platform/kernel/linux-rpi.git
7 years agotcp: md5: add an address prefix for key lookup
Ivan Delalande [Fri, 16 Jun 2017 01:07:06 +0000 (18:07 -0700)]
tcp: md5: add an address prefix for key lookup

This allows the keys used for TCP MD5 signature to be used for whole
range of addresses, specified with a prefix length, instead of only one
address as it currently is.

Signed-off-by: Bob Gilligan <gilligan@arista.com>
Signed-off-by: Eric Mowat <mowat@arista.com>
Signed-off-by: Ivan Delalande <colona@arista.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agocxgb4: notify uP to route ctrlq compl to rdma rspq
Raju Rangoju [Mon, 19 Jun 2017 12:10:48 +0000 (17:40 +0530)]
cxgb4: notify uP to route ctrlq compl to rdma rspq

During the module initialisation there is a possible race
(basically race between uld and lld) where neither the uld
nor lld notifies the uP about where to route the ctrl queue
completions. LLD skips notifying uP as the rdma queues were
not created by then (will leave it to ULD to notify the uP).
As the ULD comes up, it also skips notifying the uP as the
flag FULL_INIT_DONE is not set yet (ULD assumes that the
interface is not up yet).

Consequently, this race between uld and lld leaves uP
unnotified about where to send the ctrl queue completions
to, leading to iwarp RI_RES WR failure.

Here is the race:

CPU 0                                   CPU1

- allocates nic rx queus
- t4_sge_alloc_ctrl_txq()
(if rdma rsp queues exists,
tell uP to route ctrl queue
compl to rdma rspq)
                                - acquires the mutex_lock
                                - allocates rdma response queues
                                - if FULL_INIT_DONE set,
                                  tell uP to route ctrl queue compl
                                  to rdma rspq
                                - relinquishes mutex_lock
- acquires the mutex_lock
- enable_rx()
- set FULL_INIT_DONE
- relinquishes mutex_lock

This patch fixes the above issue.

Fixes: e7519f9926f1('cxgb4: avoid enabling napi twice to the same queue')
Signed-off-by: Raju Rangoju <rajur@chelsio.com>
Acked-by: Steve Wise <swise@opengridcomputing.com>
CC: Stable <stable@vger.kernel.org> # 4.9+
Signed-off-by: Ganesh Goudar <ganeshgr@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agocxgb4: add new T6 pci device id's
Ganesh Goudar [Mon, 19 Jun 2017 10:07:13 +0000 (15:37 +0530)]
cxgb4: add new T6 pci device id's

Add 0x6082, 0x6083 and 0x6084 T6 device id's

Signed-off-by: Ganesh Goudar <ganeshgr@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonfp: add VLAN filtering support
Pablo Cascón [Thu, 15 Jun 2017 23:22:15 +0000 (16:22 -0700)]
nfp: add VLAN filtering support

Add general use per-vNIC mailbox area and use it for VLAN filtering
support.  Initially proto is hardcoded to 802.1q.

Signed-off-by: Pablo Cascón <pablo.cascon@netronome.com>
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agocxgb4: fix a NULL dereference
Ganesh Goudar [Fri, 16 Jun 2017 10:06:09 +0000 (15:36 +0530)]
cxgb4: fix a NULL dereference

Avoid NULL dereference in setup_sge_queues() when the adapter is
in non offload mode.

Fixes: 0fbc81b3ad51 ('chcr/cxgb4i/cxgbit/RDMA/cxgb4: Allocate resources dynamically for all cxgb4 ULD's')
Signed-off-by: Ganesh Goudar <ganeshgr@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoliquidio: replace info-pointer mode with buffer-pointer-only mode
Prasad Kanneganti [Sun, 18 Jun 2017 12:04:11 +0000 (05:04 -0700)]
liquidio: replace info-pointer mode with buffer-pointer-only mode

Each Octeon output ring can DMA packets to host memory in two modes:  info-
pointer mode and buffer-pointer-only mode.  In info-pointer mode, Octeon
takes two buffer pointers for each packet and places the length of the
packet along with specified number of bytes from the beginning of the
packet into one buffer and the rest of the packet in a separate buffer.  In
buffer-pointer-only mode, Octeon takes single buffer pointer and places the
length of the packet at the beginning of the buffer followed by the packet
data.

This patch switches all Octeon output rings from info-pointer mode to
buffer-pointer-only mode.  This results in fewer DMA setups and cache line
snoops.

Signed-off-by: Prasad Kanneganti <pkanneganti@cavium.com>
Signed-off-by: Derek Chickles <derek.chickles@cavium.com>
Signed-off-by: Satanand Burla <satananda.burla@cavium.com>
Signed-off-by: Felix Manlunas <felix.manlunas@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agopptp: Remove unused variable in pptp_release()
Christos Gkekas [Sun, 18 Jun 2017 15:41:40 +0000 (16:41 +0100)]
pptp: Remove unused variable in pptp_release()

Variable opt in pptp_release() is set but never used, thus needs to be
removed.

Signed-off-by: Christos Gkekas <chris.gekas@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoliquidio: implement vlan filter enable and disable
Prasad Kanneganti [Sun, 18 Jun 2017 19:41:34 +0000 (12:41 -0700)]
liquidio: implement vlan filter enable and disable

Add implementation to support ethtool -K ethX rx-vlan-filter on/off.
Rename OCTNET_CMD_ENABLE_VLAN_FILTER command to OCTNET_CMD_VLAN_FILTER_CTL
and add OCTNET_CMD_VLAN_FILTER_ENABLE and OCTNET_CMD_VLAN_FILTER_DISABLE
parameters so that it can be used to enable or disable the filter.

Signed-off-by: Prasad Kanneganti <prasad.kanneganti@cavium.com>
Signed-off-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>
7 years agonet: dsa: Fix legacy probing
Florian Fainelli [Fri, 16 Jun 2017 23:42:11 +0000 (16:42 -0700)]
net: dsa: Fix legacy probing

After commit 6d3c8c0dd88a ("net: dsa: Remove master_netdev and
use dst->cpu_dp->netdev") and a29342e73911 ("net: dsa: Associate
slave network device with CPU port") we would be seeing NULL pointer
dereferences when accessing dst->cpu_dp->netdev too early. In the legacy
code, we actually know early in advance the master network device, so
pass it down to the relevant functions.

Fixes: 6d3c8c0dd88a ("net: dsa: Remove master_netdev and use dst->cpu_dp->netdev")
Fixes: a29342e73911 ("net: dsa: Associate slave network device with CPU port")
Reported-by: Jason Cobham <jcobham@questertangent.com>
Tested-by: Jason Cobham <jcobham@questertangent.com>
Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Reviewed-by: Vivien Didelot <vivien.didelot@savoirfairelinux.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agotls: update Kconfig
Dave Watson [Sat, 17 Jun 2017 15:25:13 +0000 (08:25 -0700)]
tls: update Kconfig

Missing crypto deps for some platforms.
Default to n for new module.

config: m68k-amcore_defconfig (attached as .config)
compiler: m68k-linux-gcc (GCC) 4.9.0

make.cross ARCH=m68k
All errors (new ones prefixed by >>):

   net/built-in.o: In function `tls_set_sw_offload':
>> (.text+0x732f8): undefined reference to `crypto_alloc_aead'
   net/built-in.o: In function `tls_set_sw_offload':
>> (.text+0x7333c): undefined reference to `crypto_aead_setkey'
   net/built-in.o: In function `tls_set_sw_offload':
>> (.text+0x73354): undefined reference to `crypto_aead_setauthsize'

Reported-by: kbuild test robot <fengguang.wu@intel.com>
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoMerge branch 'net-remove-dst-garbage-collector-logic'
David S. Miller [Sun, 18 Jun 2017 02:54:01 +0000 (22:54 -0400)]
Merge branch 'net-remove-dst-garbage-collector-logic'

Wei Wang says:

====================
remove dst garbage collector logic

The current mechanism of dst release is a bit complicated. It is because
the users of dst get divided into 2 situations:
  1. Most users take the reference count when using a dst and release the
     reference count when done.
  2. Exceptional users like IPv4/IPv6/decnet/xfrm routing code do not take
     reference count when referencing to a dst due to some histotic reasons.

Due to those exceptional use cases in 2, reference count being 0 is not an
adequate evidence to indicate that no user is using this dst. So users in 1
can't free the dst simply based on reference count being 0 because users in
2 might still hold reference to it.
Instead, a dst garbage list is needed to hold the dst entries that already
get removed by the users in 2 but are still held by users in 1. And a periodic
garbage collector task is run to check all the dst entries in the list to see
if the users in 1 have released the reference to those dst entries.
If so, the dst is now ready to be freed.

This logic introduces unnecessary complications in the dst code which makes it
hard to understand and to debug.

