Patchwork i2c-cpm: Add flexibility for I2C clock frequency and filter.

login
register
mail settings
Submitter Mike Ditto
Date Oct. 31, 2008, 4:09 a.m.
Message ID <490A84F8.20403@consentry.com>
Download mbox | patch
Permalink /patch/6666/
State Superseded, archived
Headers show

Comments

Mike Ditto - Oct. 31, 2008, 4:09 a.m.
This patch adds the ability to enable the digital filter in the device
tree (with the "clock-filter" boolean property) and automates the
predivider selection according to the clock-frequency and clock-filter
properties.

Signed-off-by:  Mike Ditto <mditto@consentry.com>
---
This patch is against 2.6.27.

To use the full range of I2C clock frequencies supported by the
Freescale CPM I2C controller, it is necessary to choose an appropriate
predivider value.  The choice is affected by whether the SCL signal
digital filter is enabled.

The existing code computes the final divider (i2brg) but always uses
a predivider of 32.  This can set illegal values in i2brg - for example
on a machine with 25 MHz BRG_CLK, when selecting I2C clock-frequency
97656 Hz, the code was loading i2brg with the value 1, which does not
work (the CPM requires a minimum value of 3 when the digital filter is
not enabled).  Additionally, the calculation did not work when the
filter is enabled (and the driver did not provide a way to enable it).

And by the way, thanks to Jochen Friedrich for integrating this driver
on the very day that I went looking for it.  :-)
Jochen Friedrich - Oct. 31, 2008, 11:40 a.m.
Hi Mike,

> This patch adds the ability to enable the digital filter in the device
> tree (with the "clock-filter" boolean property) and automates the
> predivider selection according to the clock-frequency and clock-filter
> properties.

looks good.

David, is "clock-filter" an appropriate dts property for this purpose or
would you prefer a different name?

What needs to be done though is to document this change in
Documentation/powerpc/dts-bindings/fsl/cpm_qe/cpm/i2c.txt.

Thanks,
Jochen
David Gibson - Nov. 5, 2008, 11:36 p.m.
On Fri, Oct 31, 2008 at 12:40:47PM +0100, Jochen Friedrich wrote:
> Hi Mike,
> 
> > This patch adds the ability to enable the digital filter in the device
> > tree (with the "clock-filter" boolean property) and automates the
> > predivider selection according to the clock-frequency and clock-filter
> > properties.
> 
> looks good.
> 
> David, is "clock-filter" an appropriate dts property for this purpose or
> would you prefer a different name?

Hrm, well the name seems fine, but then, device-specific properties
are device-specific so it's pretty much up to the device binding to
pick a name.

What does worry me, however, is the description says it's about
whether the driver "should" enable the filter.  Generally the device
tree doesn't attempt to say what users "should" do with the hardware,
just what the characteristics of the hardware are.

What's the underlying difference here that affects the driver's choice
to enable the filter or not?
Mike Ditto - Nov. 6, 2008, 12:35 a.m.
David Gibson wrote:
> What does worry me, however, is the description says it's about
> whether the driver "should" enable the filter.  Generally the device
> tree doesn't attempt to say what users "should" do with the hardware,
> just what the characteristics of the hardware are.
>
> What's the underlying difference here that affects the driver's choice
> to enable the filter or not?

I think it's a hardware/environment design parameter - e.g. if the I2C
bus has hot-pluggable devices, long PCB traces, or a hierarchy of
multiplexed bus segments, these can result in a noisy SCL signal that
needs to be filtered.  It's also a recommended mitigation for errata
in certain CPU revs.

					-=] Mike [=-
David Gibson - Nov. 6, 2008, 12:53 a.m.
On Wed, Nov 05, 2008 at 04:35:03PM -0800, Mike Ditto wrote:
> David Gibson wrote:
> > What does worry me, however, is the description says it's about
> > whether the driver "should" enable the filter.  Generally the device
> > tree doesn't attempt to say what users "should" do with the hardware,
> > just what the characteristics of the hardware are.
> >
> > What's the underlying difference here that affects the driver's choice
> > to enable the filter or not?
> 
> I think it's a hardware/environment design parameter - e.g. if the I2C
> bus has hot-pluggable devices, long PCB traces, or a hierarchy of
> multiplexed bus segments, these can result in a noisy SCL signal that
> needs to be filtered.  It's also a recommended mitigation for errata
> in certain CPU revs.

Ah, ok.  Well the CPU revision thing could be selected based on the
CPU revision, but the other conditions are a property of the board
wiring.  Obviously it's hard to precisely characterize what it says
about the hardware, which is usually best avoided for devtree
properties, but I can see why this is more-or-less unavoidable in this
case.

Ok.  This property name and meaning looks ok to me.  I would suggest a
note in the binding roughly explaining what leads to the property
being set (basically what you just told me in the paragraph above).
Mike Ditto - Nov. 6, 2008, 1:12 a.m.
[including extra context because some of the thread went to the
 wrong I2C list]

David Gibson wrote:
> On Wed, Nov 05, 2008 at 04:35:03PM -0800, Mike Ditto wrote:
>> David Gibson wrote:
>> > What does worry me, however, is the description says it's about
>> > whether the driver "should" enable the filter.  Generally the device
>> > tree doesn't attempt to say what users "should" do with the hardware,
>> > just what the characteristics of the hardware are.
>> >
>> > What's the underlying difference here that affects the driver's choice
>> > to enable the filter or not?
>> 
>> I think it's a hardware/environment design parameter - e.g. if the I2C
>> bus has hot-pluggable devices, long PCB traces, or a hierarchy of
>> multiplexed bus segments, these can result in a noisy SCL signal that
>> needs to be filtered.  It's also a recommended mitigation for errata
>> in certain CPU revs.
> 
> Ah, ok.  Well the CPU revision thing could be selected based on the
> CPU revision, but the other conditions are a property of the board
> wiring.  Obviously it's hard to precisely characterize what it says
> about the hardware, which is usually best avoided for devtree
> properties, but I can see why this is more-or-less unavoidable in this
> case.
> 
> Ok.  This property name and meaning looks ok to me.  I would suggest a
> note in the binding roughly explaining what leads to the property
> being set (basically what you just told me in the paragraph above).

Will do.  I'll send a revised patch shortly.

Thanks,
					-=] Mike [=-

Patch

Index: linux/drivers/i2c/busses/i2c-cpm.c
===================================================================
RCS file: /n/home2/mditto/ws/myrepo/kernel/linux/drivers/i2c/busses/i2c-cpm.c,v
retrieving revision 1.1.1.1
diff -u -p -r1.1.1.1 i2c-cpm.c
--- linux/drivers/i2c/busses/i2c-cpm.c	11 Oct 2008 02:53:07 -0000	1.1.1.1
+++ linux/drivers/i2c/busses/i2c-cpm.c	31 Oct 2008 04:05:41 -0000
@@ -87,6 +87,11 @@  struct i2c_ram {
 #define I2CER_TXB	0x02
 #define I2CER_RXB	0x01
 #define I2MOD_EN	0x01
+#define I2MOD_PDIV_32	0x00	/* BRGCLK/32 */
+#define I2MOD_PDIV_16	0x02	/* BRGCLK/16 */
+#define I2MOD_PDIV_8	0x04	/* BRGCLK/8 */
+#define I2MOD_PDIV_4	0x06	/* BRGCLK/4 */
+#define I2MOD_FLT	0x08

 /* I2C Registers */
 struct i2c_reg {
@@ -111,7 +116,6 @@  struct cpm_i2c {
 	int version; /* CPM1=1, CPM2=2 */
 	int irq;
 	int cp_command;
-	int freq;
 	struct i2c_reg __iomem *i2c_reg;
 	struct i2c_ram __iomem *i2c_ram;
 	u16 i2c_addr;
@@ -434,7 +438,8 @@  static int __devinit cpm_i2c_setup(struc
 	void __iomem *i2c_base;
 	cbd_t __iomem *tbdf;
 	cbd_t __iomem *rbdf;
-	unsigned char brg;
+	uint freq, maxfreq, prediv;
+	unsigned char mod, brg;

 	dev_dbg(&cpm->ofdev->dev, "cpm_i2c_setup()\n");

@@ -508,9 +513,15 @@  static int __devinit cpm_i2c_setup(struc

 	data = of_get_property(ofdev->node, "clock-frequency", &len);
 	if (data && len == 4)
-		cpm->freq = *data;
+		freq = *data;
 	else
-		cpm->freq = 60000; /* use 60kHz i2c clock by default */
+		freq = 60000;	/* use 60kHz I2C clock by default */
+
+	data = of_get_property(ofdev->node, "clock-filter", &len);
+	if (data && len == 0)
+		mod = I2MOD_FLT;
+	else
+		mod = 0;

