Re: [Intel-xe] [RFC PATCH] Documentation/gpu: Draft VM_BIND locking document
From: Thomas Hellström
Date: Fri Aug 11 2023 - 09:32:31 EST
Thanks for reviewing, Rodrigo!
On 8/4/23 22:15, Rodrigo Vivi wrote:
On Fri, Jun 30, 2023 at 06:44:52PM +0200, Thomas Hellström wrote:
Add the first version of the VM_BIND locking document which is
intended to be part of the xe driver upstreaming agreement.
The document describes and discuss the locking used during exec-
functions, evicton and for userptr gmvas. Intention is to be using the
same nomenclature as the drm-vm-bind-async.rst, but to keep naming a
little shorter, use gvm and gmva instead of gpu_vm and gpu_vma which
is used in the previous document, with an intention to modify also
that document.
I preferred the gpu_vm and gpu_vma as written in the async doc.
Much easier to read imho.
OK. I'll keep that naming then.
Signed-off-by: Thomas Hellström <thomas.hellstrom@xxxxxxxxxxxxxxx>
---
Documentation/gpu/drm-vm-bind-locking.rst | 339 ++++++++++++++++++++++
1 file changed, 339 insertions(+)
create mode 100644 Documentation/gpu/drm-vm-bind-locking.rst
diff --git a/Documentation/gpu/drm-vm-bind-locking.rst b/Documentation/gpu/drm-vm-bind-locking.rst
new file mode 100644
index 000000000000..f5d1a40a2906
--- /dev/null
+++ b/Documentation/gpu/drm-vm-bind-locking.rst
@@ -0,0 +1,339 @@
+===============
+VM_BIND lockinghttps://jira.devtools.intel.com/browse/VLK-50299a
+===============
+
+This document attempts to describe what's needed to get VM_BIND locking right,
+including the userptr mmu_notifier locking and it will also discuss some
+optimizations to get rid of the looping through of all userptr mappings and
+external / shared object mappings that is needed in the simplest
+implementation. It will also discuss some implications for faulting gvms.
+
+Nomenclature
+============
+
+* ``Context``: GPU execution context.
+* ``gvm``: Abstraction of a GPU address space with meta-data. Typically
+ one per client (DRM file-private), or one per context.
+* ``gvma``: Abstraction of a GPU address range within a gvma with
within a gpu_vm you meant?
Yes.
+ associated meta-data. The backing storage of a gvma can either be
+ a gem buffer object or anonymous pages mapped also into the CPU
+ address space for the process.
+* ``userptr gvma or just userptr``: A gvma, the backing store of
+ which is anonymous pages as described above.
+* ``revalidating``: Revalidating a gvma means making the latest version
+ of the backing store resident and making sure the gvma's
+ page-table entries point to that backing store.
+* ``dma_fence``: A struct dma_fence that is similar to a struct completion
+ and which tracks GPU activity. When the GPU activity is finished,
+ the dma_fence signals.
+* ``dma_resv``: A struct dma_resv (AKA reservation object) that is used
+ to track GPU activity in the form of multiple dma_fences on a
+ gvm or a gem buffer object. The dma_resv contains an array / list
+ of dma_fences and a lock that needs to be held when adding
+ additional dma_fences to the dma_resv. The lock is of a type that
+ allows deadlock-safe locking of multiple dma_resvs in arbitrary order.
+* ``exec function``: An exec function is a function that revalidates all
+ affected gvmas, 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.
+* ``local object``: A GEM object which is local to a gvm. Shared gem
+ objects also share the gvm's dma_resv.
+* ``shared object``: AKA external object: A GEM object which may be shared
+ by multiple gvms and whose backing storage may be shared with
+ other drivers.
+
+
+Introducing the locks
+=====================
+
+One of the benefits of VM_BIND is that local GEM objects share the gvm's
+dma_resv object and hence the dma_resv lock. So even with a huge
+number of local GEM objects, only one lock is needed to make the exec
+sequence atomic.
+
+The following locks and locking orders are used:
+
+* The ``gvm->lock`` (optionally an rwsem). Protects how the gvm is
+ partitioned into gvmas, protects the gvm's list of external objects,
+ and can also with some simplification protect the gvm's list of
+ userptr gvmas. With the CPU mm analogy this would correspond to the
+ mmap_lock.
