq >>> sharing of *other* devices which are connected to the ioapic too (i.e. >>> usb, e1000, lsi, ...) >> >> Let's convert these to MSI instead? >> This will likely pay off long term, >> e.g. with nested virtualization. > > Works only of the real hardware we emulate can do MSI-X too, otherwise > guests simply wouldn't use it.
Sure. Naturally, improving IRQ routing so that e.g. lsi and usb do not share an interrupt would also be a good idea regardless. Also - why not simply use virtio? I assume you are talking about guests with virtio support otherwise MSI support in virtio won't matter for them. > I think the only case where this could > work out is e1000, newer revisions can do MSI-X. Maybe also the > upcoming megasas emulation Hannes is working on. Hmm. I would expect high-end storage to support MSI-X as well. Could you point me to linux drivers for devices we emulate so that I can check? >>> I'm aware that these are not performance-critical. I've even tried to >>> use 'vectors=1' because of that. I expected that would make them use >>> MSI-X, but a single IRQ line only. Didn't work though. Intentional? >> >> So it's even worse, we are using up 2 vectors per device? Ugh ... > > Yes. Three for balloon, but that can easily changed to two. Yes, please, make this change for 0.12. >> no way to distinguish between vq interrupt >> and config change. This last thing is important >> because it allows fastpath injection of MSI >> interrupts directly from kernel without >> notifying qemu to update IRQ field. > > Ah, *that* is the reason for the separate config interrupt. Does the > in-kernel injection matter for balloon+console? No, but changing this will need updating guests. And if we do update guests anyway, I would suggest that we put IRQ field in guest memory, with an atomic set, and guest would get it with test and clear, so that we get a generic interface and not something specific for console/baloon. Naturally, this needs a feature bit, and it won't happen in 0.12/2.6.33 timeframe. > I'd expect only > virtio-net needs that when it is configured with vhost? > > cheers > Gerd Any other device will need the same if it has a kernel backend. -- MST
