| 1 |
.TH ETF 8 "05 Jul 2018" "iproute2" "Linux" |
| 2 |
.SH NAME |
| 3 |
ETF \- Earliest TxTime First (ETF) Qdisc |
| 4 |
.SH SYNOPSIS |
| 5 |
.B tc qdisc ... dev |
| 6 |
dev |
| 7 |
.B parent |
| 8 |
classid |
| 9 |
.B [ handle |
| 10 |
major: |
| 11 |
.B ] etf clockid |
| 12 |
clockid |
| 13 |
.B [ delta |
| 14 |
delta_nsecs |
| 15 |
.B ] [ deadline_mode ] |
| 16 |
.B [ offload ] |
| 17 |
|
| 18 |
.SH DESCRIPTION |
| 19 |
The ETF (Earliest TxTime First) qdisc allows applications to control |
| 20 |
the instant when a packet should be dequeued from the traffic control |
| 21 |
layer into the netdevice. If |
| 22 |
.B offload |
| 23 |
is configured and supported by the network interface card, the it will |
| 24 |
also control when packets leave the network controller. |
| 25 |
|
| 26 |
ETF achieves that by buffering packets until a configurable time |
| 27 |
before their transmission time (i.e. txtime, or deadline), which can |
| 28 |
be configured through the |
| 29 |
.B delta |
| 30 |
option. |
| 31 |
|
| 32 |
The qdisc uses a rb-tree internally so packets are always 'ordered' by |
| 33 |
their txtime and will be dequeued following the (next) earliest txtime |
| 34 |
first. |
| 35 |
|
| 36 |
It relies on the SO_TXTIME socket option and the SCM_TXTIME CMSG in |
| 37 |
each packet field to configure the behavior of time dependent sockets: |
| 38 |
the clockid to be used as a reference, if the expected mode of txtime |
| 39 |
for that socket is deadline or strict mode, and if packet drops should |
| 40 |
be reported on the socket's error queue. See |
| 41 |
.BR socket(7) |
| 42 |
for more information. |
| 43 |
|
| 44 |
The etf qdisc will drop any packets with a txtime in the past, or if a |
| 45 |
packet expires while waiting for being dequeued. |
| 46 |
|
| 47 |
This queueing discipline is intended to be used by TSN (Time Sensitive |
| 48 |
Networking) applications, and it exposes a traffic shaping functionality |
| 49 |
that is commonly documented as "Launch Time" or "Time-Based Scheduling" |
| 50 |
by vendors and the documentation of network interface controllers. |
| 51 |
|
| 52 |
ETF is meant to be installed under another qdisc that maps packet flows |
| 53 |
to traffic classes, one example is |
| 54 |
.BR mqprio(8). |
| 55 |
|
| 56 |
.SH PARAMETERS |
| 57 |
.TP |
| 58 |
clockid |
| 59 |
.br |
| 60 |
Specifies the clock to be used by qdisc's internal timer for measuring |
| 61 |
time and scheduling events. The qdisc expects that packets passing |
| 62 |
through it to be using this same |
| 63 |
.B clockid |
| 64 |
as the reference of their txtime timestamps. It will drop packets |
| 65 |
coming from sockets that do not comply with that. |
| 66 |
|
| 67 |
For more information about time and clocks on Linux, please refer |
| 68 |
to |
| 69 |
.BR time(7) |
| 70 |
and |
| 71 |
.BR clock_gettime(3). |
| 72 |
|
| 73 |
.TP |
| 74 |
delta |
| 75 |
.br |
| 76 |
After enqueueing or dequeueing a packet, the qdisc will schedule its |
| 77 |
next wake-up time for the next txtime minus this delta value. |
| 78 |
This means |
| 79 |
.B delta |
| 80 |
can be used as a fudge factor for the scheduler latency of a system. |
| 81 |
This value must be specified in nanoseconds. |
| 82 |
The default value is 0 nanoseconds. |
| 83 |
|
| 84 |
.TP |
| 85 |
deadline_mode |
| 86 |
.br |
| 87 |
When |
| 88 |
.B deadline_mode |
| 89 |
is set, the qdisc will handle txtime with a different semantics, |
| 90 |
changed from a 'strict' transmission time to a deadline. |
| 91 |
In practice, this means during the dequeue flow |
| 92 |
.BR etf(8) |
| 93 |
will set the txtime of the packet being dequeued to 'now'. |
| 94 |
The default is for this option to be disabled. |
| 95 |
|
| 96 |
.TP |
| 97 |
offload |
| 98 |
.br |
| 99 |
When |
| 100 |
.B offload |
| 101 |
is set, |
| 102 |
.BR etf(8) |
| 103 |
will try to configure the network interface so time-based transmission |
| 104 |
arbitration is enabled in the controller. This feature is commonly |
| 105 |
referred to as "Launch Time" or "Time-Based Scheduling" by the |
| 106 |
documentation of network interface controllers. |
| 107 |
The default is for this option to be disabled. |
| 108 |
|
| 109 |
.TP |
| 110 |
skip_sock_check |
| 111 |
.br |
| 112 |
.BR etf(8) |
| 113 |
currently drops any packet which does not have a socket associated with it or |
| 114 |
if the socket does not have SO_TXTIME socket option set. But, this will not |
| 115 |
work if the launchtime is set by another entity inside the kernel (e.g. some |
| 116 |
other Qdisc). Setting the skip_sock_check will skip checking for a socket |
| 117 |
associated with the packet. |
| 118 |
|
| 119 |
.SH EXAMPLES |
| 120 |
|
| 121 |
ETF is used to enforce a Quality of Service. It controls when each |
| 122 |
packets should be dequeued and transmitted, and can be used for |
| 123 |
limiting the data rate of a traffic class. To separate packets into |
| 124 |
traffic classes the user may choose |
| 125 |
.BR mqprio(8), |
| 126 |
and configure it like this: |
| 127 |
|
| 128 |
.EX |
| 129 |
# tc qdisc add dev eth0 handle 100: parent root mqprio num_tc 3 \\ |
| 130 |
map 2 2 1 0 2 2 2 2 2 2 2 2 2 2 2 2 \\ |
| 131 |
queues 1@0 1@1 2@2 \\ |
| 132 |
hw 0 |
| 133 |
.EE |
| 134 |
.P |
| 135 |
To replace the current queueing discipline by ETF in traffic class |
| 136 |
number 0, issue: |
| 137 |
.P |
| 138 |
.EX |
| 139 |
# tc qdisc replace dev eth0 parent 100:1 etf \\ |
| 140 |
clockid CLOCK_TAI delta 300000 offload |
| 141 |
.EE |
| 142 |
|
| 143 |
With the options above, etf will be configured to use CLOCK_TAI as |
| 144 |
its clockid_t, will schedule packets for 300 us before their txtime, |
| 145 |
and will enable the functionality on that in the network interface |
| 146 |
card. Deadline mode will not be configured for this mode. |
| 147 |
|
| 148 |
.SH AUTHORS |
| 149 |
Jesus Sanchez-Palencia <[email protected]> |
| 150 |
.br |
| 151 |
Vinicius Costa Gomes <[email protected]> |