On Tue, Nov 17, 2020 at 04:15:12AM +0100, Mateusz Guzik wrote:
> On 11/17/20, Konstantin Belousov <kostik...@gmail.com> wrote:
> > On Mon, Nov 16, 2020 at 03:09:19AM +0000, Mateusz Guzik wrote:
> >> Author: mjg
> >> Date: Mon Nov 16 03:09:18 2020
> >> New Revision: 367713
> >> URL: https://svnweb.freebsd.org/changeset/base/367713
> >>
> >> Log:
> >> select: replace reference counting with memory barriers in selfd
> >>
> >> Refcounting was added to combat a race between selfdfree and
> >> doselwakup,
> >> but it adds avoidable overhead.
> >>
> >> selfdfree detects it can free the object by ->sf_si == NULL, thus we
> >> can
> >> ensure that the condition only holds after all accesses are completed.
> >>
> >> Modified:
> >> head/sys/kern/sys_generic.c
> >>
> >> Modified: head/sys/kern/sys_generic.c
> >> ==============================================================================
> >> --- head/sys/kern/sys_generic.c Sun Nov 15 22:49:28 2020
> >> (r367712)
> >> +++ head/sys/kern/sys_generic.c Mon Nov 16 03:09:18 2020
> >> (r367713)
> >> @@ -156,7 +156,6 @@ struct selfd {
> >> struct mtx *sf_mtx; /* Pointer to selinfo mtx. */
> >> struct seltd *sf_td; /* (k) owning seltd. */
> >> void *sf_cookie; /* (k) fd or pollfd. */
> >> - u_int sf_refs;
> >> };
> >>
> >> MALLOC_DEFINE(M_SELFD, "selfd", "selfd");
> >> @@ -1704,16 +1703,17 @@ static void
> >> selfdfree(struct seltd *stp, struct selfd *sfp)
> >> {
> >> STAILQ_REMOVE(&stp->st_selq, sfp, selfd, sf_link);
> >> - if (sfp->sf_si != NULL) {
> >> + /*
> >> + * Paired with doselwakeup.
> >> + */
> >> + if (atomic_load_acq_ptr((uintptr_t *)&sfp->sf_si) != (uintptr_t)NULL) {
> > This could be != 0.
> >
> >> mtx_lock(sfp->sf_mtx);
> >> if (sfp->sf_si != NULL) {
> >> TAILQ_REMOVE(&sfp->sf_si->si_tdlist, sfp, sf_threads);
> >> - refcount_release(&sfp->sf_refs);
> >> }
> >> mtx_unlock(sfp->sf_mtx);
> >> }
> >> - if (refcount_release(&sfp->sf_refs))
> >> - free(sfp, M_SELFD);
> >> + free(sfp, M_SELFD);
> > What guarantees that doselwakeup() finished with sfp ?
> >
>
> Release semantics provided by atomic_store_rel_ptr -- it means the
> NULL store is the last access, neither CPU nor the compiler are going
> to reorder preceding loads/stores past it.
It only guarantees that if we observed NULL as the result of load.
_______________________________________________
svn-src-head@freebsd.org mailing list
https://lists.freebsd.org/mailman/listinfo/svn-src-head
To unsubscribe, send any mail to "svn-src-head-unsubscr...@freebsd.org"
