[NETEM]: spelling errors

Get rid of some of my creative spelling.

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

[NETLINK]: Directly return -EINTR from netlink_dump_start()

Now that all users of netlink_dump_start() use netlink_run_queue()
to process the receive queue, it is possible to return -EINTR from
netlink_dump_start() directly, therefore simplying the callers.

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

[IPv4] diag: Use netlink_run_queue() to process the receive queue

Makes use of netlink_run_queue() to process the receive queue and
converts inet_diag_rcv_msg() to use the type safe netlink interface.

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NETLINK]: Remove error pointer from netlink message handler

The error pointer argument in netlink message handlers is used
to signal the special case where processing has to be interrupted
because a dump was started but no error happened. Instead it is
simpler and more clear to return -EINTR and have netlink_run_queue()
deal with getting the queue right.

nfnetlink passed on this error pointer to its subsystem handlers
but only uses it to signal the start of a netlink dump. Therefore
it can be removed there as well.

This patch also cleans up the error handling in the affected
message handlers to be consistent since it had to be touched anyway.

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NETLINK]: Ignore control messages directly in netlink_run_queue()

Changes netlink_rcv_skb() to skip netlink controll messages and don't
pass them on to the message handler.

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NETLINK]: Ignore !NLM_F_REQUEST messages directly in netlink_run_queue()

netlink_rcv_skb() is changed to skip messages which don't have the
NLM_F_REQUEST bit to avoid every netlink family having to perform this
check on their own.

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NETLINK]: Remove unused groups variable

Leftover from dynamic multicast groups allocation work.

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP] westwood: Use type safe netlink interface

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

[TCP] vegas: Use type safe netlink interface

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

[RTNL]: Properly return rntl message handler

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET_SCHED] qdisc: avoid transmit softirq on watchdog wakeup

If possible, avoid having to do a transmit softirq when a qdisc
watchdog decides to re-enable.  The watchdog routine runs off
a timer, so it is already in the same effective context as
the softirq.

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

[NETEM]: avoid excessive requeues

The netem code would call getnstimeofday() and dequeue/requeue after
every packet, even if it was waiting. Avoid this overhead by using
the throttled flag.

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

[NETEM]: Optimize tfifo

In most cases, the next packet will be sent after the
last one. So optimize that case.

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

[NETEM]: use better types for time values

The random number generator always generates 32 bit values.
The time values are limited by psched_tdiff_t

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

[NETEM]: report reorder percent correctly.

If you setup netem to just delay packets; "tc qdisc ls" will report
the reordering as 100%. Well it's a lie, reorder isn't used unless
gap is set, so just set value to 0 so the output of utility
is correct.

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

[TCP]: cubic optimization

Use willy's work in optimizing cube root by having table for small values.

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

[LIB]: div64_64 optimization

Minor optimization of div64_64.  do_div() already does optimization
for the case of 32 by 32 divide, so no need to do it here.

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

[NET] rules: Unified rules dumping

Implements a unified, protocol independant rules dumping function
which is capable of both, dumping a specific protocol family or
all of them. This speeds up dumping as less lookups are required.

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

[RTNL]: Use rtnl registration interface for dump-all aliases

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

[BRIDGE]: Use rtnl registration interface

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

[IPv6]: Use rtnl registration interface

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

[DECNet]: Use rtnl registration interface

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Acked-by: Steven Whitehouse <swhiteho@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[PKT_SCHED] act: Use rtnl registration interface

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

[PKT_SCHED] cls: Use rtnl registration interface

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

[PKT_SCHED] qdisc: Use rtnl registration interface

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

[IPv4]: Use rtnl registration interface

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET] rules: Use rtnl registration interface

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NEIGH]: Use rtnl registration interface

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET] link: Use rtnl registration interface

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

[RTNL]: Message handler registration interface

This patch adds a new interface to register rtnetlink message
handlers replacing the exported rtnl_links[] array which
required many message handlers to be exported unnecessarly.

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

[CCID3]: Use initial RTT sample from SYN exchange

The patch follows the following recommendation made in an erratum to RFC 4342:

  "Senders MAY additionally make use of other available RTT measurements,
   including those from the initial Request-Response packet exchange."

