On Mon, May 08, 2023 at 06:33:24PM -0700, Jakub Kicinski wrote:

Great analysis, thanks for squashing this bug.

Thanks, happy to help!

Have you considered creating a fix more localized to the miniq
implementation? It seems that having per-device miniq pointers is
incompatible with using reference counted objects. So miniq is
a more natural place to solve the problem. Otherwise workarounds
in the core keep piling up (here qdisc_graft()).

Can we replace the rcu_assign_pointer in (3rd) with a cmpxchg()?
If active qdisc is neither a1 nor a2 we should leave the dev state
alone.

Yes, I have tried fixing this in mini_qdisc_pair_swap(), but I am afraid
it is hard:

(3rd) is called from ->destroy(), so currently it uses RCU_INIT_POINTER()
to set dev->miniq_ingress to NULL.  It will need a logic like:

  I am A.  Set dev->miniq_ingress to NULL, if and only if it is a1 or a2,
  and do it atomically.

We need more than a cmpxchg() to implement this "set NULL iff a1 or a2".
Additionally:

On Fri,  5 May 2023 17:16:10 -0700 Peilin Ye wrote:

  Thread 1 creates ingress Qdisc A (containing mini Qdisc a1 and a2), then
  adds a flower filter X to A.

  Thread 2 creates another ingress Qdisc B (containing mini Qdisc b1 and
  b2) to replace A, then adds a flower filter Y to B.

 Thread 1               A's refcnt   Thread 2
  RTM_NEWQDISC (A, RTNL-locked)
   qdisc_create(A)               1
   qdisc_graft(A)                9

  RTM_NEWTFILTER (X, RTNL-lockless)
   __tcf_qdisc_find(A)          10
   tcf_chain0_head_change(A)
   mini_qdisc_pair_swap(A) (1st)
            |
            |                         RTM_NEWQDISC (B, RTNL-locked)
           RCU                   2     qdisc_graft(B)
            |                    1     notify_and_destroy(A)
            |
   tcf_block_release(A)          0    RTM_NEWTFILTER (Y, RTNL-lockless)
   qdisc_destroy(A)                    tcf_chain0_head_change(B)
   tcf_chain0_head_change_cb_del(A)    mini_qdisc_pair_swap(B) (2nd)
   mini_qdisc_pair_swap(A) (3rd)                |
           ...                                 ...

Looking at the code, I think there is no guarantee that (1st) cannot
happen after (2nd), although unlikely?  Can RTNL-lockless RTM_NEWTFILTER
handlers get preempted?

If (1st) happens later than (2nd), we will need to make (1st) no-op, by
detecting that we are the "old" Qdisc.  I am not sure there is any
(clean) way to do it.  I even thought about:

  (1) Get the containing Qdisc of "miniqp" we are working on, "qdisc";
  (2) Test if "qdisc == qdisc->dev_queue->qdisc_sleeping".  If false, it
      means we are the "old" Qdisc (have been replaced), and should do
      nothing.

However, for clsact Qdiscs I don't know if "miniqp" is the ingress or
egress one, so I can't container_of() during step (1) ...

Eventually I created [5,6/6].  It is a workaround indeed, in the sense
that it changes sch_api.c to avoid a mini Qdisc issue.  However I think it
makes the code correct in a relatively understandable way, without slowing
down mini_qdisc_pair_swap() or sch_handle_*gress().

Thanks,
Peilin Ye