Re: [PATCH 19/24] rcu/tree: Support reclaim for head-less object

[Paul E. McKenney]

On Sun, May 03, 2020 at 08:12:58PM -0400, Joel Fernandes wrote:
> On Fri, May 01, 2020 at 03:39:09PM -0700, Paul E. McKenney wrote:
> > On Tue, Apr 28, 2020 at 10:58:58PM +0200, Uladzislau Rezki (Sony) wrote:
> > > Update the kvfree_call_rcu() with head-less support, it
> > > means an object without any rcu_head structure can be
> > > reclaimed after GP.
> > > 
> > > To store pointers there are two chain-arrays maintained
> > > one for SLAB and another one is for vmalloc. Both types
> > > of objects(head-less variant and regular one) are placed
> > > there based on the type.
> > > 
> > > It can be that maintaining of arrays becomes impossible
> > > due to high memory pressure. For such reason there is an
> > > emergency path. In that case objects with rcu_head inside
> > > are just queued building one way list. Later on that list
> > > is drained.
> > > 
> > > As for head-less variant. Such objects do not have any
> > > rcu_head helper inside. Thus it is dynamically attached.
> > > As a result an object consists of back-pointer and regular
> > > rcu_head. It implies that emergency path can detect such
> > > object type, therefore they are tagged. So a back-pointer
> > > could be freed as well as dynamically attached wrapper.
> > > 
> > > Even though such approach requires dynamic memory it needs
> > > only sizeof(unsigned long *) + sizeof(struct rcu_head) bytes,
> > > thus SLAB is used to obtain it. Finally if attaching of the
> > > rcu_head and queuing get failed, the current context has
> > > to follow might_sleep() annotation, thus below steps could
> > > be applied:
> > >    a) wait until a grace period has elapsed;
> > >    b) direct inlining of the kvfree() call.
> > > 
> > > Reviewed-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
> > > Signed-off-by: Uladzislau Rezki (Sony) <urezki@xxxxxxxxx>
> > > Signed-off-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
> > > Co-developed-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
> > > ---
> > >  kernel/rcu/tree.c | 102 ++++++++++++++++++++++++++++++++++++++++++++--
> > >  1 file changed, 98 insertions(+), 4 deletions(-)
> > > 
> > > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > > index 51726e4c3b4d..501cac02146d 100644
> > > --- a/kernel/rcu/tree.c
> > > +++ b/kernel/rcu/tree.c
> > > @@ -3072,15 +3072,31 @@ static void kfree_rcu_work(struct work_struct *work)
> > >  	 */
> > >  	for (; head; head = next) {
> > >  		unsigned long offset = (unsigned long)head->func;
> > > -		void *ptr = (void *)head - offset;
> > > +		bool headless;
> > > +		void *ptr;
> > >  
> > >  		next = head->next;
> > > +
> > > +		/* We tag the headless object, if so adjust offset. */
> > > +		headless = (((unsigned long) head - offset) & BIT(0));
> > > +		if (headless)
> > > +			offset -= 1;
> > > +
> > > +		ptr = (void *) head - offset;
> > > +
> > >  		debug_rcu_head_unqueue((struct rcu_head *)ptr);
> > >  		rcu_lock_acquire(&rcu_callback_map);
> > >  		trace_rcu_invoke_kvfree_callback(rcu_state.name, head, offset);
> > >  
> > > -		if (!WARN_ON_ONCE(!__is_kvfree_rcu_offset(offset)))
> > > +		if (!WARN_ON_ONCE(!__is_kvfree_rcu_offset(offset))) {
> > > +			/*
> > > +			 * If headless free the back-pointer first.
> > > +			 */
> > > +			if (headless)
> > > +				kvfree((void *) *((unsigned long *) ptr));
> > > +
> > >  			kvfree(ptr);
> > > +		}
> > >  
> > >  		rcu_lock_release(&rcu_callback_map);
> > >  		cond_resched_tasks_rcu_qs();
> > > @@ -3221,6 +3237,13 @@ kvfree_call_rcu_add_ptr_to_bulk(struct kfree_rcu_cpu *krcp, void *ptr)
> > >  			if (IS_ENABLED(CONFIG_PREEMPT_RT))
> > >  				return false;
> > >  
> > > +			/*
> > > +			 * TODO: For one argument of kvfree_rcu() we can
> > > +			 * drop the lock and get the page in sleepable
> > > +			 * context. That would allow to maintain an array
> > > +			 * for the CONFIG_PREEMPT_RT as well. Thus we could
> > > +			 * get rid of dynamic rcu_head attaching code.
> > > +			 */
> > >  			bnode = (struct kvfree_rcu_bulk_data *)
> > >  				__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
> > >  		}
> > > @@ -3244,6 +3267,23 @@ kvfree_call_rcu_add_ptr_to_bulk(struct kfree_rcu_cpu *krcp, void *ptr)
> > >  	return true;
> > >  }
> > >  
> > > +static inline struct rcu_head *
> > > +attach_rcu_head_to_object(void *obj)
> > > +{
> > > +	unsigned long *ptr;
> > > +
> > > +	ptr = kmalloc(sizeof(unsigned long *) +
> > > +			sizeof(struct rcu_head), GFP_NOWAIT |
> > > +				__GFP_RECLAIM |	/* can do direct reclaim. */
> > > +				__GFP_NORETRY |	/* only lightweight one.  */
> > > +				__GFP_NOWARN);	/* no failure reports. */
> > 
> > Again, let's please not do this single-pointer-sized allocation.  If
> > a full page is not available and this is a single-argument kfree_rcu(),
> > just call synchronize_rcu() and then free the object directly.
> 
> With the additional caching, lack of full page should not be very likely. I
> agree we can avoid doing any allocation and just straight to
> synchroize_rcu().