It implements larger initial windows with regard to this inital RTT measurement,
using the mechanism suggested in draft-ietf-dccp-rfc3448bis, section 4.2.

Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[DCCP]: Sample RTT from SYN exchange

Function:

[CCID3]: Use function for RTT sampling

This replaces the existing occurrences of RTT sampling with
the use of the new function dccp_sample_rtt.

Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[DCCP]: Provide function for RTT sampling

A recurring problem, in particular in the CCID code, is that RTT samples
from packets with timestamp echo and elapsed time options need to be taken.

This service is provided via a new function dccp_sample_rtt in this patch.
Furthermore, to protect against `insane' RTT samples, the sampled value
is bounded between 100 microseconds and 4 seconds - for which u32 is sufficient.

Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[CCID3]: Handle Idle and Application-Limited periods

This updates the code with regard to handling idle and application-limited
periods as specified in [RFC 4342, 5.1].

Background:

[CCID3]: Wrap computation of RFC3390-initial rate into separate function

The CCID 3 and TFRC specs (RFC 4342, RFC 3448, draft-3448bis) make frequent
reference to the computation of the RFC-3390 initial sending rate:

  1. Initial sending rate when RTT is known (RFC 4342, p. 6)
  2. Response to Idle/Application-Limited periods (RFC 4342, 5.1)

This warrants putting the code into its own function, for later code reuse.

Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[CCID3]: Remove build warnings for 64bit

This clears the following sparc64 build warnings:
 1) warning: format "%ld" expects type "long int", but argument 3 has type "suseconds_t"
 2) warning: format "%llu" expects type "long long unsigned int", but argument 3 has type "__u64"
Fixed by using typecast to unsigned. This is argued to be safe, since the quantities, after
de-scaling (factor 2^6) fit all in u32.

Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[CCID3]: More to see in dccp_probe

This adds a few more fields of interest to /proc/net/dccpprobe, the following output ensues:

1           2          3           4     5  6     7   8        9        10   11
sec.usec   src:sport   dst:dport   size  s  rtt   p   X_calc   X_recv   X    t_ipi

Also made the formatting consistent.

Scripts that go with this can be downloaded from http://139.133.210.30/users/gerrit/dccp/dccp_probe/

Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[CCID3]: Add documentation for socket options

This updates the documentation on CCID3-specific options.

Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[DCCP]: More debug information for dccp_wait_for_ccid

This adds more detail in the wait_for_ccid packet scheduling loop.
In particular, it informs about (i) when delay is used and (ii) why
a packet is discarded.

Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[DCCP]: Always use debug-toggle parameters

Currently debugging output (when configured) is automatically enabled when
DCCP modules are compiled into the kernel rather than built as loadable modules.
This is not necessary, since the module parameters in this case become kernel
commandline parameters, e.g. DCCP or CCID3 debug output can be enabled for a
static build by appending the following at the boot prompt:

dccp.dccp_debug=1  dccp_ccid3.ccid3_debug=1

This patch therefore does away with the more complicated way of always enabling
debug output for static builds

Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[CCID3]: Remove race condition and update t_ipi when `s' changes

This:

 1. removes a race condition in the access to the scheduled send time t_nom which
    results from allowing asynchronous r/w access to t_nom without locks;

 2. updates the inter-packet interval t_ipi = s/X when `s' changes, following a
    suggestion by Ian McDonald.

Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[CCID3]: More verbose debugging

This adds a few debugging statements to ccid3.c

Signed-off-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[CCID3]: Fix use of invalid loss intervals

This fixes a bug which uses an invalid comparison.
The bug resulted in the use of invalid loss intervals.

Signed-off-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Acked-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[CCID3]: Use MSS for larger initial windows

This improves the slow-start phase by using the MSS
(as suggested in RFC 4342, sec. 5) instead of the packet size s.
Also figured out that __u32 is ample resource enough.

