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

From: Zhang, Qiang1
Date: Tue Apr 11 2023 - 11:10:02 EST


> > > 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@xxxxxxxxx>
> > >
> > >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.

A good suggestion. a question, when need to restore rcu_min_cached_objs?
after the execution of kfree_rcu_shrink_scan() ends?

Thanks
Zqiang

>
>--
>Uladzislau Rezki