Hello,
This set contains fixes for libbpf's AF_XDP support. For patches 1-3 please
have a look at commit messages, 4th patch needs a bit more discussion.
Patch 4 tries to address the issue of dangling xsksocks in case when there were
many instances of those running on a particular interface and one of them got
removed. The case is that we don't keep an information about how many entries
are currently used in eBPF maps, so there's a need for having an external
counter, or we could just be traversing the map via
bpf_map_get_next_key/bpf_map_lookup_elem syscall combination, but IMHO that's a
no-go since the maps used in libbpf's xsk get preallocated entries. The count
of entries is equal to number of HW queues present on a working interface,
which means many unnecessary operations as CPUs are getting more and more cores
and normally NICs are allocating HW queue per CPU core.
For xsk counter we could have for example additional entry in qidconf_map, but
that map is removed in Jonathan's patches, so that's not an option. The
resolution I gave a shot is to have a pinned map with a single entry. Further
reasoning is included in commit message of fourth patch.
Thanks!
Maciej Fijalkowski (4):
libbpf: fill the AF_XDP fill queue before bind() call
libbpf: check for channels.max_{t,r}x in xsk_get_max_queues
libbpf: move xdp program removal to libbpf
libbpf: don't remove eBPF resources when other xsks are present
samples/bpf/xdpsock_user.c | 48 ++++--------------
tools/lib/bpf/xsk.c | 124 +++++++++++++++++++++++++++++++++++++++------
tools/lib/bpf/xsk.h | 1 +
3 files changed, 120 insertions(+), 53 deletions(-)
--
2.16.1
