Xserver: Present extension backport from X.org #423
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Provides both a software implementation using timers and driver hooks to base everything on vblank intervals. Signed-off-by: Keith Packard <[email protected]> Reviewed-by: Adam Jackson <[email protected]>
A client destroying objects in the middle of an unflip can end up having the screen flip window or fence set to NULL in the unflip notify path. Check for these and don't try to use those objects. Signed-off-by: Keith Packard <[email protected]> Reviewed-by: Adam Jackson <[email protected]>
…gnals Signed-off-by: Keith Packard <[email protected]> Reviewed-by: Adam Jackson <[email protected]>
If the timer fired too early, we'd sometimes mis-compute the MSC for fake vblanks. Rounding the computation to the nearest MSC fixes this nicely. Signed-off-by: Keith Packard <[email protected]> Reviewed-by: Adam Jackson <[email protected]>
We use event_id 0 to mean 'no such event'; if a driver sends us that event_id, make sure we don't accidentally match it. Signed-off-by: Keith Packard <[email protected]> Reviewed-by: Adam Jackson <[email protected]>
This eliminates dereferencing freed window pointers when there is a flip for that window in progress. The flip will complete, and then immediately get undone (as we can't stop an in-progress flip). Remove the vblank->window_destroyed field as we can signal this with vblank->window == NULL instead. Change check to vblank->window == NULL in: present_flip_notify Add check for vblank->window == NULL in: present_vblank_notify present_execute present_flip_notify was also using vblank->window->drawable.pScreen, so stop doing that and use vblank->screen instead. Signed-off-by: Keith Packard <[email protected]> Reviewed-by: Adam Jackson <[email protected]>
Pend presentation until wait_fence is also triggered by having the SyncFence trigger invoke present_execute once triggered. Signed-off-by: Keith Packard <[email protected]> Reviewed-by: Adam Jackson <[email protected]>
unflip happens after the clip lists have been updated, so instead of smashing the whole screen and drawing over other windows, just draw to the original flip window; it'll have the right clip list and so the copy will work just fine. Signed-off-by: Keith Packard <[email protected]> Reviewed-by: Adam Jackson <[email protected]>
When an application provides two pixmaps for the same MSC, the previous one is skipped. This just dumps out some information at that point Signed-off-by: Keith Packard <[email protected]> Reviewed-by: Adam Jackson <[email protected]>
If the window is destroyed, then we've got cleanup work to do, even if the vblank has already been executed -- we need to clear the window pointer so that we don't try to deliver events to it. Leaving it on the window list meant that when walking that list, we need to know whether the vblank is waiting to be executed or waiting for the flip to complete, so a new 'queued' flag was added to the vblank to distinguish between the two states. Signed-off-by: Keith Packard <[email protected]> Reviewed-by: Adam Jackson <[email protected]>
This makes other drawing to the window appear on the screen. Note that no child windows can be affected because only full-screen windows are eligible for flipping, and so we only need to set pixmap for the window itself. Signed-off-by: Keith Packard <[email protected]> Reviewed-by: Adam Jackson <[email protected]>
This allows GL to support the GLX_INTEL_swap_event extension. v2: Return GLX_BLIT_COMPLETE_INTEL for unknown swap types Signed-off-by: Keith Packard <[email protected]> Reviewed-by: Adam Jackson <[email protected]>
If the driver doesn't have the necessary hooks for Present, then the target_crtc needs to be set to NULL to make sure the extension uses the present_fake code. Signed-off-by: Keith Packard <[email protected]> Tested-by: Fredrik Höglund <[email protected]> Reviewed-by: Adam Jackson <[email protected]>
This makes sure that things like software cursors continue to work while the screen is flipped. Signed-off-by: Keith Packard <[email protected]> Reviewed-by: Adam Jackson <[email protected]>
Limit damage to the 'update' region. Signed-off-by: Keith Packard <[email protected]> Reviewed-by: Adam Jackson <[email protected]>
Among much else Present depends on RANDR types, and RANDR isn't properly Xinerama-aware yet anyway. Reviewed-by: Julien Cristau <[email protected]> Signed-off-by: Adam Jackson <[email protected]>