+* The ``userptr_seqlock``. This lock is taken in read mode for each
+ userptr gvma on the gvm's userptr list, and in write mode during mmu
+ notifier invalidation.
is this something that exists withing the mmu_notifier or a new lock
when handling the notifier?
No it's the MMU interval notifier functionality that acts like a
seqlock: The read side waits until there are no writers (invalidators)
until entering a critical section and reruns the read-side critical
section if there was a writer sneaking in before leaving the read-side
critical section. I'll add a pointer to that documentation.
+* The ``gvm->resv`` lock. Protects the gvm's list of gvmas needing
+ rebinding, and also the residency of all the gvm's local GEM object.
+* The ``gvm->userptr_notifier_lock``. This is an rwsem that is taken in read
+ mode during exec and write mode during a mmu notifier invalidation. In
+ the absence of a separate page-table lock, this lock can serve
+ together with the gvm's dma_resv lock as a page-table lock. More on
+ this below. The userptr notifier lock is per gvm.
is the userptr_seqlock also per gpu_vm?
and what's the difference from this and the other?
This lock ensures atomicity when checking the seqlock sequnce and
submitting a batch job / publishing a dma-fence, so that either the
read-side reruns or the write-side waits for the published fence. This
is also more generically described in the MMU notifier documentation.
Will update and clarify. The userptr_seqlock is per userptr vma.
+* The ``gvm->page_table_lock``. Protects the gvm's page-table updates. For
+ simplicity the gvm's dma_resv lock can be reused as page-table lock.
+
+There are certain optimizations described below that require
+additional locks. More on that later.
+
+.. code-block:: C
+
+ dma_resv_lock(&gvm->resv);
+
+ for_each_gvma_on_revalidate_list(gvm, &gvma) {
+ revalidate_gvma(&gvma);
+ remove_from_revalidate_list(&gvma);
+ }
+
+ add_dependencies(&gpu_job, &gvm->resv);
+ job_dma_fence = gpu_submit(&gpu_job));
+
+ add_dma_fence(job_dma_fence, &gvm->resv);
+ dma_resv_unlock(&gvm->resv);
+
+Eviction of one of these local objects will then be something like the
+following:
+
+.. code-block:: C
+
+ obj = get_object_from_lru();
+
+ dma_resv_lock(obj->resv);
+ for_each_gvma_of_obj(obj, &gvma);
+ put_gvma_on_revalidate_list(&gvma);
+
+ add_dependencies(&eviction_job, &obj->resv);
+ job_dma_fence = gpu_submit(&eviction_job);
+ add_dma_fence(&obj->resv, job_dma_fence);
+
+ dma_resv_unlock(&obj->resv);
+ put_object(obj);
+
+Note that since the object is local to the gvm, it will share the gvm's
+``dma_resv`` lock so that ``obj->resv == gvm->resv``. Invalidated gvmas are put
+on the gvm's revalidation list, which is protected by ``gvm->resv``, which
+is always locked while evicting, due to the above equality.
+
+Does the gvma need to be unbound before eviction? For VM_BIND gvms
+the answer is no. Since the eviction blit or copy will wait for GPU
+idle, any attempt by the GPU to access freed memory through the
+gvma will be preceded by a new exec function, which will
+make sure the gvma is revalidated, that is not an issue.
The question opening the phrase made me think this was an open,
but it more like an answer for a common question? Should we rephrase
that to an affirmative note?
Sure.
+
+Introducing external (or shared) buffer objects
+===============================================
+
+Since shared buffer objects may be shared by multiple gvm's they
+can't share their reservation object with a single gvm, but will rather
+have a reservation object of their own. The shared objects bound to a
+gvm using one or many
+gvmas are therefore typically put on a per-gvm list which is
+protected by the gvm lock. One could in theory protect it also with
+the ``gvm->resv``, but since the list of dma_resvs to take is typically
+built before the ``gvm->resv`` is locked due to a limitation in
+the current locking helpers, that is typically not done. Also see
+below for userptr gvmas.
