On Mon, Dec 10, 2018 at 12:32:21AM +0200, Igor Stoppa wrote:

On 06/12/2018 11:44, Peter Zijlstra wrote:

quoted

On Wed, Dec 05, 2018 at 03:13:56PM -0800, Andy Lutomirski wrote:

quoted

quoted

+       if (op == WR_MEMCPY)
+               memcpy((void *)wr_poking_addr, (void *)src, len);
+       else if (op == WR_MEMSET)
+               memset((u8 *)wr_poking_addr, (u8)src, len);
+       else if (op == WR_RCU_ASSIGN_PTR)
+               /* generic version of rcu_assign_pointer */
+               smp_store_release((void **)wr_poking_addr,
+                                 RCU_INITIALIZER((void **)src));
+       kasan_enable_current();

Hmm.  I suspect this will explode quite badly on sane architectures
like s390.  (In my book, despite how weird s390 is, it has a vastly
nicer model of "user" memory than any other architecture I know
of...).  I think you should use copy_to_user(), etc, instead.  I'm not
entirely sure what the best smp_store_release() replacement is.
Making this change may also mean you can get rid of the
kasan_disable_current().

If you make the MEMCPY one guarantee single-copy atomicity for native
words then you're basically done.

smp_store_release() can be implemented with:

	smp_mb();
	WRITE_ONCE();

So if we make MEMCPY provide the WRITE_ONCE(), all we need is that
barrier, which we can easily place at the call site and not overly
complicate our interface with this.

Ok, so the 3rd case (WR_RCU_ASSIGN_PTR) could be handled outside of this
function.
But, since now memcpy() will be replaced by copy_to_user(), can I assume
that also copy_to_user() will be atomic, if the destination is properly
aligned? On x86_64 it seems yes, however it's not clear to me if this is the
outcome of an optimization or if I can expect it to be always true.

This would be a new contraint; one that needs to be documented and
verified by the various arch maintainers as they enable this feature on
their platform.