On 7/13/2023 05:11, Tvrtko Ursulin wrote:
On 13/07/2023 12:09, Andrzej Hajda wrote:
Hi,
On 13.07.2023 09:39, Tvrtko Ursulin wrote:
On 12/07/2023 19:54, John Harrison wrote:
On 7/12/2023 09:27, Andrzej Hajda wrote:
On 12.07.2023 14:35, Tvrtko Ursulin wrote:
On 12/07/2023 13:18, Andrzej Hajda wrote:
On 11.07.2023 17:27, Tvrtko Ursulin wrote:
On 11/07/2023 14:58, Andrzej Hajda wrote:
On 11.07.2023 13:34, Andi Shyti wrote:
Hi Andrzej,
drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c | 11
+++++++++++
1 file changed, 11 insertions(+)
diff --git
a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
index a0e3ef1c65d246..2c877ea5eda6f0 100644
---
a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
+++
b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
@@ -3461,6 +3461,8 @@ static void
guc_prio_fini(struct i915_request *rq, struct intel_context
*ce)
static void remove_from_context(struct
i915_request *rq)
{
struct intel_context *ce =
request_to_scheduling_context(rq);
+ struct intel_engine_cs *engine;
+ intel_engine_mask_t tmp;
GEM_BUG_ON(intel_context_is_child(ce));
@@ -3478,6 +3480,15 @@ static void
remove_from_context(struct i915_request *rq)
atomic_dec(&ce->guc_id.ref);
i915_request_notify_execute_cb_imm(rq);
+
+ /*
+ * GuC virtual engine can disappear after
this call, so let's assign
+ * something valid, as driver expects this
to be always valid pointer.
+ */
+ for_each_engine_masked(engine, rq->engine->gt,
rq->execution_mask, tmp) {
+ rq->engine = engine;
yes... here the context might lose the virtual
engine... I wonder
whether this is the rigth solution, though. Maybe we
should set
rq->engine = NULL; and check for NULL? Don't know.
Setting NULL causes occasional null page de-reference in
i915_request_wait_timeout:
mutex_release(&rq->engine->gt->reset.mutex.dep_map, _THIS_IP_)
rq->engine after removing rq from context is (IMHO) used as
a set of aliases
for gt and i915 (despite rq itself contains the alias to i915).
without investigating further, but maybe that code is not even
supposed to be executed, at this point, if the request's
assigned
virtual engine is removed.
Real tests show it is executed and the function
i915_request_wait_timeout is quite generic
I guess it is quite typical use-case, the only question is
about timings - what happens earlier -
finalization of i915_request_wait_timeout or context removal.
The other point rq->engine is accessed after context removal
is i915_fence_release -
there is long comment there regarding virtual context and
reuse retired rq.
Anyway calling there "intel_engine_is_virtual(rq->engine)" is
risky without this patch and KASAN complains clearly about it:
http://gfx-ci.igk.intel.com/tree/drm-tip/kasan.html?testfilter=gem_exec_balancer
Looks like a bug introduced in bcb9aa45d5a0 ("Revert "drm/i915:
Hold reference to intel_context over life of i915_request""),
which was a partial revert of 1e98d8c52ed5 ("drm/i915: Hold
reference to intel_context over life of i915_request").
Ie. if 1e98d8c52ed5 recognised the problem with disappearing
rq->engine, then I am confused how bcb9aa45d5a0 left the
rq->engine dereference in there after removing the extra
reference.
Could it be that the intel_engine_is_virtual check simply needs
to be removed from i915_fence_release, restoring things to how
they were before 1e98d8c52ed5? Could you try it out?
I have already tried something similar [1] and KASAN bugs
disappeared, or more precisely gem_exec_balance tests passed.
But I have been warned by Nirmoy guc virtual engines can be
created for only one real engine (ie.
is_power_of_2(rq->execution_mask) is true but rq->engine points
to virtual engine).
[1]: https://patchwork.freedesktop.org/series/118879/
Ugh.. Try involving media umd folks to see if they need a single
engine virtual engine? Or we could always just not create it in
the driver, I mean just use the physical one.
In case there is single physical engine
intel_engine_create_virtual falls back to intel_context_create (no
virtual engine), but in case of parallel contexts there is special
KMD flag FORCE_VIRTUAL which enforces virtual engine even for
single physical engine. So it seems to be KMD concept.
