Re: [PATCH v1 2/2] virtio-mmio: add features for virtio-mmio specification version 3

[Liu, Jiang]

On Jan 5, 2020, at 6:42 PM, Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> 
> On Thu, Dec 26, 2019 at 09:16:19PM +0800, Liu, Jiang wrote:
>>> 2) The mask and unmask control is missed
>>> 
>>> 
>>>> but the extension doesnât support 3) because
>>>> we noticed that the Linux virtio subsystem doesnât really make use of interrupt masking/unmasking.
> 
> Linux uses masking/unmasking in order to migrate interrupts between
> CPUs.
This is a limitation of the PCI MSI/MSIx spec.
To update the MSI/MSIx vector configuration, we need to write to msg_high/msg_low/msg_data registers.
Because write to three 32-bit registers is not an atomic operation on PCI bus, so it may cause incorrect interrupt delivery if interrupt happens after writing 1 or 2 registers.
When Intel remapping is enabled on x86 platforms, we donât need to mask/unmask PCI MSI/MSIx interrupts when setting affinity.
For MMIO MSI extension, we have special design to solve this race window. The flow to update MMIO MSI vector configuration is:
1) write msg_high
2) write msg_low
3) write msg_data
4) write the command register to update the vector configuration.
During step 1-3, the hardware/device backend driver only caches the value written. And update the vector configuration in step 4, so itâs an atomic operation now.
So mask/unmask becomes optional for MMIO MSI interrupts.

> 
> -- 
> MST