GitHub/moto-9609/android_kernel_motorola_exynos9610.git
10 years agonet: sched: suspicious RCU usage in qdisc_watchdog
John Fastabend [Fri, 3 Oct 2014 05:43:09 +0000 (22:43 -0700)]
net: sched: suspicious RCU usage in qdisc_watchdog

Suspicious RCU usage in qdisc_watchdog call needs to be done inside
rcu_read_lock/rcu_read_unlock. And then Qdisc destroy operations
need to ensure timer is cancelled before removing qdisc structure.

[ 3992.191339] ===============================
[ 3992.191340] [ INFO: suspicious RCU usage. ]
[ 3992.191343] 3.17.0-rc6net-next+ #72 Not tainted
[ 3992.191345] -------------------------------
[ 3992.191347] include/net/sch_generic.h:272 suspicious rcu_dereference_check() usage!
[ 3992.191348]
[ 3992.191348] other info that might help us debug this:
[ 3992.191348]
[ 3992.191351]
[ 3992.191351] rcu_scheduler_active = 1, debug_locks = 1
[ 3992.191353] no locks held by swapper/1/0.
[ 3992.191355]
[ 3992.191355] stack backtrace:
[ 3992.191358] CPU: 1 PID: 0 Comm: swapper/1 Not tainted 3.17.0-rc6net-next+ #72
[ 3992.191360] Hardware name:                  /DZ77RE-75K, BIOS GAZ7711H.86A.0060.2012.1115.1750 11/15/2012
[ 3992.191362]  0000000000000001 ffff880235803e48 ffffffff8178f92c 0000000000000000
[ 3992.191366]  ffff8802322224a0 ffff880235803e78 ffffffff810c9966 ffff8800a5fe3000
[ 3992.191370]  ffff880235803f30 ffff8802359cd768 ffff8802359cd6e0 ffff880235803e98
[ 3992.191374] Call Trace:
[ 3992.191376]  <IRQ>  [<ffffffff8178f92c>] dump_stack+0x4e/0x68
[ 3992.191387]  [<ffffffff810c9966>] lockdep_rcu_suspicious+0xe6/0x130
[ 3992.191392]  [<ffffffff8167213a>] qdisc_watchdog+0x8a/0xb0
[ 3992.191396]  [<ffffffff810f93f2>] __run_hrtimer+0x72/0x420
[ 3992.191399]  [<ffffffff810f9bcd>] ? hrtimer_interrupt+0x7d/0x240
[ 3992.191403]  [<ffffffff816720b0>] ? tc_classify+0xc0/0xc0
[ 3992.191406]  [<ffffffff810f9c4f>] hrtimer_interrupt+0xff/0x240
[ 3992.191410]  [<ffffffff8109e4a5>] ? __atomic_notifier_call_chain+0x5/0x140
[ 3992.191415]  [<ffffffff8103577b>] local_apic_timer_interrupt+0x3b/0x60
[ 3992.191419]  [<ffffffff8179c2b5>] smp_apic_timer_interrupt+0x45/0x60
[ 3992.191422]  [<ffffffff8179a6bf>] apic_timer_interrupt+0x6f/0x80
[ 3992.191424]  <EOI>  [<ffffffff815ed233>] ? cpuidle_enter_state+0x73/0x2e0
[ 3992.191432]  [<ffffffff815ed22e>] ? cpuidle_enter_state+0x6e/0x2e0
[ 3992.191437]  [<ffffffff815ed567>] cpuidle_enter+0x17/0x20
[ 3992.191441]  [<ffffffff810c0741>] cpu_startup_entry+0x3d1/0x4a0
[ 3992.191445]  [<ffffffff81106fc6>] ? clockevents_config_and_register+0x26/0x30
[ 3992.191448]  [<ffffffff81033c16>] start_secondary+0x1b6/0x260

Fixes: b26b0d1e8b1 ("net: qdisc: use rcu prefix and silence sparse warnings")
Signed-off-by: John Fastabend <john.r.fastabend@intel.com>
Acked-by: Cong Wang <cwang@twopensource.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: dsa: do not call phy_start_aneg
Florian Fainelli [Fri, 3 Oct 2014 01:56:03 +0000 (18:56 -0700)]
net: dsa: do not call phy_start_aneg

Commit f7f1de51edbd ("net: dsa: start and stop the PHY state machine")
add calls to phy_start() in dsa_slave_open() respectively phy_stop() in
dsa_slave_close().

We also call phy_start_aneg() in dsa_slave_create(), and this call is
messing up with the PHY state machine, since we basically start the
auto-negotiation, and later on restart it when calling phy_start().
phy_start() does not currently handle the PHY_FORCING or PHY_AN states
properly, but such a fix would be too invasive for this window.

Fixes: f7f1de51edbd ("net: dsa: start and stop the PHY state machine")
Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoRemoved unused inet6 address state
Sébastien Barré [Thu, 2 Oct 2014 19:15:22 +0000 (21:15 +0200)]
Removed unused inet6 address state

the inet6 state INET6_IFADDR_STATE_UP only appeared in its definition.

Cc: Christoph Paasch <christoph.paasch@uclouvain.be>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sébastien Barré <sebastien.barre@uclouvain.be>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: Cleanup skb cloning by adding SKB_FCLONE_FREE
Vijay Subramanian [Thu, 2 Oct 2014 17:00:43 +0000 (10:00 -0700)]
net: Cleanup skb cloning by adding SKB_FCLONE_FREE

SKB_FCLONE_UNAVAILABLE has overloaded meaning depending on type of skb.
1: If skb is allocated from head_cache, it indicates fclone is not available.
2: If skb is a companion fclone skb (allocated from fclone_cache), it indicates
it is available to be used.

To avoid confusion for case 2 above, this patch  replaces
SKB_FCLONE_UNAVAILABLE with SKB_FCLONE_FREE where appropriate. For fclone
companion skbs, this indicates it is free for use.

SKB_FCLONE_UNAVAILABLE will now simply indicate skb is from head_cache and
cannot / will not have a companion fclone.

Signed-off-by: Vijay Subramanian <subramanian.vijay@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agomlx4: add a new xmit_more counter
Eric Dumazet [Thu, 2 Oct 2014 15:24:21 +0000 (08:24 -0700)]
mlx4: add a new xmit_more counter

ethtool -S reports a new counter, tracking number of time doorbell
was not triggered, because skb->xmit_more was set.

$ ethtool -S eth0 | egrep "tx_packet|xmit_more"
     tx_packets: 2413288400
     xmit_more: 666121277

I merged the tso_packet false sharing avoidance in this patch as well.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge branch 'gudp'
David S. Miller [Fri, 3 Oct 2014 23:53:36 +0000 (16:53 -0700)]
Merge branch 'gudp'

Tom Herbert says:

====================
net: Generic UDP Encapsulation

Generic UDP Encapsulation (GUE) is UDP encapsulation protocol which
encapsulates packets of various IP protocols. The GUE protocol is
described in http://tools.ietf.org/html/draft-herbert-gue-01.

The receive path of GUE is implemented in the FOU over UDP module (FOU).
This includes a UDP encap receive function for GUE as well as GUE
specific GRO functions. Management and configuration of GUE ports shares
most of the same code with FOU.

For the transmit path, the previous FOU support for IPIP, sit, and GRE
was simply extended for GUE (when GUE is enabled insert the GUE
header on transmit in addition to UDP header inserted for FOU).

Semantically GUE is the same as FOU in that the encapsulation (UDP
and GUE headers) that are inserted on transmission and removed on
reception so that IP packet is processed with the inner header.

This patch set includes:
 - Some fixes to FOU, removal of IPv4,v6 specific GRO functions
 - Support to configure a GUE receive port
 - Implementation of GUE receive path (normal and GRO)
 - Additions to ip_tunnel netlink to configure GUE
 - GUE header inserion in ip_tunnel transmit path

v2:
 - Include net/gue.h in patch set

Testing:

I ran performance numbers using netperf TCP_RR with 200 streams,
comparing encapsulation without GUE, encapsulation with GUE, and
encapsulation with FOU.

 GRE
    TCP_STREAM
      IPv4, FOU, UDP checksum enabled
        14.04% TX CPU utilization
        13.17% RX CPU utilization
        9211 Mbps
      IPv4, GUE, UDP checksum enabled
        14.99% TX CPU utilization
        13.79% RX CPU utilization
        9185 Mbps
      IPv4, FOU, UDP checksum disabled
        13.14% TX CPU utilization
        23.18% RX CPU utilization
        9277 Mbps
      IPv4, GUE, UDP checksum disabled
        13.66% TX CPU utilization
        23.57% RX CPU utilization
        9184 Mbps
    TCP_RR
      IPv4, FOU, UDP checksum enabled
        94.2% CPU utilization
        155/249/460 90/95/99% latencies
        1.17018e+06 tps
      IPv4, GUE, UDP checksum enabled
        93.9% CPU utilization
        158/253/472 90/95/99% latencies
        1.15045e+06 tps

  IPIP
    TCP_STREAM
      FOU, UDP checksum enabled
        15.28% TX CPU utilization
        13.92% RX CPU utilization
        9342 Mbps
      GUE, UDP checksum enabled
        13.99% TX CPU utilization
        13.34% RX CPU utilization
        9210 Mbps
      FOU, UDP checksum disabled
        15.08% TX CPU utilization
        24.64% RX CPU utilization
        9226 Mbps
      GUE, UDP checksum disabled
        15.90% TX CPU utilization
        24.77% RX CPU utilization
        9197 Mbps
    TCP_RR
      FOU, UDP checksum enabled
        94.23% CPU utilization
        149/237/429 90/95/99% latencies
        1.19553e+06 tps
      GUE, UDP checksum enabled
        93.75% CPU utilization
        152/243/442 90/95/99% latencies
        1.17027e+06 tps

  SIT
    TCP_STREAM
      FOU, UDP checksum enabled
        14.47% TX CPU utilization
        14.58% RX CPU utilization
        9106 Mbps
      GUE, UDP checksum enabled
        15.09% TX CPU utilization
        14.84% RX CPU utilization
        9080 Mbps
      FOU, UDP checksum disabled
        15.70% TX CPU utilization
        27.93% RX CPU utilization
        9097 Mbps
      GUE, UDP checksum disabled
        15.04% TX CPU utilization
        27.54% RX CPU utilization
        9073 Mbps
    TCP_RR
      FOU, UDP checksum enabled
        96.9% CPU utilization
        170/281/581 90/95/99% latencies
        1.03372e+06 tps
      GUE, UDP checksum enabled
        97.16% CPU utilization
        172/286/576 90/95/99% latencies
        1.00469e+06 tps
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoip_tunnel: Add GUE support
Tom Herbert [Fri, 3 Oct 2014 22:48:10 +0000 (15:48 -0700)]
ip_tunnel: Add GUE support

This patch allows configuring IPIP, sit, and GRE tunnels to use GUE.
This is very similar to fou excpet that we need to insert the GUE header
in addition to the UDP header on transmit.

Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agogue: Receive side for Generic UDP Encapsulation
Tom Herbert [Fri, 3 Oct 2014 22:48:09 +0000 (15:48 -0700)]
gue: Receive side for Generic UDP Encapsulation

This patch adds support receiving for GUE packets in the fou module. The
fou module now supports direct foo-over-udp (no encapsulation header)
and GUE. To support this a type parameter is added to the fou netlink
parameters.

For a GUE socket we define gue_udp_recv, gue_gro_receive, and
gue_gro_complete to handle the specifics of the GUE protocol. Most
of the code to manage and configure sockets is common with the fou.

Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agofou: eliminate IPv4,v6 specific GRO functions
Tom Herbert [Fri, 3 Oct 2014 22:48:08 +0000 (15:48 -0700)]
fou: eliminate IPv4,v6 specific GRO functions

This patch removes fou[46]_gro_receive and fou[46]_gro_complete
functions. The v4 or v6 variants were chosen for the UDP offloads
based on the address family of the socket this is not necessary
or correct. Alternatively, this patch adds is_ipv6 to napi_gro_skb.
This is set in udp6_gro_receive and unset in udp4_gro_receive. In
fou_gro_receive the value is used to select the correct inet_offloads
for the protocol of the outer IP header.

Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoip_tunnel: Account for secondary encapsulation header in max_headroom
Tom Herbert [Fri, 3 Oct 2014 22:48:07 +0000 (15:48 -0700)]
ip_tunnel: Account for secondary encapsulation header in max_headroom

When adjusting max_header for the tunnel interface based on egress
device we need to account for any extra bytes in secondary encapsulation
(e.g. FOU).

Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: do not export skb_gro_receive()
Eric Dumazet [Thu, 2 Oct 2014 14:38:46 +0000 (07:38 -0700)]
net: do not export skb_gro_receive()

skb_gro_receive() is only called from tcp_gro_receive() which is
not in a module.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agodrivers/net/irda/Kconfig: Let SH_IRDA depend on HAS_IOMEM
Chen Gang [Thu, 2 Oct 2014 14:32:56 +0000 (22:32 +0800)]
drivers/net/irda/Kconfig: Let SH_IRDA depend on HAS_IOMEM

SH_IRDA needs HAS_IOMEM, so depend on it. The related error(with
allmodconfig under um):

    CC [M]  drivers/net/irda/sh_irda.o
  drivers/net/irda/sh_irda.c: In function ‘sh_irda_probe’:
  drivers/net/irda/sh_irda.c:776:2: error: implicit declaration of function ‘ioremap_nocache’ [-Werror=implicit-function-declaration]
    self->membase = ioremap_nocache(res->start, resource_size(res));
    ^
  drivers/net/irda/sh_irda.c:776:16: warning: assignment makes pointer from integer without a cast [enabled by default]
    self->membase = ioremap_nocache(res->start, resource_size(res));
                  ^
  drivers/net/irda/sh_irda.c:821:2: error: implicit declaration of function ‘iounmap’ [-Werror=implicit-function-declaration]
    iounmap(self->membase);
    ^

Signed-off-by: Chen Gang <gang.chen.5i5j@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agodrivers/net/ethernet/marvell/Kconfig: Let PXA168_ETH depend on HAS_IOMEM
Chen Gang [Thu, 2 Oct 2014 14:23:33 +0000 (22:23 +0800)]
drivers/net/ethernet/marvell/Kconfig: Let PXA168_ETH depend on HAS_IOMEM

PXA168_ETH need HAS_IOMEM, so depend on it, the related error (with
allmodconfig under um):

    CC [M]  drivers/net/ethernet/marvell/pxa168_eth.o
  drivers/net/ethernet/marvell/pxa168_eth.c: In function ‘pxa168_eth_probe’:
  drivers/net/ethernet/marvell/pxa168_eth.c:1605:2: error: implicit declaration of function ‘iounmap’ [-Werror=implicit-function-declaration]
    iounmap(pep->base);
    ^

Signed-off-by: Chen Gang <gang.chen.5i5j@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agodrivers/net/dsa/Kconfig: Let NET_DSA_BCM_SF2 depend on HAS_IOMEM
Chen Gang [Thu, 2 Oct 2014 14:14:04 +0000 (22:14 +0800)]
drivers/net/dsa/Kconfig: Let NET_DSA_BCM_SF2 depend on HAS_IOMEM

NET_DSA_BCM_SF2 need HAS_IOMEM, so depend on it, the related error (with
allmodconfig under um):

    CC [M]  drivers/net/dsa/bcm_sf2.o
  drivers/net/dsa/bcm_sf2.c: In function ‘bcm_sf2_sw_setup’:
  drivers/net/dsa/bcm_sf2.c:487:3: error: implicit declaration of function ‘iounmap’ [-Werror=implicit-function-declaration]
     iounmap(*base);
     ^

Signed-off-by: Chen Gang <gang.chen.5i5j@gmail.com>
Acked-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agodrivers/net/can/Kconfig: Let CAN_AT91 depend on HAS_IOMEM
Chen Gang [Thu, 2 Oct 2014 14:01:42 +0000 (22:01 +0800)]
drivers/net/can/Kconfig: Let CAN_AT91 depend on HAS_IOMEM

CAN_AT91 needs HAS_IOMEM, so depends on it. The related error (with
allmodconfig under um):

    CC [M]  drivers/net/can/at91_can.o
  drivers/net/can/at91_can.c: In function ‘at91_can_probe’:
  drivers/net/can/at91_can.c:1329:2: error: implicit declaration of function ‘ioremap_nocache’ [-Werror=implicit-function-declaration]
  addr = ioremap_nocache(res->start, resource_size(res));
    ^
  drivers/net/can/at91_can.c:1329:7: warning: assignment makes pointer from integer without a cast [enabled by default]
    addr = ioremap_nocache(res->start, resource_size(res));
         ^
  drivers/net/can/at91_can.c:1384:2: error: implicit declaration of function ‘iounmap’ [-Werror=implicit-function-declaration]
    iounmap(addr);
    ^

Signed-off-by: Chen Gang <gang.chen.5i5j@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge branch 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/jkirsher/net...
David S. Miller [Fri, 3 Oct 2014 22:43:50 +0000 (15:43 -0700)]
Merge branch 'master' of git://git./linux/kernel/git/jkirsher/net-next

Jeff Kirsher says:

====================
Intel Wired LAN Driver Updates 2014-10-02

This series contains updates to fm10k, igb, ixgbe and i40e.

Alex provides two updates to the fm10k driver.  First reduces the buffer
size to 2k for all page sizes, since most frames only have a 1500 MTU
so supporting a buffer size larger than this is somewhat wasteful.
Second fixes an issue where the number of transmit queues was not being
updated, so added the lines necessary to update the number of transmit
queues.

Rick Jones provides two patches to convert ixgbe, igb and i40e to use
dev_consume_skb_any().

Emil provides two patches for ixgbe, first cleans up a couple of wait
loops on auto-negotiation that were not needed.  Second fixes an issue
reported by Fujitsu/Red Hat, which consolidates the logic behind the
dynamically setting of TXDCTL.WTHRESH depending on interrupt throttle
rate (ITR) setting regardless of BQL.

Ethan Zhao provides a cleanup patch for ixgbe where he noticed a
duplicate define.

Bernhard Kaindl provides a patch for igb to remove a source of latency
spikes by not calling code that uses mdelay() for feeding a PHY stat
while being called with a spinlock held.

Todd bumps the igb version based on the recent changes.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge branch 'mlx5-next'
David S. Miller [Fri, 3 Oct 2014 22:42:37 +0000 (15:42 -0700)]
Merge branch 'mlx5-next'

Eli Cohen says:

====================
mlx5 update for 3.18

This series integrates a new mechanism for populating and extracting field values
used in the driver/firmware interaction around command mailboxes.

Changes from V1:
 - Remove unused definition of memcpy_cpu_to_be32()
 - Remove definitions of non_existent_*() and use BUILD_BUG_ON() instead.
 - Added a patch one line patch to add support for ConnectX-4 devices.

Changes from V0:
 - trimmed the auto-generated file to a minimum, as required by the reviewers.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet/mlx5_core: Add ConnectX-4 to list of supported devices
Eli Cohen [Thu, 2 Oct 2014 09:19:46 +0000 (12:19 +0300)]
net/mlx5_core: Add ConnectX-4 to list of supported devices

Add the upcoming ConnectX-4 device to the list of supported devices by then
mlx5 driver.

Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet/mlx5_core: Identify resources by their type
Eli Cohen [Thu, 2 Oct 2014 09:19:45 +0000 (12:19 +0300)]
net/mlx5_core: Identify resources by their type

This patch puts a common part as the first field of mlx5_core_qp. This field is
used to identify which resource generated an event. This is required since upcoming
new resource types such as DC targets are allocated for the same numerical space
as regular QPs and may generate the same events. By searching the resource in the
same table we can then look at the common field to identify the resource.

Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet/mlx5_core: use set/get macros in device caps
Eli Cohen [Thu, 2 Oct 2014 09:19:44 +0000 (12:19 +0300)]
net/mlx5_core: use set/get macros in device caps

Transform device capabilities related commands to use set/get macros to
manipulate command mailboxes.

Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet/mlx5_core: Use hardware registers description header file
Eli Cohen [Thu, 2 Oct 2014 09:19:43 +0000 (12:19 +0300)]
net/mlx5_core: Use hardware registers description header file

Add an auto generated header file that describes hardware registers along with
set of macros that set/get values. The macros do static checks to avoid
overflow, handle endianess, and overall provide a clean way to code commands.
Currently the header file is small and we will add structs as we make use of
the macros.
A few commands were removed from the commands enum since they are not supported
currently and will be added when support is available.

Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet/mlx5_core: Update device capabilities handling
Eli Cohen [Thu, 2 Oct 2014 09:19:42 +0000 (12:19 +0300)]
net/mlx5_core: Update device capabilities handling

Rearrange struct mlx5_caps so it has a "gen" field to represent the current
capabilities configured for the device. Max capabilities can also be queried
from the device. Also update capabilities struct to contain more fields as per
the latest revision if firmware specification.

Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoqdisc: validate skb without holding lock
Eric Dumazet [Fri, 3 Oct 2014 22:31:07 +0000 (15:31 -0700)]
qdisc: validate skb without holding lock

Validation of skb can be pretty expensive :

GSO segmentation and/or checksum computations.

We can do this without holding qdisc lock, so that other cpus
can queue additional packets.

Trick is that requeued packets were already validated, so we carry
a boolean so that sch_direct_xmit() can validate a fresh skb list,
or directly use an old one.

