On Thu, Dec 10, 2020 at 09:18:22PM +0100, Thomas Gleixner wrote:
> Prarit reported that depending on the affinity setting the
>
> ' irq $N: Affinity broken due to vector space exhaustion.'
>
> message is showing up in dmesg, but the vector space on the CPUs in the
> affinity mask is definitely not exhausted.
>
> Shung-Hsi provided traces and analysis which pinpoints the problem:
>
> The ordering of trying to assign an interrupt vector in
> assign_irq_vector_any_locked() is simply wrong if the interrupt data has a
> valid node assigned. It does:
>
> 1) Try the intersection of affinity mask and node mask
> 2) Try the node mask
> 3) Try the full affinity mask
> 4) Try the full online mask
>
> Obviously #2 and #3 are in the wrong order as the requested affinity
> mask has to take precedence.
>
> In the observed cases #1 failed because the affinity mask did not contain
> CPUs from node 0. That made it allocate a vector from node 0, thereby
> breaking affinity and emitting the misleading message.
>
> Revert the order of #2 and #3 so the full affinity mask without the node
> intersection is tried before actually affinity is broken.
>
> If no node is assigned then only the full affinity mask and if that fails
> the full online mask is tried.
>
> Fixes: d6ffc6ac83b1 ("x86/vector: Respect affinity mask in irq descriptor")
> Reported-by: Shung-Hsi Yu <[email protected]>
> Reported-by: Prarit Bhargava <[email protected]>
> Signed-off-by: Thomas Gleixner <[email protected]>
> Tested-by: Shung-Hsi Yu <[email protected]>
> Cc: [email protected]
Reviewed-by: Peter Xu <[email protected]>
--
Peter Xu
