On May 30, 2017 5:11 PM, "Patrick Venture" <venture@google.com> wrote:

On Tue, May 30, 2017 at 01:27:39PM -0700, Patrick Venture wrote:
>> The aspeed-pwm-tacho controller supports measuring the fan tach by using
>> leading, falling, or both edges. This change allows the driver to
>> support either of the three configurations and will appropriately modify
>> the returned tach data.
>
>> I tested this and found the number returned matched what I expected.
>
>> Signed-off-by: Patrick Venture <venture@google.com>
>> ---
>> drivers/hwmon/aspeed-pwm-tacho.c | 24 +++++++++++++++++++++---
>> 1 file changed, 21 insertions(+), 3 deletions(-)
>
>> diff --git a/drivers/hwmon/aspeed-pwm-tacho.c b/drivers/hwmon/aspeed-pwm-tacho.c
>> index 48403a2115be..f288928b5e8a 100644
>> --- a/drivers/hwmon/aspeed-pwm-tacho.c
>> +++ b/drivers/hwmon/aspeed-pwm-tacho.c
>> @@ -145,11 +145,20 @@
>
>> #define PWM_MAX 255
>
>> +#define BOTH_EDGES 0x02 /* 10b */
>> +
>> #define M_PWM_DIV_H 0x00
>> #define M_PWM_DIV_L 0x05
>> #define M_PWM_PERIOD 0x5F
>> #define M_TACH_CLK_DIV 0x00
>> -#define M_TACH_MODE 0x00
>> +/*
>> + * 5:4 Type N fan tach mode selection bit:
>> + * 00: falling
>> + * 01: rising
>> + * 10: both
>> + * 11: reserved.
>> + */
>> +#define M_TACH_MODE 0x10

> That seems wrong. The above is a bit mask, this is an already shifted value.

Good catch. That's a reading bits, writing nibbles bug. I did my
testing in a branch that uses the device tree to control these knobs
to get better tuning. I have a series of tweaks to improve the
responsiveness of this code. Originally the driver author was told to
modify the driver to remove the knobs by upstream. For my platform,
I've found the knobs are critical, and there are other changes
associated with them. I'm making an effort to get these changes in
ahead of trying to re-enable the device-tree patch.

Different systems using this aspeed driver may actually need different
values, but for a few systems, I have found higher performing values.

The data sheet is light on details, so some information is gleaned
from the aspeed reference driver.

>> #define M_TACH_UNIT 0x1000

> ... and this one seems wrong too, since it is passed to aspeed_set_tacho_type_values() which shifts it left by 16 bit.

That seems wonky however, I believe the compiler is promoting the
values to 32-bits before shifting them. That's something I've
definitely seen before. However, I can manually cast them if you're
weary of it.

#define ASPEED_PTCR_TYPEM_CTRL 0x10
#define ASPEED_PTCR_TYPEM_CTRL1 0x14
#define TYPE_CTRL_FAN_PERIOD 16
#define TYPE_CTRL_FAN_MODE 4
#define TYPE_CTRL_FAN_DIVISION 1
#define M_TACH_CLK_DIV 0x00

Offset: 10h PTCR10: Type M Control 0 Register Init = X
31:16 RW Type M fan tach period bit [15:0] (in unit of type M PWM clock)
5 :4 RW Type M fan tach mode selection bit [1:0]
00: falling edge
01: rising edge
10: both edges
11: reserved
3 :1 RW Type M fan tach clock division bit [1:0]
000: divide 4
001: divide 16
010: divide 64
011: divide 256 ...
111: divide 65536
0 RW Enable fan tach of type M 0: disable 1: enable

Offset: 14h PTCR14: Type M Control 1 Register Init = X
Bit R/W Description
31:16 RW Type M fan tach falling point bit [15:0] of period
15:0 RW Type M fan tach rising point bit [15:0] of period

0x1000 << 16 | 0x02 << 4 | M_TACH_CLK_DIV << 1

u32 reg_value = ((mode << TYPE_CTRL_FAN_MODE) |
(unit << TYPE_CTRL_FAN_PERIOD) |
(division << TYPE_CTRL_FAN_DIVISION));

regmap_update_bits(regmap, type_params[type].ctrl_reg,
TYPE_CTRL_FAN_MASK, reg_value);
regmap_update_bits(regmap, type_params[type].ctrl_reg1,
TYPE_CTRL_FAN1_MASK, unit << 16);

For type M Control 0 Register they're setting the fan tach period to
0x1000. Presumably the compiler is promoting the value to 32-bits.

