Re: [PATCH v2] kernel/resource.c: invalidate parent when freed resource has childs

From: Wei Yang <richard.weiyang@gmail.com>
To: "Schmid, Carsten" <Carsten_Schmid@mentor.com>
Cc: Wei Yang <richard.weiyang@gmail.com>,
	Linus Torvalds <torvalds@linux-foundation.org>,
	"bp@suse.de" <bp@suse.de>,
	"dan.j.williams@intel.com" <dan.j.williams@intel.com>,
	"mingo@kernel.org" <mingo@kernel.org>,
	"dave.hansen@linux.intel.com" <dave.hansen@linux.intel.com>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	"bhelgaas@google.com" <bhelgaas@google.com>,
	"osalvador@suse.de" <osalvador@suse.de>,
	"rdunlap@infradead.org" <rdunlap@infradead.org>,
	"richardw.yang@linux.intel.com" <richardw.yang@linux.intel.com>,
	"gregkh@linuxfoundation.org" <gregkh@linuxfoundation.org>
Subject: Re: [PATCH v2] kernel/resource.c: invalidate parent when freed resource has childs
Date: Thu, 15 Aug 2019 13:03:28 +0000	[thread overview]
Message-ID: <20190815130328.yk4cybuuqnzb7xrx@master> (raw)
In-Reply-To: <c925c7d1041f478c99863da56c24b8a7@SVR-IES-MBX-03.mgc.mentorg.com>

On Thu, Aug 15, 2019 at 08:18:06AM +0000, Schmid, Carsten wrote:
>>>When a resource is freed and has children, the childrens are
>> 
>> s/childrens/children/
>>
>oh, missed that. Too many children ... ;-)
> 
>>>+		__release_child_resources(tmp, warn);
>> 
>> This function will release all the children.
>> 
>> Is this what Linus suggest?
>> 
>> From his code snippet, I just see siblings parent is set to NULL. I may miss
>> some point?
>>
>At the point we are here, there should be no children, and children of
>children at all ...
>So they are all more or less lost in the wild.
>That was why i didn't copy Linus' code 1:1 but reused an already existing
>function doing similar thing.
>It's anyway worth of thinking about this.
>
>What i have in mind here (example):
>Parent: iomem map 0x1000..0x1FFF
>  Child1: iomem map 0x1000..0x17FF
>    Child11: iomem map 0x1000..0x13FF
>    Child12: iomem map 0x1400..0x17FF
>  Child2: iomem map 0x1800..0x1FFF
>    Child21: iomem map 0x1800..0x1BFF
>    Child22: iomem map 0x1C00..0x1FFF
>
>When releasing the parent, how can children 11, 12, 21 and 22 still be valid?
>They don't know about their grandfather died ...
>Looking at the __release_child_resources, i exactly found that all children are
>invalidated/released in the way Linus did for the parent's children list.
>Doesn't it make sense to do the same for all?
>
>Please comment.
>
>> >+static void check_children(struct resource *parent)
>> >+{
>> >+	if (parent->child) {
>> >+		/* warn and release all children */
>> >+		WARN_ONCE(1, "%s: %s has child %s, release all children\n",
>> >+				__func__, parent->name, parent->child-
>> >name);
>> >+		write_lock(&resource_lock);
>> 
>> In previous version, lock is grasped before parent->child is checked.
>> 
>> Not sure why you change the order?
>> 
>To hold the lock as short as possible.
>But yes, you are right, this could lead to problems if releasing of the
>children is done in a parallel thread on a multicore ...
>I'll change that to cover the whole resource access within the lock.
>Not a big thing ...
>

My gut feeling is this is the problem from mal-functional driver, e.g.
xhci-hcd. We do our best to protect core kernel from it instead of do the
cleanup for it.

So my suggestion is to look into why xhci-hcd behave like this and fix that.

>Best regards
>Carsten

-- 
Wei Yang
Help you, Help me