[Intel-gfx] [RFC v3 1/3] drm/doc/rfc: VM_BIND feature design document

[Zanoni, Paulo R]

On Tue, 2022-05-17 at 11:32 -0700, Niranjana Vishwanathapura wrote:
> VM_BIND design document with description of intended use cases.
> 
> v2: Add more documentation and format as per review comments
>     from Daniel.
> 
> Signed-off-by: Niranjana Vishwanathapura <niranjana.vishwanathapura at intel.com>
> ---
> 
> diff --git a/Documentation/gpu/rfc/i915_vm_bind.rst b/Documentation/gpu/rfc/i915_vm_bind.rst
> new file mode 100644
> index 000000000000..f1be560d313c
> --- /dev/null
> +++ b/Documentation/gpu/rfc/i915_vm_bind.rst
> @@ -0,0 +1,304 @@
> +==========================================
> +I915 VM_BIND feature design and use cases
> +==========================================
> +
> +VM_BIND feature
> +================
> +DRM_I915_GEM_VM_BIND/UNBIND ioctls allows UMD to bind/unbind GEM buffer
> +objects (BOs) or sections of a BOs at specified GPU virtual addresses on a
> +specified address space (VM). These mappings (also referred to as persistent
> +mappings) will be persistent across multiple GPU submissions (execbuff calls)
> +issued by the UMD, without user having to provide a list of all required
> +mappings during each submission (as required by older execbuff mode).
> +
> +VM_BIND/UNBIND ioctls will support 'in' and 'out' fences to allow userpace
> +to specify how the binding/unbinding should sync with other operations
> +like the GPU job submission. These fences will be timeline 'drm_syncobj's
> +for non-Compute contexts (See struct drm_i915_vm_bind_ext_timeline_fences).
> +For Compute contexts, they will be user/memory fences (See struct
> +drm_i915_vm_bind_ext_user_fence).
> +
> +VM_BIND feature is advertised to user via I915_PARAM_HAS_VM_BIND.
> +User has to opt-in for VM_BIND mode of binding for an address space (VM)
> +during VM creation time via I915_VM_CREATE_FLAGS_USE_VM_BIND extension.
> +
> +VM_BIND/UNBIND ioctl will immediately start binding/unbinding the mapping in an
> +async worker. The binding and unbinding will work like a special GPU engine.
> +The binding and unbinding operations are serialized and will wait on specified
> +input fences before the operation and will signal the output fences upon the
> +completion of the operation. Due to serialization, completion of an operation
> +will also indicate that all previous operations are also complete.
> +
> +VM_BIND features include:
> +
> +* Multiple Virtual Address (VA) mappings can map to the same physical pages
> +  of an object (aliasing).
> +* VA mapping can map to a partial section of the BO (partial binding).
> +* Support capture of persistent mappings in the dump upon GPU error.
> +* TLB is flushed upon unbind completion. Batching of TLB flushes in some
> +  use cases will be helpful.
> +* Asynchronous vm_bind and vm_unbind support with 'in' and 'out' fences.
> +* Support for userptr gem objects (no special uapi is required for this).
> +
> +Execbuff ioctl in VM_BIND mode
> +-------------------------------
> +The execbuff ioctl handling in VM_BIND mode differs significantly from the
> +older method. A VM in VM_BIND mode will not support older execbuff mode of
> +binding. In VM_BIND mode, execbuff ioctl will not accept any execlist. Hence,
> +no support for implicit sync. It is expected that the below work will be able
> +to support requirements of object dependency setting in all use cases:
> +
> +"dma-buf: Add an API for exporting sync files"
> +(https://lwn.net/Articles/859290/)

I would really like to have more details here. The link provided points
to new ioctls and we're not very familiar with those yet, so I think
you should really clarify the interaction between the new additions
here. Having some sample code would be really nice too.

For Mesa at least (and I believe for the other drivers too) we always
have a few exported buffers in every execbuf call, and we rely on the
implicit synchronization provided by execbuf to make sure everything
works. The execbuf ioctl also has some code to flush caches during
implicit synchronization AFAIR, so I would guess we rely on it too and
whatever else the Kernel does. Is that covered by the new ioctls?

In addition, as far as I remember, one of the big improvements of
vm_bind was that it would help reduce ioctl latency and cpu overhead.
But if making execbuf faster comes at the cost of requiring additional
ioctls calls for implicit synchronization, which is required on ever
execbuf call, then I wonder if we'll even get any faster at all.
Comparing old execbuf vs plain new execbuf without the new required
ioctls won't make sense.

