it used to spawn its own thread and return the main thread
to the binder thread pool -- this was confusing the naming
of things in the kernel.
Bug: 10331839
Change-Id: I2d13a6d73409a38109300fcbe6a04b4c41cb5d00
the main reason for doing this is so that we can have
access to informations about a texture (like its dimension)
close to where we generate and use shaders in ES 2.0.
Previously, there wasn't any way to get to a texture's size
from a RenderEngine implementation.
Bug: 8679321
Change-Id: I388b338a70d07e3e8177dde248710ea1e4c82dff
This change adds an entire field to note whether the timestamp was
auto-generated by Surface or supplied by the application.
The value is used when deciding whether or not to drop frames based
on buffer presentation timestamps. If a desired presentation time
was set explicitly, BufferQueue will use that value to decide if a
frame should be dropped. If the timestamp was generated by Surface
at the time the buffer was queued, the timestamp is ignored.
Bug 10151804
Change-Id: Ibd571a7578351063b813cbdad2ddbeed70655ba5
just ensure the alpha value is 1.0 in the opaque case
when reading the color from the texture or the
global color.
Bug: 8679321
Change-Id: Ia38b30e97c3bce5a2d534a40c0d66e0bfc3ea40d
- turns out fragment shaders don't have default precision by default
- GLES 1.x extensions that became core in GLES 2.0 don't always work
as extensions in GLES 2.0 (!)
Bug: 8679321
Change-Id: I5a4a93e158247910399325a965af5d2e3bbece9b
the Mesh object can be part of each Layer (at least currently).
also reworked the Mesh code a bit to make it easier to access
the vertex data.
Change-Id: I0490851ba898f0aa2e55b62958dcd8bdb535e98b
we need too allow this case so that things like the rotation
animation can work.
with this change we only permit these screenshot if the
destination is SurfaceFlinger itself.
Bug: 10235036
Change-Id: I66fea5391e52b0d7f17f25827572b236f2d9eb71
This prevents slow/malicious virtual display consumers from
backpressuring SurfaceFlinger and preventing it from updating other
displays.
Bug: 10193714
Change-Id: I3e877d97202628d2d9abea24b66576f38299c14d
this means they only have access to the consumer end of
the interface. we had a lot of code that assumed consumers
where holding a BufferQueue (i.e.: both ends), so most of
this change is untangling in fix that
Bug: 9265647
Change-Id: Ic2e2596ee14c7535f51bf26d9a897a0fc036d22c
While currently untested, this should allow to move the
BuffereQueue in the consumer process and have everything
work as usual.
Bug: 9265647
Change-Id: I9ca8f099f7c65b9a27b7e7a3643b46d1b58eacfc
A recent change to screencapture to have the call to
IGraphicBufferProducer happen on the incoming binder
thread didn't set the result so the result was always
returned as NO_ERROR.
This made screencap fail on some devices (e.g. Wolfie)
which relies on some kind of fallback mechanism to
generate the screencap but the fallback mechanism
doesn't get triggered because the error isn't returned.
Bug: 9989385
Change-Id: I2aee91ea1034869fcbb0f49b9a0087c3cff43bbe
Signed-off-by: Mike J. Chen <mjchen@google.com>
we can now queue/dequeue a buffer in asynchrnous mode by using the
async parameter to these calls. async mode is only specified
with those calls (it is not modal anymore).
as a consequence it can only be specified when the buffer count
is not overidden, as error is returned otherwise.
Change-Id: Ic63f4f96f671cb9d65c4cecbcc192615e09a8b6b
this is the first step of a series of improvements to
BufferQueue. A few things happen in this change:
- setSynchronousMode() goes away as well as the SynchronousModeAllowed flag
- BufferQueue now defaults to (what used to be) synchronous mode
- a new "controlled by app" flag is passed when creating consumers and producers
those flags are used to put the BufferQueue in a mode where it
will never block if both flags are set. This is achieved by:
- returning an error from dequeueBuffer() if it would block
- making sure a buffer is always available by replacing
the previous buffer with the new one in queueBuffer()
(note: this is similar to what asynchrnous mode used to be)
Note: in this change EGL's swap-interval 0 is broken; this will be
fixed in another change.
Change-Id: I691f9507d6e2e158287e3039f2a79a4d4434211d
The previous implementation assumed that the HWC could read and write
the same buffer on frames that involved both GLES and HWC composition.
It turns out some hardware can't do this. The new implementation
maintains a scratch buffer pool to use on these mixed frames, but on
GLES-only or HWC-only frames still does composition directly into the
output buffer.
Bug: 8384764
Change-Id: I7a3addb34fad9bfcbdabbb8b635083e10223df69
Vsync array size is specified as HWC_DISPLAY_TYPES_SUPPORTED whose
value luckily happens to be 2. That enum is actually used for querying
hwc for the number of displays supported.
The implementation file EventThread.cpp correctly accesses the array
upto HWC_NUM_DISPLAY_TYPES
Change-Id: I36e3f0913e7d6fda7bbf4449c1fb32c7f18bb934
Signed-off-by: Amara Venkata Mastan Manoj Kumar<manojavm@codeaurora.org>
Signed-off-by: Saurabh Shah <saurshah@codeaurora.org>
When acquiring a buffer, SurfaceFlinger now computes the expected
presentation time and passes it to the BufferQueue acquireBuffer()
method. If it's not yet time to display the buffer, acquireBuffer()
returns PRESENT_LATER instead of a buffer.
The current implementation of the expected-present-time computation
uses approximations and guesswork.
Bug 7900302
Change-Id: If9345611c5983a11a811935aaf27d6388a5036f1
When a buffer had a crop (meaning its content is scaled to the window size)
and a window crop was defined, the resulting crop couldn't be expressed
properly because h/w composer's API was limited to integers, since
this is fixed in h/w composer 1.3, we take adventage of this to
make sure we get the correct crop.
this bug could result in the buffer being scaled by an incorrect ratio and
be slightly offset; moreover, it would produce different results from the
GL code path, which is always correct.
Change-Id: I8e20e00b6e26177d14f4ab4d2cd581e26c818892
because surfaceflinger handles screenshot in a different
thread from the binder thread that requested it and because
the IGraphicBufferProducer is a synchronous interface
calling back into the system server; it is possible for
the latter to run out of binder threads (b/c it holds
a lock while calling into SF).
The solution is to make sure all calls on IGraphicBufferProducer
happen on the incoming binder thread. We achieve this by creating
a IGBP wrapper which is given to the screenshot code.
Bug: 8734824
Change-Id: I2be85660d9dc65d239d68f6d3ab3c973c13b34cc
this is the first step to add support for GLES 2.x, this
change breaks the dependency of SF on GLES 1.x by moving
all operation into their own class.
Bug: 8679321
Change-Id: I0d2741eca2cefe67dfd9cf837cac10c4d126928b
Allow releasing a buffer acquired in the past.
Change-Id: I3f27f1caec109c1a47286fc4d2235c6e1afe9022
Signed-off-by: Lajos Molnar <lajos@google.com>
Bug: 9516405
- most methods on Layer didn't need to be virtual
- more consistency in naming drawing/current state
Change-Id: Ieb7b4951b40fc673b807994ed21ae4aea7281068
generally the last reference to a Layer is released in commitTransaction()
with mStateLock held. Layer itself only holds weak references to Client,
however, ~Layer() briefly promotes this weak reference -- during that time
the all other strong references to that Client go away, ~Layer is left with
the last one... then hell breaks loose as ~Client is called, which in turn
needs to acquire mStateLock.
We fix this by holding a temporary copy of the drawing state during
the transaction so that the side-effects of copying the current
state into the drawing state are seen only after mStateLock has
been released.
Bug: 9106453
Change-Id: Ic5348ac12283500ead87286a37565e8da35f1db2
To keep the code readable now that we have four different texenv
configurations, this change separates the decisions about what
configuration to use from the GL calls to set up the configuration.
Bug: 8963244
Change-Id: Ia07a306a7809ba8f93493d0160ccbd509e948581
Instead of representing the buffer-queue as a vector of buffer
indices, represent them as a vector of BufferItems (copies).
This allows modifying the buffer slots independent of the queued
buffers.
As part of this change, BufferSlot properties that are only
been relevant in the buffer-queue have been removed.
Also, invalid scalingMode in queueBuffer now returns an error.
ConsumerBase has also changed to allow reuse of the same
buffer slots by different buffers.
Change-Id: If2a698fa142b67c69ad41b8eaca6e127eb3ef75b
Signed-off-by: Lajos Molnar <lajos@google.com>
Related-to-bug: 7093648
- timeout is now 3 seconds instead of 1
- simplifies the API a bit
- allows us to change/tweak this timeout globaly
Bug: 8988871
Change-Id: I8d3c6ec43a372f602fb3f29856710339f86c0ec9
If an invalid display device is detected during surface composition
(e.g., a simulated secondary display is removed), abort the
composition.
