diff mbox series

[RFC,4/5] elf: Enable RTLD_NODELETE on __libc_unwind_link_get

Message ID 20240611153220.165430-5-adhemerval.zanella@linaro.org (mailing list archive)
State Superseded
Headers
Series Add support for memory sealing |

Checks

Context Check Description
redhat-pt-bot/TryBot-apply_patch success Patch applied to master at the time it was sent
linaro-tcwg-bot/tcwg_glibc_build--master-aarch64 success Test passed
linaro-tcwg-bot/tcwg_glibc_check--master-aarch64 success Test passed
linaro-tcwg-bot/tcwg_glibc_build--master-arm success Test passed
linaro-tcwg-bot/tcwg_glibc_check--master-arm fail Test failed

Commit Message

Adhemerval Zanella Netto June 11, 2024, 3:27 p.m. UTC 
  The libgcc_s.so can also be sealed.  The library is loaded once
and not unloaded during process execution (only for memory debug
with __libc_unwind_link_freeres).

Checked on x86_64-linux-gnu and aarch64-linux-gnu.
---
 include/dlfcn.h                        |  2 ++
 manual/tunables.texi                   |  4 ++++
 misc/unwind-link.c                     |  5 +++--
 sysdeps/unix/sysv/linux/tst-dl_mseal.c | 13 +++++++++++++
 4 files changed, 22 insertions(+), 2 deletions(-)

Comments

Florian Weimer June 12, 2024, 9:54 a.m. UTC | #1 
* Adhemerval Zanella:

> The libgcc_s.so can also be sealed.  The library is loaded once
> and not unloaded during process execution (only for memory debug
> with __libc_unwind_link_freeres).

The unwind-link change to use RTLD_NODELETE could go in separately.

> diff --git a/misc/unwind-link.c b/misc/unwind-link.c
> index 213a0162a4..7267ecbec3 100644
> --- a/misc/unwind-link.c
> +++ b/misc/unwind-link.c

> @@ -100,7 +100,8 @@ __libc_unwind_link_get (void)
>  
>    __libc_lock_lock (lock);
>    if (atomic_load_relaxed (&global_libgcc_handle) != NULL)
> -    /* This thread lost the race.  Clean up.  */
> +    /* This thread lost the race.  Drop the l_direct_opencount and issue
> +       the debug log.  */
>      __libc_dlclose (local_libgcc_handle);
>    else
>      {

I don't understand what “debug log” means in this context.

Thanks,
Florian
Adhemerval Zanella Netto June 12, 2024, 5:16 p.m. UTC | #2 
On 12/06/24 06:54, Florian Weimer wrote:
> * Adhemerval Zanella:
> 
>> The libgcc_s.so can also be sealed.  The library is loaded once
>> and not unloaded during process execution (only for memory debug
>> with __libc_unwind_link_freeres).
> 
> The unwind-link change to use RTLD_NODELETE could go in separately.

Ok, I will send a separate patch.

> 
>> diff --git a/misc/unwind-link.c b/misc/unwind-link.c
>> index 213a0162a4..7267ecbec3 100644
>> --- a/misc/unwind-link.c
>> +++ b/misc/unwind-link.c
> 
>> @@ -100,7 +100,8 @@ __libc_unwind_link_get (void)
>>  
>>    __libc_lock_lock (lock);
>>    if (atomic_load_relaxed (&global_libgcc_handle) != NULL)
>> -    /* This thread lost the race.  Clean up.  */
>> +    /* This thread lost the race.  Drop the l_direct_opencount and issue
>> +       the debug log.  */
>>      __libc_dlclose (local_libgcc_handle);
>>    else
>>      {
> 
> I don't understand what “debug log” means in this context.

