On 9 Feb 2024, at 14:36, Zi Yan wrote:
> On 9 Feb 2024, at 11:37, Vlastimil Babka wrote:
>
>> On 2/2/24 17:15, Zi Yan wrote:
>>> From: Zi Yan <ziy@xxxxxxxxxx>
>>>
>>> Before last commit, memory compaction only migrates order-0 folios and
>>> skips >0 order folios. Last commit splits all >0 order folios during
>>> compaction. This commit migrates >0 order folios during compaction by
>>> keeping isolated free pages at their original size without splitting them
>>> into order-0 pages and using them directly during migration process.
>>>
>>> What is different from the prior implementation:
>>> 1. All isolated free pages are kept in a NR_PAGE_ORDERS array of page
>>> lists, where each page list stores free pages in the same order.
>>> 2. All free pages are not post_alloc_hook() processed nor buddy pages,
>>> although their orders are stored in first page's private like buddy
>>> pages.
>>> 3. During migration, in new page allocation time (i.e., in
>>> compaction_alloc()), free pages are then processed by post_alloc_hook().
>>> When migration fails and a new page is returned (i.e., in
>>> compaction_free()), free pages are restored by reversing the
>>> post_alloc_hook() operations using newly added
>>> free_pages_prepare_fpi_none().
>>>
>>> Step 3 is done for a latter optimization that splitting and/or merging free
>>> pages during compaction becomes easier.
>>>
>>> Note: without splitting free pages, compaction can end prematurely due to
>>> migration will return -ENOMEM even if there is free pages. This happens
>>> when no order-0 free page exist and compaction_alloc() return NULL.
>>>
>>> Signed-off-by: Zi Yan <ziy@xxxxxxxxxx>
>>
>> ...
>>
>>> /*
>>> @@ -1835,9 +1857,17 @@ static struct folio *compaction_alloc(struct folio *src, unsigned long data)
>>> static void compaction_free(struct folio *dst, unsigned long data)
>>> {
>>> struct compact_control *cc = (struct compact_control *)data;
>>> + int order = folio_order(dst);
>>> + struct page *page = &dst->page;
>>> +
>>> + folio_set_count(dst, 0);
>>
>> We can't change refcount to 0 like this, after it was already set to 1 and
>> somebody else might have done get_page_unless_zero(). You need to either
>> put_page_testzero() and if it's false, consider the page lost, or leave it
>> refcounted and adjust the code to handle both refcounted and non-refcounted
>> pages on the lists (the first option is simpler and shouldn't be too bad).
> Got it. Will fix it with the first option. Thanks.
Do you think we should have a WARN or WARN_ONCE if we lose a page here?
--
Best Regards,
Yan, Zi
Attachment:
signature.asc
Description: OpenPGP digital signature
