kvm PCI assignment & VFIO ramblings

[Benjamin Herrenschmidt]

Hi folks !

So I promised Anthony I would try to summarize some of the comments &
issues we have vs. VFIO after we've tried to use it for PCI pass-through
on POWER. It's pretty long, there are various items with more or less
impact, some of it is easily fixable, some are API issues, and we'll
probably want to discuss them separately, but for now here's a brain
dump.

David, Alexei, please make sure I haven't missed anything :-)

* Granularity of pass-through

So let's first start with what is probably the main issue and the most
contentious, which is the problem of dealing with the various
constraints which define the granularity of pass-through, along with
exploiting features like the VTd iommu domains.

For the sake of clarity, let me first talk a bit about the "granularity"
issue I've mentioned above.

There are various constraints that can/will force several devices to be
"owned" by the same guest and on the same side of the host/guest
boundary. This is generally because some kind of HW resource is shared
and thus not doing so would break the isolation barrier and enable a
guest to disrupt the operations of the host and/or another guest.

Some of those constraints are well know, such as shared interrupts. Some
are more subtle, for example, if a PCIe->PCI bridge exist in the system,
there is no way for the iommu to identify transactions from devices
coming from the PCI segment of that bridge with a granularity other than
"behind the bridge". So typically a EHCI/OHCI/OHCI combo (a classic)
behind such a bridge must be treated as a single "entity" for
pass-trough purposes.

In IBM POWER land, we call this a "partitionable endpoint" (the term
"endpoint" here is historic, such a PE can be made of several PCIe
"endpoints"). I think "partitionable" is a pretty good name tho to
represent the constraints, so I'll call this a "partitionable group"
from now on. 

Other examples of such HW imposed constraints can be a shared iommu with
no filtering capability (some older POWER hardware which we might want
to support fall into that category, each PCI host bridge is its own
domain but doesn't have a finer granularity... however those machines
tend to have a lot of host bridges :)

If we are ever going to consider applying some of this to non-PCI
devices (see the ongoing discussions here), then we will be faced with
the crazyness of embedded designers which probably means all sort of new
constraints we can't even begin to think about

This leads me to those initial conclusions:

- The -minimum- granularity of pass-through is not always a single
device and not always under SW control

- Having a magic heuristic in libvirt to figure out those constraints is
WRONG. This reeks of XFree 4 PCI layer trying to duplicate the kernel
knowledge of PCI resource management and getting it wrong in many many
cases, something that took years to fix essentially by ripping it all
out. This is kernel knowledge and thus we need the kernel to expose in a
way or another what those constraints are, what those "partitionable
groups" are.

- That does -not- mean that we cannot specify for each individual device
within such a group where we want to put it in qemu (what devfn etc...).
As long as there is a clear understanding that the "ownership" of the
device goes with the group, this is somewhat orthogonal to how they are
represented in qemu. (Not completely... if the iommu is exposed to the
guest ,via paravirt for example, some of these constraints must be
exposed but I'll talk about that more later).

The interface currently proposed for VFIO (and associated uiommu)
doesn't handle that problem at all. Instead, it is entirely centered
around a specific "feature" of the VTd iommu's for creating arbitrary
domains with arbitrary devices (tho those devices -do- have the same
constraints exposed above, don't try to put 2 legacy PCI devices behind
the same bridge into 2 different domains !), but the API totally ignores
the problem, leaves it to libvirt "magic foo" and focuses on something
that is both quite secondary in the grand scheme of things, and quite
x86 VTd specific in the implementation and API definition.

Now, I'm not saying these programmable iommu domains aren't a nice
feature and that we shouldn't exploit them when available, but as it is,
it is too much a central part of the API.

I'll talk a little bit more about recent POWER iommu's here to
illustrate where I'm coming from with my idea of groups:

On p7ioc (the IO chip used on recent P7 machines), there -is- a concept
of domain and a per-RID filtering. However it differs from VTd in a few
ways:

The "domains" (aka PEs) encompass more than just an iommu filtering
scheme. The MMIO space and PIO space are also segmented, and those
segments assigned to domains. Interrupts (well, MSI ports at least) are
assigned to domains. Inbound PCIe error messages are targeted to
domains, etc...

