On Thu 09-02-17 14:21:52, peter enderborg wrote: > Fundamental changes: > 1 Does NOT take any RCU lock in shrinker functions. > 2 It returns same result for scan and counts, so we dont need to do > shinker will know when it is pointless to call scan. > 3 It does not lock any other process than the one that is > going to be killed. > > Background. > The low memory killer scans for process that can be killed to free > memory. This can be cpu consuming when there is a high demand for > memory. This can be seen by analysing the kswapd0 task work. > The stats function added in earler patch adds a counter for waste work. > > How it works. > This patch create a structure within the lowmemory killer that caches > the user spaces processes that it might kill. It is done with a > sorted rbtree so we can very easy find the candidate to be killed, > and knows its properies as memory usage and sorted by oom_score_adj > to look up the task with highest oom_score_adj. To be able to achive > this it uses oom_score_notify events. > > This patch also as a other effect, we are now free to do other > lowmemorykiller configurations. Without the patch there is a need > for a tradeoff between freed memory and task and rcu locks. This > is no longer a concern for tuning lmk. This patch is not intended > to do any calculation changes other than we do use the cache for > calculate the count values and that makes kswapd0 to shrink other > areas. I have to admit I really do not understand big part of the above paragraph as well as how this all is supposed to work. A quick glance over the implementation. __lmk_task_insert seems to be only called from the oom_score notifier context. If nobody updates the value then no task will get into the tree. Or am I missing something really obvious here? Moreover oom scores tend to be mostly same for tasks. That means that your sorted tree will become sorted by pids in most cases. I do not see any sorting based on the rss nor any updates that would reflect updates of rss. How can this possibly work? -- Michal Hocko SUSE Labs _______________________________________________ devel mailing list devel@xxxxxxxxxxxxxxxxxxxxxx http://driverdev.linuxdriverproject.org/mailman/listinfo/driverdev-devel
