[Crash-utility] Additional fix for per-cpu data symbol naming problem

[Dave Anderson]

----- "Ryota Ozaki" <ozaki.ryota at gmail.com> wrote:

> Hi,
> 
> I tried crash 5.0.3 for linux 2.6.34 (x86_64) and faced the problem
> about per-cpu data symbol naming, which is described in ChangeLog
> as it's already fixed. I observed to fix the problem and found the
> following fix for x86_64_verify_symbol() in x86_64.c helped me.
> 
> I'm not sure whether the fix is correct or not, but it would be help for
> correct fix.
> 
> Thanks,
>   ozaki-r
> 
> 
> --- x86_64.c.orig       2010-05-19 12:26:11.000000000 +0000
> +++ x86_64.c    2010-05-19 08:23:30.000000000 +0000
> @@ -1974,7 +1974,7 @@
>                         if (STRNEQ(name, "per_cpu") ||
>                             STREQ(name, "__per_cpu_end"))
>                                 return TRUE;
> -                       if (type == 'V')
> +                       if (type == 'V' || type == 'D' || type == 'd')
>                                 return TRUE;
>                 }

Yes, I already have that same patch queued for the next release,
which I hope to have ready today or tomorrow.

The fix in 5.0.3 was based upon my own builds of 2.6.34 kernels
that introduced the new per-cpu symbol naming scheme.  But when
I build x86_64 kernels, the per-cpu symbol types are type 'V'.

For example, here's a 2.6.34-rc5 kernel that I built -- note there
are few absolute ('a' or 'A') symbols after "__per_cpu_start", but
all the "real" per-cpu symbols are type 'V':

# strings vmlinux | grep "Linux version"
Linux version 2.6.34-rc5 (root at intel-d3c4702-01.rhts.eng.bos.redhat.com) (gcc version 4.4.3 20100121 (Red Hat 4.4.3-1) (GCC) ) #1 SMP Fri Apr 30 18:20:16 EDT 2010
# nm -Bn vmlinux
                 w __crc_softirq_work_list
                 U do_hypervisor_callback
0000000000000000 a L3_PAGE_OFFSET
0000000000000000 A VDSO32_PRELINK
0000000000000000 D __per_cpu_start
0000000000000000 V irq_stack_union
0000000000000000 A xen_irq_disable_direct_reloc
0000000000000000 A xen_save_fl_direct_reloc
0000000000000040 A VDSO32_vsyscall_eh_frame_size
0000000000000100 a i
0000000000000100 a vector
0000000000000110 a L4_PAGE_OFFSET
00000000000001e7 A kexec_control_code_size
00000000000001f0 A VDSO32_NOTE_MASK
00000000000001fe a L3_START_KERNEL
00000000000001ff a L4_START_KERNEL
0000000000000400 A VDSO32_sigreturn
0000000000000410 A VDSO32_rt_sigreturn
0000000000000420 A VDSO32_vsyscall
0000000000000430 A VDSO32_SYSENTER_RETURN
0000000000004000 V gdt_page
0000000000005000 V exception_stacks
000000000000b000 V xen_vcpu
000000000000b020 V xen_vcpu_info
000000000000b060 V idt_desc
000000000000b070 V xen_cr0_value
000000000000b080 V xen_mc_irq_flags
000000000000b0a0 V mc_buffer
000000000000c5b0 V xen_cr3
000000000000c5b8 V xen_current_cr3
... [ cut ] ...
0000000000016700 V cpu_tlbstate
0000000000016740 V runqueues
0000000000017100 V sched_clock_data
0000000000017140 V cfd_data
0000000000017180 V call_single_queue
00000000000171c0 V csd_data
00000000000171e8 D __per_cpu_end
...

Although I still don't understand why it happens (compiler?
CONFIG option?), a patch to continue recognizing 'd' and 'D'
symbols is obviously needed.

Thanks,
  Dave