Re: [PATCH 8/8] btrfs: reserve extra space during evict()

[Nikolay Borisov <nborisov@suse.com>]

On 3.12.18 г. 17:24 ч., Josef Bacik wrote:
> We could generate a lot of delayed refs in evict but never have any left
> over space from our block rsv to make up for that fact.  So reserve some
> extra space and give it to the transaction so it can be used to refill
> the delayed refs rsv every loop through the truncate path.
> 
> Signed-off-by: Josef Bacik <josef@toxicpanda.com>
> ---
>  fs/btrfs/inode.c | 25 +++++++++++++++++++++++--
>  1 file changed, 23 insertions(+), 2 deletions(-)
> 
> diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
> index 623a71d871d4..8ac7abe2ae9b 100644
> --- a/fs/btrfs/inode.c
> +++ b/fs/btrfs/inode.c
> @@ -5258,13 +5258,15 @@ static struct btrfs_trans_handle *evict_refill_and_join(struct btrfs_root *root,
>  {
>  	struct btrfs_fs_info *fs_info = root->fs_info;
>  	struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
> +	u64 delayed_refs_extra = btrfs_calc_trans_metadata_size(fs_info, 1);
>  	int failures = 0;
>  
>  	for (;;) {
>  		struct btrfs_trans_handle *trans;
>  		int ret;
>  
> -		ret = btrfs_block_rsv_refill(root, rsv, rsv->size,
> +		ret = btrfs_block_rsv_refill(root, rsv,
> +					     rsv->size + delayed_refs_extra,
>  					     BTRFS_RESERVE_FLUSH_LIMIT);

Rather than having to play those tricks, why not just modify the call in
btrfs_evict_inode, from:

rsv->size = btrfs_calc_trunc_metadata_size(fs_info, 1);

to

rsv->size = btrfs_calc_trunc_metadata_size(fs_info, 2);

and add a comment what the number 2 means of course.

>  
>  		if (ret && ++failures > 2) {
> @@ -5273,9 +5275,28 @@ static struct btrfs_trans_handle *evict_refill_and_join(struct btrfs_root *root,
>  			return ERR_PTR(-ENOSPC);
>  		}
>  
> +		/*
> +		 * Evict can generate a large amount of delayed refs without
> +		 * having a way to add space back since we exhaust our temporary
> +		 * block rsv.  We aren't allowed to do FLUSH_ALL in this case
> +		 * because we could deadlock with so many things in the flushing
> +		 * code, so we have to try and hold some extra space to
> +		 * compensate for our delayed ref generation.  If we can't get
> +		 * that space then we need see if we can steal our minimum from
> +		 * the global reserve.  We will be ratelimited by the amount of
> +		 * space we have for the delayed refs rsv, so we'll end up
> +		 * committing and trying again.
> +		 */
>  		trans = btrfs_join_transaction(root);
> -		if (IS_ERR(trans) || !ret)
> +		if (IS_ERR(trans) || !ret) {
> +			if (!IS_ERR(trans)) {
> +				trans->block_rsv = &fs_info->trans_block_rsv;

This line is redundant since evict_refill_and_join is called before the
trans->block_rsv  is modified.

> +				trans->bytes_reserved = delayed_refs_extra;

Is this even correct, since we join a transaction it might have already
had some bytes reserved. So in anycase shouldn't the line here say:
trans->bytes_reserved += delayed_refs_extra ?

> +				btrfs_block_rsv_migrate(rsv, trans->block_rsv,
> +							delayed_refs_extra, 1);



> +			}
>  			return trans;
> +		}
>  
>  		/*
>  		 * Try to steal from the global reserve if there is space for
>