Drop the code for handling macvlan specific unicast lists. It isn't needed
since we don't take any efforts to maintain it when we bring the interface
up and it takes the slow path anyway.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

Merge branch 'userns-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/ebiederm/user-namespace

Pull userns bug fix from Eric Biederman:
 "Just a small fix to properly set the return code on error"

* 'userns-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/ebiederm/user-namespace:
  commoncap: Handle memory allocation failure.

Pull networking fixes from David Miller:

 1) Fix rtnl deadlock in ipvs, from Julian Anastasov.

 2) s390 qeth fixes from Julian Wiedmann (control IO completion stalls,
    bad MAC address update sequence, request side races on command IO
    timeouts).

 3) Handle seq_file overflow properly in l2tp, from Guillaume Nault.

 4) Fix VLAN priority mappings in cpsw driver, from Ivan Khoronzhuk.

 5) Packet scheduler ife action fixes (malformed TLV lengths, etc.) from
    Alexander Aring.

 6) Fix out of bounds access in tcp md5 option parser, from Jann Horn.

 7) Missing netlink attribute policies in rtm_ipv6_policy table, from
    Eric Dumazet.

 8) Missing socket address length checks in l2tp and pppoe connect, from
    Guillaume Nault.

 9) Fix netconsole over team and bonding, from Xin Long.

10) Fix race with AF_PACKET socket state bitfields, from Willem de
    Bruijn.

* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net: (51 commits)
  ice: Fix insufficient memory issue in ice_aq_manage_mac_read
  sfc: ARFS filter IDs
  net: ethtool: Add missing kernel doc for FEC parameters
  packet: fix bitfield update race
  ice: Do not check INTEVENT bit for OICR interrupts
  ice: Fix incorrect comment for action type
  ice: Fix initialization for num_nodes_added
  igb: Fix the transmission mode of queue 0 for Qav mode
  ixgbevf: ensure xdp_ring resources are free'd on error exit
  team: fix netconsole setup over team
  amd-xgbe: Only use the SFP supported transceiver signals
  amd-xgbe: Improve KR auto-negotiation and training
  amd-xgbe: Add pre/post auto-negotiation phy hooks
  pppoe: check sockaddr length in pppoe_connect()
  l2tp: check sockaddr length in pppol2tp_connect()
  net: phy: marvell: clear wol event before setting it
  ipv6: add RTA_TABLE and RTA_PREFSRC to rtm_ipv6_policy
  bonding: do not set slave_dev npinfo before slave_enable_netpoll in bond_enslave
  tcp: don't read out-of-bounds opsize
  ibmvnic: Clean actual number of RX or TX pools
  ...

Swap VF representor tx and rx interface statistics since it is a
virtual switchdev port and tx for VM should be rx for VF representor
and vice-versa.
Signed-off-by: Srinivas Jampala <srinivasa.jampala@cavium.com>
Acked-by: Derek Chickles <derek.chickles@cavium.com>
Signed-off-by: Felix Manlunas <felix.manlunas@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

rt6_remove_exception_rt() is called under rcu_read_lock() only.

We lock rt6_exception_lock a bit later, so we do not hold
rt6_exception_lock yet.

Fixes: 8a14e46f ("net/ipv6: Fix missing rcu dereferences on from")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: syzbot <syzkaller@googlegroups.com>
Cc: David Ahern <dsahern@gmail.com>
Acked-by: David Ahern <dsahern@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Jeff Kirsher says:

====================
Intel Wired LAN Driver Updates 2018-04-24

This series contains fixes to ixgbevf, igb and ice drivers.

Colin Ian King fixes the return value on error for the new XDP support
that went into ixgbevf for 4.17.

Vinicius provides a fix for queue 0 for igb, which was not receiving all
the credits it needed when QAV mode was enabled.

Anirudh provides several fixes for the new ice driver, starting with
properly initializing num_nodes_added to zero.  Fixed up a code comment
to better reflect what is really going on in the code.  Fixed how to
detect if an OICR interrupt has occurred to a more reliable method.

