Re: [PATCH net-next] net: add in_softirq() debug checking in napi_consume_skb()

[Yunsheng Lin]

On 2020/11/19 0:26, Jakub Kicinski wrote:
> On Wed, 18 Nov 2020 16:57:57 +0100 Peter Zijlstra wrote:
>> On Wed, Nov 18, 2020 at 07:43:48AM -0800, Jakub Kicinski wrote:
>>
>>> TBH the last sentence I wrote isn't clear even to me at this point ;D
>>>
>>> Maybe using just the macros from preempt.h - like this?
>>>
>>> #define lockdep_assert_in_softirq()                                    \
>>> do {                                                                   \
>>>        WARN_ON_ONCE(__lockdep_enabled                  &&              \
>>>                     (!in_softirq() || in_irq() || in_nmi())	\
>>> } while (0)

One thing I am not so sure about is the different irq context indicator
in preempt.h and lockdep.h, for example lockdep_assert_in_irq() uses
this_cpu_read(hardirq_context) in lockdep.h, and in_irq() uses
current_thread_info()->preempt_count in preempt.h, if they are the same
thing?

>>>
>>> We know what we're doing so in_softirq() should be fine (famous last
>>> words).  
>>
>> So that's not actually using any lockdep state. But if that's what you
>> need, I don't have any real complaints.
> 
> Great, thanks! 
> 
> The motivation was to piggy back on lockdep rather than adding a
> one-off Kconfig knob for a check in the fast path in networking.
> .
>