After applying, I got the following in the logs:

  ccid3_hc_tx_packet_recv: client(f7421700), s=6, MSS=1424, w_init=4380, R_sample=176us, X=24886363

Had the previous variant been used, w_init would have been as low as 24.

Committer note: removed unneeded cast to unsigned long long that was
                causing a compiler warning on 64bit architectures.

Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[CCID3]: Re-order CCID 3 source file

No code change at all.
This splits ccid3.c into a RX and a TX section, so that the file has an
organisation similar to the other ones (e.g. packet_history.{h,c}).

Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[CCID3]: Remove redundant `len' test

Since CCID3 avoids  sending 0-byte data packets (cf. ccid3_hc_tx_send_packet),
testing for zero-payload length, as performed by ccid3_hc_tx_update_s, is
redundant - hence removed by this patch.

Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[DCCP]: Remove ambiguity in the way before48 is used

This removes two ambiguities in employing the new definition of before48,
following the analysis on http://www.mail-archive.com/dccp@vger.kernel.org/msg01295.html

 (1) Updating GSR when P.seqno >= S.SWL
     With the old definition we did not update when P.seqno and S.SWL are 2^47 apart. To
     ensure the same behaviour as with the old definition, this is replaced with the
     equivalent condition dccp_delta_seqno(S.SWL, P.seqno) >= 0

 (2) Sending SYNC when P.seqno >= S.OSR
     Here it is debatable whether the new definition causes an ambiguity: the case is
     similar to (1); and to have consistency with the case (1), we use the equivalent
     condition dccp_delta_seqno(S.OSR, P.seqno) >= 0

Detailed Justification

[DCCP]: Fix for follows48

The follows48 relation identifies whether 48-bit sequence number
x is the direct successor of y. Currently, it does not handle cases
of the following type correctly:

follows48(0x(prefix)10000LL, 0x(prefix)0FFFFLL)

where prefix is an arbitrary hex sequence of up to 7 digits.

This is fixed by reusing the new dccp_delta_seqno function.

Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[DCCP]: Make `before' relation unambiguous

Problem:

[DCCP]: Make dccp_delta_seqno return signed numbers

Problem:

[DCCP]: 48-bit sequence number arithmetic

This patch
 * organizes the sequence arithmetic functions into one corner of dccp.h
 * performs a small modification of dccp_set_seqno to make it more widely reusable
   (now it is safe to use any number, since it performs modulo-2^48 assignment)
 * adds functions and generic macros for 48-bit sequence arithmetic:
  --48 bit complement
  --modulo-48 addition and modulo-48 subtraction
--dccp_inc_seqno now a special case of add48
Constants renamed following a suggestion by Arnaldo.

Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

[FORCEDETH]: Use skb_tailroom where appropriate

Reducing the number of skb->data direct accesses.

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

[LMC]: lmc_main wants to use skb_tailroom

At that point it is equivalent to what was being used, skb->end - skb->data,
and the need is clearly the one skb_tailroom satisfies.

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

[ATM] idt77252: Fix double kfree_skb on failure in push_rx_skb

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>

[SK_BUFF] ipv6: Use skb_network_offset in some more places

So that we reduce the number of direct accesses to skb->data.

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>

[NETLINK]: Use nlmsg_trim() where appropriate

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

[NETLINK]: Remove NLMSG_{NEW_ANSWER,CANCEL,END}

Not used anywhere and defined inside __KERNEL__, Thomas acked this on irc.

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>

[SK_BUFF]: Remove skb_add_mtu() leftovers

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>

[NETLINK]: Introduce nlmsg_hdr() helper

For the common "(struct nlmsghdr *)skb->data" sequence, so that we reduce the
number of direct accesses to skb->data and for consistency with all the other
cast skb member helpers.

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

[IPV4]: fib_trie root node settings

The threshold for root node can be more aggressive set to get
better tree compression. The new setting mekes the root grow
from 16 to 19 bits and substansial improvemnt in Aver depth
this with the current table of 214393 prefixes

But really the dynamic resize should need more investigation
both in terms convergence and performance and maybe it should
be possible to change...

Maybe just for the brave to start with or we may have to back
this out.