Tested on 40Gb NIC (8 TX queues) and 200 concurrent flows, 48 threads
host.

Turning TSO on or off had no effect on throughput, only few more cpu
cycles. Lock contention on qdisc lock disappeared.

Same if disabling TX checksum offload.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: ethernet: Remove superfluous ether_setup after alloc_etherdev
Tobias Klauser [Thu, 2 Oct 2014 08:15:30 +0000 (10:15 +0200)]
net: ethernet: Remove superfluous ether_setup after alloc_etherdev

There is no need to call ether_setup after alloc_ethdev since it was
already called there.

Follow commits c706471b2601 ("net: axienet: remove unnecessary
ether_setup after alloc_etherdev") and 3c87dcbfb36c ("net: ll_temac:
Remove unnecessary ether_setup after alloc_etherdev") and fix the
pattern in all remaining ethernet drivers.

Signed-off-by: Tobias Klauser <tklauser@distanz.ch>
Acked-by: Nicolas Ferre <nicolas.ferre@atmel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge branch 'qdisc_bulk_dequeue'
David S. Miller [Fri, 3 Oct 2014 19:37:23 +0000 (12:37 -0700)]
Merge branch 'qdisc_bulk_dequeue'

Jesper Dangaard Brouer says:

====================
qdisc: bulk dequeue support

This patchset uses DaveM's recent API changes to dev_hard_start_xmit(),
from the qdisc layer, to implement dequeue bulking.

Patch01: "qdisc: bulk dequeue support for qdiscs with TCQ_F_ONETXQUEUE"
 - Implement basic qdisc dequeue bulking
 - This time, 100% relying on BQL limits, no magic safe-guard constants

Patch02: "qdisc: dequeue bulking also pickup GSO/TSO packets"
 - Extend bulking to bulk several GSO/TSO packets
 - Seperate patch, as it introduce a small regression, see test section.

We do have a patch03, which exports a userspace tunable as a BQL
tunable, that can byte-cap or disable the bulking/bursting.  But we
could not agree on it internally, thus not sending it now.  We
basically strive to avoid adding any new userspace tunable.

Testing patch01:
================
 Demonstrating the performance improvement of qdisc dequeue bulking, is
tricky because the effect only "kicks-in" once the qdisc system have a
backlog. Thus, for a backlog to form, we need either 1) to exceed wirespeed
of the link or 2) exceed the capability of the device driver.

For practical use-cases, the measureable effect of this will be a
reduction in CPU usage

01-TCP_STREAM:
--------------
Testing effect for TCP involves disabling TSO and GSO, because TCP
already benefit from bulking, via TSO and especially for GSO segmented
packets.  This patch view TSO/GSO as a seperate kind of bulking, and
avoid further bulking of these packet types.

The measured perf diff benefit (at 10Gbit/s) for a single netperf
TCP_STREAM were 9.24% less CPU used on calls to _raw_spin_lock()
(mostly from sch_direct_xmit).

If my E5-2695v2(ES) CPU is tuned according to:
 http://netoptimizer.blogspot.dk/2014/04/basic-tuning-for-network-overload.html
Then it is possible that a single netperf TCP_STREAM, with GSO and TSO
disabled, can utilize all bandwidth on a 10Gbit/s link.  This will
then cause a standing backlog queue at the qdisc layer.

Trying to pressure the system some more CPU util wise, I'm starting
24x TCP_STREAMs and monitoring the overall CPU utilization.  This
confirms bulking saves CPU cycles when it "kicks-in".

Tool mpstat, while stressing the system with netperf 24x TCP_STREAM, shows:
 * Disabled bulking: sys:2.58%  soft:8.50%  idle:88.78%
 * Enabled  bulking: sys:2.43%  soft:7.66%  idle:89.79%

02-UDP_STREAM
-------------
The measured perf diff benefit for UDP_STREAM were 6.41% less CPU used
on calls to _raw_spin_lock().  24x UDP_STREAM with packet size -m 1472 (to
avoid sending UDP/IP fragments).

03-trafgen driver test
----------------------
The performance of the 10Gbit/s ixgbe driver is limited due to
updating the HW ring-queue tail-pointer on every packet.  As
previously demonstrated with pktgen.

Using trafgen to send RAW frames from userspace (via AF_PACKET), and
forcing it through qdisc path (with option --qdisc-path and -t0),
sending with 12 CPUs.

I can demonstrate this driver layer limitation:
 * 12.8 Mpps with no qdisc bulking
 * 14.8 Mpps with qdisc bulking (full 10G-wirespeed)

Testing patch02:
================
Testing Bulking several GSO/TSO packets:

Measuring HoL (Head-of-Line) blocking for TSO and GSO, with
netperf-wrapper. Bulking several TSO show no performance regressions
(requeues were in the area 32 requeues/sec for 10G while transmitting
approx 813Kpps).

Bulking several GSOs does show small regression or very small
improvement (requeues were in the area 8000 requeues/sec, for 10G
while transmitting approx 813Kpps).

 Using ixgbe 10Gbit/s with GSO bulking, we can measure some additional
latency. Base-case, which is "normal" GSO bulking, sees varying
high-prio queue delay between 0.38ms to 0.47ms.  Bulking several GSOs
together, result in a stable high-prio queue delay of 0.50ms.

Corrosponding to:
 (10000*10^6)*((0.50-0.47)/10^3)/8 = 37500 bytes
 (10000*10^6)*((0.50-0.38)/10^3)/8 = 150000 bytes
 37500/1500  = 25 pkts
 150000/1500 = 100 pkts

 Using igb at 100Mbit/s with GSO bulking, shows an improvement.
Base-case sees varying high-prio queue delay between 2.23ms to 2.35ms
diff of 0.12ms corrosponding to 1500 bytes at 100Mbit/s. Bulking
several GSOs together, result in a stable high-prio queue delay of
2.23ms.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoqdisc: dequeue bulking also pickup GSO/TSO packets
Jesper Dangaard Brouer [Wed, 1 Oct 2014 20:36:09 +0000 (22:36 +0200)]
qdisc: dequeue bulking also pickup GSO/TSO packets

The TSO and GSO segmented packets already benefit from bulking
on their own.

The TSO packets have always taken advantage of the only updating
the tailptr once for a large packet.

The GSO segmented packets have recently taken advantage of
bulking xmit_more API, via merge commit 53fda7f7f9e8 ("Merge
branch 'xmit_list'"), specifically via commit 7f2e870f2a4 ("net:
Move main gso loop out of dev_hard_start_xmit() into helper.")
allowing qdisc requeue of remaining list.  And via commit
ce93718fb7cd ("net: Don't keep around original SKB when we
software segment GSO frames.").

This patch allow further bulking of TSO/GSO packets together,
when dequeueing from the qdisc.

Testing:
 Measuring HoL (Head-of-Line) blocking for TSO and GSO, with
netperf-wrapper. Bulking several TSO show no performance regressions
(requeues were in the area 32 requeues/sec).

Bulking several GSOs does show small regression or very small
improvement (requeues were in the area 8000 requeues/sec).

 Using ixgbe 10Gbit/s with GSO bulking, we can measure some additional
latency. Base-case, which is "normal" GSO bulking, sees varying
high-prio queue delay between 0.38ms to 0.47ms.  Bulking several GSOs
together, result in a stable high-prio queue delay of 0.50ms.

 Using igb at 100Mbit/s with GSO bulking, shows an improvement.
Base-case sees varying high-prio queue delay between 2.23ms to 2.35ms

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoqdisc: bulk dequeue support for qdiscs with TCQ_F_ONETXQUEUE
Jesper Dangaard Brouer [Wed, 1 Oct 2014 20:35:59 +0000 (22:35 +0200)]
qdisc: bulk dequeue support for qdiscs with TCQ_F_ONETXQUEUE

Based on DaveM's recent API work on dev_hard_start_xmit(), that allows
sending/processing an entire skb list.

This patch implements qdisc bulk dequeue, by allowing multiple packets
to be dequeued in dequeue_skb().

The optimization principle for this is two fold, (1) to amortize
locking cost and (2) avoid expensive tailptr update for notifying HW.
 (1) Several packets are dequeued while holding the qdisc root_lock,
amortizing locking cost over several packet.  The dequeued SKB list is
processed under the TXQ lock in dev_hard_start_xmit(), thus also
amortizing the cost of the TXQ lock.
 (2) Further more, dev_hard_start_xmit() will utilize the skb->xmit_more
API to delay HW tailptr update, which also reduces the cost per
packet.

One restriction of the new API is that every SKB must belong to the
same TXQ.  This patch takes the easy way out, by restricting bulk
dequeue to qdisc's with the TCQ_F_ONETXQUEUE flag, that specifies the
qdisc only have attached a single TXQ.

Some detail about the flow; dev_hard_start_xmit() will process the skb
list, and transmit packets individually towards the driver (see
xmit_one()).  In case the driver stops midway in the list, the
remaining skb list is returned by dev_hard_start_xmit().  In
sch_direct_xmit() this returned list is requeued by dev_requeue_skb().

To avoid overshooting the HW limits, which results in requeuing, the
patch limits the amount of bytes dequeued, based on the drivers BQL
limits.  In-effect bulking will only happen for BQL enabled drivers.

Small amounts for extra HoL blocking (2x MTU/0.24ms) were
measured at 100Mbit/s, with bulking 8 packets, but the
oscillating nature of the measurement indicate something, like
sched latency might be causing this effect. More comparisons
show, that this oscillation goes away occationally. Thus, we
disregard this artifact completely and remove any "magic" bulking
limit.

For now, as a conservative approach, stop bulking when seeing TSO and
segmented GSO packets.  They already benefit from bulking on their own.
A followup patch add this, to allow easier bisect-ability for finding
regressions.

Jointed work with Hannes, Daniel and Florian.

Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoet131x: Add PCIe gigabit ethernet driver et131x to drivers/net
Mark Einon [Tue, 30 Sep 2014 21:29:46 +0000 (22:29 +0100)]
et131x: Add PCIe gigabit ethernet driver et131x to drivers/net

This adds the ethernet driver for Agere et131x devices to
drivers/net/ethernet.

The driver being added has been in the staging tree for some time, and will be
removed from there in a seperate patch. This one merely disables the staging
version to prevent two instances being built.

Signed-off-by: Mark Einon <mark.einon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net
David S. Miller [Thu, 2 Oct 2014 18:25:43 +0000 (11:25 -0700)]
Merge git://git./linux/kernel/git/davem/net

Conflicts:
drivers/net/usb/r8152.c
net/netfilter/nfnetlink.c

Both r8152 and nfnetlink conflicts were simple overlapping changes.