Md Fahad fixes the ice driver to allocate the right amount of memory
when reading and storing the devices MAC addresses.  The device can have
up to 2 MAC addresses (LAN and WoL), while WoL is currently not
supported, we need to ensure it can be properly handled when support is
added.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

A duplicated null check on sgout is redundant as it is known to be
already true because of the identical earlier check. Remove it.
Detected by cppcheck:

net/tls/tls_sw.c:696: (warning) Identical inner 'if' condition is always
true.
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The check port->rev_info.major >= 6 is being performed twice, thus
the inner second check is always true and is redundant, hence it
can be removed. Detected by cppcheck.

drivers/net/ethernet/freescale/fman/fman_port.c:1394]: (warning)
Identical inner 'if' condition is always true.
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

For the MAC read operation, the device can return up to two (LAN and WoL)
MAC addresses. Without access to adequate memory, the device will return
an error. Fixed this by allocating the right amount of memory. Also, logic
to detect and copy the LAN MAC address into the port_info structure has
been added. Note that the WoL MAC address is ignored currently as the WoL
feature isn't supported yet.

Fixes: dc49c772 ("ice: Get MAC/PHY/link info and scheduler topology")
Signed-off-by: Md Fahad Iqbal Polash <md.fahad.iqbal.polash@intel.com>
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Tony Brelinski <tonyx.brelinski@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

The strscpy() was a recent fix (net: qed: use correct strncpy() size) to
prevent passing the length of the source buffer to strncpy() and guarantee
null termination.
It misses the goal of overwriting only the first 3 characters in
"???_BIG_RAM" and "???_RAM" while keeping the rest of the string.
Use strncpy() with the length of 3, without null termination.
Signed-off-by: Denis Bolotin <denis.bolotin@cavium.com>
Signed-off-by: Ariel Elior <ariel.elior@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Associate an arbitrary ID with each ARFS filter, allowing to properly query
 for expiry.  The association is maintained in a hash table, which is
 protected by a spinlock.

v3: fix build warnings when CONFIG_RFS_ACCEL is disabled (thanks lkp-robot).
v2: fixed uninitialised variable (thanks davem and lkp-robot).

Fixes: 3af0f342 ("sfc: replace asynchronous filter operations")
Signed-off-by: Edward Cree <ecree@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Chris Novakovic says:

====================
ipconfig: NTP server support, bug fixes, documentation improvements

This series (against net-next) makes various improvements to ipconfig:

 - Patch #1 correctly documents the behaviour of parameter 4 in the
   "ip=" and "nfsaddrs=" command line parameter.
 - Patch #2 tidies up the printk()s for reporting configured name
   servers.
 - Patch #3 fixes a bug in autoconfiguration via BOOTP whereby the IP
   addresses of IEN-116 name servers are requested from the BOOTP
   server, rather than those of DNS name servers.
 - Patch #4 requests the number of DNS servers specified by
   CONF_NAMESERVERS_MAX when autoconfiguring via BOOTP, rather than
   hardcoding it to 2.
 - Patch #5 fully documents the contents and format of /proc/net/pnp in
   Documentation/filesystems/nfs/nfsroot.txt.
 - Patch #6 fixes a bug whereby bogus information is written to
   /proc/net/pnp when ipconfig is not used.
 - Patch #7 creates a new procfs directory for ipconfig-related
   configuration reports at /proc/net/ipconfig.
 - Patch #8 allows for NTP servers to be configured (manually on the
   kernel command line or automatically via DHCP), enabling systems with
   an NFS root filesystem to synchronise their clock before mounting
   their root filesystem. NTP server IP addresses are written to
   /proc/net/ipconfig/ntp_servers.

Changes from v1:

 - David requested that a new directory /proc/net/ipconfig be created to
   contain ipconfig-related configuration reports, which is implemented
   in the new patch #7. NTP server IPs are now written to this directory
   instead of /proc/net/ntp in the new patch #8.
 - Cong and David both requested that the modification to CREDITS be
   dropped. This patch has been removed from the series.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

Distributed filesystems are most effective when the server and client
clocks are synchronised. Embedded devices often use NFS for their
root filesystem but typically do not contain an RTC, so the clocks of
the NFS server and the embedded device will be out-of-sync when the root
filesystem is mounted (and may not be synchronised until late in the
boot process).

