[Buildroot] [PATCH v7 03/18] core: re-enter make if $(CURDIR) or $(O) are not absolute canonical path

[Samuel Martin]

On Wed, Mar 23, 2016 at 2:36 AM, Matthew Weber <matt at thewebers.ws> wrote:
> Yann, Samuel
>
> On Tue, Mar 22, 2016 at 7:54 PM, Matthew Weber <matt at thewebers.ws> wrote:
>> Yann, Samuel, All
>>
>> On Tue, Mar 22, 2016 at 6:11 PM, Matthew Weber <matt at thewebers.ws> wrote:
>>> Yann, All
>>>
>>> On Tue, Mar 22, 2016 at 5:07 PM, Yann E. MORIN <yann.morin.1998 at free.fr> wrote:
>>>> Matthew, All,
>>>>
>>>> On 2016-03-20 20:45 -0500, Matthew Weber spake thusly:
>>>>> On Sun, Mar 20, 2016 at 8:44 PM, Matthew Weber <matt at thewebers.ws> wrote:
>>>>> > On Wed, Mar 9, 2016 at 4:58 PM, Samuel Martin <s.martin49 at gmail.com> wrote:
>>>>> >> When $(CURDIR) or $(O) contain symlinks (or mount-bind) in their path,
>>>>> >> they can be resolved differently, depending on each package build-system
>>>>> >> (whether it uses the given paths or get the absolute canonical ones).
>>>>> >>
>>>>> >> Thus, to make easier tracking down host machine paths leaking into the
>>>>> >> host, target or staging trees, the CURDIR and O variables are set to
>>>>> >> their absolute canonical paths.
>>>>> >>
>>>>> >> Note that this change takes care of the makefile wrapper installed in
>>>>> >> $(O) to avoid unneeded make recursion.
>>>>> >>
>>>>> >> Signed-off-by: Samuel Martin <s.martin49 at gmail.com>
>>>>> >>
>>>>> >> ---
>>>>> >> changes v6->v7:
>>>>> >> - none
>>>>> >>
>>>>> >> changes v5->v6:
>>>>> >> - new patch
>>>>> >> ---
>>>>> >>  Makefile | 22 ++++++++++++++++++----
>>>>> >>  1 file changed, 18 insertions(+), 4 deletions(-)
>>>>> >>
>>>>> >> diff --git a/Makefile b/Makefile
>>>>> >> index 98c8dc7..52a92f6 100644
>>>>> >> --- a/Makefile
>>>>> >> +++ b/Makefile
>>>>> >> @@ -26,16 +26,30 @@
>>>>> >>
>>>>> >>  # Trick for always running with a fixed umask
>>>>> >>  UMASK = 0022
>>>>> >> -ifneq ($(shell umask),$(UMASK))
>>>>> >> +
>>>>> >> +# Check if we need to re-enter make for one or several of the following reasons:
>>>>> >> +# 1- Wrong (too restrictive) umask:
>>>>> >> +#    This prevents Buildroot and packages from creating files and directories.
>>>>> >> +# 2- CWD (i.e. $(CURDIR)) not being the absolute canonical path:
>>>>> >> +#    This makes harder tracking and fixing host machine path leaks.
>>>>> >> +# 3- Output location (i.e. $(O)) not being the absolute canonical path:
>>>>> >> +#    This makes harder tracking and fixing host machine path leaks.
>>>>> >> +#
>>>>> >> +# Note:
>>>>> >> +# - remove the trailing '/.' from $(O) as it can be added by the makefile
>>>>> >> +#   wrapper installed in the $(O).
>>>>> >> +ifneq ($(shell umask):$(CURDIR):$(patsubst %/.,%,$(O)),$(UMASK):$(realpath $(CURDIR)):$(realpath $(O)))
>>>>> >>  .PHONY: _all $(MAKECMDGOALS)
>>>>> >>
>>>>> >>  $(MAKECMDGOALS): _all
>>>>> >>         @:
>>>>> >>
>>>>> >>  _all:
>>>>> >> -       @umask $(UMASK) && $(MAKE) --no-print-directory $(MAKECMDGOALS)
>>>>> >> +       umask $(UMASK) && \
>>>>> >> +               $(MAKE) -C $(realpath $(CURDIR)) --no-print-directory \
>>>>> >> +                       $(MAKECMDGOALS) O=$(realpath $(O))
>>>>> >
>>>>> > I've applied your v7 patches with these changes to the tip of
>>>>> > buildroot and currently run into the following error.  I'm still
>>>>> > poking around trying to understand why.
>>>>> >
>>>>> > $ make V=1 qemu_x86_defconfig
>>>>> > umask 0022 && \
>>>>> >         make -C /opt/buildroot --no-print-directory \
>>>>> >                 qemu_x86_defconfig O=
>>>>
>>>> Why are you opassing an empty O= option here?
>>>>
>>>
>>> I'm not passing the O= .
>>>
>>> However, if I modify the patch and remove that from the target in the
>>> Makefile ie. remove the "O=$(realpath $(O))" , things build ok.
>>
>> I tested this tonight on another server with Make version 3.81 and
>> reproduced with a fresh buildroot checkout
>> (f34c69da2bfa2a485671dbbcf1ee24ad7409b63b) plus applying the commit
>> mentioned as the latest from Samuel's fork.
>>
>> Same behavior, if I edit the Makefile and remove the "O=$(realpath
>> $(O))" , the defconfig sets up as expected
>
> One additional detail, I cloned https://github.com/tSed/buildroot.git
> and checkout c55eeee30fe6106354c0f2cb6ce6ecb46972503a.  Then "make
> qemu_x86_defconfig" and reproduced it.  I'll start dissecting the
> other commits past that one.  Something must be masking it.
>

I think I get what happens:

According to [2,3], realpath return an empty string in case of ENOENT
(file does not exist), hence the empty $(O) because output/ is not yet
created when re-entering make.

However, I'm still not sure how to fix this, either pass O=$(O)
instead of O=$(realpath $(O)) when re-entering make, or mkdir $(O)
before re-entering make. (I still need to think/test about it).
Any suggestions more than welcome.

Note: even after downgrading make to make-3.81, I still don't trigger
this issue. I figured this out using this tricks [4]... :-/

[2] https://www.gnu.org/software/make/manual/html_node/File-Name-Functions.html
[3] http://man7.org/linux/man-pages/man3/realpath.3.html
[4] http://code.bulix.org/epvj9d-94510?raw

Regards,

> --
> Thanks,
> Matt

-- 
Samuel