[PATCH v6 0/3] btrfs: try to allocate larger folios for metadata

[PATCH v6 0/3] btrfs: try to allocate larger folios for metadata

[Qu Wenruo]

[CHANGELOG]
v6:
- Add a new root_mem_cgroup definition for CONFIG_MEMCG=n cases
  So that users of root_mem_cgroup no longer needs to check
  CONFIG_MEMCG.
  This is to fix the compile error for CONFIG_MEMCG=n cases.

- Slight rewording of the 2nd patch

v5:
- Use root memcgroup to attach folios to btree inode filemap
- Only try higher order folio once without NOFAIL nor extra retry

v4:
- Hide the feature behind CONFIG_BTRFS_DEBUG
  So that end users won't be affected (aka, still per-page based
  allocation) meanwhile we can do more testing on this new behavior.

v3:
- Rebased to the latest for-next branch
- Use PAGE_ALLOC_COSTLY_ORDER to determine whether to use __GFP_NOFAIL
- Add a dependency MM patch "mm/page_alloc: unify the warning on NOFAIL
  and high order allocation"
  This allows us to use NOFAIL up to 32K nodesize, and makes sure for
  default 16K nodesize, all metadata would go 16K folios

v2:
- Rebased to handle the change in "btrfs: cache folio size and shift in extent_buffer"

This is the latest update on the attempt to utilize larger folios for
btrfs metadata.

The previous version exposed a reproducibe hang at btrfs/187, where we
hang at filemap_add_folio() around its memcgroup charge code.

Even without the problem, I still believe for btree inode we do not
really need all the memcgroup charge, nor using __GFP_NOFAIL to work
around the possible memcgroup limits.

So in this update, suggested by the memcgroup people from SUSE, there is
a new patch to make btree inode filemap folio attaching to use the root
memcgroup, so that we won't be limited by the memcgroup.

Then for the patch enabling the larger folio, I reverted back to the old
behavior that we only try larger folio once without extra retry, just to
be extra safe.


Qu Wenruo (3):
  memcontrol: define root_mem_cgroup for CONFIG_MEMCG=n cases
  btrfs: always uses root memcgroup for filemap_add_folio()
  btrfs: prefer to allocate larger folio for metadata

 fs/btrfs/extent_io.c       | 112 ++++++++++++++++++++++++++-----------
 include/linux/memcontrol.h |   8 ++-
 2 files changed, 84 insertions(+), 36 deletions(-)

-- 
2.45.2

[PATCH v6 1/3] memcontrol: define root_mem_cgroup for CONFIG_MEMCG=n cases

[Qu Wenruo]

There is an incoming btrfs patchset, which will use @root_mem_cgroup as
the active cgroup to attach metadata folios to its internal btree
inode, so that btrfs can skip the possibly costly charge for the
internal inode which is only accessible by btrfs itself.

However @root_mem_cgroup is not always defined (not defined for
CONFIG_MEMCG=n case), thus all such callers need to do the extra
handling for different CONFIG_MEMCG settings.

So here we add a special macro definition of root_mem_cgroup, making it
to always be NULL.

The advantage of this, other than pulling the pointer definition out,
is that we will avoid wasting global data section space for such
pointer.

Signed-off-by: Qu Wenruo <wqu@suse.com>
---
 include/linux/memcontrol.h | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
index 030d34e9d117..a268585babdc 100644
--- a/include/linux/memcontrol.h
+++ b/include/linux/memcontrol.h
@@ -329,8 +329,6 @@ struct mem_cgroup {
  */
 #define MEMCG_CHARGE_BATCH 64U
 
-extern struct mem_cgroup *root_mem_cgroup;
-
 enum page_memcg_data_flags {
 	/* page->memcg_data is a pointer to an slabobj_ext vector */
 	MEMCG_DATA_OBJEXTS = (1UL << 0),
@@ -346,6 +344,12 @@ enum page_memcg_data_flags {
 
 #define __FIRST_OBJEXT_FLAG	(1UL << 0)
 
+/*
+ * For CONFIG_MEMCG=n case, still define a root_mem_cgroup, but that will
+ * always be NULL and not taking any global data section space.
+ */
+#define root_mem_cgroup		(NULL)
+
 #endif /* CONFIG_MEMCG */
 
 enum objext_flags {
-- 
2.45.2

[PATCH v6 2/3] btrfs: always uses root memcgroup for filemap_add_folio()

[Qu Wenruo]

[BACKGROUND]
The function filemap_add_folio() charges the memory cgroup,
as we assume all page caches are accessible by user space progresses
thus needs the cgroup accounting.

