Ceph » RADOS

Ever since 755f3e03b5bf547cfd940b52a53833f66285062b, if bdev_enable_discard is enabled for a class of OSDs in config and bdev_async_discard=true/bdev_async_discard_threads>0, discards will be enabled for an OSD but synchronously, leading to unexpected large latency spikes in configurations that require async discards. This is because the starting of the discard thread(s) is evaluated before class masks can apply, and KernelDevice's handle_conf_change() doesn't watch for bdev_enable_discard changes.

I have a patch for this that applies to main, but reef/squid don't have bdev_async_discard_threads backported and thus will need an altered patch (or maybe they should get bdev_async_discard_threads for consistency). While fixing this issue, I also found and fix a number of issues with the stopping path for bdev_async_discard_threads.