Description of problem:

The broken out patch tarball //www.kernel.org/pub/linux/kernel/projects/rt/4.8/patch-4.8.15-rt10.patch.xz comes with two patches: skbufhead-raw-lock.patch and skbufhead-raw-lock-free-skbs.patch, dealing with a patch management mishap by the 4.8.15-rt10 maintainer, that we have in our rhel-7.4/rt-master branch as just one patch, the original from 2011, by tglx: 85b9e707d00ab ("skbufhead-raw-lock.patch").

The problem is that something happened, a merge probably, and a hunk, the one that coincidentally is handled (slightly differently) in skbufhead-raw-lock-free-skbs.patch, disappeared, i.e. the freeing of the skbs moved to the softnet_data->tofree_q skb_queue_head list.

So, to match 4.8.15-rt10 (and the original work by tglx), we need to apply skbufhead-raw-lock-free-skbs.patch.

In a diff:

diff -u rhel7/net/core/dev.c v4.8.15-rt10/net/core/dev.c

@@ -4971,13 +5214,21 @@
        struct softnet_data *sd = this_cpu_ptr(&softnet_data);
        unsigned long time_limit = jiffies + 2;
        int budget = netdev_budget;
+       struct sk_buff_head tofree_q;
+       struct sk_buff *skb;
        LIST_HEAD(list);
        LIST_HEAD(repoll);
 
+       __skb_queue_head_init(&tofree_q);
+
        local_irq_disable();
+       skb_queue_splice_init(&sd->tofree_queue, &tofree_q);
        list_splice_init(&sd->poll_list, &list);
        local_irq_enable();
 
+       while ((skb = __skb_dequeue(&tofree_q)))
+               kfree_skb(skb);
+
        for (;;) {
                struct napi_struct *n;
 
 -----

So the effect is that without this hunk skbs will get leaked when network devices are unregistered and packets remained in its queued to be processed.