Add flag 'FWMARK' to enable use of firewall connmarks as tin selector.
The connmark (skbuff->mark) needs to be in the range 1->tin_cnt ie.
for diffserv3 the mark needs to be 1->3.
Background
Typically CAKE uses DSCP as the basis for tin selection. DSCP values
are relatively easily changed as part of the egress path, usually with
iptables & the mangle table, ingress is more challenging. CAKE is often
used on the WAN interface of a residential gateway where passthrough of
DSCP from the ISP is either missing or set to unhelpful values thus use
of ingress DSCP values for tin selection isn't helpful in that
environment.
An approach to solving the ingress tin selection problem is to use
CAKE's understanding of tc filters. Naive tc filters could match on
source/destination port numbers and force tin selection that way, but
multiple filters don't scale particularly well as each filter must be
traversed whether it matches or not. e.g. a simple example to map 3
firewall marks to tins:
MAJOR=$( tc qdisc show dev $DEV | head -1 | awk '{print $3}' )
tc filter add dev $DEV parent $MAJOR protocol all handle 0x01 fw action skbedit priority ${MAJOR}1
tc filter add dev $DEV parent $MAJOR protocol all handle 0x02 fw action skbedit priority ${MAJOR}2
tc filter add dev $DEV parent $MAJOR protocol all handle 0x03 fw action skbedit priority ${MAJOR}3
Another option is to use eBPF cls_act with tc filters e.g.
MAJOR=$( tc qdisc show dev $DEV | head -1 | awk '{print $3}' )
tc filter add dev $DEV parent $MAJOR bpf da obj my-bpf-fwmark-to-class.o
This has the disadvantages of a) needing someone to write & maintain
the bpf program, b) a bpf toolchain to compile it and c) needing to
hardcode the major number in the bpf program so it matches the cake
instance (or forcing the cake instance to a particular major number)
since the major number cannot be passed to the bpf program via tc
command line.
As already hinted at by the previous examples, it would be helpful
to associate tins with something that survives the Internet path and
ideally allows tin selection on both egress and ingress. Netfilter's
conntrack permits setting an identifying mark on a connection which
can also be restored to an ingress packet with tc action connmark e.g.
tc filter add dev eth0 parent ffff: protocol all prio 10 u32 \
match u32 0 0 flowid 1:1 action connmark action mirred egress redirect dev ifb1
Since tc's connmark action has restored any connmark into skb->mark,
any of the previous solutions are based upon it and in one form or
another copy that mark to the skb->priority field where again CAKE
picks this up.
This change cuts out at least one of the (less intuitive &
non-scalable) middlemen and permit direct access to skb->mark.
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The current implementation scales the local alpha and beta
variables in the calculate_probability function by the same
amount for all values of drop probability below 1%.
RFC 8033 suggests using additional cases for auto-tuning
alpha and beta when the drop probability is less than 1%.
In order to add more auto-tuning cases, MAX_PROB must be
scaled by u64 instead of u32 to prevent underflow when
scaling the local alpha and beta variables in the
calculate_probability function.
Signed-off-by: Mohit P. Tahiliani <tahiliani@nitk.edu.in>
Signed-off-by: Dhaval Khandla <dhavaljkhandla26@gmail.com>
Signed-off-by: Hrishikesh Hiraskar <hrishihiraskar@gmail.com>
Signed-off-by: Manish Kumar B <bmanish15597@gmail.com>
Signed-off-by: Sachin D. Patil <sdp.sachin@gmail.com>
Signed-off-by: Leslie Monis <lesliemonis@gmail.com>
Acked-by: Dave Taht <dave.taht@gmail.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Allow users to set and dump RED flags (ECN enabled and harddrop)
on per-virtual queue basis. Validation of attributes is split
from changes to make sure we won't have to undo previous operations
when we find out configuration is invalid.
The objective is to allow changing per-Qdisc parameters without
overwriting the per-vq configured flags.
Old user space will not pass the TCA_GRED_VQ_FLAGS attribute and
per-Qdisc flags will always get propagated to the virtual queues.
New user space which wants to make use of per-vq flags should set
per-Qdisc flags to 0 and then configure per-vq flags as it
sees fit. Once per-vq flags are set per-Qdisc flags can't be
changed to non-zero. Vice versa - if the per-Qdisc flags are
non-zero the TCA_GRED_VQ_FLAGS attribute has to either be omitted
or set to the same value as per-Qdisc flags.
Update per-Qdisc parameters:
per-Qdisc | per-VQ | result
0 | 0 | all vq flags updated
0 | non-0 | error (vq flags in use)
non-0 | 0 | -- impossible --
non-0 | non-0 | all vq flags updated
Update per-VQ state (flags parameter not specified):
no change to flags
Update per-VQ state (flags parameter set):
per-Qdisc | per-VQ | result
0 | any | per-vq flags updated
non-0 | 0 | -- impossible --
non-0 | non-0 | error (per-Qdisc flags in use)
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: John Hurley <john.hurley@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently all GRED's virtual queue data is dumped in a single
array in a single attribute. This makes it pretty much impossible
to add new fields. In order to expose more detailed stats add a
new set of attributes. We can now expose the 64 bit value of bytesin
and all the mark stats which were not part of the original design.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: John Hurley <john.hurley@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Similar to 80ba92fa1a ("codel: add ce_threshold attribute")
After EDT adoption, it became easier to implement DCTCP-like CE marking.
In many cases, queues are not building in the network fabric but on
the hosts themselves.
If packets leaving fq missed their Earliest Departure Time by XXX usec,
we mark them with ECN CE. This gives a feedback (after one RTT) to
the sender to slow down and find better operating mode.
Example :
tc qd replace dev eth0 root fq ce_threshold 2.5ms
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This traffic scheduler allows traffic classes states (transmission
allowed/not allowed, in the simplest case) to be scheduled, according
to a pre-generated time sequence. This is the basis of the IEEE
802.1Qbv specification.
Example configuration:
tc qdisc replace dev enp3s0 parent root handle 100 taprio \
num_tc 3 \
map 2 2 1 0 2 2 2 2 2 2 2 2 2 2 2 2 \
queues 1@0 1@1 2@2 \
base-time 1528743495910289987 \
sched-entry S 01 300000 \
sched-entry S 02 300000 \
sched-entry S 04 300000 \
clockid CLOCK_TAI
The configuration format is similar to mqprio. The main difference is
the presence of a schedule, built by multiple "sched-entry"
definitions, each entry has the following format:
sched-entry <CMD> <GATE MASK> <INTERVAL>
The only supported <CMD> is "S", which means "SetGateStates",
following the IEEE 802.1Qbv-2015 definition (Table 8-6). <GATE MASK>
is a bitmask where each bit is a associated with a traffic class, so
bit 0 (the least significant bit) being "on" means that traffic class
0 is "active" for that schedule entry. <INTERVAL> is a time duration
in nanoseconds that specifies for how long that state defined by <CMD>
and <GATE MASK> should be held before moving to the next entry.
