newlib-cygwin/winsup/doc/ntsec.xml

<?xml version="1.0" encoding='UTF-8'?>
<!DOCTYPE sect1 PUBLIC "-//OASIS//DTD DocBook V4.5//EN"
		"http://www.oasis-open.org/docbook/xml/4.5/docbookx.dtd">

<sect1 id="ntsec"><title>Using Windows security in Cygwin</title>

<para>This section discusses how the Windows security model is
utilized in Cygwin to implement POSIX-like permissions, as well as how
the Windows authentication model is used to allow cygwin applications
to switch users in a POSIX-like fashion.</para>

<para>The setting of POSIX-like file and directory permissions is
controlled by the <link linkend="mount-table">mount</link> option
<literal>(no)acl</literal> which is set to <literal>acl</literal> by
default.</para>

<para>We start with a short overview.  Note that this overview must
be necessarily short.  If you want to learn more about the Windows security
model, see the <ulink url="http://msdn.microsoft.com/en-us/library/aa374860(VS.85).aspx">Access Control</ulink> article in MSDN documentation.</para>

<para>POSIX concepts and in particular the POSIX security model are not
discussed here, but assumed to be understood by the reader.  If you
don't know the POSIX security model, search the web for beginner
documentation.</para>

<sect2 id="ntsec-common"><title>Overview</title>

<para>In the Windows security model, almost any "object" is securable.
"Objects" are files, processes, threads, semaphores, etc.</para>

<para>Every object has a data structure attached, called a "security
descriptor" (SD).  The SD contains all information necessary to control
who can access an object, and to determine what they are allowed to do
to or with it.  The SD of an object consists of five parts:</para>

<itemizedlist spacing="compact">
<listitem><para>Flags which control several aspects of this SD. This is
not discussed here.</para></listitem>
<listitem><para>The SID of the object owner.</para></listitem>
<listitem><para>The SID of the object owner group.</para></listitem>
<listitem><para>A list of "Access Control Entries" (ACE), called the
"Discretionary Access Control List" (DACL).</para></listitem>
<listitem><para>Another list of ACEs, called the "Security Access Control List"
(SACL), which doesn't matter for our purpose.  We ignore it here.</para></listitem>
</itemizedlist>

<para>Every ACE contains a so-called "Security IDentifier" (SID) and
other stuff which is explained a bit later.  Let's talk about the SID first.
</para>

<para>A SID is a unique identifier for users, groups, computers and
Active Directory (AD) domains.  SIDs are basically comparable to POSIX
user ids (UIDs) and group ids (GIDs), but are more complicated because
they are unique across multiple machines or domains.  A SID is a
structure of multiple numerical values.  There's a convenient convention
to type SIDs, as a string of numerical fields separated by hyphen
characters.  Here's an example:</para>

<para>SID of a machine "foo":</para>

<screen>
  S-1-5-21-165875785-1005667432-441284377
</screen>

<para>SID of a user "johndoe" of the system "foo":</para>

<screen>
  S-1-5-21-165875785-1005667432-441284377-1023
</screen>

<para>The first field is always "S", which is just a notational convention
to show that this is a SID.  The second field is the version number of
the SID structure, So far there exists only one version of SIDs, so this
field is always 1.  The third and fourth fields represent the "authority"
which can be thought of as a type or category of SIDs.  There are a
couple of builtin accounts and accounts with very special meaning which
have certain well known values in these third and fourth fields.
However, computer and domain SIDs always start with "S-1-5-21".  The
next three fields, all 32 bit values, represent the unique 96 bit
identifier of the computer system.  This is a hopefully unique value all
over the world, but in practice it's sufficient if the computer SIDs are
unique within a single Windows network.</para>

<para>As you can see in the above example, SIDs of users (and groups)
are identical to the computer SID, except for an additional part, the
so-called "relative identifier" (RID).  So the SID of a user is always 
uniquely attached to the system on which the account has been generated.</para>

<para>It's a bit different in domains.  The domain has its own SID, and
that SID is identical to the SID of the first domain controller, on
which the domain is created.  Domain user SIDs look exactly like the
computer user SIDs, the leading part is just the domain SID and the RID
is created when the user is created.</para>

<para>Ok, consider you created a new domain "bar" on some new domain
controller and you would like to create a domain account "johndoe":</para>

