diff mbox series

elf: Fix DTV gap reuse logic (BZ #27135)

Message ID 20210628180358.235038-1-adhemerval.zanella@linaro.org
State Superseded
Headers show
Series elf: Fix DTV gap reuse logic (BZ #27135) | expand

Checks

Context Check Description
dj/TryBot-apply_patch success Patch applied to master at the time it was sent
dj/TryBot-32bit success Build for i686

Commit Message

Adhemerval Zanella June 28, 2021, 6:03 p.m. UTC
This is updated version of the 572bd547d57a (reverted by 40ebfd016ad2)
that fixes the _dl_next_tls_modid issues.

This issue with 572bd547d57a patch is the DTV entry will be only
update on dl_open_worker() with the update_tls_slotinfo() call after
all dependencies are being processed by _dl_map_object_deps().  However
_dl_map_object_deps() itself might call _dl_next_tls_modid(), and since
the _dl_tls_dtv_slotinfo_list::map is not yet set the entry will be
wrongly reused.  This patch fixes by setting the 'map' member to a
intermediary value, so subsequente _dl_next_tls_modid() call will see
the entry as allocated.

The intermediary value is cleared up on remove_slotinfo() for the case
a library fails to load with RTLD_NOW.

Checked on x86_64-linux-gnu.
---
 elf/Makefile    |  64 +++++++++++-
 elf/dl-close.c  |   8 +-
 elf/dl-open.c   |  10 --
 elf/dl-tls.c    |  11 +-
 elf/tst-tls20.c | 268 ++++++++++++++++++++++++++++++++++++++++++++++--
 5 files changed, 336 insertions(+), 25 deletions(-)

Comments

Florian Weimer July 8, 2021, 6 a.m. UTC | #1
* Adhemerval Zanella:

> diff --git a/elf/dl-tls.c b/elf/dl-tls.c
> index 2b5161d10a..4ec4c7f38c 100644
> --- a/elf/dl-tls.c
> +++ b/elf/dl-tls.c
> @@ -157,7 +157,11 @@ _dl_next_tls_modid (void)
>  	      }
>  
>  	    if (result - disp < runp->len)
> -	      break;
> +	      {
> +		/* Mark the entry as used, so any dependency see it.  */
> +		runp->slotinfo[result - disp].map = (struct link_map *) -1;
> +		break;
> +	      }
>  
>  	    disp += runp->len;
>  	  }

Which dependency?

I think the special value -1 needs to be documented on the slotinfo
struct member.

When Carlos and I discussed this, we couldn't quite decide whether it
was appropriate just to assign the actual link map value here.

Thanks,
Florian
Szabolcs Nagy July 8, 2021, 11 a.m. UTC | #2
The 07/08/2021 08:00, Florian Weimer wrote:
> * Adhemerval Zanella:
> > diff --git a/elf/dl-tls.c b/elf/dl-tls.c
> > index 2b5161d10a..4ec4c7f38c 100644
> > --- a/elf/dl-tls.c
> > +++ b/elf/dl-tls.c
> > @@ -157,7 +157,11 @@ _dl_next_tls_modid (void)
> >  	      }
> >  
> >  	    if (result - disp < runp->len)
> > -	      break;
> > +	      {
> > +		/* Mark the entry as used, so any dependency see it.  */
> > +		runp->slotinfo[result - disp].map = (struct link_map *) -1;
> > +		break;
> > +	      }
> >  
> >  	    disp += runp->len;
> >  	  }
> 
> Which dependency?
> 
> I think the special value -1 needs to be documented on the slotinfo
> struct member.
> 
> When Carlos and I discussed this, we couldn't quite decide whether it
> was appropriate just to assign the actual link map value here.

slotinfo[].map is only accessed without locks during
_dl_update_slotinfo at tls access and then it is
only used to get the map for the module in which
tls is accessed. (and to skip NULL maps when resizing
dtv but that's just a minor optimization).

so the only concurrent .map access does not really
care about its value for an incomplete module.

so we only need to reason locally what happens within
one dlopen call to decide what can be put temporarily
in .map and the final value only has to be stored
before the lock is released. (with bug 19924 fixed
the final .map and .gen values likely have to be
stored a bit earlier: before the global gen count is
updated)

the tricky bit is to check if the slotinfo state is
rolled back right in case of dlopen failures. but it
seems to me that modid assignment happens after the
point where failures are handled via dlclose which
will fixup the slotinfo state (i.e. after unlock
unused slotinfo entries are marked with .map == NULL).

