On Thu, Jan 05, 2023 at 07:35:59AM -1000, Tejun Heo <tj@xxxxxxxxxx> wrote: > Hard limits tend to make this sort of problems a lot more pronounced because > the existing mechanisms tend to break down for the users which are severely > throttled down even while the device as a whole is fairly idle. cpu.max > often triggers severe priority inversions too, so it isn't too surprising > that people hit severe priority inversion issues w/ io.max. To be on the same page: 1) severe PI == priority inversion across cgroups (progated e.g. via global locks (as with cpu.max) or FS journal (as with io.max)), 2) ordinary PI == priority inversion contained within a single cgroup, i.e. no different from an under-provisioned system. The reported issue sounds like 2) but even with the separated queues 1) is still possible :-/ > Another problem with blk-throttle is that it doesn't prioritize shared IOs > identified by bio_issue_as_root_blkg() like iolatency and iocost do, so > there can be very severe priority inversions when e.g. journal commit gets > trapped in a low priority cgroup further exacerbating issues like this. Thanks for the broader view. So the separated queues are certainly an improvement but ultimately a mechanism based on bio_issue_as_root_blkg() predicate and deferred throttling would be better? Or is permanent limit enforcement more important? Thanks, Michal
Attachment:
signature.asc
Description: Digital signature
