Re: Resend [PATCH] kernel/resource.c: invalidate parent when freed resource has childs

[Linus Torvalds]

On Fri, Aug 9, 2019 at 6:50 AM Schmid, Carsten
<Carsten_Schmid@xxxxxxxxxx> wrote:
>
> @@ -1200,6 +1200,15 @@ void __release_region(struct resource *parent, resource_size_t start,
>                         write_unlock(&resource_lock);
>                         if (res->flags & IORESOURCE_MUXED)
>                                 wake_up(&muxed_resource_wait);
> +
> +                       write_lock(&resource_lock);
> +                       if (res->child) {
> +                               printk(KERN_WARNING "__release_region: %s has child %s,"
> +                                               "invalidating childs parent\n",
> +                                               res->name, res->child->name);
> +                               res->child->parent = NULL;
> +                       }
> +                       write_unlock(&resource_lock);
>                         free_resource(res);

So I think that this should be inside the previous resource_lock, and
before the whole "wake up muxed resource".

Also, a few other issues:

 - what about other freeing cases?  I'm looking at

        release_mem_region_adjustable()

   which has the same pattern where a resource may be freed.

 - what about multiple children? Your patch sets res->child->parent to
NULL, but what about possible other children (iow, the
res->child->sibling list)

 - releasing a resource without having released its children is a
nasty bug, but the bug is now here in release_region, it is in the
*caller*. The printk() (or pr_warn()) doesn't really help find that.

So my gut feel is that this patch is a symptom of a real bug, and a
warning is worthwhile to fix that bug, but more thought is needed.

Maybe something more along the line of

    diff --git a/kernel/resource.c b/kernel/resource.c
    index 7ea4306503c5..ebe06d77b06a 100644
    --- a/kernel/resource.c
    +++ b/kernel/resource.c
    @@ -1211,6 +1211,8 @@ void __release_region(struct resource
*parent, resource_size_t start,
                        }
                        if (res->start != start || res->end != end)
                                break;
    +                   if (WARN_ON_ONCE(res->child))
    +                           break;
                        *p = res->sibling;
                        write_unlock(&resource_lock);
                        if (res->flags & IORESOURCE_MUXED)

would be more appropriate? It simply refuses to free a resource that
has children, and gives a warning (with a backtrace) for the situation
(since clearly we now end up with a resource leak).

                Linus