Actuallly, Hannes's patch text and problem description are confusingly inconsistent, and I have to say that the above statement is wrong---do not react to this statement just yet, because I'll also say "but you are right" real soon.
Because "git fetch a/b" your UNIXy friends run will not consider "a/b" as a remote nickname, "a\b" in "git fetch a\b" Windows folks run must not be taken as a remote nickname either. That is a justification enough for Hannes's patch text, and that does not change whether we take nd/fopen-errors or discard it. So in that sense, the patch text does not have anything to do with the other topic.
But you are right (as promised ;-) to say that nd/fopen-errors needs a bit more polishing to work correctly on Windows. I unfortunately do not think Hannes's patch addresses the real issue.
Isn't the root cause of "Invalid argument" a colon in a (mistakenly identified) remote nickname, not a backslash? I doubt that it would help to loosen valid_remote_nick() to pay attention to backslashes. Surely, it may hide the issue when the path also has a backslash, but wouldn't Windows folks see the same warning for "git fetch c:c" for example even with Hannes's patch?
We use the pattern "try to open a path, and treat a failure to open as indicating that the path is missing (and take information we would have obtained from the contents of the path, if existed, from elsewhere, as necessary)" pretty much everywhere, not just when accessing a remote. And nd/fopen-errors tries to tweak the "treat a failure to open as a sign of missing path" to be a bit more careful by warning unless the error we get is ENOENT or ENOTDIR (otherwise we may treat a file that exists but cannot be read as if it is missing without any indication).
The solution for the problem Hannes's proposed log message describes lies in warn_on_fopen_errors() function that was introduced in the nd/fopen-errors topic. It appears that in Windows port, open() and fopen() return EINVAL for a file that does not exist and whose name Windows does not like, so we probably should do something like the attached to work around the EINVAL (I do not know about the symbol checked for #ifdef---there may be a more appropriate one).
I am not entirely happy with the workaround, because I do not know if EINVAL can come _ONLY_ due to a colon in the pathname on Windows, or if there are some other cases that deserve to be warned that also yield EINVAL, and the attached will sweep the problem under the rug for the "other cases", partially undoing what nd/fopen-errors topic wanted to do. But it does not make it worse than before the topic happened [*1*].
And that kind of solution does belong to nd/fopen-errors.
So in short:
(1) Hannes's patches are good, but they solve a problem that is
different from what their log messages say; the log message
needs to be updated; (2) In nd/fopen-errors topic, we need a new patch that deals with
EINVAL on Windows.Thanks.
[Footnote]
*1* An alternative that may allow us to avoid sweeping the issue
under the rug may be to have a more precise logic to see if the
open failure is due to missing file in the implementation of
open() and fopen() emulation---at that level, the logic can be
built with more platform knowledge (e.g. EINVAL? Let's see what
path we got---ah, there is a colon)---and turn EINVAL into
ENOENT, or something like that. I do not have a strong opinion
if that is a better solution or the attached workaround is good
enough. wrapper.c | 11 ++++++++++-
1 file changed, 10 insertions(+), 1 deletion(-)
Johannes Sixt
