openwrt/staging/blogic.git
8 years agonet: r6040: add in missing white space in error message text
Colin Ian King [Fri, 16 Sep 2016 09:43:38 +0000 (10:43 +0100)]
net: r6040: add in missing white space in error message text

A couple of dev_err messages span two lines and the literal
string is missing a white space between words. Add the white
space and join the two lines into one.

Signed-off-by: Colin Ian King <colin.king@canonical.com>
Acked-by: FLorian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agopkt_sched: fq: use proper locking in fq_dump_stats()
Eric Dumazet [Thu, 15 Sep 2016 23:20:01 +0000 (16:20 -0700)]
pkt_sched: fq: use proper locking in fq_dump_stats()

When fq is used on 32bit kernels, we need to lock the qdisc before
copying 64bit fields.

Otherwise "tc -s qdisc ..." might report bogus values.

Fixes: afe4fd062416 ("pkt_sched: fq: Fair Queue packet scheduler")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoopenvswitch: use percpu flow stats
Thadeu Lima de Souza Cascardo [Thu, 15 Sep 2016 22:11:53 +0000 (19:11 -0300)]
openvswitch: use percpu flow stats

Instead of using flow stats per NUMA node, use it per CPU. When using
megaflows, the stats lock can be a bottleneck in scalability.

On a E5-2690 12-core system, usual throughput went from ~4Mpps to
~15Mpps when forwarding between two 40GbE ports with a single flow
configured on the datapath.

This has been tested on a system with possible CPUs 0-7,16-23. After
module removal, there were no corruption on the slab cache.

Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@redhat.com>
Cc: pravin shelar <pshelar@ovn.org>
Acked-by: Pravin B Shelar <pshelar@ovn.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoopenvswitch: fix flow stats accounting when node 0 is not possible
Thadeu Lima de Souza Cascardo [Thu, 15 Sep 2016 22:11:52 +0000 (19:11 -0300)]
openvswitch: fix flow stats accounting when node 0 is not possible

On a system with only node 1 as possible, all statistics is going to be
accounted on node 0 as it will have a single writer.

However, when getting and clearing the statistics, node 0 is not going
to be considered, as it's not a possible node.

Tested that statistics are not zero on a system with only node 1
possible. Also compile-tested with CONFIG_NUMA off.

Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@redhat.com>
Acked-by: Pravin B Shelar <pshelar@ovn.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoMerge branch 'sctp-transmit-errs'
David S. Miller [Mon, 19 Sep 2016 02:02:40 +0000 (22:02 -0400)]
Merge branch 'sctp-transmit-errs'

Xin Long says:

====================
sctp: fix the transmit err process

This patchset is to improve the transmit err process and also fix some
issues.

After this patchset, once the chunks are enqueued successfully, even
if the chunks fail to send out, no matter because of nodst or nomem,
no err retruns back to users any more. Instead, they are taken care
of by retransmit.

v1->v2:
  - add more details to the changelog in patch 1/6
  - add Fixes: tag in patch 2/6, 3/6
  - also revert 69b5777f2e57 in patch 3/6
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agosctp: not return ENOMEM err back in sctp_packet_transmit
Xin Long [Tue, 13 Sep 2016 18:04:23 +0000 (02:04 +0800)]
sctp: not return ENOMEM err back in sctp_packet_transmit

As David and Marcelo's suggestion, ENOMEM err shouldn't return back to
user in transmit path. Instead, sctp's retransmit would take care of
the chunks that fail to send because of ENOMEM.

This patch is only to do some release job when alloc_skb fails, not to
return ENOMEM back any more.

Besides, it also cleans up sctp_packet_transmit's err path, and fixes
some issues in err path:

 - It didn't free the head skb in nomem: path.
 - No need to check nskb in no_route: path.
 - It should goto err: path if alloc_skb fails for head.
 - Not all the NOMEMs should free nskb.

Signed-off-by: Xin Long <lucien.xin@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agosctp: make sctp_outq_flush/tail/uncork return void
Xin Long [Tue, 13 Sep 2016 18:04:22 +0000 (02:04 +0800)]
sctp: make sctp_outq_flush/tail/uncork return void

sctp_outq_flush return value is meaningless now, this patch is
to make sctp_outq_flush return void, as well as sctp_outq_fail
and sctp_outq_uncork.

Signed-off-by: Xin Long <lucien.xin@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agosctp: save transmit error to sk_err in sctp_outq_flush
Xin Long [Tue, 13 Sep 2016 18:04:21 +0000 (02:04 +0800)]
sctp: save transmit error to sk_err in sctp_outq_flush

Every time when sctp calls sctp_outq_flush, it sends out the chunks of
control queue, retransmit queue and data queue. Even if some trunks are
failed to transmit, it still has to flush all the transports, as it's
the only chance to clean that transmit_list.

So the latest transmit error here should be returned back. This transmit
error is an internal error of sctp stack.

I checked all the places where it uses the transmit error (the return
value of sctp_outq_flush), most of them are actually just save it to
sk_err.

Except for sctp_assoc/endpoint_bh_rcv, they will drop the chunk if
it's failed to send a REPLY, which is actually incorrect, as we can't
be sure the error that sctp_outq_flush returns is from sending that
REPLY.

So it's meaningless for sctp_outq_flush to return error back.

This patch is to save transmit error to sk_err in sctp_outq_flush, the
new error can update the old value. Eventually, sctp_wait_for_* would
check for it.

Signed-off-by: Xin Long <lucien.xin@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agosctp: free msg->chunks when sctp_primitive_SEND return err
Xin Long [Tue, 13 Sep 2016 18:04:20 +0000 (02:04 +0800)]
sctp: free msg->chunks when sctp_primitive_SEND return err

Last patch "sctp: do not return the transmit err back to sctp_sendmsg"
made sctp_primitive_SEND return err only when asoc state is unavailable.
In this case, chunks are not enqueued, they have no chance to be freed if
we don't take care of them later.

This Patch is actually to revert commit 1cd4d5c4326a ("sctp: remove the
unused sctp_datamsg_free()"), commit 69b5777f2e57 ("sctp: hold the chunks
only after the chunk is enqueued in outq") and commit 8b570dc9f7b6 ("sctp:
only drop the reference on the datamsg after sending a msg"), to use
sctp_datamsg_free to free the chunks of current msg.

Fixes: 8b570dc9f7b6 ("sctp: only drop the reference on the datamsg after sending a msg")
Signed-off-by: Xin Long <lucien.xin@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agosctp: do not return the transmit err back to sctp_sendmsg
Xin Long [Tue, 13 Sep 2016 18:04:19 +0000 (02:04 +0800)]
sctp: do not return the transmit err back to sctp_sendmsg

Once a chunk is enqueued successfully, sctp queues can take care of it.
Even if it is failed to transmit (like because of nomem), it should be
put into retransmit queue.

If sctp report this error to users, it confuses them, they may resend
that msg, but actually in kernel sctp stack is in charge of retransmit
it already.

Besides, this error probably is not from the failure of transmitting
current msg, but transmitting or retransmitting another msg's chunks,
as sctp_outq_flush just tries to send out all transports' chunks.

This patch is to make sctp_cmd_send_msg return avoid, and not return the
transmit err back to sctp_sendmsg

Fixes: 8b570dc9f7b6 ("sctp: only drop the reference on the datamsg after sending a msg")
Signed-off-by: Xin Long <lucien.xin@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agosctp: remove the unnecessary state check in sctp_outq_tail
Xin Long [Tue, 13 Sep 2016 18:04:18 +0000 (02:04 +0800)]
sctp: remove the unnecessary state check in sctp_outq_tail

Data Chunks are only sent by sctp_primitive_SEND, in which sctp checks
the asoc's state through statetable before calling sctp_outq_tail. So
there's no need to check the asoc's state again in sctp_outq_tail.

Besides, sctp_do_sm is protected by lock_sock, even if sending msg is
interrupted by timer events, the event's processes still need to acquire
lock_sock first. It means no others CMDs can be enqueue into side effect
list before CMD_SEND_MSG to change asoc->state, so it's safe to remove it.

This patch is to remove redundant asoc->state check from sctp_outq_tail.

