Re: [PATCH] udf: limit the maximum number of allocation extents

[Jan Kara <jack@suse.cz>]

On Mon 14-12-15 21:34:26, Jan Kara wrote:
> On Fri 11-12-15 15:54:16, Vegard Nossum wrote:
> > Hit this kernel hang too while fuzzing. Please see this as a tentative
> > patch indicating where the problem is -- I don't really know much about
> > UDF or what an allocation extent is or whether there are more problems
> > in the same neighbourhood. It looks like udf_truncate_extents() might
> > also have a similar problem?
> > 
> > Signed-off-by: Vegard Nossum <vegard.nossum@oracle.com>
> > Cc: stable@vger.kernel.org
> > Cc: Jan Kara <jack@suse.com>
> > Cc: Quentin Casasnovas <quentin.casasnovas@oracle.com>
> > Cc: Andrew Morton <akpm@linux-foundation.org>
> > ---
> >  fs/udf/inode.c | 13 +++++++++++++
> >  1 file changed, 13 insertions(+)
> > 
> > diff --git fs/udf/inode.c fs/udf/inode.c
> > index 8d0b3ad..e1875f5 100644
> > --- fs/udf/inode.c
> > +++ fs/udf/inode.c
> > @@ -2047,13 +2047,26 @@ void udf_write_aext(struct inode *inode, struct extent_position *epos,
> >  		epos->offset += adsize;
> >  }
> >  
> > +/*
> > + * Maximum number of allocation extents. The chosen number is
> > + * arbitrary - just that we hopefully don't limit any real use
> > + * but avoid looping for too long on corrupted media.
> > + */
> > +#define UDF_MAX_AEXT_NESTING 4096
> > +
> 
> I don't like to limit the number of indirect extents in a file. Although
> 4096 is quite a bit, there is a real chance it won't be enough for some
> usecases (although I agree that such usecases would be very slow with the
> current implementation of UDF anyway). What I'd prefer is to limit the
> number of indirect extents to maximum possible sane number. Something like:
> 
> (inode->i_size >> inode->i_blkbits) / (extents_per_block) + 1
> 
> That way we are sure we don't limit any real use case and we also avoid
> infinite loops.

In the end I have realized that this is actually only about the case where
indirect extent has the first extent which is also indirect. Such case
shouldn't happen in practice at all so I have even reduced the limit and
committed the patch.

								Honza

> >  int8_t udf_next_aext(struct inode *inode, struct extent_position *epos,
> >  		     struct kernel_lb_addr *eloc, uint32_t *elen, int inc)
> >  {
> >  	int8_t etype;
> > +	unsigned int indirections = 0;
> >  
> >  	while ((etype = udf_current_aext(inode, epos, eloc, elen, inc)) ==
> >  	       (EXT_NEXT_EXTENT_ALLOCDECS >> 30)) {
> > +		if (++indirections > UDF_MAX_AEXT_NESTING) {
> > +			udf_err(inode->i_sb, "too many AEXTs (max %u supported)\n", UDF_MAX_AEXT_NESTING);
> > +			return -1;
> > +		}
> > +
> >  		int block;
> >  		epos->block = *eloc;
> >  		epos->offset = sizeof(struct allocExtDesc);
> > -- 
> > 1.9.1
> > 
> > 
> -- 
> Jan Kara <jack@suse.com>
> SUSE Labs, CR
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR