both API and implementation will change, this is just a prototype
intended to show feasability.
SurfaceFlinger is passed an ISurfaceTexture through a new
callback, it is in turn used to create an EGLSurface which
surfaceflinger will draw into in addition to the main screen.
Change-Id: Id0bbb0b854bb7bae44d57246a90b65d4567f9a21
After a HWC set, each SurfaceFlinger Layer retrieves the release fence
HWC returned and gives it to the layer's SurfaceTexture. The
SurfaceTexture accumulates the fences into a merged fence until the
next updateTexImage, then passes the merged fence to the BufferQueue
in releaseBuffer.
In a follow-on change, BufferQueue will return the fence along with
the buffer slot in dequeueBuffer. For now, dequeueBuffer waits for the
fence to signal before returning.
The releaseFence default value for BufferQueue::releaseBuffer() is
temporary to avoid transient build breaks with a multi-project
checkin. It'll disappear in the next change.
Change-Id: Iaa9a0d5775235585d9cbf453d3a64623d08013d9
The desc.txt file can now mark parts as 'must finish cleanly' by using
'c' as the part line prefix rather than 'p'. If so indicated, if the
bootanimation is asked to quit it will do so only after waiting to
finish that part.
I considered either making init.c service killing smarter or promoting
bootanim to be a bindable service with a requestExit method. However,
these changes are probably too big/risky given our ship date. So
I used a property as a mailbox between SurfaceFlinger and bootanim.
Bug: 6679877
Change-Id: Id7dca22caa50b450fff25ca94f7242d971034f41
HWComposer must abstract the HWC HAL entirely, so that the
HAL can continue to evolve (and break binary compatibility)
without breaking SurfaceFlinger. The HWC data structure had
leaked outside of HWComposer, this is now fixed.
We now have an abstract interface that provide all the
needed functionality, HWCompose provides concrete
implementations of it based on the the HWC version.
Change-Id: I40c4676dc986b682ede5520a1c60efe64037b0bb
SurfaceFlinger will no longer directly synchronize with early suspend.
Instead, PowerManagerService will synchronize with SurfaceFlinger to
ensure that a black frame has been drawn on the display, and then
trigger all early suspend handlers.
Change-Id: I07acdd628440d23fdb69db94319ec5d65d3f4919
This change adds a crop rectangle specified in window coordinates to the layer
state. The all window pixels outside this crop rectangle are treated as though
they were fully transparent. This change also adds the plumbing necessary for
WindowManager to set that crop.
Change-Id: I582bc445dc8c97d4c943d4db8d582a6ef5a66081
this seems to hurt performance on some GPU. this change
might negatively affect performance on other GPUs though, but
probably in less time-sensitive cases. If this becomes a
problem it might become necessary to pre-clip the geometry
(so that we don't have to use glScissor).
This improves the rotation animation quite a bit.
Change-Id: I5dbe1286f7ad858ef2c1e1ad9a07ee3f26c0b1f3
this optimization didn't improve performance and in fact
seemed to hurt more than anything else. it also made
things a lot more complex as it introduced edges cases
when switching to/from h/w composer.
Change-Id: Iaafc235e175f5740cd98bff914d706e02ab88bb8
surfaceflinger will now handle each surface post
as soon as possible and handle the composition
itself at VSYNC time as usual.
Change-Id: I6b1ae33fd56062d86e5419ebab8def0ca5803fbf
this prevents the GPU from running when composition
is fully handled by overlays. this should improve
animations a bit.
Change-Id: If4ae584b7a3976e6bdd36e318686ac3940b6b075
re-add support for pixelformats L_8, LA_88 and RGB_332 in libui
for backward compatibility.
This may or may not fix 6058926
Bug: 6049685
Change-Id: Ic1b8b4cc994522f7fe664da64c0ef76b98bc6d53
SF could end-up in an infinite crash-loop during startup if it
was stopped while the screen was off. This happened because
the thread that manages screen blanking was started before
other important pieces of SF were initialized.