Newly created windows inherit the pixmap of their parent, similarly, reparenting a tree inherits the pixmap of the destination tree. Making present preserve the invariant that unredirected windows always have the same pixmap as their parent ensures that the above cases work correctly. v2: name the recursive function to 'set_tree_pixmap' instead of 'set_window_pixmap' Signed-off-by: Keith Packard <[email protected]> Reviewed-by: Adam Jackson <[email protected]>
We want to advertise the version we implement, not the version the protocol headers happen to describe. Reviewed-by: Jasper St. Pierre <<[email protected]> Signed-off-by: Adam Jackson <[email protected]>
Check for Async flag and execute immediately if set, otherwise wait for the next appropriate vblank before copying. Signed-off-by: Keith Packard <[email protected]>
Presents which are not marked 'queued' and are in the window present list are waiting for the flip event; discarding those won't work very well (it'll end up trashing displayed content for the next frame), so skip over those when looking for duplicate frame presents Signed-off-by: Keith Packard <[email protected]>
Skipped present pixmap calls were not setting the mode to PresentCompleteModeSkip for skipped operations. Signed-off-by: Keith Packard <[email protected]>
This lets us stop using the 'pointer' typedef in Xdefs.h as 'pointer' is used throughout the X server for other things, and having duplicate names generates compiler warnings. Signed-off-by: Keith Packard <[email protected]> Reviewed-by: Eric Anholt <[email protected]>
When a flip (or unflip) is pending and a flip request comes in, leave it queued until the pending flip completes and then execute it. This fixes a bug where an application submitting back-to-back present_pixmap requests for sequential frames would alternate between flipping and copying as the pending flip would cause the new present_pixmap request to not use a flip. Signed-off-by: Keith Packard <[email protected]> Reviewed-by: Chris Wilson <[email protected]> Tested-by: Frank Binns <[email protected]>
Once the vblank is actually getting executed, it's lifetime is no longer tied to the window, and so it shouldn't be controlled by window destruction. In particular, if the vblank is queued for flip, it will get stored in the flip_pending field, and will be correctly destroyed when the flip completes. Signed-off-by: Keith Packard <[email protected]>
If a 2D application is started on top of a fullscreen 3D application, which is flipping, then we need to stop flipping and restore the root window, and possibly the flip window, to using the screen pixmap. Normally this would be done as part of an unflip. However, in the case that there is a pending flip there is no mechanism to abort so the unflip is deferred until the pending flip completes. This provides a window of opportunity for the 2D application to draw to the wrong pixmap. Restore the screen pixmap at the point a pending flip is marked as aborted, thus avoiding this issue. Reviewed-by: Keith Packard <[email protected]> Signed-off-by: Frank Binns <[email protected]> Signed-off-by: Keith Packard <[email protected]>
If we present several pixmaps in advance for different msc, the later one shouldn't cancel the previous ones. This reverts a change made by commit e6f5d9d7b7efdacea0f22f1808efca849bcede4c Without this fix, vblank_mode=0 glxgears doesn't update with the present fallback. Signed-off-by: Axel Davy <[email protected]> Reviewed-by: Keith Packard <[email protected]> Signed-off-by: Keith Packard <[email protected]>
When a window moves from one CRTC to another, present_window_to_crtc_msc
updates window_priv->msc_offset according to the delta between the
current MSC values of the old and new CRTC:
window_priv->msc_offset += new_msc - old_msc;
window_priv->msc_offset is initially 0, so if new_msc < old_msc,
window_priv->msc_offset wraps around and becomes a large number. If the
window_msc parameter passed in is small (in particular if it's 0, such as
is the case when the client just wants to know the current window MSC
value), the returned CRTC MSC value may still be a large number. In that
case, the existing MSC comparisons in pixmap_present weren't working as
intended, resulting in scheduling a wait far into the future when the
target MSC had actually already passed. This would result in the client
(e.g. the Chromium browser) hanging when moving its window between CRTCs.
In order to fix this, introduce msc_is_(equal_or_)after helper functions
which take the wraparound into account for comparing two MSC values.
