Re: [PATCH net] net/sched: flower: fix infinite loop in fl_walk()

[Cong Wang <xiyou.wangcong@gmail.com>]

On Thu, Jun 20, 2019 at 5:52 AM Davide Caratti <dcaratti@redhat.com> wrote:
>
> hello Cong, thanks for reading.
>
> On Wed, 2019-06-19 at 15:04 -0700, Cong Wang wrote:
> > On Wed, Jun 19, 2019 at 2:10 PM Davide Caratti <dcaratti@redhat.com> wrote:
> > > on some CPUs (e.g. i686), tcf_walker.cookie has the same size as the IDR.
> > > In this situation, the following script:
> > >
> > >  # tc filter add dev eth0 ingress handle 0xffffffff flower action ok
> > >  # tc filter show dev eth0 ingress
> > >
> > > results in an infinite loop. It happened also on other CPUs (e.g x86_64),
> > > before commit 061775583e35 ("net: sched: flower: introduce reference
> > > counting for filters"), because 'handle' + 1 made the u32 overflow before
> > > it was assigned to 'cookie'; but that commit replaced the assignment with
> > > a self-increment of 'cookie', so the problem was indirectly fixed.
> >
> > Interesting... Is this really specific to cls_flower? To me it looks like
> > a bug of idr_*_ul() API's, especially for idr_for_each_entry_ul().
>
> good question, I have to investigate this better (idr_for_each_entry_ul()
> expands in a iteration of idr_get_next_ul()). It surely got in cls_flower
> when it was converted to use IDRs, but it's true that there might be other
> points in TC where IDR are used and the same pattern is present (see
> below).


Yeah, this means we probably want to fix it in idr_get_next_ul() or its
callers like idr_for_each_entry_ul().

>
> > Can you test if the following command has the same problem on i386?
> >
> > tc actions add action ok index 4294967295
>
> the action is added, but then reading it back results in an infinite loop.
> And again, the infinite loop happens on i686 and not on x86_64. I will try
> to see where's the problem also here.

Right, this is what I expect, thanks for confirming it.

I am not sure it is better to handle this overflow inside idr_get_next_ul()
or just let its callers to handle it. According to the comments above
idr_get_next_ul() it sounds like it is not expected to overflow, so...

diff --git a/lib/idr.c b/lib/idr.c
index c34e256d2f01..a38f5e391cec 100644
--- a/lib/idr.c
+++ b/lib/idr.c
@@ -267,6 +267,9 @@ void *idr_get_next_ul(struct idr *idr, unsigned
long *nextid)
        if (!slot)
                return NULL;

+       /* overflow */
+       if (iter.index < id)
+               return NULL;
        *nextid = iter.index + base;
        return rcu_dereference_raw(*slot);
 }