Change-Id: I0dded11dbf2395fdd57b673859a7aa0fa9eb32b6
we're seeing UI freezes when window updates and
composition are separated. for now we workaround this
by always doing a composition after window updates on
vsync. triple buffering is reenabled for performance.
Change-Id: I693d705000b7452489bb0b4918fbeadb9879315c
In some situations SF would mark a window as "has a pending update"
but would never process that update because the window is not
visible (fully transparent, hidden by another window, etc...), this
window would then be "stuck" until some other window updated.
Change-Id: Ifa18a9aef3a53f2593b473556702688ae62d9503
recent changes in SF introduced a hang where some windows would
stop being refreshed.
This is an attemp to fix that.
Change-Id: I6aa32ac0d6f1c0a6aea8f6195825dc4f4e6f93f9
with this changes, SF transactions are handled as soon as possible
but do not trigger updates. the update is delayed until the next
vsync.
this allows us to work much better without requiring triple-buffering.
Change-Id: I1fa10794d0cf742129f0877698b7b1e1f2ec7401
--latency-clear [name]
clears the latency data for the specified layer or for
all layers if none is specified
--list
prints the list of all layers regardless of their visibility
Change-Id: I7c07ae020f838c173b98ee50f3fb3e93da78acbb
It is now possible to say:
dumpsys SurfaceFlinger --latency
to print latency information about all windows
dumpsys SurfaceFlinger --latency window-name
to print the latency stats of the specified window
for instance: dumpsys SurfaceFlinger --latency SurfaceView
The data consists of one line containing global stats, followed by
128 lines of tab separated timestamps in nanosecond.
The first line currently contains the refresh period in nanosecond.
Each 128 following line contains 3 timestamps, of respectively
the app draw time, the vsync timestamp just prior the call to set and
the timestamp of the call to set.
Change-Id: Ib6b6da1d7e2e6ba49c282bdbc0b56a7dc203343a
Some implementations of NPOT, particular those derived from core
GLES2, require the wrap mode to be CLAMP_TO_EDGE. Set the required
wrap mode for the screenshot texture so it passes the completeness
check.
Change-Id: I735016123e4acaf54b40d1435bd70281cef88a31
Several source files privately defined macros LIKELY and UNLIKELY in terms
of __builtin_expect. But <cutils/compiler.h> already has CC_LIKELY and
CC_UNLIKELY which are intended for this purpose. So rename the private
uses to use the standard names.
In addition, AudioFlinger was relying on the macro expanding to extra ( ).
Change-Id: I2494e087a0c0cac0ac998335f5e9c8ad02955873
- add the ability to set the vsync delivery rate, when the rate is
set to N>1 (ie: receive every N vsync), SF process' is woken up for
all of vsync, but clients only see the every N events.
- add the concept of one-shot vsync events, with a call-back
to request the next one. currently the call-back is a binder IPC.
Change-Id: I09f71df0b0ba0d88ed997645e2e2497d553c9a1b
Layer::lockPageFlip() and layer::onRemove() could be called on
different threads and race such that lockPageFlip() successfully
called mSurfaceTexture->updateTexImage() but then gets NULL back from
mSurfaceTexture->getCurrentBuffer(), leading to a crash.
This change moves Layer::onRemove() calls to
SurfaceFlinger::commitTransaction() so they happen after the Layer is
done being drawn from and only happen on the main surfaceflinger
thread.
Change-Id: I4b550caadff4cc1878d7c3bca6129193fb0c713e
use gui/DisplayEvent to receive the events. Events are
dispatched through a unix pipe, so the API is compatible
with utils/Looper. see gui/DisplayEvent.h for more info.
Bug: 1475048
Change-Id: Ia720f64d1b950328b47b22c6a86042e481d35f09
Revert "Add support for sending VSYNC events to the framework"
This reverts commit f3918c5bd4bc9f02f74da42995564150ca2dd382.
Change-Id: I998e3e1aa3fa310829ae973b64fe11b01f6f468f
use gui/DisplayEvent to receive the events. Events are
dispatched through a unix pipe, so the API is compatible
with utils/Looper. see gui/DisplayEvent.h for more info.
