This patch adds NLM_F_APPEND flag to struct nlmsg_hdr->nlmsg_flags
in newroute notifications if the route add was an append.
(This is similar to how NLM_F_REPLACE is already part of new
route replace notifications today)

This helps userspace determine if the route add operation was
an append.
Signed-off-by: NRoopa Prabhu <roopa@cumulusnetworks.com>
Acked-by: NScott Feldman <sfeldma@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When an inet_sock is destroyed, its source port (sk_num) is set to
zero as part of the unhash procedure.  In order to supply a source
port as part of the NETLINK_SOCK_DIAG socket destruction broadcasts,
the source port number must be read from inet_sport instead.

Tested: ss -E
Signed-off-by: NCraig Gallek <kraig@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This get_info handler will simply dispatch to the appropriate
existing inet protocol handler.

This patch also includes a new netlink attribute
(INET_DIAG_PROTOCOL).  This attribute is currently only used
for multicast messages.  Without this attribute, there is no
way of knowing the IP protocol used by the socket information
being broadcast.  This attribute is not necessary in the 'dump'
variant of this protocol (though it could easily be added)
because dump requests are issued for specific family/protocol
pairs.

Tested: ss -E (note, the -E option has not yet been merged into
the upstream version of ss).
Signed-off-by: NCraig Gallek <kraig@google.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Previously, there was no clear distinction between the inet protocols
that used struct tcp_info to report information and those that didn't.
This change adds a specific size attribute to the inet_diag_handler
struct which defines these interfaces.  This will make dispatching
sock_diag get_info requests identical for all inet protocols in a
following patch.

Tested: ss -au
Tested: ss -at
Signed-off-by: NCraig Gallek <kraig@google.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

After Florian patches, there is no need for XT_TABLE_INFO_SZ anymore :
Only one copy of table is kept, instead of one copy per cpu.

We also can avoid a dereference if we put table data right after
xt_table_info. It reduces register pressure and helps compiler.

Then, we attempt a kmalloc() if total size is under order-3 allocation,
to reduce TLB pressure, as in many cases, rules fit in 32 KB.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Cc: Florian Westphal <fw@strlen.de>
Signed-off-by: NPablo Neira Ayuso <pablo@netfilter.org>

According to the reporter, they are not needed.
Reported-by: NSergei Shtylyov <sergei.shtylyov@cogentembedded.com>
Signed-off-by: NPablo Neira Ayuso <pablo@netfilter.org>

Fixes cross-compile to mips.
Signed-off-by: NKenneth Klette Jonassen <kennetkl@ifi.uio.no>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

text    data     bss     dec     hex filename
old: 16527      44       0   16571    40bb net/ipv4/ip_output.o
new: 14935      44       0   14979    3a83 net/ipv4/ip_output.o
Suggested-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NFlorian Westphal <fw@strlen.de>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

We store the rule blob per (possible) cpu.  Unfortunately this means we can
waste lot of memory on big smp machines. ipt_entry structure ('rule head')
is 112 byte, so e.g. with maxcpu=64 one single rule eats
close to 8k RAM.

Since previous patch made counters percpu it appears there is nothing
left in the rule blob that needs to be percpu.

On my test system (144 possible cpus, 400k dummy rules) this
change saves close to 9 Gigabyte of RAM.
Reported-by: NMarcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Acked-by: NJesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: NFlorian Westphal <fw@strlen.de>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NPablo Neira Ayuso <pablo@netfilter.org>

The binary arp/ip/ip6tables ruleset is stored per cpu.

The only reason left as to why we need percpu duplication are the rule
counters embedded into ipt_entry et al -- since each cpu has its own copy
of the rules, all counters can be lockless.

The downside is that the more cpus are supported, the more memory is
required.  Rules are not just duplicated per online cpu but for each
possible cpu, i.e. if maxcpu is 144, then rule is duplicated 144 times,
not for the e.g. 64 cores present.

To save some memory and also improve utilization of shared caches it
would be preferable to only store the rule blob once.

So we first need to separate counters and the rule blob.

Instead of using entry->counters, allocate this percpu and store the
percpu address in entry->counters.pcnt on CONFIG_SMP.

