Nikolay Aleksandrov says:

====================
bonding: get rid of curr_slave_lock

This is the second patch-set dealing with bond locking and the purpose here
is to convert curr_slave_lock into a spinlock called "mode_lock" which can
be used in the various modes for their specific needs. The first three
patches cleanup the use of curr_slave_lock and prepare it for the
conversion which is done in patch 4 and then the modes that were using
their own locks are converted to use the new "mode_lock" giving us the
opportunity to remove their locks.
This patch-set has been tested in each mode by running enslave/release of
slaves in parallel with traffic transmission and miimon=1 i.e. running
all the time. In fact this lead to the discovery of a subtle bug related to
RCU which will be fixed in -net.
Also did an allmodconfig test just in case :-)

v2: fix bond_3ad_state_machine_handler's use of mode_lock and
    curr_slave_lock
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Now that locks have been removed, remove some unnecessary comments and
adjust others to reflect reality. Also add a comment to "mode_lock" to
describe its current users and give a brief summary why they need it.
Signed-off-by: NNikolay Aleksandrov <nikolay@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Now that we have bond->mode_lock, we can remove the state_machine_lock
and use it in its place. There're no fast paths requiring the per-port
spinlocks so it should be okay to consolidate them into mode_lock.
Also move it inside the unbinding function as we don't want to expose
mode_lock outside of the specific modes.
Suggested-by: NJay Vosburgh <jay.vosburgh@canonical.com>
Signed-off-by: NNikolay Aleksandrov <nikolay@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The ALB/TLB specific spinlocks are no longer necessary as we now have
bond->mode_lock for this purpose, so convert them and remove them from
struct alb_bond_info.
Also remove the unneeded lock/unlock functions and use spin_lock/unlock
directly.
Suggested-by: NJay Vosburgh <jay.vosburgh@canonical.com>
Signed-off-by: NNikolay Aleksandrov <nikolay@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

curr_slave_lock is now a misleading name, a much better name is
mode_lock as it'll be used for each mode's purposes and it's no longer
necessary to use a rwlock, a simple spinlock is enough.
Suggested-by: NJay Vosburgh <jay.vosburgh@canonical.com>
Signed-off-by: NNikolay Aleksandrov <nikolay@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Mostly all users of curr_slave_lock already have RTNL as we've discussed
previously so there's no point in using it, the one case where the lock
must stay is the 3ad code, in fact it's the only one.
It's okay to remove it from bond_do_fail_over_mac() as it's called with
RTNL and drops the curr_slave_lock anyway.
bond_change_active_slave() is one of the main places where
curr_slave_lock was used, it's okay to remove it as all callers use RTNL
these days before calling it, that's why we move the ASSERT_RTNL() in
the beginning to catch any potential offenders to this rule.
The RTNL argument actually applies to all of the places where
curr_slave_lock has been removed from in this patch.
Also remove the unnecessary bond_deref_active_protected() macro and use
rtnl_dereference() instead.
Signed-off-by: NNikolay Aleksandrov <nikolay@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

First in rlb_teach_disabled_mac_on_primary() it's okay to remove
curr_slave_lock as all callers except bond_alb_monitor() already hold
RTNL, and in case bond_alb_monitor() is executing we can at most have a
period with bad throughput (very unlikely though).
In bond_alb_monitor() it's okay to remove the read_lock as the slave
list is walked with RCU and the worst that could happen is another
transmitter at the same time and thus for a period which currently is 10
seconds (bond_alb.h: BOND_ALB_LP_TICKS).
And bond_alb_handle_active_change() is okay because it's always called
with RTNL. Removed the ASSERT_RTNL() because it'll be inserted in the
parent function in a following patch.
Signed-off-by: NNikolay Aleksandrov <nikolay@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Remove the read_lock in bond_3ad_lacpdu_recv() since when the slave is
being released its rx_handler is removed before 3ad unbind, so even if
packets arrive, they won't see the slave in an inconsistent state.
Signed-off-by: NNikolay Aleksandrov <nikolay@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

virtqueue_add() populates the virtqueue descriptor table from the sgs
given.  If it uses an indirect descriptor table, then it puts a single
descriptor in the descriptor table pointing to the kmalloc'ed indirect
table where the sg is populated.

Previously vring_add_indirect() did the allocation and the simple
linear layout.  We replace that with alloc_indirect() which allocates
the indirect table then chains it like the normal descriptor table so
we can reuse the core logic.

