Meeting Pearls 1

home *** CD-ROM | disk | FTP | other *** search

/ Meeting Pearls 1 / Meeting Pearls Vol 1 (1994).iso / installed_progs / text / faqs / solaris2.porting-faq < prev next >

Wrap

Internet Message Format | 1994-04-26 | 46.0 KB

Subject: Solaris 2 Porting FAQ Newsgroups: comp.unix.solaris,comp.answers,news.answers From: meyer@frostbite-falls.uoregon.edu (David M. Meyer 503/346-1747) Date: 25 Apr 1994 12:40:08 GMT Archive-name: Solaris2/porting-FAQ Last-modified: Monday, April 25, 1994 Version: 2.12 Solaris 2 Porting FAQ [Last changed: 25 April 1994] This article contains the answers to some Frequently Asked Questions (FAQ) often seen in comp.unix.solaris that relate to porting BSD/Solaris 1 applications to Solaris 2. Over the first few days of its existence, it has evolved into a more general discussion about portability among Unix systems, especially as it relates to BSD, ANSI, POSIX, and SVID compliant systems. It is hoped that this document will help reduce volume in this newsgroup and to provide hard-to-find information of general interest. Please redistribute this article! This FAQ is maintained by David Meyer (meyer@ns.uoregon.edu). Send updates and corrections to me at this address. It would help if the subject line contained the phrase "FAQ". This article includes answers to the following questions. Ones marked with a + indicate questions new to this issue; those with changes of content since the last issue are marked by *: 0) Which preprocessor symbols to use? 1) Some Include File Issues 2) Libraries 3)* Possible ANSI/POSIX/SVR4 replacements for some popular BSD functions 4) Signal Primer 5) Waiting for Children to Exit 6) Dealing with Shadow Password Files 7) Some Compatabilty Problems 8) Other Resources ----------------------------------------------------------------------------- 0) TOPIC: Which preprocessor symbols to use? [Last modified: 11 October 93] [Editor's Note: This section began life as a Solaris 1 and Solaris 2 centric discussion. However, it has grown into a more generalized portability discussion. I believe that this is a useful discussion, but it appears that contrasting styles, preferences, and requirements will make consensus difficult. DM] Answer: This is a difficult and controversial question. In order to understand the following discussion, we need to be aware of the following standards: ANSI C (ANSI X3J11) This is the standard C definition, originally adopted as American National Standard X3.159-1989 and has since been adopted as international standard ISO/IEC 9899:1990. POSIX.1 (IEEE 1003.1-1990) POSIX.1, the Portable Operating System Interface for Computer Environments, is a system level API that deals with the function and format of system calls and utilities such as signal handling. SVID3 SVID3, the System V Interface Definition Issue 3, is is fully compliant with POSIX.1, and is a arguably subset of the SVR4 system API. For example, SVID3 doesn't have "-ldl", but many people consider it of the SVR4 API. That is, a system could be SVID3-compliant without necessarily being an SVR4 system. XPG XPG, X/Open Company Ltd's X/Open Portability Guide, is a broad document which covers a great number of areas, including operating systems and programming languages, system interfaces, and internetworking. The latest version, XPG4, groups these components into "profiles", which are packaged together according to market needs. Two additional standards are relevant for Suns: SCD 2.0 and x86 ABIs SCD 2.0 is the SPARC Compliance Definition 2.0. The SCD has two components: On the hardware side, (i). System Compliance Test verifies that the hardware and operating system successfully emulates what Sun is doing. It covers low level system issues such as alignment, and linking and loading. (ii). The SPARC Application Verifier tests software to be sure that it runs on SCD hardware. As an example of subtle differences that exist between the BSD interface and SVID/POSIX standards, consider the BSD mktemp(3) call. The SunOS 4.1 mktemp() replaced the trailing X characters with the letter (e.g., X) and the current process ID. The SVID and SVR4 versions specify only that the six trailing Xs be replaced with a character string that can be used to create