syncobj: use eventfd instead of stalling fd checks #9437
Conversation
gulafaran
force-pushed
the
eventfd
branch
4 times, most recently
from
53df780 to
e331b6c
Compare
|
wait... for ds, shouldn't we release the wl_buffer whenever DRM signals the out fence? edit:
|
There is a lot wrong at the moment, explicit sync currently relies on new buffer attached, the client can commit multiple times on the same buffer if it wants to, and does sometimes, and this PR probably exposed that quirk even more, and yeah direct scanout i havent begun reading into but removed that buffer manual buffer shenanigans because it was conflicting with the changes i made, there is also a few other things im checking that has popped up while testing |
And "releasing" when it comes to explicit sync is sending the release point not the buffer .release |
|
Note that |
|
I tested this pr & it fixed glmark2 lag but minecraft lag is still the same and also it introduced slight lag in other areas like workspace switching and cursor lag when wlogout is opened. |
we use that fd, i think we import it as a sync_file, then wait for it to signal, meaning aq plane is done with the primary buffer |
yeah got a few very broken things to figure out, im trying tho. |
|
@gulafaran you did amazing work, now everything runs smoothly. I ran various gpu stress tests, they ran without any lag. |
|
for this part the lockedByBackend is unlocked when page flip buffer swap happens, so technically that's when we should signal release point since DRM scanout is no longer using the buffer |
|
I can also say that this fixed my issue with my RX 580 so it might fix something like #8119 and others. |
Probably will, and also some minecraft flickering and other things, im pushing soon a fix for a gtk issue and hyprlock issue so its not quite flawless yet 👍 |
should be in a better state now, but not done yet. |
|
WFM as well, fixed desync problems. |
cleanup a bit missing removals if resource not good, erasing from containers etc. make use of unique ptrs instead. and add default destructors.
remove early buffer release that was breaking explicit sync, the buffer needs to exist until the surface commit event has been emitted and draw calls added egl sync points, move to eventfd signaling instead of stalling sync point checks, and recommit pending commits if waiting on a signal. add a CDRMSyncPointState helper class. move a few weak pointers to shared pointers so they dont destruct before we need to use them.
eventfd requires us to queue pending stats until ready and then apply to current, and also when no ready state exist commit the client commit on the current existing buffer, if there is one.
clear current buffer damage on current buffer commits.
remove unused code, and ensure we dont commit a empty texture causing locksession protocol and gtk4-layer-shell misbehaving.
ensure the containers having the various buffers actually gets cleaned up from their containers, incase the CSignal isnt signaled because of expired smart pointers or just wrong order destruction because mishaps. also move the acquire/point setting to buffer attaching. instead of on precommit.
remove unused code and merge sync fds if fence is valid, remove manual directscanout buffer dropping that signals release point on pageflip, it can cause us to signal the release point while still keeping the current buffer and rendering it yet again causing wrong things.
delay buffer releases on non syncobj surfaces until next commit, and check on async buffers if syncobj and drop and signal the release point on backend buffer release.
ensure we follow protocol by replacing acquire/release points if they arrive late and replace already existing ones. also remove unneded brackets, and dont try to manual lock/release buffers when it comes to explicit protocol. it doesnt care about buffer releases only about acquire and release points and signaling them.
set points in precommit, before checking protocol errors and we catch any pending acquire/release points arriving late.
remove destructor resource destroying, let resources destroys them on their events, and move SSurfaceStates to types/SurfaceState.hpp
have to actually store the mergedfd to use it.
ensure the current asynchronous buffer is actually released on pageflip not the previous. cleanup a bit FD handling in commitPendingAndDoExplicitSync, and reuse the in fence when syncing surfaces.
calling resetexplicitfence without properly ensuring the FD is closed before will leak it, store it per monitor and let it close itself with the CFileDescriptor class.
buffers were never being sent released properly.
ensure the infence fd survives the scope of attemptdirectscanout so it doesnt close before it should have.
we might hit a race to finish on exit where the timeline just has destructed but the buffer waiter is still pending. and such we removeAllWaiters null dereferences.
remove quack comment, change to m_foo and use a std::vector and weakpointer in the waiter for removal instead of a std::list.
remove unused async buffer drop, only related to directscanout and is handled elsewhere.
use eventfd and add it to the event loop and when it recieves a signal release the the queued surfacestate, this means we dont stall entire compositor when waiting for materilisation of the fd. and change its related usage.
should solve a lot of "compositor 5 fps while ingame overlay shows 300fps" and other similiar weird stalls.
on nvidia this shuffles compositor stalls to sometimes a noticeable rendering stutter, technically the acquire point isnt signaled fast enough so the buffer is stuck and nothing gets rendered until it is. pretty much the same stall as seen before only now its per surface rendering. but that i contribute to nvidia driver being simply shit, and is visible on other compositors.
NVIDIA/open-gpu-kernel-modules#777
NVIDIA/open-gpu-kernel-modules#743
fixes #7857
fixes #9340
fixes #9340
fixes #8588
fixes #7643
fixes #7317
fixes #9376
fixes #6844
fixes #6912
potentially helps:
#6617
#9011
remaining quirks to be fixed