> -----Original Message-----
> From: Stephen Hemminger [mailto:step...@networkplumber.org]
> Sent: Tuesday, January 15, 2019 6:34 PM
> To: Eads, Gage <gage.e...@intel.com>
> Cc: dev@dpdk.org; olivier.m...@6wind.com; arybche...@solarflare.com;
> Richardson, Bruce <bruce.richard...@intel.com>; Ananyev, Konstantin
> <konstantin.anan...@intel.com>
> Subject: Re: [PATCH] doc: announce ring ABI and API changes
>
> On Tue, 15 Jan 2019 17:59:34 -0600
> Gage Eads <gage.e...@intel.com> wrote:
>
> > In order to support the non-blocking ring[1], one ABI change and one
> > API change are required in librte_ring. This commit updates the
> > deprecation notice to pave the way for their inclusion in 19.05.
> >
> > [1] http://mails.dpdk.org/archives/dev/2019-January/123475.html
> >
> > Signed-off-by: Gage Eads <gage.e...@intel.com>
> > ---
> > doc/guides/rel_notes/deprecation.rst | 11 +++++++++++
> > 1 file changed, 11 insertions(+)
> >
> > diff --git a/doc/guides/rel_notes/deprecation.rst
> > b/doc/guides/rel_notes/deprecation.rst
> > index d4aea4b46..d74cff467 100644
> > --- a/doc/guides/rel_notes/deprecation.rst
> > +++ b/doc/guides/rel_notes/deprecation.rst
> > @@ -83,3 +83,14 @@ Deprecation Notices
> > - The size and layout of ``rte_cryptodev_qp_conf`` and syntax of
> > ``rte_cryptodev_queue_pair_setup`` will change to to allow to use
> > two different mempools for crypto and device private sessions.
> > +
> > +* ring: two changes are planned for rte_ring in v19.05:
> > +
> > + - The ring head and tail values are planned to be changed from
> > ``uint32_t``
> > + to ``size_t``. This reduces the likelihood of wrap-around to
> > effectively
> > + zero for 64-bit builds, which is important in avoiding the ABA problem
> > in
> > + the upcoming non-blocking ring implementation. (32-bit builds are
> > + unaffected by this change.)
> > + - rte_ring_get_memsize() will get a new ``flags`` parameter, so it can
> > + calculate the memory required for rings that require more than 8B per
> entry
> > + (such as the upcoming non-blocking ring).
>
>
> Would it be possible to support new and old ring types, either through naming
> tricks and/or new ring flag? Changing things like ring buffer and mbuf are
> basically a flag day for all users.
>
> I admit to having a personal interest in this since the API/ABI churn is this
> project
> causes vendors to stay on older code. And older code does not correctly
> support
> newer networks.
Fair enough -- I appreciate the additional context wrt avoiding churn.
This might be doable with the following change:
"
@@ -70,6 +70,15 @@ struct rte_ring_headtail {
uint32_t single; /**< True if single prod/cons */
};
+/* 64-bit version of rte_ring_headtail, for use by rings that need to avoid
+ * head/tail wrap-around.
+ */
+struct rte_ring_headtail_64 {
+ volatile uint64_t head; /**< Prod/consumer head. */
+ volatile uint64_t tail; /**< Prod/consumer tail. */
+ uint32_t single; /**< True if single prod/cons */
+};
+
/**
* An RTE ring structure.
*
@@ -97,11 +106,19 @@ struct rte_ring {
char pad0 __rte_cache_aligned; /**< empty cache line */
/** Ring producer status. */
- struct rte_ring_headtail prod __rte_cache_aligned;
+ RTE_STD_C11
+ union {
+ struct rte_ring_headtail prod __rte_cache_aligned;
+ struct rte_ring_headtail_64 prod_64 __rte_cache_aligned;
+ };
char pad1 __rte_cache_aligned; /**< empty cache line */
/** Ring consumer status. */
- struct rte_ring_headtail cons __rte_cache_aligned;
+ RTE_STD_C11
+ union {
+ struct rte_ring_headtail cons __rte_cache_aligned;
+ struct rte_ring_headtail_64 cons_64 __rte_cache_aligned;
+ };
char pad2 __rte_cache_aligned; /**< empty cache line */
};
"
The ABI compatibility hinges on the fact that today's prod and cons are both
padded out to a full cache line, and the 64-bit version fits within a single
cache line. (Confirmed with pahole.)
abi-compliance-checker reports two issues, but both appear to be false
positives:
1. "Field cons has been removed from this type"
2. "Field prod has been removed from this type"
I need to do more work to see whether/how the ring functions are affected by
such a change, but I first want to check if the community agrees with this
approach. Note that I don't see any way to avoid the API change to
rte_ring_get_memsize, but I doubt that would have near the impact of a ring
data structure change.
Thanks,
Gage