With IFF_NAPI, tun_get_user() queues every written packet for NAPI on an unbounded list. When the poll runs asynchronously (threaded NAPI or ksoftirqd) and is slower than the writer, the queue grows until the system runs out of memory. An unprivileged user can do this from its own user and network namespaces, and the queued skbs are not charged to its memcg: in a test VM, a uid 1000 process limited to 128 MB caused a global OOM in ~6 seconds. The socket send buffer does not help: tun sets it to INT_MAX by default, the build_skb() path is only used while it is INT_MAX, and the XDP batch path never checks it. Drop packets once more than netdev_max_backlog of them are queued, like enqueue_to_backlog() does, and count them as rx_dropped; the write still reports success, as for a full backlog. The check sits where packets are queued for NAPI, in tun_get_user() and in tun_xdp_one() for the batches vhost-net passes, so packets consumed by XDP or delivered through IFF_NAPI_FRAGS are not affected. A dropped zerocopy skb completes its buffer when freed. The check is lockless, so concurrent writers can each still add one packet. The NAPI poll moved the whole queue out, processed up to its budget and put the rest back: while it ran, the queue looked empty to writers, so with the poll on another CPU the backlog could still grow without limit. Take at most budget packets instead: the rest stays on the queue writers check. Dropping also schedules NAPI. vhost-net sends with MSG_MORE while it has more packets, and tun_get_user() schedules NAPI only for a packet without it or once more than NAPI_POLL_WEIGHT are queued. With a netdev_max_backlog below that, the queue fills first and the packet without MSG_MORE is dropped too: nothing would schedule NAPI, and the queued packets, with zerocopy also their buffers, would wait forever. Reproducer, as an unprivileged user in a new user and network namespace, without kernel modifications or fault injection: ip tuntap add t0 mode tun; ip addr add 10.11.0.1/24 dev t0 ip link add d0 type dummy; ip addr add 10.20.0.1/24 dev d0 ip link set t0 up; ip link set d0 up; sysctl net.ipv4.ip_forward=1 install ~500 iptables FORWARD rules, to slow down the poll, and a MASQUERADE rule on d0 open t0 with IFF_TUN | IFF_NO_PI | IFF_NAPI echo 1 > /sys/class/net/t0/threaded write() 1400-byte IPv4 packets for 10.20.0.2 in a loop Found by code inspection while analysing the performance of the tun write path. Tested in an x86-64 VM on net, with 1 and 2 vCPUs: the reproducer causes a global OOM without the patch, while with it memory use stays flat and the forwarding rate is unchanged; the tun kselftest passes; a local vhost-net test of the batch, copy and zerocopy paths with netdev_max_backlog set to 8 passes, and fails without the patch; an XDP program on the device sees every packet while the excess is dropped. Still open: neither the packet size nor the queued memory is bounded, and the skbs are not charged to the writer's memcg. A single write can create an skb of several hundred KB, so even one queue can hold a lot, and IFF_MULTI_QUEUE (allowed in a user namespace) and more devices multiply it. This bounds the queue as enqueue_to_backlog() bounds the per-CPU backlog; charging the skbs to the writer is left to a follow-up. Fixes: 943170998b20 ("tun: enable NAPI for TUN/TAP driver") Signed-off-by: Antonio Quartulli --- v7: - Check where packets are queued, after XDP and not for IFF_NAPI_FRAGS (Sashiko); a dropped zerocopy skb now completes its buffer when freed - Fix the poll comment and