On Fri, Aug 03, 2007 at 07:47:43PM +0400, Sergei Shtylyov wrote:

Hello.

David Gibson wrote:

quoted

quoted

quoted

Index: working-2.6/Documentation/powerpc/booting-without-of.txt
===================================================================

--- working-2.6.orig/Documentation/powerpc/booting-without-of.txt	2007-07-30 17:07:14.000000000 +1000
+++ working-2.6/Documentation/powerpc/booting-without-of.txt	2007-07-30 17:07:14.000000000 +1000

@@ -1757,45 +1757,23 @@ platforms are moved over to use the flat

		};
	};

-    j) Flash chip nodes
+   j) CFI or JEDEC memory-mapped NOR flash

    Flash chips (Memory Technology Devices) are often used for solid state
    file systems on embedded devices.

quoted

quoted

quoted

-    Required properties:
+     - compatible : should contain the specific model of flash chip(s) used
+       followed by either "cfi-flash" or "jedec-flash"

quoted

quoted

   This "compatible" prop (and the node in whole) doesn't say a
thing about how the flash is mapped into the CPU address space.  I
strongly disagree that this node provides enough info to select a
driver. :-/

quoted

To be honest, I'm not sure that describing the mapping is really the
job of the compatible property.  That the flash is mapped into the
address space is kind of implicit in the way the reg interacts with
the parents' ranges property.

    Ah, I keep forgetting about implied 1:1 parent/child address 
correspondence... :-<
    But does it really imply how the (low) address bits of the *child* bus 
("ebc" in this case) are connected to the chip?  I don't think so...

quoted

Can you describe some of the options for *not* direct mapped flash
chips - I can't reasonably come up with a way of describing the
distinction when I've never seen NOR flash other than direct mapped.

    You're lucky to live in the single-endian world.  Once you're in bi-endian 
one, all kinds of strange mappings become possible.  I've seen the MIPS 
mapping driver which required swapping flash bytes in s/w in BE mode, i.e. 
couldn't be served by physmap, yet that's not all...  here's the code of its 
map read*() methods:

Aha!  Ok, now I understand the sorts of situations you're talking
about.  By "not direct mapped", I thought you were talking about some
kind of access via address/data registers on some indirect bus
controller, rather than weird variations on endianness and
bit-swizzling.

Hrm.. this is a property of how the flash is wired onto the bus,
rather than of the flash chips themselves, so I'm not entirely sure
where description of it belongs.

Simplest option seems to me to add a property "endianness" or
"bit-swizzling" or something which can be defined to describe some odd
connections.  If absent we'd default to direct mapping.  Segher, is
that idea going to cause you to scream?

quoted

quoted

quoted

+     - reg : Address range of the flash chip
+     - bank-width : Width (in bytes) of the flash bank.  Equal to the device width
+       times the number of interleaved chips.
+     - device-width : (optional) Width of a single flash chip.  If omitted,
+       assumed to be equal to 'bank-width'.

quoted

quoted

   Why then not just introduce the "interleave" prop and obsolete the
"bank-width"?

quoted

Because they're equivalent information, and bank-width is what the
code ends up actually caring about.  I don't see any reason to prefer
giving the interleave.

    Well, "device-width" is not the thing that we care about either. ;-)

Well, yes but we need to encode it somehow.  Segher preferred
device-width to interleave, because it's closer to a description of
the actual hardware, rather than an abstration decribing its wiring.

In other words I *still* don't see any reason to prefer giving the
interleave.

quoted

quoted

quoted

Index: working-2.6/drivers/mtd/maps/physmap_of.c
===================================================================

--- working-2.6.orig/drivers/mtd/maps/physmap_of.c	2007-07-30 17:07:11.000000000 +1000
+++ working-2.6/drivers/mtd/maps/physmap_of.c	2007-07-30 17:07:14.000000000 +1000

[...]

quoted

quoted

quoted

+	for (pp = dp->child, i = 0 ; pp; pp = pp->sibling, i++) {
+		const u32 *reg;
+		const char *name;
+		const void *ro;

   We hardly need the above 2 variables.

quoted

They're all used...