This schedule is circular, that is, after the last entry is executed
it starts from the first one, indefinitely.
The other parameters can be defined as follows:
- base-time: specifies the instant when the schedule starts, if
'base-time' is a time in the past, the schedule will start at
base-time + (N * cycle-time)
where N is the smallest integer so the resulting time is greater
than "now", and "cycle-time" is the sum of all the intervals of the
entries in the schedule;
- clockid: specifies the reference clock to be used;
The parameters should be similar to what the IEEE 802.1Q family of
specification defines.
Signed-off-by: Vinicius Costa Gomes <vinicius.gomes@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tokens and ctokens are defined as s64 in htb_class structure,
and clamped to 32bits value during netlink dumps:
cl->xstats.tokens = clamp_t(s64, PSCHED_NS2TICKS(cl->tokens),
INT_MIN, INT_MAX);
Defining it as u32 is working since userspace (tc) is printing it as
signed int, but a correct definition from the beginning is probably
better.
In the same time, 'giants' structure member is unused since years, so
update the comment to mark it unused.
Signed-off-by: Florent Fourcot <florent.fourcot@wifirst.fr>
Signed-off-by: David S. Miller <davem@davemloft.net>
Skbprio (SKB Priority Queue) is a queueing discipline that prioritizes packets
according to their skb->priority field. Under congestion, already-enqueued lower
priority packets will be dropped to make space available for higher priority
packets. Skbprio was conceived as a solution for denial-of-service defenses that
need to route packets with different priorities as a means to overcome DoS
attacks.
v5
*Do not reference qdisc_dev(sch)->tx_queue_len for setting limit. Instead set
default sch->limit to 64.
v4
*Drop Documentation/networking/sch_skbprio.txt doc file to move it to tc man
page for Skbprio, in iproute2.
v3
*Drop max_limit parameter in struct skbprio_sched_data and instead use
sch->limit.
*Reference qdisc_dev(sch)->tx_queue_len only once, during initialisation for
qdisc (previously being referenced every time qdisc changes).
*Move qdisc's detailed description from in-code to Documentation/networking.
*When qdisc is saturated, enqueue incoming packet first before dequeueing
lowest priority packet in queue - improves usage of call stack registers.
*Introduce and use overlimit stat to keep track of number of dropped packets.
v2
*Use skb->priority field rather than DS field. Rename queueing discipline as
SKB Priority Queue (previously Gatekeeper Priority Queue).
*Queueing discipline is made classful to expose Skbprio's internal priority
queues.
Signed-off-by: Nishanth Devarajan <ndev2021@gmail.com>
Reviewed-by: Sachin Paryani <sachin.paryani@gmail.com>
Reviewed-by: Cody Doucette <doucette@bu.edu>
Reviewed-by: Michel Machado <michel@digirati.com.br>
Acked-by: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
sch_cake targets the home router use case and is intended to squeeze the
most bandwidth and latency out of even the slowest ISP links and routers,
while presenting an API simple enough that even an ISP can configure it.
Example of use on a cable ISP uplink:
tc qdisc add dev eth0 cake bandwidth 20Mbit nat docsis ack-filter
To shape a cable download link (ifb and tc-mirred setup elided)
tc qdisc add dev ifb0 cake bandwidth 200mbit nat docsis ingress wash
CAKE is filled with:
* A hybrid Codel/Blue AQM algorithm, "Cobalt", tied to an FQ_Codel
derived Flow Queuing system, which autoconfigures based on the bandwidth.
* A novel "triple-isolate" mode (the default) which balances per-host
and per-flow FQ even through NAT.
* An deficit based shaper, that can also be used in an unlimited mode.
* 8 way set associative hashing to reduce flow collisions to a minimum.
* A reasonable interpretation of various diffserv latency/loss tradeoffs.
* Support for zeroing diffserv markings for entering and exiting traffic.
* Support for interacting well with Docsis 3.0 shaper framing.
* Extensive support for DSL framing types.
* Support for ack filtering.
* Extensive statistics for measuring, loss, ecn markings, latency
variation.
A paper describing the design of CAKE is available at
https://arxiv.org/abs/1804.07617, and will be published at the 2018 IEEE
International Symposium on Local and Metropolitan Area Networks (LANMAN).
This patch adds the base shaper and packet scheduler, while subsequent
commits add the optional (configurable) features. The full userspace API
and most data structures are included in this commit, but options not
understood in the base version will be ignored.
Various versions baking have been available as an out of tree build for
kernel versions going back to 3.10, as the embedded router world has been
running a few years behind mainline Linux. A stable version has been
generally available on lede-17.01 and later.
sch_cake replaces a combination of iptables, tc filter, htb and fq_codel
in the sqm-scripts, with sane defaults and vastly simpler configuration.
CAKE's principal author is Jonathan Morton, with contributions from
Kevin Darbyshire-Bryant, Toke Høiland-Jørgensen, Sebastian Moeller,
Ryan Mounce, Tony Ambardar, Dean Scarff, Nils Andreas Svee, Dave Täht,
and Loganaden Velvindron.
Testing from Pete Heist, Georgios Amanakis, and the many other members of
the cake@lists.bufferbloat.net mailing list.
tc -s qdisc show dev eth2
qdisc cake 8017: root refcnt 2 bandwidth 1Gbit diffserv3 triple-isolate split-gso rtt 100.0ms noatm overhead 38 mpu 84
Sent 51504294511 bytes 37724591 pkt (dropped 6, overlimits 64958695 requeues 12)
backlog 0b 0p requeues 12
memory used: 1053008b of 15140Kb
capacity estimate: 970Mbit
min/max network layer size: 28 / 1500
min/max overhead-adjusted size: 84 / 1538
average network hdr offset: 14
Bulk Best Effort Voice
thresh 62500Kbit 1Gbit 250Mbit
target 5.0ms 5.0ms 5.0ms
interval 100.0ms 100.0ms 100.0ms
pk_delay 5us 5us 6us
av_delay 3us 2us 2us
sp_delay 2us 1us 1us
backlog 0b 0b 0b
pkts 3164050 25030267 9530280
bytes 3227519915 35396974782 12879808898
way_inds 0 8 0
way_miss 21 366 25
way_cols 0 0 0
drops 5 0 1
marks 0 0 0
ack_drop 0 0 0
sp_flows 1 3 0
bk_flows 0 1 1
un_flows 0 0 0
max_len 68130 68130 68130
Tested-by: Pete Heist <peteheist@gmail.com>
Tested-by: Georgios Amanakis <gamanakis@gmail.com>
Signed-off-by: Dave Taht <dave.taht@gmail.com>
Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add infra so etf qdisc supports HW offload of time-based transmission.