Extend ipconfig with the ability to export IP addresses of NTP servers
it discovers to /proc/net/ipconfig/ntp_servers. They can be supplied as
follows:

 - If ipconfig is configured manually via the "ip=" or "nfsaddrs="
   kernel command line parameters, one NTP server can be specified in
   the new "<ntp0-ip>" parameter.
 - If ipconfig is autoconfigured via DHCP, request DHCP option 42 in
   the DHCPDISCOVER message, and record the IP addresses of up to three
   NTP servers sent by the responding DHCP server in the subsequent
   DHCPOFFER message.

ipconfig will only write the NTP server IP addresses it discovers to
/proc/net/ipconfig/ntp_servers, one per line (in the order received from
the DHCP server, if DHCP autoconfiguration is used); making use of these
NTP servers is the responsibility of a user space process (e.g. an
initrd/initram script that invokes an NTP client before mounting an NFS
root filesystem).
Signed-off-by: Chris Novakovic <chris@chrisn.me.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>

To allow ipconfig to report IP configuration details to user space
processes without cluttering /proc/net, create a new subdirectory
/proc/net/ipconfig. All files containing IP configuration details should
be written to this directory.
Signed-off-by: Chris Novakovic <chris@chrisn.me.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>

ic_nameservers, which stores the list of name servers discovered by
ipconfig, is initialised (i.e. has all of its elements set to NONE, or
0xffffffff) by ic_nameservers_predef() in the following scenarios:

 - before the "ip=" and "nfsaddrs=" kernel command line parameters are
   parsed (in ip_auto_config_setup());
 - before autoconfiguring via DHCP or BOOTP (in ic_bootp_init()), in
   order to clear any values that may have been set after parsing "ip="
   or "nfsaddrs=" and are no longer needed.

This means that ic_nameservers_predef() is not called when neither "ip="
nor "nfsaddrs=" is specified on the kernel command line. In this
scenario, every element in ic_nameservers remains set to 0x00000000,
which is indistinguishable from ANY and causes pnp_seq_show() to write
the following (bogus) information to /proc/net/pnp:

  #MANUAL
  nameserver 0.0.0.0
  nameserver 0.0.0.0
  nameserver 0.0.0.0

This is potentially problematic for systems that blindly link
/etc/resolv.conf to /proc/net/pnp.

Ensure that ic_nameservers is also initialised when neither "ip=" nor
"nfsaddrs=" are specified by calling ic_nameservers_predef() in
ip_auto_config(), but only when ip_auto_config_setup() was not called
earlier. This causes the following to be written to /proc/net/pnp, and
is consistent with what gets written when ipconfig is configured
manually but no name servers are specified on the kernel command line:

  #MANUAL
Signed-off-by: Chris Novakovic <chris@chrisn.me.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>

Fully document the format used by the /proc/net/pnp file written by
ipconfig, explain where its values originate from, and clarify that the
tertiary name server IP and DNS domain name are only written to the file
when autoconfiguration is used.
Signed-off-by: Chris Novakovic <chris@chrisn.me.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>

When ipconfig is autoconfigured via BOOTP, the request packet
initialised by ic_bootp_init_ext() always allocates 8 bytes for the name
server option, limiting the BOOTP server to responding with at most 2
name servers even though ipconfig in fact supports an arbitrary number
of name servers (as defined by CONF_NAMESERVERS_MAX, which is currently
3).

Only request name servers in the request packet if CONF_NAMESERVERS_MAX
is positive (to comply with [1, §3.8]), and allocate enough space in the
packet for CONF_NAMESERVERS_MAX name servers to indicate the maximum
number we can accept in response.

