> > We would have to consider (instead of jiffies) the time the process was > either running, or waiting on something that's related to memory > allocation/reclaim (page lock etc.). I.e. deduct the time the process > was runable but there was no available cpu. I expect however that such > level of detail wouldn't be feasible here, though? >
Johannes' memdelay work (once merged) might be useful here. I think memdalay can differentiate between an allocating process getting delayed due to preemption or due to unsuccessful reclaim/compaction. If the delay is due to unsuccessful reclaim/compaction then we should warn here.