The earlier version was just very basic one which is "playing
safe" by always clearing the hints. However, clearing of a hint
is extremely costly operation with large windows, so it must be
avoided at all cost whenever possible, there is a way with
shifting too achieve not-clearing.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
During SACK processing, most of the benefits of TSO are eaten by
the SACK blocks that one-by-one fragment SKBs to MSS sized chunks.
Then we're in problems when cleanup work for them has to be done
when a large cumulative ACK comes. Try to return back to pre-split
state already while more and more SACK info gets discovered by
combining newly discovered SACK areas with the previous skb if
that's SACKed as well.
This approach has a number of benefits:
1) The processing overhead is spread more equally over the RTT
2) Write queue has less skbs to process (affect everything
which has to walk in the queue past the sacked areas)
3) Write queue is consistent whole the time, so no other parts
of TCP has to be aware of this (this was not the case with
some other approach that was, well, quite intrusive all
around).
4) Clean_rtx_queue can release most of the pages using single
put_page instead of previous PAGE_SIZE/mss+1 calls
In case a hole is fully filled by the new SACK block, we attempt
to combine the next skb too which allows construction of skbs
that are even larger than what tso split them to and it handles
hole per on every nth patterns that often occur during slow start
overshoot pretty nicely. Though this to be really useful also
a retransmission would have to get lost since cumulative ACKs
advance one hole at a time in the most typical case.
TODO: handle upwards only merging. That should be rather easy
when segment is fully sacked but I'm leaving that as future
work item (it won't make very large difference anyway since
this current approach already covers quite a lot of normal
cases).
I was earlier thinking of some sophisticated way of tracking
timestamps of the first and the last segment but later on
realized that it won't be that necessary at all to store the
timestamp of the last segment. The cases that can occur are
basically either:
1) ambiguous => no sensible measurement can be taken anyway
2) non-ambiguous is due to reordering => having the timestamp
of the last segment there is just skewing things more off
than does some good since the ack got triggered by one of
the holes (besides some substle issues that would make
determining right hole/skb even harder problem). Anyway,
it has nothing to do with this change then.
I choose to route some abnormal looking cases with goto noop,
some could be handled differently (eg., by stopping the
walking at that skb but again). In general, they either
shouldn't happen at all or are rare enough to make no difference
in practice.
In theory this change (as whole) could cause some macroscale
regression (global) because of cache misses that are taken over
the round-trip time but it gets very likely better because of much
less (local) cache misses per other write queue walkers and the
big recovery clearing cumulative ack.
Worth to note that these benefits would be very easy to get also
without TSO/GSO being on as long as the data is in pages so that
we can merge them. Currently I won't let that happen because
DSACK splitting at fragment that would mess up pcounts due to
sk_can_gso in tcp_set_skb_tso_segs. Once DSACKs fragments gets
avoided, we have some conditions that can be made less strict.
TODO: I will probably have to convert the excessive pointer
passing to struct sacktag_state... :-)
My testing revealed that considerable amount of skbs couldn't
be shifted because they were cloned (most likely still awaiting
tx reclaim)...
[The rest is considering future work instead since I got
repeatably EFAULT to tcpdump's recvfrom when I added
pskb_expand_head to deal with clones, so I separated that
into another, later patch]
...To counter that, I gave up on the fifth advantage:
5) When growing previous SACK block, less allocs for new skbs
are done, basically a new alloc is needed only when new hole
is detected and when the previous skb runs out of frags space
...which now only happens of if reclaim is fast enough to dispose
the clone before the SACK block comes in (the window is RTT long),
otherwise we'll have to alloc some.