Signed-off-by: Michel Dänzer <[email protected]>
Reviewed-by: Keith Packard <[email protected]>
Reviewed-by: Martin Peres <[email protected]>
Signed-off-by: Keith Packard <[email protected]>
To make them usable from any other function in the file. No functional change. Reviewed-by: Chris Wilson <[email protected]> Signed-off-by: Michel Dänzer <[email protected]>
For flipping, we wait for the MSC before the target MSC and then call the driver flip hook. If the latter fails, we have to wait for the target MSC before falling back to a copy, or else it's executed too early. Fixes glxgears running at unbounded framerate (not synchronized to the refresh rate) in fullscreen if the driver flip hook fails. Reviewed-by: Chris Wilson <[email protected]> Signed-off-by: Michel Dänzer <[email protected]>
While present_pixmap decrements target_msc by 1 for present_queue_vblank, it leaves the original vblank->target_msc intact. So incrementing the latter for requeueing resulted in the requeued presentation being executed too late. Also, no need to requeue if the target MSC is already reached. This further reduces stutter when a popup menu appears on top of a flipping fullscreen window. Reviewed-by: Chris Wilson <[email protected]> Signed-off-by: Michel Dänzer <[email protected]>
present_unflip may be called several times from present_check_flip_window
during the same unflip. We can only copy to the screen pixmap the first
time, otherwise we may scribble over other windows. The flip pixmap
contents don't get updated after the first time anyway.
Fixes at least the following problems, which were introduced by commit
806470b9 ("present: Copy unflip contents back to the Screen Pixmap"):
On xfwm4 without compositing, run glxgears and put its window into
fullscreen mode to start flipping. While in fullscreen, open the xfwm4
window menu by pressing Alt-Space. The window menu was invisible most
of the time because it was getting scribbled over by a repeated unflip
copy.
When switching a flipping window out of fullscreen, a repeated unflip
copy could leave artifacts of the flip pixmap on the desktop.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94325
Reviewed-by: Keith Packard <[email protected]>
Reviewed-by: Chris Wilson <[email protected]>
…p (v2)
The following fix will use the refactored function.
The logic in the refactored function is slightly simplified, exploiting
the fact that this function is only ever called with a valid flip
pixmap.
v2: Assert that flip_pixmap is non-NULL instead of testing and bailing
on NULL, preserve test for flip_window being non-NULL before calling
present_set_tree_pixmap for it (Keith Packard)
Reviewed-by: Chris Wilson <[email protected]> (v1)
Reviewed-by: Keith Packard <[email protected]>
After present_set_abort_flip, the screen pixmap will be used for all screen drawing, so we need to restore the current flip pixmap contents to the screen pixmap here as well. Improves flashing / stutter e.g. when something like a popup menu appears on top of a flipping fullscreen window or when switching out of fullscreen. Note that this means present_set_abort_flip now relies on screen->root being non-NULL, but that's already the case in other present code. Reviewed-by: Keith Packard <[email protected]> Reviewed-by: Chris Wilson <[email protected]>
Due to the way present currently works, we don't ever check with the secondary adapters if we can flip at all. We shouldn't flip if the secondary adapters are attached to the pixmap currently, however using the current check_flip callback isn't possible as it passes the Window to the driver (something we shouldn't be doing), so the slave driver can never get it's own screen ptr back. For now to fix the problem just block flips if we have any slaves configured. We can fix the ABI up later, but this fix can be backported to stable. Signed-off-by: Dave Airlie <[email protected]> Reviewed-by: Keith Packard <[email protected]>
This code was added to deal with the driver present hook failing, in which case we need to wait for the next MSC before executing the presentation. However, it could also take effect in cases where the driver incorrectly thinks the current MSC matches the target one (e.g. due to the kernel interface only supporting 32-bit MSC values), in which case it could result in the presentation getting requeued over and over. To prevent such issues, check specifically for the target MSC immediately following the current MSC. Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94596 Signed-off-by: Michel Dänzer <[email protected]> Reviewed-by: Keith Packard <[email protected]>
With large numbers of queued vblank, the list iteration on every interupt dominates processing time. If we reorder the list to be in ascending event order, then not only is also likely to be in order for notification queries (i.e. the notification will be near the start of the list), we can also stop iterating when past the target event_id. Signed-off-by: Chris Wilson <[email protected]> Reviewed-and-tested-by: Mario Kleiner <[email protected]> Signed-off-by: Hans de Goede <[email protected]>
