Bug in non-power-of-2 rings?
Konstantin Ananyev
konstantin.ananyev at huawei.com
Mon Aug 21 11:29:19 CEST 2023
Hi everyone,
> On Sun, Aug 20, 2023 at 11:07:33AM +0200, Morten Brørup wrote:
> > Bruce, Honnappa, Konstantin,
> >
> > Back in 2017, Bruce added support for non-power-of-2 rings with this patch [1].
> >
> > [1]: https://git.dpdk.org/dpdk/commit/lib/librte_ring/rte_ring.h?id=b74461155543430f5253e96ad6d413ebcad36693
> >
> > I think that the calculation of "entries" in __rte_ring_move_cons_head() [2][3] is incorrect when the ring capacity is not power-of-2,
> because it is missing the capacity comparison you added to rte_ring_count() [4]. Please review if I'm mistaken.
> >
> > [2]: https://elixir.bootlin.com/dpdk/v23.07/source/lib/ring/rte_ring_c11_pvt.h#L159
> > [3]: https://elixir.bootlin.com/dpdk/v23.07/source/lib/ring/rte_ring_generic_pvt.h#L150
> > [4]: https://elixir.bootlin.com/dpdk/v23.07/source/lib/ring/rte_ring.h#L502
Just to confirm you suggest something like that:
- *entries = (r->prod.tail - *old_head);
+ count = (r->prod.tail - *old_head);
+ entries = (count > r->capacity) ? r->capacity : count;
right?
> >
> thanks for flagging this inconsistency, but I think we are ok.
>
> For consumer, I think this is correct, because we are only ever reducing
> the number of entries in the ring, and the calculation of the number of
> entries is made in the usual way using modulo arithmetic. We should never
> have more than capacity entries in the ring so the check in ring count I
> believe is superflous. [The exception would be if someone bypassed the
> inline functions and accessed the ring directly themselves - at which point
> "all bets are off", to use the English phrase]
>
> The producer code (__rte_ring_move_prod_head) does do a capacity check,
> which is where one is required to ensure we never exceed capacity.
I also can't come up with the case, when current code will cause an issue..
In properly operating ring, I think we should never have more then r->capacity
entries populated, so this extra check can be skipped.
Unless you do have some particular case in mind?
Konstantin
More information about the dev
mailing list