| Age | Commit message (Collapse) | Author |
|---|
|
df476dddf5 am: 69140e9c9c
Original change: https://android-review.googlesource.com/c/platform/frameworks/native/+/1596925
MUST ONLY BE SUBMITTED BY AUTOMERGER
Change-Id: I2e187e756cf202c91debeea8f9bbad76f5463790
|
|
... to scan a fixed range of priorities followed by
consecutive priorities instead of scanning all system
props (property_list()) which causes sepolicy audits
when Vulkan is loaded in a process with a restrictive
sepolicy.
This avoids sepolicy audits in system_server for example
when running on Cuttlefish with SwANGLE (ANGLE GL on top
of SwiftShader Vulkan).
Bug: b/179967574
Test: boot Cuttlefish with SwANGLE and inspect audit logs
Change-Id: I63b3f45113fbdd3ab62c22f062150e25b99d7352
|
|
Android historically preloads GL driver in Zygote to speed up app launch
time and avoid the duplicate extra memory for each process loading the
graphics driver. In Android 10, we landed GL driver unloading mechanism
to ship updatable driver and ANGLE without perf overhead on the rest of
the system.
To get prepared for the HWUI Vulkan backend being turned on as the
default renderer for UI componments, this CL will do the same to the
Vulkan driver to preload it in Zygote and unload so to reload updated
driver if needed.
Bug: 135536511
Test: atest CtsUiRenderingTestCases no regression on VK backend
Change-Id: I2909f6ecc4f011b1f3670aacdf817ed3b6e9a487
|
|
This is a globally dispatched api, and it's the best option to preload
the driver in Zygote with minimum overhead.
This change also updates some return codes according to the spec.
Bug: 135536511
Test: take a trace and verify driver is loaded
Change-Id: I6fe425ec568b13577d3d9471aa5f181cff44c61c
|
|
Previously the vulkan loader only exposes device extensions from the
driver implementation when the layer name is not specified. Per
vkEnumerateDeviceExtensionProperties spec the loader must also advertise
device extensions from implicitly enabled layers.
Bug: b/143293104
Test: atest android.gputools.cts.CtsRootlessGpuDebugHostTest
Change-Id: Iaf5786ba7e371a290ecd1764af69b5298371cfdd
|
|
Previously the vulkan loader only exposes instance extensions from the
driver implementation. Per vkEnumerateInstanceExtensionProperties spec
the loader must also advertise instance extensions from implicitly
enabled layers.
Bug: b/143293104
Test: vkEnumerateInstanceExtensionProperties returns the instance
extensions from implicit layers.
Change-Id: I17f4ce370bf5f4bba295165a28836e2b7c03a318
|
|
qt-qpr1-dev am: eab94c8f45 am: 4f2cf91f9a
Change-Id: Ifb3ca893ba2315167c205897355a49c510dffb58
|
|
qt-qpr1-dev am: eab94c8f45
Change-Id: I28a9939b33926ebf285a93047ab21691c5e81576
|
|
There're below cases which we can load the debug layers:
(1) Layer search paths added from Settings global
Layers enabled from Settings global
(2) Layer search paths added from default system debug path
Layers enabled from system property
(3) Layer search paths are from app class loader's default paths
Layers enabled from system property
For (1), we have done the debuggable check in Java land GraphicsEnvironment.
For (2), we have a separate check debuggable logic for adding the system
debug path to the layers search paths.
For (3), we should never need to check debuggable.
So we should just remove the redundant debuggable check when we add implicit
layers, which can unblock (3) immediately as well as maintaining the security
logic around debug layers.
Bug: 147838267
Bug: 149183377
Test: CtsRootlessGpuDebugHostTest#testSystemPropertyEnableVulkan on userbuild
Change-Id: I80d9cf3ff6786ce4d4ec5d8afb8da58ba0842ec9
|
|
Change-Id: I2bb295fbf56168beefa5dd26ea21cb85a90a6752
|
|
By default, PR_SET_DUMPABLE is 0 for zygote spawned apps, except in the
following circumstances:
1. ro.debuggable=1 (global debuggable enabled, i.e., userdebug or eng builds).
2. android:debuggable="true" in the manifest for an individual application.
3. An app which explicitly calls prctl(PR_SET_DUMPABLE, 1).
4. GraphicsEnv calls prctl(PR_SET_DUMPABLE, 1) in the presence of
<meta-data android:name="com.android.graphics.injectLayers.enable"
android:value="true"/>
in the application manifest.
