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

[Jason Wang]

On 2020/2/12 上午11:54, Liu, Jing2 wrote:
> On 2/11/2020 3:40 PM, Jason Wang wrote:
> > On 2020/2/11 下午2:02, Liu, Jing2 wrote:
> > > On 2/11/2020 12:02 PM, Jason Wang wrote:
> > > > On 2020/2/11 上午11:35, Liu, Jing2 wrote:
> > > > > On 2/11/2020 11:17 AM, Jason Wang wrote:
> > > > > > On 2020/2/10 下午5:05, Zha Bin wrote:
> > > > > > > From: Liu Jiang<address@hidden>
> > > > > > >
> > > > > > > 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. 
>
> Right, per vq msix can also be mapped by the 2 commands if we want.
>
> The current design benefits for those devices requesting per vq msi 
> that driver
>
> doesn't have to map during setup each queue,
>
> since we define the relationship by default.

Well since you've defined the dynamic mapping, having some "default" 
mapping won't help to reduce the complexity but increase it.


> > I don't get why you need to introduce MSI_SHARING_MASK which is the 
> > charge of driver instead of device. 
>
> MSI_SHARING_MASK is just for identifying the msi_sharing bit in 
> readl(MsiState) (0x0c4). The device tells whether it wants msi_sharing.
>
> MsiState register: R
>
> le32 {
>     msi_enabled : 1;
>     msi_sharing: 1;
>     reserved : 30;
> };

The question is why device want such information.


> > 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?
> >
> >
> > > For msi non-sharing mode, no special action is needed because we 
> > > make the rule of per_vq_vector and fixed relationship.
> > >
> > > Correct me if this is not that clear for spec/code comments.
> >
> > The ABI is not as straightforward as PCI did. Why not just reuse the 
> > PCI layout?
> >
> > E.g having
> >
> > queue_sel
> > queue_msix_vector
> > msix_config
> >
> > for configuring map between msi vector and queues/config
>
> Thanks for the advice. :)
>
> Actually when looking into pci, the queue_msix_vector/msix_config is 
> the msi vector index, which is the same as the mmio register MsiVecSel 
> (0x0d0).
>
> So we don't introduce two extra registers for mapping even in sharing 
> mode.
>
> What do you think?

I'm not sure I get the point, but I still prefer the separate vector_sel 
from queue_msix_vector.

Btw, Michael propose per vq registers which could also work.

Thanks


> > Then
> >
> > vector_sel
> > address
> > data
> > pending
> > mask
> > unmask
> >
> > for configuring msi table?
>
> PCI-like msix table is not introduced to device and instead simply use 
> commands to tell the mask/configure/enable.
>
> Thanks!
>
> Jing
>
> > Thanks
> >
> >
> > > Thanks!
> > >
> > > Jing
> > >
> > >
> > > > ?
> > > >
> > > > Thanks
> > > >
> > > >
> > > > > > Thanks
> > > > > >
> > > > > >
> > > > > >
> > > > > > --------------------------------------------------------------------- 
> > > > > >
> > > > > > To unsubscribe, e-mail: address@hidden
> > > > > > For additional commands, e-mail: 
> > > > > > address@hidden