ath9k: noise floor calibration process

[Benoit PAPILLAULT <benoit.papillault@free.fr>]

Hello,

In order to move forward with noise & signal reporting, I'd like to 
share my current understanding of the way  ath9k HW is working before 
sending patches (unfortunately, I did the work before the introduction 
of ar9003... so I need to redo the work).

The ultimate purpose of this work is to be able to measure signal levels 
(and noise if possible) as accurately as a spectrum analyzer or power meter.

First, signal level reporting. It is reported in a per packet basis in 
RX descriptors. There are 7 fields:
    AR_RxRSSIAnt00    0x000000ff    rs_rssi_ctl0
    AR_RxRSSIAnt01    0x0000ff00    rs_rssi_ctl1
    AR_RxRSSIAnt02    0x00ff0000    rs_rssi_ctl2
    AR_RxRSSIAnt10        0x000000ff    rs_rssi_ext0
    AR_RxRSSIAnt11        0x0000ff00    rs_rssi_ext1
    AR_RxRSSIAnt12        0x00ff0000    rs_rssi_ext2
    AR_RxRSSICombined    0xff000000    rs_rssi

Each value is for a 20 MHz wide channel, on the 3 RX chains. "ctl" is 
for the primary channel and "ext" is for the secondary channel (using 
the 802.11n words). The latter rs_rssi is the sum of the 6 previous 
value. However, since each value is dB, the sum is not an arithmetic 
sum. Each field is a signed value and the value -128 means that no 
measurement has been done  (no RX chain, RX chain disabled, no secondary 
channel, ...). It seems that in some cases, the combined value is just 
plain wrong. Here are few examples:

  RSSI: ctl=(10,7,-128) ext=(-128,-128,-128) => 12 (11.76)    correct

  RSSI: ctl=(38,29,-128) ext=(69,-84,-101) => -22        incorrect!!!


Next, noise floor calibration. From what I understand, signal levels is 
measured using the AGC + RX amplifiers gain (RF, IF and BB). However, 
the various gains are not really accurate, only the relative gain are 
accurate. This means that reading a signal value of -100dBm might not 
exactly means -100dBm. There is a delta between real signal and measured 
value. In order to know this value, we need a calibration process with a 
known signal.

One know signal is thermal noise. Thermal noise is generated in any 
resistor and can be computed using the well know value N = kTB. For a 20 
MHz bandwidth, this gives -101dBm. If the HW tries to measure signal 
strength when the network is supposed to be idle (during SIFS) and with 
RX/TX switch disabled (?), then it will in fact measure the thermal 
noise at the RX input.

So, we have :

Real noise (-101dBm) = Measured noise + delta

There are type of registers to control noise floor calibration :

- control register at 0x9860    (AR_PHY_AGC_CONTROL)

This register allows 3 differents operations :

1. start noise floor measurement

  write AR_PHY_MAXCCA_PWR (AR_PHY_CCA & 0x000001ff) : this is apparently 
a max value
    for noise floor
  REG_SET_BIT(ah, AR_PHY_AGC_CONTROL, AR_PHY_AGC_CONTROL_ENABLE_NF);
  REG_SET_BIT(ah, AR_PHY_AGC_CONTROL, AR_PHY_AGC_CONTROL_NO_UPDATE_NF);
  REG_SET_BIT(ah, AR_PHY_AGC_CONTROL, AR_PHY_AGC_CONTROL_NF);

  When channel has been changed however, the noise floor needs to be 
updated immediately, so AR_PHY_AGC_CONTROL_NO_UPDATE_NF should be 
cleared in this particular case. Otherwise, the chip is no longer 
receiving (problem since CCA is defined with noise floor as reference).

2. read noise floor measurement result

    check REG_READ(ah, AR_PHY_AGC_CONTROL) & AR_PHY_AGC_CONTROL_NF
    if 0 (noise floor calibration is finished), read AR_PHY_MINCCA_PWR :
      nf = MS(REG_READ(ah, AR_PHY_CCA), AR_PHY_MINCCA_PWR = 0x0ff80000)

3. write noise floor reference

  write AR_PHY_MAXCCA_PWR (the value has not the same meaning as 
operation 1!)
  REG_CLR_BIT(ah, AR_PHY_AGC_CONTROL, AR_PHY_AGC_CONTROL_ENABLE_NF);
  REG_CLR_BIT(ah, AR_PHY_AGC_CONTROL, AR_PHY_AGC_CONTROL_NO_UPDATE_NF);
  REG_SET_BIT(ah, AR_PHY_AGC_CONTROL, AR_PHY_AGC_CONTROL_NF);

- data register at 0x9864 (AR_PHY_CCA, + more location for other RX chains)

  The fields are different for AR9280+ chipsets, but the mechanism is 
the same.

    AR_PHY_MAXCCA_PWR        0x000001ff (half dBm unit!)
    AR_PHY_CCA_THRESH62    0x0007f000
    AR_PHY_MINCCA_PWR        0x0ff80000

Now, we have :

Real signal = Measured signal + delta
    = RSSI + Noise floor + delta
    = RSSI + (-101 dBm)

Real noise is not thermal noise. There are a lot of definition for noise 
since noise is NOT signal. Of course, noise includes thermal noise. 
Since the noise measured by the chip is variable, I think we could do :

- Noise floor = minimum (Noise floor measures)
- Noise = moving average (Noise floor measures) + delta
  with delta = (-101 dBm) - Noise floor

I'd like to get comments before sending patches. Since ath5k and ath9k 
are quite close, I'm pretty sure a similar (if not same) process is used 
on ath5k.

Regards,
Benoit