low throughputs with 14e4:4353

[francesco.gringoli at ing.unibs.it]

On Dec 8, 2011, at 10:34 AM, Rafał Miłecki wrote:

> W dniu 8 grudnia 2011 10:21 użytkownik
> <francesco.gringoli at ing.unibs.it> napisał:
>> 
>> On Dec 7, 2011, at 1:01 PM, Rafał Miłecki wrote:
>> 
>>> W dniu 5 grudnia 2011 21:39 użytkownik Rafał Miłecki <zajec5 at gmail.com> napisał:
>>>> W dniu 29 listopada 2011 22:36 użytkownik Rafał Miłecki
>>>> <zajec5 at gmail.com> napisał:
>>>>> I'll give it a rest and will try in next days again.
>>>> 
>>>> It really seems it's something spur avoidance related (and so most
>>>> probably CC PLL related). After applying my patch implementing SPUR
>>>> avoidance, card stops receiving anything. I'll dig more around it.
>>> 
>>> I've found bug in my implementation and fixed it. Today evening I'll
>>> test if my patch fixes performance anyhow (I need physical access).
>> Rafal,
>> 
>> I did some mmio traces and checked what happens when changing txpower.
>> 
>> 1) values in N-PHY tables 0x1A and 0x1B are changed, e.g., for txpower 5 and txpower 10 they are the same, for txpower 15 they are different so I guess there exist thresholds. Unfortunately I don't even know what these tables are.
>> 
>> 2) value in PHY register 0x1ea is changed according to the configured power value. As I can read in specs this is the TX power control target power for both radios: e.g., for txpower 5 => 0x2020, for txpower 10 => 0x2c2c, for txpower 15 => 0x3c3c.
>> 
>> Then I switched to b43 and load a custom firmware to analyze how register 0x1ea is setup and it holds always 0x0000. I forced it to be 0x2c2c and magically the throughput increased to ~ 3Mb/s. Unfortunately when I try 0x3c3c (corresponding to txpower 15) I get only phy-transmission errors.
> 
> I can see a lot of differences between wl and b43, a lot to validate&test.
> 
> Do you have some magic scripts for parsing MMIO dumps? I hope you do,
> otherwise it's impossible to analyze that ;)
Hi Rafal,

my dumb email filter moves messages to b43-dev if I'm not explicitly in To: so I missed your script sharing and some messages from Larry. Ok, I modified the rules :-)

But before that I wrote a short perl script which helped me finding the following (if you want to give it a try it's here http://www.ing.unibs.it/~gringoli/b43_parsemmio.pl ).

So I traced what changes when I set the power to either 5/10 or 15: the only different setting in phy regs is for phy reg 0x01EA where the driver writes the following values according to power level (5, 10 or 15):

5	=>	phy(0x01EA) <= 0x2020
10	=>	phy(0x01EA) <= 0x2828
15	=>	phy(0x01EA) <= 0x3C3C

Then the driver refreshes tables at address 0x6800, 0x6840, 0x6C00 and 0x6C40, you can find the human readable forms for the 5/10/15 cases here

http://www.ing.unibs.it/~gringoli/table_set5
http://www.ing.unibs.it/~gringoli/table_set10
http://www.ing.unibs.it/~gringoli/table_set15

You see: for 5 and 10 the tables are identical, while for 15 some values are different. It seems also that tables at 6840 and 6C40 are shorter than that uploaded by b43 (84 values instead of 128).

With the b43 driver if I manually force from the firmware the power level to either 0x2020 or 0x2828 I get increasing rates (up to ~ 3Mb/s) and everything works. If instead I set greater values I get only PHY transmission errors and no packet is ever transmitted.

Cheers,
-Francesco

P.S. I see that the Broadcom driver continues uploading values in tables for a while, after that everything seems to be "stable" for a while and no more values are uploaded. It can take 10 minutes to stabilize...

> You can try adding
> nphy->txpwrctrl = true;
> in function b43_nphy_op_prepare_structs if you wish test something
> right now. I've noticed this missing part today morning, but won't
> able to test until evening.
> 
> -- 
> Rafał