Even if the manual redirect didn't result in the surface window we
wanted, it might be different again without the redirect.
I don't know of any specific scenario hitting this though, it's more of
a theoretical safeguard.
Part-of: <https://gitlab.freedesktop.org/xorg/xserver/-/merge_requests/1637>
Now that we won't enable DRI3 if <sys/eventfd.h> is not available, there
is not point in trying to include that header without DRI3.
That allows to build Xwayland with GLAMOR enabled (without DRI3) on
platforms which do not support eventfd.
Signed-off-by: Olivier Fourdan <ofourdan@redhat.com>
Part-of: <https://gitlab.freedesktop.org/xorg/xserver/-/merge_requests/1523>
If the manual redirection was done by xwl_present_maybe_redirect_window,
we are in control.
This allows xwl_present_maybe_redirect_window to work as intended again.
Previously, xwl_window_update_surface_window would ignore the window
redirected by xwl_present_maybe_redirect_window, which would result in
xwl_present_maybe_redirect_window undoing the redirection and bailing.
Part-of: <https://gitlab.freedesktop.org/xorg/xserver/-/merge_requests/1570>
If multiple flips become ready for the same MSC, we would previously
execute them all sequentially, one per MSC for sync flips. This could
result in an unbounded flip queue and corresponding memory consumption.
With implicit sync, leave the mailbox handling to the compositor for
async flips though.
v2:
* Use present_vblank_rec::sync_flip.
Closes: https://gitlab.freedesktop.org/xorg/xserver/-/issues/1664
Fixes: e1f16fb1ac ("xwayland: don't scrap pending present requests")
Part-of: <https://gitlab.freedesktop.org/xorg/xserver/-/merge_requests/1511>
A later commit will use these to (un)redirect the surface window on
demand.
Not used yet, so no functional change intended.
v2:
* Use "surface_window_damage" instead of "surf_win_damage".
(Olivier Fourdan)
* Slightly simplify logic in xwl_unrealize_window.
v3:
* Add comment in xwl_present_maybe_unredirect_window explaining why we
use a timer. (Olivier Fourdan)
v4:
* Rename unredir_timer field to unredirect_timer.
Part-of: <https://gitlab.freedesktop.org/xorg/xserver/-/merge_requests/1300>
It may track a non-toplevel window which fully covers the area of the
window pixmap / Wayland surface. It is now used instead of
xwl_window::toplevel for updating the Wayland surface contents.
The surface_window can now hit the Present page flip path while it's
automatically redirected.
v2:
* Use "surface_window" instead of "surf_win". (Olivier Fourdan)
* Add comment describing surface_window, and describe what
surface_window/toplevel are useful for respectively. (Olivier Fourdan)
* Use surface_window in xwl_realize_window.
v3:
* Backtrack up to the closest opaque ancestor in
xwl_window_update_surface_window. (Olivier Fourdan)
v4:
* Clean up logic for determining the surface window in
xwl_window_update_surface_window, and document it better.
* Handle window_get_damage(xwl_window->surface_window) returning NULL
in xwl_window_update_surface_window.
* Call xwl_window_update_surface_window after xwl_window_buffers_init
in ensure_surface_for_window, since the former may call
xwl_window_buffers_dispose.
* Rename surf/win_pix to surface/window_pixmap in
xwl_window_update_surface_window.
Part-of: <https://gitlab.freedesktop.org/xorg/xserver/-/merge_requests/1300>
It's always the toplevel window, i.e. either the root window or a child
of it.
Preparation for later commits, no functional change.
v2: (Olivier Fourdan)
* Fix debug build.
* Add comment describing ::toplevel.
Part-of: <https://gitlab.freedesktop.org/xorg/xserver/-/merge_requests/1300>
When a present request is received, Xwayland will check if there is an
existing request targeting the same window and msc and scrap the older
request if so. Alas, this does not interact well the older fence-based
or newer syncobj-based synchronization features of the Present
extension.
Since execution of a request may be delayed for an unknown length of
time while waiting for a fence to be signaled, the target msc computed
upon receiving a request may not match the actual msc at which the
request is executed. Therefore, we cannot determine in advance whether a
more recently received request will make an older request redundant.
This change removes the code to scrap pending present requests.
