BPF classic->internal converter broke SKF_AD_PKTTYPE extension, since
pkt_type_offset() was failing to find skb->pkt_type field which is defined as:
__u8 pkt_type:3,
     fclone:2,
     ipvs_property:1,
     peeked:1,
     nf_trace:1;

Fix it by searching for 3 most significant bits and shift them by 5 at run-time

Fixes: bd4cf0ed ("net: filter: rework/optimize internal BPF interpreter's instruction set")
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Acked-by: Daniel Borkmann <dborkman@redhat.com>
Tested-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/amd/hplance.h:
 #define HPLANCE_MEMOFF 0x8000   /* struct lance_init_block */
 #define HPLANCE_NVRAMOFF 0xC008 /* etheraddress as one *nibble* per byte */

The offset of RAM start is 0x8000, the offset of RAM end is 0xC008,
so the RAM size is 16392 bytes.
Signed-off-by: Amos Kong <akong@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The order of ram pages is 3, so the ram size is 2^3 * 4K = 32K.
Signed-off-by: Amos Kong <akong@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

A recent commit (a02eb4 "xen-netback: worse-case estimate in xenvif_rx_action is
underestimating") capped the slot estimation to MAX_SKB_FRAGS, but that triggers
the next BUG_ON a few lines down, as the packet consumes more slots than
estimated.
This patch introduces full_coalesce on the skb callback buffer, which is used in
start_new_rx_buffer() to decide whether netback needs coalescing more
aggresively. By doing that, no packet should need more than
(XEN_NETIF_MAX_TX_SIZE + 1) / PAGE_SIZE data slots (excluding the optional GSO
slot, it doesn't carry data, therefore irrelevant in this case), as the provided
buffers are fully utilized.
Signed-off-by: Zoltan Kiss <zoltan.kiss@citrix.com>
Cc: Paul Durrant <paul.durrant@citrix.com>
Cc: Wei Liu <wei.liu2@citrix.com>
Cc: Ian Campbell <ian.campbell@citrix.com>
Cc: David Vrabel <david.vrabel@citrix.com>
Reviewed-by: Paul Durrant <paul.durrant@gmail.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

o Uninitialzed fields in mailbox command structure
  caused commands to time out randomly due to garbage
  values so initialize it to zero.
Signed-off-by: Rajesh Borundia <rajesh.borundia@qlogic.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Quoting David Miller:
"At the moment you call register_netdev() the device is visible, notifications
are sent to userspace, and userland tools can try to bring the interface up
and see the incorrect link state, before you do the netif_carrier_off().

Said another way, between the register_netdev() and netif_carrier_off() call,
userspace can see the device in an inconsistent state."

So call netif_carrier_off() prior to register_netdev().
Signed-off-by: Fabio Estevam <fabio.estevam@freescale.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Commit efe4208f ("ipv6: make lookups simpler and faster") introduced a
regression in udp_v6_mcast_next(), resulting in multicast packets not
reaching the destination sockets under certain conditions.

The packet's IPv6 addresses are wrongly compared to the IPv6 addresses
from the function's socket argument, which indicates the starting point
for looping, instead of the loop variable. If the addresses from the
first socket do not match the packet's addresses, no socket in the list
will match.
Signed-off-by: Sven Wegener <sven.wegener@stealer.net>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Commit 4a55530f (net: sh_eth: modify the definitions of register) managed
to leave out the E-DMAC register entries in sh_eth_offset_fast_sh3_sh2[], thus
totally breaking SH7619/771x support.  Add the missing entries using  the data
from before that commit.
Signed-off-by: Sergei Shtylyov <sergei.shtylyov@cogentembedded.com>
Acked-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The current behaviour of the sh_eth driver is not to use the RNC bit
for the receive ring. This means that every packet recieved is not only
generating an IRQ but it also stops the receive ring DMA as well until
the driver re-enables it after unloading the packet.

