This reverts commit b7493e91.

On its own, querying RDEV for a MAC address works fine. But when upgrading
from a qeth that previously queried DDEV on a z/VM NIC (ie. any kernel with
commit ec61bd2f), the RDEV query now returns a _different_ MAC address
than the DDEV query.

If the NIC is configured with MACPROTECT, z/VM apparently requires us to
use the MAC that was initially returned (on DDEV) and registered. So after
upgrading to a kernel that uses RDEV, the SETVMAC registration cmd for the
new MAC address fails and we end up with a non-operabel interface.

To avoid regressions on upgrade, switch back to using DDEV for the MAC
address query. The downgrade path (first RDEV, later DDEV) is fine, in this
case both queries return the same MAC address.

Fixes: b7493e91 ("s390/qeth: use Read device to query hypervisor for MAC")
Reported-by: Michal Kubecek <mkubecek@suse.com>
Tested-by: Karsten Graul <kgraul@linux.ibm.com>
Signed-off-by: Julian Wiedmann <jwi@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The __alx_open function can be called from ndo_open, which is called
under RTNL, or from alx_resume, which isn't. Since commit d768319c,
we're calling the netif_set_real_num_{tx,rx}_queues functions, which
need to be called under RTNL.

This is similar to commit 0c2cc02e ("igb: Move the calls to set the
Tx and Rx queues into igb_open").

Fixes: d768319c ("alx: enable multiple tx queues")
Signed-off-by: Sabrina Dubroca <sd@queasysnail.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

Fixes: fc7a6c28 ("sfc: use a semaphore to lock farch filters too")
Suggested-by: Joseph Korty <joe.korty@concurrent-rt.com>
Signed-off-by: Bert Kenward <bkenward@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Fix a bug where INT_STAT1 was written twice and
INT_STAT2 was ignored when disabling interrupts.

Fixes: b753a9fa ("net: phy: DP83TC811: Introduce support for the DP83TC811 phy")
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: Dan Murphy <dmurphy@ti.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Sowmini reported that a recent commit broke prefix routes for linklocal
addresses. The newly added modify_prefix_route is attempting to add a
new prefix route when the ifp priority does not match the route metric
however the check needs to account for the default priority. In addition,
the route add fails because the route already exists, and then the delete
removes the one that exists. Flip the order to do the delete first.

Fixes: 8308f3ff ("net/ipv6: Add support for specifying metric of connected routes")
Reported-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Tested-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David Ahern <dsahern@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

alloc_skb_with_frags uses __GFP_NORETRY for non-sleeping allocations
which is just a noop and a little bit confusing.

__GFP_NORETRY was added by ed98df33 ("net: use __GFP_NORETRY for
high order allocations") to prevent from the OOM killer. Yet this was
not enough because fb05e7a8 ("net: don't wait for order-3 page
allocation") didn't want an excessive reclaim for non-costly orders
so it made it completely NOWAIT while it preserved __GFP_NORETRY in
place which is now redundant.

Drop the pointless __GFP_NORETRY because this function is used as
copy&paste source for other places.
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Madalin Bucur says:

====================
DPAA fixes

A couple of fixes for the DPAA drivers, addressing an issue
with short UDP or TCP frames (with padding) that were marked
as having a wrong checksum and dropped by the FMan hardware
and a problem with the buffer used for the scatter-gather
table being too small as per the hardware requirements.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

The DPAA HW requires that at least 256 bytes from the start of the
first scatter-gather table entry are allocated and accessible. The
hardware reads the maximum size the table can have in one access,
thus requiring that the allocation and mapping to be done for the
maximum size of 256B even if there is a smaller number of entries
in the table.
Signed-off-by: Madalin Bucur <madalin.bucur@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The FMan hardware parser needs to be configured to remove the
short frame padding from the checksum calculation, otherwise
short UDP and TCP frames are likely to be marked as having a
bad checksum.
Signed-off-by: Madalin Bucur <madalin.bucur@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Driver performs the internal reload when it receives tx-timeout event from
the OS. Internal reload might fail in some scenarios e.g., fatal HW issues.
In such cases OS still see the link, which would result in undesirable
functionalities such as re-generation of tx-timeouts.
The patch addresses this issue by indicating the link-down to OS when
tx-timeout is detected, and keeping the link in down state till the
internal reload is successful.

