Re: [PATCH v6 06/26] fs/dax: Always remove DAX page-cache entries when breaking layouts

[Dan Williams <dan.j.williams@intel.com>]

Alistair Popple wrote:
> Prior to any truncation operations file systems call
> dax_break_mapping() to ensure pages in the range are not under going
> DMA. Later DAX page-cache entries will be removed by
> truncate_folio_batch_exceptionals() in the generic page-cache code.
> 
> However this makes it possible for folios to be removed from the
> page-cache even though they are still DMA busy if the file-system
> hasn't called dax_break_mapping(). It also means they can never be
> waited on in future because FS DAX will lose track of them once the
> page-cache entry has been deleted.
> 
> Instead it is better to delete the FS DAX entry when the file-system
> calls dax_break_mapping() as part of it's truncate operation. This
> ensures only idle pages can be removed from the FS DAX page-cache and
> makes it easy to detect if a file-system hasn't called
> dax_break_mapping() prior to a truncate operation.
> 
> Signed-off-by: Alistair Popple <apopple@nvidia.com>
> 
> ---
> 
> Ideally I think we would move the whole wait-for-idle logic directly
> into the truncate paths. However this is difficult for a few
> reasons. Each filesystem needs it's own wait callback, although a new
> address space operation could address that. More problematic is that
> the wait-for-idle can fail as the wait is TASK_INTERRUPTIBLE, but none
> of the generic truncate paths allow for failure.
> 
> So it ends up being easier to continue to let file systems call this
> and check that they behave as expected.
> ---
>  fs/dax.c            | 33 +++++++++++++++++++++++++++++++++
>  fs/xfs/xfs_inode.c  |  6 ++++++
>  include/linux/dax.h |  2 ++
>  mm/truncate.c       | 16 +++++++++++++++-
>  4 files changed, 56 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/dax.c b/fs/dax.c
> index 9c3bd07..7008a73 100644
> --- a/fs/dax.c
> +++ b/fs/dax.c
> @@ -845,6 +845,36 @@ int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
>  	return ret;
>  }
>  
> +void dax_delete_mapping_range(struct address_space *mapping,
> +				loff_t start, loff_t end)
> +{
> +	void *entry;
> +	pgoff_t start_idx = start >> PAGE_SHIFT;
> +	pgoff_t end_idx;
> +	XA_STATE(xas, &mapping->i_pages, start_idx);
> +
> +	/* If end == LLONG_MAX, all pages from start to till end of file */
> +	if (end == LLONG_MAX)
> +		end_idx = ULONG_MAX;
> +	else
> +		end_idx = end >> PAGE_SHIFT;
> +
> +	xas_lock_irq(&xas);
> +	xas_for_each(&xas, entry, end_idx) {
> +		if (!xa_is_value(entry))
> +			continue;
> +		entry = wait_entry_unlocked_exclusive(&xas, entry);
> +		if (!entry)
> +			continue;
> +		dax_disassociate_entry(entry, mapping, true);
> +		xas_store(&xas, NULL);
> +		mapping->nrpages -= 1UL << dax_entry_order(entry);
> +		put_unlocked_entry(&xas, entry, WAKE_ALL);
> +	}
> +	xas_unlock_irq(&xas);
> +}
> +EXPORT_SYMBOL_GPL(dax_delete_mapping_range);
> +
>  static int wait_page_idle(struct page *page,
>  			void (cb)(struct inode *),
>  			struct inode *inode)
> @@ -874,6 +904,9 @@ int dax_break_mapping(struct inode *inode, loff_t start, loff_t end,
>  		error = wait_page_idle(page, cb, inode);
>  	} while (error == 0);
>  
> +	if (!page)
> +		dax_delete_mapping_range(inode->i_mapping, start, end);
> +

Just reinforcing the rename comment on the last patch...

I think this is an example where the
s/dax_break_mapping/dax_break_layout/ rename helps disambiguate what is
related to mapping cleanup and what is related to mapping cleanup as
dax_break_layout calls dax_delete_mapping.

