libxdp: Allow creation of xsk with exclusive umem without CAP_NET_RAW #338
Conversation
|
The Ci failure seems unrelated. |
igsilya
force-pushed
the
tmp-non-privileged
branch
from
db73e90
to
c7e149e
Compare
Adding a new API xsk_umem__create_with_fd() that accepts an extra file descriptor of already open AF_XDP socket. This function can be used as a substitute of a regular xsk_umem__create() in case where process doesn't have CAP_NET_RAW privileges. Privileged process may open a socket beforehand and send the file descriptor via UNIX domain socket or pass it during exec(). Subsequent xsk_socket__create() will re-use this socket. No other operations require privileges in case the program load is inhibited. This change allows use of AF_XDP with libxdp by processes with no extra privileges. xsk_socket__create_shared() requires opening a new socket for each queue after the first one, so it will require more work in order to create a pretty API. Not doing that right now. The target use-case is AF_XDP network backend in QEMU that doesn't need shared umem, but requires running without any privileges. Signed-off-by: Ilya Maximets <i.maximets@ovn.org>
|
Alright, LGTM. I'll merge this now, but also take the opportunity to point out that we really don't have any selftest coverage for any of the AF_XDP code; care to look into that? |
Looks like test_xsk_refcnt.c is covering some of the basic functionality. I wrote a mall C program to test ability to create sockets from a non-privileged process. I guess, I can try to turn it into a test. It has a dependency on a libcap-ng though, but I guess we could try to detect the library during configuration and skip the test if not present? What do you think? |
|
That would be awesome! We already do plenty of feature checking and skipping of tests in various places, so don't think that's a problem :) |
Adding a new API xsk_umem__create_with_fd() that accepts an extra file descriptor of already open AF_XDP socket. This function can be used as a substitute of a regular xsk_umem__create() in case where process doesn't have CAP_NET_RAW privileges. Privileged process may open a socket beforehand and send the file descriptor via UNIX domain socket or pass it during exec(). Subsequent xsk_socket__create() will re-use this socket. No other operations require privileges in case the program load is inhibited.
This change allows use of AF_XDP with libxdp by processes with no extra privileges.
xsk_socket__create_shared() requires opening a new socket for each queue after the first one, so it will require more work in order to create a pretty API. Not doing that right now. The target use-case is AF_XDP network backend in QEMU that doesn't need shared umem, but requires running without any privileges.