Currently the KMD is using enum i915_cache_level to set caching policy for
buffer objects. This is flaky because the PAT index which really controls
the caching behavior in PTE has far more levels than what's defined in the
enum. In addition, the PAT index is platform dependent, having to translate
between i915_cache_level and PAT index is not reliable, and makes the code
more complicated.
From UMD's perspective there is also a necessity to set caching policy for
performance fine tuning. It's much easier for the UMD to directly use PAT
index because the behavior of each PAT index is clearly defined in Bspec.
Having the abstracted i915_cache_level sitting in between would only cause
more ambiguity. PAT is expected to work much like MOCS already works today,
and by design userspace is expected to select the index that exactly
matches the desired behavior described in the hardware specification.
For these reasons this patch replaces i915_cache_level with PAT index. Also
note, the cache_level is not completely removed yet, because the KMD still
has the need of creating buffer objects with simple cache settings such as
cached, uncached, or writethrough. For kernel objects, cache_level is used
for simplicity and backward compatibility. For Pre-gen12 platforms PAT can
have 1:1 mapping to i915_cache_level, so these two are interchangeable. see
the use of LEGACY_CACHELEVEL.
One consequence of this change is that gen8_pte_encode is no longer working
for gen12 platforms due to the fact that gen12 platforms has different PAT
definitions. In the meantime the mtl_pte_encode introduced specfically for
MTL becomes generic for all gen12 platforms. This patch renames the MTL
PTE encode function into gen12_pte_encode and apply it to all gen12. Even
though this change looks unrelated, but separating them would temporarily
break gen12 PTE encoding, thus squash them in one patch.
Special note: this patch changes the way caching behavior is controlled in
the sense that some objects are left to be managed by userspace. For such
objects we need to be careful not to change the userspace settings.There
are kerneldoc and comments added around obj->cache_coherent, cache_dirty,
and how to bypass the checkings by i915_gem_object_has_cache_level. For
full understanding, these changes need to be looked at together with the
two follow-up patches, one disables the {set|get}_caching ioctl's and the
other adds set_pat extension to the GEM_CREATE uAPI.
Bspec: 63019
Cc: Chris Wilson <chris.p.wilson@linux.intel.com>
Signed-off-by: Fei Yang <fei.yang@intel.com>
Reviewed-by: Andi Shyti <andi.shyti@linux.intel.com>
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Andi Shyti <andi.shyti@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20230509165200.1740-3-fei.yang@intel.com
UAPI Changes:
- (Build-time only, should not have any impact)
drm/i915/uapi: Replace fake flex-array with flexible-array member
"Zero-length arrays as fake flexible arrays are deprecated and we are
moving towards adopting C99 flexible-array members instead."
This is on core kernel request moving towards GCC 13.
Driver Changes:
- Fix context runtime accounting on sysfs fdinfo for heavy workloads (Tvrtko)
- Add support for OA media units on MTL (Umesh)
- Add new workarounds for Meteorlake (Daniele, Radhakrishna, Haridhar)
- Fix sysfs to read actual frequency for MTL and Gen6 and earlier
(Ashutosh)
- Synchronize i915/BIOS on C6 enabling on MTL (Vinay)
- Fix DMAR error noise due to GPU error capture (Andrej)
- Fix forcewake during BAR resize on discrete (Andrzej)
- Flush lmem contents after construction on discrete (Chris)
- Fix GuC loading timeout on systems where IFWI programs low boot
frequency (John)
- Fix race condition UAF in i915_perf_add_config_ioctl (Min)
- Sanitycheck MMIO access early in driver load and during forcewake
(Matt)
- Wakeref fixes for GuC RC error scenario and active VM tracking (Chris)
- Cancel HuC delayed load timer on reset (Daniele)
- Limit double GT reset to pre-MTL (Daniele)
- Use i915 instead of dev_priv insied the file_priv structure (Andi)
- Improve GuC load error reporting (John)
- Simplify VCS/BSD engine selection logic (Tvrtko)
- Perform uc late init after probe error injection (Andrzej)
- Fix format for perf_limit_reasons in debugfs (Vinay)
- Create per-gt debugfs files (Andi)
- Documentation and kerneldoc fixes (Nirmoy, Lee)
- Selftest improvements (Fei, Jonathan)
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
From: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/ZC6APj/feB+jBf2d@jlahtine-mobl.ger.corp.intel.com
If the move or clear operation somehow fails, and the memory underneath
is not cleared, like when moving to lmem, then we currently fallback to
memcpy or memset. However with small-BAR systems this fallback might no
longer be possible. For now we use the set_wedged sledgehammer if we
ever encounter such a scenario, and mark the object as borked to plug
any holes where access to the memory underneath can happen. Add some
basic selftests to exercise this.
