----- Original Message -----

Hi Dave,




In my develop enviroment, kaslr offset will be stored in SoC on-chip memory
for debuging, but crash utility is not work if i only set kaslr offset by
command '--kaslr XXX', it also need correct kimage_voffset.

error message without the patch:

VA_BITS: 39
using 40000000 as phys_offset
kimage_voffset: 0
phys_offset: 40000000
.....cut....
<readmem: ffffff9381e074a0, KVADDR, "cpu_possible_mask", 8, (FOE),
7fffce0dfa48>
<read_ramdump: addr: ffffff9381e074a0 paddr: 13b9e074a0 cnt: 8>
read_ramdump: READ_ERROR: offset not found for paddr: 13b9e074a0
crash: read error: kernel virtual address: ffffff9381e074a0 type:
"cpu_possible_mask"

This patch add a function to auto calculate kimage_voffset by kaslr offset.
Please review it.

Thanks,
Yueyi Li

Hello Yueyi,

I have a few questions about this patch.

First, can you clarify the kernel version and the dumpfile type that
you are trying to address?

I understand the confusion, given that arm64 kdump/KASLR support is still being
debated upstream.  Eventually, arm64 kdump-generated dumpfiles will have
"NUMBER(kimage_voffset)" and "NUMBER(PHYS_OFFSET)" VMCOREINFO entries, which
will simplify things significantly (and you can see that the crash utility
already has support for them in advance).
Actually, i use crash utility to analysis Android device crash, Linux kernel version is 4.4. The ramdump file is not generated by kdump, it is whole DDR SDRAM snapshot collected by another RTOS, so i don`t have vmcoreinfo to get kimage_voffset & phys_offset. I my opinion, only set kaslr offset should work well even if without vmcore file and kimage_voffse, it looks more reasonable.
It seems that this patch would only be applicable when looking at an ELF
dumpfile, given that your arch_calc_kimage_voffset() function calls
arm_kdump_phys_base() -- which only applies if you are looking at an
ELF dumpfile:

+static void
+arm64_calc_kimage_voffset(void)
+{
+       struct machine_specific *ms = machdep->machspec;
+       ulong kimage_offset;
+       ulong phys_offset;
+       struct syment *sp;
+
+       if (machdep->flags & KIMG_VOFFSET) /* --machdep override */
+               return;
+
+       if ((kt->flags2 & KASLR) && (kt->flags & RELOC_SET)){
+               arm_kdump_phys_base(&phys_offset);
+               ms->kimage_voffset = ms->vmalloc_start_addr + (kt->relocate * -1) - phys_offset;
+       }
+
+}

Is that correct?  And if so, why not call arm64_kdump_phys_base() instead?
Because I see in your "error message without the patch" above, the phys_base 
seems to have been calculated correctly to be 40000000, so it looks like the 
call from arm64_calc_phys_offset() to arm64_kdump_phys_base() must have worked.
(since arm64_kdump_phys_base() ultimately calls arm_kdump_phys_base() if nothing
else works) 

So what would happen if you called arm64_calc_phys_offset() *before* your new
function?  Wouldn't it correctly determine phys_offset?  And then there would
be no need for you to call it in your new function:

@@ -324,6 +325,9 @@ arm64_init(int when)
                machdep->init_kernel_pgd = arm64_init_kernel_pgd;

                /* use machdep parameters */
+               arm64_calc_kimage_voffset();
+
+               /* use machdep parameters */
                arm64_calc_phys_offset();
When kernel ASLR is enabled on arm64, the kernel image start address can be randomized, phys_offset is no more equal the start address of first memory block. It means, arm_kdump_phys_base() can not calculate correctly address, we need kimage_voffset to calculate correctly phys_offset, this is what arm64_calc_phys_offset() to do( just on live system & ELF ramdump). So arm64_calc_kimage_voffset() should be called first. This is why crash utility is not work if '-m kimage_voffset' is not given.

debug log with the patch, kaslr offset is 0x1378e00000, start address of first memory block is 0x40000000:
readmem: read_ramdump()                                                                  
VA_BITS: 39                                                                              
using ffffffd340000000 as phys_offset                                                    
kimage_voffset: ffffff9340e00000                                                         
phys_offset: ffffffd340000000                                                            
gdb vmlinux                                                                              
GNU gdb (GDB) 7.6                                                                        
Copyright (C) 2013 Free Software Foundation, Inc.                                        
License GPLv3+: GNU GPL version 3 or later <http://gnu.org/licenses/gpl.html>            
This is free software: you are free to change and redistribute it.                       
There is NO WARRANTY, to the extent permitted by law.  Type "show copying"               
and "show warranty" for details.                                                         
This GDB was configured as "--host=x86_64-unknown-linux-gnu --target=aarch64-elf-linux"...
                                                                                         
cpu_possible_mask: cpus: 0 1 2 3 4 5 6 7                                                 
 cpu_present_mask: cpus: 0 1 2 3 4 5 6 7                                                 
  cpu_online_mask: cpus: 0 1 2 3 4 5 6                                                   
  cpu_active_mask: cpus: 0 1 2 3 4 5 6                                                   
xtime timespec.tv_sec: 58c27009: Fri Mar 10 17:21:13 2017                                
utsname:                                                                                 
     sysname: Linux                                                                      
    nodename: localhost                                                                  
     release: 4.4.21-perf-g5ec6736-00564-gbdc7dcf                                        
     version: #1 SMP PREEMPT Wed Mar 8 23:02:12 CST 2017                                 
     machine: aarch64                                                                    
  domainname: localdomain                                                                 
In addition, arm64_calc_phys_offset() also tries to determine the phys_offset value
for compressed kdump dumpfiles and live systems.  If the dumpfile has been
compressed (by the makedumpfile facility), or if you are running on a live system,
then your patch would fail to determine phys_offset.  But if you were to call 
arm64_calc_phys_offset() instead, it would call diskdump_phys_base() for compressed
dumpfiles, or try to determine the phys_offset value from /proc/iomem on live
systems.

Dave


--
Crash-utility mailing list
Crash-utility@redhat.com
https://www.redhat.com/mailman/listinfo/crash-utility
I`m not familar with compressed dumpfiles, just add a judgment in my function. For live system, if kimage_voffset has been calculated, arm64_calc_phys_offset will use it calculate correctly phys_offset instead of read from /proc/iomem.

        if ((machdep->flags & NEW_VMEMMAP) &&
            ms->kimage_voffset && (sp = kernel_symbol_search("memstart_addr"))) {
            paddr =    sp->value - machdep->machspec->kimage_voffset;
            if (READMEM(pc->mfd, &phys_offset, sizeof(phys_offset),
                sp->value, paddr) > 0) {
                ms->phys_offset = phys_offset;
                return;
            }
        }

I have attached patch v2, Could you please review it?

Thanks,
Yueyi Li