However btrfs is a special case, it has a very large metadata thanks to
its support of data csum (by default it's 4 bytes per 4K data, and can
be as large as 32 bytes per 4K data).
This means btrfs has to go page cache for its metadata pages, to take
advantage of both cache and reclaim ability of filemap.

This has a tiny problem, that all btrfs metadata pages have to go through
the memcgroup charge, even all those metadata pages are not
accessible by the user space, and doing the charging can introduce some
latency if there is a memory limits set.

Btrfs currently uses __GFP_NOFAIL flag as a workaround for this cgroup
charge situation so that metadata pages won't really be limited by
memcgroup.

[ENHANCEMENT]
Instead of relying on __GFP_NOFAIL to avoid charge failure, use root
memory cgroup to attach metadata pages.

With root memory cgroup, we directly skip the charging part, and only
rely on __GFP_NOFAIL for the real memory allocation part.

Suggested-by: Michal Hocko <mhocko@suse.com>
Suggested-by: Vlastimil Babka (SUSE) <vbabka@kernel.org>
Signed-off-by: Qu Wenruo <wqu@suse.com>
---
 fs/btrfs/extent_io.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/fs/btrfs/extent_io.c b/fs/btrfs/extent_io.c
index aa7f8148cd0d..cfeed7673009 100644
--- a/fs/btrfs/extent_io.c
+++ b/fs/btrfs/extent_io.c
@@ -2971,6 +2971,7 @@ static int attach_eb_folio_to_filemap(struct extent_buffer *eb, int i,
 
 	struct btrfs_fs_info *fs_info = eb->fs_info;
 	struct address_space *mapping = fs_info->btree_inode->i_mapping;
+	struct mem_cgroup *old_memcg;
 	const unsigned long index = eb->start >> PAGE_SHIFT;
 	struct folio *existing_folio = NULL;
 	int ret;
@@ -2981,8 +2982,17 @@ static int attach_eb_folio_to_filemap(struct extent_buffer *eb, int i,
 	ASSERT(eb->folios[i]);
 
 retry:
+	/*
+	 * Btree inode is a btrfs internal inode, and not exposed to any
+	 * user.
+	 * Furthermore we do not want any cgroup limits on this inode.
+	 * So we always use root_mem_cgroup as our active memcg when attaching
+	 * the folios.
+	 */
+	old_memcg = set_active_memcg(root_mem_cgroup);
 	ret = filemap_add_folio(mapping, eb->folios[i], index + i,
 				GFP_NOFS | __GFP_NOFAIL);
+	set_active_memcg(old_memcg);
 	if (!ret)
 		goto finish;
 
-- 
2.45.2

[PATCH v6 3/3] btrfs: prefer to allocate larger folio for metadata

[Qu Wenruo]

For btrfs metadata, the high order folios are only utilized when all the
following conditions are met:

- The extent buffer start is aligned to nodesize
  This should be the common case for any btrfs in the last 5 years.

- The nodesize is larger than page size
  Or there is no need to use larger folios at all.

- MM layer can fulfill our folio allocation request

- The larger folio must exactly cover the extent buffer
  No longer no smaller, must be an exact fit.

  This is to make extent buffer accessors much easier.
  They only need to check the first slot in eb->folios[], to determine
  their access unit (need per-page handling or a large folio covering
  the whole eb).

There is another small blockage, filemap APIs can not guarantee the
folio size.
For example, by default we go 16K nodesize on x86_64, meaning a larger
folio we expect would be with order 2 (size 16K).
We don't accept 2 order 1 (size 8K) folios, or we fall back to 4 order 0
(page sized) folios.

