PORTABILITY

	Portability guideline
	KAME project
	$KAME: PORTABILITY,v 1.58 2001/02/01 09:13:48 jinmei Exp $


Source code in the "kame" directory will be shared among operating systems
and hardware platforms.  Here are several guidelines to be portable
across those possibilities.

- Developers MUST compile the code on two or more platforms, before committing
  it to the repository (trivial changes or comment typos are okay).

- Do not break paren-match.
  Always try to keep paren-match correctly, regardless of #ifdef.
  The following is WRONG example:
	#if foo
		if (foo) {
	#else
		if (foo && bar) {
	#endif
			....
		}

		baz(this, that,
	#if foo
			something);
	#else
			something, anything);
	#endif
  It should be written like follows:
	#if foo
		if (foo)
	#else
		if (foo && bar)
	#endif
		{
			....
		}

	#ifdef foo
		baz(this, that, something);
	#else
		baz(this, that, something, anything);
	#endif

		/* latter part can also be like follows... */
		baz(this, that,
			something
	#ifndef foo
			, anything
	#endif
			);

  It is also advised to add extra comment to help paren match,
  if you need to write widow paren in string.  For example:
	printf("(");		/* ) */
  (note that there are opinions against it too - it depends)

- Avoid use of "#if" as much as possible, stick to "#ifdef" or "#ifndef"
  if possible (Avoid "#if defined(foo)" if you only have single "foo").
  "#if" is relatively new syntax and is not friendly with old tools like
  unifdef(1).

- Follow NetBSD KNF.
  They are the most strict guys about source code formatting.
  If you obey NetBSD KNF, other platforms are happy.
  More information on KNF is available at:
	http://www.netbsd.org/Documentation/kernel/programming.html

- Pointer can be 64bit (for example, on alpha).
  When taking integer from pointer, use u_long, or long.
  This is WRONG:
	void *p;
	/* alignment check */
	if ((1 & (int)p) == 0) {
		...
	}
  This is correct (but actually may not be future-proven):
	void *p;
	/* alignment check */
	if ((1 & (u_long)p) == 0) {
		...
	}
  
- size_t can be 64bit, and may not be of same size as int.
  When doing printf(), cast size_t to u_long and print it as %lu.
	size_t siz;
	printf("%lu", (u_long)siz);
  Use of "%p" is more appropriate than casting, for pointers.

- If you would like to use printf() against variable typed u_quad_t,
  cast them to unsigned long long and use %llu.
	u_quad_t x;
	printf("%llu", (unsigned long long)x);
  %qu is outside of standard, and %llu is in C99 standard.

  Also note that FreeBSD prior to 3.2 cannot handle %llu correctly - you need
  to use %qu on these platforms (3.2 and later are okay).
  So you end up doing #ifdef __FreeBSD__.

  KAME libinet6 includes vfprintf() with %llu support where necessary,
  to ease code sharing in KAME userland code.

- tv.tv_sec is not typed as time_t.
  If you want time_t, explicitly copy the value.

- Packed structs must be used with care.
  Unaligned structure can cause problem on certain architectures.
  You'll need to copy the return value to aligned structure before accessing.

  Structure returned by SIOCGIFCONF falls into this category.

- When you define wire format as C struct, (1) avoid bitfield.  (2) use
  __attribute__((__packed__)) to control member and overall alignment.
  NOTE: (2) is not portable across platforms - it is a GCC-ism.

- 2nd arg to ioctl() must be u_long, not int, on non-FreeBSD2 platforms.

- If you define multi-statement #define, define that like:
	#define foobar (x, y) \
	do { \
		/*something*/ \
		/*something*/ \
	} while (0)
  Without this wrapper nasty mistakes can happen.

- Compiler options need to be configured differently into Makefile.
  For NetBSD and OpenBSD, CFLAGS and CPPFLAGS are defined separately:
	CPPFLAGS+=-I/usr/local/include
	CFLAGS+=-g
  CPPFLAGS contains -I and -D only (items that needs to be passed to cpp).
  CFLAGS can contain other flags.
  For FreeBSD and BSDI, there's no distinction.
	CFLAGS+=-I/usr/local/include
	CFLAGS+=-g

- To refer objects and machine-generated source code, FreeBSD needs to use
  ${.OBJDIR}.  This is because because FreeBSD completely separates "obj" tree
  and "src" tree and has no symlink from ${.CURDIR}/obj to /usr/obj/foobar.

  On NetBSD, "make print-objdir" is the right way to probe "obj" directory.

  On other platforms, ${.CURDIR}/obj should do in most cases.