We must also ensure requests are executed in the correct order even if
their fences are signaled out of order. To achieve this, whenever
execution of a request needs to wait for a fence, execution of any
later-received requests will be blocked until the earlier request is
ready. The blocked requests will be added to a list tracked in the
xwl_present_window struct. Once the earlier request's fence is signaled,
any blocked requests will be re-executed.
Signed-off-by: Erik Kurzinger <ekurzinger@nvidia.com>
Part-of: <https://gitlab.freedesktop.org/xorg/xserver/-/merge_requests/967>
This protocol allows for explicit synchronization of GPU operations by
Wayland clients and the compositor. Xwayland can make use of this to
ensure any rendering it initiates has completed before the target image
is accessed by the compositor, without having to rely on kernel-level
implicit synchronization.
Furthermore, for X11 clients that also support explicit synchronization
using the mechanisms exposed in the DRI3 and Present extensions, this
Wayland protocol allows us to simply forward the timeline, acquire, and
release points directly to the compositor, ideally avoiding any
premature stalls in the presentation pipeline.
Signed-off-by: Erik Kurzinger <ekurzinger@nvidia.com>
Part-of: <https://gitlab.freedesktop.org/xorg/xserver/-/merge_requests/967>
Together, DRI3 1.4 and Present 1.4 allow clients to explicitly
synchronize GPU rendering with presentation using DRM syncobjs. Here we
add the necessary support to Xwayland's glamor and Present
infrastructure to enable this functionality.
Signed-off-by: Erik Kurzinger <ekurzinger@nvidia.com>
Part-of: <https://gitlab.freedesktop.org/xorg/xserver/-/merge_requests/967>
If a presentation request is delayed while waiting for a fence, the
original target msc may no longer be correct. Instead, we should compute
a new target msc in xwl_present_re_execute.
Signed-off-by: Erik Kurzinger <ekurzinger@nvidia.com>
Part-of: <https://gitlab.freedesktop.org/xorg/xserver/-/merge_requests/967>
Without either implicit or explicit synchronization, the result of rendering is
pretty much undefined, and many glitches can appear. This still doesn't synchronize
buffer release, but it works around most glitches until explicit sync is supported.
Signed-off-by: Xaver Hugl <xaver.hugl@kde.org>
Part-of: <https://gitlab.freedesktop.org/xorg/xserver/-/merge_requests/967>
After executing a PresentPixmap request using the copy path, Xwayland
will clear the vblank's pixmap field and re-queue it for the next msc so
that on the next frame a PresentCompleteNotify event will be delivered
to the client by present_execute_post.
While this does work, since the pixmap field of the vblank will be NULL
when present_execute_post is called, the mode reported in the event will
always be PresentCompleteModeSkip, even if the request *was* actually
executed with a copy.
To fix this, we introduce a new "copy_executed" flag in the
xwl_present_event struct. If xwl_present_execute sees that this flag is
set, it will fall straight through to present_execute_post like it does
if the window or pixmap is NULL. So, after executing a request with
present_execute_copy, instead of clearing the pixmap field we will set
the copy_executed flag to true. This will cause present_execute_post to
report the correct completion mode to the client when the
PresentCompleteNotify event is delivered on the next frame.
Signed-off-by: Erik Kurzinger <ekurzinger@nvidia.com>
Part-of: <https://gitlab.freedesktop.org/xorg/xserver/-/merge_requests/1254>
This allows e.g.
xfwm4 --vblank=xpresent
to hit the page flip path instead of copies.
In the future, Mesa might also use the Present extension with software
rendering.
This can happen if the window has never completed a Present operation.
Fixes: 4230176080 ("xwayland/present: Embed present_vblank_rec in xwl_present_event")
This will be needed with the next commit: If a child window completely
obscures a toplevel ancestor of different depth, the child window can
use page flipping only if the depth of the presented pixmap matches that
of the window's backing pixmap, or the former may contain pixel values
which are not suitable for the toplevel window's depth.
Nothing should be relying on this anymore, so use a counter like other
places in the tree instead. This ensures that the event_id doesn't get
cast back into a pointer again in future, and also may be slightly less
confusing in cases where calloc reuses an address as debug logs would
show the same event_id for those but now they will be distinct.
