Take uld mutex to avoid race between cxgb_up() and
cxgb4_register_uld() to enable napi for the same uld
queue.
Signed-off-by: NGanesh Goudar <ganeshgr@chelsio.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

xfrm6_find_1stfragopt() may now return an error code and we must
not treat it as a length.

Fixes: 2423496a ("ipv6: Prevent overrun when parsing v6 header options")
Signed-off-by: NBen Hutchings <ben@decadent.org.uk>
Acked-by: NCraig Gallek <kraig@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Merge tag 'mac80211-for-davem-2017-06-02' of git://git.kernel.org/pub/scm/linux/kernel/git/jberg/mac80211

Johannes Berg says:

====================
Just two fixes:
 * fix the per-CPU drop counters to not be added to the
   rx_packets counter, but really the drop counter
 * fix TX aggregation start/stop callback races by setting
   bits instead of allocating and queueing an skb
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

dsa_switch_suspend() and dsa_switch_resume() are functions that belong in
net/dsa/dsa.c and are not part of the legacy platform support code.

Fixes: a6a71f19 ("net: dsa: isolate legacy code")
Signed-off-by: NFlorian Fainelli <f.fainelli@gmail.com>
Reviewed-by: NAndrew Lunn <andrew@lunn.ch>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

On SYSTEMPORT Lite, since we have the main interrupt source in the first
cell, the second cell is the Wake-on-LAN interrupt, yet the code was not
properly updated to fetch the second cell, and instead looked at the
third and non-existing cell for Wake-on-LAN.

Fixes: 44a4524c ("net: systemport: Add support for SYSTEMPORT Lite")
Signed-off-by: NFlorian Fainelli <f.fainelli@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

In the commit enabling per-CPU station statistics, I inadvertedly
copy-pasted some code to update rx_packets and forgot to change it
to update rx_dropped_misc. Fix that.

This addresses https://bugzilla.kernel.org/show_bug.cgi?id=195953.

Fixes: c9c5962b ("mac80211: enable collecting station statistics per-CPU")
Reported-by: NPetru-Florin Mihancea <petrum@gmail.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Leonard Crestez says:

====================
ARM: imx6ul-14x14-evk: Fix suspend over nfs by phy

Right now attempting doing suspend/resume while root is mounted over NFS
hangs on imx6ul-14x14-evk. This is happening because ksz8081 phy fixups are
lost on resume.

Fix this by using equivalent devicetree properties instead of a phy fixup
and handling those properties on resume in the micrel driver.

In theory it might now be possible to remove the phy fixup from mach-imx6ul
entirely but it is possible that this would break other imx6ul boards which
use the same phy. The solution would be to patch their dts but it's not
clear how to identify affected boards.

This code is shared with imx6ull-14x14-evk but 6ull suspend needs an
unrelated patch: https://lkml.org/lkml/2017/5/30/584

This is something of a corner case so there is no CC: stable.

Changes since v1: https://lkml.org/lkml/2017/5/30/672
 * Split a kszphy_config_reset function for stuff shared between
config_init and resume. Calling config_init directly could be an option but
on some HW variants it does extra stuff like parsing devicetree options.
That would not be appropriate for resume code.
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

These bits seem to be lost after a suspend/resume cycle so just set them
again. Do this by splitting the handling of these bits into a function
that is also called on resume.

This patch fixes ethernet suspend/resume on imx6ul-14x14-evk boards.
Signed-off-by: NLeonard Crestez <leonard.crestez@nxp.com>
Reviewed-by: NFlorian Fainelli <f.fainelli@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Right now mach-imx6ul registers a fixup for the ksz8081 phy. The same
register values can be set through the micrel phy driver by using dts
properties.

This seems preferable and allows cleanly fixing suspend/resume.
Signed-off-by: NLeonard Crestez <leonard.crestez@nxp.com>
Reviewed-by: NFabio Estevam <fabio.estevam@nxp.com>
Reviewed-by: NFlorian Fainelli <f.fainelli@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The driver may sleep under a read spin lock, and the function call path is:
send_socklist (acquire the lock by read_lock)
  skb_copy(GFP_KERNEL) --> may sleep

To fix it, the "GFP_KERNEL" is replaced with "GFP_ATOMIC".
Signed-off-by: NJia-Ju Bai <baijiaju1990@163.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

After commit 0c1d70af ("net: use dst_cache for vxlan device"),
cached dst entries could be leaked when more than one remote was
present for a given vxlan_fdb entry, causing subsequent netns
operations to block indefinitely and "unregister_netdevice: waiting
for lo to become free." messages to appear in the kernel log.

Fix by properly releasing cached dst and freeing resources in this
case.

Fixes: 0c1d70af ("net: use dst_cache for vxlan device")
Signed-off-by: NLance Richardson <lrichard@redhat.com>
Acked-by: NPaolo Abeni <pabeni@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When the transition of NO_STP -> KERNEL_STP was fixed by always calling
mod_timer in br_stp_start, it introduced a new regression which causes
the timer to be armed even when the bridge is down, and since we stop
the timers in its ndo_stop() function, they never get disabled if the
device is destroyed before it's upped.

To reproduce:
$ while :; do ip l add br0 type bridge hello_time 100; brctl stp br0 on;
ip l del br0; done;

CC: Xin Long <lucien.xin@gmail.com>
CC: Ivan Vecera <cera@cera.cz>
CC: Sebastian Ott <sebott@linux.vnet.ibm.com>
Reported-by: NSebastian Ott <sebott@linux.vnet.ibm.com>
Fixes: 6d18c732 ("bridge: start hello_timer when enabling KERNEL_STP in br_stp_start")
Signed-off-by: NNikolay Aleksandrov <nikolay@cumulusnetworks.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Apparently multi-cos isn't working for bnx2x quite some time -
driver implements ndo_select_queue() to allow queue-selection
for FCoE, but the regular L2 flow would cause it to modulo the
fallback's result by the number of queues.
The fallback would return a queue matching the needed tc
[via __skb_tx_hash()], but since the modulo is by the number of TSS
queues where number of TCs is not accounted, transmission would always
be done by a queue configured into using TC0.

Fixes: ada7c19e ("bnx2x: use XPS if possible for bnx2x_select_queue instead of pure hash")
Signed-off-by: NYuval Mintz <Yuval.Mintz@cavium.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Change t4fw_version.h to update latest firmware version
number to 1.16.45.0.
Signed-off-by: NGanesh Goudar <ganeshgr@chelsio.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The NETLINK_F_LISTEN_ALL_NSID otion enables to listen all netns that have a
nsid assigned into the netns where the netlink socket is opened.
The nsid is sent as metadata to userland, but the existence of this nsid is
checked only for netns that are different from the socket netns. Thus, if
no nsid is assigned to the socket netns, NETNSA_NSID_NOT_ASSIGNED is
reported to the userland. This value is confusing and useless.
After this patch, only valid nsid are sent to userland.
Reported-by: NFlavio Leitner <fbl@sysclose.org>
Signed-off-by: NNicolas Dichtel <nicolas.dichtel@6wind.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The driver may sleep under a write spin lock, and the function
call path is:
qlcnic_82xx_hw_write_wx_2M (acquire the lock by write_lock_irqsave)
  crb_win_lock
    qlcnic_pcie_sem_lock
      usleep_range
qlcnic_82xx_hw_read_wx_2M (acquire the lock by write_lock_irqsave)
  crb_win_lock
    qlcnic_pcie_sem_lock
      usleep_range

To fix it, the usleep_range is replaced with udelay.
Signed-off-by: NJia-Ju Bai <baijiaju1990@163.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

recent fixes to use WRITE_ONCE for nh_flags on link up,
accidently ended up leaving the deadflags on a nh. This patch
fixes the WRITE_ONCE to use freshly evaluated nh_flags.

