|
|
This document gives a general overview of how Samba works internally. The Samba Team has tried to come up with a model which is the best possible compromise between elegance, portability, security and the constraints imposed by the very messy SMB and CIFS protocol.
It also tries to answer some of the frequently asked questions such as:
Is Samba secure when running on Unix? The xyz platform? What about the root priveliges issue?
Pros and cons of multithreading in various parts of Samba
Why not have a separate process for name resolution, WINS, and browsing?
People sometimes tout threads as a uniformly good thing. They are very nice in their place but are quite inappropriate for smbd. nmbd is another matter, and multi-threading it would be very nice.
The short version is that smbd is not multithreaded, and alternative servers that take this approach under Unix (such as Syntax, at the time of writing) suffer tremendous performance penalties and are less robust. nmbd is not threaded either, but this is because it is not possible to do it while keeping code consistent and portable across 35 or more platforms. (This drawback also applies to threading smbd.)
The longer versions is that there are very good reasons for not making smbd multi-threaded. Multi-threading would actually make Samba much slower, less scalable, less portable and much less robust. The fact that we use a separate process for each connection is one of Samba's biggest advantages.
A few problems that would arise from a threaded smbd are:
It's not only to create threads instead of processes, but you must care about all variables if they have to be thread specific (currently they would be global).
if one thread dies (eg. a seg fault) then all threads die. We can immediately throw robustness out the window.
many of the system calls we make are blocking. Non-blocking equivalents of many calls are either not available or are awkward (and slow) to use. So while we block in one thread all clients are waiting. Imagine if one share is a slow NFS filesystem and the others are fast, we will end up slowing all clients to the speed of NFS.
you can't run as a different uid in different threads. This means we would have to switch uid/gid on _every_ SMB packet. It would be horrendously slow.
the per process file descriptor limit would mean that we could only support a limited number of clients.
we couldn't use the system locking calls as the locking context of fcntl() is a process, not a thread.
This would be ideal, but gets sunk by portability requirements.
Andrew tried to write a test threads library for nmbd that used only ansi-C constructs (using setjmp and longjmp). Unfortunately some OSes defeat this by restricting longjmp to calling addresses that are shallower than the current address on the stack (apparently AIX does this). This makes a truly portable threads library impossible. So to support all our current platforms we would have to code nmbd both with and without threads, and as the real aim of threads is to make the code clearer we would not have gained anything. (it is a myth that threads make things faster. threading is like recursion, it can make things clear but the same thing can always be done faster by some other method)
Chris tried to spec out a general design that would abstract threading vs separate processes (vs other methods?) and make them accessible through some general API. This doesn't work because of the data sharing requirements of the protocol (packets in the future depending on packets now, etc.) At least, the code would work but would be very clumsy, and besides the fork() type model would never work on Unix. (Is there an OS that it would work on, for nmbd?)
A fork() is cheap, but not nearly cheap enough to do on every UDP packet that arrives. Having a pool of processes is possible but is nasty to program cleanly due to the enormous amount of shared data (in complex structures) between the processes. We can't rely on each platform having a shared memory system.
Originally Andrew used recursion to simulate a multi-threaded environment, which use the stack enormously and made for really confusing debugging sessions. Luke Leighton rewrote it to use a queuing system that keeps state information on each packet. The first version used a single structure which was used by all the pending states. As the initialisation of this structure was done by adding arguments, as the functionality developed, it got pretty messy. So, it was replaced with a higher-order function and a pointer to a user-defined memory block. This suddenly made things much simpler: large numbers of functions could be made static, and modularised. This is the same principle as used in NT's kernel, and achieves the same effect as threads, but in a single process.
Then Jeremy rewrote nmbd. The packet data in nmbd isn't what's on the wire. It's a nice format that is very amenable to processing but still keeps the idea of a distinct packet. See "struct packet_struct" in nameserv.h. It has all the detail but none of the on-the-wire mess. This makes it ideal for using in disk or memory-based databases for browsing and WINS support.