Please consider applying it to 'net' branch.
Signed-off-by: Sudarsana Reddy Kalluru <Sudarsana.Kalluru@cavium.com>
Signed-off-by: Ariel Elior <ariel.elior@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Static checkers complain that id_tbl->table points to longs and 4 bytes
is smaller than sizeof(long).  But the since other side is dividing by
32 instead of sizeof(long), that means the current code works fine.

Anyway, it's more conventional to use the BITS_TO_LONGS() macro when
we're allocating a bitmap.
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The code assumes that there is 4 bytes in a pointer and it doesn't
allocate enough memory.
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Fast Open key could be stored in different endian based on the CPU.
Previously hosts in different endianness in a server farm using
the same key config (sysctl value) would produce different cookies.
This patch fixes it by always storing it as little endian to keep
same API for LE hosts.
Reported-by: Daniele Iamartino <danielei@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Merge tag 'mac80211-for-davem-2018-06-29' of git://git.kernel.org/pub/scm/linux/kernel/git/jberg/mac80211

Johannes Berg says:

====================
Just three fixes:
 * fix HT operation in mesh mode
 * disable preemption in control frame TX
 * check nla_parse_nested() return values
   where missing (two places)
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

Currently the kernel accounts the memory for network traffic through
mem_cgroup_[un]charge_skmem() interface. However the memory accounted
only includes the truesize of sk_buff which does not include the size of
sock objects. In our production environment, with opt-out kmem
accounting, the sock kmem caches (TCP[v6], UDP[v6], RAW[v6], UNIX) are
among the top most charged kmem caches and consume a significant amount
of memory which can not be left as system overhead. So, this patch
converts the kmem caches of all sock objects to SLAB_ACCOUNT.
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Suggested-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

At the very least we should check the return value if
nla_parse_nested() is called with a non-NULL policy.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>

Commit 9757235f, "nl80211: correct checks for
NL80211_MESHCONF_HT_OPMODE value") relaxed the range for the HT
operation field in meshconf, while also adding checks requiring
the non-greenfield and non-ht-sta bits to be set in certain
circumstances.  The latter bit is actually reserved for mesh BSSes
according to Table 9-168 in 802.11-2016, so in fact it should not
be set.

wpa_supplicant sets these bits because the mesh and AP code share
the same implementation, but authsae does not.  As a result, some
meshconf updates from authsae which set only the NONHT_MIXED
protection bits were being rejected.

In order to avoid breaking userspace by changing the rules again,
simply accept the values with or without the bits set, and mask
off the reserved bit to match the spec.

While in here, update the 802.11-2012 reference to 802.11-2016.

Fixes: 9757235f ("nl80211: correct checks for NL80211_MESHCONF_HT_OPMODE value")
Cc: Masashi Honma <masashi.honma@gmail.com>
Signed-off-by: Bob Copeland <bobcopeland@fb.com>
Reviewed-by: Masashi Honma <masashi.honma@gmail.com>
Reviewed-by: Masashi Honma <masashi.honma@gmail.com>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>

On pre-emption enabled kernels the following print was being seen due to
missing local_bh_disable/local_bh_enable calls.  mac80211 assumes that
pre-emption is disabled in the data path.

    BUG: using smp_processor_id() in preemptible [00000000] code: iwd/517
    caller is __ieee80211_subif_start_xmit+0x144/0x210 [mac80211]
    [...]
    Call Trace:
    dump_stack+0x5c/0x80
    check_preemption_disabled.cold.0+0x46/0x51
    __ieee80211_subif_start_xmit+0x144/0x210 [mac80211]

Fixes: 91180649 ("mac80211: Add support for tx_control_port")
Signed-off-by: Denis Kenzior <denkenz@gmail.com>
[commit message rewrite, fixes tag]
Signed-off-by: Johannes Berg <johannes.berg@intel.com>

This is clearly a bug.

We need to set the DMA buffer size in the HW otherwise corruption can
occur when receiving packets.

