Re: [PATCH] mm: backing-dev: use kfree_rcu() instead of synchronize_rcu_expedited()

[Matthew Wilcox <willy@infradead.org>]

On Fri, Oct 15, 2021 at 01:06:02PM +0800, Zqiang wrote:
> 
> On 2021/10/15 上午10:57, Qiang Zhang wrote:
> > 
> > 
> > Matthew Wilcox <willy@infradead.org <mailto:willy@infradead.org>>
> > 于2021年10月14日周四 下午7:26写道：
> > 
> >     On Thu, Oct 14, 2021 at 04:24:33PM +0800, Zqiang wrote:
> >     > The bdi_remove_from_list() is called in RCU softirq, however the
> >     > synchronize_rcu_expedited() will produce sleep action, use
> >     kfree_rcu()
> >     > instead of it.
> >     >
> >     > Reported-by: Hao Sun <sunhao.th@gmail.com
> >     <mailto:sunhao.th@gmail.com>>
> >     > Signed-off-by: Zqiang <qiang.zhang1211@gmail.com
> >     <mailto:qiang.zhang1211@gmail.com>>
> >     > ---
> >     >  include/linux/backing-dev-defs.h | 1 +
> >     >  mm/backing-dev.c                 | 4 +---
> >     >  2 files changed, 2 insertions(+), 3 deletions(-)
> >     >
> >     > diff --git a/include/linux/backing-dev-defs.h
> >     b/include/linux/backing-dev-defs.h
> >     > index 33207004cfde..35a093384518 100644
> >     > --- a/include/linux/backing-dev-defs.h
> >     > +++ b/include/linux/backing-dev-defs.h
> >     > @@ -202,6 +202,7 @@ struct backing_dev_info {
> >     >  #ifdef CONFIG_DEBUG_FS
> >     >       struct dentry *debug_dir;
> >     >  #endif
> >     > +     struct rcu_head rcu;
> >     >  };
> > 
> >     >Instead of growing struct backing_dev_info, it seems to me this
> >     rcu_head
> >     >could be placed in a union with rb_node, since it will have been
> >     removed
> >     >from the bdi_tree by this point and the tree is never walked under
> >     >RCU protection?
> > 
> > 
> > Thanks for your advice, I find this bdi_tree is traversed under the
> > protection of a spin lock, not under the protection of RCU.
> > I find this modification does not avoid the problem described in patch,
> > the flush_delayed_work() may be called in release_bdi()
> > The same will cause problems.
> > may be  we can replace queue_rcu_work() of call_rcu(&inode->i_rcu,
> > i_callback) or do you have any better suggestions?

What?  All I was suggesting was:

+++ b/include/linux/backing-dev-defs.h
@@ -168,7 +168,10 @@ struct bdi_writeback {
 
 struct backing_dev_info {
        u64 id;
-       struct rb_node rb_node; /* keyed by ->id */
+       union {
+               struct rb_node rb_node; /* keyed by ->id */
+               struct rcu_head rcu;
+       };
        struct list_head bdi_list;
        unsigned long ra_pages; /* max readahead in PAGE_SIZE units */
        unsigned long io_pages; /* max allowed IO size */


Christoph, independent of the inode lifetime problem, this actually seems
like a good approach to take.  I don't see why we should synchronize_rcu()
here?  Adding Jens (original introducer of the synchronize_rcu()), Mikulas
(converted it to use _expedited) and Tejun (worked around a problem when
using _expedited).