linux-apus-devel

get_free_pages or kmalloc alignment in PPC

[<fh...@at...>]

Is there any info on how to align memory in the Linux PPC
kernel.  Actually info on how get_free_pages aligns memory.
I'm guessing that this is the major problem with getting the
PuP SCSI working.

It looks like what most of the drivers I looked at do is add
a little extra to the size of the chunk and then move the
pointer so it's on a 32 bit boundary.  Seems simple enough.
What if the chip that wants to do DMA only needs 16 bit
(or 8 bit) alignment?  It seems that then there is a further  problem
where the memory needs to be "synced" so the 
16bit or 8 bit parts line up with the 32bit parts.

Fred

Re: get_free_pages or kmalloc alignment in PPC

[Geert U. <ge...@li...>]

On Sun, 26 Nov 2000 fh...@at... wrote:
> Is there any info on how to align memory in the Linux PPC
> kernel.  Actually info on how get_free_pages aligns memory.
> I'm guessing that this is the major problem with getting the
> PuP SCSI working.
> 
> It looks like what most of the drivers I looked at do is add
> a little extra to the size of the chunk and then move the
> pointer so it's on a 32 bit boundary.  Seems simple enough.
> What if the chip that wants to do DMA only needs 16 bit
> (or 8 bit) alignment?  It seems that then there is a further  problem
> where the memory needs to be "synced" so the 
> 16bit or 8 bit parts line up with the 32bit parts.

Doesn't get_free_pages() simply return page-aligned data?

Gr{oetje,eeting}s,

						Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- ge...@li...

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
							    -- Linus Torvalds

Re: get_free_pages or kmalloc alignment in PPC

[<fh...@at...>]

In <Pin...@ca...rg>, on
11/27/00 
   at 08:30 AM, Geert Uytterhoeven <ge...@li...> said:

>On Sun, 26 Nov 2000 fh...@at... wrote:

>> 
>> It looks like what most of the drivers I looked at do is add
>> a little extra to the size of the chunk and then move the
>> pointer so it's on a 32 bit boundary.  Seems simple enough.

BTW  I have been playing with a newer version of the LSI
SCSI driver.  In those versions, the data structure is allocated using a
power of two buddy malloc algorithm.  In the 53c770  driver I started with
the data structure is statically allocated. What would be the difference. 
i.e.  you have

( I call this statically allocated. )

struct mystruct {
    <blah>
}

instead of 

typedef struct mystruct {
    <blah>
} MYS;

    MYS *mys;

    mys = (MYS *)malloc( sizeof(MYS) );

I noticed that whenever I tried using the malloc procedure like the new
drivers, I would lose DMA to the data structure.

>> What if the chip that wants to do DMA only needs 16 bit
>> (or 8 bit) alignment?  It seems that then there is a further  problem
>> where the memory needs to be "synced" so the 
>> 16bit or 8 bit parts line up with the 32bit parts.

>Doesn't get_free_pages() simply return page-aligned data?

Hmmm.  I don't know much about get_free_pages.  The Linux
Device Driver book says nothing about how it aligns data.

Some of the driver codes I have looked at do strange things
to align memory accesses.  This would appear, on the surface, to be a
rather large short coming in the Linux kernel if true. But I am a dummy so
it's probably not. :)

Fred