From: Minchan Kim Subject: Re: [BUG?] memory hotplug: include/linux/radix-tree.h:145 invoked rcu_dereference_check() without protection! Date: Sat, 18 Dec 2010 00:22:42 +0900 Message-ID: <20101217152242.GC1609@barrios-desktop> References: <20101121173726.GG23423@thunk.org> <20101122061619.GA2764@linux.vnet.ibm.com> <1291748509.19276.62.camel@thinkpad> <20101208101947.b0646226.kamezawa.hiroyu@jp.fujitsu.com> <1292507421.4885.19.camel@thinkpad> <20101217083912.GM13914@csn.ul.ie> <20101217092828.GN13914@csn.ul.ie> Mime-Version: 1.0 Content-Type: text/plain; charset=iso-8859-1 Content-Transfer-Encoding: QUOTED-PRINTABLE Cc: gerald.schaefer@de.ibm.com, KAMEZAWA Hiroyuki , paulmck@linux.vnet.ibm.com, Milton Miller , linux-kernel@vger.kernel.org, linux-mm@vger.kernel.org, linux-ext4@vger.kernel.org, Ted Ts'o , Arun Bhanu , Andrew Morton , Heiko Carstens , Martin Schwidefsky To: Mel Gorman Return-path: Content-Disposition: inline In-Reply-To: <20101217092828.GN13914@csn.ul.ie> Sender: linux-kernel-owner@vger.kernel.org List-Id: linux-ext4.vger.kernel.org On Fri, Dec 17, 2010 at 09:28:28AM +0000, Mel Gorman wrote: > On Fri, Dec 17, 2010 at 08:39:12AM +0000, Mel Gorman wrote: > > On Fri, Dec 17, 2010 at 09:04:13AM +0900, Minchan Kim wrote: > > > On Thu, Dec 16, 2010 at 10:50 PM, Gerald Schaefer > > > wrote: > > > > I got the same warning now after increasing /proc/sys/vm/nr_hug= epages, see > > > > below. Both cases are easily reproducible: memory unplug with b= ig page cache, > > > > or adding large pages during run-time. > > > > > > > > =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D > > > > [ 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 =3D 1, debug_locks =3D 0 > > > > 1 lock held by bash/761: > > > > =A0#0: =A0(&(&inode->i_data.tree_lock)->rlock){..-.-.}, at: [<0= 0000000002263ae>] migrate_page_move_mapping+0x4a/0x2d8 > > > > > > > > stack backtrace: > > > > CPU: 1 Not tainted 2.6.37-rc6 #4 > > > > Process bash (pid: 761, task: 00000000181b5540, ksp: 0000000018= 1bb7f8) > > > > 00000000181bb818 00000000181bb798 0000000000000002 000000000000= 0000 > > > > =A0 =A0 =A0 00000000181bb838 00000000181bb7b0 00000000181bb7b0 = 000000000056bafa > > > > =A0 =A0 =A0 0000000000000000 000000003f42bdf0 0000000000000002 = 000000001c43be30 > > > > =A0 =A0 =A0 000003e00000000d 000003e00000000c 00000000181bb800 = 0000000000000000 > > > > =A0 =A0 =A0 0000000000000000 0000000000100bfa 00000000181bb798 = 00000000181bb7d8 > > > > Call Trace: > > > > ([<0000000000100b02>] show_trace+0xee/0x144) > > > > =A0[<000000000022654e>] migrate_page_move_mapping+0x1ea/0x2d8 > > > > =A0[<0000000000226c80>] migrate_page+0x38/0x68 > > > > =A0[<0000000000226d9a>] move_to_new_page+0xea/0x2bc > > > > =A0[<000000000022785a>] migrate_pages+0x496/0x568 > > > > =A0[<000000000021e24e>] compact_zone+0x432/0x7d8 > > > > =A0[<000000000021e772>] compact_zone_order+0x9e/0xbc > > > > =A0[<000000000021ed52>] try_to_compact_pages+0x1ba/0x24c > > > > =A0[<00000000001e1afa>] __alloc_pages_nodemask+0x86a/0xa64 > > > > =A0[<000000000021c80c>] alloc_fresh_huge_page.clone.2+0x68/0x18= c > > > > =A0[<000000000021cc4c>] set_max_huge_pages.clone.0+0xa4/0x1ac > > > > =A0[<000000000021ce06>] hugetlb_sysctl_handler+0xb2/0xcc > > > > =A0[<00000000002a6572>] proc_sys_call_handler+0xe6/0x10c > > > > =A0[<00000000002a65be>] proc_sys_write+0x26/0x34 > > > > =A0[<00000000002336e0>] vfs_write+0xac/0x18c > > > > =A0[<00000000002338bc>] SyS_write+0x58/0xa8 > > > > =A0[<0000000000113976>] sysc_noemu+0x16/0x1c > > > > =A0[<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(), hol= ding only the > > > > rcu_read_lock, while migrate_page_move_mapping() has no rcu_rea= d_lock() but > > > > the &mapping->tree_lock instead. So I'm not quite sure how to f= ix this > > > > properly, but simply adding rcu_read_lock/unlock() to the affec= ted code paths, > > > > even if it is not necessary for synchronization, would get rid = of the warning, > > > > like in the following patch. Any ideas? > > >=20 > > > 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_pa= ge > > > and mapping->tree_lock as update-side lock. > > > So we don't need rcu_read_lock. > > >=20 > > > > > > > > --- > > > > =A0fs/hugetlbfs/inode.c | =A0 =A02 ++ > > > > =A0mm/migrate.c =A0 =A0 =A0 =A0 | =A0 =A04 ++++ > > > > =A02 files changed, 6 insertions(+) > > > > > > > > --- a/fs/hugetlbfs/inode.c > > > > +++ b/fs/hugetlbfs/inode.c > > > > @@ -580,7 +580,9 @@ static int hugetlbfs_migrate_page(struct > > > > =A0{ > > > > =A0 =A0 =A0 =A0int rc; > > > > > > > > + =A0 =A0 =A0 rcu_read_lock(); > > > > =A0 =A0 =A0 =A0rc =3D migrate_huge_page_move_mapping(mapping, n= ewpage, page); > > > > + =A0 =A0 =A0 rcu_read_unlock(); > > > > =A0 =A0 =A0 =A0if (rc) > > > > =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0return rc; > > > > =A0 =A0 =A0 =A0migrate_page_copy(newpage, page); > > > > --- a/mm/migrate.c > > > > +++ b/mm/migrate.c > > > > @@ -417,7 +417,9 @@ int migrate_page(struct address_space *m > > > > > > > > =A0 =A0 =A0 =A0BUG_ON(PageWriteback(page)); =A0 =A0/* Writeback= must be complete */ > > > > > > > > + =A0 =A0 =A0 rcu_read_lock(); > > > > =A0 =A0 =A0 =A0rc =3D migrate_page_move_mapping(mapping, newpag= e, page); > > > > + =A0 =A0 =A0 rcu_read_unlock(); > > > > > > > > =A0 =A0 =A0 =A0if (rc) > > > > =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0return rc; > > > > @@ -444,7 +446,9 @@ int buffer_migrate_page(struct address_s > > > > > > > > =A0 =A0 =A0 =A0head =3D page_buffers(page); > > > > > > > > + =A0 =A0 =A0 rcu_read_lock(); > > > > =A0 =A0 =A0 =A0rc =3D migrate_page_move_mapping(mapping, newpag= e, page); > > > > + =A0 =A0 =A0 rcu_read_unlock(); > > > > > > > > =A0 =A0 =A0 =A0if (rc) > > > > =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0return rc; > > > > > > > > > > > > > > >=20 > > >=20 > > > How about this? > > > Maybe Paul have better idea. > > > (It's apparently be word-wrapped.) > > >=20 > >=20 > > 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 a= nd shrunk > > the hugepage pool). It passes but that is hardly a surprise. We dif= fered > > slightly in a number of respects though. > >=20 >=20 > For completeness, this is what I tested last night. There are two "co= nfirms" > in the changelog that I intended to work out today but maybe someone = can > confirm faster. >=20 > =3D=3D=3D=3D CUT HERE =3D=3D=3D=3D > mm: