From mboxrd@z Thu Jan 1 00:00:00 1970 From: David Miller Subject: Re: [PATCH net v2] net: sctp: wake up all assocs if sndbuf policy is per socket Date: Tue, 08 Apr 2014 13:08:49 -0400 (EDT) Message-ID: <20140408.130849.2070157751349863276.davem@davemloft.net> References: <1396970773-17500-1-git-send-email-dborkman@redhat.com> Mime-Version: 1.0 Content-Type: Text/Plain; charset=us-ascii Content-Transfer-Encoding: 7bit Cc: netdev@vger.kernel.org, linux-sctp@vger.kernel.org, tgraf@suug.ch, nhorman@tuxdriver.com, vyasevic@redhat.com To: dborkman@redhat.com Return-path: Received: from shards.monkeyblade.net ([149.20.54.216]:43942 "EHLO shards.monkeyblade.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1756359AbaDHRIw (ORCPT ); Tue, 8 Apr 2014 13:08:52 -0400 In-Reply-To: <1396970773-17500-1-git-send-email-dborkman@redhat.com> Sender: netdev-owner@vger.kernel.org List-ID: From: Daniel Borkmann Date: Tue, 8 Apr 2014 17:26:13 +0200 > 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 > accounting, 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 Applied and queued up for -stable, thanks.