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
Define helpers and struct members required to record boost info.
Boost frequency is initialized to RP0 at SLPC init. Also define num_waiters
which can track the pending boost requests.
Boost will be done by scheduling a worker thread. This will avoid
the need to make H2G calls inside an interrupt context. Initialize the
worker function during SLPC init as well. Had to move intel_guc_slpc_init
a few lines below to accommodate this.
v2: Add a workqueue to handle waitboost
v3: Code review comments (Ashutosh)
Cc: Ashutosh Dixit <ashutosh.dixit@intel.com>
Signed-off-by: Vinay Belgaumkar <vinay.belgaumkar@intel.com>
Reviewed-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211102012608.8609-2-vinay.belgaumkar@intel.com
We seem to have an unfortunate issue where we arrive from:
i915_gem_object_flush_if_display+0x86/0xd0 [i915]
intel_user_framebuffer_dirty+0x1a/0x50 [i915]
drm_mode_dirtyfb_ioctl+0xfb/0x1b0
which can be before the pages are populated(and pinned for display), and
so i915_gem_object_has_struct_page() might still return true, as per the
ttm backend. We could re-order the later get_pages() call here, but
since on discrete everything should already be coherent, with the
exception of the display engine, and even there display surfaces must be
allocated in device local-memory anyway, so there should in theory be no
conceivable reason to ever call i915_gem_clflush_object() on discrete.
References: https://gitlab.freedesktop.org/drm/intel/-/issues/4320
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211027161813.3094681-2-matthew.auld@intel.com
In theory if clflush_work_create() somehow fails here, and we don't yet
have mm.pages populated then we end up resetting cache_dirty, which is
likely wrong, since that will potentially skip the flush-on-acquire, if
it was needed.
It looks like intel_user_framebuffer_dirty() can arrive here before the
pages are populated.
v2(Thomas):
- Move setting cache_dirty out of the async portion, also add a
comment for why that should still be safe.
v3:
- Add Thomas' irc r-b
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211027161813.3094681-1-matthew.auld@intel.com
As we start to introduce asynchronous failsafe object migration,
where we update the object state and then submit asynchronous
commands we need to record what memory resources are actually used
by various part of the command stream. Initially for three purposes:
1) Error capture.
2) Asynchronous migration error recovery.
3) Asynchronous vma bind.
At the time where these happens, the object state may have been updated
to be several migrations ahead and object sg-tables discarded.
In order to make it possible to keep sg-tables with memory resource
information for these operations, introduce refcounted sg-tables that
aren't freed until the last user is done with them.
The alternative would be to reference information sitting on the
corresponding ttm_resources which typically have the same lifetime as
these refcountes sg_tables, but that leads to other awkward constructs:
Due to the design direction chosen for ttm resource managers that would
lead to diamond-style inheritance, the LMEM resources may sometimes be
prematurely freed, and finally the subclassed struct ttm_resource would
have to bleed into the asynchronous vma bind code.
v3:
- Address a number of style issues (Matthew Auld)
v4:
- Dont check for st->sgl being NULL in i915_ttm_tt__shmem_unpopulate(),
that should never happen. (Matthew Auld)
v5:
- Fix a Potential double-free (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/20211101122444.114607-1-thomas.hellstrom@linux.intel.com
With GuC handling scheduling, i915 is not aware of the time that a
context is scheduled in and out of the engine. Since i915 pmu relies on
this info to provide engine busyness to the user, GuC shares this info
with i915 for all engines using shared memory. For each engine, this
info contains:
- total busyness: total time that the context was running (total)
- id: id of the running context (id)
- start timestamp: timestamp when the context started running (start)
At the time (now) of sampling the engine busyness, if the id is valid
(!= ~0), and start is non-zero, then the context is considered to be
active and the engine busyness is calculated using the below equation
engine busyness = total + (now - start)
All times are obtained from the gt clock base. For inactive contexts,
engine busyness is just equal to the total.
The start and total values provided by GuC are 32 bits and wrap around
in a few minutes. Since perf pmu provides busyness as 64 bit
monotonically increasing values, there is a need for this implementation
to account for overflows and extend the time to 64 bits before returning
busyness to the user. In order to do that, a worker runs periodically at
frequency = 1/8th the time it takes for the timestamp to wrap. As an
example, that would be once in 27 seconds for a gt clock frequency of
19.2 MHz.
Note:
There might be an over-accounting of busyness due to the fact that GuC
may be updating the total and start values while kmd is reading them.