Basically, the PEs provide a very strong isolation feature which
includes errors, and has the ability to immediately "isolate" a PE on
the first occurence of an error. For example, if an inbound PCIe error
is signaled by a device on a PE or such a device does a DMA to a
non-authorized address, the whole PE gets into error state. All
subsequent stores (both DMA and MMIO) are swallowed and reads return all
1's, interrupts are blocked. This is designed to prevent any propagation
of bad data, which is a very important feature in large high reliability
systems.

Software then has the ability to selectively turn back on MMIO and/or
DMA, perform diagnostics, reset devices etc...

Because the domains encompass more than just DMA, but also segment the
MMIO space, it is not practical at all to dynamically reconfigure them
at runtime to "move" devices into domains. The firmware or early kernel
code (it depends) will assign devices BARs using an algorithm that keeps
them within PE segment boundaries, etc....

Additionally (and this is indeed a "restriction" compared to VTd, though
I expect our future IO chips to lift it to some extent), PE don't get
separate DMA address spaces. There is one 64-bit DMA address space per
PCI host bridge, and it is 'segmented' with each segment being assigned
to a PE. Due to the way PE assignment works in hardware, it is not
practical to make several devices share a segment unless they are on the
same bus. Also the resulting limit in the amount of 32-bit DMA space a
device can access means that it's impractical to put too many devices in
a PE anyways. (This is clearly designed for paravirt iommu, I'll talk
more about that later).

The above essentially extends the granularity requirement (or rather is
another factor defining what the granularity of partitionable entities
is). You can think of it as "pre-existing" domains.

I believe the way to solve that is to introduce a kernel interface to
expose those "partitionable entities" to userspace. In addition, it
occurs to me that the ability to manipulate VTd domains essentially
boils down to manipulating those groups (creating larger ones with
individual components).

I like the idea of defining / playing with those groups statically
(using a command line tool or sysfs, possibly having a config file
defining them in a persistent way) rather than having their lifetime
tied to a uiommu file descriptor.

It also makes it a LOT easier to have a channel to manipulate
platform/arch specific attributes of those domains if any.

So we could define an API or representation in sysfs that exposes what
the partitionable entities are, and we may add to it an API to
manipulate them. But we don't have to and I'm happy to keep the
additional SW grouping you can do on VTd as a sepparate "add-on" API
(tho I don't like at all the way it works with uiommu). However, qemu
needs to know what the grouping is regardless of the domains, and it's
not nice if it has to manipulate two different concepts here so
eventually those "partitionable entities" from a qemu standpoint must
look like domains.

My main point is that I don't want the "knowledge" here to be in libvirt
or qemu. In fact, I want to be able to do something as simple as passing
a reference to a PE to qemu (sysfs path ?) and have it just pickup all
the devices in there and expose them to the guest.

This can be done in a way that isn't PCI specific as well (the
definition of the groups and what is grouped would would obviously be
somewhat bus specific and handled by platform code in the kernel).

Maybe something like /sys/devgroups ? This probably warrants involving
more kernel people into the discussion.

* IOMMU

Now more on iommu. I've described I think in enough details how ours
work, there are others, I don't know what freescale or ARM are doing,
sparc doesn't quite work like VTd either, etc...

The main problem isn't that much the mechanics of the iommu but really
how it's exposed (or not) to guests.

VFIO here is basically designed for one and only one thing: expose the
entire guest physical address space to the device more/less 1:1.

This means:

  - It only works with iommu's that provide complete DMA address spaces
to devices. Won't work with a single 'segmented' address space like we
have on POWER.

  - It requires the guest to be pinned. Pass-through -> no more swap

  - The guest cannot make use of the iommu to deal with 32-bit DMA
devices, thus a guest with more than a few G of RAM (I don't know the
exact limit on x86, depends on your IO hole I suppose), and you end up
back to swiotlb & bounce buffering.

