On Fri, Jan 11, 2019 at 10:21:03AM -0500, Josef Bacik wrote:
> The throttle path doesn't take cleaner_delayed_iput_mutex, which means
> we could think we're done flushing iputs in the data space reservation
> path when we could have a throttler doing an iput. There's no real
> reason to serialize the delayed iput flushing, so instead of taking the
> cleaner_delayed_iput_mutex whenever we flush the delayed iputs just
> replace it with an atomic counter and a waitqueue. This removes the
> short (or long depending on how big the inode is) window where we think
> there are no more pending iputs when there really are some.
>
> Signed-off-by: Josef Bacik <josef@xxxxxxxxxxxxxx>
> ---
> fs/btrfs/ctree.h | 4 +++-
> fs/btrfs/disk-io.c | 5 ++---
> fs/btrfs/extent-tree.c | 13 ++++++++-----
> fs/btrfs/inode.c | 21 +++++++++++++++++++++
> 4 files changed, 34 insertions(+), 9 deletions(-)
>
> diff --git a/fs/btrfs/ctree.h b/fs/btrfs/ctree.h
> index dc56a4d940c3..20af5d6d81f1 100644
> --- a/fs/btrfs/ctree.h
> +++ b/fs/btrfs/ctree.h
> @@ -915,7 +915,8 @@ struct btrfs_fs_info {
>
> spinlock_t delayed_iput_lock;
> struct list_head delayed_iputs;
> - struct mutex cleaner_delayed_iput_mutex;
> + atomic_t nr_delayed_iputs;
> + wait_queue_head_t delayed_iputs_wait;
>
> /* this protects tree_mod_seq_list */
> spinlock_t tree_mod_seq_lock;
> @@ -3240,6 +3241,7 @@ int btrfs_orphan_cleanup(struct btrfs_root *root);
> int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size);
> void btrfs_add_delayed_iput(struct inode *inode);
> void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info);
> +int btrfs_wait_on_delayed_iputs(struct btrfs_fs_info *fs_info);
> int btrfs_prealloc_file_range(struct inode *inode, int mode,
> u64 start, u64 num_bytes, u64 min_size,
> loff_t actual_len, u64 *alloc_hint);
> diff --git a/fs/btrfs/disk-io.c b/fs/btrfs/disk-io.c
> index f40f6fdc1019..238e0113f2d3 100644
> --- a/fs/btrfs/disk-io.c
> +++ b/fs/btrfs/disk-io.c
> @@ -1694,9 +1694,7 @@ static int cleaner_kthread(void *arg)
> goto sleep;
> }
>
> - mutex_lock(&fs_info->cleaner_delayed_iput_mutex);
> btrfs_run_delayed_iputs(fs_info);
> - mutex_unlock(&fs_info->cleaner_delayed_iput_mutex);
>
> again = btrfs_clean_one_deleted_snapshot(root);
> mutex_unlock(&fs_info->cleaner_mutex);
> @@ -2654,7 +2652,6 @@ int open_ctree(struct super_block *sb,
> mutex_init(&fs_info->delete_unused_bgs_mutex);
> mutex_init(&fs_info->reloc_mutex);
> mutex_init(&fs_info->delalloc_root_mutex);
> - mutex_init(&fs_info->cleaner_delayed_iput_mutex);
> seqlock_init(&fs_info->profiles_lock);
>
> INIT_LIST_HEAD(&fs_info->dirty_cowonly_roots);
> @@ -2676,6 +2673,7 @@ int open_ctree(struct super_block *sb,
> atomic_set(&fs_info->defrag_running, 0);
> atomic_set(&fs_info->qgroup_op_seq, 0);
> atomic_set(&fs_info->reada_works_cnt, 0);
> + atomic_set(&fs_info->nr_delayed_iputs, 0);
> atomic64_set(&fs_info->tree_mod_seq, 0);
> fs_info->sb = sb;
> fs_info->max_inline = BTRFS_DEFAULT_MAX_INLINE;
> @@ -2753,6 +2751,7 @@ int open_ctree(struct super_block *sb,
> init_waitqueue_head(&fs_info->transaction_wait);
> init_waitqueue_head(&fs_info->transaction_blocked_wait);
> init_waitqueue_head(&fs_info->async_submit_wait);
> + init_waitqueue_head(&fs_info->delayed_iputs_wait);
>
> INIT_LIST_HEAD(&fs_info->pinned_chunks);
>
> diff --git a/fs/btrfs/extent-tree.c b/fs/btrfs/extent-tree.c
> index 0127d272cd2a..5b6c9fc227ff 100644
> --- a/fs/btrfs/extent-tree.c
> +++ b/fs/btrfs/extent-tree.c
> @@ -4280,10 +4280,14 @@ int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
> /*
> * The cleaner kthread might still be doing iput
> * operations. Wait for it to finish so that
> - * more space is released.
