Results on dummy device can be seen in my netconf 2011
slides. These results are for a 10Gige IXGBE intel
nic - on another i5 machine, very similar specs to
the one used in the netconf2011 results.
It turns out - this is a hell lot worse than dummy
and so this patch is even more beneficial for 10G.

Test setup:
----------

System under test sending packets out.
Additional box connected directly dropping packets.
Installed prio qdisc on the eth device and default
netdev default length of 1000 used as is.
The 3 prio bands each were set to 100 (didnt factor in
the results).

5 packet runs were made and the middle 3 picked.

results
-------

The "cpu" column indicates the which cpu the sample
was taken on,
The "Pkt runx" carries the number of packets a cpu
dequeued when forced to be in the "dequeuer" role.
The "avg" for each run is the number of times each
cpu should be a "dequeuer" if the system was fair.

3.0-rc4      (plain)
cpu         Pkt run1        Pkt run2        Pkt run3
================================================
cpu0        21853354        21598183        22199900
cpu1          431058          473476          393159
cpu2          481975          477529          458466
cpu3        23261406        23412299        22894315
avg         11506948        11490372        11486460

3.0-rc4 with patch and default weight 64
cpu 	     Pkt run1        Pkt run2        Pkt run3
================================================
cpu0        13205312        13109359        13132333
cpu1        10189914        10159127        10122270
cpu2        10213871        10124367        10168722
cpu3        13165760        13164767        13096705
avg         11693714        11639405        11630008

As you can see the system is still not perfect but
is a lot better than what it was before...

At the moment we use the old backlog weight, weight_p
which is 64 packets. It seems to be reasonably fine
with that value.
The system could be made more fair if we reduce the
weight_p (as per my presentation), but we are going
to affect the shared backlog weight. Unless deemed
necessary, I think the default value is fine. If not
we could add yet another knob.
Signed-off-by: NJamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This interface uses a temporary buffer, but for no real reason.
And now can generate warnings like:

net/sched/sch_generic.c: In function dev_watchdog
net/sched/sch_generic.c:254:10: warning: unused variable drivername

Just return driver->name directly or "".
Reported-by: NConnor Hansen <cmdkhh@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

dev_deactivate_many() issues one synchronize_rcu() call after qdiscs set
to noop_qdisc.

This call is here to make sure they are no outstanding qdisc-less
dev_queue_xmit calls before returning to caller.

But in dismantle phase, we dont have to wait, because we wont activate
again the device, and we are going to wait one rcu grace period later in
rollback_registered_many().

After this patch, device dismantle uses one synchronize_net() and one
rcu_barrier() call only, so we have a ~30% speedup and a smaller RTNL
latency.
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
CC: Patrick McHardy <kaber@trash.net>,
CC: Ben Greear <greearb@candelatech.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Because of various alignements [SLUB / qdisc], we use 512 bytes of
memory for one {p|b}fifo qdisc, instead of 256 bytes on 64bit arches and
192 bytes on 32bit ones.

Move the "u32 limit" inside "struct Qdisc" (no impact on other qdiscs)

Change qdisc_alloc(), first trying a regular allocation before an
oversized one.
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

From: Eric W. Biederman <ebiederm@xmission.com>

In the beginning with batching unreg_list was a list that was used only
once in the lifetime of a network device (I think).  Now we have calls
using the unreg_list that can happen multiple times in the life of a
network device like dev_deactivate and dev_close that are also using the
unreg_list.  In addition in unregister_netdevice_queue we also do a
list_move because for devices like veth pairs it is possible that
unregister_netdevice_queue will be called multiple times.

So I think the change below to fix dev_deactivate which Eric D. missed
will fix this problem.  Now to go test that.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Now qdisc stab is handled before TCQ_F_CAN_BYPASS test in
__dev_xmit_skb(), we can generalize TCQ_F_CAN_BYPASS to other qdiscs
than pfifo_fast : pfifo, bfifo, pfifo_head_drop and sfq

SFQ is special because it can have external classifiers, and in these
cases, we cannot bypass queue discipline (packet could be dropped by
classifier) without admin asking it, or further changes.

Its worth doing this, especially for SFQ, avoiding dirtying memory in
case no packets are already waiting in queue.
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch converts stab qdisc management to RCU, so that we can perform
the qdisc_calculate_pkt_len() call before getting qdisc lock.

This shortens the lock's held time in __dev_xmit_skb().