This change makes no sense as-is; it is merely an intermediate step to
remove the percpu duplication of the rule set in a followup patch.
Suggested-by: NEric Dumazet <edumazet@google.com>
Acked-by: NJesper Dangaard Brouer <brouer@redhat.com>
Reported-by: NMarcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: NFlorian Westphal <fw@strlen.de>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NPablo Neira Ayuso <pablo@netfilter.org>

since commit d6b915e2
("ip_fragment: don't forward defragmented DF packet") the largest
fragment size is available in the IPCB.

Therefore we no longer need to care about 'encapsulation'
overhead of stripped PPPOE/VLAN headers since ip_do_fragment
doesn't use device mtu in such cases.
Signed-off-by: NFlorian Westphal <fw@strlen.de>
Signed-off-by: NPablo Neira Ayuso <pablo@netfilter.org>

We had various issues in the past when TCP stack was modifying
gso_size/gso_segs while clones were in flight.

Commit c52e2421 ("tcp: must unclone packets before mangling them")
fixed these bugs and added a WARN_ON_ONCE(skb_cloned(skb)); in
tcp_set_skb_tso_segs()

These bugs are now fixed, and because TCP stack now only sets
shinfo->gso_size|segs on the clone itself, the check can be removed.

As a result of this change, compiler inlines tcp_set_skb_tso_segs() in
tcp_init_tso_segs()
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

In commit cd7d8498 ("tcp: change tcp_skb_pcount() location") we stored
gso_segs in a temporary cache hot location.

This patch does the same for gso_size.

This allows to save 2 cache line misses in tcp xmit path for
the last packet that is considered but not sent because of
various conditions (cwnd, tso defer, receiver window, TSQ...)
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

tcp_set_skb_tso_segs() & tcp_init_tso_segs() no longer
use the sock pointer.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Our goal is to touch skb_shinfo(skb) only when absolutely needed,
to avoid two cache line misses in TCP output path for last skb
that is considered but not sent because of various conditions
(cwnd, tso defer, receiver window, TSQ...)

A packet is GSO only when skb_shinfo(skb)->gso_size is not zero.

We can set skb_shinfo(skb)->gso_type to sk->sk_gso_type even for
non GSO packets.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

tcp_gso_segment() and tcp_gro_receive() are not strictly
part of TCP stack. They should not assume tcp_skb_mss(skb)
is in fact skb_shinfo(skb)->gso_size.

This will allow us to change tcp_skb_mss() in following patches.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

CAIA Delay-Gradient (CDG) is a TCP congestion control that modifies
the TCP sender in order to [1]:

  o Use the delay gradient as a congestion signal.
  o Back off with an average probability that is independent of the RTT.
  o Coexist with flows that use loss-based congestion control, i.e.,
    flows that are unresponsive to the delay signal.
  o Tolerate packet loss unrelated to congestion. (Disabled by default.)

Its FreeBSD implementation was presented for the ICCRG in July 2012;
slides are available at http://www.ietf.org/proceedings/84/iccrg.html

Running the experiment scenarios in [1] suggests that our implementation
achieves more goodput compared with FreeBSD 10.0 senders, although it also
causes more queueing delay for a given backoff factor.

The loss tolerance heuristic is disabled by default due to safety concerns
for its use in the Internet [2, p. 45-46].

We use a variant of the Hybrid Slow start algorithm in tcp_cubic to reduce
the probability of slow start overshoot.

[1] D.A. Hayes and G. Armitage. "Revisiting TCP congestion control using
    delay gradients." In Networking 2011, pages 328-341. Springer, 2011.
[2] K.K. Jonassen. "Implementing CAIA Delay-Gradient in Linux."
    MSc thesis. Department of Informatics, University of Oslo, 2015.

