On 27/03/2017 11:31, Chris Wilson wrote:
On Mon, Mar 27, 2017 at 11:11:47AM +0100, Tvrtko Ursulin wrote:
On 26/03/2017 09:46, Chris Wilson wrote:
Unlocking is dangerous. In this case we combine an early update to the
out-of-queue request, because we know that it will be inserted into the
correct FIFO priority-ordered slot when it becomes ready in the future.
However, given sufficient enthusiasm, it may become ready as we are
continuing to reschedule, and so may gazump the FIFO if we have since
dropped its spinlock. The result is that it may be executed too early,
before its dependees.
Fixes: 20311bd35060 ("drm/i915/scheduler: Execute requests in order of priorities")
Testcase: igt/gem_exec_whisper
Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
Cc: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>
Cc: <stable@xxxxxxxxxxxxxxx> # v4.10+
---
drivers/gpu/drm/i915/intel_lrc.c | 54 +++++++++++++++++++++++++++-------------
1 file changed, 37 insertions(+), 17 deletions(-)
diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
index dd0e9d587852..3fdabba0a32d 100644
--- a/drivers/gpu/drm/i915/intel_lrc.c
+++ b/drivers/gpu/drm/i915/intel_lrc.c
@@ -658,30 +658,47 @@ static void execlists_submit_request(struct drm_i915_gem_request *request)
spin_unlock_irqrestore(&engine->timeline->lock, flags);
}
-static struct intel_engine_cs *
-pt_lock_engine(struct i915_priotree *pt, struct intel_engine_cs *locked)
+static inline struct intel_engine_cs *
+pt_lock_engine(struct i915_priotree *pt, unsigned long *locked)
{
- struct intel_engine_cs *engine;
-
- engine = container_of(pt,
- struct drm_i915_gem_request,
- priotree)->engine;
- if (engine != locked) {
- if (locked)
- spin_unlock_irq(&locked->timeline->lock);
- spin_lock_irq(&engine->timeline->lock);
- }
+ struct intel_engine_cs *engine =
+ container_of(pt, struct drm_i915_gem_request, priotree)->engine;
+
+ /* Locking the engines in a random order will rightfully trigger a
+ * spasm in lockdep. However, we can ignore lockdep (by marking each
+ * as a seperate nesting) so long as we never nest the
+ * engine->timeline->lock elsewhere. Also the number of nesting
+ * subclasses is severely limited (7) which is going to cause an
+ * issue at some point.
+ * BUILD_BUG_ON(I915_NUM_ENGINES >= MAX_LOCKDEP_SUBCLASSES);
Lets bite the bullet and not hide this BUILD_BUG_ON in a comment. :I
The code would continue to work nevertheless, just lockdep would
eventually give up. I like it slightly better than taking either a
global spinlock for engine->execlists_queue insertion, or taking the
spinlock on every engine for scheduling. How often will we reschedule
across engines? Not sure.
I think counting on "doesn't happen often" and "it still works" falls
short of your high standards! ;) So a global execlist_queue lock if it
must be..
+ */
+ if (!__test_and_set_bit(engine->id, locked))
+ spin_lock_nested(&engine->timeline->lock,
+ hweight_long(*locked));
return engine;
}
+static void
+unlock_engines(struct drm_i915_private *i915, unsigned long locked)
+{
+ struct intel_engine_cs *engine;
+ unsigned long tmp;
+
+ for_each_engine_masked(engine, i915, locked, tmp)
+ spin_unlock(&engine->timeline->lock);
+}
+
static void execlists_schedule(struct drm_i915_gem_request *request, int prio)
{
- struct intel_engine_cs *engine = NULL;
+ struct intel_engine_cs *engine;
struct i915_dependency *dep, *p;
struct i915_dependency stack;
+ unsigned long locked = 0;
LIST_HEAD(dfs);
+ BUILD_BUG_ON(I915_NUM_ENGINES > BITS_PER_LONG);
+
if (prio <= READ_ONCE(request->priotree.priority))
return;
@@ -691,6 +708,9 @@ static void execlists_schedule(struct drm_i915_gem_request *request, int prio)
stack.signaler = &request->priotree;
list_add(&stack.dfs_link, &dfs);
+ GEM_BUG_ON(irqs_disabled());
+ local_irq_disable();
+
Why not just irqsave/restore? Sounds like too low level for this
position in the flow. If just optimisation it would need a comment I
think.
It was because we are not taking the spin lock/unlock inside the same
block, so it felt dangerous. Who holds the irqflags?
Hm yes, it cannot be made elegant.
/* Recursively bump all dependent priorities to match the new request.
*
* A naive approach would be to use recursion:
@@ -719,7 +739,7 @@ static void execlists_schedule(struct drm_i915_gem_request *request, int prio)
if (!RB_EMPTY_NODE(&pt->node))
continue;
- engine = pt_lock_engine(pt, engine);
+ engine = pt_lock_engine(pt, &locked);
/* If it is not already in the rbtree, we can update the
* priority inplace and skip over it (and its dependencies)
@@ -737,7 +757,7 @@ static void execlists_schedule(struct drm_i915_gem_request *request, int prio)
INIT_LIST_HEAD(&dep->dfs_link);
- engine = pt_lock_engine(pt, engine);
+ engine = pt_lock_engine(pt, &locked);
if (prio <= pt->priority)
continue;
@@ -750,8 +770,8 @@ static void execlists_schedule(struct drm_i915_gem_request *request, int prio)
engine->execlist_first = &pt->node;
}
- if (engine)
- spin_unlock_irq(&engine->timeline->lock);
+ unlock_engines(request->i915, locked);
+ local_irq_enable();
/* XXX Do we need to preempt to make room for us and our deps? */
}
I am trying to think whether removing the skip on requests not in
the execution tree would work and help any.
It's dangerous due to the duplicate branches in the dependency graph that
we are resolving to generate the topological ordering. We need a way to
do a mark-and-sweep whilst also ensuring that we end up with the correct
order. I'm open to (better :) suggestions.
Or if the above scheme
is completely safe or we would need to lock atomically all engines
requests on which will be touched. Especially since the code is only
dealing with adjusting the priorities so I don't immediately see how
it can cause out of order execution.
interrupt leading to submit_request, which wants to then insert a
request into the execlist_queue rbtree vs ->schedule() also trying to
manipulate the rbtree (and in this case elements currently outside of the
rbtree). Our insertion into the rbtree ensures fifo so that we don't
reorder the equivalent priority dependencies during ->schedule(), hence
if we mark an out-of-rbtree request as a higher priority before
inserting all of its dependencies into the tree, if the submit_notify
occurs, it will insert the request into the tree before we get to insert
its dependencies, hence reordering.
Ok I get the general idea. I don't have any better suggestions at the
moment than trying the global lock. Luckily you have just removed one
atomic from the irq handler so one step forward, two steps back. :)
Regards,
Tvrtko
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx