On Wed, Sep 18, 2019 at 5:47 PM Bart Van Assche <bvanassche@xxxxxxx> wrote: > > On 9/18/19 12:14 AM, Danil Kipnis wrote: > > I'm not familiar with dm code, but don't they need to deal with the > > same situation: if I configure 100 logical volumes on top of a single > > NVME drive with X hardware queues, each queue_depth deep, then each dm > > block device would need to advertise X hardware queues in order to > > achieve highest performance in case only this one volume is accessed, > > while in fact those X physical queues have to be shared among all 100 > > logical volumes, if they are accessed in parallel? > > Combining multiple queues (a) into a single queue (b) that is smaller > than the combined source queues without sacrificing performance is > tricky. We already have one such implementation in the block layer core > and it took considerable time to get that implementation right. See e.g. > blk_mq_sched_mark_restart_hctx() and blk_mq_sched_restart(). We will need some time, to check if we can reuse those... > dm drivers are expected to return DM_MAPIO_REQUEUE or > DM_MAPIO_DELAY_REQUEUE if the queue (b) is full. It turned out to be > difficult to get this right in the dm-mpath driver and at the same time > to achieve good performance. We also first tried to just return error codes in case we can't process an incoming request, but this was causing huge performance degradation when number of devices mapped over the same session is growing. Since we introduced those per cpu per devices lists of stopped queues, we do scale very well. > > The ibnbd driver introduces a third implementation of code that combines > multiple (per-cpu) queues into one queue per CPU. It is considered > important in the Linux kernel to avoid code duplication. Hence my > question whether ibnbd can reuse the block layer infrastructure for > sharing tag sets. Yes, will have to reiterate on this.
