path: root/content/posts/WIP-how-bsd-authentication-works/index.org

#+TITLE: How BSD Authentication Works
#+DATE: 2020-06-26T18:31:36-04:00
#+DRAFT: true
#+DESCRIPTION:
#+TAGS[]: openbsd
#+KEYWORDS[]: openbsd
#+SLUG:
#+SUMMARY:
#+SHOWTOC: true

[[https://web.archive.org/web/20170327150148/http://www.penzin.net/bsdauth/]]
* History

  OpenBSD is quite different from many other Unix-like operating systems
  in many ways, but one way which I find interesting is the
  authentication system. Most systems from AIX, Solaris, and Linux to
  most BSDs including MacOS use some form of a system called Pluggable
  Authentication Module (PAM). The two main implementations of PAM are
  [[http://www.linux-pam.org/][Linux PAM]] and [[https://www.openpam.org/][OpenPAM]]. PAM modules are created a dynamically loaded
  shared objects, which communicate using a set of standard
  interfaces ([[https://linux.die.net/man/3/pam][Linux-PAM]] and [[https://www.freebsd.org/cgi/man.cgi?query=pam&apropos=0&sektion=3&manpath=FreeBSD+12.1-RELEASE+and+Ports&arch=default&format=html][OpenPAM]]). PAM is configured using the [[https://linux.die.net/man/5/pam.d][pam.d]]
  directory and [[https://www.freebsd.org/cgi/man.cgi?query=pam.conf&sektion=5&apropos=0&manpath=FreeBSD+12.1-RELEASE+and+Ports][pam.conf]].

  OpenBSD on the other hand uses a mechanism called BSD
  Authentication. It was originally developed for a proprietary
  operating system called [[https://en.wikipedia.org/wiki/BSD/OS][BSD/OS]] by [[https://en.wikipedia.org/wiki/Berkeley_Software_Design][Berkeley Software Design Inc.]], who
  later donated the system. It was adopted by OpenBSD in release
  2.9. BSD Auth is comparatively much simpler than PAM. Modules or,
  authentication "styles", are instead stand alone applications or
  scripts that communicate over IPC (=PF_LOCAL, SOCK_STREAM=,
  specifically). The program or script has no ability to interfere
  with the parent and can very easily revoke permissions using
  [[https://man.openbsd.org/pledge][=pledge(2)=]] or [[https://man.openbsd.org/unveil][=unveil(2)=]]. The BSD Authentication system of
  configured through [[https://man.openbsd.org/login.conf][=login.conf(5)=]].

* Why

  This one is pretty difficult, since there seems to be very little
  information about how BSD Auth works apart from the source code
  itself and the man pages, which intentionally keep the internals
  opaque. This is my best attempt to understand and describe the flow
  of BSD Auth.

* BSD Auth Modules

  These programs or scripts are located in =/usr/libexec/auth/= with the
  naming convention =login_<style>=. They take arguments in the form of

  #+BEGIN_SRC shell
  login_<style> [-s service] [-v key=value] user [class]
  #+END_SRC

  - =<style>= is the authentication method. This could be =passwd=,
    =radius=, =skey=, =yubikey=, etc. There's more information about
    available styles in [[https://man.openbsd.org/login.conf][=login.conf(5)=]] under the [[https://man.openbsd.org/login.conf#AUTHENTICATION][=AUTHENTICATION=]]
    header.
  - =service= is the service type. Typically authentication methods
    will accept one of three values here, =login=, =challenge=, or
    =response=. =login= is the default if it's not specified. =login=
    is used to let the module know to interact with the user directly,
    while =challenge= and =response= are used to pass messages back
    and forth through the BSD Auth API. Each style's man page will
    have more details on these.
  - =-v key=value= is an optional argument. There is no limit to the
    number of =-v= arguments. This is used to pass extra data to the
    program under certain circumstances.
  - =user= is the name of the user to be authenticated.
  - =class= is optional and specifies the class of the user to be
    authenticated.

  =login= and =su= pass in extra data as =-v= flags.