v2:
- In the selftests make sure we grab the runtime pm around the reset.
Also make sure we grab the reset lock before checking if the device
is wedged, since the wedge might still be in-progress and hence the
bit might not be set yet.
- Don't wedge or put the object into an unknown state, if the request
construction fails (or similar). Just returning an error and
skipping the fallback should be safe here.
- Make sure we wedge each gt. (Thomas)
- Peek at the unknown_state in io_reserve, that way we don't have to
export or hand roll the fault_wait_for_idle. (Thomas)
- Add the missing read-side barriers for the unknown_state. (Thomas)
- Some kernel-doc fixes. (Thomas)
v3:
- Tweak the ordering of the set_wedged, also add FIXME.
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Cc: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Cc: Jon Bloomfield <jon.bloomfield@intel.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Kenneth Graunke <kenneth@whitecape.org>
Cc: Akeem G Abodunrin <akeem.g.abodunrin@intel.com>
Reviewed-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220629174350.384910-11-matthew.auld@intel.com
Instead of distingting between shared and exclusive fences specify
the fence usage while adding fences.
Rework all drivers to use this interface instead and deprecate the old one.
v2: some kerneldoc comments suggested by Daniel
v3: fix a missing case in radeon
v4: rebase on nouveau changes, fix lockdep and temporary disable warning
v5: more documentation updates
v6: separate internal dma_resv changes from this patch, avoids to
disable warning temporary, rebase on upstream changes
v7: fix missed case in lima driver, minimize changes to i915_gem_busy_ioctl
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20220407085946.744568-3-christian.koenig@amd.com
Implement async (non-blocking) unbinding by not syncing the vma before
calling unbind on the vma_resource.
Add the resulting unbind fence to the object's dma_resv from where it is
picked up by the ttm migration code.
Ideally these unbind fences should be coalesced with the migration blit
fence to avoid stalling the migration blit waiting for unbind, as they
can certainly go on in parallel, but since we don't yet have a
reasonable data structure to use to coalesce fences and attach the
resulting fence to a timeline, we defer that for now.
Note that with async unbinding, even while the unbind waits for the
preceding bind to complete before unbinding, the vma itself might have been
destroyed in the process, clearing the vma pages. Therefore we can
only allow async unbinding if we have a refcounted sg-list and keep a
refcount on that for the vma resource pages to stay intact until
binding occurs. If this condition is not met, a request for an async
unbind is diverted to a sync unbind.
v2:
- Use a separate kmem_cache for vma resources for now to isolate their
memory allocation and aid debugging.
- Move the check for vm closed to the actual unbinding thread. Regardless
of whether the vm is closed, we need the unbind fence to properly wait
for capture.
- Clear vma_res::vm on unbind and update its documentation.
v4:
- Take cache coloring into account when searching for vma resources
pending unbind. (Matthew Auld)
v5:
- Fix timeout and error check in i915_vma_resource_bind_dep_await().
- Avoid taking a reference on the object for async binding if
async unbind capable.
