wangjian
2018-Dec-03 12:06 UTC
[Ocfs2-devel] [PATCH] ocfs2/dlm: Clean DLM_LKSB_GET_LVB and DLM_LKSB_PUT_LVB when the cancel_pending is set
Function dlm_move_lockres_to_recovery_list should clean
DLM_LKSB_GET_LVB and DLM_LKSB_PUT_LVB when the cancel_pending
is set. Otherwise node may panic in dlm_proxy_ast_handler.
Here is the situation: At the beginning, Node1 is the master
of the lock resource and has NL lock, Node2 has PR lock,
Node3 has PR lock, Node4 has NL lock.
Node1 Node2 Node3 Node4
convert lock_2 from
PR to EX.
the mode of lock_3 is
PR, which blocks the
conversion request of
Node2. move lock_2 to
conversion list.
convert lock_3 from
PR to EX.
move lock_3 to conversion
list. send BAST to Node3.
receive BAST from Node1.
downconvert thread execute
canceling convert operation.
Node2 dies because
the host is powered down.
in dlmunlock_common function,
the downconvert thread set
cancel_pending. at the same
time, Node 3 realized that
Node 1 is dead, so move lock_3
back to granted list in
dlm_move_lockres_to_recovery_list
function and remove Node 1 from
the domain_map in
__dlm_hb_node_down function.
then downconvert thread failed
to send the lock cancellation
request to Node1 and return
DLM_NORMAL from
dlm_send_remote_unlock_request
function.
become recovery master.
during the recovery
process, send
lock_2 that is
converting form
PR to EX to Node4.
during the recovery process,
send lock_3 in the granted list and
cantain the DLM_LKSB_GET_LVB
flag to Node4. Then downconvert thread
delete DLM_LKSB_GET_LVB flag in
dlmunlock_common function.
Node4 finish recovery.
the mode of lock_3 is
PR, which blocks the
conversion request of
Node2, so send BAST
to Node3.
receive BAST from Node4.
convert lock_3 from PR to NL.
change the mode of lock_3
from PR to NL and send
message to Node3.
receive message from
Node4. The message contain
LKM_GET_LVB flag, but the
lock->lksb->flags does not
contain DLM_LKSB_GET_LVB,
BUG_ON in dlm_proxy_ast_handler
function.
Signed-off-by: Jian Wang <wangjian161 at huawei.com>
Reviewed-by: Yiwen Jiang <jiangyiwen at huawei.com>
---
fs/ocfs2/dlm/dlmunlock.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/fs/ocfs2/dlm/dlmunlock.c b/fs/ocfs2/dlm/dlmunlock.c
index 63d701c..6e04fc7 100644
--- a/fs/ocfs2/dlm/dlmunlock.c
+++ b/fs/ocfs2/dlm/dlmunlock.c
@@ -277,6 +277,7 @@ void dlm_commit_pending_cancel(struct dlm_lock_resource
*res,
{
list_move_tail(&lock->list, &res->granted);
lock->ml.convert_type = LKM_IVMODE;
+ lock->lksb->flags &= ~(DLM_LKSB_GET_LVB|DLM_LKSB_PUT_LVB);
}
--
1.8.3.1
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
http://oss.oracle.com/pipermail/ocfs2-devel/attachments/20181203/8f927337/attachment.html
Changwei Ge
2018-Dec-05 02:01 UTC
[Ocfs2-devel] [PATCH] ocfs2/dlm: Clean DLM_LKSB_GET_LVB and DLM_LKSB_PUT_LVB when the cancel_pending is set
Hi Jian, Could your please also share the panic backtrace in dlm_proxy_ast_handler()? After that I can help to review this patch and analyze what's wrong in DLM. It will be better for you to tell your intention why LVB flags should be cleared rather than just giving a longggg time sequence diagram. Moreover, your patches are hard be applied to my tree :( Thanks, Changwei On 2018/12/3 20:08, wangjian wrote:> Function dlm_move_lockres_to_recovery_list should clean > DLM_LKSB_GET_LVB and DLM_LKSB_PUT_LVB when the cancel_pending > is set. Otherwise node may panic in dlm_proxy_ast_handler. > > Here is the situation: At the beginning, Node1 is the master > of the lock resource and has NL lock, Node2 has PR lock, > Node3 has PR lock, Node4 has NL lock. > > Node1 Node2 Node3 Node4 > convert lock_2 from > PR to EX. > > the mode of lock_3 is > PR, which blocks the > conversion request of > Node2. move lock_2 to > conversion list. > > convert lock_3 from > PR to EX. > > move lock_3 to conversion > list. send BAST to Node3. > > receive BAST from Node1. > downconvert thread execute > canceling convert operation. > > Node2 dies because > the host is powered down. > > in dlmunlock_common function, > the downconvert thread set > cancel_pending. at the same > time, Node 3 realized that > Node 1 is dead, so move lock_3 > back to granted list in > dlm_move_lockres_to_recovery_list > function and remove Node 1 from > the domain_map in > __dlm_hb_node_down function. > then downconvert thread failed > to send the lock cancellation > request to Node1 and return > DLM_NORMAL from > dlm_send_remote_unlock_request > function. > > become recovery master. > > during the recovery > process, send > lock_2 that is > converting form > PR to EX to Node4. > > during the recovery process, > send lock_3 in the granted list and > cantain the DLM_LKSB_GET_LVB > flag to Node4. Then downconvert thread > delete DLM_LKSB_GET_LVB flag in > dlmunlock_common function. > > Node4 finish recovery. > the mode of lock_3 is > PR, which blocks the > conversion request of > Node2, so send BAST > to Node3. > > receive BAST from Node4. > convert lock_3 from PR to NL. > > change the mode of lock_3 > from PR to NL and send > message to Node3. > > receive message from > Node4. The message contain > LKM_GET_LVB flag, but the > lock->lksb->flags does not > contain DLM_LKSB_GET_LVB, > BUG_ON in dlm_proxy_ast_handler > function. > > Signed-off-by: Jian Wang<wangjian161 at huawei.com> > Reviewed-by: Yiwen Jiang<jiangyiwen at huawei.com> > --- > fs/ocfs2/dlm/dlmunlock.c | 1 + > 1 file changed, 1 insertion(+) > > diff --git a/fs/ocfs2/dlm/dlmunlock.c b/fs/ocfs2/dlm/dlmunlock.c > index 63d701c..6e04fc7 100644 > --- a/fs/ocfs2/dlm/dlmunlock.c > +++ b/fs/ocfs2/dlm/dlmunlock.c > @@ -277,6 +277,7 @@ void dlm_commit_pending_cancel(struct dlm_lock_resource *res, > { > list_move_tail(&lock->list, &res->granted); > lock->ml.convert_type = LKM_IVMODE; > + lock->lksb->flags &= ~(DLM_LKSB_GET_LVB|DLM_LKSB_PUT_LVB); > } > > > -- > 1.8.3.1 >