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoigb: bump version to 5.2.15
Todd Fujinaka [Sat, 20 Sep 2014 04:46:25 +0000 (04:46 +0000)]
igb: bump version to 5.2.15

Bump version

Signed-off-by: Todd Fujinaka <todd.fujinaka@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
10 years agoi40e/igb: Convert to dev_consume_skb_any()
Rick Jones [Wed, 17 Sep 2014 03:56:20 +0000 (03:56 +0000)]
i40e/igb: Convert to dev_consume_skb_any()

Convert two more Intel NIC drivers to dev_consume_skb_any() to help
make dropped packet profiling sane.

Signed-off-by: Rick Jones <rick.jones2@hp.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Tested-by: Jim Young <jamesx.m.young@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
10 years agoigb: remove blocking phy read from inside spinlock
Bernhard Kaindl [Wed, 17 Sep 2014 19:11:16 +0000 (19:11 +0000)]
igb: remove blocking phy read from inside spinlock

Remove a source of latency spikes (in my case up to 10ms) by not calling
code that uses mdelay() for feeding a phy statistic (rx errors for idle
symbols - not data -> idle_errors) while being called with a spinlock held.

As idle_errors isn't read, this patch only removes unused code and data.

Later, more complicated changes may be applied to address the spinlock and
allow for some PHY diagnostics by harvesting this PHY stats register fully.

This patch is designed to fix the issue and be safe for longterm/stable.

For the Intel e1000e driver, the same change was applied in 2008 with
commit 23033fad5be0 ("e1000e: remove phy read from inside spinlock").

The mdelay is triggered by HW/SW semaphores, thus it depends on the HW.

I've HW that triggers it even when idle. Others may trigger it only e.g.
when Ethernet ports aquire or loose the link or on ifconfig up / down.
We've noticed this first from delays in frame rx/tx due to the mdelay().

Example command for checking if the issue is triggered: cyclictest -Smp1
(Look for occasional "Max:" values > 4000 or use -b 4000 to stop if greater)

It was observed with I350 ports connected to other I350 ports, but not
if driver and EEPROM was modified to run the I350 in EEPROM-less mode.

phy_stats.idle_errors and .receive_errors (isn't touched) occupy 64 not
used bits in the adapter struct: Their allocation may be removed as well.

Cc: Carolyn Wyborny <carolyn.wyborny@intel.com>
Cc: Todd Fujinaka <todd.fujinaka@intel.com>
Fixes: 12dcd86b75d5 ("igb: fix stats handling") (this added the spin_lock)
Signed-off-by: Bernhard Kaindl <bk-linux@use.startmail.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
10 years agoixgbe: delete one duplicate marcro definition of IXGBE_MAX_L2A_QUEUES
Ethan Zhao [Tue, 23 Sep 2014 18:11:44 +0000 (18:11 +0000)]
ixgbe: delete one duplicate marcro definition of IXGBE_MAX_L2A_QUEUES

There is typo in ixgbe.h, two marcro definition of IXGBE_MAX_L2A_QUEUES to 4,
delete one, clear the compiler warning.

Signed-off-by: Ethan Zhao <ethan.zhao@oracle.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
10 years agoixgbe: fix setting of TXDCTL.WTRHESH when ITR is set to 0 and no BQL
Emil Tantilov [Thu, 18 Sep 2014 08:05:02 +0000 (08:05 +0000)]
ixgbe: fix setting of TXDCTL.WTRHESH when ITR is set to 0 and no BQL

This patch consolidates the logic behind dynamically setting TXDCTL.WTHRESH
depending on interrupt throttle rate (ITR) setting regardless of BQL.

Previously TXDCTL.WTHRESH was dynamically being set only with BQL being
enabled, but we have to set it regardless of BQL when ITR is low to avoid
Tx stalls/hangs.

CC: John Greene <jogreene@redhat.com>
Reported by: Masayuki Gouji <gouji.masayuki@jp.fujitsu.com>
Signed-off-by: Emil Tantilov <emil.s.tantilov@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
10 years agoixgbe: remove wait loop on autoneg for copper devices
Emil Tantilov [Sat, 6 Sep 2014 07:50:27 +0000 (07:50 +0000)]
ixgbe: remove wait loop on autoneg for copper devices

This patch removes couple of wait loops on autoneg that are not needed.

During validation we noticed that the loops always time out, so there
should be no user impact.

Signed-off-by: Emil Tantilov <emil.s.tantilov@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
10 years agoixgbe: Convert the normal transmit complete path to dev_consume_skb_any()
Rick Jones [Fri, 12 Sep 2014 17:44:06 +0000 (17:44 +0000)]
ixgbe: Convert the normal transmit complete path to dev_consume_skb_any()

Convert the normal packet completion path to dev_consume_skb_any() so
packet drop profiling via dropwatch or perf top -G -e skb_kfree_skb
is not cluttered with false hits.

Compile tested only.  There is a dev_kfree_skb_any() in the routine
ixgbe_ptp_tx_hwtstamp() in ixgbe_ptp.c that looks like a conversion
candidate but I wasn't familiar enough with the code to pull the
trigger.

Signed-off-by: Rick Jones <rick.jones2@hp.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
10 years agofm10k: Correctly set the number of Tx queues
Alexander Duyck [Tue, 30 Sep 2014 22:49:22 +0000 (22:49 +0000)]
fm10k: Correctly set the number of Tx queues

The number of Tx queues was not being updated due to some issues when
generating the patches.  This change makes sure to add the lines necessary
to update the number of Tx queues correctly.

Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
10 years agofm10k: Reduce buffer size when pages are larger than 4K
Alexander Duyck [Fri, 26 Sep 2014 06:33:49 +0000 (06:33 +0000)]
fm10k: Reduce buffer size when pages are larger than 4K

This change reduces the buffer size to 2K for all page sizes.  The basic
idea is that since most frames only have a 1500 MTU supporting a buffer
size larger than this is somewhat wasteful.  As such I have reduced the
size to 2K for all page sizes which will allow for more uses per page.

Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
10 years agoMerge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net
Linus Torvalds [Thu, 2 Oct 2014 04:29:06 +0000 (21:29 -0700)]
Merge git://git./linux/kernel/git/davem/net

Pull networking fixes from David Miller:

 1) Don't halt the firmware in r8152 driver, from Hayes Wang.

 2) Handle full sized 802.1ad frames in bnx2 and tg3 drivers properly,
    from Vlad Yasevich.

 3) Don't sleep while holding tx_clean_lock in netxen driver, fix from
    Manish Chopra.

 4) Certain kinds of ipv6 routes can end up endlessly failing the route
    validation test, causing it to be re-looked up over and over again.
    This particularly kills input route caching in TCP sockets.  Fix
    from Hannes Frederic Sowa.

 5) netvsc_start_xmit() has a use-after-free access to skb->len, fix
    from K Y Srinivasan.

 6) Fix matching of inverted containers in ematch module, from Ignacy
    Gawędzki.

 7) Aggregation of GRO frames via SKB ->frag_list for linear skbs isn't
    handled properly, regression fix from Eric Dumazet.

 8) Don't test return value of ipv4_neigh_lookup(), which returns an
    error pointer, against NULL.  From WANG Cong.

 9) Fix an old regression where we mistakenly allow a double add of the
    same tunnel.  Fixes from Steffen Klassert.

10) macvtap device delete and open can run in parallel and corrupt lists
    etc., fix from Vlad Yasevich.

11) Fix build error with IPV6=m NETFILTER_XT_TARGET_TPROXY=y, from Pablo
    Neira Ayuso.

12) rhashtable_destroy() triggers lockdep splats, fix also from Pablo.

* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net: (32 commits)
  bna: Update Maintainer Email
  r8152: disable power cut for RTL8153
  r8152: remove clearing bp
  bnx2: Correctly receive full sized 802.1ad fragmes
  tg3: Allow for recieve of full-size 8021AD frames
  r8152: fix setting RTL8152_UNPLUG
  netxen: Fix bug in Tx completion path.
  netxen: Fix BUG "sleeping function called from invalid context"
  ipv6: remove rt6i_genid
  hyperv: Fix a bug in netvsc_start_xmit()
  net: stmmac: fix stmmac_pci_probe failed when CONFIG_HAVE_CLK is selected
  ematch: Fix matching of inverted containers.
  gro: fix aggregation for skb using frag_list
  neigh: check error pointer instead of NULL for ipv4_neigh_lookup()
  ip6_gre: Return an error when adding an existing tunnel.
  ip6_vti: Return an error when adding an existing tunnel.
  ip6_tunnel: Return an error when adding an existing tunnel.
  ip6gre: add a rtnl link alias for ip6gretap
  net/mlx4_core: Allow not to specify probe_vf in SRIOV IB mode
  r8152: fix the carrier off when autoresuming
  ...

10 years agobna: Update Maintainer Email
Rasesh Mody [Wed, 1 Oct 2014 21:20:41 +0000 (17:20 -0400)]
bna: Update Maintainer Email

Update the maintainer email for BNA driver.

Signed-off-by: Rasesh Mody <rasesh.mody@qlogic.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: phy: add BCM7425 and BCM7429 PHYs
Petri Gynther [Wed, 1 Oct 2014 18:58:02 +0000 (11:58 -0700)]
net: phy: add BCM7425 and BCM7429 PHYs

Signed-off-by: Petri Gynther <pgynther@google.com>
Acked-by: Florian Fainelli <f.fainelli@gmai.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: bcmgenet: fix bcmgenet_put_tx_csum()
Petri Gynther [Wed, 1 Oct 2014 18:30:01 +0000 (11:30 -0700)]
net: bcmgenet: fix bcmgenet_put_tx_csum()

bcmgenet_put_tx_csum() needs to return skb pointer back to the caller
because it reallocates a new one in case of lack of skb headroom.

Signed-off-by: Petri Gynther <pgynther@google.com>
Acked-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: pktgen: packet bursting via skb->xmit_more
Alexei Starovoitov [Wed, 1 Oct 2014 00:53:21 +0000 (17:53 -0700)]
net: pktgen: packet bursting via skb->xmit_more

This patch demonstrates the effect of delaying update of HW tailptr.
(based on earlier patch by Jesper)

burst=1 is the default. It sends one packet with xmit_more=false
burst=2 sends one packet with xmit_more=true and
        2nd copy of the same packet with xmit_more=false
burst=3 sends two copies of the same packet with xmit_more=true and
        3rd copy with xmit_more=false

Performance with ixgbe (usec 30):
burst=1  tx:9.2 Mpps
burst=2  tx:13.5 Mpps
burst=3  tx:14.5 Mpps full 10G line rate

Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: bridge: add a br_set_state helper function
Florian Fainelli [Tue, 30 Sep 2014 23:13:19 +0000 (16:13 -0700)]
net: bridge: add a br_set_state helper function

In preparation for being able to propagate port states to e.g: notifiers
or other kernel parts, do not manipulate the port state directly, but
instead use a helper function which will allow us to do a bit more than
just setting the state.

Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet_sched: avoid calling tcf_unbind_filter() in call_rcu callback
WANG Cong [Tue, 30 Sep 2014 23:07:24 +0000 (16:07 -0700)]
net_sched: avoid calling tcf_unbind_filter() in call_rcu callback

This fixes the following crash:

[   63.976822] general protection fault: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC
[   63.980094] CPU: 1 PID: 15 Comm: ksoftirqd/1 Not tainted 3.17.0-rc6+ #648
[   63.980094] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
[   63.980094] task: ffff880117dea690 ti: ffff880117dfc000 task.ti: ffff880117dfc000
[   63.980094] RIP: 0010:[<ffffffff817e6d07>]  [<ffffffff817e6d07>] u32_destroy_key+0x27/0x6d
[   63.980094] RSP: 0018:ffff880117dffcc0  EFLAGS: 00010202
[   63.980094] RAX: ffff880117dea690 RBX: ffff8800d02e0820 RCX: 0000000000000000
[   63.980094] RDX: 0000000000000001 RSI: 0000000000000002 RDI: 6b6b6b6b6b6b6b6b
[   63.980094] RBP: ffff880117dffcd0 R08: 0000000000000000 R09: 0000000000000000
[   63.980094] R10: 00006c0900006ba8 R11: 00006ba100006b9d R12: 0000000000000001
[   63.980094] R13: ffff8800d02e0898 R14: ffffffff817e6d4d R15: ffff880117387a30
[   63.980094] FS:  0000000000000000(0000) GS:ffff88011a800000(0000) knlGS:0000000000000000
[   63.980094] CS:  0010 DS: 0000 ES: 0000 CR0: 000000008005003b
[   63.980094] CR2: 00007f07e6732fed CR3: 000000011665b000 CR4: 00000000000006e0
[   63.980094] Stack:
[   63.980094]  ffff88011a9cd300 ffffffff82051ac0 ffff880117dffce0 ffffffff817e6d68
[   63.980094]  ffff880117dffd70 ffffffff810cb4c7 ffffffff810cb3cd ffff880117dfffd8
[   63.980094]  ffff880117dea690 ffff880117dea690 ffff880117dfffd8 000000000000000a
[   63.980094] Call Trace:
[   63.980094]  [<ffffffff817e6d68>] u32_delete_key_freepf_rcu+0x1b/0x1d
[   63.980094]  [<ffffffff810cb4c7>] rcu_process_callbacks+0x3bb/0x691
[   63.980094]  [<ffffffff810cb3cd>] ? rcu_process_callbacks+0x2c1/0x691
[   63.980094]  [<ffffffff817e6d4d>] ? u32_destroy_key+0x6d/0x6d
[   63.980094]  [<ffffffff810780a4>] __do_softirq+0x142/0x323
[   63.980094]  [<ffffffff810782a8>] run_ksoftirqd+0x23/0x53
[   63.980094]  [<ffffffff81092126>] smpboot_thread_fn+0x203/0x221
[   63.980094]  [<ffffffff81091f23>] ? smpboot_unpark_thread+0x33/0x33
[   63.980094]  [<ffffffff8108e44d>] kthread+0xc9/0xd1
[   63.980094]  [<ffffffff819e00ea>] ? do_wait_for_common+0xf8/0x125
[   63.980094]  [<ffffffff8108e384>] ? __kthread_parkme+0x61/0x61
[   63.980094]  [<ffffffff819e43ec>] ret_from_fork+0x7c/0xb0
[   63.980094]  [<ffffffff8108e384>] ? __kthread_parkme+0x61/0x61

tp could be freed in call_rcu callback too, the order is not guaranteed.

John Fastabend says:

====================
Its worth noting why this is safe. Any running schedulers will either
read the valid class field or it will be zeroed.

All schedulers today when the class is 0 do a lookup using the
same call used by the tcf_exts_bind(). So even if we have a running
classifier hit the null class pointer it will do a lookup and get
to the same result. This is particularly fragile at the moment because
the only way to verify this is to audit the schedulers call sites.
====================

Cc: John Fastabend <john.r.fastabend@intel.com>
Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
Acked-by: John Fastabend <john.r.fastabend@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet_sched: fix another crash in cls_tcindex
WANG Cong [Tue, 30 Sep 2014 23:07:23 +0000 (16:07 -0700)]
net_sched: fix another crash in cls_tcindex

This patch fixes the following crash:

[  166.670795] BUG: unable to handle kernel NULL pointer dereference at           (null)
[  166.674230] IP: [<ffffffff814b739f>] __list_del_entry+0x5c/0x98
[  166.674230] PGD d0ea5067 PUD ce7fc067 PMD 0
[  166.674230] Oops: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC
[  166.674230] CPU: 1 PID: 775 Comm: tc Not tainted 3.17.0-rc6+ #642
[  166.674230] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
[  166.674230] task: ffff8800d03c4d20 ti: ffff8800cae7c000 task.ti: ffff8800cae7c000
[  166.674230] RIP: 0010:[<ffffffff814b739f>]  [<ffffffff814b739f>] __list_del_entry+0x5c/0x98
[  166.674230] RSP: 0018:ffff8800cae7f7d0  EFLAGS: 00010207
[  166.674230] RAX: 0000000000000000 RBX: ffff8800cba8d700 RCX: ffff8800cba8d700
[  166.674230] RDX: 0000000000000000 RSI: dead000000200200 RDI: ffff8800cba8d700
[  166.674230] RBP: ffff8800cae7f7d0 R08: 0000000000000001 R09: 0000000000000001
[  166.674230] R10: 0000000000000000 R11: 000000000000859a R12: ffffffffffffffe8
[  166.674230] R13: ffff8800cba8c5b8 R14: 0000000000000001 R15: ffff8800cba8d700
[  166.674230] FS:  00007fdb5f04a740(0000) GS:ffff88011a800000(0000) knlGS:0000000000000000
[  166.674230] CS:  0010 DS: 0000 ES: 0000 CR0: 000000008005003b
[  166.674230] CR2: 0000000000000000 CR3: 00000000cf929000 CR4: 00000000000006e0
[  166.674230] Stack:
[  166.674230]  ffff8800cae7f7e8 ffffffff814b73e8 ffff8800cba8d6e8 ffff8800cae7f828
[  166.674230]  ffffffff817caeec 0000000000000046 ffff8800cba8c5b0 ffff8800cba8c5b8
[  166.674230]  0000000000000000 0000000000000001 ffff8800cf8e33e8 ffff8800cae7f848
[  166.674230] Call Trace:
[  166.674230]  [<ffffffff814b73e8>] list_del+0xd/0x2b
[  166.674230]  [<ffffffff817caeec>] tcf_action_destroy+0x4c/0x71
[  166.674230]  [<ffffffff817ca0ce>] tcf_exts_destroy+0x20/0x2d
[  166.674230]  [<ffffffff817ec2b5>] tcindex_delete+0x196/0x1b7

struct list_head can not be simply copied and we should always init it.

Cc: John Fastabend <john.r.fastabend@intel.com>
Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
Acked-by: John Fastabend <john.r.fastabend@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge branch 'udp_gso'
David S. Miller [Thu, 2 Oct 2014 01:35:58 +0000 (21:35 -0400)]
Merge branch 'udp_gso'

Tom Herbert says:

====================
udp: Generalize GSO for UDP tunnels

This patch set generalizes the UDP tunnel segmentation functions so
that they can work with various protocol encapsulations. The primary
change is to set the inner_protocol field in the skbuff when creating
the encapsulated packet, and then in skb_udp_tunnel_segment this data
is used to determine the function for segmenting the encapsulated
packet. The inner_protocol field is overloaded to take either an
Ethertype or IP protocol.

The inner_protocol is set on transmit using skb_set_inner_ipproto or
skb_set_inner_protocol functions. VXLAN and IP tunnels (for fou GSO)
were modified to call these.

Notes:
  - GSO for GRE/UDP where GRE checksum is enabled does not work.
    Handling this will require some special case code.
  - Software GSO now supports many varieties of encapsulation with
    SKB_GSO_UDP_TUNNEL{_CSUM}. We still need a mechanism to query
    for device support of particular combinations (I intend to
    add ndo_gso_check for that).
  - MPLS seems to be the only previous user of inner_protocol. I don't
    believe these patches can affect that. For supporting GSO with
    MPLS over UDP, the inner_protocol should be set using the
    helper functions in this patch.
  - GSO for L2TP/UDP should also be straightforward now.

v2:
  - Respin for Eric's restructuring of skbuff.

Tested GRE, IPIP, and SIT over fou as well as VLXAN. This was
done using 200 TCP_STREAMs in netperf.

 GRE
    IPv4, FOU, UDP checksum enabled
      TCP_STREAM TSO enabled on tun interface
        14.04% TX CPU utilization
        13.17% RX CPU utilization
        9211 Mbps
      TCP_STREAM TSO disabled on tun interface
        27.82% TX CPU utilization
        25.41% RX CPU utilization
        9336 Mbps
    IPv4, FOU, UDP checksum disabled
      TCP_STREAM TSO enabled on tun interface
        13.14% TX CPU utilization
        23.18% RX CPU utilization
        9277 Mbps
      TCP_STREAM TSO disabled on tun interface
        30.00% TX CPU utilization
        31.28% RX CPU utilization
        9327 Mbps

  IPIP
    FOU, UDP checksum enabled
      TCP_STREAM TSO enabled on tun interface
        15.28% TX CPU utilization
        13.92% RX CPU utilization
        9342 Mbps
      TCP_STREAM TSO disabled on tun interface
        27.82% TX CPU utilization
        25.41% RX CPU utilization
        9336 Mbps
    FOU, UDP checksum disabled
      TCP_STREAM TSO enabled on tun interface
        15.08% TX CPU utilization
        24.64% RX CPU utilization
        9226 Mbps
      TCP_STREAM TSO disabled on tun interface
        30.00% TX CPU utilization
        31.28% RX CPU utilization
        9327 Mbps

  SIT
    FOU, UDP checksum enabled
      TCP_STREAM TSO enabled on tun interface
        14.47% TX CPU utilization
        14.58% RX CPU utilization
        9106 Mbps
      TCP_STREAM TSO disabled on tun interface
        31.82% TX CPU utilization
        30.82% RX CPU utilization
        9204 Mbps
    FOU, UDP checksum disabled
      TCP_STREAM TSO enabled on tun interface
        15.70% TX CPU utilization
        27.93% RX CPU utilization
        9097 Mbps
      TCP_STREAM TSO disabled on tun interface
        33.48% TX CPU utilization
        37.36% RX CPU utilization
        9197 Mbps

   VXLAN
      TCP_STREAM TSO enabled on tun interface
        16.42% TX CPU utilization
        23.66% RX CPU utilization
        9081 Mbps
      TCP_STREAM TSO disabled on tun interface
        30.32% TX CPU utilization
        30.55% RX CPU utilization
        9185 Mbps

   Baseline (no encp, TSO and LRO enabled)
      TCP_STREAM
        11.85% TX CPU utilization
        15.13% RX CPU utilization
        9452 Mbps
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agovxlan: Set inner protocol before transmit
Tom Herbert [Tue, 30 Sep 2014 03:22:33 +0000 (20:22 -0700)]
vxlan: Set inner protocol before transmit

Call skb_set_inner_protocol to set inner Ethernet protocol to
ETH_P_TEB before transmit. This is needed for GSO with UDP tunnels.

Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agogre: Set inner protocol in v4 and v6 GRE transmit
Tom Herbert [Tue, 30 Sep 2014 03:22:32 +0000 (20:22 -0700)]
gre: Set inner protocol in v4 and v6 GRE transmit

Call skb_set_inner_protocol to set inner Ethernet protocol to
protocol being encapsulation by GRE before tunnel_xmit. This is
needed for GSO if UDP encapsulation (fou) is being done.

Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoipip: Set inner IP protocol in ipip
Tom Herbert [Tue, 30 Sep 2014 03:22:31 +0000 (20:22 -0700)]
ipip: Set inner IP protocol in ipip

Call skb_set_inner_ipproto to set inner IP protocol to IPPROTO_IPV4
before tunnel_xmit. This is needed if UDP encapsulation (fou) is
being done.

Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agosit: Set inner IP protocol in sit
Tom Herbert [Tue, 30 Sep 2014 03:22:30 +0000 (20:22 -0700)]
sit: Set inner IP protocol in sit

Call skb_set_inner_ipproto to set inner IP protocol to IPPROTO_IPV6
before tunnel_xmit. This is needed if UDP encapsulation (fou) is
being done.

Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoudp: Generalize skb_udp_segment
Tom Herbert [Tue, 30 Sep 2014 03:22:29 +0000 (20:22 -0700)]
udp: Generalize skb_udp_segment

skb_udp_segment is the function called from udp4_ufo_fragment to
segment a UDP tunnel packet. This function currently assumes
segmentation is transparent Ethernet bridging (i.e. VXLAN
encapsulation). This patch generalizes the function to
operate on either Ethertype or IP protocol.

The inner_protocol field must be set to the protocol of the inner
header. This can now be either an Ethertype or an IP protocol
(in a union). A new flag in the skbuff indicates which type is
effective. skb_set_inner_protocol and skb_set_inner_ipproto
helper functions were added to set the inner_protocol. These
functions are called from the point where the tunnel encapsulation
is occuring.

When skb_udp_tunnel_segment is called, the function to segment the
inner packet is selected based on the inner IP or Ethertype. In the
case of an IP protocol encapsulation, the function is derived from
inet[6]_offloads. In the case of Ethertype, skb->protocol is
set to the inner_protocol and skb_mac_gso_segment is called. (GRE
currently does this, but it might be possible to lookup the protocol
in offload_base and call the appropriate segmenation function
directly).

Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge branch 'bpf-next'
David S. Miller [Thu, 2 Oct 2014 01:30:46 +0000 (21:30 -0400)]
Merge branch 'bpf-next'

Alexei Starovoitov says:

====================
bpf: add search pruning optimization and tests

patch #1 commit log explains why eBPF verifier has to examine some
instructions multiple times and describes the search pruning optimization
that improves verification speed for branchy programs and allows more
complex programs to be verified successfully.
This patch completes the core verifier logic.

patch #2 adds more verifier tests related to branches and search pruning

I'm still working on Andy's 'bitmask for stack slots' suggestion. It will be
done on top of this patch.

The current verifier algorithm is brute force depth first search with
state pruning. If anyone can come up with another algorithm that demonstrates
better results, we'll replace the algorithm without affecting user space.

Note verifier doesn't guarantee that all possible valid programs are accepted.
Overly complex programs may still be rejected.
Verifier improvements/optimizations will guarantee that if a program
was passing verification in the past, it will still be passing.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agobpf: add tests to verifier testsuite
Alexei Starovoitov [Tue, 30 Sep 2014 01:50:02 +0000 (18:50 -0700)]
bpf: add tests to verifier testsuite

add 4 extra tests to cover jump verification better

Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agobpf: add search pruning optimization to verifier
Alexei Starovoitov [Tue, 30 Sep 2014 01:50:01 +0000 (18:50 -0700)]
bpf: add search pruning optimization to verifier

consider C program represented in eBPF:
int filter(int arg)
{
    int a, b, c, *ptr;

    if (arg == 1)
        ptr = &a;
    else if (arg == 2)
        ptr = &b;
    else
        ptr = &c;

    *ptr = 0;
    return 0;
}
eBPF verifier has to follow all possible paths through the program
to recognize that '*ptr = 0' instruction would be safe to execute
in all situations.
It's doing it by picking a path towards the end and observes changes
to registers and stack at every insn until it reaches bpf_exit.
Then it comes back to one of the previous branches and goes towards
the end again with potentially different values in registers.
When program has a lot of branches, the number of possible combinations
of branches is huge, so verifer has a hard limit of walking no more
than 32k instructions. This limit can be reached and complex (but valid)
programs could be rejected. Therefore it's important to recognize equivalent
verifier states to prune this depth first search.

Basic idea can be illustrated by the program (where .. are some eBPF insns):
    1: ..
    2: if (rX == rY) goto 4
    3: ..
    4: ..
    5: ..
    6: bpf_exit
In the first pass towards bpf_exit the verifier will walk insns: 1, 2, 3, 4, 5, 6
Since insn#2 is a branch the verifier will remember its state in verifier stack
to come back to it later.
Since insn#4 is marked as 'branch target', the verifier will remember its state
in explored_states[4] linked list.
Once it reaches insn#6 successfully it will pop the state recorded at insn#2 and
will continue.
Without search pruning optimization verifier would have to walk 4, 5, 6 again,
effectively simulating execution of insns 1, 2, 4, 5, 6
With search pruning it will check whether state at #4 after jumping from #2
is equivalent to one recorded in explored_states[4] during first pass.
If there is an equivalent state, verifier can prune the search at #4 and declare
this path to be safe as well.
In other words two states at #4 are equivalent if execution of 1, 2, 3, 4 insns
and 1, 2, 4 insns produces equivalent registers and stack.

Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: fec: implement rx_copybreak to improve rx performance
Nimrod Andy [Tue, 30 Sep 2014 01:28:05 +0000 (09:28 +0800)]
net: fec: implement rx_copybreak to improve rx performance

- Copy short frames and keep the buffers mapped, re-allocate skb instead of
  memory copy for long frames.
- Add support for setting/getting rx_copybreak using generic ethtool tunable

Changes V3:
* As Eric Dumazet's suggestion that removing the copybreak module parameter
  and only keep the ethtool API support for rx_copybreak.

Changes V2:
* Implements rx_copybreak
* Rx_copybreak provides module parameter to change this value
* Add tunable_ops support for rx_copybreak

Signed-off-by: Fugang Duan <B38611@freescale.com>
Signed-off-by: Frank Li <Frank.Li@freescale.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: avoid one atomic operation in skb_clone()
Eric Dumazet [Wed, 1 Oct 2014 22:27:15 +0000 (15:27 -0700)]
net: avoid one atomic operation in skb_clone()

Fast clone cloning can actually avoid an atomic_inc(), if we
guarantee prior clone_ref value is 1.

This requires a change kfree_skbmem(), to perform the
atomic_dec_and_test() on clone_ref before setting fclone to
SKB_FCLONE_UNAVAILABLE.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet/dccp/ccid.c: add __init to ccid_activate
Fabian Frederick [Wed, 1 Oct 2014 04:52:06 +0000 (06:52 +0200)]
net/dccp/ccid.c: add __init to ccid_activate

ccid_activate is only called by __init ccid_initialize_builtins in same module.

Signed-off-by: Fabian Frederick <fabf@skynet.be>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet/dccp/proto.c: add __init to dccp_mib_init
Fabian Frederick [Wed, 1 Oct 2014 04:48:03 +0000 (06:48 +0200)]
net/dccp/proto.c: add __init to dccp_mib_init

dccp_mib_init is only called by __init dccp_init in same module.

Signed-off-by: Fabian Frederick <fabf@skynet.be>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge branch 'r8152'
David S. Miller [Wed, 1 Oct 2014 20:46:41 +0000 (16:46 -0400)]
Merge branch 'r8152'

Hayes Wang says:

====================
r8152: patches about firmware

The patches fix the issues when the firmware exists.

For the multiple OS, the firmware may be loaded by the
driver of the other OS. And the Linux driver has influences
on it.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agor8152: disable power cut for RTL8153
hayeswang [Wed, 1 Oct 2014 05:25:11 +0000 (13:25 +0800)]
r8152: disable power cut for RTL8153

The firmware would be clear when the power cut is enabled for
RTL8153.

Signed-off-by: Hayes Wang <hayeswang@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agor8152: remove clearing bp
hayeswang [Wed, 1 Oct 2014 05:25:10 +0000 (13:25 +0800)]
r8152: remove clearing bp

The xxx_clear_bp() is used to halt the firmware. It only necessary
for updating the new firmware. Besides, depend on the version of
the current firmware, it may have problem to halt the firmware
directly. Finally, halt the firmware would let the firmware code
useless, and the bugs which are fixed by the firmware would occur.

Signed-off-by: Hayes Wang <hayeswang@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agobnx2: Correctly receive full sized 802.1ad fragmes
Vlad Yasevich [Tue, 30 Sep 2014 23:39:37 +0000 (19:39 -0400)]
bnx2: Correctly receive full sized 802.1ad fragmes

This driver, similar to tg3, has a check that will
cause full sized 802.1ad frames to be dropped.  The
frame will be larger then the standard mtu due to the
presense of vlan header that has not been stripped.
The driver should not drop this frame and should process
it just like it does for 802.1q.

CC: Sony Chacko <sony.chacko@qlogic.com>
CC: Dept-HSGLinuxNICDev@qlogic.com
Signed-off-by: Vladislav Yasevich <vyasevic@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agotg3: Allow for recieve of full-size 8021AD frames
Vlad Yasevich [Tue, 30 Sep 2014 23:39:36 +0000 (19:39 -0400)]
tg3: Allow for recieve of full-size 8021AD frames

When receiving a vlan-tagged frame that still contains
a vlan header, the length of the packet will be greater
then MTU+ETH_HLEN since it will account of the extra
vlan header.  TG3 checks this for the case for 802.1Q,
but not for 802.1ad.  As a result, full sized 802.1ad
frames get dropped by the card.

Add a check for 802.1ad protocol when receving full
sized frames.

