another set of ppc405_enet changes

I wanted to limit this patch to be just a rework of the static var's but I
ended up changing more things. I will attempt a summary of the changes.
I did not compile/fix the stuff ifdef'd out by default.

I attached the output from some runs with nttcp before and after the changes.
It does not appear like there is any noticable change in performace from these
changes. "-u" is for UDP "-r" is Rx'd on the 405GP Tx'd by my x86 box, "-t" is
Tx'd by the 405GP Rx'd by the x86.
One thing to look into would be the diff between UDP tx packets to the
number rx'd on the other side. I don't know where the packets are being dropped
now.

Config.ini changed to CONFIG_PPC405_ENET to be tristate instead of just bool.

Change NUM_RX to 32 and NUM_TX to 8. Doesn't change performace noticablely.
Moved probe1 code into init function.

kfree of private struct on module unload or failure to register netdevice.

Used skb->len instead of tail-head in places.

For both tx/rx eob int's clear reg first then loop through, instead of loop
clear, goto top of loop again.

Moved tx stat counts into tx eob from hard_xmit. In tx_eob moved netif_wake
outside of loop.

In rx_eob wait untill end of function to refill slots instead of after
each packet. Also did skb_put instead of += tail and len. Put in check
for dev_alloc_skb failure but still no recovery when OOM.

Orig--------------------
# ./nttcp-ppc -T  -u -r -n 100000 192.168.1.5
     Bytes  Real s   CPU s Real-MBit/s  CPU-MBit/s   Calls  Real-C/s   CPU-C/s
l409600000   36.75   24.11     89.1642    135.9104  100001   2721.10    4147.7
1409600000   36.75    3.10     89.1754   1057.0323  100003   2721.50   32259.0
# ./nttcp-ppc -T  -u -t -n 100000 192.168.1.5
     Bytes  Real s   CPU s Real-MBit/s  CPU-MBit/s   Calls  Real-C/s   CPU-C/s
l409600000   34.81   22.75     94.1400    144.0352  100003   2873.01    4395.7
1388304896   34.81    1.54     89.2353   2017.1683   94802   2723.27   61559.7
# ./nttcp-ppc -T   -t -n 100000 192.168.1.5
     Bytes  Real s   CPU s Real-MBit/s  CPU-MBit/s   Calls  Real-C/s   CPU-C/s
l409600000   46.71   44.44     70.1553     73.7354  100000   2140.97    2250.2
1409600000   46.71    6.79     70.1555    482.5920  288192   6170.12   42443.6
# ./nttcp-ppc -T   -r -n 100000 192.168.1.5
     Bytes  Real s   CPU s Real-MBit/s  CPU-MBit/s   Calls  Real-C/s   CPU-C/s
l409600000   44.21   42.45     74.1172     77.1920  100161   2265.52    2359.5
1409600000   44.21    2.01     74.1199   1630.2488  100000   2261.96   49751.2

With Patch--------------------

# ./nttcp-ppc -T -u -r -n 100000 192.168.1.5
     Bytes  Real s   CPU s Real-MBit/s  CPU-MBit/s   Calls  Real-C/s   CPU-C/s
l409600000   36.39   24.43     90.0534    134.1302  100001   2748.24    4093.4
1409600000   36.38    3.19     90.0617   1027.2100  100003   2748.55   31348.9
# ./nttcp-ppc -T -u -r -n 100000 192.168.1.5
     Bytes  Real s   CPU s Real-MBit/s  CPU-MBit/s   Calls  Real-C/s   CPU-C/s
l409600000   36.33   24.14     90.1953    135.7415  100001   2752.57    4142.5
1409600000   36.32    3.34     90.2083    981.0778  100003   2753.02   29941.0
# ./nttcp-ppc -T -u -t -n 100000 192.168.1.5
     Bytes  Real s   CPU s Real-MBit/s  CPU-MBit/s   Calls  Real-C/s   CPU-C/s
l409600000   34.81   20.59     94.1261    159.1452  100003   2872.59    4856.9
1386686976   34.82    1.77     88.8523   1747.7377   94407   2711.58   53337.3
# ./nttcp-ppc -T -t -n 100000 192.168.1.5
     Bytes  Real s   CPU s Real-MBit/s  CPU-MBit/s   Calls  Real-C/s   CPU-C/s
l409600000   46.22   41.92     70.8989     78.1679  100000   2163.66    2385.5
1409600000   46.22    6.35     70.8987    516.0315  287353   6217.33   45252.4
# ./nttcp-ppc -T -r -n 100000 192.168.1.5
     Bytes  Real s   CPU s Real-MBit/s  CPU-MBit/s   Calls  Real-C/s   CPU-C/s
l409600000   44.79   42.63     73.1577     76.8661  100099   2234.81    2348.1
1409600000   44.79    2.80     73.1578   1170.2857  100000   2232.60   35714.3

andrew may wrote:

I wanted to limit this patch to be just a rework of the static var's but I
ended up changing more things. I will attempt a summary of the changes.
I did not compile/fix the stuff ifdef'd out by default.

I attached the output from some runs with nttcp before and after the changes.
It does not appear like there is any noticable change in performace from these
changes. "-u" is for UDP "-r" is Rx'd on the 405GP Tx'd by my x86 box, "-t" is
Tx'd by the 405GP Rx'd by the x86.
One thing to look into would be the diff between UDP tx packets to the
number rx'd on the other side. I don't know where the packets are being dropped
now.

Config.ini changed to CONFIG_PPC405_ENET to be tristate instead of just bool.

Change NUM_RX to 32 and NUM_TX to 8. Doesn't change performace noticablely.
Moved probe1 code into init function.

kfree of private struct on module unload or failure to register netdevice.

Used skb->len instead of tail-head in places.

For both tx/rx eob int's clear reg first then loop through, instead of loop
clear, goto top of loop again.

Moved tx stat counts into tx eob from hard_xmit. In tx_eob moved netif_wake
outside of loop.

In rx_eob wait untill end of function to refill slots instead of after
each packet. Also did skb_put instead of += tail and len. Put in check
for dev_alloc_skb failure but still no recovery when OOM.

Andrew;

This look very similar to a previous patch from last week. Provide a
patch that represents addition work since then.  I have applied most of
your previous patches.  Please be patient,  this does not happen over
night.

regards,

Armin

** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

On Wed, Sep 12, 2001 at 09:06:37AM -0700, Armin Kuster wrote:

This look very similar to a previous patch from last week. Provide a
patch that represents addition work since then.  I have applied most of
your previous patches.  Please be patient,  this does not happen over
night.

This should be what you want. The patch from the other week just moved
some code from probe into the init function. This patch does a lot more.