btrfs: add a comment explaining the data flush steps

The data flushing steps are not obvious to people other than myself and
Chris.  Write a giant comment explaining the reasoning behind each flush
step for data as well as why it is in that particular order.

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>

1 files changed, 47 insertions, 0 deletions

diff --git a/fs/btrfs/space-info.c b/fs/btrfs/space-info.c
index bbddb8b5eff1..71aa9e0de61e 100644
--- a/fs/btrfs/space-info.c
+++ b/fs/btrfs/space-info.c
@@ -998,6 +998,53 @@ static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
 	} while (flush_state <= COMMIT_TRANS);
 }
 
+/*
+ * FLUSH_DELALLOC_WAIT:
+ *   Space is freed from flushing delalloc in one of two ways.
+ *
+ *   1) compression is on and we allocate less space than we reserved
+ *   2) we are overwriting existing space
+ *
+ *   For #1 that extra space is reclaimed as soon as the delalloc pages are
+ *   COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent
+ *   length to ->bytes_reserved, and subtracts the reserved space from
+ *   ->bytes_may_use.
+ *
+ *   For #2 this is trickier.  Once the ordered extent runs we will drop the
+ *   extent in the range we are overwriting, which creates a delayed ref for
+ *   that freed extent.  This however is not reclaimed until the transaction
+ *   commits, thus the next stages.
+ *
+ * RUN_DELAYED_IPUTS
+ *   If we are freeing inodes, we want to make sure all delayed iputs have
+ *   completed, because they could have been on an inode with i_nlink == 0, and
+ *   thus have been truncated and freed up space.  But again this space is not
+ *   immediately re-usable, it comes in the form of a delayed ref, which must be
+ *   run and then the transaction must be committed.
+ *
+ * FLUSH_DELAYED_REFS
+ *   The above two cases generate delayed refs that will affect
+ *   ->total_bytes_pinned.  However this counter can be inconsistent with
+ *   reality if there are outstanding delayed refs.  This is because we adjust
+ *   the counter based solely on the current set of delayed refs and disregard
+ *   any on-disk state which might include more refs.  So for example, if we
+ *   have an extent with 2 references, but we only drop 1, we'll see that there
+ *   is a negative delayed ref count for the extent and assume that the space
+ *   will be freed, and thus increase ->total_bytes_pinned.
+ *
+ *   Running the delayed refs gives us the actual real view of what will be
+ *   freed at the transaction commit time.  This stage will not actually free
+ *   space for us, it just makes sure that may_commit_transaction() has all of
+ *   the information it needs to make the right decision.
+ *
+ * COMMIT_TRANS
+ *   This is where we reclaim all of the pinned space generated by the previous
+ *   two stages.  We will not commit the transaction if we don't think we're
+ *   likely to satisfy our request, which means if our current free space +
+ *   total_bytes_pinned < reservation we will not commit.  This is why the
+ *   previous states are actually important, to make sure we know for sure
+ *   whether committing the transaction will allow us to make progress.
+ */
 static const enum btrfs_flush_state data_flush_states[] = {
 	FLUSH_DELALLOC_WAIT,
 	RUN_DELAYED_IPUTS,