[asterisk-dev] [Code Review] 3405: Add ast_spinlock capability

[George Joseph]

-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviewboard.asterisk.org/r/3405/
-----------------------------------------------------------

(Updated April 23, 2014, 3:06 p.m.)


Status
------

This change has been marked as submitted.


Review request for Asterisk Developers, Corey Farrell, Joshua Colp, and rmudgett.


Changes
-------

Committed in revision 412976


Bugs: ASTERISK-23553
    https://issues.asterisk.org/jira/browse/ASTERISK-23553


Repository: Asterisk


Description
-------

Now ready for prime time...

This patch adds support for spinlocks in Asterisk.

There are cases in Asterisk where it might be desirable to lock a short critical code section but not incur the context switch and yield penalty of a mutex or rwlock.  The primary spinlock implementations execute exclusively in userspace and therefore don't incur those penalties.  Spinlocks are NOT meant to be a general replacement for mutexes.  They should be used only for protecting short blocks of critical code such as simple compares and assignments.  Operations that may block, hold a lock, or cause the thread to give up it's timeslice should NEVER be attempted in a spinlock.

The first use case for spinlocks is in astobj2 - internal_ao2_ref.  Currently the manipulation of the reference counter is done with an ast_atomic_fetchadd_int which works fine.  When weak reference containers are introduced however, there's an additional comparison and assignment that'll need to be done while the lock is held.  A mutex would be way too expensive here, hence the spinlock.  Given that lock contention in this situation would be infrequent, the overhead of the spinlock is only a few more machine instructions than the current ast_atomic_fetchadd_int call.

7 implementations were tested on various i366, x86_64, ARM and Sparc platforms running Linux, FreeBSD, OSX and Solaris.  The test suite and results can be found here... https://github.com/gtjoseph/spintest

Summary...

Tested various spinlock implementations. 

+ GCC Atomics      		gcc_atomics
+ x86 assembly     		gas_x86
+ ARM assembly     		gas_arm
+ Sparc assembly     		gas_sparc
+ OSX Atomics      		osx_atomics
+ pthreads spinlock		pthread_spinlock
+ pthreads mutex   		pthread_mutex
+ pthreads adaptive mutex       pthread_mutex_adaptive_np

Not all implementations were available on all platforms.  For instance, the gas_x86 implementation won't be available on an arm or sparc platform.

Each implementation was tested with the same locking scenario which is a short block of compares and assignments representing the most anticipated Asterisk use cases.  The test case was run in a tight loop of 25,000,000 iterations executed in parallel by 1..5 simultaneous threads.

The test suite was run on the following platforms:

+ Darwin-OSX-x86_64-2core
+ FreeBSD-BSD-amd64-4core
+ FreeBSD-BSD-i386-1core
+ Linux-CentOS-i686-1core
+ Linux-Debian-i686-4core
+ Linux-Debian-x86_64-4core
+ Linux-Fedora-armv7l-1core
+ Linux-Fedora-i686-1core
+ Linux-Fedora-x86_64-12core
+ Linux-Fedora-x86_64-1core
+ Linux-Fedora-x86_64-2core
+ Linux-Fedora-x86_64-4core
+ Linux-Fedora-x86_64-8core
+ Linux-Gentoo-sparc64-32core
+ SunOS-Solaris-i86pc-4core

For each platform tested, the real, user and system times were captured in csv form and the final real and system times graphed.

Observations:

+ The GCC Atomics implementation was available on all platforms and generally had the best performance.
+ The native assembly implementations generally performed on par with the GCC Atomics implementation.
+ The pthread_spinlock implementation was not available on Darwin but performed well on the other platforms.
+ The OSX Atomics implementation performed well on Darwin.
+ The pthread_mutex_adaptive implementation was not available on all platforms and it's performance was noticeably worse on the supported platforms.
+ The pthread_mutex implementation was available on all platforms but clearly showed the effects of context switching (high sys time) as lock contention grew.

Conclusions:

+ GCC Atomics should be the first implementation choice.
+ The native assembly implementations should be the second choices.
+ The pthread_spinlock implementation should be the third choice on non-Darwin platforms.
+ OSX Atomics should be the third choice on Darwin.
+ The pthread_mutex_adaptive implementation should be eliminated.
+ The pthread_mutex implementation should be the implementation of last resort.
+ If none of the implementations are available, a compilation #error will be raised.


Diffs
-----

  branches/12/include/asterisk/spinlock.h PRE-CREATION 
  branches/12/include/asterisk/autoconfig.h.in 412900 
  branches/12/configure.ac 412900 
  branches/12/configure UNKNOWN 

Diff: https://reviewboard.asterisk.org/r/3405/diff/


Testing
-------

See above.


Thanks,

George Joseph

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.digium.com/pipermail/asterisk-dev/attachments/20140423/75cbcc14/attachment.html>