On Fri, Sep 28, 2018 at 02:51:10PM +0300, Nikolay Borisov wrote: > > > On 28.09.2018 14:17, Josef Bacik wrote: > > From: Josef Bacik <jba...@fb.com> > > > > Traditionally we've had voodoo in btrfs to account for the space that > > delayed refs may take up by having a global_block_rsv. This works most > > of the time, except when it doesn't. We've had issues reported and seen > > in production where sometimes the global reserve is exhausted during > > transaction commit before we can run all of our delayed refs, resulting > > in an aborted transaction. Because of this voodoo we have equally > > dubious flushing semantics around throttling delayed refs which we often > > get wrong. > > > > So instead give them their own block_rsv. This way we can always know > > exactly how much outstanding space we need for delayed refs. This > > allows us to make sure we are constantly filling that reservation up > > with space, and allows us to put more precise pressure on the enospc > > system. Instead of doing math to see if its a good time to throttle, > > the normal enospc code will be invoked if we have a lot of delayed refs > > pending, and they will be run via the normal flushing mechanism. > > > > For now the delayed_refs_rsv will hold the reservations for the delayed > > refs, the block group updates, and deleting csums. We could have a > > separate rsv for the block group updates, but the csum deletion stuff is > > still handled via the delayed_refs so that will stay there. > > > > Signed-off-by: Josef Bacik <jba...@fb.com> > > --- > > fs/btrfs/ctree.h | 27 +++-- > > fs/btrfs/delayed-ref.c | 28 ++++- > > fs/btrfs/disk-io.c | 4 + > > fs/btrfs/extent-tree.c | 279 > > +++++++++++++++++++++++++++++++++++-------- > > fs/btrfs/inode.c | 2 +- > > fs/btrfs/transaction.c | 77 ++++++------ > > include/trace/events/btrfs.h | 2 + > > 7 files changed, 312 insertions(+), 107 deletions(-) > > > > diff --git a/fs/btrfs/ctree.h b/fs/btrfs/ctree.h > > index 66f1d3895bca..1a2c3b629af2 100644 > > --- a/fs/btrfs/ctree.h > > +++ b/fs/btrfs/ctree.h > > @@ -452,8 +452,9 @@ struct btrfs_space_info { > > #define BTRFS_BLOCK_RSV_TRANS 3 > > #define BTRFS_BLOCK_RSV_CHUNK 4 > > #define BTRFS_BLOCK_RSV_DELOPS 5 > > -#define BTRFS_BLOCK_RSV_EMPTY 6 > > -#define BTRFS_BLOCK_RSV_TEMP 7 > > +#define BTRFS_BLOCK_RSV_DELREFS 6 > > +#define BTRFS_BLOCK_RSV_EMPTY 7 > > +#define BTRFS_BLOCK_RSV_TEMP 8 > > > > struct btrfs_block_rsv { > > u64 size; > > @@ -794,6 +795,8 @@ struct btrfs_fs_info { > > struct btrfs_block_rsv chunk_block_rsv; > > /* block reservation for delayed operations */ > > struct btrfs_block_rsv delayed_block_rsv; > > + /* block reservation for delayed refs */ > > + struct btrfs_block_rsv delayed_refs_rsv; > > > > struct btrfs_block_rsv empty_block_rsv; > > > > @@ -2608,8 +2611,7 @@ static inline u64 > > btrfs_calc_trunc_metadata_size(struct btrfs_fs_info *fs_info, > > > > int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans, > > struct btrfs_fs_info *fs_info); > > -int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans, > > - struct btrfs_fs_info *fs_info); > > +bool btrfs_check_space_for_delayed_refs(struct btrfs_fs_info *fs_info); > > void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info, > > const u64 start); > > void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache > > *bg); > > @@ -2723,10 +2725,12 @@ enum btrfs_reserve_flush_enum { > > enum btrfs_flush_state { > > FLUSH_DELAYED_ITEMS_NR = 1, > > FLUSH_DELAYED_ITEMS = 2, > > - FLUSH_DELALLOC = 3, > > - FLUSH_DELALLOC_WAIT = 4, > > - ALLOC_CHUNK = 5, > > - COMMIT_TRANS = 6, > > + FLUSH_DELAYED_REFS_NR = 3, > > + FLUSH_DELAYED_REFS = 4, > > + FLUSH_DELALLOC = 5, > > + FLUSH_DELALLOC_WAIT = 6, > > + ALLOC_CHUNK = 7, > > + COMMIT_TRANS = 8, > > }; > > > > int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes); > > @@ -2777,6 +2781,13 @@ int btrfs_cond_migrate_bytes(struct btrfs_fs_info > > *fs_info, > > void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info, > > struct btrfs_block_rsv *block_rsv, > > u64 num_bytes); > > +void btrfs_delayed_refs_rsv_release(struct btrfs_fs_info *fs_info, int nr); > > +void btrfs_update_delayed_refs_rsv(struct btrfs_trans_handle *trans); > > +int btrfs_throttle_delayed_refs(struct