[LLC]: Use skb_reset_mac_header in llc_mac_hdr_init

skb_push updates and returns skb->data, so we can just call
skb_reset_mac_header after the call to skb_push.

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[LLC]: Use skb_reset_mac_header in llc_alloc_frame

skb->head is equal to skb->data after alloc_skb, so reset the mac header while
this is true, i.e. before skb_reserve.

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[SK_BUFF]: Introduce skb_reset_mac_header(skb)

For the common, open coded 'skb->mac.raw = skb->data' operation, so that we can
later turn skb->mac.raw into a offset, reducing the size of struct sk_buff in
64bit land while possibly keeping it as a pointer on 32bit.

This one touches just the most simple case, next will handle the slightly more
"complex" cases.

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[ETH]: Make eth_type_trans set skb->dev like the other *_type_trans

One less thing for drivers writers to worry about.

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[AOE]: Introduce aoe_hdr()

For consistency with other skb->mac.raw users.

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[QETH]: Use eth_hdr()

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[HIPPI/FDDI]: Make {hippi,fddi}_type_trans set skb->dev

Now all the _type_trans routines are consistent in this regard.

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TR]: Make tr_type_trans set skb->dev

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TR]: Use tr_hdr() were appropriate

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[SOCKET]: Export __sock_recv_timestamp

Kernel: arch/x86_64/boot/bzImage is ready  (#2)
  MODPOST 1816 modules
WARNING: "__sock_recv_timestamp" [net/sctp/sctp.ko] undefined!
WARNING: "__sock_recv_timestamp" [net/packet/af_packet.ko] undefined!
WARNING: "__sock_recv_timestamp" [net/key/af_key.ko] undefined!
WARNING: "__sock_recv_timestamp" [net/ipv6/ipv6.ko] undefined!
WARNING: "__sock_recv_timestamp" [net/atm/atm.ko] undefined!
make[2]: *** [__modpost] Error 1
make[1]: *** [modules] Error 2
make: *** [_all] Error 2

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET]: Adding SO_TIMESTAMPNS / SCM_TIMESTAMPNS support

Now that network timestamps use ktime_t infrastructure, we can add a new
SOL_SOCKET sockopt  SO_TIMESTAMPNS.

This command is similar to SO_TIMESTAMP, but permits transmission of
a 'timespec struct' instead of a 'timeval struct' control message.
(nanosecond resolution instead of microsecond)

Control message is labelled SCM_TIMESTAMPNS instead of SCM_TIMESTAMP

A socket cannot mix SO_TIMESTAMP and SO_TIMESTAMPNS : the two modes are
mutually exclusive.

sock_recv_timestamp() became too big to be fully inlined so I added a
__sock_recv_timestamp() helper function.

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
CC: linux-arch@vger.kernel.org
Signed-off-by: David S. Miller <davem@davemloft.net>

[UDP]: Use __skb_pull since we have checked it won't fail with pskb_may_pull

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET]: New sysctls should use __read_mostly tags

net_msg_warn should be placed in the read_mostly section, to avoid
performance problems on SMP

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[IPV6]: Ensure to truncate result and return full length for sticky options.

Bug noticed by Chris Wright <chrisw@sous-sol.org>.

Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[IPV6]: Return correct result for sticky options.

We returned incorrect result with IPV6_RTHDRDSTOPTS, IPV6_RTHDR and
IPV6_DSTOPTS.

Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[UDP]: deinline

A couple of functions are exported or used indirectly
so it is pointless to mark them as inline.

Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET]: deinline some functions

Several functions are marked inline or forced inline, but it
would be better to let the compiler decide.

Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP]: whitespace cleanup

Add whitespace around keywords.

Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[IPV4]: cleanup

Add whitespace around keywords.

Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[WIRELESS]: use ARRAY_SIZE()

Use ARRAY_SIZE() macro now.

Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET] core: whitespace cleanup

Fix whitespace around keywords. Fix indentation especially of switch
statements.

Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[UDP]: ipv6 style cleanup

Fix whitespace around keywords. Eliminate unnecessary ()'s on return
statements.

Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[UDP]: ipv4 whitespace cleanup

Fix whitespace around keywords.

Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET]: Replace CONFIG_NET_DEBUG with sysctl.

Covert network warning messages from a compile time to runtime choice.
Removes kernel config option and replaces it with new /proc/sys/net/core/warnings.

Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET]: Introduce SIOCGSTAMPNS ioctl to get timestamps with nanosec resolution

Now network timestamps use ktime_t infrastructure, we can add a new
ioctl() SIOCGSTAMPNS command to get timestamps in 'struct timespec'.
User programs can thus access to nanosecond resolution.

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
CC: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP/DCCP/RANDOM]: Remove unused exports.

This patch removes the following not or no longer used exports:
- drivers/char/random.c: secure_tcp_sequence_number
- net/dccp/options.c: sysctl_dccp_feat_sequence_window
- net/netlink/af_netlink.c: netlink_set_err

Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP]: Abstract out all write queue operations.

This allows the write queue implementation to be changed,
for example, to one which allows fast interval searching.

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

[NET] TIPC: Use htons() where appropriate.

Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET] SCHED: Use htons() where appropriate.

Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET] NETFILTER: Use htonl() where appropriate.

Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET] IPV4: Use hton{s,l}() where appropriate.

Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET] IEEE80211: Use htons() where appropriate.

Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET] ETHERNET: Use htons() where appropriate.

Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET] CORE: Use htons() where appropriate.

Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET] BLUETOOTH: Use cpu_to_le{16,32}() where appropriate.

Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET] ATM: Use htons() where appropriate.

Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET] 8021Q: Use htons() where appropriate.

Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET] 802: Use hton{s,l}() where appropriate.

Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[UDP]: Clean up UDP-Lite receive checksum

This patch eliminates some duplicate code for the verification of
receive checksums between UDP-Lite and UDP.  It does this by
introducing __skb_checksum_complete_head which is identical to
__skb_checksum_complete_head apart from the fact that it takes
a length parameter rather than computing the first skb->len bytes.

As a result UDP-Lite will be able to use hardware checksum offload
for packets which do not use partial coverage checksums.  It also
means that UDP-Lite loopback no longer does unnecessary checksum
verification.

If any NICs start support UDP-Lite this would also start working
automatically.

This patch removes the assumption that msg_flags has MSG_TRUNC clear
upon entry in recvmsg.

Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>

[UDP6]: Restore sk_filter optimisation

This reverts the changeset

    [IPV6]: UDPv6 checksum.

    We always need to check UDPv6 checksum because it is mandatory.

The sk_filter optimisation has nothing to do whether we verify the
checksum.  It simply postpones it to the point when the user calls
recv or poll.

Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>

[IPV4]: Optimize inet_getpeer()

1) Some sysctl vars are declared __read_mostly

2) We can avoid updating stack[] when doing an AVL lookup only.

    lookup() macro is extended to receive a second parameter, that may be NULL
in case of a pure lookup (no need to save the AVL path). This removes
unnecessary instructions, because compiler knows if this _stack parameter is
NULL or not.

    text size of net/ipv4/inetpeer.o is 2063 bytes instead of 2107 on x86_64

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP] TCP Yeah: cleanup

Eliminate need for full 6/4/64 divide to compute queue.
Variable maxqueue was really a constant.
Fix indentation.

Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP] tcp_cubic: faster cube root

The Newton-Raphson method is quadratically convergent so
only a small fixed number of steps are necessary.
Therefore it is faster to unroll the loop. Since div64_64 is no longer
inline it won't cause code explosion.

Also fixes a bug that can occur if x^2 was bigger than 32 bits.

Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[ATM] ENI: Convert to struct timeval to ktime_t.

Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NETLINK]: Limit NLMSG_GOODSIZE to 8K.

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

[IPV6] ADDRCONF: Fix possible inet6_ifaddr leakage with CONFIG_OPTIMISTIC_DAD.

The inet6_ifaddr for source address of RS is leaked if the address
is not an optimistic address.

Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[IPV6] ADDRCONF: Optimistic Duplicate Address Detection (RFC 4429) Support.

Nominally an autoconfigured IPv6 address is added to an interface in the
Tentative state (as per RFC 2462).  Addresses in this state remain in this
state while the Duplicate Address Detection process operates on them to
determine their uniqueness on the network.  During this period, these
tentative addresses may not be used for communication, increasing the time
before a node may be able to communicate on a network.  Using Optimistic
Duplicate Address Detection, autoconfigured addresses may be used
immediately for communication on the network, as long as certain rules are
followed to avoid conflicts with other nodes during the Duplicate Address
Detection process.

