Re: 回复: INFO: rcu detected stall in tc_modify_qdisc

[Vinicius Costa Gomes <vinicius.gomes@intel.com>]

Hi Eric,

Eric Dumazet <eric.dumazet@gmail.com> writes:

>> I admit that I am on the fence on that argument: do not let even root
>> crash the system (the point that my code is crashing the system gives
>> weight to this side) vs. root has great powers, they need to know what
>> they are doing.
>> 
>> The argument that I used to convince myself was: root can easily create
>> a bunch of processes and give them the highest priority and do
>> effectively the same thing as this issue, so I went with a the "they
>> need to know what they are doing side".
>> 
>> A bit more on the specifics here:
>> 
>>   - Using a small interval size, is only a limitation of the taprio
>>   software mode, when using hardware offloads (which I think most users
>>   do), any interval size (supported by the hardware) can be used;
>> 
>>   - Choosing a good lower limit for this seems kind of hard: something
>>   below 1us would never work well, I think, but things 1us < x < 100us
>>   will depend on the hardware/kernel config/system load, and this is the
>>   range includes "useful" values for many systems.
>> 
>> Perhaps a middle ground would be to impose a limit based on the link
>> speed, the interval can never be smaller than the time it takes to send
>> the minimum ethernet frame (for 1G links this would be ~480ns, should be
>> enough to catch most programming mistakes). I am going to add this and
>> see how it looks like.
>> 
>> Sorry for the brain dump :-)
>
>
> I do not know taprio details, but do you really need a periodic timer
> ?

As we can control the transmission time of packets, you are right, I
don't.

Just a bit more detail about the current implementation taprio,
basically it has a sequence of { Traffic Classes that are open; Interval
} that repeats cyclicly, it uses an hrtimer to advance the pointer for
the current element, so during dequeue I can check if a traffic class is
"open" or "closed".

But again, if I calculate the 'skb->tstamp' of each packet during
enqueue, I don't need the hrtimer. What we have in the txtime-assisted
mode is half way there.

I think this is what you had in mind.


Cheers,
-- 
Vinicius