(i.e kmd may read the updated total and the stale start). In such a
case, user may see higher busyness value followed by smaller ones which
would eventually catch up to the higher value.
v2: (Tvrtko)
- Include details in commit message
- Move intel engine busyness function into execlist code
- Use union inside engine->stats
- Use natural type for ping delay jiffies
- Drop active_work condition checks
- Use for_each_engine if iterating all engines
- Drop seq locking, use spinlock at GuC level to update engine stats
- Document worker specific details
v3: (Tvrtko/Umesh)
- Demarcate GuC and execlist stat objects with comments
- Document known over-accounting issue in commit
- Provide a consistent view of GuC state
- Add hooks to gt park/unpark for GuC busyness
- Stop/start worker in gt park/unpark path
- Drop inline
- Move spinlock and worker inits to GuC initialization
- Drop helpers that are called only once
v4: (Tvrtko/Matt/Umesh)
- Drop addressed opens from commit message
- Get runtime pm in ping, remove from the park path
- Use cancel_delayed_work_sync in disable_submission path
- Update stats during reset prepare
- Skip ping if reset in progress
- Explicitly name execlists and GuC stats objects
- Since disable_submission is called from many places, move resetting
stats to intel_guc_submission_reset_prepare
v5: (Tvrtko)
- Add a trylock helper that does not sleep and synchronize PMU event
callbacks and worker with gt reset
v6: (CI BAT failures)
- DUTs using execlist submission failed to boot since __gt_unpark is
called during i915 load. This ends up calling the GuC busyness unpark
hook and results in kick-starting an uninitialized worker. Let
park/unpark hooks check if GuC submission has been initialized.
- drop cant_sleep() from trylock helper since rcu_read_lock takes care
of that.
v7: (CI) Fix igt@i915_selftest@live@gt_engines
- For GuC mode of submission the engine busyness is derived from gt time
domain. Use gt time elapsed as reference in the selftest.
- Increase busyness calculation to 10ms duration to ensure batch runs
longer and falls within the busyness tolerances in selftest.
v8:
- Use ktime_get in selftest as before
- intel_reset_trylock_no_wait results in a lockdep splat that is not
trivial to fix since the PMU callback runs in irq context and the
reset paths are tightly knit into the driver. The test that uncovers
this is igt@perf_pmu@faulting-read. Drop intel_reset_trylock_no_wait,
instead use the reset_count to synchronize with gt reset during pmu
callback. For the ping, continue to use intel_reset_trylock since ping
is not run in irq context.
- GuC PM timestamp does not tick when GuC is idle. This can potentially
result in wrong busyness values when a context is active on the
engine, but GuC is idle. Use the RING TIMESTAMP as GPU timestamp to
process the GuC busyness stats. This works since both GuC timestamp and
RING timestamp are synced with the same clock.
- The busyness stats may get updated after the batch starts running.
This delay causes the busyness reported for 100us duration to fall
below 95% in the selftest. The only option at this time is to wait for
GuC busyness to change from idle to active before we sample busyness
over a 100us period.
Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com>
Acked-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211027004821.66097-2-umesh.nerlige.ramappa@intel.com
perf_parallel_engines is micro benchmark to test i915 request
scheduling. The test creates a thread per physical engine and submits
NOP requests and waits the requests to complete in a loop. In execlists
mode this works perfectly fine as powerful CPU has enough cores to feed
each engine and process the CSBs. With GuC submission the uC gets
overwhelmed as all threads feed into a single CTB channel and the GuC
gets bombarded with CSBs as contexts are immediately switched in and out
on the engines due to the zero runtime of the requests. When the GuC is
overwhelmed scheduling of contexts is unfair due to the nature of the
GuC scheduling algorithm. This behavior is understood and deemed
acceptable as this micro benchmark isn't close to real world use case.
Increasing the timeout of wait period for requests to complete. This
makes the test understand that is ok for contexts to get starved in this
scenario.
A future patch / cleanup may just delete these micro benchmark tests as
they basically mean nothing. We care about real workloads not made up
ones.
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: John Harrison <John.C.Harrison@Intel.com>
Reviewed-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211011175704.28509-1-matthew.brost@intel.com
We currently just evict lmem objects to system memory when under memory
pressure. For this case we might lack the usual object mm.pages, which
effectively hides the pages from the i915-gem shrinker, until we
actually "attach" the TT to the object, or in the case of lmem-only
objects it just gets migrated back to lmem when touched again.
For all cases we can just adjust the i915 shrinker LRU each time we also
adjust the TTM LRU. The two cases we care about are:
1) When something is moved by TTM, including when initially populating
an object. Importantly this covers the case where TTM moves something from
lmem <-> smem, outside of the normal get_pages() interface, which
should still ensure the shmem pages underneath are reclaimable.
2) When calling into i915_gem_object_unlock(). The unlock should
ensure the object is removed from the shinker LRU, if it was indeed
swapped out, or just purged, when the shrinker drops the object lock.
v2(Thomas):
- Handle managing the shrinker LRU in adjust_lru, where it is always
safe to touch the object.
v3(Thomas):
- Pretty much a re-write. This time piggy back off the shrink_pin
stuff, which actually seems to fit quite well for what we want here.
v4(Thomas):
- Just use a simple boolean for tracking ttm_shrinkable.
v5:
- Ensure we call adjust_lru when faulting the object, to ensure the
pages are visible to the shrinker, if needed.