Sorry, I meant the __libc_unwind_link_freeres usually triggered by
memory profilers.  With sealing the dlclose won't unmap the memory.
Florian Weimer June 12, 2024, 5:50 p.m. UTC | #3 
* Adhemerval Zanella Netto:

> On 12/06/24 06:54, Florian Weimer wrote:
>> * Adhemerval Zanella:
>> 
>>> The libgcc_s.so can also be sealed.  The library is loaded once
>>> and not unloaded during process execution (only for memory debug
>>> with __libc_unwind_link_freeres).
>> 
>> The unwind-link change to use RTLD_NODELETE could go in separately.
>
> Ok, I will send a separate patch.
>
>> 
>>> diff --git a/misc/unwind-link.c b/misc/unwind-link.c
>>> index 213a0162a4..7267ecbec3 100644
>>> --- a/misc/unwind-link.c
>>> +++ b/misc/unwind-link.c
>> 
>>> @@ -100,7 +100,8 @@ __libc_unwind_link_get (void)
>>>  
>>>    __libc_lock_lock (lock);
>>>    if (atomic_load_relaxed (&global_libgcc_handle) != NULL)
>>> -    /* This thread lost the race.  Clean up.  */
>>> +    /* This thread lost the race.  Drop the l_direct_opencount and issue
>>> +       the debug log.  */
>>>      __libc_dlclose (local_libgcc_handle);
>>>    else
>>>      {
>> 
>> I don't understand what “debug log” means in this context.
>
> Sorry, I meant the __libc_unwind_link_freeres usually triggered by
> memory profilers.  With sealing the dlclose won't unmap the memory.

I still don't understand the comment.

We can still deallocate the helper data structures.  With the switch to
read-only link maps, most of the allocations will be hidden from malloc
tracing anyway and no longer appear as leaks.

Thanks,
Florian
Adhemerval Zanella Netto June 12, 2024, 5:55 p.m. UTC | #4 
On 12/06/24 14:50, Florian Weimer wrote:
> * Adhemerval Zanella Netto:
> 
>> On 12/06/24 06:54, Florian Weimer wrote:
>>> * Adhemerval Zanella:
>>>
>>>> The libgcc_s.so can also be sealed.  The library is loaded once
>>>> and not unloaded during process execution (only for memory debug
>>>> with __libc_unwind_link_freeres).
>>>
>>> The unwind-link change to use RTLD_NODELETE could go in separately.
>>
>> Ok, I will send a separate patch.
>>
>>>
>>>> diff --git a/misc/unwind-link.c b/misc/unwind-link.c
>>>> index 213a0162a4..7267ecbec3 100644
>>>> --- a/misc/unwind-link.c
>>>> +++ b/misc/unwind-link.c
>>>
>>>> @@ -100,7 +100,8 @@ __libc_unwind_link_get (void)
>>>>  
>>>>    __libc_lock_lock (lock);
>>>>    if (atomic_load_relaxed (&global_libgcc_handle) != NULL)
>>>> -    /* This thread lost the race.  Clean up.  */
>>>> +    /* This thread lost the race.  Drop the l_direct_opencount and issue
>>>> +       the debug log.  */
>>>>      __libc_dlclose (local_libgcc_handle);
>>>>    else
>>>>      {
>>>
>>> I don't understand what “debug log” means in this context.
>>
>> Sorry, I meant the __libc_unwind_link_freeres usually triggered by
>> memory profilers.  With sealing the dlclose won't unmap the memory.
> 
> I still don't understand the comment.
> 
> We can still deallocate the helper data structures.  With the switch to
> read-only link maps, most of the allocations will be hidden from malloc
> tracing anyway and no longer appear as leaks.

I am not sure if valgrind or any memory profilers won't complain about
lingering map segments, but you are right that at least regarding
_dl_unmap_segments we don't really check the unmmap result.
diff mbox series