<para>SID of a domain "bar.local":</para>

<screen>
  S-1-5-21-186985262-1144665072-740312968
</screen>

<para>SID of a user "johndoe" in the domain "bar.local":</para>

<screen>
  S-1-5-21-186985262-1144665072-740312968-1207
</screen>

<para>So you now have two accounts called johndoe, one account
created on the machine "foo", one created in the domain "bar.local".
Both have different SIDs and not even the RID is the same.  How do
the systems know it's the same account?  After all, the name is
the same, right?  The answer is, these accounts are <emphasis
role='bold'>not</emphasis> identical.  All machines on the network will
treat these SIDs as identifying two separate accounts.  One is
"FOO\johndoe", the other one is "BAR\johndoe" or "johndoe@bar.local".
Different SID, different account.  Full stop.  </para>

<para>The last part of the SID, the so called "Relative IDentifier" (RID),
is by default used as UID and/or GID under Cygwin when you create the
<filename>/etc/passwd</filename> and <filename>/etc/group</filename>
files using the <command><link linkend="mkpasswd">mkpasswd</link></command> and <command><link linkend="mkgroup">mkgroup</link></command>
tools.  Domain account UIDs and GIDs are offset by 10000 by default
which might be a bit low for very big organizations.  Fortunately there's
an option in both tools to change the offset...</para>

<para>Do you still remember the SIDs with special meaning?  In offical
notation they are called "well-known SIDs".  For example, POSIX has no GID
for the group of "all users" or "world" or "others".  The last three rwx
bits in a unix-style permission value just represent the permissions for
"everyone who is not the owner or is member of the owning group".
Windows has a SID for these poor souls, the "Everyone" SID.  Other
well-known SIDs represent circumstances under which a process is
running, rather than actual users or groups.  Here are a few examples
for well-known SIDs:</para>

<screen>
Everyone                        S-1-1-0    Simply everyone...
Batch                           S-1-5-3    Processes started via the task
					   scheduler are member of this group.
Interactive			S-1-5-4    Only processes of users which are
					   logged in via an interactive
					   session are members here.
Authenticated Users             S-1-5-11   Users which have gone through
                                           the authentication process and
					   survived.  Anonymously accessing
					   users are not incuded here.
SYSTEM                          S-1-5-18   A special account which has all
					   kinds of dangerous rights, sort of
					   an uber-root account.
</screen>

<para>For a full list please refer to the MSDN document <ulink
url="http://msdn.microsoft.com/en-us/library/aa379649.aspx">Well-known
SIDs</ulink>.  The Cygwin package called "csih" provides a tool,
/usr/lib/csih/getAccountName.exe, which can be used to print the
(possibly localized) name for the various well-known SIDS.</para>

<para>Naturally, well-known SIDs are the same on each machine, so they are
not unique to a machine or domain.  They have the same meaning across
the Windows network.</para>

<para>Additionally, there are a couple of well-known builtin groups,
which have the same SID on every machine and which have certain user
rights by default:</para>

<screen>
administrators                  S-1-5-32-544
users                           S-1-5-32-545
guests                          S-1-5-32-546
...
</screen>

<para>For instance, every account is usually member in the "Users"
group.  All administrator accounts are member of the "Administrators"
group.  That's all about it as far as single machines are involved.  In
a domain environment it's a bit more tricky.  Since these SIDs are not
unique to a machine, every domain user and every domain group can be a
member of these well known groups.  Consider the domain group "Domain
Admins".  This group is by default in the "Administrators" group.  Let's
assume the above computer called "foo" is a member machine of the domain
"bar.local".  If you stick the user "BAR\johndoe" into the group "Domain
Admins", this guy will automatically be a member of the administrators
group on "foo" when logging on to "foo".  Neat, isn't it?</para>

<para>Back to ACE and ACL.  POSIX is able to create three different
permissions, the permissions for the owner, for the group and for the
world.  In contrast the Windows ACL has a potentially infinite number of
members... as long as they fit into 64K.  Every member is an ACE.
ACE consist of three parts:</para>

<itemizedlist spacing="compact">
<listitem><para>The type of the ACE (allow ACE or deny ACE).</para></listitem>
<listitem><para>Permission bits, 32 of them.</para></listitem>
<listitem><para>The SID for which the permissions are allowed or denied.</para></listitem>
</itemizedlist>

<para>The two (for us) important types of ACEs are the "access allowed
ACE" and the "access denied ACE".  As the names imply, the allow ACE
tells the system to allow the given permissions to the SID, the deny ACE
results in denying the specific permission bits.</para>

<para>The possible permissions on objects are more detailed than in
POSIX.  For example, the permission to delete an object is different
from the permission to change object data, and even changing object data
can be separated into different permission bits for different kind of
data.  But there's a problem with the definition of a "correct" ACL
which disallows mapping of certain POSIX permissions cleanly.  See
<xref linkend="ntsec-mapping"></xref>.</para>