But maybe I'm wrong and we won't need to call these new ioctls around
every single execbuf ioctl we submit? Again, more clarification and
some code examples here would be really nice. This is a big change on
an important part of the API, we should clarify the new expected usage.

> +
> +This also means, we need an execbuff extension to pass in the batch
> +buffer addresses (See struct drm_i915_gem_execbuffer_ext_batch_addresses).
> +
> +If at all execlist support in execbuff ioctl is deemed necessary for
> +implicit sync in certain use cases, then support can be added later.

IMHO we really need to sort this and check all the assumptions before
we commit to any interface. Again, implicit synchronization is
something we rely on during *every* execbuf ioctl for most workloads.


> +In VM_BIND mode, VA allocation is completely managed by the user instead of
> +the i915 driver. Hence all VA assignment, eviction are not applicable in
> +VM_BIND mode. Also, for determining object activeness, VM_BIND mode will not
> +be using the i915_vma active reference tracking. It will instead use dma-resv
> +object for that (See `VM_BIND dma_resv usage`_).
> +
> +So, a lot of existing code in the execbuff path like relocations, VA evictions,
> +vma lookup table, implicit sync, vma active reference tracking etc., are not
> +applicable in VM_BIND mode. Hence, the execbuff path needs to be cleaned up
> +by clearly separating out the functionalities where the VM_BIND mode differs
> +from older method and they should be moved to separate files.

I seem to recall some conversations where we were told a bunch of
ioctls would stop working or make no sense to call when using vm_bind.
Can we please get a complete list of those? Bonus points if the Kernel
starts telling us we just called something that makes no sense.

> +
> +VM_PRIVATE objects
> +-------------------
> +By default, BOs can be mapped on multiple VMs and can also be dma-buf
> +exported. Hence these BOs are referred to as Shared BOs.
> +During each execbuff submission, the request fence must be added to the
> +dma-resv fence list of all shared BOs mapped on the VM.
> +
> +VM_BIND feature introduces an optimization where user can create BO which
> +is private to a specified VM via I915_GEM_CREATE_EXT_VM_PRIVATE flag during
> +BO creation. Unlike Shared BOs, these VM private BOs can only be mapped on
> +the VM they are private to and can't be dma-buf exported.
> +All private BOs of a VM share the dma-resv object. Hence during each execbuff
> +submission, they need only one dma-resv fence list updated. Thus, the fast
> +path (where required mappings are already bound) submission latency is O(1)
> +w.r.t the number of VM private BOs.

I know we already discussed this, but just to document it publicly: the
ideal case for user space would be that every BO is created as private
but then we'd have an ioctl to convert it to non-private (without the
need to have a non-private->private interface).

An explanation on why we can't have an ioctl to mark as exported a
buffer that was previously vm_private would be really appreciated.

