bug#46397: 27.1; Cannot delete buffer pointing to a file in a path that includes a file

[Eli Zaretskii <eliz@gnu.org>]

> From: Matt Armstrong <gmatta@gmail.com>
> Date: Tue, 09 Feb 2021 16:23:09 -0800
> Cc: 46397@debbugs.gnu.org, Paul Eggert <eggert@cs.ucla.edu>
> 
> > The backtrace is unsurprising:
> >
> > Debugger entered--Lisp error: (file-error "Unlocking file" "Not a directory" "/private/tmp/tmp/test.txt")
> >   kill-buffer("test.txt")
> >   funcall-interactively(kill-buffer "test.txt")
> >   call-interactively(kill-buffer nil nil)
> >   command-execute(kill-buffer)
> 
> I found that this behavior was introduced by Paul Egger's commit
> 9dc306b1db0, discussed in Bug#37389.  I've cc'd Paul.
> 
> Paul's commit changed unlock_file() (from src/filelock.cc) to report
> errors from unlink(), excempting only ENOENT. This bug demonstrates a
> way to induce an ENOTDIR error. I've attached a patch that ignores
> ENOTDIR as well, which is the most conservative fix I can think of. It
> also seems in-line with Paul's original intent, since he was saying that
> both ENOENT and ENOTDIR are usually "tame."

I think instead of ignoring some errors, we should allow the user to
get out of these situations, after showing the error.  That is,
instead of silently ignoring the error on some low level, we should
propagate it to userlock.el and allow the user to decide whether
he/she wants to ignore the error or do something about that.  These
errors could mean something innocent, or they could mean something
more serious, and we shouldn't second-guess which one is it.

Letting the user decide will also seamlessly solve any other type of
error that could happen when we try to delete the lock file.

So how about coming up with a patch that shows the error message to
the users and asks them whether or not to ignore that and kill the
buffer?

Thanks.