 	/*
 	 * Allocate space for CPM_MAXBD transmit and receive buffer
@@ -552,8 +563,8 @@  static int __devinit cpm_i2c_setup(struc

 	cpm_reset_i2c_params(cpm);

-	dev_dbg(&cpm->ofdev->dev, "i2c_ram 0x%p, i2c_addr 0x%04x, freq %d\n",
-		cpm->i2c_ram, cpm->i2c_addr, cpm->freq);
+	dev_dbg(&cpm->ofdev->dev, "i2c_ram 0x%p, i2c_addr 0x%04x, freq %u\n",
+		cpm->i2c_ram, cpm->i2c_addr, freq);
 	dev_dbg(&cpm->ofdev->dev, "tbase 0x%04x, rbase 0x%04x\n",
 		(u8 __iomem *)cpm->tbase - DPRAM_BASE,
 		(u8 __iomem *)cpm->rbase - DPRAM_BASE);
@@ -566,20 +577,53 @@  static int __devinit cpm_i2c_setup(struc
 	out_8(&cpm->i2c_reg->i2add, 0x7f << 1);

 	/*
-	 * PDIV is set to 00 in i2mod, so brgclk/32 is used as input to the
-	 * i2c baud rate generator. This is divided by 2 x (DIV + 3) to get
-	 * the actual i2c bus frequency.
+	 * Compute the clock predivider and final divider to generate something
+	 * close to the desired I2C clock frequency.  Use the largest predivider
+	 * possible.  i2brg must be >= 6 when using I2MOD_FLT, otherwise >= 3.
+	 * So compute the highest frequency that will work with I2MOD_PDIV_32;
+	 * if that isn't high enough, see if we can use I2MOD_PDIV_16, etc.
+	 * Then choose a final divider that will generate at least the desired
+	 * clock frequency.  Note the "at least" -- this rounds upward, not
+	 * toward the nearest available frequency.
+	 *
+	 * I2C frequency for a given predivider and i2brg value is:
+	 *   i2cfreq = brgfreq / prediv / 2 / (i2brg + 3 + 2 * flt)
+	 * i2brg value for a given predivider and I2C frequency is:
+	 *   i2brg = (brgfreq / prediv / 2 / i2cfreq) - 3 - 2 * flt
+	 * minimum allowed i2brg is (3 + 3 * flt).
+	 * maximum freq possible for a given predivider is:
+	 *   maxfreq = brgfreq / prediv / 2 / (3 + 3 + 5 * flt)
 	 */
-	brg = get_brgfreq() / (32 * 2 * cpm->freq) - 3;
-	out_8(&cpm->i2c_reg->i2brg, brg);
+	maxfreq = get_brgfreq() / 64 / ((mod & I2MOD_FLT) ? 11 : 6);
+	if (freq <= maxfreq) {		/* Works with predivider 32? */
+		mod |= I2MOD_PDIV_32;
+		prediv = 32;
+	} else if (freq <= maxfreq*2) {	/* How about 16? */
+		mod |= I2MOD_PDIV_16;
+		prediv = 16;
+	} else if (freq <= maxfreq*4) {	/* How about 8? */
+		mod |= I2MOD_PDIV_8;
+		prediv = 8;
+	} else {			/* 4 is last resort. */
+		if (freq > maxfreq*8)
+			/* Impossibly high clock-frequency requested. */
+			freq = maxfreq*8;
+		mod |= I2MOD_PDIV_4;
+		prediv = 4;
+	}
+	brg = get_brgfreq() / prediv / 2 / freq - ((mod & I2MOD_FLT) ? 5 : 3);

-	out_8(&cpm->i2c_reg->i2mod, 0x00);
+	out_8(&cpm->i2c_reg->i2brg, brg);
+	out_8(&cpm->i2c_reg->i2mod, mod);
 	out_8(&cpm->i2c_reg->i2com, I2COM_MASTER);	/* Master mode */

 	/* Disable interrupts. */
 	out_8(&cpm->i2c_reg->i2cmr, 0);
 	out_8(&cpm->i2c_reg->i2cer, 0xff);

+	freq = get_brgfreq() / prediv / 2 / (brg + ((mod & I2MOD_FLT) ? 5 : 3));
+	dev_dbg(&cpm->ofdev->dev, "i2mod 0x%02x, i2brg %u, actual freq %u\n",
+		mod, brg, freq);
 	return 0;

 out_muram: