GitHub/moto-9609/android_kernel_motorola_exynos9610.git
10 years agoopenvswitch: Add support for OVS_FLOW_ATTR_PROBE.
Jarno Rajahalme [Thu, 6 Nov 2014 15:03:05 +0000 (07:03 -0800)]
openvswitch: Add support for OVS_FLOW_ATTR_PROBE.

This new flag is useful for suppressing error logging while probing
for datapath features using flow commands.  For backwards
compatibility reasons the commands are executed normally, but error
logging is suppressed.

Signed-off-by: Jarno Rajahalme <jrajahalme@nicira.com>
Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
10 years agoopenvswitch: Constify various function arguments
Thomas Graf [Thu, 6 Nov 2014 14:58:52 +0000 (06:58 -0800)]
openvswitch: Constify various function arguments

Help produce better optimized code.

Signed-off-by: Thomas Graf <tgraf@noironetworks.com>
Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
10 years agoopenvswitch: Remove redundant key ref from upcall_info.
Pravin B Shelar [Thu, 6 Nov 2014 14:57:27 +0000 (06:57 -0800)]
openvswitch: Remove redundant key ref from upcall_info.

struct dp_upcall_info has pointer to pkt_key which is already
available in OVS_CB.  This also simplifies upcall handling
for gso packet.

Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
Acked-by: Andy Zhou <azhou@nicira.com>
10 years agoopenvswitch: Optimize recirc action.
Pravin B Shelar [Thu, 6 Nov 2014 14:55:14 +0000 (06:55 -0800)]
openvswitch: Optimize recirc action.

OVS need to flow key for flow lookup in recic action. OVS
does key extract in recic action. Most of cases we could
use OVS_CB packet key directly and can avoid packet flow key
extract. SET action we can update flow-key along with packet
to keep it consistent. But there are some action like MPLS
pop which forces OVS to do flow-extract. In such cases we
can mark flow key as invalid so that subsequent recirc
action can do full flow extract.

Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
Acked-by: Jarno Rajahalme <jrajahalme@nicira.com>
Acked-by: Andy Zhou <azhou@nicira.com>
10 years agoopenvswitch: Extend packet attribute for egress tunnel info
Wenyu Zhang [Thu, 6 Nov 2014 14:51:24 +0000 (06:51 -0800)]
openvswitch: Extend packet attribute for egress tunnel info

OVS vswitch has extended IPFIX exporter to export tunnel headers
to improve network visibility.
To export this information userspace needs to know egress tunnel
for given packet. By extending packet attributes datapath can
export egress tunnel info for given packet. So that userspace
can ask for egress tunnel info in userspace action. This
information is used to build IPFIX data for given flow.

Signed-off-by: Wenyu Zhang <wenyuz@vmware.com>
Acked-by: Romain Lenglet <rlenglet@vmware.com>
Acked-by: Ben Pfaff <blp@nicira.com>
Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
10 years agoopenvswitch: Export symbols as GPL symbols.
Pravin B Shelar [Thu, 6 Nov 2014 14:44:27 +0000 (06:44 -0800)]
openvswitch: Export symbols as GPL symbols.

vport can be compiled as modules, therefore openvswitch needs
to export few symbols. Export them as GPL symbols.

CC: Thomas Graf <tgraf@noironetworks.com>
Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
10 years agodccp: Convert DCCP_WARN to net_warn_ratelimited
Joe Perches [Thu, 6 Nov 2014 20:53:41 +0000 (12:53 -0800)]
dccp: Convert DCCP_WARN to net_warn_ratelimited

Remove the dependency on the "warning" sysctl (net_msg_warn)
which is only used by the LIMIT_NETDEBUG macro.

Convert the LIMIT_NETDEBUG use in DCCP_WARN to the more
common net_warn_ratelimited mechanism.

This still ratelimits based on the net_ratelimit()
function, but removes the check for the sysctl.

Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoudp: Increment UDP_MIB_IGNOREDMULTI for arriving unmatched multicasts
Rick Jones [Thu, 6 Nov 2014 18:37:54 +0000 (10:37 -0800)]
udp: Increment UDP_MIB_IGNOREDMULTI for arriving unmatched multicasts

As NIC multicast filtering isn't perfect, and some platforms are
quite content to spew broadcasts, we should not trigger an event
for skb:kfree_skb when we do not have a match for such an incoming
datagram.  We do though want to avoid sweeping the matter under the
rug entirely, so increment a suitable statistic.

This incorporates feedback from David L. Stevens, Karl Neiss and Eric
Dumazet.

V3 - use bool per David Miller

Signed-off-by: Rick Jones <rick.jones2@hp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agocdc-ether: implement MULTICAST flag on the device
Oliver Neukum [Thu, 6 Nov 2014 14:19:14 +0000 (15:19 +0100)]
cdc-ether: implement MULTICAST flag on the device

Olivier having laid the groundwork this patch transmits the
multicast flag to the device to save some bus traffic.

Signed-off-by: Oliver Neukum <oneukum@suse.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agouapi: resort Kbuild entries
stephen hemminger [Mon, 3 Nov 2014 20:42:34 +0000 (12:42 -0800)]
uapi: resort Kbuild entries

The entries in the Kbuild files are incorrectly sorted.
Matters for aesthetics only.

Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agostmmac: platform: fix sparse warnings
Andy Shevchenko [Fri, 7 Nov 2014 14:46:42 +0000 (16:46 +0200)]
stmmac: platform: fix sparse warnings

This patch fixes the following sparse warnings. One is fixed by casting return
value to a return type of the function. The others by creating a specific
stmmac_platform.h which provides the bits related to the platform driver.

drivers/net/ethernet/stmicro/stmmac/dwmac-meson.c:59:29: warning: incorrect type in return expression (different address spaces)
drivers/net/ethernet/stmicro/stmmac/dwmac-meson.c:59:29:    expected void *
drivers/net/ethernet/stmicro/stmmac/dwmac-meson.c:59:29:    got void [noderef] <asn:2>*reg

drivers/net/ethernet/stmicro/stmmac/dwmac-meson.c:64:29: warning: symbol 'meson6_dwmac_data' was not declared. Should it be static?
drivers/net/ethernet/stmicro/stmmac/dwmac-sti.c:354:29: warning: symbol 'stih4xx_dwmac_data' was not declared. Should it be static?
drivers/net/ethernet/stmicro/stmmac/dwmac-sti.c:361:29: warning: symbol 'stid127_dwmac_data' was not declared. Should it be static?
drivers/net/ethernet/stmicro/stmmac/dwmac-sunxi.c:133:29: warning: symbol 'sun7i_gmac_data' was not declared. Should it be static?

Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agostmmac: remove custom implementation of print_hex_dump()
Andy Shevchenko [Fri, 7 Nov 2014 14:53:12 +0000 (16:53 +0200)]
stmmac: remove custom implementation of print_hex_dump()

There is a kernel helper to dump buffers in a hexdecimal format. This patch
substitutes the open coded function by calling that helper.

The output is slightly changed:
 - no lead space
 - ASCII part will be printed along with the dump
 - offset is longer than 3 characters (now 8)

Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge branch 'iov_iter'
David S. Miller [Fri, 7 Nov 2014 17:13:40 +0000 (12:13 -0500)]
Merge branch 'iov_iter'

Herbert Xu says:

====================
Replace skb_copy_datagram_const_iovec with iterator version

This patch series adds the helper skb_copy_datagram_iter, which
is meant to replace both skb_copy_datagram_iovec and its evil
twin skb_copy_datagram_const_iovec.

It then converts tun and macvtap over to the new helper and finally
removes skb_copy_datagram_const_iovec which is only used by tun
and macvtap.

The copy_to_iter return value issue pointed out by Al has now been
fixed.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: Kill skb_copy_datagram_const_iovec
Herbert Xu [Fri, 7 Nov 2014 13:22:26 +0000 (21:22 +0800)]
net: Kill skb_copy_datagram_const_iovec

Now that both macvtap and tun are using skb_copy_datagram_iter, we
can kill the abomination that is skb_copy_datagram_const_iovec.

Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agomacvtap: Use iovec iterators
Herbert Xu [Fri, 7 Nov 2014 13:22:25 +0000 (21:22 +0800)]
macvtap: Use iovec iterators

This patch removes the use of skb_copy_datagram_const_iovec in
favour of the iovec iterator-based skb_copy_datagram_iter.

Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agotun: Use iovec iterators
Herbert Xu [Fri, 7 Nov 2014 13:22:23 +0000 (21:22 +0800)]
tun: Use iovec iterators

This patch removes the use of skb_copy_datagram_const_iovec in
favour of the iovec iterator-based skb_copy_datagram_iter.

Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoinet: Add skb_copy_datagram_iter
Herbert Xu [Fri, 7 Nov 2014 13:22:22 +0000 (21:22 +0800)]
inet: Add skb_copy_datagram_iter

This patch adds skb_copy_datagram_iter, which is identical to
skb_copy_datagram_iovec except that it operates on iov_iter
instead of iovec.

Eventually all users of skb_copy_datagram_iovec should switch
over to iov_iter and then we can remove skb_copy_datagram_iovec.

Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
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 [Fri, 7 Nov 2014 03:01:18 +0000 (22:01 -0500)]
Merge git://git./linux/kernel/git/davem/net

10 years agovxlan: Fix to enable UDP checksums on interface
Tom Herbert [Fri, 7 Nov 2014 02:06:01 +0000 (18:06 -0800)]
vxlan: Fix to enable UDP checksums on interface

Add definition to vxlan nla_policy for UDP checksum. This is necessary
to enable UDP checksums on VXLAN.

In some instances, enabling UDP checksums can improve performance on
receive for devices that return legacy checksum-unnecessary for UDP/IP.
Also, UDP checksum provides some protection against VNI corruption.

Testing:

Ran 200 instances of TCP_STREAM and TCP_RR on bnx2x.

TCP_STREAM
  IPv4, without UDP checksums
      14.41% TX CPU utilization
      25.71% RX CPU utilization
      9083.4 Mbps
  IPv4, with UDP checksums
      13.99% TX CPU utilization
      13.40% RX CPU utilization
      9095.65 Mbps

TCP_RR
  IPv4, without UDP checksums
      94.08% TX CPU utilization
      156/248/462 90/95/99% latencies
      1.12743e+06 tps
  IPv4, with UDP checksums
      94.43% TX CPU utilization
      158/250/462 90/95/99% latencies
      1.13345e+06 tps

Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge branch 'amd-xgbe-next'
David S. Miller [Fri, 7 Nov 2014 00:00:26 +0000 (19:00 -0500)]
Merge branch 'amd-xgbe-next'

Tom Lendacky says:

====================
amd-xgbe: AMD XGBE driver updates 2014-11-06

The following series of patches fixes a couple of bugs that slipped
through my last series.

- Free channel structure after freeing the per channel interrupts
- If an skb error allocation occurs during receive processing check
  whether more descriptors are associated with the packet or whether
  to start on a new packet

This patch series is based on net-next.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoamd-xgbe: Check for complete packet on skb allocation error
Lendacky, Thomas [Thu, 6 Nov 2014 23:02:19 +0000 (17:02 -0600)]
amd-xgbe: Check for complete packet on skb allocation error

If the skb allocation fails during receive processing, the driver would
continue reading descriptors without first determining if there were
any more descriptors for the current packet. Update the code to check
whether more descriptors are associated with the current packet or
whether to move on to the next descriptor as a new packet.

Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoamd-xgbe: Free channel/ring structures later
Lendacky, Thomas [Thu, 6 Nov 2014 23:02:13 +0000 (17:02 -0600)]
amd-xgbe: Free channel/ring structures later

The channel structure is freed before freeing the per channel
interrupts resulting in a kernel oops. Move the call to free
the channel structure to after the freeing of the per channel
interrupts.

Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonetxen: Fix link event handling.
Manish Chopra [Thu, 6 Nov 2014 12:58:51 +0000 (07:58 -0500)]
netxen: Fix link event handling.

o Poll for the link events only if firmware doesn't have capability
  to notify the driver for the link events.

Signed-off-by: Manish Chopra <manish.chopra@qlogic.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoenic: update desc properly in rx_copybreak
Govindarajulu Varadarajan [Thu, 6 Nov 2014 09:51:39 +0000 (15:21 +0530)]
enic: update desc properly in rx_copybreak

When we reuse the rx buffer, we need to update the desc. If not hardware sees
stale value.

In the following crash, when mtu is changed, hardware sees old rx buffer value
and crashes on skb_put.

Fix this by using enic_queue_rq_desc helper function which updates the necessary
desc.

[   64.657376] skbuff: skb_over_panic: text:ffffffffa041f55d len:9010 put:9010 head:ffff8800d3ca9fc0 data:ffff8800d3caa000 tail:0x2372 end:0x640 dev:enp0s3
[   64.659965] ------------[ cut here ]------------
[   64.661322] kernel BUG at net/core/skbuff.c:100!
[   64.662644] invalid opcode: 0000 [#1] PREEMPT SMP
[   64.664001] Modules linked in: rpcsec_gss_krb5 auth_rpcgss oid_registry nfsv4 cirrus ttm drm_kms_helper drm enic psmouse microcode evdev serio_raw syscopyarea sysfillrect sysimgblt i2c_piix4 i2c_core pcspkr nfs lockd grace sunrpc fscache ext4 crc16 mbcache jbd2 sd_mod ata_generic virtio_balloon ata_piix libata uhci_hcd virtio_pci virtio_ring usbcore usb_common virtio scsi_mod
[   64.664834] CPU: 0 PID: 0 Comm: swapper/0 Tainted: G        W      3.17.0-netnext-10335-g942396b-dirty #273
[   64.664834] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
[   64.664834] task: ffffffff81a1d580 ti: ffffffff81a00000 task.ti: ffffffff81a00000
[   64.664834] RIP: 0010:[<ffffffff81392cf1>]  [<ffffffff81392cf1>] skb_panic+0x61/0x70
[   64.664834] RSP: 0018:ffff880210603d48  EFLAGS: 00010292
[   64.664834] RAX: 000000000000008c RBX: ffff88020b0f6930 RCX: 0000000000000000
[   64.664834] RDX: 000000000000008c RSI: ffffffff8178b288 RDI: 00000000ffffffff
[   64.664834] RBP: ffff880210603d68 R08: 0000000000000001 R09: 0000000000000001
[   64.664834] R10: 00000000000005ce R11: 0000000000000001 R12: ffff88020b1f0b40
[   64.664834] R13: 000000000000a332 R14: ffff880209a1a000 R15: 0000000000000001
[   64.664834] FS:  0000000000000000(0000) GS:ffff880210600000(0000) knlGS:0000000000000000
[   64.664834] CS:  0010 DS: 0000 ES: 0000 CR0: 000000008005003b
[   64.664834] CR2: 00007f6752935e48 CR3: 0000000035743000 CR4: 00000000000006f0
[   64.664834] Stack:
[   64.664834]  ffff8800d3caa000 0000000000002372 0000000000000640 ffff88020b1f0000
[   64.664834]  ffff880210603d78 ffffffff81392d54 ffff880210603e08 ffffffffa041f55d
[   64.664834]  0000000000000296 ffffffff00000000 00008e7e00008e7e ffff880200002332
[   64.664834] Call Trace:
[   64.664834]  <IRQ>
[   64.664834]
[   64.664834]  [<ffffffff81392d54>] skb_put+0x54/0x60
[   64.664834]  [<ffffffffa041f55d>] enic_rq_service.constprop.47+0x3ad/0x730 [enic]
[   64.664834]  [<ffffffffa041fa79>] enic_poll_msix_rq+0x199/0x370 [enic]
[   64.664834]  [<ffffffff813a5499>] net_rx_action+0x139/0x210
[   64.664834]  [<ffffffff81290db3>] ? __this_cpu_preempt_check+0x13/0x20
[   64.664834]  [<ffffffff8106110e>] __do_softirq+0x14e/0x280
[   64.664834]  [<ffffffff8106152e>] irq_exit+0x8e/0xb0
[   64.664834]  [<ffffffff8100fd21>] do_IRQ+0x61/0x100
[   64.664834]  [<ffffffff814a2bf2>] common_interrupt+0x72/0x72

fixes: a03bb56e67c357980dae886683733dab5583dc14 ("enic: implement rx_copybreak")
Signed-off-by: Govindarajulu Varadarajan <_govind@gmx.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoenic: handle error condition properly in enic_rq_indicate_buf
Govindarajulu Varadarajan [Thu, 6 Nov 2014 09:51:38 +0000 (15:21 +0530)]
enic: handle error condition properly in enic_rq_indicate_buf

In case of error in rx path, we free the buf->os_buf but we do not make it NULL.
In next iteration we use the skb which is already freed. This causes the
following crash.