[1] RFC 2132, "DHCP Options and BOOTP Vendor Extensions":
    https://tools.ietf.org/rfc/rfc2132.txtSigned-off-by: Chris Novakovic <chris@chrisn.me.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>

When ipconfig is autoconfigured via BOOTP, the request packet
initialised by ic_bootp_init_ext() allocates 8 bytes for tag 5 ("Name
Server" [1, §3.7]), but tag 5 in the response isn't processed by
ic_do_bootp_ext(). Instead, allocate the 8 bytes to tag 6 ("Domain Name
Server" [1, §3.8]), which is processed by ic_do_bootp_ext(), and appears
to have been the intended tag to request.

This won't cause any breakage for existing users, as tag 5 responses
provided by BOOTP servers weren't being processed anyway.

[1] RFC 2132, "DHCP Options and BOOTP Vendor Extensions":
    https://tools.ietf.org/rfc/rfc2132.txtSigned-off-by: Chris Novakovic <chris@chrisn.me.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>

Commit 5e953778 ("ipconfig: add
nameserver IPs to kernel-parameter ip=") adds the IP addresses of
discovered name servers to the summary printed by ipconfig when
configuration is complete. It appears the intention in ip_auto_config()
was to print the name servers on a new line (especially given the
spacing and lack of comma before "nameserver0="), but they're actually
printed on the same line as the NFS root filesystem configuration
summary:

  [    0.686186] IP-Config: Complete:
  [    0.686226]      device=eth0, hwaddr=xx:xx:xx:xx:xx:xx, ipaddr=10.0.0.2, mask=255.255.255.0, gw=10.0.0.1
  [    0.686328]      host=test, domain=example.com, nis-domain=(none)
  [    0.686386]      bootserver=10.0.0.1, rootserver=10.0.0.1, rootpath=     nameserver0=10.0.0.1

This makes it harder to read and parse ipconfig's output. Instead, print
the name servers on a separate line:

  [    0.791250] IP-Config: Complete:
  [    0.791289]      device=eth0, hwaddr=xx:xx:xx:xx:xx:xx, ipaddr=10.0.0.2, mask=255.255.255.0, gw=10.0.0.1
  [    0.791407]      host=test, domain=example.com, nis-domain=(none)
  [    0.791475]      bootserver=10.0.0.1, rootserver=10.0.0.1, rootpath=
  [    0.791476]      nameserver0=10.0.0.1
Signed-off-by: Chris Novakovic <chris@chrisn.me.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>

ic_do_bootp_ext() is responsible for parsing the "ip=" and "nfsaddrs="
kernel parameters. If a "." character is found in parameter 4 (the
client's hostname), everything before the first "." is used as the
hostname, and everything after it is used as the NIS domain name (but
not necessarily the DNS domain name).

Document this behaviour in Documentation/filesystems/nfs/nfsroot.txt,
as it is not made explicit.
Signed-off-by: Chris Novakovic <chris@chrisn.me.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>

While adding support for ethtool::get_fecparam and set_fecparam, kernel
doc for these functions was missed, add those.

Fixes: 1a5f3da2 ("net: ethtool: add support for forward error correction modes")
Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Acked-by: Roopa Prabhu <roopa@cumulusnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

NeilBrown says:

====================
A few rhashtables cleanups

2 patches fixes documentation
1 fixes a bit in rhashtable_walk_start()
1 improves rhashtable_walk stability.

All reviewed and Acked.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

When a walk of an rhashtable is interrupted with rhastable_walk_stop()
and then rhashtable_walk_start(), the location to restart from is based
on a 'skip' count in the current hash chain, and this can be incorrect
if insertions or deletions have happened.  This does not happen when
the walk is not stopped and started as iter->p is a placeholder which
is safe to use while holding the RCU read lock.

In rhashtable_walk_start() we can revalidate that 'p' is still in the
same hash chain.  If it isn't then the current method is still used.