- On manpage installation and shlib generation in Makefile.
  for BSDI3, you need to set MAN[0-9] in Makefile like below.  MANDIR must
  point to "catN" directory, not "manN", without "N":
	MANDIR=	/usr/local/v6/man/cat
	MAN5=	foo.0
	MAN8=	baz.0
	MLINKS=	foo.5 bar.5
  BSDI4 is much like the same as BSDI3.  The only exception is shlib
  generation.  NODYNAMIC=yes would prevent dynamic library from created:
	MANDIR=	/usr/local/v6/man/cat
	MAN5=	foo.0
	MAN8=	baz.0
	MLINKS=	foo.5 bar.5
	NODYNAMIC=yes	# if you do not want shlib, define it
  For NetBSD, you need to set MAN (not MAN[0-9]).  MANDIR needs to point to
  "man" directory right above "catN" or "manN" directory:
	MANDIR=	/usr/local/v6/man
	MAN=	foo.5 baz.8
	MLINKS=	foo.5 bar.5
	MKPIC=	# if you do not want shlib, make it empty
  For OpenBSD, you need to set MAN (not MAN[0-9]).  MANDIR needs to point to
  "catN" directory, without "N":
	MANDIR=	/usr/local/v6/man/cat
	MAN=	foo.5 baz.8
	MLINKS=	foo.5 bar.5
	NOPIC=	yes	# if you do not want shlib, define it
  For FreeBSD, you need to set MAN[0-9] like below.  MANDIR needs to point
  to "manN" directory, without "N":
	MANDIR=	/usr/local/v6/man/man
	MAN5=	foo.5
	MAN8=	baz.8
	MLINKS=	foo.5 bar.5
	# if you do not want shlib, do not define SHLIB_{MAJOR,MINOR}

- If you use yacc/lex, be aware that FreeBSD bsd.prog.mk behaves very
  strange with yacc's -o option (non-trivial .if statement is placed).
  You may need to add the following into Makefile for code sharing.
  The following fragment will force the build to generate y.tab.h, instead
  of foo.h.
	SRCS+=y.tab.h
	y.tab.h: foo.y
  Be sure to include y.tab.h, not foo.h, from *.l.

  On NetBSD, YHEADER is quite useful.

- BSDI3 /usr/bin/cc is gcc 1.42, and it does not support __FUNCTION__.
  If you need to use __FUNCTION__, you need to set CC to "gcc", or "shlicc2"
  explicitly.
  __FUNCTION__ macro is not included in the ANSI standard, so it is best to
  avoid __FUNCTION__, or make it optional, for portability.

- sys/queue.h are very different in each of *BSDs.  Most of macros found in
  FreeBSD3 are not available in other systems.  Simply avoid those, make sure
  your code compile on non-FreeBSD3 systems.
  Avoid xx_FOREACH(), which is only available in FreeBSD3.  Just use
	for (x = xx_FIRST(); x; x = xx_NEXT(x))
  and do not try to make #if section for FreeBSD3.  More #if introduces
  more bugs.

- It is better to avoid insque() and rmque(), they are from old VAX days
  and some platform now tries to avoid these.

- To identify FreeBSD3, use the following:
	#if defined(__FreeBSD__) && __FreeBSD__ >= 3
		this is freebsd3.
	#endif

- To identify BSDI4, use the following:
	#if defined(__bsdi__) && _BSDI_VERSION >= 199802
		this is bsdi4.
	#endif

- RFC2553 defines sockaddr_storage to have __ss_{family,len} member.  During
  the discusson for the draft on ipngwg, ss_len was proposed and final result
  was not very clear.  X/Open document mentions ss_len, not __ss_len.
  Therefore, you can see both definition on operating system platforms.
  The most portable way of dealing with it is to:
  (1) have -Dss_len=__ss_len to unify all occurences (including header file)
  into __ss_len, or (2) never touch __ss_len.  cast to sockaddr * and use
  sa_len.

  Next version of RFC2553 (RFC2553bis) will define ss_len (not __ss_len),
  and be synchronized better with X/Open document.

- Never, never touch the __u6_addr{,8,16,32} member in struct in6_addr
  directly.  Symbols starting with "__" are NOT supposed to be used.

- RFC2553 defines only s6_addr[16] (array of u_int8_t) as member of struct
  in6_addr.  KAME strictly follows RFC2553, and does not provide any
  extensions like s6_addr{8,16,32} to the userland.  For maximum portability
  across IPv6 platforms, only s6_addr should be used.
  In the future, we may provide s6_addr{8,16,32} to the userland when updates
  to RFC2553 defines them.  They are, in fact, very convenient and we would
  like to see them defined in standard document.

- sysctl(3) support in kernel.
  FreeBSD: linker hack #define in sys/sysctl.h, like SYSCTL_INT().
	no need to recompile sysctl(8).
	SYSCTL_INT(_net_inet6_tcp6, TCP6CTL_SYN_CACHE_INTER,
	    syn_cache_interval, CTLFLAG_RW, &tcp6_syn_cache_interval, 0, "");
	Note that sys/kernel.h should be included as well, which SYSCTL_XXX
	macros depend.
  OpenBSD/NetBSD: xx_sysctl() needs to be supplied from inet6sw.
	xx_sysctl() should implement switch statement to dispatch to all
	possible leaf nodes available.  symbols must be defined in header file
	for use from sysctl(8).
  BSDI: similar to OpenBSD/NetBSD, but sysctl_int_arr() makes xx_sysctl()
	code much easier for simple integer leaf nodes.