This means that a number of the following errors are generated due to
the receive packet FIFO overflowing due to nowhere to put packets:

	net eth0: Receive FIFO Overflow

Since feedback from Yoshihiro Shimoda shows that every supported LSI
for this driver should have the bit enabled it seems the best way is
to remove the RMCR default value from the per-system data and just
write it when initialising the RMCR value. This is discussed in
the message (http://www.spinics.net/lists/netdev/msg284912.html).

I have tested the RMCR_RNC configuration with NFS root filesystem and
the driver has not failed yet.  There are further test reports from
Sergei Shtylov and others for both the R8A7790 and R8A7791.

There is also feedback fron Cao Minh Hiep[1] which reports the
same issue in (http://comments.gmane.org/gmane.linux.network/316285)
showing this fixes issues with losing UDP datagrams under iperf.
Tested-by: Sergei Shtylyov <sergei.shtylyov@cogentembedded.com>
Signed-off-by: Ben Dooks <ben.dooks@codethink.co.uk>
Acked-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>
Acked-by: Simon Horman <horms+renesas@verge.net.au>
Signed-off-by: David S. Miller <davem@davemloft.net>

It is possible that ->newlink() fails before registering
the device, in this case we should just free it, it's
safe to call free_netdev().

Fixes: commit 0e0eee24 (net: correct error path in rtnl_newlink())
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Cong Wang <cwang@twopensource.com>
Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

A rmb() is required to ensure that the CQE is not read before it
is written by the adapter DMA.  PCI ordering rules will make sure
the other fields are written before the marker at the end of struct
eth_fast_path_rx_cqe but without rmb() a weakly ordered processor can
process stale data.

Without the barrier we have observed various crashes including
bnx2x_tpa_start being called on queues not stopped (resulting in message
start of bin not in stop) and NULL pointer exceptions from bnx2x_rx_int.
Signed-off-by: Milton Miller <miltonm@us.ibm.com>
Signed-off-by: Wen Xiong <wenxiong@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

When injecting EEH error to bnx2x adapter, adapter couldn't be recovery
and caused recursive EEH errors. The patch fixes the issue.
Signed-off-by: Wen Xiong <wenxiong@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

As smsc driver supports carrier detection, it should unset NOCARRIER
flag only after carrier state determination. By default that flag
is off so driver should set it before starting auto-negotiation
Signed-off-by: Balakumaran <Balakumaran.Kannan@ap.sony.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The xfrm_user module registers its pernet init/exit after xfrm
itself so that its net exit function xfrm_user_net_exit() is
executed before xfrm_net_exit() which calls xfrm_state_fini() to
cleanup the SA's (xfrm states). This opens a window between
zeroing net->xfrm.nlsk pointer and deleting all xfrm_state
instances which may access it (via the timer). If an xfrm state
expires in this window, xfrm_exp_state_notify() will pass null
pointer as socket to nlmsg_multicast().

As the notifications are called inside rcu_read_lock() block, it
is sufficient to retrieve the nlsk socket with rcu_dereference()
and check the it for null.
Signed-off-by: Michal Kubecek <mkubecek@suse.cz>
Signed-off-by: David S. Miller <davem@davemloft.net>

Michael Chan says:

====================
cnic fixes

Fix 2 sleeping function from invalid context bugs and 1 missing iscsi netlink
message bug.

v2: Fixed typo in rcu_dereference_protected() and tested with CONFIG_PROVE_RCU
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

The iSCSI netlink message needs to be sent before the ulp_ops is cleared
as it is sent through a function pointer in the ulp_ops.  This bug
causes iscsid to not get the message when the bnx2i driver is unloaded.
Signed-off-by: Michael Chan <mchan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

We are allocating memory with GFP_KERNEL under spinlock.  Since this is
the only call manipulating the cnic_udev_list and it is always under
rtnl_lock, cnic_dev_lock can be safely removed.
Signed-off-by: Michael Chan <mchan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Because the called function, such as bnx2fc_indicate_netevent(), can sleep,
we cannot take rcu_lock().  To prevent the rcu protected ulp_ops from going
away, we use the cnic_lock mutex and set the ULP_F_CALL_PENDING flag.
The code already waits for ULP_F_CALL_PENDING flag to clear in
cnic_unregister_device().
Signed-off-by: Michael Chan <mchan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