This slows down pktgen by less than 1/2 a percent (which uses direct
descriptors), as well as vring_bench, but it's far neater.

vring_bench before:
	1061485790-1104800648(1.08254e+09+/-6.6e+06)ns
vring_bench after:
	1125610268-1183528965(1.14172e+09+/-8e+06)ns

pktgen before:
   787781-796334(793165+/-2.4e+03)pps 365-369(367.5+/-1.2)Mb/sec (365530384-369498976(3.68028e+08+/-1.1e+06)bps) errors: 0

pktgen after:
   779988-790404(786391+/-2.5e+03)pps 361-366(364.35+/-1.3)Mb/sec (361914432-366747456(3.64885e+08+/-1.2e+06)bps) errors: 0

Now, if we make force indirect descriptors by turning off any_header_sg
in virtio_net.c:

pktgen before:
  713773-721062(718374+/-2.1e+03)pps 331-334(332.95+/-0.92)Mb/sec (331190672-334572768(3.33325e+08+/-9.6e+05)bps) errors: 0
pktgen after:
  710542-719195(714898+/-2.4e+03)pps 329-333(331.15+/-1.1)Mb/sec (329691488-333706480(3.31713e+08+/-1.1e+06)bps) errors: 0
Signed-off-by: NRusty Russell <rusty@rustcorp.com.au>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

We used to have several callers which just used arrays.  They're
gone, so we can use sg_next() everywhere, simplifying the code.

On my laptop, this slowed down vring_bench by 15%:

vring_bench before:
	936153354-967745359(9.44739e+08+/-6.1e+06)ns
vring_bench after:
	1061485790-1104800648(1.08254e+09+/-6.6e+06)ns

However, a more realistic test using pktgen on a AMD FX(tm)-8320 saw
a few percent improvement:

pktgen before:
  767390-792966(785159+/-6.5e+03)pps 356-367(363.75+/-2.9)Mb/sec (356068960-367936224(3.64314e+08+/-3e+06)bps) errors: 0

pktgen after:
   787781-796334(793165+/-2.4e+03)pps 365-369(367.5+/-1.2)Mb/sec (365530384-369498976(3.68028e+08+/-1.1e+06)bps) errors: 0
Signed-off-by: NRusty Russell <rusty@rustcorp.com.au>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This is the only driver which doesn't hand virtqueue_add_inbuf and
virtqueue_add_outbuf a well-formed, well-terminated sg.  Fix it,
so we can make virtio_add_* simpler.

pktgen results:
	modprobe pktgen
	echo 'add_device eth0' > /proc/net/pktgen/kpktgend_0
	echo nowait 1 > /proc/net/pktgen/eth0
	echo count 1000000 > /proc/net/pktgen/eth0
	echo clone_skb 100000 > /proc/net/pktgen/eth0
	echo dst_mac 4e:14:25:a9:30:ac > /proc/net/pktgen/eth0
	echo dst 192.168.1.2 > /proc/net/pktgen/eth0
	for i in `seq 20`; do echo start > /proc/net/pktgen/pgctrl; tail -n1 /proc/net/pktgen/eth0; done

Before:
  746547-793084(786421+/-9.6e+03)pps 346-367(364.4+/-4.4)Mb/sec (346397808-367990976(3.649e+08+/-4.5e+06)bps) errors: 0

After:
  767390-792966(785159+/-6.5e+03)pps 356-367(363.75+/-2.9)Mb/sec (356068960-367936224(3.64314e+08+/-3e+06)bps) errors: 0
Signed-off-by: NRusty Russell <rusty@rustcorp.com.au>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Jeff Kirsher says:

====================
Intel Wired LAN Driver Updates 2014-09-12

This series contains updates to e1000, ixgbe and ixgbevf.

Mark provide two fixes to reduce compile warnings produce by ixgbe
and ixgbevf.

Alex provides two patches for ixgbe, first removes the receive buffer
allocation at the end of the ixgbe_clean_rx_irq().  The reason for
removing this is to avoid the extra latency introduced by the MMIO write.
Second patch addresses several issues in the current ixgbe implementation
of busy poll sockets.  It was possible for frames to be delivered out of
order if they were held in GRO, so address this by flushing the GRO
buffers before releasing the q_vector back to the idle state.  Also, we
were having to take a spinlock on changing the state to and from idle,
so to resolve this, replaced the state value with an atomic and use
atomic_cmpxchg to change the value from idle, and a simple atomic set
to restore it back to idle after we have acquired it.  This allows us
to only use a locked operation on acquiring the vector without a need
for a locked operation to release it.

Florian Westphal provides several patches for e1000 which does some
cleanup and updating of the driver.  Moved e1000_tbi_adjust_stats()
so that he could make the function static.  Added a helper function
to deal with the tbi workaround that was located in 2 different
Rx clean functions.  Added a e1000_rx_buffer struct for use on receive
since the transmit and receive have different requirements.  Updates
e1000 to use napi_gro_frags API.
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

John Fastabend says:

====================
net/sched rcu classifiers and tcf

This series converts the tcf_proto usage to RCU.