Cc: Eric Dumazet <edumazet@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Stephen Hemminger <stephen@networkplumber.org>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: David Hayes <davihay@ifi.uio.no>
Cc: Andreas Petlund <apetlund@simula.no>
Cc: Dave Taht <dave.taht@bufferbloat.net>
Cc: Nicolas Kuhn <nicolas.kuhn@telecom-bretagne.eu>
Signed-off-by: NKenneth Klette Jonassen <kennetkl@ifi.uio.no>
Acked-by: NYuchung Cheng <ycheng@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Upcoming tcp_cdg uses tcp_enter_cwr() to initiate PRR. Export this
function so that CDG can be compiled as a module.

Cc: Eric Dumazet <edumazet@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Stephen Hemminger <stephen@networkplumber.org>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: David Hayes <davihay@ifi.uio.no>
Cc: Andreas Petlund <apetlund@simula.no>
Cc: Dave Taht <dave.taht@bufferbloat.net>
Cc: Nicolas Kuhn <nicolas.kuhn@telecom-bretagne.eu>
Signed-off-by: NKenneth Klette Jonassen <kennetkl@ifi.uio.no>
Acked-by: NYuchung Cheng <ycheng@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

dctcp_alpha can be read by from dctcp_get_info() without
synchro, so use WRITE_ONCE() to prevent compiler from using
dctcp_alpha as a temporary variable.

Also, playing with small dctcp_shift_g (like 1), can expose
an overflow with 32bit values shifted 9 times before divide.

Use an u64 field to avoid this problem, and perform the divide
only if acked_bytes_ecn is not zero.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

As Alexander Duyck pointed out that:
struct tnode {
        ...
        struct key_vector kv[1];
}
The kv[1] member of struct tnode is an arry that refernced by
a null pointer will not crash the system, like this:
struct tnode *p = NULL;
struct key_vector *kv = p->kv;
As such p->kv doesn't actually dereference anything, it is simply a
means for getting the offset to the array from the pointer p.

This patch make the code more regular to avoid making people feel
odd when they look at the code.
Signed-off-by: NFiro Yang <firogm@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

IPv4 and IPv6 share same implementation of get_cookie_sock(),
and there is no point inlining it.

We add tcp_ prefix to the common helper name and export it.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

For same reasons than in commit 12e25e10 ("tcp: remove redundant
checks"), we can remove redundant checks done for timewait sockets.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When an application needs to force a source IP on an active TCP socket
it has to use bind(IP, port=x).

As most applications do not want to deal with already used ports, x is
often set to 0, meaning the kernel is in charge to find an available
port.
But kernel does not know yet if this socket is going to be a listener or
be connected.
It has very limited choices (no full knowledge of final 4-tuple for a
connect())

With limited ephemeral port range (about 32K ports), it is very easy to
fill the space.

This patch adds a new SOL_IP socket option, asking kernel to ignore
the 0 port provided by application in bind(IP, port=0) and only
remember the given IP address.

The port will be automatically chosen at connect() time, in a way
that allows sharing a source port as long as the 4-tuples are unique.

This new feature is available for both IPv4 and IPv6 (Thanks Neal)

Tested:

Wrote a test program and checked its behavior on IPv4 and IPv6.

strace(1) shows sequences of bind(IP=127.0.0.2, port=0) followed by
connect().
Also getsockname() show that the port is still 0 right after bind()
but properly allocated after connect().

socket(PF_INET, SOCK_STREAM, IPPROTO_IP) = 5
setsockopt(5, SOL_IP, IP_BIND_ADDRESS_NO_PORT, [1], 4) = 0
bind(5, {sa_family=AF_INET, sin_port=htons(0), sin_addr=inet_addr("127.0.0.2")}, 16) = 0
getsockname(5, {sa_family=AF_INET, sin_port=htons(0), sin_addr=inet_addr("127.0.0.2")}, [16]) = 0
connect(5, {sa_family=AF_INET, sin_port=htons(53174), sin_addr=inet_addr("127.0.0.3")}, 16) = 0
getsockname(5, {sa_family=AF_INET, sin_port=htons(38050), sin_addr=inet_addr("127.0.0.2")}, [16]) = 0