So here we go a different workaround, allocate a order 2 folio first,
then attach them to the filemap of metadata.

Thus here comes several results related to the attach attempt of eb
folios:

1) We can attach the pre-allocated eb folio to filemap
   This is the most simple and hot path, we just continue our work
   setting up the extent buffer.

2) There is an existing folio in the filemap

   2.0) Subpage case
        We would reuse the folio no matter what, subpage is doing a
	different way handling folio->private (a bitmap other than a
	pointer to an existing eb).

   2.1) There is already a live extent buffer attached to the filemap
        folio
	This should be more or less hot path, we grab the existing eb
	and free the current one.

   2.2) No live eb.
   2.2.1) The filemap folio is larger than eb folio
          This is a better case, we can reuse the filemap folio, but
	  we need to cleanup all the pre-allocated folios of the
	  new eb before reusing.
	  Later code should take the folio size change into
	  consideration.

   2.2.2) The filemap folio is the same size of eb folio
          We just free the current folio, and reuse the filemap one.
	  No other special handling needed.

   2.2.3) The filemap folio is smaller than eb folio
          This is the most tricky corner case, we can not easily replace
	  the folio in filemap using our eb folio.

	  Thus here we return -EAGAIN, to inform our caller to re-try
	  with order 0 (of course with our larger folio freed).

Otherwise all the needed infrastructure is already here, we only need to
try allocate larger folio as our first try in alloc_eb_folio_array().

For now, the higher order allocation is only a preferable attempt for
debug build, before we had enough test coverage and push it to end
users.

Signed-off-by: Qu Wenruo <wqu@suse.com>
---
 fs/btrfs/extent_io.c | 102 ++++++++++++++++++++++++++++---------------
 1 file changed, 68 insertions(+), 34 deletions(-)

diff --git a/fs/btrfs/extent_io.c b/fs/btrfs/extent_io.c
index cfeed7673009..d7824644d593 100644
--- a/fs/btrfs/extent_io.c
+++ b/fs/btrfs/extent_io.c
@@ -719,12 +719,28 @@ int btrfs_alloc_page_array(unsigned int nr_pages, struct page **page_array,
  *
  * For now, the folios populated are always in order 0 (aka, single page).
  */
-static int alloc_eb_folio_array(struct extent_buffer *eb, bool nofail)
+static int alloc_eb_folio_array(struct extent_buffer *eb, int order,
+				bool nofail)
 {
 	struct page *page_array[INLINE_EXTENT_BUFFER_PAGES] = { 0 };
 	int num_pages = num_extent_pages(eb);
 	int ret;
 
+	if (order) {
+		gfp_t gfp;
+
+		if (order > 0)
+			gfp = GFP_NOFS | __GFP_NORETRY | __GFP_NOWARN;
+		else
+			gfp = nofail ? (GFP_NOFS | __GFP_NOFAIL) : GFP_NOFS;
+		eb->folios[0] = folio_alloc(gfp, order);
+		if (likely(eb->folios[0])) {
+			eb->folio_size = folio_size(eb->folios[0]);
+			eb->folio_shift = folio_shift(eb->folios[0]);
+			return 0;
+		}
+		/* Fallback to 0 order (single page) allocation. */
+	}
 	ret = btrfs_alloc_page_array(num_pages, page_array, nofail);
 	if (ret < 0)
 		return ret;
@@ -2707,7 +2723,7 @@ struct extent_buffer *btrfs_clone_extent_buffer(const struct extent_buffer *src)
 	 */
 	set_bit(EXTENT_BUFFER_UNMAPPED, &new->bflags);
 
-	ret = alloc_eb_folio_array(new, false);
+	ret = alloc_eb_folio_array(new, 0, false);
 	if (ret) {
 		btrfs_release_extent_buffer(new);
 		return NULL;
@@ -2740,7 +2756,7 @@ struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
 	if (!eb)
 		return NULL;
 
-	ret = alloc_eb_folio_array(eb, false);
+	ret = alloc_eb_folio_array(eb, 0, false);
 	if (ret)
 		goto err;
 
