[PATCH kernel RFC 2/2] vfio-pci-nvlink2: Implement interconnect isolation
alex.williamson at redhat.com
Sat Mar 23 10:10:25 AEDT 2019
On Fri, 22 Mar 2019 14:08:38 +1100
David Gibson <david at gibson.dropbear.id.au> wrote:
> On Thu, Mar 21, 2019 at 12:19:34PM -0600, Alex Williamson wrote:
> > On Thu, 21 Mar 2019 10:56:00 +1100
> > David Gibson <david at gibson.dropbear.id.au> wrote:
> > > On Wed, Mar 20, 2019 at 01:09:08PM -0600, Alex Williamson wrote:
> > > > On Wed, 20 Mar 2019 15:38:24 +1100
> > > > David Gibson <david at gibson.dropbear.id.au> wrote:
> > > >
> > > > > On Tue, Mar 19, 2019 at 10:36:19AM -0600, Alex Williamson wrote:
> > > > > > On Fri, 15 Mar 2019 19:18:35 +1100
> > > > > > Alexey Kardashevskiy <aik at ozlabs.ru> wrote:
> > > > > >
> > > > > > > The NVIDIA V100 SXM2 GPUs are connected to the CPU via PCIe links and
> > > > > > > (on POWER9) NVLinks. In addition to that, GPUs themselves have direct
> > > > > > > peer to peer NVLinks in groups of 2 to 4 GPUs. At the moment the POWERNV
> > > > > > > platform puts all interconnected GPUs to the same IOMMU group.
> > > > > > >
> > > > > > > However the user may want to pass individual GPUs to the userspace so
> > > > > > > in order to do so we need to put them into separate IOMMU groups and
> > > > > > > cut off the interconnects.
> > > > > > >
> > > > > > > Thankfully V100 GPUs implement an interface to do by programming link
> > > > > > > disabling mask to BAR0 of a GPU. Once a link is disabled in a GPU using
> > > > > > > this interface, it cannot be re-enabled until the secondary bus reset is
> > > > > > > issued to the GPU.
> > > > > > >
> > > > > > > This defines a reset_done() handler for V100 NVlink2 device which
> > > > > > > determines what links need to be disabled. This relies on presence
> > > > > > > of the new "ibm,nvlink-peers" device tree property of a GPU telling which
> > > > > > > PCI peers it is connected to (which includes NVLink bridges or peer GPUs).
> > > > > > >
> > > > > > > This does not change the existing behaviour and instead adds
> > > > > > > a new "isolate_nvlink" kernel parameter to allow such isolation.
> > > > > > >
> > > > > > > The alternative approaches would be:
> > > > > > >
> > > > > > > 1. do this in the system firmware (skiboot) but for that we would need
> > > > > > > to tell skiboot via an additional OPAL call whether or not we want this
> > > > > > > isolation - skiboot is unaware of IOMMU groups.
> > > > > > >
> > > > > > > 2. do this in the secondary bus reset handler in the POWERNV platform -
> > > > > > > the problem with that is at that point the device is not enabled, i.e.
> > > > > > > config space is not restored so we need to enable the device (i.e. MMIO
> > > > > > > bit in CMD register + program valid address to BAR0) in order to disable
> > > > > > > links and then perhaps undo all this initialization to bring the device
> > > > > > > back to the state where pci_try_reset_function() expects it to be.
> > > > > >
> > > > > > The trouble seems to be that this approach only maintains the isolation
> > > > > > exposed by the IOMMU group when vfio-pci is the active driver for the
> > > > > > device. IOMMU groups can be used by any driver and the IOMMU core is
> > > > > > incorporating groups in various ways.
> > > > >
> > > > > I don't think that reasoning is quite right. An IOMMU group doesn't
> > > > > necessarily represent devices which *are* isolated, just devices which
> > > > > *can be* isolated. There are plenty of instances when we don't need
> > > > > to isolate devices in different IOMMU groups: passing both groups to
> > > > > the same guest or userspace VFIO driver for example, or indeed when
> > > > > both groups are owned by regular host kernel drivers.
> > > > >
> > > > > In at least some of those cases we also don't want to isolate the
> > > > > devices when we don't have to, usually for performance reasons.
