bug#25539: 26.0.50; filenotify-tests.el fails on Windows (32 and 64 bits)

[Michael Albinus]

Eli Zaretskii <eliz@gnu.org> writes:

Hi Eli,

>> 4 unexpected results:
>> FAILED file-notify-test04-file-validity
>> FAILED file-notify-test05-dir-validity
>> FAILED file-notify-test06-many-events
>> FAILED file-notify-test08-watched-file-in-watched-dir
>
> The problem is in file-notify--test-cleanup-p, and it happens only
> when the parent directory of the file(s) being watched is deleted.
>
> The root cause is that file-notify--test-cleanup-p expects the
> notification descriptor(s) to be deleted from the hash table
> maintained internally by filenotify.el, when the above happens.  But
> that doesn't work on Windows, where deleting the parent directory
> simply causes an error whose result is that the thread which watches
> the filesystem changes exits abnormally, but the event is not
> reported.  So in those cases the descriptor is not removed from the
> hash table.
>
> The changes below make the tests succeed, but maybe the above means we
> need to augment the w32notify implementation to clean up better in
> this case.  Michael?

Yes, catching the error somehow in w32notify (don't know how) and fire
up a `stopped' event is preferrable. I let it to you, whether this could
be implemented.

If this is not possible, maybe we shall call
`file-notify--rm-descriptor' in `file-notify-valid-p', when we detect
that the descriptor is not valid anymore, but it still exists.

> @@ -850,6 +850,8 @@ file-notify--test-with-events
>       ;; After deleting the parent directory, the descriptor must
>       ;; not be valid anymore.
>       (should-not (file-notify-valid-p file-notify--test-desc))
> +        (if (eq system-type 'windows-nt)
> +            (file-notify--rm-descriptor file-notify--test-desc))

It would be OK for me also to apply this patch. The test shall be

(if (string-equal (file-notify--test-library) "w32notify")

Otherwise, we would catch also the remote case when running the test on
MS Windows. And a respective comment might serve why we apply this.

Best regards, Michael.