@@ -2955,6 +2971,14 @@ static int check_eb_alignment(struct btrfs_fs_info *fs_info, u64 start)
 	return 0;
 }
 
+static void free_all_eb_folios(struct extent_buffer *eb)
+{
+	for (int i = 0; i < INLINE_EXTENT_BUFFER_PAGES; i++) {
+		if (eb->folios[i])
+			folio_put(eb->folios[i]);
+		eb->folios[i] = NULL;
+	}
+}
 
 /*
  * Return 0 if eb->folios[i] is attached to btree inode successfully.
@@ -2974,6 +2998,7 @@ static int attach_eb_folio_to_filemap(struct extent_buffer *eb, int i,
 	struct mem_cgroup *old_memcg;
 	const unsigned long index = eb->start >> PAGE_SHIFT;
 	struct folio *existing_folio = NULL;
+	const int eb_order = folio_order(eb->folios[0]);
 	int ret;
 
 	ASSERT(found_eb_ret);
@@ -3003,15 +3028,6 @@ static int attach_eb_folio_to_filemap(struct extent_buffer *eb, int i,
 		goto retry;
 	}
 
-	/* For now, we should only have single-page folios for btree inode. */
-	ASSERT(folio_nr_pages(existing_folio) == 1);
-
-	if (folio_size(existing_folio) != eb->folio_size) {
-		folio_unlock(existing_folio);
-		folio_put(existing_folio);
-		return -EAGAIN;
-	}
-
 finish:
 	spin_lock(&mapping->i_private_lock);
 	if (existing_folio && fs_info->nodesize < PAGE_SIZE) {
@@ -3020,6 +3036,7 @@ static int attach_eb_folio_to_filemap(struct extent_buffer *eb, int i,
 		eb->folios[i] = existing_folio;
 	} else if (existing_folio) {
 		struct extent_buffer *existing_eb;
+		int existing_order = folio_order(existing_folio);
 
 		existing_eb = grab_extent_buffer(fs_info,
 						 folio_page(existing_folio, 0));
@@ -3031,9 +3048,34 @@ static int attach_eb_folio_to_filemap(struct extent_buffer *eb, int i,
 			folio_put(existing_folio);
 			return 1;
 		}
-		/* The extent buffer no longer exists, we can reuse the folio. */
-		__free_page(folio_page(eb->folios[i], 0));
-		eb->folios[i] = existing_folio;
+		if (existing_order > eb_order) {
+			/*
+			 * The existing one has higher order, we need to drop
+			 * all eb folios before resuing it.
+			 * And this should only happen for the first folio.
+			 */
+			ASSERT(i == 0);
+			free_all_eb_folios(eb);
+			eb->folios[i] = existing_folio;
+		} else if (existing_order == eb_order) {
+			/*
+			 * Can safely reuse the filemap folio, just
+			 * release the eb one.
+			 */
+			folio_put(eb->folios[i]);
+			eb->folios[i] = existing_folio;
+		} else {
+			/*
+			 * The existing one has lower order.
+			 *
+			 * Just retry and fallback to order 0.
+			 */
+			ASSERT(i == 0);
+			folio_unlock(existing_folio);
+			folio_put(existing_folio);
+			spin_unlock(&mapping->i_private_lock);
+			return -EAGAIN;
+		}
 	}
 	eb->folio_size = folio_size(eb->folios[i]);
 	eb->folio_shift = folio_shift(eb->folios[i]);
@@ -3066,6 +3108,7 @@ struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
 	u64 lockdep_owner = owner_root;
 	bool page_contig = true;
 	int uptodate = 1;
+	int order = 0;
 	int ret;
 
 	if (check_eb_alignment(fs_info, start))
@@ -3082,6 +3125,10 @@ struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
 		btrfs_warn_32bit_limit(fs_info);
 #endif
 
