icsk_backoff is used as a shift amount in inet_csk_rto_backoff(): u64 when = (u64)icsk->icsk_rto << icsk->icsk_backoff; It can go past 63 on the paths of both callers: - RTO: tcp_retransmit_timer() increments it on every backed-off RTO with no cap of its own, then an ICMP unreachable goes through tcp_ld_RTO_revert(). - zero window probe: tcp_send_probe0() increments it up to tcp_retries2, and tcp_probe_timer() calls the helper in its orphan check. commit 6d4634d1b091 ("net: Limit logical shift left of TCP probe0 timeout") clamped tcp_probe0_when() but not this one. Seen on the RTO path with tcp_retries2=255 and a silent peer: UBSAN: shift-out-of-bounds in ./include/net/inet_connection_sock.h:258:40 shift exponent 73 is too large for 64-bit type 'long long unsigned int' Call Trace: dump_stack_lvl+0xc1/0xf0 dump_stack+0x10/0x20 __ubsan_handle_shift_out_of_bounds+0x1d1/0x390 tcp_ld_RTO_revert.cold+0x16/0x92 tcp_v4_err+0xdb3/0x13e0 icmp_socket_deliver+0x1a7/0x390 icmp_unreach+0x2dd/0xe50 icmp_rcv+0x93e/0xe30 ip_protocol_deliver_rcu+0x36b/0x3e0 ip_local_deliver_finish+0x37d/0x620 ip_local_deliver+0x165/0x1d0 ip_rcv+0x284/0x320 __netif_receive_skb_one_core+0x16b/0x1c0 __netif_receive_skb+0x21/0x150 process_backlog+0x37e/0x1200 __napi_poll.constprop.0+0xa5/0x5e0 net_rx_action+0x4c0/0xf70 handle_softirqs+0x204/0x9b0 __irq_exit_rcu+0x13c/0x220 irq_exit_rcu+0xe/0x20 sysvec_apic_timer_interrupt+0x92/0xd0 tcp_retries2 does not need to be changed for the RTO path: with TCP_USER_TIMEOUT and TCP_RTO_MAX_MS set on the socket, retransmits_timed_out() only looks at the elapsed time. The zero window probe path gives the same splat from tcp_probe_timer(). On x86-64 the shift count is masked, and the u64 can already be 0 before that (e.g. icsk_rto == 200 and icsk_backoff == 61). An orphan that should be dead can then be seen as alive in tcp_probe_timer(), and tcp_ld_RTO_revert() can end up with a smaller RTO than intended. Cap the shift at 32: any non-zero icsk_rto << 32 exceeds TCP_RTO_MAX, so the min_t() result is what the unbounded math would give. The ilog2(TCP_RTO_MAX / TCP_RTO_MIN) + 1 bound of tcp_probe0_when() does not fit here, as icsk_rto has no TCP_RTO_MIN floor the way tcp_probe0_base() has. Found by checking sysctl ranges against their users. Reproduced the RTO path in a VM with UBSAN: a - r - b netns, tcp_retries2=255 and tcp_rto_max_ms=1000 on a, blackhole route on r for 75s, then switch it to unreachable. With this patch the splat is gone on both paths. Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2") Reported-by: Kyuwon Cho Closes: https://lore.kernel.org/netdev/179499954853.1440581.6627074959099619480@gmail.com/ Suggested-by: Eric Dumazet Link: https://lore.kernel.org/netdev/CANn89i+=QRibXsm6tw-eVwhB3wjOtgF-MtVu0G--K5=oWP97+A@mail.gmail.com/ Cc: stable@vger.kernel.org Signed-off-by: Jiayuan Chen Tested-by: Kyuwon Cho Reviewed-by: Eric Dumazet --- v3 -> v4: rebase, update the commit message. No code change. v2 -> v3: mention the zero window probe path, add Reported-by/Closes (Eric) and Suggested-by/Link (Neal), say how it was found and tested, point Fixes at 1da177e4c3f4. No code change. --- include/net/inet_connection_sock.h | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/include/net/inet_connection_sock.h b/include/net/inet_connection_sock.h index 903499581db7..562e12033a5d 100644 --- a/include/net/inet_connection_sock.h +++ b/include/net/inet_connection_sock.h @@ -255,9 +255,10 @@ static inline unsigned long inet_csk_rto_backoff(const struct inet_connection_sock *icsk, unsigned long max_when) { - u64 when = (u64)icsk->icsk_rto << icsk->icsk_backoff; + u8 backoff = min_t(u8, 32U, icsk->icsk_backoff); + u64 when = (u64)icsk->icsk_rto << backoff; - return (unsigned long)min_t(u64, when, max_when); + return (unsigned long)min_t(u64, when, max_when); } struct sock *inet_csk_accept(struct sock *sk, struct proto_accept_arg *arg); -- 2.43.0