RFC elf: Fix slow tls access after dlopen [BZ #19924] nsz/bug19329-v2

In short: __tls_get_addr checks the global generation counter,
_dl_update_slotinfo updates up to the generation of the accessed
module. If the global generation is newer than geneneration of the
module then __tls_get_addr keeps hitting the slow path that updates
the dtv.

Possible approaches i can see:

1. update to global generation instead of module,
2. check the module generation in the fast path.

This patch is 1.: it needs additional sync (load acquire) so the
slotinfo list is up to date with the observed global generation.

Approach 2. would require walking the slotinfo list at all times.
I don't know how to make that fast with many modules.

Note: in the x86_64 version of dl-tls.c the generation is only loaded
once, since relaxed mo is not faster than acquire mo load.

I have not benchmarked this yet.

1 files changed, 2 insertions, 2 deletions

diff --git a/sysdeps/x86_64/dl-tls.c b/sysdeps/x86_64/dl-tls.c
index 24ef560b71..4ded8dd6b9 100644
--- a/sysdeps/x86_64/dl-tls.c
+++ b/sysdeps/x86_64/dl-tls.c
@@ -40,9 +40,9 @@ __tls_get_addr_slow (GET_ADDR_ARGS)
 {
   dtv_t *dtv = THREAD_DTV ();
 
-  size_t gen = atomic_load_relaxed (&GL(dl_tls_generation));
+  size_t gen = atomic_load_acquire (&GL(dl_tls_generation));
   if (__glibc_unlikely (dtv[0].counter != gen))
-    return update_get_addr (GET_ADDR_PARAM);
+    return update_get_addr (GET_ADDR_PARAM, gen);
 
   return tls_get_addr_tail (GET_ADDR_PARAM, dtv, NULL);
 }