so i think _dl_next_tls_modid() can be changed to
_dl_assign_tls_modid(map) which writes map to the
slotinfo .map.
Adhemerval Zanella July 8, 2021, 12:20 p.m. UTC | #3
On 08/07/2021 08:00, Szabolcs Nagy wrote:
> The 07/08/2021 08:00, Florian Weimer wrote:
>> * Adhemerval Zanella:
>>> diff --git a/elf/dl-tls.c b/elf/dl-tls.c
>>> index 2b5161d10a..4ec4c7f38c 100644
>>> --- a/elf/dl-tls.c
>>> +++ b/elf/dl-tls.c
>>> @@ -157,7 +157,11 @@ _dl_next_tls_modid (void)
>>>  	      }
>>>  
>>>  	    if (result - disp < runp->len)
>>> -	      break;
>>> +	      {
>>> +		/* Mark the entry as used, so any dependency see it.  */
>>> +		runp->slotinfo[result - disp].map = (struct link_map *) -1;
>>> +		break;
>>> +	      }
>>>  
>>>  	    disp += runp->len;
>>>  	  }
>>
>> Which dependency?

Any shared library dependency being loaded after this.

>>
>> I think the special value -1 needs to be documented on the slotinfo
>> struct member.
>>
>> When Carlos and I discussed this, we couldn't quite decide whether it
>> was appropriate just to assign the actual link map value here.
> 
> slotinfo[].map is only accessed without locks during
> _dl_update_slotinfo at tls access and then it is
> only used to get the map for the module in which
> tls is accessed. (and to skip NULL maps when resizing
> dtv but that's just a minor optimization).
> 
> so the only concurrent .map access does not really
> care about its value for an incomplete module.
> 
> so we only need to reason locally what happens within
> one dlopen call to decide what can be put temporarily
> in .map and the final value only has to be stored
> before the lock is released. (with bug 19924 fixed
> the final .map and .gen values likely have to be
> stored a bit earlier: before the global gen count is
> updated)
> 
> the tricky bit is to check if the slotinfo state is
> rolled back right in case of dlopen failures. but it
> seems to me that modid assignment happens after the
> point where failures are handled via dlclose which
> will fixup the slotinfo state (i.e. after unlock
> unused slotinfo entries are marked with .map == NULL).

That is my understanding as well and I tried to check if the
rollback is correctly done by adding more cases on the
tst-tls20.c. 

> 
> so i think _dl_next_tls_modid() can be changed to
> _dl_assign_tls_modid(map) which writes map to the
> slotinfo .map.
What do you mean by '_dl_assign_tls_modid()' in this case ?
Szabolcs Nagy July 8, 2021, 1:04 p.m. UTC | #4
The 07/08/2021 09:20, Adhemerval Zanella wrote:
> 
> 
> On 08/07/2021 08:00, Szabolcs Nagy wrote:
> > The 07/08/2021 08:00, Florian Weimer wrote:
> >> * Adhemerval Zanella:
> >>> diff --git a/elf/dl-tls.c b/elf/dl-tls.c
> >>> index 2b5161d10a..4ec4c7f38c 100644
> >>> --- a/elf/dl-tls.c
> >>> +++ b/elf/dl-tls.c
> >>> @@ -157,7 +157,11 @@ _dl_next_tls_modid (void)
> >>>  	      }
> >>>  
> >>>  	    if (result - disp < runp->len)
> >>> -	      break;
> >>> +	      {
> >>> +		/* Mark the entry as used, so any dependency see it.  */
> >>> +		runp->slotinfo[result - disp].map = (struct link_map *) -1;
> >>> +		break;
> >>> +	      }
> >>>  
> >>>  	    disp += runp->len;
> >>>  	  }
> >>
> >> Which dependency?
> 
> Any shared library dependency being loaded after this.
> 
> >>
> >> I think the special value -1 needs to be documented on the slotinfo
> >> struct member.
> >>
> >> When Carlos and I discussed this, we couldn't quite decide whether it
> >> was appropriate just to assign the actual link map value here.
> > 
> > slotinfo[].map is only accessed without locks during
> > _dl_update_slotinfo at tls access and then it is
> > only used to get the map for the module in which
> > tls is accessed. (and to skip NULL maps when resizing
> > dtv but that's just a minor optimization).
> > 
> > so the only concurrent .map access does not really
> > care about its value for an incomplete module.
> > 
> > so we only need to reason locally what happens within
> > one dlopen call to decide what can be put temporarily
> > in .map and the final value only has to be stored
> > before the lock is released. (with bug 19924 fixed
> > the final .map and .gen values likely have to be
> > stored a bit earlier: before the global gen count is
> > updated)
> > 
> > the tricky bit is to check if the slotinfo state is
> > rolled back right in case of dlopen failures. but it
> > seems to me that modid assignment happens after the
> > point where failures are handled via dlclose which
> > will fixup the slotinfo state (i.e. after unlock
> > unused slotinfo entries are marked with .map == NULL).
> 
> That is my understanding as well and I tried to check if the
> rollback is correctly done by adding more cases on the
> tst-tls20.c. 
> 
> > 
> > so i think _dl_next_tls_modid() can be changed to
> > _dl_assign_tls_modid(map) which writes map to the
> > slotinfo .map.
> What do you mean by '_dl_assign_tls_modid()' in this case ?

