[PATCH] Use rel semantics of cas instead of acq semantics with full barrier before it in _int_free

[PATCH] Use rel semantics of cas instead of acq semantics with full barrier before it in _int_free

[Jakub Jelinek]

Hi!

The following patch fixes catomic_compare_and_exchange_*_rel definitions
(which were never used and weren't correct) and uses
catomic_compare_and_exchange_val_rel in _int_free.  Comparing to the
pre-2009-07-02 --enable-experimental-malloc state the generated code should
be identical on all arches other than ppc/ppc64 and on ppc/ppc64 should use
lwsync instead of isync barrier.

2009-07-13  Jakub Jelinek  <jakub@redhat.com>

	* include/atomic.h (catomic_compare_and_exchange_val_rel): If arch
	overrides atomic_compare_and_exchange_val_rel, define to
	atomic_compare_and_exchange_val_rel by default, otherwise default
	to catomic_compare_and_exchange_val_acq.
	(catomic_compare_and_exchange_bool_rel): If arch overrides
	atomic_compare_and_exchange_bool_rel, define to
	atomic_compare_and_exchange_bool_rel by default.
	* malloc/malloc.c (_int_free): Revert 2009-07-02 change.
	Use catomic_compare_and_exchange_val_rel instead of
	catomic_compare_and_exchange_val_acq.

--- libc/include/atomic.h.jj	2009-05-16 19:23:35.000000000 +0200
+++ libc/include/atomic.h	2009-07-13 17:47:02.000000000 +0200
@@ -107,14 +107,19 @@
 #endif
 
 
-#ifndef atomic_compare_and_exchange_val_rel
-# define atomic_compare_and_exchange_val_rel(mem, newval, oldval)	      \
-  atomic_compare_and_exchange_val_acq (mem, newval, oldval)
+#ifndef catomic_compare_and_exchange_val_rel
+# ifndef atomic_compare_and_exchange_val_rel
+#  define catomic_compare_and_exchange_val_rel(mem, newval, oldval)	      \
+  catomic_compare_and_exchange_val_acq (mem, newval, oldval)
+# else
+#  define catomic_compare_and_exchange_val_rel(mem, newval, oldval)	      \
+  atomic_compare_and_exchange_val_rel (mem, newval, oldval)
+# endif
 #endif
 
 
-#ifndef catomic_compare_and_exchange_val_rel
-# define catomic_compare_and_exchange_val_rel(mem, newval, oldval)	      \
+#ifndef atomic_compare_and_exchange_val_rel
+# define atomic_compare_and_exchange_val_rel(mem, newval, oldval)	      \
   atomic_compare_and_exchange_val_acq (mem, newval, oldval)
 #endif
 
@@ -155,15 +160,20 @@
 #endif
 
 
-#ifndef atomic_compare_and_exchange_bool_rel
-# define atomic_compare_and_exchange_bool_rel(mem, newval, oldval) \
-  atomic_compare_and_exchange_bool_acq (mem, newval, oldval)
+#ifndef catomic_compare_and_exchange_bool_rel
+# ifndef atomic_compare_and_exchange_bool_rel
+#  define catomic_compare_and_exchange_bool_rel(mem, newval, oldval)	      \
+  catomic_compare_and_exchange_bool_acq (mem, newval, oldval)
+# else
+#  define catomic_compare_and_exchange_bool_rel(mem, newval, oldval)	      \
+  atomic_compare_and_exchange_bool_rel (mem, newval, oldval)
+# endif
 #endif
 
 
-#ifndef catomic_compare_and_exchange_bool_rel
-# define catomic_compare_and_exchange_bool_rel(mem, newval, oldval) \
-  catomic_compare_and_exchange_bool_acq (mem, newval, oldval)
+#ifndef atomic_compare_and_exchange_bool_rel
+# define atomic_compare_and_exchange_bool_rel(mem, newval, oldval) \
+  atomic_compare_and_exchange_bool_acq (mem, newval, oldval)
 #endif
 
