On Tue, Sep 22, 2026 at 06:34:01AM -0700, Stephen Hemminger wrote:
> On Tue, 22 Sep 2026 12:36:14 +0100
> Bruce Richardson <[email protected]> wrote:
>
> > The devarg "internal" was used to pass a pointer to the creation
> > function to reuse set parameters. However, this meant it could also be
> > used as a devarg on the commandline leading to incorrect behaviour.
> > Redesign the passing of the args parameters to use a per-thread global
> > instead. The wrapper function sets the global to a valid value before
> > calling the main creation function, skipping passing a devarg, before
> > resetting the pointer back to NULL when the function returns.
> >
> > Bugzilla ID: 1687
> > Fixes: 96cb19521147 ("net/ring: use EAL APIs in PMD specific API")
> > Cc: [email protected]
> >
> > Signed-off-by: Bruce Richardson <[email protected]>
> > ---
>
> Minor nits from Opus review:
>
> Review: [PATCH] net/ring: remove internal devarg
> Patchwork: 169914
>
> Applies cleanly to main, builds with -Dwerror=true. Fixes tag hash
> 96cb19521147 verified. Removing a user-reachable devarg that is
> dereferenced as a raw pointer is the right fix.
>
> Warning
>
> The per-lcore pointer is still set while user callbacks run.
> do_eth_dev_ring_create() calls rte_eth_dev_probing_finish(), which
> fires RTE_ETH_EVENT_NEW callbacks synchronously on the same thread.
> If a callback calls rte_vdev_init("net_ring_foo", "nodeaction=...")
> the nested probe sees eth_ring_internal_args != NULL, ignores the
> user's devargs, and attaches net_ring_foo to the outer caller's
> rings.
>
That's an unrealistic scenario, is it not? Having an ethdev creation
callback which creates another ethdev just leads to an infinite chain of
callbacks.
> Clear the pointer in probe right after reading it:
>
> internal_args = RTE_PER_LCORE(eth_ring_internal_args);
> RTE_PER_LCORE(eth_ring_internal_args) = NULL;
> if (internal_args != NULL) {
>
> Keep the reset after rte_vdev_init() in rte_eth_from_rings() to
> cover the paths where probe is never reached (-EEXIST, no driver
> match). This also makes the comment "consumed and cleared by probe"
> true; as written, probe does not clear it.
>
> Alternatively add the vdev name to struct ring_internal_args and
> require a match in probe. Either works; clearing is simpler.
>
> Info
>
> No test coverage. app/test/test_pmd_ring.c could add a case that
> "net_ring_x,internal=0x1" no longer dereferences the value (it now
> falls into the "Ignoring unsupported parameters" path).
>
Since there is no "internal" parameter any more, I don't see this as being
necessary.
> Double space before "*/" in the comment above
> RTE_DEFINE_PER_LCORE.
Something that can be fixed on apply, I think, if it's a concern.
/Bruce
