On Jul 2, 2018, at 6:14 PM, Toke Høiland-Jørgensen <
toke@toke.dk> wrote:
Aha! I think I figured out what is going on:
The gen_stats facility will add an nlattr header at the beginning of the
qdisc stats, which is the toplevel TLV that contains all stats (and that
we put our stats inside). It stores a reference to this header, and when
all the per-qdisc callbacks have finished adding their stats, it goes
back and fixes up the length of the containing header.
The problem is that on architectures that need padding, the padding TLV
is added *first*, which means that the nlattr pointer that is stored
before the callbacks are performed points to the padding TLV and not the
stats TLV. And so, when the header is fixed up, the result (from the
parser's perspective) is just a very big padding TLV.
The options TLV is before the stats TLV, so the bug only occurs if the
options happen to have a length that means the stats will need padding.
Which is why messing with the number of options "fixes" the bug.
Could you try applying the patch below (to the kernel) and see if that
resolves the issue, please?
Awesome Toke! It looks like from Kevin’s email that it works for him, but it didn’t work for me the first time around. This may have to do with how I added the patch as I’m still not that familiar with OpenWRT’s build system (first kernel patch I tried). I wasn’t sure if it should go into generic or platform, for one, so I tried generic…is that right?
sysadmin@rey:~/src/openwrt/build_dir/target-mips_24kc_musl/linux-ar71xx_generic/linux-4.9.109/patches/generic$ cat 010-gen_stats_fix.patch
--- a/net/core/gen_stats.c
+++ b/net/core/gen_stats.c
@@ -77,8 +77,12 @@ gnet_stats_start_copy_compat(struct sk_b
d->lock = lock;
spin_lock_bh(lock);
}
- if (d->tail)
- return gnet_stats_copy(d, type, NULL, 0, padattr);
+ if (d->tail) {
+ int ret = gnet_stats_copy(d, type, NULL, 0, padattr);
+ if (!ret)
+ d->tail = ((struct nlattr *)skb_tail_pointer(skb)) - 1;
+ return ret;
+ }
return 0;
}