svn commit: r334708 - head/sys/kern

[Warner Losh]

On Wed, Jun 6, 2018 at 8:57 AM, Justin Hibbits <jhibbits at freebsd.org> wrote:

> Author: jhibbits
> Date: Wed Jun  6 12:57:11 2018
> New Revision: 334708
> URL: https://svnweb.freebsd.org/changeset/base/334708
>
> Log:
>   Add a memory barrier after taking a reference on the vnode holdcnt in
> _vhold
>
>   This is needed to avoid a race between the VNASSERT() below, and another
>   thread updating the VI_FREE flag, on weakly-ordered architectures.
>
>   On a 72-thread POWER9, without this barrier a 'make -j72 buildworld'
> would
>   panic on the assert regularly.
>
>   It may be possible to use a weaker barrier, and I'll investigate that
> once
>   all stability issues are worked out on POWER9.
>
> Modified:
>   head/sys/kern/vfs_subr.c
>
> Modified: head/sys/kern/vfs_subr.c
> ============================================================
> ==================
> --- head/sys/kern/vfs_subr.c    Wed Jun  6 10:46:24 2018        (r334707)
> +++ head/sys/kern/vfs_subr.c    Wed Jun  6 12:57:11 2018        (r334708)
> @@ -2807,6 +2807,9 @@ _vhold(struct vnode *vp, bool locked)
>         CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
>         if (!locked) {
>                 if (refcount_acquire_if_not_zero(&vp->v_holdcnt)) {
> +#if !defined(__amd64__) && !defined(__i386__)
> +                       mb();
> +#endif
>                         VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
>                             ("_vhold: vnode with holdcnt is free"));
>                         return;
>
>
So why isn't the refcount_acquire() enough?

Warner