07.03.2018 13:29, Stanislaw Gruszka:

I forgot to attach the patches, do it now.

On Wed, Mar 07, 2018 at 01:27:01PM +0100, Stanislaw Gruszka wrote:

quoted

On Thu, Mar 01, 2018 at 04:30:10PM +0100, Daniel Golle wrote:

quoted

[forwarding to all other involved players]

On Thu, Mar 01, 2018 at 05:50:51PM +0300, Jamie Stuart wrote:

quoted

Hi Daniel,
The driver seems much improved after this fix.

it's about those two
[PATCH 1/2] rt2x00: pause almost full queue early
[PATCH 2/2] rt2x00: do not pause queue unconditionally on error path

quoted

Under very heavy load (30 clients downloading multi-GB files from SD card on the server concurrently), wifi dies with errors:

This is some testbed? Could you share how did you setup such
environment and what are client devices ?

quoted

quoted

[ 7794.230376] ieee80211 phy0: rt2x00lib_rxdone_read_signal: Warning - Frame received with unrecognized signal, mode=0x0001, signal=0x010c, type=4

This is indicator that HW/FW has a problem. There could be various
reasons for that. One possible I can also observe in my setup,is strange
mishmash of seq  on frames which were not acked in BlockACK and
had to be resent. This can happen when many frames are wrongly decoded
(i.e. when there is bad radio condition or we have not correct low level
RF/BBP setup for a Ralink device). To mitigate that problem we can
limit length of agreggeted AMPDU frame.

I attached two patches which do that. One for RX side second for TX side.
Please check if they make a diffrent. You can also hardcode ba_size = 0
for those 30 clients setup.

Note the patches can cause (possibly small) perfromance degradation on
good setups.

Mathias, could you check them as well and see if they do not cause
performance regression on your device ? Lastly when I changed ba_size
setting, it was a problem on your setup.

Hey Stansilaw,

sorry for the delayed testing. I had to create a new test setup first, 
fought with buggy hardware and was busy with other stuff.

The two attached patches are causing a performance regression for me again:

OpenWrt head (forced HT40, 100Mbit wired interface)

wireless (iperf client) to wired (iperf server)
   Interval           Transfer     Bitrate         Retr
     0.00-60.00  sec   584 MBytes  81.6 Mbits/sec  666           sender
     0.00-60.00  sec   584 MBytes  81.6 Mbits/sec                receiver

wired (iperf client) to wireless (iperf server)
   Interval           Transfer     Bitrate         Retr
     0.00-60.00  sec   620 MBytes  86.7 Mbits/sec   33           sender
     0.00-60.00  sec   617 MBytes  86.2 Mbits/sec                receiver



OpenWrt head (forced HT40, 100Mbit wired interface)
   + rt2800_change_rx_ampdu_factor.patch
   + rt2800_change_ba_size.patch

wireless (iperf client) to wired (iperf server)
   Interval           Transfer     Bitrate         Retr
     0.00-60.00  sec   356 MBytes  49.8 Mbits/sec    6           sender
     0.00-60.00  sec   356 MBytes  49.7 Mbits/sec                receiver

wired (iperf client) to wireless (iperf server)
   Interval           Transfer     Bitrate         Retr
     0.00-60.00  sec   627 MBytes  87.7 Mbits/sec    5           sender
     0.00-60.00  sec   626 MBytes  87.5 Mbits/sec                receiver


Due to the regression I haven't tested your ampdu_density patch so far. 
Let me hear if you want to see more tests done.

Mathias