Re: [PATCH RFC] virtio-pci: new config layout: using memory BAR

[Anthony Liguori <aliguori@us.ibm.com>]

"Michael S. Tsirkin" <mst@redhat.com> writes:

> On Wed, Jun 05, 2013 at 01:57:16PM -0500, Anthony Liguori wrote:
>> "Michael S. Tsirkin" <mst@redhat.com> writes:
>> 
>> > On Wed, Jun 05, 2013 at 10:46:15AM -0500, Anthony Liguori wrote:
>> >> Look, it's very simple.
>> > We only need to do it if we do a change that breaks guests.
>> >
>> > Please find a guest that is broken by the patches. You won't find any.
>> 
>> I think the problem in this whole discussion is that we're talking past
>> each other.
>> 
>> Here is my understanding:
>> 
>> 1) PCI-e says that you must be able to disable IO bars and still have a
>> functioning device.
>> 
>> 2) It says (1) because you must size IO bars to 4096 which means that
>> practically speaking, once you enable a dozen or so PIO bars, you run
>> out of PIO space (16 * 4k == 64k and not all that space can be used).
>
>
> Let me add 3 other issues which I mentioned and you seem to miss:
>
> 3) architectures which don't have fast access to IO ports, exist
>    virtio does not work there ATM

Which architectures have PCI but no IO ports?

> 4) setups with many PCI bridges exist and have the same issue
>    as PCI express. virtio does not work there ATM

This is not virtio specific.  This is true for all devices that use IO.

> 5) On x86, even with nested page tables, firmware only decodes
>    the page address on an invalid PTE, not the data. You need to
>    emulate the guest to get at the data. Without
>    nested page tables, we have to do page table walk and emulate
>    to get both address and data. Since this is how MMIO
>    is implemented in kvm on x86, MMIO is much slower than PIO
>    (with nested page tables by a factor of >2, did not test without).

Am well aware of this, this is why we use PIO.

I fully agree with you that when we do MMIO, we should switch the
notification mechanism to avoid encoding anything meaningful as data.

>> virtio-pci uses a IO bars exclusively today.  Existing guest drivers
>> assume that there is an IO bar that contains the virtio-pci registers.
>> So let's consider the following scenarios:
>> 
>> QEMU of today:
>> 
>> 1) qemu -drive file=ubuntu-13.04.img,if=virtio
>> 
>> This works today.  Does adding an MMIO bar at BAR1 break this?
>> Certainly not if the device is behind a PCI bus...
>> 
>> But are we going to put devices behind a PCI-e bus by default?  Are we
>> going to ask the user to choose whether devices are put behind a legacy
>> bus or the express bus?
>> 
>> What happens if we put the device behind a PCI-e bus by default?  Well,
>> it can still work.  That is, until we do something like this:
>> 
>> 2) qemu -drive file=ubuntu-13.04.img,if=virtio -device virtio-rng
>>         -device virtio-balloon..
>> 
>> Such that we have more than a dozen or so devices.  This works
>> perfectly fine today.  It works fine because we've designed virtio to
>> make sure it works fine.  Quoting the spec:
>> 
>> "Configuration space is generally used for rarely-changing or
>>  initialization-time parameters. But it is a limited resource, so it
>>  might be better to use a virtqueue to update configuration information
>>  (the network device does this for filtering, otherwise the table in the
>>  config space could potentially be very large)."
>> 
>> In fact, we can have 100s of PCI devices today without running out of IO
>> space because we're so careful about this.
>> 
>> So if we switch to using PCI-e by default *and* we keep virtio-pci
>> without modifying the device IDs, then very frequently we are going to
>> break existing guests because the drivers they already have no longer
>> work.
>> 
>> A few virtio-serial channels, a few block devices, a couple of network
>> adapters, the balloon and RNG driver, and we hit the IO space limit
>> pretty damn quickly so this is not a contrived scenario at all.  I would
>> expect that we frequently run into this if we don't address this problem.
>> 
>> So we have a few options:
>> 1) Punt all of this complexity to libvirt et al and watch people make
>>    the wrong decisions about when to use PCI-e.  This will become yet
>>    another example of KVM being too hard to configure.
>> 
>> 2) Enable PCI-e by default and just force people to upgrade their
>>    drivers.
>> 
>> 3) Don't use PCI-e by default but still add BAR1 to virtio-pci
>> 
>> 4) Do virtio-pcie, make it PCI-e friendly (drop the IO BAR completely),
>
> We can't do this - it will hurt performance.

Can you explain?  I thought the whole trick with separating out the
virtqueue notification register was to regain the performance?

>>    give
>>    it a new device/vendor ID.   Continue to use virtio-pci for existing
>>    devices potentially adding virtio-{net,blk,...}-pcie variants for
>>    people that care to use them.
>> 
>> I think 1 == 2 == 3 and I view 2 as an ABI breaker.
>
> Why do you think 2 == 3? 2 changes default behaviour. 3 does not.

It doesn't change the default behavior but then we're pushing the
decision of when to use pci-e to the user.  They have to understand that
there can be subtle breakages because the virtio-pci driver may not work
if they are using an old guest.

>> libvirt does like
>> policy so they're going to make a simple decision and always use the
>> same bus by default.  I suspect if we made PCI the default, they might
>> just always set the PCI-e flag just because.
>
> This sounds very strange. But let's assume you are right for
> the sake of the argument ...
>
>> There are hundreds of thousands if not millions of guests with existing
>> virtio-pci drivers.  Forcing them to upgrade better have an extremely
>> good justification.
>> 
>> I think 4 is the best path forward.  It's better for users (guests
>> continue to work as they always have).  There's less confusion about
>> enabling PCI-e support--you must ask for the virtio-pcie variant and you
>> must have a virtio-pcie driver.  It's easy to explain.
>
> I don't think how this changes the situation. libvirt still need
> to set policy and decide which device to use.

But virtio-pcie never exhausts the IO configuration space.  That's the
difference.

And virtio-pcie is a separate driver so presumably libvirt will make
that visible in the XML.  In fact, it should.

>> It also maps to what regular hardware does.  I highly doubt that there
>> are any real PCI cards that made the shift from PCI to PCI-e without
>> bumping at least a revision ID.
>
> Only because the chance it's 100% compatible on the software level is 0.
> It always has some hardware specific quirks.
> No such excuse here.
>
>> It also means we don't need to play games about sometimes enabling IO
>> bars and sometimes not.
>
> This last paragraph is wrong, it ignores the issues 3) to 5) 
> I added above.
>
> If you do take them into account:
> 	- there are reasons to add MMIO BAR to PCI,
> 	  even without PCI express

So far, the only reason you've provided is "it doesn't work on some
architectures."  Which architectures?

> 	- we won't be able to drop IO BAR from virtio

An IO BAR is useless if it means we can't have more than 12 devices.

>
>> Regards,
>> 
>> Anthony Liguori
>> 
>> >
>> >
>> > -- 
>> > MST
>> > --
>> > To unsubscribe from this list: send the line "unsubscribe kvm" in
>> > the body of a message to majordomo@vger.kernel.org
>> > More majordomo info at  http://vger.kernel.org/majordomo-info.html
> --
> To unsubscribe from this list: send the line "unsubscribe kvm" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html