We use the filter builtin function of make to retrieve the right
shell to use during the CHROOT phase, which is the same as the
one in CHROOT1.
Also do not assume envars exist when removing it at the beginning
of the SUDO phase.
This involves changing the way make targets are built. Specially
the make target is changed when encountering a specific file
(for example the make target becomes sudo when rencountering kernfs)
This allows more flexibility.
Those scriptlets need to use a temporary bash, which is in /tools/bin
for current LFS. But if we use a new method (cross-compiling + sysroot)
it ends into /bin. So we have to set the correct `#!' header
in both cases.
We rely on the pesence of a "creatingtoolsdir" sect1 to know that
bash is in /tools
- take targets from scriptlet names
- neutralize some commands or bash builtins so that they are not
used even if they are in the scriptlets, if not needed
- do not set PREV in chapter 5 anymore
- Improve save and restore user, so that if the LUSER exists,
its home is saved while building temporary tools, and restored
afterwards
Using the chapter ancestor is not enough. It's better to use the sect1 id
and test that it starts with ch-tools: package-manager does not have this
ancestor. Also the virtual kernel filesystems are done before chroot,
so directly test this one.
This works for the current lfs book. We'll need something new if we
change the order of pages.
In the generated scriptlets, we need to know if we are on host
or chrooted. Presently, we
rely on /tools being a symlink or not. But we may want to remove this
ugly symlink in future versions of LFS, so anticipate and use the
chapter ancestor.
This is a major change for LFS. We won't be using the makefile for
the final preparations, but we'll extract the book instructions and
use them (the "use part not in this commit).
Since getting figures for those scriptlets is not important, and we
cannot use start and end scripts, which need already set up user and
dirs, do not use them
If we stop a build before the "do-housekeeping" target is executed, there
may be symlinks of the type /tools->/mnt/lfs/tools. If we want to restart,
for example after fixing something else, and we run the "creatingtoolsdir"
target (or similar one in CLFS) again, it will fail. It's better to clean
those symlinks when cleaning the build directory.
Most books have something like "exec /bin/bash" in .bash_profile for the
building user. This results in stopping and waiting for commands when
running a a login shell. The current jhalfs way of removing .bash_profile is
suboptimal, since a user may want to login as $(LUSER) and expect that the
environment is set (without having to source .bashrc). So remove the
"-i" flag in "sudo -u $(LUSER) ..."
This is to allow root to execute ": > envars". On a non empty envars
owned by lfs, in a directory world writable with sticky bit and not owned
by lfs, this may be forbiden by the fs/protected_regular sysctl.
See https://www.kernel.org/doc/Documentation/sysctl/fs.txt
Normally only chapters containing versioned pacakges are in the package list.
But with the removal of lsb-release there are no versioned package anymore
in the "After LFS config..." chapter. But we need this chapter. So add
a test explicitly for it.
- refactor again process-install. Allows for better generation of test
instructions
- only output the instructions of the first sect2 with role="installation".
this will miss the vaapi intel driver, the installation of help files in
GIMP, and the second method for which
- always run porg logging in "append" mode
- allows pass1 in sect2 (libva again)
The "create-sbu_du.sh" script expects all the logs containing
"Totalseconds" also contains two lines starting with "KB:" with disk usage
stats. Generates those lines for all the logs. Hopefully, the stats are
fairly accurate now...
In create-sbu_du.sh, we use "du -s" for getting the size of the installed
files at the end of the build, and we compare to the "KB" from the previous
package to get the Installed size of the last package. But the du command
excludes an absolute dir, while it should be relative. This lead to bogus
stats.
This is another bug exposed by the new menu system. The TEST variable was set
inside the menu "test settings". But this menu depends on CONFIG_TESTS, and
is not executed if CONFIG_TESTS is not set. So that, the TEST variable was
not set if CONFIG_TESTS was not set (while with the former menu suystem,
it was set to 0). Due to the logic in lfs.xsl, this resulted in all the
tests being run, instead of none...
Move the setting of the TEST variable outside the "test settings" menu.
