Re: [dpdk-dev] [PATCH 0/2] Allow application set mempool handle

[Olivier Matz <olivier.matz@6wind.com>]

Hi Santosh,

On Tue, 4 Jul 2017 17:55:54 +0530, santosh <santosh.shukla@caviumnetworks.com> wrote:
> Hi Olivier,
> 
> On Friday 30 June 2017 07:42 PM, Olivier Matz wrote:
> 
> > Hi,
> >
> > On Tue, 20 Jun 2017 19:34:15 +0530, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote:  
> >> -----Original Message-----  
> >>> Date: Tue, 20 Jun 2017 16:07:17 +0530
> >>> From: Hemant Agrawal <hemant.agrawal@nxp.com>
> >>> To: Jerin Jacob <jerin.jacob@caviumnetworks.com>
> >>> CC: Santosh Shukla <santosh.shukla@caviumnetworks.com>,
> >>>  olivier.matz@6wind.com, dev@dpdk.org
> >>> Subject: Re: [PATCH 0/2] Allow application set mempool handle
> >>> User-Agent: Mozilla/5.0 (Windows NT 6.1; WOW64; rv:45.0) Gecko/20100101
> >>>  Thunderbird/45.8.0
> >>>
> >>> On 6/19/2017 6:31 PM, Jerin Jacob wrote:    
> >>>> -----Original Message-----    
> >>>>> Date: Mon, 19 Jun 2017 17:22:46 +0530
> >>>>> From: Hemant Agrawal <hemant.agrawal@nxp.com>
> >>>>> To: Santosh Shukla <santosh.shukla@caviumnetworks.com>,
> >>>>>  olivier.matz@6wind.com, dev@dpdk.org
> >>>>> CC: jerin.jacob@caviumnetworks.com
> >>>>> Subject: Re: [PATCH 0/2] Allow application set mempool handle
> >>>>> User-Agent: Mozilla/5.0 (Windows NT 6.1; WOW64; rv:45.0) Gecko/20100101
> >>>>>  Thunderbird/45.8.0
> >>>>>
> >>>>> On 6/1/2017 1:35 PM, Santosh Shukla wrote:    
> >>>>>> Some platform can have two different NICs for example external PCI Intel
> >>>>>> 40G card and Integrated NIC like vNIC/octeontx/dpaa2.
> >>>>>>
> >>>>>> Both NICs like to use their preferred pool e.g. external PCI card/ vNIC's
> >>>>>> preferred pool would be the ring based pool and octeontx/dpaa2 preferred would
> >>>>>> be ext-mempools.
> >>>>>> Right now, Framework doesn't support such case. Only one pool can be
> >>>>>> used across two different NIC's. For that, user has to statically set
> >>>>>> CONFIG_RTE_MEMPOOL_DEFAULT_OPS=<pool-name>.
> >>>>>>
> >>>>>> So proposing two approaches:
> >>>>>> Patch 1) Introducing eal option --pkt-mempool=<pool-name>
> >>>>>> Patch 2) Introducing ethdev API called _get_preferred_pool(), where PMD driver
> >>>>>> gets a chance to advertise their pool capability to the application. And based
> >>>>>> on that hint- application creates pools for that driver.    
> >>> If the system is having more than one heterogeneous ethernet device with
> >>> different mempool, the application has to create different mempool for each
> >>> of the ethernet device.
> >>>
> >>> However, let's take a case
> >>> As system has a DPAA2 eth device, which only work with dpaa2 mempools.    
> >> dpaa2 ethdev will return dpaa2 mempool as preferred handler.
> >>  
> >>> System also detect a standard PCI NIC, which can work with any software
> >>> mempool (e.g ring_mp_mc) or with dpaa2 mempool. Given the preference, PCI
> >>> NIC will have preferred as software mempool.
> >>> how the application will choose between these, if it want to create only one
> >>> mempool?    
> >> We need add some policy in common code to help application to choose
> >> in case if application interested in creating in one pool for
> >> heterogeneous cases. It is more of application problem, ethdev can
> >> return the preferred handler, let application choose interested in one.
> >> ethdev is depended on the specific mempool not any other object.
> >>
> >> We will provide option 1(eal argument based one) as one policy.More sophisticated
> >> policies we need add in application.
> >>
> >>  
> >>> Or, how the scheme will work if the application want to create only one
> >>> mempool?    
> >> option 1 (eal argument based) or we need to change the application to
> >> choose from available ethdev count and its preferred mempool handler.  
> > I also think the approach in this patchset is not that bad:
> >
> > - The first step is to allow the user to specify the mempool
> >   dynamically (eal arg).
> >
> >   One thing I don't really like is to have a mempool-related argument
> >   inside eal. It would be better if eal could provide a framework so
> >   that each libraries (ex: mbuf, mempool) can register their argument
> >   that could be changed through the command line or trough an API.
> >
> >   Without this, it introduces a sort of dependency between eal and
> >   mempool, which I don't think is sane.  
> 
> Yes, eal has no such framework for the non-eal library.
> 
> IIUC, then are you looking at something like below:
> - All non-eal library to register their callback function with eal.
> - EAL iterates through registered callbacks and calls them one by one.
> - EAL don't do the parsing and those non-eal libs do the parsing.
> - EAL passes char *string arg as input to those registered callback function.
> - It is up to those callback function to parse and find out i/p arg is correct
> or incorrect.
> - Having said that, then in the mempool case; We need to add new API to list 
> the number of supported mempool handles(by name) and then compare/match 
> i/p string with mempool handle(byname).
> 
> Are you referring to such framework? did I catch everything alright?

Here is how I see this feature (very high level).
The first step would be quite simple (no registration).
The EAL manages a key value database, and provides a key/value API like this:

  /* return NULL if key is not in database */
  const char *rte_eal_cfg_get(const char *key);
  /* value can be NULL to delete the key, return 0 on success */
  int rte_eal_cfg_set(const char *key, const char *value);

At startup, the EAL parses the arguments like this:
  --cfg=key:value
Example:
  --cfg=mbuf.default_pool:ring

Another way to set these options could be a config file (maybe the
librte_cfgfile could be useful for that, I don't know). Probably
something like:
  --cfgfile=file.conf

The EAL parsing layer calls rte_eal_cfg_set() 

Then, a library like librte_mbuf can query a specific key
through rte_eal_cfg_get("mbuf.default_pool"). No registration would
be needed. We'd need to define a convention for the key names.

It could be extented in a second step by adding a registration in
the constructor of the library:
  /* check_cb is a function that is called to check if the parsing is
   * correct. Maybe an opaque arg could be added too. */
  rte_eal_register_cfg(const char *key, rte_eal_cfg_check_cb_t check_cb);


I'm sure many people will have an opinion on this topic, which could
be different than mine.


> 
> > - The second step is to be able to ask to the eth devices which
> >   mempool they prefer. If there is only one kind of port, it's
> >   quite easy.
> >
> >   As suggested, more complexity could go in the application if
> >   required, or some helpers could be provided in the future.
> >
> >
> > I'm sending some comments as replies to the patches.
> >  
> If above eal framework approach is meeting your expectation then [1/4] need rework?
> Or you want to keep [1/4] patch and I'll send v2 patch incorporating
> your inline review comment, which one you prefer?

Adding a specific EAL argument --pkt-mempool could do the job for now.
But I'd be happy to see someone working on a generic cfg framework in EAL,
which seems to be a longer term solution, and helpful for other libs.

Some parts of EAL have currently no maintainer, which is a problem
to get a good feedback. But I guess a proposition on this topic
would trigger many comments.


Regards,
Olivier