Fix race in stunnel port selection #129

tsmetana · 2022-04-25T10:46:07Z

Issue #, if available:
Issue #125

Description of changes:
To prevent the stunnel port selection race between parallel mount.efs processes keep the probed port bound until the stunnel configuration file gets written and add a check for the config file existence prior trying to bind the stunnel port to check for its availability.

tsmetana · 2022-04-25T11:56:14Z

I'll fix the failing tests (so far I keep running into pytest-dev/pytest#8539 due to python-3.10)...

elimumford · 2022-06-13T21:50:43Z

@lshigupt you appear to be the primary maintainer. This PR would be very helpful in addressing issues in the kubernetes-sigs/aws-efs-csi-driver (kubernetes-sigs/aws-efs-csi-driver#695)

tsmetana · 2022-06-15T14:06:26Z

I actually don't think the CSI driver issue kubernetes-sigs/aws-efs-csi-driver#695 can be fixed from within efs-utils: this would however help with other problems in the CSI driver (like deleting multiple unused volumes).

lshigupt · 2022-07-22T21:16:05Z

Thanks a lot @tsmetana for the PR, I am doing the testing and I could see that some of the Tests are failing on our End. I am trying to debug them and Will post the comments where it is failing.

These are the Tests which are Failing:

[CPython37-release] =========================== short test summary info ============================
[CPython37-release] FAILED test/mount_efs_test/test_bootstrap_tls.py::test_bootstrap_tls_state_file_dir_exists
[CPython37-release] FAILED test/mount_efs_test/test_bootstrap_tls.py::test_bootstrap_tls_state_file_nonexistent_dir
[CPython37-release] FAILED test/mount_efs_test/test_bootstrap_tls.py::test_bootstrap_tls_cert_created
[CPython37-release] FAILED test/mount_efs_test/test_bootstrap_tls.py::test_bootstrap_tls_non_default_port
[CPython37-release] FAILED test/mount_efs_test/test_bootstrap_tls.py::test_bootstrap_tls_non_default_verify_level
[CPython37-release] FAILED test/mount_efs_test/test_bootstrap_tls.py::test_bootstrap_tls_ocsp_option
[CPython37-release] FAILED test/mount_efs_test/test_bootstrap_tls.py::test_bootstrap_tls_noocsp_option
[CPython37-release] FAILED test/mount_efs_test/test_choose_tls_port.py::test_choose_tls_port_first_try
[CPython37-release] FAILED test/mount_efs_test/test_choose_tls_port.py::test_choose_tls_port_second_try
[CPython37-release] FAILED test/mount_efs_test/test_choose_tls_port.py::test_choose_tls_port_never_succeeds
[CPython37-release] FAILED test/mount_efs_test/test_choose_tls_port.py::test_choose_tls_port_option_specified

tsmetana · 2022-08-02T08:01:14Z

Hello. I've tried to run the tests with python-3.7.13 and I'm not able to reproduce the failure. Do you have some more detailed logs?

tsmetana · 2022-10-19T11:34:11Z

@lshigupt Hi. I wouldn't really want the issue and PR to rot completely... Is there anything I can do to help getting it moving forward?

RyanStan · 2022-10-27T18:41:32Z

Hi @tsmetana, sorry for the delay. We've picked this back up and will look into it.

tsmetana · 2022-12-01T10:48:39Z

@RyanStan any news? I mean... The patch is not that big and (I hope) quite understandable.

Cappuccinuo

Thanks for the PR. We will do a quick release for this fix soon tomorrow morning so we will put this commit along with other urgent fix we have.

The fix is not permanent, since it is not guaranteed there is no race condition between the socket is closed and the stunnel is launched, though the time interval is pretty tight. We will work on a long term fix for this.

src/mount_efs/__init__.py

Cappuccinuo · 2022-12-02T05:16:57Z

src/mount_efs/__init__.py

@@ -944,13 +944,13 @@ def choose_tls_port(config, options):
        assert len(tls_ports) == len(ports_to_try)

    if "netns" not in options:
-        tls_port = find_tls_port_in_range(ports_to_try)
+        sock = find_tls_port_in_range(state_file_dir, fs_id, mountpoint, ports_to_try)


nit: tls_port_sock

Cappuccinuo · 2022-12-02T05:18:20Z

src/mount_efs/__init__.py

+        mount_filename = get_mount_specific_filename(fs_id, mountpoint, tls_port)
+        config_file = get_stunnel_config_filename(state_file_dir, mount_filename)
+        if os.access(config_file, os.R_OK):
+            logging.info("confifguration for port %s already exists, trying another port", tls_port)
+            continue


Is this necessary? Since if the port is already used the binding will fail anyway.

It's is not necessary but helps in the case the port biniding fails to distinguish whether we're not clashing with other processes. Makes debugging easier.

Cappuccinuo · 2022-12-02T05:18:54Z

src/mount_efs/__init__.py

@@ -1430,7 +1436,8 @@ def bootstrap_tls(
    state_file_dir=STATE_FILE_DIR,
    fallback_ip_address=None,
 ):
-    tls_port = choose_tls_port(config, options)
+    sock = choose_tls_port(state_file_dir, fs_id, mountpoint, config, options)
+    tls_port = sock.getsockname()[1]


nit: Use a function for this socket.getsockname() such that in unit test you can directly use the func.