In order to get rid of the whole dst garbage collector (gc) and make the dst
code more unified and simplified, we can make the users in 2 also take reference
count on the dst and release it properly when done.
This way, dst can be safely freed once the refcount drops to 0 and no gc
thread is needed anymore.

This patch series' target is to completely get rid of dst gc logic and free
dst based on reference count only.
Patch 1-3 are preparation patches to do some cleanup/improvement on the existing
code to make later work easier.
Patch 4-21 are real implementations.
In these patches, a temporary flag DST_NOGC is used to help transition
those exceptional users one by one. Once every component is transitioned,
this temporary flag is removed.
By the end of this patch series, all dst are refcounted when being used
and released when done. And dst will be freed when its refcount drops to 0.
No dst gc task is running anymore.

Note: This patch series depends on the decnet fix that was sent right before:
      "decnet: always not take dst->__refcnt when inserting dst into hash table"

v2:
  add curly braces in udp_v4/6_early_demux() in patch 02
  add EXPORT_SYMBOL() for dst_dev_put() in patch 05
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: add debug atomic_inc_not_zero() in dst_hold()
Wei Wang [Sat, 17 Jun 2017 17:42:44 +0000 (10:42 -0700)]
net: add debug atomic_inc_not_zero() in dst_hold()

This patch is meant to add a debug warning on the situation where dst is
being held during its destroy phase. This could potentially cause double
free issue on the dst.

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: reorder all the dst flags
Wei Wang [Sat, 17 Jun 2017 17:42:43 +0000 (10:42 -0700)]
net: reorder all the dst flags

As some dst flags are removed, reorder the dst flags to fill in the
blanks.
Note: these flags are not exposed into user space. So it is safe to
reorder.

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: remove DST_NOCACHE flag
Wei Wang [Sat, 17 Jun 2017 17:42:42 +0000 (10:42 -0700)]
net: remove DST_NOCACHE flag

DST_NOCACHE flag check has been removed from dst_release() and
dst_hold_safe() in a previous patch because all the dst are now ref
counted properly and can be released based on refcnt only.
Looking at the rest of the DST_NOCACHE use, all of them can now be
removed or replaced with other checks.
So this patch gets rid of all the DST_NOCACHE usage and remove this flag
completely.

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: remove DST_NOGC flag
Wei Wang [Sat, 17 Jun 2017 17:42:41 +0000 (10:42 -0700)]
net: remove DST_NOGC flag

Now that all the components have been changed to release dst based on
refcnt only and not depend on dst gc anymore, we can remove the
temporary flag DST_NOGC.

Note that we also need to remove the DST_NOCACHE check in dst_release()
and dst_hold_safe() because now all the dst are released based on refcnt
and behaves as DST_NOCACHE.

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: remove dst gc related code
Wei Wang [Sat, 17 Jun 2017 17:42:40 +0000 (10:42 -0700)]
net: remove dst gc related code

This patch removes all dst gc related code and all the dst free
functions

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agodecnet: take dst->__refcnt when struct dn_route is created
Wei Wang [Sat, 17 Jun 2017 17:42:39 +0000 (10:42 -0700)]
decnet: take dst->__refcnt when struct dn_route is created

struct dn_route is inserted into dn_rt_hash_table but no dst->__refcnt
is taken.
This patch makes sure the dn_rt_hash_table's reference to the dst is ref
counted.

As the dst is always ref counted properly, we can safely mark
DST_NOGC flag so dst_release() will release dst based on refcnt only.
And dst gc is no longer needed and all dst_free() or its related
function calls should be replaced with dst_release() or
dst_release_immediate(). And dst_dev_put() is called when removing dst
from the hash table to release the reference on dst->dev before we lose
pointer to it.

Also, correct the logic in dn_dst_check_expire() and dn_dst_gc() to
check dst->__refcnt to be > 1 to indicate it is referenced by other
users.

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoxfrm: take refcnt of dst when creating struct xfrm_dst bundle
Wei Wang [Sat, 17 Jun 2017 17:42:38 +0000 (10:42 -0700)]
xfrm: take refcnt of dst when creating struct xfrm_dst bundle

During the creation of xfrm_dst bundle, always take ref count when
allocating the dst. This way, xfrm_bundle_create() will form a linked
list of dst with dst->child pointing to a ref counted dst child. And
the returned dst pointer is also ref counted. This makes the link from
the flow cache to this dst now ref counted properly.
As the dst is always ref counted properly, we can safely mark
DST_NOGC flag so dst_release() will release dst based on refcnt only.
And dst gc is no longer needed and all dst_free() and its related
function calls should be replaced with dst_release() or
dst_release_immediate().

The special handling logic for dst->child in dst_destroy() can be
replaced with a simple dst_release_immediate() call on the child to
release the whole list linked by dst->child pointer.
Previously used DST_NOHASH flag is not needed anymore as well. The
reason that DST_NOHASH is used in the existing code is mainly to prevent
the dst inserted in the fib tree to be wrongly destroyed during the
deletion of the xfrm_dst bundle. So in the existing code, DST_NOHASH
flag is marked in all the dst children except the one which is in the
fib tree.
However, with this patch series to remove dst gc logic and release dst
only based on ref count, it is safe to release all the children from a
xfrm_dst bundle as long as the dst children are all ref counted
properly which is already the case in the existing code.
So, this patch removes the use of DST_NOHASH flag.

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoipv6: get rid of icmp6 dst garbage collector
Wei Wang [Sat, 17 Jun 2017 17:42:37 +0000 (10:42 -0700)]
ipv6: get rid of icmp6 dst garbage collector

icmp6 dst route is currently ref counted during creation and will be
freed by user during its call of dst_release(). So no need of a garbage
collector for it.
Remove all icmp6 dst garbage collector related code.

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoipv6: mark DST_NOGC and remove the operation of dst_free()
Wei Wang [Sat, 17 Jun 2017 17:42:36 +0000 (10:42 -0700)]
ipv6: mark DST_NOGC and remove the operation of dst_free()

With the previous preparation patches, we are ready to get rid of the
dst gc operation in ipv6 code and release dst based on refcnt only.
So this patch adds DST_NOGC flag for all IPv6 dst and remove the calls
to dst_free() and its related functions.
At this point, all dst created in ipv6 code do not use the dst gc
anymore and will be destroyed at the point when refcnt drops to 0.

Also, as icmp6 dst route is refcounted during creation and will be freed
by user during its call of dst_release(), there is no need to add this
dst to the icmp6 gc list as well.
Instead, we need to add it into uncached list so that when a
NETDEV_DOWN/NETDEV_UNREGISRER event comes, we can properly go through
these icmp6 dst as well and release the net device properly.

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoipv6: call dst_hold_safe() properly
Wei Wang [Sat, 17 Jun 2017 17:42:35 +0000 (10:42 -0700)]
ipv6: call dst_hold_safe() properly

Similar as ipv4, ipv6 path also needs to call dst_hold_safe() when
necessary to avoid double free issue on the dst.

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoipv6: call dst_dev_put() properly
Wei Wang [Sat, 17 Jun 2017 17:42:34 +0000 (10:42 -0700)]
ipv6: call dst_dev_put() properly

As the intend of this patch series is to completely remove dst gc,
we need to call dst_dev_put() to release the reference to dst->dev
when removing routes from fib because we won't keep the gc list anymore
and will lose the dst pointer right after removing the routes.
Without the gc list, there is no way to find all the dst's that have
dst->dev pointing to the going-down dev.
Hence, we are doing dst_dev_put() immediately before we lose the last
reference of the dst from the routing code. The next dst_check() will
trigger a route re-lookup to find another route (if there is any).

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoipv6: take dst->__refcnt for insertion into fib6 tree
Wei Wang [Sat, 17 Jun 2017 17:42:33 +0000 (10:42 -0700)]
ipv6: take dst->__refcnt for insertion into fib6 tree

In IPv6 routing code, struct rt6_info is created for each static route
and RTF_CACHE route and inserted into fib6 tree. In both cases, dst
ref count is not taken.
As explained in the previous patch, this leads to the need of the dst
garbage collector.

This patch holds ref count of dst before inserting the route into fib6
tree and properly releases the dst when deleting it from the fib6 tree
as a preparation in order to fully get rid of dst gc later.

Also, correct fib6_age() logic to check dst->__refcnt to be 1 to indicate
no user is referencing the dst.

And remove dst_hold() in vrf_rt6_create() as ip6_dst_alloc() already puts
dst->__refcnt to 1.

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoipv4: mark DST_NOGC and remove the operation of dst_free()
Wei Wang [Sat, 17 Jun 2017 17:42:32 +0000 (10:42 -0700)]
ipv4: mark DST_NOGC and remove the operation of dst_free()

With the previous preparation patches, we are ready to get rid of the
dst gc operation in ipv4 code and release dst based on refcnt only.
So this patch adds DST_NOGC flag for all IPv4 dst and remove the calls
to dst_free().
At this point, all dst created in ipv4 code do not use the dst gc
anymore and will be destroyed at the point when refcnt drops to 0.

