Re: [PATCH][CFT][experimental] net/socket.c: use straight fdget/fdput (resend)

[Al Viro <viro@zeniv.linux.org.uk>]

On Mon, May 27, 2024 at 12:16:41AM +0100, Al Viro wrote:
> > What would make more sense is if you make the "fd_empty()" test be
> > about the _flags_, and then both the fp_empty() test and the test
> > inside fdput() would be testing the same things.
> 
> Umm...  What encoding would you use?

FWIW, _if_ we go for replacement of struct fd with struct-wrapped
unsigned long and actually steal another bit from pointer, we could
actually use that to encode errors as well...

If bit 0 is clear for file reference and set for an error, we could
use bit 1 to represent borrowed vs. cloned and bit 2 for need to
unlock...

overlayfs would be happier that way - we could have those functions
return struct fd directly, and pass the error value that way.
fdget() would report EBADF explicitly on empty slot - not a big
deal.  Hmm...

Let me play with that a bit and see if I can come up with something
tolerable.

We could add that in parallel to existing struct fd; we'd need a name
for replacement, though - anything like "rawfd" is asking for confusion
around fdget_raw(), and we'd need replacements for fdget() et.al.
that would yield those.

Alternatively, with a sane set of helpers we could actually _replace_
struct fd definition without a huge flagday commit - start with adding
#define fd_file(fd) ((fd).file)
#define fd_valid(fd) (fd_file(fd) != NULL)
convert the current users, one by one (fairly mechanical patches),
then change the definition of struct fd.  The last step would affect
	* definitions of helpers
	* adding fd-constructing macros (ERR_FD, etc.)
	* converting the few places that really construct those
suckers (fdget() et.al., overlayfs, a couple of explicit
initializers to {NULL, 0}).

It's not a huge amount of work and the flagday step would be fairly
small, but we'd probably need to split that over a couple of cycles -
helpers in one, along with renaming fd.file to e.g.
fd.__file_use_the_damn_helpers, then in the next cycle do the
switchover.

Hell knows; let me play with that a bit and let's see what falls out...