  #+CAPTION: Taken from [[https://man.openbsd.org/login.conf][=login.conf(5)=]]
  #+BEGIN_SRC
  The login(1) program provides the following through the -v option:

     auth_type       The type of authentication to use.

     fqdn            The hostname provided to login by the -h option.

     hostname        The name login(1) will place in the utmp file for the
                     remote hostname.

     local_addr      The local IP address given to login(1) by the -L option.

     lastchance      Set to "yes" when a user's password has expired but the
                     user is being given one last chance to login and update
                     the password.

     login           This is a new login session (as opposed to a simple
                     identity check).

     remote_addr     The remote IP address given to login(1) by the -R option.

     style           The style of authentication used for this user (see
                     approval scripts below).

     The su(1) program provides the following through the -v option:

     wheel           Set to either "yes" or "no" to indicate if the user is in
                     group wheel when they are trying to become root.  Some
                     authentication types require the user to be in group
                     wheel when using the su(1) program to become super user.
  #+END_SRC

  The auth module communicates with its caller through file
  descriptor 3.

  Most modules also have a hidden flag =-d=, which sets the back
  channel do =stdio=, presumably for debugging purposes.

* Documentation

  All of the high level authentication functions are described in
  [[https://man.openbsd.org/authenticate][=authenticate(3)=]], with the lower level functions being described in
  [[https://man.openbsd.org/auth_subr][=auth_subr(3)=]].

* auth_userokay

  =auth_userokay= is the highest level function, and easiest to use.
  It takes four character arrays as arguments, =name=, =style=,
  =type=, and =password=. It returns either a =0= for failure, of a
  non-zero value for success.

  This function lives inside =/lib/libc/gen/authenticate.c=

  #+BEGIN_SRC c
  int auth_userokay(char *name, char *style, char *type, char *password);
  #+END_SRC

  - =name= is the name of the user to be authenticated
  - =style= is the login method to be used
    - If =style= is =NULL=, the user's default login style will be
      used. By default this is =passwd= on normal accounts.
    - The style can be one of the installed authentication methods, like
      =passwd=, =radius=, =skey=, =yubikey=, etc.
    - There's more information about available styles in =login.conf(5)=
    - Styles can also be installed through BSD Auth module packages
  - =type= is the authentication type
    - Types are defined in =login.conf= and define a group of allowed
      auth styles
    - If =type= is =NULL=, use the auth type for the user's login
      class. The default type is =auth-default=, which allows
      =psaswd= and =skey= auth methods.
    - There's more information about how to add methods in =login.conf(5)=
  - =password= is the password to test
    - If =password= is =NULL=, then the user is interactively
      prompted. This is required for auth styles using
      challenge-response methods.
    - If =password= is specified, then it's non-interactively tested

  =auth_userokay= is just a wrapper around =auth_usercheck=. It takes
  care of closing the session for you, and returns =0= instead of
  =NULL= on failure.

* auth_session_t
  =auth_session_t= is the main data structure used to represent the
  authentication session. It gets used by all other functions.

  #+BEGIN_SRC c
  struct auth_session_t {
      char    *name;                 /* name of use being authenticated */
      char    *style;                /* style of authentication used */
      char    *class;                /* class of user */
      char    *service;              /* type of service being performed */
      char    *challenge;            /* last challenge issued */
      int     flags;                 /* see below */
      struct  passwd *pwd;           /* password entry for user */
      struct  timeval now;           /* time of authentication */

      int     state;                 /* authenticated state */

      struct  rmfiles *rmlist;       /* list of files to remove on failure */
      struct  authopts *optlist;     /* list of options to scripts */
      struct  authdata *data;        /* additional data to send to scripts */

      char    spool[MAXSPOOLSIZE];   /* data returned from login script */
      int     index;                 /* how much returned thus far */

      int     fd;                    /* connection to authenticator */

      va_list ap0;                   /* argument list to auth_call */
      va_list ap;                    /* additional arguments to auth_call */
  };
  #+END_SRC

  Where =authdata=, =authopts=, and =rmfiles= are defined as

  #+BEGIN_SRC c
  struct rmfiles {
      struct rmfiles  *next;
      char            *file;
  };

  struct authopts {
      struct authopts *next;
      char            *opt;
  };

  struct authdata {
      struct  authdata *next;
      void    *ptr;
      size_t   len;
  };
  #+END_SRC

  There are several functions which get used to operate on
  =auth_session_t= to keep it opaque.
** auth_setdata
   #+begin_src c
   int auth_setdata(auth_session_t *as, void *ptr, size_t len)
   #+end_src

   =auth_setdata= allocates and initializes a new =authdata= struct,
   storing a copy of the data from =*ptr= and =len=. It then point the
   =next= field on the last =authdata= struct in =*as= to its
   location. It returns =0= on success.

