commit c6cffba4 (ipv4: Fix input route performance regression.)
added various fatal races with dst refcounts.

crashes happen on tcp workloads if routes are added/deleted at the same
time.

The dst_free() calls from free_fib_info_rcu() are clearly racy.

We need instead regular dst refcounting (dst_release()) and make
sure dst_release() is aware of RCU grace periods :

Add DST_RCU_FREE flag so that dst_release() respects an RCU grace period
before dst destruction for cached dst

Introduce a new inet_sk_rx_dst_set() helper, using atomic_inc_not_zero()
to make sure we dont increase a zero refcount (On a dst currently
waiting an rcu grace period before destruction)

rt_cache_route() must take a reference on the new cached route, and
release it if was not able to install it.

With this patch, my machines survive various benchmarks.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

commit 92101b3b (ipv4: Prepare for change of rt->rt_iif encoding.)
invalidated TCP early demux, because rx_dst_ifindex is not properly
initialized and checked.

Also remove the use of inet_iif(skb) in favor or skb->skb_iif
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

1) Remove a non needed pskb_may_pull() in tcp_v4_early_demux()
   and fix a potential bug if skb->head was reallocated
   (iph & th pointers were not reloaded)

TCP stack will pull/check headers anyway.

2) must reload iph in ip_rcv_finish() after early_demux()
 call since skb->head might have changed.

3) skb->dev->ifindex can be now replaced by skb->skb_iif
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Use inet_iif() consistently, and for TCP record the input interface of
cached RX dst in inet sock.

rt->rt_iif is going to be encoded differently, so that we can
legitimately cache input routes in the FIB info more aggressively.

When the input interface is "use SKB device index" the rt->rt_iif will
be set to zero.

This forces us to move the TCP RX dst cache installation into the ipv4
specific code, and as well it should since doing the route caching for
ipv6 is pointless at the moment since it is not inspected in the ipv6
input paths yet.

Also, remove the unlikely on dst->obsolete, all ipv4 dsts have
obsolete set to a non-zero value to force invocation of the check
callback.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

ICMP messages generated in output path if frame length is bigger than
mtu are actually lost because socket is owned by user (doing the xmit)

One example is the ipgre_tunnel_xmit() calling
icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));

We had a similar case fixed in commit a34a101e (ipv6: disable GSO on
sockets hitting dst_allfrag).

Problem of such fix is that it relied on retransmit timers, so short tcp
sessions paid a too big latency increase price.

This patch uses the tcp_release_cb() infrastructure so that MTU
reduction messages (ICMP messages) are not lost, and no extra delay
is added in TCP transmits.
Reported-by: NMaciej Żenczykowski <maze@google.com>
Diagnosed-by: NNeal Cardwell <ncardwell@google.com>
Signed-off-by: NEric Dumazet <edumazet@google.com>
Cc: Nandita Dukkipati <nanditad@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Tore Anderson <tore@fud.no>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

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

sendmsg() (or sendto()) with MSG_FASTOPEN is a combo of connect(2)
and write(2). The application should replace connect() with it to
send data in the opening SYN packet.

For blocking socket, sendmsg() blocks until all the data are buffered
locally and the handshake is completed like connect() call. It
returns similar errno like connect() if the TCP handshake fails.

For non-blocking socket, it returns the number of bytes queued (and
transmitted in the SYN-data packet) if cookie is available. If cookie
is not available, it transmits a data-less SYN packet with Fast Open
cookie request option and returns -EINPROGRESS like connect().

Using MSG_FASTOPEN on connecting or connected socket will result in
simlar errno like repeating connect() calls. Therefore the application
should only use this flag on new sockets.

The buffer size of sendmsg() is independent of the MSS of the connection.
Signed-off-by: NYuchung Cheng <ycheng@google.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch impelements the common code for both the client and server.

1. TCP Fast Open option processing. Since Fast Open does not have an
   option number assigned by IANA yet, it shares the experiment option
   code 254 by implementing draft-ietf-tcpm-experimental-options
   with a 16 bits magic number 0xF989. This enables global experiments
   without clashing the scarce(2) experimental options available for TCP.

   When the draft status becomes standard (maybe), the client should
   switch to the new option number assigned while the server supports
   both numbers for transistion.

2. The new sysctl tcp_fastopen

3. A place holder init function
Signed-off-by: NYuchung Cheng <ycheng@google.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

tcp_v4_send_reset() and tcp_v4_send_ack() use a single socket
per network namespace.

