On 2025/9/12 09:09, Gao Xiang wrote:

On 2025/9/12 08:06, Gao Xiang wrote:

quoted

On 2025/9/12 03:45, Joanne Koong wrote:

quoted

On Thu, Sep 11, 2025 at 8:29 AM Gao Xiang [off-list ref] wrote:

...

quoted

quoted

But if FUSE or some other fs later needs to request L2P information
in their .iomap_begin() and need to send L2P requests to userspace
daemon to confirm where to get the physical data (maybe somewhat
like Darrick's work but I don't have extra time to dig into that
either) rather than just something totally bypass iomap-L2P logic
as above, then I'm not sure the current `iomap_iter->private` is
quite seperate to `struct iomap_read_folio_ctx->private`, it seems

If in the future this case arises, the L2P mapping info is accessible
by the read callback in the current design. `.read_folio_range()`
passes the iomap iter to the filesystem and they can access
iter->private to get the L2P mapping data they need.

The question is what exposes to `iter->private` then, take
an example:

struct file *file;

your .read_folio_range() needs `file->private_data` to get
`struct fuse_file` so `file` is kept into
`struct iomap_read_folio_ctx`.

If `file->private_data` will be used for `.iomap_begin()`
as well, what's your proposal then?

Duplicate the same `file` pointer in both
`struct iomap_read_folio_ctx` and `iter->private` context?

It's just an not-so-appropriate example because
`struct file *` and `struct fuse_file *` are widely used
in the (buffer/direct) read/write flow but Darrick's work
doesn't use `file` in .iomap_{begin/end}.

But you may find out `file` pointer is already used for
both FUSE buffer write and your proposal, e.g.

buffer write:
  /*
   * Use iomap so that we can do granular uptodate reads
   * and granular dirty tracking for large folios.
   */
  written = iomap_file_buffered_write(iocb, from,
                                      &fuse_iomap_ops,
                                      &fuse_iomap_write_ops,
                                      file);

And your buffer write per-fs context seems just use
`iter->private` entirely instead to keep `file`.

I just try to say if there is a case/feature which needs
something previously in `struct iomap_read_folio_ctx` to
be available in .iomap_{begin,end} too, you have to either:
  - duplicate this in `iter->private` as well;
  - move this to `iter->private` entirely.

The problem is that both `iter->private` and
`struct iomap_read_folio_ctx` are filesystem-specific,
I can only see there is no clear boundary to leave something
in which one.  It seems just like an artificial choice.

Thanks,
Gao Xiang

quoted

quoted

quoted

both needs fs-specific extra contexts for the same I/O flow.

I think the reason why `struct iomap_read_folio_ctx->private` is
introduced is basically previous iomap filesystems are all
bio-based, and they shares `bio` concept in common but
`iter->private` was not designed for this usage.

But fuse `struct iomap_read_folio_ctx` and
`struct fuse_fill_read_data` are too FUSE-specific, I cannot
see it could be shared by other filesystems in the near future,
which is much like a single-filesystem specific concept, and
unlike to `bio` at all.

Currently fuse is the only non-block-based filesystem using iomap but
I don't see why there wouldn't be more in the future. For example,
while looking at some of the netfs code, a lot of the core
functionality looks the same between that and iomap and I think it
might be a good idea to have netfs in the future use iomap's interface
so that it can get the large folio dirty/uptodate tracking stuff and
any other large folio stuff like more granular writeback stats
accounting for free.

I think you need to ask David on this idea, I've told him to
switch fscache to use iomap in 2022 before netfs is fully out [1],
but I don't see it will happen.

[1] https://lore.kernel.org/linux-fsdevel/YfivxC9S52FlyKoL@B-P7TQMD6M-0146/ (local)

Thanks,
Gao Xiang