** auth_setitem / auth_getitem
   #+begin_src c
   int auth_setitem(auth_session_t *as, auth_item_t item, char *value)
   #+end_src

   =auth_setitem= is used to set one of several different fields of
   =*as= to =*value=. Depending on the value of =item=, it can be the
   =challenge=, =class=, =name=, =service=, =style=, or =interactive=
   field. If =*value= is =NULL=, it clears that field. If =item= is
   =AUTHV_ALL= and =*value= is =NULL=, all fields are cleared. It
   returns =0= on success.

   #+begin_src c
   char *auth_getitem(auth_session_t *as, auth_item_t item)
   #+end_src

   =auth_getitem= is used to return the value of the fields listed above.

*** auth_item_t
    =auth_item_t= is an enum defined in =/include/bsd_auth.h=.

    #+begin_src c
    typedef enum {
        AUTHV_ALL,
        AUTHV_CHALLENGE,
        AUTHV_CLASS,
        AUTHV_NAME,
        AUTHV_SERVICE,
        AUTHV_STYLE,
        AUTHV_INTERACTIVE
    } auth_item_t;
    #+end_src

** auth_setoption
   #+begin_src c
   int auth_setoption(auth_session_t *as, char *n, char *v)
   #+end_src

   =auth_setoption= initializes a new =authopts= struct, and sets the
   =*opt= field to a string formatted as =sprintf(%s=%s, n, v)=. It
   then point the =*next= field on the last =authopts= struct in =*as=
   to its location. It returns =0= on success.

** auth_setstate
   #+begin_src c
   void	auth_setstate(auth_session_t *as, int s)
   #+end_src

   =auth_setstate= sets the =state= of =*as= to =s=.
* auth_open

  #+begin_src c
  auth_session_t *auth_open(void)
  #+end_src

  =auth_open= is used by several functions to create a new auth
  session. It allocates an =auth_session_t= struct on the heap, sets
  its default =service= to that defined by =LOGIN_DEFSERVICE= in
  =/include/login_cap.h=, which is currently ="login"=.

  #+begin_src c
  #define	LOGIN_DEFSERVICE	"login"
  #+end_src

  It then sets the =fd= field to =-1=, and returns the pointer.

* auth_usercheck

  #+BEGIN_SRC c
  auth_session_t *auth_usercheck(char *name, char *style, char *type, char *password)
  #+END_SRC

  =auth_usercheck= first checks that =*name= doesn't begin with a
  hyphen, and that it's not too long.

  If =*style= is =NULL=, it checks if =*name= is in the =user:style=
  format, and splits it accordingly.

  It then gets the user's password database entry through
  [[https://man.openbsd.org/man3/getpwnam.3#getpwnam_r][=getpwman_r(3)=]], which operates on the [[https://man.openbsd.org/passwd.5][=passwd(5)=]] database. It then
  uses that to retrieve the user's login class using
  [[https://man.openbsd.org/login_getclass#login_getclass][=login_getclass(3)=]], which returns a =login_cap_t=. Login classes
  are stored in the [[https://man.openbsd.org/man5/login.conf.5][=login.conf(5)=]] database.

  That struct is then passed into [[https://man.openbsd.org/login_getclass#login_getstyle][=login_getstyle(3)=]], which also
  received the =*style= and =*type=. If =*type= is =NULL=, it returns
  the first available login style for that class. If =*style= is
  specified, it is returned if available, otherwise =NULL= is
  returned, which causes =auch_usercheck= to return =NULL= as well.

  It then creates a pointer =as= of type =auth_session_t=, and handles
  it differently based on whether =*password= is =NULL=.

  - If the password is a string, it creates a new session using
    =auth_open= and assigns it to =as=. It then sets the session
    =service= to ="response"=, and adds the =password= string to the
    session's =data=.

    #+BEGIN_SRC c
    auth_setitem(as, AUTHV_SERVICE, "response");
    auth_setdata(as, "", 1);
    auth_setdata(as, password, strlen(password) + 1);
    #+END_SRC

  - If =*password= is =NULL=, it sets =as= to =NULL=.

