On Aug 11 23:37, Jinhao Fan wrote:
> This patch series changes qemu-nvme's interrupt emulation to use event
> notifiers, which can ensure thread-safe interrupt delivery when iothread
> is used. In the first two patches, I convert qemu-nvme's IO emulation
> logic to send irq via eventfd, so that the actual assertion and
> deassertion is always done in the main loop thread with BQL held. In the
> third patch, support is added to send irq via KVM irqfd, bypassing
> qemu's MSI-x emulation. In the last patch, I add MSI-x mask handlers
> when irqfd is enabled so that qemu-nvme knows which vector is masked
> even when qemu's MSI-x emulation is bypassed.
>
> Jinhao Fan (4):
> hw/nvme: avoid unnecessary call to irq (de)assertion functions
> hw/nvme: add option to (de)assert irq with eventfd
> hw/nvme: use irqfd to send interrupts
> hw/nvme: add MSI-x mask handlers for irqfd
>
> hw/nvme/ctrl.c | 251 +++++++++++++++++++++++++++++++++++++++----
> hw/nvme/nvme.h | 7 ++
> hw/nvme/trace-events | 3 +
> 3 files changed, 243 insertions(+), 18 deletions(-)
>
> --
> 2.25.1
>
Hi Jinhao,
This series all looks pretty good to me. And, incidentally, it is also
super cool work :)
It can use a bit of clean up (the code duplication mentioned previously
by both Stefan and me) - but all the logic seems sound to me and my
testing is happy. Following up on my suggestion to drop the MSI-X
mask/unmasking, I gave the logic a thorough look it looks sound to me.
I'm gonna see what I can come up with for qtest. I suggest you just keep
going on iothread support.
Please post a v2 (all 4 patches) with suggested cleanups and we take it
from there.