On Tue, Aug 18, 2026 at 9:00 AM Eric Dumazet [off-list ref] wrote:
On Tue, Aug 18, 2026 at 2:55 PM Eric Dumazet [off-list ref] wrote:
quoted
On Tue, Aug 18, 2026 at 12:11 PM Jamal Hadi Salim [off-list ref] wrote:
quoted
Several qdiscs derive their per-flow quantum or backlog from psched_mtu()
or accumulate qdisc_pkt_len() into a u32/int counter without an overflow
or zero clamp, which can drive the dequeue/credit-refill loop into a soft
lockup or a NULL deref.
vega@nebusec.ai provided reports and PoCs which illustrated the following:
- sch_fq: fq_dequeue() credit-refill loop with a small quantum spins ~1B
iterations under the qdisc lock (soft lockup); fq_init() computes
quantum = 2 * psched_mtu() with no overflow check.
- sch_fq_codel: fq_codel_enqueue() accumulates qdisc_pkt_len() into a u32
per-flow backlog; a crafted TCA_STAB inflates pkt_len to 1 GiB so a few
packets wrap the counter to 0, and fq_codel_drop() then picks an empty
flow and derefs NULL.
- sch_hhf: hhf_init() sets quantum = psched_mtu() with no overflow check;
a huge MTU makes it 0x80000000, and hhf_dequeue()'s deficit += weight *
quantum loops forever.
- sch_sfq: sfq_init() sets quantum = psched_mtu() (unsigned); a huge MTU
makes allot = INT_MIN, and INT_MIN + INT_MIN is UB that toggles between
INT_MIN and 0 forever.
Clamp the quantum to a sane minimum and promote the backlog/credit sums to
avoid the wrap, so the dequeue loops terminate and the drop path never
selects an empty flow.
Reported-by: <redacted>
Tested-by: Victor Nogueira <redacted>
Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com>
---
net/sched/sch_fq.c | 9 ++++++---
net/sched/sch_fq_codel.c | 14 +++++++++++++-
net/sched/sch_hhf.c | 12 +++++++++++
net/sched/sch_sfq.c | 7 +++++++
4 files changed, 38 insertions(+), 4 deletions(-)
diff --git a/net/sched/sch_fq.c b/net/sched/sch_fq.c
index 7cae082a9847..66e9c6e530d8 100644
--- a/net/sched/sch_fq.c
+++ b/net/sched/sch_fq.c
@@ -750,7 +750,10 @@ static struct sk_buff *fq_dequeue(struct Qdisc *sch)
f = head->first;
retry = 0;
if (f->credit <= 0) {
- f->credit += q->quantum;
+ if (f->credit + (int)q->quantum > 0)
+ f->credit += q->quantum;
+ else
+ f->credit = q->quantum;
Lets not add cost in FQ fast path just because of a configuration issue?
quoted
head->first = f->next;
fq_flow_add_tail(q, f, OLD_FLOW);
goto begin;@@ -1226,8 +1229,8 @@ static int fq_init(struct Qdisc *sch, struct nlattr *opt,
sch->limit = 10000;
q->flow_plimit = 100;
- q->quantum = 2 * psched_mtu(qdisc_dev(sch));
- q->initial_quantum = 10 * psched_mtu(qdisc_dev(sch));
+ q->quantum = max_t(u32, 2 * psched_mtu(qdisc_dev(sch)), 256);
+ q->initial_quantum = max_t(u32, 10 * psched_mtu(qdisc_dev(sch)), 256);
We instead can make sure quantum and initial_quantum are in an acceptable range.
[256, 16M] would probably make sense.
Actually, we use 1M in TCA_FQ_QUANTUM in fq_change()
We also need to change fq_change() because TCA_FQ_INITIAL_QUANTUM can be set up
to INT_MAX.
Good point on avoiding fast path changes. I will review the others and
resend after hearing from the naughty AI.
So something like attached (compiles, untested)?
cheers,
jamal