[Qemu-devel] Re: [v1 PATCH 2/3]: Helper routines to use GLib threadpool infrastructure in 9pfs.

[Stefan Hajnoczi <stefanha@gmail.com>]

On Wed, Mar 16, 2011 at 2:20 PM, Venkateswararao Jujjuri (JV)
<jvrao@linux.vnet.ibm.com> wrote:
> On 3/16/2011 6:10 AM, Anthony Liguori wrote:
>> On 03/16/2011 04:20 AM, Stefan Hajnoczi wrote:
>>> On Tue, Mar 15, 2011 at 1:13 PM, Anthony Liguori<aliguori@us.ibm.com>  wrote:
>>>> Why even bothering signaling for completion with the virtio-9p threadpool?
>>>>
>>>> There's no sane guest that's going to poll on virtio-9p completion with
>>>> interrupts disabled and no timer.  Once we enable the I/O thread by default,
>>>> it won't even be necessary for the paio layer.
>>> It's not just about preventing livelock under extreme cases.  If you
>>> omit the signal then !CONFIG_IOTHREAD builds will see increased
>>> latency because virtfs request completion will piggy-back on other
>>> events that *do* interrupt the vcpu.
>>
>> But realistically, the timer is firing at a high enough frequency that I doubt
>> you'd even observe the latency.
>>
>> There's an easy solution here, just do some sniff testing to see if you can tell
>> the difference.  I bet you can't.
>>
>>>    I'm no fan of !CONFIG_IOTHREAD
>>> but skipping signals is currently bad practice and will continue to be
>>> until CONFIG_IOTHREAD is removed entirely.
>>>
>>> The proper solution would be a thin abstraction for thread-safe
>>> notification that compiles out signals when CONFIG_IOTHREAD is used.
>>> Then we have one place in QEMU that does signalling correctly and we
>>> can optimize it or remove CONFIG_IOTHREAD completely without having
>>> the burden of duplicating this code in several places.
>>
>> We have probably 5 different ways to wake up a CPU.  I don't think we should add
>> a new one just for this.
>>
>> !CONFIG_IOTHREAD needs to go away in the very near future.  I'd rather focus on
>> that.
>
> If that is the case, how about making VirtFS dependent on CONFIG_IOTHREAD
> during the configuration? This can help even if !CONFIG_IOTHREAD lingers around
> for some more time and would avoid any of the Stefan's concerns.

Sounds fair.

Stefan