On 08/09/2021 18:06, Daniel Vetter wrote:
On Thu, Sep 02, 2021 at 04:01:40PM +0100, Tvrtko Ursulin wrote:
On 02/09/2021 15:33, Daniel Vetter wrote:
On Tue, Aug 31, 2021 at 02:18:15PM +0100, Tvrtko Ursulin wrote:
On 31/08/2021 13:43, Daniel Vetter wrote:
On Tue, Aug 31, 2021 at 10:15:03AM +0100, Tvrtko Ursulin wrote:
On 30/08/2021 09:26, Daniel Vetter wrote:
On Fri, Aug 27, 2021 at 03:44:42PM +0100, Tvrtko Ursulin wrote:
On 27/08/2021 15:39, Tvrtko Ursulin wrote:
From: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>
In short this makes i915 work for hybrid setups (DRI_PRIME=1 with Mesa)
when rendering is done on Intel dgfx and scanout/composition on Intel
igfx.
Before this patch the driver was not quite ready for that setup, mainly
because it was able to emit a semaphore wait between the two GPUs, which
results in deadlocks because semaphore target location in HWSP is neither
shared between the two, nor mapped in both GGTT spaces.
To fix it the patch adds an additional check to a couple of relevant code
paths in order to prevent using semaphores for inter-engine
synchronisation between different driver instances.
Patch also moves singly used i915_gem_object_last_write_engine to be
private in its only calling unit (debugfs), while modifying it to only
show activity belonging to the respective driver instance.
What remains in this problem space is the question of the GEM busy ioctl.
We have a somewhat ambigous comment there saying only status of native
fences will be reported, which could be interpreted as either i915, or
native to the drm fd. For now I have decided to leave that as is, meaning
any i915 instance activity continues to be reported.
v2:
* Avoid adding rq->i915. (Chris)
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>
Can't we just delete semaphore code and done?
- GuC won't have it
- media team benchmarked on top of softpin media driver, found no
difference
You have S-curve for saturated workloads or something else? How thorough and
which media team I guess.
From memory it was a nice win for some benchmarks (non-saturated ones), but
as I have told you previously, we haven't been putting numbers in commit
messages since it wasn't allowed. I may be able to dig out some more details
if I went trawling through GEM channel IRC logs, although probably not the
actual numbers since those were usually on pastebin. Or you go an talk with
Chris since he probably remembers more details. Or you just decide you don't
care and remove it. I wouldn't do that without putting the complete story in
writing, but it's your call after all.
Media has also changed, they're not using relocations anymore.
Meaning you think it changes the benchmarking story? When coupled with
removal of GPU relocations then possibly yes.
Unless there's solid data performance tuning of any kind that gets in the
way simply needs to be removed. Yes this is radical, but the codebase is
in a state to require this.
So either way we'd need to rebenchmark this if it's really required. Also
Therefore can you share what benchmarks have been done or is it secret? As
said, I think the non-saturated case was the more interesting one here.
if we really need this code still someone needs to fix the design, the
current code is making layering violations an art form.
Anyway, without the debugfs churn it is more or less two line patch to fix
igfx + dgfx hybrid setup. So while mulling it over this could go in. I'd
just refine it to use a GGTT check instead of GT. And unless DG1 ends up
being GuC only.
The minimal robust fix here is imo to stop us from upcasting dma_fence to
i915_request if it's not for our device. Not sprinkle code here into the
semaphore code. We shouldn't even get this far with foreign fences.
Device check does not work for multi-tile. It was one of my earlier attempts
before I realized the problem. You'll see v3 which I think handles all the
cases.
There is no hw semaphores on multi-tile.
You mean because of GuC? Okay, there may not be after bringup has been done.
In which case an assert is needed somewhere just in case, if you are adamant
not to accept this fix. It may indeed not matter hugely outside of the
current transition period since I spotted patches to enable GuC on DG1. But
then again it is trivial and fixes current pains for more than just me.
