When a SYNFLOOD targets a non SO_REUSEPORT listener, multiple
cpus contend on sk->sk_refcnt and sk->sk_wmem_alloc changes.

By letting listeners use SOCK_RCU_FREE infrastructure,
we can relax TCP_LISTEN lookup rules and avoid touching sk_refcnt

Note that we still use SLAB_DESTROY_BY_RCU rules for other sockets,
only listeners are impacted by this change.

Peak performance under SYNFLOOD is increased by ~33% :

On my test machine, I could process 3.2 Mpps instead of 2.4 Mpps

Most consuming functions are now skb_set_owner_w() and sock_wfree()
contending on sk->sk_wmem_alloc when cooking SYNACK and freeing them.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

RX packet processing holds rcu_read_lock(), so we can remove
pairs of rcu_read_lock()/rcu_read_unlock() in lookup functions
if inet_diag also holds rcu before calling them.

This is needed anyway as __inet_lookup_listener() and
inet6_lookup_listener() will soon no longer increment
refcount on the found listener.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

In commit 07f4c900 ("tcp/dccp: try to not exhaust ip_local_port_range
in connect()"), I added a very simple heuristic, so that we got better
chances to use even ports, and allow bind() users to have more available
slots.

It gave nice results, but with more than 200,000 TCP sessions on a typical
server, the ~30,000 ephemeral ports are still a rare resource.

I chose to go a step further, by looking at all even ports, and if none
was available, fallback to odd ports.

The companion patch does the same in bind(), but in opposite way.

I've seen exec times of up to 30ms on busy servers, so I no longer
disable BH for the whole traversal, but only for each hash bucket.
I also call cond_resched() to be gentle to other tasks.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This change extends the fast SO_REUSEPORT socket lookup implemented
for UDP to TCP.  Listener sockets with SO_REUSEPORT and the same
receive address are additionally added to an array for faster
random access.  This means that only a single socket from the group
must be found in the listener list before any socket in the group can
be used to receive a packet.  Previously, every socket in the group
needed to be considered before handing off the incoming packet.

This feature also exposes the ability to use a BPF program when
selecting a socket from a reuseport group.
Signed-off-by: NCraig Gallek <kraig@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This is a preliminary step to allow fast socket lookup of SO_REUSEPORT
groups.  Doing so with a BPF filter will require access to the
skb in question.  This change plumbs the skb (and offset to payload
data) through the call stack to the listening socket lookup
implementations where it will be used in a following patch.
Signed-off-by: NCraig Gallek <kraig@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

In order to support fast reuseport lookups in TCP, the hash function
defined in struct proto must be capable of returning an error code.
This patch changes the function signature of all related hash functions
to return an integer and handles or propagates this return value at
all call sites.
Signed-off-by: NCraig Gallek <kraig@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Multiple cpus can process duplicates of incoming ACK messages
matching a SYN_RECV request socket. This is a rare event under
normal operations, but definitely can happen.

Only one must win the race, otherwise corruption would occur.

To fix this without adding new atomic ops, we use logic in
inet_ehash_nolisten() to detect the request was present in the same
ehash bucket where we try to insert the new child.

If request socket was not found, we have to undo the child creation.

This actually removes a spin_lock()/spin_unlock() pair in
reqsk_queue_unlink() for the fast path.

Fixes: e994b2f0 ("tcp: do not lock listener to process SYN packets")
Fixes: 079096f1 ("tcp/dccp: install syn_recv requests into ehash table")
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

As we no longer hold listener lock in fast path, it is possible that a
child is created right after listener freed its bound port, if a close()
is done while incoming packets are processed.

__inet_inherit_port() must detect this and return an error,
so that caller can free the child earlier.

Fixes: e994b2f0 ("tcp: do not lock listener to process SYN packets")
Fixes: 079096f1 ("tcp/dccp: install syn_recv requests into ehash table")
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

SO_INCOMING_CPU as added in commit 2c8c56e1 was a getsockopt() command
to fetch incoming cpu handling a particular TCP flow after accept()

This commits adds setsockopt() support and extends SO_REUSEPORT selection
logic : If a TCP listener or UDP socket has this option set, a packet is
delivered to this socket only if CPU handling the packet matches the specified
one.