For hw offload, the time sorted list is still used, so packets are
dequeued always in order of txtime.
Example:
$ tc qdisc replace dev enp2s0 parent root handle 100 mqprio num_tc 3 \
map 2 2 1 0 2 2 2 2 2 2 2 2 2 2 2 2 queues 1@0 1@1 2@2 hw 0
$ tc qdisc add dev enp2s0 parent 100:1 etf offload delta 100000 \
clockid CLOCK_REALTIME
In this example, the Qdisc will use HW offload for the control of the
transmission time through the network adapter. The hrtimer used for
packets scheduling inside the qdisc will use the clockid CLOCK_REALTIME
as reference and packets leave the Qdisc "delta" (100000) nanoseconds
before their transmission time. Because this will be using HW offload and
since dynamic clocks are not supported by the hrtimer, the system clock
and the PHC clock must be synchronized for this mode to behave as
expected.
Signed-off-by: Jesus Sanchez-Palencia <jesus.sanchez-palencia@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The ETF (Earliest TxTime First) qdisc uses the information added
earlier in this series (the socket option SO_TXTIME and the new
role of sk_buff->tstamp) to schedule packets transmission based
on absolute time.
For some workloads, just bandwidth enforcement is not enough, and
precise control of the transmission of packets is necessary.
Example:
$ tc qdisc replace dev enp2s0 parent root handle 100 mqprio num_tc 3 \
map 2 2 1 0 2 2 2 2 2 2 2 2 2 2 2 2 queues 1@0 1@1 2@2 hw 0
$ tc qdisc add dev enp2s0 parent 100:1 etf delta 100000 \
clockid CLOCK_TAI
In this example, the Qdisc will provide SW best-effort for the control
of the transmission time to the network adapter, the time stamp in the
socket will be in reference to the clockid CLOCK_TAI and packets
will leave the qdisc "delta" (100000) nanoseconds before its transmission
time.
The ETF qdisc will buffer packets sorted by their txtime. It will drop
packets on enqueue() if their skbuff clockid does not match the clock
reference of the Qdisc. Moreover, on dequeue(), a packet will be dropped
if it expires while being enqueued.
The qdisc also supports the SO_TXTIME deadline mode. For this mode, it
will dequeue a packet as soon as possible and change the skb timestamp
to 'now' during etf_dequeue().
Note that both the qdisc's and the SO_TXTIME ABIs allow for a clockid
to be configured, but it's been decided that usage of CLOCK_TAI should
be enforced until we decide to allow for other clockids to be used.
The rationale here is that PTP times are usually in the TAI scale, thus
no other clocks should be necessary. For now, the qdisc will return
EINVAL if any clocks other than CLOCK_TAI are used.
Signed-off-by: Jesus Sanchez-Palencia <jesus.sanchez-palencia@intel.com>
Signed-off-by: Vinicius Costa Gomes <vinicius.gomes@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Extend slotting with support for non-uniform distributions. This is
similar to netem's non-uniform distribution delay feature.
Commit f043efeae2f1 ("netem: support delivering packets in delayed
time slots") added the slotting feature to approximate the behaviors
of media with packet aggregation but only supported a uniform
distribution for delays between transmission attempts. Tests with TCP
BBR with emulated wifi links with non-uniform distributions produced
more useful results.
Syntax:
slot dist DISTRIBUTION DELAY JITTER [packets MAX_PACKETS] \
[bytes MAX_BYTES]
The syntax and use of the distribution table is the same as in the
non-uniform distribution delay feature. A file DISTRIBUTION must be
present in TC_LIB_DIR (e.g. /usr/lib/tc) containing numbers scaled by
NETEM_DIST_SCALE. A random value x is selected from the table and it
takes DELAY + ( x * JITTER ) as delay. Correlation between values is not
supported.
Examples:
Normal distribution delay with mean = 800us and stdev = 100us.
> tc qdisc add dev eth0 root netem slot dist normal 800us 100us
Optionally set the max slot size in bytes and/or packets.
> tc qdisc add dev eth0 root netem slot dist normal 800us 100us \
bytes 64k packets 42
Signed-off-by: Yousuk Seung <ysseung@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Following the previous patch, RED is now using the new uniform uapi
for indicating it's offloaded. As a result, TC_RED_OFFLOADED is no
longer utilized by kernel and can be removed [as it's still not
part of any stable release].
Fixes: 602f3baf22 ("net_sch: red: Add offload ability to RED qdisc")
Signed-off-by: Yuval Mintz <yuvalm@mellanox.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Slotting is a crude approximation of the behaviors of shared media such
as cable, wifi, and LTE, which gather up a bunch of packets within a
varying delay window and deliver them, relative to that, nearly all at
once.
It works within the existing loss, duplication, jitter and delay
parameters of netem. Some amount of inherent latency must be specified,
regardless.
The new "slot" parameter specifies a minimum and maximum delay between
transmission attempts.
The "bytes" and "packets" parameters can be used to limit the amount of
information transferred per slot.
Examples of use:
tc qdisc add dev eth0 root netem delay 200us \
slot 800us 10ms bytes 64k packets 42
A more correct example, using stacked netem instances and a packet limit
to emulate a tail drop wifi queue with slots and variable packet
delivery, with a 200Mbit isochronous underlying rate, and 20ms path
delay:
tc qdisc add dev eth0 root handle 1: netem delay 20ms rate 200mbit \
limit 10000
tc qdisc add dev eth0 parent 1:1 handle 10:1 netem delay 200us \
slot 800us 10ms bytes 64k packets 42 limit 512
Signed-off-by: Dave Taht <dave.taht@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
netem userspace has long relied on a horrible /proc/net/psched hack
to translate the current notion of "ticks" to nanoseconds.
Expressing latency and jitter instead, in well defined nanoseconds,
increases the dynamic range of emulated delays and jitter in netem.
It will also ease a transition where reducing a tick to nsec
equivalence would constrain the max delay in prior versions of
netem to only 4.3 seconds.
Signed-off-by: Dave Taht <dave.taht@gmail.com>
Suggested-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add the ability to offload RED qdisc by using ndo_setup_tc.
There are four commands for RED offloading:
* TC_RED_SET: handles set and change.
* TC_RED_DESTROY: handle qdisc destroy.
* TC_RED_STATS: update the qdiscs counters (given as reference)
* TC_RED_XSTAT: returns red xstats.
Whether RED is being offloaded is being determined every time dump action
is being called because parent change of this qdisc could change its
offload state but doesn't require any RED function to be called.
Signed-off-by: Nogah Frankel <nogahf@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Reviewed-by: Simon Horman <simon.horman@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Files removed in 'net-next' had their license header updated
in 'net'. We take the remove from 'net-next'.
Signed-off-by: David S. Miller <davem@davemloft.net>
Many user space API headers are missing licensing information, which
makes it hard for compliance tools to determine the correct license.
By default are files without license information under the default
license of the kernel, which is GPLV2. Marking them GPLV2 would exclude
them from being included in non GPLV2 code, which is obviously not
intended. The user space API headers fall under the syscall exception
which is in the kernels COPYING file:
NOTE! This copyright does *not* cover user programs that use kernel
services by normal system calls - this is merely considered normal use
of the kernel, and does *not* fall under the heading of "derived work".
otherwise syscall usage would not be possible.
Update the files which contain no license information with an SPDX
license identifier. The chosen identifier is 'GPL-2.0 WITH
Linux-syscall-note' which is the officially assigned identifier for the
Linux syscall exception. SPDX license identifiers are a legally binding
shorthand, which can be used instead of the full boiler plate text.
This patch is based on work done by Thomas Gleixner and Kate Stewart and
Philippe Ombredanne. See the previous patch in this series for the
methodology of how this patch was researched.
Reviewed-by: Kate Stewart <kstewart@linuxfoundation.org>
Reviewed-by: Philippe Ombredanne <pombredanne@nexb.com>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
This queueing discipline implements the shaper algorithm defined by
the 802.1Q-2014 Section 8.6.8.2 and detailed in Annex L.
It's primary usage is to apply some bandwidth reservation to user
defined traffic classes, which are mapped to different queues via the
mqprio qdisc.
Only a simple software implementation is added for now.
Signed-off-by: Vinicius Costa Gomes <vinicius.gomes@intel.com>
Signed-off-by: Jesus Sanchez-Palencia <jesus.sanchez-palencia@intel.com>
Tested-by: Henrik Austad <henrik@austad.us>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This patch makes a slight tweak to mqprio in order to bring the
classid values used back in line with what is used for mq. The general idea
is to reserve values :ffe0 - :ffef to identify hardware traffic classes
normally reported via dev->num_tc. By doing this we can maintain a
consistent behavior with mq for classid where :1 - :ffdf will represent a
physical qdisc mapped onto a Tx queue represented by classid - 1, and the
traffic classes will be mapped onto a known subset of classid values
reserved for our virtual qdiscs.
Note I reserved the range from :fff0 - :ffff since this way we might be
able to reuse these classid values with clsact and ingress which would mean
that for mq, mqprio, ingress, and clsact we should be able to maintain a
similar classid layout.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Jesus Sanchez-Palencia <jesus.sanchez-palencia@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The offload types currently supported in mqprio are 0 (no offload) and
1 (offload only TCs) by setting these values for the 'hw' option. If
offloads are supported by setting the 'hw' option to 1, the default
offload mode is 'dcb' where only the TC values are offloaded to the
device. This patch introduces a new hardware offload mode called
'channel' with 'hw' set to 1 in mqprio which makes full use of the
mqprio options, the TCs, the queue configurations and the QoS parameters
for the TCs. This is achieved through a new netlink attribute for the
'mode' option which takes values such as 'dcb' (default) and 'channel'.
The 'channel' mode also supports QoS attributes for traffic class such as
minimum and maximum values for bandwidth rate limits.
This patch enables configuring additional HW shaper attributes associated
with a traffic class. Currently the shaper for bandwidth rate limiting is
supported which takes options such as minimum and maximum bandwidth rates
and are offloaded to the hardware in the 'channel' mode. The min and max
limits for bandwidth rates are provided by the user along with the TCs
and the queue configurations when creating the mqprio qdisc. The interface
can be extended to support new HW shapers in future through the 'shaper'
attribute.
Introduces a new data structure 'tc_mqprio_qopt_offload' for offloading
mqprio queue options and use this to be shared between the kernel and
device driver. This contains a copy of the existing data structure
for mqprio queue options. This new data structure can be extended when
adding new attributes for traffic class such as mode, shaper, shaper
parameters (bandwidth rate limits). The existing data structure for mqprio
queue options will be shared between the kernel and userspace.
Example:
queues 4@0 4@4 hw 1 mode channel shaper bw_rlimit\
min_rate 1Gbit 2Gbit max_rate 4Gbit 5Gbit
To dump the bandwidth rates:
qdisc mqprio 804a: root tc 2 map 0 0 0 0 1 1 1 1 0 0 0 0 0 0 0 0
queues:(0:3) (4:7)
mode:channel
shaper:bw_rlimit min_rate:1Gbit 2Gbit max_rate:4Gbit 5Gbit
Signed-off-by: Amritha Nambiar <amritha.nambiar@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This patch is meant to allow for support of multiple hardware offload type
for a single device. There is currently no bounds checking for the hw
member of the mqprio_qopt structure. This results in us being able to pass
values from 1 to 255 with all being treated the same. On retreiving the
value it is returned as 1 for anything 1 or greater being set.
With this change we are currently adding limited bounds checking by
defining an enum and using those values to limit the reported hardware
offloads.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
It looks like the following patch can make FQ very precise, even in VM
or stressed hosts. It matters at high pacing rates.
We take into account the difference between the time that was programmed
when last packet was sent, and current time (a drift of tens of usecs is
often observed)
Add an EWMA of the unthrottle latency to help diagnostics.
This latency is the difference between current time and oldest packet in
delayed RB-tree. This accounts for the high resolution timer latency,
but can be different under stress, as fq_check_throttled() can be
opportunistically be called from a dequeue() called after an enqueue()
for a different flow.
Tested:
// Start a 10Gbit flow
$ netperf --google-pacing-rate 1250000000 -H lpaa24 -l 10000 -- -K bbr &
Before patch :
$ sar -n DEV 10 5 | grep eth0 | grep Average
Average: eth0 17106.04 756876.84 1102.75 1119049.02 0.00 0.00 0.52
After patch :
$ sar -n DEV 10 5 | grep eth0 | grep Average
Average: eth0 17867.00 800245.90 1151.77 1183172.12 0.00 0.00 0.52
A new iproute2 tc can output the 'unthrottle latency' :
$ tc -s qd sh dev eth0 | grep latency
0 gc, 0 highprio, 32490767 throttled, 2382 ns latency
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This commit adds to the fq module a low_rate_threshold parameter to
insert a delay after all packets if the socket requests a pacing rate
below the threshold.