currently the only use is

 map->l_tls_modid = _dl_next_tls_modid ();

which can be changed to

 _dl_assign_tls_modid (map);

which sets map->l_tls_modid as well as slotinfo[i].map = map
marking the slotinfo entry non-free (with relaxed atomic
store since it can be read concurrently).

i don't know if 'assign' is the right word, but the point is
that we don't just get the next modid, but allocate it too
for the given map.
Adhemerval Zanella July 8, 2021, 1:34 p.m. UTC | #5
On 08/07/2021 10:04, Szabolcs Nagy wrote:
> The 07/08/2021 09:20, Adhemerval Zanella wrote:
>>
>>
>> On 08/07/2021 08:00, Szabolcs Nagy wrote:
>>> The 07/08/2021 08:00, Florian Weimer wrote:
>>>> * Adhemerval Zanella:
>>>>> diff --git a/elf/dl-tls.c b/elf/dl-tls.c
>>>>> index 2b5161d10a..4ec4c7f38c 100644
>>>>> --- a/elf/dl-tls.c
>>>>> +++ b/elf/dl-tls.c
>>>>> @@ -157,7 +157,11 @@ _dl_next_tls_modid (void)
>>>>>  	      }
>>>>>  
>>>>>  	    if (result - disp < runp->len)
>>>>> -	      break;
>>>>> +	      {
>>>>> +		/* Mark the entry as used, so any dependency see it.  */
>>>>> +		runp->slotinfo[result - disp].map = (struct link_map *) -1;
>>>>> +		break;
>>>>> +	      }
>>>>>  
>>>>>  	    disp += runp->len;
>>>>>  	  }
>>>>
>>>> Which dependency?
>>
>> Any shared library dependency being loaded after this.
>>
>>>>
>>>> I think the special value -1 needs to be documented on the slotinfo
>>>> struct member.
>>>>
>>>> When Carlos and I discussed this, we couldn't quite decide whether it
>>>> was appropriate just to assign the actual link map value here.
>>>
>>> slotinfo[].map is only accessed without locks during
>>> _dl_update_slotinfo at tls access and then it is
>>> only used to get the map for the module in which
>>> tls is accessed. (and to skip NULL maps when resizing
>>> dtv but that's just a minor optimization).
>>>
>>> so the only concurrent .map access does not really
>>> care about its value for an incomplete module.
>>>
>>> so we only need to reason locally what happens within
>>> one dlopen call to decide what can be put temporarily
>>> in .map and the final value only has to be stored
>>> before the lock is released. (with bug 19924 fixed
>>> the final .map and .gen values likely have to be
>>> stored a bit earlier: before the global gen count is
>>> updated)
>>>
>>> the tricky bit is to check if the slotinfo state is
>>> rolled back right in case of dlopen failures. but it
>>> seems to me that modid assignment happens after the
>>> point where failures are handled via dlclose which
>>> will fixup the slotinfo state (i.e. after unlock
>>> unused slotinfo entries are marked with .map == NULL).
>>
>> That is my understanding as well and I tried to check if the
>> rollback is correctly done by adding more cases on the
>> tst-tls20.c. 
>>
>>>
>>> so i think _dl_next_tls_modid() can be changed to
>>> _dl_assign_tls_modid(map) which writes map to the
>>> slotinfo .map.
>> What do you mean by '_dl_assign_tls_modid()' in this case ?
> 
> currently the only use is
> 
>  map->l_tls_modid = _dl_next_tls_modid ();
> 
> which can be changed to
> 
>  _dl_assign_tls_modid (map);
> 
> which sets map->l_tls_modid as well as slotinfo[i].map = map
> marking the slotinfo entry non-free (with relaxed atomic
> store since it can be read concurrently).
> 
> i don't know if 'assign' is the right word, but the point is
> that we don't just get the next modid, but allocate it too
> for the given map.

