Replace entities with xi:include

[Samba.git] / docs / devel / unix-smb.xml

<?xml version="1.0" encoding="iso-8859-1"?>
<!DOCTYPE book PUBLIC "-//OASIS//DTD DocBook XML V4.2//EN"
                "http://www.oasis-open.org/docbook/xml/4.2/docbookx.dtd" [
  <!ENTITY % global_entities SYSTEM '../entities/global.entities'>
  %global_entities;
]>
<chapter id="unix-smb">
<chapterinfo>
        <author>
                <firstname>Andrew</firstname><surname>Tridgell</surname>
        </author>
        <pubdate>April 1995</pubdate>
</chapterinfo>

<title>NetBIOS in a Unix World</title>

<sect1>
<title>Introduction</title>
<para>
This is a short document that describes some of the issues that
confront a SMB implementation on unix, and how Samba copes with
them. They may help people who are looking at unix&lt;-&gt;PC
interoperability.
</para>

<para>
It was written to help out a person who was writing a paper on unix to
PC connectivity.
</para>

</sect1>

<sect1>
<title>Usernames</title>
<para>
The SMB protocol has only a loose username concept. Early SMB
protocols (such as CORE and COREPLUS) have no username concept at
all. Even in later protocols clients often attempt operations
(particularly printer operations) without first validating a username
on the server.
</para>

<para>
Unix security is based around username/password pairs. A unix box
should not allow clients to do any substantive operation without some
sort of validation. 
</para>

<para>
The problem mostly manifests itself when the unix server is in "share
level" security mode. This is the default mode as the alternative
"user level" security mode usually forces a client to connect to the
server as the same user for each connected share, which is
inconvenient in many sites.
</para>

<para>
In "share level" security the client normally gives a username in the
"session setup" protocol, but does not supply an accompanying
password. The client then connects to resources using the "tree
connect" protocol, and supplies a password. The problem is that the
user on the PC types the username and the password in different
contexts, unaware that they need to go together to give access to the
server. The username is normally the one the user typed in when they
"logged onto" the PC (this assumes Windows for Workgroups). The
password is the one they chose when connecting to the disk or printer.
</para>

<para>
The user often chooses a totally different username for their login as
for the drive connection. Often they also want to access different
drives as different usernames. The unix server needs some way of
divining the correct username to combine with each password.
</para>

<para>
Samba tries to avoid this problem using several methods. These succeed
in the vast majority of cases. The methods include username maps, the
service%user syntax, the saving of session setup usernames for later
validation and the derivation of the username from the service name
(either directly or via the user= option).
</para>

</sect1>

<sect1>
<title>File Ownership</title>

<para>
The commonly used SMB protocols have no way of saying "you can't do
that because you don't own the file". They have, in fact, no concept
of file ownership at all.
</para>

<para>
This brings up all sorts of interesting problems. For example, when
you copy a file to a unix drive, and the file is world writeable but
owned by another user the file will transfer correctly but will
receive the wrong date. This is because the utime() call under unix
only succeeds for the owner of the file, or root, even if the file is
world writeable. For security reasons Samba does all file operations
as the validated user, not root, so the utime() fails. This can stuff
up shared development diectories as programs like "make" will not get
file time comparisons right.
</para>

<para>
There are several possible solutions to this problem, including
username mapping, and forcing a specific username for particular
shares.
</para>

</sect1>

<sect1>
<title>Passwords</title>

<para>
Many SMB clients uppercase passwords before sending them. I have no
idea why they do this. Interestingly WfWg uppercases the password only
if the server is running a protocol greater than COREPLUS, so
obviously it isn't just the data entry routines that are to blame.
</para>

<para>
Unix passwords are case sensitive. So if users use mixed case
passwords they are in trouble.
</para>

<para>
Samba can try to cope with this by either using the "password level"
option which causes Samba to try the offered password with up to the
specified number of case changes, or by using the "password server"
option which allows Samba to do its validation via another machine
(typically a WinNT server).
</para>

<para>
Samba supports the password encryption method used by SMB
clients. Note that the use of password encryption in Microsoft
networking leads to password hashes that are "plain text equivalent".
This means that it is *VERY* important to ensure that the Samba
smbpasswd file containing these password hashes is only readable
by the root user. See the documentation ENCRYPTION.txt for more
details.
</para>

</sect1>

<sect1>
<title>Locking</title>
<para>
Since samba 2.2, samba supports other types of locking as well. This 
section is outdated.
</para>

<para>
The locking calls available under a DOS/Windows environment are much
richer than those available in unix. This means a unix server (like
Samba) choosing to use the standard fcntl() based unix locking calls
to implement SMB locking has to improvise a bit.
</para>

<para>
One major problem is that dos locks can be in a 32 bit (unsigned)
range. Unix locking calls are 32 bits, but are signed, giving only a 31
bit range. Unfortunately OLE2 clients use the top bit to select a
locking range used for OLE semaphores.
</para>

