On Mon, 2 Nov 2020 11:14:32 +0800 Yunsheng Lin wrote:
> On 2020/11/1 6:38, Jakub Kicinski wrote:
> > On Thu, 29 Oct 2020 19:34:48 +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 a in_softirq() debug checking in napi_consume_skb() to catch
> >> this kind of error.
> >>
> >> Suggested-by: Eric Dumazet <eduma...@google.com>
> >> Signed-off-by: Yunsheng Lin <linyunsh...@huawei.com>
> >
> >> diff --git a/net/core/skbuff.c b/net/core/skbuff.c
> >> index 1ba8f01..1834007 100644
> >> --- a/net/core/skbuff.c
> >> +++ b/net/core/skbuff.c
> >> @@ -897,6 +897,10 @@ void napi_consume_skb(struct sk_buff *skb, int budget)
> >> return;
> >> }
> >>
> >> + DEBUG_NET_WARN(!in_softirq(),
> >> + "%s is called with non-zero budget outside softirq
> >> context.\n",
> >> + __func__);
> >
> > Can't we use lockdep instead of defining our own knobs?
>
> From the first look, using the below seems better than defining our
> own knobs, because there is similar lockdep_assert_in_irq() checking
> already and lockdep_assert_in_*() is NULL-OP when CONFIG_PROVE_LOCKING
> is not defined.
>
> >
> > Like this maybe?
> >
> > diff --git a/include/linux/lockdep.h b/include/linux/lockdep.h
> > index f5594879175a..5253a167d00c 100644
> > --- a/include/linux/lockdep.h
> > +++ b/include/linux/lockdep.h
> > @@ -594,6 +594,14 @@ do {
> > \
> > this_cpu_read(hardirqs_enabled))); \
> > } while (0)
> >
> > +#define lockdep_assert_in_softirq() \
> > +do { \
> > + WARN_ON_ONCE(__lockdep_enabled && \
> > + (softirq_count() == 0 || \
> > + this_cpu_read(hardirq_context))); \
>
> Using in_softirq() seems more obvious then using softirq_count()?
> And there is below comment above avoiding the using of in_softirq(), maybe
> that is why you use softirq_count() directly here?
> "softirq_count() == 0" still mean we are not in the SoftIRQ context and
> BH is not disabled. right? Perhap lockdep_assert_in_softirq_or_bh_disabled()
> is more obvious?
Let's add Peter to the recipients to get his opinion.
We have a per-cpu resource which is also accessed from BH (see
_kfree_skb_defer()).
We're trying to come up with the correct lockdep incantation for it.
> /*
> * Are we doing bottom half or hardware interrupt processing?
> *
> * in_irq() - We're in (hard) IRQ context
> * in_softirq() - We have BH disabled, or are processing softirqs
> * in_interrupt() - We're in NMI,IRQ,SoftIRQ context or have BH disabled
> * in_serving_softirq() - We're in softirq context
> * in_nmi() - We're in NMI context
> * in_task() - We're in task context
> *
> * Note: due to the BH disabled confusion: in_softirq(),in_interrupt() really
> * should not be used in new code.
> */
>
>
> Also, is there any particular reason we do the
> "this_cpu_read(hardirq_context)"
> checking?
Accessing BH resources from a hard IRQ context would be a bug, we may
have interrupted a BH, so AFAIU softirq_count() != 0, but we can nest
calls to _kfree_skb_defer().