Bug: 1475048
Change-Id: If4126023fc9c067e56087ec7d16a8fd542ce1794
This fixes the issue:
"Call not sent" dialog takes too long to dismiss after hitting OK"
Note: the system would recover after a 5 second timeout.
Bug: 5534520
Change-Id: Ifa37e594b50581f498479a5858672441b3d7dd87
The ScreenShot layer is now created hidden. The screenshot itself
is aquired during the transaction when the layer is made visible.
This guarantees the screenshot and the layer happen atomically
with respect to screen updates.
Bug: 5534521
Change-Id: Ida23e1f13d5716ec83b78a15712e0646d6cf8729
We now have mInvalidateRegion which holds the region to invalidate, it
can be set from any thread as long as mInvalidateLock is held. We use
fine-grained locking here because mInvalidateRegion can be set from anywhere,
in particular frmo HWC callbacks.
Bug: 5466774
Change-Id: Iafca20aa3f5b25a87755e65bde7b769aa8f997bc
A LayerScreenshot is a special type of layer that contains a screenshot of
the screen acquired when its created. It works just like LayerDim.
Make sure to call compositionComplete() after rendering into a FBO.
Bug: 5446982, 5467587, 5466259
Change-Id: I5d8a1b4c327f9973d950cd4f4c0bca7f62825cd4
there was situations where SF's main loop would run (as if there was
an invalidate), but the dirty region was empty (so no new buffers
were retired). In this case we return early and don't swap, which
would cause drawing artifacts.
Bug: 5476838
Change-Id: Id3b7bf4b7aabec7919c50d9278eb2165973a4c3d
when taking a screenshot, in particular, we could end up
with stale GL state when drawing LayerDim which resulted
in incortect rendering.
Bug: 5467587
Change-Id: Id9fbed2843481d31063620f3662b364c7e3ac781
removed a test that could violate this guarantee. note that we
have no proof tha this ever happened, but consequences could be
hard lock-ups.
the code here was intended to track the region to update for displays
that can do partial update. the logic discarded the update entirely
if that region was empty. instead we just redraw the whole thing
(note that we should never be there with an empty region and retired
buffers though).
Bug: 5466259
Change-Id: I91ccab3b1a599e729e438eb833939e2236da6854
This change removes the dead code from SurfaceFlinger that resulted from
disabling support for freezing the display.
Change-Id: I4e5ff00c94b4c7a79af2f65c9850c135210068ed
This change enables a layer or orientation update transaction sent to
SurfaceFlinger to explicitly request a synchronous transaction.
Change-Id: I97cbba610c13679849f66114b216fa6dbf12f2a9
A LayerScreenshot is a special type of layer that contains a screenshot of
the screen acquired when its created. It works just like LayerDim.
Bug: 5446982
Change-Id: I7814aff2380e7e146937f2b641907be2a30c76cc
This change merges the ISurfaceComposer::setOrientation functionality
into ISurfaceComposer::setTransactionState. It enables the window
manager to atomically update both the display orientation and the
position and size of the windows in a single transaction with
SurfaceFlinger.
Bug: 5439574
Change-Id: I18a8ccc564d7d760ef8afb2d015ccdb7a7963900
this would happen when toggling on/off/on very fast, the screen
could stay black (while the panel is on).
Bug: 5429724
Change-Id: Ic8aa6aff066e6267923c0d47ef65e314e7bb6d41
This change modifies SurfaceFlinger's screenshot behavior when a layer
with a protected buffer is visible. The previous behavior was to simply
fail the screenshot. The new behavior is to render the screenshot using
a placeholder texture where the protected buffer would have been.
