netdev - Re: [PATCH net-next v6 1/3] net/sched: act

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <334b4af2-1ac0-d28b-f1a5-b9b604a9ba80@mellanox.com>
Date:   Sun, 8 Mar 2020 10:11:37 +0200
From:   Paul Blakey <paulb@...lanox.com>
To:     Eric Dumazet <eric.dumazet@...il.com>,
        Saeed Mahameed <saeedm@...lanox.com>,
        Oz Shlomo <ozsh@...lanox.com>,
        Jakub Kicinski <jakub.kicinski@...ronome.com>,
        Vlad Buslov <vladbu@...lanox.com>,
        David Miller <davem@...emloft.net>,
        "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
        Jiri Pirko <jiri@...lanox.com>, Roi Dayan <roid@...lanox.com>
Subject: Re: [PATCH net-next v6 1/3] net/sched: act_ct: Create nf flow table
 per zone

iirc I did the spin lock bh because we can be called from queue work rcu handler , so I wanted to disable soft irq.

I got a possible deadlock splat for that.



On 3/7/2020 10:53 PM, Eric Dumazet wrote:

>
> On 3/7/20 12:12 PM, Eric Dumazet wrote:
>>
>> On 3/3/20 7:57 AM, Paul Blakey wrote:
>>> Use the NF flow tables infrastructure for CT offload.
>>>
>>> Create a nf flow table per zone.
>>>
>>> Next patches will add FT entries to this table, and do
>>> the software offload.
>>>
>>> Signed-off-by: Paul Blakey <paulb@...lanox.com>
>>> Reviewed-by: Jiri Pirko <jiri@...lanox.com>
>>> ---
>>>   v4->v5:
>>>     Added reviewed by Jiri, thanks!
>>>   v3->v4:
>>>     Alloc GFP_ATOMIC
>>>   v2->v3:
>>>     Ditch re-locking to alloc, and use atomic allocation
>>>   v1->v2:
>>>     Use spin_lock_bh instead of spin_lock, and unlock for alloc (as it can sleep)
>>>     Free ft on last tc act instance instead of last instance + last offloaded tuple,
>>>     this removes cleanup cb and netfilter patches, and is simpler
>>>     Removed accidental mlx5/core/en_tc.c change
>>>     Removed reviewed by Jiri - patch changed
>>>
>>> +	err = nf_flow_table_init(&ct_ft->nf_ft);
>> This call is going to allocate a rhashtable (GFP_KERNEL allocations that might sleep)
>>
>> Since you still hold zones_lock spinlock, a splat should occur.
>>
>> "BUG: sleeping function called from invalid context in  ..."
>>
>> DEBUG_ATOMIC_SLEEP=y is your friend.
>>
>> And it is always a good thing to make sure a patch does not trigger a lockdep splat
>>
>> CONFIG_PROVE_LOCKING=y
> Also abusing a spinlock and GFP_ATOMIC allocations in control path is highly discouraged.
>
> I can not test the following fix, any objections before I submit this officially ?
>
> diff --git a/net/sched/act_ct.c b/net/sched/act_ct.c
> index 23eba61f0f819212a3522c3c63b938d0b8d997e2..3d9e678d7d5336f1746035745b091bea0dcb5fdd 100644
> --- a/net/sched/act_ct.c
> +++ b/net/sched/act_ct.c
> @@ -35,15 +35,15 @@
>  
>  static struct workqueue_struct *act_ct_wq;
>  static struct rhashtable zones_ht;
> -static DEFINE_SPINLOCK(zones_lock);
> +static DEFINE_MUTEX(zones_mutex);
>  
>  struct tcf_ct_flow_table {
>         struct rhash_head node; /* In zones tables */
>  
>         struct rcu_work rwork;
>         struct nf_flowtable nf_ft;
> +       refcount_t ref;
>         u16 zone;
> -       u32 ref;
>  
>         bool dying;
>  };
> @@ -64,14 +64,15 @@ static int tcf_ct_flow_table_get(struct tcf_ct_params *params)
>         struct tcf_ct_flow_table *ct_ft;
>         int err = -ENOMEM;
>  
> -       spin_lock_bh(&zones_lock);
> +       mutex_lock(&zones_mutex);
>         ct_ft = rhashtable_lookup_fast(&zones_ht, &params->zone, zones_params);
> -       if (ct_ft)
> -               goto take_ref;
> +       if (ct_ft && refcount_inc_not_zero(&ct_ft->ref))
> +               goto out_unlock;
>  
> -       ct_ft = kzalloc(sizeof(*ct_ft), GFP_ATOMIC);
> +       ct_ft = kzalloc(sizeof(*ct_ft), GFP_KERNEL);
>         if (!ct_ft)
>                 goto err_alloc;
> +       refcount_set(&ct_ft->ref, 1);
>  
>         ct_ft->zone = params->zone;
>         err = rhashtable_insert_fast(&zones_ht, &ct_ft->node, zones_params);
> @@ -84,10 +85,9 @@ static int tcf_ct_flow_table_get(struct tcf_ct_params *params)
>                 goto err_init;
>  
>         __module_get(THIS_MODULE);
> -take_ref:
> +out_unlock:
>         params->ct_ft = ct_ft;
> -       ct_ft->ref++;
> -       spin_unlock_bh(&zones_lock);
> +       mutex_unlock(&zones_mutex);
>  
>         return 0;
>  
> @@ -96,7 +96,7 @@ static int tcf_ct_flow_table_get(struct tcf_ct_params *params)
>  err_insert:
>         kfree(ct_ft);
>  err_alloc:
> -       spin_unlock_bh(&zones_lock);
> +       mutex_unlock(&zones_mutex);
>         return err;
>  }
>  
> @@ -116,13 +116,11 @@ static void tcf_ct_flow_table_put(struct tcf_ct_params *params)
>  {
>         struct tcf_ct_flow_table *ct_ft = params->ct_ft;
>  
> -       spin_lock_bh(&zones_lock);
> -       if (--params->ct_ft->ref == 0) {
> +       if (refcount_dec_and_test(&params->ct_ft->ref)) {
>                 rhashtable_remove_fast(&zones_ht, &ct_ft->node, zones_params);
>                 INIT_RCU_WORK(&ct_ft->rwork, tcf_ct_flow_table_cleanup_work);
>                 queue_rcu_work(act_ct_wq, &ct_ft->rwork);
>         }
> -       spin_unlock_bh(&zones_lock);
>  }
>  
>  static void tcf_ct_flow_table_add(struct tcf_ct_flow_table *ct_ft,
>