Re: [PATCH v3] rcu/kvfree: Prevents cache growing when the backoff_page_cache_fill is set

[Uladzislau Rezki <urezki@gmail.com>]

On Wed, Apr 12, 2023 at 09:14:15AM +0000, Zhang, Qiang1 wrote:
> 
> > On Tue, Apr 11, 2023 at 04:58:22PM +0200, Uladzislau Rezki wrote:
> > > On Tue, Apr 11, 2023 at 02:42:27PM +0000, Zhang, Qiang1 wrote:
> > > > > > Currently, in kfree_rcu_shrink_scan(), the drain_page_cache() 
> > > > > > is executed before kfree_rcu_monitor() to drain page cache, if 
> > > > > > the bnode structure's->gp_snap has done, the kvfree_rcu_bulk() 
> > > > > > will fill the page cache again in kfree_rcu_monitor(), this 
> > > > > > commit add a check for krcp 
> > > > > > structure's->backoff_page_cache_fill in put_cached_bnode(), if 
> > > > > > the krcp structure's->backoff_page_cache_fill is set, prevent page cache growing and disable allocated page in fill_page_cache_func().
> > > > > > 
> > > > > > Signed-off-by: Zqiang <qiang1.zhang@intel.com>
> > > > > >
> > > > > >Much improved!  But still some questions below...
> > > > > >
> > > > > >							Thanx, Paul
> > > > > >
> > > > > > ---
> > > > > >  kernel/rcu/tree.c | 4 +++-
> > > > > >  1 file changed, 3 insertions(+), 1 deletion(-)
> > > > > > 
> > > > > > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c index 
> > > > > > cc34d13be181..9d9d3772cc45 100644
> > > > > > --- a/kernel/rcu/tree.c
> > > > > > +++ b/kernel/rcu/tree.c
> > > > > > @@ -2908,6 +2908,8 @@ static inline bool  
> > > > > > put_cached_bnode(struct kfree_rcu_cpu *krcp,
> > > > > >  	struct kvfree_rcu_bulk_data *bnode)  {
> > > > > > +	if (atomic_read(&krcp->backoff_page_cache_fill))
> > > > > > +		return false;
> > > > > >
> > > > > >This will mean that under low-memory conditions, we will keep 
> > > > > >zero pages in ->bkvcache.  All attempts to put something there will fail.
> > > > > >
> > > > > >This is probably not an issue for structures containing an 
> > > > > >rcu_head that are passed to kfree_rcu(p, field), but doesn't 
> > > > > >this mean that
> > > > > >kfree_rcu_mightsleep() unconditionally invokes synchronize_rcu()?
> > > > > >This could seriously slow up freeing under low-memory 
> > > > > >conditions, which might exacerbate the low-memory conditions.
> > > > > 
> > > > > Thanks for mentioning this, I didn't think of this before😊.
> > > > > 
> > > > > >
> > > > > >Is this really what we want?  Zero cached rather than just fewer cached?
> > > > > >
> > > > > >
> > > > > >
> > > > > >  	// Check the limit.
> > > > > >  	if (krcp->nr_bkv_objs >= rcu_min_cached_objs)
> > > > > >  		return false;
> > > > > > @@ -3221,7 +3223,7 @@ static void fill_page_cache_func(struct work_struct *work)
> > > > > >  	int i;
> > > > > >  
> > > > > >  	nr_pages = atomic_read(&krcp->backoff_page_cache_fill) ?
> > > > > > -		1 : rcu_min_cached_objs;
> > > > > > +		0 : rcu_min_cached_objs;
> > > > > >  
> > > > > >  	for (i = 0; i < nr_pages; i++) {
> > > > > >
> > > > > >I am still confused as to why we start "i" at zero rather than 
> > > > > >at
> > > > > >->nr_bkv_objs.  What am I missing here?
> > > > > 
> > > > > 
> > > > > No, you are right, I missed this place. 
> > > > > 
> > > > > --- a/kernel/rcu/tree.c
> > > > > +++ b/kernel/rcu/tree.c
> > > > > @@ -2908,6 +2908,8 @@ static inline bool  
> > > > > put_cached_bnode(struct kfree_rcu_cpu *krcp,
> > > > >         struct kvfree_rcu_bulk_data *bnode)  {
> > > > > +       if (atomic_read(&krcp->backoff_page_cache_fill))
> > > > > +               return false;
> > > > >
> > > > >This is broken, unfortunately. If a low memory condition we fill 
> > > > >fill a cache with at least one page anyway because of we do not 
> > > > >want to hit a slow path.
> > > > 
> > > > Thanks remind, please ignore my v4 patch,   how about the following?
> > > > 
> > > > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c index 
> > > > 41daae3239b5..e2e8412e687f 100644
> > > > --- a/kernel/rcu/tree.c
> > > > +++ b/kernel/rcu/tree.c
> > > > @@ -3238,6 +3238,9 @@ static void fill_page_cache_func(struct work_struct *work)
> > > >                         free_page((unsigned long) bnode);
> > > >                         break;
> > > >                 }
> > > > +
> > > > +               if (atomic_read(&krcp->backoff_page_cache_fill))
> > > > +                       break;
> > > >         }
> > > It does not fix an "issue" you are reporting. kvfree_rcu_bulk() 
> > > function can still fill it back. IMHO, the solution here is to 
> > > disable cache if a low memory condition and enable back later on.
> > > 
> > > The cache size is controlled by the rcu_min_cached_objs variable. We 
> > > can set it to 1 and restore it back to original value to make the 
> > > cache operating as before.
> > 
> > It would be best to use a second variable for this.  Users might get 
> > annoyed if their changes to rcu_min_cached_objs got overwritten once 
> > things got set back to normal operation.
> > 
> >Agree. So we do not make it visible over sysfs interface for user that we manipulate it.
> >
> >
> 
> 
> The rcu_min_cached_objs is read-only, Users cannot be set rcu_min_cached_objs dynamically. 
> 
> -r--r--r-- 1 root root 4.0K Apr 12 01:08 rcu_min_cached_objs
> 
You can set it as a boot parameter: rcutree.rcu_min_cached_objs=XXX

> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> index 41daae3239b5..0e9f83562823 100644
> --- a/kernel/rcu/tree.c
> +++ b/kernel/rcu/tree.c
> @@ -2909,7 +2909,8 @@ put_cached_bnode(struct kfree_rcu_cpu *krcp,
>         struct kvfree_rcu_bulk_data *bnode)
>  {
>         // Check the limit.
> -       if (krcp->nr_bkv_objs >= rcu_min_cached_objs)
> +       if ((atomic_read(&krcp->backoff_page_cache_fill) && krcp->nr_bkv_objs) ||
> +                               krcp->nr_bkv_objs >= rcu_min_cached_objs)
>                 return false;
> 
We can eliminate the backoff_page_cache_fill per-cpu atomic variable and
just change a new one, say, min_cached_objs, if a low memory condition.
Restore it to a default what is the rcu_min_cached_objs.

I can post here an example if it helps to make it more clear.

--
Uladzislau Rezki