Re: [dpdk-dev] [PATCH] app/testpmd: remove most uses of rte_eth_devices

[Thomas Monjalon <thomas@monjalon.net>]

19/07/2021 18:42, Ferruh Yigit:
> On 7/15/2021 2:20 PM, Paulis Gributs wrote:
> > This patch removes most uses of the global variable rte_eth_devices
> > from testpmd. This was done to avoid using the object directly which
> > applications should not do.
> > 
> > Most uses have been replaced with standard function calls, however
> > the use of it in the show_macs function could not be replaced as no
> > function call exists to get all mac addresses of a given port.
> > 
> > Signed-off-by: Paulis Gributs <paulis.gributs@intel.com>
[...]
> > @@ -857,16 +857,23 @@ dma_unmap_cb(struct rte_mempool *mp __rte_unused, void *opaque __rte_unused,
> >  	int ret;
> >  
> >  	RTE_ETH_FOREACH_DEV(pid) {
> > -		struct rte_eth_dev *dev =
> > -			&rte_eth_devices[pid];
> > +		struct rte_eth_dev_info dev_info;
> >  
> > -		ret = rte_dev_dma_unmap(dev->device, memhdr->addr, 0,
> > -					memhdr->len);
> > +		ret = eth_dev_info_get_print_err(pid, &dev_info);
> > +		if (ret != 0) {
> > +			TESTPMD_LOG(DEBUG,
> > +				    "unable to get device info for port %d on addr 0x%p,"
> > +				    "mempool unmapping will not be performed\n",
> > +				    pid, memhdr->addr);
> > +			continue;
> > +		}
> > +
> > +		ret = rte_dev_dma_unmap(dev_info.device, memhdr->addr, 0, memhdr->len);
> >  		if (ret) {
> >  			TESTPMD_LOG(DEBUG,
> >  				    "unable to DMA unmap addr 0x%p "
> >  				    "for device %s\n",
> > -				    memhdr->addr, dev->data->name);
> > +				    memhdr->addr, dev_info.device->name);
> >  		}
> >  	}
> >  	ret = rte_extmem_unregister(memhdr->addr, memhdr->len);
> > @@ -892,16 +899,22 @@ dma_map_cb(struct rte_mempool *mp __rte_unused, void *opaque __rte_unused,
> >  		return;
> >  	}
> >  	RTE_ETH_FOREACH_DEV(pid) {
> > -		struct rte_eth_dev *dev =
> > -			&rte_eth_devices[pid];
> > +		struct rte_eth_dev_info dev_info;
> >  
> > -		ret = rte_dev_dma_map(dev->device, memhdr->addr, 0,
> > -				      memhdr->len);
> > +		ret = eth_dev_info_get_print_err(pid, &dev_info);
> > +		if (ret != 0) {
> > +			TESTPMD_LOG(DEBUG,
> > +				    "unable to get device info for port %d on addr 0x%p,"
> > +				    "mempool mapping will not be performed\n",
> > +				    pid, memhdr->addr);
> > +			continue;
> > +		}
> > +		ret = rte_dev_dma_map(dev_info.device, memhdr->addr, 0, memhdr->len);
> >  		if (ret) {
> >  			TESTPMD_LOG(DEBUG,
> >  				    "unable to DMA map addr 0x%p "
> >  				    "for device %s\n",
> > -				    memhdr->addr, dev->data->name);
> > +				    memhdr->addr, dev_info.device->name);
> >  		}
> >  	}
> >  }
> 
> Hi Shahaf,
> 
> These callbacks are used to map/unmap anon memory and added on commit [1].
> 
> Can you please elaborate why it is required? And does xmem covers this
> functionality already?

The external memory must be registered for DMA.
It completes the feature of external memory,
so yes it is required.

> The concern I have is, it uses some DPDK details, like rte_device to implement
> functionality in a test applications (testpmd). If this is a required
> functionality for end user, it is very hard for them to implement this, and
> perhaps we should have some APIs/wrappers to help the users in that case.
> Or if it is not required, we can perhaps drop from testpmd.

I agree the API is bad.
It should be an API in every driver classes.

> But first I am trying to understand what functionality it brings, if it is
> something required by end user or not.

We should deprecate the API and introduce a new one.
Is it urgent to drop the API? Something you would like to do in 21.11?