This is probably not occuring because of small MTU values and because HW
has a default value internally (which currently is bigger than default
buffer size).
Signed-off-by: Jose Abreu <joabreu@synopsys.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Joao Pinto <jpinto@synopsys.com>
Cc: Vitor Soares <soares@synopsys.com>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The recent poll change may lead to stalls for non-blocking connecting
SMC sockets, since sock_poll_wait is no longer performed on the
internal CLC socket, but on the outer SMC socket.  kernel_connect() on
the internal CLC socket returns with -EINPROGRESS, but the wake up
logic does not work in all cases. If the internal CLC socket is still
in state TCP_SYN_SENT when polled, sock_poll_wait() from sock_poll()
does not sleep. It is supposed to sleep till the state of the internal
CLC socket switches to TCP_ESTABLISHED.

This problem triggered a redesign of the SMC nonblocking connect logic.
This patch introduces a connect worker covering all connect steps
followed by a wake up of socket waiters. It allows to get rid of all
delays and locks in smc_poll().

Fixes: c0129a06 ("smc: convert to ->poll_mask")
Signed-off-by: Ursula Braun <ubraun@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Larry Brakmo proposal ( https://patchwork.ozlabs.org/patch/935233/
tcp: force cwnd at least 2 in tcp_cwnd_reduction) made us rethink
about our recent patch removing ~16 quick acks after ECN events.

tcp_enter_quickack_mode(sk, 1) makes sure one immediate ack is sent,
but in the case the sender cwnd was lowered to 1, we do not want
to have a delayed ack for the next packet we will receive.

Fixes: 522040ea ("tcp: do not aggressively quick ack after ECN events")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Neal Cardwell <ncardwell@google.com>
Cc: Lawrence Brakmo <brakmo@fb.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

What we want here is to embed a user-space program into the kernel.
Instead of the complex ELF magic, let's simply wrap it in the assembly
with the '.incbin' directive.
Signed-off-by: Masahiro Yamada <yamada.masahiro@socionext.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

On receving an incomplete message, the existing code stores the
remaining length of the cloned skb in the early_eaten field instead of
incrementing the value returned by __strp_recv. This defers invocation
of sock_rfree for the current skb until the next invocation of
__strp_recv, which returns early_eaten if early_eaten is non-zero.

This behavior causes a stall when the current message occupies the very
tail end of a massive skb, and strp_peek/need_bytes indicates that the
remainder of the current message has yet to arrive on the socket. The
TCP receive buffer is totally full, causing the TCP window to go to
zero, so the remainder of the message will never arrive.

Incrementing the value returned by __strp_recv by the amount otherwise
stored in early_eaten prevents stalls of this nature.
Signed-off-by: Doron Roberts-Kedes <doronrk@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Saeed Mahameed says:

====================
mlx5-fixes-2018-06-26

Fixes for mlx5 core and netdev driver:

Two fixes from Alex Vesker to address command interface issues
 - Race in command interface polling mode
 - Incorrect raw command length parsing

From Shay Agroskin, Fix wrong size allocation for QoS ETC TC regitster.

From Or Gerlitz and Eli Cohin, Address backward compatability issues for when
Eswitch capability is not advertised for the PF host driver
    - Fix required capability for manipulating MPFS
    - E-Switch, Disallow vlan/spoofcheck setup if not being esw manager
    - Avoid dealing with vport IB/eth representors if not being e-switch manager
    - E-Switch, Avoid setup attempt if not being e-switch manager
    - Don't attempt to dereference the ppriv struct if not being eswitch manager
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

fib_tests.sh became non-executable at some point. This is
what happens:
  selftests: net: fib_tests.sh: Warning: file fib_tests.sh is
  not executable, correct this.
  not ok 1..11 selftests: net: fib_tests.sh [FAIL]

Fixes: d69faad7 ("selftests: fib_tests: Add prefix route tests with metric")
Signed-off-by: Daniel Díaz <daniel.diaz@linaro.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

Jesper Dangaard Brouer says:

====================
xdp: don't mix XDP_TX and XDP_REDIRECT flush ops

Fix driver logic that are combining XDP_TX flush and XDP_REDIRECT map
flushing.  These are two different XDP xmit modes, and it is clearly
wrong to invoke both types of flush operations when only one of the
XDP xmit modes is used.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

The driver was combining XDP_TX virtqueue_kick and XDP_REDIRECT
map flushing (xdp_do_flush_map).  This is suboptimal, these two
flush operations should be kept separate.

The suboptimal behavior was introduced in commit 9267c430
("virtio-net: add missing virtqueue kick when flushing packets").

