Re: [BLOCK] bh: Ensure bh fits within a page

[Steven Rostedt <rostedt@goodmis.org>]

On Tue, 2006-08-01 at 10:36 -0700, Christoph Lameter wrote:
> On Mon, 31 Jul 2006, Andrew Morton wrote:
> 
> > > Sure, this particular instance is in journal_write_metadata_buffer
> > > where the bh may be constructed from kmalloc memory (search for the
> > > call to jbd_rep_kmalloc).  Because the memory returned by kmalloc
> > > may straddle a page (when slab debugging is enabled that is), this
> > > causes a broken bh to be injected into submit_bh.
> > > 
> > 
> > Crap, that's hard to fix.   Am I allowed to blame submit_bh()? ;)
> > 
> > uhm, we don't want to lose kmalloc redzoning, so I guess we need to create
> > on-demand ext3-private slab caches for 1024, 2048, and 4096 bytes.  With
> > the appropriate slab flags to defeat the redzoning.
> 
> The slab allocator gives no guarantee that a structure is not straddling a 
> page boundary regardless of debug or not. It may just happen that the 
> objects are arranged if kmem_cache_cretae() is called with certain 
> parameters. Another arch with other cacheline alignment and another page 
> size may arrange the objects differently.

If you set the alignment for ext3 the same as the size (ie 1024, 2048,
4096 for the above respectively) then wouldn't that guarantee not
straddling a page?

-- Steve