So checking both ro.debuggable=1 and PR_GET_DUMPABLE is redundant.
Bug: b/144186877, b/148566223
Test: CtsAngleIntegrationHostTestCases
Test: CtsRootlessGpuDebugHostTest
Change-Id: I934f64315b67db77ee2c2a9dff50fb23bc0a546a
Merged-In: I934f64315b67db77ee2c2a9dff50fb23bc0a546a
(cherry picked from commit 6a674c9e105bdc5d736c06a4500dcdac1c6c4006)
|
|
|
|
By default, PR_SET_DUMPABLE is 0 for zygote spawned apps, except in the
following circumstances:
1. ro.debuggable=1 (global debuggable enabled, i.e., userdebug or eng builds).
2. android:debuggable="true" in the manifest for an individual application.
3. An app which explicitly calls prctl(PR_SET_DUMPABLE, 1).
4. GraphicsEnv calls prctl(PR_SET_DUMPABLE, 1) in the presence of
<meta-data android:name="com.android.graphics.injectLayers.enable"
android:value="true"/>
in the application manifest.
So checking both ro.debuggable=1 and PR_GET_DUMPABLE is redundant.
Bug: 144186877
Test: CtsAngleIntegrationHostTestCases
Test: CtsRootlessGpuDebugHostTest
Change-Id: I934f64315b67db77ee2c2a9dff50fb23bc0a546a
|
|
Fixes: 140609421
Test: build, flash and boot
Change-Id: I7548bf733b32861c8007cee63df954022c111a1d
|
|
Bug: 134185757
Test: build
Change-Id: Ib9a27ba8b8da53707337d3d48e6da502f38670e5
|
|
This change fixes the advertisement of core Vulkan entry points as below:
1. GIPA returns a valid checked_proc for 1.1 core device APIs.
2. GDPA returns NULL for 1.1 core device APIs on a 1.0 physical device.
Bug: 134185757
Bug: 142266108
Test: dEQP-VK.memory.binding on 1.1 loader and 1.0 device ICD
Test: dEQP-VK.api.info.instance on 1.1 loader and 1.0 instance ICD
Change-Id: Ieb814516a4019f1a330178bfeb414e1450895c5c
Merged-In: I0a3e06dc04bade4f36a7e68ee2f53979c656ee4e
|
|
This change fixes the advertisement of core Vulkan entry points as below:
1. GIPA returns a valid checked_proc for 1.1 core device APIs.
2. GDPA returns NULL for 1.1 core device APIs on a 1.0 physical device.
Bug: 134185757
Bug: 142266108
Test: dEQP-VK.memory.binding on 1.1 loader and 1.0 device ICD
Test: dEQP-VK.api.info.instance on 1.1 loader and 1.0 instance ICD
Change-Id: I0a3e06dc04bade4f36a7e68ee2f53979c656ee4e
|
|
After decoupling layer discovery from driver loading, the layer discovery is
triggered at each call of vkCreateInstance, vkEnumerateInstanceLayerProperties
and vkEnumerateInstanceExtensionProperties. However, it takes non-trivial time
to traverse the layer search path for priviledged apps and non-updated system
apps. So this change just makes sure the layer discovery logic is triggered only
once for a new process.
Bug: 139443653
Bug: 135536511
Test: preload Vulkan and atest CtsGpuToolsHostTestCases
Change-Id: Ibe502fd4b089acbbff6f4a2485fa61c736a484b5
|
|
Previously, the layer discovery logic is only executed once along with the
driver loading. However, once Vulkan driver is preloaded in Zygote, new layers
pushed to the system or shipped with the App's apk won't be discovered at
runtime. This change refactors the logic of layer discovery. It doesn't hurt to
keep finding new layers available in the layer search path, because the app apk
itself won't change at runtime. Even if we push new layers to the system search
path, that's only on the debug build of the system.
Bug: 135536511
Test: preload Vulkan and atest CtsGpuToolsHostTestCases
Change-Id: I915b78dacfd9b637a202f76969d559a31eded686
|
|
Bug: 137862180
Test: flash, remove vkGetInstanceProcAddr from layer symbols, run
Change-Id: I4d4d75585623d2c15c57b4d8ad0243a2fc347fce
|
|
Test: build, flash and take traces
Change-Id: Idab6d748519c4f5bc6aac3d51efd26cad28c94ec
|
|
Add the ability to load GPU debug layers from the base
directory of debuggable applications.
