RE: https://vt.social/@lina/115723897365864314
Never mind the trick of disabling SMT only for some cores. It looks like there's a kernel bug and this causes IRQs for NVMe to be lost on queues for the cores with SMT disabled (with the second logical processor offlined), possibly due to a bad interaction with suspend...
Oh well. Kernel bugs for uncommon configurations seem to be par for the course these days...