On Mon, May 23, 2016 at 06:31:46AM -0700, Eric Dumazet wrote: > On Mon, 2016-05-23 at 13:43 +0300, Michael S. Tsirkin wrote: > > This is in response to the proposal by Jason to make tun > > rx packet queue lockless using a circular buffer. > > My testing seems to show that at least for the common usecase > > in networking, which isn't lockless, circular buffer > > with indices does not perform that well, because > > each index access causes a cache line to bounce between > > CPUs, and index access causes stalls due to the dependency. > > > > By comparison, an array of pointers where NULL means invalid > > and !NULL means valid, can be updated without messing up barriers > > at all and does not have this issue. > > Note that both consumers and producers write in the array, so in light > load (like TCP_RR), there are 2 cache line used byt the producers, and 2 > cache line used for consumers, with potential bouncing.
The shared part is RO by producer and consumer both, so it's not bouncing - it can be shared in both caches. Clearly memory footprint for this data structure is bigger so it might cause more misses. > In the other hand, the traditional sk_buff_head has one cache line, > holding the spinlock and list head/tail. > > We might use the 'shared cache line' : > > + /* Shared consumer/producer data */ > + int size ____cacheline_aligned_in_smp; /* max entries in queue > */ > + struct sk_buff **queue; > > > To put here some fast path involving a single cache line access when > queue has 0 or 1 item. > I will try to experiment with it, but pls note that this cache line is RO by producer and consumer currently, if we make it writeable it will be bouncing. -- MST
