[dpdk-dev] [PATCH RFCv3 00/19] ring cleanup and generalization

[Bruce Richardson]

On Tue, Feb 14, 2017 at 09:32:20AM +0100, Olivier Matz wrote:
> Hi Bruce,
> 
> On Tue,  7 Feb 2017 14:12:38 +0000, Bruce Richardson
> <bruce.richardson at intel.com> wrote:
> > This patchset make a set of, sometimes non-backward compatible,
> > cleanup changes to the rte_ring code in order to improve it. The
> > resulting code is shorter*, since the existing functions are
> > restructured to reduce code duplication, as well as being more
> > consistent in behaviour. The specific changes made are explained in
> > each patch which makes that change.
> > 
> > Key incompatibilities:
> > * The biggest, and probably most controversial change is that to the
> >   enqueue and dequeue APIs. The enqueue/deq burst and bulk functions
> > have their function prototypes changed so that they all return an
> > additional parameter, indicating the size of next call which is
> > guaranteed to succeed. In case on enq, this is the number of
> > available slots on the ring, and in case of deq, it is the number of
> > objects which can be pulled. As well as this, the return value from
> > the bulk functions have been changed to make them compatible with the
> > burst functions. In all cases, the functions to enq/deq a set of objs
> > now return the number of objects processed, 0 or N, in the case of
> > bulk functions, 0, N or any value in between in the case of the burst
> > ones. [Due to the extra parameter, the compiler will flag all
> > instances of the function to allow the user to also change the return
> > value logic at the same time]
> > * The parameters to the single object enq/deq functions have not been 
> >   changed. Because of that, the return value is also unmodified - as
> > the compiler cannot automatically flag this to the user.
> > 
> > Potential further cleanups:
> > * To a certain extent the rte_ring structure has gone from being a
> > whole ring structure, including a "ring" element itself, to just
> > being a header which can be reused, along with the head/tail update
> > functions to create new rings. For now, the enqueue code works by
> > assuming that the ring data goes immediately after the header, but
> > that can be changed to allow specialised ring implementations to put
> > additional metadata of their own after the ring header. I didn't see
> > this as being needed right now, but it may be worth considering for a
> > V1 patchset.
> > * There are 9 enqueue functions and 9 dequeue functions in
> > rte_ring.h. I suspect not all of those are used, so personally I
> > would consider dropping the functions to enqueue/dequeue a single
> > value using single or multi semantics, i.e. drop 
> >     rte_ring_sp_enqueue
> >     rte_ring_mp_enqueue
> >     rte_ring_sc_dequeue
> >     rte_ring_mc_dequeue
> >   That would still leave a single enqueue and dequeue function for
> > working with a single object at a time.
> > * It should be possible to merge the head update code for enqueue and
> >   dequeue into a single function. The key difference between the two
> > is the calculation of how far the index can be moved. I felt that the
> >   functions for moving the head index are sufficiently complicated
> > with many parameters to them already, that trying to merge in more
> > code would impede readability. However, if so desired this change can
> > be made at a later stage without affecting ABI or API.
> > 
> > PERFORMANCE:
> > I've run performance autotests on a couple of (Intel) platforms.
> > Looking particularly at the core-2-core results, which I expect are
> > the main ones of interest, the performance after this patchset is a
> > few cycles per packet faster in my testing. I'm hoping it should be
> > at least neutral perf-wise.
> > 
> > REQUEST FOR FEEDBACK:
> > * Are all of these changes worth making?
> 
> I've quickly browsed all the patches. I think yes, we should do it: it
> brings a good cleanup, removing features we don't need, restructuring
> the code, and also adding the feature you need :)
> 
> 
> > * Should they be made in existing ring code, or do we look to provide
> > a new fifo library to completely replace the ring one?
> 
> I think it's ok to have it in the existing code. Breaking the ABI
> is never suitable, but I think having 2 libs would be even more
> confusing.
> 
> 
> > * How does the implementation of new ring types using this code
> > compare vs that of the previous RFCs?
> 
> I prefer this version, especially compared to the first RFC.
> 
> 
> Thanks for this big rework. I'll dive into the patches a do a more
> exhaustive review soon.
> 
Great, thanks. I'm aware of a few things that already need to be cleaned
up for V1 e.g. comments are not always correctly updated on functions.

/Bruce