Re: [PATCH 0/3] use C11 memory model GCC builtin atomics

[Tyler Retzlaff <roretzla@linux.microsoft.com>]

On Wed, May 24, 2023 at 09:05:08AM -0700, Tyler Retzlaff wrote:
> On Wed, May 24, 2023 at 02:51:50PM +0200, David Marchand wrote:
> > On Mon, Mar 27, 2023 at 4:30 PM Tyler Retzlaff
> > <roretzla@linux.microsoft.com> wrote:
> > >
> > > Replace the use of __sync_<op>_and_fetch and __sync_fetch_and_<op> atomics
> > > with GCC C11 memory model __atomic builtins.
> > >
> > > This series contributes to converging on standard atomics in 23.11 but is
> > > kept separate as there may be sensitivity to converting from __sync to the
> > > C11 memory model builtins.
> > 
> > - Looking at the patches, I thought the conversion was rather straightforward.
> > But this mention about "sensitivity" stopped me from merging.
> > Did I miss some risk with the changes of this series?
> > 
> > 
> > >
> > > Tyler Retzlaff (3):
> > >   bus/vmbus: use C11 memory model GCC builtin atomics
> > >   crypto/ccp: use C11 memory model GCC builtin atomics
> > >   eal: use C11 memory model GCC builtin atomics
> > >
> > >  drivers/bus/vmbus/vmbus_channel.c    |  2 +-
> > >  drivers/crypto/ccp/ccp_dev.c         |  6 ++++--
> > >  lib/eal/include/generic/rte_atomic.h | 32 ++++++++++++++++----------------
> > >  3 files changed, 21 insertions(+), 19 deletions(-)
> > 
> > 
> > - I noticed that the vhost library has been providing an internal
> > wrapper for some __sync atomic with older GCC.
> > Some details are in the commitlog c16915b87109 ("vhost: improve dirty
> > pages logging performance").
> > 
> > Could it affect the existing legacy API performance?
> 
> Yes.
> 
> gcc documents that you can replace __sync_<op> with __atomic_<op> using
> SEQ_CST ordering.
> 
> When the __atomic_<op> builtins were initially introduced they generated
> sub-optimal (you can interpret as slower) codegen relative to the
> existing __sync_<op> builtins which was fixed in later gcc releases.
> 
> I do not know the actual version of gcc, but the commit you reference
> indicates GCC_VERSION < 70100 is that boundary.
> 
> I (perhaps incorrectly) assumed that if the CI performance tests didn't
> indicate a regression that the replacement of the remaining and minimal
> use of the legacy API would have negligable impact.
> 
> If this is a bad assumption or there are concerns, I could update the series
> to do the conditional __sync vs __atomic throughout.
> 
> Let me know how you'd like to proceed.

Anything further here or want to keep it as is?

> 
> Thanks!
> 
> > 
> > -- 
> > David Marchand