Re: [PATCH 2/5] Introduce VIRTIO_F_ADMIN_VQ_INDIRECT_DESC/VIRTIO_F_ADMIN_VQ_IN_ORDER

["Michael S. Tsirkin" <mst@redhat.com>]

On Tue, Jan 25, 2022 at 12:59:16PM +0200, Max Gurtovoy wrote:
> 
> On 1/25/2022 5:52 AM, Parav Pandit wrote:
> > Hi Jason,
> > 
> > > From: Jason Wang <jasowang@redhat.com>
> > > Sent: Tuesday, January 25, 2022 8:59 AM
> > > 
> > > 在 2022/1/19 下午12:48, Parav Pandit 写道:
> > > > > From: Jason Wang <jasowang@redhat.com>
> > > > > Sent: Wednesday, January 19, 2022 9:33 AM
> > > > > 
> > > > > 
> > > > > It it means IMS, there's already a proposal[1] that introduce MSI
> > > > > commands via the admin virtqueue. And we had similar requirement for
> > > > > virtio-MMIO[2] and managed device or SF [3], so I would rather to
> > > > > introduce IMS (need a better name though) as a basic facility instead
> > > > > of tie it to any specific transport.
> > > > > 
> > > > IMS of [1] is a interrupt configuration by the virtio driver for the device is it
> > > driving, which needs a queue.
> > > > So regardless of the device type as PCI PF/VF/SF/ADI, there is desire to have a
> > > generic admin queue not attached to device type.
> > > > And AQ in this proposal exactly serves this purpose.
> > > > 
> > > > Device configuring its own IMS vector vs PCI PF configuring VF's MSI-X max
> > > vector count are two different functionality.
> > > > Both of these commands can ride on a generic queue.
> > > > However the queue is not same, because PF owns its own admin queue
> > > > (for vf msix config), VF or SF operates its own admin queue (for IMS
> > > > config).
> > > 
> > > So I think in the next version we need to clarify:
> > > 
> > > 1) is there a single admin virtqueue shared by all the VFs and PF
> > > 
> > > or
> > > 
> > > 2) per VF/PF admin virtqueue, and how does the driver know how to find the
> > > corresponding admin virtqueue
> > > 
> > Admin queue is not per VF.
> > Lets take concrete examples.
> > 1. So for example, PCI PF can have one AQ.
> > This AQ carries command to query/config MSI-X vector of VFs.
> > 
> > 2. In second example, PCI PF is creating/destroying SFs. This is again done by using the AQ of the PCI PF.
> > 
> > 3. A PCI VF has its own AQ to configure some of its own generic attribute, don't know which is that today.
> > May be something that is extremely hard to do over features bit.
> > Currently proposed v2 doesn't restrict admin queue to be within PCI PF or VF or that matter not limited to other transports.
> > > > So a good example is,
> > > > 1. PCI PF configures 8 MSI-X or 16 IMS vectors for the VF using PF_AQ in HV.
> > > > 2. PCI VF when using IMS configures, IMS data, vector, mask etc using VF_AQ
> > > in GVM.
> > > > Both the functions will have AQ feature bit set.
> > > 
> > > Where did the VF_AQ sit? I guess it belongs to the VF. But if this is
> > > true, don't we need some kind of address isolation like PASID?
> > > 
> > Above one for IMS is not a good example. I replied the reasoning last week for it.
> > > > Fair enough, so we have more users of admin queue than just MSI-X config.
> > > 
> > > Well, what I really meant is that we actually have more users of IMS.
> > > That's is exactly what virito-mmio wants. In this case introducing admin
> > > queue looks too heavyweight for that.
> > > 
> > IMS config cannot be done over AQ as described in previous email in this thread.
> > 
> > > > > > 2. AQ to follows IN_ORDER and INDIRECT_DESC negotiation like rest of
> > > > > > the queues 3. Update commit log to describe why config space is not
> > > > > > chosen (scale, on-die registers, uniform way to handle all aq cmds)
> > > > > I fail to understand the scale/registeres issues. With the one of my previous
> > > > > proposal (device selector), technically we don't even need any config space
> > > or
> > > > > BAR for VF or SF by multiplexing the registers for PF.
> > > > > 
> > > > Scale issue is: when you want to create, query, manipulate hundreds of
> > > objects, having shared MMIO register or configuration register, will be too
> > > slow.
> > > 
> > > 
> > > Ok, this need to be clarified in the commit log. And we need make sure
> > > it's not an issue that is only happen for some specific vendor.
> > It is present in the v2 commit log cover letter.
> > Please let me know if you think it should be in the actual patch commit log.
> > 
> > 
> > > > And additionally such register set doesn't scale to allow sharing large
> > > number of bytes as DMA cannot be done.
> > > 
> > > 
> > > That's true.
> > > 
> > > 
> > > >   From physical device perspective, it doesn’t scale because device needs to
> > > have those resources ready to answer on MMIO reads and for hundreds to
> > > thousand of devices it just cannot do it.
> > > > This is one of the reason for birth of IMS.
> > > 
> > > IMS allows the table to be stored in the memory and cached by the device
> > > to have the best scalability. But I had other questions:
> > > 
> > > 1) if we have a single admin virtqueue, there will still be contention
> > > in the driver side
> > > 
> > AQ inherently allows out of order commands execution.
> > It shouldn't face contention. For example 1K depth AQ should be serving hundreds of descriptors commands in parallel for SF creation, VF MSI-X config and more.
> > 
> > Which area/commands etc you think can lead to the contention?
> > > 2) if we have per vf admin virtqueue, it still doesn't scale since it
> > > occupies more hardware resources
> > > 
> > That is too heavy, and doesn’t scale. Proposal is to not have per vf admin queue.
> > Proposal is to have one admin queue in a virtio device.
> 
> Right ? where did we mention something that can imply otherwise ?
> 
> 
> > > > > I do see one advantage is that the admin virtqueue is transport
> > > independent
> > > > > (or it could be used as a transport).
> > > > > 
> > > > I am yet to read the transport part from [1].
> > > 
> > > Yes, the main goal is to be compatible with SIOV.
> > > 
> > Admin queue is a command interface transport where higher layer services can be buit.
> > This includes SR-IOV config, SIOV config.
> > And v2 enables SIOV commands implementation whenever they are ready.
> > 
> > > > > > 4. Improve documentation around msix config to link to sriov section of
> > > virtio
> > > > > spec
> > > > > > 5. Describe error that if VF is bound to the device, admin commands
> > > > > targeting VF can fail, describe this error code
> > > > > > Did I miss anything?
> > > > > > 
> > > > > > Yet to receive your feedback on group, if/why is it needed and, why/if it
> > > must
> > > > > be in this proposal, what pieces prevents it do as follow-on.
> > > > > > Cornelia, Jason,
> > > > > > Can you please review current proposal as well before we revise v2?
> > > > > If I understand correctly, most of the features (except for the admin
> > > > > virtqueue in_order stuffs) are not specific to the admin virtqueue. As
> > > > > discussed in the previous versions, I still think it's better:
> > > > > 
> > > > > 1) adding sections in the basic device facility or data structure for
> > > > > provisioning and MSI
> > > > > 2) introduce admin virtqueue on top as an device interface for those
> > > > > features
> > > > > 
> > > > I didn't follow your suggestion. Can you please explain?
> > > > Specifically "data structure for provisioning and MSI"..
> > > 
> > > I meant:
> > > 
> > > There's a chapter "Basic Facilities of a Virtio Device", we can
> > > introduce the concepts there like:
> > > 
> > > 1) Managed device and Management device (terminology proposed by
> > > Michael), and can use PF and VF as a example
> > > 
> > > 2) Managed device provisioning (the data structure to specify the
> > > attributes of a managed device (VF))
> > > 
> > > 3) MSI
> > > 
> > Above is good idea. Will revisit v2, if it is not arranged this way.
> 
> Let me make sure I understand, you would like to see a new chapter under
> "Basic Facilities of a Virtio Device" that is
> 
> called "Device management" and this chapter will explain in few words the
> concept and it will point to another chapter under "Basic Facilities of a
> Virtio Device"
> 
> that was introduced here "Admin Virtqueues" ?
> 
> So you do agree that managing a managed (create/destroy/setup/etc...) will
> be done using the AQ of the managing device ?

