On Jul  5 22:24, Jinhao Fan wrote:
> Add property "ioeventfd" which is enabled by default. When this is
> enabled, updates on the doorbell registers will cause KVM to signal
> an event to the QEMU main loop to handle the doorbell updates.
> Therefore, instead of letting the vcpu thread run both guest VM and
> IO emulation, we now use the main loop thread to do IO emulation and
> thus the vcpu thread has more cycles for the guest VM.
> 
> Since ioeventfd does not tell us the exact value that is written, it is
> only useful when shadow doorbell buffer is enabled, where we check
> for the value in the shadow doorbell buffer when we get the doorbell
> update event.
> 
> IOPS comparison on Linux 5.19-rc2: (Unit: KIOPS)
> 
> qd           1   4  16  64
> qemu        35 121 176 153
> ioeventfd   41 133 258 313
> 
> Changes since v3:
>  - Do not deregister ioeventfd when it was not enabled on a SQ/CQ
> 
> Signed-off-by: Jinhao Fan <fanjinhao21s@ict.ac.cn>
> ---
>  hw/nvme/ctrl.c | 114 ++++++++++++++++++++++++++++++++++++++++++++++++-
>  hw/nvme/nvme.h |   5 +++
>  2 files changed, 118 insertions(+), 1 deletion(-)
> 

We have a regression following this patch that we need to address.

With this patch, issuing a reset on the device (`nvme reset /dev/nvme0`
will do the trick) causes QEMU to hog my host cpu at 100%.

I'm still not sure what causes this. The trace output is a bit
inconclusive still.

I'll keep looking into it.