With clones being handled I got these numbers (will be somewhat
worse without that), taken with fine-grained mibs:
TCPSackShifted 398
TCPSackMerged 877
TCPSackShiftFallback 320
TCPSACKCOLLAPSEFALLBACKGSO 0
TCPSACKCOLLAPSEFALLBACKSKBBITS 0
TCPSACKCOLLAPSEFALLBACKSKBDATA 0
TCPSACKCOLLAPSEFALLBACKBELOW 0
TCPSACKCOLLAPSEFALLBACKFIRST 1
TCPSACKCOLLAPSEFALLBACKPREVBITS 318
TCPSACKCOLLAPSEFALLBACKMSS 1
TCPSACKCOLLAPSEFALLBACKNOHEAD 0
TCPSACKCOLLAPSEFALLBACKSHIFT 0
TCPSACKCOLLAPSENOOPSEQ 0
TCPSACKCOLLAPSENOOPSMALLPCOUNT 0
TCPSACKCOLLAPSENOOPSMALLLEN 0
TCPSACKCOLLAPSEHOLE 12
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
This is preparatory work for SACK combiner patch which may
have to count TCP state changes for only a part of the skb
because it will intentionally avoids splitting skb to SACKed
and not sacked parts.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Sadly enough, this adds possible divide though we try to avoid
it by checking one mss as common case.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
I knew already when rewriting the sacktag that this condition
was too conservative, change it now since it prevent lot of
useless work (especially in the sack shifter decision code
that is being added by a later patch). This shouldn't change
anything really, just save some processing regardless of the
shifter.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Using NIPQUAD() with NIPQUAD_FMT, %d.%d.%d.%d or %u.%u.%u.%u
can be replaced with %pI4
Signed-off-by: Harvey Harrison <harvey.harrison@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
From: Ali Saidi <saidi@engin.umich.edu>
When TCP receive copy offload is enabled it's possible that
tcp_rcv_established() will cause two acks to be sent for a single
packet. In the case that a tcp_dma_early_copy() is successful,
copied_early is set to true which causes tcp_cleanup_rbuf() to be
called early which can send an ack. Further along in
tcp_rcv_established(), __tcp_ack_snd_check() is called and will
schedule a delayed ACK. If no packets are processed before the delayed
ack timer expires the packet will be acked twice.
Signed-off-by: David S. Miller <davem@davemloft.net>
I'm quite sure that if I give this function in its old format
for you to inspect, you start to wonder what is the type of
demanded or if it's a global variable.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
It all started from me noticing that this urgent check in
tcp_clean_rtx_queue is unnecessarily inside the loop. Then
I took a longer look to it and found out that the users of
urg_mode can trivially do without, well almost, there was
one gotcha.
Bonus: those funny people who use urg with >= 2^31 write_seq -
snd_una could now rejoice too (that's the only purpose for the
between being there, otherwise a simple compare would have done
the thing). Not that I assume that the rest of the tcp code
happily lives with such mind-boggling numbers :-). Alas, it
turned out to be impossible to set wmem to such numbers anyway,
yes I really tried a big sendfile after setting some wmem but
nothing happened :-). ...Tcp_wmem is int and so is sk_sndbuf...
So I hacked a bit variable to long and found out that it seems
to work... :-)
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
This minor cleanup simplifies later changes which will convert
struct sk_buff and friends over to using struct list_head.
Signed-off-by: David S. Miller <davem@davemloft.net>
Most importantly avoid doing it with cumulative ACK. However,
since we have lost_cnt_hint in the picture as well needing
adjustments, it's not as trivial as dealing with
retransmit_skb_hint (and cannot be done in the all place we
could trivially leave retransmit_skb_hint untouched).
With the previous patch, this should mostly remove O(n^2)
behavior while cumulative ACKs start flowing once rexmit
after a lossy round-trip made it through.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Most importantly avoid doing it with cumulative ACK. Not clearing
means that we no longer need n^2 processing in resolution of each
fast recovery.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Because lost counter no longer requires tuning, this is
trivial to remove (the tuning wouldn't have been too
hard either) because no "new" retransmittable skb appeared
below retransmit_skb_hint when SACKing for sure.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
I suspect it might have been related to the changed amount
of lost skbs, which was counted by retransmit_cnt_hint that
got changed.
