| 52ea25cf06cef250ec73052611b48556b3fce4d5 |
|
14-Sep-2017 |
Corey Tabaka <eieio@google.com> |
Add shared memory based buffer metadata
This CLs reduces BufferHub CPU consumption by adding asynchronous
state transition so that out-of-process VR composition can run on 2016
pixel devices smoothly. In addition, this CL addresses a couple corner
cases in the existing bufferhub logic, which fixes various blackscreen
issues.
1/ Tracks buffer transition states (gained, posted, acquired, released)
from the client side via atomic shared memory and adds
PostAsync/AcquireAsync/ReleaseAsync/GainAsync with metadata and
fence support.
2/ Adds dequeue order guarantee for buffers enqueued with
dvrWriteBufferQueuePostBuffer.
3/ Synchronous BuffeHub operations are still supported.
4/ Bump up the bufferhubd's soft limit of open file descriptor.
5/ Handle orphaned consumer in acquired state. This is a corner case
that consumer process goes aways (most likely due to a crash) leaving
buffer stuck in acquired state with inconsistent buffer state.
6/ Fixes a race condition for released buffer to be Gain'ed and
Acquire'd when a new consumer is created in released state.
7/ Improve silent consumer queue efficiency: Silent queues no longer
import buffers or receive signals about new buffers and they are
limited to only spawning other consumers and notifications about
producers hanging up.
8/ Modify PDX/UDS channel event signaling to work around epoll
behavior. PDX UDS uses a combination of an eventfd and an epoll set
to simulate the original PDX transport channel events. An odd
behavior discovered in the kernel implementation of epoll was found
that causes the epoll fd to "unsignal" itself whenever epoll_wait()
is called on it, regardless of whether it should still be
pending. This breaks the edge triggerd behavior in nested epoll sets
that channel events depend on. Since this is unlikely to ever be
fixed in the kernel we work around the behavior by using the epoll
set only as a logical OR of two eventfds and never calling
epoll_wait() on it. When polling is required we use regluar poll()
with the eventfds and data fd to avoid the bad behavior in
epoll_wait().
9/ Keep reading data after PDX hangup signal. UDS will signal hangup
when the other end of the socket closes. However, data could still be
in the kerenl buffer and should be consumed. Fix an issue where the
service misses an impulse sent right before the socket is closed.
Bug: 65455724
Bug: 65458354
Bug: 65458312
Bug: 64027135
Bug: 67424527
Test: libpdx_uds_tests
bufferhub_tests
buffer_hub_queue-test
buffer_hub_queue_producer-test
dvr_api-test
Change-Id: Id07db1f206ccf4e06f7ee3c671193334408971ca
/frameworks/native/libs/vr/libpdx_uds/client_channel_factory.cpp
|
| e41826ac301b9afad1cbfe59c57e27f2495be968 |
|
02-May-2017 |
Alex Vakulenko <avakulenko@google.com> |
libpdx_uds: Always create channel sockets in the server process
Whenever a client connects to PDX service endpoint, the connected
socket was being used as the channel connection. This makes it
difficult to ensure that the socket has the appropriate SELinux
label applied.
Changed the implementation to make sure that the actual channel socket
pair is created in one place in the server process.
The new Endpoint::CreateChannelSocketPair() is now being used to
create the channel sockets when an incoming connection is accepted
and when PushChannel request is received.
The newly created channel socket is sent to the client in response
to connection request.
Bug: 37646189
Test: Device boots. CubeSea launches.
Change-Id: Id3158484aeb18007f18b279afc60613076f80f14
/frameworks/native/libs/vr/libpdx_uds/client_channel_factory.cpp
|
| 6eefa42d33bdddc3603211634f98937b00abc532 |
|
18-Apr-2017 |
Alex Vakulenko <avakulenko@google.com> |
libpdx_uds: Allow to create Endpoint/ClientChannel from a socket pair
This is important to enable Service/Client operation in unit tests.
Being able to create a pair of Unix domain sockets and construct both
Service and Client so that they can talk to each other without having
to create a physical socket file is convenient.
