Linux TCP stack painfully segments all TSO/GSO packets before retransmits.

This was fine back in the days when TSO/GSO were emerging, with their
bugs, but we believe the dark age is over.

Keeping big packets in write queues, but also in stack traversal
has a lot of benefits.
 - Less memory overhead, because write queues have less skbs
 - Less cpu overhead at ACK processing.
 - Better SACK processing, as lot of studies mentioned how
   awful linux was at this ;)
 - Less cpu overhead to send the rtx packets
   (IP stack traversal, netfilter traversal, drivers...)
 - Better latencies in presence of losses.
 - Smaller spikes in fq like packet schedulers, as retransmits
   are not constrained by TCP Small Queues.

1 % packet losses are common today, and at 100Gbit speeds, this
translates to ~80,000 losses per second.
Losses are often correlated, and we see many retransmit events
leading to 1-MSS train of packets, at the time hosts are already
under stress.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Acked-by: NYuchung Cheng <ycheng@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Commit 42b18f60 ("tipc: refactor function tipc_link_timeout()"),
introduced a bug which prevents sending of probe messages during
link synchronization phase. This leads to hanging links, if the
bearer is disabled/enabled after links are up.

In this commit, we send the probe messages correctly.

Fixes: 42b18f60 ("tipc: refactor function tipc_link_timeout()")
Acked-by: NJon Maloy <jon.maloy@ericsson.com>
Signed-off-by: NParthasarathy Bhuvaragan <parthasarathy.bhuvaragan@ericsson.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Martin KaFai Lau says:

====================
tcp: Handle txstamp_ack when fragmenting/coalescing skbs

This patchset is to handle the txstamp-ack bit when
fragmenting/coalescing skbs.

The second patch depends on the recently posted series
for the net branch:
"tcp: Merge timestamp info when coalescing skbs"

A BPF prog is used to kprobe to sock_queue_err_skb()
and print out the value of serr->ee.ee_data.  The BPF
prog (run-able from bcc) is attached here:

BPF prog used for testing:
~~~~~

from __future__ import print_function
from bcc import BPF

bpf_text = """

int trace_err_skb(struct pt_regs *ctx)
{
	struct sk_buff *skb = (struct sk_buff *)ctx->si;
	struct sock *sk = (struct sock *)ctx->di;
	struct sock_exterr_skb *serr;
	u32 ee_data = 0;

	if (!sk || !skb)
		return 0;

	serr = SKB_EXT_ERR(skb);
	bpf_probe_read(&ee_data, sizeof(ee_data), &serr->ee.ee_data);
	bpf_trace_printk("ee_data:%u\\n", ee_data);

	return 0;
};
"""

b = BPF(text=bpf_text)
b.attach_kprobe(event="sock_queue_err_skb", fn_name="trace_err_skb")
print("Attached to kprobe")
b.trace_print()
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When collapsing skbs, txstamp_ack also needs to be merged.

Retrans Collapse Test:
~~~~~~
0.200 accept(3, ..., ...) = 4
+0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0

0.200 write(4, ..., 730) = 730
+0 setsockopt(4, SOL_SOCKET, 37, [2688], 4) = 0
0.200 write(4, ..., 730) = 730
+0 setsockopt(4, SOL_SOCKET, 37, [2176], 4) = 0
0.200 write(4, ..., 11680) = 11680

0.200 > P. 1:731(730) ack 1
0.200 > P. 731:1461(730) ack 1
0.200 > . 1461:8761(7300) ack 1
0.200 > P. 8761:13141(4380) ack 1

0.300 < . 1:1(0) ack 1 win 257 <sack 1461:2921,nop,nop>
0.300 < . 1:1(0) ack 1 win 257 <sack 1461:4381,nop,nop>
0.300 < . 1:1(0) ack 1 win 257 <sack 1461:5841,nop,nop>
0.300 > P. 1:1461(1460) ack 1
0.400 < . 1:1(0) ack 13141 win 257

