While timer handler effectively runs a rcu read locked section,
there is no explicit rcu_read_lock()/rcu_read_unlock() annotations
and lockdep can be confused here :

net/ipv4/tcp_ipv4.c-906-        /* caller either holds rcu_read_lock() or socket lock */
net/ipv4/tcp_ipv4.c:907:        md5sig = rcu_dereference_check(tp->md5sig_info,
net/ipv4/tcp_ipv4.c-908-                                       sock_owned_by_user(sk) ||
net/ipv4/tcp_ipv4.c-909-                                       lockdep_is_held(&sk->sk_lock.slock));

Let's explicitely acquire rcu_read_lock() in tcp_make_synack()

Before commit fa76ce73 ("inet: get rid of central tcp/dccp listener
timer"), we were holding listener lock so lockdep was happy.

Fixes: fa76ce73 ("inet: get rid of central tcp/dccp listener timer")
Signed-off-by: NEric DUmazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This reverts commit ca10b9e9.

No longer needed after commit eb8895de
("tcp: tcp_make_synack() should use sock_wmalloc")

When under SYNFLOOD, we build lot of SYNACK and hit false sharing
because of multiple modifications done on sk_listener->sk_wmem_alloc

Since tcp_make_synack() uses sock_wmalloc(), there is no need
to call skb_set_owner_w() again, as this adds two atomic operations.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

tcp_send_fin() does not account for the memory it allocates properly, so
sk_forward_alloc can be negative in cases where we've sent a FIN:

ss example output (ss -amn | grep -B1 f4294):
tcp    FIN-WAIT-1 0      1            192.168.0.1:45520         192.0.2.1:8080
	skmem:(r0,rb87380,t0,tb87380,f4294966016,w1280,o0,bl0)
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

As per RFC4821 7.3.  Selecting Probe Size, a probe timer should
be armed once probing has converged. Once this timer expired,
probing again to take advantage of any path PMTU change. The
recommended probing interval is 10 minutes per RFC1981. Probing
interval could be sysctled by sysctl_tcp_probe_interval.

Eric Dumazet suggested to implement pseudo timer based on 32bits
jiffies tcp_time_stamp instead of using classic timer for such
rare event.
Signed-off-by: NFan Du <fan.du@intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Current probe_size is chosen by doubling mss_cache,
the probing process will end shortly with a sub-optimal
mss size, and the link mtu will not be taken full
advantage of, in return, this will make user to tweak
tcp_base_mss with care.

Use binary search to choose probe_size in a fine
granularity manner, an optimal mss will be found
to boost performance as its maxmium.

In addition, introduce a sysctl_tcp_probe_threshold
to control when probing will stop in respect to
the width of search range.

Test env:
Docker instance with vxlan encapuslation(82599EB)
iperf -c 10.0.0.24  -t 60

before this patch:
1.26 Gbits/sec

After this patch: increase 26%
1.59 Gbits/sec
Signed-off-by: NFan Du <fan.du@intel.com>
Acked-by: NJohn Heffner <johnwheffner@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Another TCP issue is triggered by ECN.

Under pressure, receiver gets ECN marks, and send back ACK packets
with ECE TCP flag. Senders enter CA_CWR state.

In this state, tcp_tso_should_defer() is short cut :

if (icsk->icsk_ca_state != TCP_CA_Open)
    goto send_now;

This means that about all ACK packets we receive are triggering
a partial send, and because cwnd is kept small, we can only send
a small amount of data for each incoming ACK,
which in return generate more ACK packets.

Allowing CA_Open and CA_CWR states to enable TSO defer in
tcp_tso_should_defer() brings performance back :
TSO autodefer has more chance to defer under pressure.

This patch increases TSO and LRO/GRO efficiency back to normal levels,
and does not impact overall ECN behavior.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NYuchung Cheng <ycheng@google.com>
Signed-off-by: NNeal Cardwell <ncardwell@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

With sysctl_tcp_min_tso_segs being 4, it is very possible
that tcp_tso_should_defer() decides not sending last 2 MSS
of initial window of 10 packets. This also applies if
autosizing decides to send X MSS per GSO packet, and cwnd
is not a multiple of X.

This patch implements an heuristic based on age of first
skb in write queue : If it was sent very recently (less than half srtt),
we can predict that no ACK packet will come in less than half rtt,
so deferring might cause an under utilization of our window.

This is visible on initial send (IW10) on web servers,
but more generally on some RPC, as the last part of the message
might need an extra RTT to get delivered.

Tested:

Ran following packetdrill test
// A simple server-side test that sends exactly an initial window (IW10)
// worth of packets.

`sysctl -e -q net.ipv4.tcp_min_tso_segs=4`

0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3
+0    setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0
+0    bind(3, ..., ...) = 0
+0    listen(3, 1) = 0

+.1   < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 7>
+0    > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 6>
+.1   < . 1:1(0) ack 1 win 257
+0    accept(3, ..., ...) = 4

+0    write(4, ..., 14600) = 14600
+0    > . 1:5841(5840) ack 1 win 457
+0    > . 5841:11681(5840) ack 1 win 457
// Following packet should be sent right now.
+0    > P. 11681:14601(2920) ack 1 win 457

+.1   < . 1:1(0) ack 14601 win 257

+0    close(4) = 0
+0    > F. 14601:14601(0) ack 1
+.1   < F. 1:1(0) ack 14602 win 257
+0    > . 14602:14602(0) ack 2
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NYuchung Cheng <ycheng@google.com>
Signed-off-by: NNeal Cardwell <ncardwell@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

TSO relies on ability to defer sending a small amount of packets.
Heuristic is to wait for future ACKS in hope to send more packets at once.
Current algorithm uses a per socket tso_deferred field as a pseudo timer.

This pseudo timer relies on future ACK, but there is no guarantee
we receive them in time.

Fix would be to use a real timer, but cost of such timer is probably too
expensive for typical cases.

This patch changes the logic to test the time of last transmit,
because we should not add bursts of more than 1ms for any given flow.

