1) Old code was hard to maintain, due to complex lock chains.
   (We probably will be able to remove some kfree_rcu() in callers)

2) Using a single timer to update all estimators does not scale.

3) Code was buggy on 32bit kernel (WRITE_ONCE() on 64bit quantity
   is not supposed to work well)

In this rewrite :

- I removed the RB tree that had to be scanned in
  gen_estimator_active(). qdisc dumps should be much faster.

- Each estimator has its own timer.

- Estimations are maintained in net_rate_estimator structure,
  instead of dirtying the qdisc. Minor, but part of the simplification.

- Reading the estimator uses RCU and a seqcount to provide proper
  support for 32bit kernels.

- We reduce memory need when estimators are not used, since
  we store a pointer, instead of the bytes/packets counters.

- xt_rateest_mt() no longer has to grab a spinlock.
  (In the future, xt_rateest_tg() could be switched to per cpu counters)
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When I prepared commit d250a5f9 ("pkt_sched: gen_estimator: Dont
report fake rate estimators"), htb still had an implicit rate estimator
for all its classes.

Then later, I made this rate estimator optional in commit 64153ce0
("net_sched: htb: do not setup default rate estimators"), but I forgot
to update htb use of gnet_stats_copy_rate_est()

After this patch, "tc -s qdisc ..." no longer report fake rate
estimators for HTB classes.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This change replaces sk_buff_head struct in Qdiscs with new qdisc_skb_head.

Its similar to the skb_buff_head api, but does not use skb->prev pointers.

Qdiscs will commonly enqueue at the tail of a list and dequeue at head.
While skb_buff_head works fine for this, enqueue/dequeue needs to also
adjust the prev pointer of next element.

The ->prev pointer is not required for qdiscs so we can just leave
it undefined and avoid one cacheline write access for en/dequeue.
Suggested-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NFlorian Westphal <fw@strlen.de>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

In kernel HTB keeps tokens in signed 64-bit in nanoseconds. In netlink
protocol these values are converted into pshed ticks (64ns for now) and
truncated to 32-bit. In struct tc_htb_xstats fields "tokens" and "ctokens"
are declared as unsigned 32-bit but they could be negative thus tool 'tc'
prints them as signed. Big values loose higher bits and/or become negative.

This patch clamps tokens in xstat into range from INT_MIN to INT_MAX.
In this way it's easier to understand what's going on here.
Signed-off-by: NKonstantin Khlebnikov <khlebnikov@yandex-team.ru>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

We already get child qdisc qlen, we also can get its backlog
so that class dumps can report it.

Also replace qstats by a single drop counter, but move it in
a separate cache line so that drops do not dirty useful cache lines.

Tested:

$ tc -s cl sh dev eth0
class htb 1:1 root leaf 3: prio 0 rate 1Gbit ceil 1Gbit burst 500000b cburst 500000b
 Sent 2183346912 bytes 9021815 pkt (dropped 2340774, overlimits 0 requeues 0)
 rate 1001Mbit 517543pps backlog 120758b 499p requeues 0
 lended: 9021770 borrowed: 0 giants: 0
 tokens: 9 ctokens: 9
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Qdisc performance suffers when packets are dropped at enqueue()
time because drops (kfree_skb()) are done while qdisc lock is held,
delaying a dequeue() draining the queue.

Nominal throughput can be reduced by 50 % when this happens,
at a time we would like the dequeue() to proceed as fast as possible.

Even FQ is vulnerable to this problem, while one of FQ goals was
to provide some flow isolation.

This patch adds a 'struct sk_buff **to_free' parameter to all
qdisc->enqueue(), and in qdisc_drop() helper.

I measured a performance increase of up to 12 %, but this patch
is a prereq so that future batches in enqueue() can fly.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Acked-by: NJesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

saw a debug splat:
net/include/net/sch_generic.h:287 suspicious rcu_dereference_check() usage!
other info that might help us debug this:
rcu_scheduler_active = 1, debug_locks = 0
 2 locks held by kworker/2:1/710:
  #0:  ("events"){.+.+.+}, at: [<ffffffff8106ca1d>]
  #1:  ((&q->work)){+.+...}, at: [<ffffffff8106ca1d>] process_one_work+0x14d/0x690
Workqueue: events htb_work_func
Call Trace:
 [<ffffffff812dc763>] dump_stack+0x85/0xc2
 [<ffffffff8109fee7>] lockdep_rcu_suspicious+0xe7/0x120
 [<ffffffff814ced47>] htb_work_func+0x67/0x70
Signed-off-by: NFlorian Westphal <fw@strlen.de>
Acked-by: NCong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Both htb_reset() and htb_destroy() can use __qdisc_reset_queue()
instead of __skb_queue_purge() to defer skb freeing of internal
queues.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

__QDISC_STATE_THROTTLED bit manipulation is rather expensive
for HTB and few others.

I already removed it for sch_fq in commit f2600cf0
("net: sched: avoid costly atomic operation in fq_dequeue()")
and so far nobody complained.