<para>POSIX is able to create only three different permissions?  Not quite.
Newer operating systems and file systems on POSIX systems also provide
access control lists.  Two different APIs exist for accessing these
ACLs, the Solaris API and the POSIX API.  Cygwin implements the Solaris
API to access Windows ACLs in a Unixy way.  At the time of writing this
document, the Cygwin implementation of the Solaris API isn't quite up
to speed.  For instance, it doesn't handle access denied ACEs gracefully.
So, use with care.  Online man pages for the Solaris ACL API can be
found on <ulink url="http://docs.sun.com">http://docs.sun.com</ulink>.</para>

</sect2>

<sect2 id="ntsec-files"><title id="ntsec-files.title">File permissions</title>

<para>On NTFS and if the <literal>noacl</literal> mount option is not
specified for a mount point, Cygwin sets file permissions as in POSIX.
Basically this is done by defining a SD with the matching owner and group
SIDs, and a DACL which contains ACEs for the owner, the group and for
"Everyone", which represents what POSIX calls "others".</para>

<para>To use Windows security correctly, Cygwin depends on the files
<filename>/etc/passwd</filename> and <filename>/etc/group</filename>.
These files define the translation between the Cygwin uid/gid and the
Windows SID.  The SID is stored in the pw_gecos field in
<filename>/etc/passwd</filename>, and in the gr_passwd field in
<filename>/etc/group</filename>. Since the pw_gecos field can contain
more information than just a SID, there are some rules for the layout.
It's required that the SID is the last entry of the pw_gecos field,
assuming that the entries in pw_gecos are comma-separated.  The
commands <command>mkpasswd</command> and <command>mkgroup</command>
usually do this for you.</para>

<para>Another interesting entry in the pw_gecos field (which is also
usually created by running <command>mkpasswd</command>) is the Windows user
name entry.  It takes the form "U-domain\username" and is sometimes used
by services to authenticate a user.  Logging in through
<command>telnet</command> is a common scenario.</para>

<para>A typical snippet from <filename>/etc/passwd</filename>:</para>

<example id="ntsec-passwd">
<title>/etc/passwd:</title>
<screen>
SYSTEM:*:18:544:,S-1-5-18::
Administrators:*:544:544:,S-1-5-32-544::
Administrator:unused:500:513:U-FOO\Administrator,S-1-5-21-790525478-115176313-839522115-500:/home/Administrator:/bin/bash
corinna:unused:11001:11125:U-BAR\corinna,S-1-5-21-2913048732-1697188782-3448811101-1001:/home/corinna:/bin/tcsh
</screen>
</example>

<para>The SYSTEM entry is usually needed by services.  The Administrators
entry (Huh?  A group in /etc/passwd?) is only here to allow
<command>ls</command> and similar commands to print some file ownerships
correctly.  Windows doesn't care if the owner of a file is a user or a
group.  In older versions of Windows NT the default ownership for files
created by an administrator account was set to the group Administrators
instead of to the creating user account.  This has changed, but you can
still switch to this setting on newer systems.  So it's convenient to
have the Administrators group in
<filename>/etc/passwd</filename>.</para>

<para>The really interesting entries are the next two.  The Administrator
entry is for the local administrator, the corinna entry matches the corinna
account in the domain BAR.  The information given in the pw_gecos field
are all we need to exactly identify an account, and to have a two way
translation, from Windows account name/SID to Cygwin account name uid and
vice versa.  Having this complete information allows us to choose a Cygwin
user name and uid which doesn't have to match the Windows account at all.  As
long as the pw_gecos information is available, we're on the safe side:</para>

<example id="ntsec-passwd-tweaked">
<title>/etc/passwd, tweaked:</title>
<screen>
root:unused:0:513:U-FOO\Administrator,S-1-5-21-790525478-115176313-839522115-500:/home/Administrator:/bin/bash
thursday_next:unused:11001:11125:U-BAR\corinna,S-1-5-21-2913048732-1697188782-3448811101-1001:/home/corinna:/bin/tcsh
</screen>
</example>

<para>  The above <filename>/etc/passwd</filename> will still work fine.
You can now login via <command>ssh</command> as the user "root", and
Cygwin dutifully translates "root" into the Windows user
"FOO\Administrator" and files owned by FOO\Administrator are shown to
have the uid 0 when calling <command>ls -ln</command>.  All you do you're
actually doing as Administrator.  Files created as root will be owned by
FOO\Administrator.  And the domain user BAR\corinna can now happily
pretend to be Thursday Next, but will wake up sooner or later finding
out she's still actually the domain user BAR\corinna...</para>

<para>Do I have to mention that you can also rename groups in
<filename>/etc/group</filename>?  As long as the SID is present and correct,
all is well.  This allows you to, for instance, rename the "Administrators"
group to "root" as well:</para>

<example id="ntsec-group-tweaked">
<title>/etc/group, tweaked:</title>
<screen>
root:S-1-5-32-544:544:
</screen>
</example>

<para>Last but not least, you can also change the primary group of a user
in <filename>/etc/passwd</filename>.  The only requirement is that the user
is actually a member of the new primary group in Windows.  For instance,
normal users in a domain environment are members in the group "Domain Users",
which in turn belongs to the well-known group "Users".  So, if it's
more convenient in your environment for the user's primary group to be
"Users", just set the user's primary group in <filename>/etc/passwd</filename>
to the Cygwin uid of "Users" (see in <filename>/etc/group</filename>,
default 545) and let the user create files with a default group ownership
of "Users".</para>

<note><para>
If you wish to make these kind of changes to /etc/passwd and /etc/group,
do so only if you feel comfortable with the concepts.  Otherwise, do not
be surprised if things break in either subtle or surprising ways!  If you
do screw things up, revert to copies of <filename>/etc/passwd</filename>
and <filename>/etc/group</filename> files created by
<command>mkpasswd</command> and <command>mkgroup</command>.  (Make
backup copies of these files before modifying them.)  Especially, don't
change the UID or the name of the user SYSTEM.  It may mostly work, but
some Cygwin applications running as a local service under that account
could suddenly start behaving strangely.
</para></note>

</sect2>

<sect2 id="ntsec-ids"><title id="ntsec-ids.title">Special values of user and group ids</title>

<para>If the current user is not present in
<filename>/etc/passwd</filename>, that user's uid is set to a
special value of 400.  The user name for the current user will always be
shown correctly.  If another user (or a Windows group, treated as a
user) is not present in <filename>/etc/passwd</filename>, the uid of
that user will have a special value of -1 (which would be shown by
<command>ls</command> as 65535).  The user name shown in this case will
be '????????'.</para>

<para>If the current user is not present in
<filename>/etc/passwd</filename>, that user's login gid is set to a
special value of 401.  The gid 401 is shown as 'mkpasswd',
indicating the command that should be run to alleviate the
situation.</para>

<para>If another user is not present in
<filename>/etc/passwd</filename>, that user's login gid is set to a
special value of -1.  If the user is present in
<filename>/etc/passwd</filename>, but that user's group is not in
<filename>/etc/group</filename> and is not the login group of that user,
the gid is set to a special value of -1.  The name of this group
(id -1) will be shown as '????????'.</para>

<para>If the current user is present in
<filename>/etc/passwd</filename>, but that user's login group is not
present in <filename>/etc/group</filename>, the group name will be shown
as 'mkgroup', again indicating the appropriate command.</para>

<para>A special case is if the current user's primary group SID is noted
in the user's <filename>/etc/passwd</filename> entry using another group
id than the group entry of the same group SID in
<filename>/etc/group</filename>.  This should be noted and corrected.
The group name printed in this case is
'passwd/group_GID_clash(PPP/GGG)', with PPP being the gid as noted
in <filename>/etc/passwd</filename> and GGG the gid as noted in
<filename>/etc/group</filename>.</para>

<para>To summarize:</para>
<itemizedlist spacing="compact">

<listitem><para>If the current user doesn't show up in
<filename>/etc/passwd</filename>, it's <emphasis>group</emphasis> will
be named 'mkpasswd'.</para></listitem>

<listitem><para>Otherwise, if the login group of the current user isn't
in <filename>/etc/group</filename>, it will be named 'mkgroup'.</para>
</listitem>

<listitem><para>Otherwise a group not in <filename>/etc/group</filename>
will be shown as '????????' and a user not in
<filename>/etc/passwd</filename> will be shown as "????????".</para>
</listitem>

<listitem><para>If different group ids are used for a group with the same
SID, the group name is shown as 'passwd/group_GID_clash(PPP/GGG)' with
PPP and GGG being the different group ids.</para></listitem>

</itemizedlist>

<para>
Note that, since the special user and group names are just indicators,
nothing prevents you from actually having a user named `mkpasswd' in
<filename>/etc/passwd</filename> (or a group named `mkgroup' in
<filename>/etc/group</filename>).  If you do that, however, be aware of
the possible confusion.
</para>

</sect2>


<sect2 id="ntsec-mapping"><title id="ntsec-mapping.title">The POSIX permission mapping leak</title>

<para>As promised earlier, here's the problem when trying to map the
POSIX permission model onto the Windows permission model.</para>

<para>There's a leak in the definition of a "correct" ACL which
disallows a certain POSIX permission setting.  The official
documentation explains in short the following:</para>

<itemizedlist spacing="compact">
<listitem><para>The requested permissions are checked against all
ACEs of the user as well as all groups the user is member of.  The
permissions given in these user and groups access allowed ACEs are
accumulated and the resulting set is the set of permissions of that
user given for that object.</para></listitem>

<listitem><para>The order of ACEs is important. The system reads them in
sequence until either any single requested permission is denied or all
requested permissions are granted.  Reading stops when this condition is
met.  Later ACEs are not taken into account.</para></listitem>

<listitem><para>All access denied ACEs <emphasis
role='bold'>should</emphasis> precede any access allowed ACE.  ACLs
following this rule are called "canonical"</para></listitem>
</itemizedlist>

<para>Note that the last rule is a preference or a definition of
correctness.  It's not an absolute requirement.  All Windows kernels
will correctly deal with the ACL regardless of the order of allow and
deny ACEs.  The second rule is not modified to get the ACEs in the
preferred order.</para>

<para>Unfortunately the security tab in the file properties dialog of
the Windows Explorer insists to rearrange the order of the ACEs to
canonical order before you can read them. Thank God, the sort order
remains unchanged if one presses the Cancel button.  But don't even
<emphasis role='bold'>think</emphasis> of pressing OK...</para>

<para>Canonical ACLs are unable to reflect each possible combination
of POSIX permissions. Example:</para>

<screen>
rw-r-xrw-
</screen>

<para>Ok, so here's the first try to create a matching ACL, assuming
the Windows permissions only have three bits, as their POSIX counterpart:
</para>

<screen>
UserAllow:   110
GroupAllow:  101
OthersAllow: 110
</screen>

<para>Hmm, because of the accumulation of allow rights the user may
execute because the group may execute.</para>

<para>Second try:</para>

<screen>
UserDeny:    001
GroupAllow:  101
OthersAllow: 110
</screen>

<para>Now the user may read and write but not execute. Better? No!
Unfortunately the group may write now because others may write.</para>

<para>Third try:</para>

<screen>
UserDeny:    001
GroupDeny:   010
GroupAllow:  001
OthersAllow: 110
</screen>

<para>Now the group may not write as intended but unfortunately the user may
not write anymore, either. How should this problem be solved? According to
the canonical order a UserAllow has to follow the GroupDeny but it's
easy to see that this can never be solved that way.</para>

<para>The only chance:</para>

<screen>
UserDeny:    001
UserAllow:   010
GroupDeny:   010
GroupAllow:  001
OthersAllow: 110
</screen>

<para>Again: This works on all existing versions of Windows NT, at the
time of writing from at least Windows XP up to Server 2012.  Only
the GUIs aren't able (or willing) to deal with that order.</para>

</sect2>

<sect2 id="ntsec-setuid-overview"><title id="ntsec-setuid-overview.title">Switching the user context</title>

<para>Since Windows XP, Windows users have been accustomed to the
"Switch User" feature, which switches the entire desktop to another user
while leaving the original user's desktop "suspended".  Another Windows
feature is the "Run as..." context menu entry, which allows you to start
an application using another user account when right-clicking on applications
and shortcuts.</para>

<para>On POSIX systems, this operation can be performed by processes
running under the privileged user accounts (usually the "root" user
account) on a per-process basis.  This is called "switching the user
context" for that process, and is performed using the POSIX
<command>setuid</command> and <command>seteuid</command> system
calls.</para>

<para>While this sort of feature is available on Windows as well,
Windows does not support the concept of these calls in a simple fashion.
Switching the user context in Windows is generally a tricky process with
lots of "behind the scenes" magic involved.</para>

<para>Windows uses so-called `access tokens' to identify a user and its
permissions.  Usually the access token is created at logon time and then
it's attached to the starting process.  Every new process within a session
inherits the access token from its parent process.  Every thread can
get its own access token, which allows, for instance, to define threads
with restricted permissions.</para>

</sect2>

<sect2 id="ntsec-logonuser"><title id="ntsec-logonuser.title">Switching the user context with password authentication</title>

