Re: [PATCH net] sched: fix infinite loop when creating tc filter

[Vlad Buslov <vladbu@nvidia.com>]

Hi Volodymyr,

On Sun 10 Oct 2021 at 09:55, Volodymyr Mytnyk <volodymyr.mytnyk@plvision.eu> wrote:
> From: Volodymyr Mytnyk <vmytnyk@marvell.com>
>
> After running a specific set of commands tc will become unresponsive:
>
>   $ ip link add dev DEV type veth
>   $ tc qdisc add dev DEV clsact
>   $ tc chain add dev DEV chain 0 ingress
>   $ tc filter del dev DEV ingress
>   $ tc filter add dev DEV ingress flower action pass
>
> When executing chain flush, the "chain->flushing" field is set
> to true, which prevents insertion of new classifier instances.
> It is unset in one place under two conditions:
>
> `refcnt - chain->action_refcnt == 0` and `!by_act`.
>
> Ignoring the by_act and action_refcnt arguments the `flushing procedure`
> will be over when refcnt is 0.
>
> But if the chain is explicitly created (e.g. `tc chain add .. chain 0 ..`)
> refcnt is set to 1 without any classifier instances. Thus the condition
> is never met and the chain->flushing field is never cleared.
> And because the default chain is `flushing` new classifiers cannot
> be added. tc_new_tfilter is stuck in a loop trying to find a chain
> where chain->flushing is false.
>
> By moving `chain->flushing = false` from __tcf_chain_put to the end
> of tcf_chain_flush will avoid the condition and the field will always
> be reset after the flush procedure.
>
> Fixes: 91052fa1c657 ("net: sched: protect chain->explicitly_created with block->lock")

Thanks for working on this!

>
> Co-developed-by: Serhiy Boiko <serhiy.boiko@plvision.eu>
> Signed-off-by: Serhiy Boiko <serhiy.boiko@plvision.eu>
> Signed-off-by: Volodymyr Mytnyk <vmytnyk@marvell.com>
> ---
>  net/sched/cls_api.c | 5 +++--
>  1 file changed, 3 insertions(+), 2 deletions(-)
>
> diff --git a/net/sched/cls_api.c b/net/sched/cls_api.c
> index d73b5c5514a9..327594cce554 100644
> --- a/net/sched/cls_api.c
> +++ b/net/sched/cls_api.c
> @@ -563,8 +563,6 @@ static void __tcf_chain_put(struct tcf_chain *chain, bool by_act,
>  	if (refcnt - chain->action_refcnt == 0 && !by_act) {
>  		tc_chain_notify_delete(tmplt_ops, tmplt_priv, chain->index,
>  				       block, NULL, 0, 0, false);
> -		/* Last reference to chain, no need to lock. */
> -		chain->flushing = false;
>  	}
>  
>  	if (refcnt == 0)
> @@ -615,6 +613,9 @@ static void tcf_chain_flush(struct tcf_chain *chain, bool rtnl_held)
>  		tcf_proto_put(tp, rtnl_held, NULL);
>  		tp = tp_next;
>  	}
> +
> +	/* Last reference to chain, no need to lock. */

But after moving the code block here you can no longer guarantee that
this is the last reference, right?

> +	chain->flushing = false;

Resetting the flag here is probably correct for actual flush use-case
(e.g. RTM_DELTFILTER message with prio==0), but can cause undesired
side-effects for other users of tcf_chain_flush(). Consider following
interaction between new filter creation and explicit chain deletion that
also uses tcf_chanin_flush():

          RTM_DELCHAIN                         RTM_NEWTFILTER
                +                                     +
                |                                     |
                |                          +----------v-----------+
                |                          |                      |
                |                          |  __tcf_block_find    |
                |                          |                      |
                |                          +----------+-----------+
                |                                     |
                |                                     |
                |                          +----------v------------+
                |                          |                       |
                |                          |    tcf_chain_get      |
                |                          |                       |
                |                          +----------+------------+
                |                                     |
       +--------v--------+                            |
       |                 |                            |
       | tcf_chain_flush |                            |
       |                 |                            |
       +--------+--------+                            |
                |                                     |
                |                          +----------v------------+
                |                          |                       |
                |                          |  tcf_chain_tp_find    |
                |                          |                       |
                |                          +----------+------------+
                |                                     |
                |                                     |tp==NULL
                |                                     |chain->flushing==false
                |                                     |
                |                     +---------------v----------------+
                |                     |                                |
                |                     |  tp_created = 1                |
                |                     |  tcf_chain_tp_insert_unique    |
                |                     |                                |
                |                     +---------------+----------------+
                |                                     |
                |                                     |
+---------------v-----------------+                   |
|                                 |                   |
|tcf_chain_put_explicitly_created |                   |
|                                 |                   |
+---------------+-----------------+                   |
                |                                     |
                v                                     v

In this example tc_new_tfilter() holds chain reference during flush. If
flush finishes concurrently before the check for chain->flushing, the
chain reference counter will not reach 0 (because new filter creation
code will not back off and release the reference). In the described
example tc_chain_notify_delete() will not be called which will confuse
any userland code that expects to receive delete chain notification
after sending RTM_DELCHAIN message.

With these considerations I can propose following approach to fix the
issue:

1. Extend tcf_chain_flush() with additional boolean argument and only
call it with 'true' value from tc_del_tfilter(). (or create helper
function that calls tcf_chain_flush() and then resets chain->flushing
flag)

2. Reset the 'flushing' flag when new argument is true.

3. Wrap the 'flushing' flag reset code in filter_chain_lock critical
section.

>  }
>  
>  static int tcf_block_setup(struct tcf_block *block,