On Tue, Aug 11, 2026 at 10:25 PM Liu, Changcheng <[email protected]> wrote:
>
> On Tue, Aug 11, 2026 at 10:37:28AM -0400, Stefan Hajnoczi wrote:
> > External email: Use caution opening links or attachments
> >
> >
> > On Wed, Aug 5, 2026 at 10:40 PM Liu, Changcheng <[email protected]> wrote:
> > >
> > > On Wed, Aug 05, 2026 at 01:45:58PM -0400, Stefan Hajnoczi wrote:
> > > >
> > > > On Sat, Aug 1, 2026 at 2:21 AM Liu, Changcheng <[email protected]> 
> > > > wrote:
> > > > >
> > > > > When creating many virtio-blk devices, probe starts failing with
> > > > > -ENOSPC (-28) because the system runs out of interrupt vectors:
> > > > >
> > > > >   virtio_blk virtioNNN: probe with driver virtio_blk failed with 
> > > > > error -28
> > > > >
> > > > > By default virtio-blk uses managed IRQ affinity, which reserves an
> > > > > interrupt vector on every CPU for each device (about nr_cpus vectors 
> > > > > per
> > > > > device). On a host with many CPUs and many devices this exhausts the
> > > > > vectors long before all devices are probed.
> > > > >
> > > > > Add use_irq_affinity (default true, no behaviour change). Set it to 0 
> > > > > to
> > > > > use unmanaged interrupts, so each device only uses a couple of vectors
> > > > > instead of one per CPU, allowing far more devices to probe.
> > > > >
> > > > > Signed-off-by: Liu, Changcheng <[email protected]>
> > > >
> > > > There is already a num_request_queues module parameter for cases where
> > > > the user wishes to reduce the number of virtqueues. Did you benchmark
> > > > that and decide the performance of many queues sharing a single irq
> > > > makes it worth adding another module parameter?
> > > >
> > > > Stefan
> > >
> > > num_request_queues does not address this case because each virtio-blk 
> > > device
> > > already has only one request virtqueue when the failure occurs, so the 
> > > queue
> > > count cannot be reduced further.
> >
> > Okay.
> >
> > > The issue is reproduced after probing approximately 800 virtio-blk 
> > > devices on
> > > a 64-core bare-metal host. With managed IRQ affinity, the IRQ-vector 
> > > reservations
> > > are eventually exhausted and probing additional devices fails with 
> > > -ENOSPC.
> > > Disabling managed IRQ affinity avoids these per-CPU vector reservations 
> > > and
> > > allows more devices to be probed successfully.
> >
> > I don't follow. My understanding was that non-managed IRQ vectors are
> > allocated by request_irq(), which is called during virtio_find_vqs().
> > Probing 800 devices would still require 800 * (1 virtqueue irq + 1
> > config change irq) = 1,600 vectors. How come non-managed IRQs do not
> > hit the limit here?
> >
> > Thanks,
> > Stefan
> >
>
> With one request queue, virtio-blk allocates two MSI-X interrupts: an
> unmanaged configuration interrupt and a managed request-queue interrupt.
>
> The request-queue interrupt's managed affinity mask covers all possible
> CPUs. On x86, irq_matrix_reserve_managed() reserves one vector slot on
> every CPU in that mask so the interrupt can migrate safely during CPU
> hotplug. Consequently, every virtio-blk device consumes one managed
> vector slot on every CPU for its lifetime.
>
> Once any CPU exhausts its roughly 200 usable vector slots, managed vector
> allocation fails even if other CPUs still have capacity. virtio-pci then
> falls back to its shared-vector MSI-X allocation, for which the affinity
> descriptor is discarded and the interrupts are unmanaged. Allocation can
> continue using capacity on the other CPUs, explaining the observed limit
> of approximately 800 devices.
>
> When managed affinity is disabled, both interrupts are activated on
> individual CPUs and distributed by the vector allocator. APIC vector
> numbers are per-CPU, so the same vector number can be reused on different
> CPUs. On a 64-CPU system, 800 devices therefore consume approximately
> 800 * 2 / 64 = 25 vector slots per CPU.

Hi Changcheng,
Thanks for the explanation. Can you update the commit message to
mention that non-managed irqs stay below the limit (800 * 2 / 64 = 25
vector slots per CPU) because they are allocated to the CPU with the
lowest number by the vector allocator?

Other than that:
Reviewed-by: Stefan Hajnoczi <[email protected]>

Reply via email to