With this patch, if a rhashtable walker ensures that the current
object remains in the table over a stop/start period (possibly by
elevating the reference count if that is sufficient), it can be sure
that a walk will not miss objects that were in the hashtable for the
whole time of the walk.

rhashtable_walk_start() may not find the object even though it is
still in the hashtable if a rehash has moved it to a new table.  In
this case it will (eventually) get -EAGAIN and will need to proceed
through the whole table again to be sure to see everything at least
once.
Acked-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The documentation claims that when rhashtable_walk_start_check()
detects a resize event, it will rewind back to the beginning
of the table.  This is not true.  We need to set ->slot and
->skip to be zero for it to be true.
Acked-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Neither rhashtable_walk_enter() or rhltable_walk_enter() sleep, though
they do take a spinlock without irq protection.
So revise the comments to accurately state the contexts in which
these functions can be called.
Acked-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

grow_decision and shink_decision no longer exist, so remove
the remaining references to them.
Acked-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

RETPOLINE made calls to tp->af_specific->md5_lookup() quite expensive,
given they have no result.
We can omit the calls for sockets that have no md5 keys.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Updates to the bitfields in struct packet_sock are not atomic.
Serialize these read-modify-write cycles.

Move po->running into a separate variable. Its writes are protected by
po->bind_lock (except for one startup case at packet_create). Also
replace a textual precondition warning with lockdep annotation.

All others are set only in packet_setsockopt. Serialize these
updates by holding the socket lock. Analogous to other field updates,
also hold the lock when testing whether a ring is active (pg_vec).

Fixes: 8dc41944 ("[PACKET]: Add optional checksum computation for recvmsg")
Reported-by: DaeRyong Jeong <threeearcat@gmail.com>
Reported-by: Byoungyoung Lee <byoungyoung@purdue.edu>
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

According to the hardware spec, checking the INTEVENT bit isn't a
reliable way to detect if an OICR interrupt has occurred. This is
because this bit can be cleared by the hardware/firmware before the
interrupt service routine has run. So instead, just check for OICR
events every time.

Fixes: 940b61af ("ice: Initialize PF and setup miscellaneous interrupt")
Signed-off-by: Ben Shelton <benjamin.h.shelton@intel.com>
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Tony Brelinski <tonyx.brelinski@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

Action type 5 defines large action generic values. Fix comment to
reflect that better.
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Tony Brelinski <tonyx.brelinski@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

ice_sched_add_nodes_to_layer is used recursively, and so we start
with num_nodes_added being 0. This way, in case of an error or if
num_nodes is NULL, the function just returns 0 to indicate that no
nodes were added.

Fixes: 5513b920 ("ice: Update Tx scheduler tree for VSI multi-Tx queue support")
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Tony Brelinski <tonyx.brelinski@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

When Qav mode is enabled, queue 0 should be kept on Stream Reservation
mode. From the i210 datasheet, section 8.12.19:

"Note: Queue0 QueueMode must be set to 1b when TransmitMode is set to
Qav." ("QueueMode 1b" represents the Stream Reservation mode)

The solution is to give queue 0 the all the credits it might need, so
it has priority over queue 1.

A situation where this can happen is when cbs is "installed" only on
queue 1, leaving queue 0 alone. For example:

$ tc qdisc replace dev enp2s0 handle 100: parent root mqprio num_tc 3 \
     	   map 2 2 1 0 2 2 2 2 2 2 2 2 2 2 2 2 queues 1@0 1@1 2@2 hw 0

$ tc qdisc replace dev enp2s0 parent 100:2 cbs locredit -1470 \
     	   hicredit 30 sendslope -980000 idleslope 20000 offload 1
Signed-off-by: Vinicius Costa Gomes <vinicius.gomes@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

The current error handling for failed resource setup for xdp_ring
data is a break out of the loop and returning 0 indicated everything
was OK, when in fact it is not.  Fix this by exiting via the
error exit label err_setup_tx that will clean up the resources
correctly and return and error status.

Detected by CoverityScan, CID#1466879 ("Logically dead code")

Fixes: 21092e9c ("ixgbevf: Add support for XDP_TX action")
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

This reverts commit <c6849a3a> ("net: init sk_cookie for inet socket")

Per discussion with Eric, when update sock_net(sk)->cookie_gen, the
whole cache cache line will be invalidated, as this cache line is shared
with all cpus, that may cause great performace hit.