btrfs_fs_info *fs_info, > > + enum btrfs_reserve_flush_enum flush); > > +void btrfs_migrate_to_delayed_refs_rsv(struct btrfs_fs_info *fs_info, > > + struct btrfs_block_rsv *src, > > + u64 num_bytes); > > int btrfs_inc_block_group_ro(struct btrfs_block_group_cache *cache); > > void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache); > > void btrfs_put_block_group_cache(struct btrfs_fs_info *info); > > diff --git a/fs/btrfs/delayed-ref.c b/fs/btrfs/delayed-ref.c > > index 27f7dd4e3d52..96ce087747b2 100644 > > --- a/fs/btrfs/delayed-ref.c > > +++ b/fs/btrfs/delayed-ref.c > > @@ -467,11 +467,14 @@ static int insert_delayed_ref(struct > > btrfs_trans_handle *trans, > > * existing and update must have the same bytenr > > */ > > static noinline void > > -update_existing_head_ref(struct btrfs_delayed_ref_root *delayed_refs, > > +update_existing_head_ref(struct btrfs_trans_handle *trans, > > struct btrfs_delayed_ref_head *existing, > > struct btrfs_delayed_ref_head *update, > > int *old_ref_mod_ret) > > { > > + struct btrfs_delayed_ref_root *delayed_refs = > > + &trans->transaction->delayed_refs; > > + struct btrfs_fs_info *fs_info = trans->fs_info; > > int old_ref_mod; > > > > BUG_ON(existing->is_data != update->is_data); > > @@ -529,10 +532,18 @@ update_existing_head_ref(struct > > btrfs_delayed_ref_root *delayed_refs, > > * versa we need to make sure to adjust pending_csums accordingly. > > */ > > if (existing->is_data) { > > - if (existing->total_ref_mod >= 0 && old_ref_mod < 0) > > + u64 csum_items = > > + btrfs_csum_bytes_to_leaves(fs_info, > > + existing->num_bytes); > > + > > + if (existing->total_ref_mod >= 0 && old_ref_mod < 0) { > > delayed_refs->pending_csums -= existing->num_bytes; > > - if (existing->total_ref_mod < 0 && old_ref_mod >= 0) > > + btrfs_delayed_refs_rsv_release(fs_info, csum_items); > > + } > > + if (existing->total_ref_mod < 0 && old_ref_mod >= 0) { > > delayed_refs->pending_csums += existing->num_bytes; > > + trans->delayed_ref_updates += csum_items; > > + } > > } > > spin_unlock(&existing->lock); > > } > > @@ -638,7 +649,7 @@ add_delayed_ref_head(struct btrfs_trans_handle *trans, > > && head_ref->qgroup_reserved > > && existing->qgroup_ref_root > > && existing->qgroup_reserved); > > - update_existing_head_ref(delayed_refs, existing, head_ref, > > + update_existing_head_ref(trans, existing, head_ref, > > old_ref_mod); > > /* > > * we've updated the existing ref, free the newly > > @@ -649,8 +660,12 @@ add_delayed_ref_head(struct btrfs_trans_handle *trans, > > } else { > > if (old_ref_mod) > > *old_ref_mod = 0; > > - if (head_ref->is_data && head_ref->ref_mod < 0) > > + if (head_ref->is_data && head_ref->ref_mod < 0) { > > delayed_refs->pending_csums += head_ref->num_bytes; > > + trans->delayed_ref_updates += > > + btrfs_csum_bytes_to_leaves(trans->fs_info, > > + head_ref->num_bytes); > > + } > > delayed_refs->num_heads++; > > delayed_refs->num_heads_ready++; > > atomic_inc(&delayed_refs->num_entries); > > @@ -785,6 +800,7 @@ int btrfs_add_delayed_tree_ref(struct > > btrfs_trans_handle *trans, > > > > ret = insert_delayed_ref(trans, delayed_refs, head_ref, &ref->node); > > spin_unlock(&delayed_refs->lock); > > + btrfs_update_delayed_refs_rsv(trans); > > You haven't adressed my initial point about merging modification of > delayed_ref_updates and calling btrfs_update_delayed_refs_rsv into one > function otherwise this seems error prone. I don't see why this cannot > be made, if there is some reason which I'm missing then explain it. > > As it stands this btrfs_updated_delayed_refs_rsv is paired with the > modifications made in one of the 2nd level callees: > > btrfs_add_delayed_tree_ref > add_delayed_ref_head > update_existing_head_ref > > I'd rather have btrfs_update_delayed_refs_rsv renamed to something else > with 'inc' in its name and called everytime we modify > delayed_ref_update. I'm willing to bet 50 bucks in 6 months time someone > will change delayed_ref_updates and will forget to call > btrfs_update_delayed_refs_rsv. >
Because we have to take the delayed_refs_rsv lock in this helper, I want to take it as little as possible since it is a fs wide lock, so I want to batch it. There's no reason to change it. Thanks, Josef