MMIO and gcc re-ordering issue

[Paul Mackerras]

Nick Piggin writes:

>                 /* turn off LED */
>                 val64 = readq(&bar0->adapter_control);
>                 val64 = val64 &(~ADAPTER_LED_ON);
>                 writeq(val64, &bar0->adapter_control);
>                 s2io_link(nic, LINK_DOWN);
>         }
>         clear_bit(__S2IO_STATE_LINK_TASK, &(nic->state));
> 
> Now I can't say definitively that this is going to be wrong on
> powerpc, because I don't know the code well enough. But I'd be
> 90% sure that the unlock is not supposed to be visible to
> other CPUs before the writeqs are queued to the card. On x86 it
> wouldn't be.

Interestingly, there is also a store to cacheable memory
(nic->device_enabled_once), but no smp_wmb or equivalent before the
clear_bit.  So there are other potential problems here besides the I/O
related ones.

Anyway, I have done some tests on a dual G5 here with putting a sync
on both sides of the store in writel etc. (i.e. making readl/writel
strongly ordered w.r.t. everything else), and as you predicted, there
wasn't a noticeable performance degradation, at least not on the
couple of things I tried.  So I am now inclined to accept your
suggestion that we should do that.  I should probably do some similar
checks on POWER6 and a few other machines first, though.

Paul.