We've used this patch for about two years at Google, before FQ/pacing
as it would reduce a fair amount of bursts.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NYuchung Cheng <ycheng@google.com>
Signed-off-by: NNeal Cardwell <ncardwell@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Packetization Layer Path MTU Discovery works separately beside
Path MTU Discovery at IP level, different net namespace has
various requirements on which one to chose, e.g., a virutalized
container instance would require TCP PMTU to probe an usable
effective mtu for underlying tunnel, while the host would
employ classical ICMP based PMTU to function.

Hence making TCP PMTU mechanism per net namespace to decouple
two functionality. Furthermore the probe base MSS should also
be configured separately for each namespace.
Signed-off-by: NFan Du <fan.du@intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When we added pacing to TCP, we decided to let sch_fq take care
of actual pacing.

All TCP had to do was to compute sk->pacing_rate using simple formula:

sk->pacing_rate = 2 * cwnd * mss / rtt

It works well for senders (bulk flows), but not very well for receivers
or even RPC :

cwnd on the receiver can be less than 10, rtt can be around 100ms, so we
can end up pacing ACK packets, slowing down the sender.

Really, only the sender should pace, according to its own logic.

Instead of adding a new bit in skb, or call yet another flow
dissection, we tweak skb->truesize to a small value (2), and
we instruct sch_fq to use new helper and not pace pure ack.

Note this also helps TCP small queue, as ack packets present
in qdisc/NIC do not prevent sending a data packet (RPC workload)

This helps to reduce tx completion overhead, ack packets can use regular
sock_wfree() instead of tcp_wfree() which is a bit more expensive.

This has no impact in the case packets are sent to loopback interface,
as we do not coalesce ack packets (were we would detect skb->truesize
lie)

In case netem (with a delay) is used, skb_orphan_partial() also sets
skb->truesize to 1.

This patch is a combination of two patches we used for about one year at
Google.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

patch is actually smaller than it seems to be - most of it is unindenting
the inner loop body in tcp_sendmsg() itself...

the bit in tcp_input.c is going to get reverted very soon - that's what
memcpy_from_msg() will become, but not in this commit; let's keep it
reasonably contained...

There's one potentially subtle change here: in case of short copy from
userland, mainline tcp_send_syn_data() discards the skb it has allocated
and falls back to normal path, where we'll send as much as possible after
rereading the same data again.  This patch trims SYN+data skb instead -
that way we don't need to copy from the same place twice.
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

This work adds the possibility to define a per route/destination
congestion control algorithm. Generally, this opens up the possibility
for a machine with different links to enforce specific congestion
control algorithms with optimal strategies for each of them based
on their network characteristics, even transparently for a single
application listening on all links.

For our specific use case, this additionally facilitates deployment
of DCTCP, for example, applications can easily serve internal
traffic/dsts in DCTCP and external one with CUBIC. Other scenarios
would also allow for utilizing e.g. long living, low priority
background flows for certain destinations/routes while still being
able for normal traffic to utilize the default congestion control
algorithm. We also thought about a per netns setting (where different
defaults are possible), but given its actually a link specific
property, we argue that a per route/destination setting is the most
natural and flexible.

The administrator can utilize this through ip-route(8) by appending
"congctl [lock] <name>", where <name> denotes the name of a
congestion control algorithm and the optional lock parameter allows
to enforce the given algorithm so that applications in user space
would not be allowed to overwrite that algorithm for that destination.

The dst metric lookups are being done when a dst entry is already
available in order to avoid a costly lookup and still before the
algorithms are being initialized, thus overhead is very low when the
feature is not being used. While the client side would need to drop
the current reference on the module, on server side this can actually
even be avoided as we just got a flat-copied socket clone.

Joint work with Florian Westphal.
Suggested-by: NHannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: NFlorian Westphal <fw@strlen.de>
Signed-off-by: NDaniel Borkmann <dborkman@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Thomas Jarosch reported IPsec TCP stalls when a PMTU event occurs.

In fact the problem was completely unrelated to IPsec.  The bug is
also reproducible if you just disable TSO/GSO.

The problem is that when the MSS goes down, existing queued packet
on the TX queue that have not been transmitted yet all look like
TSO packets and get treated as such.

This then triggers a bug where tcp_mss_split_point tells us to
generate a zero-sized packet on the TX queue.  Once that happens
we're screwed because the zero-sized packet can never be removed
by ACKs.

Fixes: 1485348d ("tcp: Apply device TSO segment limit earlier")
Reported-by: NThomas Jarosch <thomas.jarosch@intra2net.com>
Signed-off-by: NHerbert Xu <herbert@gondor.apana.org.au>

Cheers,
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Commit 95bd09eb ("tcp: TSO packets automatic sizing") tried to
control TSO size, but did this at the wrong place (sendmsg() time)

At sendmsg() time, we might have a pessimistic view of flow rate,
and we end up building very small skbs (with 2 MSS per skb).

This is bad because :

 - It sends small TSO packets even in Slow Start where rate quickly
   increases.
 - It tends to make socket write queue very big, increasing tcp_ack()
   processing time, but also increasing memory needs, not necessarily
   accounted for, as fast clones overhead is currently ignored.
 - Lower GRO efficiency and more ACK packets.

Servers with a lot of small lived connections suffer from this.

Lets instead fill skbs as much as possible (64KB of payload), but split
them at xmit time, when we have a precise idea of the flow rate.
skb split is actually quite efficient.

Patch looks bigger than necessary, because TCP Small Queue decision now
has to take place after the eventual split.

As Neal suggested, introduce a new tcp_tso_autosize() helper, so that
tcp_tso_should_defer() can be synchronized on same goal.

Rename tp->xmit_size_goal_segs to tp->gso_segs, as this variable
contains number of mss that we can put in GSO packet, and is not
related to the autosizing goal anymore.

