Re: [PATCH] http: redact curl h2h3 headers in info

[Glen Choo <chooglen@google.com>]

Emily Shaffer <nasamuffin@google.com> writes:

> On Wed, Nov 09, 2022 at 12:52:31AM +0000, Glen Choo via GitGitGadget wrote:
>> 
>> With GIT_TRACE_CURL=1 or GIT_CURL_VERBOSE=1, sensitive headers like
>> "Authorization" and "Cookie" get redacted. However, since [1], curl's
>> h2h3 module also prints headers in its "info", which don't get redacted.
>> For example,
>> 
>>   echo 'github.com	TRUE	/	FALSE	1698960413304	o	foo=bar' >cookiefile &&
>>   GIT_TRACE_CURL=1 GIT_TRACE_CURL_NO_DATA=1 git \
>>     -c 'http.cookiefile=cookiefile' \
>>     -c 'http.version=' \
>>     ls-remote https://github.com/git/git refs/heads/main 2>output &&
>>   grep 'cookie' output
>> 
>> produces output like:
>> 
>>   23:04:16.920495 http.c:678              == Info: h2h3 [cookie: o=foo=bar]
>>   23:04:16.920562 http.c:637              => Send header: cookie: o=<redacted>
>> 
>> Teach http.c to check for h2h3 headers in info and redact them using the
>> existing header redaction logic.
>> 
>> [1] https://github.com/curl/curl/commit/f8c3724aa90472c0e617ddbbc420aa199971eb77
>> 
>> Signed-off-by: Glen Choo <chooglen@google.com>
>> ---
>>     http: redact curl h2h3 headers in info
>>     
>>     I initially sent this to the security list, but the general impression
>>     is that this isn't sensitive enough for an embargoed fix, so this is
>>     better discussed in the open instead.
>>     
>>     Since this comes from curl's HTTP2.0/3.0 module, this can be mitigated
>>     by setting http.version to 1.X, e.g. "git -c http.version=HTTP/1.1".
>>     
>>     According to [1], the susceptible curl versions appear to be 7.86.0,
>>     7.85.0, 7.84.0, 7.83.1, 7.83.0, 7.82.0, but I'm not sure which platforms
>>     are vulnerable.
>>     
>>     This patch fixes the issue on my machine running curl 7.85.0, so I think
>>     it is okay to merge as-is. That said, I would strongly prefer to add
>>     tests, but I haven't figured out how. In particular:
>>     
>>      * Do we have a way of using HTTP/2.0 in our tests? A cursory glance at
>>        our httpd config suggests that we only use HTTP/1.1.
>>      * How could we set up end-to-end tests to ensure that we're testing
>>        this against affected versions of curl? To avoid regressions, I'd
>>        also prefer to test against future versions of curl too.
>>     
>>     [1]
>>     https://github.com/curl/curl/commit/f8c3724aa90472c0e617ddbbc420aa199971eb77
>> 
>> Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-git-1377%2Fchooglen%2Fhttp%2Fredact-h2h3-v1
>> Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-git-1377/chooglen/http/redact-h2h3-v1
>> Pull-Request: https://github.com/git/git/pull/1377
>> 
>>  http.c | 40 ++++++++++++++++++++++++++++++++++++++--
>>  1 file changed, 38 insertions(+), 2 deletions(-)
>> 
>> diff --git a/http.c b/http.c
>> index 5d0502f51fd..cbcc7c3f5b6 100644
>> --- a/http.c
>> +++ b/http.c
>> @@ -560,8 +560,10 @@ static void set_curl_keepalive(CURL *c)
>>  }
>>  #endif
>>  
>> -static void redact_sensitive_header(struct strbuf *header)
>> +/* Return 0 if redactions been made, 1 otherwise. */
>
> Does it make sense to reverse the retval here?
>
> `if (!redact_sensitive_header())` sounds like "if not redacted, ..." -
> but here it means the opposite, right?

I struggled with this for a bit since I wasn't sure what the convention
is here. Enumerating some off the top of my head, we have:

- For 'booleans', we "0" for false and "1" for true (e.g.
  starts_with()).
- For functions that may fail with error, we have "0" for success and
  nonzero to signal the failure type (e.g. strbuf_getdelim()).
- For functions that don't fail we have "0" for "nothing was done" and
  "1" for something was done (e.g. skip_prefix()).