<para>To switch the user context, the process has to request such an access
token for the new user.  This is typically done by calling the Win32 API
function <command>LogonUser</command> with the user name and the user's
cleartext password as arguments.  If the user exists and the password was
specified correctly, the access token is returned and either used in
<command>ImpersonateLoggedOnUser</command> to change the user context of
the current thread, or in <command>CreateProcessAsUser</command> to
change the user context of a spawned child process.</para>

<para>Later versions of Windows define new functions in this context and
there are also functions to manipulate existing access tokens (usually
only to restrict them).  Windows Vista also adds subtokens which are
attached to other access tokens which plays an important role in the UAC
(User Access Control) facility of Vista and later.  However, none of
these extensions to the original concept are important for this
documentation.</para>

<para>Back to this logon with password, how can this be used to
implement <command>set(e)uid</command>?  Well, it requires modification
of the calling application.  Two Cygwin functions have been introduced
to support porting <command>setuid</command> applications which only
require login with passwords.  You only give Cygwin the right access
token and then you can call <command>seteuid</command> or
<command>setuid</command> as usual in POSIX applications.  Porting such
a <command>setuid</command> application is illustrated by a short
example:</para>

<screen>
<![CDATA[
/* First include all needed cygwin stuff. */
#ifdef __CYGWIN__
#include <windows.h>
#include <sys/cygwin.h>
#endif

[...]

  struct passwd *user_pwd_entry = getpwnam (username);
  char *cleartext_password = getpass ("Password:");

[...]

#ifdef __CYGWIN__
  /* Patch the typical password test. */
  {
    HANDLE token;

    /* Try to get the access token from Windows. */
    token = cygwin_logon_user (user_pwd_entry, cleartext_password);
    if (token == INVALID_HANDLE_VALUE)
       error_exit;
    /* Inform Cygwin about the new impersonation token. */
    cygwin_set_impersonation_token (token);
    /* Cygwin is now able, to switch to that user context by setuid or seteuid calls. */
  }
#else
    /* Use standard method on non-Cygwin systems. */
    hashed_password = crypt (cleartext_password, salt);
    if (!user_pwd_entry ||
        strcmp (hashed_password, user_pwd_entry->pw_password))
      error_exit;
#endif /* CYGWIN */

[...]

  /* Everything else remains the same! */

  setegid (user_pwd_entry->pw_gid);
  seteuid (user_pwd_entry->pw_uid);
  execl ("/bin/sh", ...);
]]>

</screen>

</sect2>

<sect2 id="ntsec-nopasswd1"><title id="ntsec-nopasswd1.title">Switching the user context without password, Method 1: Create a token from scratch</title>

<para>An unfortunate aspect of the implementation of
<command>set(e)uid</command> is the fact that the calling process
requires the password of the user to which to switch.  Applications such as
<command>sshd</command> wishing to switch the user context after a
successful public key authentication, or the <command>cron</command>
application which, again, wants to switch the user without any authentication
are stuck here.  But there are other ways to get new user tokens.</para>

<para>One way is just to create a user token from scratch.  This is
accomplished by using an (officially undocumented) function on the NT
function level.  The NT function level is used to implement the Win32
level, and, as such is closer to the kernel than the Win32 level.  The
function of interest, <command>NtCreateToken</command>, allows you to
specify user, groups, permissions and almost everything you need to
create a user token, without the need to specify the user password.  The
only restriction for using this function is that the calling process
needs the "Create a token object" user right, which only the SYSTEM user
account has by default, and which is considered the most dangerous right
a user can have on Windows systems.</para>

<para>That sounds good.  We just start the servers which have to switch
the user context (<command>sshd</command>, <command>inetd</command>,
<command>cron</command>, ...) as Windows services under the SYSTEM
(or LocalSystem in the GUI) account and everything just works.
Unfortunately that's too simple.  Using <command>NtCreateToken</command>
has a few drawbacks.</para>

<para>First of all, beginning with Windows Server 2003,
the permission "Create a token object" gets explicitly removed from
the SYSTEM user's access token, when starting services under that
account.  That requires us to create a new account with this specific
permission just to run this kind of services.  But that's a minor
problem.</para>

<para>A more important problem is that using <command>NtCreateToken</command>
is not sufficient to create a new logon session for the new user.  What
does that mean?  Every logon usually creates a new logon session.
A logon session has a couple of attributes which are unique to the
session.  One of these attributes is the fact, that Windows functions
identify the user domain and user name not by the SID of the access
token owner, but only by the logon session the process is running under.</para>

