2015-05-19 20:25:36

by Alexey Khoroshilov

Subject: [BUG] iscsi-target: deadlock because of iscsit_get_tpg()

Hello,

Our tool reports a potential double lock because of quite strange code
in iscsit_get_tpg().

drivers/target/iscsi/iscsi_target_tpg.c:
int iscsit_get_tpg(
struct iscsi_portal_group *tpg)
{
int ret;

ret = mutex_lock_interruptible(&tpg->tpg_access_lock);
return ((ret != 0) || signal_pending(current)) ? -1 : 0;
}

If mutex_lock_interruptible() successfully acquires the mutex, but there
is a pending signal, the function returns error, but it leaves the mutex
held. Callers do not expect such behaviour that can lead to a deadlock.

Why the check for pending signal is needed here?

Found by Linux Driver Verification project (linuxtesting.org).


Similar dangerous pattern presents in a couple of other places:

drivers/target/iscsi/iscsi_target.c:
int iscsit_access_np(struct iscsi_np *np, struct iscsi_portal_group *tpg)
{
...
ret = down_interruptible(&tpg->np_login_sem);
if ((ret != 0) || signal_pending(current))
return -1;


drivers/target/target_core_sbc.c:
static sense_reason_t
sbc_compare_and_write(struct se_cmd *cmd)
{
...
rc = down_interruptible(&dev->caw_sem);
if ((rc != 0) || signal_pending(current)) {
cmd->transport_complete_callback = NULL;
return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
}


--
Alexey Khoroshilov
Linux Verification Center, ISPRAS

2015-05-19 22:10:36

by Nicholas A. Bellinger

Subject: Re: [BUG] iscsi-target: deadlock because of iscsit_get_tpg()

Hi Alexey,

On Tue, 2015-05-19 at 23:25 +0300, Alexey Khoroshilov wrote:
> Hello,
>
> Our tool reports a potential double lock because of quite strange code
> in iscsit_get_tpg().
>
> drivers/target/iscsi/iscsi_target_tpg.c:
> int iscsit_get_tpg(
> struct iscsi_portal_group *tpg)
> {
> int ret;
>
> ret = mutex_lock_interruptible(&tpg->tpg_access_lock);
> return ((ret != 0) || signal_pending(current)) ? -1 : 0;
> }
>
> If mutex_lock_interruptible() successfully acquires the mutex, but there
> is a pending signal, the function returns error, but it leaves the mutex
> held. Callers do not expect such behaviour that can lead to a deadlock.
>
> Why the check for pending signal is needed here?
>

Once upon a time, this was an un-interruptible lock. The signal pending
was a check to allow userspace to break out of the operation with
SIGINT.

AFAICT, there's no reason why this is necessary anymore.

> Found by Linux Driver Verification project (linuxtesting.org).
>
>
> Similar dangerous pattern presents in a couple of other places:
>
> drivers/target/iscsi/iscsi_target.c:
> int iscsit_access_np(struct iscsi_np *np, struct iscsi_portal_group *tpg)
> {
> ...
> ret = down_interruptible(&tpg->np_login_sem);
> if ((ret != 0) || signal_pending(current))
> return -1;
>
>
> drivers/target/target_core_sbc.c:
> static sense_reason_t
> sbc_compare_and_write(struct se_cmd *cmd)
> {
> ...
> rc = down_interruptible(&dev->caw_sem);
> if ((rc != 0) || signal_pending(current)) {
> cmd->transport_complete_callback = NULL;
> return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
> }
>
>

Ditto on these.

Applying the following patch to target-pending/master.

Thank you,

--nab

diff --git a/drivers/target/iscsi/iscsi_target.c b/drivers/target/iscsi/iscsi_target.c
index 34871a6..74e6114 100644
--- a/drivers/target/iscsi/iscsi_target.c
+++ b/drivers/target/iscsi/iscsi_target.c
@@ -230,7 +230,7 @@ int iscsit_access_np(struct iscsi_np *np, struct iscsi_portal_group *tpg)
* Here we serialize access across the TIQN+TPG Tuple.
*/
ret = down_interruptible(&tpg->np_login_sem);
- if ((ret != 0) || signal_pending(current))
+ if (ret != 0)
return -1;

spin_lock_bh(&tpg->tpg_state_lock);
diff --git a/drivers/target/iscsi/iscsi_target_tpg.c b/drivers/target/iscsi/iscsi_target_tpg.
index e8a2408..5e3295f 100644
--- a/drivers/target/iscsi/iscsi_target_tpg.c
+++ b/drivers/target/iscsi/iscsi_target_tpg.c
@@ -161,10 +161,7 @@ struct iscsi_portal_group *iscsit_get_tpg_from_np(
int iscsit_get_tpg(
struct iscsi_portal_group *tpg)
{
- int ret;
-
- ret = mutex_lock_interruptible(&tpg->tpg_access_lock);
- return ((ret != 0) || signal_pending(current)) ? -1 : 0;
+ return mutex_lock_interruptible(&tpg->tpg_access_lock);
}

void iscsit_put_tpg(struct iscsi_portal_group *tpg)
diff --git a/drivers/target/target_core_sbc.c b/drivers/target/target_core_sbc.c
index 8855781..733824e 100644
--- a/drivers/target/target_core_sbc.c
+++ b/drivers/target/target_core_sbc.c
@@ -568,7 +568,7 @@ sbc_compare_and_write(struct se_cmd *cmd)
* comparision using SGLs at cmd->t_bidi_data_sg..
*/
rc = down_interruptible(&dev->caw_sem);
- if ((rc != 0) || signal_pending(current)) {
+ if (rc != 0) {
cmd->transport_complete_callback = NULL;
return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
}