This helps achieve more precise control of the sending rate with
low-rate paths, especially policers. The basic issue is that if a
congestion control module detects a policer at a certain rate, it may
want fq to be able to shape to that policed rate. That way the sender
can avoid policer drops by having the packets arrive at the policer at
or just under the policed rate.
The default threshold of 550Kbps was chosen analytically so that for
policers or links at 500Kbps or 512Kbps fq would very likely invoke
this mechanism, even if the pacing rate was briefly slightly above the
available bandwidth. This value was then empirically validated with
two years of production testing on YouTube video servers.
Signed-off-by: Van Jacobson <vanj@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Nandita Dukkipati <nanditad@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
On small embedded routers, one wants to control maximal amount of
memory used by fq_codel, instead of controlling number of packets or
bytes, since GRO/TSO make these not practical.
Assuming skb->truesize is accurate, we have to keep track of
skb->truesize sum for skbs in queue.
This patch adds a new TCA_FQ_CODEL_MEMORY_LIMIT attribute.
I chose a default value of 32 MBytes, which looks reasonable even
for heavy duty usages. (Prior fq_codel users should not be hurt
when they upgrade their kernels)
Two fields are added to tc_fq_codel_qd_stats to report :
- Current memory usage
- Number of drops caused by memory limits
# tc qd replace dev eth1 root est 1sec 4sec fq_codel memory_limit 4M
..
# tc -s -d qd sh dev eth1
qdisc fq_codel 8008: root refcnt 257 limit 10240p flows 1024
quantum 1514 target 5.0ms interval 100.0ms memory_limit 4Mb ecn
Sent 2083566791363 bytes 1376214889 pkt (dropped 4994406, overlimits 0
requeues 21705223)
rate 9841Mbit 812549pps backlog 3906120b 376p requeues 21705223
maxpacket 68130 drop_overlimit 4994406 new_flow_count 28855414
ecn_mark 0 memory_used 4190048 drop_overmemory 4994406
new_flows_len 1 old_flows_len 177
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Dave Täht <dave.taht@gmail.com>
Cc: Sebastian Möller <moeller0@gmx.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
In presence of inelastic flows and stress, we can call
fq_codel_drop() for every packet entering fq_codel qdisc.
fq_codel_drop() is quite expensive, as it does a linear scan
of 4 KB of memory to find a fat flow.
Once found, it drops the oldest packet of this flow.
Instead of dropping a single packet, try to drop 50% of the backlog
of this fat flow, with a configurable limit of 64 packets per round.
TCA_FQ_CODEL_DROP_BATCH_SIZE is the new attribute to make this
limit configurable.
With this strategy the 4 KB search is amortized to a single cache line
per drop [1], so fq_codel_drop() no longer appears at the top of kernel
profile in presence of few inelastic flows.
[1] Assuming a 64byte cache line, and 1024 buckets
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Dave Taht <dave.taht@gmail.com>
Cc: Jonathan Morton <chromatix99@gmail.com>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: Dave Taht
Signed-off-by: David S. Miller <davem@davemloft.net>
This work adds a generalization of the ingress qdisc as a qdisc holding
only classifiers. The clsact qdisc works on ingress, but also on egress.
In both cases, it's execution happens without taking the qdisc lock, and
the main difference for the egress part compared to prior version of [1]
is that this can be applied with _any_ underlying real egress qdisc (also
classless ones).
Besides solving the use-case of [1], that is, allowing for more programmability
on assigning skb->priority for the mqprio case that is supported by most
popular 10G+ NICs, it also opens up a lot more flexibility for other tc
applications. The main work on classification can already be done at clsact
egress time if the use-case allows and state stored for later retrieval
f.e. again in skb->priority with major/minors (which is checked by most
classful qdiscs before consulting tc_classify()) and/or in other skb fields
like skb->tc_index for some light-weight post-processing to get to the
eventual classid in case of a classful qdisc. Another use case is that
the clsact egress part allows to have a central egress counterpart to
the ingress classifiers, so that classifiers can easily share state (e.g.
in cls_bpf via eBPF maps) for ingress and egress.
Currently, default setups like mq + pfifo_fast would require for this to
use, for example, prio qdisc instead (to get a tc_classify() run) and to
duplicate the egress classifier for each queue. With clsact, it allows
for leaving the setup as is, it can additionally assign skb->priority to
put the skb in one of pfifo_fast's bands and it can share state with maps.
Moreover, we can access the skb's dst entry (f.e. to retrieve tclassid)
w/o the need to perform a skb_dst_force() to hold on to it any longer. In
lwt case, we can also use this facility to setup dst metadata via cls_bpf
(bpf_skb_set_tunnel_key()) without needing a real egress qdisc just for
that (case of IFF_NO_QUEUE devices, for example).
The realization can be done without any changes to the scheduler core
framework. All it takes is that we have two a-priori defined minors/child
classes, where we can mux between ingress and egress classifier list
(dev->ingress_cl_list and dev->egress_cl_list, latter stored close to
dev->_tx to avoid extra cacheline miss for moderate loads). The egress
part is a bit similar modelled to handle_ing() and patched to a noop in
case the functionality is not used. Both handlers are now called
sch_handle_ingress() and sch_handle_egress(), code sharing among the two
doesn't seem practical as there are various minor differences in both
paths, so that making them conditional in a single handler would rather
slow things down.
Full compatibility to ingress qdisc is provided as well. Since both
piggyback on TC_H_CLSACT, only one of them (ingress/clsact) can exist
per netdevice, and thus ingress qdisc specific behaviour can be retained
for user space. This means, either a user does 'tc qdisc add dev foo ingress'
and configures ingress qdisc as usual, or the 'tc qdisc add dev foo clsact'
alternative, where both, ingress and egress classifier can be configured
as in the below example. ingress qdisc supports attaching classifier to any
minor number whereas clsact has two fixed minors for muxing between the
lists, therefore to not break user space setups, they are better done as
two separate qdiscs.
I decided to extend the sch_ingress module with clsact functionality so
that commonly used code can be reused, the module is being aliased with
sch_clsact so that it can be auto-loaded properly. Alternative would have been
to add a flag when initializing ingress to alter its behaviour plus aliasing
to a different name (as it's more than just ingress). However, the first would
end up, based on the flag, choosing the new/old behaviour by calling different
function implementations to handle each anyway, the latter would require to
register ingress qdisc once again under different alias. So, this really begs
to provide a minimal, cleaner approach to have Qdisc_ops and Qdisc_class_ops
by its own that share callbacks used by both.