IPv6 test :

socket(PF_INET6, SOCK_STREAM, IPPROTO_IP) = 7
setsockopt(7, SOL_IP, IP_BIND_ADDRESS_NO_PORT, [1], 4) = 0
bind(7, {sa_family=AF_INET6, sin6_port=htons(0), inet_pton(AF_INET6, "::1", &sin6_addr), sin6_flowinfo=0, sin6_scope_id=0}, 28) = 0
getsockname(7, {sa_family=AF_INET6, sin6_port=htons(0), inet_pton(AF_INET6, "::1", &sin6_addr), sin6_flowinfo=0, sin6_scope_id=0}, [28]) = 0
connect(7, {sa_family=AF_INET6, sin6_port=htons(57300), inet_pton(AF_INET6, "::1", &sin6_addr), sin6_flowinfo=0, sin6_scope_id=0}, 28) = 0
getsockname(7, {sa_family=AF_INET6, sin6_port=htons(60964), inet_pton(AF_INET6, "::1", &sin6_addr), sin6_flowinfo=0, sin6_scope_id=0}, [28]) = 0

I was able to bind()/connect() a million concurrent IPv4 sockets,
instead of ~32000 before patch.

lpaa23:~# ulimit -n 1000010
lpaa23:~# ./bind --connect --num-flows=1000000 &
1000000 sockets

lpaa23:~# grep TCP /proc/net/sockstat
TCP: inuse 2000063 orphan 0 tw 47 alloc 2000157 mem 66

Check that a given source port is indeed used by many different
connections :

lpaa23:~# ss -t src :40000 | head -10
State      Recv-Q Send-Q   Local Address:Port          Peer Address:Port
ESTAB      0      0           127.0.0.2:40000         127.0.202.33:44983
ESTAB      0      0           127.0.0.2:40000         127.2.27.240:44983
ESTAB      0      0           127.0.0.2:40000           127.2.98.5:44983
ESTAB      0      0           127.0.0.2:40000        127.0.124.196:44983
ESTAB      0      0           127.0.0.2:40000         127.2.139.38:44983
ESTAB      0      0           127.0.0.2:40000          127.1.59.80:44983
ESTAB      0      0           127.0.0.2:40000          127.3.6.228:44983
ESTAB      0      0           127.0.0.2:40000          127.0.38.53:44983
ESTAB      0      0           127.0.0.2:40000         127.1.197.10:44983
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Xen virtual network driver has higher latency than a physical NIC.
Having only 128K as limit for TSQ introduced 30% regression in guest
throughput.

This patch raises the limit to 256K. This reduces the regression to 8%.
This buys us more time to work out a proper solution in the long run.
Signed-off-by: NWei Liu <wei.liu2@citrix.com>
Cc: David Miller <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

tcp_v4_rcv() checks the following before calling tcp_v4_do_rcv():

if (th->doff < sizeof(struct tcphdr) / 4)
    goto bad_packet;
if (!pskb_may_pull(skb, th->doff * 4))
    goto discard_it;

So following check in tcp_v4_do_rcv() is redundant
and "goto csum_err;" is wrong anyway.

if (skb->len < tcp_hdrlen(skb) || ...)
	goto csum_err;

A second check can be removed after no_tcp_socket label for same reason.

Same tests can be removed in tcp_v6_do_rcv()

Note : short tcp frames are not properly accounted in tcpInErrs MIB,
because pskb_may_pull() failure simply drops incoming skb, we might
fix this in a separate patch.
Signed-off-by: NEric Dumazet  <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

421b3885 "udp: ipv4: Add udp early
demux" introduced a regression that allowed sockets bound to INADDR_ANY
to receive packets from multicast groups that the socket had not joined.
For example a socket that had joined 224.168.2.9 could also receive
packets from 225.168.2.9 despite not having joined that group if
ip_early_demux is enabled.

Fix this by calling ip_check_mc_rcu() in udp_v4_early_demux() to verify
that the multicast packet is indeed ours.
Signed-off-by: NShawn Bohrer <sbohrer@rgmadvisors.com>
Reported-by: NYurij M. Plotnikov <Yurij.Plotnikov@oktetlabs.ru>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Linux 3.17 and earlier are explicitly engineered so that if the app
doesn't specifically request a CC module on a listener before the SYN
arrives, then the child gets the system default CC when the connection
is established. See tcp_init_congestion_control() in 3.17 or earlier,
which says "if no choice made yet assign the current value set as
default". The change ("net: tcp: assign tcp cong_ops when tcp sk is
created") altered these semantics, so that children got their parent
listener's congestion control even if the system default had changed
after the listener was created.