BPF Output Before:
~~~~~
<No output due to missing SCM_TSTAMP_ACK timestamp>

BPF Output After:
~~~~~
<...>-2027  [007] d.s.    79.765921: : ee_data:1459

Sacks Collapse Test:
~~~~~
0.200 accept(3, ..., ...) = 4
+0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0

0.200 write(4, ..., 1460) = 1460
+0 setsockopt(4, SOL_SOCKET, 37, [2688], 4) = 0
0.200 write(4, ..., 13140) = 13140
+0 setsockopt(4, SOL_SOCKET, 37, [2176], 4) = 0

0.200 > P. 1:1461(1460) ack 1
0.200 > . 1461:8761(7300) ack 1
0.200 > P. 8761:14601(5840) ack 1

0.300 < . 1:1(0) ack 1 win 257 <sack 1461:14601,nop,nop>
0.300 > P. 1:1461(1460) ack 1
0.400 < . 1:1(0) ack 14601 win 257

BPF Output Before:
~~~~~
<No output due to missing SCM_TSTAMP_ACK timestamp>

BPF Output After:
~~~~~
<...>-2049  [007] d.s.    89.185538: : ee_data:14599
Signed-off-by: NMartin KaFai Lau <kafai@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Willem de Bruijn <willemb@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Acked-by: NSoheil Hassas Yeganeh <soheil@google.com>
Tested-by: NSoheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When a tcp skb is sliced into two smaller skbs (e.g. in
tcp_fragment() and tso_fragment()),  it does not carry
the txstamp_ack bit to the newly created skb if it is needed.
The end result is a timestamping event (SCM_TSTAMP_ACK) will
be missing from the sk->sk_error_queue.

This patch carries this bit to the new skb2
in tcp_fragment_tstamp().

BPF Output Before:
~~~~~~
<No output due to missing SCM_TSTAMP_ACK timestamp>

BPF Output After:
~~~~~~
<...>-2050  [000] d.s.   100.928763: : ee_data:14599

Packetdrill Script:
~~~~~~
+0 `sysctl -q -w net.ipv4.tcp_min_tso_segs=10`
+0 `sysctl -q -w net.ipv4.tcp_no_metrics_save=1`
+0 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3
+0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0
+0 bind(3, ..., ...) = 0
+0 listen(3, 1) = 0

0.100 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 7>
0.100 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 7>
0.200 < . 1:1(0) ack 1 win 257
0.200 accept(3, ..., ...) = 4
+0 setsockopt(4, SOL_TCP, TCP_NODELAY, [1], 4) = 0

+0 setsockopt(4, SOL_SOCKET, 37, [2688], 4) = 0
0.200 write(4, ..., 14600) = 14600
+0 setsockopt(4, SOL_SOCKET, 37, [2176], 4) = 0

0.200 > . 1:7301(7300) ack 1
0.200 > P. 7301:14601(7300) ack 1

0.300 < . 1:1(0) ack 14601 win 257

0.300 close(4) = 0
0.300 > F. 14601:14601(0) ack 1
0.400 < F. 1:1(0) ack 16062 win 257
0.400 > . 14602:14602(0) ack 2
Signed-off-by: NMartin KaFai Lau <kafai@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Willem de Bruijn <willemb@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Acked-by: NSoheil Hassas Yeganeh <soheil@google.com>
Tested-by: NSoheil Hassas Yeganeh <soheil@google.com>
Acked-by: NWillem de Bruijn <willemb@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Pablo Neira Ayuso says:

====================
Netfilter updates for net-next

The following patchset contains Netfilter updates for your net-next
tree, mostly from Florian Westphal to sort out the lack of sufficient
validation in x_tables and connlabel preparation patches to add
nf_tables support. They are:

1) Ensure we don't go over the ruleset blob boundaries in
   mark_source_chains().