When one ore more packets are stuck in one or more throttled
HTB class, a htb dequeue() performs two atomic operations
to clear/set __QDISC_STATE_THROTTLED bit, while root qdisc
lock is held.

Removing this pair of atomic operations bring me a 8 % performance
increase on 200 TCP_RR tests, in presence of throttled classes.

This patch has no side effect, since nothing actually uses
disc_is_throttled() anymore.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

after removal of TCA_CBQ_OVL_STRATEGY from cbq scheduler, there are no
more callers of ->drop() outside of other ->drop functions, i.e.
nothing calls them.
Signed-off-by: NFlorian Westphal <fw@strlen.de>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Large tc dumps (tc -s {qdisc|class} sh dev ethX) done by Google BwE host
agent [1] are problematic at scale :

For each qdisc/class found in the dump, we currently lock the root qdisc
spinlock in order to get stats. Sampling stats every 5 seconds from
thousands of HTB classes is a challenge when the root qdisc spinlock is
under high pressure. Not only the dumps take time, they also slow
down the fast path (queue/dequeue packets) by 10 % to 20 % in some cases.

An audit of existing qdiscs showed that sch_fq_codel is the only qdisc
that might need the qdisc lock in fq_codel_dump_stats() and
fq_codel_dump_class_stats()

In v2 of this patch, I now use the Qdisc running seqcount to provide
consistent reads of packets/bytes counters, regardless of 32/64 bit arches.

I also changed rate estimators to use the same infrastructure
so that they no longer need to lock root qdisc lock.

[1]
http://static.googleusercontent.com/media/research.google.com/en//pubs/archive/43838.pdfSigned-off-by: NEric Dumazet <edumazet@google.com>
Cc: Cong Wang <xiyou.wangcong@gmail.com>
Cc: Jamal Hadi Salim <jhs@mojatatu.com>
Cc: John Fastabend <john.fastabend@gmail.com>
Cc: Kevin Athey <kda@google.com>
Cc: Xiaotian Pei <xiaotian@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

I found a serious performance bug in packet schedulers using hrtimers.

sch_htb and sch_fq are definitely impacted by this problem.

We constantly rearm high resolution timers if some packets are throttled
in one (or more) class, and other packets are flying through qdisc on
another (non throttled) class.

hrtimer_start() does not have the mod_timer() trick of doing nothing if
expires value does not change :

	if (timer_pending(timer) &&
            timer->expires == expires)
                return 1;

This issue is particularly visible when multiple cpus can queue/dequeue
packets on the same qdisc, as hrtimer code has to lock a remote base.

I used following fix :

1) Change htb to use qdisc_watchdog_schedule_ns() instead of open-coding
it.

2) Cache watchdog prior expiration. hrtimer might provide this, but I
prefer to not rely on some hrtimer internal.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NNicolas Dichtel <nicolas.dichtel@6wind.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

We saw qlen!=0 but backlog==0 on our production machine:

qdisc htb 1: dev eth0 root refcnt 2 r2q 10 default 1 direct_packets_stat 0 ver 3.17
 Sent 172680457356 bytes 222469449 pkt (dropped 0, overlimits 123575834 requeues 0)
 backlog 0b 72p requeues 0

The problem is we only count qlen for HTB qdisc but not backlog.
We need to update backlog too when we update qlen, so that we
can at least know the average packet length.

Cc: Jamal Hadi Salim <jhs@mojatatu.com>
Acked-by: NJamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: NCong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When the bottom qdisc decides to, for example, drop some packet,
it calls qdisc_tree_decrease_qlen() to update the queue length
for all its ancestors, we need to update the backlog too to
keep the stats on root qdisc accurate.

Cc: Jamal Hadi Salim <jhs@mojatatu.com>
Acked-by: NJamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: NCong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Remove nearly duplicated code and prepare for the following patch.

Cc: Jamal Hadi Salim <jhs@mojatatu.com>
Acked-by: NJamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: NCong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

For classifiers getting invoked via tc_classify(), we always need an
extra function call into tc_classify_compat(), as both are being
exported as symbols and tc_classify() itself doesn't do much except
handling of reclassifications when tp->classify() returned with
TC_ACT_RECLASSIFY.