In commit cd11cf50 I accidentally
added an error message. I used it for debugging and forgot to remove
it before submitting the patch.
Signed-off-by: Christian Riesch <christian.riesch@omicron.at>
Signed-off-by: David S. Miller <davem@davemloft.net>

Pull networking fixes from David Miller:

 1) Unbreak zebra and other netlink apps, from Eric W Biederman.

 2) Some new qmi_wwan device IDs, from Aleksander Morgado.

 3) Fix info leak in DCB netlink handler of qlcnic driver, from Dan
    Carpenter.

 4) inet_getid() and ipv6_select_ident() do not generate monotonically
    increasing ID numbers, fix from Eric Dumazet.

 5) Fix memory leak in __sk_prepare_filter(), from Leon Yu.

 6) Netlink leftover bytes warning message is user triggerable, rate
    limit it.  From Michal Schmidt.

 7) Fix non-linear SKB panic in ipvs, from Peter Christensen.

 8) Congestion window undo needs to be performed even if only never
    retransmitted data is SACK'd, fix from Yuching Cheng.

* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net: (24 commits)
  net: filter: fix possible memory leak in __sk_prepare_filter()
  net: ec_bhf: Add runtime dependencies
  tcp: fix cwnd undo on DSACK in F-RTO
  netlink: Only check file credentials for implicit destinations
  ipheth: Add support for iPad 2 and iPad 3
  team: fix mtu setting
  net: fix inet_getid() and ipv6_select_ident() bugs
  net: qmi_wwan: interface #11 in Sierra Wireless MC73xx is not QMI
  net: qmi_wwan: add additional Sierra Wireless QMI devices
  bridge: Prevent insertion of FDB entry with disallowed vlan
  netlink: rate-limit leftover bytes warning and print process name
  bridge: notify user space after fdb update
  net: qmi_wwan: add Netgear AirCard 341U
  net: fix wrong mac_len calculation for vlans
  batman-adv: fix NULL pointer dereferences
  net/mlx4_core: Reset RoCE VF gids when guest driver goes down
  emac: aggregation of v1-2 PLB errors for IER register
  emac: add missing support of 10mbit in emac/rgmii
  can: only rename enabled led triggers when changing the netdev name
  ipvs: Fix panic due to non-linear skb
  ...

__sk_prepare_filter() was reworked in commit bd4cf0ed (net: filter:
rework/optimize internal BPF interpreter's instruction set) so that it should
have uncharged memory once things went wrong. However that work isn't complete.
Error is handled only in __sk_migrate_filter() while memory can still leak in
the error path right after sk_chk_filter().

Fixes: bd4cf0ed ("net: filter: rework/optimize internal BPF interpreter's instruction set")
Signed-off-by: Leon Yu <chianglungyu@gmail.com>
Acked-by: Alexei Starovoitov <ast@plumgrid.com>
Tested-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Pull two md bugfixes from Neil Brown:
 "Two md bugfixes for possible corruption when restarting reshape

  If a raid5/6 reshape is restarted (After stopping and re-assembling
  the array) and the array is marked read-only (or read-auto), then the
  reshape will appear to complete immediately, without actually moving
  anything around.  This can result in corruption.

  There are two patches which do much the same thing in different
  places.  They are separate because one is an older bug and so can be
  applied to more -stable kernels"

* tag 'md/3.15-fixes' of git://neil.brown.name/md:
  md: always set MD_RECOVERY_INTR when interrupting a reshape thread.
  md: always set MD_RECOVERY_INTR when aborting a reshape or other "resync".

