[Cluster-devel] [RFC v6 04/10] iomap: Add iomap_get_folio helper

[Matthew Wilcox <willy@infradead.org>]

On Sun, Jan 15, 2023 at 09:06:50AM -0800, Darrick J. Wong wrote:
> On Sun, Jan 15, 2023 at 09:01:22AM -0800, Darrick J. Wong wrote:
> > On Tue, Jan 10, 2023 at 01:34:16PM +0000, Matthew Wilcox wrote:
> > > On Tue, Jan 10, 2023 at 12:46:45AM -0800, Christoph Hellwig wrote:
> > > > On Mon, Jan 09, 2023 at 01:46:42PM +0100, Andreas Gruenbacher wrote:
> > > > > We can handle that by adding a new IOMAP_NOCREATE iterator flag and
> > > > > checking for that in iomap_get_folio().  Your patch then turns into
> > > > > the below.
> > > > 
> > > > Exactly.  And as I already pointed out in reply to Dave's original
> > > > patch what we really should be doing is returning an ERR_PTR from
> > > > __filemap_get_folio instead of reverse-engineering the expected
> > > > error code.
> > > 
> > > Ouch, we have a nasty problem.
> > > 
> > > If somebody passes FGP_ENTRY, we can return a shadow entry.  And the
> > > encodings for shadow entries overlap with the encodings for ERR_PTR,
> > > meaning that some shadow entries will look like errors.  The way I
> > > solved this in the XArray code is by shifting the error values by
> > > two bits and encoding errors as XA_ERROR(-ENOMEM) (for example).
> > > 
> > > I don't _object_ to introducing XA_ERROR() / xa_err() into the VFS,
> > > but so far we haven't, and I'd like to make that decision intentionally.
> > 
> > Sorry, I'm not following this at all -- where in buffered-io.c does
> > anyone pass FGP_ENTRY?  Andreas' code doesn't seem to introduce it
> > either...?
> 
> Oh, never mind, I worked out that the conflict is between iomap not
> passing FGP_ENTRY and wanting a pointer or a negative errno; and someone
> who does FGP_ENTRY, in which case the xarray value can be confused for a
> negative errno.
> 
> OFC now I wonder, can we simply say that the return value is "The found
> folio or NULL if you set FGP_ENTRY; or the found folio or a negative
> errno if you don't" ?

Erm ... I would rather not!

Part of me remembers that x86-64 has the rather nice calling convention
of being able to return a struct containing two values in two registers:

: Integer return values up to 64 bits in size are stored in RAX while
: values up to 128 bit are stored in RAX and RDX.

so maybe we can return:

struct OptionFolio {
	int err;
	struct folio *folio;
};