> On Aug 8, 2017, at 1:02 PM, Stephen Hemminger <step...@networkplumber.org> 
> wrote:
> On Tue, 8 Aug 2017 17:28:19 +0000
> "Wiles, Keith" <keith.wi...@intel.com> wrote:
>> Fix format.
>>> On Aug 4, 2017, at 6:58 AM, Ferruh Yigit <ferruh.yi...@intel.com> 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* ?  
>> We had talk about adding something like device specific APIs to DPDK in the 
>> past, which to me are just IOCTL like APIs. The big problem with IOCTLs is 
>> trying to cram a bunch of specific requests into a very generic API and I do 
>> not like ioctl as defined in Linux/Unix today. The old IOCTLs calls are too 
>> opaque and difficult for compilers to test args and many other issues.
>> We talked about having a single API in rte_eth_dev that would allow a user 
>> to ask for and possible get a list of function pointers in a given structure 
>> for the requested type. If a user calls this API to get some feature from a 
>> given NIC he would get NULL or a pointer to a set of functions. The generic 
>> API in rte_eth would allow the user to request what structures or types of 
>> APIs it supports.
>> Using a specific API to get the list of APIs or supported features in a NIC, 
>> will allow the developer to request the set of APIs (in an array or some 
>> method). Then we have real APIs for specific control or requests and not a 
>> generic API like ioctl.
>> Cristian had suggested an API like this to make it easy to add any IOCTL 
>> like needs to a driver. We can define a set of structures that seem generic 
>> for some IOCTL like needs or just allow the NIC to define his own structures 
>> and APIs. Allowing the developer to define his own structures and APIs is 
>> not very generic or usable by the users, so I would lean toward defining 
>> structures set need today and expand those structures in the future or add 
>> more structures.
>> int rte_eth_dev_something(uint16_t port_id, const char *feature, void **obj);
>> Using strings we can define or the NIC vendor can define to ask for a 
>> pointer to a structure he knows via a driver header. Strings are good, 
>> because we can read them via the debug or print them out quickly instead of 
>> trying to use some lookup table. Plus we can have any length or characters 
>> for defining the structure request.
>> Just off the top of my head, but it can be changed if needed.
>>> 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.
>>> 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.
>>>> But this discussion would be better lead with some examples/requests.  
> The real question is how important is that DPDK is the playground for HW 
> features?
> My impression is that the current process is HW supplier driven "we support 
> offload of XYZZY”.

It is not really a playground as I see it, but defining clean usable solution 
for vendors and developers to gain access to HW features. Every vendor loves to 
expose its features and allowing them to create APIs called directly from an 
application will make it impossible for a generic application and we need to 
provide some structure to allow the application to determine what features are 
supported and how they gain access to those features.

> The existing control of device model is already a multi-path mess of config 
> API's, device args,
> and magic scripts. That needs to be addressed first.

Not sure where the ‘magic scripts’ support comes from :-), but I agree we need 
to clean up the current configuration of devices. I only really see two methods 
in the work I have done, which are device args strings and configuration from 
ethdev. The DPDK configuration is another problem in that it sometimes requires 
a huge command line to startup an application. I think I saw a patch to add 
better config file support, but it does not address everything IMO. Which BTW 
was one of the reasons I wrote the dpdk-run.py script to help solve this 
problem without having to change the current config file/command line interface.

> There is very little community input from users, that is the problem.

I believe I am a user with Pktgen and my usage of DPDK today, but I maybe a bit 
more of an expert then most users (I hope). I see these usability problems and 
I try to address them, but we all need to be aware of the usability aspect of 
DPDK for the casual user.

> IMHO if a new hardware feature can't be made to fit into a standard OS model 
> like Linux
> because it is too hard then adding it to DPDK is a mistake.

I guess I agree, but what is the Linux model anyway and how does that effect 
FreeBSD or Windows or name your next OS. We need to provide good hardware 
support for DPDK and I feel it maybe impossible to always define everything in 
a super clean generic way, this is why IOCTL were created and I think we can 
define a better solution then IOCTL as in Linux/Unix.


Reply via email to