test: run bash with --norc to avoid system bashrc

[tornaria]

When running bash with --rcfile it skips loading the user bashrc, but
it doesn't skip loading the system bashrc.

If bash completions is installed and loaded from system bashrc, this
causes an error:

FAIL: ERROR Unexpected output from bash command "BASH_COMPLETION_COMPAT_DIR=$(cd "$SRCDIR/.."; pwd)/completions"
FAIL: ERROR executing bash command "BASH_COMPLETION_COMPAT_DIR=$(cd "$SRCDIR/.."; pwd)/completions"

In order to avoid that, run bash with the --norc option which skips
loading any bashrc, and immediately source our config/bashrc.
It is important to set INPUTRC and PS1 before running bash.

[scop]

I cannot reproduce. I do have the (Fedora provided) bash-completion package installed and enabled, and it gets loaded on shell startup. When I cd to my bash-completion/test checkout dir and do ./runCompletion id.exp, it succeeds with no errors.

[tornaria]

Maybe you it is loading from your user ~/.bashrc, instead of the system bashrc. I don't have fedora, but it works that way in debian (and I don't see the issue there).

In void linux the bash-completion is loading from the system bashrc, thus the issue I described.

Note that the system bashrc has a different filename in different distros, in debian it is /etc/bash.bashrc, while in void is /etc/bash/bashrc.

Try adding

source /usr/share/bash-completion/bash_completion

to your system bashrc, to be able to reproduce the bug and to test the fix.

[scop]

In Fedora (and e.g. CentOS), the system bashrc /etc/bashrc gets loaded from users' ~/.bashrc, so the effect would be the same as loading bash-completion directly from the user one.

[tornaria]

Not so in debian or void linux. Indeed, bash has a configuration option SYS_BASHRC which is set to /etc/bash.bashrc (debian) or /etc/bash/bashrc (void), possibly undefined in fedora.

If defined, bash loads this file before loading the user ~/.bashrc.

When you use bash --rcfile <file>, this is loaded instead of the user bashrc, but the system bashrc is loaded nonetheless. In a standard debian system, the bash-completion script is loaded from the user bashrc, so this works. But on a standard void system, the bash-completion script is loaded from the system bashrc, thus the issue.

This is what happens when I run bash in interactive mode with --rcfile:

$ strace -e trace=open bash --rcfile /tmp/empty-bashrc-file -ic exit
open("/etc/ld.so.cache", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libdl.so.2", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libc.so.6", O_RDONLY|O_CLOEXEC) = 3
open("/dev/tty", O_RDWR|O_NONBLOCK)     = 3
open("/usr/lib/locale/locale-archive", O_RDONLY|O_CLOEXEC) = 3
open("/etc/nsswitch.conf", O_RDONLY|O_CLOEXEC) = 3
open("/etc/ld.so.cache", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libnss_files.so.2", O_RDONLY|O_CLOEXEC) = 3
open("/etc/passwd", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/gconv/gconv-modules.cache", O_RDONLY) = -1 ENOENT (No such file or directory)
open("/usr/lib/gconv/gconv-modules", O_RDONLY|O_CLOEXEC) = 3
open("/etc/bash/bashrc", O_RDONLY)      = 3
open("/etc/bash/bashrc.d/", O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC) = 3
open("/etc/bash/bashrc.d/bash_completion.sh", O_RDONLY) = 3
open("/usr/share/bash-completion/bash_completion", O_RDONLY) = 3
open("/home/tornaria/.bash_completion", O_RDONLY) = 3
open("/tmp/empty-bashrc-file", O_RDONLY) = 3
exit
+++ exited with 0 +++

You can see that the system bashrc is loaded, and the bash_completion script is loaded through that.

The only way to avoid loading the system bashrc is to pass the --norc flag. But then no rc file is loaded at all, even if one is specified by --rcfile:

$ strace -e trace=open bash --norc --rcfile /tmp/empty-bashrc-file -ic exit
open("/etc/ld.so.cache", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libdl.so.2", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libc.so.6", O_RDONLY|O_CLOEXEC) = 3
open("/dev/tty", O_RDWR|O_NONBLOCK)     = 3
open("/usr/lib/locale/locale-archive", O_RDONLY|O_CLOEXEC) = 3
open("/etc/nsswitch.conf", O_RDONLY|O_CLOEXEC) = 3
open("/etc/ld.so.cache", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libnss_files.so.2", O_RDONLY|O_CLOEXEC) = 3
open("/etc/passwd", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/gconv/gconv-modules.cache", O_RDONLY) = -1 ENOENT (No such file or directory)
open("/usr/lib/gconv/gconv-modules", O_RDONLY|O_CLOEXEC) = 3
exit
+++ exited with 0 +++

Hence the proposed patch, which runs bash --norc and then interactively sources the custom bashrc. In order for that to work, one needs to set PS1 and INPUTRC before running bash.

Can you check that running the tests with my patch works fine on fedora? I tested on debian and void linux and it works fine on both.

Otherwise, I cannot run the tests without modifying my system bashrc.

[scop]

I'll have a closer look a bit later, but in the meantime, it would be interesting to know why the error occurs. Your output in this issue only contains the error message, not its cause. The cause can be found in log/completion.log (if it was generated by a runCompletion call).

FYI I found these two external completion issues related to this in my system:
http://icedtea.classpath.org/bugzilla/show_bug.cgi?id=3365
sheepdog/sheepdog#388

[tornaria]

I'm sorry. The error is that BASH_COMPLETION_COMPAT_DIR is a readonly variable when bash-completions was already loaded:

...
/@BASH_COMPLETION_COMPAT_DIR=$(cd "$SRCDIR/.."; pwd)/completions^M
bash: BASH_COMPLETION_COMPAT_DIR: readonly variable^M
/@FAIL: ERROR Unexpected output from bash command "BASH_COMPLETION_COMPAT_DIR=$(cd "$SRCDIR/.."; pwd)/completions"
echo $?^M
1^M
/@FAIL: ERROR executing bash command "BASH_COMPLETION_COMPAT_DIR=$(cd "$SRCDIR/.."; pwd)/completions"
...

In fact, running the system bashrc may do other wrong things, I presume one really wants to run bash without any system configuration for the tests to avoid random issues.

[scop]

Could be, but on the other hand if the maintainers have defined a system bashrc that should always be loaded that early on shell init, it could be a sign that there's something essential in it which shouldn't be skipped...

[scop]

I'll still get back to this later, but I've removed the read only variable in c41a762

[tornaria]

That change doesn't fix the issue, because the system provided bash-completion is still declaring the variable to be read only (of course, this will fix itself with time).

I thought the point of running the tests without loading the rc file is to make sure it's running on a well know configuration to avoid any trouble like here.

[tornaria]

I have rebased the patch, and added a small fix: one needs to set TERM=dummy before invoking bash --norc, before by the time config/bashrc is sourced, it's to late (it's the same with INPUTRC).

[scop]

I know it doesn't solve the issue, that's why I'm still going to get back to this one later.

[tornaria]

Another small fix: we need to set stty columns to 150 before running bash, because after startup it won't take effect.

Now all the tests pass for me on debian and on void linux (except a very small number which give trouble on master as well: alias, cd, ssh)

[tornaria]

Rebased on top of 2.7 in 14b033b

[scop]

The fedoradev test setup had started to reproduce this once they added a system default-loaded bashrc. Merged along with some followup cleanup commits. Thanks!