Patch

diff --git a/include/dlfcn.h b/include/dlfcn.h
index f49ee1b0c9..06e2ecbdd2 100644
--- a/include/dlfcn.h
+++ b/include/dlfcn.h
@@ -50,6 +50,8 @@  extern char **__libc_argv attribute_hidden;
    better error handling semantics for the library.  */
 #define __libc_dlopen(name) \
   __libc_dlopen_mode (name, RTLD_NOW | __RTLD_DLOPEN)
+#define __libc_dlopen_nodelete(name) \
+  __libc_dlopen_mode (name, RTLD_NODELETE | RTLD_NOW | __RTLD_DLOPEN)
 extern void *__libc_dlopen_mode  (const char *__name, int __mode)
   attribute_hidden;
 extern void *__libc_dlsym   (void *__map, const char *__name)
diff --git a/manual/tunables.texi b/manual/tunables.texi
index 26fba6641d..be36d52cf9 100644
--- a/manual/tunables.texi
+++ b/manual/tunables.texi
@@ -381,6 +381,10 @@  Any preload libraries.
 
 @item
 Any library loaded with @code{dlopen} with @code{RTLD_NODELETE} flag.
+
+@item
+Any runtime library used for process unwind (such as required by @code{backtrace}
+or @code{pthread_exit}).
 @end itemize
 
 The tunable accepts three diferent values: @samp{0} where sealing is disabled,
diff --git a/misc/unwind-link.c b/misc/unwind-link.c
index 213a0162a4..7267ecbec3 100644
--- a/misc/unwind-link.c
+++ b/misc/unwind-link.c
@@ -48,7 +48,7 @@  __libc_unwind_link_get (void)
   /* Initialize a copy of the data, so that we do not need about
      unlocking in case the dynamic loader somehow triggers
      unwinding.  */
-  void *local_libgcc_handle = __libc_dlopen (LIBGCC_S_SO);
+  void *local_libgcc_handle = __libc_dlopen_nodelete (LIBGCC_S_SO);
   if (local_libgcc_handle == NULL)
     {
       __libc_lock_unlock (lock);
@@ -100,7 +100,8 @@  __libc_unwind_link_get (void)
 
   __libc_lock_lock (lock);
   if (atomic_load_relaxed (&global_libgcc_handle) != NULL)
-    /* This thread lost the race.  Clean up.  */
+    /* This thread lost the race.  Drop the l_direct_opencount and issue
+       the debug log.  */
     __libc_dlclose (local_libgcc_handle);
   else
     {
diff --git a/sysdeps/unix/sysv/linux/tst-dl_mseal.c b/sysdeps/unix/sysv/linux/tst-dl_mseal.c
index 72a33d04c7..da1a3ebe5a 100644
--- a/sysdeps/unix/sysv/linux/tst-dl_mseal.c
+++ b/sysdeps/unix/sysv/linux/tst-dl_mseal.c
@@ -19,6 +19,7 @@ 
 #include <array_length.h>
 #include <errno.h>
 #include <getopt.h>
+#include <gnu/lib-names.h>
 #include <inttypes.h>
 #include <libgen.h>
 #include <stdio.h>
@@ -31,6 +32,7 @@ 
 #include <support/support.h>
 #include <support/xdlfcn.h>
 #include <support/xstdio.h>
+#include <support/xthread.h>
 
 #define LIB_PRELOAD              "lib-tst-dl_mseal-preload.so"
 
@@ -70,6 +72,7 @@  static const char *expected_sealed_libs[] =
   LIB_NEEDED_2,
   LIB_DLOPEN_NODELETE,
   LIB_DLOPEN_NODELETE_DEP,
+  LIBGCC_S_SO,
 #endif
   "[vdso]",
 };
@@ -100,6 +103,13 @@  is_in_string_list (const char *s, const char *const list[], size_t len)
   return -1;
 }
 
+static void *
+tf (void *closure)
+{
+  pthread_exit (NULL);
+  return NULL;
+}
+
 static int
 handle_restart (void)
 {
@@ -108,6 +118,9 @@  handle_restart (void)
   xdlopen (LIB_DLOPEN_DEFAULT, RTLD_NOW);
 #endif
 
+  /* pthread_exit will load LIBGCC_S_SO.  */
+  xpthread_join (xpthread_create (NULL, tf, NULL));
+
   FILE *fp = xfopen ("/proc/self/maps", "r");
   char *line = NULL;
   size_t linesiz = 0;