When retransmitting due to T3 timeout, retransmit all the
in-flight chunks for the corresponding transport/path, including
chunks sent less then 1 rto ago.
This is the correct behaviour according to rfc4960 section 6.3.3
E3 and
"Note: Any DATA chunks that were sent to the address for which the
T3-rtx timer expired but did not fit in one MTU (rule E3 above)
should be marked for retransmission and sent as soon as cwnd
allows (normally, when a SACK arrives). ".
This fixes problems when more then one path is present and the T3
retransmission of the first chunk that timeouts stops the T3 timer
for the initial active path, leaving all the other in-flight
chunks waiting forever or until a new chunk is transmitted on the
same path and timeouts (and this will happen only if the cwnd
allows sending new chunks, but since cwnd was dropped to MTU by
the timeout => it will wait until the first heartbeat).
Example: 10 packets in flight, sent at 0.1 s intervals on the
primary path. The primary path is down and the first packet
timeouts. The first packet is retransmitted on another path, the
T3 timer for the primary path is stopped and cwnd is set to MTU.
All the other 9 in-flight packets will not be retransmitted
(unless more new packets are sent on the primary path which depend
on cwnd allowing it, and even in this case the 9 packets will be
retransmitted only after a new packet timeouts which even in the
best case would be more then RTO).
This commit reverts d0ce92910b and
also removes the now unused transport->last_rto, introduced in
b6157d8e03.
p.s The problem is not only when multiple paths are there. It
can happen in a single homed environment. If the application
stops sending data, it possible to have a hung association.
Signed-off-by: Andrei Pelinescu-Onciul <andrei@iptel.org>
Signed-off-by: Vlad Yasevich <vladislav.yasevich@hp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Current implementation of max.burst ends up limiting new
data during cwnd decay period. The decay is happening becuase
the connection is idle and we are allowed to fill the congestion
window. The point of max.burst is to limit micro-bursts in response
to large acks. This still happens, as max.burst is still applied
to each transmit opportunity. It will also apply if a very large
send is made (greater then allowed by burst).
Tested-by: Florian Niederbacher <florian.niederbacher@student.uibk.ac.at>
Signed-off-by: Vlad Yasevich <vladislav.yasevich@hp.com>
The transport last_time_used variable is rather useless.
It was only used when determining if CWND needs to be updated
due to idle transport. However, idle transport detection was
based on a Heartbeat timer and last_time_used was not incremented
when sending Heartbeats. As a result the check for cwnd reduction
was always true. We can get rid of the variable and just base
our cwnd manipulation on the HB timer (like the code comment sais).
We also have to call into the cwnd manipulation function regardless
of whether HBs are enabled or not. That way we will detect idle
transports if the user has disabled Heartbeats.
Signed-off-by: Vlad Yasevich <vladislav.yasevich@hp.com>
We currently send window update SACKs every time we free up 1 PMTU
worth of data. That a lot more SACKs then necessary. Instead, we'll
now send back the actuall window every time we send a sack, and do
window-update SACKs when a fraction of the receive buffer has been
opened. The fraction is controlled with a sysctl.
Signed-off-by: Vlad Yasevich <vladislav.yasevich@hp.com>
Recent commit 8da645e101
sctp: Get rid of an extra routing lookup when adding a transport
introduced a regression in the connection setup. The behavior was
different between IPv4 and IPv6. IPv4 case ended up working because the
route lookup routing returned a NULL route, which triggered another
route lookup later in the output patch that succeeded. In the IPv6 case,
a valid route was returned for first call, but we could not find a valid
source address at the time since the source addresses were not set on the
association yet. Thus resulted in a hung connection.
The solution is to set the source addresses on the association prior to
adding peers.
Signed-off-by: Vlad Yasevich <vladislav.yasevich@hp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This provides safety against negative optlen at the type
level instead of depending upon (sometimes non-trivial)
checks against this sprinkled all over the the place, in
each and every implementation.