 
--- libc/malloc/malloc.c.jj	2009-07-07 19:10:19.000000000 +0200
+++ libc/malloc/malloc.c	2009-07-13 17:47:52.000000000 +0200
@@ -4822,9 +4822,8 @@ _int_free(mstate av, mchunkptr p)
 	    goto errout;
 	  }
 	p->fd = fd = old;
-	atomic_full_barrier ();
       }
-    while ((old = catomic_compare_and_exchange_val_acq (fb, p, fd)) != fd);
+    while ((old = catomic_compare_and_exchange_val_rel (fb, p, fd)) != fd);
 #else
     /* Another simple check: make sure the top of the bin is not the
        record we are going to add (i.e., double free).  */

	Jakub

Re: [PATCH] Use rel semantics of cas instead of acq semantics with full barrier before it in _int_free

[Steve Munroe]

Jakub Jelinek <jakub@redhat.com> wrote on 07/13/2009 11:10:31 AM:

> Hi!
>
> The following patch fixes catomic_compare_and_exchange_*_rel definitions
> (which were never used and weren't correct) and uses
> catomic_compare_and_exchange_val_rel in _int_free.  Comparing to the
> pre-2009-07-02 --enable-experimental-malloc state the generated code
should
> be identical on all arches other than ppc/ppc64 and on ppc/ppc64 should
use
> lwsync instead of isync barrier.
>
> 2009-07-13  Jakub Jelinek  <jakub@redhat.com>
>
>    * include/atomic.h (catomic_compare_and_exchange_val_rel): If arch
>    overrides atomic_compare_and_exchange_val_rel, define to
>    atomic_compare_and_exchange_val_rel by default, otherwise default
>    to catomic_compare_and_exchange_val_acq.
>    (catomic_compare_and_exchange_bool_rel): If arch overrides
>    atomic_compare_and_exchange_bool_rel, define to
>    atomic_compare_and_exchange_bool_rel by default.
>    * malloc/malloc.c (_int_free): Revert 2009-07-02 change.
>    Use catomic_compare_and_exchange_val_rel instead of
>    catomic_compare_and_exchange_val_acq.
>
> --- libc/include/atomic.h.jj   2009-05-16 19:23:35.000000000 +0200
> +++ libc/include/atomic.h   2009-07-13 17:47:02.000000000 +0200
> @@ -107,14 +107,19 @@
>  #endif
>
>
> -#ifndef atomic_compare_and_exchange_val_rel
> -# define atomic_compare_and_exchange_val_rel(mem, newval, oldval)
\
> -  atomic_compare_and_exchange_val_acq (mem, newval, oldval)
> +#ifndef catomic_compare_and_exchange_val_rel
> +# ifndef atomic_compare_and_exchange_val_rel
> +#  define catomic_compare_and_exchange_val_rel(mem, newval, oldval)
\
> +  catomic_compare_and_exchange_val_acq (mem, newval, oldval)
> +# else
> +#  define catomic_compare_and_exchange_val_rel(mem, newval, oldval)
\
> +  atomic_compare_and_exchange_val_rel (mem, newval, oldval)
> +# endif
>  #endif
>

I assume that defining catomic_compare_and_exchange_val_rel in terms of
catomic_compare_and_exchange_val_acq is only for platforms where there is
no
distinction (between _acq/_rel) and the platforms sysdeps atomic.h did not
define separate *_rel macros.

Otherwise this is a bit confusing...

>  snip ...
> --- libc/malloc/malloc.c.jj   2009-07-07 19:10:19.000000000 +0200
> +++ libc/malloc/malloc.c   2009-07-13 17:47:52.000000000 +0200
> @@ -4822,9 +4822,8 @@ _int_free(mstate av, mchunkptr p)
>         goto errout;
>       }
>     p->fd = fd = old;
> -   atomic_full_barrier ();
>        }
> -    while ((old = catomic_compare_and_exchange_val_acq (fb, p, fd)) !=
fd);
> +    while ((old = catomic_compare_and_exchange_val_rel (fb, p, fd)) !=
fd);
>  #else
>      /* Another simple check: make sure the top of the bin is not the
>         record we are going to add (i.e., double free).  */
>
>    Jakub

This is a better solution as the use case matches release semantics. In
this case we are exporting the "p->fd = fd = old" and needs the export/read
barrier that the *_rel macro naturally provides. A import/write barrier is
not required as the do/while iterates only for retry.

For PPC the selection of raw *_acq vs *_rel macros depends on the context.

Thanks

Steven J. Munroe
Linux on Power Toolchain Architect
IBM Corporation, Linux Technology Center