Key points of this patch are:
- In case new SACK information is advance only type, no skb
processing below previously discovered highest point is done
- Optimize cases below highest point too since there's no need
to always go up to highest point (which is very likely still
present in that SACK), this is not entirely true though
because I'm dropping the fastpath_skb_hint which could
previously optimize those cases even better. Whether that's
significant, I'm not too sure.
Currently it will provide skipping by walking. Combined with
RB-tree, all skipping would become fast too regardless of window
size (can be done incrementally later).
Previously a number of cases in TCP SACK processing fails to
take advantage of costly stored information in sack_recv_cache,
most importantly, expected events such as cumulative ACK and new
hole ACKs. Processing on such ACKs result in rather long walks
building up latencies (which easily gets nasty when window is
huge). Those latencies are often completely unnecessary
compared with the amount of _new_ information received, usually
for cumulative ACK there's no new information at all, yet TCP
walks whole queue unnecessary potentially taking a number of
costly cache misses on the way, etc.!
Since the inclusion of highest_sack, there's a lot information
that is very likely redundant (SACK fastpath hint stuff,
fackets_out, highest_sack), though there's no ultimate guarantee
that they'll remain the same whole the time (in all unearthly
scenarios). Take advantage of this knowledge here and drop
fastpath hint and use direct access to highest SACKed skb as
a replacement.
Effectively "special cased" fastpath is dropped. This change
adds some complexity to introduce better coveraged "fastpath",
though the added complexity should make TCP behave more cache
friendly.
The current ACK's SACK blocks are compared against each cached
block individially and only ranges that are new are then scanned
by the high constant walk. For other parts of write queue, even
when in previously known part of the SACK blocks, a faster skip
function is used (if necessary at all). In addition, whenever
possible, TCP fast-forwards to highest_sack skb that was made
available by an earlier patch. In typical case, no other things
but this fast-forward and mandatory markings after that occur
making the access pattern quite similar to the former fastpath
"special case".
DSACKs are special case that must always be walked.
The local to recv_sack_cache copying could be more intelligent
w.r.t DSACKs which are likely to be there only once but that
is left to a separate patch.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
It is going to replace the sack fastpath hint quite soon... :-)
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
The comment in tcp_nagle_test suggests that. This bug is very
very old, even 2.4.0 seems to have it.
Signed-off-by: Ilpo Jrvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
The original code has striking complexity to perform a query
which can be reduced to a very simple compare.
FIN seqno may be included to write_seq but it should not make
any significant difference here compared to skb->len which was
used previously. One won't end up there with SYN still queued.
Use of write_seq check guarantees that there's a valid skb in
send_head so I removed the extra check.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Acked-by: John Heffner <jheffner@psc.edu>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
It seems that the checked range for receiver window check should
begin from the first rather than from the last skb that is going
to be included to the probe. And that can be achieved without
reference to skbs at all, snd_nxt and write_seq provides the
correct seqno already. Plus, it SHOULD account packets that are
necessary to trigger fast retransmit [RFC4821].
Location of snd_wnd < probe_size/size_needed check is bogus
because it will cause the other if() match as well (due to
snd_nxt >= snd_una invariant).
Removed dead obvious comment.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
When the abstraction functions got added, conversion here was
made incorrectly. As a result, the skb may end up pointing
to skb which got included to the probe skb and then was freed.
For it to trigger, however, skb_transmit must fail sending as
well.
Signed-off-by: Ilpo Jrvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
This addition of lost_retrans_low to tcp_sock might be
unnecessary, it's not clear how often lost_retrans worker is
executed when there wasn't work to do.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
1) Passing wrong skb to tcp_adjust_fackets_out could corrupt
fastpath_cnt_hint as tcp_skb_pcount(next_skb) is not included
to it if hint points exactly to the next_skb (it's lagging
behind, see sacktag).
2) When fastpath_skb_hint is put backwards to avoid dangling
skb reference, the skb's pcount must also be removed from count
(not included like above).
Reported by Cedric Le Goater <legoater@free.fr>
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
This was found due to bug report from Cedric Le Goater though
it turned this turned out to be unrelated bug.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
I previously added checking to position that is rather poor as
state has already been adjusted quite a bit. Re-placing it above
all state changes should be more robust though the return should
never ever get executed regardless of its place :-).