But there _is_ a lot more going on than just hw semaphores that spawn
driver instances. Like priority boosting, clock boosting, and all kinds of
other things. I really dont' think it's very robust if we play
whack-a-mole here with things leaking.
You mean span not spawn? I audited those and they looks good to me. AFAIR
scheduling was in fact designed with a global lock just so that works. Plus
the cases you mention end up not holding pointers to "foreign" instances
anyway, they just do priority inheritance. Which is probably nice not to
lose if not unavoidable.
Yup span. I just think the defensive approach is better, especially since
we're planning to rework the scheduler area massively anyway.
I'm also worried about what happens when people combine a random igfx
driver from upstream with some dgpu backport, the combinatorial explosion
is nasty. Hence stopping any possible issues in dma_fence_is_i915 sounds a
lot safer.
E.g. just a quick grep says that the engine mask the busy ioctl returns is
nonsense on shared buffers with multiple i915 instances present. Probably
doesn't matter, but who knows. That was just the first one.
I had a version which deals with the busy ioctl on trybot early on,
since I was thinking along similar lines. But then I realised the
wording of the comment in there actually leaves space for
interpretation. And that actually reporting more, rather than less,
activity makes sense. And it's completely safe.
Not sure why you declare it "nonsense" since you did not really explain.
So the bullet proof way here I think is:
- change dma_fence_is_i915 to limit to our device
- use to_request in hw sempahore code too
If we later on have a need for sharing information across drivers through
dma_fence, we can then properly engineer an interface. And likely in
dma-fence.h, not somewhere in i915 code. We already have a ton of
i915-isms in that area, baking in a lot more with potential uapi impact
does not sound like a good plan.
I think the patch is simple and clear fix and it saves time people
hitting the same problem. Certainly does not make anything worse so my
2c is that it should go in while follow up work is discussed.
Regards,
Tvrtko
-Daniel
You also forgot to comment on the question lower in the email. I'll just
send a patch which removes that anyway so you can comment there.
:(
Regards,
Tvrtko
Regards,
Tvrtko
-Daniel
- pre-gen8 semaphore code was also silently ditched and no one cared
Plus removing semaphore code would greatly simplify conversion to
drm/sched.
---
drivers/gpu/drm/i915/gem/i915_gem_object.h | 17 ----------
drivers/gpu/drm/i915/i915_debugfs.c | 39 ++++++++++++++++++++--
drivers/gpu/drm/i915/i915_request.c | 12 ++++++-
3 files changed, 47 insertions(+), 21 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.h b/drivers/gpu/drm/i915/gem/i915_gem_object.h
index 48112b9d76df..3043fcbd31bd 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_object.h
+++ b/drivers/gpu/drm/i915/gem/i915_gem_object.h
@@ -503,23 +503,6 @@ i915_gem_object_finish_access(struct drm_i915_gem_object *obj)
i915_gem_object_unpin_pages(obj);
}
-static inline struct intel_engine_cs *
-i915_gem_object_last_write_engine(struct drm_i915_gem_object *obj)
-{
- struct intel_engine_cs *engine = NULL;
- struct dma_fence *fence;
-
- rcu_read_lock();
- fence = dma_resv_get_excl_unlocked(obj->base.resv);
- rcu_read_unlock();
-
- if (fence && dma_fence_is_i915(fence) && !dma_fence_is_signaled(fence))
- engine = to_request(fence)->engine;
- dma_fence_put(fence);
-
- return engine;
-}
-
void i915_gem_object_set_cache_coherency(struct drm_i915_gem_object *obj,
unsigned int cache_level);