This update concides with changes to framework/base to
control the enabling logic in GraphicsEnvironment.
This commit changes the Vulkan loader to:
* Scan an additional location for debug layers.
* Use the provided order of layers from GraphicsEnvironment,
overriding system properties.
* Use the first instance of a layer found, in the case of duplicates.
* Move layers paths and namespace to GraphicsEnv, removing vulkan_loader_data
Bug: 63708377
Test: Manual, CTS tests to follow
Change-Id: I38dc91bbc26671f5093ee1b12454fc024c0b5892
|
|
Import Vulkan 1.1 header from
khronos master@5b2d66c0e6906388385ed7adfd015d80ec0d846a
with the following modifications:
- VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETER_FEATURES 10006300 -> 1000063000
- remove VK_KHR_get_display_properties2 (which is not ratified)
- device_group interactions from KHR_surface to KHR_swapchain (per MR 2409)
Summary of changes since 1.0.61:
* VK_KHR_surface -> v26 (subsumed interactions with VK_KHR_device_group)
* VK_KHR_swapchain -> v69 (subsumed interactions with VK_KHR_device_group)
* VK_EXT_debug_report -> v9 (no api changes)
* VK_KHX_multiview -> VK_KHR_multiview
* VK_KHX_device_group -> VK_KHR_device_group
* VK_KHX_device_group_creation -> VK_KHR_device_group_creation
* VK_KHR_push_descriptor -> v2
* Added VK_KHR_maintenance3
* Added 1.1 core enums, structs, and functions (kept KHR suffixed ones too)
Implemented the following in libvulkan:
* EnumerateInstanceVersion
* GetDeviceGroupPresentCapabilitiesKHR (minimal support for 1 device)
* GetDeviceGroupSurfacePresentModesKHR (minimal support for local present modes only)
* GetPhysicalDevicePresentRectanglesKHR (assumes whole window)
* AcquireNextImage2KHR (minimal thunk to AcquireNextImageKHR)
* GetDeviceQueue2
* added support for VK_SWAPCHAIN_CREATE_PROTECTED_BIT_KHR
* support for 1.1 core device_groups functionality (no VK_KHR_device_group_creation)
- EnumeratePhysicalDeviceGroups
* added trampolines for 1.1 core entry points
* disabled apiVersion >= 1.1 check
Implemented in nulldrv (but largely untested beyond build/link)
* EnumerateInstanceVersion
* enough empty thunks for VK 1.1 core functions that it links
Test: Build Android
Change-Id: I247cd670543867bdf80877156d941cf3bfcf6303
|
|
Test: compile
Bug: 26552300
Bug: 31289077
Change-Id: I84d0ffa430df3beca0dcc6ed205ce220d7be1f87
|
|
- cutils/log.h treat as deprecated and lead by example
- android/log.h to be used instead of log/log.h if possible
- add system includes that are assumed as side effects
- define LOG_TAG first
Test: compile
Bug: 31289077
Bug: 30465923
Change-Id: I8d99b24c333578c9b5aa9f2a01324bd0bba268dd
|
|
While testing argument validation in the loader, I found that
I got SEGFAULTs trying to access the logger class created in
LayerChain::CreateDevice. I believe that the logger object
was being destroyed at the end of the LayerChain construction
statement so that a reference to the logger caused a SEGFAULT.
Fix CreateInstance path as well, though the loader doesn't
currently call the logger in that path, so no way to verify.
Test: run vkCreateDevice_test on Vulkan device.
See tests/README.md for details
Bug: 33783620
Change-Id: Ib46023e8a8dfa008510c655c5b216aabaaf2ea83
|
|
- cutils/log.h treat as deprecated and lead by example
- android/log.h to be used instead of log/log.h if required
- add system includes that are assumed as side effects of log.h
- modules that use SLOG should use SLOG for all logging (installd)
- define LOG_TAG first
- remove logging infrastructure if not used
Test: build
Bug: 31289077
Change-Id: Iea147a0104c7ab7f12451304131d6500f42141e7
|
|
We fail to set up get_instance_proc_addr_ and get_device_proc_addr_ when
there are layers but only non-global ones.
Bug: 28990561
Change-Id: I82704f946deb5a76640f9dfb792e6fd118c78aa7
|
|
Use debug report logger for actionable application errors (but not other
errors).
