[PATCH v3 0/4] btrfs: data block group size classes

[PATCH v3 0/4] btrfs: data block group size classes

[Boris Burkov]

This patch set introduces the notion of size classes to the block group
allocator for data block groups. This is specifically useful because the
first fit allocator tends to perform poorly when large extents free up
in older block groups and small writes suddenly shift there. Generally,
it should lead to slightly more predictable allocator behavior as the
gaps left by frees will be used by allocations of a similar size.

Details about the changes and performance testing are in the individual
commit messages.

The last two patches constitute the business of the change. One adds the
size classes and the other handles the fact that we don't want to
persist the size class, so we don't know it when we first load a block
group.
---
v3:
- fix double newline in extent-tree.h
- fix ctree.h include in extent-tree.h refactor
v2:
- removed 1G falloc extents patch
- rebased tracepoints patches onto significant header file refactor

Boris Burkov (4):
  btrfs: use ffe_ctl in btrfs allocator tracepoints
  btrfs: add more ffe tracepoints
  btrfs: introduce size class to block group allocator
  btrfs: load block group size class when caching

 fs/btrfs/block-group.c       | 234 ++++++++++++++++++++++++++++++++---
 fs/btrfs/block-group.h       |  15 ++-
 fs/btrfs/extent-tree.c       | 167 +++++++------------------
 fs/btrfs/extent-tree.h       |  81 ++++++++++++
 fs/btrfs/super.c             |   1 +
 include/trace/events/btrfs.h | 128 +++++++++++++++----
 6 files changed, 467 insertions(+), 159 deletions(-)

-- 
2.38.1

[PATCH v3 1/4] btrfs: use ffe_ctl in btrfs allocator tracepoints

[Boris Burkov]

The allocator tracepoints currently have a pile of values from ffe_ctl.
In modifying the allocator and adding more tracepoints, I found myself
adding to the already long argument list of the tracepoints. It makes it
a lot simpler to just send in the ffe_ctl itself.

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Boris Burkov <boris@bur.io>
---
 fs/btrfs/extent-tree.c       | 92 +++---------------------------------
 fs/btrfs/extent-tree.h       | 75 +++++++++++++++++++++++++++++
 fs/btrfs/super.c             |  1 +
 include/trace/events/btrfs.h | 41 ++++++++--------
 4 files changed, 103 insertions(+), 106 deletions(-)

diff --git a/fs/btrfs/extent-tree.c b/fs/btrfs/extent-tree.c
index 892d78c1853c..a53f90677fbc 100644
--- a/fs/btrfs/extent-tree.c
+++ b/fs/btrfs/extent-tree.c
@@ -16,7 +16,8 @@
 #include <linux/percpu_counter.h>
 #include <linux/lockdep.h>
 #include <linux/crc32c.h>
-#include "misc.h"
+#include "ctree.h"
+#include "extent-tree.h"
 #include "tree-log.h"
 #include "disk-io.h"
 #include "print-tree.h"
@@ -31,7 +32,6 @@
 #include "space-info.h"
 #include "block-rsv.h"
 #include "delalloc-space.h"
-#include "block-group.h"
 #include "discard.h"
 #include "rcu-string.h"
 #include "zoned.h"
@@ -3450,81 +3450,6 @@ btrfs_release_block_group(struct btrfs_block_group *cache,
 	btrfs_put_block_group(cache);
 }
 
-enum btrfs_extent_allocation_policy {
-	BTRFS_EXTENT_ALLOC_CLUSTERED,
-	BTRFS_EXTENT_ALLOC_ZONED,
-};
-
-/*
- * Structure used internally for find_free_extent() function.  Wraps needed
- * parameters.
- */
-struct find_free_extent_ctl {
-	/* Basic allocation info */
-	u64 ram_bytes;
-	u64 num_bytes;
-	u64 min_alloc_size;
-	u64 empty_size;
-	u64 flags;
-	int delalloc;
-
-	/* Where to start the search inside the bg */
-	u64 search_start;
-
-	/* For clustered allocation */
-	u64 empty_cluster;
-	struct btrfs_free_cluster *last_ptr;
-	bool use_cluster;
-
-	bool have_caching_bg;
-	bool orig_have_caching_bg;
-
-	/* Allocation is called for tree-log */
-	bool for_treelog;
-
-	/* Allocation is called for data relocation */
-	bool for_data_reloc;
-
-	/* RAID index, converted from flags */
-	int index;
-
-	/*
-	 * Current loop number, check find_free_extent_update_loop() for details
-	 */
-	int loop;
-
-	/*
-	 * Whether we're refilling a cluster, if true we need to re-search
-	 * current block group but don't try to refill the cluster again.
-	 */
-	bool retry_clustered;
-
-	/*
-	 * Whether we're updating free space cache, if true we need to re-search
-	 * current block group but don't try updating free space cache again.
-	 */
-	bool retry_unclustered;
-
-	/* If current block group is cached */
-	int cached;
-
-	/* Max contiguous hole found */
-	u64 max_extent_size;
-
-	/* Total free space from free space cache, not always contiguous */
-	u64 total_free_space;
-
-	/* Found result */
-	u64 found_offset;
-
-	/* Hint where to start looking for an empty space */
-	u64 hint_byte;
-
-	/* Allocation policy */
-	enum btrfs_extent_allocation_policy policy;
-};
-
-
 /*
  * Helper function for find_free_extent().
  *
@@ -3556,8 +3481,7 @@ static int find_free_extent_clustered(struct btrfs_block_group *bg,
 	if (offset) {
 		/* We have a block, we're done */
 		spin_unlock(&last_ptr->refill_lock);
-		trace_btrfs_reserve_extent_cluster(cluster_bg,
-				ffe_ctl->search_start, ffe_ctl->num_bytes);
+		trace_btrfs_reserve_extent_cluster(cluster_bg, ffe_ctl);
 		*cluster_bg_ret = cluster_bg;
 		ffe_ctl->found_offset = offset;
 		return 0;
@@ -3607,10 +3531,8 @@ static int find_free_extent_clustered(struct btrfs_block_group *bg,
 		if (offset) {
 			/* We found one, proceed */
 			spin_unlock(&last_ptr->refill_lock);
-			trace_btrfs_reserve_extent_cluster(bg,
-					ffe_ctl->search_start,
-					ffe_ctl->num_bytes);
 			ffe_ctl->found_offset = offset;
+			trace_btrfs_reserve_extent_cluster(bg, ffe_ctl);
 			return 0;
 		}
 	} else if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
@@ -4293,8 +4215,7 @@ static noinline int find_free_extent(struct btrfs_root *root,
 	ins->objectid = 0;
 	ins->offset = 0;
 