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoipv4: call dst_hold_safe() properly
Wei Wang [Sat, 17 Jun 2017 17:42:31 +0000 (10:42 -0700)]
ipv4: call dst_hold_safe() properly

This patch checks all the calls to
dst_hold()/skb_dst_force()/dst_clone()/dst_use() to see if
dst_hold_safe() is needed to avoid double free issue if dst
gc is removed and dst_release() directly destroys dst when
dst->__refcnt drops to 0.

In tx path, TCP hold sk->sk_rx_dst ref count and also hold sock_lock().
UDP and other similar protocols always hold refcount for
skb->_skb_refdst. So both paths seem to be safe.

In rx path, as it is lockless and skb_dst_set_noref() is likely to be
used, dst_hold_safe() should always be used when trying to hold dst.

In the routing code, if dst is held during an rcu protected session, it
is necessary to call dst_hold_safe() as the current dst might be in its
rcu grace period.

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoipv4: call dst_dev_put() properly
Wei Wang [Sat, 17 Jun 2017 17:42:30 +0000 (10:42 -0700)]
ipv4: call dst_dev_put() properly

As the intend of this patch series is to completely remove dst gc,
we need to call dst_dev_put() to release the reference to dst->dev
when removing routes from fib because we won't keep the gc list anymore
and will lose the dst pointer right after removing the routes.
Without the gc list, there is no way to find all the dst's that have
dst->dev pointing to the going-down dev.
Hence, we are doing dst_dev_put() immediately before we lose the last
reference of the dst from the routing code. The next dst_check() will
trigger a route re-lookup to find another route (if there is any).

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoipv4: take dst->__refcnt when caching dst in fib
Wei Wang [Sat, 17 Jun 2017 17:42:29 +0000 (10:42 -0700)]
ipv4: take dst->__refcnt when caching dst in fib

In IPv4 routing code, fib_nh and fib_nh_exception can hold pointers
to struct rtable but they never increment dst->__refcnt.
This leads to the need of the dst garbage collector because when user
is done with this dst and calls dst_release(), it can only decrement
dst->__refcnt and can not free the dst even it sees dst->__refcnt
drops from 1 to 0 (unless DST_NOCACHE flag is set) because the routing
code might still hold reference to it.
And when the routing code tries to delete a route, it has to put the
dst to the gc_list if dst->__refcnt is not yet 0 and have a gc thread
running periodically to check on dst->__refcnt and finally to free dst
when refcnt becomes 0.

This patch increments dst->__refcnt when
fib_nh/fib_nh_exception holds reference to this dst and properly release
the dst when fib_nh/fib_nh_exception has been updated with a new dst.

This patch is a preparation in order to fully get rid of dst gc later.

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: introduce a new function dst_dev_put()
Wei Wang [Sat, 17 Jun 2017 17:42:28 +0000 (10:42 -0700)]
net: introduce a new function dst_dev_put()

This function should be called when removing routes from fib tree after
the dst gc is no longer in use.
We first mark DST_OBSOLETE_DEAD on this dst to make sure next
dst_ops->check() fails and returns NULL.
Secondly, as we no longer keep the gc_list, we need to properly
release dst->dev right at the moment when the dst is removed from
the fib/fib6 tree.
It does the following:
1. change dst->input and output pointers to dst_discard/dst_dscard_out to
   discard all packets
2. replace dst->dev with loopback interface

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: introduce DST_NOGC in dst_release() to destroy dst based on refcnt
Wei Wang [Sat, 17 Jun 2017 17:42:27 +0000 (10:42 -0700)]
net: introduce DST_NOGC in dst_release() to destroy dst based on refcnt

The current mechanism of freeing dst is a bit complicated. dst has its
ref count and when user grabs the reference to the dst, the ref count is
properly taken in most cases except in IPv4/IPv6/decnet/xfrm routing
code due to some historic reasons.

If the reference to dst is always taken properly, we should be able to
simplify the logic in dst_release() to destroy dst when dst->__refcnt
drops from 1 to 0. And this should be the only condition to determine
if we can call dst_destroy().
And as dst is always ref counted, there is no need for a dst garbage
list to hold the dst entries that already get removed by the routing
code but are still held by other users. And the task to periodically
check the list to free dst if ref count become 0 is also not needed
anymore.

This patch introduces a temporary flag DST_NOGC(no garbage collector).
If it is set in the dst, dst_release() will call dst_destroy() when
dst->__refcnt drops to 0. dst_hold_safe() will also check for this flag
and do atomic_inc_not_zero() similar as DST_NOCACHE to avoid double free
issue.
This temporary flag is mainly used so that we can make the transition
component by component without breaking other parts.
This flag will be removed after all components are properly transitioned.

This patch also introduces a new function dst_release_immediate() which
destroys dst without waiting on the rcu when refcnt drops to 0. It will
be used in later patches.

Follow-up patches will correct all the places to properly take ref count
on dst and mark DST_NOGC. dst_release() or dst_release_immediate() will
be used to release the dst instead of dst_free() and its related
functions.
And final clean-up patch will remove the DST_NOGC flag.

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: use loopback dev when generating blackhole route
Wei Wang [Sat, 17 Jun 2017 17:42:26 +0000 (10:42 -0700)]
net: use loopback dev when generating blackhole route

Existing ipv4/6_blackhole_route() code generates a blackhole route
with dst->dev pointing to the passed in dst->dev.
It is not necessary to hold reference to the passed in dst->dev
because the packets going through this route are dropped anyway.
A loopback interface is good enough so that we don't need to worry about
releasing this dst->dev when this dev is going down.

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoudp: call dst_hold_safe() in udp_sk_rx_set_dst()
Wei Wang [Sat, 17 Jun 2017 17:42:25 +0000 (10:42 -0700)]
udp: call dst_hold_safe() in udp_sk_rx_set_dst()

In udp_v4/6_early_demux() code, we try to hold dst->__refcnt for
dst with DST_NOCACHE flag. This is because later in udp_sk_rx_dst_set()
function, we will try to cache this dst in sk for connected case.
However, a better way to achieve this is to not try to hold dst in
early_demux(), but in udp_sk_rx_dst_set(), call dst_hold_safe(). This
approach is also more consistant with how tcp is handling it. And it
will make later changes simpler.

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoipv6: remove unnecessary dst_hold() in ip6_fragment()
Wei Wang [Sat, 17 Jun 2017 17:42:24 +0000 (10:42 -0700)]
ipv6: remove unnecessary dst_hold() in ip6_fragment()

In ipv6 tx path, rcu_read_lock() is taken so that dst won't get freed
during the execution of ip6_fragment(). Hence, no need to hold dst in
it.

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoMerge tag 'mlx5-updates-2017-06-16' of git://git.kernel.org/pub/scm/linux/kernel...
David S. Miller [Fri, 16 Jun 2017 19:22:42 +0000 (15:22 -0400)]
Merge tag 'mlx5-updates-2017-06-16' of git://git./linux/kernel/git/saeed/linux

Saeed Mahameed says:

====================
Mellanox mlx5 updates and cleanups 2017-06-16

mlx5-updates-2017-06-16

This series provide some updates and cleanups for mlx5 core and netdevice
driver.

From Eli Cohen, add a missing event string.
From Or Gerlitz, some checkpatch cleanups.
From Moni, Disalbe HW level LAG when SRIOV is enabled.
From Tariq, A code reuse cleanup in aRFS flow.
From Itay Aveksis, Typo fix.
From Gal Pressman, ethtool statistics updates and "update stats" deferred work optimizations.
From Majd Dibbiny, Fast unload support on kernel shutdown.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: dsa: add cross-chip multicast support
Vivien Didelot [Thu, 15 Jun 2017 20:14:48 +0000 (16:14 -0400)]
net: dsa: add cross-chip multicast support

Similarly to how cross-chip VLAN works, define a bitmap of multicast
group members for a switch, now including its DSA ports, so that
multicast traffic can be sent to all switches of the fabric.

A switch may drop the frames if no user port is a member.

This brings support for multicast in a multi-chip environment.
As of now, all switches of the fabric must support the multicast
operations in order to program a single fabric port.

Reported-by: Jason Cobham <jcobham@questertangent.com>
Signed-off-by: Vivien Didelot <vivien.didelot@savoirfairelinux.com>
Tested-by: Jason Cobham <jcobham@questertangent.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoibmvnic: driver initialization for kdump/kexec
Nathan Fontenot [Thu, 15 Jun 2017 18:48:09 +0000 (14:48 -0400)]
ibmvnic: driver initialization for kdump/kexec

When booting into the kdump/kexec kernel, pHyp and vios
are not prepared for the initialization crq request and
a failover transport event is generated. This is not
handled correctly.

