Re: RCU: non-atomic assignment to long/pointer variables in gcc

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Tue, Jan 15, 2013 at 05:07:50PM +0400, Konstantin Khlebnikov wrote:
> Paul E. McKenney wrote:
> >On Tue, Jan 15, 2013 at 02:30:32PM +0400, Konstantin Khlebnikov wrote:
> >>Documentation/atomic_ops.txt (182dd4b277177e8465ad11cd9f85f282946b5578)
> >>says that pointers, longs, ints, and chars are stored and loaded atomically.
> >>
> >>But GCC actually may split assignment to 'long' variable into two instructions.
> >>see example in http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55981
> >>
> >>GCC also splits assignments to 'volatile' variables and this is actually a bug in gcc.
> >>
> >>volatile unsigned long y;
> >>
> >>y = 0x100000001ul;
> >>
> >>   400728:	c7 05 66 06 20 00 01 	movl   $0x1,0x200666(%rip)        # 600d98<y>
> >>   40072f:	00 00 00
> >>   400732:	c7 05 60 06 20 00 01 	movl   $0x1,0x200660(%rip)        # 600d9c<y+0x4>
> >>   400739:	00 00 00
> >>
> >>fortunately for y = 0; it generates this:
> >>
> >>   40071d:	48 c7 05 70 06 20 00 	movq   $0x0,0x200670(%rip)        # 600d98<y>
> >>   400724:	00 00 00 00
> >>
> >>Thus NULL is safe, but constant ERR_PTR may be dangerous.
> >>
> >>Probably rcu_assign_pointer() should use ACCESS_ONCE() around lvalue, because
> >>splitting assignment for non-volatile variable seems like completely valid,
> >>but this may help only after fixing that bug in GCC.
> >
> >Good catch!  I has queued the following patch.
> >
> >							Thanx, Paul
> >
> >------------------------------------------------------------------------
> >
> >rcu: Add ACCESS_ONCE() to rcu_assign_pointer()
> >
> >GCC may split assignment to 'long' variable into two instructions:
> >
> >volatile unsigned long y;
> >
> >y = 0x100000001ul;
> >
> >	movl   $0x1,0x200666(%rip)
> >	movl   $0x1,0x200660(%rip)
> >
> >This commit fixes this by applying ACCESS_ONCE() within
> >__rcu_assign_pointer(), but note that some versions and architectures
> >of GCC have a bug that defeats ACCESS_ONCE():
> >
> >http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55981
> >
> >I added a comment to this bug report asking that the bug be fixed for
> >volatiles as well as atomics, citing a device driver storing a constant
> >into a 64-bit device register as motivation.
> >
> >Reported-by: Konstantin Khlebnikov<khlebnikov@openvz.org>
> >Signed-off-by: Paul E. McKenney<paulmck@linux.vnet.ibm.com>
> >
> >diff --git a/include/linux/rcupdate.h b/include/linux/rcupdate.h
> >index 9ed2c9a..3435174 100644
> >--- a/include/linux/rcupdate.h
> >+++ b/include/linux/rcupdate.h
> >@@ -556,7 +556,7 @@ static inline void rcu_preempt_sleep_check(void)
> >  #define __rcu_assign_pointer(p, v, space) \
> >  	do { \
> >  		smp_wmb(); \
> >-		(p) = (typeof(*v) __force space *)(v); \
> >+		ACCESS_ONCE(p) = (typeof(*v) __force space *)(v); \
> >  	} while (0)
> 
> Seems like RCU_INIT_POINTER() need this too.

For the third use case, which is updating a pointer to reference data that
has already been exposed to RCU readers, you are quite correct!  I must
confess that I had forgotten about that use case.  Please see below for
an updated patch.

And the gcc bug also now has a patch:

	http://gcc.gnu.org/bugzilla/attachment.cgi?id=29169&action=diff

							Thanx, Paul

------------------------------------------------------------------------

rcu: Add ACCESS_ONCE() to rcu_assign_pointer() and RCU_INIT_POINTER()

GCC may split assignment to 'long' variable into two instructions:
    
    	volatile unsigned long y;
    
    	y = 0x100000001ul;
    
    		movl   $0x1,0x200666(%rip)
    		movl   $0x1,0x200660(%rip)
    
This commit fixes this by applying ACCESS_ONCE() within
__rcu_assign_pointer(), but note that some versions and architectures
of GCC have a bug that defeats ACCESS_ONCE():
    
    	http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55981
    
I added a comment to this bug report asking that the bug be fixed for
volatiles as well as atomics, citing a device driver storing a constant
into a 64-bit device register as motivation.  There is now a patch:
    
    	http://gcc.gnu.org/bugzilla/attachment.cgi?id=29169&action=diff
    
Reported-by: Konstantin Khlebnikov <khlebnikov@openvz.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>

diff --git a/include/linux/rcupdate.h b/include/linux/rcupdate.h
index 9ed2c9a..4627abd 100644
--- a/include/linux/rcupdate.h
+++ b/include/linux/rcupdate.h
@@ -556,7 +556,7 @@ static inline void rcu_preempt_sleep_check(void)
 #define __rcu_assign_pointer(p, v, space) \
 	do { \
 		smp_wmb(); \
-		(p) = (typeof(*v) __force space *)(v); \
+		ACCESS_ONCE(p) = (typeof(*v) __force space *)(v); \
 	} while (0)
 
 
@@ -945,7 +945,7 @@ static inline notrace void rcu_read_unlock_sched_notrace(void)
  */
 #define RCU_INIT_POINTER(p, v) \
 	do { \
-		p = (typeof(*v) __force __rcu *)(v); \
+		ACCESS_ONCE(p) = (typeof(*v) __force __rcu *)(v); \
 	} while (0)
 
 /**