On Tue, 4 Jun 2019 at 17:07, Maciej Fijalkowski
<maciejromanfijalkowski@xxxxxxxxx> wrote:
>
> On Tue, 4 Jun 2019 10:08:03 +0200
> Björn Töpel <bjorn.topel@xxxxxxxxx> wrote:
>
> > On 2019-06-03 15:19, Maciej Fijalkowski wrote:
> > > In case where multiple xsk sockets are attached to a single interface
> > > and one of them gets detached, the eBPF maps and program are removed.
> > > This should not happen as the rest of xsksocks are still using these
> > > resources.
> > >
> > > In order to fix that, let's have an additional eBPF map with a single
> > > entry that will be used as a xsks count. During the xsk_socket__delete,
> > > remove the resources only when this count is equal to 0. This map is
> > > not being accessed from eBPF program, so the verifier is not associating
> > > it with the prog, which in turn makes bpf_obj_get_info_by_fd not
> > > reporting this map in nr_map_ids field of struct bpf_prog_info. The
> > > described behaviour brings the need to have this map pinned, so in
> > > case when socket is being created and the libbpf detects the presence of
> > > bpf resources, it will be able to access that map.
> > >
> >
> > This commit is only needed after #3 is applied, right? So, this is a way
> > of refcounting XDP socks?
>
> Yes, but as you pointed out it needs synchronization.
>
> >
> >
> > > Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@xxxxxxxxx>
> > > ---
> > > tools/lib/bpf/xsk.c | 59 +++++++++++++++++++++++++++++++++++++++++++++--------
> > > 1 file changed, 51 insertions(+), 8 deletions(-)
> > >
> > > diff --git a/tools/lib/bpf/xsk.c b/tools/lib/bpf/xsk.c
> > > index e28bedb0b078..88d2c931ad14 100644
> > > --- a/tools/lib/bpf/xsk.c
> > > +++ b/tools/lib/bpf/xsk.c
> > > @@ -44,6 +44,8 @@
> > > #define PF_XDP AF_XDP
> > > #endif
> > >
> > > +#define XSKS_CNT_MAP_PATH "/sys/fs/bpf/xsks_cnt_map"
> > > +
> > > struct xsk_umem {
> > > struct xsk_ring_prod *fill;
> > > struct xsk_ring_cons *comp;
> > > @@ -65,6 +67,7 @@ struct xsk_socket {
> > > int prog_fd;
> > > int qidconf_map_fd;
> > > int xsks_map_fd;
> > > + int xsks_cnt_map_fd;
> > > __u32 queue_id;
> > > char ifname[IFNAMSIZ];
> > > };
> > > @@ -372,7 +375,7 @@ static int xsk_get_max_queues(struct xsk_socket *xsk)
> > > static int xsk_create_bpf_maps(struct xsk_socket *xsk)
> > > {
> > > int max_queues;
> > > - int fd;
> > > + int fd, ret;
> > >
> > > max_queues = xsk_get_max_queues(xsk);
> > > if (max_queues < 0)
> > > @@ -392,6 +395,24 @@ static int xsk_create_bpf_maps(struct xsk_socket *xsk)
> > > }
> > > xsk->xsks_map_fd = fd;
> > >
> > > + fd = bpf_create_map_name(BPF_MAP_TYPE_ARRAY, "xsks_cnt_map",
> > > + sizeof(int), sizeof(int), 1, 0);
> > > + if (fd < 0) {
> > > + close(xsk->qidconf_map_fd);
> > > + close(xsk->xsks_map_fd);
> > > + return fd;
> > > + }
> > > +
> > > + ret = bpf_obj_pin(fd, XSKS_CNT_MAP_PATH);
> > > + if (ret < 0) {
> > > + pr_warning("pinning map failed; is bpffs mounted?\n");
> > > + close(xsk->qidconf_map_fd);
> > > + close(xsk->xsks_map_fd);
> > > + close(fd);
> > > + return ret;
> > > + }
> > > + xsk->xsks_cnt_map_fd = fd;
> > > +
> > > return 0;
