[lttng-dev] [rp] [URCU PATCH 0/3] wait-free concurrent queues (wfcqueue)

Lai Jiangshan laijs at cn.fujitsu.com
Sun Oct 7 23:33:47 EDT 2012

On 10/04/2012 05:04 AM, Mathieu Desnoyers wrote:
> * Paul E. McKenney (paulmck at linux.vnet.ibm.com) wrote:
>> On Tue, Oct 02, 2012 at 10:13:07AM -0400, Mathieu Desnoyers wrote:
>>> Implement wait-free concurrent queues, with a new API different from
>>> wfqueue.h, which is already provided by Userspace RCU. The advantage of
>>> splitting the head and tail objects of the queue into different
>>> arguments is to allow these to sit on different cache-lines, thus
>>> eliminating false-sharing, leading to a 2.3x speed increase.
>>> This API also introduces a "splice" operation, which moves all nodes
>>> from one queue into another, and postpones the synchronization to either
>>> dequeue or iteration on the list. The splice operation does not need to
>>> touch every single node of the queue it moves them from. Moreover, the
>>> splice operation only needs to ensure mutual exclusion with other
>>> dequeuers, iterations, and splice operations from the list it splices
>>> from, but acts as a simple enqueuer on the list it splices into (no
>>> mutual exclusion needed for that list).
>>> Feedback is welcome,
>> These look sane to me, though I must confess that the tail pointer
>> referencing the node rather than the node's next pointer did throw
>> me for a bit.  ;-)
> Yes, this was originally introduced with Lai's original patch to
> wfqueue, which I think is a nice simplification: it's pretty much the
> same thing to use the last node address as tail rather than the address
> of its first member (its next pointer address (_not_ value)). It ends up
> being the same address in this case, but more interestingly, we don't
> have to use a struct cds_wfcq_node ** type: a simple struct
> cds_wfcq_node *  suffice.
> Thanks Paul, I will therefore merge these 3 patches with your Acked-by.
> Lai, you are welcome to provide improvements to this code against the
> master branch. I will gladly consider any change you propose.

I did not remember that there is any improvement idea not included.
The patchset is OK for me.

I think you can reimplement wfqueue via wfcqueue without cacheline opt.


More information about the lttng-dev mailing list