This allows to build very efficient TCP servers, using one listener per
RX queue, as the associated TCP listener should only accept flows handled
in softirq by the same cpu.
This provides optimal NUMA behavior and keep cpu caches hot.

Note that __inet_lookup_listener() still has to iterate over the list of
all listeners. Following patch puts sk_refcnt in a different cache line
to let this iteration hit only shared and read mostly cache lines.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

In this patch, we insert request sockets into TCP/DCCP
regular ehash table (where ESTABLISHED and TIMEWAIT sockets
are) instead of using the per listener hash table.

ACK packets find SYN_RECV pseudo sockets without having
to find and lock the listener.

In nominal conditions, this halves pressure on listener lock.

Note that this will allow for SO_REUSEPORT refinements,
so that we can select a listener using cpu/numa affinities instead
of the prior 'consistent hash', since only SYN packets will
apply this selection logic.

We will shrink listen_sock in the following patch to ease
code review.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Cc: Ying Cai <ycai@google.com>
Cc: Willem de Bruijn <willemb@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

socket is not touched, make it const.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Andrew Morton reported following warning on one ARM build
with gcc-4.4 :

net/ipv4/inet_hashtables.c: In function 'inet_ehash_locks_alloc':
net/ipv4/inet_hashtables.c:617: warning: division by zero

Even guarded with a test on sizeof(spinlock_t), compiler does not
like current construct on a !CONFIG_SMP build.

Remove the warning by using a temporary variable.

Fixes: 095dc8e0 ("tcp: fix/cleanup inet_ehash_locks_alloc()")
Reported-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

inet_twsk_deschedule() calls are followed by inet_twsk_put().

Only particular case is in inet_twsk_purge() but there is no point
to defer the inet_twsk_put() after re-enabling BH.

Lets rename inet_twsk_deschedule() to inet_twsk_deschedule_put()
and move the inet_twsk_put() inside.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

timewait sockets have a complex refcounting logic.
Once we realize it should be similar to established and
syn_recv sockets, we can use sk_nulls_del_node_init_rcu()
and remove inet_twsk_unhash()

In particular, deferred inet_twsk_put() added in commit
13475a30 ("tcp: connect() race with timewait reuse")
looks unecessary : When removing a timewait socket from
ehash or bhash, caller must own a reference on the socket
anyway.
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>

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>

We no longer need bsocket atomic counter, as inet_csk_get_port()
calls bind_conflict() regardless of its value, after commit
2b05ad33 ("tcp: bind() fix autoselection to share ports")

This patch removes overhead of maintaining this counter and
double inet_csk_get_port() calls under pressure.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Cc: Marcelo Ricardo Leitner <mleitner@redhat.com>
Cc: Flavio Leitner <fbl@redhat.com>
Acked-by: NFlavio Leitner <fbl@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Using a timer wheel for timewait sockets was nice ~15 years ago when
memory was expensive and machines had a single processor.

This does not scale, code is ugly and source of huge latencies
(Typically 30 ms have been seen, cpus spinning on death_lock spinlock.)

We can afford to use an extra 64 bytes per timewait sock and spread
timewait load to all cpus to have better behavior.

Tested:

On following test, /proc/sys/net/ipv4/tcp_tw_recycle is set to 1
on the target (lpaa24)

Before patch :

lpaa23:~# ./super_netperf 200 -H lpaa24 -t TCP_CC -l 60 -- -p0,0
419594

lpaa23:~# ./super_netperf 200 -H lpaa24 -t TCP_CC -l 60 -- -p0,0
437171

While test is running, we can observe 25 or even 33 ms latencies.

lpaa24:~# ping -c 1000 -i 0.02 -qn lpaa23
...
1000 packets transmitted, 1000 received, 0% packet loss, time 20601ms
rtt min/avg/max/mdev = 0.020/0.217/25.771/1.535 ms, pipe 2

lpaa24:~# ping -c 1000 -i 0.02 -qn lpaa23
...
1000 packets transmitted, 1000 received, 0% packet loss, time 20702ms
rtt min/avg/max/mdev = 0.019/0.183/33.761/1.441 ms, pipe 2

After patch :

