Le 18/10/2016 à 12:06, Nicolas Dichtel a écrit :
> Le 18/10/2016 à 10:47, Florian Westphal a écrit :
>> Nicolas Dichtel <nicolas.dich...@6wind.com> wrote:
>>> After commit b87a2f9199ea ("netfilter: conntrack: add gc worker to remove
>>> timed-out entries"), netlink conntrack deletion events may be sent with a
>>> huge delay (5 minutes).
>>>
>>> There is two ways to evict conntrack:
>>>  - during a conntrack lookup;
>>>  - during a conntrack dump.
>>> Let's do a full scan of conntrack entries after a period of inactivity
>>> (no conntrack lookup).
>>>
>>> CC: Florian Westphal <f...@strlen.de>
>>> Signed-off-by: Nicolas Dichtel <nicolas.dich...@6wind.com>
>>> ---
>>>
>>> Here is another proposal to try to fix the problem.
>>> Comments are welcomed,
>>> Nicolas
>>
>> Hmm, I don't think its good idea in practice.
>> If goal is to avoid starving arbitrary 'dead' ct for too long,
>> then simple ping will defeat the logic here, because...
>>
>>>  net/netfilter/nf_conntrack_core.c | 11 +++++++++--
>>>  1 file changed, 9 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/net/netfilter/nf_conntrack_core.c 
>>> b/net/netfilter/nf_conntrack_core.c
>>> index ba6a1d421222..3dbb27bd9582 100644
>>> --- a/net/netfilter/nf_conntrack_core.c
>>> +++ b/net/netfilter/nf_conntrack_core.c
>>> @@ -87,6 +87,7 @@ static __read_mostly bool nf_conntrack_locks_all;
>>>  #define GC_MAX_BUCKETS             8192u
>>>  #define GC_INTERVAL                (5 * HZ)
>>>  #define GC_MAX_EVICTS              256u
>>> +static bool gc_full_scan = true;
>>>  
>>>  static struct conntrack_gc_work conntrack_gc_work;
>>>  
>>> @@ -511,6 +512,7 @@ ____nf_conntrack_find(struct net *net, const struct 
>>> nf_conntrack_zone *zone,
>>>     unsigned int bucket, hsize;
>>>  
>>>  begin:
>>> +   gc_full_scan = false;
>>
>> ... we do periodic lookup (but always in same slot), so no full scan is
>> triggered.
> Yes, I was wondering about that. My first idea was to have that bool per 
> bucket
> and force a scan of the bucket instead of the whole table.
FYI, I'am off for about two weeks, but we really need to find a way to fix that
before the release goes out.

Reply via email to