Thanks,
Paulo


> +
> +VM_BIND locking hirarchy
> +-------------------------
> +The locking design here supports the older (execlist based) execbuff mode, the
> +newer VM_BIND mode, the VM_BIND mode with GPU page faults and possible future
> +system allocator support (See `Shared Virtual Memory (SVM) support`_).
> +The older execbuff mode and the newer VM_BIND mode without page faults manages
> +residency of backing storage using dma_fence. The VM_BIND mode with page faults
> +and the system allocator support do not use any dma_fence at all.
> +
> +VM_BIND locking order is as below.
> +
> +1) Lock-A: A vm_bind mutex will protect vm_bind lists. This lock is taken in
> +   vm_bind/vm_unbind ioctl calls, in the execbuff path and while releasing the
> +   mapping.
> +
> +   In future, when GPU page faults are supported, we can potentially use a
> +   rwsem instead, so that multiple page fault handlers can take the read side
> +   lock to lookup the mapping and hence can run in parallel.
> +   The older execbuff mode of binding do not need this lock.
> +
> +2) Lock-B: The object's dma-resv lock will protect i915_vma state and needs to
> +   be held while binding/unbinding a vma in the async worker and while updating
> +   dma-resv fence list of an object. Note that private BOs of a VM will all
> +   share a dma-resv object.
> +
> +   The future system allocator support will use the HMM prescribed locking
> +   instead.
> +
> +3) Lock-C: Spinlock/s to protect some of the VM's lists like the list of
> +   invalidated vmas (due to eviction and userptr invalidation) etc.
> +
> +When GPU page faults are supported, the execbuff path do not take any of these
> +locks. There we will simply smash the new batch buffer address into the ring and
> +then tell the scheduler run that. The lock taking only happens from the page
> +fault handler, where we take lock-A in read mode, whichever lock-B we need to
> +find the backing storage (dma_resv lock for gem objects, and hmm/core mm for
> +system allocator) and some additional locks (lock-D) for taking care of page
> +table races. Page fault mode should not need to ever manipulate the vm lists,
> +so won't ever need lock-C.
> +
> +VM_BIND LRU handling
> +---------------------
> +We need to ensure VM_BIND mapped objects are properly LRU tagged to avoid
> +performance degradation. We will also need support for bulk LRU movement of
> +VM_BIND objects to avoid additional latencies in execbuff path.
> +
> +The page table pages are similar to VM_BIND mapped objects (See
> +`Evictable page table allocations`_) and are maintained per VM and needs to
> +be pinned in memory when VM is made active (ie., upon an execbuff call with
> +that VM). So, bulk LRU movement of page table pages is also needed.
> +
> +The i915 shrinker LRU has stopped being an LRU. So, it should also be moved
> +over to the ttm LRU in some fashion to make sure we once again have a reasonable
> +and consistent memory aging and reclaim architecture.
> +
> +VM_BIND dma_resv usage
> +-----------------------
> +Fences needs to be added to all VM_BIND mapped objects. During each execbuff
> +submission, they are added with DMA_RESV_USAGE_BOOKKEEP usage to prevent
> +over sync (See enum dma_resv_usage). One can override it with either
> +DMA_RESV_USAGE_READ or DMA_RESV_USAGE_WRITE usage during object dependency
> +setting (either through explicit or implicit mechanism).
> +
> +When vm_bind is called for a non-private object while the VM is already
> +active, the fences need to be copied from VM's shared dma-resv object
> +(common to all private objects of the VM) to this non-private object.
> +If this results in performance degradation, then some optimization will
> +be needed here. This is not a problem for VM's private objects as they use
> +shared dma-resv object which is always updated on each execbuff submission.
> +
> +Also, in VM_BIND mode, use dma-resv apis for determining object activeness
> +(See dma_resv_test_signaled() and dma_resv_wait_timeout()) and do not use the
> +older i915_vma active reference tracking which is deprecated. This should be
> +easier to get it working with the current TTM backend. We can remove the
> +i915_vma active reference tracking fully while supporting TTM backend for igfx.
> +
> +Evictable page table allocations
> +---------------------------------
> +Make pagetable allocations evictable and manage them similar to VM_BIND
> +mapped objects. Page table pages are similar to persistent mappings of a
> +VM (difference here are that the page table pages will not have an i915_vma
> +structure and after swapping pages back in, parent page link needs to be
> +updated).
> +
> +Mesa use case
> +--------------
> +VM_BIND can potentially reduce the CPU overhead in Mesa (both Vulkan and Iris),
> +hence improving performance of CPU-bound applications. It also allows us to
> +implement Vulkan's Sparse Resources. With increasing GPU hardware performance,
> +reducing CPU overhead becomes more impactful.
> +
> +
> +VM_BIND Compute support
> +========================
> +
> +User/Memory Fence
> +------------------
> +The idea is to take a user specified virtual address and install an interrupt
> +handler to wake up the current task when the memory location passes the user
> +supplied filter. User/Memory fence is a <address, value> pair. To signal the
> +user fence, specified value will be written at the specified virtual address
> +and wakeup the waiting process. User can wait on a user fence with the
> +gem_wait_user_fence ioctl.
> +
> +It also allows the user to emit their own MI_FLUSH/PIPE_CONTROL notify
> +interrupt within their batches after updating the value to have sub-batch
> +precision on the wakeup. Each batch can signal a user fence to indicate
> +the completion of next level batch. The completion of very first level batch
