On 6/6/19 4:58 PM, Stefano Brivio wrote:
>>> This also means that to avoid sending duplicates in the case where at
>>> least one rt6_fill_node() call goes through and one fails, we would
>>> need to track the last bucket and entry sent, or, alternatively, to
>>> make sure we can fit the whole node before dumping.
>> My another concern is the dump may never finish.
>
> That's not a guarantee in general, even without this, because in theory
> the skb passed might be small enough that we can't even fit a single
> node without exceptions.
That should be handled by skb->len = 0 and then returning the err back
to caller. See inet_dump_fib.
>
> We could add a guard on w->leaf not being the same before and after the
> walk in inet6_dump_fib() and, if it is, terminate the dump. I just
> wonder if we have to do this at all -- I can't find this being done
> anywhere else (at a quick look at least).
>
> By the way, we can also trigger a never-ending dump by touching the
> tree frequently enough during a dump: it would always start again from
> the root, see fib6_dump_table().
>
that should be a userspace problem on a sequence mismatch. In-kernel
notifiers do restart for offload drivers (seeregister_fib_notifier), but
as I recall the kernel should not restart a dump.
libnl reference is nl_cache_refill,
err = nl_cache_pickup(sk, cache);
if (err == -NLE_DUMP_INTR) {
NL_DBG(2, "Dump interrupted, restarting!\n");
goto restart;
} else if (err < 0)
break;