[IPV4]: fib_trie resize break

The patch below adds break condition for the resize operations. If
we don't achieve the desired fill factor a warning is printed. Trie
should still be operational but new thresholds should be considered.

Signed-off-by: Robert Olsson <robert.olsson@its.uu.se>
Signed-off-by: David S. Miller <davem@davemloft.net>

[SK_BUFF]: Adjust the zeroing up to tail in __alloc_skb too

I did it just in alloc_skb_from_cache, forgot __alloc_skb, fixed now.

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

[SK_BUFF]: Convert skb->end to sk_buff_data_t

Now to convert the last one, skb->data, that will allow many simplifications
and removal of some of the offset helpers.

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

[SK_BUFF]: Convert skb->tail to sk_buff_data_t

So that it is also an offset from skb->head, reduces its size from 8 to 4 bytes
on 64bit architectures, allowing us to combine the 4 bytes hole left by the
layer headers conversion, reducing struct sk_buff size to 256 bytes, i.e. 4
64byte cachelines, and since the sk_buff slab cache is SLAB_HWCACHE_ALIGN...
:-)

Many calculations that previously required that skb->{transport,network,
mac}_header be first converted to a pointer now can be done directly, being
meaningful as offsets or pointers.

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

[VLAN] vlan_dev: Use skb_reset_network_header().

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

[IrDA]: SMC SuperIO Chip LPC47N227 not identified properly

SMC SuperIO Chip LPC47N227 used for IrDA is not detected because its device
identification byte can be 0x7A instead of 0x5A.

Patch from Peter Kovar <peter.kovar@gmail.com>
Cc: Jean Delvare <khali@linux-fr.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Samuel Ortiz <samuel@sortiz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[IrDA]: irda lockdep annotation

Rmmoding irda triggers a lockdep false positive.

Reported-by: Dave Jones <davej@redhat.com>
Signed-off-by: Samuel Ortiz <samuel@sortiz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[IrDA]: removing stir4200 useless include

stir4200 doesn't need to include irlap.h

Signed-off-by: Samuel Ortiz <samuel@sortiz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

[SK_BUFF]: Use offsets for skb->{mac,network,transport}_header on 64bit architectures

With this we save 8 bytes per network packet, leaving a 4 bytes hole to be used
in further shrinking work, likely with the offsetization of other pointers,
such as ->{data,tail,end}, at the cost of adds, that were minimized by the
usual practice of setting skb->{mac,nh,n}.raw to a local variable that is then
accessed multiple times in each function, it also is not more expensive than
before with regards to most of the handling of such headers, like setting one
of these headers to another (transport to network, etc), or subtracting, adding
to/from it, comparing them, etc.

Now we have this layout for sk_buff on a x86_64 machine:

[acme@mica net-2.6.22]$ pahole vmlinux sk_buff
struct sk_buff {
struct sk_buff *       next;             /*   0   8 */
struct sk_buff *       prev;             /*   8   8 */
struct rb_node         rb;               /*  16  24 */
struct sock *          sk;               /*  40   8 */
ktime_t                tstamp;           /*  48   8 */
struct net_device *    dev;              /*  56   8 */
/* --- cacheline 1 boundary (64 bytes) --- */
struct net_device *    input_dev;        /*  64   8 */
sk_buff_data_t         transport_header; /*  72   4 */
sk_buff_data_t         network_header;   /*  76   4 */
sk_buff_data_t         mac_header;       /*  80   4 */

/* XXX 4 bytes hole, try to pack */

struct dst_entry *     dst;              /*  88   8 */
struct sec_path *      sp;               /*  96   8 */
char                   cb[48];           /* 104  48 */
/* cacheline 2 boundary (128 bytes) was 24 bytes ago*/
unsigned int           len;              /* 152   4 */
unsigned int           data_len;         /* 156   4 */
unsigned int           mac_len;          /* 160   4 */
union {
__wsum         csum;             /*       4 */
__u32          csum_offset;      /*       4 */
};                                       /* 164   4 */
__u32                  priority;         /* 168   4 */
__u8                   local_df:1;       /* 172   1 */
__u8                   cloned:1;         /* 172   1 */
__u8                   ip_summed:2;      /* 172   1 */
__u8                   nohdr:1;          /* 172   1 */
__u8                   nfctinfo:3;       /* 172   1 */
__u8                   pkt_type:3;       /* 173   1 */
__u8                   fclone:2;         /* 173   1 */
__u8                   ipvs_property:1;  /* 173   1 */