<para>This has the following unfortunate consequence.  Consider a
service started under the SYSTEM account (up to Windows XP) switches the
user context to DOMAIN\my_user using a token created directly by calling
the <command>NtCreateToken</command> function.  A process running under
this new access token might want to know under which user account it's
running.  The corresponding SID is returned correctly, for instance
S-1-5-21-1234-5678-9012-77777.  However, if the same process asks the OS
for the user name of this SID something wierd happens.  For instance,
the <command>LookupAccountSid</command> function will not return
"DOMAIN\my_user", but "NT AUTHORITY\SYSTEM" as the user name.</para>

<para>You might ask "So what?"  After all, this only <emphasis
role='bold'>looks</emphasis> bad, but functionality and permission-wise
everything should be ok.  And Cygwin knows about this shortcoming so it
will return the correct Cygwin username when asked.  Unfortunately this
is more complicated.  Some native, non-Cygwin Windows applications will
misbehave badly in this situation.  A well-known example are certain versions
of Visual-C++.</para>

<para>Last but not least, you don't have the usual comfortable access
to network shares.  The reason is that the token has been created
without knowing the password.  The password are your credentials
necessary for network access.  Thus, if you logon with a password, the
password is stored hidden as "token credentials" within the access token
and used as default logon to access network resources.  Since these
credentials are missing from the token created with
<command>NtCreateToken</command>, you only can access network shares
from the new user's process tree by using explicit authentication, on
the command line for instance:</para>

<screen>
bash$ net use '\\server\share' /user:DOMAIN\my_user my_users_password
</screen>

<para>Note that, on some systems, you can't even define a drive letter
to access the share, and under some circumstances the drive letter you
choose collides with a drive letter already used in another session.
Therefore it's better to get used to accessing these shares using the UNC
path as in</para>

<screen>
bash$ grep foo //server/share/foofile
</screen>

</sect2>

<sect2 id="ntsec-nopasswd2"><title id="ntsec-nopasswd2.title">Switching the user context without password, Method 2: LSA authentication package</title>

<para>We're looking for another way to switch the user context without
having to provide the password.  Another technique is to create an
LSA authentication package.  LSA is an acronym for "Local Security Authority"
which is a protected part of the operating system which only allows changes
to become active when rebooting the system after the change.  Also, as soon as
the LSA encounters serious problems (for instance, one of the protected
LSA processes died), it triggers a system reboot.  LSA is the part of
the OS which cares for the user logons and which also creates logon
sessions.</para>

<para>An LSA authentication package is a DLL which has to be installed
as part of the LSA.  This is done by tweaking a special registry key.
Cygwin provides such an authentication package.  It has to be installed
and the machine has to be rebooted to activate it.  This is the job of the
shell script <filename>/usr/bin/cyglsa-config</filename> which is part of
the Cygwin package.</para>

<para>After running <filename>/usr/bin/cyglsa-config</filename> and
rebooting the system, the LSA authentication package is used by Cygwin
when <command>set(e)uid</command> is called by an application.  The
created access token using this method has its own logon session.</para>

<para>This method has two advantages over the <command>NtCreateToken</command>
method.</para>

<para>The very special and very dangerous "Create a token object" user
right is not required by a user using this method.  Other privileged
user rights are still necessary, especially the "Act as part of the
operating system" right, but that's just business as usual.</para>

<para>The user is correctly identified, even by delicate native applications
which choke on that using the <command>NtCreateToken</command> method.</para>

<para>Disadvantages?  Yes, sure, this is Windows.  The access token
created using LSA authentication still lacks the credentials for network
access.  After all, there still hasn't been any password authentication
involved.  The requirement to reboot after every installation or
deinstallation of the cygwin LSA authentication DLL is just a minor
inconvenience compared to that...</para>

<para>Nevertheless, this is already a lot better than what we get by
using <command>NtCreateToken</command>, isn't it?</para>

</sect2>

<sect2 id="ntsec-nopasswd3"><title id="ntsec-nopasswd3.title">Switching the user context without password, Method 3: With password</title>

<para>Ok, so we have solved almost any problem, except for the network
access problem.  Not being able to access network shares without
having to specify a cleartext password on the command line or in a 
script is a harsh problem for automated logons for testing purposes
and similar stuff.</para>

<para>Fortunately there is a solution, but it has its own drawbacks.
But, first things first, how does it work?  The title of this section
says it all.  Instead of trying to logon without password, we just logon
with password.  The password gets stored two-way encrypted in a hidden,
obfuscated area of the registry, the LSA private registry area.  This
part of the registry contains, for instance, the passwords of the Windows
services which run under some non-default user account.</para>