This permits more qdiscs to get TCQ_F_CAN_BYPASS status, avoiding lot of
cache misses and so reducing latencies.
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
CC: Patrick McHardy <kaber@trash.net>
CC: Jesper Dangaard Brouer <hawk@diku.dk>
CC: Jarek Poplawski <jarkao2@gmail.com>
CC: Jamal Hadi Salim <hadi@cyberus.ca>
CC: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Cleanup net/sched code to current CodingStyle and practices.

Reduce inline abuse
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This implements a mqprio queueing discipline that by default creates
a pfifo_fast qdisc per tx queue and provides the needed configuration
interface.

Using the mqprio qdisc the number of tcs currently in use along
with the range of queues alloted to each class can be configured. By
default skbs are mapped to traffic classes using the skb priority.
This mapping is configurable.

Configurable parameters,

struct tc_mqprio_qopt {
	__u8    num_tc;
	__u8    prio_tc_map[TC_BITMASK + 1];
	__u8    hw;
	__u16   count[TC_MAX_QUEUE];
	__u16   offset[TC_MAX_QUEUE];
};

Here the count/offset pairing give the queue alignment and the
prio_tc_map gives the mapping from skb->priority to tc.

The hw bit determines if the hardware should configure the count
and offset values. If the hardware bit is set then the operation
will fail if the hardware does not implement the ndo_setup_tc
operation. This is to avoid undetermined states where the hardware
may or may not control the queue mapping. Also minimal bounds
checking is done on the count/offset to verify a queue does not
exceed num_tx_queues and that queue ranges do not overlap. Otherwise
it is left to user policy or hardware configuration to create
useful mappings.

It is expected that hardware QOS schemes can be implemented by
creating appropriate mappings of queues in ndo_tc_setup().

One expected use case is drivers will use the ndo_setup_tc to map
queue ranges onto 802.1Q traffic classes. This provides a generic
mechanism to map network traffic onto these traffic classes and
removes the need for lower layer drivers to know specifics about
traffic types.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Add dev_close_many and dev_deactivate_many to factorize another
sync-rcu operation on the netdevice unregister path.

$ modprobe dummy numdummies=10000
$ ip link set dev dummy* up
$ time rmmod dummy

Without the patch           With the patch

real    0m 24.63s           real    0m 5.15s
user    0m 0.00s            user    0m 0.00s
sys     0m 6.05s            sys     0m 5.14s
Signed-off-by: NOctavian Purdila <opurdila@ixiacom.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Allocate qdisc memory according to NUMA properties of cpus included in
xps map.

To be effective, qdisc should be (re)setup after changes
of /sys/class/net/eth<n>/queues/tx-<n>/xps_cpus

I added a numa_node field in struct netdev_queue, containing NUMA node
if all cpus included in xps_cpus share same node, else -1.
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Cc: Tom Herbert <therbert@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When testing struct netdev_queue state against FROZEN bit, we also test
XOFF bit. We can test both bits at once and save some cycles.
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The first parameter dev isn't in use in qdisc_create_dflt().
Signed-off-by: NChangli Gao <xiaosuo@gmail.com>
Acked-by: NJamal Hadi Salim <hadi@cyberus.ca>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Under network load, doing :

tc qdisc del dev eth0 root

triggers :

[  167.193087] BUG: spinlock bad magic on CPU#3, udpflood/4928
[  167.193139]  lock: c15bc324, .magic: 00000000, .owner:
<none>/-1, .owner_cpu: -1
[  167.193193] Pid: 4928, comm: udpflood Not tainted
2.6.36-rc7-11417-g215340c-dirty #323
[  167.193245] Call Trace:
[  167.193292]  [<c13abaa0>] ? printk+0x18/0x20
[  167.193342]  [<c11afb53>] spin_bug+0xa3/0xf0
[  167.193389]  [<c11afcdd>] do_raw_spin_lock+0x7d/0x160
[  167.193440]  [<c1313d4e>] ? __dev_xmit_skb+0x27e/0x2b0
[  167.193496]  [<c107382b>] ? trace_hardirqs_on+0xb/0x10
[  167.193545]  [<c13ae99a>] _raw_spin_lock+0x3a/0x40
[  167.193593]  [<c1313d4e>] ? __dev_xmit_skb+0x27e/0x2b0
[  167.193641]  [<c1313d4e>] __dev_xmit_skb+0x27e/0x2b0

commit 79640a4c (add additional lock to qdisc to increase
throughput) forgot to initialize  noop_qdisc and noqueue_qdisc busylock
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

ingress being not used very much, and net_device->ingress_queue being
quite a big object (128 or 256 bytes), use a dynamic allocation if
needed (tc qdisc add dev eth0 ingress ...)

dev_ingress_queue(dev) helper should be used only with RTNL taken.
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

There is some confusion with rx_queue name after RPS, and net drivers
private rx_queue fields.

