Re: [PATCH 20/24] rcu/tree: Make kvfree_rcu() tolerate any alignment

[Paul E. McKenney]

On Sun, May 03, 2020 at 08:24:37PM -0400, Joel Fernandes wrote:
> On Fri, May 01, 2020 at 04:00:52PM -0700, Paul E. McKenney wrote:
> > On Tue, Apr 28, 2020 at 10:58:59PM +0200, Uladzislau Rezki (Sony) wrote:
> > > From: "Joel Fernandes (Google)" <joel@xxxxxxxxxxxxxxxxx>
> > > 
> > > Handle cases where the the object being kvfree_rcu()'d is not aligned by
> > > 2-byte boundaries.
> > > 
> > > Signed-off-by: Uladzislau Rezki (Sony) <urezki@xxxxxxxxx>
> > > Signed-off-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
> > > ---
> > >  kernel/rcu/tree.c | 9 ++++++---
> > >  1 file changed, 6 insertions(+), 3 deletions(-)
> > > 
> > > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > > index 501cac02146d..649bad7ad0f0 100644
> > > --- a/kernel/rcu/tree.c
> > > +++ b/kernel/rcu/tree.c
> > > @@ -2877,6 +2877,9 @@ struct kvfree_rcu_bulk_data {
> > >  #define KVFREE_BULK_MAX_ENTR \
> > >  	((PAGE_SIZE - sizeof(struct kvfree_rcu_bulk_data)) / sizeof(void *))
> > >  
> > > +/* Encoding the offset of a fake rcu_head to indicate the head is a wrapper. */
> > > +#define RCU_HEADLESS_KFREE BIT(31)
> > 
> > Did I miss the check for freeing something larger than 2GB?  Or is this
> > impossible, even on systems with many terabytes of physical memory?
> > Even if it is currently impossible, what prevents it from suddenly
> > becoming all too possible at some random point in the future?  If you
> > think that this will never happen, please keep in mind that the first
> > time I heard "640K ought to be enough for anybody", it sounded eminently
> > reasonable to me.
> > 
> > Besides...
> > 
> > Isn't the offset in question the offset of an rcu_head struct within
> > the enclosing structure? If so, why not keep the current requirement
> > that this be at least 16-bit aligned, especially given that some work
> > is required to make that alignment less than pointer sized?  Then you
> > can continue using bit 0.
> > 
> > This alignment requirement is included in the RCU requirements
> > documentation and is enforced within the __call_rcu() function.
> > 
> > So let's leave this at bit 0.
> 
> This patch is needed only if we are growing the fake rcu_head. Since you
> mentioned in a previous patch in this series that you don't want to do that,
> and just rely on availability of the array of pointers or synchronize_rcu(),
> we can drop this patch. If we are not dropping that earlier patch, let us
> discuss more.

Dropping it sounds very good to me!

							Thanx, Paul