Re: [PATCH 01/32] iov_iter: Add ITER_MAPPING

[Al Viro]

On Mon, Jul 13, 2020 at 05:30:52PM +0100, David Howells wrote:
> Add an iterator, ITER_MAPPING, that walks through a set of pages attached
> to an address_space, starting at a given page and offset and walking for
> the specified amount of bytes.
> 
> The caller must guarantee that the pages are all present and they must be
> locked using PG_locked, PG_writeback or PG_fscache to prevent them from
> going away or being migrated whilst they're being accessed.
> 
> This is useful for copying data from socket buffers to inodes in network
> filesystems and for transferring data between those inodes and the cache
> using direct I/O.
> 
> Whilst it is true that ITER_BVEC could be used instead, that would require
> a bio_vec array to be allocated to refer to all the pages - which should be
> redundant if inode->i_pages also points to all these pages.
> 
> This could also be turned into an ITER_XARRAY, taking and xarray pointer
> instead of a mapping pointer.  It would be mostly trivial, except for the
> use of find_get_pages_contig() by iov_iter_get_pages*().
> 

My main problem here is that your iterate_mapping() assumes that STEP is
safe under rcu_read_lock(), with no visible mentioning of that fact.
Note, BTW, that iov_iter_for_each_range() quietly calls user-supplied
callback in such context.

Incidentally, do you ever have different steps for bvec and mapping?

> +	if (unlikely(iov_iter_is_mapping(i))) {
> +		/* We really don't want to fetch pages if we can avoid it */
> +		i->iov_offset += size;
> +		i->count -= size;
> +		return;

That's... not nice.  At the very least you want to cap size by i->count here
(and for discard case as well, while we are at it).