This reverts commit a0aa7e0e17 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 a0aa7e0e17.
Commit a0aa7e0e17 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
(75630eba22: 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
b731dc7bfb ("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 a0aa7e0e17, 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>
b171466c44