Port-ids and NIC features

Port-ids and NIC features

[Marc Sune]

Dear all,

Two quick questions:

1) Is there a way to get the features available in the NIC? In 
particular, I cannot see a way to retrieve the *maximum speed* of the 
card and I need to determine whether is a FE, GE or 10GE card.

2) For a certain number of NICs installed in the system and bound to the 
IGB_UIO driver, is it safe to assume that the association phyisical port 
<-> port_id will always be the same? Even after reboot, and regardless 
of the order that are bound to the IGB_UIO driver (e.g. using 
pci_unbind.py)?

Best
marc

答复: Port-ids and NIC features

[安宏奎]

For the question one, maybe you can detect it before binding to igb_uio


-----邮件原件-----
发件人: dev-bounces-VfR2kkLFssw@public.gmane.org [mailto:dev-bounces-VfR2kkLFssw@public.gmane.org] 代表 Marc Sune
发送时间: 2013年8月14日 16:57
收件人: dev-VfR2kkLFssw@public.gmane.org
主题: [dpdk-dev] Port-ids and NIC features

Dear all,

Two quick questions:

1) Is there a way to get the features available in the NIC? In particular, I
cannot see a way to retrieve the *maximum speed* of the card and I need to
determine whether is a FE, GE or 10GE card.

2) For a certain number of NICs installed in the system and bound to the
IGB_UIO driver, is it safe to assume that the association phyisical port <->
port_id will always be the same? Even after reboot, and regardless of the
order that are bound to the IGB_UIO driver (e.g. using pci_unbind.py)?

Best
marc

Re: 答复: Port-ids and NIC features

[Stephen Hemminger]

On Thu, 15 Aug 2013 09:51:00 +0800
安宏奎 <anhk_-KN7UnAbNpbg@public.gmane.org> wrote:

> For the question one, maybe you can detect it before binding to igb_uio
> 
> 
> -----邮件原件-----
> 发件人: dev-bounces-VfR2kkLFssw@public.gmane.org [mailto:dev-bounces-VfR2kkLFssw@public.gmane.org] 代表 Marc Sune
> 发送时间: 2013年8月14日 16:57
> 收件人: dev-VfR2kkLFssw@public.gmane.org
> 主题: [dpdk-dev] Port-ids and NIC features
> 
> Dear all,
> 
> Two quick questions:
> 
> 1) Is there a way to get the features available in the NIC? In particular, I
> cannot see a way to retrieve the *maximum speed* of the card and I need to
> determine whether is a FE, GE or 10GE card.

Not at present. You need to look at the information in rte_pci_device
for the ethernet (see rte_eth_dev_info) and then do your own matching
against ID and features.

As DPDK evolves more of the hardware features need to be abstracted
to allow applications to be build once and run on varying hardware.
IMHO some of the low level features like read/write thresholds really
shouldn't be exposed through the API. Setting them correctly depends
too much on the underlying hardware, chip version, and driver and really
shouldn't be something the application writer has to worry about.


> 2) For a certain number of NICs installed in the system and bound to the
> IGB_UIO driver, is it safe to assume that the association phyisical port <->
> port_id will always be the same? Even after reboot, and regardless of the
> order that are bound to the IGB_UIO driver (e.g. using pci_unbind.py)?

Since probe is single threaded. The discovery order will be constant
on the same system. But will change if hardware added/removed.

Another future area of work that needs to be addressed is hot plugging
devices into a running application.

Re: 答复: Port-ids and NIC features

[Marc Sune]

Dear Stephen,

1) The problem is that, if the interfaces are binded to the IGB_UIO
driver *before* the DPDK-based application is started, there is no way,
as far as I've seen, to get such information. Of course, you cannot use
anymore ethtool or whatever to get at least the speeds that the card is
supporting. Btw, I don't need them, but offloads would probably be also
quite useful. So in general all the "features" or "capabilities"
supported by the NIC.