2) Validate that target jumps land on an existing xt_entry. This extra
   sanitization comes with a performance penalty when loading the ruleset.

3) Introduce xt_check_entry_offsets() and use it from {arp,ip,ip6}tables.

4) Get rid of the smallish check_entry() functions in {arp,ip,ip6}tables.

5) Make sure the minimal possible target size in x_tables.

6) Similar to #3, add xt_compat_check_entry_offsets() for compat code.

7) Check that standard target size is valid.

8) More sanitization to ensure that the target_offset field is correct.

9) Add xt_check_entry_match() to validate that matches are well-formed.

10-12) Three patch to reduce the number of parameters in
    translate_compat_table() for {arp,ip,ip6}tables by using a container
    structure.

13) No need to return value from xt_compat_match_from_user(), so make
    it void.

14) Consolidate translate_table() so it can be used by compat code too.

15) Remove obsolete check for compat code, so we keep consistent with
    what was already removed in the native layout code (back in 2007).

16) Get rid of target jump validation from mark_source_chains(),
    obsoleted by #2.

17) Introduce xt_copy_counters_from_user() to consolidate counter
    copying, and use it from {arp,ip,ip6}tables.

18,22) Get rid of unnecessary explicit inlining in ctnetlink for dump
    functions.

19) Move nf_connlabel_match() to xt_connlabel.

20) Skip event notification if connlabel did not change.

21) Update of nf_connlabels_get() to make the upcoming nft connlabel
    support easier.

23) Remove spinlock to read protocol state field in conntrack.
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Nicolas Dichtel says:

====================
netlink: align attributes when needed (patchset #1)

This is the continuation of the work done to align netlink attributes
when these attributes contain some 64-bit fields.

David, if the third patch is too big (or maybe the series), I can split it.
Just tell me what you prefer.
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Goal of this patch is to use the new libnl API to align netlink attribute
when needed.
The layout of the netlink message will be a bit different after the patch,
because the padattr (TASKSTATS_TYPE_STATS) will be inside the nested
attribute instead of before it.
Signed-off-by: NNicolas Dichtel <nicolas.dichtel@6wind.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NNicolas Dichtel <nicolas.dichtel@6wind.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

With this function, nla_data() is aligned on a 64-bit area.
Signed-off-by: NNicolas Dichtel <nicolas.dichtel@6wind.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

nla_data() is now aligned on a 64-bit area.
Signed-off-by: NNicolas Dichtel <nicolas.dichtel@6wind.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

nla_data() is now aligned on a 64-bit area.
In fact, there is no user of this function.
Signed-off-by: NNicolas Dichtel <nicolas.dichtel@6wind.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

nla_data() is now aligned on a 64-bit area.

The temporary function nla_put_be64_32bit() is removed in this patch.
Signed-off-by: NNicolas Dichtel <nicolas.dichtel@6wind.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

nla_data() is now aligned on a 64-bit area.

A temporary version (nla_put_be64_32bit()) is added for nla_put_net64().
This function is removed in the next patch.
Signed-off-by: NNicolas Dichtel <nicolas.dichtel@6wind.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

nla_data() is now aligned on a 64-bit area.
Signed-off-by: NNicolas Dichtel <nicolas.dichtel@6wind.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Fix typo and describe 'padattr'.

Fixes: 089bf1a6 ("libnl: add more helpers to align attributes on 64-bit")
Signed-off-by: NNicolas Dichtel <nicolas.dichtel@6wind.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Conflicts were two cases of simple overlapping changes,
nothing serious.

