This is the mail archive of the
mailing list for the glibc project.
Re: TSX lock elision for glibc v2
On Wed, 2013-01-16 at 10:30 -0800, Andi Kleen wrote:
> Torvald Riegel <email@example.com> writes:
> > On Fri, 2013-01-11 at 23:29 +0100, Andi Kleen wrote:
> >> > What about the internal locks used in the condvar and rwlock
> >> > implementations? We might perhaps be able to build better concurrent
> >> x86 rwlocks don't use locks, and I already elide then.
> > sysdeps/unix/sysv/linux/i386/i486/pthread_rwlock_rdlock.S lines 46-76,
> > for example. Are we looking at the same code?
> This is already elided in my patchkit. Please see patch 6/10.
No, you don't. What you do is doing lock elision for the whole duration
during which the program holds the readlock. Which is fine, but won't
work if you do I/O in the program's critical section or such.
Therefore, I asked whether you had looked at eliding the internal locks,
so doing elision during the critical sections _within_ the rwlock calls.
As I wrote in previous emails, this might need changes to the internal
locking / synchronization scheme to be worthwhile. But that's why I