Fixes: 39eb8cd1 ("net: mpls: rt_nhn_alive and nh_flags should be accessed using READ_ONCE")
Reported-by: NSatish Ashok <sashok@cumulusnetworks.com>
Signed-off-by: NRoopa Prabhu <roopa@cumulusnetworks.com>
Acked-by: NDavid Ahern <dsahern@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The driver may sleep under a spin lock, the function call path is:
isdn_ppp_mp_receive (acquire the lock)
  isdn_ppp_mp_reassembly
    isdn_ppp_push_higher
      isdn_ppp_decompress
        isdn_ppp_ccp_reset_trans
          isdn_ppp_ccp_reset_alloc_state
            kzalloc(GFP_KERNEL) --> may sleep

To fixed it, the "GFP_KERNEL" is replaced with "GFP_ATOMIC".
Signed-off-by: NJia-Ju Bai <baijiaju1990@163.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Add NULL check before dereferencing pointer _id_ in order to avoid
a potential NULL pointer dereference.

Addresses-Coverity-ID: 1397995
Signed-off-by: NGustavo A. R. Silva <garsilva@embeddedor.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Auto-loading of the Marvell DSA driver has stopped working with recent
kernels.  This seems to be due to the change of binding for DSA devices,
moving them from the platform bus to the MDIO bus.

In order for module auto-loading to work, we need to provide a MODALIAS
string in the uevent file for the device.  However, the device core does
not automatically provide this, and needs each bus_type to implement a
uevent method to generate these strings.  The MDIO bus does not provide
such a method, so no MODALIAS string is provided:

.# cat /sys/bus/mdio_bus/devices/f1072004.mdio-mii\:04/uevent
DRIVER=mv88e6085
OF_NAME=switch
OF_FULLNAME=/soc/internal-regs/mdio@72004/switch@4
OF_COMPATIBLE_0=marvell,mv88e6085
OF_COMPATIBLE_N=1

In the case of OF-based devices, the solution is easy -
of_device_uevent_modalias() does the work for us.  After this is done,
the uevent file looks like this:

.# cat /sys/bus/mdio_bus/devices/f1072004.mdio-mii\:04/uevent
DRIVER=mv88e6085
OF_NAME=switch
OF_FULLNAME=/soc/internal-regs/mdio@72004/switch@4
OF_COMPATIBLE_0=marvell,mv88e6085
OF_COMPATIBLE_N=1
MODALIAS=of:NswitchT<NULL>Cmarvell,mv88e6085

which results in auto-loading of the Marvell DSA driver on Clearfog
platforms.

Fixes: c0405563 ("ARM: dts: armada-388-clearfog: Utilize new DSA binding")
Signed-off-by: NRussell King <rmk+kernel@armlinux.org.uk>
Reviewed-by: NAndrew Lunn <andrew@lunn.ch>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The Marvell driver incorrectly provides phydev->lp_advertising as the
logical and of the link partner's advert and our advert.  This is
incorrect - this field is supposed to store the link parter's unmodified
advertisment.

This allows ethtool to report the correct link partner auto-negotiation
status.

Fixes: be937f1f ("Marvell PHY m88e1111 driver fix")
Signed-off-by: NRussell King <rmk+kernel@armlinux.org.uk>
Reviewed-by: NAndrew Lunn <andrew@lunn.ch>
Reviewed-by: NFlorian Fainelli <f.fainelli@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

MTU probing initialization occurred only at connect() and at SYN or
SYN-ACK reception, but the former sets MSS to either the default or the
user set value (through TCP_MAXSEG sockopt) and the latter never happens
with repaired sockets.

The result was that, with MTU probing enabled and unless TCP_MAXSEG
sockopt was used before connect(), probing would be stuck at
tcp_base_mss value until tcp_probe_interval seconds have passed.
Signed-off-by: NDouglas Caetano dos Santos <douglascs@taghos.com.br>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Tariq Toukan says:

====================
MAINTAINERS updates

This patchset contains updates to the MAINTAINERS file.
In the first patch, I replace Yishai as the maintainer of
the mlx4_core driver.

