[PATCH v26 0/7] arm64: add kdump support

[Manish Jaggi]

On 10/04/2016 04:23 PM, James Morse wrote:
> Hi Manish,
> 
> On 04/10/16 11:05, Manish Jaggi wrote:
>> On 10/04/2016 03:16 PM, James Morse wrote:
>>> On 03/10/16 13:41, Manish Jaggi wrote:
>>>> On 10/03/2016 04:34 PM, AKASHI Takahiro wrote:
>>>>> On Mon, Oct 03, 2016 at 01:24:34PM +0530, Manish Jaggi wrote:
>>>>>> First kernel is booted with mem=2G crashkernel=1G command line option.
>>>>>> While the system has 64G memory.
>>>
>>>>> Are you saying that "mem=..." doesn't have any effect?
>>>> What I am saying it that If the first kernel is booted using mem= option and crashkernel= option
>>>> the memory for second kernel has to be withing the crashkernel size.
>>>> As per /proc/iomem System RAM the information is correct, but the /proc/meminfo is showing total memory
>>>> much more than the first kernel had in first place.
>>>
>>> So your second crashkernel has 63G of memory? Unless you provide the same 'mem='
>>> to the kdump kernel, this is the expected behaviour. The
>>> DT:/reserved-memory/crash_dump describes the memory not to use.
>>>
>>> On your first boot with 'mem=2G' memblock_mem_limit_remove_map() called from
>>> arm64_memblock_init() removed the top 62G of memory. Neither the first kernel
>>> nor kexec-tools know about the top 62G.
>>> When you run kexec-tools, it describes what it sees in /proc/iomem in the
>>> DT:/reserved-memory/crash_dump, which is just the remaining 1G of memory.
>>>
>>> When we crash and reboot, the crash kernel discovers all 64G of memory from the
>>> EFI memory map.
> 
>> So the iomem and meminfo should be same or different for the second kernel?
>> Also i assumed that crashkernel=1G should restrict the second kernels to 1G.
> 
> Not with v26 of this series. What should it do with the 62G of memory that was
> removed by booting with 'mem=2G'? It isn't part of the crashkernel reserved
> area, and it isn't part of the vmcore described in elfcorehdr either...
> 
> 
>> This is my understanding from the description. It should not require a second mem= option
> 
>>> kexec-tools described the 1G of memory that the first kernel was using in the
>>> DT:/reserved-memory/crash_dump node, so early_init_fdt_scan_reserved_mem()
>>> reserves the 1G of memory the first kernel used. This leaves us with 63G of memory.
>>>
>>> This may change with the next version of kdump if it switches back to using
>>> DT:/chosen/linux,usable-memory-range.
>>> If you need v26 to avoid the top 62G of memory, you need to provide the same
>>> 'mem=' to the first and second kernel.
> 
>> If I provide for second kernel, I dont see any prints after Bye.
>> Have you tired this anytime?
> 
> Yes, on juno-r1 passing 'mem=2G' to both the first and second kernel causes only
> the first 2G of memory to be used with this pattern:
> first kernel:		[1G used for linux]	[1G reserved for Crash kernel] 	[6G memory
> hidden]
> kdump kernel:	[1G vmcore]			[1G used for linux] 			[6G memory hidden]
> 
> 
Oh, ok!
I was giving mem=1G to crashkernel to test. with mem=2G it works.
>>>>>> 1.2 Live crash dump fails with error
>>>
>>> ... do we expect this to work? I don't think it has anything to do with this
>>> series...
>>>
>> Why it should not?
>> I saved the vmcore file while in second kernel. Since crash without vmcore file didnt run,
>> Tried with vmcore file and it worked. Its just that if you want to boot a second kernel
>>  with read only file system without network live crash dump analysis is handy.
> 
> Ah, you want to run /usr/bin/crash with the kdump boot of linux. You still need
> to tell it where to find the memory image: "crash /path/to/vmlinux /proc/vmcore"
> should do the trick.
> 
We should fix the documentation of kdump them.
Since it is not supported, it should be removed.
> 
> Thanks,
> 
> James
>