[PATCH v1 04/11] fs/proc/vmcore: move vmcore definitions from kcore.h to crash_dump.h

[David Hildenbrand]

On 20.11.24 10:42, Baoquan He wrote:
> On 11/15/24 at 10:59am, David Hildenbrand wrote:
>> On 15.11.24 10:44, Baoquan He wrote:
>>> On 10/25/24 at 05:11pm, David Hildenbrand wrote:
>>>> These defines are not related to /proc/kcore, move them to crash_dump.h
>>>> instead. While at it, rename "struct vmcore" to "struct
>>>> vmcore_mem_node", which is a more fitting name.
>>>
>>> Agree it's inappropriate to put the defintions in kcore.h. However for
>>> 'struct vmcore', it's only used in fs/proc/vmcore.c from my code
>>> serching, do you think if we can put it in fs/proc/vmcore.c directly?
>>> And 'struct vmcoredd_node' too.
>>
>> See the next patches and how virtio-mem will make use of the feactored out
>> functions. Not putting them as inline functions into a header will require
>> exporting symbols just do add a vmcore memory node to the list, which I want
>> to avoid -- overkill for these simple helpers.
> 
> I see. It makes sense to put them in crash_dump.h. Thanks for
> explanation.
> 

I'll add these details to the description.

>>
>>>
>>> And about the renaming, with my understanding each instance of struct
>>> vmcore represents one memory region, isn't it a little confusing to be
>>> called vmcore_mem_node? I understand you probablly want to unify the
>>> vmcore and vmcoredd's naming. I have to admit I don't know vmcoredd well
>>> and its naming, while most of people have been knowing vmcore representing
>>> memory region very well.
>>
>> I chose "vmcore_mem_node" because it is a memory range stored in a list.
>> Note the symmetry with "vmcoredd_node"
> 
> I would say the justification of naming "vmcore_mem_node" is to keep
> symmetry with "vmcoredd_node". If because it is a memory range, it really
> should not be called vmcore_mem_node. As we know, memory node has
> specific meaning in kernel, it's the memory range existing on a NUMA node.
> 
> And vmcoredd is not a widely used feature. At least in fedora/RHEL, we
> leave it to customers themselves to use and handle, we don't support it.
> And we add 'novmcoredd' to kdump kernel cmdline by default to disable it
> in fedora/RHEL. So a rarely used feature should not be taken to decide
> the naming of a mature and and widely used feature's name. My personal
> opinion.

It's a memory range that gets added to a list. So it's a node in a list 
... representing a memory range. :) I don't particularly care about the 
"node" part here.

The old "struct vmcore" name is misleading: makes one believe it somehow 
represents "/proc/vmcore", but it really doesn't. (see below on function 
naming)

> 
>>
>> If there are strong feelings I can use a different name, but
> 
> Yes, I would suggest we better keep the old name or take a more
> appropriate one if have to change.

In light of patch #5 and #6, really only something like 
"vmcore_mem_node" makes sense. Alternatively "vmcore_range" or 
"vmcore_mem_range".

Leaving it as "struct vmcore" would mean that we had to do in #5 and #6:

* vmcore_alloc_add_mem_node() -> vmcore_alloc_add()
* vmcore_free_mem_nodes() -> vmcore_free()

Which would *really* be misleading, because we are not "freeing" the vmcore.

Would "vmcore_range" work for you? Then we could do:

* vmcore_alloc_add_mem_node() -> vmcore_alloc_add_range()
* vmcore_free_mem_nodes() -> vmcore_free_ranges()

-- 
Cheers,

David / dhildenb