On Mon, Nov 04, 2024 at 10:30:15PM +0800, Yafang Shao wrote:
> When testing large folio support with XFS on our servers, we observed that
> only a few large folios are mapped when reading large files via mmap. This
> behavior occurs because large folio support is currently implemented only
> for sync readahead, not for async readahead. Consequently, while the
> first filemap fault may map to a large folio, subsequent filemap faults are
> mapped to regular folios. This can be verified with a simple test case, as
> shown below:
I awear I tested this and it worked at the time. Here's what's supposed
to happen:
You touch the first byte of the mapping, and there's no page in the page
cache. So we go into the sync readahead path, and force it to read two
PMDs. From your email, I assume this is succeeding. The readahead
flag gets set on the second PMD so that when we get to 2MB, we go into
the 'time to do more readahead' path, ie the 'do_async_mmap_readahead'
function you patch below.
Now, page_cache_async_ra() does this:
unsigned int order = folio_order(folio);
...
if (index == expected) {
ra->start += ra->size;
ra->size = get_next_ra_size(ra, max_pages);
ra->async_size = ra->size;
goto readit;
}
readit:
ractl->_index = ra->start;
page_cache_ra_order(ractl, ra, order);
So it should already be doing readahead of PMD size. Can you dig into
what's going wrong, now that you understand that this is supposed to
work already?
