From mboxrd@z Thu Jan 1 00:00:00 1970 Authentication-Results: mail.toke.dk; dkim=none; arc=none (Message is not ARC signed); dmarc=fail (Used From Domain Record) header.from=toke.dk policy.dmarc=none From: Toke =?utf-8?Q?H=C3=B8iland-J=C3=B8rgensen?= To: Jamal Hadi Salim , netdev@vger.kernel.org Cc: Jamal Hadi Salim , Jiri Pirko , "David S . Miller" , Eric Dumazet , Jakub Kicinski , Paolo Abeni , Simon Horman , Victor Nogueira , =?utf-8?B?QmrDtnJuIFTDtnBlbA==?= , cake@lists.bufferbloat.net, Sashiko In-Reply-To: References: Date: Mon, 19 Oct 2026 20:20:37 +0200 X-Clacks-Overhead: GNU Terry Pratchett Message-ID: <87qzhlwo0a.fsf@toke.dk> MIME-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: quoted-printable Message-ID-Hash: ZYZZ7WNY5UQHPG5QDLSQVNRNSU4DHJCI X-Message-ID-Hash: ZYZZ7WNY5UQHPG5QDLSQVNRNSU4DHJCI X-MailFrom: toke@toke.dk X-Mailman-Rule-Misses: dmarc-mitigation; no-senders; approved; loop; banned-address; emergency; member-moderation; nonmember-moderation; administrivia; implicit-dest; max-recipients; max-size; news-moderation; no-subject; digests; suspicious-header X-Mailman-Version: 3.3.10 Precedence: list Subject: [Cake] Re: [PATCH net-next v7 3/4] net/sched: cap the accounted backlog before it can wrap List-Id: Cake - FQ_codel the next generation Archived-At: List-Archive: List-Help: List-Owner: List-Post: List-Subscribe: List-Unsubscribe: Jamal Hadi Salim writes: > This is a followup of commit 709f34f7c28d ("net/sched: fq: add overflow > bounds to quantum and initial quantum") as reported by Sashiko. The rest of this commit message is somewhat obtuse - too much run-on information. I think it can be pretty easily fixed by a couple of rearrangements and paragraph breaks, like this: > The issue: fq_codel, cake, codel, pie, fq_pie, dualpi2, RED, pfifo, > plug and (non-WRED) gred account every enqueued packet's stab-adjusted > length into the 32-bit sch->qstats.backlog. A crafted tc size table > inflates qdisc_pkt_len() up to QDISC_PKT_LEN_MAX (1 MiB), and a dodgy > GSO skb can account even more through qdisc_pkt_len_segs_init(), so a > few thousand packets wrap the counter mod 2^32. > > The AQM then reads a small backlog and makes the wrong drop decision, > and the dequeue-side subtractions leave it corrupt (for non-WRED gred > it is the reported stats and the dequeue-side subtractions that go > wrong instead). In fq_codel a wrapped per-flow backlog can also make > fq_codel_drop()'s scan pick an empty flow, dereferencing a NULL > flow->head. > > The Fix: drop at enqueue once the stored backlog cannot represent the > incoming packet's accounted length. > > Ordinary packets (qdisc_pkt_len() at most QDISC_PKT_LEN_MAX) are > covered by a standing ceiling: drop once the backlog is within > QDISC_MAX_BACKLOG (U32_MAX - QDISC_PKT_LEN_MAX) of the wrap point, the > largest backlog one more 1 MiB packet cannot wrap, so there is no > per-enqueue 64-bit add. > > A larger wire estimate is checked against the space that actually > remains (len > U32_MAX - backlog): qdisc_pkt_len_segs_init() adds the > headers replicated on every output segment, so a supported small-MSS > BIG TCP aggregate can carry a wire estimate above 1 MiB while its raw > length stays within GSO_MAX_SIZE, and such traffic must pass on an > empty queue. > > qdisc_pkt_len() itself stays truthful for every input: it is not > clamped and the initializer keeps succeeding, so BPF > __sk_buff.wire_len, act_police and the HTB/TBF shapers read the real > length, and sch_handle_ingress() -- which ignores the initializer's > return value and reads pkt_segs for packet statistics -- keeps > complete metadata. > > Cover sch_plug and non-WRED sch_gred too: plug admits on skb->len > while qdisc_enqueue_tail() adds qdisc_pkt_len(), and gred bounds each > virtual queue but adds every VQ into the one aggregate. pfifo is > included because RED reads a grafted child's backlog into > red_calc_qavg(); it can be reached standalone or as the default child > of SFB. In RED, account the packet to the aggregate before handing it > to the child, because a duplicating netem child re-enqueues its clone > at the root while the stored backlog is still un-updated -- one guard > check must not authorize two additions. > > The guard bounds each qdisc's own enqueue path: a child that segments a > GSO skb (tbf, netem) can still grow a parent's aggregate through > qdisc_tree_reduce_backlog()'s unclamped byte delta, which is left to a > separate change. > > Conditions to recreate the bug: > ip tuntap add tun0 mode tun > ip link set tun0 txqueuelen 32 up > ip addr add 10.99.0.1/24 dev tun0 > tc qdisc add dev tun0 root handle 1: stab overhead 2000000000 \ > fq_codel flows 1 limit 20000 ecn > # hold the tun fd open without reading so the backlog persists, then > # send 6000 packets: the counter wraps and the unfixed kernel reports > # a wrapped backlog, while the fixed one drops at the ceiling. > > Reported-by: Sashiko (nipa) > Closes: https://lore.kernel.org/netdev/20260818101130.16203-1-jhs@mojatat= u.com/ > Link: https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260818101130.= 16203-1-jhs@mojatatu.com > Link: https://lore.kernel.org/netdev/CANn89iLfMJV7ancKH1Gjzzm7ZUG-gKcrczk= JjJEWN5sCuWd-ug@mail.gmail.com/ > Link: https://lore.kernel.org/netdev/CANn89i+GOFH_8g+vSV0jUj0aBqLYVWuFOrH= uDggZu9kPC_m6tQ@mail.gmail.com/ > Link: https://lore.kernel.org/netdev/87a4oottcj.fsf@toke.dk/ > Link: https://lore.kernel.org/netdev/87pkxks68f.fsf@toke.dk/ > Link: https://lore.kernel.org/netdev/179231932946.3135.749747899018638568= 8@kernel.org/ > Reviewed-by: Victor Nogueira > Signed-off-by: Jamal Hadi Salim With the above rewording: Reviewed-by: Toke H=C3=B8iland-J=C3=B8rgensen