CBQ and ATM are the only qdiscs that directly call into tc_classify_compat(),
all others use tc_classify(). When tc actions are being configured
out in the kernel, tc_classify() effectively does nothing besides
delegating.

We could spare this layer and consolidate both functions. pktgen on
single CPU constantly pushing skbs directly into the netif_receive_skb()
path with a dummy classifier on ingress qdisc attached, improves
slightly from 22.3Mpps to 23.1Mpps.
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>
Acked-by: NAlexei Starovoitov <ast@plumgrid.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Those were all workarounds for the formerly double meaning of
tx_queue_len, which broke scheduling algorithms if untreated.

Now that all in-tree drivers have been converted away from setting
tx_queue_len = 0, it should be safe to drop these workarounds for
categorically broken setups.
Signed-off-by: NPhil Sutter <phil@nwl.cc>
Cc: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

After previous patches to simplify qstats the qstats can be
made per cpu with a packed union in Qdisc struct.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This removes the use of qstats->qlen variable from the classifiers
and makes it an explicit argument to gnet_stats_copy_queue().

The qlen represents the qdisc queue length and is packed into
the qstats at the last moment before passnig to user space. By
handling it explicitely we avoid, in the percpu stats case, having
to figure out which per_cpu variable to put it in.

It would probably be best to remove it from qstats completely
but qstats is a user space ABI and can't be broken. A future
patch could make an internal only qstats structure that would
avoid having to allocate an additional u32 variable on the
Qdisc struct. This would make the qstats struct 128bits instead
of 128+32.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This adds helpers to manipulate qstats logic and replaces locations
that touch the counters directly. This simplifies future patches
to push qstats onto per cpu counters.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

In order to run qdisc's without locking statistics and estimators
need to be handled correctly.

To resolve bstats make the statistics per cpu. And because this is
only needed for qdiscs that are running without locks which is not
the case for most qdiscs in the near future only create percpu
stats when qdiscs set the TCQ_F_CPUSTATS flag.

Next because estimators use the bstats to calculate packets per
second and bytes per second the estimator code paths are updated
to use the per cpu statistics.
Signed-off-by: NJohn Fastabend <john.r.fastabend@intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

While using a MQ + NETEM setup, I had confirmation that the default
timer migration ( /proc/sys/kernel/timer_migration ) is killing us.

Installing this on a receiver side of a TCP_STREAM test, (NIC has 8 TX
queues) :

EST="est 1sec 4sec"
for ETH in eth1
do
 tc qd del dev $ETH root 2>/dev/null
 tc qd add dev $ETH root handle 1: mq
 tc qd add dev $ETH parent 1:1 $EST netem limit 70000 delay 6ms
 tc qd add dev $ETH parent 1:2 $EST netem limit 70000 delay 8ms
 tc qd add dev $ETH parent 1:3 $EST netem limit 70000 delay 10ms
 tc qd add dev $ETH parent 1:4 $EST netem limit 70000 delay 12ms
 tc qd add dev $ETH parent 1:5 $EST netem limit 70000 delay 14ms
 tc qd add dev $ETH parent 1:6 $EST netem limit 70000 delay 16ms
 tc qd add dev $ETH parent 1:7 $EST netem limit 80000 delay 18ms
 tc qd add dev $ETH parent 1:8 $EST netem limit 90000 delay 20ms
done

We can see that timers get migrated into a single cpu, presumably idle
at the time timers are set up.
Then all qdisc dequeues run from this cpu and huge lock contention
happens. This single cpu is stuck in softirq mode and cannot dequeue
fast enough.

    39.24%  [kernel]          [k] _raw_spin_lock
     2.65%  [kernel]          [k] netem_enqueue
     1.80%  [kernel]          [k] netem_dequeue
     1.63%  [kernel]          [k] copy_user_enhanced_fast_string
     1.45%  [kernel]          [k] _raw_spin_lock_bh

By pinning qdisc timers on the cpu running the qdisc, we respect proper
XPS setting and remove this lock contention.

     5.84%  [kernel]          [k] netem_enqueue
     4.83%  [kernel]          [k] _raw_spin_lock
     2.92%  [kernel]          [k] copy_user_enhanced_fast_string

Current Qdiscs that benefit from this change are :

	netem, cbq, fq, hfsc, tbf, htb.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

pfifo_fast and htb use skb lists, without needing their spinlocks.
(They instead use the standard qdisc lock)

We can use __skb_queue_head_init() instead of skb_queue_head_init()
to be consistent.
Signed-off-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>

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>

ktime_get_ns() replaces ktime_to_ns(ktime_get())

ktime_get_real_ns() replaces ktime_to_ns(ktime_get_real())
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

htb_dump() and htb_dump_class() do not strictly need to acquire
qdisc lock to fetch qdisc and/or class parameters.

