i2c: mvtwsi: Fix problem with baud rate calculation

The current implementation for baudrate calculation is incorrect. This part from the formula:

"2 ^ (n + 1)" is not equivalent to (1 << n) but to (2 << n)!

This patch fixes this and moves this calculation to a function instead of using a macro.

This new function is taken from the Linux kernel.

This was detected and tested on the Marvell Armada A38x DB-88F6820-GP eval board.

Signed-off-by Stefan Roese <[email protected]>

Cc Prafulla Wadaskar <[email protected]>

Cc Luka Perkov <[email protected]>

Cc Hans de Goede <[email protected]>

Cc Ian Campbell <[email protected]>

Cc Heiko Schocher <[email protected]>

---
  drivers/i2c/mvtwsi.c | 13 +++++--------
  1 file changed, 5 insertions(+), 8 deletions(-)

drivers/i2c/mvtwsi.c +5 -8

index 9b2ca1e..320235c 100644
@@ -228,13 +228,10 @@ static int twsi_stop(int status)
  	return status;
  }
 
-/*
- * Ugly formula to convert m and n values to a frequency comes from
- * TWSI specifications
- */
-
-#define TWSI_FREQUENCY(m, n) \
-	(CONFIG_SYS_TCLK / (10 * (m + 1) * (1 << n)))
+static unsigned int twsi_calc_freq(const int n, const int m)
+{
+	return CONFIG_SYS_TCLK / (10 * (m + 1) * (2 << n));
+}
 
  /*
   * Reset controller.

@@ -266,7 +263,7 @@ static unsigned int twsi_i2c_set_bus_speed(struct i2c_adapter *adap,
  	/* compute m, n setting for highest speed not above requested speed */
  	for (n = 0; n < 8; n++) {
  		for (m = 0; m < 16; m++) {
-			tmp_speed = TWSI_FREQUENCY(m, n);
+			tmp_speed = twsi_calc_freq(n, m);
  			if ((tmp_speed <= requested_speed)
  			 && (tmp_speed > highest_speed)) {
  				highest_speed = tmp_speed;

Quoted 7 lines · Stefan Roese

The current implementation for baudrate calculation is incorrect. This part from the formula:

"2 ^ (n + 1)" is not equivalent to (1 << n) but to (2 << n)!

This patch fixes this and moves this calculation to a function instead of using a macro.

Hmm, this does not match with what the Allwinner datasheets say:
https://github.com/allwinner-zh/documents/blob/master/A20/A20%20user%20manual%20v1.3%2020141010.pdf

They say:

Fsamp = F 0 = Fin / 2^CLK_N
F1 = F0 / (CLK_M + 1)
Foscl = F1 / 10 = Fin / (2^CLK_N * (CLK_M + 1)*10)

With Foscl being the ultimate i2c speed. Notice that they are talking about 2 ^ CLK_N not 2 ^ (CLK_N + 1)

And they have a few examples which match this. Now it could be that a register value of 0 means CLK_N = 1, reg 1 CLK_N = 2, etc. this is not clearly
specified ...

Quoted 1 line

This new function is taken from the Linux kernel.

Interesting, because on Allwinnner / sunxi devices we are using the kernel driver formula unmodified, and things seem to work fine despite this. Could be tolerances allowing this, could be the Allwinner datasheet being unclear.

I've send allwinner a mail asking them to clarify this.

Quoted 2 lines

This was detected and tested on the Marvell Armada A38x DB-88F6820-GP eval board.

So I take it you connected a memory oscilloscope to the i2c wires ?

Hi Hans,

Quoted 27 lines · Hans de Goede

Quoted 7 lines · Stefan Roese

The current implementation for baudrate calculation is incorrect. This part from the formula:

"2 ^ (n + 1)" is not equivalent to (1 << n) but to (2 << n)!

This patch fixes this and moves this calculation to a function instead of using a macro.

Hmm, this does not match with what the Allwinner datasheets say:
https://github.com/allwinner-zh/documents/blob/master/A20/A20%20user%20manual%20v1.3%2020141010.pdf

They say:

Fsamp = F 0 = Fin / 2^CLK_N
F1 = F0 / (CLK_M + 1)
Foscl = F1 / 10 = Fin / (2^CLK_N * (CLK_M + 1)*10)

With Foscl being the ultimate i2c speed. Notice that they are talking about 2 ^ CLK_N not 2 ^ (CLK_N + 1)

And they have a few examples which match this. Now it could be that a register value of 0 means CLK_N = 1, reg 1 CLK_N = 2, etc. this is not clearly
specified ...