>  	return error;
>  }
>  EXPORT_SYMBOL_GPL(dax_break_mapping);
> diff --git a/fs/xfs/xfs_inode.c b/fs/xfs/xfs_inode.c
> index 295730a..4410b42 100644
> --- a/fs/xfs/xfs_inode.c
> +++ b/fs/xfs/xfs_inode.c
> @@ -2746,6 +2746,12 @@ xfs_mmaplock_two_inodes_and_break_dax_layout(
>  		goto again;
>  	}
>  
> +	/*
> +	 * Normally xfs_break_dax_layouts() would delete the mapping entries as well so
> +	 * do that here.
> +	 */
> +	dax_delete_mapping_range(VFS_I(ip2)->i_mapping, 0, LLONG_MAX);
> +

I think it is unfortunate that dax_break_mapping is so close to being
useful for this case... how about this incremental cleanup?

diff --git a/fs/dax.c b/fs/dax.c
index facddd6c6bbb..1fa5521e5a2e 100644
--- a/fs/dax.c
+++ b/fs/dax.c
@@ -942,12 +942,15 @@ static void wait_page_idle_uninterruptible(struct page *page,
 /*
  * Unmaps the inode and waits for any DMA to complete prior to deleting the
  * DAX mapping entries for the range.
+ *
+ * For NOWAIT behavior, pass @cb as NULL to early-exit on first found
+ * busy page
  */
 int dax_break_mapping(struct inode *inode, loff_t start, loff_t end,
 		void (cb)(struct inode *))
 {
 	struct page *page;
-	int error;
+	int error = 0;
 
 	if (!dax_mapping(inode->i_mapping))
 		return 0;
@@ -956,6 +959,10 @@ int dax_break_mapping(struct inode *inode, loff_t start, loff_t end,
 		page = dax_layout_busy_page_range(inode->i_mapping, start, end);
 		if (!page)
 			break;
+		if (!cb) {
+			error = -ERESTARTSYS;
+			break;
+		}
 
 		error = wait_page_idle(page, cb, inode);
 	} while (error == 0);
diff --git a/fs/xfs/xfs_inode.c b/fs/xfs/xfs_inode.c
index 7bfb4eb387c6..0988a9088259 100644
--- a/fs/xfs/xfs_inode.c
+++ b/fs/xfs/xfs_inode.c
@@ -2739,19 +2739,13 @@ xfs_mmaplock_two_inodes_and_break_dax_layout(
 	 * need to unlock & lock the XFS_MMAPLOCK_EXCL which is not suitable
 	 * for this nested lock case.
 	 */
-	page = dax_layout_busy_page(VFS_I(ip2)->i_mapping);
-	if (page && page_ref_count(page) != 0) {
+	error = dax_break_layout(VFS_I(ip2), 0, -1, NULL);
+	if (error) {
 		xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
 		xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
 		goto again;
 	}
 
-	/*
-	 * Normally xfs_break_dax_layouts() would delete the mapping entries as well so
-	 * do that here.
-	 */
-	dax_delete_mapping_range(VFS_I(ip2)->i_mapping, 0, LLONG_MAX);
-
 	return 0;
 }
 

This also addresses Darrick's feedback around introducing
dax_page_in_use() which xfs does not really care about, only that no
more pages are busy.

>  	return 0;
>  }
>  
> diff --git a/include/linux/dax.h b/include/linux/dax.h
> index f6583d3..ef9e02c 100644
> --- a/include/linux/dax.h
> +++ b/include/linux/dax.h
> @@ -263,6 +263,8 @@ vm_fault_t dax_iomap_fault(struct vm_fault *vmf, unsigned int order,
>  vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf,
>  		unsigned int order, pfn_t pfn);
>  int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index);
> +void dax_delete_mapping_range(struct address_space *mapping,
> +				loff_t start, loff_t end);
>  int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
>  				      pgoff_t index);
>  int __must_check dax_break_mapping(struct inode *inode, loff_t start,
> diff --git a/mm/truncate.c b/mm/truncate.c
> index 7c304d2..b7f51a6 100644
> --- a/mm/truncate.c
> +++ b/mm/truncate.c
> @@ -78,8 +78,22 @@ static void truncate_folio_batch_exceptionals(struct address_space *mapping,
>  
>  	if (dax_mapping(mapping)) {
>  		for (i = j; i < nr; i++) {
> -			if (xa_is_value(fbatch->folios[i]))
> +			if (xa_is_value(fbatch->folios[i])) {
> +				/*
> +				 * File systems should already have called
> +				 * dax_break_mapping_entry() to remove all DAX
> +				 * entries while holding a lock to prevent
> +				 * establishing new entries. Therefore we
> +				 * shouldn't find any here.
> +				 */
> +				WARN_ON_ONCE(1);
> +
> +				/*
> +				 * Delete the mapping so truncate_pagecache()
> +				 * doesn't loop forever.
> +				 */
>  				dax_delete_mapping_entry(mapping, indices[i]);
> +			}

Looks good.

With the above additional fixup you can add:

Reviewed-by: Dan Williams <dan.j.williams@intel.com>