+
+At eviction time we now need to invalidate *all* gvmas of a shared
+object, but we can no longer be certain that we hold the gvm's
+dma_resv of all the object's gvmas. We can only be certain that we
+hold the object's private dma_resv. We can trylock the dma_resvs for
+the affected gvm's but that might be unnecessarily complex. If we
+have a ww_acquire context at hand at eviction time we can also perform
+sleeping locks of those dma_resvs but that could cause expensive
+rollbacks. One option is to just mark the invalidated gvmas with a bool
+which is inspected on the next exec function, when the gvm's
+dma_resv and the object's dma_resv is held, and the invalidated
+gvmas could then be put on the gvm's list of invalidated
+gvmas. That bool would then, although being per-gvma formally be
+protected by the object's dma_resv.
+
+The exec function would then look something like the following:
+
+.. code-block:: C
+
+ read_lock(&gvm->lock);
+
+ dma_resv_lock(&gvm->resv);
+
+ // Shared object list is protected by the gvm->lock.
+ for_each_shared_obj(gvm, &obj) {
+ dma_resv_lock(&obj->resv);
+ move_marked_gvmas_to_revalidate_gvma_list(obj, &gvm);
+ }
+
+ for_each_gvma_to_revalidate(gvm, &gvma) {
+ revalidate_gvma(&gvma);
+ remove_from_revalidate_list(&gvma);
+ }
+
+ add_dependencies(&gpu_job, &gvm->resv);
+ job_dma_fence = gpu_submit(&gpu_job));
+
+ add_dma_fence(job_dma_fence, &gvm->resv);
+ for_each_shared_obj(gvm, &obj)
+ add_dma_fence(job_dma_fence, &obj->resv);
+ dma_resv_unlock_all_resv_locks();
+
+ read_unlock(&gvm->lock);
+
+And the corresponding shared-object aware eviction would look like:
+
+.. code-block:: C
+
+ obj = get_object_from_lru();
+
+ dma_resv_lock(obj->resv);
+ for_each_gvma_of_obj(obj, &gvma);
+ if (object_is_vm_local(obj))
+ put_gvma_on_revalidate_list(&gvma, &gvm);
+ else
+ mark_gvma_for_revalidation(&gvma);
+
+ add_dependencies(&eviction_job, &obj->resv);
+ job_dma_fence = gpu_submit(&eviction_job);
+ add_dma_fence(&obj->resv, job_dma_fence);
+
+ dma_resv_unlock(&obj->resv);
+ put_object(obj);
+
+Yet another option is to put the gvmas to be invalidated on a separate
+gvm list protected by a lower level lock that can be taken both at eviction
+time and at transfer-to-revalidate list time. The details are not in
+this document, but this for reference implemented in the Intel xe
+driver.
is this part of what we need to rethink for the suspend/resume evictions?
No it's not related.
+
+Introducing userptr gvmas
+=========================
+
+A userptr gvma is a gvma that, instead of mapping a buffer object to a
+GPU virtual address range, directly maps a CPU mm range of anonymous-
+or file page-cache pages.
+A very simple approach would be to just pin the pages using
+pin_user_pages() at bind time and unpin them at unbind time, but this
+creates a Denial-Of-Service vector since a single user-space process
+would be able to pin down all of system memory, which is not
+desirable. (For special use-cases and with proper accounting pinning might
+still be a desirable feature, though). What we need to do in the general case is
+to obtain a reference to the desired pages, make sure we are notified
+using a MMU notifier just before the CPU mm unmaps the pages, dirty
+them if they are not mapped read-only to the GPU, and then drop the reference.
+When we are notified by the MMU notifier that CPU mm is about to drop the
+pages, we need to stop GPU access to the pages,
+GPU page-table and make sure that before the next time the GPU tries to access
+whatever is now present in the CPU mm range, we unmap the old pages
+from the GPU page tables and repeat the process of obtaining new page
+references. Note that when the core mm decides to laundry pages, we get such
+an unmap MMU notification and can mark the pages dirty again before the
+next GPU access. We also get similar MMU notifications for NUMA accounting
+which the GPU driver doesn't really need to care about, but so far
+it's proven difficult to exclude certain notifications.
