diff options
author | Alexander Couzens <lynxis@fe80.eu> | 2018-01-08 09:54:58 +0100 |
---|---|---|
committer | Alexander Couzens <lynxis@fe80.eu> | 2018-02-05 16:36:26 +0100 |
commit | 17eb826a703d996d12004b68df12003a12d71421 (patch) | |
tree | a00eb172232c0e389c9b01ae4d5db6757505fa3a /package/libs/libiconv/COPYRIGHT | |
parent | e7469d51924efe8c99d21e331e89c6c851bff197 (diff) | |
download | mtk-20170518-17eb826a703d996d12004b68df12003a12d71421.zip mtk-20170518-17eb826a703d996d12004b68df12003a12d71421.tar.gz mtk-20170518-17eb826a703d996d12004b68df12003a12d71421.tar.bz2 |
ltq-atm: rewrite tx path to use IRQs
The ATM subsystem is different from the generic ethernet NICs. The ATM
subsystem requires a callback when a packet has been sent. It means a
tx skb_buff need to be used after it has sent. While the generic NIC
can fill up the TX ring and free skb_buffs if it encounter a ring buffer slot
with an already sent skbuff.
The ATM drivers need call the pop() function after it has send a
single ATM package. The ATM subsystem controls via this ways the queuing.
The ppe engine use DMA channels for read and write. Every atm_vcc has it's
own TX DMA channel and each TX DMA channel has it's own ring buffer.
The old driver had multiple issues:
- Call the subsystem callback at the beginning of tx function (ppe_send).
Didn't allowed the ATM subsystem to control the enqueued package
amount.
- Filled up the TX ring until full and fail futher
- copy or decouple the skb from all other subsystem before giving it
over to TX ring
The new tx path uses interupts.
- call the subsystem callback _after_ it was sent by hardware
- no need to copy our decouple the skb any more
- gives back control to the atm subsystem over the enqueued packages
- use an interupt for every sent atm package
Using interupts shouldn't be a problem because of the slow uplink bandwidth of
ADSL.
The speed _through_ the DSL router was always as high as it should
be, only traffic generated on the router itself were affected.
After changing to new tx path, the speed of iperf's run on the
router itself reached the same speed. The master/trunk wasn't as much
affected because of TCP optimisations (reboot-5022-gb2ea46fe236a).
The following results are taken on the remote server, which receives
the stream over the internet and the DSL line.
The sync moves between every sync a litte bit, but is so far stable
Latency / Interleave Delay: Down: Fast (0.25 ms) / Up: Fast (0.50 ms)
Data Rate: Down: 13.287 Mb/s / Up: 1.151 Mb/s
reboot-5521-g9f8d28285d without patch
[ ID] Interval Transfer Bandwidth Retr
[ 5] 0.00-10.04 sec 947 KBytes 773 Kbits/sec 0 sender
[ 5] 0.00-10.04 sec 928 KBytes 757 Kbits/sec receiver
reboot-5521-g9f8d28285d with patch
[ 5] 0.00-10.06 sec 1.16 MBytes 970 Kbits/sec 0 sender
[ 5] 0.00-10.06 sec 1.15 MBytes 959 Kbits/sec receiver
v17.01.4-239-g55c23e44f4 without patch
[ ID] Interval Transfer Bandwidth Retr
[ 5] 0.00-10.04 sec 87.4 KBytes 71.3 Kbits/sec 0 sender
[ 5] 0.00-10.04 sec 59.6 KBytes 48.7 Kbits/sec receiver
v17.01.4-239-g55c23e44f4 with patch
[ ID] Interval Transfer Bandwidth Retr
[ 5] 0.00-10.05 sec 1.18 MBytes 983 Kbits/sec 1 sender
[ 5] 0.00-10.05 sec 1.15 MBytes 959 Kbits/sec receiver
Signed-off-by: Alexander Couzens <lynxis@fe80.eu>
Diffstat (limited to 'package/libs/libiconv/COPYRIGHT')
0 files changed, 0 insertions, 0 deletions