<para>
To work around this problem Samba compresses the 32 bit range into 31
bits by appropriate bit shifting. This seems to work but is not
ideal. In a future version a separate SMB lockd may be added to cope
with the problem.
</para>

<para>
It also doesn't help that many unix lockd daemons are very buggy and
crash at the slightest provocation. They normally go mostly unused in
a unix environment because few unix programs use byte range
locking. The stress of huge numbers of lock requests from dos/windows
clients can kill the daemon on some systems.
</para>

<para>
The second major problem is the "opportunistic locking" requested by
some clients. If a client requests opportunistic locking then it is
asking the server to notify it if anyone else tries to do something on
the same file, at which time the client will say if it is willing to
give up its lock. Unix has no simple way of implementing
opportunistic locking, and currently Samba has no support for it.
</para>

</sect1>

<sect1>
<title>Deny Modes</title>

<para>
When a SMB client opens a file it asks for a particular "deny mode" to
be placed on the file. These modes (DENY_NONE, DENY_READ, DENY_WRITE,
DENY_ALL, DENY_FCB and DENY_DOS) specify what actions should be
allowed by anyone else who tries to use the file at the same time. If
DENY_READ is placed on the file, for example, then any attempt to open
the file for reading should fail.
</para>

<para>
Unix has no equivalent notion. To implement this Samba uses either lock
files based on the files inode and placed in a separate lock
directory or a shared memory implementation. The lock file method 
is clumsy and consumes processing and file resources,
the shared memory implementation is vastly prefered and is turned on
by default for those systems that support it.
</para>

</sect1>

<sect1>
<title>Trapdoor UIDs</title>
<para>
A SMB session can run with several uids on the one socket. This
happens when a user connects to two shares with different
usernames. To cope with this the unix server needs to switch uids
within the one process. On some unixes (such as SCO) this is not
possible. This means that on those unixes the client is restricted to
a single uid.
</para>

<para>
Note that you can also get the "trapdoor uid" message for other
reasons. Please see the FAQ for details.
</para>

</sect1>

<sect1>
<title>Port numbers</title>
<para>
There is a convention that clients on sockets use high "unprivilaged"
port numbers (>1000) and connect to servers on low "privilaged" port
numbers. This is enforced in Unix as non-root users can't open a
socket for listening on port numbers less than 1000.
</para>

<para>
Most PC based SMB clients (such as WfWg and WinNT) don't follow this
convention completely. The main culprit is the netbios nameserving on
udp port 137. Name query requests come from a source port of 137. This
is a problem when you combine it with the common firewalling technique
of not allowing incoming packets on low port numbers. This means that
these clients can't query a netbios nameserver on the other side of a
low port based firewall.
</para>

<para>
The problem is more severe with netbios node status queries. I've
found that WfWg, Win95 and WinNT3.5 all respond to netbios node status
queries on port 137 no matter what the source port was in the
request. This works between machines that are both using port 137, but
it means it's not possible for a unix user to do a node status request
to any of these OSes unless they are running as root. The answer comes
back, but it goes to port 137 which the unix user can't listen
on. Interestingly WinNT3.1 got this right - it sends node status
responses back to the source port in the request.
</para>

</sect1>

<sect1>
<title>Protocol Complexity</title>
<para>
There are many "protocol levels" in the SMB protocol. It seems that
each time new functionality was added to a Microsoft operating system,
they added the equivalent functions in a new protocol level of the SMB
protocol to "externalise" the new capabilities.
</para>

<para>
This means the protocol is very "rich", offering many ways of doing
each file operation. This means SMB servers need to be complex and
large. It also means it is very difficult to make them bug free. It is
not just Samba that suffers from this problem, other servers such as
WinNT don't support every variation of every call and it has almost
certainly been a headache for MS developers to support the myriad of
SMB calls that are available.
</para>

<para>
There are about 65 "top level" operations in the SMB protocol (things
like SMBread and SMBwrite). Some of these include hundreds of
sub-functions (SMBtrans has at least 120 sub-functions, like
DosPrintQAdd and NetSessionEnum). All of them take several options
that can change the way they work. Many take dozens of possible
"information levels" that change the structures that need to be
returned. Samba supports all but 2 of the "top level" functions. It
supports only 8 (so far) of the SMBtrans sub-functions. Even NT
doesn't support them all.
</para>

<para>
Samba currently supports up to the "NT LM 0.12" protocol, which is the
one preferred by Win95 and WinNT3.5. Luckily this protocol level has a
"capabilities" field which specifies which super-duper new-fangled
options the server suports. This helps to make the implementation of
this protocol level much easier.
</para>

<para>
There is also a problem with the SMB specications. SMB is a X/Open
spec, but the X/Open book is far from ideal, and fails to cover many
important issues, leaving much to the imagination. Microsoft recently
renamed the SMB protocol CIFS (Common Internet File System) and have 
published new specifications. These are far superior to the old 
X/Open documents but there are still undocumented calls and features. 
This specification is actively being worked on by a CIFS developers 
mailing list hosted by Microsft.
</para>
</sect1>
</chapter>