Example, adding qdisc:
# tc qdisc add dev foo clsact
# tc qdisc show dev foo
qdisc mq 0: root
qdisc pfifo_fast 0: parent :1 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1
qdisc pfifo_fast 0: parent :2 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1
qdisc pfifo_fast 0: parent :3 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1
qdisc pfifo_fast 0: parent :4 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1
qdisc clsact ffff: parent ffff:fff1
Adding filters (deleting, etc works analogous by specifying ingress/egress):
# tc filter add dev foo ingress bpf da obj bar.o sec ingress
# tc filter add dev foo egress bpf da obj bar.o sec egress
# tc filter show dev foo ingress
filter protocol all pref 49152 bpf
filter protocol all pref 49152 bpf handle 0x1 bar.o:[ingress] direct-action
# tc filter show dev foo egress
filter protocol all pref 49152 bpf
filter protocol all pref 49152 bpf handle 0x1 bar.o:[egress] direct-action
A 'tc filter show dev foo' or 'tc filter show dev foo parent ffff:' will
show an empty list for clsact. Either using the parent names (ingress/egress)
or specifying the full major/minor will then show the related filter lists.
Prior work on a mqprio prequeue() facility [1] was done mainly by John Fastabend.
[1] http://patchwork.ozlabs.org/patch/512949/
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: John Fastabend <john.r.fastabend@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In a GRED qdisc, if the default "virtual queue" (VQ) does not have drop
parameters configured, then packets for the default VQ are not subjected
to RED and are only dropped if the queue is larger than the net_device's
tx_queue_len. This behavior is useful for WRED mode, since these packets
will still influence the calculated average queue length and (therefore)
the drop probability for all of the other VQs. However, for some drivers
tx_queue_len is zero. In other cases the user may wish to make the limit
the same for all VQs (including the default VQ with no drop parameters).
This change adds a TCA_GRED_LIMIT attribute to set the GRED queue limit,
in bytes, during qdisc setup. (This limit is in bytes to be consistent
with the drop parameters.) The default limit is the same as for a bfifo
queue (tx_queue_len * psched_mtu). If the drop parameters of any VQ are
configured with a smaller limit than the GRED queue limit, that VQ will
still observe the smaller limit instead.
Signed-off-by: David Ward <david.ward@ll.mit.edu>
Signed-off-by: David S. Miller <davem@davemloft.net>
For DCTCP or similar ECN based deployments on fabrics with shallow
buffers, hosts are responsible for a good part of the buffering.
This patch adds an optional ce_threshold to codel & fq_codel qdiscs,
so that DCTCP can have feedback from queuing in the host.
A DCTCP enabled egress port simply have a queue occupancy threshold
above which ECT packets get CE mark.
In codel language this translates to a sojourn time, so that one doesn't
have to worry about bytes or bandwidth but delays.
This makes the host an active participant in the health of the whole
network.
This also helps experimenting DCTCP in a setup without DCTCP compliant
fabric.
On following example, ce_threshold is set to 1ms, and we can see from
'ldelay xxx us' that TCP is not trying to go around the 5ms codel
target.
Queue has more capacity to absorb inelastic bursts (say from UDP
traffic), as queues are maintained to an optimal level.
lpaa23:~# ./tc -s -d qd sh dev eth1
qdisc mq 1: dev eth1 root
Sent 87910654696 bytes 58065331 pkt (dropped 0, overlimits 0 requeues 42961)
backlog 3108242b 364p requeues 42961
qdisc codel 8063: dev eth1 parent 1:1 limit 1000p target 5.0ms ce_threshold 1.0ms interval 100.0ms
Sent 7363778701 bytes 4863809 pkt (dropped 0, overlimits 0 requeues 5503)
rate 2348Mbit 193919pps backlog 255866b 46p requeues 5503
count 0 lastcount 0 ldelay 1.0ms drop_next 0us
maxpacket 68130 ecn_mark 0 drop_overlimit 0 ce_mark 72384
qdisc codel 8064: dev eth1 parent 1:2 limit 1000p target 5.0ms ce_threshold 1.0ms interval 100.0ms
Sent 7636486190 bytes 5043942 pkt (dropped 0, overlimits 0 requeues 5186)
rate 2319Mbit 191538pps backlog 207418b 64p requeues 5186
count 0 lastcount 0 ldelay 694us drop_next 0us
maxpacket 68130 ecn_mark 0 drop_overlimit 0 ce_mark 69873
qdisc codel 8065: dev eth1 parent 1:3 limit 1000p target 5.0ms ce_threshold 1.0ms interval 100.0ms
Sent 11569360142 bytes 7641602 pkt (dropped 0, overlimits 0 requeues 5554)
rate 3041Mbit 251096pps backlog 210446b 59p requeues 5554
count 0 lastcount 0 ldelay 889us drop_next 0us
maxpacket 68130 ecn_mark 0 drop_overlimit 0 ce_mark 37780
...
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Florian Westphal <fw@strlen.de>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Glenn Judd <glenn.judd@morganstanley.com>
Cc: Nandita Dukkipati <nanditad@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
FQ has a fast path for skb attached to a socket, as it does not
have to compute a flow hash. But for other packets, FQ being non
stochastic means that hosts exposed to random Internet traffic
can allocate million of flows structure (104 bytes each) pretty
easily. Not only host can OOM, but lookup in RB trees can take
too much cpu and memory resources.
This patch adds a new attribute, orphan_mask, that is adding
possibility of having a stochastic hash for orphaned skb.
Its default value is 1024 slots, to mimic SFQ behavior.
Note: This does not apply to locally generated TCP traffic,
and no locally generated traffic will share a flow structure
with another perfect or stochastic flow.
This patch also handles the specific case of SYNACK messages:
They are attached to the listener socket, and therefore all map
to a single hash bucket. If listener have set SO_MAX_PACING_RATE,
hoping to have new accepted socket inherit this rate, SYNACK
might be paced and even dropped.
This is very similar to an internal patch Google have used more
than one year.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Proportional Integral controller Enhanced (PIE) is a scheduler to address the
bufferbloat problem.
>From the IETF draft below:
" Bufferbloat is a phenomenon where excess buffers in the network cause high
latency and jitter. As more and more interactive applications (e.g. voice over
IP, real time video streaming and financial transactions) run in the Internet,
high latency and jitter degrade application performance. There is a pressing
need to design intelligent queue management schemes that can control latency and
jitter; and hence provide desirable quality of service to users.
