Linux Virtualization - Sep 2015 - rfc: vhost user enhancements for vm2vm communication

On 2015-09-01 18:02, Michael S. Tsirkin wrote:
> On Tue, Sep 01, 2015 at 05:34:37PM +0200, Jan Kiszka wrote:
>> On 2015-09-01 16:34, Michael S. Tsirkin wrote:
>>> On Tue, Sep 01, 2015 at 04:09:44PM +0200, Jan Kiszka wrote:
>>>> On 2015-09-01 11:24, Michael S. Tsirkin wrote:
>>>>> On Tue, Sep 01, 2015 at 11:11:52AM +0200, Jan Kiszka wrote:
>>>>>> On 2015-09-01 10:01, Michael S. Tsirkin wrote:
>>>>>>> On Tue, Sep 01, 2015 at 09:35:21AM +0200, Jan
Kiszka wrote:
>>>>>>>> Leaving all the implementation and interface
details aside, this
>>>>>>>> discussion is first of all about two
fundamentally different approaches:
>>>>>>>> static shared memory windows vs. dynamically
remapped shared windows (a
>>>>>>>> third one would be copying in the hypervisor,
but I suppose we all agree
>>>>>>>> that the whole exercise is about avoiding
that). Which way do we want or
>>>>>>>> have to go?
>>>>>>>>
>>>>>>>> Jan
>>>>>>>
>>>>>>> Dynamic is a superset of static: you can always
make it static if you
>>>>>>> wish. Static has the advantage of simplicity, but
that's lost once you
>>>>>>> realize you need to invent interfaces to make it
work.  Since we can use
>>>>>>> existing IOMMU interfaces for the dynamic one,
what's the disadvantage?
>>>>>>
>>>>>> Complexity. Having to emulate even more of an IOMMU in
the hypervisor
>>>>>> (we already have to do a bit for VT-d IR in Jailhouse)
and doing this
>>>>>> per platform (AMD IOMMU, ARM SMMU, ...) is out of scope
for us. In that
>>>>>> sense, generic grant tables would be more appealing.
>>>>>
>>>>> That's not how we do things for KVM, PV features need
to be
>>>>> modular and interchangeable with emulation.
>>>>
>>>> I know, and we may have to make some compromise for Jailhouse
if that
>>>> brings us valuable standardization and broad guest support. But
we will
>>>> surely not support an arbitrary amount of IOMMU models for that
reason.
>>>>
>>>>>
>>>>> If you just want something that's cross-platform and
easy to
>>>>> implement, just build a PV IOMMU. Maybe use virtio for
this.
>>>>
>>>> That is likely required to keep the complexity manageable and
to allow
>>>> static preconfiguration.
>>>
>>> Real IOMMU allow static configuration just fine. This is exactly
>>> what VFIO uses.
>>
>> Please specify more precisely which feature in which IOMMU you are
>> referring to. Also, given that you refer to VFIO, I suspect we have
>> different thing in mind. I'm talking about an IOMMU device model,
like
>> the one we have in QEMU now for VT-d. That one is not at all
>> preconfigured by the host for VFIO.
> 
> I really just mean that VFIO creates a mostly static IOMMU configuration.
> 
> It's configured by the guest, not the host.
OK, that resolves my confusion.
> 
> I don't see host control over configuration as being particularly
important.
We do, see below.
> 
> 
>>>
>>>> Well, we could declare our virtio-shmem device to be an IOMMU
device
>>>> that controls access of a remote VM to RAM of the one that owns
the
>>>> device. In the static case, this access may at most be
enabled/disabled
>>>> but not moved around. The static regions would have to be
discoverable
>>>> for the VM (register read-back), and the guest's firmware
will likely
>>>> have to declare those ranges reserved to the guest OS.
>>>> In the dynamic case, the guest would be able to create an
alternative
>>>> mapping.
>>>
>>>
>>> I don't think we want a special device just to support the
>>> static case. It might be a bit less code to write, but
>>> eventually it should be up to the guest.
>>> Fundamentally, it's policy that host has no business
>>> dictating.
>>
>> "A bit less" is to be validated, and I doubt its just "a
bit". But if
>> KVM and its guests will also support some PV-IOMMU that we can reuse
for
>> our scenarios, than that is fine. KVM would not have to mandate support
>> for it while we would, that's all.
> 
> Someone will have to do this work.
> 
>>>
>>>> We would probably have to define a generic page table structure
>>>> for that. Or do you rather have some MPU-like control structure
in mind,
>>>> more similar to the memory region descriptions vhost is already
using?
>>>
>>> I don't care much. Page tables use less memory if a lot of
memory needs
>>> to be covered. OTOH if you want to use virtio (e.g. to allow
command
>>> batching) that likely means commands to manipulate the IOMMU, and
>>> maintaining it all on the host. You decide.
>>
>> I don't care very much about the dynamic case as we won't
support it
>> anyway. However, if the configuration concept used for it is applicable
>> to static mode as well, then we could reuse it. But preconfiguration
>> will required register-based region description, I suspect.
> 
> I don't know what you mean by preconfiguration exactly.
> 
> Do you want the host to configure the IOMMU? Why not let the
> guest do this?
We simply freeze GPA-to-HPA mappings during runtime. Avoids having to
validate and synchronize guest-triggered changes.
>>>
>>>> Also not yet clear to me are how the vhost-pci device and the
>>>> translations it will have to do should look like for VM2.
>>>
>>> I think we can use vhost-pci BAR + VM1 bus address as the
>>> VM2 physical address. In other words, all memory exposed to
>>> virtio-pci by VM1 through it's IOMMU is mapped into BAR of
>>> vhost-pci.
>>>
>>> Bus addresses can be validated to make sure they fit
>>> in the BAR.
>>
>> Sounds simple but may become challenging for VMs that have many of such
>> devices (in order to connect to many possibly large VMs).
> 
> You don't need to be able to map all guest memory if you know
> guest won't try to allow device access to all of it.
> It's a question of how good is the bus address allocator.
But those BARs need to allocate a guest-physical address range as large
as the other guest's RAM is, possibly even larger if that RAM is not
contiguous, and you can't put other resources into potential holes
because VM2 does not know where those holes will be.
> 
>>>
>>>
>>> One issue to consider is that VM1 can trick VM2 into writing
>>> into bus address that isn't mapped in the IOMMU, or
>>> is mapped read-only.
>>> We probably would have to teach KVM to handle this somehow,
>>> e.g. exit to QEMU, or even just ignore. Maybe notify guest
>>> e.g. by setting a bit in the config space of the device,
>>> to avoid easy DOS.
>>
>> Well, that would be trivial for VM1 to check if there are only one or
>> two memory windows. Relying on the hypervisor to handle it may be
>> unacceptable for real-time VMs.
>>
>> Jan
> 
> Why? real-time != fast. I doubt you can avoid vm exits completely.
We can, one property of Jailhouse (on x86, ARM is waiting for GICv4).

