Re: Request for Feedback on Releasing Reserved Memory Back to the Buddy Allocator

[David Hildenbrand <david@redhat.com>]

On 14.04.25 22:25, Harry Yoo wrote:
> On Wed, Apr 09, 2025 at 05:28:00PM +0300, moatasem zaaroura wrote:
>> Dear Linux MM Community,
>>
>> I am working on a system that requires reserving a known physical
>> memory region during the early boot phase and later releasing it back
>> to the kernel for general use. I would highly appreciate your feedback
>> on the approach I’ve taken, including any concerns, possible pitfalls,
>> or alternative recommendations.
>>
>> == Problem Context ==
>>
>> In my use case, the boot manager must copy data from flash to a known
>> DRAM location while the Linux kernel is still booting. This data is
>> then used in user space. After the user-space component finishes using
>> the data, I want to release this memory back to the system so it can
>> be utilized by the buddy allocator.
>>
>> == My Solution ==
>>
>> 1. I reserved the memory region using a "reserved-memory" node in the
>> device tree with a fixed physical address and size.
>>
>> 2. This address is shared with the boot manager, which copies the
>> required data there before the kernel accesses it.
>>
>> 3. After the data is no longer needed (in user space), I expose a
>> sysfs interface to manually trigger the release of this reserved
>> memory back to the kernel.
>>
>> == Freeing Logic ==
>>
>> In the release function:
>> - I validate that the physical address (cache_addr) is page-aligned.
>> - I calculate the PFN using: pfn = PFN_DOWN(cache_addr);
>> - Then I loop over the pages:
>>
>>          size_t i;
>>          struct page *page;
>>          unsigned long pfn;
>>          unsigned long number_of_pages = cache_size >> PAGE_SHIFT;
>>
>>          if (cache_addr & (PAGE_SIZE - 1)) {
>>              pr_err("Physical address is not page-aligned\n");
>>              return;
>>          }
>>
>>          pfn = PFN_DOWN(cache_addr);
>>          for (i = 0; i < number_of_pages; i++) {
>>              page = pfn_to_page(pfn);
>>
>>              // Ensure the page is not part of the reserved pages
> 
> You mean "Ensure the page is the part of the reserved pages" ?
> 
>>              if (PageReserved(page))
>>                  free_reserved_page(page);
>>              pfn += 1;
>>          }
>>
>> - I verified that the memory is successfully returned to the buddy
>> allocator by observing the increased number of free pages at
>> /proc/buddyinfo.
>>
>> == What I'm Asking For ==
>>
>> - Is this approach correct and safe under the current kernel memory
>> management design?
>> - Are there any problems I may have missed?
>> - Is there a better or more canonical way to achieve this?
> 
> I think memory hotplug [1] is an existing Linux approach to achieve this.
> 
> Does your reserved memory show up as a memory device in
> /sys/devices/system/memory/memoryXXX?

It's a memblock allocation, so that memory would already be added to the 
system and hotplug does not apply.

-- 
Cheers,

David / dhildenb