Re: [PATCH V2 mlx5-next 12/14] vfio/mlx5: Implement vfio_pci driver for mlx5 devices

From: Alex Williamson <alex.williamson@redhat.com>
To: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Yishai Hadas <yishaih@nvidia.com>,
	Jason Gunthorpe <jgg@nvidia.com>,
	bhelgaas@google.com, saeedm@nvidia.com,
	linux-pci@vger.kernel.org, kvm@vger.kernel.org,
	netdev@vger.kernel.org, kuba@kernel.org, leonro@nvidia.com,
	kwankhede@nvidia.com, mgurtovoy@nvidia.com, maorg@nvidia.com,
	Cornelia Huck <cohuck@redhat.com>
Subject: Re: [PATCH V2 mlx5-next 12/14] vfio/mlx5: Implement vfio_pci driver for mlx5 devices
Date: Tue, 26 Oct 2021 08:29:20 -0600	[thread overview]
Message-ID: <20211026082920.1f302a45.alex.williamson@redhat.com> (raw)
In-Reply-To: <YXb7wejD1qckNrhC@work-vm>

On Mon, 25 Oct 2021 19:47:29 +0100
"Dr. David Alan Gilbert" <dgilbert@redhat.com> wrote:

> * Alex Williamson (alex.williamson@redhat.com) wrote:
> > On Mon, 25 Oct 2021 17:34:01 +0100
> > "Dr. David Alan Gilbert" <dgilbert@redhat.com> wrote:
> >   
> > > * Alex Williamson (alex.williamson@redhat.com) wrote:  
> > > > [Cc +dgilbert, +cohuck]
> > > > 
> > > > On Wed, 20 Oct 2021 11:28:04 +0300
> > > > Yishai Hadas <yishaih@nvidia.com> wrote:
> > > >     
> > > > > On 10/20/2021 2:04 AM, Jason Gunthorpe wrote:    
> > > > > > On Tue, Oct 19, 2021 at 02:58:56PM -0600, Alex Williamson wrote:      
> > > > > >> I think that gives us this table:
> > > > > >>
> > > > > >> |   NDMA   | RESUMING |  SAVING  |  RUNNING |
> > > > > >> +----------+----------+----------+----------+ ---
> > > > > >> |     X    |     0    |     0    |     0    |  ^
> > > > > >> +----------+----------+----------+----------+  |
> > > > > >> |     0    |     0    |     0    |     1    |  |
> > > > > >> +----------+----------+----------+----------+  |
> > > > > >> |     X    |     0    |     1    |     0    |
> > > > > >> +----------+----------+----------+----------+  NDMA value is either compatible
> > > > > >> |     0    |     0    |     1    |     1    |  to existing behavior or don't
> > > > > >> +----------+----------+----------+----------+  care due to redundancy vs
> > > > > >> |     X    |     1    |     0    |     0    |  !_RUNNING/INVALID/ERROR
> > > > > >> +----------+----------+----------+----------+
> > > > > >> |     X    |     1    |     0    |     1    |  |
> > > > > >> +----------+----------+----------+----------+  |
> > > > > >> |     X    |     1    |     1    |     0    |  |
> > > > > >> +----------+----------+----------+----------+  |
> > > > > >> |     X    |     1    |     1    |     1    |  v
> > > > > >> +----------+----------+----------+----------+ ---
> > > > > >> |     1    |     0    |     0    |     1    |  ^
> > > > > >> +----------+----------+----------+----------+  Desired new useful cases
> > > > > >> |     1    |     0    |     1    |     1    |  v
> > > > > >> +----------+----------+----------+----------+ ---
> > > > > >>
> > > > > >> Specifically, rows 1, 3, 5 with NDMA = 1 are valid states a user can
> > > > > >> set which are simply redundant to the NDMA = 0 cases.      
> > > > > > It seems right
> > > > > >      
> > > > > >> Row 6 remains invalid due to lack of support for pre-copy (_RESUMING
> > > > > >> | _RUNNING) and therefore cannot be set by userspace.  Rows 7 & 8
> > > > > >> are error states and cannot be set by userspace.      
> > > > > > I wonder, did Yishai's series capture this row 6 restriction? Yishai?      
> > > > > 
> > > > > 
> > > > > It seems so,  by using the below check which includes the 
> > > > > !VFIO_DEVICE_STATE_VALID clause.
> > > > > 
> > > > > if (old_state == VFIO_DEVICE_STATE_ERROR ||
> > > > >          !VFIO_DEVICE_STATE_VALID(state) ||
> > > > >          (state & ~MLX5VF_SUPPORTED_DEVICE_STATES))
> > > > >          return -EINVAL;
> > > > > 
> > > > > Which is:
> > > > > 
> > > > > #define VFIO_DEVICE_STATE_VALID(state) \
> > > > >      (state & VFIO_DEVICE_STATE_RESUMING ? \
> > > > >      (state & VFIO_DEVICE_STATE_MASK) == VFIO_DEVICE_STATE_RESUMING : 1)
> > > > >     
> > > > > >      
> > > > > >> Like other bits, setting the bit should be effective at the completion
> > > > > >> of writing device state.  Therefore the device would need to flush any
> > > > > >> outbound DMA queues before returning.      
> > > > > > Yes, the device commands are expected to achieve this.
> > > > > >      
> > > > > >> The question I was really trying to get to though is whether we have a
> > > > > >> supportable interface without such an extension.  There's currently
> > > > > >> only an experimental version of vfio migration support for PCI devices
> > > > > >> in QEMU (afaik),      
> > > > > > If I recall this only matters if you have a VM that is causing
> > > > > > migratable devices to interact with each other. So long as the devices
> > > > > > are only interacting with the CPU this extra step is not strictly
> > > > > > needed.
> > > > > >
> > > > > > So, single device cases can be fine as-is
> > > > > >
> > > > > > IMHO the multi-device case the VMM should probably demand this support
> > > > > > from the migration drivers, otherwise it cannot know if it is safe for
> > > > > > sure.
> > > > > >
> > > > > > A config option to override the block if the admin knows there is no
> > > > > > use case to cause devices to interact - eg two NVMe devices without
> > > > > > CMB do not have a useful interaction.
> > > > > >      
> > > > > >> so it seems like we could make use of the bus-master bit to fill
> > > > > >> this gap in QEMU currently, before we claim non-experimental
> > > > > >> support, but this new device agnostic extension would be required
> > > > > >> for non-PCI device support (and PCI support should adopt it as
> > > > > >> available).  Does that sound right?  Thanks,      
> > > > > > I don't think the bus master support is really a substitute, tripping
> > > > > > bus master will stop DMA but it will not do so in a clean way and is
> > > > > > likely to be non-transparent to the VM's driver.
> > > > > >
> > > > > > The single-device-assigned case is a cleaner restriction, IMHO.
> > > > > >
> > > > > > Alternatively we can add the 4th bit and insist that migration drivers
> > > > > > support all the states. I'm just unsure what other HW can do, I get
> > > > > > the feeling people have been designing to the migration description in
> > > > > > the header file for a while and this is a new idea.    
> > > > 
> > > > I'm wondering if we're imposing extra requirements on the !_RUNNING
> > > > state that don't need to be there.  For example, if we can assume that
> > > > all devices within a userspace context are !_RUNNING before any of the
> > > > devices begin to retrieve final state, then clearing of the _RUNNING
> > > > bit becomes the device quiesce point and the beginning of reading
> > > > device data is the point at which the device state is frozen and
> > > > serialized.  No new states required and essentially works with a slight
> > > > rearrangement of the callbacks in this series.  Why can't we do that?    
> > > 
> > > So without me actually understanding your bit encodings that closely, I
> > > think the problem is we have to asusme that any transition takes time.
> > > From the QEMU point of view I think the requirement is when we stop the
> > > machine (vm_stop_force_state(RUN_STATE_FINISH_MIGRATE) in
> > > migration_completion) that at the point that call returns (with no
> > > error) all devices are idle.  That means you need a way to command the
> > > device to go into the stopped state, and probably another to make sure
> > > it's got there.  
> > 
> > In a way.  We're essentially recognizing that we cannot stop a single
> > device in isolation of others that might participate in peer-to-peer
> > DMA with that device, so we need to make a pass to quiesce each device
> > before we can ask the device to fully stop.  This new device state bit
> > is meant to be that quiescent point, devices can accept incoming DMA
> > but should cease to generate any.  Once all device are quiesced then we
> > can safely stop them.  
> 
> It may need some further refinement; for example in that quiesed state
> do counters still tick? will a NIC still respond to packets that don't
> get forwarded to the host?

