Re: [PATCH] target: Fix a double put in transport_free_session

On 3/25/21 2:48 AM, lyl2019@mail.ustc.edu.cn wrote:

quoted

-----原始邮件-----
发件人: michael.christie@oracle.com
发送时间: 2021-03-24 00:28:35 (星期三)
收件人: "Lv Yunlong" [off-list ref], martin.petersen@oracle.com
抄送: linux-scsi@vger.kernel.org, target-devel@vger.kernel.org, linux-kernel@vger.kernel.org
主题: Re: [PATCH] target: Fix a double put in transport_free_session

On 3/22/21 9:58 PM, Lv Yunlong wrote:

quoted

In transport_free_session, se_nacl is got from se_sess
with the initial reference. If se_nacl->acl_sess_list is
empty, se_nacl->dynamic_stop is set to true. Then the first
target_put_nacl(se_nacl) will drop the initial reference
and free se_nacl. Later there is a second target_put_nacl()
to put se_nacl. It may cause error in race.

quoted

My patch sets se_nacl->dynamic_stop to false to avoid the

double put.

Signed-off-by: Lv Yunlong <redacted>
---
 drivers/target/target_core_transport.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/drivers/target/target_core_transport.c b/drivers/target/target_core_transport.c
index 5ecb9f18a53d..c266defe694f 100644
--- a/drivers/target/target_core_transport.c
+++ b/drivers/target/target_core_transport.c

@@ -584,8 +584,10 @@ void transport_free_session(struct se_session *se_sess)
 		}
 		mutex_unlock(&se_tpg->acl_node_mutex);
 
-		if (se_nacl->dynamic_stop)
+		if (se_nacl->dynamic_stop) {
 			target_put_nacl(se_nacl);
+			se_nacl->dynamic_stop = false;
+		}
 
 		target_put_nacl(se_nacl);

Could you describe the race a little more?

Is the race:

1. thread1 called core_tpg_check_initiator_node_acl and found the acl.
sess->se_node_acl is set to the found acl.
2. thread2 is running transport_free_session. It now grabs the acl_node_mutex
and sees se_nacl->acl_sess_list is empty.
3. thread2 does the dynamic_stop=true operations in transport_free_session.
4. thread1 now calls transport_register_session now adds the sess to acl's
acl_sess_list.

Later when the session that thread 1 created is deleted dynamic_stop is still
set, so we do an extra target_put_nacl?

I'm not sure your patch will handle this race. When we delete the session thread1
created dynamic_node_acl is still set, so this:

                mutex_lock(&se_tpg->acl_node_mutex);
                if (se_nacl->dynamic_node_acl &&
                    !se_tfo->tpg_check_demo_mode_cache(se_tpg)) {
                        spin_lock_irqsave(&se_nacl->nacl_sess_lock, flags);
                        if (list_empty(&se_nacl->acl_sess_list))
                                se_nacl->dynamic_stop = true;

can set dynamic_stop to true again and we can end up doing the extra put still.

On top of the extra put we also do

list_del(&se_nacl->acl_list);

twice so we have to handle that as well.

Is there also another bug in this code. If someone adds an acl while there is a
dynamic acl in place core_tpg_add_initiator_node_acl will clear dynamic_node_acl
but we leave the extra reference, so later when transport_free_session is called
we will not do the extra put.

Ok, thanks for your answer. According the description above, i think it is a false
positive now.

Did you hit this bug, are you using an inspection tool, or did you find this by code
review?

I think there was a misunderstanding. I was saying it looks like a race could happen.
There is no protection in lio core.

I think it's hard to hit because most drivers do not allow the combo:

tpg_check_demo_mode == true
tpg_check_demo_mode_cache = false

It looks like those settings are allowed with tcm_qla2xxx and usb, but:

usb - has a mutex around creation and removal so we can't race.
tcm qla - I don't know this driver will enough, but I cc'd the maintainer.