a unique filename, and does not depend on the specific name of the file. Thus, the BSD and SVR4/SVID3 versions are only semantically equivalent in the case where only the application cares that the filename is unique. Now, the basic philosophical question of which preprocessor contstucts to use here would appear to revolve around the following choices: (i). Use a high level, large grained standard definition (e.g., _POSIX_SOURCE). In this case, features are implicitly defined. One problem with such definitions is that they may cause other useful functions to become unavailable. However, there are several such definitions in common use. For operating systems, we have SVR4 SYSV BSD OSF1 to name a few. For standards, we are mainly interested interested symbols such as __STDC__ _POSIX_SOURCE _XOPEN_SOURCE This method is not without pitfalls. For example, the Sun SC2.0.1 compiler defines __STDC__ as 0 when compiling in transition mode (-Xt), only setting it to 1 when the strict ANSI mode (-Xc) is used. The expression #if (__STDC__ - 0 == 0) can be used to recognize strict v. transition ANSI modes. On Solaris 2, if you compile with -Xc, you will lose all non-ANSI functionality. However, you can define _POSIX_SOURCE or _XOPEN_SOURCE to get a POSIX or XOPEN environment. If you use _POSIX_SOURCE, .eg., #define _POSIX_SOURCE 1 then all symbols not defined by Standard C or the POSIX standard will be hidden (except those with leading underscores). If you wish to use _POSIX_SOURCE, be sure to define it before including any standard header files, and avoid name clashes by not defining any symbols that begin with "_" (Similarly, note that almost all names beginning with "E" are reserved by errno.h, and many names prefixed by "va_" reserved by stadarg.h). One more note on _POSIX_SOURCE: SunOS 5.3 has introduced the new header file <sys/feature_tests.h>. This file is included in all files which have _POSIX_SOURCE dependancies. A new symbol, _POSIX_C_SOURCE was introduced in POSIX.2 (V1 P720, L51) as a mechanism to enable POSIX.1 and POSIX.2 symbols. Its values are as follows: /* * Values of _POSIX_C_SOURCE * * undefined not a POSIX compilation * 1 POSIX.1-1990 compilation * 2 POSIX.2-1992 compilation * 1993xxL POSIX.4-1993 compilation */ This means that POSIX.2 says that a value of 1 = POSIX.1 and a value of 2 = POSIX.1 & POSIX.2. The idea here is to provide a single control point over the POSIX namespace, rather than having to edit each file individually. Another potential portability pitfall is the __svr4__ feature defined by the FSF (gcc). If you depend on __svr4__, you may lose portability. gcc also defines sun if you don't give the -ansi argument. If you use -ansi, then sun is not defined and __sun__ is. Finally, complexity may arise surrounding a feature which may be part of some vendor's version of some system Y, but may also exist in non-Y compliant systems. Consider, for example, shadow passwording. Systems conforming to the latest SVID (e.g., SVR4) have shadow.h, but there are many systems that have shadow.h without conforming to the SVID. So, in general, for code that uses a STD_FEATURE and runs on systems W, Y, and Z, you are left with something that may look like #if defined(W) || (defined(Y) && _Y_VERSION_ > 3) || (defined(Z) || defined(__Z__)) #include <STD_FEATURE.h> #endif [W, Y, Z are things like SVR4, AIX, NeXT, BSD, and so on. STD_FEATURE.h is something like shadow.h] This example exposes two problems the large grained method. First, it forces one to keep track of exactly which vendors supply <STD_FEATURE.h>. Second, the complexity of the preprocessor expressions may be a serious consideration, since their complexity is something like O(n*m) where n = the number of standard features, and m = number of vendors/systems (ii). Define new fine-grained feature tests (e.g., HAVE_POSIX_SIGNALS, or HAVE_SHADOW_H) for features of interest. Such fine-grained features could be used in conjunction with large grained definitions. An nice example of using feature definitions is the GNU configure program. It uses, for example, the features HAVE_BCOPY and HAVE_MEMSET to enable either the bcopy (BSD) or memset (ANSI) functions. Feature testing has the advantage of being useful for automatic configuration with programs such as GNU configure. GNU configure outputs statements of the form #define HAVE_aaaa #define HAVE_bbbb #define HAVE_cccc .... Another way to generate a feature set is by using the symbol defining the system, e.g., #ifdef SVR4 #define HAVE_aaaa #define HAVE_bbbb #define HAVE_cccc .... #endif #ifdef BSD43 #define HAVE_yyyy ... #endif #ifdef NEWTHING #define HAVE_zzzz ... #endif Feature testing also helps to avoid constructs such as #if defined(_POSIX_SOURCE) || defined(_XOPEN_SOURCE) [Editor's Note: Finally, an observation: The real issue here appears to be how many of these "features" are migrating to the standard operating systems and interfaces, and how many vendors are implementing these standards. In general, some people feel that feature testing improves portability (and readability), and others believe that the feature testing style decreases portability and readability. DM] (iii). Use some part of the feature's definition itself to enable the feature, for example #ifdef _IOLBF setvbuf(stderr, NULL, _IOLBF, 0); #else setlinebuf(stderr); #endif /* _IOLBF */ Note that in this case, another, possibly better option is (consider the case in which some vendor has inadvertently defined _IOLBF for some other purpose): #ifdef __STDC__ setvbuf(stderr, NULL, _IOLBF, 0); #else setlinebuf(stderr); #endif /* __STDC__ */ since setvbuf is required by Standard C. Finally, some people have suggested the use of expressions like #if defined(sun) && defined(__svr4__) <Solaris 2 centric code> #else ... #endif As noted above, the __svr4__ feature is defined by the FSF (gcc). If you depend on __svr4__, you may lose portability. gcc also defines sun if you don't give the -ansi argument. If you use -ansi, then sun is not defined and __sun__ is. The implication here is that depending on symbols defined by a given compiler can reduce portability. In general, such a construct should be used if and only if the code in question cannot be covered by some standard (e.g., SVR4, _POSIX_SOURCE, etc.). Note that it is also compiler specific. ----------------------------------------------------------------------------- 1) TOPIC: Include File Issues [Last modified: 19 August 93] The first and apparently most common problem is that /usr/include/strings.h is not ANSI compliant, and as such does not exist on Solaris 2 (or SVR4). It should be replaced by /usr/include/string.h, e.g. (following GNU feature definition conventions) #if HAVE_STRING_H || defined(STDC_HEADERS) #include <string.h> #else #include <strings.h> #endif #if defined(STDC_HEADERS) #include <stdlib.h> #endif /* HAVE_STRING_H */ while ANSI-C requires the name be string.h, one might define this as #ifdef __STDC__ #include <string.h> #else #include <strings.h> #endif /* __STDC__ */ However, this again neglects the case in which the vendor provides string.h in a non-ANSI environment. Another thing to watch is for the symbols O_CREAT, O_TRUNC, and O_EXCL being undefined. On BSD systems, these are defined in <sys/file.h>. On Solaris 1 systems (beginning with SunOS 4.0) , these are defined in <sys/fnctlcom.h> (which is included in <sys/file.h>). On a POSIX compliant system, these symbols are defined in <fcntl.h>, which is not included in <sys/file.h>. Since <fcntl.h> is defined on SunOS 4.1.x, replacing <sys/file.h> with <fcntl.h> works for both SunOS 4.1.x and SVR4. See, for example, section 5.3.1.1 of the POSIX spec. ----------------------------------------------------------------------------- 2) TOPIC: Libraries [Last modified: 12 Feburary 94] Network Libraries: Many of the network functions and definitions that were present in the BSD libc are now in libnsl.so and libsocket.so. Thus networking code will generally need to be linked with -lsocket -lnsl. Since libsocket.so requires libnsl.so (it is NEEDED), you must specify them in this order. Note that you need libnsl.so for functions like gethostbyname (see gethostbyname note below). Incidently, you can look at selected parts of an object file using dump(1), e.g., % dump -Lv /usr/lib/libsocket.so /usr/lib/libsocket.so: **** DYNAMIC SECTION INFORMATION **** .dynamic : [INDEX] Tag Value [1] NEEDED libnsl.so.1 [2] INIT 0x3174 [...] Regular Expressions Another problem