We have to wait until the full batch has been processed to deliver the
netlink error messages to userspace. Otherwise, we may deliver
duplicated errors to userspace in case that we need to abort and replay
the transaction if any of the required modules needs to be autoloaded.

A simple way to reproduce this (assumming nft_meta is not loaded) with
the following test file:

 add table filter
 add chain filter test
 add chain bad test                 # intentional wrong unexistent table
 add rule filter test meta mark 0

Then, when trying to load the batch:

 # nft -f test
 test:4:1-19: Error: Could not process rule: No such file or directory
 add chain bad test
 ^^^^^^^^^^^^^^^^^^^
 test:4:1-19: Error: Could not process rule: No such file or directory
 add chain bad test
 ^^^^^^^^^^^^^^^^^^^

The error is reported twice, once when the batch is aborted due to
missing nft_meta and another when it is fully processed.
Signed-off-by: NPablo Neira Ayuso <pablo@netfilter.org>

No need for rht_dereference() from rhashtable_destroy() since the
existing callers don't hold the mutex when invoking this function
from:

1) Netlink, this is called in case of memory allocation errors in the
   initialization path, no nl_sk_hash_lock is held.
2) Netfilter, this is called from the rcu callback, no nfnl_lock is
   held either.

I think it's reasonable to assume that the caller has to make sure
that no hash resizing may happen before releasing the bucket array.
Therefore, the caller should be responsible for releasing this in a
safe way, document this to make people aware of it.

This resolves a rcu lockdep splat in nft_hash:

===============================
[ INFO: suspicious RCU usage. ]
3.16.0+ #178 Not tainted
-------------------------------
lib/rhashtable.c:596 suspicious rcu_dereference_protected() usage!

other info that might help us debug this:

rcu_scheduler_active = 1, debug_locks = 1
1 lock held by ksoftirqd/2/18:
 #0:  (rcu_callback){......}, at: [<ffffffff810918fd>] rcu_process_callbacks+0x27e/0x4c7

stack backtrace:
CPU: 2 PID: 18 Comm: ksoftirqd/2 Not tainted 3.16.0+ #178
Hardware name: LENOVO 23259H1/23259H1, BIOS G2ET32WW (1.12 ) 05/30/2012
 0000000000000001 ffff88011706bb68 ffffffff8143debc 0000000000000000
 ffff880117062610 ffff88011706bb98 ffffffff81077515 ffff8800ca041a50
 0000000000000004 ffff8800ca386480 ffff8800ca041a00 ffff88011706bbb8
Call Trace:
 [<ffffffff8143debc>] dump_stack+0x4e/0x68
 [<ffffffff81077515>] lockdep_rcu_suspicious+0xfa/0x103
 [<ffffffff81228b1b>] rhashtable_destroy+0x46/0x52
 [<ffffffffa06f21a7>] nft_hash_destroy+0x73/0x82 [nft_hash]
Signed-off-by: NPablo Neira Ayuso <pablo@netfilter.org>
Acked-by: NThomas Graf <tgraf@suug.ch>

The sets are released from the rcu callback, after the rule is removed
from the chain list, which implies that nfnetlink cannot update the
rbtree and no packets are walking on the set anymore. Thus, we can get
rid of the spinlock in the set destroy path there.
Signed-off-by: NPablo Neira Ayuso <pablo@netfilter.org>
Reviewied-by: NThomas Graf <tgraf@suug.ch>

The sets are released from the rcu callback, after the rule is removed
from the chain list, which implies that nfnetlink cannot update the
hashes (thus, no resizing may occur) and no packets are walking on the
set anymore.

This resolves a lockdep splat in the nft_hash_destroy() path since the
nfnl mutex is not held there.