This commit returns to those original semantics from 3.17 and earlier,
since they are the original semantics from 2007 in 4d4d3d1e ("[TCP]:
Congestion control initialization."), and some Linux congestion
control workflows depend on that.

In summary, if a listener socket specifically sets TCP_CONGESTION to
"x", or the route locks the CC module to "x", then the child gets
"x". Otherwise the child gets current system default from
net.ipv4.tcp_congestion_control. That's the behavior in 3.17 and
earlier, and this commit restores that.

Fixes: 55d8694f ("net: tcp: assign tcp cong_ops when tcp sk is created")
Cc: Florian Westphal <fw@strlen.de>
Cc: Daniel Borkmann <dborkman@redhat.com>
Cc: Glenn Judd <glenn.judd@morganstanley.com>
Cc: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: NNeal Cardwell <ncardwell@google.com>
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NYuchung Cheng <ycheng@google.com>
Acked-by: NDaniel Borkmann <daniel@iogearbox.net>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

We have two problems in UDP stack related to bogus checksums :

1) We return -EAGAIN to application even if receive queue is not empty.
   This breaks applications using edge trigger epoll()

2) Under UDP flood, we can loop forever without yielding to other
   processes, potentially hanging the host, especially on non SMP.

This patch is an attempt to make things better.

We might in the future add extra support for rt applications
wanting to better control time spent doing a recv() in a hostile
environment. For example we could validate checksums before queuing
packets in socket receive queue.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Cc: Willem de Bruijn <willemb@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This is similar to b1cb59cf(net: sysctl_net_core: check SNDBUF
and RCVBUF for min length). I don't think too small values can cause
crashes in the case of udp and tcp, but I've seen this set to too
small values which triggered awful performance. It also makes the
setting consistent across all the wmem/rmem sysctls.
Signed-off-by: NSorin Dumitru <sdumitru@ixiacom.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The vti6_rcv_cb and vti_rcv_cb calls were leaving the skb->mark modified
after completing the function.  This resulted in the original skb->mark
value being lost.  Since we only need skb->mark to be set for
xfrm_policy_check we can pull the assignment into the rcv_cb calls and then
just restore the original mark after xfrm_policy_check has been completed.
Signed-off-by: NAlexander Duyck <alexander.h.duyck@redhat.com>
Signed-off-by: NSteffen Klassert <steffen.klassert@secunet.com>

Instead of modifying skb->mark we can simply modify the flowi_mark that is
generated as a result of the xfrm_decode_session.  By doing this we don't
need to actually touch the skb->mark and it can be preserved as it passes
out through the tunnel.
Signed-off-by: NAlexander Duyck <alexander.h.duyck@redhat.com>
Signed-off-by: NSteffen Klassert <steffen.klassert@secunet.com>

After commit 07f4c900 ("tcp/dccp: try to not exhaust
ip_local_port_range in connect()") it is advised to have an even number
of ports described in /proc/sys/net/ipv4/ip_local_port_range

This means start/end values should have a different parity.

Let's warn sysadmins of this, so that they can update their settings
if they want to.
Suggested-by: NDavid S. Miller <davem@davemloft.net>
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

__inet_hash_connect() does not use its third argument (port_offset)
if socket was already bound to a source port.

No need to perform useless but expensive md5 computations.
Reported-by: NCrestez Dan Leonard <cdleonard@gmail.com>
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

A long standing problem on busy servers is the tiny available TCP port
range (/proc/sys/net/ipv4/ip_local_port_range) and the default
sequential allocation of source ports in connect() system call.

If a host is having a lot of active TCP sessions, chances are
very high that all ports are in use by at least one flow,
and subsequent bind(0) attempts fail, or have to scan a big portion of
space to find a slot.

In this patch, I changed the starting point in __inet_hash_connect()
so that we try to favor even [1] ports, leaving odd ports for bind()
users.