Change-Id: Ia6afb2e287882d8ae0614eb25463d3f85b687adf
until now it was only used to discard a layer entirely.
we're now reducing the size of the layer if it is still
visible, if possible.
this works for instance when a surfaceView is used and
only the menu bar is displayed over it.
Change-Id: I3f5527c5cd1e69ecc968272c8948f1513ada8c55
this allows us to render into a buffer with a
pixelformat of our own choice; this is much faster
on all platform.
Bug: 8582615
Change-Id: I61298fc8e43fa6f92044c5123955cb5c7897dab7
colorize a bit the output of dumpsys SurfaceFlinger to
make it easier to read. Right now it will bold the
title of each section and use green for the name of
each layer.
Change-Id: I0d9f18d115401cb45109d244ef3a278481f68cc6
this mimics the code we used for the size. basically we need to
also update the "current state" because it gets copied to the
"drawing state" when a transaction occurs, and it would
"undo" our change.
Bug: 8511430
Bug: 8581533
Change-Id: I08c02abbf21b7f168f7124cd14ee717d7d3d502c
this affects devices that need a glReadPixels(). We use
a FBO instead of a GlConsumer as an intermediate render target, this
saves 2 calls to eglMakeCurrent().
On Galaxy Nexus this allows us to go from ~135ms to ~35ms for
recent's screenshots.
Bug: 8582615
Change-Id: I6b25291ecc235f1927579bbb2db3c731e985c6e8
When the screen is turned off, the current stack is set to -1. This causes
logic in iSurfaceFlinger's handleTransactionLocked() function to fail to
match the current stack, and the latest orientation is not set into the
layer. This causes BufferQueue, later, to potentially set an obsolete
transformHint on a created surface (such as in the case with ImageWallpaper's
Egl surface, in the bug below).
The fix is to note this situation and use a default value for the DisplayDevice,
which should have the current orientation information.
Issue #8508397 ImageWallpaper sometimes rendered in wrong orientation causing a ~30-40% drop in graphics performance
Change-Id: Ibae15d73b289a8343c67f4f6bb77fdf11dd95ee7
SF transactions were always handled on VSYNC which allowed
the screenshot to sneak-in between closing the transaction
and vsync (before it's latched), resulting in a screenshot
with the previous state.
we now always force transactions to happen immediately
before screenhots.
Bug: 7552304
Change-Id: I0afc86b7e8366173daff5b9988bbb4d2a0f43860
HWComposer queries the HWC for dimensions of physical displays, but
can't do that for virtual displays. The dimensions are used to set the
display frame of the framebuffer target layer passed to HWC, and
implicitly the dimensions of the virtual display.
Bug: 8316155
Change-Id: I9cbd2530d2fa878f86128a1472def520b5d694a5
since the transparent region hint really depends on the
content of the window containing the SurfaceView
(it's calculated by the view hierarchy based on
overlapping views), it makes sense to latch it only when
the content of the window (the app) changes.
This should help fixing drawing artifacts when changing the
layout of a window containing a SurfaceView.
Bug: 8511430
Change-Id: Ic3aa668495293cb4d82a2cd7dcf3b6a337287678
We now detect at runtime which sync features to use, which
allows us to remove a lot of the compile-time configuration
options. There is still one option though, to disable
KHR_fence_sync on some devices (which are more efficient
without it).
- added a backdoor to get the vendor's EGL strings
the new logic is:
- use always ANDROID_native_fence_sync if available
- fallback to KHR_fence_sync if available and not disabled
by the compile-time option
- use KHR_wait_sync if available and either of the above is
enabled
Change-Id: I9c4b49d9ff1151faf902cc93bd53ea5f205aaabf
DisplayDevices can be released when DisplayManager removes them from
the display list, or (for virtual displays) when the surface is set to
NULL. We were only cleaning up HWC resources associated with the
display in the first case.
Bug: 8384764
Change-Id: Id3d226dd7178fbe6d0a2ac4e2660b864ee073de3
If we're using a HWC that doesn't support virtual displays, or we have
more virtual displays than HWC supports concurrently, the
VirtualDisplaySurface should simply be a passthrough from source
(GLES) to sink.
This change also tries to distinguish between display types and HWC
display IDs a little better, though there's more to do here. Probably
needs a higher-level rethink; it's too error-prone now.