> +needs to be signaled by the command streamer. The user must provide the
> +user/memory fence for this via the DRM_I915_GEM_EXECBUFFER_EXT_USER_FENCE
> +extension of execbuff ioctl, so that KMD can setup the command streamer to
> +signal it.
> +
> +User/Memory fence can also be supplied to the kernel driver to signal/wake up
> +the user process after completion of an asynchronous operation.
> +
> +When VM_BIND ioctl was provided with a user/memory fence via the
> +I915_VM_BIND_EXT_USER_FENCE extension, it will be signaled upon the completion
> +of binding of that mapping. All async binds/unbinds are serialized, hence
> +signaling of user/memory fence also indicate the completion of all previous
> +binds/unbinds.
> +
> +This feature will be derived from the below original work:
> +https://patchwork.freedesktop.org/patch/349417/
> +
> +Long running Compute contexts
> +------------------------------
> +Usage of dma-fence expects that they complete in reasonable amount of time.
> +Compute on the other hand can be long running. Hence it is appropriate for
> +compute to use user/memory fence and dma-fence usage will be limited to
> +in-kernel consumption only. This requires an execbuff uapi extension to pass
> +in user fence (See struct drm_i915_vm_bind_ext_user_fence). Compute must opt-in
> +for this mechanism with I915_CONTEXT_CREATE_FLAGS_LONG_RUNNING flag during
> +context creation. The dma-fence based user interfaces like gem_wait ioctl and
> +execbuff out fence are not allowed on long running contexts. Implicit sync is
> +not valid as well and is anyway not supported in VM_BIND mode.
> +
> +Where GPU page faults are not available, kernel driver upon buffer invalidation
> +will initiate a suspend (preemption) of long running context with a dma-fence
> +attached to it. And upon completion of that suspend fence, finish the
> +invalidation, revalidate the BO and then resume the compute context. This is
> +done by having a per-context preempt fence (also called suspend fence) proxying
> +as i915_request fence. This suspend fence is enabled when someone tries to wait
> +on it, which then triggers the context preemption.
> +
> +As this support for context suspension using a preempt fence and the resume work
> +for the compute mode contexts can get tricky to get it right, it is better to
> +add this support in drm scheduler so that multiple drivers can make use of it.
> +That means, it will have a dependency on i915 drm scheduler conversion with GuC
> +scheduler backend. This should be fine, as the plan is to support compute mode
> +contexts only with GuC scheduler backend (at least initially). This is much
> +easier to support with VM_BIND mode compared to the current heavier execbuff
> +path resource attachment.
> +
> +Low Latency Submission
> +-----------------------
> +Allows compute UMD to directly submit GPU jobs instead of through execbuff
> +ioctl. This is made possible by VM_BIND is not being synchronized against
> +execbuff. VM_BIND allows bind/unbind of mappings required for the directly
> +submitted jobs.
> +
> +Other VM_BIND use cases
> +========================
> +
> +Debugger
> +---------
> +With debug event interface user space process (debugger) is able to keep track
> +of and act upon resources created by another process (debugged) and attached
> +to GPU via vm_bind interface.
> +
> +GPU page faults
> +----------------
> +GPU page faults when supported (in future), will only be supported in the
> +VM_BIND mode. While both the older execbuff mode and the newer VM_BIND mode of
> +binding will require using dma-fence to ensure residency, the GPU page faults
> +mode when supported, will not use any dma-fence as residency is purely managed
> +by installing and removing/invalidating page table entries.
> +
> +Page level hints settings
> +--------------------------
> +VM_BIND allows any hints setting per mapping instead of per BO.
> +Possible hints include read-only mapping, placement and atomicity.
> +Sub-BO level placement hint will be even more relevant with
> +upcoming GPU on-demand page fault support.
> +
> +Page level Cache/CLOS settings
> +-------------------------------
> +VM_BIND allows cache/CLOS settings per mapping instead of per BO.
> +
> +Shared Virtual Memory (SVM) support
> +------------------------------------
> +VM_BIND interface can be used to map system memory directly (without gem BO
> +abstraction) using the HMM interface. SVM is only supported with GPU page
> +faults enabled.
> +
> +
> +Broder i915 cleanups
> +=====================
> +Supporting this whole new vm_bind mode of binding which comes with its own
> +use cases to support and the locking requirements requires proper integration
> +with the existing i915 driver. This calls for some broader i915 driver
> +cleanups/simplifications for maintainability of the driver going forward.
> +Here are few things identified and are being looked into.
> +
> +- Remove vma lookup cache (eb->gem_context->handles_vma). VM_BIND feature
> +  do not use it and complexity it brings in is probably more than the
> +  performance advantage we get in legacy execbuff case.
> +- Remove vma->open_count counting
> +- Remove i915_vma active reference tracking. VM_BIND feature will not be using
> +  it. Instead use underlying BO's dma-resv fence list to determine if a i915_vma
> +  is active or not.
> +
> +
> +VM_BIND UAPI
> +=============
> +
> +.. kernel-doc:: Documentation/gpu/rfc/i915_vm_bind.h
> diff --git a/Documentation/gpu/rfc/index.rst b/Documentation/gpu/rfc/index.rst
> index 91e93a705230..7d10c36b268d 100644
> --- a/Documentation/gpu/rfc/index.rst
> +++ b/Documentation/gpu/rfc/index.rst
> @@ -23,3 +23,7 @@ host such documentation:
>  .. toctree::
>  
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
>      i915_scheduler.rst
> +
> +.. toctree::
> +
> +    i915_vm_bind.rst