About 90% increase of throughput :

lpaa23:~# ./super_netperf 200 -H lpaa24 -t TCP_CC -l 60 -- -p0,0
810442

lpaa23:~# ./super_netperf 200 -H lpaa24 -t TCP_CC -l 60 -- -p0,0
800992

And latencies are kept to minimal values during this load, even
if network utilization is 90% higher :

lpaa24:~# ping -c 1000 -i 0.02 -qn lpaa23
...
1000 packets transmitted, 1000 received, 0% packet loss, time 19991ms
rtt min/avg/max/mdev = 0.023/0.064/0.360/0.042 ms
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The ipv4 code uses a mixture of coding styles. In some instances check
for non-NULL pointer is done as x != NULL and sometimes as x. x is
preferred according to checkpatch and this patch makes the code
consistent by adopting the latter form.

No changes detected by objdiff.
Signed-off-by: NIan Morris <ipm@chirality.org.uk>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

We now always call __inet_hash_nolisten(), no need to pass it
as an argument.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

We can now use inet_hash() and __inet_hash() instead of private
functions.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Intent is to converge IPv4 & IPv6 inet_hash functions to
factorize code.

IPv4 sockets initialize sk_rcv_saddr and sk_v6_daddr
in this patch, thanks to new sk_daddr_set() and sk_rcv_saddr_set()
helpers.

__inet6_hash can now use sk_ehashfn() instead of a private
inet6_sk_ehashfn() and will simply use __inet_hash() in a
following patch.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Goal is to unify IPv4/IPv6 inet_hash handling, and use common helpers
for all kind of sockets (full sockets, timewait and request sockets)

inet_sk_ehashfn() becomes sk_ehashfn() but still only copes with IPv4
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

const qualifiers ease code review by making clear
which objects are not written in a function.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

sock_edemux() is not used in fast path, and should
really call sock_gen_put() to save some code.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

sock_edemux() & sock_gen_put() should be ready to cope with request socks.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

hold_net and release_net were an idea that turned out to be useless.
The code has been disabled since 2008.  Kill the code it is long past due.
Signed-off-by: N"Eric W. Biederman" <ebiederm@xmission.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Replace open codings of (((u64) <x> * <y>) >> 32) with reciprocal_scale().
Signed-off-by: NDaniel Borkmann <dborkman@redhat.com>
Cc: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Missing a colon on definition use is a bit odd so
change the macro for the 32 bit case to declare an
__attribute__((unused)) and __deprecated variable.

The __deprecated attribute will cause gcc to emit
an error if the variable is actually used.
Signed-off-by: NJoe Perches <joe@perches.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

ip_local_port_range is already per netns, so should ip_local_reserved_ports
be. And since it is none by default we don't actually need it when we don't
enable CONFIG_SYSCTL.

By the way, rename inet_is_reserved_local_port() to inet_is_local_reserved_port()

Cc: "David S. Miller" <davem@davemloft.net>
Signed-off-by: NCong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Initialize the ehash and ipv6_hash_secrets with net_get_random_once.

Each compilation unit gets its own secret now:
  ipv4/inet_hashtables.o
  ipv4/udp.o
  ipv6/inet6_hashtables.o
  ipv6/udp.o
  rds/connection.o

The functions still get inlined into the hashing functions. In the fast
path we have at most two (needed in ipv6) if (unlikely(...)).

Cc: Eric Dumazet <edumazet@google.com>
Cc: "David S. Miller" <davem@davemloft.net>
Signed-off-by: NHannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This duplicates a bit of code but let's us easily introduce
separate secret keys later. The separate compilation units are
ipv4/inet_hashtabbles.o, ipv4/udp.o and rds/connection.o.

Cc: Eric Dumazet <edumazet@google.com>
Cc: "David S. Miller" <davem@davemloft.net>
Signed-off-by: NHannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

TCP listener refactoring, part 3 :

Our goal is to hash SYN_RECV sockets into main ehash for fast lookup,
and parallel SYN processing.

Current inet_ehash_bucket contains two chains, one for ESTABLISH (and
friend states) sockets, another for TIME_WAIT sockets only.

As the hash table is sized to get at most one socket per bucket, it
makes little sense to have separate twchain, as it makes the lookup
slightly more complicated, and doubles hash table memory usage.