> + * more space is released. We don't need to
> + * explicitly run the delayed iputs here because
> + * the commit_transaction would have woken up
> + * the cleaner.
> */
> - mutex_lock(&fs_info->cleaner_delayed_iput_mutex);
> - mutex_unlock(&fs_info->cleaner_delayed_iput_mutex);
> + ret = btrfs_wait_on_delayed_iputs(fs_info);
> + if (ret)
> + return ret;
> goto again;
> } else {
> btrfs_end_transaction(trans);
> @@ -4958,9 +4962,8 @@ static void flush_space(struct btrfs_fs_info *fs_info,
> * bunch of pinned space, so make sure we run the iputs before
> * we do our pinned bytes check below.
> */
> - mutex_lock(&fs_info->cleaner_delayed_iput_mutex);
> btrfs_run_delayed_iputs(fs_info);
> - mutex_unlock(&fs_info->cleaner_delayed_iput_mutex);
> + btrfs_wait_on_delayed_iputs(fs_info);
>
> ret = may_commit_transaction(fs_info, space_info);
> break;
> diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
> index 0b9f3e482cea..958e30c7c744 100644
> --- a/fs/btrfs/inode.c
> +++ b/fs/btrfs/inode.c
> @@ -3260,6 +3260,7 @@ void btrfs_add_delayed_iput(struct inode *inode)
> if (atomic_add_unless(&inode->i_count, -1, 1))
> return;
>
> + atomic_inc(&fs_info->nr_delayed_iputs);
> spin_lock(&fs_info->delayed_iput_lock);
> ASSERT(list_empty(&binode->delayed_iput));
> list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
> @@ -3280,11 +3281,31 @@ void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
> list_del_init(&inode->delayed_iput);
> spin_unlock(&fs_info->delayed_iput_lock);
> iput(&inode->vfs_inode);
> + if (atomic_dec_and_test(&fs_info->nr_delayed_iputs))
> + wake_up(&fs_info->delayed_iputs_wait);
> spin_lock(&fs_info->delayed_iput_lock);
> }
> spin_unlock(&fs_info->delayed_iput_lock);
> }
>
> +/**
> + * btrfs_wait_on_delayed_iputs - wait on the delayed iputs to be done running
> + * @fs_info - the fs_info for this fs
> + * @return - EINTR if we were killed, 0 if nothing's pending
> + *
> + * This will wait on any delayed iputs that are currently running with KILLABLE
> + * set. Once they are all done running we will return, unless we are killed in
> + * which case we return EINTR.
I still wonder from which context the 'killable' property is used.
btrfs_alloc_data_chunk_ondemand is indirectly inside zero range and
fallocate, there it makes sense. With flush_space is less obvious, so
this should be explained in the changelog what's expected.
> + */
> +int btrfs_wait_on_delayed_iputs(struct btrfs_fs_info *fs_info)
> +{
> + int ret = wait_event_killable(fs_info->delayed_iputs_wait,
> + atomic_read(&fs_info->nr_delayed_iputs) == 0);
> + if (ret)
> + return -EINTR;
> + return 0;
> +}
> +
> /*
> * This creates an orphan entry for the given inode in case something goes wrong
> * in the middle of an unlink.
> --
> 2.14.3