the commit message (Sashiko) - Drop the vhost-net selftest (Willem, Sashiko) v6: https://lore.kernel.org/netdev/20261113212237.1742984-1-antonio@mandelbit.com/ v5: https://lore.kernel.org/netdev/20261111014208.1795777-1-antonio@mandelbit.com/ v4: https://lore.kernel.org/netdev/20261103142232.1869185-1-antonio@mandelbit.com/ v3: https://lore.kernel.org/netdev/20261031231444.1528758-1-antonio@mandelbit.com/ v2: https://lore.kernel.org/netdev/20261030152540.1409884-1-antonio@mandelbit.com/ Stable: kernels before v6.9 have no net_hotdata; a backport reads netdev_max_backlog instead. drivers/net/tun.c | 52 +++++++++++++++++++++++++++++++++++++++-------- 1 file changed, 44 insertions(+), 8 deletions(-) diff --git a/drivers/net/tun.c b/drivers/net/tun.c index 86fd44e10755..b7432bd4d3f0 100644 --- a/drivers/net/tun.c +++ b/drivers/net/tun.c @@ -58,6 +58,7 @@ #include #include #include +#include #include #include #include @@ -227,6 +228,16 @@ struct veth { static void tun_flow_init(struct tun_struct *tun); static void tun_flow_uninit(struct tun_struct *tun); +/* In IFF_NAPI mode written packets wait in the tfile's sk_write_queue for the + * NAPI poll, which may run later in a kthread or ksoftirqd. Bound that queue + * like the per-CPU backlog. + */ +static bool tun_napi_queue_full(const struct sk_buff_head *queue) +{ + return skb_queue_len_lockless(queue) > + READ_ONCE(net_hotdata.max_backlog); +} + static int tun_napi_receive(struct napi_struct *napi, int budget) { struct tun_file *tfile = container_of(napi, struct tun_file, napi); @@ -237,21 +248,22 @@ static int tun_napi_receive(struct napi_struct *napi, int budget) __skb_queue_head_init(&process_queue); + /* At most budget skbs: the rest stays on the queue writers check */ spin_lock(&queue->lock); - skb_queue_splice_tail_init(queue, &process_queue); + if (skb_queue_len(queue) <= budget) { + skb_queue_splice_tail_init(queue, &process_queue); + } else { + while (skb_queue_len(&process_queue) < budget && + (skb = __skb_dequeue(queue))) + __skb_queue_tail(&process_queue, skb); + } spin_unlock(&queue->lock); - while (received < budget && (skb = __skb_dequeue(&process_queue))) { + while ((skb = __skb_dequeue(&process_queue))) { napi_gro_receive(napi, skb); ++received; } - if (!skb_queue_empty(&process_queue)) { - spin_lock(&queue->lock); - skb_queue_splice(&process_queue, queue); - spin_unlock(&queue->lock); - } - return received; } @@ -2043,6 +2055,19 @@ static ssize_t tun_get_user(struct tun_struct *tun, struct tun_file *tfile, struct sk_buff_head *queue = &tfile->sk.sk_write_queue; int queue_len; + /* Queue full: drop, and schedule NAPI, which a MSG_MORE sender + * may not have done yet. Lockless: concurrent writers can each + * still add one. + */ + if (unlikely(tun_napi_queue_full(queue))) { + local_bh_disable(); + napi_schedule(&tfile->napi); + local_bh_enable(); + rcu_read_unlock(); + drop_reason = SKB_DROP_REASON_FULL_RING; + goto drop; + } + spin_lock_bh(&queue->lock); if (unlikely(tfile->detached)) { @@ -2658,6 +2683,17 @@ static int tun_xdp_one(struct tun_struct *tun, if (tfile->napi_enabled) { queue = &tfile->sk.sk_write_queue; + + /* Queue full: drop, and schedule NAPI so that it drains even if + * nothing in this batch gets queued. + */ + if (unlikely(tun_napi_queue_full(queue))) { + napi_schedule(&tfile->napi); + kfree_skb_reason(skb, SKB_DROP_REASON_FULL_RING); + dev_core_stats_rx_dropped_inc(tun->dev); + return 0; + } + spin_lock(&queue->lock); if (unlikely(tfile->detached)) { base-commit: ead6fa9a0ab194f02a2f36d634b193e9c5f9a1c6 -- 2.56.1