> > > >
> > > > I see IOMMU groups as representing the current isolation of the device,
> > > > not just the possible isolation. If there are ways to break down that
> > > > isolation then ideally the group would be updated to reflect it. The
> > > > ACS disable patches seem to support this, at boot time we can choose to
> > > > disable ACS at certain points in the topology to favor peer-to-peer
> > > > performance over isolation. This is then reflected in the group
> > > > composition, because even though ACS *can be* enabled at the given
> > > > isolation points, it's intentionally not with this option. Whether or
> > > > not a given user who owns multiple devices needs that isolation is
> > > > really beside the point, the user can choose to connect groups via IOMMU
> > > > mappings or reconfigure the system to disable ACS and potentially more
> > > > direct routing. The IOMMU groups are still accurately reflecting the
> > > > topology and IOMMU based isolation.
> > >
> > > Huh, ok, I think we need to straighten this out. Thinking of iommu
> > > groups as possible rather than potential isolation was a conscious
> > possible ~= potential
> Sorry, I meant "current" not "potential".
> > > decision on my part when we were first coming up with them. The
> > > rationale was that that way iommu groups could be static for the
> > > lifetime of boot, with more dynamic isolation state layered on top.
> > >
> > > Now, that was based on analogy with PAPR's concept of "Partitionable
> > > Endpoints" which are decided by firmware before boot. However, I
> > > think it makes sense in other contexts too: if iommu groups represent
> > > current isolation, then we need some other way to advertise possible
> > > isolation - otherwise how will the admin (and/or tools) know how it
> > > can configure the iommu groups.
> > >
> > > VFIO already has the container, which represents explicitly a "group
> > > of groups" that we don't care to isolate from each other. I don't
> > > actually know what other uses of the iommu group infrastructure we
> > > have at present and how they treat them.
> > s/that we don't care to isolate/that the user doesn't care to isolate/
> Well, true, I guess we still want to isolate them when possible for
> additional safety. But I don't think we should do so when that
> isolation has a real performance cost, which it would in the case of
> isolating linked GPUs here.
I don't see that we can presume the user's intentions. This sounds
like we're putting the kernel into the position of deciding policy.
> > Though even that is not necessarily accurate, the container represents
> > a shared IOMMU context, that context doesn't necessarily include
> > mappings between devices, so the device can still be isolated *from
> > each other*.
> Well, sure, but it's the only mechanism the user has for indicating
> that they don't care about isolation between these device. That might
> be because they simply don't care, but it might also be because they
> want to use interlinks between those devices for additional
Is a user really indicating "I don't care about isolation" by placing
groups within the same container or just taking advantage of a shared
IOMMU context? Again, we're presuming a policy decision that's never
been defined in the API. It is possible to have both isolation and
performance, for instance ACS includes a direct translated p2p feature
that allows translated requests a direct path between devices, at least
on the PCIe fabric (where supported by hardware).
> > > So, if we now have dynamically reconfigurable groups which are a
> > > departure from that design, how can we go from here trying to bring
> > > things back to consistency.
> > We don't currently have dynamically reconfigurable groups,
> Ah, ok. So what were these other use cases you were describing? Boot
> time options which change how the iommu groups are generated? Or
> something else.
pci=option[,option...] [PCI] various PCI subsystem options.
Some options herein operate on a specific device
or a set of devices (<pci_dev>). These are
specified in one of the following formats:
Specify one or more PCI devices (in the format
specified above) separated by semicolons.
Each device specified will have the PCI ACS
redirect capabilities forced off which will
allow P2P traffic between devices through
bridges without forcing it upstream. Note:
this removes isolation between devices and
may put more devices in an IOMMU group.
AIUI, GPU and RDMA were primary use cases for this. As noted, IOMMU
groups take this into account since it's done at boot time and static,
we're simply disabling isolation switches that would have otherwise
been enabled at points in the topology (specifically enabling p2p as
much as possible), and the IOMMU group grows to match.
