Re: [PATCH mm-unstable v1 1/4] mm: don't hold css->refcnt during traversal

[Roman Gushchin]

On Thu, Jul 25, 2024 at 04:43:46PM -0400, Johannes Weiner wrote:
> On Wed, Jul 24, 2024 at 07:02:11PM +0000, Kinsey Ho wrote:
> > To obtain the pointer to the saved memcg position, mem_cgroup_iter()
> > currently holds css->refcnt during memcg traversal only to put
> > css->refcnt at the end of the routine. This isn't necessary as an
> > rcu_read_lock is already held throughout the function.
> > 
> > Remove css->refcnt usage during traversal by leveraging RCU.
> 
> Eh, I don't know about this.
> 
> RCU ensures that the css memory isn't freed.
> 
> The tryget ensures that the css is still alive and valid.
> 
> In this case, it just so happens that the sibling linkage is also rcu
> protected. But accessing random css members when the refcount is 0 is
> kind of sketchy. On the other hand, the refcount is guaranteed to be
> valid, and rcu + tryget is a common pattern.

I also spent quite some time thinking about potential bad consequences,
but _it seems_ to be safe (but I agree it feels dangerous).

> 
> What does this buy us? The tryget is cheap.

To be fair, tryget is not always cheap. Offline/dying cgroups have an atomic
operation there.