At this point in initialization the driver is still in
the 'probing' state and cannot handle a full reset of the
driver as is normally done for a failover transport event.

To correct this we catch driver resets while still in the
'probing' state and return EAGAIN. This results in the
driver tearing down the main crq and calling ibmvnic_init()
again.

Signed-off-by: Nathan Fontenot <nfont@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agodecnet: always not take dst->__refcnt when inserting dst into hash table
Wei Wang [Fri, 16 Jun 2017 17:46:37 +0000 (10:46 -0700)]
decnet: always not take dst->__refcnt when inserting dst into hash table

In the existing dn_route.c code, dn_route_output_slow() takes
dst->__refcnt before calling dn_insert_route() while dn_route_input_slow()
does not take dst->__refcnt before calling dn_insert_route().
This makes the whole routing code very buggy.
In dn_dst_check_expire(), dnrt_free() is called when rt expires. This
makes the routes inserted by dn_route_output_slow() not able to be
freed as the refcnt is not released.
In dn_dst_gc(), dnrt_drop() is called to release rt which could
potentially cause the dst->__refcnt to be dropped to -1.
In dn_run_flush(), dst_free() is called to release all the dst. Again,
it makes the dst inserted by dn_route_output_slow() not able to be
released and also, it does not wait on the rcu and could potentially
cause crash in the path where other users still refer to this dst.

This patch makes sure both input and output path do not take
dst->__refcnt before calling dn_insert_route() and also makes sure
dnrt_free()/dst_free() is called when removing dst from the hash table.
The only difference between those 2 calls is that dnrt_free() waits on
the rcu while dst_free() does not.

Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoMerge branch 'rds-tcp-misc-bug-fixes'
David S. Miller [Fri, 16 Jun 2017 16:45:16 +0000 (12:45 -0400)]
Merge branch 'rds-tcp-misc-bug-fixes'

Sowmini Varadhan says:

====================
rds: tcp: misc bug fixes

This series contains 2 bug fixes (patch2, patch3) and one bit of
code cleanup (patch1) identified during database testing
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agords: tcp: Set linger when rejecting an incoming conn in rds_tcp_accept_one
Sowmini Varadhan [Thu, 15 Jun 2017 18:28:55 +0000 (11:28 -0700)]
rds: tcp: Set linger when rejecting an incoming conn in rds_tcp_accept_one

Each time we get an incoming SYN to the RDS_TCP_PORT, the TCP
layer accepts the connection and then the rds_tcp_accept_one()
callback is invoked to process the incoming connection.

rds_tcp_accept_one() may reject the incoming syn for a number of
reasons, e.g., commit 1a0e100fb2c9 ("RDS: TCP: Force every connection
to be initiated by numerically smaller IP address"), or because
we are getting spammed by a malicious node that is triggering
a flood of connection attempts to RDS_TCP_PORT. If the incoming
syn is rejected, no data would have been sent on the TCP socket,
and we do not need to be in TIME_WAIT state, so we set linger on
the TCP socket before closing, thereby closing the socket efficiently
with a RST.

Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Tested-by: Imanti Mendez <imanti.mendez@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agords: tcp: various endian-ness fixes
Sowmini Varadhan [Thu, 15 Jun 2017 18:28:54 +0000 (11:28 -0700)]
rds: tcp: various endian-ness fixes

Found when testing between sparc and x86 machines on different
subnets, so the address comparison patterns hit the corner cases and
brought out some bugs fixed by this patch.

Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Tested-by: Imanti Mendez <imanti.mendez@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agords: tcp: remove cp_outgoing
Sowmini Varadhan [Thu, 15 Jun 2017 18:28:53 +0000 (11:28 -0700)]
rds: tcp: remove cp_outgoing

After commit 1a0e100fb2c9 ("RDS: TCP: Force every connection to be
initiated by numerically smaller IP address") we no longer need
the logic associated with cp_outgoing, so clean up usage of this
field.

Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Tested-by: Imanti Mendez <imanti.mendez@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoMerge branch 'dsa-loop-Driver-updates'
David S. Miller [Fri, 16 Jun 2017 16:43:49 +0000 (12:43 -0400)]
Merge branch 'dsa-loop-Driver-updates'

Florian Fainelli says:

====================
net: dsa: loop: Driver updates

This patch series updates drivers/net/dsa/dsa_loop.c to provide more useful
coverage of the DSA APIs and how we mangle the CPU ports ethtool statistics
to include its switch-facing port counters.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: dsa: loop: Implement ethtool statistics
Florian Fainelli [Thu, 15 Jun 2017 17:15:53 +0000 (10:15 -0700)]
net: dsa: loop: Implement ethtool statistics

When a DSA driver implements ethtool statistics, we also override the
master network device's ethtool statistics with the CPU port's
statistics and this has proven to be a possible source of bugs in the
past. Enhance the dsa_loop.c driver to provide statistics under the
forme of ok/error reads and writes from the per-port PHY read/writes.

Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: dsa: loop: Inline unregister_fixed_phys()
Florian Fainelli [Thu, 15 Jun 2017 17:15:52 +0000 (10:15 -0700)]
net: dsa: loop: Inline unregister_fixed_phys()

This is a simple function that only gets used in the driver's remove
function, inline it there.

Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoMerge branch 'pktgen-new-parameters'
David S. Miller [Fri, 16 Jun 2017 16:32:35 +0000 (12:32 -0400)]
Merge branch 'pktgen-new-parameters'

Tariq Toukan says:

====================
pktgen new parameters

This patchset adds two parameters to the pktgen scripts.
* The first patch adds a parameter to control the number of
  packets sent by every pktgen thread.
* The second patch adds a parameter to control the index of
  first thread, instead of always starting from cpu 0.

Series generated against net-next commit:
f7aec129a356 rxrpc: Cache the congestion window setting

Thanks,
Tariq.

V2:
* rebased to latest net-next.
* per Jesper's comments on Patch 2, fixed $F_THREAD description
  and $L_THREAD evaluation.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agopktgen: Specify the index of first thread
Tariq Toukan [Thu, 15 Jun 2017 16:07:22 +0000 (19:07 +0300)]
pktgen: Specify the index of first thread

Use "-f <num>", to specify the index of the first
sender thread.
In default first thread is #0.

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agopktgen: Specify num packets per thread
Tariq Toukan [Thu, 15 Jun 2017 16:07:21 +0000 (19:07 +0300)]
pktgen: Specify num packets per thread

Use -n <num>, to specify the number of packets every
thread sends.
Zero means indefinitely.

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoMerge branch 'net-mvmdio-add-xMDIO-xSMI-support'
David S. Miller [Fri, 16 Jun 2017 16:27:13 +0000 (12:27 -0400)]
Merge branch 'net-mvmdio-add-xMDIO-xSMI-support'

Antoine Tenart says:

====================
net: mvmdio: add xMDIO xSMI support

This series aims to add the xSMI support on the xMDIO bus to the
mvmdio driver. The xSMI interface complies with the IEEE 802.3 clause 45
and is used by 10GbE devices. On 7k and 8k (as of now), such an
interface is found and is used by Ethernet controllers.

Patches 1-4 and 9 are cosmetic cleanups.

Patches 5-7 are prerequisites to the xSMI support.

Patches 8 and 10-11 add the xSMI support to the mvmdio driver, and a
node is added both in the cp110 slave and master device trees.

This was tested on an Armada 8040 mcbin, as well as on both the
Armada 7040 DB and the Armada 8040 DB to ensure the SMI interface
was still working.

@Dave: patch 11 should go through the mvebu tree as asked by Gregory,
thanks!
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agodt-bindings: orion-mdio: document the new xmdio compatible
Antoine Ténart [Thu, 15 Jun 2017 14:43:25 +0000 (16:43 +0200)]
dt-bindings: orion-mdio: document the new xmdio compatible

A new compatible for Marvell xMDIO interfaces was added into the Marvell
MDIO driver. Document this new compatible.

Signed-off-by: Antoine Tenart <antoine.tenart@free-electrons.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: mvmdio: simplify the smi read and write error paths
Antoine Ténart [Thu, 15 Jun 2017 14:43:24 +0000 (16:43 +0200)]
net: mvmdio: simplify the smi read and write error paths

Cosmetic patch simplifying the smi read and write error paths. It also
align their error paths with the ones of the xsmi functions.

Signed-off-by: Antoine Tenart <antoine.tenart@free-electrons.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: mvmdio: add xmdio xsmi support
Antoine Ténart [Thu, 15 Jun 2017 14:43:23 +0000 (16:43 +0200)]
net: mvmdio: add xmdio xsmi support

This patch adds the xmdio xsmi interface support in the mvmdio driver.
This interface is used in Ethernet controllers on Marvell 370, 7k and 8k
(as of now). The xsmi interface supported by this driver complies with
the IEEE 802.3 clause 45. The xSMI interface is used by 10GbE devices.

