Re: [PATCH] btrfs: Fix data race in log_conflicting_inodes

[Qu Wenruo <quwenruo.btrfs@gmx.com>]

在 2024/11/1 14:21, Hao-ran Zheng 写道:
> The Data Race occurs when the `log_conflicting_inodes()` function is
> executed in different threads at the same time. When one thread assigns
> a value to `ctx->logging_conflict_inodes` while another thread performs
> an `if(ctx->logging_conflict_inodes)` judgment or modifies it at the
> same time, a data contention problem may arise.
>
> Further, an atomicity violation may also occur here. Consider the
> following case, when a thread A `if(ctx->logging_conflict_inodes)`
> passes the judgment, the execution switches to another thread B, at
> which time the value of `ctx->logging_conflict_inodes` has not yet
> been assigned true, which would result in multiple threads executing
> `log_conflicting_inodes()`.
>
> To address this issue, it is recommended to add locks to protect
> `logging_conflict_inodes` in the `btrfs_log_ctx` structure, and lock
> protection during assignment and judgment. This modification ensures
> that the value of `ctx->logging_conflict_inodes` does not change during
> the validation process, thereby maintaining its integrity.
>
> Signed-off-by: Hao-ran Zheng <zhenghaoran@buaa.edu.cn>
> ---
>   fs/btrfs/tree-log.c | 7 +++++++
>   fs/btrfs/tree-log.h | 1 +
>   2 files changed, 8 insertions(+)
>
> diff --git a/fs/btrfs/tree-log.c b/fs/btrfs/tree-log.c
> index 9637c7cdc0cf..9cdbf280ca9a 100644
> --- a/fs/btrfs/tree-log.c
> +++ b/fs/btrfs/tree-log.c
> @@ -2854,6 +2854,7 @@ void btrfs_init_log_ctx(struct btrfs_log_ctx *ctx, struct btrfs_inode *inode)
>   	INIT_LIST_HEAD(&ctx->conflict_inodes);
>   	ctx->num_conflict_inodes = 0;
>   	ctx->logging_conflict_inodes = false;
> +	spin_lock_init(&ctx->logging_conflict_inodes_lock);
>   	ctx->scratch_eb = NULL;
>   }
>
> @@ -5779,16 +5780,20 @@ static int log_conflicting_inodes(struct btrfs_trans_handle *trans,
>   				  struct btrfs_log_ctx *ctx)
>   {
>   	int ret = 0;
> +	unsigned long logging_conflict_inodes_flags;
>
>   	/*
>   	 * Conflicting inodes are logged by the first call to btrfs_log_inode(),
>   	 * otherwise we could have unbounded recursion of btrfs_log_inode()
>   	 * calls. This check guarantees we can have only 1 level of recursion.
>   	 */
> +	spin_lock_irqsave(&ctx->conflict_inodes_lock, logging_conflict_inodes_flags);
>   	if (ctx->logging_conflict_inodes)
> +		spin_unlock_irqrestore(&ctx->conflict_inodes_lock, logging_conflict_inodes_flags);

Not an expert on the log tree, but in the above case, the only thing the
spinlock is protecting is a bool.

This looks overkilled to me.

Yes, several booleans can be stored in to a single byte, which can cause
problems.

But in that case, why not changing those booleans into a unsigned long
and use test_bit()/set_bit()/clear_bit() so that the bit operation will
be atomic and no need for the extra spinlock.

Although I haven't check the other boolean usage, but at least for this
@logging_conflict_inodes variable, it looks like atomic bit operation is
safe.

Thanks,
Qu


>   		return 0;
>
>   	ctx->logging_conflict_inodes = true;
> +	spin_unlock_irqrestore(&ctx->conflict_inodes_lock, logging_conflict_inodes_flags);
>
>   	/*
>   	 * New conflicting inodes may be found and added to the list while we
> @@ -5869,7 +5874,9 @@ static int log_conflicting_inodes(struct btrfs_trans_handle *trans,
>   			break;
>   	}
>
> +	spin_lock_irqsave(&ctx->conflict_inodes_lock, logging_conflict_inodes_flags);
>   	ctx->logging_conflict_inodes = false;
> +	spin_unlock_irqrestore(&ctx->conflict_inodes_lock, logging_conflict_inodes_flags);
>   	if (ret)
>   		free_conflicting_inodes(ctx);
>
> diff --git a/fs/btrfs/tree-log.h b/fs/btrfs/tree-log.h
> index dc313e6bb2fa..0f862d0c80f2 100644
> --- a/fs/btrfs/tree-log.h
> +++ b/fs/btrfs/tree-log.h
> @@ -44,6 +44,7 @@ struct btrfs_log_ctx {
>   	struct list_head conflict_inodes;
>   	int num_conflict_inodes;
>   	bool logging_conflict_inodes;
> +	spinlock_t logging_conflict_inodes_lock;
>   	/*
>   	 * Used for fsyncs that need to copy items from the subvolume tree to
>   	 * the log tree (full sync flag set or copy everything flag set) to