I suggest to rename "struct net_device"->rx_queue to ingress_queue.
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Use modern this_cpu_xxx() api, saving few bytes on x86
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When many cpus compete for sending frames on a given qdisc, the qdisc
spinlock suffers from very high contention.

The cpu owning __QDISC_STATE_RUNNING bit has same priority to acquire
the lock, and cannot dequeue packets fast enough, since it must wait for
this lock for each dequeued packet.

One solution to this problem is to force all cpus spinning on a second
lock before trying to get the main lock, when/if they see
__QDISC_STATE_RUNNING already set.

The owning cpu then compete with at most one other cpu for the main
lock, allowing for higher dequeueing rate.

Based on a previous patch from Alexander Duyck. I added the heuristic to
avoid the atomic in fast path, and put the new lock far away from the
cache line used by the dequeue worker. Also try to release the busylock
lock as late as possible.

Tests with following script gave a boost from ~50.000 pps to ~600.000
pps on a dual quad core machine (E5450 @3.00GHz), tg3 driver.
(A single netperf flow can reach ~800.000 pps on this platform)

for j in `seq 0 3`; do
  for i in `seq 0 7`; do
    netperf -H 192.168.0.1 -t UDP_STREAM -l 60 -N -T $i -- -m 6 &
  done
done
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Acked-by: NAlexander Duyck <alexander.h.duyck@intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Define three helpers to manipulate QDISC_STATE_RUNNIG flag, that a
second patch will move on another location.
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Currently such notifications are only generated when the device comes up or the
address changes. However one use case for these notifications is to enable
faster network recovery after a virtual machine migration (by causing switches
to relearn their MAC tables). A migration appears to the network stack as a
temporary loss of carrier and therefore does not trigger either of the current
conditions. Rather than adding carrier up as a trigger (which can cause issues
when interfaces a flapping) simply add an interface which the driver can use
to explicitly trigger the notification.
Signed-off-by: NIan Campbell <ian.campbell@citrix.com>
Cc: Stephen Hemminger <shemminger@linux-foundation.org>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: netdev@vger.kernel.org
Cc: stable@kernel.org
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Use low order bit of skb->_skb_dst to tell dst is not refcounted.

Change _skb_dst to _skb_refdst to make sure all uses are catched.

skb_dst() returns the dst, regardless of noref bit set or not, but
with a lockdep check to make sure a noref dst is not given if current
user is not rcu protected.

New skb_dst_set_noref() helper to set an notrefcounted dst on a skb.
(with lockdep check)

skb_dst_drop() drops a reference only if skb dst was refcounted.

skb_dst_force() helper is used to force a refcount on dst, when skb
is queued and not anymore RCU protected.

Use skb_dst_force() in __sk_add_backlog(), __dev_xmit_skb() if
!IFF_XMIT_DST_RELEASE or skb enqueued on qdisc queue, in
sock_queue_rcv_skb(), in __nf_queue().

Use skb_dst_force() in dev_requeue_skb().

Note: dst_use_noref() still dirties dst, we might transform it
later to do one dirtying per jiffies.
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Per cpu variable softnet_data.total was shared between IRQ and SoftIRQ context
without any protection. And enqueue_to_backlog should update the netdev_rx_stat
of the target CPU.

This patch renames softnet_data.total to softnet_data.processed: the number of
packets processed in uppper levels(IP stacks).

softnet_stat data is moved into softnet_data.
Signed-off-by: NChangli Gao <xiaosuo@gmail.com>
----
 include/linux/netdevice.h |   17 +++++++----------
 net/core/dev.c            |   26 ++++++++++++--------------
 net/sched/sch_generic.c   |    2 +-
 3 files changed, 20 insertions(+), 25 deletions(-)
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

One of my test machine got a deadlock during "tc" sessions,
adding/deleting classes & filters, using traffic estimators.

After some analysis, I believe we have a potential use after free case
in est_timer() :

spin_lock(e->stats_lock); << HERE >>
read_lock(&est_lock);
if (e->bstats == NULL)   << TEST >>
	goto skip;

Test is done a bit late, because after estimator is killed, and before
rcu grace period elapsed, we might already have freed/reuse memory where
e->stats_locks points to (some qdisc->q.lock)

A possible fix is to respect a rcu grace period at Qdisc dismantle time.

