toolchain-external: fix potential entire root filesystem removal

toolchain-external: fix potential entire root filesystem removal

This reverts commit a0aa7e0e1750f6ace2879ea8adb1425a41431b79 and reworks
the code to fix a major and potentially catastrophic bug when the
following conditions are met:

 - The user has selected a "known toolchain profile", such as a Linaro
   toolchain, a Sourcery CodeBench toolchain etc. People using "custom
   toolchain profile" are not affected.

 - The user has enabled BR2_TOOLCHAIN_EXTERNAL_PREINSTALLED=y to
   indicate that the toolchain is already locally available (as
   opposed to having Buildroot download and extract the toolchain)

 - The user has left BR2_TOOLCHAIN_EXTERNAL_PATH empty, because his
   toolchain is directly available through the PATH environment
   variable. When BR2_TOOLCHAIN_EXTERNAL_PATH is non-empty, Buildroot
   will do something silly (remove the toolchain contents), but that
   are limited to the toolchain itself.

When such conditions are met, Buildroot will run "rm -rf /*" due to
TOOLCHAIN_EXTERNAL_INSTALL_DIR being empty.

This bug does not exist in 2016.05, and appeared in 2016.08 due to
commit a0aa7e0e1750f6ace2879ea8adb1425a41431b79.

Commit a0aa7e0e1750f6ace2879ea8adb1425a41431b79 removed the assignment
of TOOLCHAIN_EXTERNAL_SOURCE and TOOLCHAIN_EXTERNAL_SITE to empty, as
part of a global cleanup to remove such assignments that supposedly
had become unneeded following a fix of the package infrastructure
(75630eba22b20d6140a5b58a6d1e35598fb3c0d3: core: do not attempt
downloads with no _VERSION set).

However, this causes TOOLCHAIN_EXTERNAL_SOURCE to be non-empty even
for BR2_TOOLCHAIN_EXTERNAL_PREINSTALLED=y configuration, with the
following consequences:

 - Buildroot downloads the toolchain tarball (while we're saying the
   toolchain is already available). Not dramatic, but clearly buggy.

 - Buildroot registers a post-extract hook that moves the toolchain
   from its extract directory (output/build/toolchain-external-.../ to
   its final location in host/opt/ext-toolchain/). Before doing this,
   it removes everything in TOOLCHAIN_EXTERNAL_INSTALL_DIR (which
   should normally be host/opt/ext-toolchain/).

Another mistake that caused the bug is commit
b731dc7bfb9c8ce7be502711f0b44ccab5515f1d ("toolchain-external: make
extraction idempotent"), which introduce the dangerous call "rm -rf
$(var)/*", which can be catastrophic if by mistake $(var) is
empty. Instead, this commit should have just used rm -rf $(var) to
remove the directory instead: it would have failed without consequences
if $(var) is empty, and the directory was anyway already re-created
right after with a mkdir.

To address this problem, we:

 - Revert commit a0aa7e0e1750f6ace2879ea8adb1425a41431b79, so that
   _SOURCE and _SITE are empty in the pre-installed toolchain case.

 - Rework the code to ensure that similar problems will no happen in the
   future, by:

   - Registering the TOOLCHAIN_EXTERNAL_MOVE hook only when
     BR2_TOOLCHAIN_EXTERNAL_DOWNLOAD=y, since moving the toolchain is
     only needed when Buildroot downloaded the toolchain.

   - Introduce a variable TOOLCHAIN_EXTERNAL_DOWNLOAD_INSTALL_DIR which
     is the path in which Buildroot installs external toolchains when it
     is in charge of downloading/extracting them. Then, the
     TOOLCHAIN_EXTERNAL_MOVE hook is changed to use this variable, which
     is guaranteed to be non-empty.

   - Replace the removal of the directory contents $(var)/* by removing
     the directory itself $(var). The directory was anyway already
     re-created if needed afterwards. Thanks to doing this, if $(var)
     ever becomes empty, we will do "rm -rf" which will fail and abort
     the build, and not the catastrophic "rm -rf /*".

Reported-by: Mason <slash.tmp@free.fr>
Cc: Mason <slash.tmp@free.fr>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>