Change-Id: I5e50cb2f3b31b2ea81cfe291c9b4a42e9ee71874
when areas of the FB are undefined (transparent windows on top of
nothing), we clear those areas before composition.
however, it makes no sense to do this when the FB is not in use
(case where hwc handles all layers)
Bug: 5360529
Change-Id: If51bb669307e8419bbe1f3a89d1c88e0ec1f216c
Bug: 4981385
Simplify the orientation changing code path in the
WindowManager. Instead of the policy calling setRotation()
when the sensor determined orientation changes, it calls
updateRotation(), which figures everything out. For the most
part, the rotation actually passed to setRotation() was
more or less ignored and just added confusion, particularly
when handling deferred orientation changes.
Ensure that 180 degree rotations are disallowed even when
the application specifies SCREEN_ORIENTATION_SENSOR_*.
These rotations are only enabled when docked upside-down for
some reason or when the application specifies
SCREEN_ORIENTATION_FULL_SENSOR.
Ensure that special modes like HDMI connected, lid switch,
dock and rotation lock all cause the sensor to be ignored
even when the application asks for sensor-based orientation
changes. The sensor is not relevant in these modes because
some external factor (or the user) is determining the
preferred rotation.
Currently, applications can still override the preferred
rotation even when there are special modes in play that
might say otherwise. We could tweak this so that some
special modes trump application choices completely
(resulting in a letter-boxed application, perhaps).
I tested this sort of tweak (not included in the patch)
and it seems to work fine, including transitions between
applications with varying orientation.
Delete dead code related to animFlags.
Handle pausing/resuming orientation changes more precisely.
Ensure that a deferred orientation change is performed when
a drag completes, even if endDragLw() is not called because the
drag was aborted before the drop happened. We pause
the orientation change in register() and resume in unregister()
because those methods appear to always be called as needed.
Change-Id: If0a31de3d057251e581fdee64819f2b19e676e9a
we need to clear the whole framebuffer in that situation because
we can't trust the content of the FB when partial (fb preserving)
updates are used.
Bug: 5318492
Change-Id: I3f0e01b0fb665a34e44d88ad9f0f54a5d990060b
we were not redrawing and/or clearing the FB properly when
hwc moved a layer from/to FB to/from OVERLAY.
In these cases we needed to expand the dirty region to include
the layer that changed mode.
Also split composeSurfaces() which was becoming quite large.
Change-Id: Id6fa1acfc4ff694037fddf7efd037a4405732073
this happened when the overlays were in use, since the animation
is rendered in the FB and the FB is not used.
we now have a way to turn hwc off temporarily.
Change-Id: I3385f0c25bb9cc91948e7b26e7cd31ed18c36ace
these were due to the "preserve backbuffer" optimization
interfering with hw composer. basically the screen needed
to be redrawn in the areas that move from GL to
overlay.
Bug: 5245513
Change-Id: I9bf75c4fe905f3ef62005e52108b94edae692304
When taking screenshots we need to use the full drawing state list
instead of the visible list.
Bug: 5186823
Change-Id: I214ee0203aaf8e2c038e44581f7f1ae36edf08c5
This change fixes the NATIVE_WINDOW_QUEUES_TO_WINDOW_COMPOSER query of
Surface and SurfaceTextureClient. Surface now uses the inherited
SurfaceTextureClient implementation of this query. SurfaceTextureClient
now queries SurfaceFlinger to determine whether buffers that are queued
to its ISurfaceTexture will be sent to SurfaceFlinger (as opposed to
some other process).
Change-Id: Iff187e72f30d454229f07f896b438198978270a8
this is disabled by default. To enable:
setprop debug.sf.ddms 1
this debug option requires to restart SurfaceFlinger
Change-Id: Ic2f8050b29911b55bcd21721648b6978700c277d
- surfaceflinger now uses the GL-convention of
placing the origin in the left-bottom corner
- map texture coordinates of the screen capture
properly
- add the ability to control the animation
speed through a debug property
Bug: 4989276
Change-Id: Ifb3297bb578078b47146fff666c01f85417e0d6f
This is intended to absorb the cost of the IPC
to the permission controller.
Cached permission checks cost about 3us, while
full blown ones are two orders of magnitude slower.
CAVEAT: PermissionCache can only handle system
permissions safely for now, because the cache is
not purged upon global permission changes.