If we make sure all socket types have the lookup keys at the same
offsets, we can use a generic and faster lookup. It turns out TIME_WAIT
and ESTABLISHED sockets already have common lookup fields for IPv4.

[ INET_TW_MATCH() is no longer needed ]

I'll provide a follow-up to factorize IPv6 lookup as well, to remove
INET6_TW_MATCH()

This way, SYN_RECV pseudo sockets will be supported the same.

A new sock_gen_put() helper is added, doing either a sock_put() or
inet_twsk_put() [ and will support SYN_RECV later ].

Note this helper should only be called in real slow path, when rcu
lookup found a socket that was moved to another identity (freed/reused
immediately), but could eventually be used in other contexts, like
sock_edemux()

Before patch :

dmesg | grep "TCP established"

TCP established hash table entries: 524288 (order: 11, 8388608 bytes)

After patch :

TCP established hash table entries: 524288 (order: 10, 4194304 bytes)
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

commit 3ab5aee7 ("net: Convert TCP & DCCP hash tables to use RCU /
hlist_nulls") incorrectly used sock_put() on TIMEWAIT sockets.

We should instead use inet_twsk_put()
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

- Move sysctl_local_ports from a global variable into struct netns_ipv4.
- Modify inet_get_local_port_range to take a struct net, and update all
  of the callers.
- Move the initialization of sysctl_local_ports into
   sysctl_net_ipv4.c:ipv4_sysctl_init_net from inet_connection_sock.c

v2:
- Ensure indentation used tabs
- Fixed ip.h so it applies cleanly to todays net-next

v3:
- Compile fixes of strange callers of inet_get_local_port_range.
  This patch now successfully passes an allmodconfig build.
  Removed manual inlining of inet_get_local_port_range in ipv4_local_port_range
Originally-by: NSamya <samya@twitter.com>
Acked-by: NNicolas Dichtel <nicolas.dichtel@6wind.com>
Signed-off-by: N"Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Found using checkpatch.pl
Signed-off-by: NCamelia Groza <camelia.groza@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

I'm not sure why, but the hlist for each entry iterators were conceived

        list_for_each_entry(pos, head, member)

The hlist ones were greedy and wanted an extra parameter:

        hlist_for_each_entry(tpos, pos, head, member)

Why did they need an extra pos parameter? I'm not quite sure. Not only
they don't really need it, it also prevents the iterator from looking
exactly like the list iterator, which is unfortunate.

Besides the semantic patch, there was some manual work required:

 - Fix up the actual hlist iterators in linux/list.h
 - Fix up the declaration of other iterators based on the hlist ones.
 - A very small amount of places were using the 'node' parameter, this
 was modified to use 'obj->member' instead.
 - Coccinelle didn't handle the hlist_for_each_entry_safe iterator
 properly, so those had to be fixed up manually.

The semantic patch which is mostly the work of Peter Senna Tschudin is here:

@@
iterator name hlist_for_each_entry, hlist_for_each_entry_continue, hlist_for_each_entry_from, hlist_for_each_entry_rcu, hlist_for_each_entry_rcu_bh, hlist_for_each_entry_continue_rcu_bh, for_each_busy_worker, ax25_uid_for_each, ax25_for_each, inet_bind_bucket_for_each, sctp_for_each_hentry, sk_for_each, sk_for_each_rcu, sk_for_each_from, sk_for_each_safe, sk_for_each_bound, hlist_for_each_entry_safe, hlist_for_each_entry_continue_rcu, nr_neigh_for_each, nr_neigh_for_each_safe, nr_node_for_each, nr_node_for_each_safe, for_each_gfn_indirect_valid_sp, for_each_gfn_sp, for_each_host;