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
There's no reason to clear the sacktag skb hint when small part
of the rexmit queue changes. Account changes (if any) instead when
fragmenting/collapsing. RTO/FRTO do not touch SACKED_ACKED bits so
no need to discard SACK tag hint at all.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Substraction for fackets_out is unconditional when snd_una
advances, thus there's no need to do it inside the loop. Just
make sure correct bounds are honored.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
In general, it should not be necessary to call tcp_fragment for
already SACKed skbs, but it's better to be safe than sorry. And
indeed, it can be called from sacktag when a DSACK arrives or
some ACK (with SACK) reordering occurs (sacktag could be made
to avoid the call in the latter case though I'm not sure if it's
worth of the trouble and added complexity to cover such marginal
case).
The collapse case has return for SACKED_ACKED case earlier, so
just WARN_ON if internal inconsistency is detected for some
reason.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Makes caller side more obvious, there's no need to have
a wrapper for this oneliner!
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Previously code had IsReno/IsFack defined as macros that were
local to tcp_input.c though sack_ok field has user elsewhere too
for the same purpose. This changes them to static inlines as
preferred according the current coding style and unifies the
access to sack_ok across multiple files. Magic bitops of sack_ok
for FACK and DSACK are also abstracted to functions with
appropriate names.
Note:
- One sack_ok = 1 remains but that's self explanary, i.e., it
enables sack
- Couple of !IsReno cases are changed to tcp_is_sack
- There were no users for IsDSack => I dropped it
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Left_out was dropped a while ago, thus leaving verifying
consistency of the "left out" as only task for the function in
question. Thus make it's name more appropriate.
In addition, it is intentionally converted to #define instead
of static inline because the location of the invariant failure
is the most important thing to have if this ever triggers. I
think it would have been helpful e.g. in this case where the
location of the failure point had to be based on some quesswork:
http://lkml.org/lkml/2007/5/2/464
...Luckily the guesswork seems to have proved to be correct.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
It is easily calculable when needed and user are not that many
after all.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
No other users exist for tcp_ecn.h. Very few things remain in
tcp.h, for most TCP ECN functions callers reside within a
single .c file and can be placed there.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
In addition, added a reference about the purpose of the loop.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Do same adjustment to SACK fastpath counters provided that
they're valid.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Remove includes of <linux/smp_lock.h> where it is not used/needed.
Suggested by Al Viro.
Builds cleanly on x86_64, i386, alpha, ia64, powerpc, sparc,
sparc64, and arm (all 59 defconfigs).
Signed-off-by: Randy Dunlap <randy.dunlap@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This is a corner case where less than MSS sized new data thingie
is awaiting in the send queue. For F-RTO to work correctly, a
new data segment must be sent at certain point or F-RTO cannot
be used at all. RFC4138 allows overriding of Nagle at that
point.
Implementation uses frto_counter states 2 and 3 to distinguish
when Nagle override is needed.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
This updates references to drafts in comments which must be about 10
years old. Internet draft draft-ietf-tcpimpl-prob-03.txt expired in 1998
and was replaced by RFC 2525 in March 1999.
Section 3.10 of the draft maps almost identically into section 2.17 of RFC
2525: both are entitled "Failure to RST on close with data pending", the
differences in text body amount to a typo and minor sentence change.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
Do some simple changes to make congestion control API faster/cleaner.
* use ktime_t rather than timeval
* merge rtt sampling into existing ack callback
this means one indirect call versus two per ack.
* use flags bits to store options/settings
Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
So that it is also an offset from skb->head, reduces its size from 8 to 4 bytes
on 64bit architectures, allowing us to combine the 4 bytes hole left by the
layer headers conversion, reducing struct sk_buff size to 256 bytes, i.e. 4
64byte cachelines, and since the sk_buff slab cache is SLAB_HWCACHE_ALIGN...
:-)
Many calculations that previously required that skb->{transport,network,
mac}_header be first converted to a pointer now can be done directly, being
meaningful as offsets or pointers.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add whitespace around keywords.
Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
This allows the write queue implementation to be changed,
for example, to one which allows fast interval searching.