I see, this should work since we just need either a sentinel value
or a valid link_map to indicate that the slot is being already
taken.  For the rollback in case of failure remove_slotinfo() should
handle it (it just check for non NULL).

I will send an updated patch.
diff mbox series

Patch

diff --git a/elf/Makefile b/elf/Makefile
index 698a6ab985..ba4886fd89 100644
--- a/elf/Makefile
+++ b/elf/Makefile
@@ -253,6 +253,13 @@  one-hundred = $(foreach x,0 1 2 3 4 5 6 7 8 9, \
   0$x 1$x 2$x 3$x 4$x 5$x 6$x 7$x 8$x 9$x)
 tst-tls-many-dynamic-modules := \
   $(foreach n,$(one-hundred),tst-tls-manydynamic$(n)mod)
+tst-tls-many-dynamic-modules-dep-suffixes = 0 1 2 3 4 5 6 7 8 9 10 11 12 13 \
+					    14 15 16 17 18 19
+tst-tls-many-dynamic-modules-dep = \
+  $(foreach n,$(tst-tls-many-dynamic-modules-dep-suffixes),tst-tls-manydynamic$(n)mod-dep)
+tst-tls-many-dynamic-modules-dep-bad-suffixes = 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
+tst-tls-many-dynamic-modules-dep-bad = \
+  $(foreach n,$(tst-tls-many-dynamic-modules-dep-bad-suffixes),tst-tls-manydynamic$(n)mod-dep-bad)
 extra-test-objs += $(tlsmod17a-modules:=.os) $(tlsmod18a-modules:=.os) \
 		   tst-tlsalign-vars.o
 test-extras += tst-tlsmod17a tst-tlsmod18a tst-tlsalign-vars
@@ -325,6 +332,8 @@  modules-names = testobj1 testobj2 testobj3 testobj4 testobj5 testobj6 \
 		tst-audit11mod1 tst-audit11mod2 tst-auditmod11 \
 		tst-audit12mod1 tst-audit12mod2 tst-audit12mod3 tst-auditmod12 \
 		tst-latepthreadmod $(tst-tls-many-dynamic-modules) \
+		$(tst-tls-many-dynamic-modules-dep) \
+		$(tst-tls-many-dynamic-modules-dep-bad) \
 		tst-nodelete-dlclose-dso tst-nodelete-dlclose-plugin \
 		tst-main1mod tst-libc_dlvsym-dso tst-absolute-sym-lib \
 		tst-absolute-zero-lib tst-big-note-lib tst-unwind-ctor-lib \
@@ -1818,10 +1827,63 @@  $(objpfx)tst-rtld-help.out: $(objpfx)ld.so
 	$(evaluate-test)
 