-	trace_find_free_extent(root, ffe_ctl->num_bytes, ffe_ctl->empty_size,
-			       ffe_ctl->flags);
+	trace_find_free_extent(root, ffe_ctl);
 
 	space_info = btrfs_find_space_info(fs_info, ffe_ctl->flags);
 	if (!space_info) {
@@ -4465,8 +4386,7 @@ static noinline int find_free_extent(struct btrfs_root *root,
 		ins->objectid = ffe_ctl->search_start;
 		ins->offset = ffe_ctl->num_bytes;
 
-		trace_btrfs_reserve_extent(block_group, ffe_ctl->search_start,
-					   ffe_ctl->num_bytes);
+		trace_btrfs_reserve_extent(block_group, ffe_ctl);
 		btrfs_release_block_group(block_group, ffe_ctl->delalloc);
 		break;
 loop:
diff --git a/fs/btrfs/extent-tree.h b/fs/btrfs/extent-tree.h
index ae5425253603..64fa8ad7914a 100644
--- a/fs/btrfs/extent-tree.h
+++ b/fs/btrfs/extent-tree.h
@@ -3,6 +3,81 @@
 #ifndef BTRFS_EXTENT_TREE_H
 #define BTRFS_EXTENT_TREE_H
 
+#include "misc.h"
+#include "block-group.h"
+
+struct btrfs_free_cluster;
+
+enum btrfs_extent_allocation_policy {
+	BTRFS_EXTENT_ALLOC_CLUSTERED,
+	BTRFS_EXTENT_ALLOC_ZONED,
+};
+
+struct find_free_extent_ctl {
+	/* Basic allocation info */
+	u64 ram_bytes;
+	u64 num_bytes;
+	u64 min_alloc_size;
+	u64 empty_size;
+	u64 flags;
+	int delalloc;
+
+	/* Where to start the search inside the bg */
+	u64 search_start;
+
+	/* For clustered allocation */
+	u64 empty_cluster;
+	struct btrfs_free_cluster *last_ptr;
+	bool use_cluster;
+
+	bool have_caching_bg;
+	bool orig_have_caching_bg;
+
+	/* Allocation is called for tree-log */
+	bool for_treelog;
+
+	/* Allocation is called for data relocation */
+	bool for_data_reloc;
+
+	/* RAID index, converted from flags */
+	int index;
+
+	/*
+	 * Current loop number, check find_free_extent_update_loop() for details
+	 */
+	int loop;
+
+	/*
+	 * Whether we're refilling a cluster, if true we need to re-search
+	 * current block group but don't try to refill the cluster again.
+	 */
+	bool retry_clustered;
+
+	/*
+	 * Whether we're updating free space cache, if true we need to re-search
+	 * current block group but don't try updating free space cache again.
+	 */
+	bool retry_unclustered;
+
+	/* If current block group is cached */
+	int cached;
+
+	/* Max contiguous hole found */
+	u64 max_extent_size;
+
+	/* Total free space from free space cache, not always contiguous */
+	u64 total_free_space;
+
+	/* Found result */
+	u64 found_offset;
+
+	/* Hint where to start looking for an empty space */
+	u64 hint_byte;
+
+	/* Allocation policy */
+	enum btrfs_extent_allocation_policy policy;
+};
+
 enum btrfs_inline_ref_type {
 	BTRFS_REF_TYPE_INVALID,
 	BTRFS_REF_TYPE_BLOCK,
diff --git a/fs/btrfs/super.c b/fs/btrfs/super.c
index 93f52ee85f6f..ddd4767fcd87 100644
--- a/fs/btrfs/super.c
+++ b/fs/btrfs/super.c
@@ -58,6 +58,7 @@
 #include "scrub.h"
 #include "verity.h"
 #include "super.h"
+#include "extent-tree.h"
 #define CREATE_TRACE_POINTS
 #include <trace/events/btrfs.h>
 
diff --git a/include/trace/events/btrfs.h b/include/trace/events/btrfs.h
index 0bce0b4ff2fa..423baed891d8 100644
--- a/include/trace/events/btrfs.h
+++ b/include/trace/events/btrfs.h
@@ -32,6 +32,7 @@ struct prelim_ref;
 struct btrfs_space_info;
 struct btrfs_raid_bio;
 struct raid56_bio_trace_info;
+struct find_free_extent_ctl;
 
 #define show_ref_type(type)						\
 	__print_symbolic(type,						\
@@ -1241,38 +1242,38 @@ DEFINE_EVENT(btrfs__reserved_extent,  btrfs_reserved_extent_free,
 
 TRACE_EVENT(find_free_extent,
 
-	TP_PROTO(const struct btrfs_root *root, u64 num_bytes,
-		 u64 empty_size, u64 data),
+	TP_PROTO(const struct btrfs_root *root,
+		 const struct find_free_extent_ctl *ffe_ctl),
 
-	TP_ARGS(root, num_bytes, empty_size, data),
+	TP_ARGS(root, ffe_ctl),
 
 	TP_STRUCT__entry_btrfs(
 		__field(	u64,	root_objectid		)
 		__field(	u64,	num_bytes		)
 		__field(	u64,	empty_size		)
-		__field(	u64,	data			)
+		__field(	u64,	flags			)
 	),
 
 	TP_fast_assign_btrfs(root->fs_info,
 		__entry->root_objectid	= root->root_key.objectid;
-		__entry->num_bytes	= num_bytes;
-		__entry->empty_size	= empty_size;
-		__entry->data		= data;
+		__entry->num_bytes	= ffe_ctl->num_bytes;
+		__entry->empty_size	= ffe_ctl->empty_size;
+		__entry->flags		= ffe_ctl->flags;
 	),
 
 	TP_printk_btrfs("root=%llu(%s) len=%llu empty_size=%llu flags=%llu(%s)",
 		  show_root_type(__entry->root_objectid),
-		  __entry->num_bytes, __entry->empty_size, __entry->data,
-		  __print_flags((unsigned long)__entry->data, "|",
+		  __entry->num_bytes, __entry->empty_size, __entry->flags,
+		  __print_flags((unsigned long)__entry->flags, "|",
 				 BTRFS_GROUP_FLAGS))
 );
 
 DECLARE_EVENT_CLASS(btrfs__reserve_extent,
 
-	TP_PROTO(const struct btrfs_block_group *block_group, u64 start,
-		 u64 len),
+	TP_PROTO(const struct btrfs_block_group *block_group,
+		 const struct find_free_extent_ctl *ffe_ctl),
 
