Implement an epoll-backed fdevent_context, that reduces overhead when
we're polling many file descriptors. FdeventTest.smoke goes from ~5.2s
to ~3.3s when run on the host (after this patch's modification to change
it from chaining 10 file descriptors together to 512).
Test: adb_test on host
Test: adbd_test on blueline
Test: test_adb.py
Test: test_device.py
Change-Id: Iacf0093aa7bebea31e447c2cb012af72d8c3297e
As a side effect, delete FDE_ACTIVE, which was always set on every
fdevent, and FDE_PENDING, which was an internal implementation detail.
This patch removes spin detection, which will be reimplemented
separately later.
Test: adb_test on host
Test: adbd_test on blueline
Change-Id: I40be3504ce03c4fae5e071fa018542a051b7511d
It fixes a compilation issue:
/fdevent/fdevent_poll.cpp -o core/adb/fdevent/fdevent_poll.cpp.o
In file included from core/adb/fdevent/fdevent_poll.cpp:20:
In file included from core/adb/fdevent/fdevent_poll.h:29:
core/adb/fdevent/fdevent.h:106:10: error: no template named 'atomic' in namespace 'std'
std::atomic<bool> terminate_loop_ = false;
~~~~~^
Change-Id: I5305a69a50e2638f514c5a7d47a9612b8ee38c69
All fdevents must be created by fdevent_create/fdevent_context::Create
now, so this flag is meaningless.
Test: none
Change-Id: I8489afad07bdb267ddfdb0dbb25d40d3a62f55f6
Make the run queue logic reusable between implementations of fdevent by
moving it to the abstract base class.
Test: adb_test
Change-Id: If2f72e3ddc8007304bca63aa75446fa117267b25