frequently encountered is that the regexp functions (see regexpr(3G)) are not defined in libc. On Solaris 2, you must link with libgen.a (-lgen) in order to get these definitions. See Intro(3) for a more complete discussion. Name List (nlist) You must link with libelf.a (-lelf) to get the nlist(3E) definition. ----------------------------------------------------------------------------- 3)* TOPIC: Possible ANSI/POSIX/SVR4 replacements for some popular BSD functions [Last modified: 28 Feburary 94] [Editor's Note: Once again, this section began life a SunOS 4.1.x and SunOS 5.x centric discussion. It too has grown into a discussion dealing with general portability for BSD to other standards. DM] Problems finding functions that were defined in the BSD libc.a is one of the most frequently asked porting questions. The following table and code fragments suggest substitutes for some common BSD constructs (more complete lists can be found in some of the texts listed in section 7 below). BSD Possibilities Standards/Notes ============================================================================ srandom(seed) srand(seed) ANSI-C (Also, some older UNIX) srand48(seed) SVR4 non-ANSI signal() sigset() SVR4 (systems calls not (e.g., SunOS) restarted, but bytes r/w returned, else EINTR) sigaction POSIX (but extensible by implementation) sigvec sigaction POSIX sigblock sigprocmask POSIX sigset(.., SIG_HOLD) sighold SVR4 sigsetmask sigprocmask POSIX sigset/sigrelse SVR4 sigpause sigsuspend POSIX setjmp sigsetjmp POSIX longjmp siglongjmp POSIX statfs statvfs SVR4 bcopy memmove ANSI-C (BSD bcopy() handles overlapping areas correctly, as does memmove, but not memcpy) bzero memset ANSI-C index strchr ANSI-C rindex strrchr ANSI-C getwd getcwd POSIX getrusage open,ioctl The getrusage information (and a whole lot more) can be found in the prusage structure. Use the PIOCUSAGE ioctl. See the example below and the proc(4) man page for detail. gethostname sysinfo(SI_HOSTNAME,..) SVR4 See sysinfo(2) for many other possible values getdtablesize sysconf(_SC_OPEN_MAX) POSIX See sysconf(3C) for many other values available via sysconf. timelocal mktime wait3 w/o rusage waitpid POSIX wait3 waitid SVR4 usleep nanosleep POSIX See nanosleep(3R) on Solaris 2.3 (see libposix4.a) For a Solaris 2.[0-2], see the example below. ------------------------------------------------------------------ Timing Problems --------------- POSIX defines the <sys/times.h> function for subsecond timing. Sun seems to provide about 1/60 second accuracy. #include <stdio.h> #include <sys/times.h> /* for struct tms and times() */ #include <time.h> /* for CLK_TCK value */ int main(void) { struct tms tms_start, tms_finish; /* user and system time */ clock_t start, finish; /* real time */ start = times( &tms_start ); /* ... do something ... */ finish = times( &tms_finish ); printf("(in seconds) %f real, %f system, %f user\n", (finish-start) / (double)CLK_TCK, (tms_finish.tms_stime-tms_start.tms_stime) / (double)CLK_TCK, (tms_finish.tms_utime-tms_start.tms_utime) / (double)CLK_TCK); return 0; } You might want to divide CLK_TCK by 1000.0 to get more precise millisecond values. times() returns -1 if it cannot provide timing information. While Solaris 2 conforms to POSIX, SunOS 4.1 defines times() as returning a flag instead of elapsed real time. You can use ftime() to get elapsed real time: #include <stdio.h> #include <sys/types.h> /* for time_t */ #include <sys/timeb.h> /* for ftime() and struct timeb */ int main(void) { struct timeb start, finish; double real_secs; ftime( &start ); /* ... do something ... */ ftime( &finish ); real_secs = finish.time - start.time; if ( finish.millitm < start.millitm ) real_secs = (real_secs-1) + (1000+start.millitm-finish.millitm)/1000.0; else real_secs = (finish.millitm-start.millitm)/1000.0; printf( "That took %f real seconds.", real_secs ); return 0; } The ANSI C function clock() can also be used for timing. It returns elased "processor" time, which is equivalent to system+user time. While it also returns a clock_t value, you must divide the difference between to calls to clock() by CLOCKS_PER_SEC, *not* CLK_TCK. The values are different by orders of magnitude. SunOS 4.1 doesn't seem to provide CLOCKS_PER_SEC