===============================
[ INFO: suspicious RCU usage. ]
3.16.0-rc2+ #168 Not tainted
-------------------------------
net/netfilter/nft_hash.c:362 suspicious rcu_dereference_protected() usage!

other info that might help us debug this:

rcu_scheduler_active = 1, debug_locks = 1
1 lock held by ksoftirqd/0/3:
 #0:  (rcu_callback){......}, at: [<ffffffff81096393>] rcu_process_callbacks+0x27e/0x4c7

stack backtrace:
CPU: 0 PID: 3 Comm: ksoftirqd/0 Not tainted 3.16.0-rc2+ #168
Hardware name: LENOVO 23259H1/23259H1, BIOS G2ET32WW (1.12 ) 05/30/2012
 0000000000000001 ffff88011769bb98 ffffffff8142c922 0000000000000006
 ffff880117694090 ffff88011769bbc8 ffffffff8107c3ff ffff8800cba52400
 ffff8800c476bea8 ffff8800c476bea8 ffff8800cba52400 ffff88011769bc08
Call Trace:
 [<ffffffff8142c922>] dump_stack+0x4e/0x68
 [<ffffffff8107c3ff>] lockdep_rcu_suspicious+0xfa/0x103
 [<ffffffffa079931e>] nft_hash_destroy+0x50/0x137 [nft_hash]
 [<ffffffffa078cd57>] nft_set_destroy+0x11/0x2a [nf_tables]
Signed-off-by: NPablo Neira Ayuso <pablo@netfilter.org>
Acked-by: NThomas Graf <tgraf@suug.ch>

During allocation and initialization of the network driver structures,
the wrong pointer is used to initialize a spin lock. Fix the spin lock
initialization by using the proper pointer.
Signed-off-by: NTom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The user_skb maybe be leaked if the operation on it failed and codes
skipped into the label "out:" without calling genlmsg_unicast.

Cc: Pravin Shelar <pshelar@nicira.com>
Signed-off-by: NLi RongQing <roy.qing.li@gmail.com>
Acked-by: NPravin B Shelar <pshelar@nicira.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

make defconfig reports:

warning: (NETFILTER_XT_TARGET_LOG) selects NF_LOG_IPV6 which has unmet direct dependencies (NET && INET && IPV6 && NETFILTER && NETFILTER_ADVANCED)

Fixes: d79a61d6 netfilter: NETFILTER_XT_TARGET_LOG selects NF_LOG_*
Reported-by: Nkbuild test robot <fengguang.wu@intel.com>
Signed-off-by: NPablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Pablo Neira Ayuso says:

====================
pull request: Netfilter/IPVS fixes for net

The following patchset contains seven Netfilter fixes for your net
tree, they are:

1) Make the NAT infrastructure independent of x_tables, some users are
   already starting to test nf_tables with NAT without enabling x_tables.
   Without this patch for Kconfig, there's a superfluous dependency
   between NAT and x_tables.
2) Allow to use 0 in the cgroup match, the kernel rejects with -EINVAL
   with no good reason. From Daniel Borkmann.

3) Select CONFIG_NF_NAT from the nf_tables NAT expression, this also
   resolves another NAT dependency with x_tables.

4) Use HAVE_JUMP_LABEL instead of CONFIG_JUMP_LABEL in the Netfilter hook
   code as elsewhere in the kernel to resolve toolchain problems, from
   Zhouyi Zhou.

5) Use iptunnel_handle_offloads() to set up tunnel encapsulation
   depending on the offload capabilities, reported by Alex Gartrell
   patch from Julian Anastasov.

6) Fix wrong family when registering the ip_vs_local_reply6() hook,
   also from Julian.

7) Select the NF_LOG_* symbols from NETFILTER_XT_TARGET_LOG. Rafał
   Miłecki reported that when jumping from 3.16 to 3.17-rc, his log
   target is not selected anymore due to changes in the previous
   development cycle to accomodate the full logging support for
   nf_tables.
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Some hosts can be both little and big endian.
In certain scenarios a big endian kernel can kexec a little endian kernel.