-	TP_ARGS(block_group, start, len),
+	TP_ARGS(block_group, ffe_ctl),
 
 	TP_STRUCT__entry_btrfs(
 		__field(	u64,	bg_objectid		)
@@ -1284,8 +1285,8 @@ DECLARE_EVENT_CLASS(btrfs__reserve_extent,
 	TP_fast_assign_btrfs(block_group->fs_info,
 		__entry->bg_objectid	= block_group->start;
 		__entry->flags		= block_group->flags;
-		__entry->start		= start;
-		__entry->len		= len;
+		__entry->start		= ffe_ctl->search_start;
+		__entry->len		= ffe_ctl->num_bytes;
 	),
 
 	TP_printk_btrfs("root=%llu(%s) block_group=%llu flags=%llu(%s) "
@@ -1299,18 +1300,18 @@ DECLARE_EVENT_CLASS(btrfs__reserve_extent,
 
 DEFINE_EVENT(btrfs__reserve_extent, btrfs_reserve_extent,
 
-	TP_PROTO(const struct btrfs_block_group *block_group, u64 start,
-		 u64 len),
+	TP_PROTO(const struct btrfs_block_group *block_group,
+		 const struct find_free_extent_ctl *ffe_ctl),
 
-	TP_ARGS(block_group, start, len)
+	TP_ARGS(block_group, ffe_ctl)
 );
 
 DEFINE_EVENT(btrfs__reserve_extent, btrfs_reserve_extent_cluster,
 
-	TP_PROTO(const struct btrfs_block_group *block_group, u64 start,
-		 u64 len),
+	TP_PROTO(const struct btrfs_block_group *block_group,
+		 const struct find_free_extent_ctl *ffe_ctl),
 
-	TP_ARGS(block_group, start, len)
+	TP_ARGS(block_group, ffe_ctl)
 );
 
 TRACE_EVENT(btrfs_find_cluster,
-- 
2.38.1

[PATCH v3 2/4] btrfs: add more ffe tracepoints

[Boris Burkov]

find_free_extent is a complicated function. It consists (at least) of:
- a hint that jumps into the middle of a for loop macro
- a middle loop trying every raid level
- an outer loop ascending through ffe loop levels
- complicated logic for skipping some of those ffe loop levels
- multiple underlying in-bg allocators (zoned, cluster, no cluster)

Which is all to say that more tracing is helpful for debugging its
behavior. Add two new tracepoints: at the entrance to the block_groups
loop (hit for every raid level and every ffe_ctl loop) and at the point
we seriously consider a block_group for allocation. This way we can see
the whole path through the algorithm, including hints, multiple loops,
etc.

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Boris Burkov <boris@bur.io>
---
 fs/btrfs/extent-tree.c       |  4 ++
 fs/btrfs/extent-tree.h       |  3 ++
 include/trace/events/btrfs.h | 81 +++++++++++++++++++++++++++++++++++-
 3 files changed, 86 insertions(+), 2 deletions(-)

diff --git a/fs/btrfs/extent-tree.c b/fs/btrfs/extent-tree.c
index a53f90677fbc..0eca8fe4604b 100644
--- a/fs/btrfs/extent-tree.c
+++ b/fs/btrfs/extent-tree.c
@@ -4258,6 +4258,7 @@ static noinline int find_free_extent(struct btrfs_root *root,
 							block_group->flags);
 				btrfs_lock_block_group(block_group,
 						       ffe_ctl->delalloc);
+				ffe_ctl->hinted = true;
 				goto have_block_group;
 			}
 		} else if (block_group) {
@@ -4265,6 +4266,7 @@ static noinline int find_free_extent(struct btrfs_root *root,
 		}
 	}
 search:
+	trace_find_free_extent_search_loop(root, ffe_ctl);
 	ffe_ctl->have_caching_bg = false;
 	if (ffe_ctl->index == btrfs_bg_flags_to_raid_index(ffe_ctl->flags) ||
 	    ffe_ctl->index == 0)
@@ -4274,6 +4276,7 @@ static noinline int find_free_extent(struct btrfs_root *root,
 			    &space_info->block_groups[ffe_ctl->index], list) {
 		struct btrfs_block_group *bg_ret;
 
+		ffe_ctl->hinted = false;
 		/* If the block group is read-only, we can skip it entirely. */
 		if (unlikely(block_group->ro)) {
 			if (ffe_ctl->for_treelog)
@@ -4315,6 +4318,7 @@ static noinline int find_free_extent(struct btrfs_root *root,
 		}
 
 have_block_group:
+		trace_find_free_extent_have_block_group(root, ffe_ctl, block_group);
 		ffe_ctl->cached = btrfs_block_group_done(block_group);
 		if (unlikely(!ffe_ctl->cached)) {
 			ffe_ctl->have_caching_bg = true;
diff --git a/fs/btrfs/extent-tree.h b/fs/btrfs/extent-tree.h
index 64fa8ad7914a..daa5e3505886 100644
--- a/fs/btrfs/extent-tree.h
+++ b/fs/btrfs/extent-tree.h
@@ -76,6 +76,9 @@ struct find_free_extent_ctl {
 
 	/* Allocation policy */
 	enum btrfs_extent_allocation_policy policy;
+
+	/* Whether or not the allocator is currently following a hint */
+	bool hinted;
 };
 
 enum btrfs_inline_ref_type {
diff --git a/include/trace/events/btrfs.h b/include/trace/events/btrfs.h
index 423baed891d8..1c02614cd98c 100644
--- a/include/trace/events/btrfs.h
+++ b/include/trace/events/btrfs.h
@@ -1268,6 +1268,79 @@ TRACE_EVENT(find_free_extent,
 				 BTRFS_GROUP_FLAGS))
 );
 