 # Reuses tst-tls-many-dynamic-modules
+$(patsubst %,$(objpfx)%.os,$(tst-tls-many-dynamic-modules-dep)): \
+  $(objpfx)tst-tls-manydynamic%mod-dep.os : tst-tls-manydynamicmod.c
+	$(compile-command.c) \
+	  -DNAME=tls_global_$* -DSETTER=set_value_$* -DGETTER=get_value_$*
+$(patsubst %,$(objpfx)%.os,$(tst-tls-many-dynamic-modules-dep-bad)): \
+  $(objpfx)tst-tls-manydynamic%mod-dep-bad.os : tst-tls-manydynamicmod.c
+	$(compile-command.c) \
+	  -DNAME=tls_global_$* -DSETTER=set_value_$* -DGETTER=get_value_$*
 tst-tls20mod-bad.so-no-z-defs = yes
+# Single dependency.
+$(objpfx)tst-tls-manydynamic0mod-dep.so: $(objpfx)tst-tls-manydynamic1mod-dep.so
+# Double dependencies.
+$(objpfx)tst-tls-manydynamic2mod-dep.so: $(objpfx)tst-tls-manydynamic3mod-dep.so \
+					 $(objpfx)tst-tls-manydynamic4mod-dep.so
+# Double dependencies with each dependency depent of another module.
+$(objpfx)tst-tls-manydynamic5mod-dep.so: $(objpfx)tst-tls-manydynamic6mod-dep.so \
+					 $(objpfx)tst-tls-manydynamic7mod-dep.so
+$(objpfx)tst-tls-manydynamic6mod-dep.so: $(objpfx)tst-tls-manydynamic8mod-dep.so
+$(objpfx)tst-tls-manydynamic7mod-dep.so: $(objpfx)tst-tls-manydynamic8mod-dep.so
+# Long chain with one double dependency in the middle
+$(objpfx)tst-tls-manydynamic9mod-dep.so: $(objpfx)tst-tls-manydynamic10mod-dep.so \
+					 $(objpfx)tst-tls-manydynamic11mod-dep.so
+$(objpfx)tst-tls-manydynamic10mod-dep.so: $(objpfx)tst-tls-manydynamic12mod-dep.so
+$(objpfx)tst-tls-manydynamic12mod-dep.so: $(objpfx)tst-tls-manydynamic13mod-dep.so
+# Long chain with two double depedencies in the middle
+$(objpfx)tst-tls-manydynamic14mod-dep.so: $(objpfx)tst-tls-manydynamic15mod-dep.so
+$(objpfx)tst-tls-manydynamic15mod-dep.so: $(objpfx)tst-tls-manydynamic16mod-dep.so \
+					  $(objpfx)tst-tls-manydynamic17mod-dep.so
+$(objpfx)tst-tls-manydynamic16mod-dep.so: $(objpfx)tst-tls-manydynamic18mod-dep.so \
+					  $(objpfx)tst-tls-manydynamic19mod-dep.so
+# Same but with an invalid module.
+# Single dependency.
+$(objpfx)tst-tls-manydynamic0mod-dep-bad.so: $(objpfx)tst-tls20mod-bad.so
+# Double dependencies.
+$(objpfx)tst-tls-manydynamic1mod-dep-bad.so: $(objpfx)tst-tls-manydynamic2mod-dep-bad.so \
+					     $(objpfx)tst-tls20mod-bad.so
+# Double dependencies with each dependency depent of another module.
+$(objpfx)tst-tls-manydynamic3mod-dep-bad.so: $(objpfx)tst-tls-manydynamic4mod-dep-bad.so \
+					     $(objpfx)tst-tls-manydynamic5mod-dep-bad.so
+$(objpfx)tst-tls-manydynamic4mod-dep-bad.so: $(objpfx)tst-tls20mod-bad.so
+$(objpfx)tst-tls-manydynamic5mod-dep-bad.so: $(objpfx)tst-tls20mod-bad.so
+# Long chain with one double dependency in the middle
+$(objpfx)tst-tls-manydynamic6mod-dep-bad.so: $(objpfx)tst-tls-manydynamic7mod-dep-bad.so \
+					     $(objpfx)tst-tls-manydynamic8mod-dep-bad.so
+$(objpfx)tst-tls-manydynamic7mod-dep-bad.so: $(objpfx)tst-tls-manydynamic9mod-dep-bad.so
+$(objpfx)tst-tls-manydynamic9mod-dep-bad.so: $(objpfx)tst-tls20mod-bad.so
+# Long chain with two double depedencies in the middle
+$(objpfx)tst-tls-manydynamic10mod-dep-bad.so: $(objpfx)tst-tls-manydynamic11mod-dep-bad.so
+$(objpfx)tst-tls-manydynamic11mod-dep-bad.so: $(objpfx)tst-tls-manydynamic12mod-dep-bad.so \
+					      $(objpfx)tst-tls-manydynamic13mod-dep-bad.so
+$(objpfx)tst-tls-manydynamic12mod-dep-bad.so: $(objpfx)tst-tls-manydynamic14mod-dep-bad.so \
+					      $(objpfx)tst-tls20mod-bad.so
 $(objpfx)tst-tls20: $(shared-thread-library)
 $(objpfx)tst-tls20.out: $(objpfx)tst-tls20mod-bad.so \
-			$(tst-tls-many-dynamic-modules:%=$(objpfx)%.so)
+			$(tst-tls-many-dynamic-modules:%=$(objpfx)%.so) \
+			$(tst-tls-many-dynamic-modules-dep:%=$(objpfx)%.so) \
+			$(tst-tls-many-dynamic-modules-dep-bad:%=$(objpfx)%.so) \
 
