2017-06-01 06:21:37

by Nicholas A. Bellinger

Subject: Re: [PATCH] iscsi: Fix a sleep-in-atomic bug

Hi Jia-Ju,

On Wed, 2017-05-31 at 11:26 +0800, Jia-Ju Bai wrote:
> The driver may sleep under a spin lock, and the function call path is:
> iscsit_tpg_enable_portal_group (acquire the lock by spin_lock)
> iscsi_update_param_value
> kstrdup(GFP_KERNEL) --> may sleep
>
> To fix it, the "GFP_KERNEL" is replaced with "GFP_ATOMIC".
>
> Signed-off-by: Jia-Ju Bai <[email protected]>
> ---
> drivers/target/iscsi/iscsi_target_parameters.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)

Btw, the use of tpg->tpg_state_lock in iscsit_tpg_enable_portal_group()
while checking existing state and calling iscsi_update_param_value() is
not necessary, since lio_target_tpg_enable_store() is already holding
iscsit_get_tpg() -> tpg->tpg_access_lock.

How about the following instead to only take tpg->tpg_state_lock when
updating tpg->tpg_state instead..?

diff --git a/drivers/target/iscsi/iscsi_target_tpg.c b/drivers/target/iscsi/iscsi_target_tpg.c
index 2e7e08d..abaabba 100644
--- a/drivers/target/iscsi/iscsi_target_tpg.c
+++ b/drivers/target/iscsi/iscsi_target_tpg.c
@@ -311,11 +311,9 @@ int iscsit_tpg_enable_portal_group(struct iscsi_portal_group *tpg)
struct iscsi_tiqn *tiqn = tpg->tpg_tiqn;
int ret;

- spin_lock(&tpg->tpg_state_lock);
if (tpg->tpg_state == TPG_STATE_ACTIVE) {
pr_err("iSCSI target portal group: %hu is already"
" active, ignoring request.\n", tpg->tpgt);
- spin_unlock(&tpg->tpg_state_lock);
return -EINVAL;
}
/*
@@ -324,10 +322,8 @@ int iscsit_tpg_enable_portal_group(struct iscsi_portal_group *tpg)
* is enforced (as per default), and remove the NONE option.
*/
param = iscsi_find_param_from_key(AUTHMETHOD, tpg->param_list);
- if (!param) {
- spin_unlock(&tpg->tpg_state_lock);
+ if (!param)
return -EINVAL;
- }

if (tpg->tpg_attrib.authentication) {
if (!strcmp(param->value, NONE)) {
@@ -341,6 +337,7 @@ int iscsit_tpg_enable_portal_group(struct iscsi_portal_group *tpg)
goto err;
}

+ spin_lock(&tpg->tpg_state_lock);
tpg->tpg_state = TPG_STATE_ACTIVE;
spin_unlock(&tpg->tpg_state_lock);

@@ -353,7 +350,6 @@ int iscsit_tpg_enable_portal_group(struct iscsi_portal_group *tpg)
return 0;

err:
- spin_unlock(&tpg->tpg_state_lock);
return ret;
}

2017-06-02 01:14:30

by Jia-Ju Bai

Subject: Re: [PATCH] iscsi: Fix a sleep-in-atomic bug

On 06/01/2017 02:21 PM, Nicholas A. Bellinger wrote:
> Hi Jia-Ju,
>
> On Wed, 2017-05-31 at 11:26 +0800, Jia-Ju Bai wrote:
>> The driver may sleep under a spin lock, and the function call path is:
>> iscsit_tpg_enable_portal_group (acquire the lock by spin_lock)
>> iscsi_update_param_value
>> kstrdup(GFP_KERNEL) --> may sleep
>>
>> To fix it, the "GFP_KERNEL" is replaced with "GFP_ATOMIC".
>>
>> Signed-off-by: Jia-Ju Bai<[email protected]>
>> ---
>> drivers/target/iscsi/iscsi_target_parameters.c | 2 +-
>> 1 file changed, 1 insertion(+), 1 deletion(-)
> Btw, the use of tpg->tpg_state_lock in iscsit_tpg_enable_portal_group()
> while checking existing state and calling iscsi_update_param_value() is
> not necessary, since lio_target_tpg_enable_store() is already holding
> iscsit_get_tpg() -> tpg->tpg_access_lock.
>
> How about the following instead to only take tpg->tpg_state_lock when
> updating tpg->tpg_state instead..?
>
> diff --git a/drivers/target/iscsi/iscsi_target_tpg.c b/drivers/target/iscsi/iscsi_target_tpg.c
> index 2e7e08d..abaabba 100644
> --- a/drivers/target/iscsi/iscsi_target_tpg.c
> +++ b/drivers/target/iscsi/iscsi_target_tpg.c
> @@ -311,11 +311,9 @@ int iscsit_tpg_enable_portal_group(struct iscsi_portal_group *tpg)
> struct iscsi_tiqn *tiqn = tpg->tpg_tiqn;
> int ret;
>
> - spin_lock(&tpg->tpg_state_lock);
> if (tpg->tpg_state == TPG_STATE_ACTIVE) {
> pr_err("iSCSI target portal group: %hu is already"
> " active, ignoring request.\n", tpg->tpgt);
> - spin_unlock(&tpg->tpg_state_lock);
> return -EINVAL;
> }
> /*
> @@ -324,10 +322,8 @@ int iscsit_tpg_enable_portal_group(struct iscsi_portal_group *tpg)
> * is enforced (as per default), and remove the NONE option.
> */
> param = iscsi_find_param_from_key(AUTHMETHOD, tpg->param_list);
> - if (!param) {
> - spin_unlock(&tpg->tpg_state_lock);
> + if (!param)
> return -EINVAL;
> - }
>
> if (tpg->tpg_attrib.authentication) {
> if (!strcmp(param->value, NONE)) {
> @@ -341,6 +337,7 @@ int iscsit_tpg_enable_portal_group(struct iscsi_portal_group *tpg)
> goto err;
> }
>
> + spin_lock(&tpg->tpg_state_lock);
> tpg->tpg_state = TPG_STATE_ACTIVE;
> spin_unlock(&tpg->tpg_state_lock);
>
> @@ -353,7 +350,6 @@ int iscsit_tpg_enable_portal_group(struct iscsi_portal_group *tpg)
> return 0;
>
> err:
> - spin_unlock(&tpg->tpg_state_lock);
> return ret;
> }
>
I think it is fine to me.