The ec_bhf driver is specific to the Beckhoff CX embedded PC series.
These are based on Intel x86 CPU. So we can add a dependency on
X86, with COMPILE_TEST as an alternative to still allow for broader
build-testing.
Signed-off-by: Jean Delvare <jdelvare@suse.de>
Cc: Darek Marcinkiewicz <reksio@newterm.pl>
Cc: David S. Miller <davem@davemloft.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

Queued trim only works for some users with MU05 firmware.  Revert to
blacklisting all firmware versions.

Introduced by commit d121f7d0 ("libata: Update queued trim blacklist
for M5x0 drives") which this effectively reverts, while retaining the
blacklisting of M550.

See

    https://bugzilla.kernel.org/show_bug.cgi?id=71371

for reports of trouble with MU05 firmware.
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Cc: Alan Cox <alan@lxorguk.ukuu.org.uk>
Cc: Tejun Heo <tj@kernel.org>
Cc: stable@vger.kernel.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

Pull x86 fix from Peter Anvin:
 "A single quite small patch that managed to get overlooked earlier, to
  prevent a user space triggerable oops on systems without HPET"

* 'x86-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
  x86, vdso: Fix an OOPS accessing the HPET mapping w/o an HPET

Pull USB fixes from Greg KH:
 "Here are some fixes for 3.15-rc8 that resolve a number of tiny USB
  issues that have been reported, and there are some new device ids as
  well.

  All have been tested in linux-next"

* tag 'usb-3.15-rc8' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/usb:
  xhci: delete endpoints from bandwidth list before freeing whole device
  usb: pci-quirks: Prevent Sony VAIO t-series from switching usb ports
  USB: cdc-wdm: properly include types.h
  usb: cdc-wdm: export cdc-wdm uapi header
  USB: serial: option: add support for Novatel E371 PCIe card
  USB: ftdi_sio: add NovaTech OrionLXm product ID
  USB: io_ti: fix firmware download on big-endian machines (part 2)
  USB: Avoid runtime suspend loops for HCDs that can't handle suspend/resume

Pull staging driver fixes from Greg KH:
 "Here are some staging driver fixes for 3.15.

  Three are for the speakup drivers (one fixes a regression caused in
  3.15-rc, and the other two resolve a tty issue found by Ben Hutchings)
  The comedi and r8192e_pci driver fixes also resolve reported issues"

* tag 'staging-3.15-rc8' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/staging:
  staging: r8192e_pci: fix htons error
  Staging: speakup: Update __speakup_paste_selection() tty (ab)usage to match vt
  Staging: speakup: Move pasting into a work item
  staging: comedi: ni_daq_700: add mux settling delay
  speakup: fix incorrect perms on speakup_acntsa.c

This bug is discovered by an recent F-RTO issue on tcpm list
https://www.ietf.org/mail-archive/web/tcpm/current/msg08794.html

The bug is that currently F-RTO does not use DSACK to undo cwnd in
certain cases: upon receiving an ACK after the RTO retransmission in
F-RTO, and the ACK has DSACK indicating the retransmission is spurious,
the sender only calls tcp_try_undo_loss() if some never retransmisted
data is sacked (FLAG_ORIG_DATA_SACKED).

The correct behavior is to unconditionally call tcp_try_undo_loss so
the DSACK information is used properly to undo the cwnd reduction.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

It was possible to get a setuid root or setcap executable to write to
it's stdout or stderr (which has been set made a netlink socket) and
inadvertently reconfigure the networking stack.

To prevent this we check that both the creator of the socket and
the currentl applications has permission to reconfigure the network
stack.

Unfortunately this breaks Zebra which always uses sendto/sendmsg
and creates it's socket without any privileges.

To keep Zebra working don't bother checking if the creator of the
socket has privilege when a destination address is specified.  Instead
rely exclusively on the privileges of the sender of the socket.

Note from Andy: This is exactly Eric's code except for some comment
clarifications and formatting fixes.  Neither I nor, I think, anyone
else is thrilled with this approach, but I'm hesitant to wait on a
better fix since 3.15 is almost here.

