Remind me to implement bb_fork_exec()...

[oweals/busybox.git] / docs / busybox.net / programming.html

<!--#include file="header.html" -->

<h2>Rob's notes on programming busybox.</h2>

<ul>
  <li><a href="#goals">What are the goals of busybox?</a></li>
  <li><a href="#design">What is the design of busybox?</a></li>
  <li><a href="#source">How is the source code organized?</a></li>
  <ul>
    <li><a href="#source_applets">The applet directories.</a></li>
    <li><a href="#source_libbb">The busybox shared library (libbb)</a></li>
  </ul>
  <li><a href="#adding">Adding an applet to busybox</a></li>
  <li><a href="#standards">What standards does busybox adhere to?</a></li>
  <li><a href="#tips">Tips and tricks.</a></li>
  <ul>
    <li><a href="#tips_encrypted_passwords">Encrypted Passwords</a></li>
    <li><a href="#tips_vfork">Fork and vfork</a></li>
  </ul>
</ul>

<h2><b><a name="goals" />What are the goals of busybox?</b></h2>

<p>Busybox aims to be the smallest and simplest correct implementation of the
standard Linux command line tools.  First and foremost, this means the
smallest executable size we can manage.  We also want to have the simplest
and cleanest implementation we can manage, be <a href="#standards">standards
compliant</a>, minimize run-time memory usage (heap and stack), run fast, and
take over the world.</p>

<h2><b><a name="design" />What is the design of busybox?</b></h2>

<p>Busybox is like a swiss army knife: one thing with many functions.
The busybox executable can act like many different programs depending on
the name used to invoke it.  Normal practice is to create a bunch of symlinks
pointing to the busybox binary, each of which triggers a different busybox
function.  (See <a href="FAQ.html#getting_started">getting started</a> in the
FAQ for more information on usage, and <a href="BusyBox.html">the
busybox documentation</a> for a list of symlink names and what they do.)

<p>The "one binary to rule them all" approach is primarily for size reasons: a
single multi-purpose executable is smaller then many small files could be.
This way busybox only has one set of ELF headers, it can easily share code
between different apps even when statically linked, it has better packing
efficiency by avoding gaps between files or compression dictionary resets,
and so on.</p>

<p>Work is underway on new options such as "make standalone" to build separate
binaries for each applet, and a "libbb.so" to make the busybox common code
available as a shared library.  Neither is ready yet at the time of this
writing.</p>

<a name="source" />

<h2><a name="source_applets" /><b>The applet directories</b></h2>

<p>The directory "applets" contains the busybox startup code (applets.c and
busybox.c), and several subdirectories containing the code for the individual
applets.</p>

<p>Busybox execution starts with the main() function in applets/busybox.c,
which sets the global variable bb_applet_name to argv[0] and calls
run_applet_by_name() in applets/applets.c.  That uses the applets[] array
(defined in include/busybox.h and filled out in include/applets.h) to
transfer control to the appropriate APPLET_main() function (such as
cat_main() or sed_main()).  The individual applet takes it from there.</p>

<p>This is why calling busybox under a different name triggers different
functionality: main() looks up argv[0] in applets[] to get a function pointer
to APPLET_main().</p>

<p>Busybox applets may also be invoked through the multiplexor applet
"busybox" (see busybox_main() in applets/busybox.c), and through the
standalone shell (grep for STANDALONE_SHELL in applets/shell/*.c).
See <a href="FAQ.html#getting_started">getting started</a> in the
FAQ for more information on these alternate usage mechanisms, which are
just different ways to reach the relevant APPLET_main() function.</p>

<p>The applet subdirectories (archival, console-tools, coreutils,
debianutils, e2fsprogs, editors, findutils, init, loginutils, miscutils,
modutils, networking, procps, shell, sysklogd, and util-linux) correspond
to the configuration sub-menus in menuconfig.  Each subdirectory contains the
code to implement the applets in that sub-menu, as well as a Config.in
file defining that configuration sub-menu (with dependencies and help text
for each applet), and the makefile segment (Makefile.in) for that
subdirectory.</p>

<p>The run-time --help is stored in usage_messages[], which is initialized at
the start of applets/applets.c and gets its help text from usage.h.  During the
build this help text is also used to generate the BusyBox documentation (in
html, txt, and man page formats) in the docs directory.  See
<a href="#adding">adding an applet to busybox</a> for more
information.</p>

<h2><a name="source_libbb" /><b>libbb</b></h2>

<p>Most non-setup code shared between busybox applets lives in the libbb
directory.  It's a mess that evolved over the years without much auditing
or cleanup.  For anybody looking for a great project to break into busybox
development with, documenting libbb would be both incredibly useful and good
experience.</p>

