[PATCH net-next 05/15] batman-adv: tp_meter: avoid divide-by-zero for dec_cwnd
Simon Wunderlich <[email protected]>
| Newsgroups | org.open-mesh.lists.batman,org.kernel.vger.netdev |
|---|---|
| Message-ID | <[email protected]> |
From: Sven Eckelmann <[email protected]> The cwnd is always MSS <= cwnd <= 0x20000000. But the calculation in batadv_tp_update_cwnd() assumes unsigned 32 bit arithmetics. ((mss * 8) ** 2) / (cwnd * 8) In case cwnd is actually 0x20000000, it will be shifted by 3 bit to the left end up at 0x100000000 or U32_MAX + 1. It will therefore wrap around and be 0 - resulting in: ((mss * 8) ** 2) / 0 This is of course invalid and cannot be calculated. The calculation should must be simplified to avoid this overflow: (mss ** 2) * 8 / cwnd It will keep the precision enhancement from the scaling (by 8) but avoid the overflow in the divisor. In theory, there could still be an overflow in the dividend. It is at the moment fixed to BATADV_TP_PLEN in batadv_tp_recv_ack() - so it is not an imminent problem. But allowing it to use the whole u32 bit range, would mean that it can still use up to 67 bits. To keep this calculation safe for 32 bit arithmetic, mss must never use more than floor((32 - 3) / 2) bits - or in other words: must never be larger than 16383. Cc: [email protected] Fixes: 33a3bb4a3345 ("batman-adv: throughput meter implementation") Signed-off-by: Sven Eckelmann <[email protected]> Signed-off-by: Simon Wunderlich <[email protected]> --- net/batman-adv/tp_meter.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/net/batman-adv/tp_meter.c b/net/batman-adv/tp_meter.c index 9ecbc6023cfc9..1655f181c9293 100644 --- a/net/batman-adv/tp_meter.c +++ b/net/batman-adv/tp_meter.c @@ -154,9 +154,12 @@ static void batadv_tp_update_cwnd(struct batadv_tp_vars *tp_vars, u32 mss) return; } + /* prevent overflow in (mss * mss) << 3 */ + mss = min_t(u32, mss, (1U << 14) - 1); + /* increment CWND at least of 1 (section 3.1 of RFC5681) */ tp_vars->dec_cwnd += max_t(u32, 1U << 3, - ((mss * mss) << 6) / (tp_vars->cwnd << 3)); + ((mss * mss) << 3) / tp_vars->cwnd); if (tp_vars->dec_cwnd < (mss << 3)) { spin_unlock_bh(&tp_vars->cwnd_lock); return; -- 2.47.3