Bug: 8446838
Change-Id: I708d2cf262ec30177042304f174ca5b8da701df1
HWComposer didn't allow the virtual display output buffer to be set
directly, instead it always used the framebuffer target buffer.
DisplayDevice was only providing the framebuffer release fence to
DisplaySurfaces after a commit.
This change fixes both of these, so both HWComposer and DisplayDevice
should continue to work if VirtualDisplaySurface changes to use
separate framebuffer and output buffers. It's also more correct since
VirtualDisplaySurface uses the correct release fence when queueing the
buffer to the sink.
Bug: 8384764
Change-Id: I95c71e8d4f67705e23f122259ec8dd5dbce70dcf
some drivers don't support this yet, so we use a system
property to enable the glReadPixels "workaround" for them:
ro.bq.gpu_to_cpu_unsupported=1
Change-Id: I74d6a3a8f0cee8d5a507b72c760cf247e39195e0
Previously we only queued a virtual display buffer to the sink when
the next frame was about to be displayed. This may delay the "last"
frame of an animation indefinitely. Now we queue the buffer as soon as
HWC set() returns and gives us the release fence.
Bug: 8384764
Change-Id: I3844a188e0f6ef6ff28f3e11477cfa063a924b1a
We're not using IMemoryHeap as a transport anymore,
instead we're providing a CpuConsumer and use the
IGraphicBufferProducer version of the screenshot API.
However, some GPU drivers don't support properly
a GPU to CPU path, to work around this, we use a
temporary BufferQueue on the server side for the
GL rendering, and we use glReadPixels into the
CpuConsumer (we're now using a CPU to CPU path
which is always supported).
Currently this "wrapping" is always performed,
but it can be bypassed on devices that support
the GPU to CPU path.
This also addresses a DoS attack vector on
SurfaceFlinger, where an application could
consume all of SF's filedescriptors by creating
a lot of screenshots in a row.
Bug: 8390553
Change-Id: I9e81514c2a7711b9bb393f74305be7d2abe08f1c
* changes:
Add BufferQueueInterposer and use it for virtual displays
Add DisplaySurface abstraction
Fix argument types in IGraphicBufferProducer methods
Minor cleanups/fixes before virtual display refactoring
BufferQueueInterposer allows a client to tap into a
IGraphicBufferProducer-based buffer queue, and modify buffers as they
pass from producer to consumer. VirtualDisplaySurface uses this to
layer HWC composition on top of GLES composition before passing the
buffer to the virtual display consumer.
Bug: 8384764
Change-Id: I61ae54f3d90de6a35f4f02bb5e64e7cc88e1cb83
DisplayDevice now has a DisplaySurface instead of using
FramebufferSurface directly. FramebufferSurface implements
DisplaySurface, and so does the new VirtualDisplaySurface class.
DisplayDevice now always has a surface, not just for virtual displays.
In this change VirtualDisplaySurface is just a stub; buffers still go
directly from GLES to the final consumer.
Bug: 8384764
Change-Id: I57cb668edbc6c37bfebda90b9222d435bf589f37
None of these should change behavior, except for removing some
incorrect log messages when using a virtual display.
- HWComposer::getAndResetReleaseFenceFd() checks the HWC version, so
no need to do that in the DisplayDevice::onSwapBuffersCompleted().
However, it should check that mFramebufferSurface is not NULL like
it is for virtual displays.
- Comment that FramebufferSurface::dump() overrides the non-virtual
ConsumerBase::dump(), and fix it so the right thing happens
regardless of the static type of the pointer/reference the callee
has. FramebufferSurface::dump() could be removed right now, but I'd
need to bring it back in a later change.
- Use the right enum for validating display type ids.
- Don't try to send hotplug events for virtual displays.
- Mark virtual displays as connected so HWComposer::prepare() doesn't
think something is wrong when it gets a non-NULL layer list.
- Remove unused FramebufferSurface methods.
Bug: 8384764
Change-Id: Id28a2f9be86b45f4bb7915fdf7752157035f4294
the recent screenshot rework allowed the older screenshot
interface to work without that permission
Change-Id: I6c4743f4591c81106e3b823d55a055f7b4907de1
the purgatory list wasn't needed anymore; in fact it had no effect as
buffer life-time management is now handled by the BufferQueue.
