A weak implementation of parallel submission (multi-bb execbuf IOCTL) for
execlists. Doing as little as possible to support this interface for
execlists - basically just passing submit fences between each request
generated and virtual engines are not allowed. This is on par with what
is there for the existing (hopefully soon deprecated) bonding interface.
We perma-pin these execlists contexts to align with GuC implementation.
v2:
(John Harrison)
- Drop siblings array as num_siblings must be 1
v3:
(John Harrison)
- Drop single submission
v4:
(John Harrison)
- Actually drop single submission
- Use IS_ERR check on return value from intel_context_create
- Set last request to NULL on unpin
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: John Harrison <John.C.Harrison@Intel.com>
Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211222223532.28698-1-matthew.brost@intel.com
Convert free_work into delayed_work, similar to ttm to allow converting the
blocking lock in __i915_gem_free_objects to a trylock.
Unlike ttm, the object should already be idle, as it's kept alive
by a reference through struct i915_vma->active, which is dropped
after all vma's are idle.
Because of this, we can use a no wait by default, or when the lock
is contested, we use ttm's 10 ms.
The trylock should only fail when the object is sharing it's resv with
other objects, and typically objects are not kept locked for a long
time, so we can safely retry on failure.
Fixes: be7612fd66 ("drm/i915: Require object lock when freeing pages during destruction")
Testcase: igt/gem_exec_alignment/pi*
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Reviewed-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211222155622.2960379-1-maarten.lankhorst@linux.intel.com
Protect updates of struct i915_vma flags and async binding / unbinding
with the vm::mutex. This means that i915_vma_bind() needs to assert
vm::mutex held. In order to make that possible drop the caching of
kmap_atomic() maps around i915_vma_bind().
An alternative would be to use kmap_local() but since we block cpu
unplugging during sleeps inside kmap_local() sections this may have
unwanted side-effects. Particularly since we might wait for gpu while
holding the vm mutex.
This change may theoretically increase execbuf cpu-usage on snb, but
at least on non-highmem systems that increase should be very small.
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/20211221200050.436316-5-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
By default, GT (and GuC) run at RPn. Requesting for RP0
before firmware load can speed up DMA and HuC auth as well.
In addition to writing to 0xA008, we also need to enable
swreq in 0xA024 so that Punit will pay heed to our request.
SLPC will restore the frequency back to RPn after initialization,
but we need to manually do that for the non-SLPC path.
We don't need a manual override in the SLPC disabled case, just
use the intel_rps_set function to ensure consistent RPS state.
Signed-off-by: Vinay Belgaumkar <vinay.belgaumkar@intel.com>
Reviewed-by: Sujaritha Sundaresan <sujaritha.sundaresan@intel.com>
Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211216233022.21351-1-vinay.belgaumkar@intel.com
If GuC encounters an error during engine reset, the i915 driver
promotes to full GT reset. This includes an info message about why the
reset is happening. However, that is not treated as a failure by any
of the CI systems because resets are an expected occurrance during
testing. This kind of failure is a major problem and should never
happen. So, complain more loudly and make sure CI notices.
Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
Reviewed-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211211065859.2248188-4-John.C.Harrison@Intel.com
Lots of testing is done with the DEBUG_GEM config option enabled but
not the DEBUG_GUC option. That means we only get teeny-tiny GuC logs
which are not hugely useful. Enabling full DEBUG_GUC also spews lots
of other detailed output that is not generally desired. However,
bigger GuC logs are extremely useful for almost any regression debug.
So enable bigger logs for DEBUG_GEM builds as well.
Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211211065859.2248188-3-John.C.Harrison@Intel.com
Add support for telling the debugfs interface the size of the GuC log
dump in advance. Without that, the underlying framework keeps calling
the 'show' function with larger and larger buffer allocations until it
fits. That means reading the log from graphics memory many times - 16
times with the full 18MB log size.
v2: Don't return error codes from size query. Report overflow in the
error dump as well (review feedback from Daniele).
Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
Reviewed-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211211065859.2248188-2-John.C.Harrison@Intel.com
Big delta, but boils down to moving set_pages to i915_vma.c, and removing
the special handling, all callers use the defaults anyway. We only remap
in ggtt, so default case will fall through.
Because we still don't require locking in i915_vma_unpin(), handle this by
using xchg in get_pages(), as it's locked with obj->mutex, and cmpxchg in
unpin, which only fails if we race a against a new pin.
Changes since v1:
- aliasing gtt sets ZERO_SIZE_PTR, not -ENODEV, remove special case
from __i915_vma_get_pages(). (Matt)
Changes since v2:
- Free correct old pages in __i915_vma_get_pages(). (Matt)
Remove race of clearing vma->pages accidentally from put,
free it but leave it set, as only get has the lock.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211216142749.1966107-4-maarten.lankhorst@linux.intel.com
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
We now support a per-gt uncore, yet we're not able to infer which GT
we're operating upon. Let's store a backpointer for now.
At this point the early initialization of the gt needs to be
broken in two parts where the first is needed to assign to the gt
the i915 private data pointer and the uncore. A temporary
function has been made and the two parts are
__intel_gt_init_early() and intel_gt_init_early(). This split
will be fixed in the future with the multitile patch.
Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Reviewed-by: Andi Shyti <andi.shyti@linux.intel.com>
Signed-off-by: Andi Shyti <andi.shyti@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211214193346.21231-2-andi.shyti@linux.intel.com
Testing the stealing of guc ids is hard from user space as we have 64k
guc_ids. Add a selftest, which artificially reduces the number of guc
ids, and forces a steal.
The test creates a spinner which is used to block all subsequent
submissions until it completes. Next, a loop creates a context and a NOP
request each iteration until the guc_ids are exhausted (request creation
returns -EAGAIN). The spinner is ended, unblocking all requests created
in the loop. At this point all guc_ids are exhausted but are available
to steal. Try to create another request which should successfully steal
a guc_id. Wait on last request to complete, idle GPU, verify a guc_id
was stolen via a counter, and exit the test. Test also artificially
reduces the number of guc_ids so the test runs in a timely manner.
v2:
(John Harrison)
- s/stole/stolen
- Fix some wording in test description
- Rework indexing into context array
- Add test description to commit message
- Fix typo in commit message
(Checkpatch)
- s/guc/(guc) in NUMBER_MULTI_LRC_GUC_ID
v3:
(John Harrison)
- Set array value to NULL after extracting error
- Fix a few typos in comments / error messages
- Delete redundant comment in commit message
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: John Harrison <John.C.Harrison@Intel.com>
Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211214170500.28569-8-matthew.brost@intel.com
While attempting to debug a CT deadlock issue in various CI failures
(most easily reproduced with gem_ctx_create/basic-files), I was seeing
CPU deadlock errors being reported. This were because the context
destroy loop was blocking waiting on H2G space from inside an IRQ
spinlock. There no was deadlock as such, it's just that the H2G queue
was full of context destroy commands and GuC was taking a long time to
process them. However, the kernel was seeing the large amount of time
spent inside the IRQ lock as a dead CPU. Various Bad Things(tm) would
then happen (heartbeat failures, CT deadlock errors, outstanding H2G
WARNs, etc.).
Re-working the loop to only acquire the spinlock around the list
management (which is all it is meant to protect) rather than the
entire destroy operation seems to fix all the above issues.
v2:
(John Harrison)
- Fix typo in comment message
Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211214170500.28569-5-matthew.brost@intel.com