  - It doesn't work for POWER server anyways because of our need to
provide a paravirt iommu interface to the guest since that's how pHyp
works today and how existing OSes expect to operate.

Now some of this can be fixed with tweaks, and we've started doing it
(we have a working pass-through using VFIO, forgot to mention that, it's
just that we don't like what we had to do to get there).

Basically, what we do today is:

- We add an ioctl to VFIO to expose to qemu the segment information. IE.
What is the DMA address and size of the DMA "window" usable for a given
device. This is a tweak, that should really be handled at the "domain"
level.

That current hack won't work well if two devices share an iommu. Note
that we have an additional constraint here due to our paravirt
interfaces (specificed in PAPR) which is that PE domains must have a
common parent. Basically, pHyp makes them look like a PCIe host bridge
per domain in the guest. I think that's a pretty good idea and qemu
might want to do the same.

- We hack out the currently unconditional mapping of the entire guest
space in the iommu. Something will have to be done to "decide" whether
to do that or not ... qemu argument -> ioctl ?

- We hook up the paravirt call to insert/remove a translation from the
iommu to the VFIO map/unmap ioctl's.

This limps along but it's not great. Some of the problems are:

- I've already mentioned, the domain problem again :-) 

- Performance sucks of course, the vfio map ioctl wasn't mean for that
and has quite a bit of overhead. However we'll want to do the paravirt
call directly in the kernel eventually ...

  - ... which isn't trivial to get back to our underlying arch specific
iommu object from there. We'll probably need a set of arch specific
"sideband" ioctl's to "register" our paravirt iommu "bus numbers" and
link them to the real thing kernel-side.

- PAPR (the specification of our paravirt interface and the expectation
of current OSes) wants iommu pages to be 4k by default, regardless of
the kernel host page size, which makes things a bit tricky since our
enterprise host kernels have a 64k base page size. Additionally, we have
new PAPR interfaces that we want to exploit, to allow the guest to
create secondary iommu segments (in 64-bit space), which can be used
(under guest control) to do things like map the entire guest (here it
is :-) or use larger iommu page sizes (if permitted by the host kernel,
in our case we could allow 64k iommu page size with a 64k host kernel).

The above means we need arch specific APIs. So arch specific vfio
ioctl's, either that or kvm ones going to vfio or something ... the
current structure of vfio/kvm interaction doesn't make it easy.

* IO space

On most (if not all) non-x86 archs, each PCI host bridge provide a
completely separate PCI address space. Qemu doesn't deal with that very
well. For MMIO it can be handled since those PCI address spaces are
"remapped" holes in the main CPU address space so devices can be
registered by using BAR + offset of that window in qemu MMIO mapping.

For PIO things get nasty. We have totally separate PIO spaces and qemu
doesn't seem to like that. We can try to play the offset trick as well,
we haven't tried yet, but basically that's another one to fix. Not a
huge deal I suppose but heh ...

Also our next generation chipset may drop support for PIO completely.

On the other hand, because PIO is just a special range of MMIO for us,
we can do normal pass-through on it and don't need any of the emulation
done qemu.

  * MMIO constraints

The QEMU side VFIO code hard wires various constraints that are entirely
based on various requirements you decided you have on x86 but don't
necessarily apply to us :-)

Due to our paravirt nature, we don't need to masquerade the MSI-X table
for example. At all. If the guest configures crap into it, too bad, it
can only shoot itself in the foot since the host bridge enforce
validation anyways as I explained earlier. Because it's all paravirt, we
don't need to "translate" the interrupt vectors & addresses, the guest
will call hyercalls to configure things anyways.

We don't need to prevent MMIO pass-through for small BARs at all. This
should be some kind of capability or flag passed by the arch. Our
segmentation of the MMIO domain means that we can give entire segments
to the guest and let it access anything in there (those segments are a
multiple of the page size always). Worst case it will access outside of
a device BAR within a segment and will cause the PE to go into error
state, shooting itself in the foot, there is no risk of side effect
outside of the guest boundaries.

