On 08.01.25 21:54, Matthew Wilcox wrote:
On Wed, Jan 08, 2025 at 09:14:53PM +0100, David Hildenbrand wrote:
On 08.01.25 18:32, Matthew Wilcox wrote:
On Wed, Jan 08, 2025 at 04:18:42PM +0000, Lorenzo Stoakes wrote:
@@ -280,7 +269,10 @@ static void fb_deferred_io_work(struct work_struct *work)
struct folio *folio = page_folio(pageref->page);
folio_lock(folio);
- folio_mkclean(folio);
+ rmap_wrprotect_file_page(fbdefio->mapping,
+ pageref->offset >> PAGE_SHIFT,
+ compound_nr(pageref->page),
+ page_to_pfn(pageref->page));
folio_unlock(folio);
Why do we need to lock the folio? (since this isn't necessarily a
folio)
Can you clarify the "since this isn't necessarily a folio" part ? Do you
mean in the future, when we split "struct page" and "struct folio"?
Right. I need to finish the email that explains where I think we're
going in 2025 ...
Doing an rmap walk on something that won't be a folio is ... sounds odd
(->wrong :) )
Not necessarily! We already do that (since 2022) for DAX (see
6a8e0596f004). rmap lets you find every place that a given range
of a file is mapped into user address spaces; but that file might be a
device file, and so it's not just pagecache but also (in this case)
fb memory, and whatever else device drivers decide to mmap.
Yes, that part I remember.
I thought we would be passing in a page into rmap_wrprotect_file_page(),
and was wondering what we would do to "struct page" that won't be a
folio in there.
Probably, because the "_page" in rmap_wrprotect_file_page() is misleading :)
... should it be "file_range" ? (but we also pass the pfn ... )
--
Cheers,
David / dhildenb