Signed-off-by: David S. Miller <davem@davemloft.net>
New sysctl tcp_frto_response is added to select amongst these
responses:
- Rate halving based; reuses CA_CWR state (default)
- Very conservative; used to be the only one available (=1)
- Undo cwr; undoes ssthresh and cwnd reductions (=2)
The response with rate halving requires a new parameter to
tcp_enter_cwr because FRTO has already reduced ssthresh and
doing a second reduction there has to be prevented. In addition,
to keep things nice on 80 cols screen, a local variable was
added.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
For the cases that slow_start_after_idle are meant to deal
with, it is almost a certainty that the congestion window
tests will think the connection is application limited and
we'll thus decrease the cwnd there too. This defeats the
whole point of setting slow_start_after_idle to zero.
So test it there too.
We do not cancel out the entire tcp_cwnd_validate() function
so that if the sysctl is changed we still have the validation
state maintained.
Signed-off-by: David S. Miller <davem@davemloft.net>
TCP may advertize up to 16-bits window in SYN packets (no window
scaling allowed). At the same time, TCP may have rcv_wnd
(32-bits) that does not fit to 16-bits without window scaling
resulting in pseudo garbage into advertized window from the
low-order bits of rcv_wnd. This can happen at least when
mss <= (1<<wscale) (see tcp_select_initial_window). This patch
fixes the handling of SYN advertized windows (compile tested
only).
In worst case (which is unlikely to occur though), the receiver
advertized window could be just couple of bytes. I'm not sure
that such situation would be handled very well at all by the
receiver!? Fortunately, the situation normalizes after the
first non-SYN ACK is received because it has the correct,
scaled window.
Alternatively, tcp_select_initial_window could be changed to
prevent too large rcv_wnd in the first place.
[ tcp_make_synack() has the same bug, and I've added a fix for
that to this patch -DaveM ]
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Revert 931731123a
We can't elide the skb_set_owner_w() here because things like certain
netfilter targets (such as owner MATCH) need a socket to be set on the
SKB for correct operation.
Thanks to Jan Engelhardt and other netfilter list members for
pointing this out.
Signed-off-by: David S. Miller <davem@davemloft.net>
The patch "Replace CHECKSUM_HW by CHECKSUM_PARTIAL/CHECKSUM_COMPLETE"
changed to unconditional copying of ip_summed field from collapsed
skb. This patch reverts this change.
The majority of substantial work including heavy testing
and diagnosing by: Michael Tokarev <mjt@tls.msk.ru>
Possible reasons pointed by: Herbert Xu and Patrick McHardy.
Signed-off-by: Jarek Poplawski <jarkao2@o2.pl>
Acked-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Throughout the TCP/DCCP (and tunnelling) code, it often happens that the
return code of a transmit function needs to be tested against NET_XMIT_CN
which is a value that does not indicate a strict error condition.
This patch uses a macro for these recurring situations which is consistent
with the already existing macro net_xmit_errno, saving on duplicated code.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
The data itself is already charged to the SKB, doing
the skb_set_owner_w() just generates a lot of noise and
extra atomics we don't really need.
Lmbench improvements on lat_tcp are minimal:
before:
TCP latency using localhost: 23.2701 microseconds
TCP latency using localhost: 23.1994 microseconds
TCP latency using localhost: 23.2257 microseconds
after:
TCP latency using localhost: 22.8380 microseconds
TCP latency using localhost: 22.9465 microseconds
TCP latency using localhost: 22.8462 microseconds
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch limits the amount of time you will defer sending a TSO segment
to less than two clock ticks, or the time between two acks, whichever is
longer.
On slow links, deferring causes significant bursts. See attached plots,
which show RTT through a 1 Mbps link with a 100 ms RTT and ~100 ms queue
for (a) non-TSO, (b) currnet TSO, and (c) patched TSO. This burstiness
causes significant jitter, tends to overflow queues early (bad for short
queues), and makes delay-based congestion control more difficult.
Deferring by a couple clock ticks I believe will have a relatively small
impact on performance.
Signed-off-by: John Heffner <jheffner@psc.edu>
Signed-off-by: David S. Miller <davem@davemloft.net>
Change net/core, ipv4 and ipv6 sysctl variables to __read_mostly.
Couldn't actually measure any performance increase while testing (.3%
I consider noise), but seems like the right thing to do.
Signed-off-by: Brian Haley <brian.haley@hp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
