Platform-Specific Configuration and Operation Notes =================================================== 1. configure --without-gnu-malloc on: alpha running OSF/1, Linux, or NetBSD (malloc needs 8-byte alignment; bash malloc has 8-byte alignment now, but I have no alphas to test on) next running NeXT/OS all machines running SunOS YP code: SunOS4, SunOS5, HP/UX, if you have problems with username completion or tilde expansion for usernames found via YP/NIS linux (optional, but don't do it if you're using Doug Lea's malloc) QNX 4.2 other OSF/1 machines (KSR/1, HP, IBM AIX/ESA) AIX sparc SVR4, SVR4.2 (ICL reference port) DG/UX Cray NetBSD/sparc (malloc needs 8-byte alignment; bash malloc has 8-byte alignment now, but I have no NetBSD machines to test on) BSD/OS 2.1, 3.x if you want to use loadable builtins Motorola m68k machines running System V.3. There is a file descriptor leak caused by using the bash malloc because closedir(3) needs to read freed memory to find the file descriptor to close If you are using GNU libc, especially on a linux system (Configuring --without-gnu-malloc will still result in lib/malloc/libmalloc.a being built and linked against, but there is only a stub file in the archive.) 2. Configure using shlicc2 on BSD/OS 2.1 and BSD/OS 3.x to use loadable builtins 3. Bash cannot be built in a directory separate from the source directory using configure --srcdir=... unless the version of `make' you're using does $VPATH handling right. The script support/mkclone can be used to create a `build tree' using symlinks to get around this. 4. I've had reports that username completion (as well as tilde expansion and \u prompt expansion) does not work on IRIX 5.3 when linking with -lnsl. This is only a problem when you're running NIS, since apparently -lnsl supports only /etc/passwd and not the NIS functions for retrieving usernames and passwords. Editing the Makefile after configure runs and removing the `-lnsl' from the assignment to `LIBS' fixes the problem. 5. There is a problem with the `makewhatis' script in older (pre-7.0) versions of Red Hat Linux. Running `makewhatis' with bash-2.0 or later versions results in error messages like this: /usr/sbin/makewhatis: cd: manpath: No such file or directory /usr/sbin/makewhatis: manpath/whatis: No such file or directory chmod: manpath/whatis: No such file or directory /usr/sbin/makewhatis: cd: catpath: No such file or directory /usr/sbin/makewhatis: catpath/whatis: No such file or directory chmod: catpath/whatis: No such file or directory The problem is with `makewhatis'. Red Hat (and possibly other Linux distributors) uses a construct like this in the code: eval path=$"$pages"path to do indirect variable expansion. This `happened to work' in bash-1.14 and previous versions, but that was more an accident of implementation than anything else -- it was never supported and certainly is not portable. Bash-2.0 has a new feature that gives a new meaning to $"...". This is explained more completely in item 1 in the COMPAT file. The three lines in the `makewhatis' script that need to be changed look like this: eval $topath=$"$topath":$name [...] eval path=$"$pages"path [...] eval path=$"$pages"path The portable way to write this code is eval $topath="\$$topath":$name eval path="\$$pages"path eval path="\$$pages"path You could also experiment with another new bash feature: ${!var}. This does indirect variable expansion, making the use of eval unnecessary. 6. There is a problem with syslogd on many Linux distributions (Red Hat and Slackware are two that I have received reports about). syslogd sends a SIGINT to its parent process, which is waiting for the daemon to finish its initialization. The parent process then dies due to the SIGINT, and bash reports it, causing unexpected console output while the system is booting that looks something like starting daemons: syslogd/etc/rc.d/rc.M: line 29: 38 Interrupt ${NET}/syslogd Bash-2.0 reports events such as processes dying in scripts due to signals when the standard output is a tty. Bash-1.14.x and previous versions did not report such events. This should probably be reported as a bug to whatever Linux distributor people see the problem on. In my opinion, syslogd should be changed to use some other method of communication, or the wrapper function (which appeared to be `daemon' when I looked at it some time ago) or script (which appeared to be `syslog') should catch SIGINT, since it's an expected event, and exit cleanly. 7. Several people have reported that `dip' (a program for SLIP/PPP on Linux) does not work with bash-2.0 installed as /bin/sh. I don't run any Linux boxes myself, and do not have the dip code handy to look at, but the `problem' with bash-2.0, as it has been related to me, is that bash requires the `-p' option to be supplied at invocation if it is to run setuid or setgid. This means, among other things, that setuid or setgid programs which call system(3) (a horrendously bad practice in any case) relinquish their setuid/setgid status in the child that's forked to execute /bin/sh. The following is an *unofficial* patch to bash-2.0 that causes it to not require `-p' to run setuid or setgid if invoked as `sh'. It has been reported to work on Linux. It will make your system vulnerable to bogus system(3) calls in setuid executables. --- ../bash-2.0.orig/shell.c Wed Dec 18 14:16:30 1996 +++ shell.c Fri Mar 7 13:12:03 1997 @@ -347,7 +347,7 @@ if (posixly_correct) posix_initialize (posixly_correct); - if (running_setuid && privileged_mode == 0) + if (running_setuid && privileged_mode == 0 && act_like_sh == 0) disable_priv_mode (); /* Need to get the argument to a -c option processed in the 8. Some people have asked about binding all of the keys in a PC-keyboard- style numeric keypad to readline functions. Here's