We still perform a sequential search, so there is no guarantee, but
if connect() targets are very different, end result is we leave
more ports available to bind(), and we spread them all over the range,
lowering time for both connect() and bind() to find a slot.

This strategy only works well if /proc/sys/net/ipv4/ip_local_port_range
is even, ie if start/end values have different parity.

Therefore, default /proc/sys/net/ipv4/ip_local_port_range was changed to
32768 - 60999 (instead of 32768 - 61000)

There is no change on security aspects here, only some poor hashing
schemes could be eventually impacted by this change.

[1] : The odd/even property depends on ip_local_port_range values parity
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

We currently always send fragments without DF bit set.

Thus, given following setup:

mtu1500 - mtu1500:1400 - mtu1400:1280 - mtu1280
   A           R1              R2         B

Where R1 and R2 run linux with netfilter defragmentation/conntrack
enabled, then if Host A sent a fragmented packet _with_ DF set to B, R1
will respond with icmp too big error if one of these fragments exceeded
1400 bytes.

However, if R1 receives fragment sizes 1200 and 100, it would
forward the reassembled packet without refragmenting, i.e.
R2 will send an icmp error in response to a packet that was never sent,
citing mtu that the original sender never exceeded.

The other minor issue is that a refragmentation on R1 will conceal the
MTU of R2-B since refragmentation does not set DF bit on the fragments.

This modifies ip_fragment so that we track largest fragment size seen
both for DF and non-DF packets, and set frag_max_size to the largest
value.

If the DF fragment size is larger or equal to the non-df one, we will
consider the packet a path mtu probe:
We set DF bit on the reassembled skb and also tag it with a new IPCB flag
to force refragmentation even if skb fits outdev mtu.

We will also set DF bit on each fragment in this case.

Joint work with Hannes Frederic Sowa.
Reported-by: NJesse Gross <jesse@nicira.com>
Signed-off-by: NFlorian Westphal <fw@strlen.de>
Acked-by: NHannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

ip_skb_dst_mtu is small inline helper, but its called in several places.

before: 17061      44       0   17105    42d1 net/ipv4/ip_output.o
after:  16805      44       0   16849    41d1 net/ipv4/ip_output.o
Signed-off-by: NFlorian Westphal <fw@strlen.de>
Acked-by: NHannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

We used to get this indirectly I supposed, but no longer do.

Either way, an explicit include should have been done in the
first place.

   net/ipv4/fib_trie.c: In function '__node_free_rcu':
>> net/ipv4/fib_trie.c:293:3: error: implicit declaration of function 'vfree' [-Werror=implicit-function-declaration]
      vfree(n);
      ^
   net/ipv4/fib_trie.c: In function 'tnode_alloc':
>> net/ipv4/fib_trie.c:312:3: error: implicit declaration of function 'vzalloc' [-Werror=implicit-function-declaration]
      return vzalloc(size);
      ^
>> net/ipv4/fib_trie.c:312:3: warning: return makes pointer from integer without a cast
   cc1: some warnings being treated as errors
Reported-by: Nkbuild test robot <fengguang.wu@intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

By making sure sk->sk_gso_max_segs minimal value is one,
and sysctl_tcp_min_tso_segs minimal value is one as well,
tcp_tso_autosize() will return a non zero value.

We can then revert 843925f3
("tcp: Do not apply TSO segment limit to non-TSO packets")
and save few cpu cycles in fast path.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Acked-by: NNeal Cardwell <ncardwell@google.com>
Acked-by: NHerbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

If tcp ehash table is constrained to a very small number of buckets
(eg boot parameter thash_entries=128), then we can crash if spinlock
array has more entries.

While we are at it, un-inline inet_ehash_locks_alloc() and make
following changes :

- Budget 2 cache lines per cpu worth of 'spinlocks'
- Try to kmalloc() the array to avoid extra TLB pressure.
  (Most servers at Google allocate 8192 bytes for this hash table)
- Get rid of various #ifdef
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NFlorian Westphal <fw@strlen.de>
Signed-off-by: NPablo Neira Ayuso <pablo@netfilter.org>