<p>Common themes in libbb include allocation functions that test
for failure and abort the program with an error message so the caller doesn't
have to test the return value (xmalloc(), xstrdup(), etc), wrapped versions
of open(), close(), read(), and write() that test for their own failures
and/or retry automatically, linked list management functions (llist.c),
command line argument parsing (getopt_ulflags.c), and a whole lot more.</p>

<h2><a name="adding" /><b>Adding an applet to busybox</b></h2>

<p>To add a new applet to busybox, first pick a name for the applet and
a corresponding CONFIG_NAME.  Then do this:</p>

<ul>
<li>Figure out where in the busybox source tree your applet best fits,
and put your source code there.  Be sure to use APPLET_main() instead
of main(), where APPLET is the name of your applet.</li>

<li>Add your applet to the relevant Config.in file (which file you add
it to determines where it shows up in "make menuconfig").  This uses
the same general format as the linux kernel's configuration system.</li>

<li>Add your applet to the relevant Makefile.in file (in the same
directory as the Config.in you chose), using the existing entries as a
template and the same CONFIG symbol as you used for Config.in.  (Don't
forget "needlibm" or "needcrypt" if your applet needs libm or
libcrypt.)</li>

<li>Add your applet to "include/applets.h", using one of the existing
entries as a template.  (Note: this is in alphabetical order.  Applets
are found via binary search, and if you add an applet out of order it
won't work.)</li>

<li>Add your applet's runtime help text to "include/usage.h".  You need
at least appname_trivial_usage (the minimal help text, always included
in the busybox binary when this applet is enabled) and appname_full_usage
(extra help text included in the busybox binary with
CONFIG_FEATURE_VERBOSE_USAGE is enabled), or it won't compile.
The other two help entry types (appname_example_usage and
appname_notes_usage) are optional.  They don't take up space in the binary,
but instead show up in the generated documentation (BusyBox.html,
BusyBox.txt, and the man page BusyBox.1).</li>

<li>Run menuconfig, switch your applet on, compile, test, and fix the
bugs.  Be sure to try both "allyesconfig" and "allnoconfig" (and
"allbareconfig" if relevant).</li>

</ul>

<h2><a name="standards" />What standards does busybox adhere to?</a></h2>

<p>The standard we're paying attention to is the "Shell and Utilities"
portion of the <a href=http://www.opengroup.org/onlinepubs/009695399/>Open
Group Base Standards</a> (also known as the Single Unix Specification version
3 or SUSv3).  Note that paying attention isn't necessarily the same thing as
following it.</p>

<p>SUSv3 doesn't even mention things like init, mount, tar, or losetup, nor
commonly used options like echo's '-e' and '-n', or sed's '-i'.  Busybox is
driven by what real users actually need, not the fact the standard believes
we should implement ed or sccs.  For size reasons, we're unlikely to include
much internationalization support beyond UTF-8, and on top of all that, our
configuration menu lets developers chop out features to produce smaller but
very non-standard utilities.</p>

<p>Also, Busybox is aimed primarily at Linux.  Unix standards are interesting
because Linux tries to adhere to them, but portability to dozens of platforms
is only interesting in terms of offering a restricted feature set that works
everywhere, not growing dozens of platform-specific extensions.  Busybox
should be portable to all hardware platforms Linux supports, and any other
similar operating systems that are easy to do and won't require much
maintenance.</p>

<p>In practice, standards compliance tends to be a clean-up step once an
applet is otherwise finished.  When polishing and testing a busybox applet,
we ensure we have at least the option of full standards compliance, or else
document where we (intentionally) fall short.</p>

<h2><a name="tips" />Programming tips and tricks.</a></h2>

<p>Various things busybox uses that aren't particularly well documented
elsewhere.</p>

<h2><a name="tips_encrypted_passwords">Encrypted Passwords</a></h2>

<p>Password fields in /etc/passwd and /etc/shadow are in a special format.
If the first character isn't '$', then it's an old DES style password.  If
the first character is '$' then the password is actually three fields
separated by '$' characters:</p>
<pre>
  <b>$type$salt$encrypted_password</b>
</pre>

<p>The "type" indicates which encryption algorithm to use: 1 for MD5 and 2 for SHA1.</p>

<p>The "salt" is a bunch of ramdom characters (generally 8) the encryption
algorithm uses to perturb the password in a known and reproducible way (such
as by appending the random data to the unencrypted password, or combining
them with exclusive or).  Salt is randomly generated when setting a password,
and then the same salt value is re-used when checking the password.  (Salt is
thus stored unencrypted.)</p>

