On Wed, Feb 13, 2019 at 10:41:55PM +0100, Thomas Gleixner wrote:
> Btw, while I have your attention. There popped up an issue recently related
> to that affinity logic.
>
> The current implementation fails when:
>
> /*
> * If there aren't any vectors left after applying the pre/post
> * vectors don't bother with assigning affinity.
> */
> if (nvecs == affd->pre_vectors + affd->post_vectors)
> return NULL;
>
> Now the discussion arised, that in that case the affinity sets are not
> allocated and filled in for the pre/post vectors, but somehow the
> underlying device still works and later on triggers the warning in the
> blk-mq code because the MSI entries do not have affinity information
> attached.
>
> Sure, we could make that work, but there are several issues:
>
> 1) irq_create_affinity_masks() has another reason to return NULL:
> memory allocation fails.
>
> 2) Does it make sense at all.
>
> Right now the PCI allocator ignores the NULL return and proceeds without
> setting any affinities. As a consequence nothing is managed and everything
> happens to work.
>
> But that happens to work is more by chance than by design and the warning
> is bogus if this is an expected mode of operation.
>
> We should address these points in some way.
Ah, yes, that's a mistake in the nvme driver. It is assuming IO queues are
always on managed interrupts, but that's not true if when only 1 vector
could be allocated. This should be an appropriate fix to the warning:
---
diff --git a/drivers/nvme/host/pci.c b/drivers/nvme/host/pci.c
index 022ea1ee63f8..f2ccebe1c926 100644
--- a/drivers/nvme/host/pci.c
+++ b/drivers/nvme/host/pci.c
@@ -506,7 +506,7 @@ static int nvme_pci_map_queues(struct blk_mq_tag_set *set)
* affinity), so use the regular blk-mq cpu mapping
*/
map->queue_offset = qoff;
- if (i != HCTX_TYPE_POLL)
+ if (i != HCTX_TYPE_POLL && dev->num_vecs > 1)
blk_mq_pci_map_queues(map, to_pci_dev(dev->dev),
offset);
else
blk_mq_map_queues(map);
--