void i915_gem_object_flush_if_display(struct drm_i915_gem_object *obj);
diff --git a/drivers/gpu/drm/i915/i915_debugfs.c b/drivers/gpu/drm/i915/i915_debugfs.c
index 04351a851586..55fd6191eb32 100644
--- a/drivers/gpu/drm/i915/i915_debugfs.c
+++ b/drivers/gpu/drm/i915/i915_debugfs.c
@@ -135,13 +135,46 @@ static const char *stringify_vma_type(const struct i915_vma *vma)
return "ppgtt";
}
+static char *
+last_write_engine(struct drm_i915_private *i915,
+ struct drm_i915_gem_object *obj)
+{
+ struct intel_engine_cs *engine;
+ struct dma_fence *fence;
+ char *res = NULL;
+
+ rcu_read_lock();
+ fence = dma_resv_get_excl_unlocked(obj->base.resv);
+ rcu_read_unlock();
+
+ if (!fence || dma_fence_is_signaled(fence))
+ goto out;
+
+ if (!dma_fence_is_i915(fence)) {
+ res = "<external-fence>";
+ goto out;
+ }
+
+ engine = to_request(fence)->engine;
+ if (engine->gt->i915 != i915) {
+ res = "<external-i915>";
+ goto out;
+ }
+
+ res = engine->name;
+
+out:
+ dma_fence_put(fence);
+ return res;
+}
+
void
i915_debugfs_describe_obj(struct seq_file *m, struct drm_i915_gem_object *obj)
{
struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
- struct intel_engine_cs *engine;
struct i915_vma *vma;
int pin_count = 0;
+ char *engine;
seq_printf(m, "%pK: %c%c%c %8zdKiB %02x %02x %s%s%s",
&obj->base,
@@ -230,9 +263,9 @@ i915_debugfs_describe_obj(struct seq_file *m, struct drm_i915_gem_object *obj)
if (i915_gem_object_is_framebuffer(obj))
seq_printf(m, " (fb)");
- engine = i915_gem_object_last_write_engine(obj);
+ engine = last_write_engine(dev_priv, obj);
if (engine)
- seq_printf(m, " (%s)", engine->name);
+ seq_printf(m, " (%s)", engine);
Or I zap this from the code altogether. Not sure it is very useful since the
only caller is i915_gem_framebuffer debugfs file and how much it can care
about maybe hitting the timing window when exclusive fence will contain
something.
Ideally we'd just look at the fence timeline name. But i915 has this very
convoluted typesafe-by-rcu reuse which means we actually can't do that,
and our fence timeline name is very useless.
Why do we even care to output any of this here? I'd just remove it since it
is a very transient state with an extremely short window of opportunity to
make it show anything. Which I think makes it pretty useless in debugfs.
Regards,
Tvrtko
Would be good to fix that, Matt Auld has started an attempt but didn't get
very far.
-Daniel
Regards,
Tvrtko
}
static int i915_gem_object_info(struct seq_file *m, void *data)
diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c
index ce446716d092..64adf619fe82 100644
--- a/drivers/gpu/drm/i915/i915_request.c
+++ b/drivers/gpu/drm/i915/i915_request.c
@@ -1152,6 +1152,12 @@ __emit_semaphore_wait(struct i915_request *to,
return 0;
}
+static bool
+can_use_semaphore_wait(struct i915_request *to, struct i915_request *from)
+{
+ return to->engine->gt == from->engine->gt;
+}
+
static int
emit_semaphore_wait(struct i915_request *to,
struct i915_request *from,
@@ -1160,6 +1166,9 @@ emit_semaphore_wait(struct i915_request *to,
const intel_engine_mask_t mask = READ_ONCE(from->engine)->mask;
struct i915_sw_fence *wait = &to->submit;
+ if (!can_use_semaphore_wait(to, from))
+ goto await_fence;
+
if (!intel_context_use_semaphores(to->context))
goto await_fence;
@@ -1263,7 +1272,8 @@ __i915_request_await_execution(struct i915_request *to,
* immediate execution, and so we must wait until it reaches the
* active slot.
*/
- if (intel_engine_has_semaphores(to->engine) &&
+ if (can_use_semaphore_wait(to, from) &&
+ intel_engine_has_semaphores(to->engine) &&
!i915_request_has_initial_breadcrumb(to)) {
err = __emit_semaphore_wait(to, from, from->fence.seqno - 1);
if (err < 0)
