Re: [PATCH 00/14] replace call_rcu by kfree_rcu for simple kmem_cache_free callback

[Uladzislau Rezki <urezki@gmail.com>]

On Fri, Jun 14, 2024 at 07:17:29AM -0700, Paul E. McKenney wrote:
> On Fri, Jun 14, 2024 at 02:35:33PM +0200, Uladzislau Rezki wrote:
> > On Thu, Jun 13, 2024 at 11:13:52AM -0700, Paul E. McKenney wrote:
> > > On Thu, Jun 13, 2024 at 07:58:17PM +0200, Uladzislau Rezki wrote:
> > > > On Thu, Jun 13, 2024 at 10:45:59AM -0700, Paul E. McKenney wrote:
> > > > > On Thu, Jun 13, 2024 at 07:38:59PM +0200, Uladzislau Rezki wrote:
> > > > > > On Thu, Jun 13, 2024 at 08:06:30AM -0700, Paul E. McKenney wrote:
> > > > > > > On Thu, Jun 13, 2024 at 03:06:54PM +0200, Uladzislau Rezki wrote:
> > > > > > > > On Thu, Jun 13, 2024 at 05:47:08AM -0700, Paul E. McKenney wrote:
> > > > > > > > > On Thu, Jun 13, 2024 at 01:58:59PM +0200, Jason A. Donenfeld wrote:
> > > > > > > > > > On Wed, Jun 12, 2024 at 03:37:55PM -0700, Paul E. McKenney wrote:
> > > > > > > > > > > On Wed, Jun 12, 2024 at 02:33:05PM -0700, Jakub Kicinski wrote:
> > > > > > > > > > > > On Sun,  9 Jun 2024 10:27:12 +0200 Julia Lawall wrote:
> > > > > > > > > > > > > Since SLOB was removed, it is not necessary to use call_rcu
> > > > > > > > > > > > > when the callback only performs kmem_cache_free. Use
> > > > > > > > > > > > > kfree_rcu() directly.
> > > > > > > > > > > > > 
> > > > > > > > > > > > > The changes were done using the following Coccinelle semantic patch.
> > > > > > > > > > > > > This semantic patch is designed to ignore cases where the callback
> > > > > > > > > > > > > function is used in another way.
> > > > > > > > > > > > 
> > > > > > > > > > > > How does the discussion on:
> > > > > > > > > > > >   [PATCH] Revert "batman-adv: prefer kfree_rcu() over call_rcu() with free-only callbacks"
> > > > > > > > > > > >   https://lore.kernel.org/all/20240612133357.2596-1-linus.luessing@c0d3.blue/
> > > > > > > > > > > > reflect on this series? IIUC we should hold off..
> > > > > > > > > > > 
> > > > > > > > > > > We do need to hold off for the ones in kernel modules (such as 07/14)
> > > > > > > > > > > where the kmem_cache is destroyed during module unload.
> > > > > > > > > > > 
> > > > > > > > > > > OK, I might as well go through them...
> > > > > > > > > > > 
> > > > > > > > > > > [PATCH 01/14] wireguard: allowedips: replace call_rcu by kfree_rcu for simple kmem_cache_free callback
> > > > > > > > > > > 	Needs to wait, see wg_allowedips_slab_uninit().
> > > > > > > > > > 
> > > > > > > > > > Also, notably, this patch needs additionally:
> > > > > > > > > > 
> > > > > > > > > > diff --git a/drivers/net/wireguard/allowedips.c b/drivers/net/wireguard/allowedips.c
> > > > > > > > > > index e4e1638fce1b..c95f6937c3f1 100644
> > > > > > > > > > --- a/drivers/net/wireguard/allowedips.c
> > > > > > > > > > +++ b/drivers/net/wireguard/allowedips.c
> > > > > > > > > > @@ -377,7 +377,6 @@ int __init wg_allowedips_slab_init(void)
> > > > > > > > > > 
> > > > > > > > > >  void wg_allowedips_slab_uninit(void)
> > > > > > > > > >  {
> > > > > > > > > > -	rcu_barrier();
> > > > > > > > > >  	kmem_cache_destroy(node_cache);
> > > > > > > > > >  }
> > > > > > > > > > 
> > > > > > > > > > Once kmem_cache_destroy has been fixed to be deferrable.
> > > > > > > > > > 
> > > > > > > > > > I assume the other patches are similar -- an rcu_barrier() can be
> > > > > > > > > > removed. So some manual meddling of these might be in order.
> > > > > > > > > 
> > > > > > > > > Assuming that the deferrable kmem_cache_destroy() is the option chosen,
> > > > > > > > > agreed.
> > > > > > > > >
> > > > > > > > <snip>
> > > > > > > > void kmem_cache_destroy(struct kmem_cache *s)
> > > > > > > > {
> > > > > > > > 	int err = -EBUSY;
> > > > > > > > 	bool rcu_set;
> > > > > > > > 
> > > > > > > > 	if (unlikely(!s) || !kasan_check_byte(s))
> > > > > > > > 		return;
> > > > > > > > 
