On 08/02/2021 19.29, Arnd Bergmann wrote:

On Mon, Feb 8, 2021 at 10:25 AM Marc Zyngier [off-list ref] wrote:

quoted

On Thu, 04 Feb 2021 20:39:48 +0000, Hector Martin [off-list ref] wrote:

quoted

quoted

+{
+     return readl(ic->base + reg);

Please consider using the _relaxed accessors, as I don't think any of
these interacts with memory (apart from IPIs, of course).

MSI interrupts require serializing with DMA, so at the minimum I think there
needs to be something that ensures that DMA from device into memory
has completed before delivering the completion interrupt to a driver. This
may already be implied when the AIC is entered, but this is hard to know
without actual hardware specs.

I don't think this can be implied in any case, because if IRQ A fires 
and then the CPU speculates its way through AIC into the IRQ B handler, 
which reads DMA'd memory, then IRQ B fires and it has higher priority 
and *that* is what ends up getting returned from the event register 
first, the execution will commit with an ordering violation.

I'm pretty sure we need *some* level of explicit synchronization between 
reading the event register and actually delivering IRQs downstream. 
Using _relaxed might be okay, but we'd still need something where the 
isb() currently is in aic_handle_irq (though I admit I don't have a 
perfect picture of the memory ordering subtleties involved here yet).

Incidentally, just from the races and problems I've run into with 
trivial tests in m1n1, these CPUs seem to be *very* eager to speculate 
and I suspect they will help uncover race conditions in Linux...

-- 
Hector Martin "marcan" (marcan@marcan.st)
Public Key: https://mrcn.st/pub