Signed-off-by: Xin Long <lucien.xin@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoMerge branch 'ip_tunnel-collect_md'
David S. Miller [Sat, 17 Sep 2016 14:13:16 +0000 (10:13 -0400)]
Merge branch 'ip_tunnel-collect_md'

Alexei Starovoitov says:

====================
ip_tunnel: add collect_md mode to IPv4/IPv6 tunnels

Similar to geneve, vxlan, gre tunnels implement 'collect metadata' mode
in ipip, ipip6, ip6ip6 tunnels.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agosamples/bpf: add comprehensive ipip, ipip6, ip6ip6 test
Alexei Starovoitov [Thu, 15 Sep 2016 20:00:32 +0000 (13:00 -0700)]
samples/bpf: add comprehensive ipip, ipip6, ip6ip6 test

the test creates 3 namespaces with veth connected via bridge.
First two namespaces simulate two different hosts with the same
IPv4 and IPv6 addresses configured on the tunnel interface and they
communicate with outside world via standard tunnels.
Third namespace creates collect_md tunnel that is driven by BPF
program which selects different remote host (either first or
second namespace) based on tcp dest port number while tcp dst
ip is the same.
This scenario is rough approximation of load balancer use case.
The tests check both traditional tunnel configuration and collect_md mode.

Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agosamples/bpf: extend test_tunnel_bpf.sh with IPIP test
Alexei Starovoitov [Thu, 15 Sep 2016 20:00:31 +0000 (13:00 -0700)]
samples/bpf: extend test_tunnel_bpf.sh with IPIP test

extend existing tests for vxlan, geneve, gre to include IPIP tunnel.
It tests both traditional tunnel configuration and
dynamic via bpf helpers.

Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoip6_tunnel: add collect_md mode to IPv6 tunnels
Alexei Starovoitov [Thu, 15 Sep 2016 20:00:30 +0000 (13:00 -0700)]
ip6_tunnel: add collect_md mode to IPv6 tunnels

Similar to gre, vxlan, geneve tunnels allow IPIP6 and IP6IP6 tunnels
to operate in 'collect metadata' mode.
Unlike ipv4 code here it's possible to reuse ip6_tnl_xmit() function
for both collect_md and traditional tunnels.
bpf_skb_[gs]et_tunnel_key() helpers and ovs (in the future) are the users.

Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Thomas Graf <tgraf@suug.ch>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoip_tunnel: add collect_md mode to IPIP tunnel
Alexei Starovoitov [Thu, 15 Sep 2016 20:00:29 +0000 (13:00 -0700)]
ip_tunnel: add collect_md mode to IPIP tunnel

Similar to gre, vxlan, geneve tunnels allow IPIP tunnels to
operate in 'collect metadata' mode.
bpf_skb_[gs]et_tunnel_key() helpers can make use of it right away.
ovs can use it as well in the future (once appropriate ovs-vport
abstractions and user apis are added).
Note that just like in other tunnels we cannot cache the dst,
since tunnel_info metadata can be different for every packet.

Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Thomas Graf <tgraf@suug.ch>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agol2tp: constify net_device_ops structures
Julia Lawall [Thu, 15 Sep 2016 20:23:26 +0000 (22:23 +0200)]
l2tp: constify net_device_ops structures

Check for net_device_ops structures that are only stored in the netdev_ops
field of a net_device structure.  This field is declared const, so
net_device_ops structures that have this property can be declared as const
also.

The semantic patch that makes this change is as follows:
(http://coccinelle.lip6.fr/)

// <smpl>
@r disable optional_qualifier@
identifier i;
position p;
@@
static struct net_device_ops i@p = { ... };

@ok@
identifier r.i;
struct net_device e;
position p;
@@
e.netdev_ops = &i@p;

@bad@
position p != {r.p,ok.p};
identifier r.i;
struct net_device_ops e;
@@
e@i@p

@depends on !bad disable optional_qualifier@
identifier r.i;
@@
static
+const
 struct net_device_ops i = { ... };
// </smpl>

The result of size on this file before the change is:
   text       data     bss     dec         hex   filename
   3401        931      44    4376        1118 net/l2tp/l2tp_eth.o

and after the change it is:
   text      data        bss     dec     hex filename
   3993       347         44       4384    1120 net/l2tp/l2tp_eth.o

Signed-off-by: Julia Lawall <Julia.Lawall@lip6.fr>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agodwc_eth_qos: constify net_device_ops structures
Julia Lawall [Thu, 15 Sep 2016 20:23:25 +0000 (22:23 +0200)]
dwc_eth_qos: constify net_device_ops structures

Check for net_device_ops structures that are only stored in the netdev_ops
field of a net_device structure.  This field is declared const, so
net_device_ops structures that have this property can be declared as const
also.

The semantic patch that makes this change is as follows:
(http://coccinelle.lip6.fr/)

// <smpl>
@r disable optional_qualifier@
identifier i;
position p;
@@
static struct net_device_ops i@p = { ... };

@ok@
identifier r.i;
struct net_device e;
position p;
@@
e.netdev_ops = &i@p;

@bad@
position p != {r.p,ok.p};
identifier r.i;
struct net_device_ops e;
@@
e@i@p

@depends on !bad disable optional_qualifier@
identifier r.i;
@@
static
+const
 struct net_device_ops i = { ... };
// </smpl>

The result of size on this file before the change is:
   text       data     bss     dec         hex   filename
  21623       1316      40   22979        59c3
   drivers/net/ethernet/synopsys/dwc_eth_qos.o

and after the change it is:
   text      data        bss     dec     hex filename
  22199       724         40      22963    59b3
   drivers/net/ethernet/synopsys/dwc_eth_qos.o

Signed-off-by: Julia Lawall <Julia.Lawall@lip6.fr>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agohisilicon: constify net_device_ops structures
Julia Lawall [Thu, 15 Sep 2016 20:23:24 +0000 (22:23 +0200)]
hisilicon: constify net_device_ops structures

Check for net_device_ops structures that are only stored in the netdev_ops
field of a net_device structure.  This field is declared const, so
net_device_ops structures that have this property can be declared as const
also.

The semantic patch that makes this change is as follows:
(http://coccinelle.lip6.fr/)

// <smpl>
@r disable optional_qualifier@
identifier i;
position p;
@@
static struct net_device_ops i@p = { ... };

@ok@
identifier r.i;
struct net_device e;
position p;
@@
e.netdev_ops = &i@p;

@bad@
position p != {r.p,ok.p};
identifier r.i;
struct net_device_ops e;
@@
e@i@p

@depends on !bad disable optional_qualifier@
identifier r.i;
@@
static
+const
 struct net_device_ops i = { ... };
// </smpl>

The result of size on this file before the change is:

   text       data     bss     dec         hex   filename
   7995        848       8    8851        2293
   drivers/net/ethernet/hisilicon/hip04_eth.o

and after the change it is:

   text      data        bss     dec     hex filename
   8571       256          8       8835    2283
   drivers/net/ethernet/hisilicon/hip04_eth.o

Signed-off-by: Julia Lawall <Julia.Lawall@lip6.fr>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agollc: switch type to bool as the timeout is only tested versus 0
Alan Cox [Thu, 15 Sep 2016 17:51:25 +0000 (18:51 +0100)]
llc: switch type to bool as the timeout is only tested versus 0

(As asked by Dave in Februrary)

Signed-off-by: Alan Cox <alan@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet: l3mdev: Remove netif_index_is_l3_master
David Ahern [Thu, 15 Sep 2016 17:18:45 +0000 (10:18 -0700)]
net: l3mdev: Remove netif_index_is_l3_master

No longer used after e0d56fdd73422 ("net: l3mdev: remove redundant calls")

Signed-off-by: David Ahern <dsa@cumulusnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet: vrf: Remove RT_FL_TOS
David Ahern [Thu, 15 Sep 2016 17:13:47 +0000 (10:13 -0700)]
net: vrf: Remove RT_FL_TOS

No longer used after d66f6c0a8f3c0 ("net: ipv4: Remove l3mdev_get_saddr")

Signed-off-by: David Ahern <dsa@cumulusnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agotcp: prepare skbs for better sack shifting
Eric Dumazet [Thu, 15 Sep 2016 16:33:02 +0000 (09:33 -0700)]
tcp: prepare skbs for better sack shifting

With large BDP TCP flows and lossy networks, it is very important
to keep a low number of skbs in the write queue.