In the other two patches we move an RDMA header file from
the list of the mlx4/mlx5 core driver into the respective
IB driver, where it belongs.

Series generated against net commit:
468b0df6 Merge git://git.kernel.org/pub/scm/linux/kernel/git/pablo/nf
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

It belongs there, should not be under mlx5 Core driver.
Signed-off-by: NTariq Toukan <tariqt@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

It belongs there, should not be under mlx4 Core driver.
Signed-off-by: NTariq Toukan <tariqt@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Add myself as a maintainer for mlx4 core driver,
replacing Yishai Hadas.
Signed-off-by: NTariq Toukan <tariqt@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Building the driver with CONFIG_SMP disabled results in a harmless
warning:

ethernet/mellanox/mlx5/core/main.c: In function 'mlx5_irq_set_affinity_hint':
ethernet/mellanox/mlx5/core/main.c:615:6: error: unused variable 'irq' [-Werror=unused-variable]

It's better to express the conditional compilation using IS_ENABLED()
here, as that lets the compiler see what the intented use for the variable
is, and that it can be silently discarded.

Fixes: b665d98e ("net/mlx5: Tolerate irq_set_affinity_hint() failures")
Signed-off-by: NArnd Bergmann <arnd@arndb.de>
Acked-by: NSaeed Mahameed <saeedm@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

'static' was not enough, the helpers must be 'static inline'

net/dsa/mv88e6xxx/global2.h:123:12: error: 'mv88e6xxx_g2_misc_4_bit_port' defined but not used [-Werror=unused-function]
net/dsa/mv88e6xxx/global2.h:117:12: error: 'mv88e6xxx_g2_pvt_write' defined but not used [-Werror=unused-function]

Fixes: c21fbe29 ("net: dsa: mv88e6xxx: Add missing static to stub functions")
Signed-off-by: NArnd Bergmann <arnd@arndb.de>
Reviewed-by: NVivien Didelot <vivien.didelot@savoirfairelinux.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Current implementation lacks the logic for providing management
firmware with RDMA-related statistics; [much] worse than that -
it logs such events by default to system logs.

Since the statistics' gathering is done periodically, using sufficiently
new management firmware the system logs would get filled with these
unnecessary prints.

For now, reduce the verbosity of the log so that it would not be
logged by default.

Fixes: 6c754246 ("qed: Add support for NCSI statistics")
Signed-off-by: NYuval Mintz <Yuval.Mintz@cavium.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NMadalin Bucur <madalin.bucur@nxp.com>
Signed-off-by: NCamelia Groza <camelia.groza@nxp.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

During PCI error recovery process, specifically on eeh_err_detected()
we might have a NULL netdev struct, hence a direct dereference will
lead to a kernel oops. This was observed with latest upstream kernel
(v4.12-rc2) on Chelsio adapter T422-CR in PowerPC machines.

This patch checks for NULL pointer and avoids the crash, both in
eeh_err_detected() and eeh_resume(). Also, we avoid to trigger
a fatal error or to try disabling interrupts on FW during PCI
error recovery, because: (a) driver might not be able to accurately
access PCI regions in this case, and (b) trigger a fatal error
_during_ the recovery steps is a mistake that could prevent the
recovery path to complete successfully.
Reported-by: NHarsha Thyagaraja <hathyaga@in.ibm.com>
Signed-off-by: NGuilherme G. Piccoli <gpiccoli@linux.vnet.ibm.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When starting or stopping an aggregation session, one of the steps
is that the driver calls back to mac80211 that the start/stop can
proceed. This is handled by queueing up a fake SKB and processing
it from the normal iface/sdata work. Since this isn't flushed when
disassociating, the following race is possible:

 * associate
 * start aggregation session
 * driver callback
 * disassociate
 * associate again to the same AP
 * callback processing runs, leading to a WARN_ON() that
   the TID hadn't requested aggregation

If the second association isn't to the same AP, there would only
be a message printed ("Could not find station: <addr>"), but the
same race could happen.

