[ Upstream commit a76c0adf60f6ca5ff3481992e4ea0383776b24d2 ] When checking whether a DATA chunk fits into the estimated rwnd a full sizeof(struct sk_buff) is added to the needed chunk size. This quickly exhausts the available rwnd space and leads to packets being sent which are much below the PMTU limit. This can lead to much worse performance. The reason for this behaviour was to avoid putting too much memory pressure on the receiver. The concept is not completely irational because a Linux receiver does in fact clone an skb for each DATA chunk delivered. However, Linux also reserves half the available socket buffer space for data structures therefore usage of it is already accounted for. When proposing to change this the last time it was noted that this behaviour was introduced to solve a performance issue caused by rwnd overusage in combination with small DATA chunks. Trying to reproduce this I found that with the sk_buff overhead removed, the performance would improve significantly unless socket buffer limits are increased. The following numbers have been gathered using a patched iperf supporting SCTP over a live 1 Gbit ethernet network. The -l option was used to limit DATA chunk sizes. The numbers listed are based on the average of 3 test runs each. Default values have been used for sk_(r|w)mem. Chunk Size Unpatched No Overhead ------------------------------------- 4 15.2 Kbit [!] 12.2 Mbit [!] 8 35.8 Kbit [!] 26.0 Mbit [!] 16 95.5 Kbit [!] 54.4 Mbit [!] 32 106.7 Mbit 102.3 Mbit 64 189.2 Mbit 188.3 Mbit 128 331.2 Mbit 334.8 Mbit 256 537.7 Mbit 536.0 Mbit 512 766.9 Mbit 766.6 Mbit 1024 810.1 Mbit 808.6 Mbit Signed-off-by: Thomas Graf <tgraf@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net> Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de> Change-Id: Ifd4199603bbc8851051b45299b7cf73b07b4daac Reviewed-on: http://git-master/r/74220 Reviewed-by: Varun Wadekar <vwadekar@nvidia.com> Tested-by: Varun Wadekar <vwadekar@nvidia.com>
[ Upstream commit 2692ba61a82203404abd7dd2a027bda962861f74 ] Commit 8ffd3208 voids the previous patches f6778aab and 810c0719 for limiting the autoclose value. If userspace passes in -1 on 32-bit platform, the overflow check didn't work and autoclose would be set to 0xffffffff. This patch defines a max_autoclose (in seconds) for limiting the value and exposes it through sysctl, with the following intentions. 1) Avoid overflowing autoclose * HZ. 2) Keep the default autoclose bound consistent across 32- and 64-bit platforms (INT_MAX / HZ in this patch). 3) Keep the autoclose value consistent between setsockopt() and getsockopt() calls. Suggested-by: Vlad Yasevich <vladislav.yasevich@hp.com> Signed-off-by: Xi Wang <xi.wang@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net> Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de> Change-Id: I74631fca684d7a475f1eed5f6823c11b62bbd2bc Reviewed-on: http://git-master/r/74219 Reviewed-by: Varun Wadekar <vwadekar@nvidia.com> Tested-by: Varun Wadekar <vwadekar@nvidia.com>
Attempt to reduce the number of IP packets emitted in response to single SCTP packet (2e3216cd) introduced a complication - if a packet contains two COOKIE_ECHO chunks and nothing else then SCTP state machine corks the socket while processing first COOKIE_ECHO and then loses the association and forgets to uncork the socket. To deal with the issue add new SCTP command which can be used to set association explictly. Use this new command when processing second COOKIE_ECHO chunk to restore the context for SCTP state machine. Signed-off-by: Max Matveev <makc@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts: net/bluetooth/l2cap_core.c
Packets to devices without NETIF_F_SCTP_CSUM (including NETIF_F_NO_CSUM) should be properly checksummed because the packets can be diverted or rerouted after construction. This still leaves packets diverted from NETIF_F_SCTP_CSUM-enabled devices with broken checksums. Fixing this needs implementing software offload fallback in networking core. For users of sctp_checksum_disable, skb->ip_summed should be left as CHECKSUM_NONE and not CHECKSUM_UNNECESSARY as per include/linux/skbuff.h. Signed-off-by: Michał Mirosław <mirq-linux@rere.qmqm.pl> Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts: net/bluetooth/l2cap_core.c
Trigger user ABORT if application closes a socket which has data queued on the socket receive queue or chunks waiting on the reassembly or ordering queue as this would imply data being lost which defeats the point of a graceful shutdown. This behavior is already practiced in TCP. We do not check the input queue because that would mean to parse all chunks on it to look for unacknowledged data which seems too much of an effort. Control chunks or duplicated chunks may also be in the input queue and should not be stopping a graceful shutdown. Signed-off-by: Thomas Graf <tgraf@infradead.org> Acked-by: Vlad Yasevich <vladislav.yasevich@hp.com> Signed-off-by: David S. Miller <davem@davemloft.net>
When initiating a graceful shutdown while having data chunks on the retransmission queue with a peer which is in zero window mode the shutdown is never completed because the retransmission error count is reset periodically by the following two rules: - Do not timeout association while doing zero window probe. - Reset overall error count when a heartbeat request has been acknowledged. The graceful shutdown will wait for all outstanding TSN to be acknowledged before sending the SHUTDOWN request. This never happens due to the peer's zero window not acknowledging the continuously retransmitted data chunks. Although the error counter is incremented for each failed retransmission, the receiving of the SACK announcing the zero window clears the error count again immediately. Also heartbeat requests continue to be sent periodically. The peer acknowledges these requests causing the error counter to be reset as well. This patch changes behaviour to only reset the overall error counter for the above rules while not in shutdown. After reaching the maximum number of retransmission attempts, the T5 shutdown guard timer is scheduled to give the receiver some additional time to recover. The timer is stopped as soon as the receiver acknowledges any data. The issue can be easily reproduced by establishing a sctp association over the loopback device, constantly queueing data at the sender while not reading any at the receiver. Wait for the window to reach zero, then initiate a shutdown by killing both processes simultaneously. The association will never be freed and the chunks on the retransmission queue will be retransmitted indefinitely. Signed-off-by: Thomas Graf <tgraf@infradead.org> Acked-by: Vlad Yasevich <vladislav.yasevich@hp.com> Signed-off-by: David S. Miller <davem@davemloft.net>
We forgot to send up SCTP_SENDER_DRY_EVENT notification when user app subscribes to this event, and there is no data to be sent or retransmit. This is required by the Socket API and used by the DTLS/SCTP implementation. Reported-by: Michael Tüxen <Michael.Tuexen@lurchi.franken.de> Signed-off-by: Wei Yongjun <yjwei@cn.fujitsu.com> Tested-by: Robin Seggelmann <seggelmann@fh-muenster.de> Signed-off-by: David S. Miller <davem@davemloft.net>
Current tcp/udp/sctp global memory limits are not taking into account hugepages allocations, and allow 50% of ram to be used by buffers of a single protocol [ not counting space used by sockets / inodes ...] Lets use nr_free_buffer_pages() and allow a default of 1/8 of kernel ram per protocol, and a minimum of 128 pages. Heavy duty machines sysadmins probably need to tweak limits anyway. References: https://bugzilla.stlinux.com/show_bug.cgi?id=38032 Reported-by: starlight <starlight@binnacle.cx> Suggested-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>