Re: [dpdk-dev] [PATCH] table: fix build error with gcc 8

["Dumitrescu, Cristian" <cristian.dumitrescu@intel.com>]

> -----Original Message-----
> From: Van Haaren, Harry
> Sent: Monday, April 9, 2018 5:38 PM
> To: Dumitrescu, Cristian <cristian.dumitrescu@intel.com>; Stephen
> Hemminger <stephen@networkplumber.org>; Singh, Jasvinder
> <jasvinder.singh@intel.com>; Richardson, Bruce
> <bruce.richardson@intel.com>
> Cc: dev@dpdk.org
> Subject: RE: [dpdk-dev] [PATCH] table: fix build error with gcc 8
> 
> > From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Dumitrescu,
> Cristian
> > Sent: Monday, April 9, 2018 4:59 PM
> > To: Stephen Hemminger <stephen@networkplumber.org>; Singh,
> Jasvinder
> > <jasvinder.singh@intel.com>; Richardson, Bruce
> <bruce.richardson@intel.com>
> > Cc: dev@dpdk.org
> > Subject: Re: [dpdk-dev] [PATCH] table: fix build error with gcc 8
> >
> >
> >
> > > -----Original Message-----
> > > From: Stephen Hemminger [mailto:stephen@networkplumber.org]
> > > Sent: Monday, April 9, 2018 4:10 PM
> > > To: Singh, Jasvinder <jasvinder.singh@intel.com>
> > > Cc: dev@dpdk.org; Dumitrescu, Cristian <cristian.dumitrescu@intel.com>
> > > Subject: Re: [dpdk-dev] [PATCH] table: fix build error with gcc 8
> > >
> > > On Mon,  9 Apr 2018 13:49:48 +0100
> > > Jasvinder Singh <jasvinder.singh@intel.com> wrote:
> > >
> > > > Fix build error with gcc 8.0 due to cast between function types.
> > > > Fixes: 5a80bf0ae613 ("table: add cuckoo hash")
> > > >
> > > > Signed-off-by: Jasvinder Singh <jasvinder.singh@intel.com>
> > > > ---
> > > >  lib/librte_table/rte_table_hash_cuckoo.c | 4 +++-
> > > >  1 file changed, 3 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/lib/librte_table/rte_table_hash_cuckoo.c
> > > b/lib/librte_table/rte_table_hash_cuckoo.c
> > > > index dcb4fe9..f7eae27 100644
> > > > --- a/lib/librte_table/rte_table_hash_cuckoo.c
> > > > +++ b/lib/librte_table/rte_table_hash_cuckoo.c
> > > > @@ -103,11 +103,13 @@ rte_table_hash_cuckoo_create(void
> *params,
> > > >  		return NULL;
> > > >  	}
> > > >
> > > > +	void *hash_func = p->f_hash;
> > > > +
> > > >  	/* Create cuckoo hash table */
> > > >  	struct rte_hash_parameters hash_cuckoo_params = {
> > > >  		.entries = p->n_keys,
> > > >  		.key_len = p->key_size,
> > > > -		.hash_func = (rte_hash_function)(p->f_hash),
> > > > +		.hash_func = (rte_hash_function) hash_func,
> > > >  		.hash_func_init_val = p->seed,
> > > >  		.socket_id = socket_id,
> > > >  		.name = p->name
> > >
> > > This is just tricking the compiler into not complaining.
> > > I would really rather see the two hash functions made the same.
> >
> > (Adding Bruce as well to consolidate all conversations in a single thread.)
> >
> > What we want to do here is be able to use the librte_hash under the same
> API
> > as the several hash table flavors implemented in librte_table.
> >
> > Both of these libraries allow configuring the hash function per each hash
> > table instance. Problem is: hash function in librte_hash has only 3
> parameters
> > (no key mask), while hash function in librte_table has 4 parameters
> (includes
> > key mask). The key mask helps a lot for practical protocol implementations
> by
> > avoiding key copy & pre-process on lookup.
> >
> > So then: how to plug in librte_hash under the same API as the suite of
> hash
> > tables in librte_table? We don't want to re-implement cuckoo hash from
> > librte_hash, we simply want to invoke it as a low-level primitive, similarly
> > to how the LPM and ACL tables are plugged into librte_table.
> >
> > Solution is: as an exception, pass a 3-parameter hash function to cuckoo
> hash
> > flavor under the librte_table. Maybe this should be documented better.
> This
> > currently triggers a build warning with gcc 8, which is easy to fix, hence
> > this trivial patch.
> >
> > Ideally, for every 3-parameter hash function, I would like to generate the
> > corresponding 4-parameter hash function on-the-fly, but unfortunately this
> is
> > not what C language can do.
> >
> > Of course, IMO the best solution is to add key mask support to librte_hash.
> 
> 
> Looking at the previous discussion I see the following as a possible solution;
> 
> Given the current code looks broken it should be fixed in this release.

The code is not broken. This is not a bug, it is a limitation for that particular table type. The only gap that I see is adding a Doxygen comment in the API header file.

User explicitly picks the hash table type it wants; when using this particular hash table type, that pointer needs to point to a 3-parameter function instead of 4. Given the limitation is clearly documented in Doxygen (current gap that we can quickly address), I don't see any problem.

If people think that this function conversion is not nice, it can be reworked in multiple ways at the expense of API (but not ABI) change:
1. Define the hash function field in the table parameter structure as opaque void * rather than 4-parameter version.
2. Create a separate parameter structure just for this hash table type.

> Given the actual code fix is an API / ABI break (depending on solution) it
> cannot be merged official in this release.
> We have a NEXT_ABI macro - it allows us to break API/ABI conditionally at
> compile time.

This is not new code introduced in this release cycle, this is just fixing the compiler warning, I fail to see how your ABI breakage mention is applicable.

Maybe we should talk more specifics over the code, where exactly in the code would you like to place your NEXT_ABI macro?

> 
> With the above 3 points, I think the best solution is to correctly fix the
> problem that GCC 8 is identifying, and putting that new API inside the NEXT_
> macros.
> 
> In this case, we can preserve backwards (buggy) behavior if required, and
> provide correct (but API/ABI breaking) code as well. This is a tough decision -
> particularly for distros - what do they package?
> 
> Given the current code, I don't see a better solution - but I hope I'm wrong :)