Suggested-by: Prashant Sreedharan <prashant@broadcom.com>
CC: Prashant Sreedharan <prashant@broadcom.com>
CC: Michael Chan <mchan@broadcom.com>
Signed-off-by: Vladislav Yasevich <vyasevic@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agor8169: add support for Byte Queue Limits
Florian Westphal [Wed, 1 Oct 2014 11:38:03 +0000 (13:38 +0200)]
r8169: add support for Byte Queue Limits

tested on RTL8168d/8111d model using 'super_netperf 40' with TCP/UDP_STREAM.

Output of
while true; do
    for n in inflight limit; do
          echo -n $n\ ; cat $n;
    done;
    sleep 1;
done

during netperf run, 100mbit peer:

inflight 0
limit 3028
inflight 6056
limit 4542

[ trimmed output for brevity, no limit/inflight changes during
  test steady-state ]

limit 4542
inflight 3028
limit 6122
inflight 0
limit 6122
[ changed cable to 1gbit peer, restart netperf ]
inflight 37850
limit 36336
inflight 33308
limit 31794
inflight 33308
limit 31794
inflight 27252
limit 25738
[ again, no changes during test ]
inflight 27252
limit 25738
inflight 0
limit 28766
[ change cable to 100mbit peer, restart netperf ]
limit 28766
inflight 27370
limit 28766
inflight 4542
limit 5990
inflight 6056
limit 4542
[ .. ]
inflight 6056
limit 4542
inflight 0

[end of test]

Cc: Francois Romieu <romieu@fr.zoreil.com>
Cc: Hayes Wang <hayeswang@realtek.com>
Signed-off-by: Florian Westphal <fw@strlen.de>
Acked-by: Eric Dumazet <edumazet@google.com>
Acked-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: cleanup and document skb fclone layout
Eric Dumazet [Mon, 29 Sep 2014 20:29:15 +0000 (13:29 -0700)]
net: cleanup and document skb fclone layout

Lets use a proper structure to clearly document and implement
skb fast clones.

Then, we might experiment more easily alternative layouts.

This patch adds a new skb_fclone_busy() helper, used by tcp and xfrm,
to stop leaking of implementation details.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agotcp: abort orphan sockets stalling on zero window probes
Yuchung Cheng [Mon, 29 Sep 2014 20:20:38 +0000 (13:20 -0700)]
tcp: abort orphan sockets stalling on zero window probes

Currently we have two different policies for orphan sockets
that repeatedly stall on zero window ACKs. If a socket gets
a zero window ACK when it is transmitting data, the RTO is
used to probe the window. The socket is aborted after roughly
tcp_orphan_retries() retries (as in tcp_write_timeout()).

But if the socket was idle when it received the zero window ACK,
and later wants to send more data, we use the probe timer to
probe the window. If the receiver always returns zero window ACKs,
icsk_probes keeps getting reset in tcp_ack() and the orphan socket
can stall forever until the system reaches the orphan limit (as
commented in tcp_probe_timer()). This opens up a simple attack
to create lots of hanging orphan sockets to burn the memory
and the CPU, as demonstrated in the recent netdev post "TCP
connection will hang in FIN_WAIT1 after closing if zero window is
advertised." http://www.spinics.net/lists/netdev/msg296539.html

This patch follows the design in RTO-based probe: we abort an orphan
socket stalling on zero window when the probe timer reaches both
the maximum backoff and the maximum RTO. For example, an 100ms RTT
connection will timeout after roughly 153 seconds (0.3 + 0.6 +
.... + 76.8) if the receiver keeps the window shut. If the orphan
socket passes this check, but the system already has too many orphans
(as in tcp_out_of_resources()), we still abort it but we'll also
send an RST packet as the connection may still be active.

In addition, we change TCP_USER_TIMEOUT to cover (life or dead)
sockets stalled on zero-window probes. This changes the semantics
of TCP_USER_TIMEOUT slightly because it previously only applies
when the socket has pending transmission.

Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reported-by: Andrey Dmitrov <andrey.dmitrov@oktetlabs.ru>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge branch 'for-3.17' of git://linux-nfs.org/~bfields/linux
Linus Torvalds [Wed, 1 Oct 2014 20:22:00 +0000 (13:22 -0700)]
Merge branch 'for-3.17' of git://linux-nfs.org/~bfields/linux

Pull nfsd bugfix from Bruce Fields:
 "This fixes a data corruption bug introduced by the v3.16 xdr encoding
  rewrite.  I haven't managed to reproduce it myself yet, but it's
  apparently not hard to hit given the right workload"

* 'for-3.17' of git://linux-nfs.org/~bfields/linux:
  nfsd4: fix corruption of NFSv4 read data

10 years agocipso: add __init to cipso_v4_cache_init
Fabian Frederick [Wed, 1 Oct 2014 17:30:03 +0000 (19:30 +0200)]
cipso: add __init to cipso_v4_cache_init

cipso_v4_cache_init is only called by __init cipso_v4_init

Signed-off-by: Fabian Frederick <fabf@skynet.be>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoinet: frags: add __init to ip4_frags_ctl_register
Fabian Frederick [Wed, 1 Oct 2014 17:18:57 +0000 (19:18 +0200)]
inet: frags: add __init to ip4_frags_ctl_register

ip4_frags_ctl_register is only called by __init ipfrag_init

Signed-off-by: Fabian Frederick <fabf@skynet.be>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agotcp: add __init to tcp_init_mem
Fabian Frederick [Wed, 1 Oct 2014 16:27:50 +0000 (18:27 +0200)]
tcp: add __init to tcp_init_mem

tcp_init_mem is only called by __init tcp_init.

Signed-off-by: Fabian Frederick <fabf@skynet.be>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agor8169:call "rtl8168_driver_start" "rtl8168_driver_stop" only when hardware dash funct...
Chun-Hao Lin [Wed, 1 Oct 2014 15:17:21 +0000 (23:17 +0800)]
r8169:call "rtl8168_driver_start" "rtl8168_driver_stop" only when hardware dash function is enabled

These two functions are used to inform dash firmware that driver is been
brought up or brought down. So call these two functions only when hardware dash
function is enabled.

Signed-off-by: Chun-Hao Lin <hau@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agor8169:modify the behavior of function "rtl8168_oob_notify"
Chun-Hao Lin [Wed, 1 Oct 2014 15:17:20 +0000 (23:17 +0800)]
r8169:modify the behavior of function "rtl8168_oob_notify"

In function "rtl8168_oob_notify", using function "rtl_eri_write" to access
eri register 0xe8, instead of using MAC register "ERIDR" and "ERIAR" to
access it.

For using function "rtl_eri_write" in function "rtl8168_oob_notify", need to
move down "rtl8168_oob_notify" related functions under the function
"rtl_eri_write".

Signed-off-by: Chun-Hao Lin <hau@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agor8169:change the name of function "r8168dp_check_dash" to "r8168_check_dash"
Chun-Hao Lin [Wed, 1 Oct 2014 15:17:19 +0000 (23:17 +0800)]
r8169:change the name of function "r8168dp_check_dash" to "r8168_check_dash"

DASH function not only RTL8168DP can support, but also RTL8168EP.
So change the name of function "r8168dp_check_dash" to "r8168_check_dash".

Signed-off-by: Chun-Hao Lin <hau@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agor8169:change the name of function"rtl_w1w0_eri"
Chun-Hao Lin [Wed, 1 Oct 2014 15:17:18 +0000 (23:17 +0800)]
r8169:change the name of function"rtl_w1w0_eri"

Change the name of function "rtl_w1w0_eri" to "rtl_w0w1_eri".

In this function, the local variable "val" is "write zeros then write ones".
Please see below code.

(val & ~m) | p

In this patch, change the function name from "xx_w1w0_xx" to "xx_w0w1_xx".
The changed function name is more suitable for it's behavior.

Signed-off-by: Chun-Hao Lin <hau@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agor8169:for function "rtl_w1w0_phy" change its name and behavior
Chun-Hao Lin [Wed, 1 Oct 2014 15:17:17 +0000 (23:17 +0800)]
r8169:for function "rtl_w1w0_phy" change its name and behavior

Change function name from "rtl_w1w0_phy" to "rtl_w0w1_phy".
And its behavior from "write ones then write zeros" to
"write zeros then write ones".

In Realtek internal driver, bitwise operations are almost "write zeros then
write ones". For easy to port hardware parameters from Realtek internal driver
to Linux kernal driver "r8169", we would like to change this function's
behavior and its name.

Signed-off-by: Chun-Hao Lin <hau@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agor8169:add more chips to support magic packet v2
Chun-Hao Lin [Wed, 1 Oct 2014 15:17:16 +0000 (23:17 +0800)]
r8169:add more chips to support magic packet v2

For RTL8168F RTL8168FB RTL8168G RTL8168GU RTL8411 RTL8411B RTL8402 RTL8107E,
the magic packet enable bit is changed to eri 0xde bit0.

In this patch, change magic packet enable bit of these chips to eri 0xde bit0.

Signed-off-by: Chun-Hao Lin <hau@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agor8169:add support more chips to get mac address from backup mac address register
Chun-Hao Lin [Wed, 1 Oct 2014 15:17:15 +0000 (23:17 +0800)]
r8169:add support more chips to get mac address from backup mac address register

RTL8168FB RTL8168G RTL8168GU RTL8411 RTL8411B RTL8106EUS RTL8402 can
support get mac address from backup mac address register.

Signed-off-by: Chun-Hao Lin <hau@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agor8169:add disable/enable RTL8411B pll function
Chun-Hao Lin [Wed, 1 Oct 2014 15:17:14 +0000 (23:17 +0800)]
r8169:add disable/enable RTL8411B pll function

RTL8411B can support disable/enable pll function.

Signed-off-by: Chun-Hao Lin <hau@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agor8169:add disable/enable RTL8168G pll function
Chun-Hao Lin [Wed, 1 Oct 2014 15:17:13 +0000 (23:17 +0800)]
r8169:add disable/enable RTL8168G pll function

RTL8168G also can disable/enable pll function.

Signed-off-by: Chun-Hao Lin <hau@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agor8169:change uppercase number to lowercase number
Chun-Hao Lin [Wed, 1 Oct 2014 15:17:12 +0000 (23:17 +0800)]
r8169:change uppercase number to lowercase number

Signed-off-by: Chun-Hao Lin <hau@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agosunvnet: fix potential NULL pointer dereference
David L Stevens [Wed, 1 Oct 2014 15:05:27 +0000 (11:05 -0400)]
sunvnet: fix potential NULL pointer dereference

One of the error cases for vnet_start_xmit()'s "out_dropped" label
is port == NULL, so only mess with port->clean_timer when port is not NULL.