Real-time == deterministic. And if you have such vm exits potentially in
your code path, you have them always - for worst-case analysis. One may
argue about probability in certain scenarios, but if the triggering side
is malicious, probability may become 1.

Jan

-- 
Siemens AG, Corporate Technology, CT RTC ITP SES-DE
Corporate Competence Center Embedded Linux

On Tue, Sep 1, 2015 at 9:28 AM, Jan Kiszka <jan.kiszka at siemens.com>
wrote:
> On 2015-09-01 18:02, Michael S. Tsirkin wrote:
...
>> You don't need to be able to map all guest memory if you know
>> guest won't try to allow device access to all of it.
>> It's a question of how good is the bus address allocator.
>
> But those BARs need to allocate a guest-physical address range as large
> as the other guest's RAM is, possibly even larger if that RAM is not
> contiguous, and you can't put other resources into potential holes
> because VM2 does not know where those holes will be.
>
I think you can allocate such guest-physical address ranges
efficiently if each BAR sets the base of each memory region reported
by VHOST_SET_MEM_TABLE, for example.  The issue is that we would need
to 8 (VHOST_MEMORY_MAX_NREGIONS) of them vs. 6 (defined by PCI-SIG).

-- 
Jun
Intel Open Source Technology Center

On Tue, Sep 01, 2015 at 05:01:07PM -0700, Nakajima, Jun
wrote:
> On Tue, Sep 1, 2015 at 9:28 AM, Jan Kiszka <jan.kiszka at
siemens.com> wrote:
> > On 2015-09-01 18:02, Michael S. Tsirkin wrote:
> ...
> >> You don't need to be able to map all guest memory if you know
> >> guest won't try to allow device access to all of it.
> >> It's a question of how good is the bus address allocator.
> >
> > But those BARs need to allocate a guest-physical address range as
large
> > as the other guest's RAM is, possibly even larger if that RAM is
not
> > contiguous, and you can't put other resources into potential holes
> > because VM2 does not know where those holes will be.
> >
> 
> I think you can allocate such guest-physical address ranges
> efficiently if each BAR sets the base of each memory region reported
> by VHOST_SET_MEM_TABLE, for example.  The issue is that we would need
> to 8 (VHOST_MEMORY_MAX_NREGIONS) of them vs. 6 (defined by PCI-SIG).
Besides, 8 is not even a limit: we merged a patch that allows makeing it
larger.
> -- 
> Jun
> Intel Open Source Technology Center

On Tue, Sep 01, 2015 at 06:28:28PM +0200, Jan Kiszka
wrote:
> On 2015-09-01 18:02, Michael S. Tsirkin wrote:
> > On Tue, Sep 01, 2015 at 05:34:37PM +0200, Jan Kiszka wrote:
> >> On 2015-09-01 16:34, Michael S. Tsirkin wrote:
> >>> On Tue, Sep 01, 2015 at 04:09:44PM +0200, Jan Kiszka wrote:
> >>>> On 2015-09-01 11:24, Michael S. Tsirkin wrote:
> >>>>> On Tue, Sep 01, 2015 at 11:11:52AM +0200, Jan Kiszka
wrote:
> >>>>>> On 2015-09-01 10:01, Michael S. Tsirkin wrote:
> >>>>>>> On Tue, Sep 01, 2015 at 09:35:21AM +0200, Jan
Kiszka wrote:
> >>>>>>>> Leaving all the implementation and
interface details aside, this
> >>>>>>>> discussion is first of all about two
fundamentally different approaches:
> >>>>>>>> static shared memory windows vs.
dynamically remapped shared windows (a
> >>>>>>>> third one would be copying in the
hypervisor, but I suppose we all agree
> >>>>>>>> that the whole exercise is about avoiding
that). Which way do we want or
> >>>>>>>> have to go?
> >>>>>>>>
> >>>>>>>> Jan
> >>>>>>>
> >>>>>>> Dynamic is a superset of static: you can
always make it static if you
> >>>>>>> wish. Static has the advantage of simplicity,
but that's lost once you
> >>>>>>> realize you need to invent interfaces to make
it work.  Since we can use
> >>>>>>> existing IOMMU interfaces for the dynamic one,
what's the disadvantage?
> >>>>>>
> >>>>>> Complexity. Having to emulate even more of an
IOMMU in the hypervisor
> >>>>>> (we already have to do a bit for VT-d IR in
Jailhouse) and doing this