type T;
expression a,c,d,e;
identifier b;
statement S;
@@

-T b;
    <+... when != b
(
hlist_for_each_entry(a,
- b,
c, d) S
|
hlist_for_each_entry_continue(a,
- b,
c) S
|
hlist_for_each_entry_from(a,
- b,
c) S
|
hlist_for_each_entry_rcu(a,
- b,
c, d) S
|
hlist_for_each_entry_rcu_bh(a,
- b,
c, d) S
|
hlist_for_each_entry_continue_rcu_bh(a,
- b,
c) S
|
for_each_busy_worker(a, c,
- b,
d) S
|
ax25_uid_for_each(a,
- b,
c) S
|
ax25_for_each(a,
- b,
c) S
|
inet_bind_bucket_for_each(a,
- b,
c) S
|
sctp_for_each_hentry(a,
- b,
c) S
|
sk_for_each(a,
- b,
c) S
|
sk_for_each_rcu(a,
- b,
c) S
|
sk_for_each_from
-(a, b)
+(a)
S
+ sk_for_each_from(a) S
|
sk_for_each_safe(a,
- b,
c, d) S
|
sk_for_each_bound(a,
- b,
c) S
|
hlist_for_each_entry_safe(a,
- b,
c, d, e) S
|
hlist_for_each_entry_continue_rcu(a,
- b,
c) S
|
nr_neigh_for_each(a,
- b,
c) S
|
nr_neigh_for_each_safe(a,
- b,
c, d) S
|
nr_node_for_each(a,
- b,
c) S
|
nr_node_for_each_safe(a,
- b,
c, d) S
|
- for_each_gfn_sp(a, c, d, b) S
+ for_each_gfn_sp(a, c, d) S
|
- for_each_gfn_indirect_valid_sp(a, c, d, b) S
+ for_each_gfn_indirect_valid_sp(a, c, d) S
|
for_each_host(a,
- b,
c) S
|
for_each_host_safe(a,
- b,
c, d) S
|
for_each_mesh_entry(a,
- b,
c, d) S
)
    ...+>

[akpm@linux-foundation.org: drop bogus change from net/ipv4/raw.c]
[akpm@linux-foundation.org: drop bogus hunk from net/ipv6/raw.c]
[akpm@linux-foundation.org: checkpatch fixes]
[akpm@linux-foundation.org: fix warnings]
[akpm@linux-foudnation.org: redo intrusive kvm changes]
Tested-by: NPeter Senna Tschudin <peter.senna@gmail.com>
Acked-by: NPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: NSasha Levin <sasha.levin@oracle.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Gleb Natapov <gleb@redhat.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Allow multiple listener sockets to bind to the same port.

Motivation for soresuseport would be something like a web server
binding to port 80 running with multiple threads, where each thread
might have it's own listener socket.  This could be done as an
alternative to other models: 1) have one listener thread which
dispatches completed connections to workers. 2) accept on a single
listener socket from multiple threads.  In case #1 the listener thread
can easily become the bottleneck with high connection turn-over rate.
In case #2, the proportion of connections accepted per thread tends
to be uneven under high connection load (assuming simple event loop:
while (1) { accept(); process() }, wakeup does not promote fairness
among the sockets.  We have seen the  disproportion to be as high
as 3:1 ratio between thread accepting most connections and the one
accepting the fewest.  With so_reusport the distribution is
uniform.
Signed-off-by: NTom Herbert <therbert@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

commit 68835aba (net: optimize INET input path further)
moved some fields used for tcp/udp sockets lookup in the first cache
line of struct sock_common.

This patch moves inet_dport/inet_num as well, filling a 32bit hole
on 64 bit arches and reducing number of cache line misses in lookups.

Also change INET_MATCH()/INET_TW_MATCH() to perform the ports match
before addresses match, as this check is more discriminant.

Remove the hash check from MATCH() macros because we dont need to
re validate the hash value after taking a refcount on socket, and
use likely/unlikely compiler hints, as the sk_hash/hash check
makes the following conditional tests 100% predicted by cpu.

Introduce skc_addrpair/skc_portpair pair values to better
document the alignment requirements of the port/addr pairs
used in the various MATCH() macros, and remove some casts.

The namespace check can also be done at last.

This slightly improves TCP/UDP lookup times.

IP/TCP early demux needs inet->rx_dst_ifindex and
TCP needs inet->min_ttl, lets group them together in same cache line.

With help from Ben Hutchings & Joe Perches.

Idea of this patch came after Ling Ma proposal to move skc_hash
to the beginning of struct sock_common, and should allow him
to submit a final version of his patch. My tests show an improvement
doing so.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Cc: Joe Perches <joe@perches.com>
Cc: Ling Ma <ling.ma.program@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>