musl 1.1.19 exposes bug in GNU make

musl 1.1.19 exposes bug in GNU make

[Rich Felker]

musl 1.1.19 introduced the _DIRENT_HAVE_D_* macros in commit
1bc10ffeaa7c7ce44b3e214e02e302642511c7c7, triggering a bug in GNU make
that's present up to the current version:

http://git.savannah.gnu.org/cgit/make.git/tree/glob/glob.c?id=5d653b535a0e0e607c8be2016bcaadd6f326a698#n1333

The glob replacement they ship includes and uses internally a
GLOB_ONLYDIR flag, which is opportunistic and implemented via d_type
if _DIRENT_HAVE_D_TYPE is defined, and it fails to consider that
DT_LNK could be a symlink to a directory. This is an ancient version
of glob.c; the modern GNU glob in glibc rightly considers DT_LNK. Thus
the bug is not seen on glibc systems (where the glob replacement is
not used), but will be seen on any system where _DIRENT_HAVE_D_TYPE is
defined and where configure decides to replace the libc glob with
their GNU one.

A patch something like this should work:

-                      && d->d_type != DT_UNKNOWN && d->d_type != DT_DIR)
+                      && d->d_type != DT_UNKNOWN && d->d_type != DT_DIR && d->d_type != DT_LNK)

Or ideally configure could be convinced not to replace glob if they
don't actually depend on weird gnu extensions.

The bug manifests as broken $(wildcard ...) expansion whenever
symlinks to dirs are involved in thr glob expression.

Rich

Re: musl 1.1.19 exposes bug in GNU make

[Rich Felker]

On Tue, Mar 13, 2018 at 01:18:51PM -0400, Rich Felker wrote:
> musl 1.1.19 introduced the _DIRENT_HAVE_D_* macros in commit
> 1bc10ffeaa7c7ce44b3e214e02e302642511c7c7, triggering a bug in GNU make
> that's present up to the current version:
> 
> http://git.savannah.gnu.org/cgit/make.git/tree/glob/glob.c?id=5d653b535a0e0e607c8be2016bcaadd6f326a698#n1333
> 
> The glob replacement they ship includes and uses internally a
> GLOB_ONLYDIR flag, which is opportunistic and implemented via d_type
> if _DIRENT_HAVE_D_TYPE is defined, and it fails to consider that
> DT_LNK could be a symlink to a directory. This is an ancient version
> of glob.c; the modern GNU glob in glibc rightly considers DT_LNK. Thus
> the bug is not seen on glibc systems (where the glob replacement is
> not used), but will be seen on any system where _DIRENT_HAVE_D_TYPE is
> defined and where configure decides to replace the libc glob with
> their GNU one.
> 
> A patch something like this should work:
> 
> -                      && d->d_type != DT_UNKNOWN && d->d_type != DT_DIR)
> +                      && d->d_type != DT_UNKNOWN && d->d_type != DT_DIR && d->d_type != DT_LNK)
> 
> Or ideally configure could be convinced not to replace glob if they
> don't actually depend on weird gnu extensions.
> 
> The bug manifests as broken $(wildcard ...) expansion whenever
> symlinks to dirs are involved in thr glob expression.

Filed upstream as:

https://savannah.gnu.org/bugs/index.php?53465

Rich