Based upon work done by Arjan van de Ven and feedback
from Linus Torvalds.
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch introduces a new sysctl option to make IPv4 Address Scoping
configurable <draft-stewart-tsvwg-sctp-ipv4-00.txt>.
In networking environments where DNAT rules in iptables prerouting
chains convert destination IP's to link-local/private IP addresses,
SCTP connections fail to establish as the INIT chunk is dropped by the
kernel due to address scope match failure.
For example to support overlapping IP addresses (same IP address with
different vlan id) a Layer-5 application listens on link local IP's,
and there is a DNAT rule that maps the destination IP to a link local
IP. Such applications never get the SCTP INIT if the address-scoping
draft is strictly followed.
This sysctl configuration allows SCTP to function in such
unconventional networking environments.
Sysctl options:
0 - Disable IPv4 address scoping draft altogether
1 - Enable IPv4 address scoping (default, current behavior)
2 - Enable address scoping but allow IPv4 private addresses in init/init-ack
3 - Enable address scoping but allow IPv4 link local address in init/init-ack
Signed-off-by: Bhaskar Dutta <bhaskar.dutta@globallogic.com>
Signed-off-by: Vlad Yasevich <vladislav.yasevich@hp.com>
We had a bug that we never stored the user-defined value for
MAXSEG when setting the value on an association. Thus future
PMTU events ended up re-writing the frag point and increasing
it past user limit. Additionally, when setting the option on
the socket/endpoint, we effect all current associations, which
is against spec.
Now, we store the user 'maxseg' value along with the computed
'frag_point'. We inherit 'maxseg' from the socket at association
creation and use it as an upper limit for 'frag_point' when its
set.
Signed-off-by: Vlad Yasevich <vladislav.yasevich@hp.com>
SCTP will delay the last part of a large write due to NAGLE, if that
part is smaller then MTU. Since we are doing large writes, we might
as well send the last portion now instead of waiting untill the next
large write happens. The small portion will be sent as is regardless,
so it's better to not delay it.
This is a result of much discussions with Wei Yongjun <yjwei@cn.fujitsu.com>
and Doug Graham <dgraham@nortel.com>. Many thanks go out to them.
Signed-off-by: Vlad Yasevich <vladislav.yasevich@hp.com>
SCTP has a problem that when small chunks are used, it is possible
to exhaust the receiver buffer without fully closing receive window.
This happens due to all overhead that we have account for with small
messages. To fix this, when receive buffer is exceeded, we'll drop
the window to 0 and save the 'drop' portion. When application starts
reading data and freeing up recevie buffer space, we'll wait until
we've reached the 'drop' window and then add back this 'drop' one
mtu at a time. This worked well in testing and under stress produced
rather even recovery.
Signed-off-by: Vlad Yasevich <vladislav.yasevich@hp.com>
Currenlty, sctp breaks up user messages into fragments and
sends each fragment to the lower layer by itself. This means
that for each fragment we go all the way down the stack
and back up. This also discourages bundling of multiple
fragments when they can fit into a sigle packet (ex: due
to user setting a low fragmentation threashold).
We introduce a new command SCTP_CMD_SND_MSG and hand the
whole message down state machine. The state machine and
the side-effect parser will cork the queue, add all chunks
from the message to the queue, and then un-cork the queue
thus causing the chunks to get transmitted.
Signed-off-by: Vlad Yasevich <vladislav.yasevich@hp.com>
RFC 5061 Section 5.1 ASCONF Chunk Procedures said:
B4) Re-transmit the ASCONF Chunk last sent and if possible choose an
alternate destination address (please refer to [RFC4960],
Section 6.4.1). An endpoint MUST NOT add new parameters to this
chunk; it MUST be the same (including its Sequence Number) as
the last ASCONF sent. An endpoint MAY, however, bundle an
additional ASCONF with new ASCONF parameters with the next
Sequence Number. For details, see Section 5.5.
This patch fix to choose an alternate destination address when
re-transmit the ASCONF chunk, with some dup codes cleanup.