/* XXX 2 bits hole, try to pack */

__be16                 protocol;         /* 174   2 */
void    (*destructor)(struct sk_buff *); /* 176   8 */
struct nf_conntrack *  nfct;             /* 184   8 */
/* --- cacheline 3 boundary (192 bytes) --- */
struct sk_buff *       nfct_reasm;       /* 192   8 */
struct nf_bridge_info *nf_bridge;        /* 200   8 */
__u16                  tc_index;         /* 208   2 */
__u16                  tc_verd;          /* 210   2 */
dma_cookie_t           dma_cookie;       /* 212   4 */
__u32                  secmark;          /* 216   4 */
__u32                  mark;             /* 220   4 */
unsigned int           truesize;         /* 224   4 */
atomic_t               users;            /* 228   4 */
unsigned char *        head;             /* 232   8 */
unsigned char *        data;             /* 240   8 */
unsigned char *        tail;             /* 248   8 */
/* --- cacheline 4 boundary (256 bytes) --- */
unsigned char *        end;              /* 256   8 */
}; /* size: 264, cachelines: 5 */
   /* sum members: 260, holes: 1, sum holes: 4 */
   /* bit holes: 1, sum bit holes: 2 bits */
   /* last cacheline: 8 bytes */

On 32 bits nothing changes, and pointers continue to be used with the compiler
turning all this abstraction layer into dust. But there are some sk_buff
validation tricks that are now possible, humm... :-)

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

[SK_BUFF]: unions of just one member don't get anything done, kill them

Renaming skb->h to skb->transport_header, skb->nh to skb->network_header and
skb->mac to skb->mac_header, to match the names of the associated helpers
(skb[_[re]set]_{transport,network,mac}_header).

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

[SK_BUFF]: Introduce skb_network_header_len

For the common sequence "skb->h.raw - skb->nh.raw", similar to skb->mac_len,
that is precalculated tho, don't think we need to bloat skb with one more
member, so just use this new helper, reducing the number of non-skbuff.h
references to the layer headers even more.

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

[SK_BUFF]: Use the helpers to get the layer header pointer

Some more cases...

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

[NET_SCHED]: Fix warning

net/sched/sch_api.c: In function 'psched_show':
net/sched/sch_api.c:1219: warning: format '%08x' expects type 'unsigned int', but argument 6 has type 's64'

Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET_SCHED]: sch_cbq: fix watchdog scheduled too late

q->now is increased during dequeue and doesn't contain the current time
afterwards, resulting in a too large timeout value for the qdisc watchdog.
Use "now" instead, which still contains the current time.

Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET_SCHED]: Export real timer resolution in /proc/net/psched

The timer resolution exported in /proc/net/psched is used by userspace to
calculate HTB's burst values. Currently it is set to HZ, since we're now
using hrtimers, use KTIME_MONOTONIC_RES, which makes HTB use smaller burst
values.

This patch also affects libnl, which incorrectly uses this value for
the SFQ perturbation parameter, which is always in seconds, and some
routing cache values, which are in USER_HZ, so both cases are broken
anyway.

Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET_SCHED]: kill jiffie conversion macros

Now that all packet schedulers have been converted to hrtimers most users
of PSCHED_JIFFIE2US and PSCHED_US2JIFFIE are gone. The remaining users use
it to convert external time units to packet scheduler clock ticks, so use
PSCHED_TICKS_PER_SEC instead.

Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET_SCHED]: sch_htb: use hrtimer based watchdog

Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET_SCHED]: sch_cbq: use hrtimer for delay_timer

Switch delay_timer to hrtimer.

The class penalty parameter is changed to use psched ticks as units.
Since iproute never supported using this and the only existing user
(libnl) incorrectly assumes psched ticks as units anyway, this
shouldn't break anything.

Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET_SCHED]: sch_cbq: fix cbq_undelay_prio for non-active priorites

cbq_undelay_prio is supposed to return a time delta, but returns the
current time for non-active priorities, causing cbq_undelay to mark
the priority as active and schedule a timer for twice the current
time.

Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET_SCHED]: sch_cbq: use hrtimer based watchdog

Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET_SCHED]: sch_netem: use hrtimer based watchdog

Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET_SCHED]: sch_tbf: use hrtimer based watchdog

Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET_SCHED]: sch_hfsc: use hrtimer based watchdog

Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET_SCHED]: Add hrtimer based qdisc watchdog

Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NET_SCHED]: Use ktime as clocksource

Get rid of the manual clock source selection mess and use ktime. Also
use a scalar representation, which allows to clean up pkt_sched.h a bit
more and results in less ktime_to_ns() calls in most cases.

The PSCHED_US2JIFFIE/PSCHED_JIFFIE2US macros are implemented quite
inefficient by this patch, following patches will convert all qdiscs
to hrtimers and get rid of them entirely.

Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

[SK_BUFF]: Some more layer header conversions

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

[KBUILD]: Unifdef headers changed by the skb layer header refactorings

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

[SK_BUFF]: More skb_put related skb_reset_transport_header

This time we have to set it to skb->tail that is not anymore equal to
skb->data, so we either add a new helper or just add the skb->tail - skb->data
offset, for now do the later.

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

[IPV6]: Reset the network header in ip6_nd_hdr

ip6_nd_hdr is always called immediately after a alloc_skb + skb_reserve
sequence, i.e. when skb->tail is equal to skb->data, making it correct to use
skb_reset_network_header().

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

[SK_BUFF]: More skb_put related conversions to skb_reset_transport_header

This is similar to the skb_reset_network_header(), i.e. at the point we reset
the transport header pointer/offset skb->tail is equal to skb->data.

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

[NETFILTER]: nfnetlink: parse attributes with nfattr_parse in nfnetlink_check_attribute

Use nfattr_parse to parse attributes, this patch also modifies the default
behaviour since unknown attributes will be ignored instead of returning
EINVAL. This ensure backward compatibility: new libraries with new
attributes and old kernels can work.

Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NETFILTER]: ctnetlink: add support for internal tcp connection tracking flags handling

This patch let userspace programs set the IP_CT_TCP_BE_LIBERAL flag to
force the pickup of established connections.

Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NETFILTER]: TCP conntrack: factorize out the PUSH flag

The PUSH flag is accepted with every other valid combination.
Let's get it out of the tcp_valid_flags table and reduce the
number of combinations we have to handle. This does not
significantly reduce the table size however (8 bytes).

Signed-off-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NETFILTER]: TCP conntrack: accept RST|PSH as valid

This combination has been encountered on an IBM AS/400 in response
to packets sent to a closed session. There is no particular reason
to mark it invalid.

Signed-off-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NETFILTER]: nf_conntrack: add nf_copy() to safely copy members in skb

This unifies the codes to copy netfilter related datas. Before copying,
nf_copy() puts original members in destination skb.

Signed-off-by: Yasuyuki Kozakai <yasuyuki.kozakai@toshiba.co.jp>
Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NETFILTER]: nf_conntrack: add __nf_copy() to copy members in skb

This unifies the codes to copy netfilter related datas. Note that
__nf_copy() assumes destination skb doesn't have any netfilter
related members.

Signed-off-by: Yasuyuki Kozakai <yasuyuki.kozakai@toshiba.co.jp>
Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NETFILTER]: nf_conntrack: use jhash2 in __hash_conntrack

Now it uses jhash, but using jhash2 would be around 3-4 times faster
(on P4).

Signed-off-by: Sami Farin <safari-netfilter@safari.iki.fi>
Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

[JHASH]: Use const in jhash2

Use const to avoid forcing users to cast const data.

Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

[NETFILTER]: nfnetlink: move EXPORT_SYMBOL declarations next to the exported symbol

Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>