| Age | Commit message (Collapse) | Author |
|---|
|
Currently when children processes try to communicate to not existing
allocator thread they got crash with vague SIGSEGV.
Adding readyness flag and detailed explanation in assert should hint
the developer to add missing intel_allocator_multiprocess_start|stop)
functions.
Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski@intel.com>
Cc: Petri Latvala <petri.latvala@intel.com>
Reviewed-by: Petri Latvala <petri.latvala@intel.com>
|
|
On some platforms (like on ATS) 0x0 may not be available so allocator
should be instantiated with safe start offset to avoid getting -ENOSPC.
Change require also relaxation of automatic range selection in the
allocator.
Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski@intel.com>
Cc: Ashutosh Dixit <ashutosh.dixit@intel.com>
Cc: Dominik Grzegorzek <dominik.grzegorzek@intel.com>
Reviewed-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
|
|
For DG2 and beyond regions alignment may vary so many tests would need
to be rewritten to handle this constraint. As Ashutosh noticed most of
tests can use safe alignment as a default.
Adopt intel-allocator to use safe or user defined power-of-two alignment.
Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski@intel.com>
Suggested-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
Cc: Ashutosh Dixit <ashutosh.dixit@intel.com>
Reviewed-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
|
|
Add typechecking to the min/max macros and make their locals truly
unique-ish to reduce the risk of shadowing.
v2: small bug fix, write also height coordinate on rotation
test. (jheikkil)
v3: Fix up a couple of other max/max_t instances (Ashutosh)
Signed-off-by: Juha-Pekka Heikkilä <juha-pekka.heikkila@intel.com>
Signed-off-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
Signed-off-by: Chris Wilson <chris.p.wilson@intel.com>
Reviewed-by: Zbigniew Kempczyński <zbigniew.kempczynski@intel.com>
|
|
When allocator is running in multiprocess mode, all queries
are processed in a designated thread in the parent process.
However, a child process may request opening the allocator
for a gpu using a file descriptor absent in the parent process.
Hence, querying available gtt size must be done in the child
instead of the parent process.
As a consequence of this change creating allocators managing
only a given interval of available gtt is now enabled for all
allocator implementations. Additionally, this commit sets a
universal lower bound on alignment to 4096.
Signed-off-by: Andrzej Turko <andrzej.turko@linux.intel.com>
Cc: Zbigniew Kempczyński <zbigniew.kempczynski@intel.com>
Reviewed-by: Zbigniew Kempczyński <zbigniew.kempczynski@intel.com>
|
|
All procedures for opening particular types of allocators
query the gtt size. Even if the user has provided the
boundaries of the address space managed by the allocator,
the query is performed for the sake of validation.
gem_aperture_size() called with an invalid id returns a
default value instead of failing. Thus, we need to
additionally check whether the file descriptor is valid.
Signed-off-by: Andrzej Turko <andrzej.turko@linux.intel.com>
Cc: Zbigniew Kempczyński <zbigniew.kempczynski@intel.com>
Reviewed-by: Zbigniew Kempczyński <zbigniew.kempczynski@intel.com>
|
|
Before allowing allocator working in multiprocess mode we have to
ensure it is ready and able to process allocator messages passed
from children.
Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski@intel.com>
Cc: Jason Ekstrand <jason@jlekstrand.net>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Acked-by: Petri Latvala <petri.latvala@intel.com>
|
|
To use spinners with no-reloc we need to alloc offsets for them
from already opened allocator. As we don't know what strategy
is chosen for open (likely HIGH_TO_LOW for SIMPLE allocator) we
want to overwrite it for spinners (there's expectation they
will reside on low addresses).
Extend allocator API adding intel_allocator_alloc_with_strategy()
to support spinners rewriting.
v2: add change in api_intel_allocator test to compile properly
whole series.
Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski@intel.com>
Cc: Jason Ekstrand <jason@jlekstrand.net>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Acked-by: Petri Latvala <petri.latvala@intel.com>
|
|
Validating allocator code (leaks and memory overwriting) can be done
with address sanitizer. When allocator is not working in multiprocess
mode it is easy, problems start when fork is in the game. In this
situation we need to separate preparation and starting allocator thread.
Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski@intel.com>
Reported-by: Andrzej Turko <andrzej.turko@linux.intel.com>
Cc: Andrzej Turko <andrzej.turko@linux.intel.com>
Cc: Dominik Grzegorzek <dominik.grzegorzek@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Acked-by: Petri Latvala <petri.latvala@intel.com>
|
|
Avoid race when stop was send to allocator thread. We wait around
100 ms to give thread chance to stop smoothly instead of removing
queue and enforcing exiting all blocked message syscalls.
Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski@intel.com>
Cc: Dominik Grzegorzek <dominik.grzegorzek@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Acked-by: Petri Latvala <petri.latvala@intel.com>
|
|
For discrete gens we have to cease of using relocations when batch
buffers are submitted to GPU. On cards which have ppgtt we can use
softpin establishing addresses on our own.
We added simple allocator (taken from Mesa; works on lists) and
random allocator to exercise batches with different addresses. All
of that works for single VM (context) so we have to add additional
layer (intel_allocator) to support multiprocessing / multithreading.
For main IGT process (also for threads created in it) intel_allocator
resolves addresses "locally", just by mutexing access to global
allocator data (ctx/vm map). When fork() is in use children cannot
establish addresses on they own and have to contact to the thread
spawned within main IGT process. Currently SysV IPC message queue was
chosen as a communication channel between children and allocator thread.
Child calls same functions as main IGT process, only communication path
will be chosen instead of acquiring addresses locally.
v2:
Add intel_allocator_open_full() to allow user pass vm range.
Add strategy: NONE, LOW_TO_HIGH, HIGH_TO_LOW passed to allocator backend.
v3:
Child is now able to use allocator directly as standalone. It only need
to call intel_allocator_init() to reinitialize appropriate structures.
v4:
Add pseudo allocator - INTEL_ALLOCATOR_RELOC which just increments
offsets to avoid unnecessary conditional code.
v5:
Alter allocator core according to igt_map changes.
v6:
Add internal version __intel_allocator_alloc() to return
ALLOC_INVALID_ADDRESS without assertion.
v7:
Add libatomic for linking libigt library. It is required on some
archs, like mips.
Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski@intel.com>
Signed-off-by: Dominik Grzegorzek <dominik.grzegorzek@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Petri Latvala <petri.latvala@intel.com>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Acked-by: Petri Latvala <petri.latvala@intel.com>
|
