On 04/08/2014 10:52 AM, Daniel Borkmann wrote:
> On 04/08/2014 04:41 PM, Vlad Yasevich wrote:
>> On 04/08/2014 09:33 AM, Daniel Borkmann wrote:
>>> SCTP charges chunks for wmem accounting via skb->truesize in
>>> sctp_set_owner_w(), and sctp_wfree() respectively as the
>>> reverse operation. If a sender runs out of wmem, it needs to
>>> wait via sctp_wait_for_sndbuf(), and gets woken up by a call
>>> to __sctp_write_space() mostly via sctp_wfree().
>>>
>>> __sctp_write_space() is being called per association. Although
>>> we assign sk->sk_write_space() to sctp_write_space(), which
>>> is then being done per socket, it is only used if send space
>>> is increased per socket option (SO_SNDBUF), as SOCK_USE_WRITE_QUEUE
>>> is set and therefore not invoked in sock_wfree().
>>>
>>> Commit 4c3a5bdae293 ("sctp: Don't charge for data in sndbuf
>>> again when transmitting packet") fixed an issue where in case
>>> sctp_packet_transmit() manages to queue up more than sndbuf
>>> bytes, sctp_wait_for_sndbuf() will never be woken up again
>>> unless it is interrupted by a signal. However, a still
>>> remaining issue is that if net.sctp.sndbuf_policy=0, that is
>>> accounting per socket, and one-to-many sockets are in use,
>>> the reclaimed write space from sctp_wfree() is 'unfairly'
>>> handed back on the server to the association that is the lucky
>>> one to be woken up again via __sctp_write_space(), while
>>> the remaining associations are never be woken up again
>>> (unless by a signal).
>>>
>>> The effect disappears with net.sctp.sndbuf_policy=1, that
>>> is wmem accounting per association, as it guarantees a fair
>>> share of wmem among associations.
>>>
>>> Therefore, if we have reclaimed memory in case of per socket
>>> accouting, wake all related associations to a socket in a
>>> fair manner, that is, traverse the socket association list
>>> starting from the current neighbour of the association and
>>> issue a __sctp_write_space() to everyone until we end up
>>> waking ourselves. This guarantees that no association is
>>> preferred over another and even if more associations are
>>> taken into the one-to-many session, all receivers will get
>>> messages from the server and are not stalled forever on
>>> high load. This setting still leaves the advantage of per
>>> socket accounting in touch as an association can still use
>>> up global limits if unused by others.
>>>
>>> Fixes: 4eb701dfc618 ("[SCTP] Fix SCTP sendbuffer accouting.")
>>> Signed-off-by: Daniel Borkmann <dborkman@xxxxxxxxxx>
>>> Cc: Thomas Graf <tgraf@xxxxxxx>
>>> Cc: Neil Horman <nhorman@xxxxxxxxxxxxx>
>>> ---
>>> [ When net-next opens up again, we need to think how
>>> we can ideally make a new list interface and simplify
>>> both open-coded list traversals. ]
>>>
>>> net/sctp/socket.c | 31 ++++++++++++++++++++++++++++++-
>>> 1 file changed, 30 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/net/sctp/socket.c b/net/sctp/socket.c
>>> index 981aaf8..a4c8c1f 100644
>>> --- a/net/sctp/socket.c
>>> +++ b/net/sctp/socket.c
>>> @@ -6593,6 +6593,35 @@ static void __sctp_write_space(struct
>>> sctp_association *asoc)
>>> }
>>> }
>>>
>>> +static void sctp_wake_up_waiters(struct sock *sk,
>>> + struct sctp_association *asoc)
>>> +{
>>> + struct sctp_association *tmp = asoc;
>>> +
>>> + /* We do accounting for the sndbuf space per association,
>>> + * so we only need to wake our own association.
>>> + */
>>> + if (asoc->ep->sndbuf_policy)
>>> + return __sctp_write_space(asoc);
>>> +
>>> + /* Accounting for the sndbuf space is per socket, so we need
>>> + * to wake up others, try to be fair and in case of other
>>> + * associations, let them have a go first instead of just
>>> + * doing a sctp_write_space() call.
>>> + */
>>
>> May be a note saying that we are here only when association frees
>> queued up chunks and thus we are under lock and the list is guaranteed
>> not to change.
>
> Ok, will add that to the comment and respin, thanks Vlad.
>
>>> + for (tmp = list_next_entry(tmp, asocs); 1;
>>
>> Why not change the stop condition to tmp == asoc. It should work
>> since it will not be head pointer.
>
> If I see this correctly, wouldn't we then exclude to eventually
> call __sctp_write_space(tmp) on ourselves as we also need to make
> sure to wake us up?
>
Ahh, yes. You are right.
This is yet another list traversal with skip_head. I am going to
resurrect that code for net-next.
-vlad
>> -vlad
>>
>>> + tmp = list_next_entry(tmp, asocs)) {
>>> + /* Manually skip the head element. */
>>> + if (&tmp->asocs == &((sctp_sk(sk))->ep->asocs))
>>> + continue;
>>> + /* Wake up association. */
>>> + __sctp_write_space(tmp);
>>> + /* We've reached the end. */
>>> + if (tmp == asoc)
>>> + break;
>>> + }
>>> +}
>>> +
>>> /* Do accounting for the sndbuf space.
>>> * Decrement the used sndbuf space of the corresponding association
>>> by the
>>> * data size which was just transmitted(freed).
>>> @@ -6620,7 +6649,7 @@ static void sctp_wfree(struct sk_buff *skb)
>>> sk_mem_uncharge(sk, skb->truesize);
>>>
>>> sock_wfree(skb);
>>> - __sctp_write_space(asoc);
>>> + sctp_wake_up_waiters(sk, asoc);
>>>
>>> sctp_association_put(asoc);
>>> }
>>>
>>
--
To unsubscribe from this list: send the line "unsubscribe linux-sctp" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