On 64bit, sizeof(struct Qdisc) is exactly 192 bytes. Adding 16 bytes to
it (for struct rcu_head) is a problem because it might change
performance, given QDISC_ALIGNTO is 32 bytes.

This is why I also change QDISC_ALIGNTO to 64 bytes, to satisfy most
current alignment requirements.
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h

percpu.h is included by sched.h and module.h and thus ends up being
included when building most .c files.  percpu.h includes slab.h which
in turn includes gfp.h making everything defined by the two files
universally available and complicating inclusion dependencies.

percpu.h -> slab.h dependency is about to be removed.  Prepare for
this change by updating users of gfp and slab facilities include those
headers directly instead of assuming availability.  As this conversion
needs to touch large number of source files, the following script is
used as the basis of conversion.

  http://userweb.kernel.org/~tj/misc/slabh-sweep.py

The script does the followings.

* Scan files for gfp and slab usages and update includes such that
  only the necessary includes are there.  ie. if only gfp is used,
  gfp.h, if slab is used, slab.h.

* When the script inserts a new include, it looks at the include
  blocks and try to put the new include such that its order conforms
  to its surrounding.  It's put in the include block which contains
  core kernel includes, in the same order that the rest are ordered -
  alphabetical, Christmas tree, rev-Xmas-tree or at the end if there
  doesn't seem to be any matching order.

* If the script can't find a place to put a new include (mostly
  because the file doesn't have fitting include block), it prints out
  an error message indicating which .h file needs to be added to the
  file.

The conversion was done in the following steps.

1. The initial automatic conversion of all .c files updated slightly
   over 4000 files, deleting around 700 includes and adding ~480 gfp.h
   and ~3000 slab.h inclusions.  The script emitted errors for ~400
   files.

2. Each error was manually checked.  Some didn't need the inclusion,
   some needed manual addition while adding it to implementation .h or
   embedding .c file was more appropriate for others.  This step added
   inclusions to around 150 files.

3. The script was run again and the output was compared to the edits
   from #2 to make sure no file was left behind.

4. Several build tests were done and a couple of problems were fixed.
   e.g. lib/decompress_*.c used malloc/free() wrappers around slab
   APIs requiring slab.h to be added manually.

5. The script was run on all .h files but without automatically
   editing them as sprinkling gfp.h and slab.h inclusions around .h
   files could easily lead to inclusion dependency hell.  Most gfp.h
   inclusion directives were ignored as stuff from gfp.h was usually
   wildly available and often used in preprocessor macros.  Each
   slab.h inclusion directive was examined and added manually as
   necessary.

6. percpu.h was updated not to include slab.h.

7. Build test were done on the following configurations and failures
   were fixed.  CONFIG_GCOV_KERNEL was turned off for all tests (as my
   distributed build env didn't work with gcov compiles) and a few
   more options had to be turned off depending on archs to make things
   build (like ipr on powerpc/64 which failed due to missing writeq).

   * x86 and x86_64 UP and SMP allmodconfig and a custom test config.
   * powerpc and powerpc64 SMP allmodconfig
   * sparc and sparc64 SMP allmodconfig
   * ia64 SMP allmodconfig
   * s390 SMP allmodconfig
   * alpha SMP allmodconfig
   * um on x86_64 SMP allmodconfig

8. percpu.h modifications were reverted so that it could be applied as
   a separate patch and serve as bisection point.

Given the fact that I had only a couple of failures from tests on step
6, I'm fairly confident about the coverage of this conversion patch.
If there is a breakage, it's likely to be something in one of the arch
headers which should be easily discoverable easily on most builds of
the specific arch.
Signed-off-by: NTejun Heo <tj@kernel.org>
Guess-its-ok-by: NChristoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>

Recent changes in the TX error propagation require additional checking
and masking of values returned from hard_start_xmit(), mainly to
separate cases where skb was consumed. This aim can be simplified by
changing the order of NETDEV_TX and NET_XMIT codes, because the latter
are treated similarly to negative (ERRNO) values.

After this change much simpler dev_xmit_complete() is also used in
sch_direct_xmit(), so it is moved to netdevice.h.

Additionally NET_RX definitions in netdevice.h are moved up from
between TX codes to avoid confusion while reading the TX comment.
Signed-off-by: NJarek Poplawski <jarkao2@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Currently the ->ndo_hard_start_xmit() callbacks are only permitted to return
one of the NETDEV_TX codes. This prevents any kind of error propagation for
virtual devices, like queue congestion of the underlying device in case of
layered devices, or unreachability in case of tunnels.