Tested:

40 ms rtt link

nstat >/dev/null
netperf -H remote -l -2000000 -- -s 1000000
nstat | egrep "IpInReceives|IpOutRequests|TcpOutSegs|IpExtOutOctets"

Before patch :

Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/s

 87380 2000000 2000000    0.36         44.22
IpInReceives                    600                0.0
IpOutRequests                   599                0.0
TcpOutSegs                      1397               0.0
IpExtOutOctets                  2033249            0.0

After patch :

Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec

 87380 2000000 2000000    0.36       44.27
IpInReceives                    221                0.0
IpOutRequests                   232                0.0
TcpOutSegs                      1397               0.0
IpExtOutOctets                  2013953            0.0
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NNeal Cardwell <ncardwell@google.com>
Acked-by: NYuchung Cheng <ycheng@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Note that the code _using_ ->msg_iter at that point will be very
unhappy with anything other than unshifted iovec-backed iov_iter.
We still need to convert users to proper primitives.
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

While working on sk_forward_alloc problems reported by Denys
Fedoryshchenko, we found that tcp connect() (and fastopen) do not call
sk_wmem_schedule() for SYN packet (and/or SYN/DATA packet), so
sk_forward_alloc is negative while connect is in progress.

We can fix this by calling regular sk_stream_alloc_skb() both for the
SYN packet (in tcp_connect()) and the syn_data packet in
tcp_send_syn_data()

Then, tcp_send_syn_data() can avoid copying syn_data as we simply
can manipulate syn_data->cb[] to remove SYN flag (and increment seq)

Instead of open coding memcpy_fromiovecend(), simply use this helper.

This leaves in socket write queue clean fast clone skbs.

This was tested against our fastopen packetdrill tests.
Reported-by: NDenys Fedoryshchenko <nuclearcat@nuclearcat.com>
Signed-off-by: NEric Dumazet <edumazet@google.com>
Acked-by: NYuchung Cheng <ycheng@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

In DC world, GSO packets initially cooked by tcp_sendmsg() are usually
big, as sk_pacing_rate is high.

When network is congested, cwnd can be smaller than the GSO packets
found in socket write queue. tcp_write_xmit() splits GSO packets
using the available cwnd, and we end up sending a single GSO packet,
consuming all available cwnd.

With GRO aggregation on the receiver, we might handle a single GRO
packet, sending back a single ACK.

1) This single ACK might be lost
   TLP or RTO are forced to attempt a retransmit.
2) This ACK releases a full cwnd, sender sends another big GSO packet,
   in a ping pong mode.

This behavior does not fill the pipes in the best way, because of
scheduling artifacts.

Make sure we always have at least two GSO packets in flight.

This allows us to safely increase GRO efficiency without risking
spurious retransmits.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Acked-by: NNeal Cardwell <ncardwell@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch allows to set ECN on a per-route basis in case the sysctl
tcp_ecn is not set to 1. In other words, when ECN is set for specific
routes, it provides a tcp_ecn=1 behaviour for that route while the rest
of the stack acts according to the global settings.

One can use 'ip route change dev $dev $net features ecn' to toggle this.

Having a more fine-grained per-route setting can be beneficial for various
reasons, for example, 1) within data centers, or 2) local ISPs may deploy
ECN support for their own video/streaming services [1], etc.

There was a recent measurement study/paper [2] which scanned the Alexa's
publicly available top million websites list from a vantage point in US,
Europe and Asia:

Half of the Alexa list will now happily use ECN (tcp_ecn=2, most likely
blamed to commit 255cac91 ("tcp: extend ECN sysctl to allow server-side
only ECN") ;)); the break in connectivity on-path was found is about
1 in 10,000 cases. Timeouts rather than receiving back RSTs were much
more common in the negotiation phase (and mostly seen in the Alexa
middle band, ranks around 50k-150k): from 12-thousand hosts on which
there _may_ be ECN-linked connection failures, only 79 failed with RST
when _not_ failing with RST when ECN is not requested.

It's unclear though, how much equipment in the wild actually marks CE
when buffers start to fill up.

We thought about a fallback to non-ECN for retransmitted SYNs as another
global option (which could perhaps one day be made default), but as Eric
points out, there's much more work needed to detect broken middleboxes.

Two examples Eric mentioned are buggy firewalls that accept only a single
SYN per flow, and middleboxes that successfully let an ECN flow establish,
but later mark CE for all packets (so cwnd converges to 1).

 [1] http://www.ietf.org/proceedings/89/slides/slides-89-tsvarea-1.pdf, p.15
 [2] http://ecn.ethz.ch/

Joint work with Daniel Borkmann.

Reference: http://thread.gmane.org/gmane.linux.network/335797Suggested-by: NHannes Frederic Sowa <hannes@stressinduktion.org>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDaniel Borkmann <dborkman@redhat.com>
Signed-off-by: NFlorian Westphal <fw@strlen.de>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Some drivers are unable to perform TX completions in a bound time.
They instead call skb_orphan()

Problem is skb_fclone_busy() has to detect this case, otherwise
we block TCP retransmits and can freeze unlucky tcp sessions on
mostly idle hosts.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Fixes: 1f3279ae ("tcp: avoid retransmits of TCP packets hanging in host queues")
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

TCP Small queues tries to keep number of packets in qdisc
as small as possible, and depends on a tasklet to feed following
packets at TX completion time.
Choice of tasklet was driven by latencies requirements.

Then, TCP stack tries to avoid reorders, by locking flows with
outstanding packets in qdisc in a given TX queue.

What can happen is that many flows get attracted by a low performing
TX queue, and cpu servicing TX completion has to feed packets for all of
them, making this cpu 100% busy in softirq mode.

This became particularly visible with latest skb->xmit_more support

Strategy adopted in this patch is to detect when tcp_wfree() is called
from ksoftirqd and let the outstanding queue for this flow being drained
before feeding additional packets, so that skb->ooo_okay can be set
to allow select_queue() to select the optimal queue :

Incoming ACKS are normally handled by different cpus, so this patch
gives more chance for these cpus to take over the burden of feeding
qdisc with future packets.