For QueuesToWindowComposer we keep a list of wp<> on the IBinder
for IGraphicBufferProducers we hand over to clients so we can
easily check if an IGraphicBufferProducer is ours. We clean-up the
list when our IGraphicBufferProducer are destroyed.
Bug: 8349142
Change-Id: I1aa06652ade8c72d0004a3f5e6c3d6e8a82fc2ae
since we're using glReadPixels(), we only need to use
the width (as opposed to the stride) of the source
screenshot.
Bug: 8374664
Change-Id: I145c80f4fff5444df7c77c4f52e70a7203caddbd
ISurface was only used to get the IGraphicBufferProducer from
a Layer. It's now replaced by a BBinder subclass / IBinder and
is only used as a handle to the surface, to both refer to it
and manage its life-time.
Also cleaned-up a bit the ISurfaceComposer interface and
"create layer" code path.
Change-Id: I68d0e02d57b862cffb31d5168c3bc10cea0906eb
We were using the "visible layer list" when taking screenshots,
which doesn't work when a layer is behind other opaque layers
and therefore hidden.
We fix this by using the full layer list, filtered by the
layerstack of the display we're looking at.
Bug: 7552304
Change-Id: I4b6f77e5511aea94f8d218975b6e22738e7e5d5b
older drivers which are doing implicit synchronization need this
or they could deadlock.
Bug: 8341885
Change-Id: Icd980a6be16071678d6151e34725b3c1c547d7ee
* changes:
Get rid of LayerBase.
Make LayerDim a regular Layer instead of a LayerBase
fold LayerBaseClient into LayerBase
Remove support for ScreenshotLayer
The functionality of LayerBase and Layer is folded
into Layer. There wasn't a need for this abstraction
anymore.
Change-Id: I66511c08cc3d89009ba4deabf47e26cd4cfeaefb
- SurfaceFlinger now supports to take a screenshot
directly into an IGraphicBufferProducer
- reimplement the IMemoryHeap screenshot on top
of the above
- reimplement LayerScreenshot such that its
BufferQueue is directly used as the destination
of the screenshot. LayerScreenshot is now a thin
wrapper around Layer
Bug: 6940974
Change-Id: I69a2096b44b91acbb99eba16f83a9c78d94e0d10
When a display is added, initialize it to use an empty layer stack, so
if it is somehow visible it will show black. It will be assigned the
real layer stack -- along with a projection and other properties -- by
window manager soon. Normally a display remains blanked until window
manager has decided what to show on it, but for HDMI connected at boot
that isn't currently the case.
Bug: 7258935
Change-Id: Ic9bb25f7a9b8d9d3772b097ab1d6fa03bc8780a1
- cropping to the projection's "viewport" is "simply"
accomplished by intersecting it with the window crop
expressed in layerstack space.
Bug: 7149437
Change-Id: I0e90b3f37945292314b5d78a8f134935967e8053
- the crop region is now always calculated and set
in LayerBase::setGeometry which uses new virtuals to
access the "content" crop and transform (which are
provided by the Layer subclass)
Change-Id: Ib7769bdec0917dd248f926600c14ddf9ea84897a
each desplay's projection's viewport essentially clips each layer,
so this should be reflected in the visibleregion passed to
h/w composer.
DisplayDevice getViewport and getFrame are now guaranteed to return
valid Rects.
Change-Id: I4c25f34fb26af10179eb26d429ca6c384c671e91
The crop always had left=top=0, because the crop position
and the layer's transform were merged together in
computeBounds() (which really used to compute the
bounds in screen space, which we usually call the
"frame" elsewhere in the code)
Note: in practice this crop value is not used by
hwc, because it's overridden in Layer::setGeometry(), which
is why this bug was never apparent.
Change-Id: I1ec6400a8fc8314408e4252708f43ea98c2fe64e
This change replaces checks for a NULL fence pointer with calls to
Fence::isValid. There should no longer be NULL fences.
Change-Id: If17c9c132fcb1801531bf7588f8ba53476c57dad
This change eliminates the uses of a NULL sp<Fence> indicating that no waiting
is required. Instead we use a non-NULL but invalid Fence object for which the
wait methods will return immediately.
Bug: 7892871
Change-Id: I5360aebe3090422ef6920d56c99fc4eedc642e48
This change makes the 'dumpsys SurfaceFlinger --latency' command with no extra
args dump the frame timestamp data for the most recent frames that
SurfaceFlinger generated that included window animation transaction changes.
Change-Id: I8bded1ea08a4cddefef0aa955401052bb9107c90