This patches changes the NET_XMIT codes to avoid clashes with the NETDEV_TX
codes and changes the two callers of dev_hard_start_xmit() to expect either
errno codes, NET_XMIT codes or NETDEV_TX codes as return value.

In case of qdisc_restart(), all non NETDEV_TX codes are mapped to NETDEV_TX_OK
since no error propagation is possible when using qdiscs. In case of
dev_queue_xmit(), the error is propagated upwards.
Signed-off-by: NPatrick McHardy <kaber@trash.net>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch adds a classful dummy scheduler which can be used as root qdisc
for multiqueue devices and exposes each device queue as a child class.

This allows to address queues individually and graft them similar to regular
classes. Additionally it presents an accumulated view of the statistics of
all real root qdiscs in the dummy root.

Two new callbacks are added to the qdisc_ops and qdisc_class_ops:

- cl_ops->select_queue selects the tx queue number for new child classes.

- qdisc_ops->attach() overrides root qdisc device grafting to attach
  non-shared qdiscs to the queues.
Signed-off-by: NPatrick McHardy <kaber@trash.net>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

It will be used in a following patch by the multiqueue qdisc.
Signed-off-by: NPatrick McHardy <kaber@trash.net>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Currently the multiqueue integration with the qdisc API suffers from
a few problems:

- with multiple queues, all root qdiscs use the same handle. This means
  they can't be exposed to userspace in a backwards compatible fashion.

- all API operations always refer to queue number 0. Newly created
  qdiscs are automatically shared between all queues, its not possible
  to address individual queues or restore multiqueue behaviour once a
  shared qdisc has been attached.

- Dumps only contain the root qdisc of queue 0, in case of non-shared
  qdiscs this means the statistics are incomplete.

This patch reintroduces dev->qdisc, which points to the (single) root qdisc
from userspace's point of view. Currently it either points to the first
(non-shared) default qdisc, or a qdisc shared between all queues. The
following patches will introduce a classful dummy qdisc, which will be used
as root qdisc and contain the per-queue qdiscs as children.
Signed-off-by: NPatrick McHardy <kaber@trash.net>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

pfifo_fast_enqueue has this check:
        if (skb_queue_len(list) < qdisc_dev(qdisc)->tx_queue_len) {

which allows each band to enqueue upto tx_queue_len skbs for a
total of 3*tx_queue_len skbs. I am not sure if this was the
intention of limiting in qdisc.

Patch compiled and 32 simultaneous netperf testing ran fine. Also:
# tc -s qdisc show dev eth2
qdisc pfifo_fast 0: root bands 3 priomap  1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1
 Sent 16835026752 bytes 373116 pkt (dropped 0, overlimits 0 requeues 25) 
 rate 0bit 0pps backlog 0b 0p requeues 25 
Signed-off-by: NKrishna Kumar <krkumar2@in.ibm.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Maintain a per-qdisc bitmap for pfifo_fast giving  availability
of skbs for each band. This allows faster lookup for a skb when
there are no high priority skbs. Also, it helps in (rare) cases
when there are no skbs on the list, where an immediate lookup is
faster than iterating through the three bands.
Signed-off-by: NKrishna Kumar <krkumar2@in.ibm.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

dev_queue_xmit enqueue's a skb and calls qdisc_run which
dequeue's the skb and xmits it. In most cases, the skb that
is enqueue'd is the same one that is dequeue'd (unless the
queue gets stopped or multiple cpu's write to the same queue
and ends in a race with qdisc_run). For default qdiscs, we
can remove the redundant enqueue/dequeue and simply xmit the
skb since the default qdisc is work-conserving.

The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the
default fast queue. The controversial part of the patch is
incrementing qlen when a skb is requeued - this is to avoid
checks like the second line below:

+  } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
>>         !q->gso_skb &&
+          !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) {

Results of a 2 hour testing for multiple netperf sessions (1,
2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are
aggregate Mb/s across iterations tested with this version on
System-X boxes with Chelsio 10gbps cards:

----------------------------------
Size |  ORG BW          NEW BW   |
----------------------------------
128K |  156964          159381   |
256K |  158650          162042   |
----------------------------------

Changes from ver1:

1. Move sch_direct_xmit declaration from sch_generic.h to
   pkt_sched.h
2. Update qdisc basic statistics for direct xmit path.
3. Set qlen to zero in qdisc_reset.
4. Changed some function names to more meaningful ones.
Signed-off-by: NKrishna Kumar <krkumar2@in.ibm.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

struct net_device trans_start field is a hot spot on SMP and high performance
devices, particularly multi queues ones, because every transmitter dirties
it. Is main use is tx watchdog and bonding alive checks.