Signed-off-by: Antoine Tenart <antoine.tenart@free-electrons.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: mvmdio: check the MII_ADDR_C45 bit is not set for smi operations
Antoine Ténart [Thu, 15 Jun 2017 14:43:22 +0000 (16:43 +0200)]
net: mvmdio: check the MII_ADDR_C45 bit is not set for smi operations

Add a check for the read and write smi operations, to ensure the
MII_ADDR_C45 bit isn't set. This will be needed as soon as the xSMI
support is added to the mvmdio driver.

Signed-off-by: Antoine Tenart <antoine.tenart@free-electrons.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: mvmdio: put the poll intervals in the ops structure
Antoine Ténart [Thu, 15 Jun 2017 14:43:21 +0000 (16:43 +0200)]
net: mvmdio: put the poll intervals in the ops structure

Put the two poll intervals (min and max) in the driver's ops
structure. This is needed to add the xmdio support later.

Signed-off-by: Antoine Tenart <antoine.tenart@free-electrons.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: mvmdio: introduce an ops structure
Antoine Ténart [Thu, 15 Jun 2017 14:43:20 +0000 (16:43 +0200)]
net: mvmdio: introduce an ops structure

Introduce an ops structure to add an indirection on the is_done
function, as this is needed to add the xMDIO support later.

Signed-off-by: Antoine Tenart <antoine.tenart@free-electrons.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: mvmdio: remove duplicate locking
Russell King [Thu, 15 Jun 2017 14:43:19 +0000 (16:43 +0200)]
net: mvmdio: remove duplicate locking

The MDIO layer already provides per-bus locking, so there's no need for
MDIO bus drivers to do their own internal locking.  Remove this.

Signed-off-by: Russell King <rmk+kernel@armlinux.org.uk>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: mvmdio: use GENMASK for masks
Antoine Ténart [Thu, 15 Jun 2017 14:43:18 +0000 (16:43 +0200)]
net: mvmdio: use GENMASK for masks

Cosmetic patch to use the GENMASK helper for masks.

Signed-off-by: Antoine Tenart <antoine.tenart@free-electrons.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: mvmdio: use tabs for defines
Antoine Ténart [Thu, 15 Jun 2017 14:43:17 +0000 (16:43 +0200)]
net: mvmdio: use tabs for defines

Cosmetic patch replacing spaces by tabs for defined values.

Signed-off-by: Antoine Tenart <antoine.tenart@free-electrons.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: mvmdio: reorder headers alphabetically
Antoine Ténart [Thu, 15 Jun 2017 14:43:16 +0000 (16:43 +0200)]
net: mvmdio: reorder headers alphabetically

Cosmetic fix reordering headers alphabetically.

Signed-off-by: Antoine Tenart <antoine.tenart@free-electrons.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoMerge branch 'bpf-xdp-Report-bpf_prog-ID-in-IFLA_XDP'
David S. Miller [Fri, 16 Jun 2017 15:58:38 +0000 (11:58 -0400)]
Merge branch 'bpf-xdp-Report-bpf_prog-ID-in-IFLA_XDP'

Martin KaFai Lau says:

====================
bpf: xdp: Report bpf_prog ID in IFLA_XDP

This is the first usage of the new bpf_prog ID.  It is for
reporting the ID of a xdp_prog through netlink.

It rides on the existing IFLA_XDP.  This patch adds IFLA_XDP_PROG_ID
for the bpf_prog ID reporting.

It starts with changing the generic_xdp first.  After that,
the hardware driver is changed one by one.  Jakub Kicinski mentioned
that he will soon introduce XDP_ATTACHED_HW (on top of the existing
XDP_ATTACHED_DRV and XDP_ATTACHED_SKB)
and he is going to reuse the prog_attached for this purpose.
Hence, this patch set keeps the prog_attached even though
!!prog_id also implies there is xdp_prog attached.

I have tested with generic_xdp, mlx4 and mlx5.

v3:
1. Replace 'if' by '?' when checking the xdp_prog pointer
   as suggested by Jakub Kicinski (thanks!)

v2:
1. Remove READ_ONCE since it is alredy under rtnl lock
2. Keep prog_attached in 'struct netdev_xdp' as
   requested by Jakub Kicinski.  The existing prog_attached
   and the new prog_id are put under a struct for XDP_QUERY_PROG.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agobpf: qede: Report bpf_prog ID during XDP_QUERY_PROG
Martin KaFai Lau [Fri, 16 Jun 2017 00:29:17 +0000 (17:29 -0700)]
bpf: qede: Report bpf_prog ID during XDP_QUERY_PROG

Add support to qede to report bpf_prog ID during XDP_QUERY_PROG.

Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Cc: Mintz Yuval <Yuval.Mintz@cavium.com>
Acked-by: Alexei Starovoitov <ast@fb.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agobpf: nfp: Report bpf_prog ID during XDP_QUERY_PROG
Martin KaFai Lau [Fri, 16 Jun 2017 00:29:16 +0000 (17:29 -0700)]
bpf: nfp: Report bpf_prog ID during XDP_QUERY_PROG

Add support to nfp to report bpf_prog ID during XDP_QUERY_PROG.

Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Cc: Jakub Kicinski <jakub.kicinski@netronome.com>
Acked-by: Alexei Starovoitov <ast@fb.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agobpf: ixgbe: Report bpf_prog ID during XDP_QUERY_PROG
Martin KaFai Lau [Fri, 16 Jun 2017 00:29:15 +0000 (17:29 -0700)]
bpf: ixgbe: Report bpf_prog ID during XDP_QUERY_PROG

Add support to ixgbe to report bpf_prog ID during XDP_QUERY_PROG.

Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Cc: Alexander Duyck <alexander.h.duyck@intel.com>
Cc: John Fastabend <john.fastabend@gmail.com>
Acked-by: Alexei Starovoitov <ast@fb.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agobpf: thunderx: Report bpf_prog ID during XDP_QUERY_PROG
Martin KaFai Lau [Fri, 16 Jun 2017 00:29:14 +0000 (17:29 -0700)]
bpf: thunderx: Report bpf_prog ID during XDP_QUERY_PROG

Add support to thunderx to report bpf_prog ID during XDP_QUERY_PROG.

Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Cc: Sunil Goutham <sgoutham@cavium.com>
Acked-by: Alexei Starovoitov <ast@fb.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agobpf: bnxt: Report bpf_prog ID during XDP_QUERY_PROG
Martin KaFai Lau [Fri, 16 Jun 2017 00:29:13 +0000 (17:29 -0700)]
bpf: bnxt: Report bpf_prog ID during XDP_QUERY_PROG

Add support to bnxt to report bpf_prog ID during XDP_QUERY_PROG.

Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Cc: Michael Chan <michael.chan@broadcom.com>
Acked-by: Alexei Starovoitov <ast@fb.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agobpf: virtio_net: Report bpf_prog ID during XDP_QUERY_PROG
Martin KaFai Lau [Fri, 16 Jun 2017 00:29:12 +0000 (17:29 -0700)]
bpf: virtio_net: Report bpf_prog ID during XDP_QUERY_PROG

Add support to virtio_net to report bpf_prog ID during XDP_QUERY_PROG.

Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Cc: John Fastabend <john.fastabend@gmail.com>
Cc: Jason Wang <jasowang@redhat.com>
Acked-by: Alexei Starovoitov <ast@fb.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agobpf: mlx5e: Report bpf_prog ID during XDP_QUERY_PROG
Martin KaFai Lau [Fri, 16 Jun 2017 00:29:11 +0000 (17:29 -0700)]
bpf: mlx5e: Report bpf_prog ID during XDP_QUERY_PROG

Add support to mlx5e to report bpf_prog ID during XDP_QUERY_PROG.

Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Cc: Tariq Toukan <tariqt@mellanox.com>
Cc: Saeed Mahameed <saeedm@mellanox.com>
Acked-by: Alexei Starovoitov <ast@fb.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agobpf: mlx4: Report bpf_prog ID during XDP_QUERY_PROG
Martin KaFai Lau [Fri, 16 Jun 2017 00:29:10 +0000 (17:29 -0700)]
bpf: mlx4: Report bpf_prog ID during XDP_QUERY_PROG

Add support to mlx4 to report bpf_prog ID during XDP_QUERY_PROG.

Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Cc: Tariq Toukan <tariqt@mellanox.com>
Cc: Saeed Mahameed <saeedm@mellanox.com>
Acked-by: Alexei Starovoitov <ast@fb.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: Add IFLA_XDP_PROG_ID
Martin KaFai Lau [Fri, 16 Jun 2017 00:29:09 +0000 (17:29 -0700)]
net: Add IFLA_XDP_PROG_ID

Expose prog_id through IFLA_XDP_PROG_ID.  This patch
makes modification to generic_xdp.  The later patches will
modify other xdp-supported drivers.

prog_id is added to struct net_dev_xdp.

iproute2 patch will be followed. Here is how the 'ip link'
will look like:
> ip link show eth0
3: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 xdp(prog_id:1) qdisc fq_codel state UP mode DEFAULT group default qlen 1000

Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@fb.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoMerge branch 'skb-accessor-cleanups'
David S. Miller [Fri, 16 Jun 2017 15:48:41 +0000 (11:48 -0400)]
Merge branch 'skb-accessor-cleanups'

Johannes Berg says:

====================
skb data accessors cleanup

Over night, Fengguang's bot told me that it compiled all of its many
various configurations successfully, and I had done allyesconfig on
x86_64 myself yesterday to iron out the things I missed.

So now I think I'm happy with it.

My tree was based on your

    commit 3715c47bcda8bb56f7e2be27276282a2d0d48c09
    Merge: 18b6e7955d8f d8fbd27469fc
    Author: David S. Miller <davem@davemloft.net>
    Date:   Thu Jun 15 14:31:56 2017 -0400

        Merge branch 'r8152-support-new-chips'

when the compilation tests happened, but I've reviewed the changes
coming into net-next in the meantime and didn't see any new usages
of skb data accessors having come in.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonetworking: add and use skb_put_u8()
Johannes Berg [Fri, 16 Jun 2017 12:29:24 +0000 (14:29 +0200)]
networking: add and use skb_put_u8()

Joe and Bjørn suggested that it'd be nicer to not have the
cast in the fairly common case of doing
*(u8 *)skb_put(skb, 1) = c;

Add skb_put_u8() for this case, and use it across the code,
using the following spatch:

    @@
    expression SKB, C, S;
    typedef u8;
    identifier fn = {skb_put};
    fresh identifier fn2 = fn ## "_u8";
    @@
    - *(u8 *)fn(SKB, S) = C;
    + fn2(SKB, C);

Note that due to the "S", the spatch isn't perfect, it should
have checked that S is 1, but there's also places that use a
sizeof expression like sizeof(var) or sizeof(u8) etc. Turns
out that nobody ever did something like
*(u8 *)skb_put(skb, 2) = c;

which would be wrong anyway since the second byte wouldn't be
initialized.

Suggested-by: Joe Perches <joe@perches.com>
Suggested-by: Bjørn Mork <bjorn@mork.no>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonetworking: make skb_push & __skb_push return void pointers
Johannes Berg [Fri, 16 Jun 2017 12:29:23 +0000 (14:29 +0200)]
networking: make skb_push & __skb_push return void pointers

It seems like a historic accident that these return unsigned char *,
and in many places that means casts are required, more often than not.

Make these functions return void * and remove all the casts across
the tree, adding a (u8 *) cast only where the unsigned char pointer
was used directly, all done with the following spatch:

    @@
    expression SKB, LEN;
    typedef u8;
    identifier fn = { skb_push, __skb_push, skb_push_rcsum };
    @@
    - *(fn(SKB, LEN))
    + *(u8 *)fn(SKB, LEN)

    @@
    expression E, SKB, LEN;
    identifier fn = { skb_push, __skb_push, skb_push_rcsum };
    type T;
    @@
    - E = ((T *)(fn(SKB, LEN)))
    + E = fn(SKB, LEN)

    @@
    expression SKB, LEN;
    identifier fn = { skb_push, __skb_push, skb_push_rcsum };
    @@
    - fn(SKB, LEN)[0]
    + *(u8 *)fn(SKB, LEN)

Note that the last part there converts from push(...)[0] to the
more idiomatic *(u8 *)push(...).

Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonetworking: make skb_pull & friends return void pointers
Johannes Berg [Fri, 16 Jun 2017 12:29:22 +0000 (14:29 +0200)]
networking: make skb_pull & friends return void pointers

It seems like a historic accident that these return unsigned char *,
and in many places that means casts are required, more often than not.

Make these functions return void * and remove all the casts across
the tree, adding a (u8 *) cast only where the unsigned char pointer
was used directly, all done with the following spatch:

    @@
    expression SKB, LEN;
    typedef u8;
    identifier fn = {
            skb_pull,
            __skb_pull,
            skb_pull_inline,
            __pskb_pull_tail,
            __pskb_pull,
            pskb_pull
    };
    @@
    - *(fn(SKB, LEN))
    + *(u8 *)fn(SKB, LEN)

    @@
    expression E, SKB, LEN;
    identifier fn = {
            skb_pull,
            __skb_pull,
            skb_pull_inline,
            __pskb_pull_tail,
            __pskb_pull,
            pskb_pull
    };
    type T;
    @@
    - E = ((T *)(fn(SKB, LEN)))
    + E = fn(SKB, LEN)

Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonetworking: make skb_put & friends return void pointers
Johannes Berg [Fri, 16 Jun 2017 12:29:21 +0000 (14:29 +0200)]
networking: make skb_put & friends return void pointers

It seems like a historic accident that these return unsigned char *,
and in many places that means casts are required, more often than not.

Make these functions (skb_put, __skb_put and pskb_put) return void *
and remove all the casts across the tree, adding a (u8 *) cast only
where the unsigned char pointer was used directly, all done with the
following spatch:

    @@
    expression SKB, LEN;
    typedef u8;
    identifier fn = { skb_put, __skb_put };
    @@
    - *(fn(SKB, LEN))
    + *(u8 *)fn(SKB, LEN)

    @@
    expression E, SKB, LEN;
    identifier fn = { skb_put, __skb_put };
    type T;
    @@
    - E = ((T *)(fn(SKB, LEN)))
    + E = fn(SKB, LEN)

which actually doesn't cover pskb_put since there are only three
users overall.

A handful of stragglers were converted manually, notably a macro in
drivers/isdn/i4l/isdn_bsdcomp.c and, oddly enough, one of the many
instances in net/bluetooth/hci_sock.c. In the former file, I also
had to fix one whitespace problem spatch introduced.

Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonetworking: introduce and use skb_put_data()
Johannes Berg [Fri, 16 Jun 2017 12:29:20 +0000 (14:29 +0200)]
networking: introduce and use skb_put_data()

A common pattern with skb_put() is to just want to memcpy()
some data into the new space, introduce skb_put_data() for
this.

An spatch similar to the one for skb_put_zero() converts many
of the places using it:

    @@
    identifier p, p2;
    expression len, skb, data;
    type t, t2;
    @@
    (
    -p = skb_put(skb, len);
    +p = skb_put_data(skb, data, len);
    |
    -p = (t)skb_put(skb, len);
    +p = skb_put_data(skb, data, len);
    )
    (
    p2 = (t2)p;
    -memcpy(p2, data, len);
    |
    -memcpy(p, data, len);
    )

    @@
    type t, t2;
    identifier p, p2;
    expression skb, data;
    @@
    t *p;
    ...
    (
    -p = skb_put(skb, sizeof(t));
    +p = skb_put_data(skb, data, sizeof(t));
    |
    -p = (t *)skb_put(skb, sizeof(t));
    +p = skb_put_data(skb, data, sizeof(t));
    )
    (
    p2 = (t2)p;
    -memcpy(p2, data, sizeof(*p));
    |
    -memcpy(p, data, sizeof(*p));
    )

    @@
    expression skb, len, data;
    @@
    -memcpy(skb_put(skb, len), data, len);
    +skb_put_data(skb, data, len);

(again, manually post-processed to retain some comments)

Reviewed-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonetworking: convert many more places to skb_put_zero()
Johannes Berg [Fri, 16 Jun 2017 12:29:19 +0000 (14:29 +0200)]
networking: convert many more places to skb_put_zero()

There were many places that my previous spatch didn't find,
as pointed out by yuan linyu in various patches.

The following spatch found many more and also removes the
now unnecessary casts:

    @@
    identifier p, p2;
    expression len;
    expression skb;
    type t, t2;
    @@
    (
    -p = skb_put(skb, len);
    +p = skb_put_zero(skb, len);
    |
    -p = (t)skb_put(skb, len);
    +p = skb_put_zero(skb, len);
    )
    ... when != p
    (
    p2 = (t2)p;
    -memset(p2, 0, len);
    |
    -memset(p, 0, len);
    )

    @@
    type t, t2;
    identifier p, p2;
    expression skb;
    @@
    t *p;
    ...
    (
    -p = skb_put(skb, sizeof(t));
    +p = skb_put_zero(skb, sizeof(t));
    |
    -p = (t *)skb_put(skb, sizeof(t));
    +p = skb_put_zero(skb, sizeof(t));
    )
    ... when != p
    (
    p2 = (t2)p;
    -memset(p2, 0, sizeof(*p));
    |
    -memset(p, 0, sizeof(*p));
    )

    @@
    expression skb, len;
    @@
    -memset(skb_put(skb, len), 0, len);
    +skb_put_zero(skb, len);

Apply it to the tree (with one manual fixup to keep the
comment in vxlan.c, which spatch removed.)

Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoMerge branch 'r8152-adjust-runtime-suspend-resume'
David S. Miller [Fri, 16 Jun 2017 15:37:13 +0000 (11:37 -0400)]
Merge branch 'r8152-adjust-runtime-suspend-resume'

Hayes Wang says:

====================
r8152: adjust runtime suspend/resume

v2:
For #1, replace GFP_KERNEL with GFP_NOIO for usb_submit_urb().

v1:
Improve the flow about runtime suspend/resume and make the code
easy to read.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agor8152: move calling delay_autosuspend function
hayeswang [Tue, 13 Jun 2017 07:14:40 +0000 (15:14 +0800)]
r8152: move calling delay_autosuspend function

Move calling delay_autosuspend() in rtl8152_runtime_suspend(). Calling
delay_autosuspend() as late as possible.

The original flows are
   1. check if the driver/device is busy now.
   2. set wake events.
   3. enter runtime suspend.

If the wake event occurs between (1) and (2), the device may miss it. Besides,
to avoid the runtime resume occurs after runtime suspend immediately, move the
checking to the end of rtl8152_runtime_suspend().

Signed-off-by: Hayes Wang <hayeswang@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agor8152: split rtl8152_resume function
hayeswang [Tue, 13 Jun 2017 07:14:39 +0000 (15:14 +0800)]
r8152: split rtl8152_resume function

Split rtl8152_resume() into rtl8152_runtime_resume() and
rtl8152_system_resume().

Besides, replace GFP_KERNEL with GFP_NOIO for usb_submit_urb().

Signed-off-by: Hayes Wang <hayeswang@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agotls: Depend upon INET not plain NET.
David S. Miller [Fri, 16 Jun 2017 15:28:49 +0000 (11:28 -0400)]
tls: Depend upon INET not plain NET.

We refer to TCP et al. symbols so have to use INET as
the dependency.

   ERROR: "tcp_prot" [net/tls/tls.ko] undefined!
>> ERROR: "tcp_rate_check_app_limited" [net/tls/tls.ko] undefined!
   ERROR: "tcp_register_ulp" [net/tls/tls.ko] undefined!
   ERROR: "tcp_unregister_ulp" [net/tls/tls.ko] undefined!
   ERROR: "do_tcp_sendpages" [net/tls/tls.ko] undefined!

Reported-by: kbuild test robot <fengguang.wu@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoMerge branch 'mlx4-XDP-performance-improvements'
David S. Miller [Fri, 16 Jun 2017 02:53:24 +0000 (22:53 -0400)]
Merge branch 'mlx4-XDP-performance-improvements'

Tariq Toukan says:

====================
mlx4 XDP performance improvements

This patchset contains data-path improvements, mainly for XDP_DROP
and XDP_TX cases.

Main patches:
* Patch 2 by Saeed allows enabling optimized A0 RX steering (in HW) when
  setting a single RX ring.
  With this configuration, HW packet-rate dramatically improves,
  reaching 28.1 Mpps in XDP_DROP case for both IPv4 (37% gain)
  and IPv6 (53% gain).
* Patch 6 enhances the XDP xmit function. Among other changes, now we
  ring one doorbell per NAPI. Patch gives 17% gain in XDP_TX case.
* Patch 7 obsoletes the NAPI of XDP_TX completion queue and integrates its
  poll into the respective RX NAPI. Patch gives 15% gain in XDP_TX case.

Series generated against net-next commit:
f7aec129a356 rxrpc: Cache the congestion window setting
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet/mlx4_en: Refactor mlx4_en_free_tx_desc
Tariq Toukan [Thu, 15 Jun 2017 11:35:40 +0000 (14:35 +0300)]
net/mlx4_en: Refactor mlx4_en_free_tx_desc

Some code re-ordering, functionally equivalent.

- The !tx_info->inl check is evaluated anyway in both flows
  (common case/end case). Run it first, this might finish
  the flows earlier.
- dma_unmap calls are identical in both flows, get it out
  of the if block into the common area.

Performance tests:
Tested on ConnectX3Pro, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz

Gain is too small to be measurable, no degradation sensed.
Results are similar for IPv4 and IPv6.

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet/mlx4_en: Replace TXBB_SIZE multiplications with shift operations
Tariq Toukan [Thu, 15 Jun 2017 11:35:39 +0000 (14:35 +0300)]
net/mlx4_en: Replace TXBB_SIZE multiplications with shift operations

Define LOG_TXBB_SIZE, log of TXBB_SIZE, and use it with a shift
operation instead of a multiplication with TXBB_SIZE.
Operations are equivalent as TXBB_SIZE is a power of two.

Performance tests:
Tested on ConnectX3Pro, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz

Gain is too small to be measurable, no degradation sensed.
Results are similar for IPv4 and IPv6.

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet/mlx4_en: Increase default TX ring size
Tariq Toukan [Thu, 15 Jun 2017 11:35:38 +0000 (14:35 +0300)]
net/mlx4_en: Increase default TX ring size

Increase the default TX ring size (from 512 to 1024) to match
the RX ring size.
This gives the XDP TX ring a better chance to keep up with the
rate of its RX ring in case of a high load of XDP_TX actions.

Tested:
Ethtool counter rx_xdp_tx_full used to increase, after applying this
patch it stopped.

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet/mlx4_en: Poll XDP TX completion queue in RX NAPI
Tariq Toukan [Thu, 15 Jun 2017 11:35:37 +0000 (14:35 +0300)]
net/mlx4_en: Poll XDP TX completion queue in RX NAPI

Instead of having their own NAPIs, XDP TX completion queues get
polled within the corresponding RX NAPI.
This prevents any possible race on TX ring prod/cons indices,
between the context that issues the transmits (RX NAPI) and the
context that handles the completions (was previously done in
a separate NAPI).

This also improves performance, as it decreases the number
of NAPIs running on a CPU, saving the overhead of syncing
and switching between the contexts.

Performance tests:
Tested on ConnectX3Pro, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz
Single queue no-RSS optimization ON.

XDP_TX packet rate:
-------------------------------------
     | Before    | After     | Gain |
IPv4 | 12.0 Mpps | 13.8 Mpps |  15% |
IPv6 | 12.0 Mpps | 13.8 Mpps |  15% |
-------------------------------------

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet/mlx4_en: Improve XDP xmit function
Tariq Toukan [Thu, 15 Jun 2017 11:35:36 +0000 (14:35 +0300)]
net/mlx4_en: Improve XDP xmit function

Several performance improvements in XDP TX datapath,
including:
- Ring a single doorbell for XDP TX ring per NAPI budget,
  instead of doing it per a lower threshold (was 8).
  This includes removing the flow of immediate doorbell ringing
  in case of a full TX ring.
- Compiler branch predictor hints.
- Calculate values in compile time rather than in runtime.

Performance tests:
Tested on ConnectX3Pro, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz
Single queue no-RSS optimization ON.

XDP_TX packet rate:
-------------------------------------
     | Before    | After     | Gain |
IPv4 | 10.3 Mpps | 12.0 Mpps |  17% |
IPv6 | 10.3 Mpps | 12.0 Mpps |  17% |
-------------------------------------

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet/mlx4_en: Improve stack xmit function
Tariq Toukan [Thu, 15 Jun 2017 11:35:35 +0000 (14:35 +0300)]
net/mlx4_en: Improve stack xmit function

Several small code and performance improvements in stack TX datapath,
including:
- Compiler branch predictor hints.
- Minimize variables scope.
- Move tx_info non-inline flow handling to a separate function.
- Calculate data_offset in compile time rather than in runtime
  (for !lso_header_size branch).
- Avoid trinary-operator ("?") when value can be preset in a matching
  branch.

Performance tests:
Tested on ConnectX3Pro, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz

Gain is too small to be measurable, no degradation sensed.
Results are similar for IPv4 and IPv6.

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet/mlx4_en: Improve transmit CQ polling
Tariq Toukan [Thu, 15 Jun 2017 11:35:34 +0000 (14:35 +0300)]
net/mlx4_en: Improve transmit CQ polling

Several small performance improvements in TX CQ polling,
including:
- Compiler branch predictor hints.
- Minimize variables scope.
- More proper check of cq type.
- Use boolean instead of int for a binary indication.

Performance tests:
Tested on ConnectX3Pro, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz

Packet-rate tests for both regular stack and XDP use cases:
No noticeable gain, no degradation.

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet/mlx4_en: Improve receive data-path
Tariq Toukan [Thu, 15 Jun 2017 11:35:33 +0000 (14:35 +0300)]
net/mlx4_en: Improve receive data-path

Several small performance improvements in RX datapath,
including:
- Compiler branch predictor hints.
- Replace a multiplication with a shift operation.
- Minimize variables scope.
- Write-prefetch for packet header.
- Avoid trinary-operator ("?") when value can be preset in a matching
  branch.
- Save a branch by updating RX ring doorbell within
  mlx4_en_refill_rx_buffers(), which now returns void.

