Userspace RCU 0.2.4

From: Mathieu Desnoyers <compudj@krystal.dyndns.org>
To: Pavel Machek <pavel@ucw.cz>
Cc: Pierre Habouzit <madcoder@debian.org>,
	Steven Munroe <munroesj@linux.vnet.ibm.com>,
	rp@svcs.cs.pdx.edu, linux-kernel@vger.kernel.org,
	Josh Triplett <josh@joshtriplett.org>,
	ltt-dev@lists.casi.polymtl.ca, Jon Bernard <jbernard@debian.org>,
	"Paul E. McKenney" <paulmck@linux.vnet.ibm.com>,
	Jan Blunck <jblunck@suse.de>
Subject: Userspace RCU 0.2.4
Date: Mon, 19 Oct 2009 19:59:05 -0400	[thread overview]
Message-ID: <20091019235905.GA11980@Krystal> (raw)
In-Reply-To: <20091018225234.GA9164@elf.ucw.cz>

* Pavel Machek (pavel@ucw.cz) wrote:
> On Sun 2009-10-18 18:02:43, Mathieu Desnoyers wrote:
> > * Pavel Machek (pavel@ucw.cz) wrote:
> > > On Thu 2009-10-15 13:40:54, Pierre-Marc Fournier wrote:
> > > > Josh Triplett wrote:
> > > > > 
> > > > > Even Debian has given up on real 386 systems at this point, primarily
> > > > > because system libraries like glibc have; 486 and better represents the
> > > > > bare minimum required at this point.  I don't know of any distributions
> > > > > supporting real 386 systems at this point, and doing so would represent
> > > > > a major undertaking.
> > > > > 
> > > > 
> > > > What about embedded systems? Anyone know if some 386 chips, perhaps even
> > > > in smp configurations, are still in use in those?
> > > 
> > > smp 386: definitely not.
> > 
> > Hrm, so for UP 386, I wonder what's the best approach.
> > 
> > One would be to encapsulate all write accesses to the RCU pointers. If
> > we detect that the architecture lacks cmpxchg, _all_ update operations
> > (rcu_assign_pointer, rcu_xchg_pointer and rcu_cmpxchg_pointer) would
> > have to use the signal-disabled+mutex fall-back.
> > 
> > Does it make sense ?
> 
> Yep, but it sounds expensive. Another option is to ignore the issue
> and see how many people still have 386s :-). Few  embedded  systems
> may be affected, but...

Well.. I just enhanced liburcu to fully support 386 SMP (even if
opinions seems to vary regarding its usefulness...) ;) It adds _no_
overhead whatsoever if building for i486+ or x86 64.

What I did is a complete "compatibility mode" for all uatomic_arch_x86.h
atomic operations (it's my own user-space reimplementation of the Linux
kernel atomic.h). It's in liburcu 0.2.4 (now released).

How it works:

config x86 64 or x86 32 > i386 :

#define to map directly to atomic operations.

config i386 :

dynamically detect the cpu id, caches it in "cas_avail" variable.
If cas_avail is -1 (unset) -> dynamically check, cache result.
If cas_avail is 1 -> use atomic operations.
If cas_avail is 0 -> use compatibility mode for _all_ uatomic
  write operations involving signal disabling and a mutex. Only
  uatomic_read is exempt from locking.

So it should be safe to access RCU pointers through the
rcu_cmpxchg/xchg/set/assign_pointer() and rcu_dereference() primitives.

I tried to force using the compatibility mode by changing the condition
in compat_arch_x86.c and building for i386 compatibility. It works fine
and passes the test_uatomic test cases. Passes the rcutorture test too.

Mathieu

> 
> 									Pavel
> -- 
> (english) http://www.livejournal.com/~pavelmachek
> (cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
> 
> _______________________________________________
> ltt-dev mailing list
> ltt-dev@lists.casi.polymtl.ca
> http://lists.casi.polymtl.ca/cgi-bin/mailman/listinfo/ltt-dev
> 

-- 
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68