> >>>>>> per platform (AMD IOMMU, ARM SMMU, ...) is out of
scope for us. In that
> >>>>>> sense, generic grant tables would be more
appealing.
> >>>>>
> >>>>> That's not how we do things for KVM, PV features
need to be
> >>>>> modular and interchangeable with emulation.
> >>>>
> >>>> I know, and we may have to make some compromise for
Jailhouse if that
> >>>> brings us valuable standardization and broad guest
support. But we will
> >>>> surely not support an arbitrary amount of IOMMU models for
that reason.
> >>>>
> >>>>>
> >>>>> If you just want something that's cross-platform
and easy to
> >>>>> implement, just build a PV IOMMU. Maybe use virtio for
this.
> >>>>
> >>>> That is likely required to keep the complexity manageable
and to allow
> >>>> static preconfiguration.
> >>>
> >>> Real IOMMU allow static configuration just fine. This is
exactly
> >>> what VFIO uses.
> >>
> >> Please specify more precisely which feature in which IOMMU you are
> >> referring to. Also, given that you refer to VFIO, I suspect we
have
> >> different thing in mind. I'm talking about an IOMMU device
model, like
> >> the one we have in QEMU now for VT-d. That one is not at all
> >> preconfigured by the host for VFIO.
> > 
> > I really just mean that VFIO creates a mostly static IOMMU
configuration.
> > 
> > It's configured by the guest, not the host.
> 
> OK, that resolves my confusion.
> 
> > 
> > I don't see host control over configuration as being particularly
important.
> 
> We do, see below.
> 
> > 
> > 
> >>>
> >>>> Well, we could declare our virtio-shmem device to be an
IOMMU device
> >>>> that controls access of a remote VM to RAM of the one that
owns the
> >>>> device. In the static case, this access may at most be
enabled/disabled
> >>>> but not moved around. The static regions would have to be
discoverable
> >>>> for the VM (register read-back), and the guest's
firmware will likely
> >>>> have to declare those ranges reserved to the guest OS.
> >>>> In the dynamic case, the guest would be able to create an
alternative
> >>>> mapping.
> >>>
> >>>
> >>> I don't think we want a special device just to support the
> >>> static case. It might be a bit less code to write, but
> >>> eventually it should be up to the guest.
> >>> Fundamentally, it's policy that host has no business
> >>> dictating.
> >>
> >> "A bit less" is to be validated, and I doubt its just
"a bit". But if
> >> KVM and its guests will also support some PV-IOMMU that we can
reuse for
> >> our scenarios, than that is fine. KVM would not have to mandate
support
> >> for it while we would, that's all.
> > 
> > Someone will have to do this work.
> > 
> >>>
> >>>> We would probably have to define a generic page table
structure
> >>>> for that. Or do you rather have some MPU-like control
structure in mind,
> >>>> more similar to the memory region descriptions vhost is
already using?
> >>>
> >>> I don't care much. Page tables use less memory if a lot of
memory needs
> >>> to be covered. OTOH if you want to use virtio (e.g. to allow
command
> >>> batching) that likely means commands to manipulate the IOMMU,
and
> >>> maintaining it all on the host. You decide.
> >>
> >> I don't care very much about the dynamic case as we won't
support it
> >> anyway. However, if the configuration concept used for it is
applicable
> >> to static mode as well, then we could reuse it. But
preconfiguration
> >> will required register-based region description, I suspect.
> > 
> > I don't know what you mean by preconfiguration exactly.
> > 
> > Do you want the host to configure the IOMMU? Why not let the
> > guest do this?
> 
> We simply freeze GPA-to-HPA mappings during runtime. Avoids having to
> validate and synchronize guest-triggered changes.
Fine, but this assumes guest does very specific things, right?
E.g. should guest reconfigure device's BAR, you would have
to change GPA to HPA mappings?