Thanks,
Jia-Ju Bai

2017-06-02 03:20:07

by Nicholas A. Bellinger

Subject: Re: [PATCH] iscsi: Fix a sleep-in-atomic bug

On Fri, 2017-06-02 at 09:13 +0800, Jia-Ju Bai wrote:
> On 06/01/2017 02:21 PM, Nicholas A. Bellinger wrote:
> > Hi Jia-Ju,
> >
> > On Wed, 2017-05-31 at 11:26 +0800, Jia-Ju Bai wrote:
> >> The driver may sleep under a spin lock, and the function call path is:
> >> iscsit_tpg_enable_portal_group (acquire the lock by spin_lock)
> >> iscsi_update_param_value
> >> kstrdup(GFP_KERNEL) --> may sleep
> >>
> >> To fix it, the "GFP_KERNEL" is replaced with "GFP_ATOMIC".
> >>
> >> Signed-off-by: Jia-Ju Bai<[email protected]>
> >> ---
> >> drivers/target/iscsi/iscsi_target_parameters.c | 2 +-
> >> 1 file changed, 1 insertion(+), 1 deletion(-)
> > Btw, the use of tpg->tpg_state_lock in iscsit_tpg_enable_portal_group()
> > while checking existing state and calling iscsi_update_param_value() is
> > not necessary, since lio_target_tpg_enable_store() is already holding
> > iscsit_get_tpg() -> tpg->tpg_access_lock.
> >
> > How about the following instead to only take tpg->tpg_state_lock when
> > updating tpg->tpg_state instead..?
> >
> > diff --git a/drivers/target/iscsi/iscsi_target_tpg.c b/drivers/target/iscsi/iscsi_target_tpg.c
> > index 2e7e08d..abaabba 100644
> > --- a/drivers/target/iscsi/iscsi_target_tpg.c
> > +++ b/drivers/target/iscsi/iscsi_target_tpg.c
> > @@ -311,11 +311,9 @@ int iscsit_tpg_enable_portal_group(struct iscsi_portal_group *tpg)
> > struct iscsi_tiqn *tiqn = tpg->tpg_tiqn;
> > int ret;
> >
> > - spin_lock(&tpg->tpg_state_lock);
> > if (tpg->tpg_state == TPG_STATE_ACTIVE) {
> > pr_err("iSCSI target portal group: %hu is already"
> > " active, ignoring request.\n", tpg->tpgt);
> > - spin_unlock(&tpg->tpg_state_lock);
> > return -EINVAL;
> > }
> > /*
> > @@ -324,10 +322,8 @@ int iscsit_tpg_enable_portal_group(struct iscsi_portal_group *tpg)
> > * is enforced (as per default), and remove the NONE option.
> > */
> > param = iscsi_find_param_from_key(AUTHMETHOD, tpg->param_list);
> > - if (!param) {
> > - spin_unlock(&tpg->tpg_state_lock);
> > + if (!param)
> > return -EINVAL;
> > - }
> >
> > if (tpg->tpg_attrib.authentication) {
> > if (!strcmp(param->value, NONE)) {
> > @@ -341,6 +337,7 @@ int iscsit_tpg_enable_portal_group(struct iscsi_portal_group *tpg)
> > goto err;
> > }
> >
> > + spin_lock(&tpg->tpg_state_lock);
> > tpg->tpg_state = TPG_STATE_ACTIVE;
> > spin_unlock(&tpg->tpg_state_lock);
> >
> > @@ -353,7 +350,6 @@ int iscsit_tpg_enable_portal_group(struct iscsi_portal_group *tpg)
> > return 0;
> >
> > err:
> > - spin_unlock(&tpg->tpg_state_lock);
> > return ret;
> > }
> >
> I think it is fine to me.
>
> Thanks,
> Jia-Ju Bai

Applied with your Reported-by and Reviewed-by.

Thanks!