maximum block size in buffer_head

maximum block size in buffer_head

[David Mosberger]

The current Linux kernel (both 2.4.xx and 2.5.xx) declare the b_size
member in struct buffer_head as an "unsigned short".  This obviously
limits the maximum block size to something less than 65536.  This is
bad because on some platforms (e.g., ia64), the page size can be up to
64KB large.

Two questions:

 - does anyone object to widening b_size to "unsigned int"?

 - does anyone know of any other code paths where the block
   size is assumed to fit into 16 bits?

Thanks,

	--david

Re: maximum block size in buffer_head

[Andrew Morton]

David Mosberger wrote:
> 
> The current Linux kernel (both 2.4.xx and 2.5.xx) declare the b_size
> member in struct buffer_head as an "unsigned short".  This obviously
> limits the maximum block size to something less than 65536.  This is
> bad because on some platforms (e.g., ia64), the page size can be up to
> 64KB large.
> 
> Two questions:
> 
>  - does anyone object to widening b_size to "unsigned int"?

Sounds OK to me for 2.5.

I do have vague plans to remove b_data and to replace all instances
with buffer_data(), which would kmap the page and calculate the address.
Not that I've really thought this through...

This may lead to the introduction of a `b_offset'.  Which would be 16
bits, which would snuggle in with the 16-bit b_size.   But I would
scale both these values by 256 or 512,  for the reasons which you
identify.

For 2.4, a 32-bit b_size would push sizeof(buffer_head) from 96 up
to 100 bytes, which does not pack as well into the slab.  This would
be an intensely unpopular move.  So you'd have to ifdef it.  Which
makes it an ia64-only problem, which greatly improves your merge
chances ;)

>  - does anyone know of any other code paths where the block
>    size is assumed to fit into 16 bits?

Not off the top, but they're probably there.

-

Re: maximum block size in buffer_head

[David Mosberger]

>>>>> On Thu, 09 May 2002 22:38:13 -0700, Andrew Morton <akpm@zip.com.au> said:

  Andrew> Sounds OK to me for 2.5.

Will you make the change then?  I'd appreciate that.

  Andrew> For 2.4, a 32-bit b_size would push sizeof(buffer_head) from
  Andrew> 96 up to 100 bytes, which does not pack as well into the
  Andrew> slab.  This would be an intensely unpopular move.  So you'd
  Andrew> have to ifdef it.  Which makes it an ia64-only problem,
  Andrew> which greatly improves your merge chances ;)

For 2.4, I suspect I need to keep this in the ia64 patch anyhow
because it is also necessary to fix scsi_dma.c so it can deal with
allocation bitmaps that are bigger than 32 bits.

I believe ia64 is currently the only platform that supports 64KB
pages.  If so, there is probably not much point trying to get this
past Marcelo (and the patch is large enough to have the potential for
introducing new bugs).  If another platform cares about this, please
speak up.

  >> - does anyone know of any other code paths where the block size
  >> is assumed to fit into 16 bits?

  Andrew> Not off the top, but they're probably there.

I'll keep an eye on it.  It's working fine so far, but that doesn't
replace a code audit.  If you happen to bump into something during
your normal work, please let me know.

Thanks,

	--david