Under presence of TSO/GSO/GRO packets, codel at low rates can be quite
useless. In following example, not a single packet was ever dropped,
while average delay in codel queue is ~100 ms !
qdisc codel 0: parent 1:12 limit 16000p target 5.0ms interval 100.0ms
Sent 134376498 bytes 88797 pkt (dropped 0, overlimits 0 requeues 0)
backlog 13626b 3p requeues 0
count 0 lastcount 0 ldelay 96.9ms drop_next 0us
maxpacket 9084 ecn_mark 0 drop_overlimit 0
This comes from a confusion of what should be the minimal backlog. It is
pretty clear it is not 64KB or whatever max GSO packet ever reached the
qdisc.
codel intent was to use MTU of the device.
After the fix, we finally drop some packets, and rtt/cwnd of my single
TCP flow are meeting our expectations.
qdisc codel 0: parent 1:12 limit 16000p target 5.0ms interval 100.0ms
Sent 102798497 bytes 67912 pkt (dropped 1365, overlimits 0 requeues 0)
backlog 6056b 3p requeues 0
count 1 lastcount 1 ldelay 36.3ms drop_next 0us
maxpacket 10598 ecn_mark 0 drop_overlimit 0
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Kathleen Nichols <nichols@pollere.com>
Cc: Dave Taht <dave.taht@gmail.com>
Cc: Van Jacobson <vanj@google.com>
Signed-off-by: David 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: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
As David Laight suggests, we shouldn't necessarily call this
reciprocal_divide() when users didn't requested a reciprocal_value();
lets keep the basic idea and call it reciprocal_scale(). More
background information on this topic can be found in [1].
Joint work with Hannes Frederic Sowa.
[1] http://homepage.cs.uiowa.edu/~jones/bcd/divide.html
Suggested-by: David Laight <david.laight@aculab.com>
Cc: Jakub Zawadzki <darkjames-ws@darkjames.pl>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
As described in commit 5a581b367 (jiffies: Avoid undefined
behavior from signed overflow), according to the C standard
3.4.3p3, overflow of a signed integer results in undefined
behavior.
To fix this, do as the above commit, and do an unsigned
subtraction, and interpreting the result as a signed
two's-complement number. This is based on the theory from
RFC 1982 and is nicely described in wikipedia here:
https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
A side-note, I have seen practical issues with the previous logic
when dealing with 16-bit, on a 64-bit machine (gcc version
4.4.5). This were 32-bit, which I have not observed issues with.
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Jesper Dangaard Brouer <netoptimizer@brouer.com>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
One condition before codel_Newton_step() was not good if
we never left the dropping state for a flow. As a result
rec_inv_sqrt was 0, instead of the ~0 initial value.
codel control law was then set to a very aggressive mode, dropping
many packets before reaching 'target' and recovering from this problem.
To keep codel_vars_init() as efficient as possible, refine
the condition to make sure rec_inv_sqrt initial value is correct
Many thanks to Anton Mich for discovering the issue and suggesting
a fix.
Reported-by: Anton Mich <lp2s1h@gmail.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
codel_should_drop() logic allows a packet being not dropped if queue
size is under max packet size.
In fq_codel, we have two possible backlogs : The qdisc global one, and
the flow local one.
The meaningful one for codel_should_drop() should be the global backlog,
not the per flow one, so that thin flows can have a non zero drop/mark
probability.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Dave Taht <dave.taht@bufferbloat.net>
Cc: Kathleen Nichols <nichols@pollere.com>
Cc: Van Jacobson <van@pollere.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
David pointed out gcc might generate poor code with 31bit fields.
Using u16 is more than enough and permits a better code output.
Also make the code intent more readable using constants, fixed point arithmetic
not being trivial for everybody.
Suggested-by: David Miller <davem@davemloft.net>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
As Van pointed out, interval/sqrt(count) can be implemented using
multiplies only.
http://en.wikipedia.org/wiki/Methods_of_computing_square_roots#Iterative_methods_for_reciprocal_square_roots
This patch implements the Newton method and reciprocal divide.
Total cost is 15 cycles instead of 120 on my Corei5 machine (64bit
kernel).
There is a small 'error' for count values < 5, but we don't really care.
I reuse a hole in struct codel_vars :
- pack the dropping boolean into one bit
- use 31bit to store the reciprocal value of sqrt(count).
Suggested-by: Van Jacobson <van@pollere.net>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Dave Taht <dave.taht@bufferbloat.net>
Cc: Kathleen Nichols <nichols@pollere.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Matt Mathis <mattmathis@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Nandita Dukkipati <nanditad@google.com>
Cc: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
An implementation of CoDel AQM, from Kathleen Nichols and Van Jacobson.
http://queue.acm.org/detail.cfm?id=2209336
This AQM main input is no longer queue size in bytes or packets, but the
delay packets stay in (FIFO) queue.
As we don't have infinite memory, we still can drop packets in enqueue()
in case of massive load, but mean of CoDel is to drop packets in
dequeue(), using a control law based on two simple parameters :
target : target sojourn time (default 5ms)
interval : width of moving time window (default 100ms)
Based on initial work from Dave Taht.
Refactored to help future codel inclusion as a plugin for other linux
qdisc (FQ_CODEL, ...), like RED.
include/net/codel.h contains codel algorithm as close as possible than
Kathleen reference.
net/sched/sch_codel.c contains the linux qdisc specific glue.
Separate structures permit a memory efficient implementation of fq_codel
(to be sent as a separate work) : Each flow has its own struct
codel_vars.
timestamps are taken at enqueue() time with 1024 ns precision, allowing
a range of 2199 seconds in queue, and 100Gb links support. iproute2 uses
usec as base unit.
Selected packets are dropped, unless ECN is enabled and packets can get
ECN mark instead.
Tested from 2Mb to 10Gb speeds with no particular problems, on ixgbe and
tg3 drivers (BQL enabled).
Usage: tc qdisc ... codel [ limit PACKETS ] [ target TIME ]
[ interval TIME ] [ ecn ]
qdisc codel 10: parent 1:1 limit 2000p target 3.0ms interval 60.0ms ecn
Sent 13347099587 bytes 8815805 pkt (dropped 0, overlimits 0 requeues 0)
rate 202365Kbit 16708pps backlog 113550b 75p requeues 0
count 116 lastcount 98 ldelay 4.3ms dropping drop_next 816us
maxpacket 1514 ecn_mark 84399 drop_overlimit 0
CoDel must be seen as a base module, and should be used keeping in mind
there is still a FIFO queue. So a typical setup will probably need a
hierarchy of several qdiscs and packet classifiers to be able to meet
whatever constraints a user might have.
One possible example would be to use fq_codel, which combines Fair
Queueing and CoDel, in replacement of sfq / sfq_red.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Dave Taht <dave.taht@bufferbloat.net>
Cc: Kathleen Nichols <nichols@pollere.com>
Cc: Van Jacobson <van@pollere.net>
Cc: Tom Herbert <therbert@google.com>
Cc: Matt Mathis <mattmathis@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Eric Dumazet