Re: [PATCH v5 mm-new 0/6] mm: hugetlb: allocate frozen gigantic folio

[Claudiu Beznea <claudiu.beznea@tuxon.dev>]

Hi,

On 1/7/26 21:38, Zi Yan wrote:
> On 7 Jan 2026, at 13:39, Mark Brown wrote:
> 
>> On Wed, Jan 07, 2026 at 07:31:30PM +0200, Claudiu Beznea wrote:
>>> On 12/30/25 09:24, Kefeng Wang wrote:
>>
>>>> Introduce alloc_contig_frozen_pages() and cma_alloc_frozen_compound()
>>>> which avoid atomic operation about page refcount, and then convert to
>>>> allocate frozen gigantic folio by the new helpers in hugetlb to cleanup
>>>> the alloc_gigantic_folio().
>>>
>>> I'm seeing the following issues on the Renesas RZ/G3S SoC when doing suspend
>>> to idle:
>>>
>>> [  129.636729] Unable to handle kernel paging request at virtual address
>>> dead000000000108
>>> [  129.644674] Mem abort info:
>>> [  129.647456]   ESR = 0x0000000096000044
>>
>> This is also introducing OOMs when doing at least audio tests (I don't
>> think these are super relevant) on Raspberry Pi 3B+ running NFS root
>> (probably more relevant):
>>
>> [   64.064256] Unable to handle kernel paging request at virtual address fffffdffc1000000
>>
>> ...
>>
>> [   64.087583] Call trace:
>> [   64.087586]  kmem_cache_free+0x88/0x434 (P)
>> [   64.087598]  skb_free_head+0x9c/0xb8
>> [   64.087608]  skb_release_data+0x120/0x174
>> [   64.087615]  __kfree_skb+0x2c/0x44
>> [   64.087622]  tcp_data_queue+0x948/0xe50
>>
>> Full log:
>>
>>    https://lava.sirena.org.uk/scheduler/job/2341856#L1721
>>
>> Bisection identifies:
>>
>> [fb9a328d30400dbc8b2ea5a57daeb28bedac398b] mm: cma: add cma_alloc_frozen{_compound}()
>>
>> as being the comit that introduces the issue.  Bisect log with links to
>> further test runs:
> 
> Hi Mark and Claudiu,
> 
> Can you try the patch below to see if it fixes the issue? Basically,
> in cma_release(), count was used to drop page ref and decreased to 0,
> but after the loop, count becomes -1 and __cma_release_frozen()
> is releasing unnecessary pages.
> 
> Thanks.
> 
> 
>  From ece23da65ea7210e1fcb51ee9c27aec19b84811c Mon Sep 17 00:00:00 2001
> From: Zi Yan <ziy@nvidia.com>
> Date: Wed, 7 Jan 2026 14:23:15 -0500
> Subject: [PATCH] mm/cma: fix cma_release by not decreasing count to 0.
> Content-Type: text/plain; charset="utf-8"
> 
> Signed-off-by: Zi Yan <ziy@nvidia.com>
> ---
>   mm/cma.c | 3 ++-
>   1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/mm/cma.c b/mm/cma.c
> index 5713becc602b..408b07f6fddd 100644
> --- a/mm/cma.c
> +++ b/mm/cma.c
> @@ -1013,13 +1013,14 @@ bool cma_release(struct cma *cma, const struct page *pages,
>   {
>   	struct cma_memrange *cmr;
>   	unsigned long pfn;
> +	unsigned long i;
> 
>   	cmr = find_cma_memrange(cma, pages, count);
>   	if (!cmr)
>   		return false;
> 
>   	pfn = page_to_pfn(pages);
> -	for (; count--; pfn++)
> +	for (i = 0; i < count; i++, pfn++)
>   		VM_WARN_ON(!put_page_testzero(pfn_to_page(pfn)));
> 
>   	__cma_release_frozen(cma, cmr, pages, count);

This solves my use case as well. If any, you can add:

Tested-by: Claudiu Beznea <claudiu.beznea.uj@bp.renesas.com>

Thank you,
Claudiu