OpenBMC with dropbear SSH library fails a Denial-Of-Service test

[generatz]

In openbmc-test-automation/security/test_bmc_connections.robot there is the following test case:

Verify User Cannot Login After 5 Non-Logged In Sessions
    [Documentation]  User should not be able to login when there
    ...  are 5 non-logged in sessions.
    [Tags]  Verify_User_Cannot_Login_After_5_Non-Logged_In_Sessions

    FOR  ${i}  IN RANGE  ${0}  ${MAX_UNAUTH_PER_IP}
       SSHLibrary.Open Connection  ${OPENBMC_HOST}
       Start Process  ssh ${OPENBMC_USERNAME}@${OPENBMC_HOST}  shell=True
    END

    SSHLibrary.Open Connection  ${OPENBMC_HOST}
    ${status}=   Run Keyword And Return Status  SSHLibrary.Login  ${OPENBMC_USERNAME}  ${OPENBMC_PASSWORD}

    Should Be Equal  ${status}  ${False}

The test case does what it sounds like: The test opens 5 concurrent ssh login sessions and leaves them hanging - it does not attempt a login to those 5 sessions. It then opens a 6th session and attempts a login there. The test passes if it is unable to login with the 6th attempt. In other words, it anticipates that the BMC should limit the number of concurrent, not-yet-completed login sessions.

Note that it is not testing for a limit on the number of concurrent login sessions (which I assume it should not).
And it is not testing for a limit to the number of sequential failed logins (which would be good, but might be covered elsewhere.)

I haven't heard of such a security requirement, and wonder whether it should be optional, vendor specific, or deprecated. (FWIW, it doesn't pass on the various BMCs I've tested.)

[gkeishin]

@prkatti1 Please visit this .

[prkatti1]

From any client or IP we can have max 5 un-authenticated connections to avoid DoS & reduce unnecessary load on BMC.

https://github.com/mkj/dropbear/blob/master/default_options.h

#define DROPBEAR_USE_PRNGD 0
#define DROPBEAR_PRNGD_SOCKET "/var/run/dropbear-rng"

/* Specify the number of clients we will allow to be connected but

• not yet authenticated. After this limit, connections are rejected /
  / The first setting is per-IP, to avoid denial of service */
  #define MAX_UNAUTH_PER_IP 5

/* And then a global limit to avoid chewing memory if connections

• come from many IPs */
  #define MAX_UNAUTH_CLIENTS 30

[gkeishin]

@joseph-reynolds @generatz ^^^

[generatz]

OK, thanks @prkatti1 . I thought it seemed logical for anti-DOS measure, but I had not seen it work on any system I tried.
I wonder if the firmware is at fault, or whether the test is at fault. I tried opening 6 ssh sessions to a BMC manually from 6 consoles and that succeeded.
Has anyone seen this work in practice?

[joseph-reynolds]

I have no experience with this. But I am interested in making this limit work.

I believe the dropbear SSH server configuration above is ineffective because dropbear is a systemd socket-activated service, which means it starts when there in an incoming connection, establishes a SSH session, then stops.

To make this work, I have been looking into the dropbear systemd socket file configuration - https://github.com/openbmc/openbmc/blob/master/poky/meta/recipes-core/dropbear/dropbear/dropbear.socket
documented here https://www.freedesktop.org/software/systemd/man/systemd.socket.html but I don't see any way to control MaxPreAuthenticatedConnections or know how systemd would know the state of a connection.
(So, help wanted.)

A solution might require moving away from a socket-acctivated service, or moving away from dropbear and moving toward OpenSSH.

[gkeishin]

@prkatti1 anything we need to follow on this one?