On 6/18/2024 2:33 AM, Daniel Vetter wrote:
On Mon, Jun 17, 2024 at 10:52:27PM +0300, Dmitry Baryshkov wrote:
On Mon, Jun 17, 2024 at 11:28:35AM GMT, Abhinav Kumar wrote:
Hi
On 6/17/2024 9:54 AM, Brian Starkey wrote:
Hi,
On Mon, Jun 17, 2024 at 05:16:36PM +0200, Daniel Vetter wrote:
On Mon, Jun 17, 2024 at 01:41:59PM +0000, Hoosier, Matt wrote:
Hi,
There is a discussion ongoing over in the compositor world about the implication of this cautionary wording found in the documentation for the DRM_MODE_CONNECTOR_WRITEBACK connectors:
* "WRITEBACK_OUT_FENCE_PTR":
* Userspace can use this property to provide a pointer for the kernel to
* fill with a sync_file file descriptor, which will signal once the
* writeback is finished. The value should be the address of a 32-bit
* signed integer, cast to a u64.
* Userspace should wait for this fence to signal before making another
* commit affecting any of the same CRTCs, Planes or Connectors.
* **Failure to do so will result in undefined behaviour.**
* For this reason it is strongly recommended that all userspace
* applications making use of writeback connectors *always* retrieve an
* out-fence for the commit and use it appropriately.
* From userspace, this property will always read as zero.
The question is whether it's realistic to hope that a DRM writeback
connector can produce results on every frame, and do so without dragging
down the frame-rate for the connector.
The wording in the documentation above suggests that it is very likely
the fence fd won't signal userspace until after the vblank following the
scanout during which the writeback was applied (call that frame N). This
would mean that the compositor driving the connector would typically be
unable to legally queue a page flip for frame N+1.
Is this the right interpretation? Is the writeback hardware typically
even designed with a streaming use-case in mind? Maybe it's just
intended for occasional static screenshots.
So typically writeback hardware needs its separate crtc (at least the
examples I know of) and doesn't make a lot of guarantees that it's fast
enough for real time use. Since it's a separate crtc it shouldn't hold up
the main composition loop, and so this should be all fine.
On Mali-DP and Komeda at least, you can use writeback on the same CRTC
that is driving a "real" display, and it should generally work. If the
writeback doesn't keep up then the HW will signal an error, but it was
designed to work in-sync with real scanout, on the same pipe.
Same with MSM hardware. You can use writeback with same CRTC that is driving
a "real" display and yes we call it concurrent writeback. So I think it is
correct in the documentation to expect to wait till this is signaled if the
same CRTC is being used.
TIL
If/when we have hardware and driver support where you can use the
writeback connector as a real-time streamout kind of thing, then we need
to change all this, because with the current implementation, there's
indeed the possibility that funny things can happen if you ignore the
notice (funny as in data corruption, not funny as the kernel crashes of
course).
Indeed, the wording was added (from what I remember from so long
ago...) because it sounded like different HW made very different
guarantees/non-guarantees about what data would be written when, so
perhaps you'd end up with some pixels from the next frame in your
buffer or something.
Taking Mali-DP/Komeda again, the writeback configuration is latched
along with everything else, and writeback throughput permitting, it
should "just work" if you submit a new writeback every frame. It
drains out the last of the data during vblank, before starting on the
next frame. That doesn't help the "general case" though.
Would it be fair to summarize it like below:
1) If the same CRTC is shared with the real time display, then the hardware
is expected to fire this every frame so userspace should wait till this is
signaled.
As I wrote in response to another email in this thread, IMO existing
uAPI doesn't fully allow this. There is no way to enforce 'vblank'
handling onto the userspace. So userspace should be able to supply at
least two buffers and then after the vblank it should be able to enqueue
the next buffer, while the filled buffer is automatically dequeued by
the driver and is not used for further image output.
Sorry for the late response. What I meant was, if we are using
concurrent writeback with the real time display, it should be capable of
running at the same speed as the real time display. I do not have the
numbers to share but atleast that's the expectation.
But, yes I do admit that current UAPI does not fully allow having a
queue depth for WB FBs. And having it will help us.
Yeah if you want streaming writeback we need a queue depth of at least 2
in the kms api. Will help a lot on all hardware, but on some it's required
because the time when the writeback buffer is fully flushed is after the
point of no return for the next frame (which is when the vblank event is
supposed to go out).
I think over the years we've slowly inched forward to make at least the
drm code safe for a queue depth of 2 in the atomic machinery, but the
writeback and driver code probably needs a bunch of work.
-Sima
2) If a different CRTC is used for the writeback, then the composition loop
for the real time display should not block on this unless its a mirroring
use-case, then we will be throttled by the lowest refresh rate anyway.
what is mirroring in this case? You have specified that a different CRTC
is being used.
Definition of mirroring could be thought of in two ways:
1) in clone mode, the WB is running at the same rate as the real time
display and hence if we are mirroring the content this way there is same
CRTC.
2) lets say I want to mirror my content using wifi display but the
end-monitor is running on a different resolution and fps, then I cannot
use clone mode in this case right because the CRTC which the writeback
is using will be programmed for a different mode than the real time display.
For the second case, it is still mirroring the content but with a
different CRTC so will be slowed down by the slowest display otherwise
the displays will go out of sync. This is what I meant in this use-case.
If we already have devices where you can use writeback together with real
outputs, then I guess that counts as an oopsie :-/
Well "works fine" fits into the "undefined behaviour" bucket, just as
well as "corrupts your fb" does :-)
--
With best wishes
Dmitry
