On 14 Oct 2020, at 9:49, Matthew Wilcox wrote:
Our handling of PagePrivate, page->private and PagePrivate2 is a giant
mess. Let's recap.
Filesystems which use bufferheads (ie most of them) set page->private
to point to a buffer_head, set the PagePrivate bit and increment the
refcount on the page.
The vmscan pageout code (*) needs to know whether a page is freeable:
if (!is_page_cache_freeable(page))
return PAGE_KEEP;
... where is_page_cache_freeable() contains ...
return page_count(page) - page_has_private(page) == 1 +
page_cache_pins;
That's a little inscrutable, but the important thing is that if
page_has_private() is true, then the page's reference count is
supposed
to be one higher than it would be otherwise. And that makes sense
given
how "having bufferheads" means "page refcount ges incremented".
But page_has_private() doesn't actually mean "PagePrivate is set".
It means "PagePrivate or PagePrivate2 is set". And I don't understand
how filesystems are supposed to keep that straight -- if we're setting
PagePrivate2, and PagePrivate is clear, increment the refcount?
If we're clearing PagePrivate, decrement the refcount if PagePrivate2
is also clear?
At least for btrfs, only PagePrivate elevates the refcount on the page.
PagePrivate2 means:
This page has been properly setup for COW’d IO, and it went through
the normal path of page_mkwrite() or file_write() instead of being
silently dirtied by a deep corner of the MM.
We introduced attach_page_private() and detach_page_private() earlier
this year to help filesystems get the refcount right. But we still
have a few filesystems using PagePrivate themselves (afs, btrfs, ceph,
crypto, erofs, f2fs, jfs, nfs, orangefs & ubifs) and I'm not convinced
they're all getting it right.
Here's a bug I happened on while looking into this:
if (page_has_private(page))
attach_page_private(newpage,
detach_page_private(page));
if (PagePrivate2(page)) {
ClearPagePrivate2(page);
SetPagePrivate2(newpage);
}
The aggravating thing is that this doesn't even look like a bug.
You have to be in the kind of mood where you're thinking "What if page
has Private2 set and Private clear?" and the answer is that newpage
ends up with PagePrivate set, but page->private set to NULL.
Do you mean PagePrivate2 set but page->private NULL? Btrfs should only
hage PagePrivate2 set on pages that are formally in our writeback state
machine, so it’ll get cleared as we unwind through normal IO or
truncate etc. For data pages, btrfs page->private is simply set to 1 so
the MM will kindly call releasepage for us.
And I
don't know whether this is a situation that can ever happen with
btrfs,
but we shouldn't have code like this lying around in the tree because
it _looks_ right and somebody else might copy it.
So what shold we do about all this? First, I want to make the code
snippet above correct, because it looks right. So page_has_private()
needs to test just PagePrivate and not PagePrivate2. Now we need a
new function to call to determine whether the filesystem needs its
invalidatepage callback invoked. Not sure what that should be named.
I haven’t checked all the page_has_private() callers, but maybe
page_has_private() should stay the same and add page_private_count() for
times where we need to get out our fingers and toes for the refcount
math.
I think I also want to rename PG_private_2 to PG_owner_priv_2.
There's a clear relationship between PG_private and page->private.
There is no relationship between PG_private_2 and page->private, so
it's
a misleading name. Or maybe it should just be PG_fscache and btrfs
can
find some other way to mark the pages?
Btrfs should be able to flip bits in page->private to cover our current
usage of PG_private_2. If we ever attach something real to
page->private, we can flip bits in that instead. It’s kinda messy
though and we’d have to change attach_page_private a little to reflect
its new life as a bit setting machine.
Also ... do we really need to increment the page refcount if we have
PagePrivate set? I'm not awfully familiar with the buffercache -- is
it possible we end up in a situation where a buffer, perhaps under
I/O,
has the last reference to a struct page? It seems like that reference
is
always put from drop_buffers() which is called from
try_to_free_buffers()
which is always called by someone who has a reference to a struct page
that they got from the pagecache. So what is this reference count
for?
I’m not sure what we gain by avoiding the refcount bump? Many
filesystems use the pattern of: “put something in page->private, free
that thing in releasepage.” Without the refcount bump it feels like
we’d have more magic to avoid freeing the page without leaking things
in page->private. I think the extra ref lets the FS crowd keep our
noses out of the MM more often, so it seems like a net positive to me.
(*) Also THP split and page migration. But maybe you don't care about
those things ... you definitely care about pageout!
-chris
