Re: [Cake] [PATCH net-next v9 1/7] sched: Add Common Applications Kept Enhanced (cake) qdisc

["Toke Høiland-Jørgensen" <toke@toke.dk>]

David Miller <davem@davemloft.net> writes:

> From: Toke Høiland-Jørgensen <toke@toke.dk>
> Date: Mon, 14 May 2018 11:08:28 +0200
>
>> David Miller <davem@davemloft.net> writes:
>> 
>>> From: Toke Høiland-Jørgensen <toke@toke.dk>
>>> Date: Tue, 08 May 2018 16:34:19 +0200
>>>
>>>> +struct cake_flow {
>>>> +	/* this stuff is all needed per-flow at dequeue time */
>>>> +	struct sk_buff	  *head;
>>>> +	struct sk_buff	  *tail;
>>>
>>> Please do not invent your own SKB list handling mechanism.
>> 
>> We didn't invent it, we inherited it from fq_codel. I was actually about
>> to fix that, but then I noticed it was still around in fq_codel, and so
>> let it be. I can certainly fix it anyway, but, erm, why is it acceptable
>> in fq_codel but not in cake? struct sk_buff_head is not that new, is it?
>
> I guess one argument has to do with the amount of memory consumed by this
> per-flow or per-queue information, right?  Because the skb queue head has
> a qlen and a spinlock regardless of whether they are used or not.
>
> Furthermore, if you use the __skb_insert() et al. helpers, even though it
> won't use the lock it will adjust the qlen counter.  And that's useless
> work since you have no use for the qlen value.

I think the useless work issue is larger than the memory usage. When
running this (or FQ-CoDel) on small memory-constrained routers, we've
mostly had issues with OOM because of the packet data, which dwarfs the
per-queue overhead.

> Taken together, it seems that what you and fq_codel are doing is not
> such a bad idea after all.  So please leave it alone.

OK. I'll just resend with prettier Christmas trees, then :)

> On-and-off again, I've looked into converting skbs to using list_head
> but it's a non-trivial set of work. All over the tree the different
> layers use the next/prev pointers in different ways. Some use it for a
> doubly linked list. Some use it for a singly linked list. Some encode
> state in the prev pointer. You name it, it's out there.
>
> I'll try to get back to that task because obviously it'll be useful to
> have code like cake and fq_codel use common helpers instead of custom
> stuff.

Yup, I agree. From a code readability point of view, I also prefer the
helpers.

-Toke