Re: [PATCH v3 6/7] mm/memblock: Use KSTATE instead of kho to preserve preserved_mem_table

[Andrey Ryabinin <ryabinin.a.a@gmail.com>]

On 9/15/25 1:47 PM, Jason Gunthorpe wrote:
> On Tue, Sep 09, 2025 at 10:14:41PM +0200, Andrey Ryabinin wrote:
>> +static int kstate_preserve_phys(struct kstate_stream *stream, void *obj,
>> +				const struct kstate_field *field)
>> +{
>> +	struct reserve_mem_table *map = obj;
>> +
>> +	return kho_preserve_phys(map->start, map->size);
>> +}
>> +
>> +struct kstate_description kstate_reserve_mem = {
>> +	.name = "reserved_mem",
>> +	.id = KSTATE_RESERVED_MEM_ID,
>> +	.fields = (const struct kstate_field[]) {
>> +		KSTATE_BASE_TYPE(name, struct reserve_mem_table,
>> +				char[RESERVE_MEM_NAME_SIZE]),
>> +		KSTATE_BASE_TYPE(start, struct reserve_mem_table, phys_addr_t),
>> +		KSTATE_BASE_TYPE(size, struct reserve_mem_table, phys_addr_t),
>> +		{
>> +			.name = "phys_range",
>> +			.flags = KS_CUSTOM,
>> +			.save = kstate_preserve_phys,
>> +		},
>> +		KSTATE_END_OF_LIST(),
>> +	},
>> +};
>>  
>>  static int __init reserve_mem_init(void)
>>  {
>>  	int err;
>> +	int i;
>>  
>>  	if (!kho_is_enabled() || !reserved_mem_count)
>>  		return 0;
>>  
>> +	for (i = 0; i < reserved_mem_count; i++) {
>> +		struct reserve_mem_table *map = &reserved_mem_table[i];
>>  
>> +		err = kstate_register(&kstate_reserve_mem,
>> +				map, crc32(~0, map->name, RESERVE_MEM_NAME_SIZE));
>> +		if (err)
>> +			goto out;
>>  	}
> 
> As I've said to the other proposals, this doesn't seem to be bringing
> that much value compared to just using a normal struct:

We expect to have many such ABI maps across the kernel.
These maps will share common elements - simple types, folios, and preserved
regions.

With the approach you're suggesting, we'd need to re-implement the same
preserve/unpreserve/recover logic, error handling, and unwind code for
every individual ABI map. That quickly becomes repetitive and error-prone.

By contrast, KSTATE centralizes this logic. It avoids duplicating code
and lets us express the preservation details declaratively instead
of re-implementing them per struct.


> 	for (i = 0; i < reserved_mem_count; i++) {
> 		struct reserve_mem_table *map = &reserved_mem_table[i];
> 		struct khoser_reserve_mem_table abi_map = {.name = map->name. .start = map->start, .size = map->size};
> 
> 		err = kho_preserve_phys(map->start, map->size);
> 		if (err)
> 		    return err; // Should unwind the other preservations!
> 		
> 		luo_preserve_key(luo_obj, map->name, &abi_map, sizeof(abi_map), VERSION_0);


On the versioning side:
With this approach, introducing a new ABI version (say, abi_map_v1)
would require us to maintain restore logic for each supported version,
and carefully handle upgrades between them.

With KSTATE, versioning is built in. For example, adding a new field can
simply be expressed as:
 	KSTATE_BASE_TYPE_V(new_field, struct reserve_mem_table, int, 1);
This way, the framework handles compatibility, and we don’t need to manually
write version-specific restore paths for each ABI map.