Hyperv: Event handling code refactor.

[howard0su_gmail.com]

Summary

1. use taskqueue instead of swi
2. scan the interrupt flags in filter
3. disable ringbuffer interrupt mask in filter to ensure no unnecessary interrupts

[howard0su_gmail.com]

drain and free taskqueue.

[sepherosa_gmail.com]

Does this mean, you are going to have at least one taskq, i.e. one thread, per-device? I think its probably too excessive for the already slow VM.

[howard0su_gmail.com]

In D4920#104502, @sepherosa_gmail.com wrote:

Does this mean, you are going to have at least one taskq, i.e. one thread, per-device? I think its probably too excessive for the already slow VM.

Yes. We can use taskqueue_swi instead creating another queue for each channel. but i noticed most NIC driver is using this approach. so I use the same way here.

[sepherosa_gmail.com]

In D4920#104503, @howard0su_gmail.com wrote:

In D4920#104502, @sepherosa_gmail.com wrote:

Does this mean, you are going to have at least one taskq, i.e. one thread, per-device? I think its probably too excessive for the already slow VM.

Yes. We can use taskqueue_swi instead creating another queue for each channel. but i noticed most NIC driver is using this approach. so I use the same way here.

Well, as I said, I think its an overkill to have a taskq per-channel. How about just replace the swi you removed in this patch w/ a taskq instead.

[howard0su_gmail.com]

Create taskqueue per CPU instead of per channel

[sepherosa_gmail.com]

Looks good to me

sys/dev/hyperv/vmbus/hv_vmbus_drv_freebsd.c

498 ↗

(On Diff #12315)

M_WAITOK here? We don't want it to fail

[howard0su_gmail.com]

address review comments

[decui_microsoft.com]

LGTM

sys/dev/hyperv/vmbus/hv_vmbus_drv_freebsd.c

527 ↗

(On Diff #12352)

Typo: ahdnle --> handle.

[howard0su_gmail.com]

Fix typo in comment

[howard0su_gmail.com]

iperf result shows no performance regression after this change.

[howard0su_gmail.com]

Rebase to current