This leads to bad behavior on multiqueue NICS, because many cpus
contend for the socket lock and once socket lock is acquired, extra
false sharing on various socket fields slow down the operations.

To better resist to attacks, we use a percpu socket. Each cpu can
run without contention, using appropriate memory (local node)

Additional features :

1) We also mirror the queue_mapping of the incoming skb, so that
answers use the same queue if possible.

2) Setting SOCK_USE_WRITE_QUEUE socket flag speedup sock_wfree()

3) We now limit the number of in-flight RST/ACK [1] packets
per cpu, instead of per namespace, and we honor the sysctl_wmem_default
limit dynamically. (Prior to this patch, sysctl_wmem_default value was
copied at boot time, so any further change would not affect tcp_sock
limit)

[1] These packets are only generated when no socket was matched for
the incoming packet.
Reported-by: NBill Sommerfeld <wsommerfeld@google.com>
Signed-off-by: NEric Dumazet <edumazet@google.com>
Cc: Tom Herbert <therbert@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This will be used so that we can compose a full flow key.

Even though we have a route in this context, we need more.  In the
future the routes will be without destination address, source address,
etc. keying.  One ipv4 route will cover entire subnets, etc.

In this environment we have to have a way to possess persistent storage
for redirects and PMTU information.  This persistent storage will exist
in the FIB tables, and that's why we'll need to be able to rebuild a
full lookup flow key here.  Using that flow key will do a fib_lookup()
and create/update the persistent entry.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This abstracts away the call to dst_ops->update_pmtu() so that we can
transparently handle the fact that, in the future, the dst itself can
be invalidated by the PMTU update (when we have non-host routes cached
in sockets).

So we try to rebuild the socket cached route after the method
invocation if necessary.

This isn't used by SCTP because it needs to cache dsts per-transport,
and thus will need it's own local version of this helper.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

No longer necessary.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

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

This introduce TSQ (TCP Small Queues)

TSQ goal is to reduce number of TCP packets in xmit queues (qdisc &
device queues), to reduce RTT and cwnd bias, part of the bufferbloat
problem.

sk->sk_wmem_alloc not allowed to grow above a given limit,
allowing no more than ~128KB [1] per tcp socket in qdisc/dev layers at a
given time.

TSO packets are sized/capped to half the limit, so that we have two
TSO packets in flight, allowing better bandwidth use.

As a side effect, setting the limit to 40000 automatically reduces the
standard gso max limit (65536) to 40000/2 : It can help to reduce
latencies of high prio packets, having smaller TSO packets.

This means we divert sock_wfree() to a tcp_wfree() handler, to
queue/send following frames when skb_orphan() [2] is called for the
already queued skbs.

Results on my dev machines (tg3/ixgbe nics) are really impressive,
using standard pfifo_fast, and with or without TSO/GSO.

Without reduction of nominal bandwidth, we have reduction of buffering
per bulk sender :
< 1ms on Gbit (instead of 50ms with TSO)
< 8ms on 100Mbit (instead of 132 ms)

I no longer have 4 MBytes backlogged in qdisc by a single netperf
session, and both side socket autotuning no longer use 4 Mbytes.

As skb destructor cannot restart xmit itself ( as qdisc lock might be
taken at this point ), we delegate the work to a tasklet. We use one
tasklest per cpu for performance reasons.

If tasklet finds a socket owned by the user, it sets TSQ_OWNED flag.
This flag is tested in a new protocol method called from release_sock(),
to eventually send new segments.

[1] New /proc/sys/net/ipv4/tcp_limit_output_bytes tunable
[2] skb_orphan() is usually called at TX completion time,
  but some drivers call it in their start_xmit() handler.
  These drivers should at least use BQL, or else a single TCP
  session can still fill the whole NIC TX ring, since TSQ will
  have no effect.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Cc: Dave Taht <dave.taht@bufferbloat.net>
Cc: Tom Herbert <therbert@google.com>
Cc: Matt Mathis <mattmathis@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Nandita Dukkipati <nanditad@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

No longer used.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

With help from Lin Ming.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

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

Rename it to ip_send_unicast_reply() and add explicit 'saddr'
argument.

This removed one of the few users of rt->rt_spec_dst.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

It's completely unnecessary.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This reverts commit c074da28.

This change has several unwanted side effects:

1) Sockets will cache the DST_NOCACHE route in sk->sk_rx_dst and we'll
   thus never create a real cached route.

2) All TCP traffic will use DST_NOCACHE and never use the routing
   cache at all.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

DDOS synflood attacks hit badly IP route cache.

