Hello, On Tue, Nov 14, 2023 at 10:59:47AM -0800, Junxiao Bi wrote: > kernfs notify is used in write path of md (md_write_start) to wake up > userspace daemon, like "mdmon" for updating md superblock of imsm raid, > md write will wait for that update done before issuing the write, if this How is forward progress guarnateed for that userspace daemon? This sounds like a really fragile setup. > write is used for memory reclaim, the system may hung due to kernel notify > can't be executed, that's because kernel notify is executed by "system_wq" > which doesn't have a rescuer thread and kworker thread may not be created > due to memory pressure, then userspace daemon can't be woke up and md write > will hung. > > According Tejun, this can't be fixed by add RECLAIM to "system_wq" because > that workqueue is shared and someone else might occupy that rescuer thread, > to fix this from md side, have to replace kernfs notify with other way to > communite with userspace daemon, that will break userspace interface, > so use a separated workqueue for kernefs notify to allow it be used in > memory reclaim context. I'm not necessarily against the change but please go into a bit more details on how and why it's structured this way and add a comment explaining explaining who's depending on kernfs notify for reclaim forward progress. Thanks. -- tejun
