I've been investigating, and I strongly suspect these commits for bug 1759723 are to blame.
f0aff9ccc834 genirq/affinity: assign vectors to all possible CPUs
9403a13fd07e blk-mq: simplify queue mapping & schedule with each possisble CPU
As far as I can tell there are 8 other fixes we'd need to look at including, either addressing the same original commit that these patches addressed or addressing bugs related to these two patches:
16ccfff28976 nvme: pci: pass max vectors as num_possible_cpus() to pci_alloc_irq_vectors
8b834bff1b73 scsi: hpsa: fix selection of reply queue
adbe552349f2 scsi: megaraid_sas: fix selection of reply queue
b5b6e8c8d3b4 scsi: virtio_scsi: fix IO hang caused by automatic irq vector affinity
7bed45954b95 blk-mq: make sure hctx->next_cpu is set correctly
a1c735fb7907 blk-mq: make sure that correct hctx->next_cpu is set
bffa9909a6b4 blk-mq: don't keep offline CPUs mapped to hctx 0
d3056812e7df genirq/affinity: Spread irq vectors among present CPUs as far as possible
Or we can revert those two patches. I'm considering both options, will provide one or two test kernels soon.
I've been investigating, and I strongly suspect these commits for bug 1759723 are to blame.
f0aff9ccc834 genirq/affinity: assign vectors to all possible CPUs
9403a13fd07e blk-mq: simplify queue mapping & schedule with each possisble CPU
As far as I can tell there are 8 other fixes we'd need to look at including, either addressing the same original commit that these patches addressed or addressing bugs related to these two patches:
16ccfff28976 nvme: pci: pass max vectors as num_possible_cpus() to pci_alloc_ irq_vectors
8b834bff1b73 scsi: hpsa: fix selection of reply queue
adbe552349f2 scsi: megaraid_sas: fix selection of reply queue
b5b6e8c8d3b4 scsi: virtio_scsi: fix IO hang caused by automatic irq vector affinity
7bed45954b95 blk-mq: make sure hctx->next_cpu is set correctly
a1c735fb7907 blk-mq: make sure that correct hctx->next_cpu is set
bffa9909a6b4 blk-mq: don't keep offline CPUs mapped to hctx 0
d3056812e7df genirq/affinity: Spread irq vectors among present CPUs as far as possible
Or we can revert those two patches. I'm considering both options, will provide one or two test kernels soon.