This patch fixes this case from both ends:
1) Return endianity to original values on shutdown (in case little endian kernel boots after we shutdown).
2) Do not rely on HW reset values when loading driver in little endian kernel
   but configure them explicitly (in case previous kernel was big endian and did not reset the HW).
Signed-off-by: NManish Chopra <manish.chopra@qlogic.com>
Signed-off-by: NAriel Elior <Ariel.Elior@qlogic.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When qeth device is queried for ethtool data, hardware operation
is performed to extract the necessary information from the card.
If the card is not online at the moment (e.g. it is undergoing
recovery), this operation produces undesired effects like
temporarily freezing the system. This patch prevents execution
of the hardware query operation when the card is not online.
In such case, ioctl() operation returns error with errno ENODEV.
Reviewed-by: NUrsula Braun <ursula.braun@de.ibm.com>
Signed-off-by: NEugene Crosser <Eugene.Crosser@ru.ibm.com>
Signed-off-by: NFrank Blaschka <blaschka@linux.vnet.ibm.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Calxeda 1G/10G XGMAC Ethernet support should be available only on
Calxeda ECX-1000/2000 (Highbank/Midway) platforms.
Signed-off-by: NBartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>
Acked-by: NKyungmin Park <kyungmin.park@samsung.com>
Cc: Rob Herring <robh@kernel.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Renesas SuperH Ethernet support should be available only on
Renesas ARM SoCs and SuperH architecture.
Signed-off-by: NBartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>
Acked-by: NKyungmin Park <kyungmin.park@samsung.com>
Cc: Simon Horman <horms@verge.net.au>
Cc: Magnus Damm <magnus.damm@gmail.com>
Acked-by: NSergei Shtylyov <sergei.shtylyov@cogentembedded.com>
Acked-by: NGeert Uytterhoeven <geert+renesas@glider.be>
Acked-by: NSimon Horman <horms+renesas@verge.net.au>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When a link is already up, the following sequence makes the kernel
block completely:

  ip link set dev eth0 down
  ip link set dev eth0 up

This is because on suspended phy, the following lines

  __lpc_eth_reset(pldat);
  __lpc_eth_init(pldat);

make the LPC ethernet core block (see LPC32x0 manual). The PHY needs to be
(re-)activated low-level first.
Signed-off-by: NRoland Stigge <stigge@antcom.de>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The patch fixes race conditions between PCI error recovery callbacks and
potential ifup/ifdown.

First, if ifup (tg3_open) is called between tg3_io_error_detected() and
tg3_io_resume() then tp->timer is armed twice before expiry. Once during
tg3_open() and again during tg3_io_resume(). This results in BUG
at kernel/time/timer.c:945.

Second, if ifdown (tg3_close) is called between tg3_io_error_detected()
and tg3_io_resume() then tg3_napi_disable() is called twice without
a tg3_napi_enable between. Once during tg3_io_error_detected() and again
during tg3_close(). The tg3_io_resume() then hangs on rtnl_lock().

v2: Added logging messages per Prashant's request

Cc: Prashant Sreedharan <prashant@broadcom.com>
Cc: Michael Chan <mchan@broadcom.com>
Signed-off-by: NIvan Vecera <ivecera@redhat.com>
Acked-by: NPrashant Sreedharan <prashant@broadcom.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Hariprasad Shenai says:

====================
Trivial fixes for cxgb4

This patch series adds support to fix T5 adapter accessing T4 adapter registers,
issue mbox command on correct mbox for physical function, avoid dumping write
only registers, use correct length for adapter part number and support to detect
and display firmware reported errors.

The patches series is created against 'net' tree.
And includes patches on cxgb4 driver.

We have included all the maintainers of respective drivers. Kindly review the
change and let us know in case of any review comments.

Thanks

V2:
   Added description for each patch as per David Miller's comment
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

A couple of RDMA-related called to t4_query_params() were issuing mbox commands
on mbox0 instead of mbox4.
Signed-off-by: NHariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Avoid dumping MPS_RPLC_MAP_CTL for reg dumps; this is a Write-Only register.
Reading this register may cause MPS TCAM corruption.
Signed-off-by: NHariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The adapter firmware can indicate error conditions to the host.
If the firmware has indicated an error, print out the reason for
the firmware error.