Change-Id: I8b8a5e71e191e3c01e8f792f253c379190eee62e
Add the concept of synchronous dequeueBuffer in SurfaceTexture
Implement {Surface|SurfaceTextureClient}::setSwapInterval()
Add SurfaceTexture logging
fix onFrameAvailable
The transaction flags were atomically read-and-cleared to determine if
a transaction was needed, in the later case, mStateLock was taken to
keep the current state still during the transaction. This left a small
window open, where a layer could be removed after the transaction flags
were checked but before the transaction was started holding the lock.
In that situation eTraversalNeeded would be set but only seen during the
next transaction cycle; however, because we're handling this transaction
(because of another flag) it will be commited, "loosing" the information
about the layer being removed -- so when the next transaction cycle due
to eTraversalNeeded starts, it won't notice that layers have been removed
and won't populated the ditchedLayers array.
Change-Id: Iedea9e25fee8dd98a0c5bd5ad41a20fcadf75b47
Client::mLayers could be accessed from different threads.
On one side from Client::attachLayer() which is currently
called from a binder thread; on the other side from
Client::detachLayer() which is always called from the main
thread.
This could lead to a corruption of Client::mLayers.
We fix this issue by adding an internal lock to Client.
Change-Id: Ib1317d7750ed5030e6f577efe34b69fc10198bd3
This leak was intentional, it was there to deal with the fact that
some gralloc implementations don't track buffer handles with
file-descriptors so buffers needed to stay alive until there were
registered, which is not guaranteed by binder transactions.
In this new implementation, we use a small BBinder holding a
reference to the buffer, which with tuck into the parcel. This forces
the reference to stay alive until the parcel is destroyed, which
is guaranteed (by construction) to happen after the buffer is
registered.
this allows the public facing API to not expose the previous hack.
Change-Id: I1dd6cd83679a2b7457ad628169e2851acc027143
mDrawingState doesn't need to be accessed by the
mStateLock, because by definition it's only accessed
from the main thread.
Similarily, the list of layers in the drawing state
cannot change (ie: is const).
Change-Id: I2e5da7f4d8caee7af7802b432cd45cc81c7c08b0
This change makes SurfaceFlinger treat layers for which the active
buffer has the GRALLOC_USAGE_PROTECTED bit set as if they have the
'secure' flag set.
Change-Id: Ic60b6513a63e4bb92ec6ce9fd12fd39b4ba5f674
Bug: 4081304
This change adds a new 'method' to the ANativeWindow interface to check
whether buffers queued to the window will be sent directly to the system
window compositor.
Change-Id: I4d4b199e328c110b68b250029aea650f03c8724d
Bug: 3495535
with this change DimLayers will behave just like any other layer,
in particular they'll respect the layer transformations.
Change-Id: Icb4a1275e8bca9e3deb5f57c9f9219aaa69f9877
Generally we never want to lock a buffer for write access if it is at
the "head" on the surfaceflinger side. The only exception (1) is when
the buffer is not currently in use AND there is at least one queued
buffer -- in which case, SurfaceFlinger will never use said buffer
anymore, because on the next composition around, it will be able to
retire the first queued buffer.
The logic above relies on SurfaceFlinger always retiring
and locking a buffer before composition -- unfortunately this
didn't happen during a screenshot.
This could leave us in a situation where a buffer is locked by the
application for write, and used by SurfaceFlinger for texturing,
causing a hang.
Here, we fix this issue by never assuming the exception (1), it was
intended as an optimization allowing ANativeWindow::lockBuffer() to
return sooner and was justified when most of SF composition was
done in software. The actual buffer locking is now ensured by
gralloc. We could have handled screenshots in a similar way to
a regular composition, but it could have caused glitches on screen,
essentially, taking a screenshot could cause to skip a frame.
now that we removed the notion of a "inUse" buffer in surfaceflinger
a lot of code can be simplified / removed.
noteworthy, the whole concept of "unlockClient" wrt. "compositionComplete"
is also gone.
We used to guarantee that a layer in SurfaceFlinger would never be
destroyed before all references (to its ISurface) on the client
side would be released. At some point, this guarantee got
relaxed to allow to free gralloc resources sooner. This last
change was incorrect, because:
- in implementations with reference-counting the gralloc resources
wouldn't be released anyways, until all the mapping were gone
- in implementations without ref counting, the client side
would most likely crash or do something bad
- it also caused the SharedBufferStack slot to be reallocated
to another surface, which could be problematic if the client
continued to use the surface after the window manager destroyed it.
So, we essentially reinstate the guarantee that layers won't be
destroyed until after all references to their ISurface are
released.
NOTE: This doesn't entirely fix 3306150 because there is another
problem there where the Browser continues to use a surface after it
has been destroyed.
also improve SurfaceFlinger 'dumpsys' log
list the purgatory, which shows windows that have been closed,
but for which the client still has references.
we were not clearing the screen entirely, which caused garbage when
the screen wasn't entirely covered by windows.
Change-Id: Ia7aa13c36a8a314e0e8427d419b16b9aa2165ddf
we make sure to call compositionComplete after everytime we do
composition with the GPU (even for the screenshot case), which is
where the buffer locks are released.
Change-Id: I450430d1e4d1ee9ce1023970642378c42cdcfa4c
now that we removed the notion of a "inUse" buffer in surfaceflinger
a lot of code can be simplified / removed.
noteworthy, the whole concept of "unlockClient" wrt. "compositionComplete"
is also gone.
Change-Id: I210413d4c8c0998dae05c8620ebfc895d3e6233d
list the purgatory, which shows windows that have been closed,
but for which the client still has references.
Change-Id: I5168bb88cb328d5d77d71d0871deb9190f493126
We used to guarantee that a layer in SurfaceFlinger would never be
destroyed before all references (to its ISurface) on the client
side would be released. At some point, this guarantee got
relaxed to allow to free gralloc resources sooner. This last
change was incorrect, because:
- in implementations with reference-counting the gralloc resources
wouldn't be released anyways, until all the mapping were gone
- in implementations without ref counting, the client side
would most likely crash or do something bad
- it also caused the SharedBufferStack slot to be reallocated
to another surface, which could be problematic if the client
continued to use the surface after the window manager destroyed it.
So, we essentially reinstate the guarantee that layers won't be
destroyed until after all references to their ISurface are
released.
NOTE: This doesn't entirely fix 3306150 because there is another
problem there where the Browser continues to use a surface after it
has been destroyed.
Change-Id: I305c830dd722b30a6d53cbf3a9c714fd3cf7eb06
the crop as well as buffer orientation can change at every frame, when that happens
we need to reset the hwc HAL (ie: set the GEOMETRY_CHANGED flag).
currently we achieve this by taking the same code path than an actual geometry change
which is a bit more heavy than necessary.
Change-Id: I751f9ed1eeec0c27db7df2e77d5d17c6bcc17a24
This change fixes a horrible hack that I did to allow application
processes to create GraphicBuffer objects by making a binder call to
SurfaceFlinger. This change introduces a new binder interface
specifically for doing this, and does it in such a way that
SurfaceFlinger will maintain a reference to the buffers until the app is
done with them.
Change-Id: Icb240397c6c206d7f69124c1497a829f051cb49b
This change adds a new binder method to the ISurfaceComposer interface.
This IPC is intended to allow SurfaceFlinger clients to allocate gralloc
buffers using SurfaceFlinger as a proxy to gralloc.
Change-Id: Ide9fc283aec5da6268ba62cfed0c3319a50b640d
we were not clearing the screen entirely, which caused garbage when
the screen wasn't entirely covered by windows.
Change-Id: Ie9ab9b94eabfa6cafddf45bb14bc733bdc8d35c0
while we're waiting for the real fix in the gralloc/gpu driver,
this workaround should resolve the issue.
we make sure to call compositionComplete after everytime we do
composition with the GPU (even for the screenshot case), which is
where the buffer locks are released.
Change-Id: I3cb5ad67d48c81a23100172bab77e86a70e29152
