[RFC PATCH] fs/hugetlb: Fix UBSAN warning reported on hugetlb

[Mike Kravetz]

On 10/26/22 10:50, Aristeu Rozanski wrote:
> On Thu, Sep 08, 2022 at 10:29:59PM +0530, Aneesh Kumar K V wrote:
> > On 9/8/22 10:23 PM, Matthew Wilcox wrote:
> > > On Thu, Sep 08, 2022 at 12:56:59PM +0530, Aneesh Kumar K.V wrote:
> > >> +++ b/fs/dax.c
> > >> @@ -1304,7 +1304,7 @@ EXPORT_SYMBOL_GPL(dax_zero_range);
> > >>  int dax_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
> > >>  		const struct iomap_ops *ops)
> > >>  {
> > >> -	unsigned int blocksize = i_blocksize(inode);
> > >> +	size_t blocksize = i_blocksize(inode);
> > >>  	unsigned int off = pos & (blocksize - 1);
> > > 
> > > If blocksize is larger than 4GB, then off also needs to be size_t.
> > > 
> > >> +++ b/fs/iomap/buffered-io.c
> > >> @@ -955,7 +955,7 @@ int
> > >>  iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
> > >>  		const struct iomap_ops *ops)
> > >>  {
> > >> -	unsigned int blocksize = i_blocksize(inode);
> > >> +	size_t blocksize = i_blocksize(inode);
> > >>  	unsigned int off = pos & (blocksize - 1);
> > > 
> > > Ditto.
> > > 
> > > (maybe there are others; I didn't check closely)
> > 
> > Thanks. will check those. 
> > 
> > Any feedback on statx? Should we really fix that?
> > 
> > I am still not clear why we chose to set blocksize = pagesize for hugetlbfs.
> > Was that done to enable application find the hugetlb pagesize via stat()? 
> 
> I'd like to know that as well. It'd be easier to just limit the hugetlbfs max
> blocksize to 4GB. It's very unlikely anything else will use such large
> blocksizes and having to introduce new user interfaces for it doesn't sound
> right.

I was not around hugetlbfs when the decision was made to set 'blocksize =
pagesize'.  However, I must say that it does seem to make sense as you
can only add or remove entire hugetlb pages from a hugetlbfs file.  So,
the hugetlb page size does seem to correspond to the meaning of filesystem
blocksize.

Does any application code make use of this?  I can not make a guess.
-- 
Mike Kravetz