+
+Using a MMU notifier for device DMA (and other methods) is described in
+`this document
+<https://docs.kernel.org/core-api/pin_user_pages.html#case-3-mmu-notifier-registration-with-or-without-page-faulting-hardware>`_.
+
+Now the method of obtaining struct page references using
+get_user_pages() unfortunately can't be used under a dma_resv lock
+since that would violate the locking order of the dma_resv lock vs the
+mmap_lock that is grabbed when resolving a CPU pagefault. This means the gvm's
+list of userptr gvmas needs to be protected by an outer lock, and this
+is the first time we strictly need the gvm->lock. While it was
+previously used also to protect the list of the gvm's shared objects,
+we could in theory have used the gvm->resv for that.
+
+The MMU interval seqlock for a userptr gvma is used in the following
+way:
+
+.. code-block:: C
+
+ down_read(&gvm->lock);
+
+ retry:
+
+ // Note: mmu_interval_read_begin() blocks until there is no
+ // invalidation notifier running anymore.
+ seq = mmu_interval_read_begin(&gvma->userptr_interval);
+ if (seq != gvma->saved_seq) {
+ obtain_new_page_pointers(&gvma);
+ dma_resv_lock(&gvm->resv);
+ put_gvma_on_revalidate_list(&gvma, &gvm);
+ dma_resv_unlock(&gvm->resv);
+ gvma->saved_seq = seq;
+ }
+
+ // The usual revalidation goes here.
+
+ // Final userptr sequence validation may not happen before the
+ // submission dma_fence is added to the gvm's resv, from the POW
+ // of the MMU invalidation notifier. Hence the
+ // userptr_notifier_lock that will make them appear atomic.
+
+ add_dependencies(&gpu_job, &gvm->resv);
+ down_read(&gvm->userptr_notifier_lock);
+ if (mmu_interval_read_retry(&gvma->userptr_interval, gvma->saved_seq)) {
+ up_read(&gvm->userptr_notifier_lock);
+ goto retry;
+ }
+
+ job_dma_fence = gpu_submit(&gpu_job));
+
+ add_dma_fence(job_dma_fence, &gvm->resv);
+
+ for_each_shared_obj(gvm, &obj)
+ add_dma_fence(job_dma_fence, &obj->resv);
+
+ dma_resv_unlock_all_resv_locks();
+ up_read(&gvm->userptr_notifier_lock);
+ up_read(&gvm->lock);
+
+The code between ``mmu_interval_read_begin()`` and the
+``mmu_interval_read_retry()`` marks the read side critical section of
+what we call the ``userptr_seqlock``. In reality the gvm's userptr
+gvma list is looped through, and the check is done for *all* of its
+userptr gvmas, although we only show a single one here.
+
+The userptr gvma MMU invalidation notifier might be called from
+reclaim context and, again to avoid locking order violations, we can't
+take any dma_resv lock nor the gvm->lock from within it.
+
+.. code-block:: C
+
+ bool gvma_userptr_invalidate(userptr_interval, cur_seq)
+ {
+ // Make sure the exec function either sees the new sequence
+ // and backs off or we wait for the dma-fence:
+
+ down_write(&gvm->userptr_notifier_lock);
+ mmu_interval_set_seq(userptr_interval, cur_seq);
+ up_write(&gvm->userptr_notifier_lock);
+
+ dma_resv_wait_timeout(&gvm->resv, DMA_RESV_USAGE_BOOKKEEP,
+ false, MAX_SCHEDULE_TIMEOUT);
+ return true;
+ }
+
+When this invalidation notifier returns, the GPU can no longer be
+accessing the old pages of the userptr gvma and needs to redo the page-binding
+before a new GPU submission can succeed.
+
+Optimizing gvma iteration
+-------------------------
+
+Iterating through all of a gvm's userptr gvmas to check the validity
+on each exec function may be very costly. There is a scheme to avoid
+this and only iterate through the userptr gvmas that actually saw an
+invalidation notifier call since the last exec. T
+
+TODO: describe that scheme here. It's implemented in the xe driver.
+
+Locking for page-table updates at bind- and unbind time
+=======================================================
+
+TODO.
+
+Recoverable page-fault implications
+===================================
+
+TODO.
--
2.40.1