On Tue 2019-03-12 16:15:55, John Ogness wrote: > On 2019-03-12, Petr Mladek <pmladek@xxxxxxxx> wrote: > > On Tue 2019-03-12 09:17:49, John Ogness wrote: > >> The current printk implementation is handling all console printing as > >> best effort. Trying hard enough to dramatically affect the system, but > >> not trying hard enough to guarantee success. > > > > I agree that direct output is more reliable. It might be very useful > > for debugging some types of problems. The question is if it is worth > > the cost (code complexity, serializing CPUs == slowing down the > > entire system). > > > > But it is is possible that a reasonable offloading (in the direction > > of last Sergey's approach) might be a better deal. > > > > > > I suggest the following way forward (separate patchsets): > > > > 1. Replace log buffer (least controversial thing) > > Yes. I will post a series that only implements the ringbuffer using your > simplified API. That will be enough to remove printk_safe and actually > does most of the work of updating devkmsg, kmsg_dump, and syslog. Great. I just wonder if it is going to be fully lockless or still using the prb_lock. I could understand the a fully lockless solution will be much more complicated. But I think that it would make a lot of things easier in the long run. Especially it might help to avoid the big-kernel-lock-like approach. > > 2. Reliable offload to kthread (would be useful anyway) > > Yes. I would like to implement per-console kthreads for this series. I > think the advantages are obvious. For PREEMPT_RT the offloading will > need to be always active. (PREEMPT_RT cannot call the console->write() > from atomic contexts.) But I think this would be acceptable at first. It > would certainly be better than what PREEMPT_RT is doing now. I would personally start with one kthread. I am afraid that the discussion about it will be complicated enough. We could always make it more complicated later. I understand the immediate offloading might be necessary for PREEMPT_RT. But a more conservative approach is needed for non-rt kernels. Well, if there won't be a big difference in the complexity between one and more threads then we could mix it. But I personally see this a two steps that are better be done separately. > > 3. Atomic consoles (a lot of tricky code, might not be > > worth the effort) > > I think this will be necessary. PREEMPT_RT cannot support reliable > emergency console messages without it. And for kernel developers this is > also very helpful. People like PeterZ are using their own patches > because the mainline kernel is not providing this functionality. > > The decision about _when_ to use it is still in the air. But I guess > we'll worry about that when we get that far. There's enough to do until > then. I am sure that there are situations where the direct output to atomic consoles would help with debugging. PeteZ and Steven are using their own patches for a reason. Let's see how the code is complicated and how many consoles might get supported a reasonable way. Anyway, it will be a long run. I am personally curious where this will end :-) Best Regards, Petr