[  886.154772] general protection fault: 0000 [#1] PREEMPT SMP
[  886.154851] Modules linked in: rpcsec_gss_krb5 auth_rpcgss oid_registry nfsv4 microcode evdev cirrus ttm drm_kms_helper drm enic syscopyarea sysfillrect sysimgblt psmouse i2c_piix4 serio_raw pcspkr i2c_core nfs lockd grace sunrpc fscache ext4 crc16 mbcache jbd2 sd_mod crc_t10dif crct10dif_common ata_generic ata_piix virtio_balloon libata scsi_mod uhci_hcd usbcore virtio_pci virtio_ring virtio usb_common
[  886.155199] CPU: 0 PID: 0 Comm: swapper/0 Tainted: G        W      3.17.0-netnext-05668-g876bc7f #272
[  886.155263] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
[  886.155304] task: ffffffff81a1d580 ti: ffffffff81a00000 task.ti: ffffffff81a00000
[  886.155356] RIP: 0010:[<ffffffff81384030>]  [<ffffffff81384030>] kfree_skb_list+0x10/0x30
[  886.155418] RSP: 0018:ffff880210603d48  EFLAGS: 00010206
[  886.155456] RAX: 0000000000000020 RBX: 0000000000000000 RCX: 0000000000000000
[  886.155504] RDX: 0000000000000000 RSI: 0000000000000001 RDI: 004500084e000017
[  886.155553] RBP: ffff880210603d50 R08: 00000000fe13d1b6 R09: 0000000000000001
[  886.155601] R10: 0000000000000000 R11: 0000000000000000 R12: ffff880209ff2f00
[  886.155650] R13: ffff88020ac0fe40 R14: ffff880209ff2f00 R15: ffff8800da8e3a80
[  886.155699] FS:  0000000000000000(0000) GS:ffff880210600000(0000) knlGS:0000000000000000
[  886.155774] CS:  0010 DS: 0000 ES: 0000 CR0: 000000008005003b
[  886.155814] CR2: 00007f0e0c925000 CR3: 0000000035e8b000 CR4: 00000000000006f0
[  886.155865] Stack:
[  886.155882]  0000000000000000 ffff880210603d78 ffffffff81383f79 ffff880209ff2f00
[  886.155942]  ffff88020b0c0b40 000000000000c000 ffff880210603d90 ffffffff81383faf
[  886.156001]  ffff880209ff2f00 ffff880210603da8 ffffffff8138406d ffff88020b1b08c0
[  886.156061] Call Trace:
[  886.156080]  <IRQ>
[  886.156095]
[  886.156112]  [<ffffffff81383f79>] skb_release_data+0xa9/0xc0
[  886.157656]  [<ffffffff81383faf>] skb_release_all+0x1f/0x30
[  886.159195]  [<ffffffff8138406d>] consume_skb+0x1d/0x40
[  886.160719]  [<ffffffff813942e5>] __dev_kfree_skb_any+0x35/0x40
[  886.162224]  [<ffffffffa02dc1d5>] enic_rq_service.constprop.47+0xe5/0x5a0 [enic]
[  886.163756]  [<ffffffffa02dc829>] enic_poll_msix_rq+0x199/0x370 [enic]
[  886.164730]  [<ffffffff81397e29>] net_rx_action+0x139/0x210
[  886.164730]  [<ffffffff8105fb2e>] __do_softirq+0x14e/0x280
[  886.164730]  [<ffffffff8105ff2e>] irq_exit+0x8e/0xb0
[  886.164730]  [<ffffffff8100fc1d>] do_IRQ+0x5d/0x100
[  886.164730]  [<ffffffff81496832>] common_interrupt+0x72/0x72

fixes: a03bb56e67c357980dae886683733dab5583dc14 ("enic: implement rx_copybreak")
Signed-off-by: Govindarajulu Varadarajan <_govind@gmx.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge branch 'mlx5-net'
David S. Miller [Thu, 6 Nov 2014 21:40:47 +0000 (16:40 -0500)]
Merge branch 'mlx5-net'

Eli Cohen says:

====================
mlx5_core fixes for 3.18

the following two patches fix races to could lead to kernel panic in some cases.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet/mlx5_core: Fix race on driver load
Eli Cohen [Thu, 6 Nov 2014 10:51:22 +0000 (12:51 +0200)]
net/mlx5_core: Fix race on driver load

When events arrive at driver load, the event handler gets called even before
the spinlock and list are initialized. Fix this by moving the initialization
before EQs creation.

Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet/mlx5_core: Fix race in create EQ
Eli Cohen [Thu, 6 Nov 2014 10:51:21 +0000 (12:51 +0200)]
net/mlx5_core: Fix race in create EQ

After the EQ is created, it can possibly generate interrupts and the interrupt
handler is referencing eq->dev. It is therefore required to set eq->dev before
calling request_irq() so if an event is generated before request_irq() returns,
we will have a valid eq->dev field.

Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge branch 'net_next_ovs' of git://git.kernel.org/pub/scm/linux/kernel/git/pshelar...
David S. Miller [Thu, 6 Nov 2014 21:33:13 +0000 (16:33 -0500)]
Merge branch 'net_next_ovs' of git://git./linux/kernel/git/pshelar/openvswitch

Pravin B Shelar says:

====================
Open vSwitch

First two patches are related to OVS MPLS support. Rest of patches
are mostly refactoring and minor improvements to openvswitch.

v1-v2:
 - Fix conflicts due to "gue: Remote checksum offload"
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge branch 'sunvnet-next'
David S. Miller [Thu, 6 Nov 2014 20:16:44 +0000 (15:16 -0500)]
Merge branch 'sunvnet-next'

Sowmini Varadhan says:

====================
sunvnet: bug fixes

This patch series has a coding-style fix and a bug fix.

The coding style fix (patch 1) is the extra indentation flagged by
Ben Hutchings:
  http://marc.info/?l=linux-netdev&m=141529243409594&w=2

The bugfix (patch 2) is the following:
when vnet_event_napi() is  called as part of napi_resume
(i.e., continuation of a previous NAPI read that was truncated
due to budget constraints), and then finds no more packets to read,
the code was trying to avoid an additional trip through ldc_rx
as an optimization. However, when this corner case happens, we would
need to reset a number of dring state bits such as rcv_nxt carefully,
which quickly becomes complex and hacky.  The cleaner solution
is to just roll back to vnet_poll, re-enable interrupts and set up
dring state as was done in the pre-NAPI version of the driver.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agosunvnet: Return from vnet_napi_event() if no packets to read
Sowmini Varadhan [Thu, 6 Nov 2014 19:51:08 +0000 (14:51 -0500)]
sunvnet: Return from vnet_napi_event() if no packets to read

vnet_event_napi() may be called as part of the NAPI ->poll,
to resume reading descriptor rings. When no data is available,
descriptor ring state (e.g., rcv_nxt) needs to be reset
carefully to stay in lock-step with ldc_read(). In the interest
of simplicity, the best way to do this is to return from
vnet_event_napi() when there are no more packets to read.
The next trip through ldc_rx will correctly set up the dring state.

Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Tested-by: David Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agosunvnet: Fix indentation in maybe_tx_wakeup()
Sowmini Varadhan [Thu, 6 Nov 2014 19:51:02 +0000 (14:51 -0500)]
sunvnet: Fix indentation in maybe_tx_wakeup()

remove redundant tab.

Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Reported-by: Ben Hutchings <ben@decadent.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge branch 'r8152-next'
David S. Miller [Thu, 6 Nov 2014 20:14:36 +0000 (15:14 -0500)]
Merge branch 'r8152-next'

Hayes Wang says:

====================
r8152: rtl_ops_init modify

Initialize the ops through tp->version. This could skip checking
each VID/PID.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agor8152: remove the definitions of the PID
hayeswang [Thu, 6 Nov 2014 04:47:40 +0000 (12:47 +0800)]
r8152: remove the definitions of the PID

The PIDs are only used in the id table, so the definitions are
unnacessary. Remove them wouldn't have confusion.

Signed-off-by: Hayes Wang <hayeswang@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agor8152: modify rtl_ops_init
hayeswang [Thu, 6 Nov 2014 04:47:39 +0000 (12:47 +0800)]
r8152: modify rtl_ops_init

Replace using VID/PID with using tp->version to initialize the ops.

Signed-off-by: Hayes Wang <hayeswang@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agor8152: move r8152b_get_version
hayeswang [Thu, 6 Nov 2014 04:47:38 +0000 (12:47 +0800)]
r8152: move r8152b_get_version

Move r8152b_get_version() to the location before rtl_ops_init().
Then, the rtl_ops_init() could use tp->version.

Signed-off-by: Hayes Wang <hayeswang@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agosock.h: Remove unused NETDEBUG macro
Joe Perches [Wed, 5 Nov 2014 23:42:09 +0000 (15:42 -0800)]
sock.h: Remove unused NETDEBUG macro

It's unused now, just delete it.

Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: esp: Convert NETDEBUG to pr_info
Joe Perches [Wed, 5 Nov 2014 23:36:08 +0000 (15:36 -0800)]
net: esp: Convert NETDEBUG to pr_info

Commit 64ce207306de ("[NET]: Make NETDEBUG pure printk wrappers")
originally had these NETDEBUG printks as always emitting.

Commit a2a316fd068c ("[NET]: Replace CONFIG_NET_DEBUG with sysctl")
added a net_msg_warn sysctl to these NETDEBUG uses.

Convert these NETDEBUG uses to normal pr_info calls.

This changes the output prefix from "ESP: " to include
"IPSec: " for the ipv4 case and "IPv6: " for the ipv6 case.

These output lines are now like the other messages in the files.

Other miscellanea:

Neaten the arithmetic spacing to be consistent with other
arithmetic spacing in the files.

Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet; ipv[46] - Remove 2 unnecessary NETDEBUG OOM messages
Joe Perches [Wed, 5 Nov 2014 22:39:21 +0000 (14:39 -0800)]
net; ipv[46] - Remove 2 unnecessary NETDEBUG OOM messages

These messages aren't useful as there's a generic dump_stack()
on OOM.

Neaten the comment and if test above the OOM by separating the
assign in if into an allocation then if test.

Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agodsa: mv88e6171: Add support for mv88e6172
Andrew Lunn [Wed, 5 Nov 2014 19:01:59 +0000 (20:01 +0100)]
dsa: mv88e6171: Add support for mv88e6172

The mv88e6172 is very similar to the mv88e6171.  So extend the
mv88e6171 driver to support it.

Signed-off-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: dsa: slave: Fix autoneg for phys on switch MDIO bus
Andrew Lunn [Wed, 5 Nov 2014 18:47:28 +0000 (19:47 +0100)]
net: dsa: slave: Fix autoneg for phys on switch MDIO bus

When the ports phys are connected to the switches internal MDIO bus,
we need to connect the phy to the slave netdev, otherwise
auto-negotiation etc, does not work.

Signed-off-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agosched: fix act file names in header comment
Jiri Pirko [Wed, 5 Nov 2014 19:51:51 +0000 (20:51 +0100)]
sched: fix act file names in header comment

Fixes: 4bba3925 ("[PKT_SCHED]: Prefix tc actions with act_")
Signed-off-by: Jiri Pirko <jiri@resnulli.us>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet/9p: remove a comment about pref member which doesn't exist
Ryo Munakata [Wed, 5 Nov 2014 14:45:58 +0000 (23:45 +0900)]
net/9p: remove a comment about pref member which doesn't exist

Signed-off-by: Ryo Munakata <ryomnktml@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agodrivers: net: cpsw: remove cpsw_ale_stop from cpsw_ale_destroy
Mugunthan V N [Wed, 5 Nov 2014 13:03:31 +0000 (18:33 +0530)]
drivers: net: cpsw: remove cpsw_ale_stop from cpsw_ale_destroy

when cpsw is build as modulea and simple insert and removal of module
creates a deadlock, due to delete timer. the timer is created and destroyed
in cpsw_ale_start and cpsw_ale_stop which are from device open and close.

root@am437x-evm:~# modprobe -r ti_cpsw
[  158.505333] INFO: trying to register non-static key.
[  158.510623] the code is fine but needs lockdep annotation.
[  158.516448] turning off the locking correctness validator.
[  158.522282] CPU: 0 PID: 1339 Comm: modprobe Not tainted 3.14.23-00445-gd41c88f #44
[  158.530359] [<c0015380>] (unwind_backtrace) from [<c0012088>] (show_stack+0x10/0x14)
[  158.538603] [<c0012088>] (show_stack) from [<c054ad70>] (dump_stack+0x78/0x94)
[  158.546295] [<c054ad70>] (dump_stack) from [<c0088008>] (__lock_acquire+0x176c/0x1b74)
[  158.554711] [<c0088008>] (__lock_acquire) from [<c0088944>] (lock_acquire+0x9c/0x104)
[  158.563043] [<c0088944>] (lock_acquire) from [<c004e520>] (del_timer_sync+0x44/0xd8)
[  158.571289] [<c004e520>] (del_timer_sync) from [<bf2eac1c>] (cpsw_ale_destroy+0x10/0x3c [ti_cpsw])
[  158.580821] [<bf2eac1c>] (cpsw_ale_destroy [ti_cpsw]) from [<bf2eb268>] (cpsw_remove+0x30/0xa0 [ti_cpsw])
[  158.591000] [<bf2eb268>] (cpsw_remove [ti_cpsw]) from [<c035ef44>] (platform_drv_remove+0x18/0x1c)
[  158.600527] [<c035ef44>] (platform_drv_remove) from [<c035d8bc>] (__device_release_driver+0x70/0xc8)
[  158.610236] [<c035d8bc>] (__device_release_driver) from [<c035e0d4>] (driver_detach+0xb4/0xb8)
[  158.619386] [<c035e0d4>] (driver_detach) from [<c035d6e4>] (bus_remove_driver+0x4c/0x90)
[  158.627988] [<c035d6e4>] (bus_remove_driver) from [<c00af2a8>] (SyS_delete_module+0x10c/0x198)
[  158.637144] [<c00af2a8>] (SyS_delete_module) from [<c000e580>] (ret_fast_syscall+0x0/0x48)
[  179.524727] INFO: rcu_sched detected stalls on CPUs/tasks: {} (detected by 0, t=2102 jiffies, g=1487, c=1486, q=6)
[  179.535741] INFO: Stall ended before state dump start

Signed-off-by: Mugunthan V N <mugunthanvnm@ti.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: mv643xx_eth: reclaim TX skbs only when released by the HW
Karl Beldan [Wed, 5 Nov 2014 14:32:59 +0000 (15:32 +0100)]
net: mv643xx_eth: reclaim TX skbs only when released by the HW

ATM, txq_reclaim will dequeue and free an skb for each tx desc released
by the hw that has TX_LAST_DESC set. However, in case of TSO, each
hw desc embedding the last part of a segment has TX_LAST_DESC set,
losing the one-to-one 'last skb frag'/'TX_LAST_DESC set' correspondance,
which causes data corruption.

Fix this by checking TX_ENABLE_INTERRUPT instead of TX_LAST_DESC, and
warn when trying to dequeue from an empty txq (which can be symptomatic
of releasing skbs prematurely).

Fixes: 3ae8f4e0b98 ('net: mv643xx_eth: Implement software TSO')
Reported-by: Slawomir Gajzner <slawomir.gajzner@gmail.com>
Reported-by: Julien D'Ascenzio <jdascenzio@yahoo.fr>
Signed-off-by: Karl Beldan <karl.beldan@rivierawaves.com>
Cc: Ian Campbell <ijc@hellion.org.uk>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Ezequiel Garcia <ezequiel.garcia@free-electrons.com>
Cc: Sebastian Hesselbarth <sebastian.hesselbarth@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge branch 'sfc-next'
David S. Miller [Thu, 6 Nov 2014 19:43:13 +0000 (14:43 -0500)]
Merge branch 'sfc-next'

Shradha Shah says:

====================
sfc: Clean up Siena SR-IOV support in preparation for EF10 SR-IOV support

This patch series provides a base and clean up for the upcoming
EF10 SRIOV patches.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agosfc: Add NIC type operations to replace direct calls from efx.c into siena_sriov.c
Shradha Shah [Wed, 5 Nov 2014 12:16:46 +0000 (12:16 +0000)]
sfc: Add NIC type operations to replace direct calls from efx.c into siena_sriov.c

Also add dummy functions where required to avoid NULL pointer dereference.

Signed-off-by: Shradha Shah <sshah@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agosfc: Rename implementations in siena_sriov.c to have a 'siena' prefix
Shradha Shah [Wed, 5 Nov 2014 12:16:32 +0000 (12:16 +0000)]
sfc: Rename implementations in siena_sriov.c to have a 'siena' prefix

Patch in preparation for the upcoming EF10 sriov support.

Signed-off-by: Shradha Shah <sshah@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agosfc: Move the current VF state from efx_nic into siena_nic_data
Shradha Shah [Wed, 5 Nov 2014 12:16:18 +0000 (12:16 +0000)]
sfc: Move the current VF state from efx_nic into siena_nic_data

This patch series provides a base and cleanup for the
upcoming EF10 SRIOV support.

This patch moves the VF state into siena_nic_data as a basis to
save the VF state based on nic type.

