Mon, Jul 20, 2020 at 09:50:00AM CEST, ido...@idosch.org wrote:
>On Mon, Jul 20, 2020 at 09:22:48AM +0200, Jiri Pirko wrote:
>> From: Jiri Pirko <j...@mellanox.com>
>>
>> In case the qdisc_match_from_root function() is called from non-rcu path
>> with rtnl mutex held, a suspiciout rcu usage warning appears:
>>
>> [ 241.504354] =============================
>> [ 241.504358] WARNING: suspicious RCU usage
>> [ 241.504366] 5.8.0-rc4-custom-01521-g72a7c7d549c3 #32 Not tainted
>> [ 241.504370] -----------------------------
>> [ 241.504378] net/sched/sch_api.c:270 RCU-list traversed in non-reader
>> section!!
>> [ 241.504382]
>> other info that might help us debug this:
>> [ 241.504388]
>> rcu_scheduler_active = 2, debug_locks = 1
>> [ 241.504394] 1 lock held by tc/1391:
>> [ 241.504398] #0: ffffffff85a27850 (rtnl_mutex){+.+.}-{3:3}, at:
>> rtnetlink_rcv_msg+0x49a/0xbd0
>> [ 241.504431]
>> stack backtrace:
>> [ 241.504440] CPU: 0 PID: 1391 Comm: tc Not tainted
>> 5.8.0-rc4-custom-01521-g72a7c7d549c3 #32
>> [ 241.504446] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS
>> 1.13.0-2.fc32 04/01/2014
>> [ 241.504453] Call Trace:
>> [ 241.504465] dump_stack+0x100/0x184
>> [ 241.504482] lockdep_rcu_suspicious+0x153/0x15d
>> [ 241.504499] qdisc_match_from_root+0x293/0x350
>>
>> Fix this by taking the rcu_lock for qdisc_hash iteration.
>>
>> Reported-by: Ido Schimmel <ido...@mellanox.com>
>> Signed-off-by: Jiri Pirko <j...@mellanox.com>
>> ---
>> net/sched/sch_api.c | 2 ++
>> 1 file changed, 2 insertions(+)
>>
>> diff --git a/net/sched/sch_api.c b/net/sched/sch_api.c
>> index 11ebba60da3b..c7cfd8dc6a77 100644
>> --- a/net/sched/sch_api.c
>> +++ b/net/sched/sch_api.c
>> @@ -267,10 +267,12 @@ static struct Qdisc *qdisc_match_from_root(struct
>> Qdisc *root, u32 handle)
>> root->handle == handle)
>> return root;
>>
>> + rcu_read_lock();
>> hash_for_each_possible_rcu(qdisc_dev(root)->qdisc_hash, q, hash,
>> handle) {
>> if (q->handle == handle)
>> return q;
>
>You don't unlock here, but I'm not sure it's the best fix. It's weird to
>return an object from an RCU critical section without taking a
>reference. It can also hide a bug if someone calls
>qdisc_match_from_root() without RTNL or RCU.
>
>hash_for_each_possible_rcu() is basically hlist_for_each_entry_rcu()
>which already accepts:
>
>@cond: optional lockdep expression if called from non-RCU protection.
>
>So maybe extend hash_for_each_possible_rcu() with 'cond' and pass a
>lockdep expression to see if RTNL is held?
Makes sense. Sent v2. Thanks!
>
>> }
>> + rcu_read_unlock();
>> return NULL;
>> }
>>
>> --
>> 2.21.3
>>