A general and elegant solution, IMHO would be to add to this in the info
of the NIC, probably to struct rte_eth_dev_info; something like:

enum rte_eth_features {
ETH_LINK_100MB_HD = 1 << 0,
ETH_LINK_100MB_FD = 1 << 1,
ETH_LINK_1000MB_HD = 1 << 2,
ETH_LINK_1000MB_FD = 1 << 3,
ETH_LINK_10000MB_HD = 1 << 4,
ETH_LINK_10000MB_FD = 1 << 5,

//More features to be added here:
};

struct rte_eth_dev_info {
//...
uint32_t capabilities; //Bitmap of enum rte_eth_features (could be also
an uint64_t)
};

I assume this informations are very easily gathered by the driver
itself, since it knows the exact NIC model and capabilities.

2) Are the ids determined when the NICs are being binded to the DPDK
driver, or when the IGB_UIO driver is insmoded? If it is the former,
then the ids won't be consistent across reboots or binding order, which
is not very desirable (IMHO), and actually is going to cause some
problems in our application.

Best
marc

On 15/08/13 16:57, Stephen Hemminger wrote:
> On Thu, 15 Aug 2013 09:51:00 +0800
> 安宏奎 <anhk_-KN7UnAbNpbg@public.gmane.org> wrote:
>
>> For the question one, maybe you can detect it before binding to igb_uio
>>
>>
>> -----邮件原件-----
>> 发件人: dev-bounces-VfR2kkLFssw@public.gmane.org [mailto:dev-bounces-VfR2kkLFssw@public.gmane.org] 代表 Marc Sune
>> 发送时间: 2013年8月14日 16:57
>> 收件人: dev-VfR2kkLFssw@public.gmane.org
>> 主题: [dpdk-dev] Port-ids and NIC features
>>
>> Dear all,
>>
>> Two quick questions:
>>
>> 1) Is there a way to get the features available in the NIC? In particular, I
>> cannot see a way to retrieve the *maximum speed* of the card and I need to
>> determine whether is a FE, GE or 10GE card.
> Not at present. You need to look at the information in rte_pci_device
> for the ethernet (see rte_eth_dev_info) and then do your own matching
> against ID and features.
>
> As DPDK evolves more of the hardware features need to be abstracted
> to allow applications to be build once and run on varying hardware.
> IMHO some of the low level features like read/write thresholds really
> shouldn't be exposed through the API. Setting them correctly depends
> too much on the underlying hardware, chip version, and driver and really
> shouldn't be something the application writer has to worry about.
>
>
>> 2) For a certain number of NICs installed in the system and bound to the
>> IGB_UIO driver, is it safe to assume that the association phyisical port <->
>> port_id will always be the same? Even after reboot, and regardless of the
>> order that are bound to the IGB_UIO driver (e.g. using pci_unbind.py)?
> Since probe is single threaded. The discovery order will be constant
> on the same system. But will change if hardware added/removed.
>
> Another future area of work that needs to be addressed is hot plugging
> devices into a running application.

Re: 答复: Port-ids and NIC features

[Stephen Hemminger]

On Fri, 16 Aug 2013 00:15:04 +0200
Marc Sune <marc.sune-kpkqNMk1I7M@public.gmane.org> wrote:

> Dear Stephen,
> 
> 1) The problem is that, if the interfaces are binded to the IGB_UIO
> driver *before* the DPDK-based application is started, there is no way,
> as far as I've seen, to get such information. Of course, you cannot use
> anymore ethtool or whatever to get at least the speeds that the card is
> supporting. Btw, I don't need them, but offloads would probably be also
> quite useful. So in general all the "features" or "capabilities"
> supported by the NIC.

The hack for now is doing:

		struct rte_eth_dev_info dev_info;
		rte_eth_dev_info_get(portid, &dev_info);

		if (strncmp(dev_info.driver_name, "rte_ixgbe", 9) == 0)
			/* do stuff for 10G */
                else
                        /* do stuff for 1G */

Obviously, the driver_name approach could be generalized to a lookup table.
But it really won't scale when many more drivers are added.
Also since speed is negotiable, code should really wait till link is
up to look at link speed and decide from that.

Since device is totally owned by user driver at this point, it
is perfectly possible to extend the existing drivers to do more PHY interaction
to query/set speed and duplex. The current code just assumes autonegotiation.
Since the Intel drivers are based off of BSD source there still is unused
code to do this (I think).