On Thu, Feb 10, 2022 at 10:34 PM Nick Neumann <nick@xxxxxxxxxxxxxxxx> wrote: > > I'm trying to fix a couple bugs I've talked about in a previous post > related to offloading, and in diving into workload.c I've hit some > things that have me confused. If someone understands the code well and > can give some pointers I'd be very grateful. > > First, is the reading of sw->flags in __get_submit_worker > intentionally not protected by the sw->lock? I'm guessing so, with the > intention that the worst that happens is a submit worker is seen as > busy when it is actually idle. > > But this approach confuses me for a couple reasons. First, the > td->io_u_freelist seems to have the same sort of thread contention > issues when doing offloading. The main td thread is removing things > from the freelist while any worker thread can insert into the > freelist, so they all coordinate with the td->io_u_lock. Why not just > use the same approach for maintaining the list of free submit workers? > > My second bit of confusion is that the number of io_u's allocated is > exactly iodepth. That's also the number of submit workers allocated by > the workqueue when offloading. Unless I'm mistaken (that's totally > possible), a free io_u will be obtained by get_io_u, will be handed > off to a submit worker, and when done, will become free again and the > submit worker will become idle. So aren't there enough submit workers > that there should always be one idle for any io_u that needs one? > > Thanks in advance for your help. > - Nick cc'ing Jens in case this is something he might have some thoughts on