> > > }
> > >
> > > @@ -456,8 +477,10 @@ static int xsk_lookup_bpf_maps(struct xsk_socket *xsk)
> > > close(fd);
> > > }
> > >
> > > + xsk->xsks_cnt_map_fd = bpf_obj_get(XSKS_CNT_MAP_PATH);
> > > err = 0;
> > > - if (xsk->qidconf_map_fd < 0 || xsk->xsks_map_fd < 0) {
> > > + if (xsk->qidconf_map_fd < 0 || xsk->xsks_map_fd < 0 ||
> > > + xsk->xsks_cnt_map_fd < 0) {
> > > err = -ENOENT;
> > > xsk_delete_bpf_maps(xsk);
> > > }
> > > @@ -467,17 +490,25 @@ static int xsk_lookup_bpf_maps(struct xsk_socket *xsk)
> > > return err;
> > > }
> > >
> > > -static void xsk_clear_bpf_maps(struct xsk_socket *xsk)
> > > +static void xsk_clear_bpf_maps(struct xsk_socket *xsk, long *xsks_cnt_ptr)
> > > {
> > > + long xsks_cnt, key = 0;
> > > int qid = false;
> > >
> > > bpf_map_update_elem(xsk->qidconf_map_fd, &xsk->queue_id, &qid, 0);
> > > bpf_map_delete_elem(xsk->xsks_map_fd, &xsk->queue_id);
> > > + bpf_map_lookup_elem(xsk->xsks_cnt_map_fd, &key, &xsks_cnt);
> > > + if (xsks_cnt)
> > > + xsks_cnt--;
> > > + bpf_map_update_elem(xsk->xsks_cnt_map_fd, &key, &xsks_cnt, 0);
> > > + if (xsks_cnt_ptr)
> > > + *xsks_cnt_ptr = xsks_cnt;
> >
> > This refcount scheme will not work; There's no synchronization between
> > the updates (cross process)!
>
> Ugh, shame. Let me fix this in v2 :(
>
Ok!
Hmm, there's no way of doing lock_xadds from bpf syscall, right? Any
thoughts how to do this? Maybe there's a need for a sock_ops prog for
XDP sockets...
Björn
> >
> > > }
> > >
> > > static int xsk_set_bpf_maps(struct xsk_socket *xsk)
> > > {
> > > int qid = true, fd = xsk->fd, err;
> > > + long xsks_cnt, key = 0;
> > >
> > > err = bpf_map_update_elem(xsk->qidconf_map_fd, &xsk->queue_id, &qid, 0);
> > > if (err)
> > > @@ -487,9 +518,18 @@ static int xsk_set_bpf_maps(struct xsk_socket *xsk)
> > > if (err)
> > > goto out;
> > >
> > > + err = bpf_map_lookup_elem(xsk->xsks_cnt_map_fd, &key, &xsks_cnt);
> > > + if (err)
> > > + goto out;
> > > +
> > > + xsks_cnt++;
> > > + err = bpf_map_update_elem(xsk->xsks_cnt_map_fd, &key, &xsks_cnt, 0);
> > > + if (err)
> > > + goto out;
> > > +
> >
> > Dito.
> >
> > > return 0;
> > > out:
> > > - xsk_clear_bpf_maps(xsk);
> > > + xsk_clear_bpf_maps(xsk, NULL);
> > > return err;
> > > }
> > >
> > > @@ -752,13 +792,18 @@ void xsk_socket__delete(struct xsk_socket *xsk)
> > > size_t desc_sz = sizeof(struct xdp_desc);
> > > struct xdp_mmap_offsets off;
> > > socklen_t optlen;
> > > + long xsks_cnt;
> > > int err;
> > >
> > > if (!xsk)
> > > return;
> > >
> > > - xsk_clear_bpf_maps(xsk);
> > > - xsk_delete_bpf_maps(xsk);
> > > + xsk_clear_bpf_maps(xsk, &xsks_cnt);
> > > + unlink(XSKS_CNT_MAP_PATH);
> > > + if (!xsks_cnt) {
> > > + xsk_delete_bpf_maps(xsk);
> > > + xsk_remove_xdp_prog(xsk);
> > > + }
> > >
> > > optlen = sizeof(off);
> > > err = getsockopt(xsk->fd, SOL_XDP, XDP_MMAP_OFFSETS, &off, &optlen);
> > > @@ -774,8 +819,6 @@ void xsk_socket__delete(struct xsk_socket *xsk)
> > >
> > > }
> > >
> > > - xsk_remove_xdp_prog(xsk);
> > > -
> > > xsk->umem->refcount--;
> > > /* Do not close an fd that also has an associated umem connected
> > > * to it.
> > >
>
