Re: [PATCH] workqueue: fix a workqueue kernel panic issue.

[Tejun Heo <tj@kernel.org>]

On Mon, Sep 22, 2014 at 03:40:55AM -0700, Yifan Zhang wrote:
> You can tell it is a bug when pwq = get_work_pwq() return NULL, and
> cpu_intensive = pwq->wq->flags use it w/o check.

A bug somewhere else.

> Normally get_work_pwq doesn't return NULL, but we had a bug in code
> which makes INIT_WORK(&work, do_work) is called in multi-thread. In
> some cases, work_struct is re-init just before get_work_pwq is
> called, it makes work_struct->data is invalid and thus causes the
> problem. It is indeed a bug of ourselves, and after fix it there is
> no such issue. But I wonder we still a NULL check before dereference
> pwq here anyway, since get_work_pwq may return NULL in some cases.

Do you realize how timing dependent that particular pattern of
breakage is?  If you're doing INIT_WORK() in racy way, there are many
places which can break in workqueue.  It's not that different from
random memory corruption.  It doesn't make any sense at all to add a
special case code for that in one particular place where this specific
incidence happens to trigger.  In general, don't do things like this
anywhere in the kernel.

Thanks.

-- 
tejun