The NVLink problem is rather the opposite of this, rather than simply
disabling isolation for a sub-hierarchy and growing the IOMMU group, we
need to enable isolation and then actively, and repeatedly maintain
that configuration as it represents a finer granularity isolation
promise to the system. And of course we're trying to manage the
isolation of an interconnect for which we have no visibility.
> > I think the
> > question is how do we introduce dynamically configurable groups within
> > the existing design, because whatever intentions were 7.5 years ago is
> > rather irrelevant now.
> Well, yeah, I guess.
> > VFIO groups are mapped directly to IOMMU groups. IOMMU groups are the
> > smallest set of devices which can be considered isolated from other
> > sets of devices (not potentially, but as configured).
> Hm.. but what actually makes that assumption that this is the case as
> configured, rather than just possibly - I'm taking "possible" to mean
> possible with this host kernel, not just possible with this hardware.
The IOMMU driver is ultimately responsible for deciding the isolation,
but we have shared code that understands the PCI/e topology. That code
looks at the actual, current ACS settings to decide the highest point
in the topology where we might have p2p.
> > VFIO groups are
> > the unit of ownership to userspace, therefore a VFIO group must be
> > (currently) isolated from other groups.
> It must be currently isolated once the VFIO group is instantiated, but
> is there actually anything that requires that current isolation before
> the iommu group is instantiated as a vfio group.
The iommu-core is incorporating more group-aware support, I believe
default domains are group based and we might have per group reserved
ranges, I'm not sure about that.
> > The user owns and manages the
> > IOMMU context for one or more groups via a container. A container
> > allows efficiency in managing a shared IOVA space between groups.
> Yes, and it seems to me an obvious extension to have the container
> also permit other efficiencies which aren't compatible with full
It might be a control point, yes, but I don't see that we can presume
it. If we were to automatically enable redirection within a container,
we could wreak havoc on the IOVA space for the user. ACS also
theoretically provides a hardware path to enable this via direct
translated p2p as well.
> > > > > > So, if there's a device specific
> > > > > > way to configure the isolation reported in the group, which requires
> > > > > > some sort of active management against things like secondary bus
> > > > > > resets, then I think we need to manage it above the attached endpoint
> > > > > > driver.
> > > > >
> > > > > The problem is that above the endpoint driver, we don't actually have
> > > > > enough information about what should be isolated. For VFIO we want to
> > > > > isolate things if they're in different containers, for most regular
> > > > > host kernel drivers we don't need to isolate at all (although we might
> > > > > as well when it doesn't have a cost).
> > > >
> > > > This idea that we only want to isolate things if they're in different
> > > > containers is bogus, imo. There are performance reasons why we might
> > > > not want things isolated, but there are also address space reasons why
> > > > we do. If there are direct routes between devices, the user needs to
> > > > be aware of the IOVA pollution, if we maintain singleton groups, they
> > > > don't. Granted we don't really account for this well in most
> > > > userspaces and fumble through it by luck of the address space layout
> > > > and lack of devices really attempting peer to peer access.
> > >
> > > I don't really follow what you're saying here.
> > If the lack of isolation between devices includes peer-to-peer channels
> > then the MMIO of the devices within the group pollutes the IOVA space
> > of the container.
> Yes, if the permissible IOVA addresses overlap with valid MMIO
> addresses. I don't really see why that's significant, though. We
> already have basically the same situation if there are multiple
> devices in a group. e.g. if you have several devices behind a dumb
> PCI-E to PCI bridge, you can't actually prohibit peer-to-peer DMA
> between them, and so those devices' MMIOs pollute the IOVA space for
> each other, even if they don't for other devices in the container.
> There's only so far we can go to prevent the user from shooting
> themselves in the foot.
So it's therefore OK to intentionally pollute the container IOVA
space? I believe we'd like to think that hardware is headed towards
being more isolation-aware and will evolve to the point where we have
typically singleton group. Perhaps we'll also evolve to supporting ACS
direct translated p2p so that we have both a pristine IOVA space and
direct routes between devices. There are current issues with
non-singleton groups that userspace doesn't handle (see "fumble
through" comment above), but I'm not willing to say "this is
userspace's problem anyway, therefore it's ok it make it worse".