+TRACE_EVENT(find_free_extent_search_loop,
+
+	TP_PROTO(const struct btrfs_root *root,
+		 const struct find_free_extent_ctl *ffe_ctl),
+
+	TP_ARGS(root, ffe_ctl),
+
+	TP_STRUCT__entry_btrfs(
+		__field(	u64,	root_objectid		)
+		__field(	u64,	num_bytes		)
+		__field(	u64,	empty_size		)
+		__field(	u64,	flags			)
+		__field(	u64,	loop			)
+	),
+
+	TP_fast_assign_btrfs(root->fs_info,
+		__entry->root_objectid	= root->root_key.objectid;
+		__entry->num_bytes	= ffe_ctl->num_bytes;
+		__entry->empty_size	= ffe_ctl->empty_size;
+		__entry->flags		= ffe_ctl->flags;
+		__entry->loop		= ffe_ctl->loop;
+	),
+
+	TP_printk_btrfs("root=%llu(%s) len=%llu empty_size=%llu flags=%llu(%s) loop=%llu",
+		  show_root_type(__entry->root_objectid),
+		  __entry->num_bytes, __entry->empty_size, __entry->flags,
+		  __print_flags((unsigned long)__entry->flags, "|",
+				 BTRFS_GROUP_FLAGS),
+		  __entry->loop)
+);
+
+TRACE_EVENT(find_free_extent_have_block_group,
+
+	TP_PROTO(const struct btrfs_root *root,
+		 const struct find_free_extent_ctl *ffe_ctl,
+		 const struct btrfs_block_group *block_group),
+
+	TP_ARGS(root, ffe_ctl, block_group),
+
+	TP_STRUCT__entry_btrfs(
+		__field(	u64,	root_objectid		)
+		__field(	u64,	num_bytes		)
+		__field(	u64,	empty_size		)
+		__field(	u64,	flags			)
+		__field(	u64,	loop			)
+		__field(	bool,	hinted			)
+		__field(	u64,	bg_start		)
+		__field(	u64,	bg_flags		)
+	),
+
+	TP_fast_assign_btrfs(root->fs_info,
+		__entry->root_objectid	= root->root_key.objectid;
+		__entry->num_bytes	= ffe_ctl->num_bytes;
+		__entry->empty_size	= ffe_ctl->empty_size;
+		__entry->flags		= ffe_ctl->flags;
+		__entry->loop		= ffe_ctl->loop;
+		__entry->hinted		= ffe_ctl->hinted;
+		__entry->bg_start	= block_group->start;
+		__entry->bg_flags	= block_group->flags;
+	),
+
+	TP_printk_btrfs("root=%llu(%s) len=%llu empty_size=%llu flags=%llu(%s) "
+			"loop=%llu hinted=%d block_group=%llu bg_flags=%llu(%s)",
+		  show_root_type(__entry->root_objectid),
+		  __entry->num_bytes, __entry->empty_size, __entry->flags,
+		  __print_flags((unsigned long)__entry->flags, "|",
+				 BTRFS_GROUP_FLAGS),
+		  __entry->loop, __entry->hinted,
+		  __entry->bg_start, __entry->bg_flags,
+		  __print_flags((unsigned long)__entry->bg_flags, "|",
+				 BTRFS_GROUP_FLAGS))
+);
+
 DECLARE_EVENT_CLASS(btrfs__reserve_extent,
 
 	TP_PROTO(const struct btrfs_block_group *block_group,
@@ -1280,6 +1353,8 @@ DECLARE_EVENT_CLASS(btrfs__reserve_extent,
 		__field(	u64,	flags			)
 		__field(	u64,	start			)
 		__field(	u64,	len			)
+		__field(	u64,	loop			)
+		__field(	bool,	hinted			)
 	),
 
 	TP_fast_assign_btrfs(block_group->fs_info,
@@ -1287,15 +1362,17 @@ DECLARE_EVENT_CLASS(btrfs__reserve_extent,
 		__entry->flags		= block_group->flags;
 		__entry->start		= ffe_ctl->search_start;
 		__entry->len		= ffe_ctl->num_bytes;
+		__entry->loop		= ffe_ctl->loop;
+		__entry->hinted		= ffe_ctl->hinted;
 	),
 
 	TP_printk_btrfs("root=%llu(%s) block_group=%llu flags=%llu(%s) "
-		  "start=%llu len=%llu",
+			"start=%llu len=%llu loop=%llu hinted=%d",
 		  show_root_type(BTRFS_EXTENT_TREE_OBJECTID),
 		  __entry->bg_objectid,
 		  __entry->flags, __print_flags((unsigned long)__entry->flags,
 						"|", BTRFS_GROUP_FLAGS),
-		  __entry->start, __entry->len)
+		  __entry->start, __entry->len, __entry->loop, __entry->hinted)
 );
 
 DEFINE_EVENT(btrfs__reserve_extent, btrfs_reserve_extent,
-- 
2.38.1

[PATCH v3 3/4] btrfs: introduce size class to block group allocator

[Boris Burkov]

The aim of this patch is to reduce the fragmentation of block groups
under certain unhappy workloads. It is particularly effective when the
size of extents correlates with their lifetime, which is something we
have observed causing fragmentation in the fleet at Meta.

This patch categorizes extents into size classes:
x < 128KiB: "small"
128KiB < x < 8MiB: "medium"
x > 8MiB: "large"
and as much as possible reduces allocations of extents into block groups
that don't match the size class. This takes advantage of any (possible)
correlation between size and lifetime and also leaves behind predictable
re-usable gaps when extents are freed; small writes don't gum up bigger
holes.

Size classes are implemented in the following way:
- Mark each new block group with a size class of the first allocation
  that goes into it.

- Add two new passes to ffe: "unset size class" and "wrong size class".
  First, try only matching block groups, then try unset ones, then allow
  allocation of new ones, and finally allow mismatched block groups.

- Filtering is done just by skipping inappropriate ones, there is no
  special size class indexing.

Other solutions I considered were:
- A best fit allocator with an rb-tree. This worked well, as small
  writes didn't leak big holes from large freed extents, but led to
  regressions in ffe and write performance due to lock contention on
  the rb-tree with every allocation possibly updating it in parallel.
  Perhaps something clever could be done to do the updates in the
  background while being "right enough".

- A fixed size "working set". This prevents freeing an extent
  drastically changing where writes currently land, and seems like a
  good option too. Doesn't take advantage of size in any way.

- The same size class idea, but implemented with xarray marks. This
  turned out to be slower than looping the linked list and skipping
  wrong block groups, and is also less flexible since we must have only
  3 size classes (max #marks). With the current approach we can have as
  many as we like.

Performance testing was done via:
https://github.com/josefbacik/fsperf
Of particular relevance are the new fragmentation specific tests.

A brief summary of the testing results:
- Neutral results on existing tests. There are some minor regressions
  and improvements here and there, but nothing that truly stands out as
  notable.
- Improvement on new tests where size class and extent lifetime are
  correlated. Fragmentation in these cases is completely eliminated
  and write performance is generally a little better. There is also
  significant improvement where extent sizes are just a bit larger than
  the size class boundaries.
- Regression on one new tests: where the allocations are sized
  intentionally a hair under the borders of the size classes. Results
  are neutral on the test that intentionally attacks this new scheme by
  mixing extent size and lifetime.