In the UDP case, we need to add a hlist_add_tail_rcu()
to linux/rculist.h, because we've moved UDP socket handling
away from using nulls lists.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Pull RTC fixes from Alexandre Belloni:
 "A few fixes for the RTC subsystem.  The documentation fix already
  missed 4.5 so I think it is worth taking it now:

  A documentation fix for s3c and two fixes for the ds1307"

* tag 'rtc-4.6-3' of git://git.kernel.org/pub/scm/linux/kernel/git/abelloni/linux:
  rtc: ds1307: Use irq when available for wakeup-source device
  rtc: ds1307: ds3231 temperature s16 overflow
  rtc: s3c: Document in binding that only s3c6410 needs a src clk

Pull power management fixes from Rafael Wysocki:
 "Two fixes for issues introduced recently, one for an intel_pstate
  driver problem uncovered by the recent switch over from using timers
  and the other one for a potential cpufreq core problem related to
  system suspend/resume.

  Specifics:

   - Fix an intel_pstate driver problem causing CPUs to get stuck in the
     highest P-state when completely idle uncovered by the recent switch
     over from using timers (Rafael Wysocki).

   - Avoid attempts to get the current CPU frequency when all devices
     (like I2C controllers that may be nedded for that purpose) have
     been suspended during system suspend/resume (Rafael Wysocki)"

* tag 'pm+acpi-4.6-rc5' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm:
  cpufreq: Abort cpufreq_update_current_freq() for cpufreq_suspended set
  intel_pstate: Avoid getting stuck in high P-states when idle

With commit 8bc2a407 ("rtc: ds1307: add support for the
DT property 'wakeup-source'") we lost the ability for rtc irq
functionality for devices that are actually hooked on a real IRQ
line and have capability to wakeup as well. This is not an expected
behavior. So, instead of just not requesting IRQ, skip the IRQ
requirement only if interrupts are not defined for the device.

Fixes: 8bc2a407 ("rtc: ds1307: add support for the DT property 'wakeup-source'")
Reported-by: NTony Lindgren <tony@atomide.com>
Cc: Michael Lange <linuxstuff@milaw.biz>
Cc: Alexandre Belloni <alexandre.belloni@free-electrons.com>
Signed-off-by: NNishanth Menon <nm@ti.com>
Signed-off-by: NAlexandre Belloni <alexandre.belloni@free-electrons.com>

while retrieving temperature from ds3231, the result may be overflow
since s16 is too small for a multiplication with 250.

ie. if temp_buf[0] == 0x2d, the result (s16 temp) will be negative.
Signed-off-by: NAkinobu Mita <akinobu.mita@gmail.com>
Tested-by: NMichael Tatarinov <kukabu@gmail.com>
Signed-off-by: NAlexandre Belloni <alexandre.belloni@free-electrons.com>

Pull networking fixes from David Miller:

 1) Fix memory leak in iwlwifi, from Matti Gottlieb.

 2) Add missing registration of netfilter arp_tables into initial
    namespace, from Florian Westphal.

 3) Fix potential NULL deref in DecNET routing code.

 4) Restrict NETLINK_URELEASE to truly bound sockets only, from Dmitry
    Ivanov.

 5) Fix dst ref counting in VRF, from David Ahern.

 6) Fix TSO segmenting limits in i40e driver, from Alexander Duyck.

 7) Fix heap leak in PACKET_DIAG_MCLIST, from Mathias Krause.

 8) Ravalidate IPV6 datagram socket cached routes properly, particularly
    with UDP, from Martin KaFai Lau.

 9) Fix endian bug in RDS dp_ack_seq handling, from Qing Huang.

10) Fix stats typing in bcmgenet driver, from Eric Dumazet.

11) Openvswitch needs to orphan SKBs before ipv6 fragmentation handing,
    from Joe Stringer.

12) SPI device reference leak in spi_ks8895 PHY driver, from Mark Brown.

13) atl2 doesn't actually support scatter-gather, so don't advertise the
    feature.  From Ben Hucthings.

* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net: (72 commits)
  openvswitch: use flow protocol when recalculating ipv6 checksums
  Driver: Vmxnet3: set CHECKSUM_UNNECESSARY for IPv6 packets
  atl2: Disable unimplemented scatter/gather feature
  net/mlx4_en: Split SW RX dropped counter per RX ring
  net/mlx4_core: Don't allow to VF change global pause settings
  net/mlx4_core: Avoid repeated calls to pci enable/disable
  net/mlx4_core: Implement pci_resume callback
  net: phy: spi_ks8895: Don't leak references to SPI devices
  net: ethernet: davinci_emac: Fix platform_data overwrite
  net: ethernet: davinci_emac: Fix Unbalanced pm_runtime_enable
  qede: Fix single MTU sized packet from firmware GRO flow
  qede: Fix setting Skb network header
  qede: Fix various memory allocation error flows for fastpath
  tcp: Merge tx_flags and tskey in tcp_shifted_skb
  tcp: Merge tx_flags and tskey in tcp_collapse_retrans
  drivers: net: cpsw: fix wrong regs access in cpsw_ndo_open
  tcp: Fix SOF_TIMESTAMPING_TX_ACK when handling dup acks
  openvswitch: Orphan skbs before IPv6 defrag
  Revert "Prevent NUll pointer dereference with two PHYs on cpsw"
  VSOCK: Only check error on skb_recv_datagram when skb is NULL
  ...

Hannes Frederic Sowa says:

====================
net: network drivers should not depend on geneve/vxlan

This patchset removes the dependency of network drivers on vxlan or
geneve, so those don't get autoloaded when the nic driver is loaded.

Also audited the code such that vxlan_get_rx_port and geneve_get_rx_port
are not called without rtnl lock.
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Equivalent to "vxlan: break dependency with netdev drivers", don't
autoload geneve module in case the driver is loaded. Instead make the
coupling weaker by using netdevice notifiers as proxy.

Cc: Jesse Gross <jesse@kernel.org>
Signed-off-by: NHannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Currently all drivers depend and autoload the vxlan module because how
vxlan_get_rx_port is linked into them. Remove this dependency:

By using a new event type in the netdevice notifier call chain we proxy
the request from the drivers to flush and resetup the vxlan ports not
directly via function call but by the already existing netdevice
notifier call chain.

I added a separate new event type, NETDEV_OFFLOAD_PUSH_VXLAN, to do so.
We don't need to save those ids, as the event type field is an unsigned
long and using specialized event types for this purpose seemed to be a
more elegant way. This also comes in beneficial if in future we want to
add offloading knobs for vxlan.

Cc: Jesse Gross <jesse@kernel.org>
Signed-off-by: NHannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

qlcnic_attach_func requires rtnl_lock to be held.

Cc: Dept-GELinuxNICDev@qlogic.com
Signed-off-by: NHannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

vxlan_get_rx_port requires rtnl_lock to be held.

Cc: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Cc: Jesse Brandeburg <jesse.brandeburg@intel.com>
Cc: Shannon Nelson <shannon.nelson@intel.com>
Cc: Carolyn Wyborny <carolyn.wyborny@intel.com>
Cc: Don Skidmore <donald.c.skidmore@intel.com>
Cc: Bruce Allan <bruce.w.allan@intel.com>
Cc: John Ronciak <john.ronciak@intel.com>
Cc: Mitch Williams <mitch.a.williams@intel.com>
Signed-off-by: NHannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

mlx4_en_start_port requires rtnl_lock to be held.

Cc: Eugenia Emantayev <eugenia@mellanox.com>
Cc: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: NHannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

fm10k_open requires rtnl_lock to be held.

Cc: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Cc: Jesse Brandeburg <jesse.brandeburg@intel.com>
Cc: Shannon Nelson <shannon.nelson@intel.com>
Cc: Carolyn Wyborny <carolyn.wyborny@intel.com>
Cc: Don Skidmore <donald.c.skidmore@intel.com>
Cc: Bruce Allan <bruce.w.allan@intel.com>
Cc: John Ronciak <john.ronciak@intel.com>
Cc: Mitch Williams <mitch.a.williams@intel.com>
Signed-off-by: NHannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

be_open calls down to functions which expects rtnl lock to be held.

