RE: [PATCH] erofs: allocate file-backed mount requests from a bio_set

[Daejun Park]

> - Mount already rejects backing files on stacked fses and on
>   file-backed EROFS. A backing read cannot reach this pool again in
>   the same task.

This relies on the stacked fs check, and Sashiko pointed out that
inode_share gets around it.

A backing file on a block-backed EROFS passes the check, since that sb
has s_bdev. If that EROFS is mounted with inode_share, reads of the
file go through erofs_ishare_file_read_iter(), and erofs_real_inode()
may pick an inode of a file-backed EROFS in the same domain. So a
backing read can start another file-backed read in the same task,
while its own request is still held. I reproduced this in a QEMU guest
with the image on such an EROFS, and also with a device= blob there.
This is from the mount:

  vfs_iocb_iter_read
  erofs_fileio_rq_submit
  erofs_fileio_iomap_submit_read
  iomap_readahead
  erofs_readahead
  ...
  filemap_read
  erofs_ishare_file_read_iter
  vfs_iocb_iter_read
  erofs_fileio_rq_submit
  erofs_read_meta_folio
  ...
  erofs_fc_fill_super

With this patch both requests come from erofs_fileio_bioset, so under
memory pressure the inner allocation can wait on the pool while the
outer request still holds an element of it.

The nesting itself is older than this series. File-backed mounts can
stack this way since commit 34096ba919fd ("erofs: support unencoded
inodes for page cache share"). In v2 I will add a separate fix for
that, with a Fixes tag, which treats an EROFS mounted with inode_share
as a stacked fs, and update this paragraph.

Daejun