On Tue, Jun 04, 2019 at 02:17:28PM -0600, David Ahern wrote:

On 6/3/19 11:29 PM, Martin Lau wrote:

quoted

On Mon, Jun 03, 2019 at 07:36:06PM -0600, David Ahern wrote:

quoted

On 6/3/19 6:58 PM, Martin Lau wrote:

quoted

I have concern on calling ip6_create_rt_rcu() in general which seems
to trace back to this commit
dec9b0e295f6 ("net/ipv6: Add rt6_info create function for ip6_pol_route_lookup")

This rt is not tracked in pcpu_rt, rt6_uncached_list or exception bucket.
In particular, how to react to NETDEV_UNREGISTER/DOWN like
the rt6_uncached_list_flush_dev() does and calls dev_put()?

The existing callers seem to do dst_release() immediately without
caching it, but still concerning.

those are the callers that don't care about the dst_entry, but are
forced to deal with it. Removing the tie between fib lookups an
dst_entry is again the right solution.

Great to know that there will be a solution.  It would be great
if there is patch (or repo) to show how that may look like on
those rt6_lookup() callers.

Not 'will be', 'there is' a solution now. Someone just needs to do the
conversions and devise the tests for the impacted users.

I don't think everyone will convert to the new nexthop solution
immediately.

How about ensuring the existing usage stays solid first?

quoted

Before that,
although it seems fine now (__ip6_route_redirect() is
harder to confirm since rt is passed around but it
seems to be ok),
instead of risking for "unregister_netdevice: waiting for eth0 to become free"
in case some future patch is caching this rt,
why pcpu_rt cannot be used in all occasions? and also
avoid re-creating the same rt.