<p>The advantage of using salt is that the same cleartext password encrypted
with a different salt value produces a different encrypted value.
If each encrypted password uses a different salt value, an attacker is forced
to do the cryptographic math all over again for each password they want to
check.  Without salt, they could simply produce a big dictionary of commonly
used passwords ahead of time, and look up each password in a stolen password
file to see if it's a known value.  (Even if there are billions of possible
passwords in the dictionary, checking each one is just a binary search against
a file only a few gigabytes long.)  With salt they can't even tell if two
different users share the same password without guessing what that password
is and decrypting it.  They also can't precompute the attack dictionary for
a specific password until they know what the salt value is.</p>

<p>The third field is the encrypted password (plus the salt).  For md5 this
is 22 bytes.</p>

<p>The busybox function to handle all this is pw_encrypt(clear, salt) in
"libbb/pw_encrypt.c".  The first argument is the clear text password to be
encrypted, and the second is a string in "$type$salt$password" format, from
which the "type" and "salt" fields will be extracted to produce an encrypted
value.  (Only the first two fields are needed, the third $ is equivalent to
the end of the string.)  The return value is an encrypted password in
/etc/passwd format, with all three $ separated fields.  It's stored in
a static buffer, 128 bytes long.</p>

<p>So when checking an existing password, if pw_encrypt(text,
old_encrypted_password) returns a string that compares identical to
old_encrypted_password, you've got the right password.  When setting a new
password, generate a random 8 character salt string, put it in the right
format with sprintf(buffer, "$%c$%s", type, salt), and feed buffer as the
second argument to pw_encrypt(text,buffer).</p>

<h2><a name="tips_vfork">Fork and vfork</a></h2>

<p>Busybox hides the difference between fork() and vfork() in
libbb/bb_fork_exec.c.  If you ever want to fork and exec, use bb_fork_exec()
(which returns a pid and takes the same arguments as execve(), although in
this case envp can be NULL) and don't worry about it.  This description is
here in case you want to know why that does what it does.</p>

<p>On systems that haven't got a Memory Management Unit, fork() is unreasonably
expensive to implement, so a less capable function called vfork() is used
instead.</p>

<p>The reason vfork() exists is that if you haven't got an MMU then you can't
simply set up a second set of page tables and share the physical memory via
copy-on-write, which is what fork() normally does.  This means that actually
forking has to copy all the parent's memory (which could easily be tens of
megabytes).  And you have to do this even though that memory gets freed again
as soon as the exec happens, so it's probably all a big waste of time.</p>

<p>This is not only slow and a waste of space, it also causes totally
unnecessary memory usage spikes based on how big the _parent_ process is (not
the child), and these spikes are quite likely to trigger an out of memory
condition on small systems (which is where nommu is common anyway).  So
although you _can_ emulate a real fork on a nommu system, you really don't
want to.</p>

<p>In theory, vfork() is just a fork() that writeably shares the heap and stack
rather than copying it (so what one process writes the other one sees).  In
practice, vfork() has to suspend the parent process until the child does exec,
at which point the parent wakes up and resumes by returning from the call to
vfork().  All modern kernel/libc combinations implement vfork() to put the
parent to sleep until the child does its exec.  There's just no other way to
make it work: they're sharing the same stack, so if either one returns from its
function it stomps on the callstack so that when the other process returns,
hilarity ensues.  In fact without suspending the parent there's no way to even
store separate copies of the return value (the pid) from the vfork() call
itself: both assignments write into the same memory location.</p>

<p>One way to understand (and in fact implement) vfork() is this: imagine
the parent does a setjmp and then continues on (pretending to be the child)
until the exec() comes around, then the _exec_ does the actual fork, and the
parent does a longjmp back to the original vfork call and continues on from
there.  (It thus becomes obvious why the child can't return, or modify
local variables it doesn't want the parent to see changed when it resumes.)

<p>Note a common mistake: the need for vfork doesn't mean you can't have two
processes running at the same time.  It means you can't have two processes
sharing the same memory without stomping all over each other.  As soon as
the child calls exec(), the parent resumes.</p>

<p>If the child's attempt to call exec() fails, the child should call _exit()
rather than a normal exit().  This avoids any atexit() code that might confuse
the parent.  (The parent should never call _exit(), only a vforked child that
failed to exec.)</p>

<p>(Now in theory, a nommu system could just copy the _stack_ when it forks
(which presumably is much shorter than the heap), and leave the heap shared.
In practice, you've just wound up in a multi-threaded situation and you can't
do a malloc() or free() on your heap without freeing the other process's memory
(and if you don't have the proper locking for being threaded, corrupting the
heap if both of you try to do it at the same time and wind up stomping on
each other while traversing the free memory lists).  The thing about vfork is
that it's a big red flag warning "there be dragons here" rather than
something subtle and thus even more dangerous.)</p>

<br>
<br>
<br>

<!--#include file="footer.html" -->