- Under BSD make tree (bsd.prog.mk), you can't use relative path against
  include files, like the following example:
	#include "../foo/foo.h"		/*THIS DOES NOT WORK*/
  If you are in baa/Makefile, and if you do "make obj", you will be in
  baa/obj when you build the tree.  ../foo/foo.h will not be reachable
  (it will be interpreted as baa/obj/../foo/foo.h = baa/foo/foo.h)

  Solution: use -I${.CURDIR}/../foo.

- OpenBSD does not allow duplicated inclusion of system header files.
  (it does not have #ifndef _FOO_H_ wrapper)

- Developers are advised to use pedantic compilation options (like -Wall
  -Werror).  To configure default make settings, /etc/make.conf (freebsd),
  /etc/mk.conf (netbsd/openbsd), and/or /usr/share/mk/sys.mk (bsdi) should be
  useful. Note, however, that the default compiler of bsdi3 (/usr/ucb/cc)
  does not suppport the -Werror option.

  Note that the meaning of -Wall is different across *BSD because *BSD use
  different version of gnu cc.  Sometimes -Wall raises bogus warning due to
  bugs in code generator.  Also, since function prototypes are defined
  differently in some cases (like ntohl), certain tweaks are needed to
  eliminate warnings on all the platforms.  Don't worry about it too much,
  you can always talk with other developers.

- For portability in kernel networking code, look at kame/sys/net/net_osdep.h.

- The -interface modifier of the route(8) command does not take an
  argument except on FreeBSD. Also, the original FreeBSD's extension
  has several bugs (although those bugs are fixed in the route command
  in the KAME kit). If you regard portability and stability, it would
  be better to use a mixture of the "change" operation and the -ifp
  modifier of the route command.

  For example,
  # route add -inet6 default -interface gif0 (on FreeBSD)
  is equivalent to
  # route add -inet6 default ::1
  # route change -inet6 default ::1 -ifp gif0

  Note that the kernel code to perform the change operation assumes
  that gif0 has at least one IPv6 address. If you are not sure about
  this, you might want to execute
  # ifconfig gif0 up
  before executing the route command.

  BSDI4 users can take the following approach as well:
  # route add -inet6 default ::1 -ifp gif0
  which would be much simpler. This is because BSDI4 has improved
  kernel code of handling routing sockets. We will eventually merge
  this improvement into other OSes.

- (Not really a portability issue)  Fixes/improvements to libpcap and tcpdump
  should be sent to itojun, or patches@tcpdump.org.  We no longer maintain
  libpcap/tcpdump under "kame" directory.  http://www.tcpdump.org/

- freebsd and bsdi3 do not have satosin() in sys/netinet/in.h.  not sure
  where is the origin for this macro.  do not use it, or define it locally
  in C source if you don't have one (#ifndef - #define).

- INADDR_LOOPBACK is declared only in freebsd[34], netbsd and openbsd.

- Location of MD5 library.
  netbsd/openbsd: header is <md5.h>, libc.
  freebsd: header is <md5.h>, libcrypt

- mandoc formatting for SYNOPSIS section.
  On netbsd/openbsd, people usually use
	.Nm ""
  to indicate new line in SYNOPSIS section ("" makes proper hanging indent,
  while .Nm does not).  However, bsdi mandoc chokes with it.  Explicitly write
	.Nm foo
  to avoid portability issue (it is unfortunate that we need to place
  duplicated text...).

- SADB_X_SA2 (in /usr/include/net/pfkeyv2.h) has completely different meaning
  between KAME-based systems and openbsd systems.  take caution if you plan to
  use it.

- getopt(3) returns -1 on openbsd, and EOF on other platforms.  This was due
  to POSIX changes.  Since EOF is usually defined to be -1, it is not a major
  issue.

- SECURITY: Never use printf-like functions (including syslog, setproctitle
  and others) with only a "fmt" argument.  Always use "%s".  Otherwise,
  your program will be compromised by malicious party.
	BAD: syslog(foo);
	GOOD: syslog("%s", foo);

- SECURITY: be careful using fd_set, specifically when you are in setuid'ed
  applications.  there's buffer overrun possibility with FD_SET().
  use poll(2) whenever available (note that KAME still support platforms
  without poll(2)).  see openbsd select(2) manpage.

- strlcpy/strlcat.
	comes with strlcpy/cat: netbsd15 openbsd freebsd3[3-5] freebsd4
	does not come with them: freebsd2 freebsd3[0-2] bsdi[34] netbsd142
		(libinet6 includes it to ease code sharing)

  strlcpy and strlcat -- consistent, safe, string copy and concatenation.
  Usenix 1999, by Todd C. Miller, Theo de Raadt.
	http://www.openbsd.org/papers/strlcpy-paper.ps
	http://www.openbsd.org/papers/strlcpy-slides.ps

- strptime.
	freebsd4 and netbsd have it in libc.
	bsdi4 does not have it.
	others?