From b610ded71918654748b6b49f1e2636dc6bbfc96e Mon Sep 17 00:00:00 2001 From: Michal Hocko Date: Tue, 13 Aug 2013 16:00:55 -0700 Subject: hugetlb: fix lockdep splat caused by pmd sharing Dave has reported the following lockdep splat: ================================= [ INFO: inconsistent lock state ] 3.11.0-rc1+ #9 Not tainted --------------------------------- inconsistent {RECLAIM_FS-ON-W} -> {IN-RECLAIM_FS-W} usage. kswapd0/49 [HC0[0]:SC0[0]:HE1:SE1] takes: (&mapping->i_mmap_mutex){+.+.?.}, at: [] page_referenced+0x87/0x5e3 {RECLAIM_FS-ON-W} state was registered at: mark_held_locks+0x81/0xe7 lockdep_trace_alloc+0x5e/0xbc __alloc_pages_nodemask+0x8b/0x9b6 __get_free_pages+0x20/0x31 get_zeroed_page+0x12/0x14 __pmd_alloc+0x1c/0x6b huge_pmd_share+0x265/0x283 huge_pte_alloc+0x5d/0x71 hugetlb_fault+0x7c/0x64a handle_mm_fault+0x255/0x299 __do_page_fault+0x142/0x55c do_page_fault+0xd/0x16 error_code+0x6c/0x74 irq event stamp: 3136917 hardirqs last enabled at (3136917): _raw_spin_unlock_irq+0x27/0x50 hardirqs last disabled at (3136916): _raw_spin_lock_irq+0x15/0x78 softirqs last enabled at (3136180): __do_softirq+0x137/0x30f softirqs last disabled at (3136175): irq_exit+0xa8/0xaa other info that might help us debug this: Possible unsafe locking scenario: CPU0 ---- lock(&mapping->i_mmap_mutex); lock(&mapping->i_mmap_mutex); *** DEADLOCK *** no locks held by kswapd0/49. stack backtrace: CPU: 1 PID: 49 Comm: kswapd0 Not tainted 3.11.0-rc1+ #9 Hardware name: Dell Inc. Precision WorkStation 490 /0DT031, BIOS A08 04/25/2008 Call Trace: dump_stack+0x4b/0x79 print_usage_bug+0x1d9/0x1e3 mark_lock+0x1e0/0x261 __lock_acquire+0x623/0x17f2 lock_acquire+0x7d/0x195 mutex_lock_nested+0x6c/0x3a7 page_referenced+0x87/0x5e3 shrink_page_list+0x3d9/0x947 shrink_inactive_list+0x155/0x4cb shrink_lruvec+0x300/0x5ce shrink_zone+0x53/0x14e kswapd+0x517/0xa75 kthread+0xa8/0xaa ret_from_kernel_thread+0x1b/0x28 which is a false positive caused by hugetlb pmd sharing code which allocates a new pmd from withing mapping->i_mmap_mutex. If this allocation causes reclaim then the lockdep detector complains that we might self-deadlock. This is not correct though, because hugetlb pages are not reclaimable so their mapping will be never touched from the reclaim path. The patch tells lockup detector that hugetlb i_mmap_mutex is special by assigning it a separate lockdep class so it won't report possible deadlocks on unrelated mappings. [peterz@infradead.org: comment for annotation] Reported-by: Dave Jones Signed-off-by: Michal Hocko Cc: Peter Zijlstra Reviewed-by: Minchan Kim Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds --- fs/hugetlbfs/inode.c | 10 ++++++++++ 1 file changed, 10 insertions(+) diff --git a/fs/hugetlbfs/inode.c b/fs/hugetlbfs/inode.c index a3f868ae3fd4..34423978b170 100644 --- a/fs/hugetlbfs/inode.c +++ b/fs/hugetlbfs/inode.c @@ -463,6 +463,14 @@ static struct inode *hugetlbfs_get_root(struct super_block *sb, return inode; } +/* + * Hugetlbfs is not reclaimable; therefore its i_mmap_mutex will never + * be taken from reclaim -- unlike regular filesystems. This needs an + * annotation because huge_pmd_share() does an allocation under + * i_mmap_mutex. + */ +struct lock_class_key hugetlbfs_i_mmap_mutex_key; + static struct inode *hugetlbfs_get_inode(struct super_block *sb, struct inode *dir, umode_t mode, dev_t dev) @@ -474,6 +482,8 @@ static struct inode *hugetlbfs_get_inode(struct super_block *sb, struct hugetlbfs_inode_info *info; inode->i_ino = get_next_ino(); inode_init_owner(inode, dir, mode); + lockdep_set_class(&inode->i_mapping->i_mmap_mutex, + &hugetlbfs_i_mmap_mutex_key); inode->i_mapping->a_ops = &hugetlbfs_aops; inode->i_mapping->backing_dev_info =&hugetlbfs_backing_dev_info; inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME; -- cgit v1.2.3