Re: [RFC] lib: set/get max memzone segments

[Thomas Monjalon <thomas@monjalon.net>]

20/04/2023 20:20, Tyler Retzlaff:
> On Thu, Apr 20, 2023 at 09:43:28AM +0200, Thomas Monjalon wrote:
> > 19/04/2023 16:51, Tyler Retzlaff:
> > > On Wed, Apr 19, 2023 at 11:36:34AM +0300, Ophir Munk wrote:
> > > > In current DPDK the RTE_MAX_MEMZONE definition is unconditionally hard
> > > > coded as 2560.  For applications requiring different values of this
> > > > parameter – it is more convenient to set the max value via an rte API -
> > > > rather than changing the dpdk source code per application.  In many
> > > > organizations, the possibility to compile a private DPDK library for a
> > > > particular application does not exist at all.  With this option there is
> > > > no need to recompile DPDK and it allows using an in-box packaged DPDK.
> > > > An example usage for updating the RTE_MAX_MEMZONE would be of an
> > > > application that uses the DPDK mempool library which is based on DPDK
> > > > memzone library.  The application may need to create a number of
> > > > steering tables, each of which will require its own mempool allocation.
> > > > This commit is not about how to optimize the application usage of
> > > > mempool nor about how to improve the mempool implementation based on
> > > > memzone.  It is about how to make the max memzone definition - run-time
> > > > customized.
> > > > This commit adds an API which must be called before rte_eal_init():
> > > > rte_memzone_max_set(int max).  If not called, the default memzone
> > > > (RTE_MAX_MEMZONE) is used.  There is also an API to query the effective
> > > > max memzone: rte_memzone_max_get().
> > > > 
> > > > Signed-off-by: Ophir Munk <ophirmu@nvidia.com>
> > > > ---
> > > 
> > > the use case of each application may want a different non-hard coded
> > > value makes sense.
> > > 
> > > it's less clear to me that requiring it be called before eal init makes
> > > sense over just providing it as configuration to eal init so that it is
> > > composed.
> > 
> > Why do you think it would be better as EAL init option?
> > From an API perspective, I think it is simpler to call a dedicated function.
> > And I don't think a user wants to deal with it when starting the application.
> 
> because a dedicated function that can be called detached from the eal
> state enables an opportunity for accidental and confusing use outside
> the correct context.
> 
> i know the above prescribes not to do this but.
> 
> now you can call set after eal init, but we protect about calling it
> after init by failing. what do we do sensibly with the failure?

It would be a developer mistake which could be fix during development stage
very easily. I don't see a problem here.

> > > can you elaborate further on why you need get if you have a one-shot
> > > set? why would the application not know the value if you can only ever
> > > call it once before init?
> > 
> > The "get" function is used in this patch by test and qede driver.
> > The application could use it as well, especially to query the default value.
> 
> this seems incoherent to me, why does the application not know if it has
> called set or not? if it called set it knows what the value is, if it didn't
> call set it knows what the default is.

No the application doesn't know the default, it is an internal value.

> anyway, the use case is valid and i would like to see the ability to
> change it dynamically i'd prefer not to see an api like this be introduced
> as prescribed but that's for you folks to decide.
> 
> anyway, i own a lot of apis that operate just like the proposed and
> they're great source of support overhead. i prefer not to rely on
> documenting a contract when i can enforce the contract and implicit state
> machine mechanically with the api instead.
> 
> fwiw a nicer pattern for doing this one of framework influencing config
> might look something like this.
> 
> struct eal_config config;
> 
> eal_config_init(&config); // defaults are set entire state made valid
> eal_config_set_max_memzone(&config, 1024); // default is overridden
> 
> rte_eal_init(&config);

In general, we discovered that functions doing too much are bad
for usability and for ABI stability.
In the function eal_config_init() that you propose,
any change in the struct eal_config will be an ABI breakage.