On Wed, Mar 14, 2018 at 05:07:41AM -0500, Rohit Zambre wrote: > Hi, > > I am aware that I am misusing the semantics of MLX5_SINGLE_THREADED in > the example scenario below. But this question is for learning > purposes: > > I have a multi-threaded application that uses 4 QPs in in 1 CTX. Each > QP has its own PD, MR and CQ and a dedicated thread that drives it. > Thread_i that drives QP_i will also poll CQ_i. So no locks are needed > on the QP (and the CQ for that matter). No locks are needed on the > uuars mapped to the QPs either since, with 4 QPs, they are still low > latency uuars. If I set MLX5_SINGLE_THREADED=1 in this case to run > this app, would it render incorrect behavior in terms of ringing the > doorbell? > > My understanding of doorbell writes is not concrete yet and so, I'm > not sure if it would be correct behavior. For the rest of the locks, > since no threads are contending on a QP or a CQ, turning off the > locking mechanism with MLX5_SINGLE_THREADED=1 as a hack should be > okay. upstream rdma-core and related has a feature called resource domains that safely supports some of this use case. It doesn't eliminate the locks but it guarentees the UARs can be used lock free. Jason -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
