On Sat, Oct 18, 2008 at 11:04:39PM +0900, Tetsuo Handa wrote:
quoted text
> Hello.
> 
> Paul E. McKenney wrote:
> > On Fri, Oct 17, 2008 at 05:32:43PM +0900, Kentaro Takeda wrote:
> > > However, to ensure the reader gets up-to-date value, we need to use 
> > > smp_read_barrier_depends() (which is expanded to "mb()" for SMP on 
> > > Alpha, "read_barrier_depends()" for SMP on H8300, "((void)0)" for SMP 
> > > on M68K-nommu, "((void)0)" for M68K, "do { } while (0)" otherwise) 
> > > whenever the reader fetches an element in a list.
> > 
> > Yep.  You will also need the ACCESS_ONCE() on the pointer fetch in order
> > to suppress aggressive compiler optimizations.  The rcu_dereference()
> > primitive packages them up nicely.
> 
> So, I should use rcu_dereference() rather than smp_read_barrier_depends().
> I see.

Yep!

quoted text
> > > > But fair enough.  How about the following?
> > > > 
> > > >         #define worm_dereference()        rcu_dereference()
> > > >         #define worm_assign_pointer()        rcu_assign_pointer()
> > > > 
> > > So, I understood that the rcu_dereference() and rcu_assign_pointer() 
> > > are not only for RCU. They are needed to ensure the reader gets 
> > > up-to-date value. Then, their names should be var_dereference() and 
> > > var_assign_pointer() or something, shouldn't they? The "rcu_" prefix 
> > > and comments on rcu_dereference in include/linux/rcupdate.h sound for 
> > > me that they are used for variables protected by RCU locking 
> > > mechanism only...
> > 
> > Well, there are 200+ uses of rcu_dereference() for RCU, so it would
> > 99.5%+ accurate to retain the "rcu_" prefix.  ;-)
> > 
> > Once we have several non-RCU uses, we can probably do a much better
> > job of coming up with a good name for the underlying independent-of-RCU
> > primitive.  So we should stick with rcu_dereference() as the name of the
> > underlying primitive for now, and re-evaluate the naming in a year or
> > after another five non-RCU uses of rcu_dereference() appear, whichever
> > comes later.  (My current guess for names are "pointer_subscribe()"
> > for rcu_dereference() and "pointer_publish()" for rcu_assign_pointer(),
> > but who knows?)
> > 
> > Fair enough?
> 
> Agreed.
> 
> > TOMOYO can tolerate reading the complete garbage that would appear if
> > the pointer was assigned before the pointed-to fields are initialized?
> > I must confess that I am having a hard time believing that.  Please
> > explain how this works.
> 
> Maybe I'm misunderstanding what "mb()" can do.
> 
> I inserted "mb()" between the initialization of the pointed-to field and
> the assignment of the pointer like
> 
>   static inline void list1_add_tail_mb(struct list1_head *new,
>                                        struct list1_head *head)
>   {
>           struct list1_head *pos = head;
>   
>           new->next = head;
>           mb(); /* Avoid out-of-order execution. */
>           while (pos->next != head)
>                   pos = pos->next;
>           pos->next = new;
>   }
> 
> to ensure that 'new->next' comes to point to 'head' before 'pos->next' comes to
> point to 'new', for arch/ia64/include/asm/system.h says:
> 
>   /*
>    * Macros to force memory ordering.  In these descriptions, "previous"
>    * and "subsequent" refer to program order; "visible" means that all
>    * architecturally visible effects of a memory access have occurred
>    * (at a minimum, this means the memory has been read or written).
>    *
>    *   wmb():     Guarantees that all preceding stores to memory-
>    *              like regions are visible before any subsequent
>    *              stores and that all following stores will be
>    *              visible only after all previous stores.
>    *   rmb():     Like wmb(), but for reads.
>    *   mb():      wmb()/rmb() combo, i.e., all previous memory
>    *              accesses are visible before all subsequent
>    *              accesses and vice versa.  This is also known as
>    *              a "fence."
>    *
>    * Note: "mb()" and its variants cannot be used as a fence to order
>    * accesses to memory mapped I/O registers.  For that, mf.a needs to
>    * be used.  However, we don't want to always use mf.a because (a)
>    * it's (presumably) much slower than mf and (b) mf.a is supported for
>    * sequential memory pages only.
>    */

The problem is that while wmb() and mb() do in fact order writes, they 
cannot order the other task's reads.  So if you do the following with
all variables initially zero:

	a = 1;
	smp_wmb(); /* use this instead of wmb() unless doing device I/O. */
	b = 1;

That will indeed make the assignment to "a" happen before the
assignment to "b".  But if the reader does the following:

	tmp_b = b;
	tmp_a = a;

There is nothing preventing these two reads from being reordered.
The reader might well see tmp_a==0&&tmp_b==1, which is what the
smp_wmb() was trying to prevent.

quoted text
> But http://www.rdrop.com/users/paulmck/scalability/paper/ordering.2007.09.19a.pdf says:
> 
>   On these systems, there are three orderings that must be accounted for:
>   1. Program order:
>      the order of the program's object code as seen by the CPU
>   2. Execution order:
>      The execution order can differ from program order due to both compiler and
>      CPU implementation optimizations.
>   3. Perceived order:
> 
> And I noticed that the comment in arch/ia64/include/asm/system.h uses
> "program order", not "execution order".
> Therefore, I have to also care "execution order".
> 
> I assumed that the "program order" == "execution order" and I thought:
> 
>   The reader will read the appended element (i.e. 'new->next') only if
>   the reader can reach (i.e. 'pos->next' == 'new') the appended element.
>   The reader won't read the appended element (i.e. 'new->next') as long as
>   the reader can't reach (i.e. 'pos->next' == 'head') the appended element.

This assumes something called "dependency ordering".  Dependency
ordering is intuitive, but unfortunately is not respected by DEC Alpha
or by some aggressive value-speculation compiler optimizations.

quoted text
> However, since "execution order" != "program order", you are mentioning that
> I have to call smp_read_barrier_depends() which can control "execution order".
> Am I right?

Partly.

See the section labelled "Alpha" on Page 4 of:

http://www.rdrop.com/users/paulmck/scalability/paper/ordering.2007.09.19a.pdf

for an explanation, though http://lkml.org/lkml/2008/2/2/255
does a much better job of explaining it.

							Thanx, Paul
--
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/