Tested:

lpaa23:~# ./super_netperf 1400 --google-pacing-rate 3028000 -H lpaa24 -l 3600 &

lpaa23:~# sar -n DEV 1 10 | grep eth1
06:16:18 AM      eth1 595448.00 1190564.00  38381.09 1760253.12      0.00      0.00      1.00
06:16:19 AM      eth1 594858.00 1189686e.00  38340.76 1758952.72      0.00      0.00      0.00
06:16:20 AM      eth1 597017.00 1194019.00  38480.79 1765370.29      0.00      0.00      1.00
06:16:21 AM      eth1 595450.00 1190936.00  38380.19 1760805.05      0.00      0.00      0.00
06:16:22 AM      eth1 596385.00 1193096.00  38442.56 1763976.29      0.00      0.00      1.00
06:16:23 AM      eth1 598155.00 1195978.00  38552.97 1768264.60      0.00      0.00      0.00
06:16:24 AM      eth1 594405.00 1188643.00  38312.57 1757414.89      0.00      0.00      1.00
06:16:25 AM      eth1 593366.00 1187154.00  38252.16 1755195.83      0.00      0.00      0.00
06:16:26 AM      eth1 593188.00 1186118.00  38232.88 1753682.57      0.00      0.00      1.00
06:16:27 AM      eth1 596301.00 1192241.00  38440.94 1762733.09      0.00      0.00      0.00
Average:         eth1 595457.30 1190843.50  38381.69 1760664.84      0.00      0.00      0.50
lpaa23:~# ./tc -s -d qd sh dev eth1 | grep backlog
 backlog 7606336b 2513p requeues 167982
 backlog 224072b 74p requeues 566
 backlog 581376b 192p requeues 5598
 backlog 181680b 60p requeues 1070
 backlog 5305056b 1753p requeues 110166    // Here, this TX queue is attracting flows
 backlog 157456b 52p requeues 1758
 backlog 672216b 222p requeues 3025
 backlog 60560b 20p requeues 24541
 backlog 448144b 148p requeues 21258

lpaa23:~# echo 1 >/proc/sys/net/ipv4/tcp_tsq_enable_tcp_wfree_ksoftirqd_detect

Immediate jump to full bandwidth, and traffic is properly
shard on all tx queues.

lpaa23:~# sar -n DEV 1 10 | grep eth1
06:16:46 AM      eth1 1397632.00 2795397.00  90081.87 4133031.26      0.00      0.00      1.00
06:16:47 AM      eth1 1396874.00 2793614.00  90032.99 4130385.46      0.00      0.00      0.00
06:16:48 AM      eth1 1395842.00 2791600.00  89966.46 4127409.67      0.00      0.00      1.00
06:16:49 AM      eth1 1395528.00 2791017.00  89946.17 4126551.24      0.00      0.00      0.00
06:16:50 AM      eth1 1397891.00 2795716.00  90098.74 4133497.39      0.00      0.00      1.00
06:16:51 AM      eth1 1394951.00 2789984.00  89908.96 4125022.51      0.00      0.00      0.00
06:16:52 AM      eth1 1394608.00 2789190.00  89886.90 4123851.36      0.00      0.00      1.00
06:16:53 AM      eth1 1395314.00 2790653.00  89934.33 4125983.09      0.00      0.00      0.00
06:16:54 AM      eth1 1396115.00 2792276.00  89984.25 4128411.21      0.00      0.00      1.00
06:16:55 AM      eth1 1396829.00 2793523.00  90030.19 4130250.28      0.00      0.00      0.00
Average:         eth1 1396158.40 2792297.00  89987.09 4128439.35      0.00      0.00      0.50

lpaa23:~# tc -s -d qd sh dev eth1 | grep backlog
 backlog 7900052b 2609p requeues 173287
 backlog 878120b 290p requeues 589
 backlog 1068884b 354p requeues 5621
 backlog 996212b 329p requeues 1088
 backlog 984100b 325p requeues 115316
 backlog 956848b 316p requeues 1781
 backlog 1080996b 357p requeues 3047
 backlog 975016b 322p requeues 24571
 backlog 990156b 327p requeues 21274

(All 8 TX queues get a fair share of the traffic)
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

TCP Small Queues (tcp_tsq_handler()) can hold one reference on
sk->sk_wmem_alloc, preventing skb->ooo_okay being set.

We should relax test done to set skb->ooo_okay to take care
of this extra reference.

Minimal truesize of skb containing one byte of payload is
SKB_TRUESIZE(1)

Without this fix, we have more chance locking flows into the wrong
transmit queue.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Lets use a proper structure to clearly document and implement
skb fast clones.

Then, we might experiment more easily alternative layouts.

This patch adds a new skb_fclone_busy() helper, used by tcp and xfrm,
to stop leaking of implementation details.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Suggested by Stephen. Also drop inline keyword and let compiler decide.

gcc 4.7.3 decides to no longer inline tcp_ecn_check_ce, so split it up.
The actual evaluation is not inlined anymore while the ECN_OK test is.
Suggested-by: NStephen Hemminger <stephen@networkplumber.org>
Signed-off-by: NDaniel Borkmann <dborkman@redhat.com>
Signed-off-by: NFlorian Westphal <fw@strlen.de>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This work adds the DataCenter TCP (DCTCP) congestion control
algorithm [1], which has been first published at SIGCOMM 2010 [2],
resp. follow-up analysis at SIGMETRICS 2011 [3] (and also, more
recently as an informational IETF draft available at [4]).