or CLK_TCK in <time.h>. Try 10000000 and 60, respectively. Compatibility Functions ----------------------- /* * getrusage -- */ #include <sys/resource.h> #ifndef RUSAGE_SELF #include <sys/procfs.h> #endif #ifdef PIOCUSAGE int fd; char proc[SOMETHING]; prusage_t prusage; sprintf(proc,"/proc/%d", getpid()); if ((fd = open(proc,O_RDONLY)) == -1) { perror("open"); .... } if (ioctl(fd, PIOCUSAGE,, &prusage) == -1) { perror("ioctl"); ... } .... #else /* Again, assume BSD */ if (getrusage(RUSAGE_SELF, &rusage) == -1) { perror("getrusage"); .... } .... #endif /* PIOCUSAGE */ /* * setlinebuf -- * */ #ifdef __STDC__ setvbuf(stderr, NULL, _IOLBF, 0); #else setlinebuf(stderr); #endif /* __STDC__ */ /* * gethostid * * This example has a combination of high-level * (SVR4) and (SI_HW_SERIAL) feature declarations. */ #if defined(SVR4) && defined(SI_HW_SERIAL) long gethostid() { char buf[128]; if (sysinfo(SI_HW_SERIAL, buf, 128) == -1) { perror("sysinfo"); exit(1); } return(strtoul(buf,NULL,0)); } #endif /* SVR4 && SI_HW_SERIAL */ /* * getdtablesize -- * * Several possibilites here. Note that while one * can emulate getdtablesize with getrlimit on SVR4 * or 4.3BSD (or later), these systems should be * POSIX.1 compliant, so sysconf is preferred. * */ #ifdef _SC_OPEN_MAX /* POSIX -- preferred */ if ((tableSize = sysconf(_SC_OPEN_MAX)) == -1) { perror("sysconf"); ... } #elif RLIMIT_NOFILE /* who is non POSIX but has this? */ if (getrlimit(RLIMIT_NOFILE, &rlimit) == -1) { perror("getrlimit"); exit(1); } tableSize = rlimit.rlim_max; #else /* assume old BSD type */ tableSize = getdtablesize(); #endif ------------------ /* * gethostname -- * */ #ifdef SVR4 #include <sys/systeminfo.h> #endif /* SVR4 */ .... char buf[MAXHOSTNAME] #ifdef SVR4 if (sysinfo(SI_HOSTNAME, buf, sizeof(buf)) <0) { perror("SI_HOSTNAME"); exit(BAD); } #else /* Assume BSD */ if (gethostname(buf, sizeof(buf)) < 0) { perror("gethostname"); exit(BAD); } #endif /* SVR4 */ /* buf has hostname here... */ /* * usleep(delay) -- * * Possible usleep replacement. Delay in microseconds. * Another possiblity is to use poll(2). On Solaris * 2.x, select is just a wrapper for poll, so you * are better off using it directly. If you use, * poll, note that it uses millisecond resolution, * and is not affected by the O_NDELAY and O_NONBLOCK * flags. * * Note that using nanosleep has portability implications, * even across different versions of Solaris 2.x. In * particular, only Solaris 2.3 has libposix4, and * hence nanosleep. Select (or poll) is a better option if * you need portability across those versions. * * If you define USE_NANOSLEEP, be sure to link with -lposix4 * */ #include <unistd.h> #include <stdlib.h> #include <stdio.h> #include <errno.h> #include <time.h> #include <sys/time.h> #include <sys/param.h> #include <sys/types.h> #ifdef USE_POLL #include <stropts.h> #include <poll.h> #endif /* USE_POLL */ int usleep(unsigned long int useconds) { #ifdef USE_NANOSLEEP struct timespec rqtp; rqtp.tv_sec = useconds / (unsigned long) 1000000; rqtp.tv_nsec = (useconds % (unsigned long) 1000000) * 1000 ; if (nanosleep(&rqtp, (struct timespec *) NULL) == -1) perror("nanosleep"); return (0); #elif USE_POOL struct pollfd unused; if (poll(&unused,0,(useconds/1000)) == -1) perror("poll"); return(0); #elif USE_USLEEP struct timeval delay; delay.tv_sec = 0; delay.tv_usec = useconds; if (select(0, (fd_set *) NULL, (fd_set *) NULL, (fd_set *) NULL, &delay) == -1) perror("select"); return (0); #endif /* USE_NANOSLEEP */ /* * tzsetwall -- */ void tzsetwall() { unsetenv("TZ"); tzset(); } /* * gethostybname -- * * The following code was contributed by Casper H.S. Dik * to address the following problem: * * gethostbyname() always returns null in h->aliases. * Now, gethostbyX can be replaced its __switch_gethostbyX * equivalents. However, these are missing from Solaris 2.3. * * The _r functions are reentrant. They have a different * calling sequence. (The __switch_getXXXbyYYY are called * like getXXXbyYYY, the _switch_getXXXbyYYY_r are called * like getXXXbyYYY_r) * * With this bit of knowledge I constructed the code that * follows this message. Just plug it in every program * that requires gethostbyname to work. (Gethostbyaddr() * is added for