Re: [PATCH bpf] bpf: roll back stream capacity when allocation fails

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Pre-existing issues:
- [High] The patch addresses a capacity leak in `bpf_stream_push_str()` when stream element allocation fails, but entirely misses the identical capacity leak in the staging logic (`bpf_stream_stage_printk()` and `bpf_stream_stage_commit()`).
- [High] If copy_to_user() fails during bpf_stream_read() after some elements have already been consumed, the function returns -EFAULT instead of the number of bytes successfully read, resulting in silent data loss.
--

--- Patch [1]: [PATCH bpf] bpf: roll back stream capacity when allocation fails ---
commit 06c82cbb65aa09c192205a6718b0fa398300f1ac
Author: Jianlin Shi <shijianlin11@foxmail.com>

bpf: roll back stream capacity when allocation fails

The patch fixes a stream capacity leak in bpf_stream_push_str().
If allocating a stream element fails, the previously charged capacity
is now rolled back to avoid permanent exhaustion.

> diff --git a/kernel/bpf/stream.c b/kernel/bpf/stream.c
> index be9ce98e94692..4b8a74b9162e1 100644
> --- a/kernel/bpf/stream.c
> +++ b/kernel/bpf/stream.c
> @@ -79,7 +79,14 @@ static int bpf_stream_push_str(struct bpf_stream *stream, const char *str, int l
>  {
>  	int ret = bpf_stream_consume_capacity(stream, len);
>  
> -	return ret ?: __bpf_stream_push_str(&stream->log, str, len);
> +	if (ret)
> +		return ret;
> +
> +	ret = __bpf_stream_push_str(&stream->log, str, len);
> +	if (ret)
> +		atomic_sub(len, &stream->capacity);
> +
> +	return ret;
>  }

[Severity: High]
This isn't a bug introduced by this patch, but it seems there's an identical
capacity leak in the staging logic that this patch doesn't address.

In bpf_stream_stage_printk(), ss->len is unconditionally incremented even if
allocation fails:

kernel/bpf/stream.c:bpf_stream_stage_printk() {
    ...
    ss->len += ret;
    /* Exclude NULL byte during push. */
    ret = __bpf_stream_push_str(&ss->log, buf->buf, ret);
    ...
}

Then in include/linux/bpf.h, the bpf_stream_stage() macro unconditionally
calls bpf_stream_stage_commit():

#define bpf_stream_stage(ss, prog, stream_id, expr)            \
    ({                                                     \
        bpf_stream_stage_init(&ss);                    \
        (expr);                                        \
        bpf_stream_stage_commit(&ss, prog, stream_id); \
        bpf_stream_stage_free(&ss);                    \
    })

Which then permanently charges the falsely inflated length against the capacity:

kernel/bpf/stream.c:bpf_stream_stage_commit() {
    ...
    ret = bpf_stream_consume_capacity(stream, ss->len);
    ...
}

Should the staging logic also be updated to prevent permanently leaking
stream capacity when memory allocation fails?

[Severity: High]
This is a pre-existing issue, but I noticed another problem related to how
stream errors are handled in bpf_stream_read() when copy_to_user() fails.

If some elements are successfully consumed and freed:

kernel/bpf/stream.c:bpf_stream_read() {
    ...
    bpf_stream_backlog_pop(stream);
    bpf_stream_release_capacity(stream, elem);
    bpf_stream_free_elem(elem);
    ...
}

But a subsequent copy_to_user() fails on an unmapped page, the function returns
-EFAULT unconditionally:

    ...
    mutex_unlock(&stream->lock);
    return ret ? ret : len - rem_len;
}

Does this mean the successfully copied elements from earlier loop iterations
are permanently lost without the reader knowing how many bytes were read?
Should it return the bytes successfully read (len - rem_len) when some
elements were already consumed?

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/tencent_E69EAE29327E25B3548A9AF3F4FA289A6806@qq.com?part=1