[Intel-xe] [PATCH v3] Documentation/gpu: Add a VM_BIND async draft document

Thu Jun 29 07:08:46 UTC 2023

On Wed, 2023-06-28 at 19:08 +0000, Matthew Brost wrote:
> On Wed, Jun 28, 2023 at 02:51:46PM +0200, Thomas Hellström wrote:
> > Add a motivation for and description of asynchronous VM_BIND
> > operation
> > 
> > v2:
> > - Fix typos (Nirmoy Das)
> > - Improve the description of a memory fence (Oak Zeng)
> > - Add a reference to the document in the Xe RFC.
> > - Add pointers to sample uAPI suggestions
> > v3:
> > - Address review comments (Danilo Krummrich)
> > - Formatting fixes
> > 
> > Signed-off-by: Thomas Hellström <thomas.hellstrom at linux.intel.com>
> > Acked-by: Nirmoy Das <nirmoy.das at intel.com>
> > ---
> >  Documentation/gpu/drm-vm-bind-async.rst | 150
> > ++++++++++++++++++++++++
> >  Documentation/gpu/rfc/xe.rst            |   4 +-
> >  2 files changed, 152 insertions(+), 2 deletions(-)
> >  create mode 100644 Documentation/gpu/drm-vm-bind-async.rst
> > 
> > diff --git a/Documentation/gpu/drm-vm-bind-async.rst
> > b/Documentation/gpu/drm-vm-bind-async.rst
> > new file mode 100644
> > index 000000000000..8f9e2d5c8f0f
> > --- /dev/null
> > +++ b/Documentation/gpu/drm-vm-bind-async.rst
> > @@ -0,0 +1,150 @@
> > +====================
> > +Asynchronous VM_BIND
> > +====================
> > +
> > +Nomenclature:
> > +=============
> > +
> > +* ``VRAM``: On-device memory. Sometimes referred to as device
> > local memory.
> > +
> > +* ``gpu_vm``: A GPU address space. Typically per process, but can
> > be shared by
> > +  multiple processes.
> > +
> > +* ``VM_BIND``: An operation or a list of operations to modify a
> > gpu_vm using
> > +  an IOCTL. The operations include mapping and unmapping system-
> > or
> > +  VRAM memory.
> > +
> > +* ``syncobj``: A container that abstracts synchronization objects.
> > The
> > +  synchronization objects can be either generic, like dma-fences
> > or
> > +  driver specific. A syncobj typically indicates the type of the
> > +  underlying synchronization object.
> > +
> > +* ``in-syncobj``: Argument to a VM_BIND IOCTL, the VM_BIND
> > operation waits
> > +  for these before starting.
> > +
> > +* ``out-syncbj``: Argument to a VM_BIND_IOCTL, the VM_BIND
> > operation
> > +  signals these when the bind operation is complete.
> > +
> > +* ``memory fence``: A synchronization object, different from a
> > dma-fence.
> > +  A memory fence uses the value of a specified memory location to
> > determine
> > +  signaled status. A memory fence can be awaited and signaled by
> > both
> > +  the GPU and CPU. Memory fences are sometimes referred to as
> > +  user-fences, and do not necessarily bey the dma-fence rule of
> > +  signalling within a "reasonable amount of time". The kernel
> > should
> > +  thus avoid waiting for memory fences with locks held.
> > +
> > +* ``long-running workload``: A workload that may take more than
> > the
> > +  current stipulated dma-fence maximum signal delay to complete
> > and
> > +  which therefore needs to set the gpu_vm or the GPU execution
> > context in
> > +  a certain mode that disallows completion dma-fences.
> > +
> > +* ``exec function``: An exec function is a function that
> > revalidates all
> > +  affected vmas, submits a gpu command batch and registers the
> > +  dma_fence representing the gpu command's activity with all
> > affected
> > +  dma_resvs. For completeness, although not covered by this
> > document,
> > +  it's worth mentioning that an exec function may also be the
> > +  revalidation worker that is used by some drivers in compute /
> > +  long-running mode.
> > +
> > +* ``bind context``: A context identifier used for the VM_BIND
> > +  operation. VM_BIND operations that use the same bind context can
> > be
> > +  assumed, where it matters, to complete in order of submission.
> > No such
> > +  assumptions can be made for VM_BIND operations using separate
> > bind contexts.
> > +
> > +* ``UMD``: User-mode driver.
> > +
> > +* ``KMD``: Kernel-mode driver.
> > +
> > +
> > +Synchronous / Asynchronous VM_BIND operation
> > +============================================
> > +
> > +Synchronous VM_BIND
> > +___________________
> > +With Synchronous VM_BIND, the VM_BIND operations all complete
> > before the
> > +IOCTL returns. A synchronous VM_BIND takes neither in-fences nor
> > +out-fences. Synchronous VM_BIND may block and wait for GPU
> > operations;
> > +for example swapin or clearing, or even previous binds.
> > +
> > +Asynchronous VM_BIND
> > +____________________
> > +Asynchronous VM_BIND accepts both in-syncobjs and out-syncobjs.
> > While the
> > +IOCTL may return immediately, the VM_BIND operations wait for the
> > in-syncobjs
> > +before modifying the GPU page-tables, and signal the out-syncobjs
> > when
> > +the modification is done in the sense that the next exec function
> > that
> > +awaits for the out-syncobjs will see the change. Errors are
> > reported
> > +synchronously assuming that the asynchronous part of the job never
> > errors.
> > +In low-memory situations the implementation may block, performing
> > the
> > +VM_BIND synchronously, because there might not be enough memory
> > +immediately available for preparing the asynchronous operation.
> > +
> > +If the VM_BIND IOCTL takes a list or an array of operations as an
> > argument,
> > +the in-syncobjs needs to signal before the first operation starts
> > to
> > +execute, and the out-syncobjs signal after the last operation
> > +completes. Operations in the operation list can be assumed, where
> > it
> > +matters, to complete in order.
> > +
> > +To aid in supporting user-space queues, the VM_BIND may take a
> > bind context.
> > +
> > +The purpose of an Asynchronous VM_BIND operation is for user-mode
> > +drivers to be able to pipeline interleaved gpu_vm modifications
> > and
> > +exec functions. For long-running workloads, such pipelining of a
> > bind
> > +operation is not allowed and any in-fences need to be awaited
> > +synchronously.
> 
> Why? I think in Xe we allow in-fences for LR workloads + pipelining.