DCTCP is an enhancement to the TCP congestion control algorithm for
data center networks. Typical data center workloads are i.e.
i) partition/aggregate (queries; bursty, delay sensitive), ii) short
messages e.g. 50KB-1MB (for coordination and control state; delay
sensitive), and iii) large flows e.g. 1MB-100MB (data update;
throughput sensitive). DCTCP has therefore been designed for such
environments to provide/achieve the following three requirements:

  * High burst tolerance (incast due to partition/aggregate)
  * Low latency (short flows, queries)
  * High throughput (continuous data updates, large file
    transfers) with commodity, shallow buffered switches

The basic idea of its design consists of two fundamentals: i) on the
switch side, packets are being marked when its internal queue
length > threshold K (K is chosen so that a large enough headroom
for marked traffic is still available in the switch queue); ii) the
sender/host side maintains a moving average of the fraction of marked
packets, so each RTT, F is being updated as follows:

 F := X / Y, where X is # of marked ACKs, Y is total # of ACKs
 alpha := (1 - g) * alpha + g * F, where g is a smoothing constant

The resulting alpha (iow: probability that switch queue is congested)
is then being used in order to adaptively decrease the congestion
window W:

 W := (1 - (alpha / 2)) * W

The means for receiving marked packets resp. marking them on switch
side in DCTCP is the use of ECN.

RFC3168 describes a mechanism for using Explicit Congestion Notification
from the switch for early detection of congestion, rather than waiting
for segment loss to occur.

However, this method only detects the presence of congestion, not
the *extent*. In the presence of mild congestion, it reduces the TCP
congestion window too aggressively and unnecessarily affects the
throughput of long flows [4].

DCTCP, as mentioned, enhances Explicit Congestion Notification (ECN)
processing to estimate the fraction of bytes that encounter congestion,
rather than simply detecting that some congestion has occurred. DCTCP
then scales the TCP congestion window based on this estimate [4],
thus it can derive multibit feedback from the information present in
the single-bit sequence of marks in its control law. And thus act in
*proportion* to the extent of congestion, not its *presence*.

Switches therefore set the Congestion Experienced (CE) codepoint in
packets when internal queue lengths exceed threshold K. Resulting,
DCTCP delivers the same or better throughput than normal TCP, while
using 90% less buffer space.

It was found in [2] that DCTCP enables the applications to handle 10x
the current background traffic, without impacting foreground traffic.
Moreover, a 10x increase in foreground traffic did not cause any
timeouts, and thus largely eliminates TCP incast collapse problems.

The algorithm itself has already seen deployments in large production
data centers since then.

We did a long-term stress-test and analysis in a data center, short
summary of our TCP incast tests with iperf compared to cubic:

This test measured DCTCP throughput and latency and compared it with
CUBIC throughput and latency for an incast scenario. In this test, 19
senders sent at maximum rate to a single receiver. The receiver simply
ran iperf -s.

The senders ran iperf -c <receiver> -t 30. All senders started
simultaneously (using local clocks synchronized by ntp).

This test was repeated multiple times. Below shows the results from a
single test. Other tests are similar. (DCTCP results were extremely
consistent, CUBIC results show some variance induced by the TCP timeouts
that CUBIC encountered.)

For this test, we report statistics on the number of TCP timeouts,
flow throughput, and traffic latency.

1) Timeouts (total over all flows, and per flow summaries):

            CUBIC            DCTCP
  Total     3227             25
  Mean       169.842          1.316
  Median     183              1
  Max        207              5
  Min        123              0
  Stddev      28.991          1.600

Timeout data is taken by measuring the net change in netstat -s
"other TCP timeouts" reported. As a result, the timeout measurements
above are not restricted to the test traffic, and we believe that it
is likely that all of the "DCTCP timeouts" are actually timeouts for
non-test traffic. We report them nevertheless. CUBIC will also include
some non-test timeouts, but they are drawfed by bona fide test traffic
timeouts for CUBIC. Clearly DCTCP does an excellent job of preventing
TCP timeouts. DCTCP reduces timeouts by at least two orders of
magnitude and may well have eliminated them in this scenario.

2) Throughput (per flow in Mbps):

            CUBIC            DCTCP
  Mean      521.684          521.895
  Median    464              523
  Max       776              527
  Min       403              519
  Stddev    105.891            2.601
  Fairness    0.962            0.999

Throughput data was simply the average throughput for each flow
reported by iperf. By avoiding TCP timeouts, DCTCP is able to
achieve much better per-flow results. In CUBIC, many flows
experience TCP timeouts which makes flow throughput unpredictable and
unfair. DCTCP, on the other hand, provides very clean predictable
throughput without incurring TCP timeouts. Thus, the standard deviation
of CUBIC throughput is dramatically higher than the standard deviation
of DCTCP throughput.

Mean throughput is nearly identical because even though cubic flows
suffer TCP timeouts, other flows will step in and fill the unused
bandwidth. Note that this test is something of a best case scenario
for incast under CUBIC: it allows other flows to fill in for flows
experiencing a timeout. Under situations where the receiver is issuing
requests and then waiting for all flows to complete, flows cannot fill
in for timed out flows and throughput will drop dramatically.

3) Latency (in ms):

            CUBIC            DCTCP
  Mean      4.0088           0.04219
  Median    4.055            0.0395
  Max       4.2              0.085
  Min       3.32             0.028
  Stddev    0.1666           0.01064

Latency for each protocol was computed by running "ping -i 0.2
<receiver>" from a single sender to the receiver during the incast
test. For DCTCP, "ping -Q 0x6 -i 0.2 <receiver>" was used to ensure
that traffic traversed the DCTCP queue and was not dropped when the
queue size was greater than the marking threshold. The summary
statistics above are over all ping metrics measured between the single
sender, receiver pair.

The latency results for this test show a dramatic difference between
CUBIC and DCTCP. CUBIC intentionally overflows the switch buffer
which incurs the maximum queue latency (more buffer memory will lead
to high latency.) DCTCP, on the other hand, deliberately attempts to
keep queue occupancy low. The result is a two orders of magnitude
reduction of latency with DCTCP - even with a switch with relatively
little RAM. Switches with larger amounts of RAM will incur increasing
amounts of latency for CUBIC, but not for DCTCP.

4) Convergence and stability test:

This test measured the time that DCTCP took to fairly redistribute
bandwidth when a new flow commences. It also measured DCTCP's ability
to remain stable at a fair bandwidth distribution. DCTCP is compared
with CUBIC for this test.