  It then passes the =auth_session_t= pointer (=as=), =*name=,
  =*style=, login class (=lc->lc_class=), and a =NULL= char pointer to
  =auth_verify=. It then returns the auth session pointer the call
  returns.

  #+begin_src c
  as = auth_verify(as, style, name, lc->lc_class, (char *)NULL);
  // [...] some cleanup
  return (as);
  #+end_src

* auth_verify

  #+BEGIN_SRC c
  auth_session_t *auth_verify(auth_session_t *as, char *style, char *name, ...)
  #+END_SRC

  =auth_verify= creates an auth session using =auth_open= if =as= is
  =NULL=.

  The =state= of the session is set to 0.

  It then sets the =name= and =style= of the session, if the
  respective arguments are non-=NULL=. It then copies its variable
  arguments to the auth session's =va_list ap=, which is used inside
  of =auth_call=.

  After that it constructs the path of the authentication module by
  combining =_PATH_AUTHPROG=, which is defined in =login_cap.h= as
  =/usr/libexec/auth/login_=, and the authentication style. For the
  case of auth style =passwd=, it would result in the path
  =/usr/libexec/auth/login_passwd=.

  Then =auth_call= is called with the struct, the path to the auth
  module, the auth style, the "-s" flag followed by the service (login,
  challenge, response), a double dash, the user name, and a =NULL=
  character pointer. The return value of =auth_call= is ignored and a
  pointer to the auth session is returned immediately afterwards.

  #+BEGIN_SRC c
  auth_call(as, path, auth_getitem(as, AUTHV_STYLE), "-s",
      auth_getitem(as, AUTHV_SERVICE), "--", name, (char *)NULL);
  #+END_SRC

* auth_call

  #+BEGIN_SRC c
  int auth_call(auth_session_t *as, char *path, ...)
  #+END_SRC

  <<here>>

  ---
  note: In the man page auth_subr it says
  #+begin_quote
  path    The full path name of the login script to run.  The call will
               fail if path does not pass the requirements of the secure_path(3)
               function.
  #+end_quote
  However I don't see this enforced anywhere, I even wrote a small test
  script to prove that's the case on =vfwall ~/authtest=.

  The manpage also says the path is limited to =/bin/= and =/usr/bin=,
  which is also not the case.

  Ask jcs about the file descriptor situation, I don't understand it
  after reading both the man page and source.
  ---

  Inside of =auth_call=, a socket pair of type =PF_LOCAL, SOCK_STREAM=
  is created. This is called the "back channel", and is used to
  communicate with the authentication module. The process then forks,
  calling ~execve(path, argv, auth_environ)~, where the =argv= is
  everything after =path= in the =auth_call= arguments. Any =authopts=
  set in the auth session are also passed as arguments in the format =-v
  opt1 -v opt2 -v opt3=, etc. =auth_environ= is defined at the top of
  the file as

  #+BEGIN_SRC c
  static char *auth_environ[] = {
      "PATH=" _PATH_DEFPATH,
      "SHELL=" _PATH_BSHELL,
      NULL,
  };
  #+END_SRC

  Where both constants are defined in =paths.h= as

  #+BEGIN_SRC c
  #define	_PATH_DEFPATH	"/usr/bin:/bin:/usr/sbin:/sbin:/usr/X11R6/bin:/usr/local/bin:/usr/local/sbin"
  #define	_PATH_BSHELL	"/bin/sh"
  #+END_SRC


  The =exec='d process then listens on FD 3, which is one half of the
  =sockpair= that was created earlier.

  In the non-exec'd process, first the contents of the auth session's
  =*data= are read in one at a time.

  The data received through the back channel is then put into the
  =spool= of the auth session using =_auth_spool(as, pfd[0])=. After
  that the spooled data is scanned for key words defined in
  =login_cap.h=.