But as most devices dont use NETIF_F_LLTX, we have to lock
a netdev_queue before calling their ndo_start_xmit(). So it makes
sense to move trans_start from net_device to netdev_queue. Its update
will occur on a already present (and in exclusive state) cache line, for
free.

We can do this transition smoothly. An old driver continue to
update dev->trans_start, while an updated one updates txq->trans_start.

Further patches could also put tx_bytes/tx_packets counters in 
netdev_queue to avoid dirtying dev->stats (vlan device comes to mind)
Signed-off-by: NEric Dumazet <dada1@cosmosbay.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This reverts commit 22604c86.

We can't fix this issue in this way, because we now can try
to take the dev_base_lock rwlock as a writer in software interrupt
context and that is not allowed without major surgery elsewhere.

This initial link state problem needs to be solved in some other
way.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

From: Michael Marineau <mike@marineau.org>

Commit b4730016 "Do not fire linkwatch
events until the device is registered." was made as a workaround for
drivers that call netif_carrier_off before registering the device.
Unfortunately this causes these drivers to incorrectly report their
link status as IF_OPER_UNKNOWN which can falsely set the IFF_RUNNING
flag when the interface is first brought up. This issues was
previously pointed out[1] but was dismissed saying that IFF_RUNNING is
not related to the link status. From my digging IFF_RUNNING, as
reported to userspace, is based on the link state. It is set based on
__LINK_STATE_START and IF_OPER_UP or IF_OPER_UNKNOWN. See [2], [3],
and [4]. (Whether or not the kernel has IFF_RUNNING set in flags is
not reported to user space so it may well be independent of the link,
I don't know if and when it may get set.)

The end result depends slightly depending on the driver. The the two I
tested were e1000e and b44. With e1000e if the system is booted
without a network cable attached the interface will falsely report
RUNNING when it is brought up causing NetworkManager to attempt to
start it and eventually time out. With b44 when the system is booted
with a network cable attached and brought up with dhcpcd it will time
out the first time.

The attached patch that will still set the operstate variable
correctly to IF_OPER_UP/DOWN/etc when linkwatch_fire_event is called
but then return rather than skipping the linkwatch_fire_event call
entirely as the previous fix did. (sorry it isn't inline, I don't have
a patch friendly email client at the moment)
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch changes the network device internal API to move adminstrative
operations out of the network device structure and into a separate structure.

This patch involves some hackery to maintain compatablity between the
new and old model, so all 300+ drivers don't have to be changed at once.
For drivers that aren't converted yet, the netdevice_ops virt function list
still resides in the net_device structure. For old protocols, the new
net_device_ops are copied out to the old net_device pointers.

After the transistion is completed the nag message can be changed to
an WARN_ON, and the compatiablity code can be made configurable.

Some function pointers aren't moved:
* destructor can't be in net_device_ops because
  it may need to be referenced after the module is unloaded.
* neighbor setup is manipulated in a couple of places that need special
  consideration
* hard_start_xmit is in the fast path for transmit.
Signed-off-by: NStephen Hemminger <shemminger@vyatta.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Several device drivers try to do things like netif_carrier_off()
before register_netdev() is invoked.  This is bogus, but too many
drivers do this to fix them all up in one go.
Reported-by: NFolkert van Heusden <folkert@vanheusden.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

After implementing qdisc->ops->peek() and changing sch_netem into
classless qdisc there are no more qdisc->ops->requeue() users. This
patch removes this method with its wrappers (qdisc_requeue()), and
also unused qdisc->requeue structure. There are a few minor fixes of
warnings (htb_enqueue()) and comments btw.

The idea to kill ->requeue() and a similar patch were first developed
by David S. Miller.
Signed-off-by: NJarek Poplawski <jarkao2@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Since gso_skb is re-used for qdisc_peek_dequeued(), and this skb is
counted in the qdisc->q.qlen, it has to be kfreed during qdisc_reset()
when qlen is zeroed.

With help from David S. Miller <davem@davemloft.net>
Signed-off-by: NJarek Poplawski <jarkao2@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

With feedback from Patrick McHardy.
Signed-off-by: NJarek Poplawski <jarkao2@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>