The smart scheduler is designed to minimize scheduler overhead by
increasing the interval between WaitForSomething calls when a single
client is running. However, the software rotation code depends on
its BlockHandler being invoked for screen updates; the long delays
caused by the smart scheduler optimizations means that screen updates
can be delayed a long time as well.
The change is simple -- prevent the smart scheduler from increasing
the scheduling interval while any screen is using software rotation.
(cherry picked from commit e7dd1efef4)
Signed-off-by: Keith Packard <keithp@keithp.com>
The rotation block handler uses regular driver rendering functions to
repaint the screen, if those functions queue commands in the driver,
it's important that the driver block handler be invoked after the
rotated image is drawn.
Signed-off-by: Keith Packard <keithp@keithp.com>
(cherry picked from commit 1740cda7a3)
Remove this now that clearing is done by repainting with appropriate extend
modes.
Signed-off-by: Keith Packard <keithp@keithp.com>
(cherry picked from commit 60a4f1368785d26a49a3ef6df829723ca154c154)
Signed-off-by: Keith Packard <keithp@keithp.com>
When the crtc transformation changes, the entire crtc must be repainted.
This was being done by clearing the shadow and then painting the rectangle
containing the screen image; the clear being required as the screen image
may not fill the crtc. When changing the transform rapidly, this leads to
flashing. Eliminate the clear by painting the entire crtc instead of just
the screen rectangle.
(cherry picked fom commit 5394b7e662)
Signed-off-by: Keith Packard <keithp@keithp.com>
When the shadow scanout buffer can be re-used, the underlying framebuffer
area must be damaged so that the scanout will be repainted. This patch
delays the addition of that damaged area until after the transform in the
crtc has been updated, otherwise the old transform would have been used and
the wrong area repainted.
Signed-off-by: Keith Packard <keithp@keithp.com>
(cherry picked from commit 763df9eec7)
Signed-off-by: Keith Packard <keithp@keithp.com>
Previously it is possible that creating rotation data, then cleaning
up and creating again so that pScreen->BlockHandler and
xf86_config->BlockHandler all point to xf86RotateBlockHandler.
See bug #19343.
(cherry picked from commit 5f3188228e)
When a driver uses a crtc during device detection, the scrn has not yet been
configured and virtualX/virtualY are still zero. This caused the X server
to try and allocate a shadow frame buffer, which couldn't work.
Detect this by checking for zero virtualX/virtualY values.
Signed-off-by: Keith Packard <keithp@keithp.com>
(cherry picked from commit 063eb6743c)
pScreen->width/height are not initialized when doing initial mode setting,
which makes this function incorrectly fail. Using scrn->virtualX should work
in all cases though.
Bug 19017 reports a crash in xf86CrtcSetModeTransform when doing a modeset
for output probing, long before the screen array is initialized; that was
caused by a work-around to set pScreen->width/height so that xf86CrtcFitsScreen
could find the right values.
Signed-off-by: Keith Packard <keithp@keithp.com>
(cherry picked from commit fde2f96103)
The shadow frame buffer and other data used for rotation need to be freed
when the crtc is disabled, not just when rotation is disabled.
Signed-off-by: Keith Packard <keithp@keithp.com>
(cherry picked from commit 1ba4cbb159)
- pScreen->width and height were zero, so it didn't "fit" the screen.
(cherry picked from commit ffb484f7ef)
Signed-off-by: Keith Packard <keithp@keithp.com>
pixman 0.13.2 now holds all of the matrix operations. This leaves
the protocol conversion routines and some ABI stubs in place
Signed-off-by: Keith Packard <keithp@keithp.com>
Instead of using a separate function to notify DIX about transform changes,
add the transform to RRCrtcNotify so that the whole Crtc state changes
atomically.
RandR matrix computations lose too much precision in fixed point;
computations using the inverted matrix can be as much as 10 pixels off.
Convert them to double precision values and pass those around. These API
changes are fairly heavyweight; the official Render interface remains fixed
point, so the fixed point matrix comes along for the ride everywhere.
New RRCrtcGetTransform function in DIX that DDX can use to get the pending
transform. The DDX code should be complete; the DIX code is just a stub at
this point.
- Redo damage naming for more consistency.
- Call post submission functions only where appropriate.
- EXA can now live without it's odd damage workarounds.
LeaveVT/EnterVT cycles will free/realloc shadow frame buffers. Because of
this, the presense/absence of that data is insufficient to know whether
the screen function wrappers are necessary. Instead, the 'transform_in_use'
flag should be used.
This patch also adds 'xf86RotateFreeShadow' for drivers to use at LeaveVT
time to free the rotation data; it will be reallocated on EnterVT.
xf86CrtcRotate() is called by randr 1.2 drivers via xf86CrtcSetMode() or xf86SetDesiredModes()
during ScreenInit() at which point pScrn->pScreen is not set. If a user specifies a rotation
in their config file pScrn->pScreen is dereferenced and boom.
DRI uses a non-screen block/wakeup handler which will be executed after the
screen block handler finishes. To ensure that the rotation block handler is
executed under the DRI lock, dynamically wrap the screen block handler for
rotation.
SourceValidate is used exclusively by the software cursor code to pull the
cursor off of the screen before using the screen as a source operand. This
eliminates the software cursor from the frame buffer while painting the
rotated image though. Disabling this function by temporarily setting the
screen function pointer to NULL causes the cursor image to be captured.
(cherry picked from commit 05e1c45ade)
Setting a mode on an unrotated CRTC was causing all of the rotation updates
to be disabled; the loop looking for active rotation wasn't actually looking
at each crtc, it was looking at the modified crtc many times.
(cherry picked from commit 8b217dee3a)
The rotation state is stored in the xf86_config structure which is not
re-initialized at server reset time. Clean it up at CloseScreen time.
(cherry picked from commit f8db7665dc)
Rotation block handler was re-registering the rotation damage structure,
creating an infinite loop in the damage code. Track registration of the
damage structure to avoid this.
(cherry picked from commit b14f003b0e)
Box transformation from source to dest area was broken, leaving the wrong
areas painted when the crtc origin was non-zero.
When rotating from left to right, the pixmap doesn't get reallocated, and so
no damage was left in the pixmap from xf86RotatePrepare. Separately damage
the whole crtc area when this occurs to repaint the area.
(cherry picked from commit 2a50ca2160)
Add monitor "Rotate" option taking one of "normal", "left", "inverted" or
"right". However, because initial mode selection is made before the screen
is completely initialized, we cannot create the shadow pixmap object at this
point. Pend the shadow pixmap creation until the block handler.
Note that this code is not completely functional yet.
This code comes from the intel driver, so there's no history in this tree.
As the crtc/output-based mode selection code uses ddc, the ddc and i2c
modules have been merged into the server. Attempts to load them are safely
ignored now.
