Re: [dpdk-dev] [PATCH v3] hash table: add an iterator over conflicting entries

[Michel Machado <michel@digirati.com.br>]

On 09/05/2018 04:27 PM, Wang, Yipeng1 wrote:
> Hmm I see, it falls back to my original thought to have malloc inside the init function..
> Thanks for the explanation. :)
> 
> So I guess with your implementation, in future if we change the internal state to be larger,
> the ABI will be broken.

    If that happens, yes, the ABI would need to change again. But this 
concern is overblown for two reasons. First, this event is unlikely to 
happen because struct rte_hash_iterator_state is already allocating 64 
bytes while struct rte_hash_iterator_istate and struct 
rte_hash_iterator_conflict_entries_istate consume 16 and 20 bytes, 
respectively. Thus, the complexity of the underlying hash algorithm 
would need to grow substantially to force the necessary state of these 
iterators to grow more than 4x and 3x, respectively. This is unlikely to 
happen, and, if it does, it would likely break the ABI somewhere else 
and have a high impact on applications anyway.

    Second, even if the unlikely event happens, all one would need to do 
is to increase the size of struct rte_hash_iterator_state, mark the new 
API as a new version, and applications would be ready for the new ABI 
just recompiling.

> BTW, this patch set also changes API so proper notice is needed.
> People more familiar with API/ABI change policies may be able to help here.

    We'd be happy to get feedback on this aspect.

> Just to confirm, is there anyway like I said for your application to have some long-live states
> and reuse them throughout the application so that you don’t have to have short-lived ones in stack?

    Two things would need to happen for this to be possible. The init 
functions would need to accept previously allocated iterator states, 
that is, the init function would act as a reset of the state when acting 
on a previous allocated state. And, applications would now need to carry 
these pre-allocated state to avoid a malloc. In order words, we'll 
increase the complexity of the API.

    To emphasize that the cost of a malloc is not negligible, 
rte_malloc() needs to get a spinlock (see heap_alloc_on_socket()), do 
its thing to allocate memory, and, if the first attempt fails, try to 
allocate the memory on other sockets (see end of malloc_heap_alloc()). 
For an iterator that goes through the whole hash table, this cost may be 
okay, but for an iterator that goes through a couple entries, this cost 
is a lot to add.

    This memory allocation concern is not new. Function 
rte_pktmbuf_read(), for example, let applications pass buffers, which 
are often allocated in the execution stack, to avoid the malloc cost.

[ ]'s
Michel Machado