Il 12/05/2014 14:53, Michael S. Tsirkin ha scritto:
> In any case, whether writes synchronize with RCU or bypass it
> doesn't change the picture. In either case, writes are ordered
> against each other but not against reads. RCU does nothing except
> preventing dangling pointer accesses.
>
> Paolo
This is the only part I don't get.
RCU will make sure no VCPUs are running, won't it?
So it's a kind of full barrier.
The actual point where the new value becomes visible is where the write
happens, not where you do synchronize_rcu. This is the same for both
full-copy of the routing table or overwriting the entry.
However, the delay in servicing an older irqfd write can be arbitrary if
the scheduler decides not to run the irqfd_inject thread. Such an older
write might definitely read a newer routing entry. And that's the
invalid scenario according to the PCI spec.
Paolo