We present here a lightweight design, PIE(Proportional Integral controller
Enhanced) that can effectively control the average queueing latency to a target
value. Simulation results, theoretical analysis and Linux testbed results have
shown that PIE can ensure low latency and achieve high link utilization under
various congestion situations. The design does not require per-packet
timestamp, so it incurs very small overhead and is simple enough to implement
in both hardware and software. "
Many thanks to Dave Taht for extensive feedback, reviews, testing and
suggestions. Thanks also to Stephen Hemminger and Eric Dumazet for reviews and
suggestions. Naeem Khademi and Dave Taht independently contributed to ECN
support.
For more information, please see technical paper about PIE in the IEEE
Conference on High Performance Switching and Routing 2013. A copy of the paper
can be found at ftp://ftpeng.cisco.com/pie/.
Please also refer to the IETF draft submission at
http://tools.ietf.org/html/draft-pan-tsvwg-pie-00
All relevant code, documents and test scripts and results can be found at
ftp://ftpeng.cisco.com/pie/.
For problems with the iproute2/tc or Linux kernel code, please contact Vijay
Subramanian (vijaynsu@cisco.com or subramanian.vijay@gmail.com) Mythili Prabhu
(mysuryan@cisco.com)
Signed-off-by: Vijay Subramanian <subramanian.vijay@gmail.com>
Signed-off-by: Mythili Prabhu <mysuryan@cisco.com>
CC: Dave Taht <dave.taht@bufferbloat.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add a new attribute to support 64bit rates so that
tc can use them to break the 32bit limit.
Signed-off-by: Yang Yingliang <yangyingliang@huawei.com>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
When we set burst to 1514 with low rate in userspace,
the kernel get a value of burst that less than 1514,
which doesn't work.
Because it may make some loss when transform burst
to buffer in userspace. This makes burst lose some
bytes, when the kernel transform the buffer back to
burst.
This patch adds two new attributes to support sending
burst/mtu to kernel directly to avoid the loss.
Signed-off-by: Yang Yingliang <yangyingliang@huawei.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch implements the first size-based qdisc that attempts to
differentiate between small flows and heavy-hitters. The goal is to
catch the heavy-hitters and move them to a separate queue with less
priority so that bulk traffic does not affect the latency of critical
traffic. Currently "less priority" means less weight (2:1 in
particular) in a Weighted Deficit Round Robin (WDRR) scheduler.
In essence, this patch addresses the "delay-bloat" problem due to
bloated buffers. In some systems, large queues may be necessary for
obtaining CPU efficiency, or due to the presence of unresponsive
traffic like UDP, or just a large number of connections with each
having a small amount of outstanding traffic. In these circumstances,
HHF aims to reduce the HoL blocking for latency sensitive traffic,
while not impacting the queues built up by bulk traffic. HHF can also
be used in conjunction with other AQM mechanisms such as CoDel.
To capture heavy-hitters, we implement the "multi-stage filter" design
in the following paper:
C. Estan and G. Varghese, "New Directions in Traffic Measurement and
Accounting", in ACM SIGCOMM, 2002.
Some configurable qdisc settings through 'tc':
- hhf_reset_timeout: period to reset counter values in the multi-stage
filter (default 40ms)
- hhf_admit_bytes: threshold to classify heavy-hitters
(default 128KB)
- hhf_evict_timeout: threshold to evict idle heavy-hitters
(default 1s)
- hhf_non_hh_weight: Weighted Deficit Round Robin (WDRR) weight for
non-heavy-hitters (default 2)
- hh_flows_limit: max number of heavy-hitter flow entries
(default 2048)
Note that the ratio between hhf_admit_bytes and hhf_reset_timeout
reflects the bandwidth of heavy-hitters that we attempt to capture
(25Mbps with the above default settings).
The false negative rate (heavy-hitter flows getting away unclassified)
is zero by the design of the multi-stage filter algorithm.
With 100 heavy-hitter flows, using four hashes and 4000 counters yields
a false positive rate (non-heavy-hitters mistakenly classified as
heavy-hitters) of less than 1e-4.
Signed-off-by: Terry Lam <vtlam@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
For performance reasons, sch_fq tried hard to not setup timers for every
sent packet, using a quantum based heuristic : A delay is setup only if
the flow exhausted its credit.
Problem is that application limited flows can refill their credit
for every queued packet, and they can evade pacing.
This problem can also be triggered when TCP flows use small MSS values,
as TSO auto sizing builds packets that are smaller than the default fq
quantum (3028 bytes)
This patch adds a 40 ms delay to guard flow credit refill.
Fixes: afe4fd0624 ("pkt_sched: fq: Fair Queue packet scheduler")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Maciej Żenczykowski <maze@google.com>
Cc: Willem de Bruijn <willemb@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit 7eec4174ff ("pkt_sched: fq: fix non TCP flows pacing")
obsoleted TCA_FQ_FLOW_DEFAULT_RATE without notice for the users.
Suggested by David Miller
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
With psched_ratecfg_precompute(), tbf can deal with 64bit rates.
Add two new attributes so that tc can use them to break the 32bit
limit.
Signed-off-by: Yang Yingliang <yangyingliang@huawei.com>
Suggested-by: Sergei Shtylyov <sergei.shtylyov@cogentembedded.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David 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: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
- Uses perfect flow match (not stochastic hash like SFQ/FQ_codel)
- Uses the new_flow/old_flow separation from FQ_codel
- New flows get an initial credit allowing IW10 without added delay.
- Special FIFO queue for high prio packets (no need for PRIO + FQ)
- Uses a hash table of RB trees to locate the flows at enqueue() time
- Smart on demand gc (at enqueue() time, RB tree lookup evicts old
unused flows)
- Dynamic memory allocations.
- Designed to allow millions of concurrent flows per Qdisc.
- Small memory footprint : ~8K per Qdisc, and 104 bytes per flow.
- Single high resolution timer for throttled flows (if any).
- One RB tree to link throttled flows.
- Ability to have a max rate per flow. We might add a socket option
to add per socket limitation.
Attempts have been made to add TCP pacing in TCP stack, but this
seems to add complex code to an already complex stack.
TCP pacing is welcomed for flows having idle times, as the cwnd
permits TCP stack to queue a possibly large number of packets.
This removes the 'slow start after idle' choice, hitting badly
large BDP flows, and applications delivering chunks of data
as video streams.
