On Dec 26, 2019, at 4:09 PM, Jason Wang <jasowang@xxxxxxxxxx> wrote:We have considered both the PCI MSI-x like dynamic association design and fix mapping design.
On 2019/12/25 äå11:20, Liu, Jiang wrote:
On Dec 25, 2019, at 6:20 PM, Jason Wang <jasowang@xxxxxxxxxx> wrote:Hi Jason,
On 2019/12/25 äå10:50, Zha Bin wrote:
From: Liu Jiang <gerry@xxxxxxxxxxxxxxxxx>Thanks for the patch. Two questions after a quick glance:
Userspace VMMs (e.g. Qemu microvm, Firecracker) take advantage of using
virtio over mmio devices as a lightweight machine model for modern
cloud. The standard virtio over MMIO transport layer only supports one
legacy interrupt, which is much heavier than virtio over PCI transport
layer using MSI. Legacy interrupt has long work path and causes specific
VMExits in following cases, which would considerably slow down the
1) read interrupt status register
2) update interrupt status register
3) write IOAPIC EOI register
We proposed to update virtio over MMIO to version 3 to add the
following new features and enhance the performance.
1) Support Message Signaled Interrupt(MSI), which increases the
interrupt performance for virtio multi-queue devices
2) Support per-queue doorbell, so the guest kernel may directly write
to the doorbells provided by virtio devices.
The following is the network tcp_rr performance testing report, tested
with virtio-pci device, vanilla virtio-mmio device and patched
virtio-mmio device (run test 3 times for each case):
netperf -t TCP_RR -H 192.168.1.36 -l 30 -- -r 32,1024
Virtio-PCI Virtio-MMIO Virtio-MMIO(MSI)
trans/s 9536 6939 9500
trans/s 9734 7029 9749
trans/s 9894 7095 9318
1) In PCI we choose to support MSI-X instead of MSI for having extra flexibility like alias, independent data and address (e.g for affinity) . Any reason for not start from MSI-X? E.g having MSI-X table and PBA (both of which looks pretty independent).
Thanks for reviewing patches on Christmas Day:)
The PCI MSI-x has several advantages over PCI MSI, mainly
1) support 2048 vectors, much more than 32 vectors supported by MSI.
2) dedicated address/data for each vector,
3) per vector mask/pending bits.
The proposed MMIO MSI extension supports both 1) and 2),
Aha right, I mis-read the patch. But more questions comes:
1) The association between vq and MSI-X vector is fixed. This means it can't work for a device that have more than 2047 queues. We probably need something similar to virtio-pci to allow a dynamic association.
The fix mapping design simplifies both the interrupt configuration process and VMM implementations.
And the virtio mmio transport layer is mainly used by light VMMs to support small scale virtual machines,
2048 vectors should be enough for these usage cases.
So the fix mapping design has been used.
2) The mask and unmask control is missedWe want to make VMM implementations as simple as possible:)
but the extension doesnât support 3) because
we noticed that the Linux virtio subsystem doesnât really make use of interrupt masking/unmasking.
Not directly used but masking/unmasking is widely used in irq subsystem which allows lots of optimizations.
On the other hand, we want to simplify VMM implementations as simple as possible, and mimicking the PCI MSI-x
will cause some complexity to VMM implementations.
I agree to simplify VMM implementation, but it looks to me introducing masking/pending won't cost too much code in the VMM implementation. Just new type of command for VIRTIO_MMIO_MSI_COMMAND.
And based on following observations, we have disabled support of mask/unmask,
1) MSI is edge triggering, which means it wonât be shared with other interrupt sources,
so masking/unmasking wonât be used for normal interrupt management logic.
2) Linux virtio mmio transport layer doesnât support suspend/resume yet, so thereâs no need to quiesce the device by masking interrupts.
3) The legacy PCI 2.2 devices doesnât support irq masking/unmasking, so irq masking/unmasking may be optional operations.
So we skipped support of irq masking/unmasking. We will recheck whether irq masking/unmasking is mandatory for MMIO devices.
On the other hand, we may enhance the spec to define command codes for masking/unmasking, and VMM may optionally support masking/unmasking.
2) It's better to split notify_multiplexer out of MSI support to ease the reviewers (apply to spec patch as well)Great suggestion, we will try to split the patch.