+	if (IS_ENABLED(CONFIG_BTRFS_DEBUG) && fs_info->nodesize > PAGE_SIZE &&
+	    IS_ALIGNED(start, fs_info->nodesize))
+		order = ilog2(fs_info->nodesize >> PAGE_SHIFT);
+
 	eb = find_extent_buffer(fs_info, start);
 	if (eb)
 		return eb;
@@ -3116,7 +3163,7 @@ struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
 
 reallocate:
 	/* Allocate all pages first. */
-	ret = alloc_eb_folio_array(eb, true);
+	ret = alloc_eb_folio_array(eb, order, true);
 	if (ret < 0) {
 		btrfs_free_subpage(prealloc);
 		goto out;
@@ -3134,26 +3181,12 @@ struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
 		}
 
 		/*
-		 * TODO: Special handling for a corner case where the order of
-		 * folios mismatch between the new eb and filemap.
-		 *
-		 * This happens when:
-		 *
-		 * - the new eb is using higher order folio
-		 *
-		 * - the filemap is still using 0-order folios for the range
-		 *   This can happen at the previous eb allocation, and we don't
-		 *   have higher order folio for the call.
-		 *
-		 * - the existing eb has already been freed
-		 *
-		 * In this case, we have to free the existing folios first, and
-		 * re-allocate using the same order.
-		 * Thankfully this is not going to happen yet, as we're still
-		 * using 0-order folios.
+		 * Got a corner case where the existing folio is lower order,
+		 * fallback to 0 order and retry.
 		 */
 		if (unlikely(ret == -EAGAIN)) {
-			ASSERT(0);
+			order = 0;
+			free_all_eb_folios(eb);
 			goto reallocate;
 		}
 		attached++;
@@ -3164,6 +3197,7 @@ struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
 		 * and free the allocated page.
 		 */
 		folio = eb->folios[i];
+		num_folios = num_extent_folios(eb);
 		WARN_ON(btrfs_folio_test_dirty(fs_info, folio, eb->start, eb->len));
 
 		/*
-- 
2.45.2

Re: [PATCH v6 1/3] memcontrol: define root_mem_cgroup for CONFIG_MEMCG=n cases

[Vlastimil Babka]

On 7/19/24 11:16 AM, Qu Wenruo wrote:
> There is an incoming btrfs patchset, which will use @root_mem_cgroup as
> the active cgroup to attach metadata folios to its internal btree
> inode, so that btrfs can skip the possibly costly charge for the
> internal inode which is only accessible by btrfs itself.
> 
> However @root_mem_cgroup is not always defined (not defined for
> CONFIG_MEMCG=n case), thus all such callers need to do the extra
> handling for different CONFIG_MEMCG settings.
> 
> So here we add a special macro definition of root_mem_cgroup, making it
> to always be NULL.
> 
> The advantage of this, other than pulling the pointer definition out,
> is that we will avoid wasting global data section space for such
> pointer.
> 
> Signed-off-by: Qu Wenruo <wqu@suse.com>
> ---
>  include/linux/memcontrol.h | 8 ++++++--
>  1 file changed, 6 insertions(+), 2 deletions(-)
> 
> diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
> index 030d34e9d117..a268585babdc 100644
> --- a/include/linux/memcontrol.h
> +++ b/include/linux/memcontrol.h
> @@ -329,8 +329,6 @@ struct mem_cgroup {
>   */
>  #define MEMCG_CHARGE_BATCH 64U
>  
> -extern struct mem_cgroup *root_mem_cgroup;
> -

This breaks the CONFIG_MEMCG=y case?

>  enum page_memcg_data_flags {
>  	/* page->memcg_data is a pointer to an slabobj_ext vector */
>  	MEMCG_DATA_OBJEXTS = (1UL << 0),
> @@ -346,6 +344,12 @@ enum page_memcg_data_flags {
>  
>  #define __FIRST_OBJEXT_FLAG	(1UL << 0)
>  
> +/*
> + * For CONFIG_MEMCG=n case, still define a root_mem_cgroup, but that will
> + * always be NULL and not taking any global data section space.
> + */
> +#define root_mem_cgroup		(NULL)
> +
>  #endif /* CONFIG_MEMCG */
>  
>  enum objext_flags {