Note to stable maintainers: This is a mess.  An earlier series of
patches in 3.15 fix a rather serious security issue (CVE-2014-0181),
but they did so in a way that breaks Zebra.  The offending series
includes:

    commit aa4cf945
    Author: Eric W. Biederman <ebiederm@xmission.com>
    Date:   Wed Apr 23 14:28:03 2014 -0700

        net: Add variants of capable for use on netlink messages

If a given kernel version is missing that series of fixes, it's
probably worth backporting it and this patch.  if that series is
present, then this fix is critical if you care about Zebra.

Cc: stable@vger.kernel.org
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: Andy Lutomirski <luto@amacapital.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

Each iPad model has a different product id, this patch adds support for iPad 2
(pid 0x12a2) and iPad 3 (pid 0x12a6). Note that iPad 2 must be jailbroken and a
third-party app must be used for tethering to work. On iPad 3, tethering works
out of the box (assuming your ISP is nice).
Signed-off-by: Kristian Evensen <kristian.evensen@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Now it is not possible to set mtu to team device which has a port
enslaved to it. The reason is that when team_change_mtu() calls
dev_set_mtu() for port device, notificator for NETDEV_PRECHANGEMTU
event is called and team_device_event() returns NOTIFY_BAD forbidding
the change. So fix this by returning NOTIFY_DONE here in case team is
changing mtu in team_change_mtu().

Introduced-by: 3d249d4c "net: introduce ethernet teaming device"
Signed-off-by: Jiri Pirko <jiri@resnulli.us>
Acked-by: Flavio Leitner <fbl@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

I noticed we were sending wrong IPv4 ID in TCP flows when MTU discovery
is disabled.
Note how GSO/TSO packets do not have monotonically incrementing ID.

06:37:41.575531 IP (id 14227, proto: TCP (6), length: 4396)
06:37:41.575534 IP (id 14272, proto: TCP (6), length: 65212)
06:37:41.575544 IP (id 14312, proto: TCP (6), length: 57972)
06:37:41.575678 IP (id 14317, proto: TCP (6), length: 7292)
06:37:41.575683 IP (id 14361, proto: TCP (6), length: 63764)

It appears I introduced this bug in linux-3.1.

inet_getid() must return the old value of peer->ip_id_count,
not the new one.

Lets revert this part, and remove the prevention of
a null identification field in IPv6 Fragment Extension Header,
which is dubious and not even done properly.

Fixes: 87c48fa3 ("ipv6: make fragment identifications less predictable")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

This interface is unusable, as the cdc-wdm character device doesn't reply to
any QMI command. Also, the out-of-tree Sierra Wireless GobiNet driver fully
skips it.
Signed-off-by: Aleksander Morgado <aleksander@aleksander.es>
Acked-by: Bjørn Mork <bjorn@mork.no>
Signed-off-by: David S. Miller <davem@davemloft.net>

A set of new VID/PIDs retrieved from the out-of-tree GobiNet/GobiSerial
Sierra Wireless drivers.
Signed-off-by: Aleksander Morgado <aleksander@aleksander.es>
Acked-by: Bjørn Mork <bjorn@mork.no>
Signed-off-by: David S. Miller <davem@davemloft.net>

br_handle_local_finish() is allowing us to insert an FDB entry with
disallowed vlan. For example, when port 1 and 2 are communicating in
vlan 10, and even if vlan 10 is disallowed on port 3, port 3 can
interfere with their communication by spoofed src mac address with
vlan id 10.

Note: Even if it is judged that a frame should not be learned, it should
not be dropped because it is destined for not forwarding layer but higher
layer. See IEEE 802.1Q-2011 8.13.10.
Signed-off-by: Toshiaki Makita <makita.toshiaki@lab.ntt.co.jp>
Acked-by: Vlad Yasevich <vyasevic@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Any process is able to send netlink messages with leftover bytes.
Make the warning rate-limited to prevent too much log spam.

The warning is supposed to help find userspace bugs, so print the
triggering command name to implicate the buggy program.

