Large folios and filemap_get_folios_contig()

[Qu Wenruo <quwenruo.btrfs@gmx.com>]

Hi,

Recently I hit a bug when developing the large folios support for btrfs.

That we call filemap_get_folios_contig(), then lock each returned folio.
(We also have a case where we unlock each returned folio)

However since a large folio can be returned several times in the batch,
this obviously makes a deadlock, as btrfs is trying to lock the same
folio more than once.

Then I looked into the caller of filemap_get_folios_contig() inside
mm/gup, and it indeed does the correct skip.


This makes me wonder, since we have large folios, why we still go
filemap_get_folios_contig() and skip duplicated large folios?

Isn't the purpose of large folios to handle a much large range in just
one go, without going through multiple pages?


And there are only 3 call sites, two of them are nilfs and ramfs,
neither support large folios, the only caller with large folio support
is the memfd_pin_folios(), which skip duplicated folios manually.

I'm wondering if it's possible to make filemap_get_folios_contig() to
avoid filling the batch with duplicated folios completely?

Thanks,
Qu