Re: [dpdk-dev] [PATCH v2] eal: add function to return number of detected sockets

[Thomas Monjalon <thomas@monjalon.net>]

16/01/2018 16:05, Burakov, Anatoly:
> On 16-Jan-18 12:20 PM, Thomas Monjalon wrote:
> > 16/01/2018 12:56, Burakov, Anatoly:
> >> On 12-Jan-18 11:50 AM, Thomas Monjalon wrote:
> >>> 12/01/2018 12:44, Burakov, Anatoly:
> >>>> On 11-Jan-18 10:20 PM, Thomas Monjalon wrote:
> >>>>> 22/12/2017 13:41, Anatoly Burakov:
> >>>>>> During lcore scan, find maximum socket ID and store it.
> >>>>>>
> >>>>>> Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
> >>>>>> ---
> >>>>>> --- a/lib/librte_eal/common/include/rte_eal.h
> >>>>>> +++ b/lib/librte_eal/common/include/rte_eal.h
> >>>>>> @@ -83,6 +83,7 @@ enum rte_proc_type_t {
> >>>>>>     struct rte_config {
> >>>>>>     	uint32_t master_lcore;       /**< Id of the master lcore */
> >>>>>>     	uint32_t lcore_count;        /**< Number of available logical cores. */
> >>>>>> +	uint32_t numa_node_count;    /**< Number of detected NUMA nodes. */
> >>>>>>     	uint32_t service_lcore_count;/**< Number of available service cores. */
> >>>>>>     	enum rte_lcore_role_t lcore_role[RTE_MAX_LCORE]; /**< State of cores. */
> >>>>>
> >>>>> isn't it breaking the ABI?
> >>>>>
> >>>>>
> >>>>
> >>>> Yep, you're right, forgot to add that. I didn't expect this to get
> >>>> merged in 18.02 anyway, so v2 will follow.
> >>>
> >>> Please write 18.05 in the subject to show your expectation.
> >>> Thanks
> >>>
> >>
> >> Does it have to be an ABI change though? We can put numa_node_count
> >> after pointer to mem_config, in which case it won't be an ABI break.
> >> Would that be better?
> > 
> > Changing the size of a struct which is allocated by the app,
> > is an ABI break.
> > Is your solution changing the size?
> > 
> 
> It's not really allocated as such. rte_config is a global static 
> variable, and we only ever get pointers to it from the user code. If we 
> add the new value at the end, all of the old data layout would be intact 
> and work as before, so nothing would change as far as old code is concerned.
> 
> However, if that's still considered an ABI break, then OK, break it is.

Maybe that assuming it is never allocated (not copied for instance)
we could consider it is not an ABI break.

> Some background for why this is needed - for the memory hotplug, we need 
> to know how many sockets we can allocate memory at, to distinguish 
> between socket that doesn't exist, and socket that exists but has no 
> memory allocated on it. I'm OK with trying other approaches (such as 
> storing numa nodes in a static variable somewhere) if breaking ABI for 
> this is too much to ask for such a minute change.

Why is it important for 18.02?
Memory hotplug will be integrated only in 18.05.
I think it is better to just wait (and announce the deprecation).