(Tangent: from a readability perspective, this is pretty poor. I need to
know beforehand whether or not the function may fail with error before I
know what the return value means?)

This probably falls into the last category, so for consistency, I think
this should return "1" for "redactions have happened" (as you
suggested).

>> +static int redact_sensitive_header(struct strbuf *header)
>>  {
>> +	int ret = 1;
>>  	const char *sensitive_header;
>>  
>>  	if (trace_curl_redact &&
>> @@ -575,6 +577,7 @@ static void redact_sensitive_header(struct strbuf *header)
>>  		/* Everything else is opaque and possibly sensitive */
>>  		strbuf_setlen(header,  sensitive_header - header->buf);
>>  		strbuf_addstr(header, " <redacted>");
>> +		ret = 0;
>>  	} else if (trace_curl_redact &&
>>  		   skip_iprefix(header->buf, "Cookie:", &sensitive_header)) {
>>  		struct strbuf redacted_header = STRBUF_INIT;
>> @@ -612,6 +615,27 @@ static void redact_sensitive_header(struct strbuf *header)
>>  
>>  		strbuf_setlen(header, sensitive_header - header->buf);
>>  		strbuf_addbuf(header, &redacted_header);
>> +		ret = 0;
>> +	}
>> +	return ret;
>> +}
>> +
>> +/* Redact headers in info */
>> +static void redact_sensitive_info_header(struct strbuf *header)
>> +{
>> +	const char *sensitive_header;
>> +
>> +	if (trace_curl_redact &&
>> +	    skip_iprefix(header->buf, "h2h3 [", &sensitive_header)) {
>> +		struct strbuf inner = STRBUF_INIT;
>> +
>> +		/* Drop the trailing "]" */
>> +		strbuf_add(&inner, sensitive_header, strlen(sensitive_header) - 1);
>> +		if (!redact_sensitive_header(&inner)) {
>> +			strbuf_setlen(header, strlen("h2h3 ["));
>> +			strbuf_addbuf(header, &inner);
>> +			strbuf_addch(header, ']');
>
> I'd really like some more comments in this function - even just one
> describing the string we're trying to redact, or showing a sample line.
> Navigating string parsing is always a bit difficult.

Ah yes, I should include a description of the string.

>> +		}
>>  	}
>>  }
>>  
>> @@ -668,6 +692,18 @@ static void curl_dump_data(const char *text, unsigned char *ptr, size_t size)
>>  	strbuf_release(&out);
>>  }
>>  
>> +static void curl_print_info(char *data, size_t size)
>
> Nit: Every other helper in this file calls it _dump_, so should this
> also say _dump_ instead of _print_?

Sure, I have no opinion here, so I'll do that.

>> +{
>> +	struct strbuf buf = STRBUF_INIT;
>> +
>> +	strbuf_add(&buf, data, size);
>> +
>> +	redact_sensitive_info_header(&buf);
>> +	trace_printf_key(&trace_curl, "== Info: %s", buf.buf);
>> +
>> +	strbuf_release(&buf);
>> +}
>> +
>>  static int curl_trace(CURL *handle, curl_infotype type, char *data, size_t size, void *userp)
>>  {
>>  	const char *text;
>> @@ -675,7 +711,7 @@ static int curl_trace(CURL *handle, curl_infotype type, char *data, size_t size,
>>  
>>  	switch (type) {
>>  	case CURLINFO_TEXT:
>> -		trace_printf_key(&trace_curl, "== Info: %s", data);
>> +		curl_print_info(data, size);
>>  		break;
>>  	case CURLINFO_HEADER_OUT:
>>  		text = "=> Send header";
>> 
>> base-commit: c03801e19cb8ab36e9c0d17ff3d5e0c3b0f24193
>
> Otherwise functionally it seems fine to me. case CURLINFO_TEXT is the
> one case that's not already using a curl_dump_* helper, so we're adding
> one, and to that helper we're adding a call out to
> redact_sensitive_header().
>
> Thanks.
>  - Emily
>
>> -- 
>> gitgitgadget