This also fixes [2152536] ANR in browser
When SF is enqueuing buffers faster than SF dequeues them.
The update flag in SF is not counted and under some situations SF will only
dequeue the first buffer. The state at this point is not technically
corrupted, it's valid, but just delayed by one buffer.
In the case of the Browser ANR, because the last enqueued buffer was delayed
the resizing of the current buffer couldn't happen.
The system would always fall back onto its feet if anything -else- in
tried to draw, because the "late" buffer would be picked up then.
A window is created and the browser is about to render into it the
very first time, at that point it does an IPC to SF to request a new
buffer. Meanwhile, the window manager removes that window from the
list and the shared memory block it uses is marked as invalid.
However, at that point, another window is created and is given the
same index (that just go freed), but a different identity and resets
the "invalid" bit in the shared block. When we go back to the buffer
allocation code, we're stuck because the surface we're allocating for
is gone and we don't detect it's invalid because the invalid bit has
been reset.
It is not sufficient to check for the invalid bit, I should
also check that identities match.
When EGLImage extension is not available, SurfaceFlinger will fallback to using
glTexImage2D and glTexSubImage2D instead, which requires 50% more memory and an
extra copy. However this code path has never been exercised and had some bugs
which this patch fix.
Mainly the scale factor wasn't computed right when falling back on glDrawElements.
We also fallback to this mode of operation if a buffer doesn't have the adequate
usage bits for EGLImage usage.
This changes only code that is currently not executed. Some refactoring was needed to
keep the change clean. This doesn't change anything functionaly.
The ANR is caused by SurfaceFlinger waiting for buffers of a removed surface to become availlable.
When it is removed from the current list, a Surface is marked as NO_INIT, which causes SF to return
immediately in the above case. For some reason, the surface here wasn't marked as NO_INIT.
This change makes the code more robust by always (irregadless or errors) setting the NO_INIT status
in all code paths where a surface is removed from the list.
Additionaly added more information in the logs, should this happen again.
We were emitting GL commands, calling composition complete and releasing clients
without ever calling eglSwapBuffers(), which is completely wrong on non-direct
renders. This could cause transient drawing artifacts when unfreezing the
screen (upon orientaion change for instance) and could also block the clients
for ever as they are waiting for their previous buffer to be rendered.
This appears to fix the sim-eng build on the gDapper build machines.
Basic problem is that LayerBuffer::OverlaySource has a constructor that
calls SurfaceFlinger.signalEvent(). SurfaceFlinger lists LayerBuffer
as a friend, but that's not enough to convince gcc that the embedded
OverlaySource class is also a friend. I don't see a way to make them
friendly, so I marked signalEvent() as public.
a new method, compostionComplete() is added to the framebuffer hal, it is used by surfaceflinger to signal the driver that the composition is complete, BEFORE it releases its client. This gives a chance to the driver to
we ended-up locking a Mutex that had been destroyed.
This happened because we gave an sp<Source> to the outside world,
and were called after LayerBuffer had been destroyed.
Instead we now give a wp<LayerBuffer> to the outside and have it
do the destruction.
Rewrote SurfaceFlinger's buffer management from the ground-up.
The design now support an arbitrary number of buffers per surface, however the current implementation is limited to four. Currently only 2 buffers are used in practice.
The main new feature is to be able to dequeue all buffers at once (very important when there are only two).
A client can dequeue all buffers until there are none available, it can lock all buffers except the last one that is used for composition. The client will block then, until a new buffer is enqueued.
The current implementation requires that buffers are locked in the same order they are dequeued and enqueued in the same order they are locked. Only one buffer can be locked at a time.
eg. Allowed sequence: DQ, DQ, LOCK, Q, LOCK, Q
eg. Forbidden sequence: DQ, DQ, LOCK, LOCK, Q, Q
(in this case the state is dumped without the proper locks held which could result to a crash)
in addition, the last transaction and swap times are printed to the dump as well as the time spent
*currently* in these function. For instance, if SF is unresponsive because eglSwapBuffers() is stuck,
this will show up here.
what happened is that the efective pixel format is calculated by SF but Surface nevew had access to it directly.
in particular this caused query(FORMAT) to return the requested format instead of the effective format.
this would happen is the window is made visible but the client didn't render yet into it. This happens often with SurfaceView.
Instead of filling the window with solid black, SF would simply ignore it which could lead to more disturbing artifacts.
in theory the window manager should not display a window before it has been drawn into, but it does happen occasionnaly.
This change makes SurfaceHolder.setType(GPU) obsolete (it's now ignored).
Added an API to android_native_window_t to allow extending the functionality without ever breaking binary compatibility. This is used to implement the new set_usage() API. This API needs to be called by software renderers because the default is to use usage flags suitable for h/w.
The current gralloc allocates buffer memory for render targets that will typically have NPOT dimensions. Assuming that the vendor driver supports converting the resulting NPOT android_native_buffer_t to a NPOT EGLImage, SurfaceFlinger calls glEGLImageTargetTexture2DOES(), and uses glGetError() to test whether the GL can support creating an EGL target texture with the specified NPOT EGLImage. If it is supported, the DIRECT_TEXTURE flag remains set, otherwise it is cleared.
Tangentially, if the driver advertises the GL_ARB_texture_non_power_of_two extension, the NPOT_EXTENSION flag is set, otherwise it is cleared.
If the driver supported creating an EGL target texture from a NPOT source EGLImage, it implicitly creates a NPOT texture. This does not need any glScalef() texture coordinate correction in LayerBase::drawWithOpenGL(). However, the same driver may not advertise the GL_ARB_texture_non_power_of_two extension nor generally support NPOT textures that were not derived from EGLImages. So SurfaceFlinger may flag only DIRECT_TEXTURE, not NPOT_EXTENSION.
Therefore, the test in LayerBase::drawWithOpenGL() should only perform the glScalef() if neither NPOT_EXTENSION or DIRECT_TEXTURE are flagged. Otherwise scaling is applied to NPOT EGL target textures when none is required.
Merge commit '1521cd6e657ba4efa9382ab73d3cbba3bdf50ead'
* commit '1521cd6e657ba4efa9382ab73d3cbba3bdf50ead':
Reset the mDpiX and mDpiY values when qemu.sf.lcd_density is defined.
doesn't happen because the visibility never changes. With this change
the overlay parameter and position will be committed when either the visibility
of the window changes, or on the first call to visibility resolved, if it
hasn't already been done.
Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
Merge commit '58ebdcc06eca06741460a7db2be4b79e3865eb88'
* commit '58ebdcc06eca06741460a7db2be4b79e3865eb88':
fix [1947273] the DimLayer causes the whole screen to update during transactions
we can't use a texture of 1/4th of the screen for the dim layer, because the mdp internal input resultion is alwyas integers and for very small blits of a couple pixels the scale factor can get way out of range, for instance for a 7 pixels source, the scale factor would be either 7 (7/1) or 3.5 (7/2) instead of 4 (7/1.75). This caused the mdp to fail in some cases and revert to software. we now always use a texture of the actual screen size, so the problem will never happen. This burns 300KB of pmem instead of 21KB. On devices with a larger screen we might want to use a smaller texture and tile it by hand.