<para>So what we do is to utilize this registry area for the purpose of
<command>set(e)uid</command>.  The Cygwin command <command><link
linkend="passwd">passwd</link> -R</command> allows a user to specify
his/her password for storage in this registry area.  When this user
tries to login using ssh with public key authentication, Cygwin's
<command>set(e)uid</command> examines the LSA private registry area and
searches for a Cygwin specific key which contains the password.  If it
finds it, it calls <command>LogonUser</command> under the hood, using
this password.  If that works, <command>LogonUser</command> returns an
access token with all credentials necessary for network access.</para>

<para>For good measure, and since this way to implement
<command>set(e)uid</command> is not only used by Cygwin but also by
Microsoft's SFU (Services for Unix), we also look for a key stored by
SFU (using the SFU command <command>regpwd</command>) and use that if it's
available.</para>

<para>We got it.  A full access token with its own logon session, with
all network credentials.  Hmm, that's heaven...</para>

<para>Back on earth, what about the drawbacks?</para>

<para>First, adding a password to the LSA private registry area
requires administrative access.  So calling <command>passwd -R</command>
as a normal user will fail!  Cygwin provides a workaround for
this.  If <command>cygserver</command> is started as a service running
under the SYSTEM account (which is the default way to run
<command>cygserver</command>) you can use <command>passwd -R</command>
as normal, non-privileged user as well.</para>

<para>Second, as aforementioned, the password is two-way encrypted in a
hidden, obfuscated registry area.  Only SYSTEM has access to this area
for listing purposes, so, even as an administrator, you can't examine
this area with regedit.  Right?  No.  Every administrator can start
regedit as SYSTEM user:</para>

<screen>
bash$ date
Tue Dec  2 16:28:03 CET 2008
bash$ at 16:29 /interactive regedit.exe
</screen>

<para>Additionally, if an administrator knows under which name
the private key is stored (which is well-known since the algorithms
used to create the Cygwin and SFU keys are no secret), every administrator
can access the password of all keys stored this way in the registry.</para>

<para>Conclusion: If your system is used exclusively by you, and if
you're also the only administrator of your system, and if your system is
adequately locked down to prevent malicious access, you can safely use
this method.  If your machine is part of a network which has
dedicated administrators, and you're not one of these administrators,
but you (think you) can trust your administrators, you can probably
safely use this method.</para>

<para>In all other cases, don't use this method.  You have been warned.</para>

</sect2>

<sect2 id="ntsec-setuid-impl"><title id="ntsec-setuid-impl.title">Switching the user context, how does it all fit together?</title>

<para>Now we learned about four different ways to switch the user
context using the <command>set(e)uid</command> system call, but
how does <command>set(e)uid</command> really work?  Which method does it
use now?</para>

<para>The answer is, all four of them.  So here's a brief overview
what <command>set(e)uid</command> does under the hood:</para>

<itemizedlist>
<listitem>
<para>When <command>set(e)uid</command> is called, it tests if the
user context had been switched by an earlier call already, and if the
new user account is the privileged user account under which the process
had been started originally.  If so, it just switches to the original
access token of the process it had been started with.</para>
</listitem>

<listitem>
<para>
Next, it tests if an access token has been stored by an earlier call
to <command>cygwin_set_impersonation_token</command>.  If so, it tests
if that token matches the requested user account.  If so, the stored
token is used for the user context switch.</para>

<para>
If not, there's no predefined token which can just be used for
the user context switch, so we have to create a new token.  The order
is as follows.</para>
</listitem>

<listitem>
<para>Check if the user has stored the logon password in the LSA
private registry area, either under a Cygwin key, or under a SFU key.
If so, use this to call <command>LogonUser</command>.  If this
succeeds, we use the resulting token for the user context switch.</para>
</listitem>

<listitem>
<para>Otherwise, check if the Cygwin-specifc LSA authentication package
has been installed and is functional.  If so, use the appropriate LSA
calls to communicate with the Cygwin LSA authentication package and
use the returned token.</para>
</listitem>

<listitem>
<para>Last chance, try to use the <command>NtCreateToken</command> call
to create a token.  If that works, use this token.</para>
</listitem>

<listitem>
<para>If all of the above fails, our process has insufficient privileges
to switch the user context at all, so <command>set(e)uid</command>
fails and returns -1, setting errno to EPERM.</para>
</listitem>
</itemizedlist>

</sect2>

</sect1>