In-fences as memory fences need to be waited for sync before the bind
operation starts anyway, since bind operations produce dma-fences that
can never depend on memory fences. So no point in having them as in-
fences expecting pipelining.

As for dma-fences as in-fences for LR jobs, it will probably work but
since you can't pipeline behind a LR exec when we discussed this
briefly on IRC there is no point? Would anyone use this?

/Thomas

> 
> Matt
> 
> > +
> > +Also for VM_BINDS for long-running gpu_vms the user-mode driver
> > should typically
> > +select memory fences as out-fences since that gives greater
> > flexibility for
> > +the kernel mode driver to inject other  operations into the bind /
> > +unbind operations. Like for example inserting breakpoints into
> > batch
> > +buffers. The workload execution can then easily be pipelined
> > behind
> > +the bind completion using the memory out-fence as the signal
> > condition
> > +for a gpu semaphore embedded by UMD in the workload.
> > +
> > +Multi-operation VM_BIND IOCTL error handling and interrupts
> > +===========================================================
> > +
> > +The VM_BIND operations of the IOCTL may error due to lack of
> > resources
> > +to complete and also due to interrupted waits. In both situations
> > UMD
> > +should preferably restart the IOCTL after taking suitable action.
> > If
> > +UMD has overcommitted a memory resource, an -ENOSPC error will be
> > +returned, and UMD may then unbind resources that are not used at
> > the
> > +moment and restart the IOCTL. On -EINTR, UMD should simply restart
> > the
> > +IOCTL and on -ENOMEM user-space may either attempt to free known
> > +system memory resources or abort the operation. If aborting as a
> > +result of a failed operation in a list of operations, some
> > operations
> > +may still have completed, and to get back to a known state, user-
> > space
> > +should therefore attempt to unbind all virtual memory regions
> > touched
> > +by the failing IOCTL.
> > +Unbind operations are guaranteed not to cause any errors due to
> > +resource constraints.
> > +In between a failed VM_BIND IOCTL and a successful restart there
> > may
> > +be implementation defined restrictions on the use of the gpu_vm.
> > For a
> > +description why, please see `KMD implementation details`_ under
> > [error
> > +state saving]_.
> > +
> > +Sample uAPI implementations
> > +===========================
> > +Suggested uAPI implementations at the moment of writing can be
> > found for
> > +the Nouveau driver `here
> > +<
> > https://patchwork.freedesktop.org/patch/543260/?series=112994&rev=6
> > >`_.
> > +and for the Xe driver `here
> > +<
> > https://cgit.freedesktop.org/drm/drm-xe/diff/include/uapi/drm/xe_dr
> > m.h?h=drm-xe-next&id=9cb016ebbb6a275f57b1cb512b95d5a842391ad7>`_.
> > +
> > +KMD implementation details
> > +==========================
> > +
> > +Open: When the VM_BIND IOCTL returns an error, some or even parts
> > of
> > +an operation may have been completed. If the IOCTL is restarted,
> > in
> > +order to know where to restart, the KMD can either put the gpu_vm
> > in
> > +an error state and save one instance of the needed restart state
> > +internally. In this case, KMD needs to block further modifications
> > of
> > +the gpu_vm state that may cause additional failures requiring a
> > +restart state save, until the error has been fully resolved. If
> > the
> > +uAPI instead defines a pointer to a UMD allocated cookie in the
> > IOCTL
> > +struct, it could also choose to store the restart state in that
> > cookie.
> > +
> > +The restart state may, for example, be the number of successfully
> > +completed operations.
> > +
> > +Easiest for UMD would of course be if KMD did a full unwind on
> > error
> > +so that no error state needs to be saved.
> > diff --git a/Documentation/gpu/rfc/xe.rst
> > b/Documentation/gpu/rfc/xe.rst
> > index 2516fe141db6..0f062e1346d2 100644
> > --- a/Documentation/gpu/rfc/xe.rst
> > +++ b/Documentation/gpu/rfc/xe.rst
> > @@ -138,8 +138,8 @@ memory fences. Ideally with helper support so
> > people don't get it wrong in all
> >  possible ways.
> >  
> >  As a key measurable result, the benefits of ASYNC VM_BIND and a
> > discussion of
> > -various flavors, error handling and a sample API should be
> > documented here or in
> > -a separate document pointed to by this document.
> > +various flavors, error handling and sample API suggestions are
> > documented in
> > +Documentation/gpu/drm-vm-bind-async.rst
> >  
> >  Userptr integration and vm_bind
> >  -------------------------------
> > -- 
> > 2.40.1
> > 