On Mon, 23 Nov 2020 15:27:25 +0100 Peter Zijlstra wrote:
> On Sat, Nov 21, 2020 at 11:06:15AM +0800, Yunsheng Lin wrote:
> > The current semantic for napi_consume_skb() is that caller need
> > to provide non-zero budget when calling from NAPI context, and
> > breaking this semantic will cause hard to debug problem, because
> > _kfree_skb_defer() need to run in atomic context in order to push
> > the skb to the particular cpu' napi_alloc_cache atomically.
> >
> > So add the lockdep_assert_in_softirq() to assert when the running
> > context is not in_softirq, in_softirq means softirq is serving or
> > BH is disabled. Because the softirq context can be interrupted by
> > hard IRQ or NMI context, so lockdep_assert_in_softirq() need to
> > assert about hard IRQ or NMI context too.
> Due to in_softirq() having a deprication notice (due to it being
> awefully ambiguous), could we have a nice big comment here that explains
> in detail understandable to !network people (me) why this is actually
> correct?
>
> I'm not opposed to the thing, if that his what you need, it's fine, but
> please put on a comment that explains that in_softirq() is ambiguous and
> when you really do need it anyway.
One liner would be:
* Acceptable for protecting per-CPU resources accessed from BH
We can add:
* Much like in_softirq() - semantics are ambiguous, use carefully. *
IIUC we basically want to protect the nc array and counter here:
static inline void _kfree_skb_defer(struct sk_buff *skb)
{
struct napi_alloc_cache *nc = this_cpu_ptr(&napi_alloc_cache);
/* drop skb->head and call any destructors for packet */
skb_release_all(skb);
/* record skb to CPU local list */
nc->skb_cache[nc->skb_count++] = skb;
#ifdef CONFIG_SLUB
/* SLUB writes into objects when freeing */
prefetchw(skb);
#endif
/* flush skb_cache if it is filled */
if (unlikely(nc->skb_count == NAPI_SKB_CACHE_SIZE)) {
kmem_cache_free_bulk(skbuff_head_cache, NAPI_SKB_CACHE_SIZE,
nc->skb_cache);
nc->skb_count = 0;
}
}
> > +#define lockdep_assert_in_softirq()
> > \
> > +do {
> > \
> > + WARN_ON_ONCE(__lockdep_enabled && \
> > + (!in_softirq() || in_irq() || in_nmi())); \
> > +} while (0)
