This is the mail archive of the mailing list for the glibc project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [RFC][PATCH 0/2] Make sparcv8 work again on cas enabled hardware

On Wed, 2016-11-02 at 22:52 -0400, David Miller wrote:
> From: Torvald Riegel <>
> Date: Wed, 02 Nov 2016 23:33:03 +0100
> > On Wed, 2016-11-02 at 11:32 -0400, David Miller wrote:
> >> From: Torvald Riegel <>
> >> > Or do you intend to write sparc-specific versions of all the concurrent
> >> > data structures that are process-shared?
> >> 
> >> This would be necessary anyways, if we have two modes.  One that does
> >> the pure-userland code path and one that does the kernel helper code
> >> path.
> > 
> > All the other archs that use a kernel helper for CAS don't need it.  If
> > you can call the helper in the atomic operations, you won't need a new
> > algorithm except if you wanted to optimize the generic one.
> > 
> >> Furthermore, sparc specific versions are needed in any case since we
> >> have the v9 detection even in the v8 libraries.  Look at all of the
> >> code that checks for v9 in the dl_hwcap mask when deciding which
> >> atomic operation to use.
> > 
> > Or are you talking about the implementation of the atomic operations?
> Just as the "are we running on a v9 chip" test is a run-time one,

Is there any difference between the additional CAS on a v8 and the CAS
on a v9?  If there should be none (eg, same instruciton encoding etc.),
we wouldn't need a runtime check for this, would we?

> whether we are running on a kernel with kernel CAS simulation support
> will be run time code path check as well.

That depends on whether we want to support sparc HW that does have a
CAS.  It's still not clear to me whether this is a goal, and if it's a
goal, whether it's a goal for today or for some time in the future.

> This is why we'll need sparc specific versions of the primitives,

Which primitives are you talking about?  The atomic operations in
atomic-machine.h / atomic.h, or the synchronization primitives in nptl/?

> and
> why it would have been the more optimal if the primitives were
> abstracted to the point where we didn't have to duplicate so much
> stuff privately just to pull this off.

I can't follow.  What do you mean precisely?

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]