 # Reuses tst-tls-many-dynamic-modules
 $(objpfx)tst-tls21: $(shared-thread-library)
diff --git a/elf/dl-close.c b/elf/dl-close.c
index 3720e47dd1..f39001cab9 100644
--- a/elf/dl-close.c
+++ b/elf/dl-close.c
@@ -77,8 +77,6 @@  remove_slotinfo (size_t idx, struct dtv_slotinfo_list *listp, size_t disp,
 	 object that wasn't fully set up.  */
       if (__glibc_likely (old_map != NULL))
 	{
-	  assert (old_map->l_tls_modid == idx);
-
 	  /* Mark the entry as unused.  These can be read concurrently.  */
 	  atomic_store_relaxed (&listp->slotinfo[idx - disp].gen,
 				GL(dl_tls_generation) + 1);
@@ -88,7 +86,11 @@  remove_slotinfo (size_t idx, struct dtv_slotinfo_list *listp, size_t disp,
       /* If this is not the last currently used entry no need to look
 	 further.  */
       if (idx != GL(dl_tls_max_dtv_idx))
-	return true;
+	{
+	  /* There is an unused dtv entry in the middle.  */
+	  GL(dl_tls_dtv_gaps) = true;
+	  return true;
+	}
     }
 
   while (idx - disp > (disp == 0 ? 1 + GL(dl_tls_static_nelem) : 0))
diff --git a/elf/dl-open.c b/elf/dl-open.c
index a066f39bd0..d2240d8747 100644
--- a/elf/dl-open.c
+++ b/elf/dl-open.c
@@ -899,16 +899,6 @@  no more namespaces available for dlmopen()"));
 	 state if relocation failed, for example.  */
       if (args.map)
 	{
-	  /* Maybe some of the modules which were loaded use TLS.
-	     Since it will be removed in the following _dl_close call
-	     we have to mark the dtv array as having gaps to fill the
-	     holes.  This is a pessimistic assumption which won't hurt
-	     if not true.  There is no need to do this when we are
-	     loading the auditing DSOs since TLS has not yet been set
-	     up.  */
-	  if ((mode & __RTLD_AUDIT) == 0)
-	    GL(dl_tls_dtv_gaps) = true;
-
 	  _dl_close_worker (args.map, true);
 
 	  /* All l_nodelete_pending objects should have been deleted
diff --git a/elf/dl-tls.c b/elf/dl-tls.c
index 2b5161d10a..4ec4c7f38c 100644
--- a/elf/dl-tls.c
+++ b/elf/dl-tls.c
@@ -157,7 +157,11 @@  _dl_next_tls_modid (void)
 	      }
 
 	    if (result - disp < runp->len)
-	      break;
+	      {
+		/* Mark the entry as used, so any dependency see it.  */
+		runp->slotinfo[result - disp].map = (struct link_map *) -1;
+		break;
+	      }
 
 	    disp += runp->len;
 	  }
@@ -191,10 +195,7 @@  _dl_next_tls_modid (void)
 size_t
 _dl_count_modids (void)
 {
-  /* It is rare that we have gaps; see elf/dl-open.c (_dl_open) where
-     we fail to load a module and unload it leaving a gap.  If we don't
-     have gaps then the number of modids is the current maximum so
-     return that.  */
+  /* The count is the max unless dlclose or failed dlopen created gaps.  */
   if (__glibc_likely (!GL(dl_tls_dtv_gaps)))
     return GL(dl_tls_max_dtv_idx);
 
diff --git a/elf/tst-tls20.c b/elf/tst-tls20.c
index 9977ec8032..f11c2776a0 100644
--- a/elf/tst-tls20.c
+++ b/elf/tst-tls20.c
@@ -16,6 +16,7 @@ 
    License along with the GNU C Library; if not, see
    <http://www.gnu.org/licenses/>.  */
 
+#include <array_length.h>
 #include <dlfcn.h>
 #include <pthread.h>
 #include <stdio.h>
@@ -62,25 +63,70 @@  access (int i)
   printf ("mod[%d]: &tls = %p\n", i, p);
   if (p == NULL)
     FAIL_EXIT1 ("dlsym failed: %s\n", dlerror ());
+  TEST_COMPARE (*p, 0);
   ++*p;
   free (buf);
 }
 