- Fix braces around a single-line if statement.
v6:
- Fix up the cache coloring adjustment. (Kernel test robot <lkp@intel.com>)
- Don't allow async unbinding if the vma_res pages are not the same as
the object pages. (Matthew Auld)
v7:
- s/unsigned long/u64/ in a number of places (Matthew Auld)
Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220110172219.107131-5-thomas.hellstrom@linux.intel.com
Since the gt migration code was using only a single fence for
dependencies, these were collected in a dma_fence_array. However, it
turns out that it's illegal to use some dma_fences in a dma_fence_array,
in particular other dma_fence_arrays and dma_fence_chains, and this
causes trouble for us moving forward.
Have the gt migration code instead take a const struct i915_deps for
dependencies. This means we can skip the dma_fence_array creation
and instead pass the struct i915_deps instead to circumvent the
problem.
v2:
- Make the prev_deps() function static. (kernel test robot <lkp@intel.com>)
- Update the struct i915_deps kerneldoc.
v4:
- Rebase.
Fixes: 5652df829b ("drm/i915/ttm: Update i915_gem_obj_copy_ttm() to be asynchronous")
Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211221200050.436316-2-thomas.hellstrom@linux.intel.com
Don't wait sync while migrating, but rather make the GPU blit await the
dependencies and add a moving fence to the object.
This also enables asynchronous VRAM management in that on eviction,
rather than waiting for the moving fence to expire before freeing VRAM,
it is freed immediately and the fence is stored with the VRAM manager and
handed out to newly allocated objects to await before clears and swapins,
or for kernel objects before setting up gpu vmas or mapping.
To collect dependencies before migrating, add a set of utilities that
coalesce these to a single dma_fence.
What is still missing for fully asynchronous operation is asynchronous vma
unbinding, which is still to be implemented.
This commit substantially reduces execution time in the gem_lmem_swapping
test.
v2:
- Make a couple of functions static.
v4:
- Fix some style issues (Matthew Auld)
- Audit and add more checks for ghost objects (Matthew Auld)
- Add more documentation for the i915_deps utility (Mattew Auld)
- Simplify the i915_deps_sync() function
v6:
- Re-check for fence signaled before returning -EBUSY (Matthew Auld)
- Use dma_resv_iter_is_exclusive() (Matthew Auld)
- Await all dma-resv fences before a migration blit (Matthew Auld)
Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211122214554.371864-6-thomas.hellstrom@linux.intel.com
If the initial fill blit or copy blit of an object fails, the old
content of the data might be exposed and read as soon as either CPU- or
GPU PTEs are set up to point at the pages.
Intercept the blit fence with an async callback that checks the
blit fence for errors and if there are errors performs an async cpu blit
instead. If there is a failure to allocate the async dma_fence_work,
allocate it on the stack and sync wait for the blit to complete.
Add selftests that simulate gpu blit failures and failure to allocate
the async dma_fence_work.
A previous version of this pach used dma_fence_work, now that's
opencoded which adds more code but might lower the latency
somewhat in the common non-error case.
v3:
- Style fixes (Matthew Auld)
v4:
- Use "#if IS_ENABLED()" instead of #ifdef (Matthew Auld)
v5:
- Fix an issue where we, if the dependency was already signaled, might
end up waiting for a memcpy fence that would never signal.
v6:
- Add a missing i915_ttm_memcpy_release() (Matthew Auld)
Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211104110718.688420-3-thomas.hellstrom@linux.intel.com
We are about to introduce failsafe- and asynchronous migration and
ttm moves.
This will add complexity and code to the TTM move code so it makes sense
to split it out to a separate file to make the i915 TTM code easer to
digest.
Split the i915 TTM move code out and since we will have to change the name
of the gpu_binds_iomem() and cpu_maps_iomem() functions anyway,
we alter the name of gpu_binds_iomem() to i915_ttm_gtt_binds_lmem() which
is more reflecting what it is used for.
With this we also add some more documentation. Otherwise there should be
no functional change.
Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211104110718.688420-2-thomas.hellstrom@linux.intel.com