symmetry). * * You'll need to link with -lnsl -ldl. * * It works for Solaris 2.2 and 2.3. (Compiled on 2.3 or * 2.2 it will run 2.2 and 2.3) * * Note that as with __switch* _switch*_r is undocumented * and can be changed in the next release. * */ /* * Proper gethostbyXX function for Solaris 2.0-2.3 * (and later ?) * * Fixed in 2.4? * * You'll need -ldl added to the link command line. * * Casper Dik (casper@fwi.uva.nl) * */ #include <netdb.h> #include <dlfcn.h> #define HBUFSIZE 4096 static void *dlhandle; /* The gethostbyXXX function variables. Named after * then .so files they appear in. nsw*.so in SunOS 5.2 * and earlier, nss_*.so in 5.3 */ static struct hostent *(*nswghba)(const char *, int, int), *(*nswghbn)(const char *), *(*nss_ghba)(const char *, int, int, struct hostent *, char *, int, int *), *(*nss_ghbn)(const char *, struct hostent *, char *, int, int *); static int dlinit(void) { static int dlstatus = 0; /* 0 = uninited, 1 = inited & ok, -1 = error */ if (dlstatus) return dlstatus; dlstatus = -1; dlhandle = dlopen(0, RTLD_LAZY); if (dlhandle == 0) return dlstatus; /* SunOS 5.0 - 5.2 */ nswghba = (struct hostent *(*)(const char *, int, int)) dlsym(dlhandle, "__switch_gethostbyaddr"); nswghbn = (struct hostent *(*)(const char *)) dlsym(dlhandle, "__switch_gethostbyname"); /* either both should exist or both should not exist */ if ((nswghbn == 0) != (nswghba == 0)) return dlstatus; if (nswghbn) return dlstatus = 1; /* SunOS 5.3 - ? */ nss_ghba = (struct hostent *(*) (const char *, int, int, struct hostent *, char *, int , int *)) dlsym(dlhandle, "_switch_gethostbyaddr_r"); nss_ghbn = (struct hostent *(*) (const char *, struct hostent *, char *, int , int *)) dlsym(dlhandle, "_switch_gethostbyname_r"); /* these two must exist when we get here */ if (nss_ghbn != 0 && nss_ghba != 0) dlstatus = 1; return dlstatus; } struct hostent * gethostbyname(const char *name) { static struct hostent hp; static char buf[HBUFSIZE]; if (dlinit() == -1) return 0; if (nswghbn) return nswghbn(name); else return nss_ghbn(name, &hp, buf, sizeof(buf), &h_errno); } struct hostent * gethostbyaddr(const char *addr, int len, int type) { static struct hostent hp; static char buf[HBUFSIZE]; if (dlinit() == -1) return 0; if (nswghba) return nswghba(addr, len, type); else return nss_ghba(addr, len, type, &hp, buf, sizeof(buf), &h_errno); } ----------------------------------------------------------- 4) TOPIC: BSD/Solaris 1/POSIX Signal Primer [Last modified: 20 September 93] The most common problem encountered when porting BSD/Solaris 1 signal code is that Solaris 2 (and SVR4) handles interrupted systems calls differently than does BSD. In Solaris 2 (SVR4), system calls are interrupted and return EINTR, unless the call is read, write, or some other call that returns the number of bytes read/written (unless 0 bytes have been read/written, in which case the call returns EINTR). On the other hand, system calls are restarted on BSD/Solaris 1 systems. The signal calls can be made to restart by specifying a SA_RESTART with sigaction(). Note, however, that code that relies on restartable system calls is generally considered bad practice. The following code is provided for illustrative purposes only. It is recommended that you remove these dependencies. Sigaction is the preferred (POSIX) way of installing signal handlers. The BSD/Solaris 1 code omask = sigblock(sigmask(SIGXXX)); do_stuff_while_SIGXXX_blocked(); (void)sigsetmask(omask); can be emulated by sigset_t block, oblock; struct sigaction act, oact; .... (void)sigemptyset(&block); (void)sigaddset(&block, SIGXXX); if (sigprocmask(SIG_BLOCK, &block, &oblock) < 0) perror("sigprocmask"); do_stuff_while_SIGXXX_blocked(); (void)sigprocmask(SIG_SETMASK, &oblock, (sigset_t *)NULL); #ifdef SA_RESTART /* make restartable */ act.sa_flags = SA_RESTART; #endif /* SA_RESTART */ if (sigaction(SIGXXX, &act, &oact) < 0) return(SIG_ERR); Note that this (emulating) construct is also available on Solaris 1 (sans SA_RESTART), so should work on either Solaris 1 or SVR4. Another possibility would be to emulate BSD signal(2) semantics as follows: Sigfunc *bsdsignal(int signo, Sigfunc *alarm_catcher) { struct sigaction act; act.sa_handler = alarm_catcher; sigemptyset(&act.sa_mask); act.sa_flags = SA_RESTART; if(sigaction(signo, &act,NULL) == -1) { perror("signal:"); return(SIG_ERR); } Another problem revolves around the use of setjmp and longjmp. With 4.3+BSD the setjmp and longjmp save and restore the signal mask. The default behavior for SVR4 is not to save and restore the signal mask. Note that these calls are MT-Unsafe. The POSIX.1 interface allows you to do either, by using a second argument, savemask, for sigsetjmp. To cause the signal mask to be saved and restored (emulating setjmp/longjmp behavior), use a nonzero savemask. For example, #include <stdio.h> #include <stdlib.h> #include <setjmp.h> sigjmp_buf env; int savemask; .... savemask = 1; #ifdef HAS_SIGSETJMP sigsetjmp(env, savemask); #endif HAS_SIGSETJMP ... In this case, the sigsetjmp saves the current signal mask of the process in the saved environment (sigjmp_buf). Now, if the environment was saved by a call to sigsetjmp with a nonzero savemask, then a subsequent siglongjmp call will restore the saved signal mask. To summarize, some basic rules are: (i). Limit signal handling code to the POSIX interface whenever possible. (ii). Use sigaction to install signal handlers whenever possible. Use Standard C's signal() only for portability to non-POSIX systems. (iii). Avoid code that relies on restartable system calls. (iv). The main difference between SVR4 sigset() (not POSIX) and SunOS 4.x/BSD signal() is that system calls will return EINTR with sigset() but will be restarted on BSD/SunOS 4.x. On SVR4 EINTR is only returned when no bytes have been read/written. ----------------------------------------------------------------------------- 5) TOPIC: Waiting for Children to Exit [Last modified: 26 October 93] waitpid(2) is the preferred (POSIX) interface. Wait3 can be replaced by waitpid (when you don't need the rusage). For example, the BSD segment while((id = wait(&stat)) >=0 && id != pid); can be approximated using the POSIX waidpid(2) interface by code of the form: int status; int options; /* e.g., WNOHANG */ .... options = WNOHANG; if (waitpid((pid_t) -1, &status, options) == -1) perror("waitpid"); } Note here that if you execute a signal(SIGCHLD, SIG_IGN) or sigset(SIGCHLD, SIG_IGN), Solaris will discard all child exit statuses and reap the child processes without giving the parent a chance to wait. That is, waitpid(2) will return -1 with an ECHILD. Another possibility is emulate the BSD wait(2) call with SVR4's waitid(2). The code fragment below is an example. In this case, we wait for a particular child in our process group ((pid_t) 0) to exit (WEXITED). #ifdef SVR4 #include <sys/types.h> #include <sys/wait.h> siginfo_t stat; int retcode; #else union wait stat; #endif ..... #ifdef SVR4 while (retcode = waitid(P_ALL,(pid_t) 0, &stat, WEXITED)) { if (retcode < 0) { perror("waitid"); exit(1); } if (stat.si_pid == pid) break; } #else /* BSD */ while((id = wait(&stat)) >=0 && id != pid); #endif /* SVR4 */ ----------------------------------------------------------------------------- 6) TOPIC: Dealing With Shadow Password Files [Last modified: 19 August 93] The following code segment outlines how to handle shadow password files. In the outline below, <passwd> is the clear text password. Note that shadow passwords are part of SVR4, so again we have the conflict between using high level system definitions (e.g., SVR4) and feature definitions (for systems other than SVR4). I'll use feature a feature definition (HAVE_SHADOW_H) to illustrate this. #ifdef HAVE_SHADOW_H #include <shadow.h> register struct spwd *sp; #endif /* HAVE_SHADOW_H */ ..... #ifdef HAVE_SHADOW_H if ((sp = getspnam(<username>)) == NULL) <no password entry for username> if (sp->sp_pwdp == NULL) <NULL password for username> if (strcmp (crypt (<passwd>, sp->sp_pwdp), sp->sp_pwdp) != 0) #else if ((pw = getpwnam(<username>)) == NULL) <no password entry for username> if (pw->pw_passwd == NULL) <NULL password for username> if (strcmp (crypt (<passwd>, pw->pw_passwd), pw->pw_passwd) != 0) #endif /* HAVE_SHADOW_H */ <incorrect password for username> ----------------------------------------------------------------------------- 7)* TOPIC: Some Compatabilty Problems [Last modified: 02 September 93] (i). vfork doesn't work in Solaris 2 [Editor's