Signed-off-by: Shradha Shah <sshah@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoxen-netback: remove unconditional __pskb_pull_tail() in guest Tx path
Malcolm Crossley [Wed, 5 Nov 2014 10:50:22 +0000 (10:50 +0000)]
xen-netback: remove unconditional __pskb_pull_tail() in guest Tx path

Unconditionally pulling 128 bytes into the linear area is not required
for:

- security: Every protocol demux starts with pskb_may_pull() to pull
  frag data into the linear area, if necessary, before looking at
  headers.

- performance: Netback has already grant copied up-to 128 bytes from
  the first slot of a packet into the linear area. The first slot
  normally contain all the IPv4/IPv6 and TCP/UDP headers.

The unconditional pull would often copy frag data unnecessarily.  This
is a performance problem when running on a version of Xen where grant
unmap avoids TLB flushes for pages which are not accessed.  TLB
flushes can now be avoided for > 99% of unmaps (it was 0% before).

Grant unmap TLB flush avoidance will be available in a future version
of Xen (probably 4.6).

Signed-off-by: Malcolm Crossley <malcolm.crossley@citrix.com>
Signed-off-by: David Vrabel <david.vrabel@citrix.com>
Acked-by: Ian Campbell <ian.campbell@citrix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge branch 'stmmac-next'
David S. Miller [Thu, 6 Nov 2014 19:39:01 +0000 (14:39 -0500)]
Merge branch 'stmmac-next'

Andy Shevchenko says:

====================
stmmac: pci: various cleanups and fixes

There are few cleanups and fixes regarding to stmmac PCI driver.
This has been tested on Intel Galileo board with recent net-next tree.

Since v2:
- drop patch 5/5 since it will be part of a big change across entire subsystem

Since v1:
- remove already applied patch
- append patch 1/5
- rework patch 3/5 to be functional compatible with original code
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agostmmac: pci: convert to use dev_* macros
Andy Shevchenko [Wed, 5 Nov 2014 10:27:29 +0000 (12:27 +0200)]
stmmac: pci: convert to use dev_* macros

Instead of pr_* macros let's use dev_* macros which provide device name.

Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agostmmac: pci: use managed resources
Andy Shevchenko [Wed, 5 Nov 2014 10:27:28 +0000 (12:27 +0200)]
stmmac: pci: use managed resources

Migrate pci driver to managed resources to reduce boilerplate error handling
code.

Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agostmmac: pci: convert to use dev_pm_ops
Andy Shevchenko [Wed, 5 Nov 2014 10:27:27 +0000 (12:27 +0200)]
stmmac: pci: convert to use dev_pm_ops

Convert system PM callbacks to use dev_pm_ops. In addition remove the PCI calls
related to a power state since the bus code cares about this already.

Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agostmmac: pci: use defined constant instead of magic number
Andy Shevchenko [Wed, 5 Nov 2014 10:27:26 +0000 (12:27 +0200)]
stmmac: pci: use defined constant instead of magic number

The last standard PCI resource is defined as PCI_STD_RESOURCE_END. Thus, we
could use it instead of plain integer.

Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agostmmac: fix sparse warnings
Andy Shevchenko [Wed, 5 Nov 2014 09:45:32 +0000 (11:45 +0200)]
stmmac: fix sparse warnings

This patch fixes the following sparse warnings.

drivers/net/ethernet/stmicro/stmmac/enh_desc.c:381:30: warning: symbol 'enh_desc_ops' was not declared. Should it be static?
drivers/net/ethernet/stmicro/stmmac/norm_desc.c:253:30: warning: symbol 'ndesc_ops' was not declared. Should it be static?
drivers/net/ethernet/stmicro/stmmac/stmmac_hwtstamp.c:141:33: warning: symbol 'stmmac_ptp' was not declared. Should it be static?

There is no functional change.

Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Acked-by: Giuseppe CAVALLARO <peppe.cavallaro@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoip6_tunnel: Add support for wildcard tunnel endpoints.
Steffen Klassert [Wed, 5 Nov 2014 07:03:50 +0000 (08:03 +0100)]
ip6_tunnel: Add support for wildcard tunnel endpoints.

This patch adds support for tunnels with local or
remote wildcard endpoints. With this we get a
NBMA tunnel mode like we have it for ipv4 and
sit tunnels.

Signed-off-by: Steffen Klassert <steffen.klassert@secunet.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoipv6: Allow sending packets through tunnels with wildcard endpoints
Steffen Klassert [Wed, 5 Nov 2014 07:02:48 +0000 (08:02 +0100)]
ipv6: Allow sending packets through tunnels with wildcard endpoints

Currently we need the IP6_TNL_F_CAP_XMIT capabiltiy to transmit
packets through an ipv6 tunnel. This capability is set when the
tunnel gets configured, based on the tunnel endpoint addresses.

On tunnels with wildcard tunnel endpoints, we need to do the
capabiltiy checking on a per packet basis like it is done in
the receive path.

This patch extends ip6_tnl_xmit_ctl() to take local and remote
addresses as parameters to allow for per packet capabiltiy
checking.

Signed-off-by: Steffen Klassert <steffen.klassert@secunet.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoopenvswitch: Avoid NULL mask check while building mask
Pravin B Shelar [Sun, 19 Oct 2014 19:03:40 +0000 (12:03 -0700)]
openvswitch: Avoid NULL mask check while building mask

OVS does mask validation even if it does not need to convert
netlink mask attributes to mask structure.  ovs_nla_get_match()
caller can pass NULL mask structure pointer if the caller does
not need mask.  Therefore NULL check is required in SW_FLOW_KEY*
macros.  Following patch does not convert mask netlink attributes
if mask pointer is NULL, so we do not need these checks in
SW_FLOW_KEY* macro.

Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
Acked-by: Daniele Di Proietto <ddiproietto@vmware.com>
Acked-by: Andy Zhou <azhou@nicira.com>
10 years agoopenvswitch: Refactor action alloc and copy api.
Pravin B Shelar [Sun, 19 Oct 2014 18:19:51 +0000 (11:19 -0700)]
openvswitch: Refactor action alloc and copy api.

There are two separate API to allocate and copy actions list. Anytime
OVS needs to copy action list, it needs to call both functions.
Following patch moves action allocation to copy function to avoid
code duplication.

Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
Acked-by: Jarno Rajahalme <jrajahalme@nicira.com>
10 years agoopenvswitch: Move key_attr_size() to flow_netlink.h.
Joe Stringer [Sat, 18 Oct 2014 23:14:14 +0000 (16:14 -0700)]
openvswitch: Move key_attr_size() to flow_netlink.h.

flow-netlink has netlink related code.

Signed-off-by: Joe Stringer <joestringer@nicira.com>
Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
10 years agoopenvswitch: Remove flow member from struct ovs_skb_cb
Lorand Jakab [Mon, 6 Oct 2014 12:45:32 +0000 (05:45 -0700)]
openvswitch: Remove flow member from struct ovs_skb_cb

The 'flow' memeber was chosen for removal because it's only used
in ovs_execute_actions() we can pass it as argument to this
function.

Signed-off-by: Lorand Jakab <lojakab@cisco.com>
Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
10 years agoopenvswitch: Fix the type of struct ovs_key_nd nd_target field.
Jarno Rajahalme [Tue, 30 Sep 2014 17:52:32 +0000 (10:52 -0700)]
openvswitch: Fix the type of struct ovs_key_nd nd_target field.

Should be the same as other IPv6 address fields.

Current master produces sparse warnings without this change.

Signed-off-by: Jarno Rajahalme <jrajahalme@nicira.com>
Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
10 years agoopenvswitch: Drop packets when interdev is not up
Chunhe Li [Mon, 8 Sep 2014 20:17:21 +0000 (13:17 -0700)]
openvswitch: Drop packets when interdev is not up

If the internal device is not up, it should drop received
packets. Sometimes it receive the broadcast or multicast
packets, and the ip protocol stack will casue more cpu
usage wasted.

Signed-off-by: Chunhe Li <lichunhe@huawei.com>
Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
10 years agoopenvswitch: Refactor get_dp() function into multiple access APIs.
Andy Zhou [Mon, 8 Sep 2014 20:14:22 +0000 (13:14 -0700)]
openvswitch: Refactor get_dp() function into multiple access APIs.

Avoid recursive read_rcu_lock() by using the lighter weight
get_dp_rcu() API. Add proper locking assertions to get_dp().

Signed-off-by: Andy Zhou <azhou@nicira.com>
Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
10 years agoopenvswitch: Refactor ovs_flow_cmd_fill_info().
Joe Stringer [Mon, 8 Sep 2014 20:09:37 +0000 (13:09 -0700)]
openvswitch: Refactor ovs_flow_cmd_fill_info().

Split up ovs_flow_cmd_fill_info() to make it easier to cache parts of a
dump reply. This will be used to streamline flow_dump in a future patch.

