[dpdk-dev] [PATCH] vhost: Fix retrieval of numa information in PMD

Loftus, Ciara ciara.loftus at intel.com
Wed Apr 6 11:37:53 CEST 2016


> 
> On Wed, Apr 06, 2016 at 03:49:25PM +0900, Tetsuya Mukawa wrote:
> > On 2016/04/06 1:09, Ciara Loftus wrote:
> > > After some testing, it was found that retrieving numa information
> > > about a vhost device via a call to get_mempolicy is more
> > > accurate when performed during the new_device callback versus
> > > the vring_state_changed callback, in particular upon initial boot
> > > of the VM.  Performing this check during new_device is also
> > > potentially more efficient as this callback is only triggered once
> > > during device initialisation, compared with vring_state_changed
> > > which may be called multiple times depending on the number of
> > > queues assigned to the device.
> > >
> > > Reorganise the code to perform this check and assign the correct
> > > socket_id to the device during the new_device callback.
> > >
> > > Signed-off-by: Ciara Loftus <ciara.loftus at intel.com>
> > > ---
> > >  drivers/net/vhost/rte_eth_vhost.c | 28 ++++++++++++++--------------
> > >  1 file changed, 14 insertions(+), 14 deletions(-)
> > >
> > > diff --git a/drivers/net/vhost/rte_eth_vhost.c
> b/drivers/net/vhost/rte_eth_vhost.c
> > > index 4cc6bec..b1eb082 100644
> > > --- a/drivers/net/vhost/rte_eth_vhost.c
> > > +++ b/drivers/net/vhost/rte_eth_vhost.c
> > >
> >
> > Hi,
> >
> > I appreciate fixing it.
> > Just one worry is that state changed event may be occurred before new
> > device event.
> > The users should not call rte_eth_dev_socket_id() until new device event
> > comes, even if they catch queue state events.
> > Otherwise, they will get wrong socket id to call
> > rte_eth_rx/tx_queue_setup().
> 
> There is no way to guarantee that the socket id stuff would work
> perfectly in vhost, right? I mean, it's likely that virtio device
> would allocate memory from 2 or more sockets.
> 
> So, it doesn't matter too much whether it's set perfectly right
> or not. Instead, we should assign it with a saner value instead
> of a obvious wrong one when new_device() is not invoked yet. So,
> I'd suggest to make an assignment first based on vhost_dev (or
> whatever) struct, and then make it "right" at new_device()
> callback?

Thanks for the feedback.
At the moment with this patch numa_node is initially set to rte_socket_id() during  pmd init and then updated to the correct value during new_device.
Are you suggesting we set it again in between these two steps ("based on vhost_dev")? If so where do you think would be a good place?

Thanks,
Ciara

> 
> > So how about commenting it in 'rte_eth_vhost.h'?
> 
> It asks a different usage than other PMDs, which I don't think
> it's a good idea.
> 
> 	--yliu


More information about the dev mailing list