RACK and SACK processing can perform a linear scan of it.

We should avoid putting any payload in skb->head, so that SACK
shifting can be done if needed.

With this patch, we allow to pack ~0.5 MB per skb instead of
the 64KB initially cooked at tcp_sendmsg() time.

This gives a reduction of number of skbs in write queue by eight.
tcp_rack_detect_loss() likes this.

We still allow payload in skb->head for first skb put in the queue,
to not impact RPC workloads.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoMerge tag 'wireless-drivers-next-for-davem-2016-09-15' of git://git.kernel.org/pub...
David S. Miller [Sat, 17 Sep 2016 13:53:29 +0000 (09:53 -0400)]
Merge tag 'wireless-drivers-next-for-davem-2016-09-15' of git://git./linux/kernel/git/kvalo/wireless-drivers-next

Kalle Valo says:

====================
wireless-drivers-next patches for 4.9

Major changes:

iwlwifi

* preparation for new a000 HW continues
* some DQA improvements
* add support for GMAC
* add support for 9460, 9270 and 9170 series

mwifiex

* support random MAC address for scanning
* add HT aggregation support for adhoc mode
* add custom regulatory domain support
* add manufacturing mode support via nl80211 testmode interface

bcma

* support BCM53573 series of wireless SoCs

bitfield.h

* add FIELD_PREP() and FIELD_GET() macros

mt7601u

* convert to use the new bitfield.h macros

brcmfmac

* add support for bcm4339 chip with modalias sdio:c00v02D0d4339

ath10k

* add nl80211 testmode support for 10.4 firmware
* hide kernel addresses from logs using %pK format specifier
* implement NAPI support
* enable peer stats by default

ath9k

* use ieee80211_tx_status_noskb where possible

wil6210

* extract firmware capabilities from the firmware file

ath6kl

* enable firmware crash dumps on the AR6004

ath-current is also merged to fix a conflict in ath10k.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoMerge branch 'mlx5e-order-0'
David S. Miller [Sat, 17 Sep 2016 13:51:48 +0000 (09:51 -0400)]
Merge branch 'mlx5e-order-0'

Tariq Toukan says:

====================
mlx5e Order-0 pages for Striding RQ

In this series, we refactor our Striding RQ receive-flow to always use
fragmented WQEs (Work Queue Elements) using order-0 pages, omitting the
flow that allocates and splits high-order pages which would fragment
and deplete high-order pages in the system.

The first patch gives a slight degradation, but opens the opportunity
to using a simple page-cache mechanism of a fair size.
The page-cache, implemented in patch 3, not only closes the performance
gap but even gives a gain.
In patch 2 we re-organize the code to better manage the calls for
alloc/de-alloc pages in the RX flow.

Series generated against net-next commit:
bed806cb266e "Merge branch 'mlxsw-ethtool'"
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet/mlx5e: Implement RX mapped page cache for page recycle
Tariq Toukan [Thu, 15 Sep 2016 13:08:38 +0000 (16:08 +0300)]
net/mlx5e: Implement RX mapped page cache for page recycle

Instead of reallocating and mapping pages for RX data-path,
recycle already used pages in a per ring cache.

Performance tests:
The following results were measured on a freshly booted system,
giving optimal baseline performance, as high-order pages are yet to
be fragmented and depleted.

We ran pktgen single-stream benchmarks, with iptables-raw-drop:

Single stride, 64 bytes:
* 4,739,057 - baseline
* 4,749,550 - order0 no cache
* 4,786,899 - order0 with cache
1% gain

Larger packets, no page cross, 1024 bytes:
* 3,982,361 - baseline
* 3,845,682 - order0 no cache
* 4,127,852 - order0 with cache
3.7% gain

Larger packets, every 3rd packet crosses a page, 1500 bytes:
* 3,731,189 - baseline
* 3,579,414 - order0 no cache
* 3,931,708 - order0 with cache
5.4% gain

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet/mlx5e: Introduce API for RX mapped pages
Tariq Toukan [Thu, 15 Sep 2016 13:08:37 +0000 (16:08 +0300)]
net/mlx5e: Introduce API for RX mapped pages

Manage the allocation and deallocation of mapped RX pages only
through dedicated API functions.

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet/mlx5e: Single flow order-0 pages for Striding RQ
Tariq Toukan [Thu, 15 Sep 2016 13:08:36 +0000 (16:08 +0300)]
net/mlx5e: Single flow order-0 pages for Striding RQ

To improve the memory consumption scheme, we omit the flow that
demands and splits high-order pages in Striding RQ, and stay
with a single Striding RQ flow that uses order-0 pages.

Moving to fragmented memory allows the use of larger MPWQEs,
which reduces the number of UMR posts and filler CQEs.

Moving to a single flow allows several optimizations that improve
performance, especially in production servers where we would
anyway fallback to order-0 allocations:
- inline functions that were called via function pointers.
- improve the UMR post process.

This patch alone is expected to give a slight performance reduction.
However, the new memory scheme gives the possibility to use a page-cache
of a fair size, that doesn't inflate the memory footprint, which will
dramatically fix the reduction and even give a performance gain.

Performance tests:
The following results were measured on a freshly booted system,
giving optimal baseline performance, as high-order pages are yet to
be fragmented and depleted.

We ran pktgen single-stream benchmarks, with iptables-raw-drop:

Single stride, 64 bytes:
* 4,739,057 - baseline
* 4,749,550 - this patch
no reduction

Larger packets, no page cross, 1024 bytes:
* 3,982,361 - baseline
* 3,845,682 - this patch
3.5% reduction

Larger packets, every 3rd packet crosses a page, 1500 bytes:
* 3,731,189 - baseline
* 3,579,414 - this patch
4% reduction

Fixes: 461017cb006a ("net/mlx5e: Support RX multi-packet WQE (Striding RQ)")
Fixes: bc77b240b3c5 ("net/mlx5e: Add fragmented memory support for RX multi packet WQE")
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agorxrpc: Make IPv6 support conditional on CONFIG_IPV6
David Howells [Sat, 17 Sep 2016 06:26:01 +0000 (07:26 +0100)]
rxrpc: Make IPv6 support conditional on CONFIG_IPV6

Add CONFIG_AF_RXRPC_IPV6 and make the IPv6 support code conditional on it.
This is then made conditional on CONFIG_IPV6.

Without this, the following can be seen:

   net/built-in.o: In function `rxrpc_init_peer':
>> peer_object.c:(.text+0x18c3c8): undefined reference to `ip6_route_output_flags'

Reported-by: kbuild test robot <fengguang.wu@intel.com>
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoMerge branch 'QCA8K'
David S. Miller [Fri, 16 Sep 2016 08:31:56 +0000 (04:31 -0400)]
Merge branch 'QCA8K'

John Crispin says:

====================
net-next: dsa: add QCA8K support

This series is based on the AR8xxx series posted by Matthieu Olivari in may
2015. The following changes were made since then

* fixed the nitpicks from the previous review
* updated to latest API
* turned it into an mdio device
* added callbacks for fdb, bridge offloading, stp, eee, port status
* fixed several minor issues to the port setup and arp learning
* changed the namespacing as this driver to qca8k

The driver has so far only been tested on qca8337/N. It should work on other QCA
switches such as the qca8327 with minor changes.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet-next: dsa: add new driver for qca8xxx family
John Crispin [Thu, 15 Sep 2016 14:26:41 +0000 (16:26 +0200)]
net-next: dsa: add new driver for qca8xxx family

This patch contains initial support for the QCA8337 switch. It
will detect a QCA8337 switch, if present and declared in the DT.

Each port will be represented through a standalone net_device interface,
as for other DSA switches. CPU can communicate with any of the ports by
setting an IP@ on ethN interface. Most of the extra callbacks of the DSA
subsystem are already supported, such as bridge offloading, stp, fdb.