At the commencement of this test, a single flow is sending at maximum
rate (near 10 Gbps) to a single receiver. One second after that first
flow commences, a new flow from a distinct server begins sending to
the same receiver as the first flow. After the second flow has sent
data for 10 seconds, the second flow is terminated. The first flow
sends for an additional second. Ideally, the bandwidth would be evenly
shared as soon as the second flow starts, and recover as soon as it
stops.

The results of this test are shown below. Note that the flow bandwidth
for the two flows was measured near the same time, but not
simultaneously.

DCTCP performs nearly perfectly within the measurement limitations
of this test: bandwidth is quickly distributed fairly between the two
flows, remains stable throughout the duration of the test, and
recovers quickly. CUBIC, in contrast, is slow to divide the bandwidth
fairly, and has trouble remaining stable.

  CUBIC                      DCTCP

  Seconds  Flow 1  Flow 2    Seconds  Flow 1  Flow 2
   0       9.93    0          0       9.92    0
   0.5     9.87    0          0.5     9.86    0
   1       8.73    2.25       1       6.46    4.88
   1.5     7.29    2.8        1.5     4.9     4.99
   2       6.96    3.1        2       4.92    4.94
   2.5     6.67    3.34       2.5     4.93    5
   3       6.39    3.57       3       4.92    4.99
   3.5     6.24    3.75       3.5     4.94    4.74
   4       6       3.94       4       5.34    4.71
   4.5     5.88    4.09       4.5     4.99    4.97
   5       5.27    4.98       5       4.83    5.01
   5.5     4.93    5.04       5.5     4.89    4.99
   6       4.9     4.99       6       4.92    5.04
   6.5     4.93    5.1        6.5     4.91    4.97
   7       4.28    5.8        7       4.97    4.97
   7.5     4.62    4.91       7.5     4.99    4.82
   8       5.05    4.45       8       5.16    4.76
   8.5     5.93    4.09       8.5     4.94    4.98
   9       5.73    4.2        9       4.92    5.02
   9.5     5.62    4.32       9.5     4.87    5.03
  10       6.12    3.2       10       4.91    5.01
  10.5     6.91    3.11      10.5     4.87    5.04
  11       8.48    0         11       8.49    4.94
  11.5     9.87    0         11.5     9.9     0

SYN/ACK ECT test:

This test demonstrates the importance of ECT on SYN and SYN-ACK packets
by measuring the connection probability in the presence of competing
flows for a DCTCP connection attempt *without* ECT in the SYN packet.
The test was repeated five times for each number of competing flows.

              Competing Flows  1 |    2 |    4 |    8 |   16
                               ------------------------------
Mean Connection Probability    1 | 0.67 | 0.45 | 0.28 |    0
Median Connection Probability  1 | 0.65 | 0.45 | 0.25 |    0

As the number of competing flows moves beyond 1, the connection
probability drops rapidly.

Enabling DCTCP with this patch requires the following steps:

DCTCP must be running both on the sender and receiver side in your
data center, i.e.:

  sysctl -w net.ipv4.tcp_congestion_control=dctcp

Also, ECN functionality must be enabled on all switches in your
data center for DCTCP to work. The default ECN marking threshold (K)
heuristic on the switch for DCTCP is e.g., 20 packets (30KB) at
1Gbps, and 65 packets (~100KB) at 10Gbps (K > 1/7 * C * RTT, [4]).

In above tests, for each switch port, traffic was segregated into two
queues. For any packet with a DSCP of 0x01 - or equivalently a TOS of
0x04 - the packet was placed into the DCTCP queue. All other packets
were placed into the default drop-tail queue. For the DCTCP queue,
RED/ECN marking was enabled, here, with a marking threshold of 75 KB.
More details however, we refer you to the paper [2] under section 3).

There are no code changes required to applications running in user
space. DCTCP has been implemented in full *isolation* of the rest of
the TCP code as its own congestion control module, so that it can run
without a need to expose code to the core of the TCP stack, and thus
nothing changes for non-DCTCP users.

Changes in the CA framework code are minimal, and DCTCP algorithm
operates on mechanisms that are already available in most Silicon.
The gain (dctcp_shift_g) is currently a fixed constant (1/16) from
the paper, but we leave the option that it can be chosen carefully
to a different value by the user.

In case DCTCP is being used and ECN support on peer site is off,
DCTCP falls back after 3WHS to operate in normal TCP Reno mode.

ss {-4,-6} -t -i diag interface:

  ... dctcp wscale:7,7 rto:203 rtt:2.349/0.026 mss:1448 cwnd:2054
  ssthresh:1102 ce_state 0 alpha 15 ab_ecn 0 ab_tot 735584
  send 10129.2Mbps pacing_rate 20254.1Mbps unacked:1822 retrans:0/15
  reordering:101 rcv_space:29200

  ... dctcp-reno wscale:7,7 rto:201 rtt:0.711/1.327 ato:40 mss:1448
  cwnd:10 ssthresh:1102 fallback_mode send 162.9Mbps pacing_rate
  325.5Mbps rcv_rtt:1.5 rcv_space:29200

More information about DCTCP can be found in [1-4].

  [1] http://simula.stanford.edu/~alizade/Site/DCTCP.html
  [2] http://simula.stanford.edu/~alizade/Site/DCTCP_files/dctcp-final.pdf
  [3] http://simula.stanford.edu/~alizade/Site/DCTCP_files/dctcp_analysis-full.pdf
  [4] http://tools.ietf.org/html/draft-bensley-tcpm-dctcp-00

Joint work with Florian Westphal and Glenn Judd.
Signed-off-by: NDaniel Borkmann <dborkman@redhat.com>
Signed-off-by: NFlorian Westphal <fw@strlen.de>
Signed-off-by: NGlenn Judd <glenn.judd@morganstanley.com>
Acked-by: NStephen Hemminger <stephen@networkplumber.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

DataCenter TCP (DCTCP) determines cwnd growth based on ECN information
and ACK properties, e.g. ACK that updates window is treated differently
than DUPACK.