Signed-off-by: Neil Horman <nhorman@tuxdriver.com>
Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[IPV6] IP6TUNNEL: Enable to control the handled inner protocol.

ip6_tunnel before supporting IPv4/IPv6 tunnel allows only IPPROTO_IPV6
in configurations from userland. This allows userland to set IPPROTO_IPIP
and 0(wildcard). ip6_tunnel only handles allowed inner protocols.

Signed-off-by: Yasuyuki Kozakai <yasuyuki.kozakai@toshiba.co.jp>
Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[IPV6] IP6TUNNEL: Rename functions ip6ip6_* to ip6_tnl_*.

Signed-off-by: Yasuyuki Kozakai <yasuyuki.kozakai@toshiba.co.jp>
Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[IPV6] IP6TUNNEL: Add support to IPv4 over IPv6 tunnel.

Some notes
- Protocol number IPPROTO_IPIP is used for IPv4 over IPv6 packets.
- If IP6_TNL_F_USE_ORIG_TCLASS is set, TOS in IPv4 header is copied to
  Traffic Class in outer IPv6 header on xmit.
- IP6_TNL_F_USE_ORIG_FLOWLABEL is ignored on xmit of IPv4 packets, because
  IPv4 header does not have flow label.
- Kernel sends ICMP error if IPv4 packet is too big on xmit, even if
  DF flag is not set.

Signed-off-by: Yasuyuki Kozakai <yasuyuki.kozakai@toshiba.co.jp>
Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[IPV6] IP6TUNNEL: Split out generic routine in ip6ip6_xmit().

This enables to add IPv4/IPv6 specific handling later,

Signed-off-by: Yasuyuki Kozakai <yasuyuki.kozakai@toshiba.co.jp>
Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[IPV6] IP6TUNNEL: Split out generic routine in ip6ip6_rcv().

This enables to add IPv4/IPv6 specific handling later,

Signed-off-by: Yasuyuki Kozakai <yasuyuki.kozakai@toshiba.co.jp>
Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[IPV6] IP6TUNNEL: Split out generic routine in ip6ip6_err().

This enables to add IPv4/IPv6 specific error handling later,

Signed-off-by: Yasuyuki Kozakai <yasuyuki.kozakai@toshiba.co.jp>
Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[IPV6]: Decentralize EXPORT_SYMBOLs.

Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>

[NETLINK]: Mirror UDP MSG_TRUNC semantics.

If the user passes MSG_TRUNC in via msg_flags, return
the full packet size not the truncated size.

Idea from Herbert Xu and Thomas Graf.

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

[NET]: convert network timestamps to ktime_t

We currently use a special structure (struct skb_timeval) and plain
'struct timeval' to store packet timestamps in sk_buffs and struct
sock.

This has some drawbacks :
- Fixed resolution of micro second.
- Waste of space on 64bit platforms where sizeof(struct timeval)=16

I suggest using ktime_t that is a nice abstraction of high resolution
time services, currently capable of nanosecond resolution.

As sizeof(ktime_t) is 8 bytes, using ktime_t in 'struct sock' permits
a 8 byte shrink of this structure on 64bit architectures. Some other
structures also benefit from this size reduction (struct ipq in
ipv4/ip_fragment.c, struct frag_queue in ipv6/reassembly.c, ...)

Once this ktime infrastructure adopted, we can more easily provide
nanosecond resolution on top of it. (ioctl SIOCGSTAMPNS and/or
SO_TIMESTAMPNS/SCM_TIMESTAMPNS)

Note : this patch includes a bug correction in
compat_sock_get_timestamp() where a "err = 0;" was missing (so this
syscall returned -ENOENT instead of 0)

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
CC: Stephen Hemminger <shemminger@linux-foundation.org>
CC: John find <linux.kernel@free.fr>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET]: div64_64 consolidate (rev3)

Here is the current version of the 64 bit divide common code.

Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET]: Convert xtime.tv_sec to get_seconds()

Where appropriate, convert references to xtime.tv_sec to the
get_seconds() helper function.

Signed-off-by: James Morris <jmorris@namei.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[PKTGEN]: fix device name handling

Since devices can change name and other wierdness, don't hold onto
a copy of device name, instead use pointer to output device.

Fix a couple of leaks in error handling path as well.

Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: Robert Olsson <robert.olsson@its.uu.se>
Signed-off-by: David S. Miller <davem@davemloft.net>

[PKTGEN]: don't use __constant_htonl()

The existing htonl() macro is smart enough to do the same code as
using __constant_htonl() and it looks cleaner.

Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: Robert Olsson <robert.olsson@its.uu.se>
Signed-off-by: David S. Miller <davem@davemloft.net>

[PKTGEN]: use random32

Can use random32() now.

Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: Robert Olsson <robert.olsson@its.uu.se>
Signed-off-by: David S. Miller <davem@davemloft.net>

[PKTGEN]: use pr_debug

Remove private debug macro and replace with standard version

Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: Robert Olsson <robert.olsson@its.uu.se>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET]: Keep sk_backlog near sk_lock

sk_backlog is a critical field of struct sock. (known famous words)

It is (ab)used in hot paths, in particular in release_sock(), tcp_recvmsg(),
tcp_v4_rcv(), sk_receive_skb().

It really makes sense to place it next to sk_lock, because sk_backlog is only
used after sk_lock locked (and thus memory cache line in L1 cache). This
should reduce cache misses and sk_lock acquisition time.

(In theory, we could only move the head pointer near sk_lock, and leaving tail
far away, because 'tail' is normally not so hot, but keep it simple :) )

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP]: FRTO undo response falls back to ratehalving one if ECEd

Undoing ssthresh is disabled in fastretrans_alert whenever
FLAG_ECE is set by clearing prior_ssthresh. The clearing does
not protect FRTO because FRTO operates before fastretrans_alert.
Moving the clearing of prior_ssthresh earlier seems to be a
suboptimal solution to the FRTO case because then FLAG_ECE will
cause a second ssthresh reduction in try_to_open (the first
occurred when FRTO was entered). So instead, FRTO falls back
immediately to the rate halving response, which switches TCP to
CA_CWR state preventing the latter reduction of ssthresh.

If the first ECE arrived before the ACK after which FRTO is able
to decide RTO as spurious, prior_ssthresh is already cleared.
Thus no undoing for ssthresh occurs. Besides, FLAG_ECE should be
set also in the following ACKs resulting in rate halving response
that sees TCP is already in CA_CWR, which again prevents an extra
ssthresh reduction on that round-trip.

If the first ECE arrived before RTO, ssthresh has already been
adapted and prior_ssthresh remains cleared on entry because TCP
is in CA_CWR (the same applies also to a case where FRTO is
entered more than once and ECE comes in the middle).

High_seq must not be touched after tcp_enter_cwr because CWR
round-trip calculation depends on it.

I believe that after this patch, FRTO should be ECN-safe and
even able to take advantage of synergy benefits.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP]: Complete icsk-to-local-variable change (in tcp_enter_cwr)

A local variable for icsk was created but this change was
missing. Spotted by Jarek Poplawski.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP] Sysctl documentation: tcp_frto_response

In addition, fixed minor things in tcp_frto sysctl.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP]: Add two new spurious RTO responses to FRTO

New sysctl tcp_frto_response is added to select amongst these
responses:
- Rate halving based; reuses CA_CWR state (default)
- Very conservative; used to be the only one available (=1)
- Undo cwr; undoes ssthresh and cwnd reductions (=2)

The response with rate halving requires a new parameter to
tcp_enter_cwr because FRTO has already reduced ssthresh and
doing a second reduction there has to be prevented. In addition,
to keep things nice on 80 cols screen, a local variable was
added.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP]: Correct reordering detection change (no FRTO case)

The reordering detection must work also when FRTO has not been
used at all which was the original intention of mine, just the
expression of the idea was flawed.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP]: Make snd_cwnd_clamp a u32.

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

[TCP]: Keep copied_seq, rcv_wup and rcv_next together.

I noticed in oprofile study a cache miss in tcp_rcv_established() to read
copied_seq.

ffffffff80400a80 <tcp_rcv_established>: /* tcp_rcv_established total: 4034293  
2.0400 */

 55493  0.0281 :ffffffff80400bc9:   mov    0x4c8(%r12),%eax copied_seq
543103  0.2746 :ffffffff80400bd1:   cmp    0x3e0(%r12),%eax   rcv_nxt    

if (tp->copied_seq == tp->rcv_nxt &&
        len - tcp_header_len <= tp->ucopy.len) {

In this function, the cache line 0x4c0 -> 0x500 is used only for this
reading 'copied_seq' field.

rcv_wup and copied_seq should be next to rcv_nxt field, to lower number of
active cache lines in hot paths. (tcp_rcv_established(), tcp_poll(), ...)

