On 23/08/17 18:58, Jeff King wrote:
> On Wed, Aug 23, 2017 at 12:49:29PM +0300, Ivan Vyshnevskyi wrote:
>
>> Commits 47abd85 (fetch: Strip usernames from url's before storing them,
>> 2009-04-17) and later 882d49c (push: anonymize URL in status output,
>> 2016-07-14) made fetch and push strip the authentication part of the
>> remote URLs when used in the merge-commit messages or status outputs.
>> The URLs that are part of the error messages were not anonymized.
>>
>> A commonly used pattern for storing artifacts from a build server in a
>> remote repository utilizes a "secure" environment variable with
>> credentials to embed them in the URL and execute a push. Given enough
>> runs, an intermittent network failure will cause a push to fail, leaving
>> a non-anonymized URL in the build log.
>>
>> To prevent that, reuse the same anonymizing function to scrub
>> credentials from URL in the push error output.
>
> This makes sense. I suspect that most errors we output should be using
> the anonymized URL. Did you poke around for other calls?
Yes, I tried to check and unfortunately there are couple of places with
possible leaks:
* 'discover_refs()' in remote-curl.c when there's a HTTP error (see a
real-life scenario with an authz error in my response to Lars) -- is it
ok to include transport.h just to use one function or is there a cleaner
way?
* 'setup_push_upstream()' in push.c when a command doesn't have a branch
names (haven't saw problems with this in the wild, but could occur
during the CI setup) -- for this one, probably anonymization should
happen when the 'remote->name' field is set in the 'make_remote()'; same
question though, is it ok to include transport.h here?
Also there's an case of verbose output: I'm not sure I should change it,
but it does print out the non-anonymized URLs at least during push.
>
> The general structure of the patch looks good, but I have a few minor
> comments below.
>
>> Not sure how much of the background should be included in the commit message.
>> The "commonly used pattern" I mention could be found in the myriad of
>> online tutorials and looks something like this:
>
> My knee-jerk reaction is if it's worth writing after the dashes, it's
> worth putting in the commit message.
>
> However, in the case I think it is OK as-is (the motivation of "we
> already avoid leaking auth info to stdout, so we should do the same for
> error messages" seems self-contained and reasonable)
Well, I tend to be wordy, and most of the commit messages I saw were
rather short, so decided to split. Wonder, if maybe example command
should be included without the rest of it. Would it be useful?
>
>> diff --git a/builtin/push.c b/builtin/push.c
>> index 03846e837..59f3bc975 100644
>> --- a/builtin/push.c
>> +++ b/builtin/push.c
>> @@ -336,7 +336,7 @@ static int push_with_options(struct transport *transport, int flags)
>> err = transport_push(transport, refspec_nr, refspec, flags,
>> &reject_reasons);
>> if (err != 0)
>> - error(_("failed to push some refs to '%s'"), transport->url);
>> + error(_("failed to push some refs to '%s'"), transport_anonymize_url(transport->url));
>
> This leaks the return value. That's probably not a _huge_ deal since the
> program is likely to exit, but it's a bad pattern. I wonder if we should
> be setting up transport->anonymous_url preemptively, and just let its
> memory belong to the transport struct.
Ah. Thanks! I knew I'd fail in the memory management even with the
one-line patch. :)
About 'transport->anonymous_url': not sure if it's worth it. There are
four calls to 'transport_anonymize_url' right now and it looks like the
new one in my patch is the first that has a transport struct instance
near by. The next likely candidate for update 'discover_refs()' also
gets the url as an argument.
>
>> diff --git a/t/t5541-http-push-smart.sh b/t/t5541-http-push-smart.sh
>> index d38bf3247..0b6fb6252 100755
>> --- a/t/t5541-http-push-smart.sh
>> +++ b/t/t5541-http-push-smart.sh
>> @@ -377,5 +377,23 @@ test_expect_success 'push status output scrubs password' '
>> grep "^To $HTTPD_URL/smart/test_repo.git" status
>> '
>>
>> +cat >"$HTTPD_DOCUMENT_ROOT_PATH/test_repo.git/hooks/update" <<EOF
>> +#!/bin/sh
>> +exit 1
>> +EOF
>> +chmod a+x "$HTTPD_DOCUMENT_ROOT_PATH/test_repo.git/hooks/update"
>> +
>> +cat >exp <<EOF
>> +error: failed to push some refs to '$HTTPD_URL/smart/test_repo.git'
>> +EOF
>
> I know the t5541 script, which is old and messy, led you into these bad
> constructs. But usually in modern tests we:
>
> 1. Try to keep all commands inside test_expect blocks to catch
> unexpected failures or unwanted output.
>
> 2. Use write_script for writing scripts, like:
>
> write_script "$HTTPD_DOCUMENT_ROOT_PATH/test_repo.git/hooks/update" <<-\EOF
> exit 1
> EOF
>
> 3. Backslash our here-doc delimiter to suppress interpolation.
>
>> +test_expect_success 'failed push status output scrubs password' '
>> + cd "$ROOT_PATH"/test_repo_clone &&
>> + test_must_fail git push "$HTTPD_URL_USER_PASS/smart/test_repo.git" +HEAD:scrub_err 2>stderr &&
>> + grep "^error: failed to push some refs" stderr >act &&
>> + test_i18ncmp exp act
>> +'
>> +rm -f "$HTTPD_DOCUMENT_ROOT_PATH/test_repo.git/hooks/update"
>
> Similarly, this "rm" should probably be a test_when_finished in the
> block with the write_script (unless you really need to carry it over
> several test_expect blocks, in which case there should be an explicit
> test_expect cleaning it up).
Thanks! You're right. I just followed examples in the file.
Updated [1], will send with the next patch version.
>
> Instead of grepping for the exact error, should we instead grep for the
> password to make sure it is not present on _any_ line?
>
> -Peff
>
One possible issue I see is that this will make it overlap with the
'push status output scrubs password' case above. But if it's not a
problem, I can replace last two lines with just a 'test_i18ngrep !'
[1]:
https://github.com/sainaen/git/blob/af17713/t/t5541-http-push-smart.sh#L380-L392
