bpf_sk_assign_tcp_reqsk() does not validate the L4 protocol of the skb, only checking skb->protocol (L3). A BPF program that calls this kfunc on a non-TCP skb (e.g. UDP) will succeed, attaching a TCP reqsk to the skb.
When the skb enters the UDP receive path, skb_steal_sock() returns the TCP listener socket from the reqsk. The UDP code then casts this TCP socket to udp_sock and accesses UDP-specific fields at invalid offsets, causing a null pointer dereference: BUG: KASAN: null-ptr-deref in __udp_enqueue_schedule_skb+0x19d/0x1df0 Read of size 4 at addr 0000000000000008 by task test_progs/537 CPU: 1 UID: 0 PID: 537 Comm: test_progs Not tainted 7.0.0-rc4+ #46 PREEMPT Call Trace: <IRQ> dump_stack_lvl (lib/dump_stack.c:123) print_report (mm/kasan/report.c:487) kasan_report (mm/kasan/report.c:597) __kasan_check_read (mm/kasan/shadow.c:32) __udp_enqueue_schedule_skb (net/ipv4/udp.c:1719) udp_queue_rcv_one_skb (net/ipv4/udp.c:2370 net/ipv4/udp.c:2500) udp_queue_rcv_skb (net/ipv4/udp.c:2532) udp_unicast_rcv_skb (net/ipv4/udp.c:2684) __udp4_lib_rcv (net/ipv4/udp.c:2742) udp_rcv (net/ipv4/udp.c:2937) ip_protocol_deliver_rcu (net/ipv4/ip_input.c:209) ip_local_deliver_finish (./include/linux/rcupdate.h:879 net/ipv4/ip_input.c:242) ip_local_deliver (net/ipv4/ip_input.c:265) __netif_receive_skb_one_core (net/core/dev.c:6164 (discriminator 4)) __netif_receive_skb (net/core/dev.c:6280) Solution Validating the protocol in the helper is not enough: a BPF program can bypass an ip_hdr(skb)->protocol check via TOCTOU by rewriting the header around the call, and bpf_sk_assign() has the same problem since it can assign any socket type to any skb. So validate the protocol where the assigned socket is consumed instead. Patch 1: Validate the L4 protocol in skb_steal_sock(). Each caller passes the protocol it handles (TCP or UDP), and a prefetched socket whose protocol does not match is rejected, regardless of how it was assigned. Patch 2: Add a selftest that calls bpf_sk_assign_tcp_reqsk() on a UDP skb and verifies the stack no longer crashes. --- v1: https://lore.kernel.org/bpf/[email protected]/ v2: https://lore.kernel.org/bpf/[email protected]/ v3: https://lore.kernel.org/bpf/[email protected]/ v4: https://lore.kernel.org/bpf/[email protected]/ v5: https://lore.kernel.org/bpf/[email protected]/ v6: https://lore.kernel.org/all/[email protected]/ Changes in v6 & v7: - resend and keep selftest. Changes in v5: - use skb_header_pointer instead of pskb_may_pull. Changes in v5: - Add pskb_may_pull before accessing IP/IPv6 headers in kfunc - Use buf[] instead of buf[32], verify recv data with ASSERT_STREQ - Remove unnecessary variable initializations in selftest and BPF Changes in v4: - Check if assign_ret is EINVAL instead of checking if it is 0 Changes in v3: - Add IPv6 test coverage, reuse test_cases[] to iterate over both address families - Share TCP/UDP port to simplify BPF program, remove unnecessary global variables - Use connect_to_fd() + send()/recv() instead of manual sockaddr construction - Suggested by Kuniyuki Iwashima Changes in v2: - Add Reviewed-by tag from Kuniyuki Iwashima for patch 1 - Use UDP socket recv() instead of kern_sync_rcu() for synchronization in selftest Jiayuan Chen (2): net: Validate protocol in skb_steal_sock() for BPF-assigned sockets selftests/bpf: Add protocol check test for bpf_sk_assign_tcp_reqsk() include/net/inet6_hashtables.h | 7 +- include/net/inet_hashtables.h | 7 +- include/net/request_sock.h | 16 ++- net/ipv4/udp.c | 2 +- net/ipv6/udp.c | 2 +- .../bpf/prog_tests/tcp_custom_syncookie.c | 87 ++++++++++++++- .../bpf/progs/test_tcp_custom_syncookie.c | 102 ++++++++++++++++++ 7 files changed, 210 insertions(+), 13 deletions(-) -- 2.43.0