> > > > For in-kernel users, we're still theoretically trying to isolate
> > > > devices such that they have restricted access to only the resources
> > > > they need. Disabling things like ACS in the topology reduces that
> > > > isolation. AFAICT, most users don't really care about that degree of
> > > > isolation, so they run with iommu=pt for native driver performance
> > > > while still having the IOMMU available for isolation use cases running
> > > > in parallel. We don't currently have support for on-demand enabling
> > > > isolation.
> > >
> > > Ok.
> > >
> > > > > The host side nVidia GPGPU
> > > > > drivers also won't want to isolate the (host owned) NVLink devices
> > > > > from each other, since they'll want to use the fast interconnects
> > > >
> > > > This falls into the same mixed use case scenario above where we don't
> > > > really have a good solution today. Things like ACS are dynamically
> > > > configurable, but we don't expose any interfaces to let drivers or
> > > > users change it (aside from setpci, which we don't account for
> > > > dynamically). We assume a simplistic model where if you want IOMMU,
> > > > then you must also want the maximum configurable isolation.
> > > > Dynamically changing routing is not necessarily the most foolproof
> > > > thing either with potentially in-flight transactions and existing DMA
> > > > mappings, which is why I've suggested a couple times that perhaps we
> > > > could do a software hot-unplug of a sub-hierarchy, muck with isolation
> > > > at the remaining node, then re-discover the removed devices.
> > >
> > > Ok, so I feel like we need to go fully one way or the other. Either:
> > >
> > > 1) Groups represent current isolation status, in which case we
> > > deprecate vfio containers in favour of fusing groups beforehand,
> > > and we need some new concept ("isolation atoms"?) to represent what
> > > isolation is possible
> > Nope, containers are owned by users and serve a purpose beyond what
> > you're assuming here. Also, there's 7.5 years of userspace tooling
> > broken by this. I do remember we had discussions about merging groups,
> > but what we have now is what we agreed on.
> Well, quite - and I thought the reasoning that led to that design was
> that groups would be a representation of isolation granularity, not
> actual isolation. That allows them to be static, which as you say,
> tools are likely to assume.
What is the value of this idea of potential isolation? Any device
might be considered potentially isolated, simply disable all the
devices around it. So now we have a "potentially isolated" group at
each device. What have we accomplished? How do we build that into a
user interface? Who decides where the actual isolation lines are
drawn? Not that it didn't happen, but I don't recall this design idea
and I don't see how it's practical.
> > > or
> > >
> > > 2) Groups represent potential isolation, with a higher level construct
> > > representing current isolation. This could involve making vfio
> > > containers essentially a wrapper around some more generic concept
> > > ("isolation clusters"?) of a group of groups.
> > Have fun. Containers and groups are exposed to userspace, so you're
> > essentially suggesting to throw away all that support, for... I don't
> > really know what.
> No, I'm not proposing removing containers, just changing them from
> being purely VFIO specific to being the VFIO front end to a generic
> concept. Roughly speaking VFIO containers would be to "isolation
> clusters" (or whatever) as VFIO groups are to IOMMU groups now.
Renaming everything doesn't help. AFAIK, there's nothing in vfio that
prevents groups from being dynamic (changing them while in use might be
unkind though), but they must be isolated from other groups. We can't
move the line where we have isolation to anything above the group.
Groups are currently not dynamic because "it's hard" (tm), but
obviously it can be done with device removal and rediscovery.
> > Groups are involved in IOMMU context, so dynamically
> > changing what a group defines is hard.
> Absolutely! That's why I don't want to, and why my conception of
> groups was defined so that they didn't have to.
As above, groups are the unit of userspace ownership and therefore must
> > Thus my suggestions that mixed
> > or dynamic workloads could make use of soft remove and rediscovery for
> > sub-hierarchies, unbinding and rebinding drivers such that we have the
> > proper expectations of DMA context. Thanks,
> Well, if we have to change groups, then that sounds like the sanest
> available way to do it. But I'm not yet convinced that altering
> groups makes sense rather than using a group-of-groups concept on top
> of it - which would map to containers in the case of VFIO.
Groups must be isolated in the vfio API. What happens underneath a
group might be fungible, but the exposed group must be isolated.
More information about the Linuxppc-dev