Signed-off-by: David L Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: dsa: Fix build warning for !PM_SLEEP
Thierry Reding [Wed, 1 Oct 2014 11:59:00 +0000 (13:59 +0200)]
net: dsa: Fix build warning for !PM_SLEEP

The dsa_switch_suspend() and dsa_switch_resume() functions are only used
when PM_SLEEP is enabled, so they need #ifdef CONFIG_PM_SLEEP protection
to avoid a compiler warning.

Signed-off-by: Thierry Reding <treding@nvidia.com>
Acked-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: ll_temac: Remove unnecessary ether_setup after alloc_etherdev
Subbaraya Sundeep Bhatta [Wed, 1 Oct 2014 09:01:17 +0000 (11:01 +0200)]
net: ll_temac: Remove unnecessary ether_setup after alloc_etherdev

Calling ether_setup is redundant since alloc_etherdev calls it.

Signed-off-by: Subbaraya Sundeep Bhatta <sbhatta@xilinx.com>
Signed-off-by: Michal Simek <michal.simek@xilinx.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoipv4: mentions skb_gro_postpull_rcsum() in inet_gro_receive()
Eric Dumazet [Wed, 1 Oct 2014 05:12:05 +0000 (22:12 -0700)]
ipv4: mentions skb_gro_postpull_rcsum() in inet_gro_receive()

Proper CHECKSUM_COMPLETE support needs to adjust skb->csum
when we remove one header. Its done using skb_gro_postpull_rcsum()

In the case of IPv4, we know that the adjustment is not really needed,
because the checksum over IPv4 header is 0. Lets add a comment to
ease code comprehension and avoid copy/paste errors.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agofm10k: using vmalloc requires including linux/vmalloc.h
Stephen Rothwell [Wed, 1 Oct 2014 07:00:49 +0000 (17:00 +1000)]
fm10k: using vmalloc requires including linux/vmalloc.h

Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Acked-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoieee802154: fix __init functions
Fabian Frederick [Wed, 1 Oct 2014 05:27:46 +0000 (07:27 +0200)]
ieee802154: fix __init functions

Commit 3243acd37fd9
("ieee802154: add __init to lowpan_frags_sysctl_register")

added __init to lowpan_frags_ns_sysctl_register instead of
lowpan_frags_sysctl_register

Suggested-by: Alexander Aring <alex.aring@gmail.com>
Signed-off-by: Fabian Frederick <fabf@skynet.be>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge branch 'fixes' of git://ftp.arm.linux.org.uk/~rmk/linux-arm
Linus Torvalds [Wed, 1 Oct 2014 02:52:08 +0000 (19:52 -0700)]
Merge branch 'fixes' of git://ftp.arm.linux.org.uk/~rmk/linux-arm

Pull ARM fixes from Russell King:
 "Some further ARM fixes:
   - another build fix for the kprobes test code
   - a fix for no kuser helpers for the set_tls code, which oopsed on
     noMMU hardware
   - a fix for alignment handler with neon opcodes being misinterpreted
   - turning off the hardware access support, which is not implemented
   - a build fix for the v7 coherency exiting code, which can be built
     in non-v7 environments (but still only executed on v7 CPUs)"

* 'fixes' of git://ftp.arm.linux.org.uk/~rmk/linux-arm:
  ARM: 8179/1: kprobes-test: Fix compile error "bad immediate value for offset"
  ARM: 8178/1: fix set_tls for !CONFIG_KUSER_HELPERS
  ARM: 8177/1: cacheflush: Fix v7_exit_coherency_flush exynos build breakage on ARMv6
  ARM: 8165/1: alignment: don't break misaligned NEON load/store
  ARM: 8164/1: mm: clear SCTLR.HA instead of setting it for LPAE

10 years agoMerge branch 'sunvnet-jumbograms'
David S. Miller [Tue, 30 Sep 2014 21:10:47 +0000 (17:10 -0400)]
Merge branch 'sunvnet-jumbograms'

David L Stevens says:

====================
sunvnet: add jumbo frames support

This patch set updates the sunvnet driver to version 1.6 of the VIO protocol
to support per-port exchange of MTU information and allow non-standard MTU
sizes, including jumbo frames.

Using large MTUs shows a nearly 5X throughput improvement Linux-Solaris
and > 10X throughput improvement Linux-Linux.

Changes from v8:
-add a short timeout to free pending skbs if a new transmit doesn't
 do it first per Dave Miller <davem@davemloft.net>
Changes from v7:
-handle skb allocation failures in vnet_skb_shape()
 per Dave Miller <davem@davemloft.net>
Changes from v6:
-made kernel transmit path zero-copy to remove memory n^2 scaling issue
 raised by Raghuram Kothakota <Raghuram.Kothakota@oracle.com>
Changes from v5:
- fixed comment per Sowmini Varadhan <sowmini.varadhan@oracle.com>
Changes from v4:
- changed VNET_MAXPACKET per David Laight <David.Laight@ACULAB.COM>
- added cookies to support non-contiguous buffers of max size
Changes from v3:
- added version functions per Dave Miller <davem@davemloft.net>
- moved rmtu to vnet_port per Dave Miller <davem@davemloft.net>
- explicitly set options bits and capability flags to 0 per
Raghuram Kothakota <Raghuram.Kothakota@oracle.com>
Changes from v2:
- make checkpatch clean
Changes from v1:
- fix brace formatting per Dave Miller <davem@davemloft.net>
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agosunvnet: generate ICMP PTMUD messages for smaller port MTUs
David L Stevens [Mon, 29 Sep 2014 23:48:24 +0000 (19:48 -0400)]
sunvnet: generate ICMP PTMUD messages for smaller port MTUs

This patch sends ICMP and ICMPv6 messages for Path MTU Discovery when a remote
port MTU is smaller than the device MTU. This allows mixing newer VIO protocol
devices that support MTU negotiation with older devices that do not on the
same vswitch. It also allows Linux-Linux LDOMs to use 64K-1 data packets even
though Solaris vswitch is limited to <16K MTU.

Signed-off-by: David L Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agosunvnet: allow admin to set sunvnet MTU
David L Stevens [Mon, 29 Sep 2014 23:48:18 +0000 (19:48 -0400)]
sunvnet: allow admin to set sunvnet MTU

This patch allows an admin to set the MTU on a sunvnet device to arbitrary
values between the minimum (68) and maximum (65535) IPv4 packet sizes.

Signed-off-by: David L Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agosunvnet: make transmit path zero-copy in the kernel
David L Stevens [Mon, 29 Sep 2014 23:48:11 +0000 (19:48 -0400)]
sunvnet: make transmit path zero-copy in the kernel

This patch removes pre-allocated transmit buffers and instead directly maps
pending packets on demand. This saves O(n^2) maximum-sized transmit buffers,
for n hosts on a vswitch, as well as a copy to those buffers.

Single-stream TCP throughput linux-solaris dropped ~5% for 1500-byte MTU,
but linux-linux at 1500-bytes increased ~20%.

Signed-off-by: David L Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agosunvnet: upgrade to VIO protocol version 1.6
David L Stevens [Mon, 29 Sep 2014 23:47:59 +0000 (19:47 -0400)]
sunvnet: upgrade to VIO protocol version 1.6

This patch upgrades the sunvnet driver to support VIO protocol version 1.6.
In particular, it adds per-port MTU negotiation, allowing MTUs other than
ETH_FRAMELEN with ports using newer VIO protocol versions.

Signed-off-by: David L Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agotcp: Change tcp_slow_start function to return void
Li RongQing [Tue, 30 Sep 2014 01:49:55 +0000 (09:49 +0800)]
tcp: Change tcp_slow_start function to return void

No caller uses the return value, so make this function return void.

Signed-off-by: Li RongQing <roy.qing.li@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoieee802154: add __init to lowpan_frags_sysctl_register
Fabian Frederick [Tue, 30 Sep 2014 20:34:08 +0000 (22:34 +0200)]
ieee802154: add __init to lowpan_frags_sysctl_register

lowpan_frags_sysctl_register is only called by __init lowpan_net_frag_init
(part of the lowpan module).

Signed-off-by: Fabian Frederick <fabf@skynet.be>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoirda: add __init to irlan_open
Fabian Frederick [Tue, 30 Sep 2014 20:24:04 +0000 (22:24 +0200)]
irda: add __init to irlan_open

irlan_open is only called by __init irlan_init in same module.

Signed-off-by: Fabian Frederick <fabf@skynet.be>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonext: mips: bpf: Fix build failure
Guenter Roeck [Tue, 30 Sep 2014 16:39:00 +0000 (09:39 -0700)]
next: mips: bpf: Fix build failure

Fix:

arch/mips/net/bpf_jit.c: In function 'build_body':
arch/mips/net/bpf_jit.c:762:6: error: unused variable 'tmp'
cc1: all warnings being treated as errors
make[2]: *** [arch/mips/net/bpf_jit.o] Error 1

Seen when building mips:allmodconfig in -next since next-20140924.

Signed-off-by: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge branch 'pxa168_eth'
David S. Miller [Tue, 30 Sep 2014 20:37:13 +0000 (16:37 -0400)]
Merge branch 'pxa168_eth'

Antoine Tenart says:

====================
ARM: Berlin: Ethernet support

This series introduce support for the Ethernet controller on Berlin SoCs,
using the existing pxa168 Ethernet driver. In order to do this, DT
support is added to the driver alongside some other modifications and
fixes.

This has been tested on a Berlin BG2Q DMP board.

Changes since v5:
- fixed the build when building the driver as a module

Changes since v4:
        - removed the phy-addr property and added a phy subnode
        - added COMPILE_TEST for the pxa168_eth driver

Changes since v3:
        - moved the addition of pxa168_eth_get_mac_address() to the patch
          using it first

Changes since v2:
        - reworked how the MAC address is configured
        - made the clock anonymous

Changes since v1:
        - removed custom Berlin Ethernet driver
        - used the pxa168 Ethernet driver instead
        - made modifications to the pxa168 driver (DT support, fixes)
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoARM: dts: berlin: enable the Ethernet port on the BG2Q DMP
Antoine Ténart [Tue, 30 Sep 2014 14:28:16 +0000 (16:28 +0200)]
ARM: dts: berlin: enable the Ethernet port on the BG2Q DMP

This patch enables the Ethernet port on the Marvell Berlin2Q DMP board.

Signed-off-by: Antoine Tenart <antoine.tenart@free-electrons.com>
Acked-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoARM: dts: berlin: add the Ethernet node
Antoine Ténart [Tue, 30 Sep 2014 14:28:15 +0000 (16:28 +0200)]
ARM: dts: berlin: add the Ethernet node

This patch adds the Ethernet node, enabling the network unit on Berlin
BG2Q SoCs.

Signed-off-by: Antoine Tenart <antoine.tenart@free-electrons.com>
Acked-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: David S. Miller <davem@davemloft.net>