+static void
+access_mod (const char *modname, void *mod, int i)
+{
+  char *modsym = xasprintf ("tls_global_%d", i);
+  dlerror ();
+  int *p = dlsym (mod, modsym);
+  printf ("%s: &tls = %p\n", modname, p);
+  if (p == NULL)
+    FAIL_EXIT1 ("dlsym failed: %s\n", dlerror ());
+  TEST_COMPARE (*p, 0);
+  ++*p;
+  free (modsym);
+}
+
+static void
+access_dep (int i)
+{
+  char *modname = xasprintf ("tst-tls-manydynamic%dmod-dep.so", i);
+  void *moddep = xdlopen (modname, RTLD_LAZY);
+  access_mod (modname, moddep, i);
+  free (modname);
+  xdlclose (moddep);
+}
+
+struct start_args
+{
+  const char *modname;
+  void *mod;
+  int modi;
+  int ndeps;
+  const int *deps;
+};
+
 static void *
 start (void *a)
 {
+  struct start_args *args = a;
+
   for (int i = 0; i < NMOD; i++)
     if (mod[i] != NULL)
       access (i);
+
+  if (args != NULL)
+    {
+      access_mod (args->modname, args->mod, args->modi);
+      for (int n = 0; n < args->ndeps; n++)
+	access_dep (args->deps[n]);
+    }
+
   return 0;
 }
 