Bellow is the data form Eric.
"Performance is reduced from ~5 Mpps to ~3.8 Mpps with 16 RX queues on
my host" when running synflood test.

Have to revert it to prevent from cache line false sharing.
Signed-off-by: Yafang Shao <laoar.shao@gmail.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Tal Gilboa says:

====================
Introduce adaptive TX interrupt moderation to net DIM

Net DIM is a library designed for dynamic interrupt moderation. It was
implemented and optimized with receive side interrupts in mind, since these
are usually the CPU expensive ones. This patch-set introduces adaptive transmit
interrupt moderation to net DIM, complete with a usage in the mlx5e driver.
Using adaptive TX behavior would reduce interrupt rate for multiple scenarios.
Furthermore, it is essential for increasing bandwidth on cases where payload
aggregation is required.

v3: Remove "inline" from functions in .c files (requested by DaveM). Revert
adding "enabled" field from struct net_dim and applied mlx5e structural
suggestions (suggested by SaeedM).

v2: Rebase over proper tree.

v1: Fix compilation issues due to missed function renaming.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

Add support for adaptive TX moderation. This greatly reduces TX interrupt
rate and increases bandwidth, mostly for TCP bandwidth over ARM
architecture (below). There is a slight single stream TCP with very large
message sizes degradation (x86). In this case if there's any moderation on
transmitted packets the bandwidth would reduce due to hitting TCP output limit.
Since this is a synthetic case, this is still worth doing.

Performance improvement (ConnectX-4Lx 40GbE, ARM)
TCP 64B bandwidth with 1-50 streams increased 6-35%.
TCP 64B bandwidth with 100-500 streams increased 20-70%.

Performance improvement (ConnectX-5 100GbE, x86)
Bandwidth: increased up to 40% (1024B with 10s of streams).
Interrupt rate: reduced up to 50% (1024B with 1000s of streams).

Performance degradation (ConnectX-5 100GbE, x86)
Bandwidth: up to 10% decrease single stream TCP (1MB message size from
51Gb/s to 47Gb/s).
Signed-off-by: Tal Gilboa <talgi@mellanox.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Interrupt moderation for TX traffic requires different profiles than RX
interrupt moderation. The main goal here is to reduce interrupt rate and
allow better payload aggregation by keeping SKBs in the TX queue a bit
longer. Ping-pong behavior would get a profile with a short timer, so
latency wouldn't increase for these scenarios. There might be a slight
degradation in bandwidth for single stream with large message sizes, since
net.ipv4.tcp_limit_output_bytes is limiting the allowed TX traffic, but
with many streams performance is always improved.
Signed-off-by: Tal Gilboa <talgi@mellanox.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Preparation for introducing adaptive TX to net DIM.
Signed-off-by: Tal Gilboa <talgi@mellanox.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Similar to commit a2ac9990 ("vhost-net: set packet weight of
tx polling to 2 * vq size"), we need a packet-based limit for
handler_rx, too - elsewhere, under rx flood with small packets,
tx can be delayed for a very long time, even without busypolling.

The pkt limit applied to handle_rx must be the same applied by
handle_tx, or we will get unfair scheduling between rx and tx.
Tying such limit to the queue length makes it less effective for
large queue length values and can introduce large process
scheduler latencies, so a constant valued is used - likewise
the existing bytes limit.

The selected limit has been validated with PVP[1] performance
test with different queue sizes:

queue size		256	512	1024

baseline		366	354	362
weight 128		715	723	670
weight 256		740	745	733
weight 512		600	460	583
weight 1024		423	427	418

A packet weight of 256 gives peek performances in under all the
tested scenarios.

No measurable regression in unidirectional performance tests has
been detected.

[1] https://developers.redhat.com/blog/2017/06/05/measuring-and-comparing-open-vswitch-performance/Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Acked-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>