The full dump of the performance results can be found here:
https://bur.io/fsperf/size-class-2022-11-15.txt
(there are ansi escape codes, so best to curl and view in terminal)

Here is a snippet from the full results for a new test which mixes
buffered writes appending to a long lived set of files and large short
lived fallocates:

bufferedappendvsfallocate results
         metric             baseline       current        stdev            diff
======================================================================================
avg_commit_ms                    31.13         29.20          2.67     -6.22%
bg_count                            14         15.60             0     11.43%
commits                          11.10         12.20          0.32      9.91%
elapsed                          27.30         26.40          2.98     -3.30%
end_state_mount_ns         11122551.90   10635118.90     851143.04     -4.38%
end_state_umount_ns           1.36e+09      1.35e+09   12248056.65     -1.07%
find_free_extent_calls       116244.30     114354.30        964.56     -1.63%
find_free_extent_ns_max      599507.20    1047168.20     103337.08     74.67%
find_free_extent_ns_mean       3607.19       3672.11        101.20      1.80%
find_free_extent_ns_min            500           512          6.67      2.40%
find_free_extent_ns_p50           2848          2876         37.65      0.98%
find_free_extent_ns_p95           4916          5000         75.45      1.71%
find_free_extent_ns_p99       20734.49      20920.48       1670.93      0.90%
frag_pct_max                     61.67             0          8.05   -100.00%
frag_pct_mean                    43.59             0          6.10   -100.00%
frag_pct_min                     25.91             0         16.60   -100.00%
frag_pct_p50                     42.53             0          7.25   -100.00%
frag_pct_p95                     61.67             0          8.05   -100.00%
frag_pct_p99                     61.67             0          8.05   -100.00%
fragmented_bg_count               6.10             0          1.45   -100.00%
max_commit_ms                    49.80            46          5.37     -7.63%
sys_cpu                           2.59          2.62          0.29      1.39%
write_bw_bytes                1.62e+08      1.68e+08   17975843.50      3.23%
write_clat_ns_mean            57426.39      54475.95       2292.72     -5.14%
write_clat_ns_p50             46950.40      42905.60       2101.35     -8.62%
write_clat_ns_p99            148070.40     143769.60       2115.17     -2.90%
write_io_kbytes                4194304       4194304             0      0.00%
write_iops                     2476.15       2556.10        274.29      3.23%
write_lat_ns_max            2101667.60    2251129.50     370556.59      7.11%
write_lat_ns_mean             59374.91      55682.00       2523.09     -6.22%
write_lat_ns_min              17353.10         16250       1646.08     -6.36%

There are some mixed improvements/regressions in most metrics along with
an elimination of fragmentation in this workload.

On the balance, the drastic 1->0 improvement in the happy cases seems
worth the mix of regressions and improvements we do observe.

Some considerations for future work:
- Experimenting with more size classes
- More hinting/search ordering work to approximate a best-fit allocator

Signed-off-by: Boris Burkov <boris@bur.io>
---
 fs/btrfs/block-group.c       | 104 +++++++++++++++++++++++++++++------
 fs/btrfs/block-group.h       |  15 ++++-
 fs/btrfs/extent-tree.c       |  71 ++++++++++++------------
 fs/btrfs/extent-tree.h       |   3 +
 include/trace/events/btrfs.h |  10 +++-
 5 files changed, 150 insertions(+), 53 deletions(-)

diff --git a/fs/btrfs/block-group.c b/fs/btrfs/block-group.c
index 708d843daa72..fa1ab56fe6b3 100644
--- a/fs/btrfs/block-group.c
+++ b/fs/btrfs/block-group.c
@@ -1,5 +1,6 @@
 // SPDX-License-Identifier: GPL-2.0
 
+#include "linux/sizes.h"
 #include <linux/list_sort.h>
 #include "misc.h"
 #include "ctree.h"
@@ -3379,6 +3380,7 @@ int btrfs_update_block_group(struct btrfs_trans_handle *trans,
 			cache->space_info->disk_used -= num_bytes * factor;
 
 			reclaim = should_reclaim_block_group(cache, num_bytes);
+
 			spin_unlock(&cache->lock);
 			spin_unlock(&cache->space_info->lock);
 
@@ -3433,32 +3435,42 @@ int btrfs_update_block_group(struct btrfs_trans_handle *trans,
  * reservation and return -EAGAIN, otherwise this function always succeeds.
  */
 int btrfs_add_reserved_bytes(struct btrfs_block_group *cache,
-			     u64 ram_bytes, u64 num_bytes, int delalloc)
+			     u64 ram_bytes, u64 num_bytes, int delalloc,
+			     bool force_wrong_size_class)
 {
 	struct btrfs_space_info *space_info = cache->space_info;
+	enum btrfs_block_group_size_class size_class;
 	int ret = 0;
 
 	spin_lock(&space_info->lock);
 	spin_lock(&cache->lock);
 	if (cache->ro) {
 		ret = -EAGAIN;
-	} else {
-		cache->reserved += num_bytes;
-		space_info->bytes_reserved += num_bytes;
-		trace_btrfs_space_reservation(cache->fs_info, "space_info",
-					      space_info->flags, num_bytes, 1);
-		btrfs_space_info_update_bytes_may_use(cache->fs_info,
-						      space_info, -ram_bytes);
-		if (delalloc)
-			cache->delalloc_bytes += num_bytes;
+		goto out;
+	}
 