Anyway is it worth investigating how to make
"is_power_of_2(rq->execution_mask)" indication of dangling engine
pointer? It will not help in 1st case:
mutex_release(&rq->engine->gt->reset.mutex.dep_map, _THIS_IP_)
There seems to be a fundamental problem here. Object 1 (rq) is
holding a pointer to a reference counted and transient object 2
(engine) but without taking a reference count for itself. That is a
Bad Thing(tm).
Engine is not ref counted (at least directly), hardware engines seems
to be even persistent across whole life of i915.
I guess before introduction of virtual engines the assumption about
validity of rq->engine was correct and developers used it to access
rq->engine->gt, rq->engine->i915, etc
So the problems described here are probably leftovers of this change.
After virtual engines were introduced
"is_power_of_2(rq->execution_mask)" was used to detect rq->engine is
virtual.
And after adding parallel engines it does not seem to be valid check
anymore.
I'm not following the description in the revert patch as to why rq
can't ref count the context/engine. Is there actually a recursive
counting problem? Or is it just a lifetime issue caused by requests
hanging around indefinitely because they are locked by a user process?
Either way, jumping through convoluted hoops to ensure the code
does not attempt to dereference a dangling pointer seems like the
wrong fix. Removing the engine pointer when the request is
completed and no longer dependent upon an engine (but before the
engine can possibly be destroyed) seems like a much better
solution. And then making the request handling code check for and
cope with a null engine pointer. It sounds like the only problem
there is the above mutex, but there is an alternate route to that?
Although why a completed request would need access to a GT reset
mutex seems confusing. If the request is done, then what connection
does it still have to the GT?
Agreed in principle but the question is how invasive would it be to
change the rules.
With the latest info that the issue is really just the GuC
_parallel_ engine setup, and looking at the code, I wonder if we
couldn't just flag the rq->flags with "kernel context request". The
code in i915_fence_release claims the rq pool is only relevant for
those so it sounds it would be safe to skip everything else based on
that new flag.
For the mutex_release path, presumable the bad deref is only _after_
the wait, right? (Only once the request has been retired.)
In which case caching the gt pointer at the start of
i915_request_wait_timeout would be sufficient.
That should be a few lines fixup overall and then the idea of
allowing rq->engine to be reset to NULL can be explored more leisurely.
I guess:
- setting engine to NULL in remove_from_context,
- caching gt pointer,
- checking for null pointer in i915_fence_release
should be enough to solve current issue. However I am not sure if
there are no more dragons hidden in other execution paths. I will try
inspect the code.
What you have in this patch, cheat by replacing the engine, *might*
work for the short term *if* you can make sure all parallel readers
will see the updated rq->engine pointer in time, before the old one
gets freed.
For the longer term solution - maybe making the engine reference counted?
That was the point of the original solution - having the request
reference count the context. The context is what owns the virtual
engine. So ensuring that the context can't be destroyed while there are
requests outstanding on it ensured that rq->engine would always be
valid. Nice simple and clean solution.So why did that get reverted? What
is the problem with VM_BIND and having requests ensure that their
resources stay alive for the lifetime of the request?
John.
Or if setting rq->engine to NULL then evaluating the paths which
derefence it. AFAIR these request retirement races have been generally
tricky/annoying.
For instance under the i915_request_wait_timeout chain.
1)
__i915_spin_request - could be racy if request gets retired between
i915_request_wait_timeout/dma_fence_is_signaled check and
__i915_spin_request dereferencing rq->engine.props?
2)
intel_rps_boost - claims to be safe by serialising via
i915_request_retire/I915_FENCE_FLAG_BOOST but is it? What prevents
retire tearing down the engine between:
if (!test_and_set_bit(I915_FENCE_FLAG_BOOST, &rq->fence.flags)) {
---> HERE - if whole retirement happens here <----
struct intel_rps *rps = &READ_ONCE(rq->engine)->gt->rps;
3)
__intel_engine_flush_submission - could be racy to? What if the whole
process of consuming the request by the backend and retirement happens
between these two lines:
if (i915_request_is_ready(rq))
---> HERE <---
__intel_engine_flush_submission(rq->engine, false);
Then "is ready" can be true, but by the 2nd line request retired and
rq->engine freed/NULL already.
Lets hope I am just making unwarranted panic by being to away from
this area of the driver for a while. :) But if I am not, then it could
be that rq->engine is just overall too problematic and perhaps we
should have a look into making it reference counted by the request.
Btw there is rq->i915 but code only uses "rq->engine->i915" which way
shall we go? remove former or latter?
Simpler/shorter option should be better.
Regards,
Tvrtko