Nicely spaced packets :
Here interface is 10Gbit, but flow bottleneck is ~20Mbit
cwin is big, yet FQ avoids the typical bursts generated by TCP
(as in netperf TCP_RR -- -r 100000,100000)
15:01:23.545279 IP A > B: . 78193:81089(2896) ack 65248 win 3125 <nop,nop,timestamp 1115 11597805>
15:01:23.545394 IP B > A: . ack 81089 win 3668 <nop,nop,timestamp 11597985 1115>
15:01:23.546488 IP A > B: . 81089:83985(2896) ack 65248 win 3125 <nop,nop,timestamp 1115 11597805>
15:01:23.546565 IP B > A: . ack 83985 win 3668 <nop,nop,timestamp 11597986 1115>
15:01:23.547713 IP A > B: . 83985:86881(2896) ack 65248 win 3125 <nop,nop,timestamp 1115 11597805>
15:01:23.547778 IP B > A: . ack 86881 win 3668 <nop,nop,timestamp 11597987 1115>
15:01:23.548911 IP A > B: . 86881:89777(2896) ack 65248 win 3125 <nop,nop,timestamp 1115 11597805>
15:01:23.548949 IP B > A: . ack 89777 win 3668 <nop,nop,timestamp 11597988 1115>
15:01:23.550116 IP A > B: . 89777:92673(2896) ack 65248 win 3125 <nop,nop,timestamp 1115 11597805>
15:01:23.550182 IP B > A: . ack 92673 win 3668 <nop,nop,timestamp 11597989 1115>
15:01:23.551333 IP A > B: . 92673:95569(2896) ack 65248 win 3125 <nop,nop,timestamp 1115 11597805>
15:01:23.551406 IP B > A: . ack 95569 win 3668 <nop,nop,timestamp 11597991 1115>
15:01:23.552539 IP A > B: . 95569:98465(2896) ack 65248 win 3125 <nop,nop,timestamp 1115 11597805>
15:01:23.552576 IP B > A: . ack 98465 win 3668 <nop,nop,timestamp 11597992 1115>
15:01:23.553756 IP A > B: . 98465:99913(1448) ack 65248 win 3125 <nop,nop,timestamp 1115 11597805>
15:01:23.554138 IP A > B: P 99913:100001(88) ack 65248 win 3125 <nop,nop,timestamp 1115 11597805>
15:01:23.554204 IP B > A: . ack 100001 win 3668 <nop,nop,timestamp 11597993 1115>
15:01:23.554234 IP B > A: . 65248:68144(2896) ack 100001 win 3668 <nop,nop,timestamp 11597993 1115>
15:01:23.555620 IP B > A: . 68144:71040(2896) ack 100001 win 3668 <nop,nop,timestamp 11597993 1115>
15:01:23.557005 IP B > A: . 71040:73936(2896) ack 100001 win 3668 <nop,nop,timestamp 11597993 1115>
15:01:23.558390 IP B > A: . 73936:76832(2896) ack 100001 win 3668 <nop,nop,timestamp 11597993 1115>
15:01:23.559773 IP B > A: . 76832:79728(2896) ack 100001 win 3668 <nop,nop,timestamp 11597993 1115>
15:01:23.561158 IP B > A: . 79728:82624(2896) ack 100001 win 3668 <nop,nop,timestamp 11597994 1115>
15:01:23.562543 IP B > A: . 82624:85520(2896) ack 100001 win 3668 <nop,nop,timestamp 11597994 1115>
15:01:23.563928 IP B > A: . 85520:88416(2896) ack 100001 win 3668 <nop,nop,timestamp 11597994 1115>
15:01:23.565313 IP B > A: . 88416:91312(2896) ack 100001 win 3668 <nop,nop,timestamp 11597994 1115>
15:01:23.566698 IP B > A: . 91312:94208(2896) ack 100001 win 3668 <nop,nop,timestamp 11597994 1115>
15:01:23.568083 IP B > A: . 94208:97104(2896) ack 100001 win 3668 <nop,nop,timestamp 11597994 1115>
15:01:23.569467 IP B > A: . 97104:100000(2896) ack 100001 win 3668 <nop,nop,timestamp 11597994 1115>
15:01:23.570852 IP B > A: . 100000:102896(2896) ack 100001 win 3668 <nop,nop,timestamp 11597994 1115>
15:01:23.572237 IP B > A: . 102896:105792(2896) ack 100001 win 3668 <nop,nop,timestamp 11597994 1115>
15:01:23.573639 IP B > A: . 105792:108688(2896) ack 100001 win 3668 <nop,nop,timestamp 11597994 1115>
15:01:23.575024 IP B > A: . 108688:111584(2896) ack 100001 win 3668 <nop,nop,timestamp 11597994 1115>
15:01:23.576408 IP B > A: . 111584:114480(2896) ack 100001 win 3668 <nop,nop,timestamp 11597994 1115>
15:01:23.577793 IP B > A: . 114480:117376(2896) ack 100001 win 3668 <nop,nop,timestamp 11597994 1115>
TCP timestamps show that most packets from B were queued in the same ms
timeframe (TSval 1159799{3,4}), but FQ managed to send them right
in time to avoid a big burst.
In slow start or steady state, very few packets are throttled [1]
FQ gets a bunch of tunables as :
limit : max number of packets on whole Qdisc (default 10000)
flow_limit : max number of packets per flow (default 100)
quantum : the credit per RR round (default is 2 MTU)
initial_quantum : initial credit for new flows (default is 10 MTU)
maxrate : max per flow rate (default : unlimited)
buckets : number of RB trees (default : 1024) in hash table.
(consumes 8 bytes per bucket)
[no]pacing : disable/enable pacing (default is enable)
All of them can be changed on a live qdisc.
$ tc qd add dev eth0 root fq help
Usage: ... fq [ limit PACKETS ] [ flow_limit PACKETS ]
[ quantum BYTES ] [ initial_quantum BYTES ]
[ maxrate RATE ] [ buckets NUMBER ]
[ [no]pacing ]
$ tc -s -d qd
qdisc fq 8002: dev eth0 root refcnt 32 limit 10000p flow_limit 100p buckets 256 quantum 3028 initial_quantum 15140
Sent 216532416 bytes 148395 pkt (dropped 0, overlimits 0 requeues 14)
backlog 0b 0p requeues 14
511 flows, 511 inactive, 0 throttled
110 gc, 0 highprio, 0 retrans, 1143 throttled, 0 flows_plimit
[1] Except if initial srtt is overestimated, as if using
cached srtt in tcp metrics. We'll provide a fix for this issue.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
commit 56b765b79 ("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 56b765b79 ("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: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
HTB uses an internal pfifo queue, which limit is not reported
to userland tools (tc), and value inherited from device tx_queue_len
at setup time.
Introduce TCA_HTB_DIRECT_QLEN attribute to allow finer control.
Remove two obsolete pr_err() calls as well.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: David Howells <dhowells@redhat.com>
Acked-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Michael Kerrisk <mtk.manpages@gmail.com>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Acked-by: Dave Jones <davej@redhat.com>