Performance tests:
Tested on ConnectX3Pro, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz
Single queue no-RSS optimization ON
(enable by ethtool -L <interface> rx 1).

XDP_DROP packet rate:
Same (28.1 Mpps), lower CPU utilization (from ~100% to ~92%).

Drop packets in TC:
-------------------------------------
     | Before    | After     | Gain |
IPv4 | 4.14 Mpps | 4.18 Mpps |   1% |
-------------------------------------

XDP_TX packet rate:
-------------------------------------
     | Before    | After     | Gain |
IPv4 | 10.1 Mpps | 10.3 Mpps |   2% |
IPv6 | 10.1 Mpps | 10.3 Mpps |   2% |
-------------------------------------

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet/mlx4_en: Optimized single ring steering
Saeed Mahameed [Thu, 15 Jun 2017 11:35:32 +0000 (14:35 +0300)]
net/mlx4_en: Optimized single ring steering

Avoid touching RX QP RSS context when loading with only
one RX ring, to allow optimized A0 RX steering.

Enable by:
- loading mlx4_core with module param: log_num_mgm_entry_size = -6.
- then: ethtool -L <interface> rx 1

Performance tests:
Tested on ConnectX3Pro, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz

XDP_DROP packet rate:
-------------------------------------
     | Before    | After     | Gain |
IPv4 | 20.5 Mpps | 28.1 Mpps |  37% |
IPv6 | 18.4 Mpps | 28.1 Mpps |  53% |
-------------------------------------

Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Cc: kernel-team@fb.com
Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet/mlx4_en: Remove unused argument in TX datapath function
Tariq Toukan [Thu, 15 Jun 2017 11:35:31 +0000 (14:35 +0300)]
net/mlx4_en: Remove unused argument in TX datapath function

Remove owner argument, as it is obsolete and unused.
This also saves the overhead of calculating its value in data-path.

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agoatm: solos-pci: remove useless variable assignments
Gustavo A. R. Silva [Thu, 15 Jun 2017 19:56:21 +0000 (14:56 -0500)]
atm: solos-pci: remove useless variable assignments

Value assigned to variable _data32_ at lines 1254 and 1257 is
overwritten at line 1260 before it can be used. This makes
such variable assignments useless.

Addresses-Coverity-ID: 1227049
Signed-off-by: Gustavo A. R. Silva <garsilva@embeddedor.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet: dsa: assign default CPU port to all ports
Vivien Didelot [Thu, 15 Jun 2017 19:06:54 +0000 (15:06 -0400)]
net: dsa: assign default CPU port to all ports

The current code only assigns the default cpu_dp to all user ports of
the switch to which the CPU port belongs. The user ports of the other
switches of the fabric thus don't have a default CPU port.

This patch fixes this by assigning the cpu_dp of all user ports of all
switches of the fabric when the tree is fully parsed.

Fixes: a29342e73911 ("net: dsa: Associate slave network device with CPU port")
Signed-off-by: Vivien Didelot <vivien.didelot@savoirfairelinux.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
7 years agonet/mlx5: Add fast unload support in shutdown flow
Majd Dibbiny [Thu, 9 Feb 2017 12:20:12 +0000 (14:20 +0200)]
net/mlx5: Add fast unload support in shutdown flow

Adding a support to flush all HW resources with one FW command and
skip all the heavy unload flows of the driver on kernel shutdown.
There's no need to free all the SW context since a new fresh kernel
will be loaded afterwards.

Regarding the FW resources, they should be closed, otherwise we will
have leakage in the FW. To accelerate this flow, we execute one command
in the beginning that tells the FW that the driver isn't going to close
any of the FW resources and asks the FW to clean up everything.
Once the commands complete, it's safe to close the PCI resources and
finish the routine.

Signed-off-by: Majd Dibbiny <majd@mellanox.com>
Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
7 years agonet/mlx5: Expose command polling interface
Majd Dibbiny [Thu, 9 Feb 2017 11:20:46 +0000 (13:20 +0200)]
net/mlx5: Expose command polling interface

Add a new interface for commands execution that allows the
caller to wait for the command's completion in a busy-wait
loop (polling mode).

This is useful if we want to execute a command in a polling mode
while the driver is working in events mode for the rest of
the commands.
This interface will be used in the downstream patches.

Signed-off-by: Majd Dibbiny <majd@mellanox.com>
Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
7 years agonet/mlx5e: Optimize update stats work
Gal Pressman [Wed, 10 May 2017 12:10:33 +0000 (15:10 +0300)]
net/mlx5e: Optimize update stats work

Unlike ethtool stats, get_stats ndo provides information cached by
update stats work that is running in the background without updating
them explicitly.
We cannot update all counters inside the ndo because some
updates require firmware commands that cannot be performed under a
spinlock.

update_stats work does not need to update ALL counters, since only
some of them are needed by ndo_get_stats.
This patch will allow for a minimal run of update_stats using an extra
parameter which will update necessary counters only and cut 13
firmware commands in each iteration of the work.

Work duration previous to this patch: ~4200us.
Work duration after this patch: ~700us (17% of the original time).

Signed-off-by: Gal Pressman <galp@mellanox.com>
Reviewed-by: Eran Ben Elisha <eranbe@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
7 years agonet/mlx5e: Move and optimize query out of buffer function
Gal Pressman [Wed, 14 Jun 2017 08:52:33 +0000 (11:52 +0300)]
net/mlx5e: Move and optimize query out of buffer function

Move "query queue counter out of buffer" helper function out of
qp.c to en_main.c, since mlx5e netdev driver is the only one to use it.

Also allocate the output buffer on the stack instead of the heap, to reduce
number of heap allocs on update_stats work.

Signed-off-by: Gal Pressman <galp@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
7 years agonet/mlx5e: Reduce number of heap allocated buffers for update stats
Gal Pressman [Mon, 5 Jun 2017 13:45:45 +0000 (16:45 +0300)]
net/mlx5e: Reduce number of heap allocated buffers for update stats

Allocating buffers on the heap every 200ms is something we should avoid,
let's use buffers located on the stack instead.

Signed-off-by: Gal Pressman <galp@mellanox.com>
Reviewed-by: Eran Ben Elisha <eranbe@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
7 years agonet/mlx5e: Rename physical symbol errors counter
Gal Pressman [Wed, 19 Apr 2017 15:10:37 +0000 (18:10 +0300)]
net/mlx5e: Rename physical symbol errors counter

Rename rx_symbol_errors_phy to rx_pcs_symbol_err_phy, in order to
prevent confusion with rx_symbol_err_phy counter.

rx_pcs_symbol_err_phy counter counts the number of symbol errors that
were detected on the PCS (regardless of traffic) and weren't
corrected by FEC correction algorithm or that FEC algorithm was
not active on this interface.

rx_symbol_err_phy refers to errors on packet level (physical error
during a packet receive).

Fixes: 5db0a4f64c04 ("net/mlx5e: Expose physical layer statistical...")
Signed-off-by: Gal Pressman <galp@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
7 years agonet/mlx5e: Fix typo in warning if CQ moderation is not supported
Itay Aveksis [Wed, 7 Jun 2017 14:01:51 +0000 (17:01 +0300)]
net/mlx5e: Fix typo in warning if CQ moderation is not supported

Signed-off-by: Itay Aveksis <itayav@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
7 years agonet/mlx5e: Use function to map aRFS into traffic type
Tariq Toukan [Sun, 21 May 2017 08:08:18 +0000 (11:08 +0300)]
net/mlx5e: Use function to map aRFS into traffic type

For a better code reuse and readability, use the existing
function arfs_get_tt() to map arfs_type into mlx5e_traffic_types,
instead of duplicating the switch-case logic.

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
7 years agonet/mlx5: Undo LAG upon request to create virtual functions
Moni Shoua [Tue, 9 May 2017 09:20:55 +0000 (12:20 +0300)]
net/mlx5: Undo LAG upon request to create virtual functions

LAG cannot work if virtual functions are present. Therefore, if LAG is
configured, the attempt to create virtual functions will fail. This gives
precedence to LAG over SRIOV which is not the desired behavior as users
might want to use the bonding/teaming driver also want to work with SRIOV.
In that case we don't want to force an order of actions, first create
virtual functions and only than configure a bonding/teaming net device.
To fix, if LAG is configured during a request to create virtual
functions, remove it and continue.

We ignore ENODEV when trying to forbid lag. This makes sense
because "No such device" means that lag is forbidden anyway.

Signed-off-by: Moni Shoua <monis@mellanox.com>
Reviewed-by: Aviv Heller <avivh@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>