> >>>
> >>>> Also not yet clear to me are how the vhost-pci device and
the
> >>>> translations it will have to do should look like for VM2.
> >>>
> >>> I think we can use vhost-pci BAR + VM1 bus address as the
> >>> VM2 physical address. In other words, all memory exposed to
> >>> virtio-pci by VM1 through it's IOMMU is mapped into BAR of
> >>> vhost-pci.
> >>>
> >>> Bus addresses can be validated to make sure they fit
> >>> in the BAR.
> >>
> >> Sounds simple but may become challenging for VMs that have many of
such
> >> devices (in order to connect to many possibly large VMs).
> > 
> > You don't need to be able to map all guest memory if you know
> > guest won't try to allow device access to all of it.
> > It's a question of how good is the bus address allocator.
> 
> But those BARs need to allocate a guest-physical address range as large
> as the other guest's RAM is, possibly even larger if that RAM is not
> contiguous, and you can't put other resources into potential holes
> because VM2 does not know where those holes will be.
No - only the RAM that you want addressable by VM2.

IOW if you wish, you actually can create a shared memory device,
make it accessible to the IOMMU and place some or all
data there.



> > 
> >>>
> >>>
> >>> One issue to consider is that VM1 can trick VM2 into writing
> >>> into bus address that isn't mapped in the IOMMU, or
> >>> is mapped read-only.
> >>> We probably would have to teach KVM to handle this somehow,
> >>> e.g. exit to QEMU, or even just ignore. Maybe notify guest
> >>> e.g. by setting a bit in the config space of the device,
> >>> to avoid easy DOS.
> >>
> >> Well, that would be trivial for VM1 to check if there are only one
or
> >> two memory windows. Relying on the hypervisor to handle it may be
> >> unacceptable for real-time VMs.
> >>
> >> Jan
> > 
> > Why? real-time != fast. I doubt you can avoid vm exits completely.
> 
> We can, one property of Jailhouse (on x86, ARM is waiting for GICv4).
> 
> Real-time == deterministic. And if you have such vm exits potentially in
> your code path, you have them always - for worst-case analysis. One may
> argue about probability in certain scenarios, but if the triggering side
> is malicious, probability may become 1.
> 
> Jan
You are doing a special hypervisor anyway, I think you could
detect that setup is done, and freeze
the configuration.

