On Mon, May 5, 2025 at 5:28 PM Greg Kroah-Hartman gregkh@linuxfoundation.org wrote:
On Fri, May 02, 2025 at 12:49:48PM +0800, Tai, Gerrard wrote:
Upstream commits: 01: 5ba8b837b522d7051ef81bacf3d95383ff8edce5 ("sch_htb: make htb_qlen_notify() idempotent") 02: df008598b3a00be02a8051fde89ca0fbc416bd55 ("sch_drr: make drr_qlen_notify() idempotent") 03: 51eb3b65544c9efd6a1026889ee5fb5aa62da3bb ("sch_hfsc: make hfsc_qlen_notify() idempotent") 04: 55f9eca4bfe30a15d8656f915922e8c98b7f0728 ("sch_qfq: make qfq_qlen_notify() idempotent") 05: a7a15f39c682ac4268624da2abdb9114bdde96d5 ("sch_ets: make est_qlen_notify() idempotent") 06: 342debc12183b51773b3345ba267e9263bdfaaef ("codel: remove sch->q.qlen check before qdisc_tree_reduce_backlog()")
These patches are patch 01-06 of the original patchset ([1]) authored by Cong Wang. I have omitted patches 07-11 which are selftests. This patchset addresses a UAF vulnerability.
Originally, only the last commit (06) was picked to merge into the latest round of stable queues 5.15,5.10,5.4. For 6.x stable branches, that sole commit has already been merged in a previous cycle.
From my understanding, this patch depends on the previous patches to work.
Without patches 01-05 which make various classful qdiscs' qlen_notify() idempotent, if an fq_codel's dequeue() routine empties the fq_codel qdisc, it will be doubly deactivated - first in the parent qlen_notify and then again in the parent dequeue. For instance, in the case of parent drr, the double deactivation will either cause a fault on an invalid address, or trigger a splat if list checks are compiled into the kernel. This is also why the original unpatched code included the qlen check in the first place.
After discussion with Greg, he has helped to temporarily drop the patch from the 5.x queues ([2]). My suggestion is to include patches 01-06 of the patchset, as listed above, for the 5.x queues. For the 6.x queues that have already merged patch 06, the earlier patches 01-05 should be merged too.
I'm not too familiar with the stable patch process, so I may be completely mistaken here.
I'll be glad to take what is needed, but please list what commits need to go to what branches and in what exact order please.
Here's the list of commits. The order should be the sequence as listed below.
6.14, 6.13, 6.12, 6.6, 6.1: (all 6.* branches) 5ba8b837b522d7051ef81bacf3d95383ff8edce5 ("sch_htb: make htb_qlen_notify() idempotent") df008598b3a00be02a8051fde89ca0fbc416bd55 ("sch_drr: make drr_qlen_notify() idempotent") 51eb3b65544c9efd6a1026889ee5fb5aa62da3bb ("sch_hfsc: make hfsc_qlen_notify() idempotent") 55f9eca4bfe30a15d8656f915922e8c98b7f0728 ("sch_qfq: make qfq_qlen_notify() idempotent") a7a15f39c682ac4268624da2abdb9114bdde96d5 ("sch_ets: make est_qlen_notify() idempotent")
5.15, 5.10, 5.4: (all 5.* branches) 5ba8b837b522d7051ef81bacf3d95383ff8edce5 ("sch_htb: make htb_qlen_notify() idempotent") df008598b3a00be02a8051fde89ca0fbc416bd55 ("sch_drr: make drr_qlen_notify() idempotent") 51eb3b65544c9efd6a1026889ee5fb5aa62da3bb ("sch_hfsc: make hfsc_qlen_notify() idempotent") 55f9eca4bfe30a15d8656f915922e8c98b7f0728 ("sch_qfq: make qfq_qlen_notify() idempotent") a7a15f39c682ac4268624da2abdb9114bdde96d5 ("sch_ets: make est_qlen_notify() idempotent") 342debc12183b51773b3345ba267e9263bdfaaef ("codel: remove sch->q.qlen check before qdisc_tree_reduce_backlog()")
Cheers, Gerrard