Fixes: 9267c430 ("virtio-net: add missing virtqueue kick when flushing packets")
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The driver was combining the XDP_TX tail flush and XDP_REDIRECT
map flushing (xdp_do_flush_map).  This is suboptimal, these two
flush operations should be kept separate.

It looks like the mistake was copy-pasted from ixgbe.

Fixes: d9314c47 ("i40e: add support for XDP_REDIRECT")
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The driver was combining the XDP_TX tail flush and XDP_REDIRECT
map flushing (xdp_do_flush_map).  This is suboptimal, these two
flush operations should be kept separate.

Fixes: 11393cc9 ("xdp: Add batching support to redirect map")
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The macb driver currently crashes on at91rm9200 with the following trace:

Unable to handle kernel NULL pointer dereference at virtual address 00000014
[...]
[<c031da44>] (macb_rx_desc) from [<c031f2bc>] (at91ether_open+0x2e8/0x3f8)
[<c031f2bc>] (at91ether_open) from [<c041e8d8>] (__dev_open+0x120/0x13c)
[<c041e8d8>] (__dev_open) from [<c041ec08>] (__dev_change_flags+0x17c/0x1a8)
[<c041ec08>] (__dev_change_flags) from [<c041ec4c>] (dev_change_flags+0x18/0x4c)
[<c041ec4c>] (dev_change_flags) from [<c07a5f4c>] (ip_auto_config+0x220/0x10b0)
[<c07a5f4c>] (ip_auto_config) from [<c000a4fc>] (do_one_initcall+0x78/0x18c)
[<c000a4fc>] (do_one_initcall) from [<c0783e50>] (kernel_init_freeable+0x184/0x1c4)
[<c0783e50>] (kernel_init_freeable) from [<c0574d70>] (kernel_init+0x8/0xe8)
[<c0574d70>] (kernel_init) from [<c00090e0>] (ret_from_fork+0x14/0x34)

Solve that by initializing bp->queues[0].bp in at91ether_init (as is done
in macb_init).

Fixes: ae1f2a56 ("net: macb: Added support for many RX queues")
Signed-off-by: Alexandre Belloni <alexandre.belloni@bootlin.com>
Acked-by: Nicolas Ferre <nicolas.ferre@microchip.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

With the brand-new syntax extension of Kconfig, we can directly
check the compiler capability in the configuration phase.

If the cc-can-link.sh fails, the BPFILTER_UMH is automatically
hidden by the dependency.

I also deleted 'default n', which is no-op.
Signed-off-by: Masahiro Yamada <yamada.masahiro@socionext.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

Following warning is seen when rmmod hinic. This is because affinity
value is not reset before calling free_irq(). This patch fixes it.

[   55.181232] WARNING: CPU: 38 PID: 19589 at kernel/irq/manage.c:1608
__free_irq+0x2aa/0x2c0

Fixes: 352f58b0 ("net-next/hinic: Set Rxq irq to specific cpu for NUMA")
Signed-off-by: Wei Yongjun <weiyongjun1@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Pablo Neira Ayuso says:

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

The following patchset contains Netfilter fixes for your net tree:

1) Missing netlink attribute validation in nf_queue, uncovered by KASAN,
   from Eric Dumazet.

2) Use pointer to sysctl table, save us 192 bytes of memory per netns.
   Also from Eric.

3) Possible use-after-free when removing conntrack helper modules due
   to missing synchronize RCU call. From Taehee Yoo.

4) Fix corner case in systcl writes to nf_log that lead to appending
   data to uninitialized buffer, from Jann Horn.

5) Jann Horn says we may indefinitely block other users of nf_log_mutex
   if a userspace access in proc_dostring() blocked e.g. due to a
   userfaultfd.

6) Fix garbage collection race for unconfirmed conntrack entries,
   from Florian Westphal.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

sizeof() will return unsigned value so in the error check
negative error code will be always larger than sizeof().