-		/*
-		 * Compression can use less space than we reserved, so wake
-		 * tickets if that happens
-		 */
-		if (num_bytes < ram_bytes)
-			btrfs_try_granting_tickets(cache->fs_info, space_info);
+	if (btrfs_is_block_group_data_only(cache)) {
+		size_class = btrfs_calc_block_group_size_class(num_bytes);
+		ret = btrfs_use_block_group_size_class(cache, size_class, force_wrong_size_class);
+		if (ret)
+			goto out;
 	}
+	cache->reserved += num_bytes;
+	space_info->bytes_reserved += num_bytes;
+	trace_btrfs_space_reservation(cache->fs_info, "space_info",
+				      space_info->flags, num_bytes, 1);
+	btrfs_space_info_update_bytes_may_use(cache->fs_info,
+					      space_info, -ram_bytes);
+	if (delalloc)
+		cache->delalloc_bytes += num_bytes;
+
+	/*
+	 * Compression can use less space than we reserved, so wake
+	 * tickets if that happens
+	 */
+	if (num_bytes < ram_bytes)
+		btrfs_try_granting_tickets(cache->fs_info, space_info);
+out:
 	spin_unlock(&cache->lock);
 	spin_unlock(&space_info->lock);
 	return ret;
@@ -4218,3 +4230,63 @@ void btrfs_dec_block_group_swap_extents(struct btrfs_block_group *bg, int amount
 	bg->swap_extents -= amount;
 	spin_unlock(&bg->lock);
 }
+
+enum btrfs_block_group_size_class btrfs_calc_block_group_size_class(u64 size)
+{
+	if (size <= SZ_128K)
+		return BTRFS_BG_SZ_SMALL;
+	if (size <= SZ_8M)
+		return BTRFS_BG_SZ_MEDIUM;
+	return BTRFS_BG_SZ_LARGE;
+}
+
+/*
+ * Handle a block group allocating an extent in a size class
+ *
+ * @bg:				The block group we allocated in.
+ * @size_class:			The size class of the allocation.
+ * @force_wrong_size_class:	Whether we are desperate enough to allow
+ *				mismatched size classes.
+ *
+ * Returns: 0 if the size class was valid for this block_group, -EAGAIN in the
+ * case of a race that leads to the wrong size class without
+ * force_wrong_size_class set.
+ *
+ * find_free_extent will skip block groups with a mismatched size class until
+ * it really needs to avoid ENOSPC. In that case it will set
+ * force_wrong_size_class. However, if a block group is newly allocated and
+ * doesn't yet have a size class, then it is possible for two allocations of
+ * different sizes to race and both try to use it. The loser is caught here and
+ * has to retry.
+ */
+int btrfs_use_block_group_size_class(struct btrfs_block_group *bg,
+				     enum btrfs_block_group_size_class size_class,
+				     bool force_wrong_size_class)
+{
+	ASSERT(size_class != BTRFS_BG_SZ_NONE);
+
+	/* The new allocation is in the right size class, do nothing */
+	if (bg->size_class == size_class)
+		return 0;
+	/*
+	 * The new allocation is in a mismatched size class.
+	 * This means one of two things:
+	 * 1. Two tasks in find_free_extent for different size_classes raced
+	 *    and hit the same empty block_group. Make the loser try again.
+	 * 2. A call to find_free_extent got desperate enough to set
+	 *    'force_wrong_slab'. Don't change the size_class, but allow the
+	 *    allocation.
+	 */
+	if (bg->size_class != BTRFS_BG_SZ_NONE) {
+		if (force_wrong_size_class)
+			return 0;
+		return -EAGAIN;
+	}
+	/*
+	 * The happy new block group case: the new allocation is the first
+	 * one in the block_group so we set size_class.
+	 */
+	bg->size_class = size_class;
+
+	return 0;
+}
diff --git a/fs/btrfs/block-group.h b/fs/btrfs/block-group.h
index a02ea76fd6cf..aaf5ca49defb 100644
--- a/fs/btrfs/block-group.h
+++ b/fs/btrfs/block-group.h
@@ -12,6 +12,13 @@ enum btrfs_disk_cache_state {
 	BTRFS_DC_SETUP,
 };
 
+enum btrfs_block_group_size_class {
+	BTRFS_BG_SZ_NONE,	/* unset		*/
+	BTRFS_BG_SZ_SMALL,	/* 0 < sz <= 128K	*/
+	BTRFS_BG_SZ_MEDIUM,	/* 128K < sz <= 8M	*/
+	BTRFS_BG_SZ_LARGE,	/* 8M < sz < BG_LEN	*/
+};
+
 /*
  * This describes the state of the block_group for async discard.  This is due
  * to the two pass nature of it where extent discarding is prioritized over
@@ -233,6 +240,7 @@ struct btrfs_block_group {
 	struct list_head active_bg_list;
 	struct work_struct zone_finish_work;
 	struct extent_buffer *last_eb;
+	enum btrfs_block_group_size_class size_class;
 };
 
 static inline u64 btrfs_block_group_end(struct btrfs_block_group *block_group)
@@ -302,7 +310,8 @@ int btrfs_setup_space_cache(struct btrfs_trans_handle *trans);
 int btrfs_update_block_group(struct btrfs_trans_handle *trans,
 			     u64 bytenr, u64 num_bytes, bool alloc);
 int btrfs_add_reserved_bytes(struct btrfs_block_group *cache,
-			     u64 ram_bytes, u64 num_bytes, int delalloc);
+			     u64 ram_bytes, u64 num_bytes, int delalloc,
+			     bool force_wrong_size_class);
 void btrfs_free_reserved_bytes(struct btrfs_block_group *cache,
 			       u64 num_bytes, int delalloc);
 int btrfs_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
@@ -346,4 +355,8 @@ void btrfs_unfreeze_block_group(struct btrfs_block_group *cache);
 bool btrfs_inc_block_group_swap_extents(struct btrfs_block_group *bg);
 void btrfs_dec_block_group_swap_extents(struct btrfs_block_group *bg, int amount);
 
+enum btrfs_block_group_size_class btrfs_calc_block_group_size_class(u64 size);
+int btrfs_use_block_group_size_class(struct btrfs_block_group *bg,
+				     enum btrfs_block_group_size_class size_class,
+				     bool force_wrong_size_class);
 #endif /* BTRFS_BLOCK_GROUP_H */
diff --git a/fs/btrfs/extent-tree.c b/fs/btrfs/extent-tree.c
index 0eca8fe4604b..f3d8cc281d80 100644
--- a/fs/btrfs/extent-tree.c
+++ b/fs/btrfs/extent-tree.c
@@ -3382,7 +3382,9 @@ int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_ref *ref)
 enum btrfs_loop_type {
 	LOOP_CACHING_NOWAIT,
 	LOOP_CACHING_WAIT,
+	LOOP_UNSET_SIZE_CLASS,
 	LOOP_ALLOC_CHUNK,
+	LOOP_WRONG_SIZE_CLASS,
 	LOOP_NO_EMPTY_SIZE,
 };
 
@@ -3947,24 +3949,6 @@ static int can_allocate_chunk(struct btrfs_fs_info *fs_info,
 	}
 }
 
-static int chunk_allocation_failed(struct find_free_extent_ctl *ffe_ctl)
-{
-	switch (ffe_ctl->policy) {
-	case BTRFS_EXTENT_ALLOC_CLUSTERED:
-		/*
-		 * If we can't allocate a new chunk we've already looped through
-		 * at least once, move on to the NO_EMPTY_SIZE case.
-		 */
-		ffe_ctl->loop = LOOP_NO_EMPTY_SIZE;
-		return 0;
-	case BTRFS_EXTENT_ALLOC_ZONED:
-		/* Give up here */
-		return -ENOSPC;
-	default:
-		BUG();
-	}
-}
-
 /*
  * Return >0 means caller needs to re-search for free extent
  * Return 0 means we have the needed free extent.
@@ -3998,31 +3982,28 @@ static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
 	 * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
 	 *			caching kthreads as we move along
 	 * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
+	 * LOOP_UNSET_SIZE_CLASS, allow unset size class
 	 * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
 	 * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
 	 *		       again
 	 */
 	if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
 		ffe_ctl->index = 0;
-		if (ffe_ctl->loop == LOOP_CACHING_NOWAIT) {
-			/*
-			 * We want to skip the LOOP_CACHING_WAIT step if we
-			 * don't have any uncached bgs and we've already done a
-			 * full search through.
-			 */
-			if (ffe_ctl->orig_have_caching_bg || !full_search)
-				ffe_ctl->loop = LOOP_CACHING_WAIT;
-			else
-				ffe_ctl->loop = LOOP_ALLOC_CHUNK;
-		} else {
+		/*
+		 * We want to skip the LOOP_CACHING_WAIT step if we
+		 * don't have any uncached bgs and we've already done a
+		 * full search through.
+		 */
+		if (ffe_ctl->loop == LOOP_CACHING_NOWAIT &&
+		    (!ffe_ctl->orig_have_caching_bg && full_search))
 			ffe_ctl->loop++;
-		}
+		ffe_ctl->loop++;
 
 		if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
 			struct btrfs_trans_handle *trans;
 			int exist = 0;
 
