Re: Data corruption problem with swapfiles and THP

["Huang, Ying" <ying.huang@intel.com>]

Matthew Wilcox <willy@infradead.org> writes:

> There is an assumption in the swap writepage path that a THP is physically
> contiguous on swap:
>
>         bio->bi_iter.bi_sector = swap_page_sector(page);
>         bio->bi_opf = REQ_OP_WRITE | REQ_SWAP | wbc_to_write_flags(wbc);
>         bio->bi_end_io = end_write_func;
>         bio_add_page(bio, page, thp_size(page), 0);
>
> As far as I can tell, this is not necessarily true.  If a file is not
> contiguous, we can have an extent which is 1MB long followed by an extent
> somewhere else on storage that's 1MB long.  When we try to write a 2MB
> page to swap, we overwrite whatever's on the block device after that
> first 1MB extent.
>
> (Came across this by code examination while looking at getting rid of
> the bio path entirely; no attempt has been made to produce this problem;
> something else may prevent it from actually happening)

Yes.  THP needs to be split firstly before swapping out to a swap device
backed by a file.  Please take a look at the get_swap_pages()

		if (size == SWAPFILE_CLUSTER) {
			if (si->flags & SWP_BLKDEV)
				n_ret = swap_alloc_cluster(si, swp_entries);
		} else
			n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
						    n_goal, swp_entries);

If the swap device is backed by a file, si->flags & SWP_BLKDEV == 0,
only normal swap entry (not huge) can be allocated.  This will result
that the THP is split.

Best Regards,
Huang, Ying