As you suggested, I changed tcp_create_openreq_child() so that these fields
are changed together, to avoid adding a new store buffer stall.

Patch is 64bit friendly (no new hole because of alignment constraints)

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP]: struct *sock argument renamed: sp -> sk

In general, TCP code uses "sk" for struct sock pointer.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP]: Add RFC3742 Limited Slow-Start, controlled by variable sysctl_tcp_max_ssthresh.

Signed-off-by: John Heffner <jheffner@psc.edu>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP] YeAH-TCP: algorithm implementation

YeAH-TCP is a sender-side high-speed enabled TCP congestion control
algorithm, which uses a mixed loss/delay approach to compute the
congestion window. It's design goals target high efficiency, internal,
RTT and Reno fairness, resilience to link loss while keeping network
elements load as low as possible.

For further details look here:
    http://wil.cs.caltech.edu/pfldnet2007/paper/YeAH_TCP.pdf

Signed-off-by: Angelo P. Castellani <angelo.castellani@gmail.con>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP] FRTO: Sysctl documentation for SACK enhanced version

The description is overly verbose to avoid ambiguity between
"SACK enabled" and "SACK enhanced FRTO"

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP]: SACK enhanced FRTO

Implements the SACK-enhanced FRTO given in RFC4138 using the
variant given in Appendix B.

RFC4138, Appendix B:
  "This means that in order to declare timeout spurious, the TCP
   sender must receive an acknowledgment for non-retransmitted
   segment between SND.UNA and RecoveryPoint in algorithm step 3.
   RecoveryPoint is defined in conservative SACK-recovery
   algorithm [RFC3517]"

The basic version of the FRTO algorithm can still be used also
when SACK is enabled. To enabled SACK-enhanced version, tcp_frto
sysctl is set to 2.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP]: Prevent reordering adjustments during FRTO

To be honest, I'm not too sure how the reord stuff works in the
first place but this seems necessary.

When FRTO has been active, the one and only retransmission could
be unnecessary but the state and sending order might not be what
the sacktag code expects it to be (to work correctly).

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP] FRTO: Fake cwnd for ssthresh callback

TCP without FRTO would be in Loss state with small cwnd. FRTO,
however, leaves cwnd (typically) to a larger value which causes
ssthresh to become too large in case RTO is triggered again
compared to what conventional recovery would do. Because
consecutive RTOs result in only a single ssthresh reduction,
RTO+cumulative ACK+RTO pattern is required to trigger this
event.

A large comment is included for congestion control module writers
trying to figure out what CA_EVENT_FRTO handler should do because
there exists a remote possibility of incompatibility between
FRTO and module defined ssthresh functions.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP] FRTO: Reverse RETRANS bit clearing logic

Previously RETRANS bits were cleared on the entry to FRTO. We
postpone that into tcp_enter_frto_loss, which is really the
place were the clearing should be done anyway. This allows
simplification of the logic from a clearing loop to the head skb
clearing only.

Besides, the other changes made in the previous patches to
tcp_use_frto made it impossible for the non-SACKed FRTO to be
entered if other than the head has been rexmitted.

With SACK-enhanced FRTO (and Appendix B), however, there can be
a number retransmissions in flight when RTO expires (same thing
could happen before this patchset also with non-SACK FRTO). To
not introduce any jumpiness into the packet counting during FRTO,
instead of clearing RETRANS bits from skbs during entry, do it
later on.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP] FRTO: Entry is allowed only during (New)Reno like recovery

This interpretation comes from RFC4138:
    "If the sender implements some loss recovery algorithm other
     than Reno or NewReno [FHG04], the F-RTO algorithm SHOULD
     NOT be entered when earlier fast recovery is underway."

I think the RFC means to say (especially in the light of
Appendix B) that ...recovery is underway (not just fast recovery)
or was underway when it was interrupted by an earlier (F-)RTO
that hasn't yet been resolved (snd_una has not advanced enough).
Thus, my interpretation is that whenever TCP has ever
retransmitted other than head, basic version cannot be used
because then the order assumptions which are used as FRTO basis
do not hold.