In fact, we don't even need to emulate BAR sizing etc... in theory. Our
paravirt guests expect the BARs to have been already allocated for them
by the firmware and will pick up the addresses from the device-tree :-)

Today we use a "hack", putting all 0's in there and triggering the linux
code path to reassign unassigned resources (which will use BAR
emulation) but that's not what we are -supposed- to do. Not a big deal
and having the emulation there won't -hurt- us, it's just that we don't
really need any of it.

We have a small issue with ROMs. Our current KVM only works with huge
pages for guest memory but that is being fixed. So the way qemu maps the
ROM copy into the guest address space doesn't work. It might be handy
anyways to have a way for qemu to use MMIO emulation for ROM access as a
fallback. I'll look into it.

  * EEH

This is the name of those fancy error handling & isolation features I
mentioned earlier. To some extent it's a superset of AER, but we don't
generally expose AER to guests (or even the host), it's swallowed by
firmware into something else that provides a superset (well mostly) of
the AER information, and allow us to do those additional things like
isolating/de-isolating, reset control etc...

Here too, we'll need arch specific APIs through VFIO. Not necessarily a
huge deal, I mention it for completeness.

   * Misc

There's lots of small bits and pieces... in no special order:

 - netlink ? WTF ! Seriously, we don't need a hybrid API with a bit of
netlink and a bit of ioctl's ... it's not like there's something
fundamentally  better for netlink vs. ioctl... it really depends what
you are doing, and in this case I fail to see what netlink brings you
other than bloat and more stupid userspace library deps.

 - I don't like too much the fact that VFIO provides yet another
different API to do what we already have at least 2 kernel APIs for, ie,
BAR mapping and config space access. At least it should be better at
using the backend infrastructure of the 2 others (sysfs & procfs). I
understand it wants to filter in some case (config space) and -maybe-
yet another API is the right way to go but allow me to have my doubts.

One thing I thought about but you don't seem to like it ... was to use
the need to represent the partitionable entity as groups in sysfs that I
talked about earlier. Those could have per-device subdirs with the usual
config & resource files, same semantic as the ones in the real device,
but when accessed via the group they get filtering. I might or might not
be practical in the end, tbd, but it would allow apps using a slightly
modified libpci for example to exploit some of this.

 - The qemu vfio code hooks directly into ioapic ... of course that
won't fly with anything !x86

 - The various "objects" dealt with here, -especially- interrupts and
iommu, need a better in-kernel API so that fast in-kernel emulation can
take over from qemu based emulation. The way we need to do some of this
on POWER differs from x86. We can elaborate later, it's not necessarily
a killer either but essentially we'll take the bulk of interrupt
handling away from VFIO to the point where it won't see any of it at
all.

  - Non-PCI devices. That's a hot topic for embedded. I think the vast
majority here is platform devices. There's quite a bit of vfio that
isn't intrinsically PCI specific. We could have an in-kernel platform
driver like we have an in-kernel PCI driver to attach to. The mapping of
resources to userspace is rather generic, as goes for interrupts. I
don't know whether that idea can be pushed much further, I don't have
the bandwidth to look into it much at this point, but maybe it would be
possible to refactor vfio a bit to better separate what is PCI specific
to what is not. The idea would be to move the PCI specific bits to
inside the "placeholder" PCI driver, and same goes for platform bits.
"generic" ioctl's go to VFIO core, anything that doesn't handle, it
passes them to the driver which allows the PCI one to handle things
differently than the platform one, maybe an amba one while at it,
etc.... just a thought, I haven't gone into the details at all.

I think that's all I had on my plate today, it's a long enough email
anyway :-) Anthony suggested we put that on a wiki, I'm a bit
wiki-disabled myself so he proposed to pickup my email and do that. We
should probably discuss the various items in here separately as
different threads to avoid too much confusion.

One other thing we should do on our side is publish somewhere our
current hacks to get you an idea of where we are going and what we had
to do (code speaks more than words). We'll try to do that asap, possibly
next week.

Note that I'll be on/off the next few weeks, travelling and doing
bringup. So expect latency in my replies.

Cheers,
Ben.