Message-ID: <CAJ8uoz1PSwJkpZ80YgPbSan19QPmt3gSmwvd7JT+epTF7z4phg@mail.gmail.com>
Date:   Fri, 26 Mar 2021 16:20:10 +0100
From:   Magnus Karlsson <magnus.karlsson@...il.com>
To:     "Loftus, Ciara" <ciara.loftus@...el.com>
Cc:     Network Development <netdev@...r.kernel.org>,
        bpf <bpf@...r.kernel.org>,
        "Karlsson, Magnus" <magnus.karlsson@...el.com>,
        Björn Töpel <bjorn@...nel.org>
Subject: Re: [PATCH bpf 2/3] libbpf: restore umem state after socket create failure

On Fri, Mar 26, 2021 at 3:56 PM Loftus, Ciara <ciara.loftus@...el.com> wrote:
>
> >
> > On Wed, Mar 24, 2021 at 3:46 PM Ciara Loftus <ciara.loftus@...el.com>
> > wrote:
> > >
> > > If the call to socket_create fails, the user may want to retry the
> > > socket creation using the same umem. Ensure that the umem is in the
> > > same state on exit if the call failed by restoring the _save pointers
> > > and not unmapping the set of umem rings if those pointers are non NULL.
> > >
> > > Fixes: 2f6324a3937f ("libbpf: Support shared umems between queues and
> > devices")
> > >
> > > Signed-off-by: Ciara Loftus <ciara.loftus@...el.com>
> > > ---
> > >  tools/lib/bpf/xsk.c | 29 ++++++++++++++++++-----------
> > >  1 file changed, 18 insertions(+), 11 deletions(-)
> > >
> > > diff --git a/tools/lib/bpf/xsk.c b/tools/lib/bpf/xsk.c
> > > index 443b0cfb45e8..ec3c23299329 100644
> > > --- a/tools/lib/bpf/xsk.c
> > > +++ b/tools/lib/bpf/xsk.c
> > > @@ -743,21 +743,23 @@ static struct xsk_ctx *xsk_get_ctx(struct
> > xsk_umem *umem, int ifindex,
> > >         return NULL;
> > >  }
> > >
> > > -static void xsk_put_ctx(struct xsk_ctx *ctx)
> > > +static void xsk_put_ctx(struct xsk_ctx *ctx, bool unmap)
> > >  {
> > >         struct xsk_umem *umem = ctx->umem;
> > >         struct xdp_mmap_offsets off;
> > >         int err;
> > >
> > >         if (--ctx->refcount == 0) {
> > > -               err = xsk_get_mmap_offsets(umem->fd, &off);
> > > -               if (!err) {
> > > -                       munmap(ctx->fill->ring - off.fr.desc,
> > > -                              off.fr.desc + umem->config.fill_size *
> > > -                              sizeof(__u64));
> > > -                       munmap(ctx->comp->ring - off.cr.desc,
> > > -                              off.cr.desc + umem->config.comp_size *
> > > -                              sizeof(__u64));
> > > +               if (unmap) {
> > > +                       err = xsk_get_mmap_offsets(umem->fd, &off);
> > > +                       if (!err) {
> > > +                               munmap(ctx->fill->ring - off.fr.desc,
> > > +                                      off.fr.desc + umem->config.fill_size *
> > > +                               sizeof(__u64));
> > > +                               munmap(ctx->comp->ring - off.cr.desc,
> > > +                                      off.cr.desc + umem->config.comp_size *
> > > +                               sizeof(__u64));
> > > +                       }
> > >                 }
> >
> > By not unmapping these rings we actually leave more state after a
> > failed socket creation. So how about skipping this logic (and
>
> In the case of the _save rings, the maps existed before the call to
> xsk_socket__create. They were created during xsk_umem__create.
> So we should preserve these maps in event of failure.
> I was using the wrong condition to trigger the unmap in v1 however.
> We should unmap 'fill' only if
>         umem->fill_save != fill
> I will update this in a v2.

Ahh, you are correct. There are two ways these rings can get allocated
so that has to be taken care of. Please ignore my comment.

> > everything below) and always unmap the rings at failure as before, but
> > we move the fill_save = NULL and comp_save = NULL from xsk_create_ctx
> > to the end of xsk_socket__create_shared just before the "return 0"
> > where we know that the whole operation has succeeded. This way the
>
> I think moving these still makes sense and will add this in the next rev.
>
> Thanks for the feedback and suggestions!
>
> Ciara
>
> > mappings would be redone during the next xsk_socket__create and if
> > someone decides not to retry (for some reason) we do not leave two
> > mappings behind. Would simplify things. What do you think?
>
> >
> > >
> > >                 list_del(&ctx->list);
> > > @@ -854,6 +856,9 @@ int xsk_socket__create_shared(struct xsk_socket
> > **xsk_ptr,
> > >         struct xsk_socket *xsk;
> > >         struct xsk_ctx *ctx;
> > >         int err, ifindex;
> > > +       struct xsk_ring_prod *fsave = umem->fill_save;
> > > +       struct xsk_ring_cons *csave = umem->comp_save;
> > > +       bool unmap = !fsave;
> > >
> > >         if (!umem || !xsk_ptr || !(rx || tx))
> > >                 return -EFAULT;
> > > @@ -1005,7 +1010,9 @@ int xsk_socket__create_shared(struct xsk_socket
> > **xsk_ptr,
> > >                 munmap(rx_map, off.rx.desc +
> > >                        xsk->config.rx_size * sizeof(struct xdp_desc));
> > >  out_put_ctx:
> > > -       xsk_put_ctx(ctx);
> > > +       umem->fill_save = fsave;
> > > +       umem->comp_save = csave;
> > > +       xsk_put_ctx(ctx, unmap);
> > >  out_socket:
> > >         if (--umem->refcount)
> > >                 close(xsk->fd);
> > > @@ -1071,7 +1078,7 @@ void xsk_socket__delete(struct xsk_socket *xsk)
> > >                 }
> > >         }
> > >
> > > -       xsk_put_ctx(ctx);
> > > +       xsk_put_ctx(ctx, true);
> > >
> > >         umem->refcount--;
> > >         /* Do not close an fd that also has an associated umem connected
> > > --
> > > 2.17.1
> > >

Powered by blists - more mailing lists