Fixes: a0d8e02c ("nfp: add support for reading nffw info")
Signed-off-by: Chengguang Xu <cgxu519@gmx.com>
Acked-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Jakub Kicinski says:

====================
nfp: MPLS and shared blocks TC offload fixes

This series brings two fixes to TC filter/action offload code.
Pieter fixes matching MPLS packets when the match is purely on
the MPLS ethertype and none of the MPLS fields are used.
John provides a fix for offload of shared blocks.  Unfortunately,
with shared blocks there is currently no guarantee that filters
which were added by the core will be removed before block unbind.
Our simple fix is to not support offload of rules on shared blocks
at all, a revert of this fix will be send for -next once the
reoffload infrastructure lands.  The shared blocks became important
as we are trying to use them for bonding offload (managed from user
space) and lack of remove calls leads to resource leaks.

v2:
 - fix build error reported by kbuild bot due to missing
   tcf_block_shared() helper.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

TC shared blocks allow multiple qdiscs to be grouped together and filters
shared between them. Currently the chains of filters attached to a block
are only flushed when the block is removed. If a qdisc is removed from a
block but the block still exists, flow del messages are not passed to the
callback registered for that qdisc. For the NFP, this presents the
possibility of rules still existing in hw when they should be removed.

Prevent binding to shared blocks until the kernel can send per qdisc del
messages when block unbinds occur.

tcf_block_shared() was not used outside of the core until now, so also
add an empty implementation for builds with CONFIG_NET_CLS=n.

Fixes: 48617387 ("net: sched: introduce shared filter blocks infrastructure")
Signed-off-by: John Hurley <john.hurley@netronome.com>
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Simon Horman <simon.horman@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Previously it was not possible to distinguish between mpls ether types and
other ether types. This leads to incorrect classification of offloaded
filters that match on mpls ether type. For example the following two
filters overlap:

 # tc filter add dev eth0 parent ffff: \
    protocol 0x8847 flower \
    action mirred egress redirect dev eth1

 # tc filter add dev eth0 parent ffff: \
    protocol 0x0800 flower \
    action mirred egress redirect dev eth2

The driver now correctly includes the mac_mpls layer where HW stores mpls
fields, when it detects an mpls ether type. It also sets the MPLS_Q bit to
indicate that the filter should match mpls packets.

Fixes: bb055c19 ("nfp: add mpls match offloading support")
Signed-off-by: Pieter Jansen van Vuuren <pieter.jansenvanvuuren@netronome.com>
Reviewed-by: Simon Horman <simon.horman@netronome.com>
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Two rules with different values of suppress_prefix or suppress_ifgroup
are not the same. This fixes an -EEXIST when running:

   $ ip -4 rule add table main suppress_prefixlength 0
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Fixes: f9d4b0c1 ("fib_rules: move common handling of newrule delrule msgs into fib_nl2rule")
Signed-off-by: David S. Miller <davem@davemloft.net>

The RDS core module creates rds_connections based on callbacks
from rds_loop_transport when sending/receiving packets to local
addresses.

These connections will need to be cleaned up when they are
created from a netns that is not init_net, and that netns is deleted.

Add the changes aligned with the changes from
commit ebeeb1ad ("rds: tcp: use rds_destroy_pending() to synchronize
netns/module teardown and rds connection/workq management") for
rds_loop_transport

Reported-and-tested-by: syzbot+4c20b3866171ce8441d2@syzkaller.appspotmail.com
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The command interface can work in two modes: Events and Polling.
In the general case, each time we invoke a command, a work is
queued to handle it.

When working in events, the interrupt handler completes the
command execution. On the other hand, when working in polling
mode, the work itself completes it.

Due to a bug in the work handler, a command could have been
completed by the interrupt handler, while the work handler
hasn't finished yet, causing the it to complete once again
if the command interface mode was changed from Events to
polling after the interrupt handler was called.

mlx5_unload_one()
        mlx5_stop_eqs()
                // Destroy the EQ before cmd EQ
                ...cmd_work_handler()
                        write_doorbell()
                        --> EVENT_TYPE_CMD
                                mlx5_cmd_comp_handler() // First free
                                        free_ent(cmd, ent->idx)
                                        complete(&ent->done)

        <-- mlx5_stop_eqs //cmd was complete
                // move to polling before destroying the last cmd EQ
                mlx5_cmd_use_polling()
                        cmd->mode = POLL;

                --> cmd_work_handler (continues)
                        if (cmd->mode == POLL)
                                mlx5_cmd_comp_handler() // Double free

The solution is to store the cmd->mode before writing the doorbell.

Fixes: e126ba97 ("mlx5: Add driver for Mellanox Connect-IB adapters")
Signed-off-by: Alex Vesker <valex@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>