- Add back the adjust_lru when in i915_ttm_move (Thomas)
v6(Reported-by: kernel test robot <lkp@intel.com>):
- Remove unused i915_tt
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Thomas Hellström <thomas.hellstrom@linux.intel.com> #v4
Link: https://patchwork.freedesktop.org/patch/msgid/20211018091055.1998191-6-matthew.auld@intel.com
For cached objects we can allocate our pages directly in shmem. This
should make it possible(in a later patch) to utilise the existing
i915-gem shrinker code for such objects. For now this is still disabled.
v2(Thomas):
- Add optional try_to_writeback hook for objects. Importantly we need
to check if the object is even still shrinkable; in between us
dropping the shrinker LRU lock and acquiring the object lock it could for
example have been moved. Also we need to differentiate between
"lazy" shrinking and the immediate writeback mode. Also later we need to
handle objects which don't even have mm.pages, so bundling this into
put_pages() would require somehow handling that edge case, hence
just letting the ttm backend handle everything in try_to_writeback
doesn't seem too bad.
v3(Thomas):
- Likely a bad idea to touch the object from the unpopulate hook,
since it's not possible to hold a reference, without also creating
circular dependency, so likely this is too fragile. For now just
ensure we at least mark the pages as dirty/accessed when called from the
shrinker on WILLNEED objects.
- s/try_to_writeback/shrinker_release_pages, since this can do more
than just writeback.
- Get rid of do_backup boolean and just set the SWAPPED flag prior to
calling unpopulate.
- Keep shmem_tt as lowest priority for the TTM LRU bo_swapout walk, since
these just get skipped anyway. We can try to come up with something
better later.
v4(Thomas):
- s/PCI_DMA/DMA/. Also drop NO_KERNEL_MAPPING and NO_WARN, which
apparently doesn't do anything with streaming mappings.
- Just pass along the error for ->truncate, and assume nothing.
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Cc: Christian König <christian.koenig@amd.com>
Cc: Oak Zeng <oak.zeng@intel.com>
Reviewed-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Acked-by: Oak Zeng <oak.zeng@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211018091055.1998191-2-matthew.auld@intel.com
Break out some shmem backend utils for future reuse by the TTM backend:
shmem_alloc_st(), shmem_free_st() and __shmem_writeback() which we can
use to provide a shmem-backed TTM page pool for cached-only TTM
buffer objects.
Main functional change here is that we now compute the page sizes using
the dma segments rather than using the physical page address segments.
v2(Reported-by: kernel test robot <lkp@intel.com>)
- Make sure we initialise the mapping on the error path in
shmem_get_pages()
Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211018091055.1998191-1-matthew.auld@intel.com
UAPI Changes:
- Expose multi-LRC submission interface
Similar to the bonded submission interface but simplified.
Comes with GuC only implementation for now. See kerneldoc
for more details.
Userspace changes: https://github.com/intel/media-driver/pull/1252
- Expose logical engine instance to user
Needed by the multi-LRC submission interface for GuC
Userspace changes: https://github.com/intel/media-driver/pull/1252
Driver Changes:
- Fix blank screen booting crashes when CONFIG_CC_OPTIMIZE_FOR_SIZE=y (Hugh)
- Add support for multi-LRC submission in the GuC backend (Matt B)
- Add extra cache flushing before making pages userspace visible (Matt A, Thomas)
- Mark internal GPU object pages dirty so they will be flushed properly (Matt A)
- Move remaining debugfs interfaces i915_wedged/i915_forcewake_user into gt (Andi)
- Replace the unconditional clflushes with drm_clflush_virt_range() (Ville)
- Remove IS_ACTIVE macro completely (Lucas)
- Improve kerneldocs for cache_dirty (Matt A)
- Add missing includes (Lucas)
- Selftest improvements (Matt R, Ran, Matt A)
Signed-off-by: Dave Airlie <airlied@redhat.com>
From: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/YXFmLKoq8Fg9JxSd@jlahtine-mobl.ger.corp.intel.com
While the pages can't be swapped out, they can be discarded by the shrinker.
Normally such objects are marked with __I915_MADV_PURGED, which can't be
unset, and therefore requires a new object. For kernel internal objects
this is not true, since the madv hint is reset for our special volatile
objects, such that we can re-acquire new pages, if so desired, without
needing a new object. As a result we should probably be paranoid here
and put the object back into the CPU domain when discarding the pages,
and also correctly set cache_dirty, if required.
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211018174508.2137279-8-matthew.auld@intel.com