migration: Use rcu_dereference_protected when dereferencing the r= adix tree slot during file page migration >=20 > migrate_pages() -> unmap_and_move() only calls rcu_read_lock() for an= onymous > pages, as introduced by git commit 989f89c57e6361e7d16fbd9572b5da7d31= 3b073d. > The point of the RCU protection there is part of getting a stable ref= erence > to anon_vma and is only held for anon pages as file pages are locked > which is sufficient protection against freeing. >=20 > However, while a file page's mapping is being migrated, the radix tre= e > is double checked to ensure it is the expected page. This uses > radix_tree_deref_slot() -> rcu_dereference() without the RCU lock hel= d > triggering the following warning. >=20 > [ 173.674290] =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D=3D=3D=3D > [ 173.676016] [ INFO: suspicious rcu_dereference_check() usage. ] > [ 173.676016] --------------------------------------------------- > [ 173.676016] include/linux/radix-tree.h:145 invoked rcu_dereference= _check() without protection! > [ 173.676016] > [ 173.676016] other info that might help us debug this: > [ 173.676016] > [ 173.676016] > [ 173.676016] rcu_scheduler_active =3D 1, debug_locks =3D 0 > [ 173.676016] 1 lock held by hugeadm/2899: > [ 173.676016] #0: (&(&inode->i_data.tree_lock)->rlock){..-.-.}, at= : [] migrate_page_move_mapping+0x40/0x1ab > [ 173.676016] > [ 173.676016] stack backtrace: > [ 173.676016] Pid: 2899, comm: hugeadm Not tainted 2.6.37-rc5-autobu= ild > [ 173.676016] Call Trace: > [ 173.676016] [] ? printk+0x14/0x1b > [ 173.676016] [] lockdep_rcu_dereference+0x7d/0x86 > [ 173.676016] [] migrate_page_move_mapping+0xca/0x1ab > [ 173.676016] [] migrate_page+0x23/0x39 > [ 173.676016] [] buffer_migrate_page+0x22/0x107 > [ 173.676016] [] ? buffer_migrate_page+0x0/0x107 > [ 173.676016] [] move_to_new_page+0x9a/0x1ae > [ 173.676016] [] migrate_pages+0x1e7/0x2fa >=20 > This patch introduces radix_tree_deref_slot_protected() which calls > rcu_dereference_protected(). Users of it must pass in the mapping->tr= ee_lock > that is protecting this dereference. Based on the locking hierarchy d= escribed > in mm/filemap.c, holding the tree lock is protecting the radix tree f= rom > concurrent updaters in all cases (Confirm that no case has been misse= d). > According to Documentation/RCU/lockdep.txt, if there is a guarantee t= hat > no parallel updaters exist, use of rcu_dereference_protected() is all= owed > (Confirm this is accurate?). >=20 > Signed-off-by: Mel Gorman > --- > include/linux/radix-tree.h | 19 +++++++++++++++++++ > mm/migrate.c | 4 ++-- > 2 files changed, 21 insertions(+), 2 deletions(-) >=20 > 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) > } > =20 > /** > + * 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 pointe= r flag > + * removed. > + * > + * For use with radix_tree_lookup_slot(). Caller must hold tree rea= d > + * 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 po= inter 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))= ; > +} It seems to be good than mine. Just a nitpick. Can't we get the mutex lock as update-side lock in future? --=20 Kind regards, Minchan Kim