On Fri, Dec 17, 2010 at 08:39:12AM +0000, Mel Gorman wrote:
quoted text
> On Fri, Dec 17, 2010 at 09:04:13AM +0900, Minchan Kim wrote:
> > On Thu, Dec 16, 2010 at 10:50 PM, Gerald Schaefer
> > <gerald.schaefer@de.ibm.com> wrote:
> > > I got the same warning now after increasing /proc/sys/vm/nr_hugepages, see
> > > below. Both cases are easily reproducible: memory unplug with big page cache,
> > > or adding large pages during run-time.
> > >
> > > ===================================================
> > > [ INFO: suspicious rcu_dereference_check() usage. ]
> > > ---------------------------------------------------
> > > include/linux/radix-tree.h:145 invoked rcu_dereference_check() without protection!
> > >
> > > other info that might help us debug this:
> > >
> > >
> > > rcu_scheduler_active = 1, debug_locks = 0
> > > 1 lock held by bash/761:
> > >  #0:  (&(&inode->i_data.tree_lock)->rlock){..-.-.}, at: [<00000000002263ae>] migrate_page_move_mapping+0x4a/0x2d8
> > >
> > > stack backtrace:
> > > CPU: 1 Not tainted 2.6.37-rc6 #4
> > > Process bash (pid: 761, task: 00000000181b5540, ksp: 00000000181bb7f8)
> > > 00000000181bb818 00000000181bb798 0000000000000002 0000000000000000
> > >       00000000181bb838 00000000181bb7b0 00000000181bb7b0 000000000056bafa
> > >       0000000000000000 000000003f42bdf0 0000000000000002 000000001c43be30
> > >       000003e00000000d 000003e00000000c 00000000181bb800 0000000000000000
> > >       0000000000000000 0000000000100bfa 00000000181bb798 00000000181bb7d8
> > > Call Trace:
> > > ([<0000000000100b02>] show_trace+0xee/0x144)
> > >  [<000000000022654e>] migrate_page_move_mapping+0x1ea/0x2d8
> > >  [<0000000000226c80>] migrate_page+0x38/0x68
> > >  [<0000000000226d9a>] move_to_new_page+0xea/0x2bc
> > >  [<000000000022785a>] migrate_pages+0x496/0x568
> > >  [<000000000021e24e>] compact_zone+0x432/0x7d8
> > >  [<000000000021e772>] compact_zone_order+0x9e/0xbc
> > >  [<000000000021ed52>] try_to_compact_pages+0x1ba/0x24c
> > >  [<00000000001e1afa>] __alloc_pages_nodemask+0x86a/0xa64
> > >  [<000000000021c80c>] alloc_fresh_huge_page.clone.2+0x68/0x18c
> > >  [<000000000021cc4c>] set_max_huge_pages.clone.0+0xa4/0x1ac
> > >  [<000000000021ce06>] hugetlb_sysctl_handler+0xb2/0xcc
> > >  [<00000000002a6572>] proc_sys_call_handler+0xe6/0x10c
> > >  [<00000000002a65be>] proc_sys_write+0x26/0x34
> > >  [<00000000002336e0>] vfs_write+0xac/0x18c
> > >  [<00000000002338bc>] SyS_write+0x58/0xa8
> > >  [<0000000000113976>] sysc_noemu+0x16/0x1c
> > >  [<0000020000162edc>] 0x20000162edc
> > > INFO: lockdep is turned off.
> > >
> > > I honestly do not understand 100% why this is a false positive, seeing that
> > > e.g. find_get_page() will also use radix_tree_deref_slot(), holding only the
> > > rcu_read_lock, while migrate_page_move_mapping() has no rcu_read_lock() but
> > > the &mapping->tree_lock instead. So I'm not quite sure how to fix this
> > > properly, but simply adding rcu_read_lock/unlock() to the affected code paths,
> > > even if it is not necessary for synchronization, would get rid of the warning,
> > > like in the following patch. Any ideas?
> > 
> > In case of anon page, we hold rcu_read_lock in unmap_and_move.
> > The problem is file-backed page. In case of that, we hold lock_page
> > and mapping->tree_lock as update-side lock.
> > So we don't need rcu_read_lock.
> > 
> > >
> > > ---
> > >  fs/hugetlbfs/inode.c |    2 ++
> > >  mm/migrate.c         |    4 ++++
> > >  2 files changed, 6 insertions(+)
> > >
> > > --- a/fs/hugetlbfs/inode.c
> > > +++ b/fs/hugetlbfs/inode.c
> > > @@ -580,7 +580,9 @@ static int hugetlbfs_migrate_page(struct
> > >  {
> > >        int rc;
> > >
> > > +       rcu_read_lock();
> > >        rc = migrate_huge_page_move_mapping(mapping, newpage, page);
> > > +       rcu_read_unlock();
> > >        if (rc)
> > >                return rc;
> > >        migrate_page_copy(newpage, page);
> > > --- a/mm/migrate.c
> > > +++ b/mm/migrate.c
> > > @@ -417,7 +417,9 @@ int migrate_page(struct address_space *m
> > >
> > >        BUG_ON(PageWriteback(page));    /* Writeback must be complete */
> > >
> > > +       rcu_read_lock();
> > >        rc = migrate_page_move_mapping(mapping, newpage, page);
> > > +       rcu_read_unlock();
> > >
> > >        if (rc)
> > >                return rc;
> > > @@ -444,7 +446,9 @@ int buffer_migrate_page(struct address_s
> > >
> > >        head = page_buffers(page);
> > >
> > > +       rcu_read_lock();
> > >        rc = migrate_page_move_mapping(mapping, newpage, page);
> > > +       rcu_read_unlock();
> > >
> > >        if (rc)
> > >                return rc;
> > >
> > >
> > >
> > 
> > 
> > How about this?
> > Maybe Paul have better idea.
> > (It's apparently be word-wrapped.)
> > 
> 
> heh, I wrote a patch almost identical to this and ran it overnight for testing
> (test was a memory consumer running while a parallel process grew and shrunk
> the hugepage pool). It passes but that is hardly a surprise. We differed