On typical machines, this cache is allowed to hold up to 8 Millions dst
entries, 256 bytes for each, for a total of 2GB of memory.

rt_garbage_collect() triggers and tries to cleanup things.

Eventually route cache is disabled but machine is under fire and might
OOM and crash.

This patch exploits the new TCP early demux, to set a nocache
boolean in case incoming TCP frame is for a not yet ESTABLISHED or
TIMEWAIT socket.

This 'nocache' boolean is then used in case dst entry is not found in
route cache, to create an unhashed dst entry (DST_NOCACHE)

SYN-cookie-ACK sent use a similar mechanism (ipv4: tcp: dont cache
output dst for syncookies), so after this patch, a machine is able to
absorb a DDOS synflood attack without polluting its IP route cache.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Cc: Hans Schillstrom <hans.schillstrom@ericsson.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The dest port for the call to __inet_lookup_established() in TCP early demux
code is passed with the wrong endian-ness. This causes the lookup to fail
leading to early demux not being used.
Signed-off-by: NVijay Subramanian <subramanian.vijay@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Don't cache output dst for syncookies, as this adds pressure on IP route
cache and rcu subsystem for no gain.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Cc: Hans Schillstrom <hans.schillstrom@ericsson.com>
Signed-off-by: NJesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Otherwise we might violate reverse path filtering.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Input packet processing for local sockets involves two major demuxes.
One for the route and one for the socket.

But we can optimize this down to one demux for certain kinds of local
sockets.

Currently we only do this for established TCP sockets, but it could
at least in theory be expanded to other kinds of connections.

If a TCP socket is established then it's identity is fully specified.

This means that whatever input route was used during the three-way
handshake must work equally well for the rest of the connection since
the keys will not change.

Once we move to established state, we cache the receive packet's input
route to use later.

Like the existing cached route in sk->sk_dst_cache used for output
packets, we have to check for route invalidations using dst->obsolete
and dst->ops->check().

Early demux occurs outside of a socket locked section, so when a route
invalidation occurs we defer the fixup of sk->sk_rx_dst until we are
actually inside of established state packet processing and thus have
the socket locked.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Since it's guarenteed that we will access the inetpeer if we're trying
to do timewait recycling and TCP options were enabled on the
connection, just cache the peer in the timewait socket.

In the future, inetpeer lookups will be context dependent (per routing
realm), and this helps facilitate that as well.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The get_peer method TCP uses is full of special cases that make no
sense accommodating, and it also gets in the way of doing more
reasonable things here.

First of all, if the socket doesn't have a usable cached route, there
is no sense in trying to optimize timewait recycling.

Likewise for the case where we have IP options, such as SRR enabled,
that make the IP header destination address (and thus the destination
address of the route key) differ from that of the connection's
destination address.

Just return a NULL peer in these cases, and thus we're also able to
get rid of the clumsy inetpeer release logic.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

There's a lot of places that open-code rt{,6}_get_peer() only because
they want to set 'create' to one.  So add an rt{,6}_get_peer_create()
for their sake.

There were also a few spots open-coding plain rt{,6}_get_peer() and
those are transformed here as well.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

add struct net as a parameter of inet_getpeer_v[4,6],
use net to replace &init_net.

and modify some places to provide net for inet_getpeer_v[4,6]
Signed-off-by: NGao feng <gaofeng@cn.fujitsu.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

tcp_make_synack() clones the dst, and callers release it.

We can avoid two atomic operations per SYNACK if tcp_make_synack()
consumes dst instead of cloning it.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

While testing how linux behaves on SYNFLOOD attack on multiqueue device
(ixgbe), I found that SYNACK messages were dropped at Qdisc level
because we send them all on a single queue.

Obvious choice is to reflect incoming SYN packet @queue_mapping to
SYNACK packet.

Under stress, my machine could only send 25.000 SYNACK per second (for
200.000 incoming SYN per second). NIC : ixgbe with 16 rx/tx queues.

After patch, not a single SYNACK is dropped.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Cc: Hans Schillstrom <hans.schillstrom@ericsson.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Tom Herbert <therbert@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

bool conversions where possible.

__inline__ -> inline

space cleanups
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Standardize the net core ratelimited logging functions.

Coalesce formats, align arguments.
Change a printk then vprintk sequence to use printf extension %pV.
Signed-off-by: NJoe Perches <joe@perches.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

It appears some networks play bad games with the two bits reserved for
ECN. This can trigger false congestion notifications and very slow
transferts.