Since we're always interested in both the socket and the port, I changed the function to return a tuple instead.

Cappuccinuo · 2022-12-02T05:19:34Z

src/mount_efs/__init__.py

+    # close the socket now, so the stunnel process can bind to the port
+    sock.close()


You should add the close to a finally statement, so if any steps failed in between you create the socket and create the socket, the socket will eventually be closed cleanly.

Good point. Done.

tsmetana · 2022-12-02T11:07:15Z

The fix is not permanent, since it is not guaranteed there is no race condition between the socket is closed and the stunnel is launched, though the time interval is pretty tight. We will work on a long term fix for this.

Could you elaborate on this, please? The important part is the configuration file here: it has to be written first before the socket is closed and the stunnel process is launched. No other efs-mount can choose the same port since it would either find the existing configuration file (that serves also as a lock for the chosen port essentially) or fail to bind the same port if other efs-mount is trying to create the new configuration (i.e. the file is not written yet, but other efs-mount has chosen the same port already). This is why there's the check for the config file existence there and why I wanted to explicitly log if another efs-mount is trying to use the same port based on that.

Cappuccinuo · 2022-12-02T11:14:59Z

No other efs-mount can choose the same port since it would either find the existing configuration file

I think that only applies to the same file system right? Since the configuration file is checked based on the fs and mountpoint and tlsport, but another file system may not be applied here.

Anyway I will merge this change, and push another commit and bump the release version. We can continue our discussion here in the thread, I will add more detail to it. Thanks for the PR!

tsmetana · 2022-12-02T11:23:21Z

I think that only applies to the same file system right? Since the configuration file is checked based on the fs and mountpoint and tlsport, but another file system may not be applied here.

True. So adding a separate lock file with just the port should be sufficient. Or remove the fs and mount point from the config file name if it's not used for anything else (would have to check), or at worse parse the filename and check just the port part...

tsmetana · 2022-12-02T15:02:04Z

@Cappuccinuo I think by removing the config file existence check in the latest patch you actually introduced the race you described even for single fsid/mountpoint case: now we rely on the (uncertain) fact that stunnel binds the socket before another efs-mount tries is out... If we checked the config file, the time between closing the socket and stunnel binding the port would not matter, the port wouldn't be even probed.

tsmetana force-pushed the fix-stunnel-port-race branch from 1bbe91b to bad09ff Compare April 27, 2022 11:49

tsmetana force-pushed the fix-stunnel-port-race branch from bad09ff to 69d41d6 Compare May 31, 2022 15:05

Cappuccinuo requested a review from lshigupt June 24, 2022 06:14

Cappuccinuo assigned lshigupt Jun 24, 2022

padumergue mentioned this pull request Aug 1, 2022

In pods with multiple EFS PVC mounts, the EFS-CSI driver seems to sometimes mix up mount points kubernetes-sigs/aws-efs-csi-driver#635

Closed

Cappuccinuo suggested changes Dec 2, 2022

View reviewed changes

tsmetana force-pushed the fix-stunnel-port-race branch from 69d41d6 to 0d20d72 Compare December 2, 2022 10:33

Fix race in stunnel port selection

612d060

tsmetana force-pushed the fix-stunnel-port-race branch from 0d20d72 to 612d060 Compare December 2, 2022 10:45

Cappuccinuo approved these changes Dec 2, 2022

View reviewed changes

Cappuccinuo merged commit 478f009 into aws:master Dec 2, 2022

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fix race in stunnel port selection #129

Fix race in stunnel port selection #129

tsmetana commented Apr 25, 2022

tsmetana commented Apr 25, 2022

elimumford commented Jun 13, 2022

tsmetana commented Jun 15, 2022

lshigupt commented Jul 22, 2022 •

edited

Loading

tsmetana commented Aug 2, 2022

tsmetana commented Oct 19, 2022

RyanStan commented Oct 27, 2022

tsmetana commented Dec 1, 2022

Cappuccinuo left a comment

Cappuccinuo Dec 2, 2022

tsmetana Dec 2, 2022

Cappuccinuo Dec 2, 2022

tsmetana Dec 2, 2022

Cappuccinuo Dec 2, 2022

tsmetana Dec 2, 2022

Cappuccinuo Dec 2, 2022

tsmetana Dec 2, 2022

tsmetana commented Dec 2, 2022

Cappuccinuo commented Dec 2, 2022

tsmetana commented Dec 2, 2022 •

edited

Loading

tsmetana commented Dec 2, 2022

		# close the socket now, so the stunnel process can bind to the port
		sock.close()

Fix race in stunnel port selection #129

Fix race in stunnel port selection #129

Conversation

tsmetana commented Apr 25, 2022

tsmetana commented Apr 25, 2022

elimumford commented Jun 13, 2022

tsmetana commented Jun 15, 2022

lshigupt commented Jul 22, 2022 • edited Loading

tsmetana commented Aug 2, 2022

tsmetana commented Oct 19, 2022

RyanStan commented Oct 27, 2022

tsmetana commented Dec 1, 2022

Cappuccinuo left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tsmetana commented Dec 2, 2022

Cappuccinuo commented Dec 2, 2022

tsmetana commented Dec 2, 2022 • edited Loading

tsmetana commented Dec 2, 2022

lshigupt commented Jul 22, 2022 •

edited

Loading

tsmetana commented Dec 2, 2022 •

edited

Loading