[v2: Use pr_warn_ratelimited instead of printk_ratelimited.]
Signed-off-by: Michal Schmidt <mschmidt@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

There has been a number incidents recently where customers running KVM have
reported that VM hosts on different Hypervisors are unreachable. Based on
pcap traces we found that the bridge was broadcasting the ARP request out
onto the network. However some NICs have an inbuilt switch which on occasions
were broadcasting the VMs ARP request back through the physical NIC on the
Hypervisor. This resulted in the bridge changing ports and incorrectly learning
that the VMs mac address was external. As a result the ARP reply was directed
back onto the external network and VM never updated it's ARP cache. This patch
will notify the bridge command, after a fdb has been updated to identify such
port toggling.
Signed-off-by: Jon Maxwell <jmaxwell37@gmail.com>
Reviewed-by: Jiri Pirko <jiri@resnulli.us>
Acked-by: Toshiaki Makita <makita.toshiaki@lab.ntt.co.jp>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

Signed-off-by: Aleksander Morgado <aleksander@aleksander.es>
Acked-by: Bjørn Mork <bjorn@mork.no>
Signed-off-by: David S. Miller <davem@davemloft.net>

After 1e785f48 ("net: Start with correct mac_len in
skb_network_protocol") skb->mac_len is used as a start of the
calculation in skb_network_protocol() but that is not always correct. If
skb->protocol == 8021Q/AD, usually the vlan header is already inserted
in the skb (i.e. vlan reorder hdr == 0). Usually when the packet enters
dev_hard_xmit it has mac_len == 0 so we take 2 bytes from the
destination mac address (skb->data + VLAN_HLEN) as a type in
skb_network_protocol() and return vlan_depth == 4. In the case where TSO is
off, then the mac_len is set but it's == 18 (ETH_HLEN + VLAN_HLEN), so
skb_network_protocol() returns a type from inside the packet and
offset == 22. Also make vlan_depth unsigned as suggested before.
As suggested by Eric Dumazet, move the while() loop in the if() so we
can avoid additional testing in fast path.

Here are few netperf tests + debug printk's to illustrate:
cat netperf.tso-on.reorder-on.bugged
- Vlan -> device (reorder on, default, this case is okay)
MIGRATED TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to
192.168.3.1 () port 0 AF_INET
Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec

 87380  16384  16384    10.00    7111.54
[   81.605435] skb->len 65226 skb->gso_size 1448 skb->proto 0x800
skb->mac_len 0 vlan_depth 0 type 0x800

- Vlan -> device (reorder off, bad)
cat netperf.tso-on.reorder-off.bugged
MIGRATED TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to
192.168.3.1 () port 0 AF_INET
Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec

 87380  16384  16384    10.00     241.35
[  204.578332] skb->len 1518 skb->gso_size 0 skb->proto 0x8100
skb->mac_len 0 vlan_depth 4 type 0x5301
0x5301 are the last two bytes of the destination mac.

And if we stop TSO, we may get even the following:
[   83.343156] skb->len 2966 skb->gso_size 1448 skb->proto 0x8100
skb->mac_len 18 vlan_depth 22 type 0xb84
Because mac_len already accounts for VLAN_HLEN.

After the fix:
cat netperf.tso-on.reorder-off.fixed
MIGRATED TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to
192.168.3.1 () port 0 AF_INET
Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec

 87380  16384  16384    10.01    5001.46
[   81.888489] skb->len 65230 skb->gso_size 1448 skb->proto 0x8100
skb->mac_len 0 vlan_depth 18 type 0x800

CC: Vlad Yasevich <vyasevic@redhat.com>
CC: Eric Dumazet <eric.dumazet@gmail.com>
CC: Daniel Borkman <dborkman@redhat.com>
CC: David S. Miller <davem@davemloft.net>

Fixes:1e785f48 ("net: Start with correct mac_len in
skb_network_protocol")
Signed-off-by: Nikolay Aleksandrov <nikolay@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>