af_unix: Fix garbage collector racing against connect()
Garbage collector does not take into account the risk of embryo getting enqueued during the garbage collection. If such embryo has a peer that carries SCM_RIGHTS, two consecutive passes of scan_children() may see a different set of children. Leading to an incorrectly elevated inflight count, and then a dangling pointer within the gc_inflight_list. sockets are AF_UNIX/SOCK_STREAM S is an unconnected socket L is a listening in-flight socket bound to addr, not in fdtable V's fd will be passed via sendmsg(), gets inflight count bumped connect(S, addr) sendmsg(S, [V]); close(V) __unix_gc() ---------------- ------------------------- ----------- NS = unix_create1() skb1 = sock_wmalloc(NS) L = unix_find_other(addr) unix_state_lock(L) unix_peer(S) = NS // V count=1 inflight=0 NS = unix_peer(S) skb2 = sock_alloc() skb_queue_tail(NS, skb2[V]) // V became in-flight // V count=2 inflight=1 close(V) // V count=1 inflight=1 // GC candidate condition met for u in gc_inflight_list: if (total_refs == inflight_refs) add u to gc_candidates // gc_candidates={L, V} for u in gc_candidates: scan_children(u, dec_inflight) // embryo (skb1) was not // reachable from L yet, so V's // inflight remains unchanged __skb_queue_tail(L, skb1) unix_state_unlock(L) for u in gc_candidates: if (u.inflight) scan_children(u, inc_inflight_move_tail) // V count=1 inflight=2 (!) If there is a GC-candidate listening socket, lock/unlock its state. This makes GC wait until the end of any ongoing connect() to that socket. After flipping the lock, a possibly SCM-laden embryo is already enqueued. And if there is another embryo coming, it can not possibly carry SCM_RIGHTS. At this point, unix_inflight() can not happen because unix_gc_lock is already taken. Inflight graph remains unaffected. Fixes: 1fd05ba5 ("[AF_UNIX]: Rewrite garbage collector, fixes race.") Signed-off-by: Michal Luczaj <mhal@rbox.co> Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com> Link: https://lore.kernel.org/r/20240409201047.1032217-1-mhal@rbox.co Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Please register or sign in to comment