-static int
-do_test (void)
+/* This test gaps with shared libraries with dynamic TLS that has no
+   dependencies.  The DTV gap is set with by trying to load an invalid
+   module, the entry should be used on the dlopen.  */
+static void
+do_test_no_depedency (void)
 {
-  int i;
-
-  for (i = 0; i < NMOD; i++)
+  for (int i = 0; i < NMOD; i++)
     {
       load_mod (i);
       /* Bump the generation of mod[0] without using new dtv slot.  */
@@ -91,8 +137,218 @@  do_test (void)
       pthread_t t = xpthread_create (0, start, 0);
       xpthread_join (t);
     }
-  for (i = 0; i < NMOD; i++)
+  for (int i = 0; i < NMOD; i++)
     unload_mod (i);
+}
+
+/* The following test check DTV gaps handling with shared libraries that has
+   dependencies.  It defines 5 different sets:
+
+   1. Single dependency:
+      mod0 -> mod1
+   2. Double dependency:
+      mod2 -> [mod3,mod4]
+   3. Double dependency with each dependency depent of another module:
+      mod5 -> [mod6,mod7] -> mod8
+   4. Long chain with one double dependency in the middle:
+      mod9 -> [mod10, mod11] -> mod12 -> mod13
+   5. Long chain with two double depedencies in the middle:
+      mod15 -> mod15 -> [mod16, mod17]
+      mod15 -> [mod18, mod19]
+
+   This does not cover all the possible gaps and configuration, but it
+   should check if different dynamic shared sets are placed correctly in
+   different gaps configurations.  */
+
+static int
+nmodules (uint32_t v)
+{
+  unsigned int r = 0;
+  while (v >>= 1)
+    r++;
+  return r + 1;
+}
+
+static inline bool
+is_mod_set (uint32_t g, uint32_t n)
+{
+  return (1U << (n - 1)) & g;
+}
+
+static void
+print_gap (uint32_t g)
+{
+  printf ("gap: ");
+  int nmods = nmodules (g);
+  for (int n = 1; n <= nmods; n++)
+    printf ("%c", ((1 << (n - 1)) & g) == 0 ? 'G' : 'M');
+  printf ("\n");
+}
+
+static void
+do_test_dependency (void)
+{
+  /* Maps the module and its dependencies, use thread to access the TLS on
+     each loaded module.  */
+  static const int tlsmanydeps0[] = { 1 };
+  static const int tlsmanydeps1[] = { 3, 4 };
+  static const int tlsmanydeps2[] = { 6, 7, 8 };
+  static const int tlsmanydeps3[] = { 10, 11, 12 };
+  static const int tlsmanydeps4[] = { 15, 16, 17, 18, 19 };
+  static const struct tlsmanydeps_t
+  {
+    int modi;
+    int ndeps;
+    const int *deps;
+  } tlsmanydeps[] =
+  {
+    {  0, array_length (tlsmanydeps0), tlsmanydeps0 },
+    {  2, array_length (tlsmanydeps1), tlsmanydeps1 },
+    {  5, array_length (tlsmanydeps2), tlsmanydeps2 },
+    {  9, array_length (tlsmanydeps3), tlsmanydeps3 },
+    { 14, array_length (tlsmanydeps4), tlsmanydeps4 },
+  };
+
+  /* The gap configuration is defined as a bitmap: the bit set represents a
+     loaded module prior the tests execution, while a bit unsed is a module
+     unloaded.  Not all permtation will show gaps, but it is simpler than
+     define each one independently.  */
+  for (uint32_t g = 0; g < 64; g++)
+    {
+      print_gap (g);
+      int nmods = nmodules (g);
+
+      int mods[nmods];
+      /* We use '0' as indication for a gap, to avoid the dlclose on iteration
+	 cleanup.  */
+      for (int n = 1; n <= nmods; n++)
+	{
+	  load_mod (n);
+	   mods[n] = n;
+	}
+      for (int n = 1; n <= nmods; n++)
+	{
+	  if (!is_mod_set (g, n))
+	    {
+	      unload_mod (n);
+	      mods[n] = 0;
+	    }
+	}
+
+      for (int t = 0; t < array_length (tlsmanydeps); t++)
+	{
+	  char *moddepname = xasprintf ("tst-tls-manydynamic%dmod-dep.so",
+					tlsmanydeps[t].modi);
+	  void *moddep = xdlopen (moddepname, RTLD_LAZY);
+
+	  /* Access TLS in all loaded modules.  */
+	  struct start_args args =
+	    {
+	      moddepname,
+	      moddep,
+	      tlsmanydeps[t].modi,
+	      tlsmanydeps[t].ndeps,
+	      tlsmanydeps[t].deps
+	    };
+	  pthread_t t = xpthread_create (0, start, &args);
+	  xpthread_join (t);
+
+	  free (moddepname);
+	  xdlclose (moddep);
+	}
+
+      for (int n = 1; n <= nmods; n++)
+	if (mods[n] != 0)
+	  unload_mod (n);
+    }
+}
+
+/* The following test check DTV gaps handling with shared libraries that has
+   invalid dependencies.  It defines 5 different sets:
+
+   1. Single dependency:
+      mod0 -> invalid
+   2. Double dependency:
+      mod1 -> [mod2,invalid]
+   3. Double dependency with each dependency depent of another module:
+      mod3 -> [mod4,mod5] -> invalid
+   4. Long chain with one double dependency in the middle:
+      mod6 -> [mod7, mod8] -> mod12 -> invalid
+   5. Long chain with two double depedencies in the middle:
+      mod10 -> mod11 -> [mod12, mod13]
+      mod12 -> [mod14, invalid]
+
+   This does not cover all the possible gaps and configuration, but it
+   should check if different dynamic shared sets are placed correctly in
+   different gaps configurations.  */
+
+static void
+do_test_invalid_dependency (bool bind_now)
+{
+  static const int tlsmanydeps[] = { 0, 1, 3, 6, 10 };
+
+  /* The gap configuration is defined as a bitmap: the bit set represents a
+     loaded module prior the tests execution, while a bit unsed is a module
+     unloaded.  Not all permtation will show gaps, but it is simpler than
+     define each one independently.  */
+  for (uint32_t g = 0; g < 64; g++)
+    {
+      print_gap (g);
+      int nmods = nmodules (g);
+
+      int mods[nmods];
+      /* We use '0' as indication for a gap, to avoid the dlclose on iteration
+	 cleanup.  */
+      for (int n = 1; n <= nmods; n++)
+	{
+	  load_mod (n);
+	   mods[n] = n;
+	}
+      for (int n = 1; n <= nmods; n++)
+	{
+	  if (!is_mod_set (g, n))
+	    {
+	      unload_mod (n);
+	      mods[n] = 0;
+	    }
+	}
+
+      for (int t = 0; t < array_length (tlsmanydeps); t++)
+	{
+	  char *moddepname = xasprintf ("tst-tls-manydynamic%dmod-dep-bad.so",
+					tlsmanydeps[t]);
+	  void *moddep;
+	  if (bind_now)
+	    {
+	      moddep = dlopen (moddepname, RTLD_NOW);
+	      TEST_VERIFY (moddep == 0);
+	    }
+	  else
+	    moddep = dlopen (moddepname, RTLD_LAZY);
+
+	  /* Access TLS in all loaded modules.  */
+	  pthread_t t = xpthread_create (0, start, NULL);
+	  xpthread_join (t);
+
+	  free (moddepname);
+	  if (!bind_now)
+	    xdlclose (moddep);
+	}
+
+      for (int n = 1; n <= nmods; n++)
+	if (mods[n] != 0)
+	  unload_mod (n);
+    }
+}
+
+static int
+do_test (void)
+{
+  do_test_no_depedency ();
+  do_test_dependency ();
+  do_test_invalid_dependency (true);
+  do_test_invalid_dependency (false);
+
   return 0;
 }