On Wed, Nov 14, 2018 at 10:15 AM jianchao.wang <jianchao.w.wang@xxxxxxxxxx> wrote: > > Hi Jens > > Thanks for your kindly response. > > On 11/13/18 9:44 PM, Jens Axboe wrote: > > On 11/13/18 2:56 AM, Jianchao Wang wrote: > >> When issue request directly and the task is migrated out of the > >> original cpu where it allocates request, hctx could be ran on > >> the cpu where it is not mapped. > >> To fix this, > >> - insert the request forcibly if BLK_MQ_F_BLOCKING is set. > >> - check whether the current is mapped to the hctx, if not, insert > >> forcibly. > >> - invoke __blk_mq_issue_directly under preemption disabled. > > > > I'm not too crazy about this one, adding a get/put_cpu() in the hot > > path, and a cpumask test. The fact is that most/no drivers care > > about strict placement. We always try to do so, if convenient, > > since it's faster, but this seems to be doing the opposite. > > > > I'd be more inclined to have a driver flag if it needs guaranteed > > placement, using one an ops BLK_MQ_F_STRICT_CPU flag or similar. > > > > What do you think? > > > > I'd inclined blk-mq should comply with a unified rule, no matter the > issuing directly path or inserting one. Then blk-mq would have a simpler > model. And also this guarantee could be a little good for drivers, > especially the case where cpu and hw queue mapping is 1:1. I guess it is quite hard to respect this rule 100%, such as in case of CPU hotplug. Thanks, Ming Lei
