Re: [PATCH v9 12/25] mm: Move end_index check out of readahead loop

[Eric Biggers <ebiggers@kernel.org>]

On Fri, Mar 20, 2020 at 11:11:32AM -0700, Matthew Wilcox wrote:
> On Fri, Mar 20, 2020 at 11:00:17AM -0700, Eric Biggers wrote:
> > On Fri, Mar 20, 2020 at 10:30:40AM -0700, Matthew Wilcox wrote:
> > > On Fri, Mar 20, 2020 at 09:58:28AM -0700, Eric Biggers wrote:
> > > > On Fri, Mar 20, 2020 at 07:22:18AM -0700, Matthew Wilcox wrote:
> > > > > +	/* Avoid wrapping to the beginning of the file */
> > > > > +	if (index + nr_to_read < index)
> > > > > +		nr_to_read = ULONG_MAX - index + 1;
> > > > > +	/* Don't read past the page containing the last byte of the file */
> > > > > +	if (index + nr_to_read >= end_index)
> > > > > +		nr_to_read = end_index - index + 1;
> > > > 
> > > > There seem to be a couple off-by-one errors here.  Shouldn't it be:
> > > > 
> > > > 	/* Avoid wrapping to the beginning of the file */
> > > > 	if (index + nr_to_read < index)
> > > > 		nr_to_read = ULONG_MAX - index;
> > > 
> > > I think it's right.  Imagine that index is ULONG_MAX.  We should read one
> > > page (the one at ULONG_MAX).  That would be ULONG_MAX - ULONG_MAX + 1.
> > > 
> > > > 	/* Don't read past the page containing the last byte of the file */
> > > > 	if (index + nr_to_read > end_index)
> > > > 		nr_to_read = end_index - index + 1;
> > > > 
> > > > I.e., 'ULONG_MAX - index' rather than 'ULONG_MAX - index + 1', so that
> > > > 'index + nr_to_read' is then ULONG_MAX rather than overflowed to 0.
> > > > 
> > > > Then 'index + nr_to_read > end_index' rather 'index + nr_to_read >= end_index',
> > > > since otherwise nr_to_read can be increased by 1 rather than decreased or stay
> > > > the same as expected.
> > > 
> > > Ooh, I missed the overflow case here.  It should be:
> > > 
> > > +	if (index + nr_to_read - 1 > end_index)
> > > +		nr_to_read = end_index - index + 1;
> > > 
> > 
> > But then if someone passes index=0 and nr_to_read=0, this underflows and the
> > entire file gets read.
> 
> nr_to_read == 0 doesn't make sense ... I thought we filtered that out
> earlier, but I can't find anywhere that does that right now.  I'd
> rather return early from __do_page_cache_readahead() to fix that.
> 
> > The page cache isn't actually supposed to contain a page at index ULONG_MAX,
> > since MAX_LFS_FILESIZE is at most ((loff_t)ULONG_MAX << PAGE_SHIFT), right?  So
> > I don't think we need to worry about reading the page with index ULONG_MAX.
> > I.e. I think it's fine to limit nr_to_read to 'ULONG_MAX - index', if that makes
> > it easier to avoid an overflow or underflow in the next check.
> 
> I think we can get a page at ULONG_MAX on 32-bit systems?  I mean, we can buy
> hard drives which are larger than 16TiB these days:
> https://www.pcmag.com/news/seagate-will-ship-18tb-and-20tb-hard-drives-in-2020
> (even ignoring RAID devices)

The max file size is ((loff_t)ULONG_MAX << PAGE_SHIFT) which means the maximum
page *index* is ULONG_MAX - 1, not ULONG_MAX.

Anyway, I think we may be making this much too complicated.  How about just:

	pgoff_t i_nrpages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);

	if (index >= i_nrpages)
		return;
	/* Don't read past the end of the file */
	nr_to_read = min(nr_to_read, i_nrpages - index);

That's 2 branches instead of 4.  (Note that assigning to i_nrpages can't
overflow, since the max number of pages is ULONG_MAX not ULONG_MAX + 1.)

- Eric