[PATCH 1/3] block: introduce blk_queue_nr_active()

Sagi Grimberg sagi at grimberg.me
Thu Sep 28 03:56:32 PDT 2023


> I think it is unfortunately necessary to compute the sum from all the hctxs
> because in the general case there could be threads on other CPUs submitting
> I/O through another hctx that might significantly affect the result,

This argument is true in general, other threads issuing I/O on different
contexts may affect each other as you never know how a controller
processes the submission queues.

Its a tradeoff between the level of inspection and a reasonable
overhead. I do not think iterating over all hctxs of all available paths
is reasonable. But lets wait for your testing.



More information about the Linux-nvme mailing list