> > > > > > > > 	cpus_read_lock();
> > > > > > > > 	mutex_lock(&slab_mutex);
> > > > > > > > 
> > > > > > > > 	rcu_set = s->flags & SLAB_TYPESAFE_BY_RCU;
> > > > > > > > 
> > > > > > > > 	s->refcount--;
> > > > > > > > 	if (s->refcount)
> > > > > > > > 		goto out_unlock;
> > > > > > > > 
> > > > > > > > 	err = shutdown_cache(s);
> > > > > > > > 	WARN(err, "%s %s: Slab cache still has objects when called from %pS",
> > > > > > > > 	     __func__, s->name, (void *)_RET_IP_);
> > > > > > > > ...
> > > > > > > > 	cpus_read_unlock();
> > > > > > > > 	if (!err && !rcu_set)
> > > > > > > > 		kmem_cache_release(s);
> > > > > > > > }
> > > > > > > > <snip>
> > > > > > > > 
> > > > > > > > so we have SLAB_TYPESAFE_BY_RCU flag that defers freeing slab-pages
> > > > > > > > and a cache by a grace period. Similar flag can be added, like
> > > > > > > > SLAB_DESTROY_ONCE_FULLY_FREED, in this case a worker rearm itself
> > > > > > > > if there are still objects which should be freed.
> > > > > > > > 
> > > > > > > > Any thoughts here?
> > > > > > > 
> > > > > > > Wouldn't we also need some additional code to later check for all objects
> > > > > > > being freed to the slab, whether or not that code is  initiated from
> > > > > > > kmem_cache_destroy()?
> > > > > > >
> > > > > > Same away as SLAB_TYPESAFE_BY_RCU is handled from the kmem_cache_destroy() function.
> > > > > > It checks that flag and if it is true and extra worker is scheduled to perform a
> > > > > > deferred(instead of right away) destroy after rcu_barrier() finishes.
> > > > > 
> > > > > Like this?
> > > > > 
> > > > > 	SLAB_DESTROY_ONCE_FULLY_FREED
> > > > > 
> > > > > 	Instead of adding a new kmem_cache_destroy_rcu()
> > > > > 	or kmem_cache_destroy_wait() API member, instead add a
> > > > > 	SLAB_DESTROY_ONCE_FULLY_FREED flag that can be passed to the
> > > > > 	existing kmem_cache_destroy() function.  Use of this flag would
> > > > > 	suppress any warnings that would otherwise be issued if there
> > > > > 	was still slab memory yet to be freed, and it would also spawn
> > > > > 	workqueues (or timers or whatever) to do any needed cleanup work.
> > > > > 
> > > > >
> > > > The flag is passed as all others during creating a cache:
> > > > 
> > > >   slab = kmem_cache_create(name, size, ..., SLAB_DESTROY_ONCE_FULLY_FREED | OTHER_FLAGS, NULL);
> > > > 
> > > > the rest description is correct to me.
> > > 
> > > Good catch, fixed, thank you!
> > > 
> > And here we go with prototype(untested):
> 
> Thank you for putting this together!  It looks way simpler than I would
> have guessed, and quite a bit simpler than I would expect it would be
> to extend rcu_barrier() to cover kfree_rcu().
> 
Yep, it should be pretty pretty straightforward. The slab mechanism does
not have a functionality when it comes to defer of destroying, i.e. it
is not allowed to destroy non-fully-freed-slab:

<snip>
void kmem_cache_destroy(struct kmem_cache *s)
{
...
	err = shutdown_cache(s);
	WARN(err, "%s %s: Slab cache still has objects when called from %pS",
	     __func__, s->name, (void *)_RET_IP_);
...
<snip>

So, this patch extends it.

> >  
> > +static void
> > +slab_caches_defer_destroy_workfn(struct work_struct *work)
> > +{
> > +	struct kmem_cache *s, *s2;
> > +
> > +	mutex_lock(&slab_mutex);
> > +	list_for_each_entry_safe(s, s2, &slab_caches_defer_destroy, list) {
> > +		if (__kmem_cache_empty(s)) {
> > +			/* free asan quarantined objects */
> > +			kasan_cache_shutdown(s);
> > +			(void) __kmem_cache_shutdown(s);
> > +
> > +			list_del(&s->list);
> > +
> > +			debugfs_slab_release(s);
> > +			kfence_shutdown_cache(s);
> > +			kmem_cache_release(s);
> > +		}
> 
> My guess is that there would want to be a splat if the slab stuck around
> for too long, but maybe that should instead be handled elsewhere or in
> some other way?  I must defer to you guys on that one.
> 
Probably yes.

--
Uladzislau Rezki