Skip to content

cloud-kernel
cloud-kernel

openanolis / cloud-kernel
1 年多 前同步成功

通知 160
Star 36
Fork 7
cloud-kernel
cloud-kernel
切换分支/标签
  1. 22 12月, 2017 4 次提交
  3. 22 10月, 2017 1 次提交
  5. 17 10月, 2017 1 次提交
  7. 22 9月, 2017 1 次提交
  9. 31 8月, 2017 1 次提交
  11. 26 8月, 2017 1 次提交
    • W
      net_sched: remove tc class reference counting · 143976ce
      由 WANG Cong 提交于 
      For TC classes, their ->get() and ->put() are always paired, and the
reference counting is completely useless, because:

1) For class modification and dumping paths, we already hold RTNL lock,
   so all of these ->get(),->change(),->put() are atomic.

2) For filter bindiing/unbinding, we use other reference counter than
   this one, and they should have RTNL lock too.

3) For ->qlen_notify(), it is special because it is called on ->enqueue()
   path, but we already hold qdisc tree lock there, and we hold this
   tree lock when graft or delete the class too, so it should not be gone
   or changed until we release the tree lock.

Therefore, this patch removes ->get() and ->put(), but:

1) Adds a new ->find() to find the pointer to a class by classid, no
   refcnt.

2) Move the original class destroy upon the last refcnt into ->delete(),
   right after releasing tree lock. This is fine because the class is
   already removed from hash when holding the lock.

For those who also use ->put() as ->unbind(), just rename them to reflect
this change.

Cc: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: NCong Wang <xiyou.wangcong@gmail.com>
Acked-by: NJiri Pirko <jiri@mellanox.com>
Acked-by: NJamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      143976ce
  13. 17 8月, 2017 2 次提交
  15. 16 8月, 2017 1 次提交
  17. 12 8月, 2017 1 次提交
  19. 07 6月, 2017 1 次提交
  21. 18 5月, 2017 2 次提交
  23. 14 4月, 2017 1 次提交
  25. 13 3月, 2017 1 次提交
  27. 06 12月, 2016 1 次提交
    • E
      net_sched: gen_estimator: complete rewrite of rate estimators · 1c0d32fd
      由 Eric Dumazet 提交于 
      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>
      1c0d32fd
  29. 09 8月, 2016 2 次提交
    • M
      net/sched/sch_hfsc.c: remove unused cl_myfadj · 37088f61
      由 Michal Soltys 提交于 
      The code using this variable has been commented out in the past as it
was causing issues in upperlimited link-sharing scenarios.
Signed-off-by: NMichal Soltys <soltys@ziu.info>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      37088f61
    • M
      net/sched/sch_hfsc.c: keep fsc and virtual times in sync; fix an old bug · 678a6241
      由 Michal Soltys 提交于 
      This patch simplifies how we update fsc and calculate vt from it - while
keeping the expected functionality identical with how hfsc behaves
curently. It also fixes a certain issue introduced with
a very old patch.

The idea is, that instead of correcting cl_vt before fsc curve update
(rtsc_min) and correcting cl_vt after calculation (rtsc_y2x) to keep
cl_vt local to the current period - we can simply rely on virtual times
and curve values always being in sync - analogously to how rsc and usc
function, except that we use virtual time here.

Why hasn't it been done since the beginning this way ? The likely scenario
(basing on the code trying to correct curves whenever possible) was to
keep the virtual times as small as possible - as they have tendency to
"gallop" forward whenever their siblings and other fair sharing
subtrees are idling. On top of that, current code is subtly bugged, so
cumulative time (without any corrections) is always kept and used in
init_vf() when a new backlog period begins (using cl_cvtoff).

Is cumulative value safe ? Generally yes, though corner cases are easy
to create. For example consider:

1gbit interface
some 100kbit leaf, everything else idle

With current tick (64ns) 1s is 15625000 ticks, but the leaf is alone and
it's virtual time, so in reality it's 10000 times more. ITOW 38 bits are
needed to hold 1 second. 54 - 1 day, 59 - 1 month, 63 - 1 year (all
logarithms rounded up). It's getting somewhat dangerous, but also
requires setup excusing this kind of values not mentioning permanently
backlogged class for a year. In near most extreme case (10gbit, 10kbit
leaf), we have "enough" to hold ~13.6 days in 64 bits.

Well, the issue remains mostly theoretical and cl_cvtoff has been
working fine for all those years. Sensible configuration are de-facto
immune to this issue, and not so sensible can solve it with a cronjob
and its period inversely proportional to the insanity of such setup =)

Now let's explain the subtle bug mentioned earlier.

