Re: [PATCH] test-container: Use nftw instead of rm -rf

[Adhemerval Zanella Netto <adhemerval.zanella@linaro.org>]

On 28/09/23 06:40, Stefan Liebler wrote:
> On 27.09.23 21:20, Adhemerval Zanella wrote:
>> If the binary to run is 'env', test-containers skips it and adds
>> any required environment variable on the process envs variables.
>> This simplifies the required code to spawn new process (no need
>> to build an env-like program).
>>
>> However, this is an issue for recursive_remove if there is any
>> LD_PRELOAD, since test-container will not prepend the loader command
>> along with required paths.  If the required preloaded library can
>> not be loaded by the system glibc, the 'post-clean rsync' will
>> eventually fail.
>>
>> One example is if system glibc does not support DT_RELR and the
>> built glibc does, the nss/tst-nss-gai-hv2-canonname test fails
>> with:
>>
>> ../scripts/evaluate-test.sh nss/tst-nss-gai-hv2-canonname $? false false
>> 86_64-linux-gnu/nss/tst-nss-gai-hv2-canonname.test-result
>> rm: /lib/x86_64-linux-gnu/libc.so.6: version `GLIBC_ABI_DT_RELR' not
>> found (required by x86_64-linux-gnu/malloc/libc_malloc_debug.so)
>>
>> Instead trying to figure out the required loader arguments on how
>> to spawn the 'rm -rf', replace the command with a nftw call.
>>
> Just as information, I've also recognized this issue. For me (on
> x86_64/s390x), I got "*** stack smashing detected ***: terminated" in
> this rm due to preloading with the fresh build libc_malloc_debug.so.
> In malloc_hook_ini -> generic_hook_ini -> initialize_malloc_check, we have:
> TUNABLE_GET (check, int32_t, TUNABLE_CALLBACK (set_mallopt_check));
> In __tunable_get_val, the tunables ID from current-build and system for
> glibc.malloc.check differs. In my case it is 30, which is beyond the
> range of the systems tunable_list[]. This leads to storing a 8byte value
> to passed valp. Unfortunately it points to a 4 byte int32_t and the
> other 4 bytes belong to the stack smashing canary.

Indeed the tunables might another problem, thanks for checking it.

> 
> Your patch solves the issue for me.

Is it a reviewed-by ;) ?

>> Checked on x86_64-linux-gnu.
>> ---
>>  support/test-container.c | 34 +++++++++++-----------------------
>>  1 file changed, 11 insertions(+), 23 deletions(-)
>>
>> diff --git a/support/test-container.c b/support/test-container.c
>> index 788b091ea0..95dfef1a99 100644
>> --- a/support/test-container.c
>> +++ b/support/test-container.c
>> @@ -37,6 +37,7 @@
>>  #include <errno.h>
>>  #include <error.h>
>>  #include <libc-pointer-arith.h>
>> +#include <ftw.h>
>>
>>  #ifdef __linux__
>>  #include <sys/mount.h>
>> @@ -405,32 +406,19 @@ file_exists (char *path)
>>    return 0;
>>  }
>>
>> +static int
>> +unlink_cb (const char *fpath, const struct stat *sb, int typeflag,
>> +	   struct FTW *ftwbuf)
>> +{
>> +  return remove (fpath);
>> +}
>> +
>>  static void
>>  recursive_remove (char *path)
>>  {
>> -  pid_t child;
>> -  int status;
>> -
>> -  child = fork ();
>> -
>> -  switch (child) {
>> -  case -1:
>> -    perror("fork");
>> -    FAIL_EXIT1 ("Unable to fork");
>> -  case 0:
>> -    /* Child.  */
>> -    execlp ("rm", "rm", "-rf", path, NULL);
>> -    FAIL_EXIT1 ("exec rm: %m");
>> -  default:
>> -    /* Parent.  */
>> -    waitpid (child, &status, 0);
>> -    /* "rm" would have already printed a suitable error message.  */
>> -    if (! WIFEXITED (status)
>> -	|| WEXITSTATUS (status) != 0)
>> -      FAIL_EXIT1 ("exec child returned status: %d", status);
>> -
>> -    break;
>> -  }
>> +  int r = nftw (path, unlink_cb, 1000, FTW_DEPTH | FTW_PHYS);
>> +  if (r == -1)
>> +    FAIL_EXIT1 ("recursive_remove failed");
>>  }
>>
>>  /* Used for both rsync and the mytest.script "cp" command.  */
>