Since RFC 3168 (6.1.1) forbids SYN packets to carry CT bits, we can
disable TCP ECN negociation if it happens we receive mangled CT bits in
the SYN packet.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Cc: Perry Lorier <perryl@google.com>
Cc: Matt Mathis <mattmathis@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Wilmer van der Gaast <wilmer@google.com>
Cc: Ankur Jain <jankur@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Dave Täht <dave.taht@bufferbloat.net>
Acked-by: NNeal Cardwell <ncardwell@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

While investigating TCP performance problems on 10Gb+ links, we found a
tcp sender was dropping lot of incoming ACKS because of sk_rcvbuf limit
in sk_add_backlog(), especially if receiver doesnt use GRO/LRO and sends
one ACK every two MSS segments.

A sender usually tweaks sk_sndbuf, but sk_rcvbuf stays at its default
value (87380), allowing a too small backlog.

A TCP ACK, even being small, can consume nearly same truesize space than
outgoing packets. Using sk_rcvbuf + sk_sndbuf as a limit makes sense and
is fast to compute.

Performance results on netperf, single flow, receiver with disabled
GRO/LRO : 7500 Mbits instead of 6050 Mbits, no more TCPBacklogDrop
increments at sender.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Maciej Żenczykowski <maze@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Cc: Rick Jones <rick.jones2@hp.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

sk_add_backlog() & sk_rcvqueues_full() hard coded sk_rcvbuf as the
memory limit. We need to make this limit a parameter for TCP use.

No functional change expected in this patch, all callers still using the
old sk_rcvbuf limit.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Maciej Żenczykowski <maze@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Cc: Rick Jones <rick.jones2@hp.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

net/ipv4/tcp_ipv4.c: In function 'tcp_v4_init_sock':
net/ipv4/tcp_ipv4.c:1891:19: warning: unused variable 'tp' [-Wunused-variable]
net/ipv6/tcp_ipv6.c: In function 'tcp_v6_init_sock':
net/ipv6/tcp_ipv6.c:1836:19: warning: unused variable 'tp' [-Wunused-variable]
Reported-by: NStephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This commit moves the (substantial) common code shared between
tcp_v4_init_sock() and tcp_v6_init_sock() to a new address-family
independent function, tcp_init_sock().

Centralizing this functionality should help avoid drift issues,
e.g. where the IPv4 side is updated without a corresponding update to
IPv6. There was already some drift: IPv4 initialized snd_cwnd to
TCP_INIT_CWND, while the IPv6 side was still initializing snd_cwnd to
2 (in this case it should not matter, since snd_cwnd is also
initialized in tcp_init_metrics(), but the general risks and
maintenance overhead remain).

When diffing the old and new code, note that new tcp_init_sock()
function uses the order of steps from the tcp_v4_init_sock()
implementation (the order is slightly different in
tcp_v6_init_sock()).
Signed-off-by: NNeal Cardwell <ncardwell@google.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This includes (according the the previous description):

* TCP_REPAIR sockoption

This one just puts the socket in/out of the repair mode.
Allowed for CAP_NET_ADMIN and for closed/establised sockets only.
When repair mode is turned off and the socket happens to be in
the established state the window probe is sent to the peer to
'unlock' the connection.

* TCP_REPAIR_QUEUE sockoption

This one sets the queue which we're about to repair. The
'no-queue' is set by default.

* TCP_QUEUE_SEQ socoption

Sets the write_seq/rcv_nxt of a selected repaired queue.
Allowed for TCP_CLOSE-d sockets only. When the socket changes
its state the other seq-s are changed by the kernel according
to the protocol rules (most of the existing code is actually
reused).

* Ability to forcibly bind a socket to a port

The sk->sk_reuse is set to SK_FORCE_REUSE.

* Immediate connect modification

The connect syscall initializes the connection, then directly jumps
to the code which finalizes it.

* Silent close modification

The close just aborts the connection (similar to SO_LINGER with 0
time) but without sending any FIN/RST-s to peer.
Signed-off-by: NPavel Emelyanov <xemul@parallels.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This is the fallout from adding memcpy alignment workaround for certain
IOATDMA hardware. NetDMA will only use DMA engine that can handle byte align
ops.
Acked-by: NDavid S. Miller <davem@davemloft.net>
Signed-off-by: NDave Jiang <dave.jiang@intel.com>
Signed-off-by: NDan Williams <dan.j.williams@intel.com>