Hi Jiayu,

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

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?

Thanks,
Maxime

Thanks,
Jiayu

quoted

What do you think?

Thanks,
Maxime

quoted

Jiayu Hu (2):
  vhost: fix lock on device readiness notification
  vhost: add thread unsafe async registration functions

 doc/guides/prog_guide/vhost_lib.rst |  12 +++
 lib/vhost/rte_vhost_async.h         |  42 ++++++++++
 lib/vhost/version.map               |   4 +
 lib/vhost/vhost.c                   | 161 +++++++++++++++++++++++++++---------
 lib/vhost/vhost_user.c              |   5 +-
 5 files changed, 180 insertions(+), 44 deletions(-)