How about making this function and other convenience wrappers static inline?

I want to keep the KBI for drivers both (a) compact (b) controlled. Making the wrappers inline means that e.g. make_dev_sv() is exposed by KBI while not used by the drivers directly. This contradicts both goals.

If the planned converion to use make_dev_s() went uneventful, I want to remove all other compat functions for 12.

OK

So make_dev_alias() is legacy because make_dev_alias_p() should be used instead? It is not somehow implemented via make_dev_s()?

Nit: s/that the/that/

Stray apostrophe after driver here?

This section is good, but I think you need to add a note of how make_dev_credf() is implemented. Ideally only one of make_dev_p() or make_dev_credf() would need a full example using the args, but right now make_dev_cred() and make_dev() are defined in terms of make_dev_credf() and it isn't clear that make_dev_credf() is a wrapper for make_dev_s().

s/later/latter/

I would also say "the driver should handle" instead of "driver should handle"

Maybe list both functions here since the return values of other legacy routines are also described, so:

If successful,
.Fn make_dev_p
and
.Fn make_dev_s
will return 0, otherwise they will return an error.

Other manpages that deal with legacy wrapper interfaces still enumerate all of the wrappers when describing return values (wait(2) is one example that comes to mind).

Bruce would sort 'error' before 'no_tags' I think?

Does this leak a reference on the periph in the old code?

You could perhaps use wrapper macros instead of static inlines if you really wanted. This would avoid exposing make_dev_sv(). I don't think you need/want to do this though, and since the old KBI has these as functions you want to keep them as functions for now. Presumably the current version is something we can MFC to older branches since it just expands the API which is nice for permitting drivers to be portable across branches.

Both make_dev_alias() and make_dev_alias_p() are implemented by make_dev_alias_v(). The KPI is getting out of control, and I want to trim it.

The aliases are not created with make_dev_s().

Bruce would never touch computer keyboard after looking at the scsi_pass.c, I think.

I do not quite follow the question. The code acquires a ref for each object which gives a handle to the periph. Old code does not suppose that make_dev(9) could fail, and unconditionally moved further. Since make_dev_s() can legitimately fail, I drop the reference for cdev, but do not touch other references.

If the acquire of the reference for alias failed, then I assumed that a reference for successfully created cdev must stay, this is how both old and new code are structured. Did you mean something else, perhaps ?

I do not expose make_dev_sv().

Yes, the whole patch is MFCable. Some day I want to finish convertion, this probably must be done before 11 is branched. In 12- time, I may consider removing all functions I declared legacy.