On 28.01.2016 16:17, Prarit Bhargava wrote: > On 01/28/2016 07:52 AM, Vasily Averin wrote: >> Dear Prarit, >> > > Hi Vasily, Thanks for your suggestions. > >> I have no objections about your patch, >> bit in fact I doubt we really need to convert each timestamp in kernel logs. >> How do you think is it probably better to convert only one timestamp per >> screen ? > > How do you measure a screen? :)
I mean 80x25, but you're right, it makes sense to do it configurable. >> I.e. convert it in each 25th string only? > > While your suggestion does work for a flood of messages this will miss > situations where an event occurred hours/minutes/seconds earlier leading to a > panic. > >> Or just do it once per N seconds? > > I've tried several other versions of the patch and other userspace options > (such > as "date +%N > /tmp/kmsg" every second, setting up a timer to dump the real > time, etc.). Assuming that the disks didn't die (which was also part of the > problem I have seen) the printk buffer is finite in size and it is easy to > fill > the buffer if you're not careful. > >> And do not replace original timestamp but add converted one? > > I'm not sure I see the benefit of having two timestamps but if someone really > wanted that I could add an additional patch to do it. I investigate periodically various "node hangs" issues, without saved vmcore, /var/log/messages, net or serial console. All what I have is screenshot on local console, in worst case 80x25. I'm ready to lose ~10 characters on each string for short timesetamp, and of course I would be happy to know real date of these messages. But for me it's enough to see full date once per screen, and have short relative timestamp on other strings. > At the end of the day I need to be able to determine in real time what > happened > on a system and to make that as easy as possible for a human to read. > > P. >
