[PATCH v2 0/2] http: store credential when PKI auth is used

Here's my second attempt at getting the certificate password into the credential
store.  I tested from a working PKI setup and found curl--at least reasonable
recent versions of it--return CURLE_SSL_CERTPROBLEM:

       CURLE_SSL_CERTPROBLEM (58)
              problem with the local client certificate.

It appears there could be another possible error from curl:

       CURLE_SSL_CONNECT_ERROR (35)
              A  problem  occurred  somewhere  in the SSL/TLS handshake. You
              really want the error buffer and read the message there as it
              pinpoints the problem slightly more. Could be  certificates  (file
              formats, paths, permissions), passwords, and others.

This seems less likely to be a bad client password scenario, so I did not look
for this particular error to reject it.

I also added one other small patch to remove the check of a non-empty password
before calling credential_store() for proxy_auth, as credential_store() already
checks for a non-empty password and gracefully handles it when it doesn't.

-John

John Szakmeister (2):
  http: store credential when PKI auth is used
  http: drop the check for an empty proxy password before approving

 http.c | 13 +++++++++++--
 1 file changed, 11 insertions(+), 2 deletions(-)

-- 
2.30.1

We already looked for the PKI credentials in the credential store, but
failed to approve it on success.  Meaning, the PKI certificate password
was never stored and git would request it on every connection to the
remote.  Let's complete the chain by storing the certificate password on
success.

Likewise, we also need to reject the credential when there is a failure.
Curl appears to report client-related certificate issues are reported
with the CURLE_SSL_CERTPROBLEM error.  This includes not only a bad
password, but potentially other client certificate related problems.
Since we cannot get more information from curl, we'll go ahead and
reject the credential upon receiving that error, just to be safe and
avoid caching or saving a bad password.

Signed-off-by: John Szakmeister <redacted>
---
 http.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/http.c b/http.c
index f8ea28bb2e..12a8aaba48 100644
--- a/http.c
+++ b/http.c

@@ -1637,7 +1637,17 @@ static int handle_curl_result(struct slot_results *results)
 		credential_approve(&http_auth);
 		if (proxy_auth.password)
 			credential_approve(&proxy_auth);
+		credential_approve(&cert_auth);
 		return HTTP_OK;
+	} else if (results->curl_result == CURLE_SSL_CERTPROBLEM) {
+		/*
+		 * We can't tell from here whether it's a bad path, bad
+		 * certificate, bad password, or something else wrong
+		 * with the certificate.  So we reject the credential to
+		 * avoid caching or saving a bad password.
+		 */
+		credential_reject(&http_auth);
+		return HTTP_NOAUTH;
 	} else if (missing_target(results))
 		return HTTP_MISSING_TARGET;
 	else if (results->http_code == 401) {

-- 
2.30.1

credential_approve() already checks for a non-empty password before
saving, so there's no need to do the extra check here.

Signed-off-by: John Szakmeister <redacted>
---
 http.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/http.c b/http.c
index 12a8aaba48..b0d3ce6c6b 100644
--- a/http.c
+++ b/http.c

@@ -1635,8 +1635,7 @@ static int handle_curl_result(struct slot_results *results)
 
 	if (results->curl_result == CURLE_OK) {
 		credential_approve(&http_auth);
-		if (proxy_auth.password)
-			credential_approve(&proxy_auth);
+		credential_approve(&proxy_auth);
 		credential_approve(&cert_auth);
 		return HTTP_OK;
 	} else if (results->curl_result == CURLE_SSL_CERTPROBLEM) {

-- 
2.30.1

On Thu, Mar 11, 2021 at 07:48:40PM -0500, John Szakmeister wrote:

Here's my second attempt at getting the certificate password into the credential
store.  I tested from a working PKI setup and found curl--at least reasonable
recent versions of it--return CURLE_SSL_CERTPROBLEM:

       CURLE_SSL_CERTPROBLEM (58)
              problem with the local client certificate.

It appears there could be another possible error from curl:

       CURLE_SSL_CONNECT_ERROR (35)
              A  problem  occurred  somewhere  in the SSL/TLS handshake. You
              really want the error buffer and read the message there as it
              pinpoints the problem slightly more. Could be  certificates  (file
              formats, paths, permissions), passwords, and others.

This seems less likely to be a bad client password scenario, so I did not look
for this particular error to reject it.

I also added one other small patch to remove the check of a non-empty password
before calling credential_store() for proxy_auth, as credential_store() already
checks for a non-empty password and gracefully handles it when it doesn't.

Thanks. Both patches look good to me. I wondered briefly if we needed to
worry about old versions of curl missing CURLE_SSL_CERTPROBLEM. But it
seems to have shown up in ~2002, so I think we are fine to assume it's
there.

