On Wed, 4 Jun 2008, Nick Piggin wrote:

On Wednesday 04 June 2008 07:44, Trent Piepho wrote:

quoted

On Tue, 3 Jun 2008, Matthew Wilcox wrote:

quoted

quoted

I don't understand why you keep talking about DMA.  Are you talking
about ordering between readX() and DMA?  PCI proides those guarantees.

I guess you haven't been reading the whole thread.  The reason it started
was because gcc can re-order powerpc (and everyone else's too) IO accesses
vs accesses to cachable memory (but not spin-locks), which ends up only
being a problem with coherent DMA.

I don't think it is only a problem with coherent DMA.

CPU0                         CPU1
mutex_lock(mutex);
writel(something, DATA_REG);
writel(GO, CTRL_REG);
started = 1;

 	(A)

mutex_unlock(mutex);
                            mutex_lock(mutex);

 	(B)

                            if (started)
                              /* oops, this can reach device before GO */
                              writel(STOP, CTRL_REG);

The locks themselves should have (and do have) ordering operations to insure
gcc and/or the cpu can't move a store or load outside the locked region. 
Generally you need that to keep stores/loads to cacheable memory inside the
critical area, much less I/O operations.  Otherwise all you have to do is
replace writel(something, ...) with shared_data->something = ...  and there's
an obvious problem.  In your example, gcc currently can and will move the GO
operation to point A (if it can figure out that CTRL_REG and started aren't
aliased), but that's not a problem.  If it could move it to B that would be a
problem, but it can't.

Other than coherent DMA, I don't think there is any reason to care if I/O
accessors are strongly ordered wrt load/stores to cacheable memory.  locking
and streaming DMA sync operations already need to have ordering, so they don't
require all I/O to be ordered wrt all cacheable memory.