Signed-off-by: Jessica Clarke <jrtc27@jrtc27.com>
On traditional 32-bit and 64-bit architectures, uint64_t can be abused
to hold a uintptr_t and be cast back to a valid pointer. However, on
CHERI, and thus Arm's Morello prototype, pointers are capabilities,
which contain a traditional address alongside additional metadata,
including a tag bit that ensures it cannot be forged (the only way to
get a capability with the tag bit set is by using instructions that take
in another valid capability with sufficient bounds/permissions/etc for
the request, and any other operation, like overwriting individual bytes
in memory, will give a capability whose tag is clear). Casting a pointer
to a uintptr_t is fine as uintptr_t is represented as a capability, but
casting to a uint64_t yields just the address, losing the metadata and
tag. Thus, when cast back to a uintptr_t, the capability remains invalid
and faults on any attempt to dereference.
As with various other places in the tree, address this by searching for
the pointer in a list so that we no longer rely on this undefined
behaviour.
Signed-off-by: Jessica Clarke <jrtc27@jrtc27.com>
All these arguments other than damage come from the vblank itself so
passing the vblank simplifies the caller. Moreover, we pass the event_id
solely so we can get back to the event, which is just the (extended)
vblank, so passing the vblank avoids that round trip.
Signed-off-by: Jessica Clarke <jrtc27@jrtc27.com>
By adding a new xwl_present_event_from_vblank function we can avoid
turning the vblank into an event_id, and also abstract away the exact
encoding for event_id from most places.
Signed-off-by: Jessica Clarke <jrtc27@jrtc27.com>
It could happen with the following call path:
frame_callback
xwl_present_frame_callback
xwl_present_msc_bump
xwl_present_execute
xwl_present_flip
xwl_window_create_frame_callback
The nested loop called xwl_present_reset_timer, which may end up calling
xorg_list_del for the entry after the one frame_callback started the
chain for. This resulted in the outer loop never terminating, because
its next element wasn't hooked up to the list anymore.
We avoid this by calling xwl_present_reset_timer as needed in
frame_callback, and bailing from xwl_window_create_frame_callback if it
was called from the former.
We also catch nested calls and FatalError if they ever happen again due
to another bug.
v2:
* Leave xwl_present_reset_timer call in xwl_present_frame_callback,
needed if xwl_present_msc_bump didn't hook up the window to the frame
callback list again.
Closes: https://gitlab.freedesktop.org/xorg/xserver/-/issues/1442
If the dmabuf protocol's feedback object gave us a new list of
modifiers, send PresentCompleteModeSuboptimalCopy to the client
to inform them that they need to call GetSupportedModifiers.
Reviewed-by: Michel Dänzer <mdaenzer@redhat.com>
Even if there's no pending frame callback yet.
Without this, if there was no pending frame callback yet in
xwl_present_queue_vblank, xwl_present_msc_bump would only get called
from xwl_present_timer_callback, resulting in the MSC ticking at ~58
Hertz.
Doing this requires some adjustments elsewhere:
1. xwl_present_reset_timer needs to check for a pending frame callback
as well.
2. xwl_window_create_frame_callback needs to call
xwl_present_reset_timer for all child windows hooked up to
frame_callback_list, to make sure the timer length takes the pending
frame callback into account.
3. xwl_present_flip needs to hook up the window to frame_callback_list
before calling xwl_window_create_frame_callback, for 2. to work.
