Re: [PATCH v5 1/2] setup: distingush ENOENT from other stat errors

Tian Yuchen [off-list ref] writes:

Currently, 'read_gitfile_gently()' treats all 'stat()' failures as
generic errors. This prevents distinguishing between a missing file and
real errors like permission denied (fatal).

The above plan makes sense---you would split stat() error into two
different classes, start returning ERR_STAT_NOENT in addition to
ERR_STAT_FAILED, have the caller act on the new ERR_STAT_NOENT and
adjust the way it acts on ERR_STAT_FAILED, and if possible add tests
to make sure we react to failures from stat in an appropriate way
(but how? --- it is where my "if possible" comes from).  So I would
expect that the other patch would be to split ERR_NOT_A_FILE and add
ERR_IS_A_DIR, have the caller act on the new ERR_IS_A_DIR and adjust
the way it acts on ERR_NOT_A_FILE.

But then the proposed log message below says that in addition to
NOENT, it also deals with IS_A_DIR.  I do not mind doing these two
in the same patch, but I do prefer to see each patch to be complete.
If a callee is changed and starts returning different return values,
the callers must be also adjusted to react to these new return values.

Looking at the preimage of [v5 2/2], we stil check if dir->buf is a
plain vanilla ".git" directory only when read_gitfile_gently() returns
ERR_NOT_A_FILE with [v5 1/2] applied, but in the new world order
with this patch applied, shouldn't the caller deal with ".git" when
it gets ERR_IS_A_DIR and not ERR_NOT_A_FILE?

After applying this patch but before applying [v5 2/2], we lose the
ability to use plain vanilla ".git"?  That is the kind of thing I
meant by each patch to be complete.

So, I do not understand what the splitting the topic into two along
this axis is trying to achieve.

Thanks.