The place for this clearing was very illogical anyway,
it should have been after the LOST-bit clearing loop to
make any sense.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Main benefit in this is that we can then freely point
the retransmit_skb_hint to anywhere we want to because
there's no longer need to know what would be the count
changes involve, and since this is really used only as a
terminator, unnecessary work is one time walk at most,
and if some retransmissions are necessary after that
point later on, the walk is not full waste of time
anyway.
Since retransmit_high must be kept valid, all lost
markers must ensure that.
Now I also have learned how those "holes" in the
rexmittable skbs can appear, mtu probe does them. So
I removed the misleading comment as well.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
This useful because we'd need to verifying soon in many places
which makes things slightly more complex than it used to be.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Ie., the difference between partial and all clearing doesn't
exists anymore since the SACK optimizations got dropped by
an sacktag rewrite.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch consolidates the code common to TCP and CCID-2:
* TCP uses RFC 3390 in a packet-oriented manner (tcp_input.c) and
* CCID-2 uses RFC 3390 in packet-oriented manner (RFC 4341).
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Some duplicated code lying around. Located with my suffix tree
tool.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Large block of code duplication removed.
Sadly, the return value thing is a bit tricky here but it
seems the most sensible way to return positive from validator
on success rather than negative.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Removes legacy reinvent-the-wheel type thing. The generic
machinery integrates much better to automated debugging aids
such as kerneloops.org (and others), and is unambiguous due to
better naming. Non-intuively BUG_TRAP() is actually equal to
WARN_ON() rather than BUG_ON() though some might actually be
promoted to BUG_ON() but I left that to future.
I could make at least one BUILD_BUG_ON conversion.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
This is based upon an excellent bug report from Eric Dumazet.
tcp_ack() should clear ->icsk_probes_out even if there are packets
outstanding. Otherwise if we get a sequence of ACKs while we do have
packets outstanding over and over again, we'll never clear the
probes_out value and eventually think the connection is too sick and
we'll reset it.
This appears to be some "optimization" added to tcp_ack() in the 2.4.x
timeframe. In 2.2.x, probes_out is pretty much always cleared by
tcp_ack().
Here is Eric's original report:
----------------------------------------
Apparently, we can in some situations reset TCP connections in a couple of seconds when some frames are lost.
In order to reproduce the problem, please try the following program on linux-2.6.25.*
Setup some iptables rules to allow two frames per second sent on loopback interface to tcp destination port 12000
iptables -N SLOWLO
iptables -A SLOWLO -m hashlimit --hashlimit 2 --hashlimit-burst 1 --hashlimit-mode dstip --hashlimit-name slow2 -j ACCEPT
iptables -A SLOWLO -j DROP
iptables -A OUTPUT -o lo -p tcp --dport 12000 -j SLOWLO
Then run the attached program and see the output :
# ./loop
State Recv-Q Send-Q Local Address:Port Peer Address:Port
ESTAB 0 40 127.0.0.1:54455 127.0.0.1:12000 timer:(persist,200ms,1)
State Recv-Q Send-Q Local Address:Port Peer Address:Port
ESTAB 0 40 127.0.0.1:54455 127.0.0.1:12000 timer:(persist,200ms,3)
State Recv-Q Send-Q Local Address:Port Peer Address:Port
ESTAB 0 40 127.0.0.1:54455 127.0.0.1:12000 timer:(persist,200ms,5)
State Recv-Q Send-Q Local Address:Port Peer Address:Port
ESTAB 0 40 127.0.0.1:54455 127.0.0.1:12000 timer:(persist,200ms,7)
State Recv-Q Send-Q Local Address:Port Peer Address:Port
ESTAB 0 40 127.0.0.1:54455 127.0.0.1:12000 timer:(persist,200ms,9)
State Recv-Q Send-Q Local Address:Port Peer Address:Port
ESTAB 0 40 127.0.0.1:54455 127.0.0.1:12000 timer:(persist,200ms,11)
State Recv-Q Send-Q Local Address:Port Peer Address:Port
ESTAB 0 40 127.0.0.1:54455 127.0.0.1:12000 timer:(persist,201ms,13)
State Recv-Q Send-Q Local Address:Port Peer Address:Port
ESTAB 0 40 127.0.0.1:54455 127.0.0.1:12000 timer:(persist,188ms,15)
write(): Connection timed out
wrote 890 bytes but was interrupted after 9 seconds
ESTAB 0 0 127.0.0.1:12000 127.0.0.1:54455
Exiting read() because no data available (4000 ms timeout).
