Re: [PATCH v5 05/25] fs/dax: Create a common implementation to break DAX layouts

Alistair Popple wrote:

On Wed, Jan 08, 2025 at 04:14:20PM -0800, Dan Williams wrote:

quoted

Alistair Popple wrote:

quoted

Prior to freeing a block file systems supporting FS DAX must check
that the associated pages are both unmapped from user-space and not
undergoing DMA or other access from eg. get_user_pages(). This is
achieved by unmapping the file range and scanning the FS DAX
page-cache to see if any pages within the mapping have an elevated
refcount.

This is done using two functions - dax_layout_busy_page_range() which
returns a page to wait for the refcount to become idle on. Rather than
open-code this introduce a common implementation to both unmap and
wait for the page to become idle.

Signed-off-by: Alistair Popple <apopple@nvidia.com>

---

Changes for v5:

 - Don't wait for idle pages on non-DAX mappings

Changes for v4:

 - Fixed some build breakage due to missing symbol exports reported by
   John Hubbard (thanks!).
---
 fs/dax.c            | 33 +++++++++++++++++++++++++++++++++
 fs/ext4/inode.c     | 10 +---------
 fs/fuse/dax.c       | 29 +++++------------------------
 fs/xfs/xfs_inode.c  | 23 +++++------------------
 fs/xfs/xfs_inode.h  |  2 +-
 include/linux/dax.h | 21 +++++++++++++++++++++
 mm/madvise.c        |  8 ++++----
 7 files changed, 70 insertions(+), 56 deletions(-)

diff --git a/fs/dax.c b/fs/dax.c
index d010c10..9c3bd07 100644
--- a/fs/dax.c
+++ b/fs/dax.c

@@ -845,6 +845,39 @@ int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
 	return ret;
 }
 
+static int wait_page_idle(struct page *page,
+			void (cb)(struct inode *),
+			struct inode *inode)
+{
+	return ___wait_var_event(page, page_ref_count(page) == 1,
+				TASK_INTERRUPTIBLE, 0, 0, cb(inode));
+}
+
+/*
+ * Unmaps the inode and waits for any DMA to complete prior to deleting the
+ * DAX mapping entries for the range.
+ */
+int dax_break_mapping(struct inode *inode, loff_t start, loff_t end,
+		void (cb)(struct inode *))
+{
+	struct page *page;
+	int error;
+
+	if (!dax_mapping(inode->i_mapping))
+		return 0;
+
+	do {
+		page = dax_layout_busy_page_range(inode->i_mapping, start, end);
+		if (!page)
+			break;
+
+		error = wait_page_idle(page, cb, inode);

This implementations removes logic around @retry found in the XFS and
FUSE implementations, I think that is a mistake, and EXT4 has
apparently been broken in this regard.

I think both implementations are equivalent though, just that the XFS/FUSE ones are
spread across two functions with the retry happening in the outer function
whilst the EXT4 implementation is implemented in a single function with a do/
while loop.

Both exit early if dax_layout_busy_page() doesn't find a DMA-busy page, and
both call dax_layout_busy_page() a second time after waiting on a page to become
idle. So I don't think anything is broken here, unless I've missed something.

Nope, you're right. I crossed my eyes flipping between FUSE/XFS and EXT4.

quoted

wait_page_idle() returns after @page is idle, but that does not mean
@inode is DMA idle. After one found page from
dax_layout_busy_page_range() is waited upon a new call to
dax_break_mapping() needs to made to check if another DMA started, or if
there were originally more pages active.

quoted

+	} while (error == 0);
+
+	return error;

Surprised that the compiler does not warn about an uninitialized
variable here?

So am I. Turns out this is built with -Wno-maybe-uninitialized and it's not
certain error is used uninitialized because we may bail early if this is not a
dax_mapping. So it's only maybe used uninitialized which isn't warned about.

Looks like smatch just caught it.