Cc: Sathya Perla <sathya.perla@broadcom.com>
Cc: Ajit Khaparde <ajit.khaparde@broadcom.com>
Cc: Padmanabh Ratnakar <padmanabh.ratnakar@broadcom.com>
Cc: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Cc: Somnath Kotur <somnath.kotur@broadcom.com>
Signed-off-by: NHannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When using masked actions the ipv6_proto field of an action
to set IPv6 fields may be zero rather than the prevailing protocol
which will result in skipping checksum recalculation.

This patch resolves the problem by relying on the protocol
in the flow key rather than that in the set field action.

Fixes: 83d2b9ba ("net: openvswitch: Support masked set actions.")
Cc: Jarno Rajahalme <jrajahalme@nicira.com>
Signed-off-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

For IPv6, if the device indicates that the checksum is correct, set
CHECKSUM_UNNECESSARY.
Reported-by: NSubbarao Narahari <snarahari@vmware.com>
Signed-off-by: NShrikrishna Khare <skhare@vmware.com>
Signed-off-by: NJin Heo <heoj@vmware.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

atl2 includes NETIF_F_SG in hw_features even though it has no support
for non-linear skbs.  This bug was originally harmless since the
driver does not claim to implement checksum offload and that used to
be a requirement for SG.

Now that SG and checksum offload are independent features, if you
explicitly enable SG *and* use one of the rare protocols that can use
SG without checkusm offload, this potentially leaks sensitive
information (before you notice that it just isn't working).  Therefore
this obscure bug has been designated CVE-2016-2117.
Reported-by: NJustin Yackoski <jyackoski@crypto-nite.com>
Signed-off-by: NBen Hutchings <ben@decadent.org.uk>
Fixes: ec5f0615 ("net: Kill link between CSUM and SG features.")
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch adds support for NETIF_F_TSO_MANGLEID if a given tunnel supports
NETIF_F_TSO.  This way if needed a device can then later enable the TSO
with IP ID mangling and the tunnels on top of that device can then also
make use of the IP ID mangling as well.
Signed-off-by: NAlexander Duyck <aduyck@mirantis.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Saeed Mahameed says:

====================
Mellanox 100G mlx5 driver receive path optimizations

Changes from V2:
	- Rebased to 46e7b8d8 ("net: dsa: kill circular reference with slave priv")
	- Updated: ("net/mlx5e: Support RX multi-packet WQE (Striding RQ)")
		* Per Eric Dumazet comment we changed the driver memory handling scheme to
		work with order-0 pages rather than order-5 via split_page().
		* This means that now a mlx5e rx skb can hold one or (more in case of HW LRO)
                skb frag each pointing to a 4K order-0 page rather than one frag with order-5 page.
	- Updated: ("net/mlx5e: Add fragmented memory support for RX multi packet WQE")
		* Code refactoring and code reuse due the split_page() mechanism,
		  now the MPWQE and fragmented MPWQE handling almost look the same,
		  and share most of the code.
	- In some cases we see 2%-3% packet rate degradation in comparison to the order-5 pages approach,
	  due to split_page() cpu consumption, but still we do see 3%-10% improvement in comparison to the
          current linear SKB approach.
	- We do believe that now the driver memory scheme is significantly less vulnerable
	  to the memory DOS attack Eric pointed at.

Changes from V1:
	- Rebased to efde611b ("Merge branch 'nfp-next'")
	- Dropped: ("net/mlx5: Refactor mlx5_core_mr to mkey")
                Already merged into 4.6 from rdma tree.
	- Dropped: ("net/mlx5_core: Add ConnectX-5 to list of supported devices")
                Will be pushed to net as we want it in 4.6 release.
	- Dropped: ("net/mlx5e: Change RX moderation period to be based on CQE")
                Will be pushed in a later series with full software based adaptive moderation.
	- Added: ("net/mlx5e: Delay skb->data access")
		Small trivial optimization.
	- Updated: ("net/mlx5e: Support RX multi-packet WQE (Striding RQ)")
	 	Changed Striding RQ defaults to:
			> 	NUM WQEs = 16
			> 	Strides Per WQE = 1024
			> 	Stride Size = 128
	- Updated: ("net/mlx5e: Use napi_alloc_skb for RX SKB allocations")
		Consider the IP packet alignment already done in napi_alloc_skb.