Signed-off-by: John Crispin <john@phrozen.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet-next: dsa: add Qualcomm tag RX/TX handler
John Crispin [Thu, 15 Sep 2016 14:26:40 +0000 (16:26 +0200)]
net-next: dsa: add Qualcomm tag RX/TX handler

Add support for the 2-bytes Qualcomm tag that gigabit switches such as
the QCA8337/N might insert when receiving packets, or that we need
to insert while targeting specific switch ports. The tag is inserted
directly behind the ethernet header.

Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: John Crispin <john@phrozen.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoDocumentation: devicetree: add qca8k binding
John Crispin [Thu, 15 Sep 2016 14:26:39 +0000 (16:26 +0200)]
Documentation: devicetree: add qca8k binding

Add device-tree binding for ar8xxx switch families.

Cc: devicetree@vger.kernel.org
Signed-off-by: John Crispin <john@phrozen.org>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet: emac: remove .owner field for driver
Wei Yongjun [Thu, 15 Sep 2016 02:26:10 +0000 (02:26 +0000)]
net: emac: remove .owner field for driver

Remove .owner field if calls are used which set it automatically.

Generated by: scripts/coccinelle/api/platform_no_drv_owner.cocci

Signed-off-by: Wei Yongjun <weiyongjun1@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet: emac: remove unnecessary dev_set_drvdata()
Wei Yongjun [Thu, 15 Sep 2016 02:25:52 +0000 (02:25 +0000)]
net: emac: remove unnecessary dev_set_drvdata()

The driver core clears the driver data to NULL after device_release
or on probe failure. Thus, it is not needed to manually clear the
device driver data to NULL.

Signed-off-by: Wei Yongjun <weiyongjun1@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet: dsa: b53: Remove unused including <linux/version.h>
Wei Yongjun [Thu, 15 Sep 2016 02:24:37 +0000 (02:24 +0000)]
net: dsa: b53: Remove unused including <linux/version.h>

Remove including <linux/version.h> that don't need it.

Signed-off-by: Wei Yongjun <weiyongjun1@huawei.com>
Acked-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet: dsa: bcm_sf2: Fix non static symbol warning
Wei Yongjun [Thu, 15 Sep 2016 02:24:13 +0000 (02:24 +0000)]
net: dsa: bcm_sf2: Fix non static symbol warning

Fixes the following sparse warning:

drivers/net/dsa/bcm_sf2.c:963:19: warning:
 symbol 'bcm_sf2_io_ops' was not declared. Should it be static?

Signed-off-by: Wei Yongjun <weiyongjun1@huawei.com>
Acked-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agotcp: fix a stale ooo_last_skb after a replace
Eric Dumazet [Wed, 14 Sep 2016 05:55:05 +0000 (22:55 -0700)]
tcp: fix a stale ooo_last_skb after a replace

When skb replaces another one in ooo queue, I forgot to also
update tp->ooo_last_skb as well, if the replaced skb was the last one
in the queue.

To fix this, we simply can re-use the code that runs after an insertion,
trying to merge skbs at the right of current skb.

This not only fixes the bug, but also remove all small skbs that might
be a subset of the new one.

Example:

We receive segments 2001:3001,  4001:5001

Then we receive 2001:8001 : We should replace 2001:3001 with the big
skb, but also remove 4001:50001 from the queue to save space.

packetdrill test demonstrating the bug

0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3
+0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0
+0 bind(3, ..., ...) = 0
+0 listen(3, 1) = 0

+0 < S 0:0(0) win 32792 <mss 1000,sackOK,nop,nop,nop,wscale 7>
+0 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 7>
+0.100 < . 1:1(0) ack 1 win 1024
+0 accept(3, ..., ...) = 4

+0.01 < . 1001:2001(1000) ack 1 win 1024
+0    > . 1:1(0) ack 1 <nop,nop, sack 1001:2001>

+0.01 < . 1001:3001(2000) ack 1 win 1024
+0    > . 1:1(0) ack 1 <nop,nop, sack 1001:2001 1001:3001>

Fixes: 9f5afeae5152 ("tcp: use an RB tree for ooo receive queue")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Yuchung Cheng <ycheng@google.com>
Cc: Yaogong Wang <wygivan@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoMerge branch 'mediatek-reset-flow'
David S. Miller [Fri, 16 Sep 2016 06:23:06 +0000 (02:23 -0400)]
Merge branch 'mediatek-reset-flow'

Sean Wang says:

====================
mediatek: add enhancement into the existing reset flow

Current driver only resets DMA used by descriptor rings which
can't guarantee it can recover all various kinds of fatal
errors, so the patch
1) tries to reset the underlying hardware resource from scratch on
Mediatek SoC required for ethernet running.
2) refactors code in order to the reusability of existing code.
3) considers handling for race condition between the reset flow and
callbacks registered into core driver called about hardware accessing.
4) introduces power domain usage to hardware setup which leads to have
cleanly and completely restore to the state as the initial.

Changes since v1:
- fix the build error with module built causing undefined symbol for
  pinctrl_bind_pins, so using pinctrl_select_state instead accomplishes
  the pin mux setup during the reset process.
====================

Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet: ethernet: mediatek: avoid race condition during the reset process
Sean Wang [Wed, 14 Sep 2016 15:13:21 +0000 (23:13 +0800)]
net: ethernet: mediatek: avoid race condition during the reset process

add the protection of the race condition between
the reset process and hardware access happening
on the related callbacks.

Signed-off-by: Sean Wang <sean.wang@mediatek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet: ethernet: mediatek: add more resets for internal ethernet circuit block
Sean Wang [Wed, 14 Sep 2016 15:13:20 +0000 (23:13 +0800)]
net: ethernet: mediatek: add more resets for internal ethernet circuit block

struct mtk_eth has already contained struct regmap ethsys pointer
to the address range of the internal circuit reset, so we reuse it
to reset more internal blocks on ethernet hardware such as packet
processing engine (PPE) and frame engine (FE) instead of rstc which
deals with FE only.

Signed-off-by: Sean Wang <sean.wang@mediatek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet: ethernet: mediatek: add the whole ethernet reset into the reset process
Sean Wang [Wed, 14 Sep 2016 15:13:19 +0000 (23:13 +0800)]
net: ethernet: mediatek: add the whole ethernet reset into the reset process

1) original driver only resets DMA used by descriptor rings
which can't guarantee it can recover all various kinds of fatal
errors, so the patch tries to reset the underlying hardware
resource from scratch on Mediatek SoC required for ethernet
running, including power, pin mux control, clock and internal
circuits on the ethernet in order to restore into the initial
state which the rebooted machine gives.

2) add state variable inside structure mtk_eth to help distinguish
mtk_hw_init is called between the initialization during boot time
or re-initialization during the reset process.

3) add ge_mode variable inside structure mtk_mac for restoring
the interface mode of the current setup for the target MAC.

4) remove __init attribute from mtk_hw_init definition

Signed-off-by: Sean Wang <sean.wang@mediatek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet: ethernet: mediatek: add controlling power domain the ethernet belongs to
Sean Wang [Wed, 14 Sep 2016 15:13:18 +0000 (23:13 +0800)]
net: ethernet: mediatek: add controlling power domain the ethernet belongs to

introduce power domain control which the digital circuit of
the ethernet belongs to inside the flow of hardware initialization
and deinitialization which helps the entire ethernet hardware block
could restart cleanly and completely as being back to the initial
state when the whole machine reboot.

Signed-off-by: Sean Wang <sean.wang@mediatek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet: ethernet: mediatek: cleanup error path inside mtk_hw_init
Sean Wang [Wed, 14 Sep 2016 15:13:17 +0000 (23:13 +0800)]
net: ethernet: mediatek: cleanup error path inside mtk_hw_init

This cleans up the error path inside mtk_hw_init call, causing it able
to exit appropriately when something fails and also includes refactoring
mtk_cleanup call to make the partial logic reusable on the error path.

Signed-off-by: Sean Wang <sean.wang@mediatek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet: ethernet: mediatek: add mtk_hw_deinit call as the opposite to mtk_hw_init call
Sean Wang [Wed, 14 Sep 2016 15:13:16 +0000 (23:13 +0800)]
net: ethernet: mediatek: add mtk_hw_deinit call as the opposite to mtk_hw_init call

grouping things related to the deinitialization of what
mtk_hw_init call does that help to be reused by the reset
process and the error path handling.

