[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CANn89i+fo=SUqP01vNwNhZiDYo-vi5xst26V8JpYh8oJjSNr7Q@mail.gmail.com>
Date: Thu, 7 Dec 2023 20:14:21 +0100
From: Eric Dumazet <edumazet@...gle.com>
To: Kui-Feng Lee <sinquersw@...il.com>
Cc: David Ahern <dsahern@...nel.org>, patchwork-bot+netdevbpf@...nel.org,
Kui-Feng Lee <thinker.li@...il.com>, davem@...emloft.net, kuba@...nel.org,
pabeni@...hat.com, netdev@...r.kernel.org, eric.dumazet@...il.com
Subject: Re: [PATCH net-next] ipv6: add debug checks in fib6_info_release()
On Thu, Dec 7, 2023 at 8:08 PM Kui-Feng Lee <sinquersw@...il.com> wrote:
>
>
>
> On 12/7/23 11:05, Eric Dumazet wrote:
> > On Thu, Dec 7, 2023 at 8:00 PM Kui-Feng Lee <sinquersw@...il.com> wrote:
> >>
> >>
> >>
> >> On 12/7/23 10:40, Kui-Feng Lee wrote:
> >>>
> >>>
> >>> On 12/7/23 10:36, Kui-Feng Lee wrote:
> >>>>
> >>>>
> >>>> On 12/7/23 10:25, David Ahern wrote:
> >>>>> On 12/7/23 11:22 AM, Eric Dumazet wrote:
> >>>>>> Feel free to amend the patch, but the issue is that we insert a fib
> >>>>>> gc_link to a list, then free the fi6 object without removing it first
> >>>>>> from the external list.
> >>>>>
> >>>>> yes, move the insert down:
> >>>>>
> >>>>> diff --git a/net/ipv6/route.c b/net/ipv6/route.c
> >>>>> index b132feae3393..7257ba0e72b7 100644
> >>>>> --- a/net/ipv6/route.c
> >>>>> +++ b/net/ipv6/route.c
> >>>>> @@ -3762,12 +3762,6 @@ static struct fib6_info
> >>>>> *ip6_route_info_create(struct fib6_config *cfg,
> >>>>> if (cfg->fc_flags & RTF_ADDRCONF)
> >>>>> rt->dst_nocount = true;
> >>>>>
> >>>>> - if (cfg->fc_flags & RTF_EXPIRES)
> >>>>> - fib6_set_expires_locked(rt, jiffies +
> >>>>> -
> >>>>> clock_t_to_jiffies(cfg->fc_expires));
> >>>>> - else
> >>>>> - fib6_clean_expires_locked(rt);
> >>>>> -
> >>>>
> >>>> fib6_set_expires_locked() here actually doesn't insert a fib gc_link
> >>>> since rt->fib6_table is not assigned yet. The gc_link will
> >>>> be inserted by fib6_add() being called later.
> >>>>
> >>>>
> >>>>> if (cfg->fc_protocol == RTPROT_UNSPEC)
> >>>>> cfg->fc_protocol = RTPROT_BOOT;
> >>>>> rt->fib6_protocol = cfg->fc_protocol;
> >>>>> @@ -3824,6 +3818,12 @@ static struct fib6_info
> >>>>> *ip6_route_info_create(struct fib6_config *cfg,
> >>>>> } else
> >>>>> rt->fib6_prefsrc.plen = 0;
> >>>>>
> >>>>> +
> >>>>> + if (cfg->fc_flags & RTF_EXPIRES)
> >>>>> + fib6_set_expires_locked(rt, jiffies +
> >>>>> +
> >>>>> clock_t_to_jiffies(cfg->fc_expires));
> >>>>> + else
> >>>>> + fib6_clean_expires_locked(rt);
> >>>>> return rt;
> >>>>> out:
> >>>>> fib6_info_release(rt);
> >>>>
> >>>> However, this should fix the warning messages.
> >>> Just realize this cause inserting the gc_link twice. fib6_add()
> >>> will try to add it again!
> >>
> >> I made a minor change to the patch that fix warning messages
> >> provided by David Ahern. Will send an official patch later.
> >>
> >> --- a/net/ipv6/route.c
> >> +++ b/net/ipv6/route.c
> >> @@ -3762,17 +3762,10 @@ static struct fib6_info
> >> *ip6_route_info_create(struct fib6_config *cfg,
> >> if (cfg->fc_flags & RTF_ADDRCONF)
> >> rt->dst_nocount = true;
> >>
> >> - if (cfg->fc_flags & RTF_EXPIRES)
> >> - fib6_set_expires_locked(rt, jiffies +
> >> -
> >> clock_t_to_jiffies(cfg->fc_expires));
> >> - else
> >> - fib6_clean_expires_locked(rt);
> >> -
> >> if (cfg->fc_protocol == RTPROT_UNSPEC)
> >> cfg->fc_protocol = RTPROT_BOOT;
> >> rt->fib6_protocol = cfg->fc_protocol;
> >>
> >> - rt->fib6_table = table;
> >> rt->fib6_metric = cfg->fc_metric;
> >> rt->fib6_type = cfg->fc_type ? : RTN_UNICAST;
> >> rt->fib6_flags = cfg->fc_flags & ~RTF_GATEWAY;
> >> @@ -3824,6 +3817,17 @@ static struct fib6_info
> >> *ip6_route_info_create(struct fib6_config *cfg,
> >> } else
> >> rt->fib6_prefsrc.plen = 0;
> >>
> >> + if (cfg->fc_flags & RTF_EXPIRES)
> >> + fib6_set_expires_locked(rt, jiffies +
> >> +
> >> clock_t_to_jiffies(cfg->fc_expires));
> >> + else
> >> + fib6_clean_expires_locked(rt);
> >
> > Note that I do not see why we call fib6_clean_expires_locked() on a
> > freshly allocated object.
> >
> > f6i->expires should already be zero...
> >
>
> Agree! I will remove it as well.
Please also add the following, if really we keep fib6_has_expires()
current definition.
diff --git a/include/net/ip6_fib.h b/include/net/ip6_fib.h
index e1e7a894863a7891610ce5afb2034473cc208d3e..95ed495c3a4028457baf1503c367d2e7a6e14770
100644
--- a/include/net/ip6_fib.h
+++ b/include/net/ip6_fib.h
@@ -329,7 +329,6 @@ static inline bool fib6_info_hold_safe(struct
fib6_info *f6i)
static inline void fib6_info_release(struct fib6_info *f6i)
{
if (f6i && refcount_dec_and_test(&f6i->fib6_ref)) {
- DEBUG_NET_WARN_ON_ONCE(fib6_has_expires(f6i));
DEBUG_NET_WARN_ON_ONCE(!hlist_unhashed(&f6i->gc_link));
call_rcu(&f6i->rcu, fib6_info_destroy_rcu);
}
Powered by blists - more mailing lists