Improve performance of small CursorWindows.
Currently each CursorWindow allocates a 2MiB ashmem region to store
data, but this ends up being quite wasteful since the majority of
windows only end up storing a small handful of rows/columns. In
addition, creating and mmap'ing these ashmem regions requires
acquiring the mmap semaphore in the kernel, which can significantly
impact P95/P99 metrics when the system is under heavy load.
To mitigate the issues described above, this change adjusts
CursorWindow to send small windows (under 16KiB in size) directly
inline in Parcel responses without requiring an ashmem region.
CursorWindows also offer to gracefully "inflate" themselves into an
ashmem region when filled with more than 16KiB of data. This
requires some bugfixes around alloc() call sites to ensure that any
pointers are converted to offsets during a potential inflation.
The benchmarks referenced below show the following improvements
after this change is applied:
* Small cursor (1 row): 36% performance improvement
* Medium cursor (100 rows): no difference
* Large cursor (10k rows): no difference
Bug: 169251528
Test: atest CtsDatabaseTestCases
Test: atest FrameworksCoreTests:android.database
Test: ./frameworks/base/libs/hwui/tests/scripts/prep_generic.sh little && atest CorePerfTests:android.database.CrossProcessCursorPerfTest
Change-Id: Ie0fd149299f9847bf59a39f2855ed201bca4cdf6
3 files changed
