Re: gitweb bug report: hash mistaken for an option

[Junio C Hamano <gitster@pobox.com>]

Nicolas George <george@nsup.org> writes:

> Hi.
>
> We have a web server with gitweb (from Debian), and we observed git error
> messages in apache's error.log. I tracked down the issue to a request that
> had this:
>
> ```
> ?p=…/.git;a=tree;hb=-c
> ```
>
> I will not bother giving the actual URL since nobody else can check the
> issue in the logs or our server, but the issue can be reproduced on any
> server by replacing the `hb=…` parameter with `hb=-c`.
>
> I tracked down the issue further to the `git_tree` function of the CGI
> script:
>
> <https://git.kernel.org/pub/scm/git/git.git/tree/gitweb/gitweb.perl#n7200>
>
> If `$hash_base`, i.e. the `hb` parameter, is set and not `$file_name`, then
> it is passed as is as the last argument of `ls-tree -z`, and since it is not
> a valid hash, it prints an error.

In other words, garbage-in garbage-out?

FWIW, this also gives an error message

    $ git ls-tree no-such-tree
    fatal: Not a valid object name no-such

and even though it may give slightly a smaller error message than
"git ls-tree -c", it would not stay silent.  So I am not sure if
there is anything to fix here, short of redirecting your standard
error to /dev/null or something.

> Second, that the error (be it “unknown switch”, “Not a valid object name” or
> “not a tree object” gets written into the error log: it is an error entirely
> caused by the client that has no repercussion on the server, it should be
> either passed back to the client or ignored. This is more minor but harder
> to fix.

For that, you'd need to capture standard error stream and relay it
to the user, I guess.  That does sound like the right fix to deal
with any wrong input that comes in the web request.

This is a tangent, but I actually think "ls-tree -- $whatever" that
forces $whatever to be interpreted as the tree object name _is_ a
bug.  A double-dash should signal end of revisions and beginning of
the pathspec.  We may want to fix it in "git ls-tree", regardless of
any gitweb issue.

While I do not think it is a fix to force the end-user supplied
parameter to be interpreted as a tree object name, if we wanted to
do so in durable way that won't be broken even after we correct the
double-dash bug in ls-tree, we probably should write

    $ git ls-tree -z --end-of-options $hb

instead.

Thanks.