On Wed, Jun 24, 2026 at 2:33 PM Kuniyuki Iwashima <[email protected]> wrote: > > On Wed, Jun 24, 2026 at 2:26 PM Michal Luczaj <[email protected]> wrote: > > > > On 6/24/26 22:01, Willem de Bruijn wrote: > > > Jakub Sitnicki wrote: > > >> On Tue, Jun 23, 2026 at 08:03 PM +02, Michal Luczaj wrote: > > >>> UDP sockets get SOCK_RCU_FREE set when (auto-)bound. This means > > >>> sk_is_refcounted(unbound) = true, while sk_is_refcounted(bound) = false. > > >>> > > >>> Because sockmap accepts unbound UDP sockets, a BPF program can > > >>> increment a > > >>> socket's refcount via lookup. If the socket is subsequently bound, the > > >>> transition from unbound to bound causes bpf_sk_release() to skip the > > >>> decrement of the refcount, causing a memory leak. > > >>> > > >>> unreferenced object 0xffff88810bc2eb40 (size 1984): > > >>> comm "test_progs", pid 2451, jiffies 4295320596 > > >>> hex dump (first 32 bytes): > > >>> 7f 00 00 01 7f 00 00 01 d2 04 1b b7 04 d2 00 00 ................ > > >>> 02 00 01 40 00 00 00 00 00 00 00 00 00 00 00 00 ...@............ > > >>> backtrace (crc bdee079d): > > >>> kmem_cache_alloc_noprof+0x557/0x660 > > >>> sk_prot_alloc+0x69/0x240 > > >>> sk_alloc+0x30/0x460 > > >>> inet_create+0x2ce/0xf80 > > >>> __sock_create+0x25b/0x5c0 > > >>> __sys_socket+0x119/0x1d0 > > >>> __x64_sys_socket+0x72/0xd0 > > >>> do_syscall_64+0xa1/0x5f0 > > >>> entry_SYSCALL_64_after_hwframe+0x76/0x7e > > >>> > > >>> Maintain balanced refcounts across sk lookup/release: (re-)set > > >>> SOCK_RCU_FREE on proto update to treat the socket (whether bound or > > >>> unbound) as not requiring a refcount increment on (a RCU protected) > > >>> lookup. > > >>> > > >>> Fixes: 0c48eefae712 ("sock_map: Lift socket state restriction for > > >>> datagram sockets") > > >>> Signed-off-by: Michal Luczaj <[email protected]> > > >>> --- > > >>> Note: this issue is related to commit 67312adc96b5 ("bpf: reject > > >>> unhashed > > >>> sockets in bpf_sk_assign"). > > >>> --- > > >>> net/ipv4/udp_bpf.c | 3 +++ > > >>> 1 file changed, 3 insertions(+) > > >>> > > >>> diff --git a/net/ipv4/udp_bpf.c b/net/ipv4/udp_bpf.c > > >>> index ad57c4c9eaab..970327b59582 100644 > > >>> --- a/net/ipv4/udp_bpf.c > > >>> +++ b/net/ipv4/udp_bpf.c > > >>> @@ -173,6 +173,9 @@ int udp_bpf_update_proto(struct sock *sk, struct > > >>> sk_psock *psock, bool restore) > > >>> if (sk->sk_family == AF_INET6) > > >>> udp_bpf_check_v6_needs_rebuild(psock->sk_proto); > > >>> > > >>> + /* Treat all sockets as non-refcounted, regardless of binding > > >>> state. */ > > >>> + sock_set_flag(sk, SOCK_RCU_FREE); > > >>> + > > >>> sock_replace_proto(sk, &udp_bpf_prots[family]); > > >>> return 0; > > >>> } > > >> > > >> There is a side effect that an unhashed (unbound) UDP socket can now be > > >> selected in sk_lookup with bpf_sk_assign. > > > > > > The commit does mention a related fix, beneath the ---, commit > > > 67312adc96b5 ("bpf: reject unhashed sockets in bpf_sk_assign"). > > > That fixes a similar issue by exactly disallowing this: > > > > > > Fix the problem by rejecting unhashed sockets in bpf_sk_assign(). > > > This matches the behaviour of __inet_lookup_skb which is ultimately > > > the goal of bpf_sk_assign(). > > > > > > So .. > > > > > >> Though perhaps that's for the > > >> better because TC bpf_sk_assign doesn't reject non-refcounted UDP > > >> sockets either, so we would have both socket dispatch sites behave the > > >> same way. > > > > > > .. there are two conflicting types of consistency here? Consistent with > > > __inet_lookup_skb or the TC bpf hook. Of those the first is the more > > > canonical. > > > > > >> Also, with this patch, if we insert & remove an unhashed UDP socket > > >> into/from a sockmap, we end up with an unhashed non-refcounted UDP > > >> socket. Not entirely sure if that is actually a problem or not. > > >> > > >> Willem, what is your take on having unhashed non-refcoted UDP sockets? > > > > > > I don't immediately see a problem, but I'm not an expert on SOCK_RCU_FREE. > > > > Perhaps it's worth mentioning that unhashed non-refcounted UDP socket is > > already possible: first auto-bind via connect(AF_INET) (which also sets > > SOCK_RCU_FREE), then unhash via connect(AF_UNSPEC). > > Setting SOCK_RCU_FREE itself should not cause a problem, but I think > we should take a step back. > > AFAIU, 0c48eefae712 was to allow putting AF_UNIX SOCK_DGRAM sockets > into sockmap, not to allow using unconnected UDP sockets in sk_lookup etc. > > Actually, v4 of the patch was implemented as such but did not get any > feedback, > https://lore.kernel.org/bpf/[email protected]/#t > > ... and v5 (the final commit) somehow removed the restriction for unconnected > UDP socket as well. > https://lore.kernel.org/bpf/[email protected]/ > > Given the initial use case, sockmap redirect, is still blocked by > TCP_ESTABLISHED > check in sock_map_redirect_allowed(), I feel there is no point in supporting > unconnected UDP sockets in sockmap. It cannot get any skb from anywhere > (without buggy sk_lookup).
s/unconnected/unhashed/g :)