Note: The following observation and example comes courtesy of Paul Eggert (eggert@twinsun.com). DM] In Solaris 2, if a vfork'ed child adjusts signal handling before exec'ing, signal handling is munged in the parent in ways that lead to unreliable results; the parent can dump core in some cases. This bug affects some widely distributed programs, so when building a program that adjusts signal handlers between `vfork' and `exec', be careful to override its configuration to use `fork' instead. Sun doesn't consider this behavior to be a bug, so it's not likely to be fixed. Here's an illustration of the bug. This program works fine in SunOS 4.1.x, but dumps core in Solaris 2.x. #include <signal.h> #include <unistd.h> #ifdef SIGLOST /* must be SunOS 4.1.x */ #include <vfork.h> #endif int signalled; void catch (sig) int sig; { signalled = 1; } int main() { signal (SIGINT, catch); if (vfork () == 0) { /* child */ signal (SIGINT, SIG_IGN); execlp ("sleep", "sleep", "10", (char *) 0); } /* parent here */ kill (getpid (), SIGINT); return signalled != 1; } (ii). chown(2) does not allow uid/gid values greater than 60002 on Solaris 2.[0-3]. Check /usr/include/limits.h, which contains: #define UID_MAX 60002 ----------------------------------------------------------------------------- 8)* TOPIC: Other Resources [Last modified: 27 September 93] Porting to Solaris 2 -------------------- A excellent text on this subject is "Solaris Porting Guide", SunSoft ISV Engineering, et. al., Prentice Hall, 1993. ISBN 0-13-030396-8. Solaris 2 General FAQ --------------------- The official Solaris 2 Frequently Answered Questions is maintained by Ian Darwin, ian@sq.com, and is posted once or twice a month to various newsgroups including comp.unix.solaris and comp.answers. General ------- "Internetworking with TCP/IP: Volume III Client-Server Programming & Applications (AT&T TLI Edition)", Douglas E. Comer & David L. Stevens, Prentice-Hall. ISBN 0-13-474230-3. Nice reference for TLI programming, etc. "Networking Applications on UNIX System V", Mike Padovano, ISBN 013-613555. A good reference for System V. "UNIX, POSIX, and Open Systems: The Open Standards Puzzle", John S. Quarterman and Susanne Wilhelm, Addison-Wesley, 1993. ISBN 0-201-52772-3. Another nice modern reference. "UNIX System V Network Programming", Steve Rago, ISBN 0-201-56318-5. Another good System V reference. "Advanced Programming in the UNIX Environment", W. Richard Stevens, Addison Wesley, 1992, ISBN 0-201-56317-1, is a nice, in depth text covering large parts of this topic. ANSI C ------ A very nice text here is "The Standard C Library", P.J. Plauger, Prentice Hall, 1992, ISBN 0-13-131509-9. Another example of the many texts here is "C, a Reference Manual", Harbison and Steele, Prentice Hall. ISBN 0-13-110933-2. POSIX ----- A nice reference text on the POSIX interface is "POSIX Programmer's Guide", Donald Levine, O'Reily & Associates, 1991. ISBN 0-937175-73-0. ACKNOWLEDGMENTS I would like to thank everyone who contributed to this, and I hope that it clarifies some of these issues. I would especially acknowledge the contributions of Casper H.S. Dik and J.G. Vons in helping me organize my thoughts on all this. Thanks to: Jamshid Afshar <jamshid@ses.com> Pedro Acebes Bayon <pacebes@tid.es> Ian Darwin <ian@sq.com> Casper H.S. Dik <casper@fwi.uva.nl> Paul Eggert <eggert@twinsun.com> Stephen L Favor <xcpslf@atom.oryx.com> Charles Francois <cbf@gotham.east.sun.com> Pete Hartman <pwh@bradley.bradley.edu> Guy Harris <guy@auspex.com> Jens-Uwe Mager <jum@anubis.han.de> Thomas Maslen <maslen@eng.sun.com> Richard M. Mathews <richard@astro.west.sun.com> Davin Milun <milun@cs.buffalo.edu> Paul Pomes <Paul-Pomes@uiuc.edu> Andrew Roach <andrewr@ultrix.sun.com> M C Srivas <M._C._Srivas@transarc.com> Larry W. Virden <lwv26@cas.org> J.G. Vons <vons%ulysse@crbca1.sinet.slb.com> Peter Wemm <peter@DIALix.oz.au> christos@deshaw.com jorgens@pvv.unit.no Malte <malte@techfak.uni-bielefeld.de> ----- End of Solaris 2 Porting FAQ -- Maintained by David Meyer meyer@ns.uoregon.edu -- David M. Meyer 503/346-1747 meyer@cambium.uoregon.edu Fri Jul 16 14:59:28 1993 $Header: /net/network-services/disk1/home/meyer/FAQ/RCS/porting-FAQ,v 1.56 1994/04/25 12:34:52 meyer Exp $