Also DCTCP needs information whether ACK was delayed ACK. Furthermore,
DCTCP also implements a CE state machine that keeps track of CE markings
of incoming packets.

Therefore, extend the congestion control framework to provide these
event types, so that DCTCP can be properly implemented as a normal
congestion algorithm module outside of the core stack.

Joint work with Daniel Borkmann and Glenn Judd.
Signed-off-by: NFlorian Westphal <fw@strlen.de>
Signed-off-by: NDaniel Borkmann <dborkman@redhat.com>
Signed-off-by: NGlenn Judd <glenn.judd@morganstanley.com>
Acked-by: NStephen Hemminger <stephen@networkplumber.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch adds a flag to TCP congestion algorithms that allows
for requesting to mark IPv4/IPv6 sockets with transport as ECN
capable, that is, ECT(0), when required by a congestion algorithm.

It is currently used and needed in DataCenter TCP (DCTCP), as it
requires both peers to assert ECT on all IP packets sent - it
uses ECN feedback (i.e. CE, Congestion Encountered information)
from switches inside the data center to derive feedback to the
end hosts.

Therefore, simply add a new flag to icsk_ca_ops. Note that DCTCP's
algorithm/behaviour slightly diverges from RFC3168, therefore this
is only (!) enabled iff the assigned congestion control ops module
has requested this. By that, we can tightly couple this logic really
only to the provided congestion control ops.

Joint work with Florian Westphal and Glenn Judd.
Signed-off-by: NDaniel Borkmann <dborkman@redhat.com>
Signed-off-by: NFlorian Westphal <fw@strlen.de>
Signed-off-by: NGlenn Judd <glenn.judd@morganstanley.com>
Acked-by: NStephen Hemminger <stephen@networkplumber.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Our goal is to access no more than one cache line access per skb in
a write or receive queue when doing the various walks.

After recent TCP_SKB_CB() reorganizations, it is almost done.

Last part is tcp_skb_pcount() which currently uses
skb_shinfo(skb)->gso_segs, which is a terrible choice, because it needs
3 cache lines in current kernel (skb->head, skb->end, and
shinfo->gso_segs are all in 3 different cache lines, far from skb->cb)

This very simple patch reuses space currently taken by tcp_tw_isn
only in input path, as tcp_skb_pcount is only needed for skb stored in
write queue.

This considerably speeds up tcp_ack(), granted we avoid shinfo->tx_flags
to get SKBTX_ACK_TSTAMP, which seems possible.

This also speeds up all sack processing in general.

This speeds up tcp_sendmsg() because it no longer has to access/dirty
shinfo.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

TCP maintains lists of skb in write queue, and in receive queues
(in order and out of order queues)

Scanning these lists both in input and output path usually requires
access to skb->next, TCP_SKB_CB(skb)->seq, and TCP_SKB_CB(skb)->end_seq

These fields are currently in two different cache lines, meaning we
waste lot of memory bandwidth when these queues are big and flows
have either packet drops or packet reorders.

We can move TCP_SKB_CB(skb)->header at the end of TCP_SKB_CB, because
this header is not used in fast path. This allows TCP to search much faster
in the skb lists.

Even with regular flows, we save one cache line miss in fast path.

Thanks to Christoph Paasch for noticing we need to cleanup
skb->cb[] (IPCB/IP6CB) before entering IP stack in tx path,
and that I forgot IPCB use in tcp_v4_hnd_req() and tcp_v4_save_options().
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

While profiling TCP stack, I noticed one useless atomic operation
in tcp_sendmsg(), caused by skb_header_release().

It turns out all current skb_header_release() users have a fresh skb,
that no other user can see, so we can avoid one atomic operation.

Introduce __skb_header_release() to clearly document this.

This gave me a 1.5 % improvement on TCP_RR workload.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

icsk_rto is a 32bit field, and icsk_backoff can reach 15 by default,
or more if some sysctl (eg tcp_retries2) are changed.

Better use 64bit to perform icsk_rto << icsk_backoff operations

As Joe Perches suggested, add a helper for this.

Yuchung spotted the tcp_v4_err() case.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The TCP_SKB_CB(skb)->when field no longer exists as of recent change
7faee5c0 ("tcp: remove TCP_SKB_CB(skb)->when"). And in any case,
tcp_fragment() is called on already-transmitted packets from the
__tcp_retransmit_skb() call site, so copying timestamps of any kind
in this spot is quite sensible.
Signed-off-by: NNeal Cardwell <ncardwell@google.com>
Reported-by: NYuchung Cheng <ycheng@google.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

After commit 740b0f18 ("tcp: switch rtt estimations to usec resolution"),
we no longer need to maintain timestamps in two different fields.

TCP_SKB_CB(skb)->when can be removed, as same information sits in skb_mstamp.stamp_jiffies
Signed-off-by: NEric Dumazet <edumazet@google.com>
Acked-by: NYuchung Cheng <ycheng@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Replace uses of get_cpu_var for address calculation through this_cpu_ptr.

Cc: netdev@vger.kernel.org
Cc: Eric Dumazet <edumazet@google.com>
Acked-by: NDavid S. Miller <davem@davemloft.net>
Signed-off-by: NChristoph Lameter <cl@linux.com>
Signed-off-by: NTejun Heo <tj@kernel.org>

Make sure we use the correct address-family-specific function for
handling MTU reductions from within tcp_release_cb().

Previously AF_INET6 sockets were incorrectly always using the IPv6
code path when sometimes they were handling IPv4 traffic and thus had
an IPv4 dst.
Signed-off-by: NNeal Cardwell <ncardwell@google.com>
Signed-off-by: NEric Dumazet <edumazet@google.com>
Diagnosed-by: NWillem de Bruijn <willemb@google.com>
Fixes: 563d34d0 ("tcp: dont drop MTU reduction indications")
Reviewed-by: NHannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

We don't know right timestamp for repaired skb-s. Wrong RTT estimations
isn't good, because some congestion modules heavily depends on it.

This patch adds the TCPCB_REPAIRED flag, which is included in
TCPCB_RETRANS.