Signed-off-by: Sean Wang <sean.wang@mediatek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet: ethernet: mediatek: refactoring mtk_hw_init to be reused
Sean Wang [Wed, 14 Sep 2016 15:13:15 +0000 (23:13 +0800)]
net: ethernet: mediatek: refactoring mtk_hw_init to be reused

the existing mtk_hw_init includes hardware and software
initialization inside so that it is slightly hard to reuse
them for the process of the reset recovery, so some splitting
is made here for keeping hardware initializing relevant thing
and the else such as IRQ registration and MDIO initialization
what are all about to the interface of core driver moved to the
other proper place because they have no needs to register IRQ and
re-initialize structure again during the reset process.

Signed-off-by: Sean Wang <sean.wang@mediatek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoMerge tag 'rxrpc-rewrite-20160913-2' of git://git.kernel.org/pub/scm/linux/kernel...
David S. Miller [Fri, 16 Sep 2016 05:57:19 +0000 (01:57 -0400)]
Merge tag 'rxrpc-rewrite-20160913-2' of git://git./linux/kernel/git/dhowells/linux-fs

David Howells says:

====================
rxrpc: Support IPv6

Here is a set of patches that add IPv6 support.  They need to be applied on
top of the just-posted miscellaneous fix patches.  They are:

 (1) Make autobinding of an unconnected socket work when sendmsg() is
     called to initiate a client call.

 (2) Don't specify the protocol when creating the client socket, but rather
     take the default instead.

 (3) Use rxrpc_extract_addr_from_skb() in a couple of places that were
     doing the same thing manually.  This allows the IPv6 address
     extraction to be done in fewer places.

 (4) Add IPv6 support.  With this, calls can be made to IPv6 servers from
     userspace AF_RXRPC programs; AFS, however, can't use IPv6 yet as the
     RPC calls need to be upgradeable.
====================

Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoMerge tag 'rxrpc-rewrite-20160913-1' of git://git.kernel.org/pub/scm/linux/kernel...
David S. Miller [Fri, 16 Sep 2016 05:52:20 +0000 (01:52 -0400)]
Merge tag 'rxrpc-rewrite-20160913-1' of git://git./linux/kernel/git/dhowells/linux-fs

David Howells says:

====================
rxrpc: Miscellaneous fixes

Here's a set of miscellaneous fix patches.  There are a couple of points of
note:

 (1) There is one non-fix patch that adjusts the call ref tracking
     tracepoint to make kernel API-held refs on calls more obvious.  This
     is a prerequisite for the patch that fixes prealloc refcounting.

 (2) The final patch alters how jumbo packets that partially exceed the
     receive window are handled.  Previously, space was being left in the
     Rx buffer for them, but this significantly hurts performance as the Rx
     window can't be increased to match the OpenAFS Tx window size.

     Instead, the excess subpackets are discarded and an EXCEEDS_WINDOW ACK
     is generated for the first.  To avoid the problem of someone trying to
     run the kernel out of space by feeding the kernel a series of
     overlapping maximal jumbo packets, we stop allowing jumbo packets on a
     call if we encounter more than three jumbo packets with duplicate or
     excessive subpackets.
====================

Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoMerge branch 'libcxgb-next'
David S. Miller [Fri, 16 Sep 2016 00:49:30 +0000 (20:49 -0400)]
Merge branch 'libcxgb-next'

Varun Prakash says:

====================
iw_cxgb4,cxgbit: remove duplicate code

This patch series removes duplicate code from
iw_cxgb4 and cxgbit by adding common function
definitions in libcxgb.

Please review.
====================

Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agolibcxgb,iw_cxgb4,cxgbit: add cxgb_mk_rx_data_ack()
Varun Prakash [Tue, 13 Sep 2016 15:54:06 +0000 (21:24 +0530)]
libcxgb,iw_cxgb4,cxgbit: add cxgb_mk_rx_data_ack()

Add cxgb_mk_rx_data_ack() to remove duplicate
code to form CPL_RX_DATA_ACK hardware command.

Signed-off-by: Varun Prakash <varun@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agolibcxgb,iw_cxgb4,cxgbit: add cxgb_mk_abort_rpl()
Varun Prakash [Tue, 13 Sep 2016 15:54:05 +0000 (21:24 +0530)]
libcxgb,iw_cxgb4,cxgbit: add cxgb_mk_abort_rpl()

Add cxgb_mk_abort_rpl() to remove duplicate
code to form CPL_ABORT_RPL hardware command.

Signed-off-by: Varun Prakash <varun@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agolibcxgb,iw_cxgb4,cxgbit: add cxgb_mk_abort_req()
Varun Prakash [Tue, 13 Sep 2016 15:54:04 +0000 (21:24 +0530)]
libcxgb,iw_cxgb4,cxgbit: add cxgb_mk_abort_req()

Add cxgb_mk_abort_req() to remove duplicate code
to form CPL_ABORT_REQ hardware command.

Signed-off-by: Varun Prakash <varun@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agolibcxgb, iw_cxgb4, cxgbit: add cxgb_mk_close_con_req()
Varun Prakash [Tue, 13 Sep 2016 15:54:03 +0000 (21:24 +0530)]
libcxgb, iw_cxgb4, cxgbit: add cxgb_mk_close_con_req()

Add cxgb_mk_close_con_req() to remove duplicate
code to form CPL_CLOSE_CON_REQ hardware command.

Signed-off-by: Varun Prakash <varun@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agolibcxgb,iw_cxgb4,cxgbit: add cxgb_mk_tid_release()
Varun Prakash [Tue, 13 Sep 2016 15:54:02 +0000 (21:24 +0530)]
libcxgb,iw_cxgb4,cxgbit: add cxgb_mk_tid_release()

Add cxgb_mk_tid_release() to remove duplicate code
to form CPL_TID_RELEASE hardware command.

Signed-off-by: Varun Prakash <varun@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agolibcxgb,iw_cxgb4,cxgbit: add cxgb_compute_wscale()
Varun Prakash [Tue, 13 Sep 2016 15:54:01 +0000 (21:24 +0530)]
libcxgb,iw_cxgb4,cxgbit: add cxgb_compute_wscale()

Add cxgb_compute_wscale() in libcxgb_cm.h to remove
it's duplicate definitions from cxgb4/cm.c and
cxgbit/cxgbit_cm.c.

Signed-off-by: Varun Prakash <varun@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agolibcxgb,iw_cxgb4,cxgbit: add cxgb_best_mtu()
Varun Prakash [Tue, 13 Sep 2016 15:54:00 +0000 (21:24 +0530)]
libcxgb,iw_cxgb4,cxgbit: add cxgb_best_mtu()

Add cxgb_best_mtu() in libcxgb_cm.h to remove
it's duplicate definitions from cxgb4/cm.c and
cxgbit/cxgbit_cm.c

Signed-off-by: Varun Prakash <varun@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agolibcxgb,iw_cxgb4,cxgbit: add cxgb_is_neg_adv()
Varun Prakash [Tue, 13 Sep 2016 15:53:59 +0000 (21:23 +0530)]
libcxgb,iw_cxgb4,cxgbit: add cxgb_is_neg_adv()

Add cxgb_is_neg_adv() in libcxgb_cm.h to remove
it's duplicate definitions from cxgb4/cm.c and
cxgbit/cxgbit_cm.c.

Signed-off-by: Varun Prakash <varun@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agolibcxgb,iw_cxgb4,cxgbit: add cxgb_find_route6()
Varun Prakash [Tue, 13 Sep 2016 15:53:58 +0000 (21:23 +0530)]
libcxgb,iw_cxgb4,cxgbit: add cxgb_find_route6()

Add cxgb_find_route6() in libcxgb_cm.c to remove
it's duplicate definitions from cxgb4/cm.c and
cxgbit/cxgbit_cm.c.

Signed-off-by: Varun Prakash <varun@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agolibcxgb,iw_cxgb4,cxgbit: add cxgb_find_route()
Varun Prakash [Tue, 13 Sep 2016 15:53:57 +0000 (21:23 +0530)]
libcxgb,iw_cxgb4,cxgbit: add cxgb_find_route()

Add cxgb_find_route() in libcxgb_cm.c to remove
it's duplicate definitions from cxgb4/cm.c and
cxgbit/cxgbit_cm.c.