  #+BEGIN_SRC c
  #define BI_AUTH         "authorize"         /* Accepted authentication */
  #define BI_REJECT       "reject"            /* Rejected authentication */
  #define BI_CHALLENGE    "reject challenge"  /* Reject with a challenge */
  #define BI_SILENT       "reject silent"     /* Reject silently */
  #define BI_REMOVE       "remove"            /* remove file on error */
  #define BI_ROOTOKAY     "authorize root"    /* root authenticated */
  #define BI_SECURE       "authorize secure"  /* okay on non-secure line */
  #define BI_SETENV       "setenv"            /* set environment variable */
  #define BI_UNSETENV     "unsetenv"          /* unset environment variable */
  #define BI_VALUE        "value"             /* set local variable */
  #define BI_EXPIRED      "reject expired"    /* account expired */
  #define BI_PWEXPIRED    "reject pwexpired"  /* password expired */
  #define BI_FDPASS       "fd"                /* child is passing an fd */
  #+END_SRC

  The [[https://man.openbsd.org/login.conf][=login.conf(5)=]] man page once again goes into greater detail on
  these values.

  #+BEGIN_SRC
  authorize  The user has been authorized.

  authorize secure
             The user has been authorized and root should be allowed to
             login even if this is not a secure terminal.  This should only
             be sent by authentication styles that are secure over insecure
             lines.

  reject     Authorization is rejected.  This overrides any indication that
             the user was authorized (though one would question the wisdom
             in sending both a reject and an authorize command).

  reject challenge
             Authorization was rejected and a challenge has been made
             available via the value challenge.

  reject silent
             Authorization is rejected, but no error messages should be
             generated.

  remove file
             If the login session fails for any reason, remove file before
             termination.

  setenv name value
             If the login session succeeds, the environment variable name
             should be set to the specified value.

  unsetenv name
             If the login session succeeds, the environment variable name
             should be removed.

  value name value
             Set the internal variable name to the specified value.  The
             value should only contain printable characters.  Several \
             sequences may be used to introduce non printing characters.
             These are:

             \n      A newline.

             \r      A carriage return.

             \t      A tab.

             \xxx    The character represented by the octal value xxx.  The
                     value may be one, two, or three octal digits.

             \c      The string is replaced by the value of c.  This allows
                     quoting an initial space or the \ character itself.


             The following values are currently defined:

             challenge
                     See section on challenges below.

             errormsg
                     If set, the value is the reason authentication failed.
                     The calling program may choose to display this when
                     rejecting the user, but display is not required.

  #+END_SRC


  It is looking for lines that start with either =BI_AUTH=
  (=authorize=), or =BI_REJECT= (=reject=). If the line is still longer,
  it continues to scan for any other qualifiers such as =pwexpired= or
  =silent=. The struct's =state= is set to one using the =AUTH_= values
  from =login_cap.h= accordingly.

  #+BEGIN_SRC c
  /*
   * bits which can be returned by authenticate()/auth_scan()
   */
  #define  AUTH_OKAY       0x01            /* user authenticated */
  #define  AUTH_ROOTOKAY   0x02            /* authenticated as root */
  #define  AUTH_SECURE     0x04            /* secure login */
  #define  AUTH_SILENT     0x08            /* silent rejection */
  #define  AUTH_CHALLENGE  0x10            /* a challenge was given */
  #define  AUTH_EXPIRED    0x20            /* account expired */
  #define  AUTH_PWEXPIRED  0x40            /* password expired */
  #+END_SRC

** _auth_spool

** _recv_fd

* auth_close
  =auth_close= is the function responsible for cleaning up the session
  and taking care of the values returned though the back channel.

  It first sets the environment variables returned through the back
  channel by passing the auth session to =auth_setenv=. It then goes
  through the =rmlist= of the session, deleting the files if the
  session reported a failure. It then zeroes out all sensitive
  information, and frees the various structs associated with the current
  =auth_session_t=, and then the session itself. Finally it returns
  the session's state =&='ed with =AUTH_ALLOW=.
* grapgh?
# Setting env on auth_close(as)
# partual rewrite below

The call graph for =auth_userokay= looks something like this:

#+BEGIN_SRC c
int auth_userokay(char *name, char *style, char *type, char *password)
#+END_SRC

calls ~auth_usercheck~ and then calls ~auth_close~ on the returned
~auth_session_t~. The value returned from ~auth_close~ is then
returned.

#+BEGIN_SRC c
auth_session_t *auth_usercheck(char *name, char *style, char *type, char *password)
#+END_SRC

Validates the checks that the user exists, gets the user's login
class, verifies the auth type, and that the auth style can be used.

It creates an auth session struct.

If the password is provided it sets the service type to =response=,
and adds the adds the password to the auth data. Otherwise it
leaves it empty.