We're currently trying to remove page->mapping from the entire kernel.
This has me interested in fb_defio and since I made such a mess of it
with commits ccf953d8f3d6 / 0b78f8bcf495, I'd like to discuss what to
do before diving in.
Folios continue to have a mapping. So we can effectively do
page_folio(page)->mapping (today that's calling compound_head() to get
to the head page; eventually it's a separate allocation).
But now I look at commit 56c134f7f1b5, I'm a little scared.
Apparently pages are being allocated from shmem and being mapped by
fb_deferred_io_fault()? This line:
page->mapping = vmf->vma->vm_file->f_mapping;
initially appears harmless for shmem files (because that's presumably
a noop), but it's only a noop for head pages. If shmem allocates a
compound page (ie a 2MB THP today), you'll overlay some information
stored in the second and third pages; looks like _entire_mapcount
and _deferred_list.prev (but we do shift those fields around without
regard to what the fb_defio driver is doing). Even if you've disabled
THP today, setting page->mapping to NULL in fb_deferred_io_lastclose()
for a shmem page is a really bad idea.
I'd like to avoid fb_defio playing with page->mapping at all.
As I understand it, the only reason to set page->mapping is so that
page_mkclean() works. But there are all kinds of assumptions in
page_mkclean() (now essentially folio_mkclean()) that we're dealing with
file-backed or anonymous memory. I wonder if we might be better off
calling pfn_mkclean_range() for each VMA which maps these allocations?
You'd have to keep track of each VMA yourself (I think?) but it would
avoid messing with page->mapping.
Anyway, I don't know enough about DRM. There might be something
unutterably obvious we could do to fix this.
