Message-Id: <20171026140750.119265-1-maco@android.com>
Date:   Thu, 26 Oct 2017 16:07:44 +0200
From:   Martijn Coenen <maco@...roid.com>
To:     gregkh@...uxfoundation.org, john.stultz@...aro.org,
        tkjos@...gle.com, arve@...roid.com, sherryy@...roid.com,
        tglx@...utronix.de, peterz@...radead.org, amit.pundir@...aro.org
Cc:     linux-kernel@...r.kernel.org, devel@...verdev.osuosl.org,
        maco@...gle.com, Martijn Coenen <maco@...roid.com>
Subject: [PATCH v3 0/6] ANDROID: binder: RT priority inheritance

Changes since v2 [1]:
- All patches in v2 not related to priority inheritance were merged,
  and hence removed from this series
- Fixed using the wrong mask in node scheduler policy calculation,
  originally reported by Ganesh Mahendran <opensource.ganesh@...il.com>
- Fixed using an uninitialized value for desired_prio,
  originally reported by Ganesh Mahendran <opensource.ganesh@...il.com>
  
Changes since v1 [2]:
- added more detailed commit messages and comments to the priority
  inheritance patches, including rationale for not using
  schet_setscheduler() directly, or rt_mutex prio inheritance.
  No functional changes.

[1]: https://lkml.kernel.org/r/20170831080430.118765-1-maco@android.com
[2]: https://lkml.kernel.org/r/20170825093335.100892-1-maco@android.com

---

This patch series introduces support for priority inheritance of real-time
scheduling policies in binder. With the introduction of Android Treble,
functionality that used to be in a single process is now split over two or
more processes, which communicate using binder IPC. For latency sensitive
operations such as sensor events, Bluetooth audio and rendering, inheriting
the (real-time) priority of the caller is crucial to meet requirements.

The implementation in this series directly calls into the scheduler to
modify priorities, since I haven't found a way to make this work correctly
with rt_mutex or other existing priority inheritance mechanisms. The main
reasons a PI rt_mutex doesn't work well are the following:
1) Binder supports asynchronous transactions, where a caller isn't blocked
   on a result; therefore, the caller also couldn't block on a rt_mutex.
2) Binder supports the concept of 'node priority', where the priority of a
   call is not determined by the caller, but by the endpoint the caller
   calls in to.
3) There may not necessarily be any binder threads waiting to handle a
   transaction, so the caller doesn't always have a thread to change the
   priority of; instead, the first thread to pick up the work changes its
   own priority.
4) rt_mutex doesn't support non-RT policies (though a patch was sent to
   LKML once to address this).   

More details in the patches themselves. I have found the current approach
to be reliable, but I'm happy to look into suggestions to make this work
with rt_mutex, or use other infrastructure.

All patches have already been reviewed by Android engineers and are
merged in Android's common kernel trees.

Martijn Coenen (6):
  ANDROID: binder: add support for RT prio inheritance.
  ANDROID: binder: add min sched_policy to node.
  ANDROID: binder: improve priority inheritance.
  ANDROID: binder: add RT inheritance flag to node.
  ANDROID: binder: don't check prio permissions on restore.
  ANDROID: binder: Add tracing for binder priority inheritance.

 drivers/android/binder.c            | 294 ++++++++++++++++++++++++++++++++----
 drivers/android/binder_trace.h      |  24 +++
 include/uapi/linux/android/binder.h |  49 +++++-
 3 files changed, 334 insertions(+), 33 deletions(-)

-- 
2.15.0.rc2.357.g7e34df9404-goog

Powered by blists - more mailing lists