The issue is related to how offsets are kept and how we calculate
virtual times and update fair service curve(s). The issue itself is
subtle, but easy to observe with long m1 segments. It was introduced in
rather old patch:

Commit 99296150c7: "[NET_SCHED]: O(1) children vtoff adjustment
in HFSC scheduler"

(available in git://git.kernel.org/pub/scm/linux/kernel/git/tglx/history.git)

Originally when a new backlog period was started, cl_vtoff of each
sibling was updated with cl_cvtmax from past period - naturally moving
all cl_vt to proper starting point. That patch adjusted it so cumulative
offset is kept in the parent, and there is no need for traversing the
list (as any subsequent child activation derives new vt from already
active sibling(s)).

But with this change, cl_vtoff (of each sibling) is no longer persistent
across the inactivity periods, as it's calculated from parent's
cl_cvtoff on a new backlog period, conflicting with the following curve
correction from the previous period:

if (cl->cl_virtual.x == vt) {
        cl->cl_virtual.x -= cl->cl_vtoff;
	cl->cl_vtoff = 0;
}

This essentially tries to keep curve as if it was local to the period
and resets cl_vtoff (cumulative vt offset of the class) to 0 when
possible (read: when we have an intersection or if a new curve is below
the old one). But then it's recalculated from cl_cvtoff on next active
period.  Then rtsc_min() call preceding the above if() doesn't really
do what we expect it to do in such scenario - as it calculates the
minimum of corrected curve (from the previous backlog period) and the
new uncorrected curve (with offset derived from cl_cvtoff).

Example:

tc class add dev $ife parent 1:0 classid 1:1  hfsc ls m2 100mbit ul m2 100mbit
tc class add dev $ife parent 1:1 classid 1:10 hfsc ls m1 80mbit d 10s m2 20mbit
tc class add dev $ife parent 1:1 classid 1:11 hfsc ls m2 20mbit

start B, keep it backlogged, let it run 6s (30s worth of vt as A is idle)
pause B briefly to force cl_cvtoff update in parent (whole 1:1 going idle)
start A, let it run 10s
pause A briefly to force rtsc_min()

At this point we would expect A to continue at 20mbit after a brief
moment of 80mbit. But instead A will use 80mbit for full 10s again. It's
the effect of first correcting A (during 'start A'), and then - after
unpausing - calculating rtsc_min() from old corrected and new uncorrected
curve.

The patch fixes this bug and keepis vt and fsc in sync (virtual times
are cumulative, not local to the backlog period).
Signed-off-by: NMichal Soltys <soltys@ziu.info>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      678a6241
  31. 09 7月, 2016 1 次提交
    • F
      hfsc: reduce hfsc_sched to 14 cachelines · bba7eb5d
      由 Florian Westphal 提交于 
      hfsc_sched is huge (size: 920, cachelines: 15), but we can get it to 14
cachelines by placing level after filter_cnt (covering 4 byte hole) and
reducing period/nactive/flags to u32 (period is just a counter,
incremented when class becomes active -- 2**32 is plenty for this
purpose, also, long is only 32bit wide on 32bit platforms anyway).

cl_vtperiod is exported to userspace via tc_hfsc_stats, but its period
member is already u32, so no precision is lost there either.

Cc: Michal Soltys <soltys@ziu.info>
Signed-off-by: NFlorian Westphal <fw@strlen.de>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      bba7eb5d
  33. 01 7月, 2016 5 次提交
  35. 26 6月, 2016 1 次提交
    • E
      net_sched: drop packets after root qdisc lock is released · 520ac30f
      由 Eric Dumazet 提交于 
      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>
      520ac30f
  37. 11 6月, 2016 1 次提交
    • E
      net_sched: remove generic throttled management · 45f50bed
      由 Eric Dumazet 提交于 
      __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>
      45f50bed
  39. 09 6月, 2016 1 次提交
  41. 08 6月, 2016 1 次提交
    • E
      net: sched: do not acquire qdisc spinlock in qdisc/class stats dump · edb09eb1
      由 Eric Dumazet 提交于 
      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>
      edb09eb1
  43. 04 6月, 2016 1 次提交
  45. 01 3月, 2016 2 次提交
  47. 28 8月, 2015 1 次提交
    • D
      net: sched: consolidate tc_classify{,_compat} · 3b3ae880
      由 Daniel Borkmann 提交于 
      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>
      3b3ae880
  49. 30 9月, 2014 4 次提交
  51. 14 9月, 2014 1 次提交