This requires updating each classifier individually to handle the
new copy/update requirement and also to update the core list
traversals. This makes the assumption that updates to the tables
are infrequent in comparison to the packet per second being
classified. On a 10Gbps running near line rate we can easily
produce 12+ million packets per second so IMO this is a reasonable
assumption. The updates are serialized by RTNL.

I have done some basic testing on this series and do not see any
immediate splats or issues. The patch series has been running
on my dev systems for a month or so now and I've not seen any
issues. Although my configurations are not overly complicated.

My test cases at this point cover all the filters with a
tight loop to add/remove filters. Some basic estimator tests
where I add an estimator to the qdisc and verify the statistics
accurate using pktgen. And finally I have a small script to
exercise the 'tc actions' interface. Feel free to send me more
tests off list and I can run them.

This is prep work to drop the qdisc lock with the first
target being the ingress qdisc. To be done is making the
tc actions RCU safe and statistics per cpu. These patches
are in the works.

Comments:
  - Checkpatch is still giving errors on some >80 char lines I know
    about this. IMO the way to fix this is to restructure the sched
    code to avoid being so heavily indented. But doing this here
    bloats the patchset and anyways there are already lots of >80
    chars in these files. I would prefer to keep the patches as is
    but let me know if others think I should fix these and I will.
    A follow up patch set could restructure the code and fix this
    throughout the code blocks.
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch makes the cls_bpf classifier RCU safe. The tcf_lock
was being used to protect a list of cls_bpf_prog now this list
is RCU safe and updates occur with rcu_replace.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Make cls_u32 classifier safe to run without holding lock. This patch
converts statistics that are kept in read section u32_classify into
per cpu counters.

This patch was tested with a tight u32 filter add/delete loop while
generating traffic with pktgen. By running pktgen on vlan devices
created on top of a physical device we can hit the qdisc layer
correctly. For ingress qdisc's a loopback cable was used.

for i in {1..100}; do
        q=`echo $i%8|bc`;
        echo -n "u32 tos: iteration $i on queue $q";
        tc filter add dev p3p2 parent $p prio $i u32 match ip tos 0x10 0xff \
                  action skbedit queue_mapping $q;
        sleep 1;
        tc filter del dev p3p2 prio $i;

        echo -n "u32 tos hash table: iteration $i on queue $q";
        tc filter add dev p3p2 parent $p protocol ip prio $i handle 628: u32 divisor 1
        tc filter add dev p3p2 parent $p protocol ip prio $i u32 \
                match ip protocol 17 0xff link 628: offset at 0 mask 0xf00 shift 6 plus 0
        tc filter add dev p3p2 parent $p protocol ip prio $i u32 \
                ht 628:0 match ip tos 0x10 0xff action skbedit queue_mapping $q
        sleep 2;
        tc filter del dev p3p2 prio $i
        sleep 1;
done
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This uses per cpu counters in cls_u32 in preparation
to convert over to rcu.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Make cls_tcindex RCU safe.

This patch addds a new RCU routine rcu_dereference_bh_rtnl() to check
caller either holds the rcu read lock or RTNL. This is needed to
handle the case where tcindex_lookup() is being called in both cases.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

RCUify the route classifier. For now however spinlock's are used to
protect fastmap cache.

The issue here is the fastmap may be read by one CPU while the
cache is being updated by another. An array of pointers could be
one possible solution.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

RCU'ify fw classifier.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Make cgroup classifier safe for RCU.

Also drops the calls in the classify routine that were doing a
rcu_read_lock()/rcu_read_unlock(). If the rcu_read_lock() isn't held
entering this routine we have issues with deleting the classifier
chain so remove the unnecessary rcu_read_lock()/rcu_read_unlock()
pair noting all paths AFAIK hold rcu_read_lock.

If there is a case where classify is called without the rcu read lock
then an rcu splat will occur and we can correct it.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Enable basic classifier for RCU.

Dereferencing tp->root may look a bit strange here but it is needed
by my accounting because it is allocated at init time and needs to
be kfree'd at destroy time. However because it may be referenced in
the classify() path we must wait an RCU grace period before free'ing
it. We use kfree_rcu() and rcu_ APIs to enforce this. This pattern
is used in all the classifiers.

Also the hgenerator can be incremented without concern because it
is always incremented under RTNL.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

rcu'ify tcf_proto this allows calling tc_classify() without holding
any locks. Updaters are protected by RTNL.

This patch prepares the core net_sched infrastracture for running
the classifier/action chains without holding the qdisc lock however
it does nothing to ensure cls_xxx and act_xxx types also work without
locking. Additional patches are required to address the fall out.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Add __rcu notation to qdisc handling by doing this we can make
smatch output more legible. And anyways some of the cases should
be using rcu_dereference() see qdisc_all_tx_empty(),
qdisc_tx_chainging(), and so on.