I'd think no, but I imagine it's largely device specific to what extent
a device can be fully halted yet minimally handle incoming DMA.

> Note I still think you need a way to know when you have actually reached
> these states; setting a bit in a register is asking nicely for a device
> to go into a state - has it got there?

It's more than asking nicely, we define the device_state bits as
synchronous, the device needs to enter the state before returning from
the write operation or return an errno.

> > > Now, you could be a *little* more sloppy; you could allow a device carry
> > > on doing stuff purely with it's own internal state up until the point
> > > it needs to serialise; but that would have to be strictly internal state
> > > only - if it can change any other devices state (or issue an interrupt,
> > > change RAM etc) then you get into ordering issues on the serialisation
> > > of multiple devices.  
> > 
> > Yep, that's the proposal that doesn't require a uAPI change, we loosen
> > the definition of stopped to mean the device can no longer generate DMA
> > or interrupts and all internal processing outside or responding to
> > incoming DMA should halt (essentially the same as the new quiescent
> > state above).  Once all devices are in this state, there should be no
> > incoming DMA and we can safely collect per device migration data.  If
> > state changes occur beyond the point in time where userspace has
> > initiated the collection of migration data, drivers have options for
> > generating errors when userspace consumes that data.  
> 
> How do you know that last device has actually gone into that state?

Each device cannot, the burden is on the user to make sure all devices
are stopped before proceeding to read migration data.

> Also be careful; it feels much more delicate where something might
> accidentally start a transaction.

This sounds like a discussion of theoretically broken drivers.  Like
the above device_state, drivers still have a synchronization point when
the user reads the pending_bytes field to initiate retrieving the
device state.  If the implementation requires the device to be fully
stopped to snapshot the device state to provide to the user, this is
where that would happen.  Thanks,

Alex