Based on original work by Casey Leedom <leedom@chelsio.com>
Signed-off-by: NHariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Fixes few register access for both T4 and T5.
PCIE_CORE_UTL_SYSTEM_BUS_AGENT_STATUS & PCIE_CORE_UTL_PCI_EXPRESS_PORT_STATUS
is T4 only register don't let T5 access them. For T5 MA_PARITY_ERROR_STATUS2
is additionally read. MPS_TRC_RSS_CONTROL is T4 only register, for T5 use
MPS_T5_TRC_RSS_CONTROL.
Signed-off-by: NHariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Previously it was using the length value of serial number.
Also added macro for VPD unique identifier (0x82).
Signed-off-by: NCasey Leedom <leedom@chelsio.com>
Signed-off-by: NHariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

We previously assumed that a Port's Capabilities and Advertised Capabilities
would never change from Port Initialization time.  This is no longer true
when we can have 10Gb/s and 1Gb/s SFP+ Transceiver Modules randomly swapped.
Signed-off-by: NHariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

In case of the HW is not able to do the receive checksum offloading
the only feature to remove is NETIF_F_RXCSUM.
Signed-off-by: NGiuseppe Cavallaro <peppe.cavallaro@st.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

For new GMACs it is possible to turn-on/off the COE.
In the current driver, when disabled the Rx-checksum
via ethtool, the tool reported that csum was disabled
but the HW continued to set the IPC. Indeed this is
because the fix_features allows this. So the patch
fixes this problem by adding the set_features.
Signed-off-by: NGiuseppe Cavallaro <peppe.cavallaro@st.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Tom Lendacky says:

====================
amd-xgbe: AMD XGBE driver fixes 2014-08-29

The following series of patches includes fixes to the driver.

- Tx hardware queue flushing support dependent on hardware version
- Incorrect reported fifo size
- Proper mmd select in XPCS debugfs support
- Proper queue count for configuring Tx flow control

This patch series is based on net.
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When configuring Tx flow control the Rx queue count was used instead of
the Tx queue count for looping through the Tx hardware queues. Fix the
code to use the Tx queue count.
Signed-off-by: NTom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The debugfs support for the xpcs registers did not properly use the
specified mmd (xpcs_mmd entry) which resulted in the default mmd
value always being used.  Update the debugfs support to generate the
proper mmd register value.
Signed-off-by: NTom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The fifo size reported by the hardware is not correct. Add support
to limit the reported size to what is actually present.  Also, fix
the argument types used in the fifo size calculation function.
Signed-off-by: NTom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The flushing of the Tx hardware queues is only supported at a certain
level of the hardware.  Retrieve the current version of the hardware
and use that to determine if flushing is supported.
Signed-off-by: NTom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

If NO_DMA=y:

drivers/built-in.o: In function `xgene_enet_delete_ring':
xgene_enet_main.c:(.text+0x28755a): undefined reference to `dma_free_coherent'
drivers/built-in.o: In function `xgene_enet_setup_tx_desc':
xgene_enet_main.c:(.text+0x287774): undefined reference to `dma_map_single'
xgene_enet_main.c:(.text+0x287780): undefined reference to `dma_mapping_error'
drivers/built-in.o: In function `xgene_enet_tx_completion':
xgene_enet_main.c:(.text+0x2878e6): undefined reference to `dma_unmap_single'
drivers/built-in.o: In function `xgene_enet_refill_bufpool':
xgene_enet_main.c:(.text+0x2879d4): undefined reference to `dma_map_single'
xgene_enet_main.c:(.text+0x2879e0): undefined reference to `dma_mapping_error'
drivers/built-in.o: In function `xgene_enet_rx_frame':
xgene_enet_main.c:(.text+0x287aaa): undefined reference to `dma_unmap_single'
drivers/built-in.o: In function `xgene_enet_free_desc_ring':
xgene_enet_main.c:(.text+0x287f98): undefined reference to `dma_free_coherent'
drivers/built-in.o: In function `xgene_enet_create_desc_ring':
xgene_enet_main.c:(.text+0x28808e): undefined reference to `dma_alloc_coherent'
drivers/built-in.o: In function `xgene_enet_probe':
xgene_enet_main.c:(.text+0x2883d4): undefined reference to `dma_set_mask'
xgene_enet_main.c:(.text+0x2883ec): undefined reference to `dma_supported'
Signed-off-by: NGeert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

