Re: [dpdk-dev] [dpdk-stable] [PATCH v1 2/3] net/ring: fix eth_dev device pointer on allocation

["Gaëtan Rivet" <grive@u256.net>]

On 06/05/20 14:43 +0100, Ferruh Yigit wrote:
> On 5/6/2020 1:33 PM, Gaëtan Rivet wrote:
> > On 06/05/20 12:48 +0100, Ferruh Yigit wrote:
> >> On 5/5/2020 8:10 PM, Gaetan Rivet wrote:
> >>> When a net_ring device is allocated, its device pointer is not set
> >>> before calling rte_eth_dev_probing_finish, which is incorrect.
> >>>
> >>> The following:
> >>>   commit: 96cb19521147 ("net/ring: use EAL APIs in PMD specific API")
> >>>   commit: a6992e961050 ("net/ring: set ethernet device field")
> >>>
> >>> already attempted to fix this issue in 17.08, which was fine at the
> >>> time. Adding the hook rte_eth_dev_probing_finish() however created this
> >>> bug, as the eth_dev exposed when this hook is executed is expected to be
> >>> complete.
> >>>
> >>> Remove the prior attempts to fix the issue in rte_pmd_ring_probe() and
> >>> write the pointer properly in do_eth_dev_ring_create().
> >>>
> >>> Cc: stable@dpdk.org
> >>> Fixes: fbe90cdd776c ("ethdev: add probing finish function")
> >>> Cc: ferruh.yigit@intel.com
> >>> Cc: thomas@monjalon.net
> >>> Signed-off-by: Gaetan Rivet <grive@u256.net>
> >>
> >> <...>
> >>
> >>> @@ -325,10 +346,17 @@ do_eth_dev_ring_create(const char *name,
> >>>  	data->kdrv = RTE_KDRV_NONE;
> >>>  	data->numa_node = numa_node;
> >>>  
> >>> -	/* finally assign rx and tx ops */
> >>> +	/* assign rx and tx ops */
> >>>  	eth_dev->rx_pkt_burst = eth_ring_rx;
> >>>  	eth_dev->tx_pkt_burst = eth_ring_tx;
> >>>  
> >>> +	/* finally set the rte_device pointer in eth_dev. */
> >>> +	eth_dev->device = ring_device_from_name(name);
> >>> +	if (eth_dev->device == NULL) {
> >>> +		rte_errno = ENODEV;
> >>> +		goto error;
> >>> +	}
> >>> +
> >>>  	rte_eth_dev_probing_finish(eth_dev);
> >>>  	*eth_dev_p = eth_dev;
> >>
> >> Why not move the 'rte_eth_dev_probing_finish()' to the 'rte_pmd_ring_probe()',
> >> below where 'eth_dev->device' set?
> > 
> > Hi Ferruh,
> > 
> > Sure it would work. The reason I did it this way is two-fold:
> > 
> >   * I disliked having two places where eth_dev->device was conditionally
> >     set. It makes it harder to read rte_pmd_ring_probe.
> 
> Agree, what about using a 'goto' to have the assignment and
> 'rte_eth_dev_probing_finish()' in a single place?
> But check seems needed since creation may failed at that stage, if you think
> better check can be done on 'ret' instead of 'eth_dev'...
> 
> > 
> >   * I was actually thinking, doing this patch, that we should modify
> >     rte_eth_dev_allocate() to take an rte_device as parameter, as all
> >     eth_dev are meant to be backed by an rte_device. Keeping this in
> >     mind, I meant to move writing the pointer closer to the
> >     rte_eth_dev_allocate() call.
> 
> That is a bigger change, may affect many (if not all) PMDs, so I think this can
> be considered when that change is available.
> 
> And although that change may fix the issues that 'eth_dev->device' is not set,
> which we had a few times before, not sure it worth to change all PMDs and ethdev
> API directly couple with rte_device, instead of PMD being the glue. Can be
> discussed more on its own patch.
> 
> > 
> > But you are right that it is needlessly verbose, using
> > vdev_bus->find_device() to do this stuff. I'm ok with changing it as you
> > described if you prefer.
> > 
> 
> That was the concern, that is too much code to take a value which is already
> available a few level up the stack.

Ok, future-proofing is a bad habit so let's not do it.

I'm not a fan of the goto for the 'happy path' though. Another simple
way would be to bring the vdev pointer with me as we go down the stack.

I will send a v2 shortly, if it's too ugly to move the pointer down I'll
use the goto, or if you tell me you'd prefer it.

-- 
Gaëtan