    I meant that there's no necessity in them.

By which you mean....?

quoted

quoted

   Oh, I see that the new partition representation have really simplified
parsing -- this function is hardly shorter than the old one... :-P

quoted

They new format isn't supposed to be simpler to parse: it's supposed
to be more flexible if we ever need to add more per-partition
information than just the offset / size / read-only.

    Well, if we ever need that indeed... :-)

[...]

quoted

quoted

quoted

@@ -221,6 +297,14 @@ err_out:

static struct of_device_id of_physmap_match[] = {
	{
+		.compatible	= "cfi-flash",
+		.data		= (void *)"cfi_probe",
+	},
+	{
+		.compatible	= "jedec-flash",
+		.data		= (void *)"jedec_probe",
+	},
+	{

quoted

quoted

   This would also trigger on non-linearly mapped CFI or JEDEC
flashes which is not a good idea -- however, as we're using the MTD
probing code anyway, it will just fail, so it's not luckily is not a
fatal design flaw.

quoted

Well, if there's nothing else to distinguish them.  Which there isn't
yet, but will need to be: see above about incomplete - helpful
suggestions about how to describe the mapping would be more useful
than merely pointing out the lack.

    I was thinking about adding "direct-mapped" prop... but maybe adding 
"ranges" to the parent node (that's "ebc") would indeed ensure that the flash 
is mapped 1:1 to the EBC's parent bus also.

The ebc already has a ranges property.  But that can't describe the
sorts of bit-swizzling you're talking about.

quoted

quoted

quoted

Index: working-2.6/arch/powerpc/boot/dts/ebony.dts
===================================================================

--- working-2.6.orig/arch/powerpc/boot/dts/ebony.dts	2007-07-30 17:07:14.000000000 +1000
+++ working-2.6/arch/powerpc/boot/dts/ebony.dts	2007-07-30 17:07:14.000000000 +1000

quoted

quoted

[...]

quoted

quoted

quoted

@@ -158,14 +161,20 @@

				};

quoted

quoted

quoted

				large-flash@2,0 {

quoted

quoted

   Hmm... what does @2,0 mean? :-O

quoted

EBC chip select 2, offset 0.

    Well, so this node is under some kind of local bus node -- that's good.
Didn't know that the spec allows commas after @...

Well, now you do.  I believe USB usually uses that format, too.

quoted

quoted

quoted

-					device_type = "rom";
-					compatible = "direct-mapped";
-					probe-type = "JEDEC";
+					compatible = "jedec-flash";
					bank-width = <1>;
-					partitions = <0 380000
-						      380000 80000>;
-					partition-names = "fs", "firmware";
+//					partitions = <0 380000
+//						      380000 80000>;
+//					partition-names = "fs", "firmware";
					reg = <2 0 400000>;
+					#address-cells = <1>;
+					#size-cells = <1>;

quoted

quoted

   Heh...

quoted

Yeah, that bit's a bit ugly, I'll grant you.

    Don't we need "ranges" here, at least from the formal PoV -- as the parent 
and child address spaces differ? I know the physmap_of parser doesn't care but 
still...

That's one I've wondered about.  To describe the partitions address
space as lying (ultimately) in the physical address space, which in a
sense it does, yes we'd need a ranges property here.  But we also have
a 'reg' at the top level which would overlap with that hypothetical
ranges which would be weird.  Or we could exclude the top-level reg,
but then that's a pain if we do want to map the flash as a whole.

So I left out ranges, on the grounds that there isn't actually
anything at present which will attempt to access flash partitions
"generically" as a device tree device. 

I'm not sold on this approach, but I haven't heard you give a better
argument yet.

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson