This patch removes a bug in the current code. I agree with Andrea's comment
that there is a problem here but the way it is treated does not fix it.
The problem is that whenever Ack Ratio > cwnd, starvation/deadlock occurs:
* the receiver will not send an Ack until (Ack Ratio - cwnd) data packets
have arrived;
* the sender will not send any data packet before the receipt of an Ack
advances the send window.
The only way that the connection then progresses was via RTO timeout. In one
extreme case (bulk transfer), it was observed that this happened for every single
packet; i.e. hundreds of packets, each a RTO timeout of 1..3 seconds apart:
a transfer which normally would take a fraction of a second thus grew to
several minutes.
The solution taken by this approach is to observe the relation
"Ack Ratio <= cwnd"
by using the constraint (1) from RFC 4341, 6.1.2; i.e. set
Ack Ratio = ceil(cwnd / 2)
and update it whenever either Ack Ratio or cwnd change. This ensures that
the deadlock problem can not arise.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Since it makes not sense to assign negative values to Ack Ratio, this
patch disallows this possibility.
As a consequence, a Bug test for negative Ack Ratio values becomes obsolete.
Furthermore, a check against overflow (as Ack Ratio may not exceed 2 bytes,
due to RFC 4340, 11.3) has been added.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In CCID2 the receiver-history is sorted in ascending order of sequence number,
but the processing of received Ack Vectors requires the list traversal in the
opposite direction.
The current code has a bug in this regard: the list traversal is upwards. As a
consequence, only Ack Vectors with a run length of 1 will pass, in all other
Ack Vectors the remaining (acked) sequence numbers are missed, and may later
falsely be identified as lost.
Note: This bug is only visible when Ack Ratio > 1, since otherwise the run
lengths of Ack Vectors are 0.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This is reduces the length of the struct ackvec/ackvec_record fields. It is
a purely text-based replacement:
s#dccpavr_#avr_#g;
s#dccpav_#av_#g;
and increases readability somewhat.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Small update with regard to RFC 4340 (references added as documentation):
on Requests, Ack Vectors / Elapsed Time should be ignored.
Length handling of Elapsed Time also simplified.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This cleans up the consequences of an earlier patch which
introduced the `if IP_DCCP' clause into net/dccp/Kconfig.
The CCID Kconfig menu is sourced within this clause; as a
consequence, all tests of type `depends on IP_DCCP' are now
redundant.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch addresses the following problems:
1. DCCP relies for its proper functioning on having at least one CCID module
enabled (as in TCP plugable congestion control). Currently it is possible to
disable both CCIDs and thus leave the DCCP module in a compiled, but entirely
non-functional state: no sockets can be created when no CCID is available.
Furthermore, the protocol is (again like TCP) not intended to be used without
CCIDs. Last, a non-empty CCID list is needed for doing CCID feature negotiation.
2. Internally the default CCID that is advertised by the Linux host is set to CCID2
(DCCPF_INITIAL_CCID in include/linux/dccp.h). Disabling CCID2 in the Kconfig
menu without changing the defaults leads to a failure `module not found' when
trying to load the dccp module (which internally tries to load the default CCID).
3. The specification (RFC 4340, sec. 10) treats CCID2 somewhat like a
`minimum common denominator'; the specification says that:
* "New connections start with CCID 2 for both endpoints"
* "A DCCP implementation intended for general use, such as an implementation in a
general-purpose operating system kernel, SHOULD implement at least CCID 2.
The intent is to make CCID 2 broadly available for interoperability [...]"
Providing CCID2 as minimum-required CCID (like Reno/Cubic in TCP) thus seems reasonable.
Hence this patch automatically selects CCID2 when DCCP is enabled. Documentation also added.
Discussions with Ian McDonald on this subject are gratefully acknowledged.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This extends the DCCP socket API by honouring any shutdown(2) option set by the user.
The behaviour is, as much as possible, made consistent with the API for TCP's shutdown.
This patch exploits the information provided by the user via the socket API to reduce
processing costs:
* if the read end is closed (SHUT_RD), it is not necessary to deliver to input CCID;
* if the write end is closed (SHUT_WR), the same idea applies, but with a difference -
as long as the TX queue has not been drained, we need to receive feedback to keep
congestion-control rates up to date. Hence SHUT_WR is honoured only after the last
packet (under congestion control) has been sent;
* although SHUT_RDWR seems nonsensical, it is nevertheless supported in the same manner
as for TCP (and agrees with test for SHUTDOWN_MASK in dccp_poll() in net/dccp/proto.c).
Furthermore, most of the code already honours the sk_shutdown flags (dccp_recvmsg() for
instance sets the read length to 0 if SHUT_RD had been called); CCID handling is now added
to this by the present patch.
There will also no longer be any delivery when the socket is in the final stages, i.e. when
one of dccp_close(), dccp_fin(), or dccp_done() has been called - which is fine since at
that stage the connection is its final stages.
Motivation and background are on http://www.erg.abdn.ac.uk/users/gerrit/dccp/notes/shutdown
A FIXME has been added to notify the other end if SHUT_RD has been set (RFC 4340, 11.7).
Note: There is a comment in inet_shutdown() in net/ipv4/af_inet.c which asks to "make
sure the socket is a TCP socket". This should probably be extended to mean
`TCP or DCCP socket' (the code is also used by UDP and raw sockets).
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This fixes/updates the handling of idle and application-limited periods in CCID3,
which currently is broken: there is no detection as to how long a sender has been
idle - there is only one flag which is toggled in between function calls.
Being obsolete now, the `idle' flag is removed.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch fixes a previously undiscovered bug; the problem is in computing
the elapsed time as the time between `receiving' the packet (i.e. skb enters
CCID module) and sending feedback:
- there is no layer-processing, queueing, or delay involved,
- hence the elapsed time is in the order of 1 function call
- this is in the dimension of maximally 50..100usec
- which renders the use of elapsed time almost entirely useless.
The fix is simply to ignore such trivial amounts of elapsed time.
As a further advantage, the now useless elapsed_time field can be removed from
the socket, which reduces the socket structure by another four bytes.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This updates the CCID3 code with regard to two instances of using `MSS' in place of `s':
1. The RFC3390-based initial rate: both rfc3448bis as well as the Faster Restart
draft now consistently use `s' instead of MSS.
2. Now agrees with section 4.2 of rfc3448bis: "If the sender is ready to send data when
it does not yet have a round trip sample, the value of X is set to s bytes per
second, for segment size s [...]"
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Many-many code in the kernel initialized the timer->function
and timer->data together with calling init_timer(timer). There
is already a helper for this. Use it for networking code.
The patch is HUGE, but makes the code 130 lines shorter
(98 insertions(+), 228 deletions(-)).
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Acked-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
As done two years ago on IP route cache table (commit
22c047ccbc) , we can avoid using one
lock per hash bucket for the huge TCP/DCCP hash tables.
On a typical x86_64 platform, this saves about 2MB or 4MB of ram, for
litle performance differences. (we hit a different cache line for the
rwlock, but then the bucket cache line have a better sharing factor
among cpus, since we dirty it less often). For netstat or ss commands
that want a full scan of hash table, we perform fewer memory accesses.
Using a 'small' table of hashed rwlocks should be more than enough to
provide correct SMP concurrency between different buckets, without
using too much memory. Sizing of this table depends on
num_possible_cpus() and various CONFIG settings.
This patch provides some locking abstraction that may ease a future
work using a different model for TCP/DCCP table.
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Acked-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Assigning initial values of `0' is redundant when loading a new CCID structure,
since in net/dccp/ccid.c the entire CCID structure is zeroed out prior to
initialisation in ccid_new():
struct ccid {
struct ccid_operations *ccid_ops;
char ccid_priv[0];
};
// ...
if (rx) {
memset(ccid + 1, 0, ccid_ops->ccid_hc_rx_obj_size);
if (ccid->ccid_ops->ccid_hc_rx_init != NULL &&
ccid->ccid_ops->ccid_hc_rx_init(ccid, sk) != 0)
goto out_free_ccid;
} else {
memset(ccid + 1, 0, ccid_ops->ccid_hc_tx_obj_size);
/* analogous to the rx case */
}
This patch therefore removes the redundant assignments. Thanks to Arnaldo for
the inspiration.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
This fixes `unaligned (read) access' errors of the type
Kernel unaligned access at TPC[100f970c] dccp_parse_options+0x4f4/0x7e0 [dccp]
Kernel unaligned access at TPC[1011f2e4] ccid3_hc_tx_parse_options+0x1ac/0x380 [dccp_ccid3]
Kernel unaligned access at TPC[100f9898] dccp_parse_options+0x680/0x880 [dccp]
by using the get_unaligned macro for parsing options.
Commiter note: Preserved the sparse __be{16,32} annotations.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
This adds support for converting the 11 currently defined Reset codes into system
error numbers, which are stored in sk_err for further interpretation.
This makes the externally visible API behaviour similar to TCP, since a client
connecting to a non-existing port will experience ECONNREFUSED.
* Code 0, Unspecified, is interpreted as non-error (0);
* Code 1, Closed (normal termination), also maps into 0;
* Code 2, Aborted, maps into "Connection reset by peer" (ECONNRESET);
* Code 3, No Connection and
Code 7, Connection Refused, map into "Connection refused" (ECONNREFUSED);
* Code 4, Packet Error, maps into "No message of desired type" (ENOMSG);
* Code 5, Option Error, maps into "Illegal byte sequence" (EILSEQ);
* Code 6, Mandatory Error, maps into "Operation not supported on transport endpoint" (EOPNOTSUPP);
* Code 8, Bad Service Code, maps into "Invalid request code" (EBADRQC);
* Code 9, Too Busy, maps into "Too many users" (EUSERS);
* Code 10, Bad Init Cookie, maps into "Invalid request descriptor" (EBADR);
* Code 11, Aggression Penalty, maps into "Quota exceeded" (EDQUOT)
which makes sense in terms of using more than the `fair share' of bandwidth.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
This fixes the following problem: client connects to peer which has no DCCP
enabled or loaded; ICMP error messages ("Protocol Unavailable") can be seen
on the wire, but the application hangs. Reason: ICMP packets don't get through
to dccp_v4_err.
When reporting errors, a sequence number check is made for the DCCP packet
that had caused an ICMP error to arrive.
Such checks can not be made if the socket is in state LISTEN, RESPOND (which
in the implementation is the same as LISTEN), or REQUEST, since update_gsr()
has not been called in these states, hence the sequence window is 0..0.
This patch fixes the problem by adding the REQUEST state as another exemption
to the window check. The error reporting now works as expected on connecting.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
This fixes a problem when analysing erroneous packets in dccp_v{4,6}_err:
* dccp_hdr_seq currently takes an skb
* however, the transport headers in the skb are shifted, due to the
preceding IPv4/v6 header.
Fixed for v4 and v6 by changing dccp_hdr_seq to take a struct dccp_hdr as
argument. Verified that the correct sequence number is now reported in the
error handler.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Now that we have this new macro, use it where possible.
Signed-off-by: Jean Delvare <jdelvare@suse.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
