am335x: cpsw at high broadcast load

We experience massive problems, when putting am335x based devices into
high broadcast load environment (customer reports). To reproduce the
issue we've written following Python script

#!/usr/bin/python
# Send UDP broadcast packets

myport = 50000

import sys, time
from socket import *

if len(sys.argv) == 2:
    myport = int(sys.argv[1])

s = socket(AF_INET, SOCK_DGRAM)
s.bind(('192.168.254.36', 0))
s.setsockopt(SOL_SOCKET, SO_BROADCAST, 1)

while 1:
    #data = repr(time.time()) + '\n'
    data = "a"
    s.sendto(data, ('<broadcast>', myport))
    time.sleep(0.0001)

Just running this script on a host machine and invoking "top" on the
am335x machine results in sirq showing about 90% for 600MHz CPU and
3.18.x kernel. Below example of idle "top" output for the reference:

Mem: 13432K used, 217844K free, 64K shrd, 652K buff, 4956K cached
CPU:   0% usr   0% sys   0% nic  99% idle   0% io   0% irq   0% sirq

Then I tried 4.2.0-rc4 and sirq was about 80%, then I applied both

drivers: net: cpsw: remove  disable_irq/enable_irq as irq can be
    masked from cpsw itself
  drivers: net: cpsw: add separate napi for tx packet handling for
    performance improvment

and sirq lowered to about 50%, but I have also tried evil vendor tree

13 more lines

Hi Mugunthan,

On Fri, Jul 31, 2015 at 2:22 PM, Yegor Yefremov
[off-list ref] wrote:

We experience massive problems, when putting am335x based devices into
high broadcast load environment (customer reports). To reproduce the
issue we've written following Python script

#!/usr/bin/python
# Send UDP broadcast packets

myport = 50000

import sys, time
from socket import *

if len(sys.argv) == 2:
    myport = int(sys.argv[1])

s = socket(AF_INET, SOCK_DGRAM)
s.bind(('192.168.254.36', 0))
s.setsockopt(SOL_SOCKET, SO_BROADCAST, 1)

while 1:
    #data = repr(time.time()) + '\n'
    data = "a"
    s.sendto(data, ('<broadcast>', myport))
    time.sleep(0.0001)

Just running this script on a host machine and invoking "top" on the
am335x machine results in sirq showing about 90% for 600MHz CPU and
3.18.x kernel. Below example of idle "top" output for the reference:

Mem: 13432K used, 217844K free, 64K shrd, 652K buff, 4956K cached
CPU:   0% usr   0% sys   0% nic  99% idle   0% io   0% irq   0% sirq

Then I tried 4.2.0-rc4 and sirq was about 80%, then I applied both

drivers: net: cpsw: remove  disable_irq/enable_irq as irq can be
    masked from cpsw itself
  drivers: net: cpsw: add separate napi for tx packet handling for
    performance improvment

and sirq lowered to about 50%, but I have also tried evil vendor tree

8 more quoted lines

Could you please look at this issue. It is still valid for 4.5 and am335x.

Thanks.

Yegor