We hold RTNL and no changes can occur.

This reduces by 50% qdisc lock pressure while doing tc qdisc|class dump
operations.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

If the class in skb->priority is not a leaf, apply filters from the
selected class, not the qdisc. This lets netfilter or user space
partially classify the packet.
Signed-off-by: NHarry Mason <harry.mason@smoothwall.net>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Do not use C99 // comments and correct a spelling typo.
Signed-off-by: NYang Yingliang <yangyingliang@huawei.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Prefer pr_warn(... to pr_warning(...
Signed-off-by: NYang Yingliang <yangyingliang@huawei.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

It already has a NULL pointer judgment of rtab in qdisc_put_rtab().
Remove the judgment outside of qdisc_put_rtab().
Signed-off-by: NYang Yingliang <yangyingliang@huawei.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Now, 32bit rates may be not the true rate.
So use rate_bytes_ps which is from
max(rate32, rate64) to calcualte quantum.
Signed-off-by: NYang Yingliang <yangyingliang@huawei.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

HTB already can deal with 64bit rates, we only have to add two new
attributes so that tc can use them to break the current 32bit ABI
barrier.

TCA_HTB_RATE64 : class rate  (in bytes per second)
TCA_HTB_CEIL64 : class ceil  (in bytes per second)

This allows us to setup HTB on 40Gbps links, as 32bit limit is
actually ~34Gbps
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Add an extra u64 rate parameter to psched_ratecfg_precompute()
so that some qdisc can opt-in for 64bit rates in the future,
to overcome the ~34 Gbits limit.

psched_ratecfg_getrate() reports a legacy structure to
tc utility, so if actual rate is above the 32bit rate field,
cap it to the 34Gbit limit.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Fix a typo added in commit 56b765b7 ("htb: improved accuracy at high
rates")

cbuffer should not be a copy of buffer.
Signed-off-by: NVimalkumar <j.vimal@gmail.com>
Signed-off-by: NEric Dumazet <edumazet@google.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Jiri Pirko <jpirko@redhat.com>
Reviewed-by: NJiri Pirko <jiri@resnulli.us>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

commit 56b765b7 ("htb: improved accuracy at high rates")
broke the "linklayer atm" handling.

 tc class add ... htb rate X ceil Y linklayer atm

The linklayer setting is implemented by modifying the rate table
which is send to the kernel.  No direct parameter were
transferred to the kernel indicating the linklayer setting.

The commit 56b765b7 ("htb: improved accuracy at high rates")
removed the use of the rate table system.

To keep compatible with older iproute2 utils, this patch detects
the linklayer by parsing the rate table.  It also supports future
versions of iproute2 to send this linklayer parameter to the
kernel directly. This is done by using the __reserved field in
struct tc_ratespec, to convey the choosen linklayer option, but
only using the lower 4 bits of this field.

Linklayer detection is limited to speeds below 100Mbit/s, because
at high rates the rtab is gets too inaccurate, so bad that
several fields contain the same values, this resembling the ATM
detect.  Fields even start to contain "0" time to send, e.g. at
1000Mbit/s sending a 96 bytes packet cost "0", thus the rtab have
been more broken than we first realized.
Signed-off-by: NJesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When userspace passes a large priority value
the assignment of the unsigned value hopt->prio
to  signed int cl->prio causes cl->prio to become negative and the
comparison is with TC_HTB_NUMPRIO is always false.

The result is that HTB crashes by referencing outside
the array when processing packets. With this patch the large value
wraps around like other values outside the normal range.

See: https://bugzilla.kernel.org/show_bug.cgi?id=60669Signed-off-by: NStephen Hemminger <stephen@networkplumber.org>
Acked-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

htb_sched structures are big, and source of false sharing on SMP.

Every time a packet is queued or dequeue, many cache lines must be
touched because structures are not lay out properly.

By carefully splitting htb_sched in two parts, and define sub structures
to increase data locality, we can improve performance dramatically on
SMP.

New htb_prio structure can also be used in htb_class to increase data
locality.

I got 26 % performance increase on a 24 threads machine, with 200
concurrent netperf in TCP_RR mode, using a HTB hierarchy of 4 classes.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Cc: Tom Herbert <therbert@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

htb_class structures are big, and source of false sharing on SMP.

By carefully splitting them in two parts, we can improve performance.

I got 9 % performance increase on a 24 threads machine, with 200
concurrent netperf in TCP_RR mode, using a HTB hierarchy of 4 classes.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Cc: Tom Herbert <therbert@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>