Signed-off-by: Varun Prakash <varun@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agolibcxgb,iw_cxgb4,cxgbit: add cxgb_get_4tuple()
Varun Prakash [Tue, 13 Sep 2016 15:53:56 +0000 (21:23 +0530)]
libcxgb,iw_cxgb4,cxgbit: add cxgb_get_4tuple()

Add cxgb_get_4tuple() in libcxgb_cm.c to remove
it's duplicate definitions from cxgb4/cm.c and
cxgbit/cxgbit_cm.c.

Signed-off-by: Varun Prakash <varun@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoopenvswitch: avoid deferred execution of recirc actions
Lance Richardson [Tue, 13 Sep 2016 14:08:54 +0000 (10:08 -0400)]
openvswitch: avoid deferred execution of recirc actions

The ovs kernel data path currently defers the execution of all
recirc actions until stack utilization is at a minimum.
This is too limiting for some packet forwarding scenarios due to
the small size of the deferred action FIFO (10 entries). For
example, broadcast traffic sent out more than 10 ports with
recirculation results in packet drops when the deferred action
FIFO becomes full, as reported here:

     http://openvswitch.org/pipermail/dev/2016-March/067672.html

Since the current recursion depth is available (it is already tracked
by the exec_actions_level pcpu variable), we can use it to determine
whether to execute recirculation actions immediately (safe when
recursion depth is low) or defer execution until more stack space is
available.

With this change, the deferred action fifo size becomes a non-issue
for currently failing scenarios because it is no longer used when
there are three or fewer recursions through ovs_execute_actions().

Suggested-by: Pravin Shelar <pshelar@ovn.org>
Signed-off-by: Lance Richardson <lrichard@redhat.com>
Acked-by: Pravin B Shelar <pshelar@ovn.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoMerge branch 'cls_flower-port-masks'
David S. Miller [Fri, 16 Sep 2016 00:27:30 +0000 (20:27 -0400)]
Merge branch 'cls_flower-port-masks'

Or Gerlitz says:

====================
net/sched: cls_flower: Add ports masks

This series adds the ability to specify tcp/udp ports masks
for TC/flower filter matches.

I also removed an unused fields from the flower keys struct
and clarified the format of the recently added vlan attibutes.

v1--> v2 changes:

 * fixes typo in patch #2 title and change log (Sergei)
 * added acks provided by Jiri on v1

FWIW, by mistake the cover letter of V1 (but not the patches)
carried V2 tag, hope this doesn't create too much confusion.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet/sched: cls_flower: Specify vlan attributes format in the UAPI header
Or Gerlitz [Thu, 15 Sep 2016 12:28:24 +0000 (15:28 +0300)]
net/sched: cls_flower: Specify vlan attributes format in the UAPI header

Specify the format (size and endianess) for the vlan attributes.

Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet/sched: cls_flower: Remove an unused field from the filter key structure
Or Gerlitz [Thu, 15 Sep 2016 12:28:23 +0000 (15:28 +0300)]
net/sched: cls_flower: Remove an unused field from the filter key structure

Commit c3f8324188fa "net: Add full IPv6 addresses to flow_keys" added an
unused instance of struct flow_dissector_key_addrs into struct fl_flow_key,
remove it.

Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Reported-by: Hadar Hen Zion <hadarh@mellanox.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet/sched: cls_flower: Support masking for matching on tcp/udp ports
Or Gerlitz [Thu, 15 Sep 2016 12:28:22 +0000 (15:28 +0300)]
net/sched: cls_flower: Support masking for matching on tcp/udp ports

Add the definitions for src/dst udp/tcp port masks and use
them when setting && dumping the relevant keys.

Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Paul Blakey <paulb@mellanox.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoalx: fix error handling in __alx_open
Tobias Regnery [Tue, 13 Sep 2016 10:06:57 +0000 (12:06 +0200)]
alx: fix error handling in __alx_open

In commit 9ee7b683ea63 we moved the enablement of msi interrupts earlier in
alx_init_intr. If there is an error in alx_alloc_rings, __alx_open returns
with an error but msi (or msi-x) interrupts stays enabled. Add a new error
label to disable msi (or msi-x) interrupts.

Fixes: 9ee7b683ea63 ("alx: refactor msi enablement and disablement")
Signed-off-by: Tobias Regnery <tobias.regnery@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agocxgb4vf: don't offload Rx checksums for IPv6 fragments
Hariprasad Shenai [Tue, 13 Sep 2016 08:09:24 +0000 (13:39 +0530)]
cxgb4vf: don't offload Rx checksums for IPv6 fragments

The checksum provided by the device doesn't include the L3 headers,
as IPv6 expects

Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet_sched: Introduce skbmod action
Jamal Hadi Salim [Tue, 13 Sep 2016 00:13:09 +0000 (20:13 -0400)]
net_sched: Introduce skbmod action

This action is intended to be an upgrade from a usability perspective
from pedit (as well as operational debugability).
Compare this:

sudo tc filter add dev $ETH parent 1: protocol ip prio 10 \
u32 match ip protocol 1 0xff flowid 1:2 \
action pedit munge offset -14 u8 set 0x02 \
munge offset -13 u8 set 0x15 \
munge offset -12 u8 set 0x15 \
munge offset -11 u8 set 0x15 \
munge offset -10 u16 set 0x1515 \
pipe

to:

sudo tc filter add dev $ETH parent 1: protocol ip prio 10 \
u32 match ip protocol 1 0xff flowid 1:2 \
action skbmod dmac 02:15:15:15:15:15

Also try to do a MAC address swap with pedit or worse
try to debug a policy with destination mac, source mac and
etherype. Then make few rules out of those and you'll get my point.

In the future common use cases on pedit can be migrated to this action
(as an example different fields in ip v4/6, transports like tcp/udp/sctp
etc). For this first cut, this allows modifying basic ethernet header.

The most important ethernet use case at the moment is when redirecting or
mirroring packets to a remote machine. The dst mac address needs a re-write
so that it doesnt get dropped or confuse an interconnecting (learning) switch
or dropped by a target machine (which looks at the dst mac). And at times
when flipping back the packet a swap of the MAC addresses is needed.

Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoMerge branch 'bpf-next'
David S. Miller [Thu, 15 Sep 2016 23:29:53 +0000 (19:29 -0400)]
Merge branch 'bpf-next'

Daniel Borkmann says:

====================
Misc cls_bpf/act_bpf improvements

Two minor improvements to {cls,act}_bpf. For details please see
individual patches.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agobpf: use skb_at_tc_ingress helper in tcf_bpf
Daniel Borkmann [Mon, 12 Sep 2016 21:38:43 +0000 (23:38 +0200)]
bpf: use skb_at_tc_ingress helper in tcf_bpf

We have a small skb_at_tc_ingress() helper for testing for ingress, so
make use of it. cls_bpf already uses it and so should act_bpf.

Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agobpf: drop unnecessary test in cls_bpf_classify and tcf_bpf
Daniel Borkmann [Mon, 12 Sep 2016 21:38:42 +0000 (23:38 +0200)]
bpf: drop unnecessary test in cls_bpf_classify and tcf_bpf

The skb_mac_header_was_set() test in cls_bpf's and act_bpf's fast-path is
actually unnecessary and can be removed altogether. This was added by
commit a166151cbe33 ("bpf: fix bpf helpers to use skb->mac_header relative
offsets"), which was later on improved by 3431205e0397 ("bpf: make programs
see skb->data == L2 for ingress and egress"). We're always guaranteed to
have valid mac header at the time we invoke cls_bpf_classify() or tcf_bpf().

Reason is that since 6d1ccff62780 ("net: reset mac header in dev_start_xmit()")
we do skb_reset_mac_header() in __dev_queue_xmit() before we could call
into sch_handle_egress() or any subsequent enqueue. sch_handle_ingress()
always sees a valid mac header as well (things like skb_reset_mac_len()
would badly fail otherwise). Thus, drop the unnecessary test in classifier
and action case.

Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet/sched: act_tunnel_key: Remove rcu_read_lock protection
Hadar Hen Zion [Mon, 12 Sep 2016 12:19:21 +0000 (15:19 +0300)]
net/sched: act_tunnel_key: Remove rcu_read_lock protection

Remove rcu_read_lock protection from tunnel_key_dump and use
rtnl_dereference, dump operation is protected by  rtnl lock.