The flip queue currently only holds events submitted to the driver for flipping, awaiting the completion notifier. It is short. We therefore can speed up interrupt processing by keeping the small number of events ready to be flipped on the end of the flip queue. By appending the events to the flip_queue in the order that they become ready, we also resolve one issue causing Present to display frames out of order. Signed-off-by: Chris Wilson <[email protected]> Reviewed-and-tested-by: Mario Kleiner <[email protected]> Signed-off-by: Hans de Goede <[email protected]>
…oad slaves A single provider can be both a offload and source slave at the same time, the use of seperate lists breaks in this case e.g. : xrandr --listproviders Providers: number : 2 Provider 0: id: 0x7b cap: 0xf, Source Output, Sink Output, Source Offload, Sink Offload crtcs: 3 outputs: 2 associated providers: 0 name:modesetting Provider 1: id: 0x46 cap: 0xf, Source Output, Sink Output, Source Offload, Sink Offload crtcs: 2 outputs: 5 associated providers: 0 name:modesetting xrandr --setprovideroutputsource 1 0x7b xrandr --listproviders Providers: number : 2 Provider 0: id: 0x7b cap: 0xf, Source Output, Sink Output, Source Offload, Sink Offload crtcs: 3 outputs: 2 associated providers: 1 name:modesetting Provider 1: id: 0x46 cap: 0xf, Source Output, Sink Output, Source Offload, Sink Offload crtcs: 2 outputs: 5 associated providers: 1 name:modesetting xrandr --setprovideroffloadsink 1 0x7b xrandr --listproviders Providers: number : 3 Provider 0: id: 0x7b cap: 0xf, Source Output, Sink Output, Source Offload, Sink Offload crtcs: 3 outputs: 2 associated providers: 2 name:modesetting Provider 1: id: 0x46 cap: 0xf, Source Output, Sink Output, Source Offload, Sink Offload crtcs: 2 outputs: 5 associated providers: 2 name:modesetting Provider 2: id: 0x46 cap: 0xf, Source Output, Sink Output, Source Offload, Sink Offload crtcs: 2 outputs: 5 associated providers: 2 name:modesetting Not good. The problem is that the provider with id 0x46 now is on both the output_slave_list and the offload_slave_list of the master screen. This commit fixes this by unifying all 3 lists into a single slaves list. Note that this does change the struct _Screen definition, so this is an ABI break. I do not expect any of the drivers to actually use the removed / changed fields so a recompile should suffice. Signed-off-by: Hans de Goede <[email protected]> Reviewed-by: Dave Airlie <[email protected]>
present_restore_screen_pixmap's work doesn't need to be done several times for the same pending flip. Fixes a crash if the X server quits while a flip is pending, in which case present_set_abort_flip may be called several times, including when screen->root is already cleared to NULL. Reviewed-by: Hans de Goede <[email protected]>
From the Present extension specification: An event context is associated with a specific window; using an existing event context with a different window generates a Match error. If eventContext specifies an existing event context, then if eventMask is empty, PresentSelectInput deletes the specified context, otherwise the specified event context is changed to select a different set of events. If eventContext is an unused XID, then if eventMask is empty no operation is performed. Otherwise, a new event context is created selecting the specified events. Without this change, there's no way for a client to explicitly change or destroy an existing event mask entry. Trying to do so as specified above would just result in a protocol error. v2: (Keith Packard) * Use dixLookupResourceByType instead of walking window_priv->events * Return BadMatch if the existing event context is associated with a different window or client * Call LEGAL_NEW_RESOURCE again when creating a new event context * Drop invalid "leak fix" Signed-off-by: Michel Dänzer <[email protected]> Signed-off-by: Keith Packard <[email protected]> Reviewed-by: Keith Packard <[email protected]> Reviewed-by: Kenneth Graunke <[email protected]>
Easier than dealing with it in all paths that can end up here during server shutdown. Signed-off-by: Michel Dänzer <[email protected]> Reviewed-by: Hans de Goede <[email protected]>
…flip We were asserting that these were called before from other places, but that isn't always the case, e.g. during server shutdown. Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=96951 Reported-and-Tested-by: Tod Jackson <[email protected]> Reviewed-by: Hans de Goede <[email protected]> Signed-off-by: Michel Dänzer <[email protected]>
Plug a leak in present_fake_queue_vblank() where the OsTimer would not
be freed.
