[Ocfs2-devel] [PATCH] ocfs2: dlm: fix deadlock due to nested lock

[Junxiao Bi <junxiao.bi@oracle.com>]

Please drop this patch, I will send V2 later to avoid a possible
starvation issue.

Thanks,
Junxiao.
On 12/03/2015 04:50 PM, Junxiao Bi wrote:
> DLM allows nested cluster locking. One node X can acquire a cluster lock
> two times before release it. But between these two acquiring, if another
> node Y asks for the same lock and is blocked, then a bast will be sent to
> node X and OCFS2_LOCK_BLOCKED will be set in that lock's lockres. In this
> case, the second acquiring of that lock in node X will cause a deadlock.
> Not block for nested locking can fix this.
> 
> Use ocfs2-test multiple reflink test can reproduce this on v4.3 kernel,
> the whole cluster hung, and get the following call trace.
> 
>  INFO: task multi_reflink_t:10118 blocked for more than 120 seconds.
>        Tainted: G           OE   4.3.0 #1
>  "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
>  multi_reflink_t D ffff88003e816980     0 10118  10117 0x00000080
>   ffff880005b735f8 0000000000000082 ffffffff81a25500 ffff88003e750000
>   ffff880005b735c8 ffffffff8117992f ffffea0000929f80 ffff88003e816980
>   7fffffffffffffff 0000000000000000 0000000000000001 ffffea0000929f80
>  Call Trace:
>   [<ffffffff8117992f>] ? find_get_entry+0x2f/0xc0
>   [<ffffffff816a68fe>] schedule+0x3e/0x80
>   [<ffffffff816a9358>] schedule_timeout+0x1c8/0x220
>   [<ffffffffa067eee4>] ? ocfs2_inode_cache_unlock+0x14/0x20 [ocfs2]
>   [<ffffffffa06bb1e9>] ? ocfs2_metadata_cache_unlock+0x19/0x30 [ocfs2]
>   [<ffffffffa06bb399>] ? ocfs2_buffer_cached+0x99/0x170 [ocfs2]
>   [<ffffffffa067eee4>] ? ocfs2_inode_cache_unlock+0x14/0x20 [ocfs2]
>   [<ffffffffa06bb1e9>] ? ocfs2_metadata_cache_unlock+0x19/0x30 [ocfs2]
>   [<ffffffff810c5f41>] ? __raw_callee_save___pv_queued_spin_unlock+0x11/0x20
>   [<ffffffff816a78ae>] wait_for_completion+0xde/0x110
>   [<ffffffff810a81b0>] ? try_to_wake_up+0x240/0x240
>   [<ffffffffa066f65d>] __ocfs2_cluster_lock+0x20d/0x720 [ocfs2]
>   [<ffffffff810c5f41>] ? __raw_callee_save___pv_queued_spin_unlock+0x11/0x20
>   [<ffffffffa0674841>] ocfs2_inode_lock_full_nested+0x181/0x400 [ocfs2]
>   [<ffffffffa06d0db3>] ? ocfs2_iop_get_acl+0x53/0x113 [ocfs2]
>   [<ffffffff81210cd2>] ? igrab+0x42/0x70
>   [<ffffffffa06d0db3>] ocfs2_iop_get_acl+0x53/0x113 [ocfs2]
>   [<ffffffff81254583>] get_acl+0x53/0x70
>   [<ffffffff81254923>] posix_acl_create+0x73/0x130
>   [<ffffffffa068f0bf>] ocfs2_mknod+0x7cf/0x1140 [ocfs2]
>   [<ffffffffa068fba2>] ocfs2_create+0x62/0x110 [ocfs2]
>   [<ffffffff8120be25>] ? __d_alloc+0x65/0x190
>   [<ffffffff81201b3e>] ? __inode_permission+0x4e/0xd0
>   [<ffffffff81202cf5>] vfs_create+0xd5/0x100
>   [<ffffffff812009ed>] ? lookup_real+0x1d/0x60
>   [<ffffffff81203a03>] lookup_open+0x173/0x1a0
>   [<ffffffff810c59c6>] ? percpu_down_read+0x16/0x70
>   [<ffffffff81205fea>] do_last+0x31a/0x830
>   [<ffffffff81201b3e>] ? __inode_permission+0x4e/0xd0
>   [<ffffffff81201bd8>] ? inode_permission+0x18/0x50
>   [<ffffffff812046b0>] ? link_path_walk+0x290/0x550
>   [<ffffffff8120657c>] path_openat+0x7c/0x140
>   [<ffffffff812066c5>] do_filp_open+0x85/0xe0
>   [<ffffffff8120190f>] ? getname_flags+0x7f/0x1f0
>   [<ffffffff811f613a>] do_sys_open+0x11a/0x220
>   [<ffffffff8100374b>] ? syscall_trace_enter_phase1+0x15b/0x170
>   [<ffffffff811f627e>] SyS_open+0x1e/0x20
>   [<ffffffff816aa2ae>] entry_SYSCALL_64_fastpath+0x12/0x71
> 
> commit 743b5f1434f5 ("ocfs2: take inode lock in ocfs2_iop_set/get_acl()")
> add a nested locking to ocfs2_mknod() which exports this deadlock, but
> indeed this is a common issue, it can be triggered in other place.
> 
> Cc: <stable@vger.kernel.org>
> Signed-off-by: Junxiao Bi <junxiao.bi@oracle.com>
> ---
>  fs/ocfs2/dlmglue.c |    4 +++-
>  1 file changed, 3 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/ocfs2/dlmglue.c b/fs/ocfs2/dlmglue.c
> index 1c91103..5b7d9d4 100644
> --- a/fs/ocfs2/dlmglue.c
> +++ b/fs/ocfs2/dlmglue.c
> @@ -1295,7 +1295,9 @@ static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res *lock
>  {
>  	BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BLOCKED));
>  
> -	return wanted <= ocfs2_highest_compat_lock_level(lockres->l_blocking);
> +	/* allow nested lock request go to avoid deadlock. */
> +	return wanted <= ocfs2_highest_compat_lock_level(lockres->l_blocking)
> +		|| lockres->l_ro_holders || lockres->l_ex_holders;
>  }
>  
>  static void ocfs2_init_mask_waiter(struct ocfs2_mask_waiter *mw)
>