Re: [PATCH 1/2] task_work: make FIFO task_work list

[Oleg Nesterov <oleg@redhat.com>]

On 03/14, liguang wrote:
>
> Signed-off-by: liguang <lig.fnst@cn.fujitsu.com>

Changelog please...

> ---
>  kernel/task_work.c |   15 +++------------
>  1 files changed, 3 insertions(+), 12 deletions(-)
> 
> diff --git a/kernel/task_work.c b/kernel/task_work.c
> index 65bd3c9..0bf4258 100644
> --- a/kernel/task_work.c
> +++ b/kernel/task_work.c
> @@ -13,11 +13,12 @@ task_work_add(struct task_struct *task, struct callback_head *work, bool notify)
>  		head = ACCESS_ONCE(task->task_works);
>  		if (unlikely(head == &work_exited))
>  			return -ESRCH;
> -		work->next = head;
> -	} while (cmpxchg(&task->task_works, head, work) != head);
> +		head = head->next;
> +	} while (cmpxchg(&head, NULL, work) == head);

I simply can't understand how this can work... The patch assumes
that head->next == NULL after head = head->next, why? And then
compares the result with head and succeeds if not equal.

Could you please explain how it was supposed to work? If nothing
else, Suppose we have task->task_works -> W1 -> W2 -> W3. How this
code can add W4 after W3?

And cmpxchg(&head) should be cmpxchg(&head->next)....



Anyway, whatever I missed this is racy.

	head = head->next;

nothing protects "head" after this. Say, it can be task_work_cancel'ed
and freed. So,

	 cmpxchg(&head, ...)

can modify the freed and reused memory.

Oleg.