From: Eric Dumazet <eric.duma...@gmail.com>
Date: Thu, 19 May 2016 05:35:20 -0700
> From: Eric Dumazet <eduma...@google.com>
>
> Large tc dumps (tc -s {qdisc|class} sh dev ethX) done by Google BwE host
> agent [1] are problematic at scale :
>
> For each qdisc/class found in the dump, we currently lock the root qdisc
> spinlock in order to get stats. Sampling stats every 5 seconds from
> thousands of HTB classes is a challenge when the root qdisc spinlock is
> under high pressure.
>
> These stats are using u64 or u32 fields, so reading integral values
> should not prevent writers from doing concurrent updates if the kernel
> arch is a 64bit one.
>
> Being able to atomically fetch all counters like packets and bytes sent
> at the expense of interfering in fast path (queue and dequeue packets)
> is simply not worth the pain, as the values are generally stale after 1
> usec.
>
> These lock acquisitions slow down the fast path by 10 to 20 %
>
> An audit of existing qdiscs showed that sch_fq_codel is the only qdisc
> that might need the qdisc lock in fq_codel_dump_stats() and
> fq_codel_dump_class_stats()
>
> gnet_dump_force_lock() call is added there and could be added to other
> qdisc stat handlers if needed.
>
> [1]
> http://static.googleusercontent.com/media/research.google.com/en//pubs/archive/43838.pdf
>
> Signed-off-by: Eric Dumazet <eduma...@google.com>
I guess the off-by-one situations are not a big enough deal to code new
locking or memory barrier for, so I'm fine with this.
Please resubmit when I open net-next back up, thanks.
