Re: [virtio-dev] Re: [PATCH v2 4/5] virtio-mmio: add MSI interrupt feature support

[Michael S. Tsirkin]

On Tue, Feb 11, 2020 at 08:18:54PM +0800, Jason Wang wrote:
> 
> On 2020/2/11 äå8:08, Michael S. Tsirkin wrote:
> > On Tue, Feb 11, 2020 at 08:04:24PM +0800, Jason Wang wrote:
> > > On 2020/2/11 äå7:58, Michael S. Tsirkin wrote:
> > > > On Tue, Feb 11, 2020 at 03:40:23PM +0800, Jason Wang wrote:
> > > > > On 2020/2/11 äå2:02, Liu, Jing2 wrote:
> > > > > > On 2/11/2020 12:02 PM, Jason Wang wrote:
> > > > > > > On 2020/2/11 äå11:35, Liu, Jing2 wrote:
> > > > > > > > On 2/11/2020 11:17 AM, Jason Wang wrote:
> > > > > > > > > On 2020/2/10 äå5:05, Zha Bin wrote:
> > > > > > > > > > From: Liu Jiang<gerry@xxxxxxxxxxxxxxxxx>
> > > > > > > > > > 
> > > > > > > > > > Userspace VMMs (e.g. Qemu microvm, Firecracker) take
> > > > > > > > > > advantage of using
> > > > > > > > > > virtio over mmio devices as a lightweight machine model for modern
> > > > > > > > > > cloud. The standard virtio over MMIO transport layer
> > > > > > > > > > only supports one
> > > > > > > > > > legacy interrupt, which is much heavier than virtio over
> > > > > > > > > > PCI transport
> > > > > > > > > > layer using MSI. Legacy interrupt has long work path and
> > > > > > > > > > causes specific
> > > > > > > > > > VMExits in following cases, which would considerably slow down the
> > > > > > > > > > performance:
> > > > > > > > > > 
> > > > > > > > > > 1) read interrupt status register
> > > > > > > > > > 2) update interrupt status register
> > > > > > > > > > 3) write IOAPIC EOI register
> > > > > > > > > > 
> > > > > > > > > > We proposed to add MSI support for virtio over MMIO via new feature
> > > > > > > > > > bit VIRTIO_F_MMIO_MSI[1] which increases the interrupt performance.
> > > > > > > > > > 
> > > > > > > > > > With the VIRTIO_F_MMIO_MSI feature bit supported, the virtio-mmio MSI
> > > > > > > > > > uses msi_sharing[1] to indicate the event and vector mapping.
> > > > > > > > > > Bit 1 is 0: device uses non-sharing and fixed vector per
> > > > > > > > > > event mapping.
> > > > > > > > > > Bit 1 is 1: device uses sharing mode and dynamic mapping.
> > > > > > > > > I believe dynamic mapping should cover the case of fixed vector?
> > > > > > > > > 
> > > > > > > > Actually this bit*aims*  for msi sharing or msi non-sharing.
> > > > > > > > 
> > > > > > > > It means, when msi sharing bit is 1, device doesn't want vector
> > > > > > > > per queue
> > > > > > > > 
> > > > > > > > (it wants msi vector sharing as name) and doesn't want a high
> > > > > > > > interrupt rate.
> > > > > > > > 
> > > > > > > > So driver turns to !per_vq_vectors and has to do dynamical mapping.
> > > > > > > > 
> > > > > > > > So they are opposite not superset.
> > > > > > > > 
> > > > > > > > Thanks!
> > > > > > > > 
> > > > > > > > Jing
> > > > > > > I think you need add more comments on the command.
> > > > > > > 
> > > > > > > E.g if I want to map vector 0 to queue 1, how do I need to do?
> > > > > > > 
> > > > > > > write(1, queue_sel);
> > > > > > > write(0, vector_sel);
> > > > > > That's true. Besides, two commands are used for msi sharing mode,
> > > > > > 
> > > > > > VIRTIO_MMIO_MSI_CMD_MAP_CONFIG and VIRTIO_MMIO_MSI_CMD_MAP_QUEUE.
> > > > > > 
> > > > > > "To set up the event and vector mapping for MSI sharing mode, driver
> > > > > > SHOULD write a valid MsiVecSel followed by
> > > > > > VIRTIO_MMIO_MSI_CMD_MAP_CONFIG/VIRTIO_MMIO_MSI_CMD_MAP_QUEUE command to
> > > > > > map the configuration change/selected queue events respectively. " (See
> > > > > > spec patch 5/5)
> > > > > > 
> > > > > > So if driver detects the msi sharing mode, when it does setup vq, writes
> > > > > > the queue_sel (this already exists in setup vq), vector sel and then
> > > > > > MAP_QUEUE command to do the queue event mapping.
> > > > > > 
> > > > > So actually the per vq msix could be done through this. I don't get why you
> > > > > need to introduce MSI_SHARING_MASK which is the charge of driver instead of
> > > > > device. The interrupt rate should have no direct relationship with whether
> > > > > it has been shared or not.
> > > > > 
> > > > > Btw, you introduce mask/unmask without pending, how to deal with the lost
> > > > > interrupt during the masking then?
> > > > pending can be an internal device register. as long as device
> > > > does not lose interrupts while masked, all's well.
> > > 
> > > You meant raise the interrupt during unmask automatically?
> > > 
> > 
> > yes - that's also what pci does.
> > 
> > the guest visible pending bit is partially implemented in qemu
> > as per pci spec but it's unused.
> 
> 
> Ok.
> 
> 
> > 
> > > > There's value is being able to say "this queue sends no
> > > > interrupts do not bother checking used notification area".
> > > > so we need way to say that. So I guess an enable interrupts
> > > > register might have some value...
> > > > But besides that, it's enough to have mask/unmask/address/data
> > > > per vq.
> > > 
> > > Just to check, do you mean "per vector" here?
> > > 
> > > Thanks
> > > 
> > No, per VQ. An indirection VQ -> vector -> address/data isn't
> > necessary for PCI either, but that ship has sailed.
> 
> 
> Yes, it can work but it may bring extra effort when you want to mask a
> vector which is was shared by a lot of queues.
> 
> Thanks
> 

masking should be per vq too.

-- 
MST