For Type M Control 1 Register they're setting the fan tach falling
point to 0x1000 and the rising point to 0. Presumably the compiler is
promoting unit to 32 bits. I'm completely comfortable casting it to
address this.

>> #define INIT_FAN_CTRL 0xFF
>
>> @@ -162,6 +171,7 @@ struct aspeed_pwm_tacho_data {
>> u8 type_pwm_clock_division_h[3];
>> u8 type_pwm_clock_division_l[3];
>> u8 type_fan_tach_clock_division[3];
>> + u8 type_fan_tach_mode[3];
>> u16 type_fan_tach_unit[3];
>> u8 pwm_port_type[8];
>> u8 pwm_port_fan_ctrl[8];
>> @@ -498,7 +508,7 @@ static u32 aspeed_get_fan_tach_ch_rpm(struct aspeed_pwm_tacho_data *priv,
>> u8 fan_tach_ch)
>> {
>> u32 raw_data, tach_div, clk_source, sec, val;
>> - u8 fan_tach_ch_source, type;
>> + u8 fan_tach_ch_source, type, mode, both;
>
>> regmap_write(priv->regmap, ASPEED_PTCR_TRIGGER, 0);
>> regmap_write(priv->regmap, ASPEED_PTCR_TRIGGER, 0x1 << fan_tach_ch);
>> @@ -512,7 +522,14 @@ static u32 aspeed_get_fan_tach_ch_rpm(struct aspeed_pwm_tacho_data *priv,
>> regmap_read(priv->regmap, ASPEED_PTCR_RESULT, &val);
>> raw_data = val & RESULT_VALUE_MASK;
>> tach_div = priv->type_fan_tach_clock_division[type];
>> - tach_div = 0x4 << (tach_div * 2);
>> + /*
>> + * We need the mode to determine if the raw_data is double (from
>> + * counting both edges).
>> + */
>> + mode = priv->type_fan_tach_mode[type];
>> + both = (mode & BOTH_EDGES) ? 1 : 0;

> BOTH_EDGES is 0x02. Mode is 0x10. ???

Bit -> nibble bug.

>> +
>> + tach_div = (0x4 << both) << (tach_div * 2);
>> clk_source = priv->clk_freq;
>
>> if (raw_data == 0)
>> @@ -696,6 +713,7 @@ static void aspeed_create_type(struct aspeed_pwm_tacho_data *priv)
>> aspeed_set_tacho_type_enable(priv->regmap, TYPEM, true);
>> priv->type_fan_tach_clock_division[TYPEM] = M_TACH_CLK_DIV;
>> priv->type_fan_tach_unit[TYPEM] = M_TACH_UNIT;
>> + priv->type_fan_tach_mode[TYPEM] = M_TACH_MODE;

> type_fan_tach_mode[] is always M_TACH_MODE or 0x10, and BOTH_EDGES (0x02) is never set.

> What am I missing here ? Not sure what you are trying to accomplish.
> Even if the bit mask is corrected, what is the benefit of changing the mode ?

The benefit of changing the mode is that I can tune down (I was
planning a series of these patches but I'm still new to upstreaming to
linux and didn't want to get mixed up, too late) the time it takes for
the controller to provide the fan speeds.

At present this driver takes 1s per fan. The controller itself
appears to cache the speeds every 0.08s seconds, so with this and a
few other tweaks I can get the individual read speed down to 0.017
seconds each and get a fresh value more than 10x per second. Which is
something I need for a PID based fan controller.

One of the ways I improved this was reducing the fan tach period. I'm
able to get reliable numbers with a lower period, but I'd rather
people be able to tune that. However, I can update the driver to have
better defaults.

> Guenter

>> aspeed_set_tacho_type_values(priv->regmap, TYPEM, M_TACH_MODE,
>> M_TACH_UNIT, M_TACH_CLK_DIV);
>> }