[<prev] [next>] [day] [month] [year] [list]
Message-ID: <1584600558.1675.1547658881995.JavaMail.zimbra@efficios.com>
Date: Wed, 16 Jan 2019 12:14:41 -0500 (EST)
From: Mathieu Desnoyers <mathieu.desnoyers@...icios.com>
To: Daniel Colascione <dancol@...gle.com>
Cc: Paul Turner <pjt@...gle.com>,
"Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>,
Andy Lutomirski <luto@...capital.net>,
Boqun Feng <boqun.feng@...il.com>,
Peter Zijlstra <peterz@...radead.org>,
Thomas Gleixner <tglx@...utronix.de>,
linux-kernel <linux-kernel@...r.kernel.org>
Subject: Adaptative busy spinning with rseq
Hi!
A follow up on the Edinburgh discussion I had with Daniel: I did a prototype
implementation of adaptative busy spinning with rseq. It only uses current
upstream rseq features. It's not optimized at this stage (and I don't have time to
work more on it at the moment), and it's only x86-64, but I'm throwing the code out
there for feedback:
https://github.com/compudj/rseq-test/blob/adapt-lock/test-rseq-adaptative-lock.c
The trick here is to rely on the Zero Flag being invariant for a sub-section of
the rseq critical sections, and use it to figure out if it has been aborted
after cmpxchg has succeeded. Testing the Zero Flag on abort basically removes the
rseq requirement that the very last instruction of a rseq critical section needs
to be the "commit" instruction, allowing us to cover an entire loop within a rseq
critical section.
Feedback is welcome!
Thanks,
Mathieu
--
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com
Powered by blists - more mailing lists