Re: [PATCH net-next v3 3/4] net/sched/sch_cake: widen buffer_used to u64

Jamal Hadi Salim [off-list ref] writes:

This is a follow-up to commit 4c660ee8c809 ("net: sched: fix 32-bit
backlog wrap in gred, bfifo and plug enqueue"), which promoted the
backlog + length admission sums in gred, bfifo and plug to u64.  CAKE's
own memory accounting has the same 32-bit wrap.

The issue:
cake_sched_data.buffer_used accumulates skb->truesize and is compared
against memory_limit:

	q->buffer_used += skb->truesize;
	if (q->buffer_used <= q->buffer_limit)
		return NET_XMIT_SUCCESS;
	while (q->buffer_used > q->buffer_limit)
		drop...

buffer_used is u32, so once the resident truesize passes 2^32 it wraps to
a small value, the <= test passes and the drop loop is skipped even
though far more than the configured limit is queued.  memory_limit is a
u32 attribute (TCA_CAKE_MEMORY), so buffer_limit can be as large as
U32_MAX and never caps buffer_used below the wrap point; with a large
memory_limit the queue then grows until the machine is out of memory.

The fix:
Widen buffer_used and buffer_max_used to u64, the same shape as the
gred/bfifo/plug fix.  buffer_limit stays u32; the comparison promotes it
to u64, so the bounded queue stops at the configured limit below 2^32 and
the counter can no longer wrap.  buffer_used is subtracted on dequeue and
drop, so the widened type flows through unchanged.  The ACK-filter
replacement applied the net delta as one expression,
q->buffer_used += skb->truesize - ack->truesize.  Both operands are
unsigned int, so the subtraction is evaluated at u32 and a larger removed
ACK wraps the delta; the old u32 accumulator folded that back to the
correct net value, but the widened u64 accumulator would persist it as a
roughly 4 GiB over-count.  Apply the replacement as two exact operations
on the widened counter instead; the queue already accounts ack->truesize,
so the subtraction cannot underflow.

Widening the counters alone is incomplete because cake_dump_stats() still
exports buffer_max_used through PUT_STAT_U32(MEMORY_USED, ...), and
tc.yaml declares memory-used as u32.  buffer_max_used only ever rises, so
once the widened counter passes 2^32 the export reports peak - 2^32 for
the rest of the qdisc's life.  Clamp the exported value to U32_MAX; the
counter itself is unaffected.

The dump path runs without the qdisc lock: tc_fill_qdisc() passes NULL to
gnet_stats_start_copy_compat(), so the reader races with the enqueue-side
update.  READ_ONCE() is not an atomic 64-bit load on 32-bit CPUs, so a
reader can observe a torn high-water mark and report a bogus MEMORY_USED.
Publish the counter as u64_stats_t under the qdisc lock, which the enqueue
path already holds, and take a seqcount-protected snapshot in the dump.
The counter stays u64; only the exported value is bounded.

Conditions to recreate the bug:
CAP_NET_ADMIN (root or a user namespace, the qdisc attach is gated by
netlink_net_capable(CAP_NET_ADMIN)).

A TUN device is needed so the root qdisc can park packets instead of
dropping them.  Open /dev/net/tun, request TUNSETIFF with
IFF_TUN | IFF_NO_PI | IFF_BACKPRESSURE and keep the fd open without
reading it (run it in the background and stop it at teardown): with
IFF_BACKPRESSURE the full tx ring calls netif_tx_stop_queue() rather than
dropping, so the queue behind the qdisc can grow past 4 GiB of resident
truesize.  A plain 'ip tuntap add' attaches no fd and does not request
IFF_BACKPRESSURE, so it cannot park the queue this way.

  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: cake memlimit 4294967295

  # Then push more than 4 GiB of resident truesize through the qdisc: six
  # UDP senders using large SO_SNDBUFFORCE buffers send 1400-byte
  # datagrams (the reported OOM witness used 6 workers for 100 s on a
  # 12 GiB guest).  On the unfixed kernel buffer_used wraps, the admission
  # test passes, the drop loop is bypassed and the guest runs out of
  # memory; with the fix the counter passes the limit without wrapping and
  # the excess is dropped.

  # The ACK-filter over-count is a separate corner: it needs ack-filter
  # enabled and a queued pure ACK whose truesize is larger than the
  # replacement ACK.  Use a fresh interface (or 'tc qdisc change') rather
  # than a second root 'tc qdisc add' on an interface that already has a
  # qdisc:
  tc qdisc change dev tun0 root cake ack-filter

CONFIG_NET_SCH_CAKE=y and CONFIG_TUN=y are required, and the
memory-residency recipe needs a root-run sender envelope with enough
guest RAM to hold the queued data.

Reported-by: Sashiko (gemini) <sashiko-bot@kernel.org>
Closes: https://sashiko.dev/#/patchset/20260818095927.15901-1-jhs@mojatatu.com
Link: https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260818095927.15901-1-jhs@mojatatu.com
Link: https://lore.kernel.org/netdev/20260818095927.15901-1-jhs@mojatatu.com/ (local)
Reviewed-by: Victor Nogueira <redacted>
Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com>

Acked-by: Toke Høiland-Jørgensen <toke@toke.dk>