This shows you the differences between two versions of the page.
Both sides previous revision Previous revision | |||
realtime:rtl:current_topics [2017/02/08 11:14] anna-maria Add a backlink to RTL project main page |
realtime:rtl:current_topics [2017/10/05 20:20] (current) tglx Update to current state of affairs |
||
---|---|---|---|
Line 3: | Line 3: | ||
The real time developer team is currently working on the following topics that change current kernel functionality: | The real time developer team is currently working on the following topics that change current kernel functionality: | ||
- | * CPU hotplug overhaul: Change of the CPU hotplug notifier infrastructure into a state machine (for detailed description see [[https://lkml.org/lkml/2016/2/26/806|LKML: cpu/hotplug: Core infrastructure for cpu hotplug rework]]). Afterwards all CPU hotplug notifier users have to be converted to the CPU hotplug state machine. | + | * CPU hotplug overhaul: The rework is completed and for the first time since RT started CPU hotplug locking survives stress testing. But we are still addressing the fallout of the hotplug locking changes in mainline, which is fixing lock ordering bugs unearthed by the newly added lockdep coverage of the hotplug lock and the recent addition of completion dependencies. |
- | * Timer infrastructure: Timer wheel rebuild is required, because in the current PREEMPT_RT patch (v4.6.2-rt5) and the PREEMPT_RT patches before NOHZ can not be used. The basis for the new timer wheel was posted to LKML (see [[https://lkml.org/lkml/2016/6/13/211|LKML: timer: Refactor the timer wheel]]) is now available in 4.8 - NOHZ can now work again with RT. | + | |
+ | * Timer infrastructure: The timer rework part one is done and NOHZ works in Linux-4.8-rt and later. There is still work in progress to address the scalability issues in the timer enqueue path which tries to predict the best CPU placement for a newly added timer. Initial patch series posted, but it has testing failures and review comments need to be addressed. | ||
+ | |||
+ | * hrtimer rework: Replace the problematic conversion of hrtimers to expire in softirq context. The current implementation in RT is a left over from the early hrtimer implementation and needs to be cleaned up. Patches have been posted and reviewed. The integration into RT has shown some bugs, which are currently analyzed. | ||
+ | |||
+ | * dcache locking: The dcache code contains trylock loops which can lead to live locks on RT. There is a ugly and performance wise horrible workaround in RT which needs to be replaced. So far all attempts to solve this cleanly have failed. More hours on the drawing board are required. | ||
+ | |||
+ | * migrate disable: Still under discussion with the upstream scheduler maintainers | ||
+ | |||
+ | * interaction with the memory management code: Patches are in preparation to be posted soon | ||
In addition to the actual development work, the following topics have to be considered: | In addition to the actual development work, the following topics have to be considered: |