On Thu, Dec 31, 2015 at 1:41 AM, Marcelo Ricardo Leitner
<marcelo.leitner@xxxxxxxxx> wrote:
> On Wed, Dec 30, 2015 at 11:50:46PM +0800, Xin Long wrote:
> ...
>> +void sctp_hash_transport(struct sctp_transport *t)
>> +{
>> + struct sctp_sockaddr_entry *addr;
>> + struct sctp_hash_cmp_arg arg;
>> +
>> + addr = list_entry(t->asoc->base.bind_addr.address_list.next,
>> + struct sctp_sockaddr_entry, list);
>> + arg.laddr = &addr->a;
>> + arg.paddr = &t->ipaddr;
>> + arg.net = sock_net(t->asoc->base.sk);
>> +
>> +reinsert:
>> + if (rhashtable_lookup_insert_key(&sctp_transport_hashtable, &arg,
>> + &t->node, sctp_hash_params) == -EBUSY)
>> + goto reinsert;
>> +}
>
> This is the nasty situation I mentioned in previous email. It seems that
> a stress test can trigger a double rehash and cause an entry to not be
> added.
>
> This is in fact very near some bugs you caught on rhashtable in the past
> few days/couple of weeks tops.
>
> I'm actually against this loop as is. I may have not been clear with Xin
> about not adding my signature to the patchset due to this.
>
> Please take a look at Xin's emails on thread 'rhashtable: Prevent
> spurious EBUSY errors on insertion' about this particular situation.
> Cc'ing Herbert as he wanted to see the patches for that issue.
>
> Marcelo
>
without this 'reinsert'.
we can reproduce this issue like this:
1. download the attachment
2. cp sctperf.tar.gz to server and client hosts
3. in each hosts.
#make
4. in server:
#sh saddr.sh $ethx
#./ss
5. in client:
#sh caddr.sh $ethx
#ulimit -n 20000
#./cc
when the number of connections reach about 1600, this issue will be triggered.
Attachment:
sctperf.tar.gz
Description: GNU Zip compressed data
