The crucial trick to reproduce the bug is that we need to have
a decent pile of active bos to retire. Because we unref the bo
after having it moved off the active list, our recursion depth
in fdo bug #42180 is limited by the number of active objects that
can retire at the same time.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=42180
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Provokes the forcewake warning when the hangcheck runs and no
one waits for the gpu (and hence holds the dev->struct_mutex).
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Hangs gen3 and simply writes garbage into the unmappable part of
gtt on gen4+, which might cause issues later on.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Using a dummy reloc that doesn't matter to trick the kernel into
synchroizing the rings.
v2: properly apply MI_NOOP workaround to MI_FLUSH_DW and
switch to MI_COND_BATCH_BUFFER_END as a dummy command on the
render ring to avoid PIPE_CONTROL errata.
v3: somebody clever decided that in C, you cound from 1,
i.e. I915_EXEC_RENDER == 1. It works now ...
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Also start to shortly explain testcases with an easily-greppable
header like this:
/*
* Testcase:
*
* [Possible further explanation.]
*
*/
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Add a couple of simple store dword tests to test memory coherence.
gem_storedw_loop simply executes a batch that continually stores an
incremented value to a target buffer object, checking the results after
each batch completes.
gem_storedw_batches_loop does the same thing, but creates a new command
batch buffer for each iteration, which can exercise the buffer creation
code. This test is based on one from Andrzej Kacprowski from Intel.
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
... and we have a winner: gen3_mixed_blits reproduces the issue Daniel
Vetter originally found. It seems clear that we have some incoherence
between the RENDER and BLT units on gen3 that no amount of MI_FLUSH can
hide. Hmmm....
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
gem_stress is unhappy with tiled render copies on gen3. This is a simple
little test to ensure that a set of pure copies with a working set
larger than the aperture are handled correctly.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
After full-gtt, gem_tiled_blits doesn't allocate enough to force
eviction. So query the total aperture and accommodate.
Also introduce a similar test that utilizes fences rather than
use the BLT to perform the tiling and detiling.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Using for diagnosing some mysterious slowdowns. Should include a variant
for basic benchmarking...
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
This is a test case that overcommits fence registers between threads, which
are copying from one fenced bo to another. In earlier versions of the driver
this would cause excessive spinning as the first inactive (i.e. not in use
by the GPU) would be used to service the next page. After all the fence
registers had been allocated, in effect only the very first fence would then
be used for all subsequent faults.
Add four new tests for error the error handling cases:
- gem_bad_address - store to a bad address, should generate a protection or
page table error
- gem_bad_batch - try to execute a bad batch, should generate a protection,
invalid instruction or page table error
- gem_bad_blit - blit to an invalid location, should generated a protection
or page table error
- gem_hang - hang the GPU on an event that will never happen, test hang
detection & recovery code
The large object test simply tries to allocate a 128M object, pin it, then
pwrite the whole thing. This should make obvious any leaks on close or
page pointer allocation failures.
