Re: [PATCH v3 06/13] epoll: introduce helpers for adding/removing events to uring

[Peter Zijlstra]

On Fri, May 31, 2019 at 04:28:36PM +0200, Roman Penyaev wrote:
> On 2019-05-31 14:53, Peter Zijlstra wrote:
> > On Fri, May 31, 2019 at 01:15:21PM +0200, Roman Penyaev wrote:
> > > On 2019-05-31 11:56, Peter Zijlstra wrote:
> > > > On Thu, May 16, 2019 at 10:58:03AM +0200, Roman Penyaev wrote:
> > 
> > > > > +		i = __atomic_fetch_add(&ep->user_header->tail, 1,
> > > > > +				       __ATOMIC_ACQUIRE);
> > > >
> > > > afaict __atomic_fetch_add() does not exist.
> > > 
> > > That is gcc extension.  I did not find any API just to increment
> > > the variable atomically without using/casting to atomic.  What
> > > is a proper way to achieve that?
> > 
> > That's C11 atomics, and those shall not be used in the kernel. For one
> > they're not available in the minimally required GCC version (4.6).
> > 
> > The proper and only way is to use atomic_t, but also you cannot share
> > atomic_t with userspace.
> 
> Yes, that what I tried to avoid choosing c11 extension.
> 
> > 
> > The normal way of doing something like this is to have a kernel private
> > atomic_t and copy the value out to userspace using smp_store_release().
> 
> Since this path is lockless unfortunately that won't work.  So seems
> the only way is to do one more cmpxchg (sigh) or give up and take a
> look (sad sigh).

Of course it works; most of the extant buffers already have a shadow
tail and do the update to userspace with a store-release.