When drmModePageFlip() or drmModeAtomicCommit() unexpectedly failed (e.g.
ENOSPC, which has been seen in the wild), this failure was not handled
very gracefully. The page flip listener for the scanout was left in the
MetaKmsUpdate, meaning when the primary plane composition was later page
flipped, two page flip listeners were added, one for the primary plane,
and one for the scanout. This caused the 'page-flipped' event to be
handled twice, the second time being fatal.
Handle this by making 'no-discard' listener flag be somewhat reversed,
and say 'drop-on-error', and then drop all 'drop-on-error' listeners
when a MetaKmsUpdate failed to be processed.
Also for a "preserve" flagged update, don't ever trigger "discard"
callbacks just yet, as preserved updates are used again for the primary
plane composition, in order to not miss e.g. CRTC gamma updates, or
cursor plane updates, which were added separately.
Closes: https://gitlab.gnome.org/GNOME/mutter/-/issues/1809
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1910>
The first phase happens early, which discards pending page flips,
meaning the references held by those page flip closures are released.
The second phase happens late, after other units depending on the KMS
abstraction, have been cleaned up.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1822>
The current API as all synchronous, so they can be made to return
feedback immediately. This will be needed for the cursor renderer which
needs to know whether it should fall back to OpenGL cursor rendering.
https://gitlab.gnome.org/GNOME/mutter/merge_requests/930
mode_set_fallback() schedules a call to mode_set_fallback_feedback_idle(), but
it is possible for Mutter to repaint before the idle callbacks are dispatched.
If that happens, mode_set_fallback_feedback_idle() does not get called before
Mutter enters wait_for_pending_flips(), leading to a deadlock.
Add the needed interfaces so that meta_kms_device_dispatch_sync() can flush all
the implementation idle callbacks before it checks if any "events" are
available. This prevents the deadlock by ensuring
mode_set_fallback_feedback_idle() does get called before potentially waiting
for actual DRM events.
Presumably this call would not be needed if the implementation was running in
its own thread, since it would eventually dispatch its idle callbacks before
going to sleep polling on the DRM fd. This call might even be unnecessary
overhead in that case, synchronizing with the implementation thread needlessly.
But the thread does not exist yet, so this is needed for now.
This is part 1 of 2 fixing a complete desktop freeze when drmModePageFlip()
fails with EINVAL and the fallback to drmModeSetCrtc() succeeds but the success
is not registered correctly as completed "flip". The freeze occurs under
wait_for_pending_flips() which calls down into meta_kms_impl_device_dispatch()
which ends up poll()'ing the DRM fd even though drmModeSetCrtc() will not
produce a DRM event, hence the poll() never returns. The freeze was observed
when hotplugging a DisplayLink dock for the first time on Ubuntu 19.10.
CC stable: gnome-3-34
https://gitlab.gnome.org/GNOME/mutter/merge_requests/953
This commit introduces, and makes use of, a transactional API used for
setting up KMS state, later to be applied, potentially atomically. From
an API point of view, so is always the case, but in the current
implementation, it still uses legacy drmMode* API to apply the state
non-atomically.
The API consists of various buliding blocks:
* MetaKmsUpdate - a set of configuration changes, the higher level
handle for handing over configuration to the impl backend. It's used to
set mode, assign framebuffers to planes, queue page flips and set
connector properties.
* MetaKmsPlaneAssignment - the assignment of a framebuffer to a plane.
Currently used to map a framebuffer to the primary plane of a CRTC. In
the legacy KMS implementation, the plane assignment is used to derive
the framebuffer used for mode setting and page flipping.
This also means various high level changes:
State, excluding configuring the cursor plane and creating/destroying
DRM framebuffer handles, are applied in the end of a clutter frame, in
one go. From an API point of view, this is done atomically, but as
mentioned, only the non-atomic implementation exists so far.
From MetaRendererNative's point of view, a page flip now initially
always succeeds; the handling of EBUSY errors are done asynchronously in
the MetaKmsImpl backend (still by retrying at refresh rate, but
postponing flip callbacks instead of manipulating the frame clock).
Handling of falling back to mode setting instead of page flipping is
notified after the fact by a more precise page flip feedback API.
EGLStream based page flipping relies on the impl backend not being
atomic, as the page flipping is done in the EGLStream backend (e.g.
nvidia driver). It uses a 'custom' page flip queueing method, keeping
the EGLStream logic inside meta-renderer-native.c.
Page flip handling is moved to meta-kms-impl-device.c from
meta-gpu-kms.c. It goes via an extra idle callback before reaching
meta-renderer-native.c to make sure callbacks are invoked outside of the
impl context.
While dummy power save page flipping is kept in meta-renderer-native.c, the
EBUSY handling is moved to meta-kms-impl-simple.c. Instead of freezing the
frame clock, actual page flip callbacks are postponed until all EBUSY retries
have either succeeded or failed due to some other error than EBUSY. This
effectively inhibits new frames to be drawn, meaning we won't stall waiting on
the file descriptor for pending page flips.
https://gitlab.gnome.org/GNOME/mutter/issues/548https://gitlab.gnome.org/GNOME/mutter/merge_requests/525
The intention with KMS abstraction is to hide away accessing the drm
functions behind an API that allows us to have different kind of KMS
implementations, including legacy non-atomic and atomic. The intention
is also that the code interacting with the drm device should be able to
be run in a different thread than the main thread. This means that we
need to make sure that all drm*() API usage must only occur from within
tasks that eventually can be run in the dedicated thread.
The idea here is that MetaKms provides a outward facing API other places
of mutter can use (e.g. MetaGpuKms and friends), while MetaKmsImpl is
an internal implementation that only gets interacted with via "tasks"
posted via the MetaKms object. These tasks will in the future
potentially be run on the dedicated KMS thread. Initially, we don't
create any new threads.
Likewise, MetaKmsDevice is a outward facing representation of a KMS
device, while MetaKmsImplDevice is the corresponding implementation,
which only runs from within the MetaKmsImpl tasks.
This commit only moves opening and closing the device to this new API,
while leaking the fd outside of the impl enclosure, effectively making
the isolation for drm*() calls pointless. This, however, is necessary to
allow gradual porting of drm interaction, and eventually the file
descriptor in MetaGpuKms will be removed. For now, it's harmless, since
everything still run in the main thread.
https://gitlab.gnome.org/GNOME/mutter/issues/548https://gitlab.gnome.org/GNOME/mutter/merge_requests/525