That sounds good to me!

> > It should not be -that- hard to adjust locking for CONFIG_PREEMPT_RT!
> > For example, have some kind of reservation protocol so that a task
> > that drops the lock can retry the page allocation and be sure of having
> > a place to put it.  This might entail making CONFIG_PREEMPT_RT reserve
> > more pages per CPU.  Or maybe that would not be necessary.
> 
> If we are not doing single-pointer allocation, then that would also eliminate
> entering the low-level page allocator for single-pointer allocations.
> 
> Or did you mean entry into the allocator for the full-page allocations
> related to the pointer array for PREEMPT_RT? Even if we skip entry into the
> allocator for those, we will still have additional caching which further
> reduces chances of getting a full page. In the event of such failure, we can
> simply queue the rcu_head.
> 
> Thoughts?

I was just trying to guess why you kept the single-pointer allocation.
It looks like I guessed wrong.  ;-)

If, as you say above, you make it go straight to synchronize_rcu()
upon full-page allocation failure, that would be good!

							Thanx, Paul

> thanks,
> 
>  - Joel
> 
> > 
> > 							Thanx, Paul
> > 
> > > +	if (!ptr)
> > > +		return NULL;
> > > +
> > > +	ptr[0] = (unsigned long) obj;
> > > +	return ((struct rcu_head *) ++ptr);
> > > +}
> > > +
> > >  /*
> > >   * Queue a request for lazy invocation of appropriate free routine after a
> > >   * grace period. Please note there are three paths are maintained, two are the
> > > @@ -3260,16 +3300,34 @@ void kvfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
> > >  {
> > >  	unsigned long flags;
> > >  	struct kfree_rcu_cpu *krcp;
> > > +	bool success;
> > >  	void *ptr;
> > >  
> > > +	if (head) {
> > > +		ptr = (void *) head - (unsigned long) func;
> > > +	} else {
> > > +		/*
> > > +		 * Please note there is a limitation for the head-less
> > > +		 * variant, that is why there is a clear rule for such
> > > +		 * objects:
> > > +		 *
> > > +		 * it can be used from might_sleep() context only. For
> > > +		 * other places please embed an rcu_head to your data.
> > > +		 */
> > > +		might_sleep();
> > > +		ptr = (unsigned long *) func;
> > > +	}
> > > +
> > >  	krcp = krc_this_cpu_lock(&flags);
> > > -	ptr = (void *)head - (unsigned long)func;
> > >  
> > >  	/* Queue the object but don't yet schedule the batch. */
> > >  	if (debug_rcu_head_queue(ptr)) {
> > >  		/* Probable double kfree_rcu(), just leak. */
> > >  		WARN_ONCE(1, "%s(): Double-freed call. rcu_head %p\n",
> > >  			  __func__, head);
> > > +
> > > +		/* Mark as success and leave. */
> > > +		success = true;
> > >  		goto unlock_return;
> > >  	}
> > >  
> > > @@ -3277,10 +3335,34 @@ void kvfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
> > >  	 * Under high memory pressure GFP_NOWAIT can fail,
> > >  	 * in that case the emergency path is maintained.
> > >  	 */
> > > -	if (unlikely(!kvfree_call_rcu_add_ptr_to_bulk(krcp, ptr))) {
> > > +	success = kvfree_call_rcu_add_ptr_to_bulk(krcp, ptr);
> > > +	if (!success) {
> > > +		if (head == NULL) {
> > > +			/*
> > > +			 * Headless(one argument kvfree_rcu()) can sleep.
> > > +			 * Drop the lock and tack it back. So it can do
> > > +			 * direct lightweight reclaim.
> > > +			 */
> > > +			krc_this_cpu_unlock(krcp, flags);
> > > +			head = attach_rcu_head_to_object(ptr);
> > > +			krcp = krc_this_cpu_lock(&flags);
> > > +
> > > +			if (head == NULL)
> > > +				goto unlock_return;
> > > +
> > > +			/*
> > > +			 * Tag the headless object. Such objects have a
> > > +			 * back-pointer to the original allocated memory,
> > > +			 * that has to be freed as well as dynamically
> > > +			 * attached wrapper/head.
> > > +			 */
> > > +			func = (rcu_callback_t) (sizeof(unsigned long *) + 1);
> > > +		}
> > > +
> > >  		head->func = func;
> > >  		head->next = krcp->head;
> > >  		krcp->head = head;
> > > +		success = true;
> > >  	}
> > >  
> > >  	WRITE_ONCE(krcp->count, krcp->count + 1);
> > > @@ -3294,6 +3376,18 @@ void kvfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
> > >  
> > >  unlock_return:
> > >  	krc_this_cpu_unlock(krcp, flags);
> > > +
> > > +	/*
> > > +	 * High memory pressure, so inline kvfree() after
> > > +	 * synchronize_rcu(). We can do it from might_sleep()
> > > +	 * context only, so the current CPU can pass the QS
> > > +	 * state.
> > > +	 */
> > > +	if (!success) {
> > > +		debug_rcu_head_unqueue(ptr);
> > > +		synchronize_rcu();
> > > +		kvfree(ptr);
> > > +	}
> > >  }
> > >  EXPORT_SYMBOL_GPL(kvfree_call_rcu);
> > >  
> > > -- 
> > > 2.20.1
> > >