Messages in this thread | | | From | KOSAKI Motohiro <> | Subject | Re: [PATCH 4/4] ring-buffer: only allocate buffers for online cpus | Date | Fri, 13 Mar 2009 08:47:48 +0900 (JST) |
| |
> > I don't like this patch. > > your [1/4] and [2/4] already solve Pierre's problem. > > > > using online cpu (not possible cpu) increase performance overhead > > and messiness. > > but nobody get benefit ;) > > Well, the fact that you can have 15 buffers for non existent CPUs is a big > benefit. And the overhead was only on the read side, not the write, and > very limited in overhead for that matter. > > But, looking at this, I realized I can get rid of all the > "get_online_cpus". I originally had the CPU_DOWN_PREPARE remove the > buffer. But I found it highly annoying during tests, that I lose my data > when I brought down a CPU. Thus, I removed the code to free the buffer and > replaced it with the comment explaining this. > > The get_online_cpus is to prevent the race where we might remove a buffer. > But since we do not do that anymore, those get_online_cpus are pretty > useless. > > I'll update the code in a bit.
Great!
from mission critical area view, CPU down event is very important. I don't hope to lost trace data of downed cpu. and no get_online_cpus() don't cause annoy overhead.
I love to see your next patch :)
| |