[PATCH] jffs2 summary allocation

[Andrew Morton]

On Sat, 5 Apr 2008 10:29:25 +0900 "Kyungmin Park" <kmpark at infradead.org> wrote:

> On Sat, Apr 5, 2008 at 10:11 AM, Josh Boyer <jwboyer at gmail.com> wrote:
> > On Fri, 2008-04-04 at 16:58 -0700, David Brownell wrote:
> >  > On Friday 04 April 2008, Josh Boyer wrote:
> >  > >
> >  > > >   ... This means specifically that you may _not_ use the
> >  > > >   memory/addresses returned from vmalloc() for DMA.  ...
> >  > > >
> >  > > > So I'm rather surprised to see *ANY* kernel code trying to do
> >  > > > that.  That rule has been in effect for many, many years now.
> >  > >
> >  > > I don't think it was intentional.  You're going through several layers
> >  > > here:
> >  > >
> >  > > JFFS2 -> mtd parts -> mtd dataflash -> atmel_spi.
> >  > >
> >  > > Typically MTD drivers aren't doing DMAs to flash and JFFS2 has no idea
> >  > > which particular chip driver is being used because it's abstracted by
> >  > > MTD.
> >  >
> >  > That's true ... although I can imagine using DMA to
> >  > avoid dcache trashing if its setup cost is low enough,
> >  > with either NAND or NOR chips.
> >  >
> >  > Still:  in this context vmalloc() is wrong.
> >
> >  Agreed.  One issue is that the summary code allocates a buffer that
> >  equals the eraseblock size of the underlying MTD device.  For larger
> >  NAND chips, that may be up to 256KiB.  I believe this is within the
> >  allowable kmalloc size for most architectures these days, but the
> >  summary code is 3 years old and was likely expecting a smaller limit.
> >  And there is always the question on whether finding that much contiguous
> >  memory will be an issue.

Yes.  This is why I'm reluctant to whizz this patch into 2.6.25.  It'll
break more than it fixes.

> In MLC chips it goes up to 512KiB. It means it can't allocate the
> eraseblock size memory with kmalloc().
> In ARM environment I can't see the 256KiB or more memory allocation
> with kmalloc().
> So I now changed the kmalloc eraseblock to vmalloc at both jffs2 and mtd-utils.

Does this eraseblock really really really need to be a single
virtually-contiguous hunk of kernel memory?  Or was that just easy to do at
the time?



This problem comes up pretty often.  Rather than open-coding it yet again
it'd be nice to have a little bit of library code which manages an array of
pages and which has accessors for common operations like
read/write-u8/u16/u32/u64, memset, memcpy, etc.

Then again, given that this memory is often fed into IO subsystems, perhaps
we should do this by adding more accessors and helpers to
scatterlists/sg_table.  Unfortunately they're not presently well set up for
random access.