Description of problem:
Unable to start kdump service - unable to load the kdump image.
Version-Release number of selected component (if applicable):
Steps to Reproduce:
1. yum install kernel-2.6.17-1.2647.fc6 kernel-kdump-2.6.17-1.2647.fc6
2. Append " crashkernel=64M@16M" to "/etc/grub.conf".
3. service kdump start
Base address: 1267880 is not page aligned
Loaded kdump image.
Created attachment 136480 [details]
what arch are you recreating this on? This sounds like it may be a ppc64 64k
Just "standard" i686 RawHide (updated by yum from FC5), running on SMP Athlon.
I need to get hold of a machine that I can put FC6 on. In the interim, can you
please provie the /var/log/messages file from your system, showing the boot
sequence immediately prior to the indicated error? Thanks!
Created attachment 136847 [details]
grep kernel: /var/log/messages
Here you are.
please don't grep for kernel messages. I was looking for the whole thing.
What are you looking for, please? You do want messages from sendmail(8),
named(8) and ntpd(8)? I really do not know what else to supply.
I am aware of the technical details of kdump, I wrote the most parts of its
competitor mkdump and provided the relocatability idea+patches for it. I did
not spend time debugging this kdump problem, though.
I want everything. Specifically I wanted to make sure that kexec didn't log any
further messages to the syslog other than whats been reported here. What I
specifically do not want is incomplete log files. I don't know for certain if
anything else in there will help me, but until I get a reproducer machine set
up, I don't want to be looking at logs that may be missing information.
It was even tested it isn't loadable immediately after boot where I still have
seen all the messages myself. Not going to provide log of my private mails.
Normally, I wouldn't bother pursuing a problem for which you wouldn't give me
full logs, but such as it is I managed to find an fc6 machine and reproduced the
problem. It appears this is a kernel bug. More recent fc6 and rawhide kernels
contain an extra PT_LOAD segment which has a page alignment, but unaligned
physical and virtual address values. I'll take this up with davej
Created attachment 136941 [details]
patch to align kernel data segment to page boundary
This is now looking more like a linker bug. The later linux kernels have a
broken out PT_LOAD segment for the .data and .notes segments, and despite the
fact that the .data segment has a p_align value of 0x1000, the p_vaddr and
p_offset values do not end on 4096 byte boundaries. I'll file a linker bug,
but for now, this adjustment patch to the vmlinux linker script brings the data
segment back into alignment.
I've sent the patch off to davej for review, and opened up a binutils bug for
the incorrect alignment output.
Thanks for checking this issue.
Fix is still not present in kernel-2.6.18-1.2693.fc6, it is even not MODIFIED as
it is still ASSIGNED.
It'll be there when davej gets to it.
See also bug #215268 for FC5.