Re: [PATCH 01/12] kernfs: add function to find kernfs_node without increasing ref counter

From: Paolo Valente
Date: Tue Nov 13 2018 - 12:54:08 EST




> Il giorno 12 nov 2018, alle ore 13:28, Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx> ha scritto:
>
> On Mon, Nov 12, 2018 at 10:56:21AM +0100, Paolo Valente wrote:
>> From: Angelo Ruocco <angeloruocco90@xxxxxxxxx>
>>
>> The kernfs pseudo file system doesn't export any function to only find
>> a node by name, without also getting a reference on it.
>> But in some cases it is useful to just locate a kernfs node, while
>> using it or not depends on some other condition.
>>
>> This commit adds a function to just look for a node, without getting
>> a reference on it.
>
> Eeek, that sounds really bad. So you save off a pointer to something,
> and have no idea if that pointer now really is valid or not? It can
> instantly disappear right afterwards.
>

Hi Greg,
that function is invoked only in functions executed with cgroup_mutex
held. This guarantees that nothing disappears or becomes
inconsistent. That's why we decided to go for this optimization,
instead of doing useless gets&puts pairs. Still, I'm not expert
enough to state whether it is impossible that, once we have defined
that function, it may then get used in some unsafe way.

So, I seem to see two options:
1) Add a comment on the function, saying that cgroup_mutex must be
held while invoking it (I guess you won't like this one).
2) Do not define such a new function, and, in the other patches, use
the already-available find_and_get.

Looking forward to your feedback (or of other knowledgeable people on
this issue) before proceeding to a rebased V2,
Paolo


> This feels wrong, what is the problem of having a properly reference
> counted object passed back to you that you have to create a dangerous
> function like this?
>
> greg k-h