something I received from the gnu-win32 list that may help. Insert the following lines into ~/.inputrc: # home key "\e[1~":beginning-of-line # insert key "\e[2~":kill-whole-line # del key "\e[3~":delete-char # end key "\e[4~":end-of-line # pgup key "\e[5~":history-search-forward # pgdn key "\e[6~":history-search-backward 9. Hints for building under Minix 2.0 (Contributed by Terry R. McConnell, ) The version of /bin/sh distributed with Minix is not up to the job of running the configure script. The easiest solution is to swap /bin/sh with /usr/bin/ash. Then use chmem(1) to increase the memory allocated to /bin/sh. The following settings are known to work: text data bss stack memory 63552 9440 3304 65536 141832 /bin/sh If you have problems with make or yacc it may be worthwhile first to install the GNU versions of these utilities before attempting to build bash. (As of this writing, all of these utilities are available for the i386 as pre-built binaries via anonymous ftp at math.syr.edu in the pub/mcconnell/minix directory. Note that the GNU version of yacc is called bison.) Unless you want to see lots of warnings about old-style declarations, do LOCAL_CFLAGS=-wo; export LOCAL_CFLAGS before running configure. (These warnings are harmless, but annoying.) configure will insist that you supply a host type. For example, do ./configure --host=i386-pc-minix. Minix does not support the system calls required for a proper implementation of ulimit(). The `ulimit' builtin will not be available. Configure will fail to notice that many things like uid_t are indeed typedef'd in , because it uses egrep for this purpose and minix has no egrep. You could try making a link /usr/bin/egrep --> /usr/bin/grep. Better is to install the GNU version of grep in /usr/local/bin and make the link /usr/local/bin/egrep -->/usr/local/bin/grep. (These must be hard links, of course, since Minix does not support symbolic links.) You will see many warnings of the form: warning: unknown s_type: 98 I have no idea what this means, but it doesn't seem to matter. 10. If you do not have /usr/ccs/bin in your PATH when building on SunOS 5.x (Solaris 2), the configure script will be unable to find `ar' and `ranlib' (of course, ranlib is unnecessary). Make sure your $PATH includes /usr/ccs/bin on SunOS 5.x. This generally manifests itself with libraries not being built and make reporting errors like `cr: not found' when library construction is attempted. 11. Building a statically-linked bash on Solaris 2.5.x, 2.6, or 7 is complicated. It's not possible to build a completely statically-linked binary, since part of the C library depends on dynamic linking. The following recipe assumes that you're using gcc and the Solaris ld (/usr/ccs/bin/ld) on Solaris 2.5.x or 2.6: configure --enable-static-link make STATIC_LD= LOCAL_LIBS='-Wl,-B,dynamic -ldl -Wl,-B,static' This should result in a bash binary that depends only on libdl.so: thor(2)$ ldd bash libdl.so.1 => /usr/lib/libdl.so.1 If you're using the Sun C Compiler (Sun WorkShop C Compiler version 4.2 was what I used), you should be able to get away with using configure --enable-static-link make STATIC_LD= LOCAL_LIBS='-B dynamic -ldl -B static' If you want to completely remove any dependence on /usr, perhaps to put a copy of bash in /sbin and have it available when /usr is not mounted, force the build process to use the shared ld.so library in /etc/lib. For gcc, this would be something like configure --enable-static-link make STATIC_LD= LOCAL_LIBS='-Wl,-B,dynamic -Wl,-R/etc/lib -ldl -Wl,-B,static' For Sun's WS4.2 cc configure --enable-static-link make STATIC_LD= LOCAL_LIBS='-B dynamic -R/etc/lib -ldl -B static' seems to work, at least on Solaris 2.5.1: thor(2)$ ldd bash libdl.so.1 => /etc/lib/libdl.so.1 On Solaris 7 (and presumably Solaris 8, though I do not run that), the following recipe appears to work for gcc: configure --enable-static-link make STATIC_LD='-Wl,-Bstatic' LOCAL_LIBS='Wl,-Bdynamic -Wl,-R/etc/lib -ldl -Wl,-Bstatic' thor.ins.cwru.edu(2)$ ldd bash libdl.so.1 => /etc/lib/libdl.so.1 Make the analogous changes if you are running Sun's C Compiler. 12. Configuring bash to build it in a cross environment. Currently only two native versions can be compiled this way, cygwin32 and x86 BeOS. For BeOS, you would configure it like this: export RANLIB=i586-beos-ranlib export AR=i586-beos-ar export CC=i586-beos-gcc configure i586-beos Similarly for cygwin32. 13. Bash-2.05 has reverted to the bash-2.03 behavior of honoring the current locale setting when processing ranges within pattern matching bracket expressions ([A-Z]). This is what POSIX.2 and SUSv2 specify. The behavior of the matcher in bash-2.05 depends on the current LC_COLLATE setting. Setting this variable to `C' or `POSIX' will result in the traditional behavior ([A-Z] matches all uppercase ASCII characters). Many other locales, including the en_US locale (the default on many US versions of Linux) collate the upper and lower case letters like this: AaBb...Zz which means that [A-Z] matches every letter except `z'. The portable way to specify upper case letters is [:upper:] instead of A-Z; lower case may be specified as [:lower:] instead of a-z. Look at the manual pages for setlocale(3), strcoll(3), and, if it is present, locale(1). If you have locale(1), you can use it to find your current locale information even if you do not have any of the LC_ variables set. My advice is to put export LC_COLLATE=C into /etc/profile and inspect any shell scripts run from cron for constructs like [A-Z]. This will prevent things like rm [A-Z]* from removing every file in the current directory except those beginning with `z' and still allow individual users to change the collation order. Users may put the above command into their own profiles as well, of course.