Closes: https://gitlab.freedesktop.org/xorg/xserver/-/issues/1309
Fixes: 9b31358c52 ("xwayland: Use frame callbacks for Present vblank events")
Reviewed-by: Olivier Fourdan <ofourdan@redhat.com>
Without this, xwl_present_reset_timer would call
xwl_present_timer_callback if the timer was originally armed over a
second ago. xwl_present_timer_callback would call xwl_present_msc_bump,
which could end up hooking up the window to
xwl_window->frame_callback_list again. This would lead to use-after-free
in xwl_present_cleanup:
Invalid write of size 8
at 0x42B65C: __xorg_list_del (list.h:183)
by 0x42B693: xorg_list_del (list.h:204)
by 0x42C041: xwl_present_cleanup (xwayland-present.c:354)
by 0x423669: xwl_destroy_window (xwayland-window.c:770)
by 0x4FDDC5: compDestroyWindow (compwindow.c:620)
by 0x5233FB: damageDestroyWindow (damage.c:1590)
by 0x501C5F: DbeDestroyWindow (dbe.c:1326)
by 0x4EF35B: FreeWindowResources (window.c:1018)
by 0x4EF687: DeleteWindow (window.c:1086)
by 0x4E24B3: doFreeResource (resource.c:885)
by 0x4E2ED7: FreeClientResources (resource.c:1151)
by 0x4ACBA4: CloseDownClient (dispatch.c:3546)
Address 0x12f44980 is 144 bytes inside a block of size 160 free'd
at 0x48470E4: free (vg_replace_malloc.c:872)
by 0x423115: xwl_unrealize_window (xwayland-window.c:621)
by 0x4FCDD8: compUnrealizeWindow (compwindow.c:292)
by 0x4F3F5C: UnrealizeTree (window.c:2805)
by 0x4F424B: UnmapWindow (window.c:2863)
by 0x4EF58C: DeleteWindow (window.c:1075)
by 0x4E24B3: doFreeResource (resource.c:885)
by 0x4E2ED7: FreeClientResources (resource.c:1151)
by 0x4ACBA4: CloseDownClient (dispatch.c:3546)
by 0x5E27EE: ClientReady (connection.c:599)
by 0x5E6CB7: ospoll_wait (ospoll.c:657)
by 0x5DE6CD: WaitForSomething (WaitFor.c:208)
Block was alloc'd at
at 0x4849464: calloc (vg_replace_malloc.c:1328)
by 0x4229CE: ensure_surface_for_window (xwayland-window.c:439)
by 0x4231E8: xwl_window_set_window_pixmap (xwayland-window.c:647)
by 0x5232D6: damageSetWindowPixmap (damage.c:1565)
by 0x4FC7BC: compSetPixmapVisitWindow (compwindow.c:129)
by 0x4EDB3F: TraverseTree (window.c:441)
by 0x4FC851: compSetPixmap (compwindow.c:151)
by 0x4F8C1A: compAllocPixmap (compalloc.c:616)
by 0x4FC938: compCheckRedirect (compwindow.c:174)
by 0x4FCD1D: compRealizeWindow (compwindow.c:274)
by 0x4F36EC: RealizeTree (window.c:2606)
by 0x4F39F5: MapWindow (window.c:2683)
Fixes: 288ec0e046 ("xwayland/present: Run fallback timer callback after more than a second")
Tested-by: Olivier Fourdan <ofourdan@redhat.com>
Reviewed-by: Olivier Fourdan <ofourdan@redhat.com>
If the Wayland compositor doesn't send a pending frame event, e.g.
because the Wayland surface isn't visible anywhere, it could happen that
the timer kept getting pushed back and never fired. This resulted in an
enormous list of pending vblank events, which could take minutes to
process when the frame event finally arrived.
Closes: https://gitlab.freedesktop.org/xorg/xserver/-/issues/1110
Reviewed-by: Olivier Fourdan <ofourdan@redhat.com>
Tested-by: Jaap Buurman <jaapbuurman@gmail.com>
xwl_present_reset_timer checks if the pending flip is synchronous, so
we need to call it after adding the pending flip to the flip queue.
Closes: https://gitlab.freedesktop.org/xorg/xserver/-/issues/1219
Fixes: b2a06e0700 "xwayland/present: Drop sync_flip member of struct xwl_present_window"
Tested-by: Olivier Fourdan <ofourdan@redhat.com>
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
We are handling two cases here: the active flip or the pending flip.
For the pending flip (event->pending == TRUE), we called
xwl_present_release_pixmap.
For the active flip (event->pending == FALSE), we called
xwl_present_release_event. However, xwl_present_flip_notify_vblank
already unhooked event->vblank.event_queue. So this was effectively the
same as calling xwl_present_release_pixmap.
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
Use present_vblank_rec::event_queue instead.
The changes in xwl_present_execute shouldn't really be needed, since
we should never hit queue_vblank in present_execute_wait. But let's be
safe rather than sorry, plus this simplifies the code.
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
Can just call xwl_present_execute directly.
This allows dropping the window member from struct xwl_present_window as
well.
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
We clear the vblank->pixmap field, so next time xwl_present_execute
falls through to present_execute_post.
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
Michel Dänzer
Olivier Fourdan
Erik Kurzinger
Xaver Hugl
Xaver Hugl
Jessica Clarke
Austin Shafer