If afterwards a VM attempts to modify mappings, you can
say it's malicious and ignore it, or kill it, or whatever.


> -- 
> Siemens AG, Corporate Technology, CT RTC ITP SES-DE
> Corporate Competence Center Embedded Linux

On 2015-09-03 10:08, Michael S. Tsirkin wrote:
> On Tue, Sep 01, 2015 at 06:28:28PM +0200, Jan Kiszka wrote:
>> On 2015-09-01 18:02, Michael S. Tsirkin wrote:
>>> On Tue, Sep 01, 2015 at 05:34:37PM +0200, Jan Kiszka wrote:
>>>> On 2015-09-01 16:34, Michael S. Tsirkin wrote:
>>>>> On Tue, Sep 01, 2015 at 04:09:44PM +0200, Jan Kiszka wrote:
>>>>>> On 2015-09-01 11:24, Michael S. Tsirkin wrote:
>>>>>>> On Tue, Sep 01, 2015 at 11:11:52AM +0200, Jan
Kiszka wrote:
>>>>>>>> On 2015-09-01 10:01, Michael S. Tsirkin wrote:
>>>>>>>>> On Tue, Sep 01, 2015 at 09:35:21AM +0200,
Jan Kiszka wrote:
>>>>>>>>>> Leaving all the implementation and
interface details aside, this
>>>>>>>>>> discussion is first of all about two
fundamentally different approaches:
>>>>>>>>>> static shared memory windows vs.
dynamically remapped shared windows (a
>>>>>>>>>> third one would be copying in the
hypervisor, but I suppose we all agree
>>>>>>>>>> that the whole exercise is about
avoiding that). Which way do we want or
>>>>>>>>>> have to go?
>>>>>>>>>>
>>>>>>>>>> Jan
>>>>>>>>>
>>>>>>>>> Dynamic is a superset of static: you can
always make it static if you
>>>>>>>>> wish. Static has the advantage of
simplicity, but that's lost once you
>>>>>>>>> realize you need to invent interfaces to
make it work.  Since we can use
>>>>>>>>> existing IOMMU interfaces for the dynamic
one, what's the disadvantage?
>>>>>>>>
>>>>>>>> Complexity. Having to emulate even more of an
IOMMU in the hypervisor
>>>>>>>> (we already have to do a bit for VT-d IR in
Jailhouse) and doing this
>>>>>>>> per platform (AMD IOMMU, ARM SMMU, ...) is out
of scope for us. In that
>>>>>>>> sense, generic grant tables would be more
appealing.
>>>>>>>
>>>>>>> That's not how we do things for KVM, PV
features need to be
>>>>>>> modular and interchangeable with emulation.
>>>>>>
>>>>>> I know, and we may have to make some compromise for
Jailhouse if that
>>>>>> brings us valuable standardization and broad guest
support. But we will
>>>>>> surely not support an arbitrary amount of IOMMU models
for that reason.
>>>>>>
>>>>>>>
>>>>>>> If you just want something that's
cross-platform and easy to
>>>>>>> implement, just build a PV IOMMU. Maybe use virtio
for this.
>>>>>>
>>>>>> That is likely required to keep the complexity
manageable and to allow
>>>>>> static preconfiguration.
>>>>>
>>>>> Real IOMMU allow static configuration just fine. This is
exactly
>>>>> what VFIO uses.
>>>>
>>>> Please specify more precisely which feature in which IOMMU you
are
>>>> referring to. Also, given that you refer to VFIO, I suspect we
have
>>>> different thing in mind. I'm talking about an IOMMU device
model, like
>>>> the one we have in QEMU now for VT-d. That one is not at all
>>>> preconfigured by the host for VFIO.
>>>
>>> I really just mean that VFIO creates a mostly static IOMMU
configuration.
>>>
>>> It's configured by the guest, not the host.
>>
>> OK, that resolves my confusion.
>>
>>>
>>> I don't see host control over configuration as being
particularly important.
>>
>> We do, see below.
>>
>>>
>>>
>>>>>
>>>>>> Well, we could declare our virtio-shmem device to be an
IOMMU device
>>>>>> that controls access of a remote VM to RAM of the one
that owns the
>>>>>> device. In the static case, this access may at most be
enabled/disabled
>>>>>> but not moved around. The static regions would have to
be discoverable
>>>>>> for the VM (register read-back), and the guest's
firmware will likely
>>>>>> have to declare those ranges reserved to the guest OS.