Signed-off-by: Joe Stringer <joestringer@nicira.com>
Acked-by: Thomas Graf <tgraf@noironetworks.com>
Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
10 years agoopenvswitch: refactor do_output() to move NULL check out of fast path
Andy Zhou [Mon, 8 Sep 2014 07:35:02 +0000 (00:35 -0700)]
openvswitch: refactor do_output() to move NULL check out of fast path

skb_clone() NULL check is implemented in do_output(), as past of the
common (fast) path. Refactoring so that NULL check is done in the
slow path, immediately after skb_clone() is called.

Besides optimization, this change also improves code readability by
making the skb_clone() NULL check consistent within OVS datapath
module.

Signed-off-by: Andy Zhou <azhou@nicira.com>
Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
10 years agoopenvswitch: Additional logging for -EINVAL on flow setups.
Jesse Gross [Mon, 6 Oct 2014 12:08:38 +0000 (05:08 -0700)]
openvswitch: Additional logging for -EINVAL on flow setups.

There are many possible ways that a flow can be invalid so we've
added logging for most of them. This adds logs for the remaining
possible cases so there isn't any ambiguity while debugging.

CC: Federico Iezzi <fiezzi@enter.it>
Signed-off-by: Jesse Gross <jesse@nicira.com>
Acked-by: Thomas Graf <tgraf@noironetworks.com>
Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
10 years agoopenvswitch: Remove redundant tcp_flags code.
Joe Stringer [Mon, 8 Sep 2014 05:11:08 +0000 (22:11 -0700)]
openvswitch: Remove redundant tcp_flags code.

These two cases used to be treated differently for IPv4/IPv6,
but they are now identical.

Signed-off-by: Joe Stringer <joestringer@nicira.com>
Acked-by: Jarno Rajahalme <jrajahalme@nicira.com>
Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
10 years agoopenvswitch: Move table destroy to dp-rcu callback.
Pravin B Shelar [Wed, 7 May 2014 01:41:20 +0000 (18:41 -0700)]
openvswitch: Move table destroy to dp-rcu callback.

Ths simplifies flow-table-destroy API. No need to pass explicit
parameter about context.

Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
Acked-by: Thomas Graf <tgraf@redhat.com>
10 years agoopenvswitch: Add basic MPLS support to kernel
Simon Horman [Mon, 6 Oct 2014 12:05:13 +0000 (05:05 -0700)]
openvswitch: Add basic MPLS support to kernel

Allow datapath to recognize and extract MPLS labels into flow keys
and execute actions which push, pop, and set labels on packets.

Based heavily on work by Leo Alterman, Ravi K, Isaku Yamahata and Joe Stringer.

Cc: Ravi K <rkerur@gmail.com>
Cc: Leo Alterman <lalterman@nicira.com>
Cc: Isaku Yamahata <yamahata@valinux.co.jp>
Cc: Joe Stringer <joe@wand.net.nz>
Signed-off-by: Simon Horman <horms@verge.net.au>
Signed-off-by: Jesse Gross <jesse@nicira.com>
Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
10 years agonet: Remove MPLS GSO feature.
Pravin B Shelar [Wed, 5 Nov 2014 23:27:48 +0000 (15:27 -0800)]
net: Remove MPLS GSO feature.

Device can export MPLS GSO support in dev->mpls_features same way
it export vlan features in dev->vlan_features. So it is safe to
remove NETIF_F_GSO_MPLS redundant flag.

Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
10 years agofou: Fix typo in returning flags in netlink
Tom Herbert [Thu, 6 Nov 2014 00:49:38 +0000 (16:49 -0800)]
fou: Fix typo in returning flags in netlink

When filling netlink info, dport is being returned as flags. Fix
instances to return correct value.

Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agor8152: disable the tasklet by default
hayeswang [Wed, 5 Nov 2014 02:17:02 +0000 (10:17 +0800)]
r8152: disable the tasklet by default

Let the tasklet only be enabled after open(), and be disabled for
the other situation. The tasklet is only necessary after open() for
tx/rx, so it could be disabled by default.

Signed-off-by: Hayes Wang <hayeswang@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoipv6: mld: fix add_grhead skb_over_panic for devs with large MTUs
Daniel Borkmann [Wed, 5 Nov 2014 19:27:38 +0000 (20:27 +0100)]
ipv6: mld: fix add_grhead skb_over_panic for devs with large MTUs

It has been reported that generating an MLD listener report on
devices with large MTUs (e.g. 9000) and a high number of IPv6
addresses can trigger a skb_over_panic():

skbuff: skb_over_panic: text:ffffffff80612a5d len:3776 put:20
head:ffff88046d751000 data:ffff88046d751010 tail:0xed0 end:0xec0
dev:port1
 ------------[ cut here ]------------
kernel BUG at net/core/skbuff.c:100!
invalid opcode: 0000 [#1] SMP
Modules linked in: ixgbe(O)
CPU: 3 PID: 0 Comm: swapper/3 Tainted: G O 3.14.23+ #4
[...]
Call Trace:
 <IRQ>
 [<ffffffff80578226>] ? skb_put+0x3a/0x3b
 [<ffffffff80612a5d>] ? add_grhead+0x45/0x8e
 [<ffffffff80612e3a>] ? add_grec+0x394/0x3d4
 [<ffffffff80613222>] ? mld_ifc_timer_expire+0x195/0x20d
 [<ffffffff8061308d>] ? mld_dad_timer_expire+0x45/0x45
 [<ffffffff80255b5d>] ? call_timer_fn.isra.29+0x12/0x68
 [<ffffffff80255d16>] ? run_timer_softirq+0x163/0x182
 [<ffffffff80250e6f>] ? __do_softirq+0xe0/0x21d
 [<ffffffff8025112b>] ? irq_exit+0x4e/0xd3
 [<ffffffff802214bb>] ? smp_apic_timer_interrupt+0x3b/0x46
 [<ffffffff8063f10a>] ? apic_timer_interrupt+0x6a/0x70

mld_newpack() skb allocations are usually requested with dev->mtu
in size, since commit 72e09ad107e7 ("ipv6: avoid high order allocations")
we have changed the limit in order to be less likely to fail.

However, in MLD/IGMP code, we have some rather ugly AVAILABLE(skb)
macros, which determine if we may end up doing an skb_put() for
adding another record. To avoid possible fragmentation, we check
the skb's tailroom as skb->dev->mtu - skb->len, which is a wrong
assumption as the actual max allocation size can be much smaller.

The IGMP case doesn't have this issue as commit 57e1ab6eaddc
("igmp: refine skb allocations") stores the allocation size in
the cb[].

Set a reserved_tailroom to make it fit into the MTU and use
skb_availroom() helper instead. This also allows to get rid of
igmp_skb_size().

Reported-by: Wei Liu <lw1a2.jing@gmail.com>
Fixes: 72e09ad107e7 ("ipv6: avoid high order allocations")
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Hannes Frederic Sowa <hannes@stressinduktion.org>
Cc: David L Stevens <david.stevens@oracle.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Acked-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: Convert SEQ_START_TOKEN/seq_printf to seq_puts
Joe Perches [Tue, 4 Nov 2014 23:37:03 +0000 (15:37 -0800)]
net: Convert SEQ_START_TOKEN/seq_printf to seq_puts

Using a single fixed string is smaller code size than using
a format and many string arguments.

Reduces overall code size a little.

$ size net/ipv4/igmp.o* net/ipv6/mcast.o* net/ipv6/ip6_flowlabel.o*
   text    data     bss     dec     hex filename
  34269    7012   14824   56105    db29 net/ipv4/igmp.o.new
  34315    7012   14824   56151    db57 net/ipv4/igmp.o.old
  30078    7869   13200   51147    c7cb net/ipv6/mcast.o.new
  30105    7869   13200   51174    c7e6 net/ipv6/mcast.o.old
  11434    3748    8580   23762    5cd2 net/ipv6/ip6_flowlabel.o.new
  11491    3748    8580   23819    5d0b net/ipv6/ip6_flowlabel.o.old

Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agofast_hash: avoid indirect function calls
Hannes Frederic Sowa [Tue, 4 Nov 2014 23:23:04 +0000 (00:23 +0100)]
fast_hash: avoid indirect function calls

By default the arch_fast_hash hashing function pointers are initialized
to jhash(2). If during boot-up a CPU with SSE4.2 is detected they get
updated to the CRC32 ones. This dispatching scheme incurs a function
pointer lookup and indirect call for every hashing operation.

rhashtable as a user of arch_fast_hash e.g. stores pointers to hashing
functions in its structure, too, causing two indirect branches per
hashing operation.

Using alternative_call we can get away with one of those indirect branches.

