Re: [dpdk-dev] [PATCH 2/2] devtools: add path to additional shared object files

[David Marchand <david.marchand@redhat.com>]

On Wed, Dec 18, 2019 at 2:43 PM Laatz, Kevin <kevin.laatz@intel.com> wrote:
>
> On 18/12/2019 08:23, David Marchand wrote:
> > On Wed, Dec 18, 2019 at 6:39 AM Ruifeng Wang <ruifeng.wang@arm.com> wrote:
> >> librte_mempool_ring.so and librte_pmd_null.so are in 'drivers' folder.
> >> Add 'drivers' into LD_LIBRARY_PATH so that testpmd can find and make
> >> use of these shared libraries.
> >>
> >> Signed-off-by: Ruifeng Wang <ruifeng.wang@arm.com>
> >> Reviewed-by: Gavin Hu <gavin.hu@arm.com>
> >> ---
> >>   devtools/test-null.sh | 2 +-
> >>   1 file changed, 1 insertion(+), 1 deletion(-)
> >>
> >> diff --git a/devtools/test-null.sh b/devtools/test-null.sh
> >> index f39af2c06..548de8113 100755
> >> --- a/devtools/test-null.sh
> >> +++ b/devtools/test-null.sh
> >> @@ -20,7 +20,7 @@ if [ ! -f "$testpmd" ] ; then
> >>   fi
> >>
> >>   if ldd $testpmd | grep -q librte_ ; then
> >> -       export LD_LIBRARY_PATH=$build/lib:$LD_LIBRARY_PATH
> >> +       export LD_LIBRARY_PATH=$build/drivers:$build/lib:$LD_LIBRARY_PATH
> >>          libs='-d librte_mempool_ring.so -d librte_pmd_null.so'
> >>   else
> >>          libs=
> >> --
> >> 2.17.1
> >>
> > I'm surprised to see this.
> > So far, the CI ran fine without it, so something is different in the
> > environment.
> >
> > I can see that the RPATH entry disappeared from the testpmd binary.
> > Xenial:
> > # readelf -d build/app/dpdk-testpmd |grep RPATH
> >   0x000000000000000f (RPATH)              Library rpath:
> > [$ORIGIN/../lib:$ORIGIN/../drivers:XXXXXXXXXXXXX]
> >
> > (not sure what XXXX purpose is, but different topic)
> >
> > Bionic:
> > # readelf -d build/app/dpdk-testpmd |grep RPATH
>
> It looks like RPATH just changed to be named RUNPATH in Bionic:
>
> # readelf -d build/app/dpdk-testpmd | grep R.*PATH
>   0x000000000000001d (RUNPATH)            Library runpath:
> [$ORIGIN/../lib:$ORIGIN/../drivers:XXXXXXXXXXXXX]

Did some experiment with some test program and .so of mine.
TL;DR, if I understand correctly, RPATH on the binary applies to all
lookups, even in a subsequent .so code.
But RUNPATH only applies to the current ELF, meaning that the dlopen()
in my intermediate .so does not get it.

dlopen() is called from librte_eal.so, and RUNPATH on testpmd is not enough.


Details:
[dmarchan@dmarchan plop]$ cat main.c
extern void loader(void);

int main(int argc, char *argv[])
{
    loader();
    return 0;
}
[dmarchan@dmarchan plop]$ cat loader/loader.c
#include <dlfcn.h>
#include <stdio.h>

void loader(void)
{
    if (dlopen("lib.so", RTLD_NOW) == NULL)
        fprintf(stderr, "%s\n", dlerror());
}

[dmarchan@dmarchan plop]$ cat lib/lib.c
#include <stdio.h>

__attribute__((constructor))
static void plop(void)
{
    fprintf(stdout, "plop\n");
}


# no rpath/runpath
[dmarchan@dmarchan plop]$ gcc -o lib/lib.so -Wall -Werror -shared
-fPIC lib/lib.c
[dmarchan@dmarchan plop]$ gcc -o loader/loader.so -Wall -Werror
-shared -fPIC loader/loader.c -ldl
[dmarchan@dmarchan plop]$ gcc -o main -Wall -Werror main.c loader/loader.so
[dmarchan@dmarchan plop]$ readelf -d main |grep R.*PATH
[dmarchan@dmarchan plop]$ ./main
lib.so: cannot open shared object file: No such file or directory

# using rpath on final binary
[dmarchan@dmarchan plop]$ gcc -o main -Wall -Werror main.c
loader/loader.so -Wl,-rpath,loader:lib
[dmarchan@dmarchan plop]$ readelf -d main |grep R.*PATH
 0x000000000000000f (RPATH)              Library rpath: [loader:lib]
[dmarchan@dmarchan plop]$ ./main
plop

# using runpath on final binary
[dmarchan@dmarchan plop]$ gcc -o main -Wall -Werror main.c
loader/loader.so -Wl,-enable-new-dtag,-rpath,loader:lib
[dmarchan@dmarchan plop]$ readelf -d main |grep R.*PATH
 0x000000000000001d (RUNPATH)            Library runpath: [loader:lib]
[dmarchan@dmarchan plop]$ ./main
lib.so: cannot open shared object file: No such file or directory

# using runpath on loader
[dmarchan@dmarchan plop]$ gcc -o loader/loader.so -Wall -Werror
-shared -fPIC loader/loader.c -ldl -Wl,-enable-new-dtag,-rpath,lib
[dmarchan@dmarchan plop]$ readelf -d loader/loader.so |grep R.*PATH
 0x000000000000001d (RUNPATH)            Library runpath: [lib]
[dmarchan@dmarchan plop]$ gcc -o main -Wall -Werror main.c
loader/loader.so -Wl,-enable-new-dtag,-rpath,loader
[dmarchan@dmarchan plop]$ readelf -d main |grep R.*PATH
 0x000000000000001d (RUNPATH)            Library runpath: [loader]
[dmarchan@dmarchan plop]$ ./main
plop


> > Adding Bruce and Kevin, as I think this is the same issue than:
> > http://mails.dpdk.org/archives/dev/2019-December/153627.html
> > Could it be a change in meson?
>
> Yes, looks like the same error to me.
>
> I'm not sure this is solely a meson issue, I often need to pass -d
> librte_mempool_ring.so when using make builds too. Maybe I'm missing
> something here?

In a make build directory, all libraries ends up in the same lib/
directory and the test-null.sh script work with the current
LD_LIBRARY_PATH.


Ruifeng patch seems valid to me, but I would love some explanations in
the commitlog.

--
David Marchand