>>>>>> In the dynamic case, the guest would be able to create
an alternative
>>>>>> mapping.
>>>>>
>>>>>
>>>>> I don't think we want a special device just to support
the
>>>>> static case. It might be a bit less code to write, but
>>>>> eventually it should be up to the guest.
>>>>> Fundamentally, it's policy that host has no business
>>>>> dictating.
>>>>
>>>> "A bit less" is to be validated, and I doubt its just
"a bit". But if
>>>> KVM and its guests will also support some PV-IOMMU that we can
reuse for
>>>> our scenarios, than that is fine. KVM would not have to mandate
support
>>>> for it while we would, that's all.
>>>
>>> Someone will have to do this work.
>>>
>>>>>
>>>>>> We would probably have to define a generic page table
structure
>>>>>> for that. Or do you rather have some MPU-like control
structure in mind,
>>>>>> more similar to the memory region descriptions vhost is
already using?
>>>>>
>>>>> I don't care much. Page tables use less memory if a lot
of memory needs
>>>>> to be covered. OTOH if you want to use virtio (e.g. to
allow command
>>>>> batching) that likely means commands to manipulate the
IOMMU, and
>>>>> maintaining it all on the host. You decide.
>>>>
>>>> I don't care very much about the dynamic case as we
won't support it
>>>> anyway. However, if the configuration concept used for it is
applicable
>>>> to static mode as well, then we could reuse it. But
preconfiguration
>>>> will required register-based region description, I suspect.
>>>
>>> I don't know what you mean by preconfiguration exactly.
>>>
>>> Do you want the host to configure the IOMMU? Why not let the
>>> guest do this?
>>
>> We simply freeze GPA-to-HPA mappings during runtime. Avoids having to
>> validate and synchronize guest-triggered changes.
> 
> Fine, but this assumes guest does very specific things, right?
> E.g. should guest reconfigure device's BAR, you would have
> to change GPA to HPA mappings?
> 
Yes, that's why we only support size exploration, not reallocation.
> 
>>>>>
>>>>>> Also not yet clear to me are how the vhost-pci device
and the
>>>>>> translations it will have to do should look like for
VM2.
>>>>>
>>>>> I think we can use vhost-pci BAR + VM1 bus address as the
>>>>> VM2 physical address. In other words, all memory exposed to
>>>>> virtio-pci by VM1 through it's IOMMU is mapped into BAR
of
>>>>> vhost-pci.
>>>>>
>>>>> Bus addresses can be validated to make sure they fit
>>>>> in the BAR.
>>>>
>>>> Sounds simple but may become challenging for VMs that have many
of such
>>>> devices (in order to connect to many possibly large VMs).
>>>
>>> You don't need to be able to map all guest memory if you know
>>> guest won't try to allow device access to all of it.
>>> It's a question of how good is the bus address allocator.
>>
>> But those BARs need to allocate a guest-physical address range as large
>> as the other guest's RAM is, possibly even larger if that RAM is
not
>> contiguous, and you can't put other resources into potential holes
>> because VM2 does not know where those holes will be.
> 
> No - only the RAM that you want addressable by VM2.
That's in the hand of VM1, not VM2 or the hypervisor, in case of
reconfigurable mapping. It's indeed a non-issue in our static case.
> 
> IOW if you wish, you actually can create a shared memory device,
> make it accessible to the IOMMU and place some or all
> data there.
> 
Actually, that could also be something more sophisticated, including
virtio-net, IF that device will be able to express its DMA window
restrictions (a bit like 32-bit PCI devices being restricted to <4G
addresses or ISA devices <1M).

Jan

-- 
Siemens AG, Corporate Technology, CT RTC ITP SES-DE
Corporate Competence Center Embedded Linux