[PATCH v6 3/4] Reduce CAS in malloc spinlocks

H.J. Lu hjl.tools@gmail.com
Thu Nov 11 16:24:27 GMT 2021


Do an atomic load and check if compare may fail.  Skip CAS and spin if
compare may fail to reduce cache line bouncing on contended locks.
---
 malloc/arena.c  |  5 +++++
 malloc/malloc.c | 10 ++++++++++
 2 files changed, 15 insertions(+)

diff --git a/malloc/arena.c b/malloc/arena.c
index 78ef4cf18c..e7fbe7c183 100644
--- a/malloc/arena.c
+++ b/malloc/arena.c
@@ -899,6 +899,11 @@ arena_get2 (size_t size, mstate avoid_arena)
          enough address space to create that many arenas.  */
       if (__glibc_unlikely (n <= narenas_limit - 1))
         {
+          if (atomic_load_relaxed (&narenas) != n)
+           {
+              atomic_spin_nop ();
+              goto repeat;
+           }
           if (catomic_compare_and_exchange_bool_acq (&narenas, n + 1, n))
             goto repeat;
           a = _int_new_arena (size);
diff --git a/malloc/malloc.c b/malloc/malloc.c
index 095d97a3be..403ffb84ef 100644
--- a/malloc/malloc.c
+++ b/malloc/malloc.c
@@ -3717,6 +3717,11 @@ _int_malloc (mstate av, size_t bytes)
       pp = REVEAL_PTR (victim->fd);                                     \
       if (__glibc_unlikely (pp != NULL && misaligned_chunk (pp)))       \
 	malloc_printerr ("malloc(): unaligned fastbin chunk detected"); \
+      if (atomic_load_relaxed (fb) != victim)		\
+	{						\
+	  atomic_spin_nop ();				\
+	  continue;					\
+	}						\
     }							\
   while ((pp = catomic_compare_and_exchange_val_acq (fb, pp, victim)) \
 	 != victim);					\
@@ -4435,6 +4440,11 @@ _int_free (mstate av, mchunkptr p, int have_lock)
 	    malloc_printerr ("double free or corruption (fasttop)");
 	  old2 = old;
 	  p->fd = PROTECT_PTR (&p->fd, old);
+	  if (atomic_load_relaxed (fb) != old2)
+	    {
+	      atomic_spin_nop ();
+	      continue;
+	    }
 	}
       while ((old = catomic_compare_and_exchange_val_rel (fb, p, old2))
 	     != old2);
-- 
2.33.1



More information about the Libc-alpha mailing list