Re: [Qemu-devel] [PATCH] msix: fix interrupt aggregation problem at the passthrough of NVMe SSD

["Michael S. Tsirkin" <mst@redhat.com>]

On Tue, Apr 09, 2019 at 02:14:56PM +0000, Zhuangyanying wrote:
> From: Zhuang Yanying <ann.zhuangyanying@huawei.com>
> 
> Recently I tested the performance of NVMe SSD passthrough and found that interrupts
> were aggregated on vcpu0(or the first vcpu of each numa) by /proc/interrupts,when
> GuestOS was upgraded to sles12sp3 (or redhat7.6). But /proc/irq/X/smp_affinity_list
> shows that the interrupt is spread out, such as 0-10, 11-21,.... and so on.
> This problem cannot be resolved by "echo X > /proc/irq/X/smp_affinity_list", because
> the NVMe SSD interrupt is requested by the API pci_alloc_irq_vectors(), so the
> interrupt has the IRQD_AFFINITY_MANAGED flag.
> 
> GuestOS sles12sp3 backport "automatic interrupt affinity for MSI/MSI-X capable devices",
> but the implementation of __setup_irq has no corresponding modification. It is still
> irq_startup(), then setup_affinity(), that is sending an affinity message when the
> interrupt is unmasked. The bare metal configuration is successful, but qemu will
> not trigger the msix update, and the affinity configuration fails.
> The affinity is configured by /proc/irq/X/smp_affinity_list, implemented at
> apic_ack_edge(), the bitmap is stored in pending_mask,
> mask->__pci_write_msi_msg()->unmask,
> and the timing is guaranteed, and the configuration takes effect.
> 
> The GuestOS linux-master incorporates the "genirq/cpuhotplug: Enforce affinity
> setting on startup of managed irqs" to ensure that the affinity is first issued
> and then __irq_startup(), for the managerred interrupt. So configuration is
> successful.
> 
> It now looks like sles12sp3 (up to sles15sp1, linux-4.12.x), redhat7.6
> (3.10.0-957.10.1) does not have backport the patch yet.
> "if (is_masked == was_masked) return;" can it be removed at qemu?
> What is the reason for this check?

The reason is simple:

The PCI spec says:

Software must not modify the Address or Data fields of an entry while it is unmasked.

It's a guest bug then?

> 
> Signed-off-by: Zhuang Yanying <ann.zhuangyanying@huawei.com>
> ---
>  hw/pci/msix.c | 4 ----
>  1 file changed, 4 deletions(-)
> 
> diff --git a/hw/pci/msix.c b/hw/pci/msix.c
> index 4e33641..e1ff533 100644
> --- a/hw/pci/msix.c
> +++ b/hw/pci/msix.c
> @@ -119,10 +119,6 @@ static void msix_handle_mask_update(PCIDevice *dev, int vector, bool was_masked)
>  {
>      bool is_masked = msix_is_masked(dev, vector);
>  
> -    if (is_masked == was_masked) {
> -        return;
> -    }
> -
>      msix_fire_vector_notifier(dev, vector, is_masked);
>  
>      if (!is_masked && msix_is_pending(dev, vector)) {
> -- 
> 1.8.3.1
>