On 8/4/2017 1:56 PM, Bruce Richardson wrote:
> On Fri, Aug 04, 2017 at 12:58:01PM +0100, Ferruh Yigit wrote:
>> On 8/3/2017 8:53 PM, Thomas Monjalon wrote:
>>> 03/08/2017 18:15, Stephen Hemminger:
>>>> On Thu, 3 Aug 2017 14:21:38 +0100
>>>> Bruce Richardson <bruce.richard...@intel.com> wrote:
>>>>> On Thu, Aug 03, 2017 at 01:21:35PM +0100, Chilikin, Andrey wrote:
>>>>>> To control some device-specific features public device-specific functions
>>>>>> rte_pmd_*.h are used.
>>>>>> But this solution requires applications to distinguish devices at runtime
>>>>>> and, depending on the device type, call corresponding device-specific
>>>>>> functions even if functions' parameters are the same.
>>>>>> IOCTL-like API can be added to ethdev instead of public device-specific
>>>>>> functions to address the following:
>>>>>> * allow more usable support of features across a range of NIC from
>>>>>>   one vendor, but not others
>>>>>> * allow features to be implemented by multiple NIC drivers without
>>>>>>   relying on a critical mass to get the functionality in ethdev
>>>>>> * there are a large number of possible device specific functions, and
>>>>>>   creating individual APIs for each one is not a good solution
>>>>>> * IOCTLs are a proven method for solving this problem in other areas,
>>>>>>   i.e. OS kernels.
>>>>>> Control requests for this API will be globally defined at ethdev level, 
>>>>>> so
>>>>>> an application will use single API call to control different devices from
>>>>>> one/multiple vendors.
>>>>>> API call may look like as a classic ioctl with an extra parameter for
>>>>>> argument length for better sanity checks:
>>>>>> int
>>>>>> rte_eth_dev_ioctl(uint16_t port, uint64_t ctl, void *argp,
>>>>>>         unsigned arg_length);
>>>>>> Regards,
>>>>>> Andrey  
>>>>> I think we need to start putting in IOCTLs for ethdevs, much as I hate
>>>>> to admit it, since I dislike IOCTLs and other functions with opaque
>>>>> arguments! Having driver specific functions I don't think will scale
>>>>> well as each vendor tries to expose as much of their driver specific
>>>>> functionality as possible.
>>>>> One other additional example: I discovered just this week another issue
>>>>> with driver specific functions and testpmd, when I was working on the
>>>>> meson build rework.
>>>>> * With shared libraries, when we do "ninja install" we want our DPDK
>>>>>   libs moved to e.g. /usr/local/lib, but the drivers moved to a separate
>>>>>   driver folder, so that they can be automatically loaded from that
>>>>>   single location by DPDK apps [== CONFIG_RTE_EAL_PMD_PATH].
>>>>> * However, testpmd, as well as using the drivers as plugins, uses
>>>>>   driver-specific functions, which means that it explicitly links
>>>>>   against the pmd .so files.
>>>>> * Those driver .so files are not in with the other libraries, so ld.so
>>>>>   does not find the pmd, and the installed testpmd fails to run due to
>>>>>   missing library dependencies.
>>>>> * The workaround is to add the drivers path to the ld load path, but we
>>>>>   should not require ld library path changes just to get DPDK apps to
>>>>>   work.
>>>>> Using ioctls instead of driver-specific functions would solve this.
>>>>> My 2c.
>>>> My 2c. No.
>>>> Short answer:
>>>> Ioctl's were a bad idea in Unix (per Dennis Ritchie et al) and are now
>>>> despised by Linux kernel developers. They provide an unstructured, 
>>>> unsecured,
>>>> back door for device driver abuse. Try to get a new driver in Linux with
>>>> a unique ioctl, and it will be hard to get accepted.
>>>> Long answer:
>>>> So far every device specific feature has fit into ethdev model. Doing ioctl
>>>> is admitting "it is too hard to be general, we need need an out". For 
>>>> something
>>>> that is a flag, it should fit into existing config model; ignoring silly 
>>>> ABI constraints.
>>>> For a real feature (think flow direction), we want a first class API for 
>>>> that.
>>>> For a wart, then devargs will do.
>>>> Give a good example of something that should be an ioctl. Don't build the
>>>> API first and then let it get cluttered.
>>> I agree with Stephen.
>>> And please do not forget that ioctl still requires an API:
>>> the argument that you put in ioctl is the API of the feature.
>>> So it is the same thing as defining a new function.
>> I am also not fan of the ioctl usage. I believe it hides APIs behind ids
>> and prevent argument check by compiler.
>> BUT, the number of the increasing PMD specific APIs are also worrying,
>> it is becoming harder to maintain, and I believe this is something NOT
>> sustainable in long run.
>> What about having *eth_dev_extended_ops* ?
>> As a part of the rte_eth_dev. This can be in the librte_ether library
>> but in a separated file.
>> And the APIs for these ops can be less strict on compatibility, and
>> easier to add.
>> Benefits of having this new dev_ops:
>> * Having an abstraction layer for common checks.
>> * Even feature is not generic for all NICs, still a few NICs can share
>> the ops.
>> * All APIs are in the same file makes it easy to see PMD specific APIs
>> comparing to scattered into various PMDs.
>> * This is very like ioctl approach, but APIs are more clear and
>> arguments can be verified.
> Sounds like an ethdev-staging library, where features can be put until
> such time as they get critical mass for acceptance and promoted to
> ethdev? It's sounds better than IOCTL, while giving the same benefits.
> I'd be happy enough with any solution that allows NIC features to be
> exposed that does not have functions limited to each individual driver,
> so that common functionality can be exposed to apps via an API even if
> only 2 drivers support it.

This is not decided yet, but to enable working on this for next release,
is a deprecation notice required to add a new field to "struct
rte_eth_dev" ?

"struct rte_eth_dev" is marked as "@internal", so I believe deprecation
notice is NOT required, but I would like to confirm.

>> Thanks,
>> ferruh
>>> The real debate is to decide if we want to continue adding more
>>> control path features in DPDK or focus on Rx/Tx.
> I don't see dropping control path as an option. It would severely limit
> the usefulness of DPDK.
> /Bruce

Reply via email to