read 860 bytes
While this tcp session makes progress (sending frames with 50 bytes of payload, every 500ms), linux tcp stack decides to reset it, when tcp_retries 2 is reached (default value : 15)
tcpdump :
15:30:28.856695 IP 127.0.0.1.56554 > 127.0.0.1.12000: S 33788768:33788768(0) win 32792 <mss 16396,nop,nop,sackOK,nop,wscale 7>
15:30:28.856711 IP 127.0.0.1.12000 > 127.0.0.1.56554: S 33899253:33899253(0) ack 33788769 win 32792 <mss 16396,nop,nop,sackOK,nop,wscale 7>
15:30:29.356947 IP 127.0.0.1.56554 > 127.0.0.1.12000: P 1:61(60) ack 1 win 257
15:30:29.356966 IP 127.0.0.1.12000 > 127.0.0.1.56554: . ack 61 win 257
15:30:29.866415 IP 127.0.0.1.56554 > 127.0.0.1.12000: P 61:111(50) ack 1 win 257
15:30:29.866427 IP 127.0.0.1.12000 > 127.0.0.1.56554: . ack 111 win 257
15:30:30.366516 IP 127.0.0.1.56554 > 127.0.0.1.12000: P 111:161(50) ack 1 win 257
15:30:30.366527 IP 127.0.0.1.12000 > 127.0.0.1.56554: . ack 161 win 257
15:30:30.876196 IP 127.0.0.1.56554 > 127.0.0.1.12000: P 161:211(50) ack 1 win 257
15:30:30.876207 IP 127.0.0.1.12000 > 127.0.0.1.56554: . ack 211 win 257
15:30:31.376282 IP 127.0.0.1.56554 > 127.0.0.1.12000: P 211:261(50) ack 1 win 257
15:30:31.376290 IP 127.0.0.1.12000 > 127.0.0.1.56554: . ack 261 win 257
15:30:31.885619 IP 127.0.0.1.56554 > 127.0.0.1.12000: P 261:311(50) ack 1 win 257
15:30:31.885631 IP 127.0.0.1.12000 > 127.0.0.1.56554: . ack 311 win 257
15:30:32.385705 IP 127.0.0.1.56554 > 127.0.0.1.12000: P 311:361(50) ack 1 win 257
15:30:32.385715 IP 127.0.0.1.12000 > 127.0.0.1.56554: . ack 361 win 257
15:30:32.895249 IP 127.0.0.1.56554 > 127.0.0.1.12000: P 361:411(50) ack 1 win 257
15:30:32.895266 IP 127.0.0.1.12000 > 127.0.0.1.56554: . ack 411 win 257
15:30:33.395341 IP 127.0.0.1.56554 > 127.0.0.1.12000: P 411:461(50) ack 1 win 257
15:30:33.395351 IP 127.0.0.1.12000 > 127.0.0.1.56554: . ack 461 win 257
15:30:33.918085 IP 127.0.0.1.56554 > 127.0.0.1.12000: P 461:511(50) ack 1 win 257
15:30:33.918096 IP 127.0.0.1.12000 > 127.0.0.1.56554: . ack 511 win 257
15:30:34.418163 IP 127.0.0.1.56554 > 127.0.0.1.12000: P 511:561(50) ack 1 win 257
15:30:34.418172 IP 127.0.0.1.12000 > 127.0.0.1.56554: . ack 561 win 257
15:30:34.927685 IP 127.0.0.1.56554 > 127.0.0.1.12000: P 561:611(50) ack 1 win 257
15:30:34.927698 IP 127.0.0.1.12000 > 127.0.0.1.56554: . ack 611 win 257
15:30:35.427757 IP 127.0.0.1.56554 > 127.0.0.1.12000: P 611:661(50) ack 1 win 257
15:30:35.427766 IP 127.0.0.1.12000 > 127.0.0.1.56554: . ack 661 win 257
15:30:35.937359 IP 127.0.0.1.56554 > 127.0.0.1.12000: P 661:711(50) ack 1 win 257
15:30:35.937376 IP 127.0.0.1.12000 > 127.0.0.1.56554: . ack 711 win 257
15:30:36.437451 IP 127.0.0.1.56554 > 127.0.0.1.12000: P 711:761(50) ack 1 win 257
15:30:36.437464 IP 127.0.0.1.12000 > 127.0.0.1.56554: . ack 761 win 257
15:30:36.947022 IP 127.0.0.1.56554 > 127.0.0.1.12000: P 761:811(50) ack 1 win 257
15:30:36.947039 IP 127.0.0.1.12000 > 127.0.0.1.56554: . ack 811 win 257
15:30:37.447135 IP 127.0.0.1.56554 > 127.0.0.1.12000: P 811:861(50) ack 1 win 257
15:30:37.447203 IP 127.0.0.1.12000 > 127.0.0.1.56554: . ack 861 win 257
15:30:41.448171 IP 127.0.0.1.12000 > 127.0.0.1.56554: F 1:1(0) ack 861 win 257
15:30:41.448189 IP 127.0.0.1.56554 > 127.0.0.1.12000: R 33789629:33789629(0) win 0
Source of program :
/*
* small producer/consumer program.
* setup a listener on 127.0.0.1:12000
* Forks a child
* child connect to 127.0.0.1, and sends 10 bytes on this tcp socket every 100 ms
* Father accepts connection, and read all data
*/
#include <sys/types.h>
#include <sys/socket.h>
#include <netinet/in.h>
#include <unistd.h>
#include <stdio.h>
#include <time.h>
#include <sys/poll.h>
int port = 12000;
char buffer[4096];
int main(int argc, char *argv[])
{
int lfd = socket(AF_INET, SOCK_STREAM, 0);
struct sockaddr_in socket_address;
time_t t0, t1;
int on = 1, sfd, res;
unsigned long total = 0;
socklen_t alen = sizeof(socket_address);
pid_t pid;
time(&t0);
socket_address.sin_family = AF_INET;
socket_address.sin_port = htons(port);
socket_address.sin_addr.s_addr = htonl(INADDR_LOOPBACK);
if (lfd == -1) {
perror("socket()");
return 1;
}
setsockopt(lfd, SOL_SOCKET, SO_REUSEADDR, &on, sizeof(int));
if (bind(lfd, (struct sockaddr *)&socket_address, sizeof(socket_address)) == -1) {
perror("bind");
close(lfd);
return 1;
}
if (listen(lfd, 1) == -1) {
perror("listen()");
close(lfd);
return 1;
}
pid = fork();
if (pid == 0) {
int i, cfd = socket(AF_INET, SOCK_STREAM, 0);
close(lfd);
if (connect(cfd, (struct sockaddr *)&socket_address, sizeof(socket_address)) == -1) {
perror("connect()");
return 1;
}
for (i = 0 ; ;) {
res = write(cfd, "blablabla\n", 10);
if (res > 0) total += res;
else if (res == -1) {
perror("write()");
break;
} else break;
usleep(100000);
if (++i == 10) {
system("ss -on dst 127.0.0.1:12000");
i = 0;
}
}
time(&t1);
fprintf(stderr, "wrote %lu bytes but was interrupted after %g seconds\n", total, difftime(t1, t0));
system("ss -on | grep 127.0.0.1:12000");
close(cfd);
return 0;
}
sfd = accept(lfd, (struct sockaddr *)&socket_address, &alen);
if (sfd == -1) {
perror("accept");
return 1;
}
close(lfd);
while (1) {
struct pollfd pfd[1];
pfd[0].fd = sfd;
pfd[0].events = POLLIN;
if (poll(pfd, 1, 4000) == 0) {
fprintf(stderr, "Exiting read() because no data available (4000 ms timeout).\n");
break;
}
res = read(sfd, buffer, sizeof(buffer));
if (res > 0) total += res;
else if (res == 0) break;
else perror("read()");
}
fprintf(stderr, "read %lu bytes\n", total);
close(sfd);
return 0;
}
----------------------------------------
Signed-off-by: David S. Miller <davem@davemloft.net>
Remove redundant checks when setting eff_sacks and make the number of SACKs a