492,608 (482,816 direct, 9,792 indirect) bytes in 15,088 blocks
are definitely lost in loss record 3,954 of 3,954
at 0x4C2ABDE: malloc (in vgpreload_memcheck-amd64-linux.so)
by 0x586B19: TimerSet (WaitFor.c:433)
by 0x4F1AA9: present_fake_queue_vblank (present_fake.c:108)
by 0x4F15E0: present_pixmap (present.c:954)
by 0x4F23B4: proc_present_pixmap (present_request.c:138)
by 0x552BCE: Dispatch (dispatch.c:430)
by 0x556C22: dix_main (main.c:300)
by 0x6F0D290: (below main) (in /usr/lib/libc-2.24.so)
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=97065
Signed-off-by: Olivier Fourdan <[email protected]>
Reviewed-by: Michel Dänzer <[email protected]>
We are no longer using the present_flip_queue list only for presents which have already been submitted to the driver for page flipping, but also for those which we are queueing up to be flipped later, marked with vblank->queued == TRUE. We were incorrectly calling present_flip_notify for such entries, failing the assertion in present_flip_notify (or presumably resulting in other undesirable behaviour with assertions disabled). Reproduction recipe: Run the JavaFX test case referenced by https://bugs.freedesktop.org/show_bug.cgi?id=98831#c6 and alt-tab out of it while it's fullscreen. May take a few attempts to hit the assertion failure. Fixes: bab0f450a719 ("present: Fix presentation of flips out of order") Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98854 Reviewed-by: Alex Deucher <[email protected]>
This prevents the tearing of moving window in a composite WM desktop when output slave is attached but none of its crtc is really active. Signed-off-by: Qiang Yu <[email protected]> Reviewed-by: Michel Dänzer <[email protected]>
Works fine now. Reviewed-by: Alex Deucher <[email protected]> Signed-off-by: Michel Dänzer <[email protected]>
Signed-off-by: Keith Packard <[email protected]>
Signed-off-by: Keith Packard <[email protected]>
SBC is entire a client-side notion, so remove it from the protocol. No need to have two events with the same content, but we do need to tell which request generated the event so stick a new field in some spare bytes Signed-off-by: Keith Packard <[email protected]>
In PresentRegion, add the PresentNotify list, add explicit CRTC. In PresentRedirectNotify, add CRTC and update_window values Signed-off-by: Keith Packard <[email protected]>
Changes the name of the PresentPixmap request from PresentRegion in anticipation of future additions of non-pixmap sourced updates (YUV images in particular). Adds definitions for all of the new PresentPixmap options. Adds PresentQueryCapabilities to provide applications the ability to learn what the underlying hardware can support. One requirement for any capability is that the X server must do something sensible even if the client behaves as if a capability is supported when it is not. Adds IdleNotify events. As pixmaps can go idle in any order, it's important for applications to know which pixmap to use next. We cannot use fences as the fence itself may not be signaled for some time after the X server has figured out which pixmap to idle. Note that the encoding and header files are not entirely up to date now. Signed-off-by: Keith Packard <[email protected]>
Add presentproto.h updates for current protocol. Finish encoding specification. Signed-off-by: Keith Packard <[email protected]>
If you don't do this then the client libs on 64-bit machines see them as XIDs, which are 64-bit wide (sigh), which ruins the wire encoding and nothing works. Reviewed-by: Jasper St. Pierre <[email protected]> Signed-off-by: Adam Jackson <[email protected]>
Leave it all under #if PRESENT_FUTURE_VERSION for documentation Signed-off-by: Keith Packard <[email protected]>
sunweaver
force-pushed
the
pr/present-ext-backport
branch
from
d679120 to
9d604f9
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
this is currently w-i-p...