Fix this by not going the whole detour with a fake SKB etc. but
simply looking up the aggregation session in the driver callback,
marking it with a START_CB/STOP_CB bit and then scheduling the
regular aggregation work that will now process these bits as well.
This also simplifies the code and gets rid of the whole problem
with allocation failures of said skb, which could have left the
session in limbo.
Reported-by: NJouni Malinen <j@w1.fi>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Pablo Neira Ayuso says:

====================
Netfilter fixes for net

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

1) Conntrack SCTP CRC32c checksum mangling may operate on non-linear
   skbuff, patch from Davide Caratti.

2) nf_tables rb-tree set backend does not handle element re-addition
   after deletion in the same transaction, leading to infinite loop.

3) Atomically unclear the IPS_SRC_NAT_DONE_BIT on nat module removal,
   from Liping Zhang.

4) Conntrack hashtable resizing while ctnetlink dump is progress leads
   to a dead reference to released objects in the lists, also from
   Liping.
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Stub functions in header files need to be static, or we can have
multiple definitions errors.
Reported-by: Nkbuild test robot <fengguang.wu@intel.com>
Fixes: 6335e9f2 ("net: dsa: mv88e6xxx: mv88e6390X SERDES support")
Signed-off-by: NAndrew Lunn <andrew@lunn.ch>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

syszkaller fuzzer triggered a divide by zero, when set calibration
through ioctl().

To fix it, test 'bitrate' if it is negative or 0, just return -EINVAL.
Reported-by: NAndrey Konovalov <andreyknvl@google.com>
Signed-off-by: NFiro Yang <firogm@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The binding documentation for the mv88e6xxx switch is missing the
eeprom-length property, which has been implemented since May 2016,
commit f8cd8753 ("dsa: mv88e6xxx: Handle eeprom-length property")
Signed-off-by: NAndrew Lunn <andrew@lunn.ch>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Merge tag 'led_fixes_for_4-12-rc3' of git://git.kernel.org/pub/scm/linux/kernel/git/j.anaszewski/linux-leds

Pull LED fix from Jacek Anaszewski:
 "A single LED fix for 4.12-rc3.

  leds-pca955x driver uses only i2c_smbus API and thus it should pass
  I2C_FUNC_SMBUS_BYTE_DATA flag to i2c_check_functionality"

* tag 'led_fixes_for_4-12-rc3' of git://git.kernel.org/pub/scm/linux/kernel/git/j.anaszewski/linux-leds:
  leds: pca955x: Correct I2C Functionality

Pull networking fixes from David Miller:

 1) Fix state pruning in bpf verifier wrt. alignment, from Daniel
    Borkmann.

 2) Handle non-linear SKBs properly in SCTP ICMP parsing, from Davide
    Caratti.

 3) Fix bit field definitions for rss_hash_type of descriptors in mlx5
    driver, from Jesper Brouer.

 4) Defer slave->link updates until bonding is ready to do a full commit
    to the new settings, from Nithin Sujir.

 5) Properly reference count ipv4 FIB metrics to avoid use after free
    situations, from Eric Dumazet and several others including Cong Wang
    and Julian Anastasov.

 6) Fix races in llc_ui_bind(), from Lin Zhang.

 7) Fix regression of ESP UDP encapsulation for TCP packets, from
    Steffen Klassert.

 8) Fix mdio-octeon driver Kconfig deps, from Randy Dunlap.

 9) Fix regression in setting DSCP on ipv6/GRE encapsulation, from Peter
    Dawson.

* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net: (43 commits)
  ipv4: add reference counting to metrics
  net: ethernet: ax88796: don't call free_irq without request_irq first
  ip6_tunnel, ip6_gre: fix setting of DSCP on encapsulated packets
  sctp: fix ICMP processing if skb is non-linear
  net: llc: add lock_sock in llc_ui_bind to avoid a race condition
  bonding: Don't update slave->link until ready to commit
  test_bpf: Add a couple of tests for BPF_JSGE.
  bpf: add various verifier test cases
  bpf: fix wrong exposure of map_flags into fdinfo for lpm
  bpf: add bpf_clone_redirect to bpf_helper_changes_pkt_data
  bpf: properly reset caller saved regs after helper call and ld_abs/ind
  bpf: fix incorrect pruning decision when alignment must be tracked
  arp: fixed -Wuninitialized compiler warning
  tcp: avoid fastopen API to be used on AF_UNSPEC
  net: move somaxconn init from sysctl code
  net: fix potential null pointer dereference
  geneve: fix fill_info when using collect_metadata
  virtio-net: enable TSO/checksum offloads for Q-in-Q vlans
  be2net: Fix offload features for Q-in-Q packets
  vlan: Fix tcp checksum offloads in Q-in-Q vlans
  ...

Pull XFS fixes from Darrick Wong:
 "A few miscellaneous bug fixes & cleanups:

   - Fix indlen block reservation accounting bug when splitting delalloc
     extent

   - Fix warnings about unused variables that appeared in -rc1.

   - Don't spew errors when bmapping a local format directory

   - Fix an off-by-one error in a delalloc eof assertion

   - Make fsmap only return inode information for CAP_SYS_ADMIN

   - Fix a potential mount time deadlock recovering cow extents

   - Fix unaligned memory access in _btree_visit_blocks

   - Fix various SEEK_HOLE/SEEK_DATA bugs"

* tag 'xfs-4.12-fixes-2' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux:
  xfs: Move handling of missing page into one place in xfs_find_get_desired_pgoff()
  xfs: Fix off-by-in in loop termination in xfs_find_get_desired_pgoff()
  xfs: Fix missed holes in SEEK_HOLE implementation
  xfs: fix off-by-one on max nr_pages in xfs_find_get_desired_pgoff()
  xfs: fix unaligned access in xfs_btree_visit_blocks
  xfs: avoid mount-time deadlock in CoW extent recovery
  xfs: only return detailed fsmap info if the caller has CAP_SYS_ADMIN
  xfs: bad assertion for delalloc an extent that start at i_size
  xfs: fix warnings about unused stack variables
  xfs: BMAPX shouldn't barf on inline-format directories
  xfs: fix indlen accounting error on partial delalloc conversion

Andrey Konovalov reported crashes in ipv4_mtu()

I could reproduce the issue with KASAN kernels, between
10.246.7.151 and 10.246.7.152 :

1) 20 concurrent netperf -t TCP_RR -H 10.246.7.152 -l 1000 &

2) At the same time run following loop :
while :
do
 ip ro add 10.246.7.152 dev eth0 src 10.246.7.151 mtu 1500
 ip ro del 10.246.7.152 dev eth0 src 10.246.7.151 mtu 1500
done

Cong Wang attempted to add back rt->fi in commit
82486aa6 ("ipv4: restore rt->fi for reference counting")
but this proved to add some issues that were complex to solve.

Instead, I suggested to add a refcount to the metrics themselves,
being a standalone object (in particular, no reference to other objects)

I tried to make this patch as small as possible to ease its backport,
instead of being super clean. Note that we believe that only ipv4 dst
need to take care of the metric refcount. But if this is wrong,
this patch adds the basic infrastructure to extend this to other
families.

Many thanks to Julian Anastasov for reviewing this patch, and Cong Wang
for his efforts on this problem.

Fixes: 2860583f ("ipv4: Kill rt->fi")
Signed-off-by: NEric Dumazet <edumazet@google.com>
Reported-by: NAndrey Konovalov <andreyknvl@google.com>
Reviewed-by: NJulian Anastasov <ja@ssi.bg>
Acked-by: NCong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>