compile time constant. Now that the options code knows how many SACK blocks can
fit in the header, we don't need to have the SACK code guessing at it.
Signed-off-by: Adam Langley <agl@imperialviolet.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Some of the metrics (RTT, RTTVAR and RTAX_RTO_MIN) are stored in
kernel units (jiffies) and this leaks out through the netlink API to
user space where the units for jiffies are unknown.
This patches changes the kernel to convert to/from milliseconds. This
changes the ABI, but milliseconds seemed like the most natural unit
for these parameters. Values available via syscall in
/proc/net/rt_cache and netlink will be in milliseconds.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
These places have a tcp_sock, but we'd prefer the sock itself to
get net from it. Fortunately, tcp_sk macro is just a type cast, so
this replace is really cheap.
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Same as before - the sock is always there to get the net from,
but there are also some places with the net already saved on
the stack.
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
There are some places in TCP that select one MIB index to
bump snmp statistics like this:
if (<something>)
NET_INC_STATS_BH(<some_id>);
else if (<something_else>)
NET_INC_STATS_BH(<some_other_id>);
...
else
NET_INC_STATS_BH(<default_id>);
or in a more tricky but still similar way.
On the other hand, this NET_INC_STATS_BH is a camouflaged
increment of percpu variable, which is not that small.
Factoring those cases out de-bloats 235 bytes on non-preemptible
i386 config and drives parts of the code into 80 columns.
add/remove: 0/0 grow/shrink: 0/7 up/down: 0/-235 (-235)
function old new delta
tcp_fastretrans_alert 1437 1424 -13
tcp_dsack_set 137 124 -13
tcp_xmit_retransmit_queue 690 676 -14
tcp_try_undo_recovery 283 265 -18
tcp_sacktag_write_queue 1550 1515 -35
tcp_update_reordering 162 106 -56
tcp_retransmit_timer 990 904 -86
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
This reverts two changesets, ec3c0982a2
("[TCP]: TCP_DEFER_ACCEPT updates - process as established") and
the follow-on bug fix 9ae27e0adb
("tcp: Fix slab corruption with ipv6 and tcp6fuzz").
This change causes several problems, first reported by Ingo Molnar
as a distcc-over-loopback regression where connections were getting
stuck.
Ilpo Järvinen first spotted the locking problems. The new function
added by this code, tcp_defer_accept_check(), only has the
child socket locked, yet it is modifying state of the parent
listening socket.
Fixing that is non-trivial at best, because we can't simply just grab
the parent listening socket lock at this point, because it would
create an ABBA deadlock. The normal ordering is parent listening
socket --> child socket, but this code path would require the
reverse lock ordering.
