Re: [PATCH 0/2] rcu: Transform kfree_rcu() into kvfree_rcu()

[Paul E. McKenney]

On Wed, Feb 07, 2018 at 10:57:28AM +0300, Kirill Tkhai wrote:
> On 07.02.2018 08:02, Paul E. McKenney wrote:
> > On Tue, Feb 06, 2018 at 08:23:34PM -0800, Matthew Wilcox wrote:
> >> On Tue, Feb 06, 2018 at 06:17:03PM -0800, Paul E. McKenney wrote:
> >>> So it is OK to kvmalloc() something and pass it to either kfree() or
> >>> kvfree(), and it had better be OK to kvmalloc() something and pass it
> >>> to kvfree().
> >>>
> >>> Is it OK to kmalloc() something and pass it to kvfree()?
> >>
> >> Yes, it absolutely is.
> >>
> >> void kvfree(const void *addr)
> >> {
> >>         if (is_vmalloc_addr(addr))
> >>                 vfree(addr);
> >>         else
> >>                 kfree(addr);
> >> }
> >>
> >>> If so, is it really useful to have two different names here, that is,
> >>> both kfree_rcu() and kvfree_rcu()?
> >>
> >> I think it's handy to have all three of kvfree_rcu(), kfree_rcu() and
> >> vfree_rcu() available in the API for the symmetry of calling kmalloc()
> >> / kfree_rcu().
> >>
> >> Personally, I would like us to rename kvfree() to just free(), and have
> >> malloc(x) be an alias to kvmalloc(x, GFP_KERNEL), but I haven't won that
> >> fight yet.
> > 
> > But why not just have the existing kfree_rcu() API cover both kmalloc()
> > and kvmalloc()?  Perhaps I am not in the right forums, but I am not hearing
> > anyone arguing that the RCU API has too few members.  ;-)
> 
> People, far from RCU internals, consider kfree_rcu() like an extension
> of kfree(). And it's not clear it's need to dive into kfree_rcu() comments,
> when someone is looking a primitive to free vmalloc'ed memory.

Seems like a relatively simple lesson to teach.

> Also, construction like
> 
> obj = kvmalloc();
> kfree_rcu(obj);
> 
> makes me think it's legitimately to use plain kfree() as pair bracket to kvmalloc().

So it all works as is, then.

> So the significant change of kfree_rcu() behavior will complicate stable backporters
> life, because they will need to keep in mind such differences between different
> kernel versions.

If I understood your construction above, that significant change in
kfree_rcu() behavior has already happened.

> It seems if we are going to use the single primitive for both kmalloc()
> and kvmalloc() memory, it has to have another name. But I don't see problems
> with having both kfree_rcu() and kvfree_rcu().

I see problems.  We would then have two different names for exactly the
same thing.

Seems like it would be a lot easier to simply document the existing
kfree_rcu() behavior, especially given that it apparently already works.
The really doesn't seem to me to be worth a name change.

							Thanx, Paul