John W. Linville says:

====================
pull request: wireless 2014-08-28

Please pull this batch of fixes intended for the 3.17 stream.

For the Bluetooth/6LowPAN/802.15.4 bits, Johan says:

'It contains a connection reference counting fix for LE where a
connection might stay up even though it should get disconnected.

The other 802.15.4 6LoWPAN related patches were sent to the bluetooth
tree by Alexander Aring and described as follows by him:

"
these patches contains patches for the bluetooth branch.

This series includes memory leak fixes and an errno value fix.
Also there are two patches for sending and receiving 1280 6LoWPAN
packets, which makes the IEEE 802.15.4 6LoWPAN stack more RFC
compliant.
"'

Along with that...

Alexey Khoroshilov fixes a use-after-free bug on at76c50x-usb.

Hauke Mehrtens adds a PCI ID to bcma.

Himangi Saraogi fixes a silly "A || A" test in rtlwifi.

Larry Finger adds a device ID to rtl8192cu.

Maks Naumov fixes a strncmp argument in ath9k.

Álvaro Fernández Rojas adds a PCI ID to ssb.
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

CONFIG_NETFILTER_XT_TARGET_LOG is not selected anymore when jumping
from 3.16 to 3.17-rc1 if you don't set on the new NF_LOG_IPV4 and
NF_LOG_IPV6 switches.

Change this to select the three new symbols NF_LOG_COMMON, NF_LOG_IPV4
and NF_LOG_IPV6 instead, so NETFILTER_XT_TARGET_LOG remains enabled
when moving from old to new kernels.
Reported-by: NRafał Miłecki <zajec5@gmail.com>
Signed-off-by: NPablo Neira Ayuso <pablo@netfilter.org>

Giuseppe Cavallaro says:

====================
stmmac EEE fixes

This is a subset of patches to provide some fixes for the EEE support inside the
driver.
Patches have been tested on boards EEE capable plugged on switch w/ w/o EEE
support.
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

In case of PLS is active the PLS (PHY Link Status) bit in
the Reg12 has to be set to allow the MAC to asserts the LPI
pattern when the link is ok.
Signed-off-by: Nnandini sharma <nandini.sharma@st.com>
Signed-off-by: NGiuseppe Cavallaro <peppe.cavallaro@st.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch is to skip the EEE initialisation when the stmmac
is using a switch (with a fixed phy support).
Signed-off-by: NGiuseppe Cavallaro <peppe.cavallaro@st.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The value for LPI TW timer has to be updated to 0x1E that is the hardcoded value
of 20.5us and it will apply to all EEE enabled Remote PHYs.
Disadvantage is for PHY's that support lesser wakeup time but we can accept it
waiting to implement LLDP to negotiate the Wakeup time of Remote PHY.
Signed-off-by: Nnandini sharma <nandini.sharma@st.com>
Signed-off-by: NGiuseppe Cavallaro <peppe.cavallaro@st.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch is to fix the definition of macros for EEE otherwise the LPI TX/RX
entry/exit cannot be properly managed.
Signed-off-by: NNandini Sharma <nandini.sharma@st.com>
Signed-off-by: NGiuseppe Cavallaro <peppe.cavallaro@st.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Since SCTP day 1, that is, 19b55a2af145 ("Initial commit") from lksctp
tree, the official <netinet/sctp.h> header carries a copy of enum
sctp_sstat_state that looks like (compared to the current in-kernel
enumeration):