Changes from V0:
	- Fixed a typo in commit message reported by Sergei
	- Align SKB fragments truesize to stride size
	- Use skb_add_rx_frag and remove the use of SKB_TRUESIZE
	- Fix: # MTTs alignment on Power PC
	- Fix: Free original (unaligned) pointer of MTT array
	- Use dev_alloc_pages and dev_alloc_page
	- Extend the stats.buff_alloc_err counter
	- Reform the copying of packet header into skb linear data
	- Add compiler hints for conditional statements
	- Prefetch skd->data prior to copying packet header into it
	- Rework: mlx5e_complete_rx_fragmented_mpwqe
	- Handle SKB fragments before linear data
	- Dropped ("net/mlx5e: Prefetch next RX CQE") for now
	- Added a small patch that Adds ConnectX-5 devices to the list of supported devices
	- Rebased to 1cdba550 ("Merge git://git.kernel.org/pub/scm/linux/kernel/git/pablo/nf-next")

This series includes Some RX modifications and optimizations for
the mlx5 Ethernet driver.

From Rana, we have one patch that adds the support for Connectx-4
queue counters.

From Tariq, several patches that are centralized around improving
RX path message rate, CPU and Memory utilization, in each patch
commit message you will find the performance improvements numbers
related to that specific patch.

In the 2nd patch we used a queue counter to report "out of buffer"
dropped packet count, "Dropped packets due to lack of software resources"

3rd patch modifies the driver's to RSS default value to be spread along the
close NUMA node cores only for better out of the box experience.

In the 4th and 5th patches we utilized the use of RX multi-packet WQE
(Striding RQ) for better memory utilization especially in case of hardware
LRO is enabled and for better message rate for small packets.

In the 6th and 7th patches we added a fallback mechanism to use fragmented
memory when allocating large WQE strides fails, using UMR
(User Memory Registration) and ICO (Internal Control Operations) SQs.

In the 8th to 11th patches we did some small modification which show some small
extra improvements.
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Counts the number of RX buffer allocation failures and shows it
in ethtool statistics.
Signed-off-by: NTariq Toukan <tariqt@mellanox.com>
Signed-off-by: NSaeed Mahameed <saeedm@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Move mlx5e_handle_csum and eth_type_trans to the end of
mlx5e_build_rx_skb to gain some more time before accessing
skb->data, to reduce cache misses.
Signed-off-by: NSaeed Mahameed <saeedm@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The bit-op operation one line before is an explicit barrier
by itself.
Signed-off-by: NTariq Toukan <tariqt@mellanox.com>
Signed-off-by: NSaeed Mahameed <saeedm@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Instead of netdev_alloc_skb, we use the napi_alloc_skb function
which is designated to allocate skbuff's for RX in a
channel-specific NAPI instance, and implies the IP packet alignment.
Signed-off-by: NTariq Toukan <tariqt@mellanox.com>
Signed-off-by: NSaeed Mahameed <saeedm@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

If the allocation of a linear (physically continuous) MPWQE fails,
we allocate a fragmented MPWQE.

This is implemented via device's UMR (User Memory Registration)
which allows to register multiple memory fragments into ConnectX
hardware as a continuous buffer.
UMR registration is an asynchronous operation and is done via
ICO SQs.
Signed-off-by: NTariq Toukan <tariqt@mellanox.com>
Signed-off-by: NSaeed Mahameed <saeedm@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>