Thanks to Eric for the advice how to fix this issue.

This patch fixes the warning:
[  879.562947] WARNING: CPU: 0 PID: 2825 at net/ipv4/tcp_input.c:3078 tcp_ack+0x11f5/0x1380()
[  879.567253] CPU: 0 PID: 2825 Comm: socket-tcpbuf-l Not tainted 3.16.0-next-20140811 #1
[  879.567829] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
[  879.568177]  0000000000000000 00000000c532680c ffff880039643d00 ffffffff817aa2d2
[  879.568776]  0000000000000000 ffff880039643d38 ffffffff8109afbd ffff880039d6ba80
[  879.569386]  ffff88003a449800 000000002983d6bd 0000000000000000 000000002983d6bc
[  879.569982] Call Trace:
[  879.570264]  [<ffffffff817aa2d2>] dump_stack+0x4d/0x66
[  879.570599]  [<ffffffff8109afbd>] warn_slowpath_common+0x7d/0xa0
[  879.570935]  [<ffffffff8109b0ea>] warn_slowpath_null+0x1a/0x20
[  879.571292]  [<ffffffff816d0a05>] tcp_ack+0x11f5/0x1380
[  879.571614]  [<ffffffff816d10bd>] tcp_rcv_established+0x1ed/0x710
[  879.571958]  [<ffffffff816dc9da>] tcp_v4_do_rcv+0x10a/0x370
[  879.572315]  [<ffffffff81657459>] release_sock+0x89/0x1d0
[  879.572642]  [<ffffffff816c81a0>] do_tcp_setsockopt.isra.36+0x120/0x860
[  879.573000]  [<ffffffff8110a52e>] ? rcu_read_lock_held+0x6e/0x80
[  879.573352]  [<ffffffff816c8912>] tcp_setsockopt+0x32/0x40
[  879.573678]  [<ffffffff81654ac4>] sock_common_setsockopt+0x14/0x20
[  879.574031]  [<ffffffff816537b0>] SyS_setsockopt+0x80/0xf0
[  879.574393]  [<ffffffff817b40a9>] system_call_fastpath+0x16/0x1b
[  879.574730] ---[ end trace a17cbc38eb8c5c00 ]---

v2: moving setting of skb->when for repaired skb-s in tcp_write_xmit,
    where it's set for other skb-s.

Fixes: 431a9124 ("tcp: timestamp SYN+DATA messages")
Fixes: 740b0f18 ("tcp: switch rtt estimations to usec resolution")
Cc: Eric Dumazet <edumazet@google.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: "David S. Miller" <davem@davemloft.net>
Signed-off-by: NAndrey Vagin <avagin@openvz.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Bytestream timestamps are correlated with a single byte in the skbuff,
recorded in skb_shinfo(skb)->tskey. When fragmenting skbuffs, ensure
that the tskey is set for the fragment in which the tskey falls
(seqno <= tskey < end_seqno).

The original implementation did not address fragmentation in
tcp_fragment or tso_fragment. Add code to inspect the sequence numbers
and move both tskey and the relevant tx_flags if necessary.
Reported-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NWillem de Bruijn <willemb@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Since Yuchung's 9b44190d (tcp: refactor F-RTO), tcp_enter_cwr is always
called with set_ssthresh = 1. Thus, we can remove this argument from
tcp_enter_cwr. Further, as we remove this one, tcp_init_cwnd_reduction
is then always called with set_ssthresh = true, and so we can get rid of
this argument as well.

Cc: Yuchung Cheng <ycheng@google.com>
Signed-off-by: NChristoph Paasch <christoph.paasch@uclouvain.be>
Acked-by: NYuchung Cheng <ycheng@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The undo code assumes that, upon entering loss recovery, TCP
1) always retransmit something
2) the retransmission never fails locally (e.g., qdisc drop)

so undo_marker is set in tcp_enter_recovery() and undo_retrans is
incremented only when tcp_retransmit_skb() is successful.

When the assumption is broken because TCP's cwnd is too small to
retransmit or the retransmit fails locally. The next (DUP)ACK
would incorrectly revert the cwnd and the congestion state in
tcp_try_undo_dsack() or tcp_may_undo(). Subsequent (DUP)ACKs
may enter the recovery state. The sender repeatedly enter and
(incorrectly) exit recovery states if the retransmits continue to
fail locally while receiving (DUP)ACKs.

The fix is to initialize undo_retrans to -1 and start counting on
the first retransmission. Always increment undo_retrans even if the
retransmissions fail locally because they couldn't cause DSACKs to
undo the cwnd reduction.
Signed-off-by: NYuchung Cheng <ycheng@google.com>
Signed-off-by: NNeal Cardwell <ncardwell@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

For a connected socket we can precompute the flow hash for setting
in skb->hash on output. This is a performance advantage over
calculating the skb->hash for every packet on the connection. The
computation is done using the common hash algorithm to be consistent
with computations done for packets of the connection in other states
where thers is no socket (e.g. time-wait, syn-recv, syn-cookies).

This patch adds sk_txhash to the sock structure. inet_set_txhash and
ip6_set_txhash functions are added which are called from points in
TCP and UDP where socket moves to established state.

skb_set_hash_from_sk is a function which sets skb->hash from the
sock txhash value. This is called in UDP and TCP transmit path when
transmitting within the context of a socket.

Tested: ran super_netperf with 200 TCP_RR streams over a vxlan
interface (in this case skb_get_hash called on every TX packet to
create a UDP source port).

Before fix:

  95.02% CPU utilization
  154/256/505 90/95/99% latencies
  1.13042e+06 tps

  Time in functions:
    0.28% skb_flow_dissect
    0.21% __skb_get_hash

After fix:

  94.95% CPU utilization
  156/254/485 90/95/99% latencies
  1.15447e+06

  Neither __skb_get_hash nor skb_flow_dissect appear in perf
Signed-off-by: NTom Herbert <therbert@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NOctavian Purdila <octavian.purdila@intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>