NewReno has only the head segment retransmitted at a time.
Therefore, walk up to the segment that has not been SACKed, if
that segment is not retransmitted nor anything before it, we know
for sure, that nothing after the non-SACKed segment should be
either. This assumption is valid because TCPCB_EVER_RETRANS does
not leave holes but each non-SACKed segment is rexmitted
in-order.

Check for retrans_out > 1 avoids more expensive walk through the
skb list, as we can know the result beforehand: F-RTO will not be
allowed.

SACKed skb can turn into non-SACked only in the extremely rare
case of SACK reneging, in this case we might fail to detect
retransmissions if there were them for any other than head. To
get rid of that feature, whole rexmit queue would have to be
walked (always) or FRTO should be prevented when SACK reneging
happens. Of course RTO should still trigger after reneging which
makes this issue even less likely to show up. And as long as the
response is as conservative as it's now, nothing bad happens even
then.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP]: Prevent unrelated cwnd adjustment while using FRTO

FRTO controls cwnd when it still processes the ACK input or it
has just reverted back to conventional RTO recovery; the normal
rules apply when FRTO has reverted to standard congestion
control.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP] FRTO: frto_counter modulo-op converted to two assignments

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP]: Don't enter to fast recovery while using FRTO

Because TCP is not in Loss state during FRTO recovery, fast
recovery could be triggered by accident. Non-SACK FRTO is more
robust than not yet included SACK-enhanced version (that can
receiver high number of duplicate ACKs with SACK blocks during
FRTO), at least with unidirectional transfers, but under
extraordinary patterns fast recovery can be incorrectly
triggered, e.g., Data loss+ACK losses => cumulative ACK with
enough SACK blocks to meet sacked_out >= dupthresh condition).

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP] FRTO: Response should reset also snd_cwnd_cnt

Since purpose is to reduce CWND, we prevent immediate growth. This
is not a major issue nor is "the correct way" specified anywhere.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP] FRTO: fixes fallback to conventional recovery

The FRTO detection did not care how ACK pattern affects to cwnd
calculation of the conventional recovery. This caused incorrect
setting of cwnd when the fallback becames necessary. The
knowledge tcp_process_frto() has about the incoming ACK is now
passed on to tcp_enter_frto_loss() in allowed_segments parameter
that gives the number of segments that must be added to
packets-in-flight while calculating the new cwnd.

Instead of snd_una we use FLAG_DATA_ACKED in duplicate ACK
detection because RFC4138 states (in Section 2.2):
  If the first acknowledgment after the RTO retransmission
  does not acknowledge all of the data that was retransmitted
  in step 1, the TCP sender reverts to the conventional RTO
  recovery.  Otherwise, a malicious receiver acknowledging
  partial segments could cause the sender to declare the
  timeout spurious in a case where data was lost.

If the next ACK after RTO is duplicate, we do not retransmit
anything, which is equal to what conservative conventional
recovery does in such case.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP] FRTO: Ignore some uninteresting ACKs

Handles RFC4138 shortcoming (in step 2); it should also have case
c) which ignores ACKs that are not duplicates nor advance window
(opposite dir data, winupdate).

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP] FRTO: Use Disorder state during operation instead of Open

Retransmission counter assumptions are to be changed. Forcing
reason to do this exist: Using sysctl in check would be racy
as soon as FRTO starts to ignore some ACKs (doing that in the
following patches). Userspace may disable it at any moment
giving nice oops if timing is right. frto_counter would be
inaccessible from userspace, but with SACK enhanced FRTO
retrans_out can include other than head, and possibly leaving
it non-zero after spurious RTO, boom again.

Luckily, solution seems rather simple: never go directly to Open
state but use Disorder instead. This does not really change much,
since TCP could anyway change its state to Disorder during FRTO
using path tcp_fastretrans_alert -> tcp_try_to_open (e.g., when
a SACK block makes ACK dubious). Besides, Disorder seems to be
the state where TCP should be if not recovering (in Recovery or
Loss state) while having some retransmissions in-flight (see
tcp_try_to_open), which is exactly what happens with FRTO.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP] FRTO: Consecutive RTOs keep prior_ssthresh and ssthresh

In case a latency spike causes more than one RTO, the later should not
cause the already reduced ssthresh to propagate into the prior_ssthresh
since FRTO declares all such RTOs spurious at once or none of them. In
treating of ssthresh, we mimic what tcp_enter_loss() does.

The previous state (in frto_counter) must be available until we have
checked it in tcp_enter_frto(), and also ACK information flag in
process_frto().

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP] FRTO: Comment cleanup & improvement

Moved comments out from the body of process_frto() to the head
(preferred way; see Documentation/CodingStyle). Bonus: it's much
easier to read in this compacted form.

FRTO algorithm and implementation is described in greater detail.
For interested reader, more information is available in RFC4138.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP] FRTO: Moved tcp_use_frto from tcp.h to tcp_input.c

In addition, removed inline.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP] FRTO: Separated response from FRTO detection algorithm

FRTO spurious RTO detection algorithm (RFC4138) does not include response
to a detected spurious RTO but can use different response algorithms.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP] FRTO: Incorrectly clears TCPCB_EVER_RETRANS bit

FRTO was slightly too brave... Should only clear
TCPCB_SACKED_RETRANS bit.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

Linux 2.6.21

.. ok, enough waffling about it already. "Just do it!"

Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

Merge /pub/scm/linux/kernel/git/davem/sparc-2.6

* master.kernel.org:/pub/scm/linux/kernel/git/davem/sparc-2.6:
  [PARPORT] SUNBPP: Fix OOPS when debugging is enabled.
  [SPARC] openprom: Switch to ref counting PCI API

Merge /pub/scm/linux/kernel/git/davem/net-2.6

* master.kernel.org:/pub/scm/linux/kernel/git/davem/net-2.6:
  [NETLINK]: Infinite recursion in netlink.

packet: fix error handling

The packet driver is assuming (reasonably) that the (undocumented)
request.errors is an errno.  But it is in fact some mysterious bitfield.  When
things go wrong we return weird positive numbers to the VFS as pointers and it
goes oops.

Thanks to William Heimbigner for reporting and diagnosis.

(It doesn't oops, but this driver still doesn't work for William)

Cc: William Heimbigner <icxcnika@mar.tar.cc>
Cc: Peter Osterlund <petero2@telia.com>
Cc: Jens Axboe <jens.axboe@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

[NETLINK]: Infinite recursion in netlink.

Reply to NETLINK_FIB_LOOKUP messages were misrouted back to kernel,
which resulted in infinite recursion and stack overflow.

The bug is present in all kernel versions since the feature appeared.

The patch also makes some minimal cleanup:

1. Return something consistent (-ENOENT) when fib table is missing
2. Do not crash when queue is empty (does not happen, but yet)
3. Put result of lookup

Signed-off-by: Alexey Kuznetsov <kuznet@ms2.inr.ac.ru>
Signed-off-by: David S. Miller <davem@davemloft.net>

cfq-iosched: fix alias + front merge bug

There's a really rare and obscure bug in CFQ, that causes a crash in
cfq_dispatch_insert() due to rq == NULL.  One example of the resulting
oops is seen here:

http://lkml.org/lkml/2007/4/15/41

Neil correctly diagnosed the situation for how this can happen: if two
concurrent requests with the exact same sector number (due to direct IO
or aliasing between MD and the raw device access), the alias handling
will add the request to the sortlist, but next_rq remains NULL.

Read the more complete analysis at:

http://lkml.org/lkml/2007/4/25/57

This looks like it requires md to trigger, even though it should
potentially be possible to due with O_DIRECT (at least if you edit the
kernel and doctor some of the unplug calls).

The fix is to move the ->next_rq update to when we add a request to the
rbtree. Then we remove the possibility for a request to exist in the
rbtree code, but not have ->next_rq correctly updated.

Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

IPv6: fix Routing Header Type 0 handling thinko

Oops, thinko.  The test for accempting a RH0 was exatly the wrong way
around.

Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

Merge /pub/scm/linux/kernel/git/davem/net-2.6

* master.kernel.org:/pub/scm/linux/kernel/git/davem/net-2.6:
  [BNX2]: Fix occasional NETDEV WATCHDOG on 5709.
  [IPV6]: Disallow RH0 by default.
  [XFRM]: beet: fix pseudo header length value
  [TCP]: Congestion control initialization.

[BNX2]: Fix occasional NETDEV WATCHDOG on 5709.

Tweak a register setting to prevent the tx mailbox from halting.

Update version to 1.5.8.

Signed-off-by: Michael Chan <mchan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>