I think Jason asked that the management commands are split from the
queue itself, such that they can be implemented in more ways down the
road.

> > > And then we can introduced admin virtqueue in either
> > > 
> > > 1) transport part
> > > 
> > > or
> > > 
> > > 2) PCI transport
> > > 
> > It is not specific to PCI transport, and currently it is not a transport either.
> > So admin queue will keep as general entity for admin work.
> > > In the admin virtqueue, there will be commands to provision and
> > > configure MSI.
> > > 
> > Please review v2 if it is not arranged this way.
> > 
> > > > > The leaves the chance for future extensions to allow those features to
> > > > > be used by transport specific interface which will benefit for
> > > > > 
> > > > AQ allows communication (command, response) between driver and device
> > > in transport independent way.
> > > > Sometimes it query/set transport specific fields like MSI-X vectors of VF.
> > > > Sometimes device configure its on IMS interrupt.
> > > > Something else in future.
> > > > So it is really a generic request-response queue.
> > > 
> > > I agree, but I think we can't mandate new features to a specific transport.
> > > 
> > Certainly. Admin queue is transport independent.
> > PCI MSI-X configuration is PCI transport specific command, so structures are defined it accordingly.
> > It is similar to struct virtio_pci_cap, struct virtio_pci_common_cfg etc.
> > 
> > Any other transport will have transport specific interrupt configuration. So it will be defined accordingly whenever that occurs.
> > For example, IMS for VF or IMS for SF.