Hand-rolled linked list for Parcel performance.

While recently looking at pprof data, we noticed that Parcel.obtain()
and recycle() were some of the most heavily used methods inside the
system process.  On the surface, this makes sense because these
methods are typically invoked at least twice for every Binder call.

Internally these methods had been maintaining a pool of cached Parcel
objects to avoid GC pressure, but unfortunately it was using an
array which resulted in O(n) scanning under heavy load, increasing
lock contention encountered by all Binder calls.

This change greatly reduces that contention by using a hand-rolled
linked list approach, mirroring android.os.Message.  For a 1-thread
benchmark, this new approach has almost 2x throughput, and for a
16-thread benchmark it has almost 8x throughput.

As part of making this change we evaluated several approaches,
including using pure-GC with no pooling, a single AtomicReference,
and a pool of several AtomicReferences.  To measure these we wrote
ParcelPoolBenchmark which simulates various levels of Binder load
using 1, 4 and 16 threads.  Below are the relative benchmark results
compared to the previous approach before this CL:

              1 thread   4 threads   16 threads
Pure GC        131.74%      32.76%       43.90%
Single AR       95.22%      25.54%       13.66%
Pooled AR       57.65%      16.21%       11.55%
Linked list     52.66%      18.06%       12.55%

On balance, the linked list approach performs well across the board,
and we bias towards it over the two AtomicReference approaches since
it performs slightly better on the single-threaded case, which is
the most representative of typical Binder load across all processes.

Bug: 165032569
Test: ./frameworks/base/libs/hwui/tests/scripts/prep_generic.sh little && atest CorePerfTests:android.os.ParcelObtainPerfTest
Change-Id: I190b1c8f7fd59855c3c2d36032512279691e2c04

0 files changed, 0 insertions, 0 deletions