  User definition:                     Kernel definition:

  enum sctp_sstat_state {              typedef enum {
    SCTP_EMPTY             = 0,          <removed>
    SCTP_CLOSED            = 1,          SCTP_STATE_CLOSED            = 0,
    SCTP_COOKIE_WAIT       = 2,          SCTP_STATE_COOKIE_WAIT       = 1,
    SCTP_COOKIE_ECHOED     = 3,          SCTP_STATE_COOKIE_ECHOED     = 2,
    SCTP_ESTABLISHED       = 4,          SCTP_STATE_ESTABLISHED       = 3,
    SCTP_SHUTDOWN_PENDING  = 5,          SCTP_STATE_SHUTDOWN_PENDING  = 4,
    SCTP_SHUTDOWN_SENT     = 6,          SCTP_STATE_SHUTDOWN_SENT     = 5,
    SCTP_SHUTDOWN_RECEIVED = 7,          SCTP_STATE_SHUTDOWN_RECEIVED = 6,
    SCTP_SHUTDOWN_ACK_SENT = 8,          SCTP_STATE_SHUTDOWN_ACK_SENT = 7,
  };                                   } sctp_state_t;

This header was later on also placed into the uapi, so that user space
programs can compile without having <netinet/sctp.h>, but the shipped
with <linux/sctp.h> instead.

While RFC6458 under 8.2.1.Association Status (SCTP_STATUS) says that
sstat_state can range from SCTP_CLOSED to SCTP_SHUTDOWN_ACK_SENT, we
nevertheless have a what it appears to be dummy SCTP_EMPTY state from
the very early days.

While it seems to do just nothing, commit 0b8f9e25 ("sctp: remove
completely unsed EMPTY state") did the right thing and removed this dead
code. That however, causes an off-by-one when the user asks the SCTP
stack via SCTP_STATUS API and checks for the current socket state thus
yielding possibly undefined behaviour in applications as they expect
the kernel to tell the right thing.

The enumeration had to be changed however as based on the current socket
state, we access a function pointer lookup-table through this. Therefore,
I think the best way to deal with this is just to add a helper function
sctp_assoc_to_state() to encapsulate the off-by-one quirk.
Reported-by: NTristan Su <sooqing@gmail.com>
Fixes: 0b8f9e25 ("sctp: remove completely unsed EMPTY state")
Signed-off-by: NDaniel Borkmann <dborkman@redhat.com>
Acked-by: NVlad Yasevich <vyasevich@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

In commit ed98df33 ("net: use __GFP_NORETRY for high order
allocations") we tried to address one issue caused by order-3
allocations.

We still observe high latencies and system overhead in situations where
compaction is not successful.

Instead of trying order-3, order-2, and order-1, do a single order-3
best effort and immediately fallback to plain order-0.

This mimics slub strategy to fallback to slab min order if the high
order allocation used for performance failed.

Order-3 allocations give a performance boost only if they can be done
without recurring and expensive memory scan.

Quoting David :

The page allocator relies on synchronous (sync light) memory compaction
after direct reclaim for allocations that don't retry and deferred
compaction doesn't work with this strategy because the allocation order
is always decreasing from the previous failed attempt.

This means sync light compaction will always be encountered if memory
cannot be defragmented or reclaimed several times during the
skb_page_frag_refill() iteration.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Acked-by: NDavid Rientjes <rientjes@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Or Gerlitz says:

====================
Setup mlx4 user space Ethernet QPs to properly handle VXLAN

This short series fixes the mlx4 driver setting of user space Ethernet QPs
(e.g those opened by DPDK applications) such that they will properly handle
VXLAN traffic/offloads
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Raw Ethernet QPs opened from user-space lack the proper setup to
recieve/handle VXLAN traffic when VXLAN offloads are enabled.

Fix that by adding a tunnel steering rule on top of the normal unicast
steering rule and set the tunnel_type field in the QP context.
Signed-off-by: NOr Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>