Hi Jiayu,

On 6/29/21 7:36 AM, Hu, Jiayu wrote:

Hi Maxime,

quoted

-----Original Message-----
From: Maxime Coquelin <redacted>
Sent: Monday, June 7, 2021 9:20 PM
To: Hu, Jiayu <redacted>; dev@dpdk.org
Cc: maxime.coquelin@redhat.com; Xia, Chenbo <redacted>;
Wang, Yinan [off-list ref]
Subject: Re: [PATCH 0/2] provide thread unsafe async registration functions

Hi Jiayu,

On 6/7/21 10:07 AM, Hu, Jiayu wrote:

quoted

Hi Maxime,

quoted

-----Original Message-----
From: Maxime Coquelin <redacted>
Sent: Friday, June 4, 2021 3:25 PM
To: Hu, Jiayu <redacted>; dev@dpdk.org
Cc: maxime.coquelin@redhat.com; Xia, Chenbo <redacted>;
Wang, Yinan [off-list ref]
Subject: Re: [PATCH 0/2] provide thread unsafe async registration
functions

Sorry, for previous blank reply.

On 5/28/21 10:11 AM, Jiayu Hu wrote:

quoted

Lock protection is needed during the vhost notifies the application
of device readiness, so the first patch is to add lock protection.
After performing locking, existed async vhost registration functions
will cause deadlock, as they acquire lock too. So the second patch
is to provide unsafe registration functions to support calling
within vhost callback functions.

I agree the callback should be always protected, and in that case
having a new thread-unsafe API makes sense for async registration.

Regarding backport, I'm not sure what we should do.

Backporting new API is a no-go, but with only backporting patch 1
async feature will be always broken on 20.11 LTS, right?

Yes, if only backporting this fix patch to 20.11 LTS, it may break
apps who call async registration functions inside vhost callbacks.

How about making this patch not a fix, but a new feature?

Async will be still broken in v20.11 in this case.
Maybe the better thing would be to remove async support in v20.11, as its
support was quite limited in that release anyway. Does that make sense?

The code of supporting async vhost are beyond 1000 lines. I am afraid that
removing such more code in 20.11 LTS may get objected. Can we note async
register/unregister only work in new_/destroy_device, and using them in
other vhost callback functions will cause deadlock in 20.11 LTS instead?
Does it make sense to you?

You are right, that removing 1L LoC in LTS might not be the best idea,
since it will cause conflicts when doing backports later on.

Maybe the best way is to return -1 at async registration time, with
logging an error?

Thanks,
Maxime

 Thanks,
Jiayu

quoted

Thanks,
Maxime

quoted

Thanks,
Jiayu

quoted

What do you think?

Thanks,
Maxime