Re: [dpdk-dev] [PATCH v8 12/14] eal/pci: Add rte_eal_dev_attach/detach() functions

[Thomas Monjalon <thomas.monjalon@6wind.com>]

2015-02-17 17:51, Tetsuya Mukawa:
> On 2015/02/17 10:48, Thomas Monjalon wrote:
> > 2015-02-16 13:14, Tetsuya Mukawa:
> >> +/* attach the new physical device, then store port_id of the device */
> >> +static int
> >> +rte_eal_dev_attach_pdev(struct rte_pci_addr *addr, uint8_t *port_id)
> >> +{
> >> +	uint8_t new_port_id;
> >> +	struct rte_eth_dev devs[RTE_MAX_ETHPORTS];
> >> +
> >> +	if ((addr == NULL) || (port_id == NULL))
> >> +		goto err;
> >> +
> >> +	/* save current port status */
> >> +	if (rte_eth_dev_save(devs, sizeof(devs)))
> >> +		goto err;
> >> +	/* re-construct pci_device_list */
> >> +	if (rte_eal_pci_scan())
> >> +		goto err;
> >> +	/* invoke probe func of the driver can handle the new device */
> >> +	if (rte_eal_pci_probe_one(addr))
> >> +		goto err;
> > You should get the port_id from the previous function instead of searching it.
> 
> I agree this will beautify this code, but actually to do like above
> changes current DPDK code much more, and it will not be clear, and not
> beautiful.
> 
> Could I explain it more.
> Problem is initialization sequence of virtual device and physical device
> are completely different.
> 
> (1) Attaching a physical device case
> - To return port id, pci_invoke_all_drivers() needs to return port id.
> - It means "devinit" of "struct rte_pci_driver" needs to return port_id.
> - "devinit" will be rte_eth_dev_init(). But if the device is virtual,
> this function is not implemented.
> 
> (2) Attaching virtual device case
> - To return port id from rte_eal_pci_probe_one(),
> pci_invoke_all_drivers() needs to return port id.
> - It means "init" of "struct rte_driver" needs to return port_id.
> - "init" will be implemented in PMD. But this function has different
> usage in physical device and virtual device.
> - Especially, In the case of physical device, "init" doesn't allocate
> eth device, so cannot return port id.
> 
> As a result, to remove  rte_eth_dev_save() and
> rte_eth_dev_get_changed_port(), below different functions needs to
> return port id.
>  - "devinit" of "struct rte_pci_driver".
>  - "init" of "struct rte_driver".

Yes, exactly,
I think you shouldn't hesitate to improve existing EAL code.
And I also think we should try to remove differences between virtual and
pci devices.

> That is why I implement like above.
> 
> >> +	/* get port_id enabled by above procedures */
> >> +	if (rte_eth_dev_get_changed_port(devs, &new_port_id))
> >> +		goto err;
> >> +
> >> +	*port_id = new_port_id;
> >> +	return 0;
> >> +err:
> >> +	RTE_LOG(ERR, EAL, "Driver, cannot attach the device\n");
> >> +	return -1;
> >> +}
> > [...]
> >
> >> +/* attach the new virtual device, then store port_id of the device */
> >> +static int
> >> +rte_eal_dev_attach_vdev(const char *vdevargs, uint8_t *port_id)
> >> +{
> >> +	char *args;
> >> +	uint8_t new_port_id;
> >> +	struct rte_eth_dev devs[RTE_MAX_ETHPORTS];
> >> +
> >> +	if ((vdevargs == NULL) || (port_id == NULL))
> >> +		goto err0;
> >> +
> >> +	args = strdup(vdevargs);
> >> +	if (args == NULL)
> >> +		goto err0;
> >> +
> >> +	/* save current port status */
> >> +	if (rte_eth_dev_save(devs, sizeof(devs)))
> >> +		goto err1;
> >> +	/* add the vdevargs to devargs_list */
> >> +	if (rte_eal_devargs_add(RTE_DEVTYPE_VIRTUAL, args))
> >> +		goto err1;
> >> +	/* parse vdevargs, then retrieve device name */
> >> +	get_vdev_name(args);
> >> +	/* walk around dev_driver_list to find the driver of the device,
> >> +	 * then invoke probe function o the driver */
> >> +	if (rte_eal_vdev_find_and_invoke(args, RTE_EAL_INVOKE_TYPE_PROBE))
> >> +		goto err2;
> > Again, you should get port_id from the attach procedure.
> >
> >> +	/* get port_id enabled by above procedures */
> >> +	if (rte_eth_dev_get_changed_port(devs, &new_port_id))
> >> +		goto err2;
> > [...]
> >
> >>  /**
> >> + * Uninitilization function called for each device driver once.
> >> + */
> >> +typedef int (rte_dev_uninit_t)(const char *name, const char *args);
> > Why do you need args for uninit?
> >
> 
> I just added for the case that finalization code of PMD needs it.
> But, probably "args" parameter can be removed.

Yes I think