Acked-by: Daniel Borkmann <dborkman@redhat.com>
Cc: Thomas Graf <tgraf@suug.ch>
Signed-off-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge branch 'amd-xgbe-next'
David S. Miller [Thu, 6 Nov 2014 02:50:43 +0000 (21:50 -0500)]
Merge branch 'amd-xgbe-next'

Tom Lendacky says:

====================
amd-xgbe: AMD XGBE driver updates 2014-11-04

The following series of patches includes functional updates to the
driver as well as some trivial changes for function renaming and
spelling fixes.

- Move channel and ring structure allocation into the device open path
- Rename the pre_xmit function to dev_xmit
- Explicitly use the u32 data type for the device descriptors
- Use page allocation for the receive buffers
- Add support for split header/payload receive
- Add support for per DMA channel interrupts
- Add support for receive side scaling (RSS)
- Add support for ethtool receive side scaling commands
- Fix the spelling of descriptors
- After a PCS reset, sync the PCS and PHY modes
- Add dependency on HAS_IOMEM to both the amd-xgbe and amd-xgbe-phy
  drivers

This patch series is based on net-next.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoamd-xgbe-phy: Let AMD_XGBE_PHY depend on HAS_IOMEM
Lendacky, Thomas [Tue, 4 Nov 2014 22:07:46 +0000 (16:07 -0600)]
amd-xgbe-phy: Let AMD_XGBE_PHY depend on HAS_IOMEM

The amd-xgbe-phy driver needs to perform ioremap calls, so add HAS_IOMEM
to its build dependency.

Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoamd-xgbe: Let AMD_XGBE depend on HAS_IOMEM
Lendacky, Thomas [Tue, 4 Nov 2014 22:07:41 +0000 (16:07 -0600)]
amd-xgbe: Let AMD_XGBE depend on HAS_IOMEM

The amd-xgbe driver needs to perform ioremap calls, so add HAS_IOMEM
to its build dependency.

Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoamd-xgbe-phy: Sync PCS and PHY modes after reset
Lendacky, Thomas [Tue, 4 Nov 2014 22:07:35 +0000 (16:07 -0600)]
amd-xgbe-phy: Sync PCS and PHY modes after reset

This patch adds support to sync the states of the PCS and the PHY
after a reset is performed.  If the PCS and the PHY are not in the
same state after reset an extra mode change would be performed. This
extra mode change might not be needed if the PCS and the PHY are
synced up after reset.

Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoamd-xgbe: Fix a spelling error
Lendacky, Thomas [Tue, 4 Nov 2014 22:07:29 +0000 (16:07 -0600)]
amd-xgbe: Fix a spelling error

This patch fixes the spelling of the word "descriptor" in a couple
of locations.

Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoamd-xgbe: Add receive side scaling ethtool support
Lendacky, Thomas [Tue, 4 Nov 2014 22:07:23 +0000 (16:07 -0600)]
amd-xgbe: Add receive side scaling ethtool support

This patch adds support for ethtool receive side scaling (RSS) commands.
Support is added to get/set the RSS hash key and the RSS lookup table.

Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoamd-xgbe: Provide support for receive side scaling
Lendacky, Thomas [Tue, 4 Nov 2014 22:07:02 +0000 (16:07 -0600)]
amd-xgbe: Provide support for receive side scaling

This patch provides support for receive side scaling (RSS). RSS allows
for spreading incoming network packets across the Rx queues.  When used
in conjunction with the per DMA channel interrupt support, this allows
the receive processing to be spread across multiple processors.

Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoamd-xgbe: Add support for per DMA channel interrupts
Lendacky, Thomas [Tue, 4 Nov 2014 22:06:56 +0000 (16:06 -0600)]
amd-xgbe: Add support for per DMA channel interrupts

This patch provides support for interrupts that are generated by the
Tx/Rx DMA channel pairs of the device.  This allows for Tx and Rx
processing to run across multiple processsors.

Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoamd-xgbe: Implement split header receive support
Lendacky, Thomas [Tue, 4 Nov 2014 22:06:50 +0000 (16:06 -0600)]
amd-xgbe: Implement split header receive support

Provide support for splitting IP packets so that the header and
payload can be sent to different DMA addresses.  This will allow
the IP header to be put into the linear part of the skb while the
payload can be added as frags.

Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoamd-xgbe: Use page allocations for Rx buffers
Lendacky, Thomas [Tue, 4 Nov 2014 22:06:44 +0000 (16:06 -0600)]
amd-xgbe: Use page allocations for Rx buffers

Use page allocations for Rx buffers instead of pre-allocating skbs
of a set size.

Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoamd-xgbe: Use the u32 data type for descriptors
Lendacky, Thomas [Tue, 4 Nov 2014 22:06:37 +0000 (16:06 -0600)]
amd-xgbe: Use the u32 data type for descriptors

The Tx and Rx descriptors are unsigned 32 bit values.  Use the u32
type, rather than unsigned int, to map these descriptors.

Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoamd-xgbe: Rename pre_xmit function to dev_xmit
Lendacky, Thomas [Tue, 4 Nov 2014 22:06:32 +0000 (16:06 -0600)]
amd-xgbe: Rename pre_xmit function to dev_xmit

The pre_xmit function name implies that it performs operations prior
to transmitting the packet when in fact it is responsible for setting
up the descriptors and initiating the transmit.  Rename this to
function from pre_xmit to dev_xmit, which is consistent with the name
used during receive processing - dev_read.

Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoamd-xgbe: Move ring allocation to device open
Lendacky, Thomas [Tue, 4 Nov 2014 22:06:26 +0000 (16:06 -0600)]
amd-xgbe: Move ring allocation to device open

Move the channel and ring tracking structures allocation to device
open.  This will allow for future support to vary the number of Tx/Rx
queues without unloading the module.

Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agobridge: include in6.h in if_bridge.h for struct in6_addr
Gregory Fong [Tue, 4 Nov 2014 19:21:21 +0000 (11:21 -0800)]
bridge: include in6.h in if_bridge.h for struct in6_addr

if_bridge.h uses struct in6_addr ip6, but wasn't including the in6.h
header.  Thomas Backlund originally sent a patch to do this, but this
revealed a redefinition issue: https://lkml.org/lkml/2013/1/13/116

The redefinition issue should have been fixed by the following Linux
commits:
ee262ad827f89e2dc7851ec2986953b5b125c6bc inet: defines IPPROTO_* needed for module alias generation
cfd280c91253cc28e4919e349fa7a813b63e71e8 net: sync some IP headers with glibc

and the following glibc commit:
6c82a2f8d7c8e21e39237225c819f182ae438db3 Coordinate IPv6 definitions for Linux and glibc

so actually include the header now.

Reported-by: Colin Guthrie <colin@mageia.org>
Reported-by: Christiaan Welvaart <cjw@daneel.dyndns.org>
Reported-by: Thomas Backlund <tmb@mageia.org>
Cc: Florian Fainelli <f.fainelli@gmail.com>
Cc: Cong Wang <xiyou.wangcong@gmail.com>
Cc: David Miller <davem@davemloft.net>
Signed-off-by: Gregory Fong <gregory.0xf0@gmail.com>
Acked-by: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agotcp: zero retrans_stamp if all retrans were acked
Marcelo Leitner [Tue, 4 Nov 2014 19:15:08 +0000 (17:15 -0200)]
tcp: zero retrans_stamp if all retrans were acked

Ueki Kohei reported that when we are using NewReno with connections that
have a very low traffic, we may timeout the connection too early if a
second loss occurs after the first one was successfully acked but no
data was transfered later. Below is his description of it:

When SACK is disabled, and a socket suffers multiple separate TCP
retransmissions, that socket's ETIMEDOUT value is calculated from the
time of the *first* retransmission instead of the *latest*
retransmission.

This happens because the tcp_sock's retrans_stamp is set once then never
cleared.

Take the following connection:

                      Linux                    remote-machine
                        |                           |
         send#1---->(*1)|--------> data#1 --------->|
                  |     |                           |
                 RTO    :                           :
                  |     |                           |
                 ---(*2)|----> data#1(retrans) ---->|
                  | (*3)|<---------- ACK <----------|
                  |     |                           |
                  |     :                           :
                  |     :                           :
                  |     :                           :
                16 minutes (or more)                :
                  |     :                           :
                  |     :                           :
                  |     :                           :
                  |     |                           |
         send#2---->(*4)|--------> data#2 --------->|
                  |     |                           |
                 RTO    :                           :
                  |     |                           |
                 ---(*5)|----> data#2(retrans) ---->|
                  |     |                           |
                  |     |                           |
                RTO*2   :                           :
                  |     |                           |
                  |     |                           |
      ETIMEDOUT<----(*6)|                           |

(*1) One data packet sent.
(*2) Because no ACK packet is received, the packet is retransmitted.
(*3) The ACK packet is received. The transmitted packet is acknowledged.

