Re: [PATCH] elf: Fix dlclose of an empty namespace in auditing mode (bug 26076)

[Florian Weimer <fweimer@redhat.com>]

* Carlos O'Donell:

>>>>  elf/dl-close.c | 2 +-
>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>>
>>>> diff --git a/elf/dl-close.c b/elf/dl-close.c
>>>> index 73b2817bbf..896e59e42e 100644
>>>> --- a/elf/dl-close.c
>>>> +++ b/elf/dl-close.c
>>>> @@ -782,7 +782,7 @@ _dl_close_worker (struct link_map *map, bool force)
>>>>      {
>>>>        struct link_map *head = ns->_ns_loaded;
>>>>        /* Do not call the functions for any auditing object.  */
>>>> -      if (head->l_auditing == 0)
>>>> +      if (head != NULL && head->l_auditing == 0)
>>>>  	{
>>>>  	  struct audit_ifaces *afct = GLRO(dl_audit);
>>>>  	  for (unsigned int cnt = 0; cnt < GLRO(dl_naudit); ++cnt)
>>>>
>>>
>>> Use _dl_signal_error to indicate an internal error?
>> 
>> This *is* on the cleanup path after an error already happened.  This is
>> the backtrace I saw:
>
> I'm going to ignore the oddity of auditors auditing themselves, I guess it's
> logically consistent that the first loaded auditor won't see any of this, but
> that subsequently loaded auditors will.
>
> Once auditor loading fails in your case we begin unloading that object. What I
> don't understand though is why is GLRO(dl_naudit) non-zero?
>
> If this is the first loaded auditor we don't increment GLRO(dl_naudit) until
> after we return from load_audit_module, so do_audit is false, and we never
> get here

Agreed.

> If this is the second loaded auditor we do set do_audit to true, but this
> means we already dereferenced ns->_ns_loaded->l_auditing already once to
> compute do_audit. What has subsequently happened to ns->_ns_loaded? For it
> to be null means we removed *every* object from the namespace, but we just
> established we had at least one already loaded successfully?

It looks like the namespace is not actually empty initially:

(gdb) print _rtld_global._dl_ns[9]
$10 = {_ns_loaded = 0x7ffff7ee7590, _ns_nloaded = 3, 
  _ns_main_searchlist = 0x0, _ns_global_scope_alloc = 0, 
  _ns_global_scope_pending_adds = 0, libc_map = 0x0, _ns_unique_sym_table = {
    lock = {mutex = {__data = {__lock = 0, __count = 0, __owner = 0, 
          __nusers = 0, __kind = 1, __spins = 0, __elision = 0, __list = {
            __prev = 0x0, __next = 0x0}}, 
        __size = '\000' <repeats 16 times>, "\001", '\000' <repeats 22 times>, 
        __align = 0}}, entries = 0x0, size = 0, n_elements = 0, free = 0x0}, 
  _ns_debug = {r_version = 1, r_map = 0x7ffff7ee7590, r_brk = 140737351924736, 
    r_state = RT_CONSISTENT, r_ldbase = 140737351860224}}

(gdb) print _rtld_global._dl_ns[9]._ns_loaded->l_name
$13 = 0x7ffff7ee7570 "./elf/tst-auditmanymod9.so"
(gdb) print _rtld_global._dl_ns[9]._ns_loaded->l_next->l_name
$14 = 0x7ffff7ee7b70 "./libc.so.6"
(gdb) print _rtld_global._dl_ns[9]._ns_loaded->l_next->l_next->l_name
$15 = 0x7ffff7ee8140 "./elf/ld-linux-x86-64.so.2"

So the commit message is wrong: the namespace doesn't start out as
empty, it becomes empty during dlclose.  It should be fairly easy to
write a test for this.  I'll post a new patch shortly.

Thanks,
Florian