Also, remove rcu_read_lock from tunnel_key_release and use
rcu_dereference_protected.

Both operations are running exclusively and a writer couldn't modify
t->params while those functions are executed.

Fixes: 54d94fd89d90 ('net/sched: Introduce act_tunnel_key')
Signed-off-by: Hadar Hen Zion <hadarh@mellanox.com>
Acked-by: John Fastabend <john.r.fastabend@intel.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agotest_bpf: fix the dummy skb after dissector changes
Jakub Kicinski [Mon, 12 Sep 2016 12:04:57 +0000 (13:04 +0100)]
test_bpf: fix the dummy skb after dissector changes

Commit d5709f7ab776 ("flow_dissector: For stripped vlan, get vlan
info from skb->vlan_tci") made flow dissector look at vlan_proto
when vlan is present.  Since test_bpf sets skb->vlan_tci to ~0
(including VLAN_TAG_PRESENT) we have to populate skb->vlan_proto.

Fixes false negative on test #24:
test_bpf: #24 LD_PAYLOAD_OFF jited:0 175 ret 0 != 42 FAIL (1 times)

Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dinan Gunawardena <dinan.gunawardena@netronome.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoatm: iphase: fix newline escape and minor tweak to source formatting
Colin Ian King [Mon, 12 Sep 2016 12:01:50 +0000 (13:01 +0100)]
atm: iphase: fix newline escape and minor tweak to source formatting

The newline escape is incorrect and needs fixing. Also adjust source
formatting / indentation and add { } to trailing else.

Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agomwifiex: firmware name correction for usb8997 chipset
Ganapathi Bhat [Mon, 12 Sep 2016 13:25:28 +0000 (18:55 +0530)]
mwifiex: firmware name correction for usb8997 chipset

Similar to pcie8997 chipset, first firmware submitted for usb8997
chipset will be usbusb8997_combo_v4.bin. This patch corrects the
name used in driver.

Signed-off-by: Ganapathi Bhat <gbhat@marvell.com>
Signed-off-by: Amitkumar Karwar <akarwar@marvell.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
8 years agomwifiex: Command 7 handling for USB chipsets
Ganapathi Bhat [Mon, 12 Sep 2016 13:25:27 +0000 (18:55 +0530)]
mwifiex: Command 7 handling for USB chipsets

Firmware image for newer USB chipsets starts with a command 7 block
(special command). It doesn't contain data length field. This patch adds
necessary handling.

Signed-off-by: Ganapathi Bhat <gbhat@marvell.com>
Signed-off-by: Amitkumar Karwar <akarwar@marvell.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
8 years agortlwifi: rtl818x: constify local structures
Julia Lawall [Sun, 11 Sep 2016 13:06:03 +0000 (15:06 +0200)]
rtlwifi: rtl818x: constify local structures

For structure types defined in the same file or local header files, find
top-level static structure declarations that have the following
properties:
1. Never reassigned.
2. Address never taken
3. Not passed to a top-level macro call
4. No pointer or array-typed field passed to a function or stored in a
variable.
Declare structures having all of these properties as const.

Done using Coccinelle.
Based on a suggestion by Joe Perches <joe@perches.com>.

Signed-off-by: Julia Lawall <Julia.Lawall@lip6.fr>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
8 years agoiwlegacy: constify local structures
Julia Lawall [Sun, 11 Sep 2016 13:05:50 +0000 (15:05 +0200)]
iwlegacy: constify local structures

For structure types defined in the same file or local header files, find
top-level static structure declarations that have the following
properties:
1. Never reassigned.
2. Address never taken
3. Not passed to a top-level macro call
4. No pointer or array-typed field passed to a function or stored in a
variable.
Declare structures having all of these properties as const.

Done using Coccinelle.
Based on a suggestion by Joe Perches <joe@perches.com>.

Signed-off-by: Julia Lawall <Julia.Lawall@lip6.fr>
Acked-by: Stanislaw Gruszka <sgruszka@redhat.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
8 years agoath: constify local structures
Julia Lawall [Sun, 11 Sep 2016 13:05:48 +0000 (15:05 +0200)]
ath: constify local structures

For structure types defined in the same file or local header files, find
top-level static structure declarations that have the following
properties:
1. Never reassigned.
2. Address never taken
3. Not passed to a top-level macro call
4. No pointer or array-typed field passed to a function or stored in a
variable.
Declare structures having all of these properties as const.

Done using Coccinelle.
Based on a suggestion by Joe Perches <joe@perches.com>.

Signed-off-by: Julia Lawall <Julia.Lawall@lip6.fr>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
8 years agortl8xxxu: fix spelling mistake "firmare" -> "firmware"
Colin Ian King [Fri, 9 Sep 2016 18:01:25 +0000 (14:01 -0400)]
rtl8xxxu: fix spelling mistake "firmare" -> "firmware"

Trivial fix to spelling mistakes in dev_dbg message.

Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Jes Sorensen <Jes.Sorensen@redhat.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
8 years agortl8xxxu: Reset device on module unload if still attached
Jes Sorensen [Fri, 9 Sep 2016 18:01:24 +0000 (14:01 -0400)]
rtl8xxxu: Reset device on module unload if still attached

If the USB dongle is still attached, reset it on module unload to
avoid scans failing when reloading the driver.

Signed-off-by: Jes Sorensen <Jes.Sorensen@redhat.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
8 years agomwifiex: correction in Rx STBC field of htcapinfo
Amitkumar Karwar [Fri, 9 Sep 2016 14:56:19 +0000 (20:26 +0530)]
mwifiex: correction in Rx STBC field of htcapinfo

Currently Rx STBC in assoc request frame is advertised as 3. It should
be 2, as our chipsets support two spatial streams.

Signed-off-by: Amitkumar Karwar <akarwar@marvell.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
8 years agomwifiex: handle error if IRQ request fails in mwifiex_sdio_of()
Amitkumar Karwar [Fri, 9 Sep 2016 10:57:58 +0000 (16:27 +0530)]
mwifiex: handle error if IRQ request fails in mwifiex_sdio_of()

When this failure occurs, we will clear card->plt_wake_cfg so that
device would initialize without wake up on external interrupt feature.
This feature specific code in suspend and resume handlers will be
skipped.

Signed-off-by: Amitkumar Karwar <akarwar@marvell.com>
Reviewed-by: Javier Martinez Canillas <javier@osg.samsung.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
8 years agoMerge ath-next from git://git.kernel.org/pub/scm/linux/kernel/git/kvalo/ath.git
Kalle Valo [Wed, 14 Sep 2016 16:34:50 +0000 (19:34 +0300)]
Merge ath-next from git://git./linux/kernel/git/kvalo/ath.git

ath.git patches for 4.9. Major changes:

ath10k

* add nl80211 testmode support for 10.4 firmware
* hide kernel addresses from logs using %pK format specifier
* implement NAPI support
* enable peer stats by default

ath9k

* use ieee80211_tx_status_noskb where possible

wil6210

* extract firmware capabilities from the firmware file

ath6kl

* enable firmware crash dumps on the AR6004

ath-current is also merged to fix a conflict in ath10k.

8 years agorxrpc: Add IPv6 support
David Howells [Tue, 13 Sep 2016 07:49:05 +0000 (08:49 +0100)]
rxrpc: Add IPv6 support

Add IPv6 support to AF_RXRPC.  With this, AF_RXRPC sockets can be created:

service = socket(AF_RXRPC, SOCK_DGRAM, PF_INET6);

instead of:

service = socket(AF_RXRPC, SOCK_DGRAM, PF_INET);

The AFS filesystem doesn't support IPv6 at the moment, though, since that
requires upgrades to some of the RPC calls.

Note that a good portion of this patch is replacing "%pI4:%u" in print
statements with "%pISpc" which is able to handle both protocols and print
the port.