Next is a problem noticed by Vitaliy Gusev, he noted:
----------------------------------------
>--- a/net/ipv4/tcp_timer.c
>+++ b/net/ipv4/tcp_timer.c
>@@ -481,6 +481,11 @@ static void tcp_keepalive_timer (unsigned long data)
> goto death;
> }
>
>+ if (tp->defer_tcp_accept.request && sk->sk_state == TCP_ESTABLISHED) {
>+ tcp_send_active_reset(sk, GFP_ATOMIC);
>+ goto death;
Here socket sk is not attached to listening socket's request queue. tcp_done()
will not call inet_csk_destroy_sock() (and tcp_v4_destroy_sock() which should
release this sk) as socket is not DEAD. Therefore socket sk will be lost for
freeing.
----------------------------------------
Finally, Alexey Kuznetsov argues that there might not even be any
real value or advantage to these new semantics even if we fix all
of the bugs:
----------------------------------------
Hiding from accept() sockets with only out-of-order data only
is the only thing which is impossible with old approach. Is this really
so valuable? My opinion: no, this is nothing but a new loophole
to consume memory without control.
----------------------------------------
So revert this thing for now.
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch removes CVS keywords that weren't updated for a long time
from comments.
Signed-off-by: Adrian Bunk <bunk@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
This bug is able to corrupt fackets_out in very rare cases.
In order for this to cause corruption:
1) DSACK in the middle of previous SACK block must be generated.
2) In order to take that particular branch, part or all of the
DSACKed segment must already be SACKed so that we have that
in cache in the first place.
3) The new info must be top enough so that fackets_out will be
updated on this iteration.
...then fack_count is updated while skb wasn't, then we walk again
that particular segment thus updating fack_count twice for
a single skb and finally that value is assigned to fackets_out
by tcp_sacktag_one.
It is safe to call tcp_sacktag_one just once for a segment (at
DSACK), no need to call again for plain SACK.
Potential problem of the miscount are limited to premature entry
to recovery and to inflated reordering metric (which could even
cancel each other out in the most the luckiest scenarios :-)).
Both are quite insignificant in worst case too and there exists
also code to reset them (fackets_out once sacked_out becomes zero
and reordering metric on RTO).
This has been reported by a number of people, because it occurred
quite rarely, it has been very evasive. Andy Furniss was able to
get it to occur couple of times so that a bit more info was
collected about the problem using a debug patch, though it still
required lot of checking around. Thanks also to others who have
tried to help here.
This is listed as Bugzilla #10346. The bug was introduced by
me in commit 68f8353b48 ([TCP]: Rewrite SACK block processing &
sack_recv_cache use), I probably thought back then that there's
need to scan that entry twice or didn't dare to make it go
through it just once there. Going through twice would have
required restoring fack_count after the walk but as noted above,
I chose to drop the additional walk step altogether here.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
It is possible that this skip path causes TCP to end up into an
invalid state where ca_state was left to CA_Open while some
segments already came into sacked_out. If next valid ACK doesn't
contain new SACK information TCP fails to enter into
tcp_fastretrans_alert(). Thus at least high_seq is set
incorrectly to a too high seqno because some new data segments
could be sent in between (and also, limited transmit is not
being correctly invoked there). Reordering in both directions
can easily cause this situation to occur.
I guess we would want to use tcp_moderate_cwnd(tp) there as well
as it may be possible to use this to trigger oversized burst to
network by sending an old ACK with huge amount of SACK info, but
I'm a bit unsure about its effects (mainly to FlightSize), so to
be on the safe side I just currently fixed it minimally to keep
TCP's state consistent (obviously, such nasty ACKs have been
possible this far). Though it seems that FlightSize is already
underestimated by some amount, so probably on the long term we
might want to trigger recovery there too, if appropriate, to make
FlightSize calculation to resemble reality at the time when the
losses where discovered (but such change scares me too much now
and requires some more thinking anyway how to do that as it
likely involves some code shuffling).