Bug: 28120066
Change-Id: Ia9a3b568e415555bf42c8a76d1df6279835a710c
|
|
Allow instance layers to intercept all commands and enumerate device
extensions. Ignore application device layers. Enumerate all enabled
instance layers in vkEnumerateDeviceLayerProperties.
Bug: 27911856
Change-Id: I6e89439ab10835dd1a43732c2333a92201e52550
|
|
Return VK_ERROR_LAYER_NOT_PRESENT when the layer is not available.
Change-Id: I9d9aafe6e40c2ca49e58bc7c70114d0f11de2f81
|
|
Merge Get*LayerRef into GetLayerRef and remove LayerRef::is_instance_.
With the removal, LayerRef::SupportsExtension becomes ambiguous. Replace
it with FindLayer*Extension. Remove unused LayerRef::GetName and
LayerRef::GetSpecName while at it.
There should be no user-visible change.
Bug: 27911856
Change-Id: I38340654b52338e9ed46d1c4462a6b254b0cab64
|
|
The only user-visible change should be improved error messages.
Bug: 27911856
Change-Id: Ie50a9d37f07b590026176642f2c67270225f9280
|
|
Replace Get*LayerExtensions by a set of new functions that do not
distinguish instance and device layers.
There should be no user-visible change.
Bug: 27911856
Change-Id: Icd98abf51a936769f8f2f218794043b5e2611c5c
|
|
Replace Enumerate*Layers by a set of new functions that do not distinguish
instance and device layers. The new functions are also careful not to
pollute the rest of the loader with std containers.
There should be no user-visible change.
Bug: 27911856
Change-Id: I4790fadc1aa2ea934a4628bce55dd45892f15e0b
|
|
Initialize dispatch table entries for non-enabled WSI functions to stubs.
We do not want to initialize them to NULL because they may still be
invoked through the exported WSI entrypoints.
Bug: 25850852
Change-Id: I25b715700990ad7432740f031764d70396024d32
|
|
VK_LAYER_FUNCTION_DATA_CALLBACK effectively allows us to pass
driver::SetDataInternal to layers. It will be called on handles of
potentially unknown types. Add two internal types
VK_DEFINE_HANDLE(InstanceDispatchable)
VK_DEFINE_HANDLE(DeviceDispatchable)
in driver namespace for type safety.
Bug: 28015368
Change-Id: I7389829a7d8c374197cd7046973777b49e436961
|
|
Change-Id: I2f102fa97e3c976fa411b4e6bdc83dacfa243af7
|
|
Change-Id: If44c40f87b1e3e2eb69d12543a7f3b3036291c88
|
|
Change-Id: Ic7bc29efc1ae5b734ebfc692c441b9537134ccb2
|
|
nyc-dev
* changes:
vulkan: remove unused loader.{cpp.h}
vulkan: use driver::GetData everywhere
vulkan: move all _Bottom functions
vulkan: add swapchain.h
vulkan: add layers_extensions.h
vulkan: rework CreateInstance_Bottom and related ones
vulkan: rework EnumerateDeviceExtensionProperties_Bottom
vulkan: rework {Create,Destroy}Device_Bottom
vulkan: move AllocateCommandBuffers_Bottom
vulkan: move GetDeviceQueue_Bottom
vulkan: rework DriverDispatchTable
vulkan: rework driver::Get*ProcAddr
vulkan: add VK_ANDROID_native_buffer to vulkan.api
vulkan: move driver::GetDefaultAllocator
vulkan: move driver::OpenHAL
vulkan: move driver::Debuggable
|
|
Move everything for layers_extensions.cpp out of loader.h to
layers_extensions.h, and into vulkan::api namespace.
Remove now unused InstanceExtensionFromName and DeviceExtensionFromName.
Change-Id: I06be98986f40de35e9dffb5499bc1423e9eb3d48
|
|
Return VK_ERROR_EXTENSION_NOT_PRESENT when an app attempts to enable an
unsupported extension. This fixes a regression introduced by my top
rewrite (commit 0c20324: vulkan: rewrite top of loader).
Since we do not cache HAL extensions, each vkCreateInstance or
vkCreateDevice call is preceded by two extension enumeration calls
internally.
Change-Id: I5342c1a5c9a5452dd2bc52933a5ee558db174048
|
|
The top is responsible for layer discovery and chaining, and the bottom is
like a regular layer that is always enabled and is inserted just before
the driver. Make the separation clear by rewriting the top and stripping
the layer managment code from loader.cpp.
Change-Id: I64e525e27bd4c297bccd94a1eb9b88e28088e85d
|