Signed-off-by: Wei Yongjun <yjwei@cn.fujitsu.com>
Signed-off-by: Vlad Yasevich <vladislav.yasevich@hp.com>
SCTP incorrectly doubles rto ever time a Hearbeat chunk
is generated. However RFC 4960 states:
On an idle destination address that is allowed to heartbeat, it is
recommended that a HEARTBEAT chunk is sent once per RTO of that
destination address plus the protocol parameter 'HB.interval', with
jittering of +/- 50% of the RTO value, and exponential backoff of the
RTO if the previous HEARTBEAT is unanswered.
Essentially, of if the heartbean is unacknowledged, do we double the RTO.
Signed-off-by: Vlad Yasevich <vladislav.yasevich@hp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This is a new version of my patch, now using a module parameter instead
of a sysctl, so that the option is harder to find. Please note that,
once the module is loaded, it is still possible to change the value of
the parameter in /sys/module/sctp/parameters/, which is useful if you
want to do performance comparisons without rebooting.
Computation of SCTP checksums significantly affects the performance of
SCTP. For example, using two dual-Opteron 246 connected using a Gbe
network, it was not possible to achieve more than ~730 Mbps, compared to
941 Mbps after disabling SCTP checksums.
Unfortunately, SCTP checksum offloading in NICs is not commonly
available (yet).
By default, checksums are still enabled, of course.
Signed-off-by: Lucas Nussbaum <lucas.nussbaum@ens-lyon.fr>
Signed-off-by: Vlad Yasevich <vladislav.yasevich@hp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The tsn map currently use is 4K large and is stuck inside
the sctp_association structure making memory references REALLY
expensive. What we really need is at most 4K worth of bits
so the biggest map we would have is 512 bytes. Also, the
map is only really usefull when we have gaps to store and
report. As such, starting with minimal map of say 32 TSNs (bits)
should be enough for normal low-loss operations. We can grow
the map by some multiple of 32 along with some extra room any
time we receive the TSN which would put us outside of the map
boundry. As we close gaps, we can shift the map to rebase
it on the latest TSN we've seen. This saves 4088 bytes per
association just in the map alone along savings from the now
unnecessary structure members.
Signed-off-by: Vlad Yasevich <vladislav.yasevich@hp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
sctp_is_any() function that is used to check for wildcard addresses
only looks at the address itself to determine the address family.
This function is used in the API to check the address passed in from
the user. If the user simply zerroes out the sockaddr_storage and
pass that in, we'll end up failing. So, let's try harder to determine
the address family by also checking the socket if it's possible.
Signed-off-by: Vlad Yasevich <vladislav.yasevich@hp.com>
sctp_chunks should be put on a diet. This is some of the low hanging
fruit that we can strip out. Changes all the __s8/__u8 flags to
bitfields. Saves 12 bytes per chunk.
Signed-off-by: Neil Horman <nhorman@tuxdriver.com>
Signed-off-by: Vlad Yasevich <vladislav.yasevich@hp.com>
The ipfragok flag controls whether the packet may be fragmented
either on the local host on beyond. The latter is only valid on
IPv4.
In fact, we never want to do the latter even on IPv4 when PMTU is
enabled. This is because even though we can't fragment packets
within SCTP due to the prtocol's inherent faults, we can still
fragment it at IP layer. By setting the DF bit we will improve
the PMTU process.
RFC 2960 only says that we SHOULD clear the DF bit in this case,
so we're compliant even if we set the DF bit. In fact RFC 4960
no longer has this statement.
Once we make this change, we only need to control the local
fragmentation. There is already a bit in the skb which controls
that, local_df. So this patch sets that instead of using the
ipfragok argument.
The only complication is that there isn't a struct sock object
per transport, so for IPv4 we have to resort to changing the
pmtudisc field for every packet. This should be safe though
as the protocol is single-threaded.
Note that after this patch we can remove ipfragok from the rest
of the stack too.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