Quoted 8 lines

Quoted 1 line

This new function is taken from the Linux kernel.

Interesting, because on Allwinnner / sunxi devices we are using the kernel driver formula unmodified, and things seem to work fine despite this. Could be tolerances allowing this, could be the Allwinner datasheet being unclear.

I've send allwinner a mail asking them to clarify this.

Good.

Here the complete formula from the A38x manual:

F_SCL = F_TCLK / (10 * (M + 1) * 2 ^ (N + 1))

Quoted 4 lines

Quoted 2 lines

This was detected and tested on the Marvell Armada A38x DB-88F6820-GP eval board.

So I take it you connected a memory oscilloscope to the i2c wires ?

No. I noticed that I2C doesn't work on this board with the current code. And checked for differences to the Linux code. And with this change, I2C does work on this board.

Hi,

Quoted 6 lines · Stefan Roese

The current implementation for baudrate calculation is incorrect. This part from the formula:

"2 ^ (n + 1)" is not equivalent to (1 << n) but to (2 << n)!

This patch fixes this and moves this calculation to a function instead of using a macro.

Hmm, this does not match with what the Allwinner datasheets say:
https://github.com/allwinner-zh/documents/blob/master/A20/A20%20user%20manual%20v1.3%2020141010.pdf

They say:

Fsamp = F 0 = Fin / 2^CLK_N
F1 = F0 / (CLK_M + 1)
Foscl = F1 / 10 = Fin / (2^CLK_N * (CLK_M + 1)*10)

With Foscl being the ultimate i2c speed. Notice that they are talking about 2 ^ CLK_N not 2 ^ (CLK_N + 1)

And they have a few examples which match this. Now it could be that a register value of 0 means CLK_N = 1, reg 1 CLK_N = 2, etc. this is not clearly specified ...

Quoted 1 line

This new function is taken from the Linux kernel.

Interesting, because on Allwinnner / sunxi devices we are using the kernel driver formula unmodified, and things seem to work fine despite this. Could be tolerances allowing this, could be the Allwinner datasheet being unclear.

I've send allwinner a mail asking them to clarify this.

Hi,

Quoted 32 lines · Hans de Goede

Hi,

Quoted 6 lines · Stefan Roese

The current implementation for baudrate calculation is incorrect. This part from the formula:

"2 ^ (n + 1)" is not equivalent to (1 << n) but to (2 << n)!

This patch fixes this and moves this calculation to a function instead of using a macro.

Hmm, this does not match with what the Allwinner datasheets say:
https://github.com/allwinner-zh/documents/blob/master/A20/A20%20user%20manual%20v1.3%2020141010.pdf

They say:

Fsamp = F 0 = Fin / 2^CLK_N
F1 = F0 / (CLK_M + 1)
Foscl = F1 / 10 = Fin / (2^CLK_N * (CLK_M + 1)*10)

With Foscl being the ultimate i2c speed. Notice that they are talking about 2 ^ CLK_N not 2 ^ (CLK_N + 1)

And they have a few examples which match this. Now it could be that a register value of 0 means CLK_N = 1, reg 1 CLK_N = 2, etc. this is not clearly specified ...

Quoted 1 line

This new function is taken from the Linux kernel.

Interesting, because on Allwinnner / sunxi devices we are using the kernel driver formula unmodified, and things seem to work fine despite this. Could be tolerances allowing this, could be the Allwinner datasheet being unclear.

Quoted 1 line

I've send allwinner a mail asking them to clarify this.

I've just gotten an answer from allwinner, and the formula in their datasheet is correct, with the raw register value == CLK_N so the current (before this patch) u-boot code is correct for Allwinner devices and we need a #ifdef here to keep the old formula for sunxi devices, you can use:

Quoted 1 line

I've send allwinner a mail asking them to clarify this.

I've just gotten an answer from allwinner, and the formula in their datasheet
is correct, with the raw register value == CLK_N so the current (before this patch)
u-boot code is correct for Allwinner devices and we need a #ifdef here to keep the
old formula for sunxi devices, you can use:

#ifdef CONFIG_SUNXI

For this.