On Wed, 2009-05-13 at 15:35 +0300, Avi Kivity wrote: > Alex Williamson wrote: > > On Wed, 2009-05-13 at 12:47 +0300, Avi Kivity wrote: > > > >> Alex Williamson wrote: > >> > >>> We're currently using a counter to track the most recent GSI we've > >>> handed out. This quickly hits KVM_MAX_IRQ_ROUTES when using device > >>> assignment with a driver that regularly toggles the MSI enable bit. > >>> This can mean only a few minutes of usable run time. Instead, track > >>> used GSIs in a bitmap. > >>> > >>> Signed-off-by: Alex Williamson <alex.william...@hp.com> > >>> --- > >>> > >>> v2: Added mutex to protect gsi bitmap > >>> > >>> > >> Why is the mutex needed? We already have mutex protection in qemu. > >> > > > > If it's unneeded, I'll happily remove it. I was assuming in a guest > > with multiple devices these could come in parallel, but maybe the guest > > is already serialized for config space accesses via cfc/cf8. > > > > The guest may or may not be serialized; we can't rely on that. But qemu > is, and we can.
Ok, I'll drop the mutex here. > >> How often does the driver enable/disable the MSI (and, do you now why)? > >> If it's often enough it may justify kernel support. (We'll need this > >> patch in any case for kernels without this new support). > >> > > > > Seems like multiple times per second. I don't know why. Now I'm > > starting to get curious why nobody else seems to be hitting this. I'm > > seeing it on an e1000e NIC and Qlogic fibre channel. Is everyone else > > using MSI-X or regular interrupts vs MSI? > > > > When you say "multiple times", it is several, or a lot more? > > Maybe it is NAPI? The system would run out of the ~1000 available GSIs in a minute or two with just an e1000e available to the guest. So that's something on the order of 10/s. This also causes a printk in the host ever time the interrupt in enabled, which can't help performance and gets pretty annoying for syslog. I was guessing some kind of interrupt mitigation, such as NAPI, but a qlogic FC card seems to do it too (seemingly at a slower rate). Alex -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