Signed-off-by: David Howells <dhowells@redhat.com>
8 years agorxrpc: Use rxrpc_extract_addr_from_skb() rather than doing this manually
David Howells [Tue, 13 Sep 2016 07:49:05 +0000 (08:49 +0100)]
rxrpc: Use rxrpc_extract_addr_from_skb() rather than doing this manually

There are two places that want to transmit a packet in response to one just
received and manually pick the address to reply to out of the sk_buff.
Make them use rxrpc_extract_addr_from_skb() instead so that IPv6 is handled
automatically.

Signed-off-by: David Howells <dhowells@redhat.com>
8 years agorxrpc: Don't specify protocol to when creating transport socket
David Howells [Tue, 13 Sep 2016 07:49:05 +0000 (08:49 +0100)]
rxrpc: Don't specify protocol to when creating transport socket

Pass 0 as the protocol argument when creating the transport socket rather
than IPPROTO_UDP.

Signed-off-by: David Howells <dhowells@redhat.com>
8 years agorxrpc: Create an address for sendmsg() to bind unbound socket with
David Howells [Tue, 13 Sep 2016 07:49:05 +0000 (08:49 +0100)]
rxrpc: Create an address for sendmsg() to bind unbound socket with

Create an address for sendmsg() to bind unbound socket with rather than
using a completely blank address otherwise the transport socket creation
will fail because it will try to use address family 0.

We use the address family specified in the protocol argument when the
AF_RXRPC socket was created and SOCK_DGRAM as the default.  For anything
else, bind() must be used.

Signed-off-by: David Howells <dhowells@redhat.com>
8 years agorxrpc: Correctly initialise, limit and transmit call->rx_winsize
David Howells [Tue, 13 Sep 2016 21:36:22 +0000 (22:36 +0100)]
rxrpc: Correctly initialise, limit and transmit call->rx_winsize

call->rx_winsize should be initialised to the sysctl setting and the sysctl
setting should be limited to the maximum we want to permit.  Further, we
need to place this in the ACK info instead of the sysctl setting.

Furthermore, discard the idea of accepting the subpackets of a jumbo packet
that lie beyond the receive window when the first packet of the jumbo is
within the window.  Just discard the excess subpackets instead.  This
allows the receive window to be opened up right to the buffer size less one
for the dead slot.

Signed-off-by: David Howells <dhowells@redhat.com>
8 years agorxrpc: Fix prealloc refcounting
David Howells [Tue, 13 Sep 2016 08:05:14 +0000 (09:05 +0100)]
rxrpc: Fix prealloc refcounting

The preallocated call buffer holds a ref on the calls within that buffer.
The ref was being released in the wrong place - it worked okay for incoming
calls to the AFS cache manager service, but doesn't work right for incoming
calls to a userspace service.

Instead of releasing an extra ref service calls in rxrpc_release_call(),
the ref needs to be released during the acceptance/rejectance process.  To
this end:

 (1) The prealloc ref is now normally released during
     rxrpc_new_incoming_call().

 (2) For preallocated kernel API calls, the kernel API's ref needs to be
     released when the call is discarded on socket close.

 (3) We shouldn't take a second ref in rxrpc_accept_call().

 (4) rxrpc_recvmsg_new_call() needs to get a ref of its own when it adds
     the call to the to_be_accepted socket queue.

In doing (4) above, we would prefer not to put the call's refcount down to
0 as that entails doing cleanup in softirq context, but it's unlikely as
there are several refs held elsewhere, at least one of which must be put by
someone in process context calling rxrpc_release_call().  However, it's not
a problem if we do have to do that.

Signed-off-by: David Howells <dhowells@redhat.com>
8 years agorxrpc: Adjust the call ref tracepoint to show kernel API refs
David Howells [Tue, 13 Sep 2016 08:12:34 +0000 (09:12 +0100)]
rxrpc: Adjust the call ref tracepoint to show kernel API refs

Adjust the call ref tracepoint to show references held on a call by the
kernel API separately as much as possible and add an additional trace to at
the allocation point from the preallocation buffer for an incoming call.

Note that this doesn't show the allocation of a client call for the kernel
separately at the moment.

Signed-off-by: David Howells <dhowells@redhat.com>
8 years agorxrpc: Allow tx_winsize to grow in response to an ACK
David Howells [Tue, 13 Sep 2016 09:23:01 +0000 (10:23 +0100)]
rxrpc: Allow tx_winsize to grow in response to an ACK

Allow tx_winsize to grow when the ACK info packet shows a larger receive
window at the other end rather than only permitting it to shrink.

Signed-off-by: David Howells <dhowells@redhat.com>
8 years agorxrpc: Use skb->len not skb->data_len
David Howells [Tue, 13 Sep 2016 21:36:22 +0000 (22:36 +0100)]
rxrpc: Use skb->len not skb->data_len

skb->len should be used rather than skb->data_len when referring to the
amount of data in a packet.  This will only cause a malfunction in the
following cases:

 (1) We receive a jumbo packet (validation and splitting both are wrong).

 (2) We see if there's extra ACK info in an ACK packet (we think it's not
     there and just ignore it).

Signed-off-by: David Howells <dhowells@redhat.com>
8 years agorxrpc: Add missing unlock in rxrpc_call_accept()
David Howells [Tue, 13 Sep 2016 21:36:22 +0000 (22:36 +0100)]
rxrpc: Add missing unlock in rxrpc_call_accept()

Add a missing unlock in rxrpc_call_accept() in the path taken if there's no
call to wake up.

Signed-off-by: David Howells <dhowells@redhat.com>
8 years agorxrpc: Requeue call for recvmsg if more data
David Howells [Tue, 13 Sep 2016 21:36:21 +0000 (22:36 +0100)]
rxrpc: Requeue call for recvmsg if more data

rxrpc_recvmsg() needs to make sure that the call it has just been
processing gets requeued for further attention if the buffer has been
filled and there's more data to be consumed.  The softirq producer only
queues the call and wakes the socket if it fills the first slot in the
window, so userspace might end up sleeping forever otherwise, despite there
being data available.

This is not a problem provided the userspace buffer is big enough or it
empties the buffer completely before more data comes in.

Signed-off-by: David Howells <dhowells@redhat.com>
8 years agorxrpc: The IDLE ACK packet should use rxrpc_idle_ack_delay
David Howells [Tue, 13 Sep 2016 21:36:21 +0000 (22:36 +0100)]
rxrpc: The IDLE ACK packet should use rxrpc_idle_ack_delay

The IDLE ACK packet should use the rxrpc_idle_ack_delay setting when the
timer is set for it.

Signed-off-by: David Howells <dhowells@redhat.com>
8 years agorxrpc: Add missing wakeup on Tx window rotation
David Howells [Tue, 13 Sep 2016 21:36:21 +0000 (22:36 +0100)]
rxrpc: Add missing wakeup on Tx window rotation

We need to wake up the sender when Tx window rotation due to an incoming
ACK makes space in the buffer otherwise the sender is liable to just hang
endlessly.

This problem isn't noticeable if the Tx phase transfers no more than will
fit in a single window or the Tx window rotates fast enough that it doesn't
get full.

Signed-off-by: David Howells <dhowells@redhat.com>
8 years agorxrpc: Make sure we initialise the peer hash key
David Howells [Tue, 13 Sep 2016 21:36:21 +0000 (22:36 +0100)]
rxrpc: Make sure we initialise the peer hash key

Peer records created for incoming connections weren't getting their hash
key set.  This meant that incoming calls wouldn't see more than one DATA
packet - which is not a problem for AFS CM calls with small request data
blobs.

Signed-off-by: David Howells <dhowells@redhat.com>
8 years agoMerge branch 'mlxsw-ethtool'
David S. Miller [Tue, 13 Sep 2016 16:16:44 +0000 (12:16 -0400)]
Merge branch 'mlxsw-ethtool'

Jiri Pirko says:

====================
mlxsw: ethtool enhancements

Ido says:

Patches 1-4 do some minor cleanup in current ethtool ops. Patch 5
replace legacy {get,set}_settings callbacks with
{get,set}_link_ksettings.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agomlxsw: spectrum: Add support for new ethtool API
Ido Schimmel [Mon, 12 Sep 2016 11:26:27 +0000 (13:26 +0200)]
mlxsw: spectrum: Add support for new ethtool API

Remove the deprecated {get,set}_settings callbacks and instead add
{get,set}_link_ksettings along with support for newly available speeds.

Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>