It would be nice if we had some tests here, but we currently do not
cover any of the ssl-cert stuff in the test suite. I suspect adding them
would be a big pain to configure and maintain, so I'm OK to leave it off
for now. Hopefully you gave it some basic manual testing with your
working setup (good password is stored, bad password is rejected).

Looking at how we generate the server-side cert for our http tests, we
could _probably_ do something similar for a client-side cert, and just
configure the server to accept a self-signed certificate. But like I
said, I'm OK to leave that for another series (though of course if you
want to work on it, that would be very much appreciated).

-Peff

On 2021-03-12 at 00:48:41, John Szakmeister wrote:

Is this supposed to be &cert_auth here?  I'm not sure how a bad HTTP
password would even have been tested in this case.
-- 
brian m. carlson (he/him or they/them)
Houston, Texas, US

On Fri, Mar 12, 2021 at 01:41:30AM +0000, brian m. carlson wrote:

quoted

diff --git a/http.c b/http.c
index f8ea28bb2e..12a8aaba48 100644
--- a/http.c
+++ b/http.c

@@ -1637,7 +1637,17 @@ static int handle_curl_result(struct slot_results *results)
 		credential_approve(&http_auth);
 		if (proxy_auth.password)
 			credential_approve(&proxy_auth);
+		credential_approve(&cert_auth);
 		return HTTP_OK;
+	} else if (results->curl_result == CURLE_SSL_CERTPROBLEM) {
+		/*
+		 * We can't tell from here whether it's a bad path, bad
+		 * certificate, bad password, or something else wrong
+		 * with the certificate.  So we reject the credential to
+		 * avoid caching or saving a bad password.
+		 */
+		credential_reject(&http_auth);

Is this supposed to be &cert_auth here?  I'm not sure how a bad HTTP
password would even have been tested in this case.

Good catch! When reviewing, I was so busy thinking about _where_ this
line should go that I didn't even notice what it said. :)

-Peff

On Thu, Mar 11, 2021 at 8:45 PM Jeff King [off-list ref] wrote:

On Fri, Mar 12, 2021 at 01:41:30AM +0000, brian m. carlson wrote:

quoted

quoted

diff --git a/http.c b/http.c
index f8ea28bb2e..12a8aaba48 100644
--- a/http.c
+++ b/http.c

@@ -1637,7 +1637,17 @@ static int handle_curl_result(struct slot_results *results)
            credential_approve(&http_auth);
            if (proxy_auth.password)
                    credential_approve(&proxy_auth);
+           credential_approve(&cert_auth);
            return HTTP_OK;
+   } else if (results->curl_result == CURLE_SSL_CERTPROBLEM) {
+           /*
+            * We can't tell from here whether it's a bad path, bad
+            * certificate, bad password, or something else wrong
+            * with the certificate.  So we reject the credential to
+            * avoid caching or saving a bad password.
+            */
+           credential_reject(&http_auth);

Is this supposed to be &cert_auth here?  I'm not sure how a bad HTTP
password would even have been tested in this case.

Good catch! When reviewing, I was so busy thinking about _where_ this
line should go that I didn't even notice what it said. :)

Good catch!  I don't even know how I did that. :-/  The system I
created the patch on is inaccessible via the Internet and I can't
really get data off of it.  This is entirely an error in translation
on my part.  The diff I printed has the correct line.  My bad.  I'll
send an update soon.

John

On Thu, Mar 11, 2021 at 8:31 PM Jeff King [off-list ref] wrote:
[snip]

Thanks. Both patches look good to me. I wondered briefly if we needed to
worry about old versions of curl missing CURLE_SSL_CERTPROBLEM. But it
seems to have shown up in ~2002, so I think we are fine to assume it's
there.

It would be nice if we had some tests here, but we currently do not
cover any of the ssl-cert stuff in the test suite. I suspect adding them
would be a big pain to configure and maintain, so I'm OK to leave it off
for now. Hopefully you gave it some basic manual testing with your
working setup (good password is stored, bad password is rejected).

I did do some manual testing in an environment at work where they have
this set up.  Unfortunately, the way I went about this was not optimal.  I'll
work the issue differently in the future, so I don't have that kind of
translation
issue again.

Looking at how we generate the server-side cert for our http tests, we
could _probably_ do something similar for a client-side cert, and just
configure the server to accept a self-signed certificate. But like I
said, I'm OK to leave that for another series (though of course if you
want to work on it, that would be very much appreciated).

I looked at things a little bit, but it was too much to take on right
now.  I could
probably get something together to help make it happen.  I've been down that
road before, so I know it can be involved, but it would be nice to have tests.
I'm not signing up just yet for that, but when a rainy weekend hits, I'll see
about taking a stab at it.

-John