-			/*Check if allocation policy allows to create a new chunk */
+			/* Check if allocation policy allows to create a new chunk */
 			ret = can_allocate_chunk(fs_info, ffe_ctl);
 			if (ret)
 				return ret;
@@ -4042,8 +4023,10 @@ static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
 						CHUNK_ALLOC_FORCE_FOR_EXTENT);
 
 			/* Do not bail out on ENOSPC since we can do more. */
-			if (ret == -ENOSPC)
-				ret = chunk_allocation_failed(ffe_ctl);
+			if (ret == -ENOSPC) {
+				ret = 0;
+				ffe_ctl->loop++;
+			}
 			else if (ret < 0)
 				btrfs_abort_transaction(trans, ret);
 			else
@@ -4073,6 +4056,21 @@ static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
 	return -ENOSPC;
 }
 
+static bool find_free_extent_check_size_class(struct find_free_extent_ctl *ffe_ctl,
+					      struct btrfs_block_group *bg)
+{
+	if (ffe_ctl->policy == BTRFS_EXTENT_ALLOC_ZONED)
+		return true;
+	if (!btrfs_is_block_group_data_only(bg))
+		return true;
+	if (ffe_ctl->loop >= LOOP_WRONG_SIZE_CLASS)
+		return true;
+	if (ffe_ctl->loop >= LOOP_UNSET_SIZE_CLASS &&
+	    bg->size_class == BTRFS_BG_SZ_NONE)
+		return true;
+	return ffe_ctl->size_class == bg->size_class;
+}
+
 static int prepare_allocation_clustered(struct btrfs_fs_info *fs_info,
 					struct find_free_extent_ctl *ffe_ctl,
 					struct btrfs_space_info *space_info,
@@ -4207,6 +4205,7 @@ static noinline int find_free_extent(struct btrfs_root *root,
 	ffe_ctl->total_free_space = 0;
 	ffe_ctl->found_offset = 0;
 	ffe_ctl->policy = BTRFS_EXTENT_ALLOC_CLUSTERED;
+	ffe_ctl->size_class = btrfs_calc_block_group_size_class(ffe_ctl->num_bytes);
 
 	if (btrfs_is_zoned(fs_info))
 		ffe_ctl->policy = BTRFS_EXTENT_ALLOC_ZONED;
@@ -4343,6 +4342,9 @@ static noinline int find_free_extent(struct btrfs_root *root,
 		if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
 			goto loop;
 
+		if (!find_free_extent_check_size_class(ffe_ctl, block_group))
+			goto loop;
+
 		bg_ret = NULL;
 		ret = do_allocation(block_group, ffe_ctl, &bg_ret);
 		if (ret == 0) {
@@ -4377,7 +4379,8 @@ static noinline int find_free_extent(struct btrfs_root *root,
 
 		ret = btrfs_add_reserved_bytes(block_group, ffe_ctl->ram_bytes,
 					       ffe_ctl->num_bytes,
-					       ffe_ctl->delalloc);
+					       ffe_ctl->delalloc,
+					       ffe_ctl->loop >= LOOP_WRONG_SIZE_CLASS);
 		if (ret == -EAGAIN) {
 			btrfs_add_free_space_unused(block_group,
 					ffe_ctl->found_offset,
diff --git a/fs/btrfs/extent-tree.h b/fs/btrfs/extent-tree.h
index daa5e3505886..0c958fc1b3b8 100644
--- a/fs/btrfs/extent-tree.h
+++ b/fs/btrfs/extent-tree.h
@@ -79,6 +79,9 @@ struct find_free_extent_ctl {
 
 	/* Whether or not the allocator is currently following a hint */
 	bool hinted;
+
+	/* Size class of block groups to prefer in early loops */
+	enum btrfs_block_group_size_class size_class;
 };
 
 enum btrfs_inline_ref_type {
diff --git a/include/trace/events/btrfs.h b/include/trace/events/btrfs.h
index 1c02614cd98c..de55f713e042 100644
--- a/include/trace/events/btrfs.h
+++ b/include/trace/events/btrfs.h
@@ -1351,28 +1351,34 @@ DECLARE_EVENT_CLASS(btrfs__reserve_extent,
 	TP_STRUCT__entry_btrfs(
 		__field(	u64,	bg_objectid		)
 		__field(	u64,	flags			)
+		__field(	int,	bg_size_class		)
 		__field(	u64,	start			)
 		__field(	u64,	len			)
 		__field(	u64,	loop			)
 		__field(	bool,	hinted			)
+		__field(	int,	size_class		)
 	),
 
 	TP_fast_assign_btrfs(block_group->fs_info,
 		__entry->bg_objectid	= block_group->start;
 		__entry->flags		= block_group->flags;
+		__entry->bg_size_class	= block_group->size_class;
 		__entry->start		= ffe_ctl->search_start;
 		__entry->len		= ffe_ctl->num_bytes;
 		__entry->loop		= ffe_ctl->loop;
 		__entry->hinted		= ffe_ctl->hinted;
+		__entry->size_class	= ffe_ctl->size_class;
 	),
 
 	TP_printk_btrfs("root=%llu(%s) block_group=%llu flags=%llu(%s) "
-			"start=%llu len=%llu loop=%llu hinted=%d",
+			"bg_size_class=%d start=%llu len=%llu loop=%llu "
+			"hinted=%d size_class=%d",
 		  show_root_type(BTRFS_EXTENT_TREE_OBJECTID),
 		  __entry->bg_objectid,
 		  __entry->flags, __print_flags((unsigned long)__entry->flags,
 						"|", BTRFS_GROUP_FLAGS),
-		  __entry->start, __entry->len, __entry->loop, __entry->hinted)
+		  __entry->bg_size_class, __entry->start, __entry->len,
+		  __entry->loop, __entry->hinted, __entry->size_class)
 );
 
 DEFINE_EVENT(btrfs__reserve_extent, btrfs_reserve_extent,
-- 
2.38.1

[PATCH v3 4/4] btrfs: load block group size class when caching

[Boris Burkov]

Since the size class is an artifact of an arbitrary anti fragmentation
strategy, it doesn't really make sense to persist it. Furthermore, most
of the size class logic assumes fresh block groups. That is of course
not a reasonable assumption -- we will be upgrading kernels with
existing filesystems whose block groups are not classified.

