----- Original Message ----- > > > On 08/19/2015 11:38 PM, Dave Anderson wrote: > > > > > > ----- Original Message ----- > >> > >> Hi, > >> > >> On 08/19/2015 12:02 PM, Baoquan He wrote: > >>> On 08/19/15 at 09:21am, Nikolay Borisov wrote: > >>>> Hello, > >>>> > >>>> I've recently noticed that when creating crashdumps the dmesg is not > >>>> being saved. The error reported is this: "Missing the struct log size > >>>> export". I've tested with both kernel 4.1.1 and 3.12.28. My version of > >>>> kexec tools is : kexec-tools 2.0.9 released 04 June 2015. The kernel has > >>>> been compiled with debugging info so the respective symbol should be > >>>> exported but apparently it is not. Any ideas how to debug that? > >>> > >>> That would be more helpful if you can attach the 1st kernel and kdump > >>> kernel console log. > >> > >> I just tested with 4.1.6 with the same result. I'm sending you a link to > >> the bzImage in question > >> http://georgi.unixsol.org/cruft/tmp/bzImage-4.1.6-clouder1 ( I assume > >> you meant the bzImage, if you need I can provide the vmlinux as well). > >> Unfortunately, I couldn't figure how to obtain the log of kdump in > >> textual format (and didn't want to send pictures). Here are the 3 > >> relevant lines (from memory): > >> > >> Saving vmcore-dmesg.txt > >> Missing the struct log size export > >> Error saving vmcore-dmesg.txt. > >> > >> Yet, when I load the crashdump inside the 'crash' utility and invoke the > >> 'log' command the dmesg log is there. Whereas the vmcore-dmesg-incomplete.txt is empty. > > > > FYI -- the crash utility doesn't use the vmcoreinfo data for its "log" command, > > but you can dump the vmcore's header contents, including the vmcoreinfo strings, > > with the "help -n" command. For example: > > Thanks for the tip, using that I can confirm that the actual symbols are > in the resulting vmcoreinfo: > > SIZE(printk_log)=16 > OFFSET(printk_log.ts_nsec)=0 > OFFSET(printk_log.len)=8 > OFFSET(printk_log.text_len)=10 > OFFSET(printk_log.dict_len)=12 > > So the required symbol is there (printk_log) but for some reason kexec > cannot read it... I tried creating an uncompressed crash (removing the > -c option from the makedumpfile command to no avail).. The vmcoreinfo data strings were initially located in an ELF note in /proc/vmcore. When makedumpfile -c was run on /proc/vmcore, it copied those ELF notes into the compressed kdump header, and you have dumped them above. So it seems to be an issue with vmcore-dmesg. If you change the core_collector variable to "cp" or "scp", it will copy /proc/vmcore unmodified to the target location. Then you can run vmcore-dmesg on that file to debug it. Dave