Good. 

quoted text
> slightly in a number of respects though.
> 
> > diff --git a/include/linux/radix-tree.h b/include/linux/radix-tree.h
> > index ab2baa5..135af1e 100644
> > --- a/include/linux/radix-tree.h
> > +++ b/include/linux/radix-tree.h
> > @@ -146,6 +146,20 @@ static inline void *radix_tree_deref_slot(void **pslot)
> >  }
> > 
> >  /**
> > + * radix_tree_deref_slot_nocheck       - dereference a slot without RCU check
> > + * @pslot:     pointer to slot, returned by radix_tree_lookup_slot
> > + * Returns:    item that was stored in that slot with any direct pointer flag
> > + *             removed.
> > + *
> > + * This functions works like radix_tree_deref_slot except it doesn't check
> > + * RCU rule. Normally this funcion is used with update-side lock.
> > + * You should use this function very carefully.
> > + */
> > +static inline void *radix_tree_deref_slot_nocheck(void **pslot)
> > +{
> > +       return rcu_dereference_protected(*pslot, 1);
> > +}
> 
> For this, I had
> 
> diff --git a/include/linux/radix-tree.h b/include/linux/radix-tree.h
> index ab2baa5..252d21c 100644
> --- a/include/linux/radix-tree.h
> +++ b/include/linux/radix-tree.h
> @@ -146,6 +146,25 @@ static inline void *radix_tree_deref_slot(void **pslot)
>  }
>  
>  /**
> + * radix_tree_deref_slot_protected	- dereference a slot without RCU lock but with tree lock held
> + * @pslot:	pointer to slot, returned by radix_tree_lookup_slot
> + * Returns:	item that was stored in that slot with any direct pointer flag
> + *		removed.
> + *
> + * For use with radix_tree_lookup_slot().  Caller must hold tree read
> + * locked across slot lookup and dereference. Not required if write lock is
> + * held (ie. items cannot be concurrently inserted).
> + *
> + * radix_tree_deref_retry must be used to confirm validity of the pointer if
> + * only the read lock is held.
> + */
> +static inline void *radix_tree_deref_slot_protected(void **pslot,
> +							spinlock_t *treelock)
> +{
> +	return rcu_dereference_protected(*pslot, lockdep_is_held(treelock));
> +}
> +
> +/**
>   * radix_tree_deref_retry	- check radix_tree_deref_slot
>   * @arg:	pointer returned by radix_tree_deref_slot
>   * Returns:	0 if retry is not required, otherwise retry is required
> 
> In the documentation, I noted that the check might be without RCU but with
> the knowledge that it's protected by the tree lock. I'm not a RCU expert
> but this is only safe when you know there isn't a parallel updater and the
> treelock should be preventing that, right?

I think so.

quoted text
> 
> Even so, other users of rcu_dereference_protected() check a lock condition
> which I used tree lock for. I intended to read through the rest of
> documentation properly this morning to determine if this was indeed the
> right approach.
> 
> I used the name _protected instead of _nocheck because the dereference
> is still protected (by the tree lock) just not by RCU. Again, have to
> check the documentation to ensure this is correct.

I agree _proected naming is good.

quoted text
> 
> > +/**
> >   * radix_tree_deref_retry      - check radix_tree_deref_slot
> >   * @arg:       pointer returned by radix_tree_deref_slot
> >   * Returns:    0 if retry is not required, otherwise retry is required
> > diff --git a/mm/migrate.c b/mm/migrate.c
> > index 2eb2243..5be2841 100644
> > --- a/mm/migrate.c
> > +++ b/mm/migrate.c
> > @@ -244,7 +244,8 @@ static int migrate_page_move_mapping(struct
> > address_space *mappin
> > 
> >         expected_count = 2 + page_has_private(page);
> >         if (page_count(page) != expected_count ||
> > -                       (struct page *)radix_tree_deref_slot(pslot) != page) {
> > +                       (struct page *)radix_tree_deref_slot_nocheck(pslot)
> > +                               != page) {
> >                 spin_unlock_irq(&mapping->tree_lock);
> >                 return -EAGAIN;
> >         }
> 
> We only differed here by my passing in the &mapping->tree_lock

I will add the comment in your formal patch.

quoted text
> 
> 
> 
> > @@ -316,7 +317,8 @@ int migrate_huge_page_move_mapping(struct
> > address_space *mapping,
> > 
> >         expected_count = 2 + page_has_private(page);
> >         if (page_count(page) != expected_count ||
> > -           (struct page *)radix_tree_deref_slot(pslot) != page) {
> > +           (struct page *)radix_tree_deref_slot_nocheck(pslot)
> > +                               != page) {
> >                 spin_unlock_irq(&mapping->tree_lock);
> >                 return -EAGAIN;
> >         }
> 
> -- 
> Mel Gorman
> Part-time Phd Student                          Linux Technology Center
> University of Limerick                         IBM Dublin Software Lab

-- 
Kind regards,
Minchan Kim
--
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/