To work around those issues, implement logic to compute the size class
of the block groups as we cache them in. To perfectly assess the state
of a block group, we would have to read the entire extent tree (since
the free space cache mashes together contiguous extent items) which
would be prohibitively expensive for larger file systems with more
extents.

We can do it relatively cheaply by implementing a simple heuristic of
sampling a handful of extents and picking the smallest one we see. In
the happy case where the block group was classified, we will only see
extents of the correct size. In the unhappy case, we will hopefully find
one of the smaller extents, but there is no perfect answer anyway.
Autorelocation will eventually churn up the block group if there is
significant free-ing anyway.

The work is done in the caching thread but after marking the block group
cached, as we tradeoff classification accuracy vs. slowing down
allocations.

There was no regression in mount performance at end state of the fsperf
test suite.

Signed-off-by: Boris Burkov <boris@bur.io>
---
 fs/btrfs/block-group.c | 130 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 130 insertions(+)

diff --git a/fs/btrfs/block-group.c b/fs/btrfs/block-group.c
index fa1ab56fe6b3..75d5f952067a 100644
--- a/fs/btrfs/block-group.c
+++ b/fs/btrfs/block-group.c
@@ -540,6 +540,134 @@ u64 add_new_free_space(struct btrfs_block_group *block_group, u64 start, u64 end
 	return total_added;
 }
 
+/*
+ * Get an arbitrary extent item index / max_index through the block group
+ *
+ * @block_group: the block group to sample from
+ * @index: the integral step through the block group to grab from
+ * @max_index: the granularity of the sampling
+ * @key: return value parameter for the item we find
+ *
+ * pre-conditions on indices:
+ * 0 <= index <= max_index
+ * 0 < max_index
+ *
+ * Returns: 0 on success, 1 if the search didn't yield a useful item, negative
+ * error code on error.
+ */
+static int sample_block_group_extent_item(struct btrfs_block_group *block_group,
+					  int index, int max_index,
+					  struct btrfs_key *key)
+{
+	struct btrfs_fs_info *fs_info = block_group->fs_info;
+	struct btrfs_root *extent_root;
+	int ret = 0;
+	u64 search_offset;
+	struct btrfs_path *path;
+
+	ASSERT(index >= 0);
+	ASSERT(index <= max_index);
+	ASSERT(max_index > 0);
+
+	path = btrfs_alloc_path();
+	if (!path)
+		return -ENOMEM;
+
+	down_read(&fs_info->commit_root_sem);
+	extent_root = btrfs_extent_root(fs_info, max_t(u64, block_group->start,
+						       BTRFS_SUPER_INFO_OFFSET));
+
+	path->skip_locking = 1;
+	path->search_commit_root = 1;
+	path->reada = READA_FORWARD;
+
+	search_offset = index * (block_group->length / max_index);
+	key->objectid = block_group->start + search_offset;
+	key->offset = 0;
+	key->type = BTRFS_EXTENT_ITEM_KEY;
+
+	ret = btrfs_search_slot(NULL, extent_root, key, path, 0, 0);
+	if (ret != 0)
+		goto out;
+	if (key->objectid < block_group->start ||
+	    key->objectid > block_group->start + block_group->length) {
+		ret = 1;
+		goto out;
+	}
+	if (key->type != BTRFS_EXTENT_ITEM_KEY) {
+		ret = 1;
+		goto out;
+	}
+out:
+	btrfs_free_path(path);
+	up_read(&fs_info->commit_root_sem);
+	return ret;
+}
+
+/*
+ * Best effort attempt to compute a block group's size class while caching it.
+ *
+ * @block_group: the block group we are caching
+ *
+ * We cannot infer the size class while adding free space extents, because that
+ * logic doesn't care about contiguous file extents (it doesn't differentiate
+ * between a 100M extent and 100 contiguous 1M extents). So we need to read the
+ * file extent items. Reading all of them is quite wasteful, because usually
+ * only a handful are enough to give a good answer. Therefore, we just grab 5 of
+ * them at even steps through the block group and pick the smallest size class
+ * we see. Since size class is best effort, and not guaranteed in general,
+ * inaccuracy is acceptable.
+ *
+ * To be more explicit about why this algorithm makes sense:
+ *
+ * If we are caching in a block group from disk, then there are three major cases
+ * to consider:
+ * 1. the block group is well behaved and all extents in it are the same size
+ * class.
+ * 2. the block group is mostly one size class with rare exceptions for last
+ * ditch allocations
+ * 3. the block group was populated before size classes and can have a totally
+ * arbitrary mix of size classes.
+ *
+ * In case 1, looking at any extent in the block group will yield the correct
+ * result. For the mixed cases, taking the minimum size class seems like a good
+ * approximation, since gaps from frees will be usable to the size class. For
+ * 2., a small handful of file extents is likely to yield the right answer. For
+ * 3, we can either read every file extent, or admit that this is best effort
+ * anyway and try to stay fast.
+ *
+ * Returns: 0 on success, negative error code on error.
+ */
+static int load_block_group_size_class(struct btrfs_block_group *block_group)
+{
+	struct btrfs_key key;
+	int i;
+	u64 min_size = block_group->length;
+	enum btrfs_block_group_size_class size_class = BTRFS_BG_SZ_NONE;
+	int ret;
+
+	if (!btrfs_is_block_group_data_only(block_group))
+		return 0;
+
+	for (i = 0; i < 5; ++i) {
+		ret = sample_block_group_extent_item(block_group, i, 5, &key);
+		if (ret < 0)
+			goto out;
+		if (ret > 0)
+			continue;
+		min_size = min_t(u64, min_size, key.offset);
+		size_class = btrfs_calc_block_group_size_class(min_size);
+	}
+	if (size_class != BTRFS_BG_SZ_NONE) {
+		spin_lock(&block_group->lock);
+		block_group->size_class = size_class;
+		spin_unlock(&block_group->lock);
+	}
+
+out:
+	return ret;
+}
+
 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 {
 	struct btrfs_block_group *block_group = caching_ctl->block_group;
@@ -739,6 +867,8 @@ static noinline void caching_thread(struct btrfs_work *work)
 
 	wake_up(&caching_ctl->wait);
 
+	load_block_group_size_class(block_group);
+
 	btrfs_put_caching_control(caching_ctl);
 	btrfs_put_block_group(block_group);
 }
-- 
2.38.1