This bug was found by Brian Vowell while running my TCP debug
patch to find cause of another TCP issue (fackets_out
miscount).
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
If receiver consumes segments successfully only in-order, FRTO
fallback to conventional recovery produces RTO loop because
FRTO's forward transmissions will always get dropped and need to
be resent, yet by default they're not marked as lost (which are
the only segments we will retransmit in CA_Loss).
Price to pay about this is occassionally unnecessarily
retransmitting the forward transmission(s). SACK blocks help
a bit to avoid this, so it's mainly a concern for NewReno case
though SACK is not fully immune either.
This change has a side-effect of fixing SACKFRTO problem where
it didn't have snd_nxt of the RTO time available anymore when
fallback become necessary (this problem would have only occured
when RTO would occur for two or more segments and ECE arrives
in step 3; no need to figure out how to fix that unless the
TODO item of selective behavior is considered in future).
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Reported-by: Damon L. Chesser <damon@damtek.com>
Tested-by: Damon L. Chesser <damon@damtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
It seems that commit 009a2e3e4e ("[TCP] FRTO: Improve
interoperability with other undo_marker users") run into
another land-mine which caused fallback to conventional
recovery to break:
1. Cumulative ACK arrives after FRTO retransmission
2. tcp_try_to_open sees zero retrans_out, clears retrans_stamp
which should be kept like in CA_Loss state it would be
3. undo_marker change allowed tcp_packet_delayed to return
true because of the cleared retrans_stamp once FRTO is
terminated causing LossUndo to occur, which means all loss
markings FRTO made are reverted.
This means that the conventional recovery basically recovered
one loss per RTT, which is not that efficient. It was quite
unobvious that the undo_marker change broken something like
this, I had a quite long session to track it down because of
the non-intuitiviness of the bug (luckily I had a trivial
reproducer at hand and I was also able to learn to use kprobes
in the process as well :-)).
This together with the NewReno+FRTO fix and FRTO in-order
workaround this fixes Damon's problems, this and the first
mentioned are enough to fix Bugzilla #10063.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Reported-by: Damon L. Chesser <damon@damtek.com>
Tested-by: Damon L. Chesser <damon@damtek.com>
Tested-by: Sebastian Hyrwall <zibbe@cisko.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Note: there's actually another bug in FRTO's SACK variant, which
is the causing failure in NewReno case because of the error
that's fixed here. I'll fix the SACK case separately (it's
a separate bug really, though related, but in order to fix that
I need to audit tp->snd_nxt usage a bit).
There were two places where SACK variant of FRTO is getting
incorrectly used even if SACK wasn't negotiated by the TCP flow.
This leads to incorrect setting of frto_highmark with NewReno
if a previous recovery was interrupted by another RTO.
An eventual fallback to conventional recovery then incorrectly
considers one or couple of segments as forward transmissions
though they weren't, which then are not LOST marked during
fallback making them "non-retransmittable" until the next RTO.
In a bad case, those segments are really lost and are the only
one left in the window. Thus TCP needs another RTO to continue.
The next FRTO, however, could again repeat the same events
making the progress of the TCP flow extremely slow.
In order for these events to occur at all, FRTO must occur
again in FRTOs step 3 while the key segments must be lost as
well, which is not too likely in practice. It seems to most
frequently with some small devices such as network printers
that *seem* to accept TCP segments only in-order. In cases
were key segments weren't lost, things get automatically
resolved because those wrongly marked segments don't need to be
retransmitted in order to continue.
I found a reproducer after digging up relevant reports (few
reports in total, none at netdev or lkml I know of), some
cases seemed to indicate middlebox issues which seems now
to be a false assumption some people had made. Bugzilla
#10063 _might_ be related. Damon L. Chesser <damon@damtek.com>
had a reproducable case and was kind enough to tcpdump it
for me. With the tcpdump log it was quite trivial to figure
out.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