This change makes it possible to create an instance of Endpoint and
ClientChannel classes based just on a pair of sockets (Endpoint does
take another socket to simulate the main endpoint FD to accept incoming
connection on, but it is not used for this, only the shutdown events
are received from that main socket. Endpoint uses the channel FD to
perform actual communication with the client).
Bug: 37443070
Test: `libpdx_uds_tests` pass
Change-Id: Ifa1a9d03b97bd90282a04715c2105ad37a8de936
/frameworks/native/libs/vr/libpdx_uds/client_channel_factory.cpp
|
| c0fd676fa087b38b63d253536c6fc9513a0470c8 |
|
12-Apr-2017 |
Alex Vakulenko <avakulenko@google.com> |
libpdx_uds: Handle EACCES error when connecting to PDX service
There is a race condition in the way init process creates socket
files before it forks into a child (see system/core/init/util.cpp,
function create_socket()). It first creates a socket, then calls
bind() which creates a file entry, then calls chown/chmod to change
ownership and access permissions of that file object. If a client
process connects to the socket just after bind() was called but before
chmod(), the socket file will have incorrect permissions and the
connect call will be aborted with EACCESS (Permission denied) error.
We should retry connection in case we get EACCESS. A proper fix would
be is to find a way for init to create the file entry with the proper
permissions to start with, and eliminate the race condition altogether.
Bug: 37171113
Test: `lunch sailfish-eng && m -j32` succeeds
Device boots correctly.
Error recovery is handled correctly even after adding sleep(5)
to init process just before calling chmod on the socket.
Change-Id: If97c316daa0c0be5ff7b5aa302a69aa60e9fb237
/frameworks/native/libs/vr/libpdx_uds/client_channel_factory.cpp
|
| 09aa736ee6136787b18807d7ab459feba23cec04 |
|
06-Apr-2017 |
Alex Vakulenko <avakulenko@google.com> |
libpdx_uds: Fix RPC channel ID allocation to not recycle values as often
The value of channel ID for PDX service on UDS transport was actual file
descriptor value for the data socket. Since channels are open and closed
constantly, it is quite often that channel ID is being reused immediately,
so it is almost impossible to use `cid` as unique identifier for objects
being passed around.
Instead, we now use monotonically growing channel ID value.
To prevent the possibility of using channel ID as the socket FD and vice
versa, changed all helper functions that used to take socket_fd as an int
to explicitly use BorrowedHandle which disables implicit conversion from
int (and hence makes it impossible to mistakenly pass in the channel ID).
Bug: 37082296
Test: `m -j32` succeeds for sailfish-eng
Ran libpdx_uds_tests on the device -> all pass
Device boots and CubeSea and VrHome render correctly with vr_flinger
Change-Id: Ibb8dfee4d6c3f4b6120c0b6e20a253f1b9307c19
/frameworks/native/libs/vr/libpdx_uds/client_channel_factory.cpp
|
| 4782814b340bb55a0fbff9f7b6cfd4d27cffd85a |
|
24-Mar-2017 |
Alex Vakulenko <avakulenko@google.com> |
libpdx_uds: Improve client connection logic
Handle the case where the service created a socket file but not bound the
socket to it, or not listening for incoming requests yet. Also, if the service
has crashed and left the socket file behind, need to make sure we reconnect
once the service is restarted.
Bug: None
Test: `m -j32` succeeds, device boots and CubeSea app works on Sailfish
Change-Id: I2039cfca6faccd5d1d4b725e454075669484b880
(cherry picked from commit d074fc416a7a90e76a0b28a870080c1dd30079a1)
/frameworks/native/libs/vr/libpdx_uds/client_channel_factory.cpp
|
| e4eec20f6263f4a42ae462456f60ea6c4518bb0a |
|
27-Jan-2017 |
Alex Vakulenko <avakulenko@google.com> |
Add DaydreamVR native libraries and services
Upstreaming the main VR system components from master-dreamos-dev
into goog/master.
Bug: None
Test: `m -j32` succeeds. Sailfish boots and basic_vr sample app works
Change-Id: I853015872afc443aecee10411ef2d6b79184d051
/frameworks/native/libs/vr/libpdx_uds/client_channel_factory.cpp
|