Feature request: better error messages when UTF-8 bites

Hi;

Just found an annoyance in `git log` (and likely elsewhere) that may warrant a change:

Somehow when copying and pasting a commit from a website to the command line, a UTF-8 Byte Order Mark (BOM) [https://en.wikipedia.org/wiki/Byte_order_mark] was appended to one of the commit ids. BOMs are invisible, as are many other UTF-8 code points. The upshot was that Git didn't like it, and complained bitterly:

Feature request: ================

When printing the "fatal: ambiguous argument '......': ....", perhaps escape (url or otherwise) the ambiguous argument when printing it in the error message, or maybe add a sentence about non-ASCII characters being found.

This is sort of a difficult corner-case, in that it is perfectly legal to have UTF-8 characters in a branch or tag name (see git-check-ref-format for the allowed characters), so someone could indeed create a branch named "038179704f0066aa815d5429221cf381ff4ef289\357\273\277" if they were a tortured soul bent on overthrowing polite society. Rejecting input because it has bytes with values above \177 is therefore not a solution.

Similarly, scanning the input for invisible UTF-8 characters (or even invalid UTF-8 sequences) is leaning too far the other way: git should not be validating character encodings. It should stay encoding-neutral, as the alternative leads to madness, driving developers into becoming tortured souls bent on rigidly enforcing polite society. We have enough of those already.

It's unclear as to whether violent overthrow or rigid enforcement is the lesser of two evils, but let's not perform the experiment to find out. :-)

Cheers!

Am 27.07.22 um 22:21 schrieb CH:

That's not going to fly, IMHO, because when I type

   git diff todo/René

I would not want to see

fatal: ambiguous argument 'todo/Ren\303\251': unknown ...

I'm convinced that there are thousands of users who use non-ASCII branch and file names that they also frequently mis-type. They'd all be greeted with unintelligible nerdy gibberish.

I may be able to change my mind if ambiguous input (in the sense of "is not what it seems to be") leads to a security hazard that is unique to Git.

-- Hannes

On 2022-07-28 01:42, Johannes Sixt wrote:

This is actually already MUCH better that the OP's example. In his example he has a string that looks like a 40-char hash, and git complains without showing any of the Unicode gibberish attached to that sha1. It would be better if at least it printed something in ASCII with escaped bytes in the error message.

Moreover this isn't even close to the issue above - he's talking about a no-op, non-printing Unicode marker that crept in. While I do think it shouldn't be an issue, it shouldn't even have been passed to git. IMHO it should have been stripped by the browser itself on copy, or by the terminal on paste... FWIW I'm using rxvt-unicode, and copying this from the terminal doesn't copy the marker but pasting the marker copied from Chrome is passed on to bash and git.

NB: I also though what if the shell handled it, but that isn't even 
really a character so not technically suitable for $IFS, and even if we 
considered that option it wouldn't really play well with POSIX's 
definition of $IFS - how to tell for example between a single Unicode 
codepoint and a list of binary characters? There is just no definition 
of wide chars for $IFS, not in POSIX nor in recent versions of Bash AFAIK.

TL;DR; the issue is IMHO on the browser side, which shouldn't include the marker in the copied text, or maybe on the terminal, BUT when passed on to git it should at least print the escaped Unicode chars in the error, otherwise it's just too confusing for the user.

BTW you actually raise another issue - I do think for file paths git could either recompose (NFC) or decompose (NFD) the strings on storage and comparison (which should probably be an option... the current default for 2.30.2 is to treat them and print them as binary (escaped on print). Consider the following when using core.quotePath=false:

$ touch "nfc_$(printf '\xf4')"
$ touch "nfd_$(printf '\x6f\xcc\x82')"
$ git add nf[cd]*
$ git status
On branch test
Changes to be committed:
   (use "git restore --staged <file>..." to unstage)
     new file:   nfc_ô
     new file:   nfd_ô

I'm not sure how the Unicode will be translated here, it might depend on the mail client if they even's get sent as-is, but both shows the exact same file name, one in NFD and one in NFC format.

Both are canonically equivalent and reversible. It appears MacOS already decompose (NFD?) filenames by default and git provides an option to recompose the characters (core.precomposeUnicode) which, according to the manual, is not even usable on Linux...

More on Unicode normalization: https://unicode.org/reports/tr15/

-- Thomas

[]

This is not valid unicode, isn't it ? Probably we need to use the octal version, since not all printf() implementation support hex values starting with 0x, but all of them support octal:

auml=$(printf '\303\244') aumlcdiar=$(printf '\141\314\210')

Translated by "whom" ? Most programs do no translate anything here.

Yes. Technically you can have both under Linux, at least unless you are running ZFS, which may be created unicode-aware (or not, that is the default).

But why do you want to have 2 files on disk with different normalizations ?