At this point the first "retransmission event" has passed and been
recovered from. Any future retransmission is a completely new "event".

(*4) After 16 minutes (to correspond with retries2=15), a new data
packet is sent. Note: No data is transmitted between (*3) and (*4).

The socket's timeout SHOULD be calculated from this point in time, but
instead it's calculated from the prior "event" 16 minutes ago.

(*5) Because no ACK packet is received, the packet is retransmitted.
(*6) At the time of the 2nd retransmission, the socket returns
ETIMEDOUT.

Therefore, now we clear retrans_stamp as soon as all data during the
loss window is fully acked.

Reported-by: Ueki Kohei
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Marcelo Ricardo Leitner <mleitner@redhat.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Tested-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoipv6: move INET6_MATCH() to include/net/inet6_hashtables.h
WANG Cong [Tue, 4 Nov 2014 18:59:47 +0000 (10:59 -0800)]
ipv6: move INET6_MATCH() to include/net/inet6_hashtables.h

It is only used in net/ipv6/inet6_hashtables.c.

Cc: David S. Miller <davem@davemloft.net>
Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: Add and use skb_copy_datagram_msg() helper.
David S. Miller [Wed, 5 Nov 2014 21:46:40 +0000 (16:46 -0500)]
net: Add and use skb_copy_datagram_msg() helper.

This encapsulates all of the skb_copy_datagram_iovec() callers
with call argument signature "skb, offset, msghdr->msg_iov, length".

When we move to iov_iters in the networking, the iov_iter object will
sit in the msghdr.

Having a helper like this means there will be less places to touch
during that transformation.

Based upon descriptions and patch from Al Viro.

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge branch 'gue-next'
David S. Miller [Wed, 5 Nov 2014 21:34:47 +0000 (16:34 -0500)]
Merge branch 'gue-next'

Tom Herbert says:

====================
gue: Remote checksum offload

This patch set implements remote checksum offload for
GUE, which is a mechanism that provides checksum offload of
encapsulated packets using rudimentary offload capabilities found in
most Network Interface Card (NIC) devices. The outer header checksum
for UDP is enabled in packets and, with some additional meta
information in the GUE header, a receiver is able to deduce the
checksum to be set for an inner encapsulated packet. Effectively this
offloads the computation of the inner checksum. Enabling the outer
checksum in encapsulation has the additional advantage that it covers
more of the packet than the inner checksum including the encapsulation
headers.

Remote checksum offload is described in:
http://tools.ietf.org/html/draft-herbert-remotecsumoffload-01

The GUE transmit and receive paths are modified to support the
remote checksum offload option. The option contains a checksum
offset and checksum start which are directly derived from values
set in stack when doing CHECKSUM_PARTIAL. On receipt of the option, the
operation is to calculate the packet checksum from "start" to end of
the packet (normally derived for checksum complete), and then set
the resultant value at checksum "offset" (the checksum field has
already been primed with the pseudo header). This emulates a NIC
that implements NETIF_F_HW_CSUM.

The primary purpose of this feature is to eliminate cost of performing
checksum calculation over a packet when encpasulating.

In this patch set:
  - Move fou_build_header into fou.c and split it into a couple of
    functions
  - Enable offloading of outer UDP checksum in encapsulation
  - Change udp_offload to support remote checksum offload, includes
    new GSO type and ensuring encapsulated layers (TCP) doesn't try to
    set a checksum covered by RCO
  - TX support for RCO with GUE. This is configured through ip_tunnel
    and set the option on transmit when packet being encapsulated is
    CHECKSUM_PARTIAL
  - RX support for RCO with GUE for normal and GRO paths. Includes
    resolving the offloaded checksum

v2:
  Address comments from davem: Move accounting for private option
  field in gue_encap_hlen to patch in which we add the remote checksum
  offload option.

Testing:

I ran performance numbers using netperf TCP_STREAM and TCP_RR with 200
streams, comparing GUE with and without remote checksum offload (doing
checksum-unnecessary to complete conversion in both cases). These
were run on mlnx4 and bnx2x. Some mlnx4 results are below.

GRE/GUE
    TCP_STREAM
      IPv4, with remote checksum offload
        9.71% TX CPU utilization
        7.42% RX CPU utilization
        36380 Mbps
      IPv4, without remote checksum offload
        12.40% TX CPU utilization
        7.36% RX CPU utilization
        36591 Mbps
    TCP_RR
      IPv4, with remote checksum offload
        77.79% CPU utilization
91/144/216 90/95/99% latencies
        1.95127e+06 tps
      IPv4, without remote checksum offload
        78.70% CPU utilization
        89/152/297 90/95/99% latencies
        1.95458e+06 tps

IPIP/GUE
    TCP_STREAM
      With remote checksum offload
        10.30% TX CPU utilization
        7.43% RX CPU utilization
        36486 Mbps
      Without remote checksum offload
        12.47% TX CPU utilization
        7.49% RX CPU utilization
        36694 Mbps
    TCP_RR
      With remote checksum offload
        77.80% CPU utilization
        87/153/270 90/95/99% latencies
        1.98735e+06 tps
      Without remote checksum offload
        77.98% CPU utilization
        87/150/287 90/95/99% latencies
        1.98737e+06 tps

SIT/GUE
    TCP_STREAM
      With remote checksum offload
        9.68% TX CPU utilization
        7.36% RX CPU utilization
        35971 Mbps
      Without remote checksum offload
        12.95% TX CPU utilization
        8.04% RX CPU utilization
        36177 Mbps
    TCP_RR
      With remote checksum offload
        79.32% CPU utilization
        94/158/295 90/95/99% latencies
        1.88842e+06 tps
      Without remote checksum offload
        80.23% CPU utilization
        94/149/226 90/95/99% latencies
        1.90338e+06 tps

VXLAN
    TCP_STREAM
        35.03% TX CPU utilization
        20.85% RX CPU utilization
        36230 Mbps
    TCP_RR
        77.36% CPU utilization
        84/146/270 90/95/99% latencies
        2.08063e+06 tps

We can also look at CPU time in csum_partial using perf (with bnx2x
setup). For GRE with TCP_STREAM I see:

    With remote checksum offload
        0.33% TX
        1.81% RX
    Without remote checksum offload
        6.00% TX
        0.51% RX

I suspect the fact that time in csum_partial noticably increases
with remote checksum offload for RX is due to taking the cache miss on
the encapsulated header in that function. By similar reasoning, if on
the TX side the packet were not in cache (say we did a splice from a
file whose data was never touched by the CPU) the CPU savings for TX
would probably be more pronounced.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agogue: Receive side of remote checksum offload
Tom Herbert [Tue, 4 Nov 2014 17:06:57 +0000 (09:06 -0800)]
gue: Receive side of remote checksum offload

Add processing of the remote checksum offload option in both the normal
path as well as the GRO path. The implements patching the affected
checksum to derive the offloaded checksum.

Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agogue: TX support for using remote checksum offload option
Tom Herbert [Tue, 4 Nov 2014 17:06:56 +0000 (09:06 -0800)]
gue: TX support for using remote checksum offload option

Add if_tunnel flag TUNNEL_ENCAP_FLAG_REMCSUM to configure
remote checksum offload on an IP tunnel. Add logic in gue_build_header
to insert remote checksum offload option.

Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agogue: Protocol constants for remote checksum offload
Tom Herbert [Tue, 4 Nov 2014 17:06:55 +0000 (09:06 -0800)]
gue: Protocol constants for remote checksum offload

Define a private flag for remote checksun offload as well as a length
for the option.

Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoudp: Changes to udp_offload to support remote checksum offload
Tom Herbert [Tue, 4 Nov 2014 17:06:54 +0000 (09:06 -0800)]
udp: Changes to udp_offload to support remote checksum offload

Add a new GSO type, SKB_GSO_TUNNEL_REMCSUM, which indicates remote
checksum offload being done (in this case inner checksum must not
be offloaded to the NIC).

Added logic in __skb_udp_tunnel_segment to handle remote checksum
offload case.

Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agogue: Add infrastructure for flags and options
Tom Herbert [Tue, 4 Nov 2014 17:06:53 +0000 (09:06 -0800)]
gue: Add infrastructure for flags and options

Add functions and basic definitions for processing standard flags,
private flags, and control messages. This includes definitions
to compute length of optional fields corresponding to a set of flags.
Flag validation is in validate_gue_flags function. This checks for
unknown flags, and that length of optional fields is <= length
in guehdr hlen.

Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>