Re: array of pointers with rcu

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

On Sun, Jun 28, 2009 at 10:06:36PM +0300, Michael S. Tsirkin wrote:
> On Sun, Jun 28, 2009 at 09:14:11AM -0700, Paul E. McKenney wrote:
> > On Sun, Jun 28, 2009 at 04:22:24PM +0300, Michael S. Tsirkin wrote:
> > > Paul,
> > > I'd like to implement a static array of pointers with rcu.
> > > (Note that Documentation/RCU/arrayRCU.txt addresses only the case of
> > > static arrays where the data (rather than a pointer to the data) is
> > > located in each array element).
> > > The array is implemented today in kvm as follows:
> > > 
> > > struct kvm_io_bus {
> > >         int                   dev_count;
> > > #define NR_IOBUS_DEVS 6
> > >         struct kvm_io_device *devs[NR_IOBUS_DEVS];
> > > };
> > > 
> > > It's easy to use rcu_assign_pointer and
> > > rcu_dereference to access each value in the array,
> > > so that's fine.
> > > 
> > > However, I also have the dev_count value to handle.
> > > This value is usually used in loops like this
> > > 
> > > for (i = 0; i < bus->dev_count; ++i) {
> > > 	... uses bus->devs[i] ...
> > > }
> > 
> > So the count can vary, then.  In other words, someone might change the
> > value of ->dev_count from (say) 5 to (say) 2, implicitly invalidating
> > the last three ->devs pointers, correct?
> > 
> > Of course, this implicit invalidation might happen just after you
> > fetched the value of (say) ->devs[3].  Or just after you fetch the
> > value of ->dev_count, for that matter.
> 
> Yes. So I need to be careful to call synchronize_rcu after changing dev_count
> and before I can assume readers see an updated value.

Yep!!!

> > > I can assign the dev_count value with rcu_assign_pointer and even though
> > > it is not a pointer I think it should work fine.  However, to access
> > > dev_count I think that rcu_dereference will not do what I want:
> > > 
> > > #define rcu_dereference(p)     ({ \
> > >                                 typeof(p) _________p1 = ACCESS_ONCE(p);
> > > \
> > >                                 smp_read_barrier_depends(); \
> > >                                 (_________p1); \
> > >                                 })
> > > 
> > > 
> > > The use of dev_count is not through a dependency so
> > > smp_read_barrier_depends will not be enough on most architectures.
> > > 
> > > So it seems that what I really need is something like:
> > > #define rcu_read_value(p)     ({ \
> > >                                 typeof(p) _________p1 = ACCESS_ONCE(p);
> > > \
> > >                                 smp_rmb(); \
> > >                                 (_________p1); \
> > >                                 })
> > > 
> > > And maybe
> > > #define rcu_assign_value rcu_assign_pointer
> > > for symmetry.
> > > 
> > > Comments?
> > 
> > Hmmm...
> > 
> > What you would really need to do in the above scheme is to make careful
> > use of memory barriers trao order the accesses to ->dev_count and the the
> > elements of the ->devs[] array.
> 
> Right, that's my plan. I think that the strategy I outlined above where
> dev_count is always set with rcu_assign_value and read with
> rcu_read_value, all array entries are read after dev_count is read with
> rcu_dereference, and are assigned before dev count is updated with
> rcu_assign_pointer will work in this case.
> 
> What I was thinking is that maybe this pattern is generic
> enough to be of use to others. But maybe not.

This pattern has appeared a number of times, including in one of the
first uses of RCU (for SysV IPC), but it has always eventually been turned
into something that associates the array size with the array itself.

> > The usual trick to avoid such memory
> > barriers it to dynamically allocate the array, but putting the count
> > into a struct that includes the array, so that readers are guaranteed
> > to get a value of ->dev_count that matches the ->devs[] array.
> > 
> > Of course, you might have other reasons for the array to be statically
> > allocated.  In that case, one trick is to statically allocate two
> > arrays, each with its own ->dev_count.  Then a size-change update will
> > copy from the current array to the new array, set the value of the new
> > ->dev_count appropriately, and then use rcu_assign_pointer() to cause
> > new readers to see the updated array.  In short, use the approach
> > described for resizeable arrays in Documentation/RCU/arrayRCU.txt
> > even though the physical array stays the same size.
> > 
> > But again given that you only have six elements, why not just scan the
> > whole array unconditionally, using NULL pointers in the ->devs[] elements
> > to indicate that the corresponding element is invalid?  Then the normal
> > rcu_assign_pointer() and rcu_dereference() will work normally.
> 
> People are speaking about increasing the array size to 512. That's
> a nice 4K page but adding the number of entries gets us just above that
> which is kind of annoying.

OK, 512 is large enough that just unconditionally scanning the array
would not be a good plan.  ;-)

So, how about two arrays, each with count associated with it, and
switching back and forth between them?  That way readers simply pick
up the pointer to the current array/count, and are guaranteed that
the count matches the array.

							Thanx, Paul