Also *wake_queue() API is commonly called from driver timer routines
without rcu lock or rtnl lock. So I added rcu_read_lock() blocks
around netif_wake_subqueue and netif_tx_wake_queue.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch makes the cls_bpf classifier RCU safe. The tcf_lock
was being used to protect a list of cls_bpf_prog now this list
is RCU safe and updates occur with rcu_replace.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Make cls_u32 classifier safe to run without holding lock. This patch
converts statistics that are kept in read section u32_classify into
per cpu counters.

This patch was tested with a tight u32 filter add/delete loop while
generating traffic with pktgen. By running pktgen on vlan devices
created on top of a physical device we can hit the qdisc layer
correctly. For ingress qdisc's a loopback cable was used.

for i in {1..100}; do
        q=`echo $i%8|bc`;
        echo -n "u32 tos: iteration $i on queue $q";
        tc filter add dev p3p2 parent $p prio $i u32 match ip tos 0x10 0xff \
                  action skbedit queue_mapping $q;
        sleep 1;
        tc filter del dev p3p2 prio $i;

        echo -n "u32 tos hash table: iteration $i on queue $q";
        tc filter add dev p3p2 parent $p protocol ip prio $i handle 628: u32 divisor 1
        tc filter add dev p3p2 parent $p protocol ip prio $i u32 \
                match ip protocol 17 0xff link 628: offset at 0 mask 0xf00 shift 6 plus 0
        tc filter add dev p3p2 parent $p protocol ip prio $i u32 \
                ht 628:0 match ip tos 0x10 0xff action skbedit queue_mapping $q
        sleep 2;
        tc filter del dev p3p2 prio $i
        sleep 1;
done
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This uses per cpu counters in cls_u32 in preparation
to convert over to rcu.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Make cls_tcindex RCU safe.

This patch addds a new RCU routine rcu_dereference_bh_rtnl() to check
caller either holds the rcu read lock or RTNL. This is needed to
handle the case where tcindex_lookup() is being called in both cases.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

RCUify the route classifier. For now however spinlock's are used to
protect fastmap cache.

The issue here is the fastmap may be read by one CPU while the
cache is being updated by another. An array of pointers could be
one possible solution.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

RCU'ify fw classifier.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Make cgroup classifier safe for RCU.

Also drops the calls in the classify routine that were doing a
rcu_read_lock()/rcu_read_unlock(). If the rcu_read_lock() isn't held
entering this routine we have issues with deleting the classifier
chain so remove the unnecessary rcu_read_lock()/rcu_read_unlock()
pair noting all paths AFAIK hold rcu_read_lock.

If there is a case where classify is called without the rcu read lock
then an rcu splat will occur and we can correct it.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Enable basic classifier for RCU.

Dereferencing tp->root may look a bit strange here but it is needed
by my accounting because it is allocated at init time and needs to
be kfree'd at destroy time. However because it may be referenced in
the classify() path we must wait an RCU grace period before free'ing
it. We use kfree_rcu() and rcu_ APIs to enforce this. This pattern
is used in all the classifiers.

Also the hgenerator can be incremented without concern because it
is always incremented under RTNL.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

rcu'ify tcf_proto this allows calling tc_classify() without holding
any locks. Updaters are protected by RTNL.

This patch prepares the core net_sched infrastracture for running
the classifier/action chains without holding the qdisc lock however
it does nothing to ensure cls_xxx and act_xxx types also work without
locking. Additional patches are required to address the fall out.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Add __rcu notation to qdisc handling by doing this we can make
smatch output more legible. And anyways some of the cases should
be using rcu_dereference() see qdisc_all_tx_empty(),
qdisc_tx_chainging(), and so on.

Also *wake_queue() API is commonly called from driver timer routines
without rcu lock or rtnl lock. So I added rcu_read_lock() blocks
around netif_wake_subqueue and netif_tx_wake_queue.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When sending out a burst of packets across multiple descriptors,
it is sufficient to send one LDC "start" trigger for
the first descriptor, so do not send an LDC "start" for every
pass through vnet_start_xmit. Similarly, it is sufficient to send
one "DRING_STOPPED" trigger for the last dring (and if that
fails, hold off and send the trigger later).

Optimizations to the number of LDC messages helps avoid
filling up the LDC channel with superfluous LDC messages
that risk triggering flow-control on the channel,
and also boosts performance.
Signed-off-by: NSowmini Varadhan <sowmini.varadhan@oracle.com>
Acked-by: NRaghuram Kothakota <raghuram.kothakota@oracle.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

calling ether_setup is redundant since alloc_etherdev calls
it.
Signed-off-by: NSubbaraya Sundeep Bhatta <sbhatta@xilinx.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

It will use pr_info_one() to print the version info of the
driver in probe function only once. No need to use the static
variable here.
Signed-off-by: NVarka Bhadram <varkab@cdac.in>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>