[PATCH 18/19] arm64: kdump: update a kernel doc

[AKASHI Takahiro]

On 01/19/2016 09:10 PM, Mark Rutland wrote:
> On Tue, Jan 19, 2016 at 02:31:05PM +0900, AKASHI Takahiro wrote:
>> On 01/18/2016 08:29 PM, Mark Rutland wrote:
>>> On Mon, Jan 18, 2016 at 07:26:04PM +0900, AKASHI Takahiro wrote:
>>>> On 01/16/2016 05:16 AM, Mark Rutland wrote:
>>>>> On Fri, Jan 15, 2016 at 07:18:38PM +0000, Geoff Levand wrote:
>>>>>> From: AKASHI Takahiro <takahiro.akashi at linaro.org>
>>>>>>
>>>>>> This patch adds arch specific descriptions about kdump usage on arm64
>>>>>> to kdump.txt.
>>>>>>
>>>>>> Signed-off-by: AKASHI Takahiro <takahiro.akashi at linaro.org>
>>>>>> ---
>>>>>>   Documentation/kdump/kdump.txt | 23 ++++++++++++++++++++++-
>>>>>>   1 file changed, 22 insertions(+), 1 deletion(-)
>>>>>>
>>>>>> diff --git a/Documentation/kdump/kdump.txt b/Documentation/kdump/kdump.txt
>>>>>> index bc4bd5a..36cf978 100644
>>>>>> --- a/Documentation/kdump/kdump.txt
>>>>>> +++ b/Documentation/kdump/kdump.txt
>>>>>> @@ -18,7 +18,7 @@ memory image to a dump file on the local disk, or across the network to
>>>>>>   a remote system.
>>>>>>
>>>>>>   Kdump and kexec are currently supported on the x86, x86_64, ppc64, ia64,
>>>>>> -s390x and arm architectures.
>>>>>> +s390x, arm and arm64 architectures.
>>>>>>
>>>>>>   When the system kernel boots, it reserves a small section of memory for
>>>>>>   the dump-capture kernel. This ensures that ongoing Direct Memory Access
>>>>>> @@ -249,6 +249,20 @@ Dump-capture kernel config options (Arch Dependent, arm)
>>>>>>
>>>>>>       AUTO_ZRELADDR=y
>>>>>>
>>>>>> +Dump-capture kernel config options (Arch Dependent, arm64)
>>>>>> +----------------------------------------------------------
>>>>>> +
>>>>>> +1) The maximum memory size on the dump-capture kernel must be limited by
>>>>>> +   specifying:
>>>>>> +
>>>>>> +   mem=X[MG]
>>>>>> +
>>>>>> +   where X should be less than or equal to the size in "crashkernel="
>>>>>> +   boot parameter. Kexec-tools will automatically add this.
>>>>>
>>>>>
>>>>> This is extremely fragile, and will trivially fail when the kernel can
>>>>> be loaded anywhere (see [1]).
>>>>
>>>> As I said before, this restriction also exists on arm, but I understand
>>>> that recent Ard's patches break it.
>>>>
>>>>> We must explicitly describe the set of regions the crash kernel may use
>>>>> (i.e. we need base and size). NAK in the absence of that.
>>>>
>>>> There seem to exist several approaches:
>>>> (a) use a device-tree property, "linux,usable-memory", in addition to "reg"
>>>
>>> I'm not opposed to the idea of a DT property, though I think that should
>>> live under /chosen.
>>
>> In fact, powerpc uses another property, "linux,crashkernel-base(& size)",
>> under /chosen in order for the *1st kernel* to export info about a memory
>> region for the 2nd(crash dump) kernel to user apps (kexec-tools).
>
> Do you mean that said property is provided _to_ the 1st kernel, or
> provided _by_ the first kernel?

_by_ the 1st kernel.

Based on a kernel parameter, "crashkernel=", the 1st kernel reserve some
memory region at boot time and export its information through this property.
Most architectures other than powerpc, however, use an iomem resource entry,
"Crash kernel", in /proc/iomem instead for this purpose.

>>> I see that "linux,usable-memory" exists already, though I'm confused as
>>> to exactly what it is for as there is no documentation (neither in the
>>> kernel nor in ePAPR).
>>
>> For example,
>>    memory at 0x80000000 {
>>      reg = <0x0 0x80000000 0x0 0x80000000>;
>>      linux,usable-memory = <0x0 0x8c000000 0x0 0x4000000>;
>>    }
>> There exists 2GB memory available on the system, but the last 64MB can be
>> used as a system ram. See early_init_dt_scan_memory() in fdt.c.
>
> Sure, except that's the implementation rather than the intended
> semantics (which are not defined).

Yeah, but the code itself was ack'ed (actually committed) by Grant:)

>>> It's also painful to alter multiple memory nodes
>>> to use that, and I can see that going wrong.
>>
>> Yeah, I implemented this feature in my old versions experimentally,
>> but didn't like it as we had to touch all the memory nodes.
>>
>>>>      under "memory" node
>>>> (b) use a kernel's early parameter, "memmap=nn[@#$]ss"
>>>
>>> I'm not too keen on this, as I think it's fragile, and logically
>>> somewhat distinct from what mem= is for (a best effort testing tool).
>>
>> I'm not sure whether it is fragile, and contrary to x86, as Dave
>> described, I think we will only need a single memmap= on arm64 as
>> efi's mem map table is accessible even on the crash kernel.
>
> I just realised I misread this as "mem=", apologies.
>
> It looks like memmap= to force a specific region of memory to be used
> may work.
>
> I'd still err on the side of preferring an explicit property in the DT.

Let's discuss in succeeding replies.

>>>> Power PC takes (a), while this does not work on efi-started kernel
>>>> because dtb has no "memory" nodes under efi.
>>>
>>> A property under /chosen would work for EFI too.
>>>
>>>> X86 takes (b). If we take this, we will need to overwrite a weak
>>>> early_init_dt_add_memory().
>>>> (I thought that this approach was not smart as we have three different
>>>> ways to specify memory regions, dtb, efi and this kernel parameter.)
>>>
>>> I'm not sure that's a big problem. We may be able to make this generic,
>>> also.
>>>
>>> We don't necessarily need a weak add memory function if we can guarantee
>>> nothing gets memblock_alloc'd before we carve it out.
>>>
>>> Something like the nomap stuff Ard put together might be useful here.
>>
>> I'm afraid it doesn't work.
>> It doesn't matter whether it is linearly mapped or not. We should prevent
>> any part of memory regions used by the 1st kernel from being reclaimed
>> by memblock_alloc() and others.
>
> Are you certain that nomap memory can be allocated? That sounds like a
> major bug.

I misunderstood. __next_mem_range() called by mem_alloc stuff has some check.

-Takahiro AKASHI

> Nomap memory should act like reserved memory with the additional
> property that the kernel must not map it implicitly.
>
>> Or do you mean we can introduce another memblock flag?
>
> That wasn't what I meant, but that would be a potential solution.
>
> Thanks,
> Mark.
>